Sequential Rationality and Subgame Perfect Equilibrium

Subgame Perfect Equilibrium
Strategic form of the game

O 1 L R
1, 3
O 1, 3 1, 3
Microeconomics II T 2, 1 0, 0
T B B 0, 2 0, 1
Extensive Form Games III
2 Set of Nash equilibria
Levent Koçkesen N(Γ) = {(T, L), (O, R)}
L R L R
Koç University What is the set of SPE?
2, 1 0, 0 0, 2 0, 1 SPE(Γ) = {(T, L), (O, R)}
(O, R) equilibrium is not plausible: R is strictly dominated for player 2

SPE does not test for sequential rationality at every non-singleton
information set
Levent Koçkesen (Koç University) Extensive Form Games III 1 / 57 Levent Koçkesen (Koç University) Extensive Form Games III 2 / 57
Sequential Rationality Beliefs and Assessments

In this game sequential rationality of player 2 requires playing L Unless noted otherwise, from now on we restrict our analysis to extensive
How about in the following game? form games
1, 3
O 1 with perfect recall, and
T B in which every information set is finite
2
L R L R
Definition (Belief System)
A belief system is a collection of probability distributions µ = (µ(.|I))I∈I ,
2, 1 0, 0 0, 1 0, 2
where µ(h|I) denotes the probability assigned to h ∈ I. We denote the set
Depends on player 2’s beliefs of all possible beliefs by M.
We have to introduce beliefs in addition to strategies
Definition (Assessment)
... and require sequential rationality at every information set given
beliefs An assessment (β, µ) ∈ B × M is a behavioral strategy profile combined
This immediately eliminates (O, R) equilibrium for the game in the with a belief system.
previous slide
Outcomes and Expected Payoffs Outcomes and Expected Payoffs
Take any assessment (β, µ) ∈ B × M

Conditional probability assigned to history h given that history h′ has Conditional probability assigned to history h given that information
been reached: P β (h|h′ ) set I has been reached: P β,µ (h|I)
◮ If there is no sequence (a1 , a2 , . . . , ak ) such that X
h = (h′ , a1 , a2 , . . . , ak ), then P β (h|h′ ) = 0 P β,µ (h|I) = µ(h′ |I)P β (h|h′ )
◮ If there is a sequence (a1 , a2 , . . . , ak ) such that h′ ∈I
h = (h′ , a1 , a2 , . . . , ak ), then let h0 = h′ , hl = (h′ , a1 , a2 , . . . , al ), for
any l = 1, 2, . . . , k − 1, and define ◮ Note that there is at most one h′ ∈ I such that P β (h|h′ ) > 0.
k−1
Y Expected payoff of the assessment (β, µ) conditional on reaching
P β (h|h′ ) = βι(hl ) (al+1 |hl ) information set I to the player who is moving at that information set
l=0
X
Uι(I) (β, µ|I) = P β,µ (z|I)ui (z)
Perfect recall implies that each hl lies in a different information set z∈Z
and hence events {hl+1 occurs conditional on hl occuring} are
independent and multiplying their probabilities makes sense.
Sequential Rationality Is Sequential Rationality Sufficient?

Consider the following game
O 1
2, 1
L R
Definition (Sequential Rationality) 2
An assessment (β, µ) ∈ B × M is sequentially rational if for all I ∈ I

l r l r

′
βι(I) ∈ argmax Uι(I) (βι(I) , β−ι(I) ), µ|I
′
βι(I) ∈Bι(I) 0, 1 3, 0 1, 0 0, 1
The following assessment is sequentially rational

β1 (L|{a0 }) = 1, β2 (r|{L, R}) = 1, µ(R|{L, R}) = 1
Yet, the strategy profile is not even a Nash equilibrium of this game
There is something unsatisfactory about the beliefs
We would like beliefs to be consistent with strategies
Weak Sequential Equilibrium Is Weak Sequential Equilibrium Good Enough?
For any β ∈ B and h ∈ H let I(h) be the information set that contains h, Consider the following game
and X 1
P β (I(h)) = P β (h′ )
O E
h′ ∈I(h)
1
0, 20
Definition (Weak Consistency) F

2
A
An assessment (β, µ) ∈ B × M is weakly consistent if the belief system is

F A F A
derived from the strategies using Bayes’ rule whenever possible. In other
words, for every h ∈ H such that P β (I(h)) 6= 0 −5, 0 8, −1 −1, 8 10, 10
P β (h) The following assessment is a weak sequential equilibrium

µ(h|I) =
P β (I(h))
β1 (O|{a0 }) = 1, β1 (A|{E}) = 1, β2 (F |{(E, F ), (E, A)}) = 1
µ((E, F )|{(E, F ), (E, A)}) = 1
Definition (Weak Sequential Equilibrium)
Weak consistency is trivially satisfied since player 2’s information set
An assessment (β, µ) ∈ B × M is a weak sequential equilibrium if it
is off-the-equilibrium
satisfies sequential rationality and weak consistency.
However, the strategy profile is not a SPE
Consistency Sequential Equilibrium

If a behavioral strategy profile β assigns a strictly positive probability
to every action, we say β is completely mixed and write β ≫ 0.
Clearly, if β ≫ 0, then P β (I(h)) > 0 for all h ∈ H
For every β ≫ 0, there is a unique belief system µ such that (β, µ) is
weakly consistent
Definition (Sequential Equilibrium)
Definition (Consistency) An assessment (β, µ) ∈ B × M is a sequential equilibrium if it satisfies
An assessment (β, µ) ∈ B × M is consistent if there exists a sequence sequential rationality and consistency.
(β n , µn ) → (β, µ) such that β n ≫ 0 and (β n , µn ) is weakly consistent for
all n.
The intuition is that probability of an event conditional on zero

probability events must approximate the probability that is derived
from “nearby” completely mixed strategies.
For on-the-path information sets, weak consistency uniquely
determines beliefs and implies consistency.
Sequential Equilibrium Sequential Equilibrium
Consider again
1 What is the set of sequential equilibria for this game?
O E
1
0, 20 1
F A
2
O E
F A F A 1
0, 20
−5, 0 8, −1 −1, 8 10, 10
F A
2
and the weak sequential equilibrium F A F A
β1 (O|{a0 }) = 1, β1 (A|{E}) = 1, β2 (F |{(E, F ), (E, A)}) = 1 −5, 0 8, −1 −1, 8 10, 10
µ((E, F )|{(E, F ), (E, A)}) = 1

There are three possibilities
1 β (F |{(E, F ), (E, A)}) = 1
2
For any weakly consistent (β n , µn ) we must have ◮ sequential rationality implies β1 (A|{E}) = 1
µn ((E, F )|{(E, F ), (E, A)}) = β1n (F |{E}) ◮ consistency implies µ((E, A)|{(E, F ), (E, A)}) = 1
The convergence condition requires that β1n (F |{E}) → 0 and hence ◮ sequential rationality implies β2 (F |{(E, F ), (E, A)}) = 0, contradiction
µn ((E, F )|{(E, F ), (E, A)}) → 0,
But µ((E, F )|{(E, F ), (E, A)}) = 1 in the assessment
Sequential Equilibrium Sequential Equilibrium

1
O E
1 1
0, 20
F A
O E
2
1
0, 20
F A F A
F A
2
−5, 0 8, −1 −1, 8 10, 10
F A F A
2 β2 (F |{(E, F ), (E, A)}) = 0 −5, 0 8, −1 −1, 8 10, 10
◮ sequential rationality implies β1 (A|{E}) = 1

◮ consistency implies µ((E, A)|{(E, F ), (E, A)}) = 1
◮ sequential rationality implies β2 (F |{(E, F ), (E, A)}) = 0 3 β2 (F |{(E, F ), (E, A)}) ∈ (0, 1)
◮ sequential rationality implies β1 (E|{a0 }) = 1 ◮ sequential rationality of player 1 implies β1 (A|{E}) = 1
(β n , µn ) below satisfies weak consistency and convergence criteria ◮ consistency implies µ((E, A)|{(E, F ), (E, A)}) = 1
1
◮ sequential rationality of player 2 implies β2 (A|{(E, F ), (E, A)}) = 1, a
β1n (E|{a0 }) = β1n (A|{E}) = β2n (A|{(E, F ), (E, A)}) = 1 − contradiction
n+1
1
µn ((E, A)|{(E, F ), (E, A)}) = 1 −
n+1
Selten’s Horse Selten’s Horse
1 C 2 c
1, 1, 1
1 C 2 c
D d 1, 1, 1
hαi hβi
3 D d
hµi hαi
3
hβi
hµi
L R L R L R L R
hγi hγi hγi hγi
3, 3, 2 0, 0, 0 4, 4, 0 0, 0, 1
3, 3, 2 0, 0, 0 4, 4, 0 0, 0, 1
γ=1
Let µ = µ(D|{D, (C, d)}), α = β1 (D|{a0 }), β = β2 (d|{C}), γ =
β3 (L|{D, (C, d)}) Sequential rationality (SR) of player 2 implies β = 1
There are three possible types of sequential equilibria: SR of player 1 implies α = 0
1 γ=1 Bayes’ rule (BR) implies µ = 0
2 γ=0 But then SR of player 3 implies γ = 0, a contradiction
3 γ ∈ (0, 1)
Let’s investigate each.

For any µ ∈ (0, 1/3], let
µ 1−µ n 1 1
αn = , βn = ,γ = , µn =
1 C 2 c
1, 1, 1
n+1 n+1 n+1 1
µ − 1−µ
n+1
D d
hµi
hαi
3
hβi For µ = 0, let
L R L R 1 1 1 n+1
hγi hγi
αn = , βn = , γn = , µn =
3, 3, 2 0, 0, 0 4, 4, 0 0, 0, 1 (n + 1)2 n+1 n+1 n + (n + 1)2
Note that αn → α, β n → β, γ n → γ, µn → µ and
γ=0
Sequential rationality (SR) of player 2 implies β = 0 αn
µn =
αn + (1 − αn )β n
SR of player 1 implies α = 0
BR has no bite, but SR of player 3 implies 1 − µ ≥ 2µ or µ ≤ 1/3 so that consistency is satisfied. Therefore, the following is a class of SE
whose members differ only in off-the-equilibrium beliefs.
{(α, β, γ, µ) : α = β = γ = 0, µ ≤ 1/3}
So let µ = 1/3, α = 0, β = 0, γ ∈ (0, 1/4] and
1 2
C c
1, 1, 1 1 2 1
αn = , βn = , γ n = γ, µn = 2
D
hαi
d
hβi
3(n + 1) 3(n + 1) 3 − 3(n+1)
3
hµi
L R L R Note that αn → α, β n → β, γ n → γ, µn → 1/3 and

hγi hγi
αn
3, 3, 2 0, 0, 0 4, 4, 0 0, 0, 1
µn =
αn + (1 − αn )β n
γ ∈ (0, 1) so that consistency is satisfied.
Sequential rationality (SR) of player 3 implies µ = 1/3 Combining this with the class of SE found previously, we have
Let’s consider the three possibilities:
Proposition
1 β = 1: SR of 1 implies α = 0. But then BR implies µ = 0, a
contradiction. The set of SE behavioral strategy profile is given by
2 β ∈ (0, 1): SR of 2 implies γ = 1/4, which, by SR of 1, implies {(α, β, γ, µ) : α = β = 0, γ ∈ [0, 1/4]}
α = 0. But then BR implies µ = 0, a contradiction
3 β = 0: SR of 2 implies γ ≤ 1/4, which, by SR of 1, implies α = 0. Each such strategy profile can be combined with some µ ≤ 1/3 to
constitute a SE.
Properties of Sequential Equilibrium Properties of Sequential Equilibrium
Proposition (One-Deviation Property)

Proposition (Kreps and Wilson (82))
A consistent assessment (β, µ) is sequentially rational if, and only if, for all
Every finite extensive form game with perfect recall has a sequential I ∈ I and all β ′ ∈ B satisfying βι(I)
′ (.|I ′ ) = β ′ ′
ι(I) (.|I ) for all I 6= I, the
equilibrium. following is true
Uι(I) (β, µ|I) ≥ Uι(I) (β ′ , µ|I)
Proposition (Kreps and Wilson (82))
Every sequential equilibrium is a subgame perfect equilibrium. Proof.
See Hendon et al (1996), Games and Economic Behavior
Perfect Bayesian Equilibrium Games with Observed Actions and Perfect Bayesian
Equilibrium
In an extensive form game with incomplete information and observed
actions
Consistency is difficult to work with which makes sequential Each player has a type that is revealed only to the player at the
equilibrium difficult to use in applications beginning of the game
A widely used alternative: perfect Bayesian equilibrium After each history in the game, players simultaneously choose actions,
◮ Developed by Fudenberg and Tirole (1991) for a particular class of which are revealed before players choose actions again
extensive form games
⋆ Extensive form games with incomplete information and observed actions Perfect Bayesian equilibrium requires that
◮ Puts more restrictions on off-the-equilibrium beliefs than weak Posterior beliefs about a player’s type is derived from prior beliefs
sequential equilibrium but less restrictions than sequential equilibrium using Bayes’ rule even at off-the-equilibrium information sets, as long
as the player’s action has positive probability conditional on the
history
Beliefs about a player’s type depend only on that player’s actions
Beliefs are such that the players’ types are independently distributed
Games with Observed Actions Games with Observed Actions

4 Θi : a finite set of types for player i
◮ Θ = ×i∈N Θi : set of type profiles
An extensive form game with incomplete information and observed actions 5 p: Nature’s probability distribution over Θ
is composed of the following elements: ◮ Nature moves only once at the beginning of the game choosing a type
profile θ according to p
1 N : a finite set of players ◮
Q
We assume types are independent: p(θ) = i∈N pi (θi )
2 H: a set of histories ⋆ pi is the marginal distribution over Θi
⋆ Assume pi (θi ) > 0, ∀θi ∈ Θi
3 ι : H \ Z → 2N : the player function
6 An information partition I
◮ Players in ι(h) move simultaneously after history h
◮ There is a collection of sets {Ai (h)}i∈ι(h) such that
◮ Each player observes only her own type and the history so far
◮ Player i’s information set after history h is given by
A(h) = {a : (h, a) ∈ H} = ×i∈ι(h) Ai (h) ′
I(θi , h) = {((θi , θ−i ′
), h) : θ−i ∈ Θ−i }
⋆ Player i chooses from Ai (h) after history h 7 For each i ∈ N , a von Neumann-Morgenstern payoff function
ui : Θ × Z → R
Games with Observed Actions Definition (Perfect Bayesian Equilibrium)
An assessment (β, µ) ∈ B × M is a perfect Bayesian equilibrium (PBE) if
it is sequentially rational and for all h ∈ H \ Z
1 Strong Bayes’ rule: For all i ∈ ι(h) and ai ∈ Ai (h), if there exists
θi′ ∈ Θi with µ(θi′ |h) > 0 and βi (ai |θi′ , h) > 0 (i.e., ai has positive
An assessment is given by a behavioral strategy profile and belief system probability conditional on h), then for any θi ∈ Θi
(β, µ), where
βi (ai |θi , h) denotes the probability with which player i of type θi µ(θi |h)βi (ai |θi , h)
µ(θi |h, a) = P
chooses action ai after history h θ̃i ∈Θi µ(θ̃i |h)βi (ai |θ̃i , h)
µ(θi |h) denotes the (common) probability assigned, after history h, to 2 Independence: For all θ ∈ Θ
player i being of type θi
Y
µ(θ|h) = µ(θi |h)
i∈N
3 Action determined beliefs: For all θi ∈ Θi

1 If i ∈
/ ι(h) and a ∈ A(h), then µ(θi |h, a) = µ(θi |h)
2 If i ∈ ι(h) and a, a′ ∈ A(h) and ai = a′i , then µ(θi |h, a) = µ(θi |h, a′ )
The following figure illustrates the application of the Bayes’ rule

PBE Beliefs
Nature
Strong Bayes rule:
x y
◮ The denominator is the probability assigned to player i playing action 1 hpi h1 − pi 1
ai conditional upon reaching history h, whereas the numerator is the L R L R

probability assigned to player i playing action ai and having type θi
conditional upon reaching history h.
2 2
◮ It applies even if hµi h1 − µi hpi h1 − pi
D D D D
⋆ h has zero probability under β, and
1 1 1 1
⋆ h has positive probability but, under β, a is a zero probability event
l r l r l r l r
after h. This, of course, must be due to some other player deviating 2 2
hµ1 i h1 − µ1 i h0i h1i
from a, as there is at least one type of player i who put positive
probability on ai . 2 2
hµ2 i h1 − µ2 i h1i h0i
◮ It says that once beliefs are formed about player i at an off-the-path
information set, they should, together with βi , form the basis for
updating at subsequent information sets until player i deviates from βi The standard Bayes’ rule leaves µ, µ1 , µ2 free. But the stronger rule
again. In that sense it prevents belief reversals. defined above says that
Action determined beliefs: Other players’ actions cannot give µ is free
information about a player’s type. It fits our assumption that players’ µ1 = µ
strategies are independent. µ2 is free
Perfect Bayesian Equilibrium PBE vs. SE
Proposition Consider the following (portion of a) game
Every SE is a PBE. Player 1 has three types l, m, r, player 2 has only one type
the game has reached a history (h) where player 2 gives probability 1
Proposition to type r
If each player has at most two types, or the game has at most two stages,
Strategies of the players are indicated by arrows, and beliefs in angled
then the sets of PBE and SE coincide.
parentheses
l m r
Proof.
See Fudenberg and Tirole (1991), Journal of Economic Theory h0i 2 h0i 2 h1i
D D D
1 1 1
Proposition (One-Deviation Property)
L L L
An assessment (β, µ) that satisfies the strong Bayes rule is sequentially h0i
M
h1i
M
h0i
M
rational if, and only if, for all I ∈ I and all β ′ ∈ B satisfying R R R
h1i h0i h0i
′ (a|I) = β
βι(I) ι(I) (a|I) whenever a ∈ / A(I), the following is true
h0i h0i h1i
′
Uι(I) (β, µ|I) ≥ Uι(I) (β , µ|I)
Proof.
See Hendon et al (1996), Games and Economic Behavior
l m r
l m r
h0i 2 h0i 2 h1i
D D D
1 1 1
h0i 2 h0i 2 h1i L L L
h0i h1i h0i
M M M
D D D R R R
1 1 1 h1i h0i h0i
h0i h0i h1i

L L L
h0i h1i h0i
M M M
R R R Bayes’ rule implies
h1i h0i h0i
µn (m|h)β n (L|m, h)
h0i h0i h1i µn (m|h, L) =
µn (m|h)β n (L|m, h) + µn (l|h)β n (L|l, h) + µn (r|h)β n (L|r, h)
Do these beliefs satisfy the conditions for PBE? Divide the numerator and denominator by µn (m|h)
Since there is no θ1 such that µ(θ1 |h) > 0 and β1 (L|θ1 , h) > 0 beliefs β n (L|m, h)
following L are free. µn (m|h, L) = µn (l|h) n µn (r|h) n
β n (L|m, h) + µn (m|h) β (L|l, h) + µn (m|h) β (L|r, h)
The same is true for beliefs that follow M .
(l|h) n
Is the assessment consistent? Does there exist completely mixed Consistency and µn (m|h, L) → 1, β n (L|l, h) → 1 imply that µµn (m|h) → 0.
(β n , µn ) → (β, µ) that satisfy Bayes’ rule for every n? Similarly, consistency and µn (l|h, M ) → 1, β n (M |m, h) → 1 imply that
µn (m|h)
n
µ (l|h) → 0, a contradiction.
PBE vs. SE An Application: Reputation in Chain-Store Game
Consider the following simple entry game with a > 1 and 0 < b < 1
Player 1 is a potential entrant who may enter (E) or stay out (O)
Player M is the incumbent monopolist who may fight (F ) or
This is an example proving that not every PBE beliefs are consistent, accommodate (A)
and hence not every PBE is a SE. The following is the game tree, where the first payoff number is that
of the monopolist
Fudenberg and Tirole (1991) extend the definition to relative
probabilities and establish equivalence between PBE and SE when 1
beliefs about relative probabilities satisfy the three conditions given in
the definition of PBE. O E
Kohlberg and Reny (1997) provide another characterization of M
consistency that does not use sequences. a, 0
F A
−1, b − 1 0, b
Unique SPE is (E, A)

An Application: Reputation in Chain-Store Game An Application: Reputation in Chain-Store Game

There are two possible types of the monopolist: Tough (τ ) and Weak
Now consider the K market version: the same game is played in markets (ω)
1, 2, . . . , K against different entrants in sequence Tough monopolist prefers to fight in any single market
First, market K is played against entrant K, second against entrant Weak has the preferences given before
K − 1, and so on. Nature determines the type of the monopolist: Tough with probability
Monopolist’s payoff is the sum of his payoffs in K markets. ε ∈ (0, 1), Weak with 1 − ε
Assume that there is no integer k such that ε = bk
Entrant k observes the history of market interaction up to and
Monopolist knows his type but entrants don’t
including market k + 1
Here is the K = 1 version
What is the SPE?
Nature
In the unique SPE every entrant enters and the monopolist
accommodates in every market independent of the history ω τ
h1 − εi 1 hεi
Couldn’t the monopolist build a reputation for being a “tough”
O E O E
competitor by fighting in earlier markets so as to deter entry later on?
M M
Formalization of this idea requires introducing incomplete information a, 0 a, 0

F A F A
−1, b − 1 0, b 0, b − 1 −1, b
K=1 Nature
ω τ
h1 − εi 1 hεi
Nature
O E O E
ω τ
M M
h1 − εi 1 hεi
a, 0 a, 0
F A F A
O E O E
M M
−1, b − 1 0, b 0, b − 1 −1, b
a, 0 a, 0
F A F A
−1, b − 1 0, b 0, b − 1 −1, b Perfect Bayesian Equilibrium

Sequential Rationality (SR) ⇒ βM (F |ω, E) = 0, βM (F |τ, E) = 1
Action Determined Beliefs (ADB) ⇒ µ(τ |a0 ) = ε
N = {1, M }
ADB ⇒ µ(τ |E) = ε
H = {a0 , O, E, (E, F ), (E, A)} Expected payoff to entering is
ι(a0 ) = 1, ι(E) = M
(1 − ε)b + ε(b − 1) = b − ε
ΘM = {ω, τ }, Θ1 = {x}
p(τ ) = ε Therefore, SR ⇒ (
0, ε>b
For all h : ι(h) = 1, I(x, h) = {((x, θM ), h) : θM ∈ {ω, τ }} β1 (E|a0 ) =
For all h : ι(h) = M, I(θM , h) = {((θM , x), h)}, θM ∈ {ω, τ } 1, ε<b
Let’s find PBE for the two markets version: Market 2 is played first and
Market 1 second. Also assume b2 < ε < b k=2
Let’s organize our search for PBE according to the way tough monopolist
k=1 plays. There are three possibilities:
For any market 2 history h, SR ⇒ βM (F |ω, h, E) = 0, βM (F |τ, h, E) = 1 1 βM (F |τ, E) = 0
 2 βM (F |τ, E) = 1
0,
 µ(τ |h) > b
3 βM (F |τ, E) ∈ (0, 1)
β1 (E|h) = ∈ [0, 1], µ(τ |h) = b
 We will investigate each in turn.
1, µ(τ |h) < b

1 βM (F |τ, E) = 0
a. βM (F |ω, E) = 1
⋆ SBR and ADB ⇒ µ(τ |E, F ) = 0, µ(τ |E, A) = 1
⋆ UM ((F, β−M ), µ|ω, E) = −1 + 0 < 0 + a = UM ((A, β−M ), µ|ω, E), 2 βM (F |τ, E) = 1
contradiction ( ) a. βM (F |ω, E) = 1
b. βM (F |ω, E) = 0 ⋆ SBR and ADB ⇒ µ(τ |E, F ) = ε, µ(τ |E, A) ∈ [0, 1]
⋆ SBR and ADB ⇒ µ(τ |E, F ) ∈ [0, 1], µ(τ |E, A) = ε ⋆ UM ((F, β−M ), µ|ω, E) = −1 + 0 < 0 ≤ UM ((A, β−M ), µ|ω, E),
⋆ UM ((A, β−M ), µ|τ, E) = −1 + 0 < 0 ≤ UM ((F, β−M ), µ|τ, E), b. βM (F |ω, E) = 0
c. βM (F |ω, E) ∈ (0, 1) ⋆ SBR and ADB ⇒ µ(τ |E, F ) = 1, µ(τ |E, A) = 0
⋆ SBR and ADB ⇒ µ(τ |E, F ) = 0 and ⋆ UM ((A, β−M ), µ|ω, E) = 0 + 0 < −1 + a = UM ((F, β−M ), µ|ω, E),
ε×1
µ(τ |E, A) = >ε
ε × 1 + (1 − ε) × βM (A|ω, E)
⋆ UM ((F, β−M ), µ|ω, E) = −1 + 0 < 0 ≤ UM ((A, β−M ), µ|ω, E),
Therefore, the following is a PBE of the two-market game

2 βM (F |τ, E) = 1 (cont’d)
c. βM (F |ω, E) ∈ (0, 1) (1 − b)ε
⋆ SBR and ADB ⇒ µ(τ |E, A) = 0 and β2 (O|a0 ) = 1, βM (F |τ, E) = 1, βM (F |ω, E) =
b(1 − ε)
ε×1
µ(τ |E, F ) = >ε
ε × 1 + (1 − ε) × βM (F |ω, E)
βM (F |ω, h, E) = 0, βM (F |τ, h, E) = 1
⋆ SR ⇒ UM ((A, β−M ), µ|ω, E) = 0 + 0 = −1 + β1 (O|E, F )a ⇒
β1 (O|E, F ) = 1/a 
⋆ SR ⇒ UM ((A, β−M ), µ|τ, E) = −1 + 0 < 0 + ( a1 × a) = ε, h ∈ {a0 , O}

UM ((F, β−M ), µ|τ, E), X µ(τ |h) = 0, h = (E, A)
⋆ SR and β1 (O|E, F ) = 1/a ⇒ b = µ(τ |E, F ) ⇒ 
b, h = (E, F )

(1 − b)ε
βM (F |ω, E) = (
b(1 − ε) 0, h ∈ {O, (E, A)}
β1 (O|h) =
⋆ Entrant 2’s expected payoff to E 1/a, h = (E, F )

(1 − b)ε ε
ε(b − 1) + (1 − ε) b − =b−
b(1 − ε) b In this equilibrium, even the weak monopolist fights earlier with
positive probability so as to deter entry later on.
Therefore, since we assumed ε > b2 , SR ⇒ β2 (E|a0 ) = 0
In this sense we may say that the weak monopolist “builds”
3 βM (F |τ, E) ∈ (0, 1) is left as an exercise.
reputation for being tough.
Let k(h) be the market number that corresponds to history h. For any
Reputation in Chain-Store Game h : ι(h) = M
βM (F |τ, h) =1

0,

 k(h) = 1
Now consider the K-market version. We will show that an equilibrium βM (F |ω, h) = 1, k(h) > 1 and µ(τ |h) ≥ bk(h)−1
k(h)−1 )µ(τ |h)
with the following features exists: (1−b

, k(h) > 1 and µ(τ |h) < bk(h)−1


(1−µ(τ |h))bk(h)−1
Tough monopolist always fights
For any h : ι(h) = k, k = K, K − 1, . . . , 1
Weak monopolist fights in the earlier markets and starts 
k
accommodating with positive probability towards the end. 1, µ(τ |h) > b

Entrants stay out in the earlier markets βk (O|h) = a1 , µ(τ |h) = bk

0, µ(τ |h) < bk

Let k = K, K − 1, . . . , 2, 1 and hk ∈ {O, (E, A), (E, F )} stand for the
market k sequence of events. µ(τ |a0 ) = ε and beliefs at market k − 1 as a function of beliefs at market
k and hk are given by

µ(τ |h),
 hk = O
µ(τ |h, hk ) = max{bk−1 , µ(τ |h)}, hk = (E, F ) and µ(τ |h) > 0

0, hk = (E, A) or µ(τ |h) = 0

Let’s verify that this is a PBE.

Beliefs Entrant k
Action determined beliefs (ADB) implies µ(τ |a0 ) = ε and
µ(τ |h, O) = µ(τ |h). If µ(τ |h) ≥ bk−1 , then βM (F |θ, h) = 1, θ = τ, ω, so βk (O|h) = 1 is
If, µ(τ |h) = 0 we either have µ(τ |h, hk ) = 0 by strong Bayes rule optimal.
(SBR) or beliefs are free. If bk < µ(τ |h) < bk−1 , then probability of fight in market k is
Since βM (A|τ, h) = 0, hk = (E, A) implies µ(τ |h, hk ) = 0, by SBR,
or µ(τ |h, hk ) is free. (1 − bk−1 )µ(τ |h) µ(τ |h)
µ(τ |h) + (1 − µ(τ |h)) = k−1
If µ(τ |h) > 0 and hk = (E, F ), SBR implies (1 − µ(τ |h))bk−1 b
µ(τ |h)βM (F |τ, h, E) and the expected payoff to entering is
µ(τ |h, E, F ) =
µ(τ |h)βM (F |τ, h, E) + (1 − µ(τ |h))βM (F |ω, h, E) 
< 0, µ(τ |h) > bk
◮ If µ(τ |h) ≥ bk−1 , then βM (F |ω, h, E) = 1, and hence bk − µ(τ |h) 
µ(τ |h, E, F ) = µ(τ |h). = 0, µ(τ |h) = bk
bk−1 
If, µ(τ |h) < bk−1 , then > 0, µ(τ |h) < bk
◮

(1 − bk−1 )µ(τ |h)

βM (F |ω, h, E) = and hence the entrant’s strategy is sequentially rational.
(1 − µ(τ |h))bk−1
k−1
and hence µ(τ |h, E, F ) = b .
Therefore, beliefs satisfy the three conditions for PBE.
Tough monopolist Weak monopolist (cont’d)
By accommodating the tough monopolist gets −1 today and 0 afterwards, Optimality in the case µ(τ |h) < bk(h)−1 can be proved by induction.
whereas by fighting he gets 0 today and at worst 0 in the future. In market 2, µ(τ |h, E, F ) = b and therefore βk−1 (O|h, E, F ) = 1/a.
Therefore, always fighting is optimal. The monopolist gets an expected payoff of −1 + a1 a = 0 by fighting.
Therefore, mixing is optimal.
Weak monopolist Now suppose mixing is optimal at market k when µ(τ |h) < bk−1 .
If k(h) = 1, there is no tomorrow and A is optimal. So, suppose This implies that, if µ(τ |h) < bk−1 , A and F bring the same expected
k(h) > 1 and note that playing A brings a total payoff of zero. total payoff (of zero) starting at k.
If µ(τ |h) > bk(h)−1 , then µ(τ |h, E, F ) = µ(τ |h) > bk(h)−1 and Now suppose that µ(τ |h) < bk in market k + 1. If the monopolist
therefore βk−1 (O|h, E, F ) = 1, and the monopolist’s payoff is at least fights in this market, µ(τ |h, E, F ) = bk in market k. This implies that
−1 + a > 0. In this case fighting is optimal. βk (O|h, E, F ) = 1/a. Furthermore, since µ(τ |h, E, F ) = bk < bk−1 ,
If µ(τ |h) = bk(h)−1 , then µ(τ |h, E, F ) = bk(h)−1 and therefore if entry occurs the monopolist gets a total payoff of zero starting from
βk−1 (O|h, E, F ) = 1/a. This implies that the payoff to F is at least market k, by the induction hypothesis. Therefore, the payoff to
−1 + a1 a = 0, and hence fighting is again optimal. fighting is −1 + a1 a + (1 − a1 ) × 0 = 0 in market k + 1. This proves
that if µ(τ |h) < bk , then mixing is optimal in market k + 1.
Let min{k : bk < ε} = 4. The following figure illustrates the evolution of beliefs and the equilibrium path.
K The game starts in market K with belief µK = ε. Since µK > bK , entrant stays out and the next period belief
µK−1 = ε. If the out-of-equilibrium outcome (E, F ) ((E, A)) were to occur µK−1 = ε (µK−1 = 0). The
Reputation in Chain-Store Game
game continues in this manner until market 4.
4 In market 4 we still have µ4 = ε > b4 , and therefore entrant stays out. However, since µ4 < b3 , weak monopolist
would mix between E and F if entry were to occur.
3 In market 3 µ3 = ε < b3 , and entrant enters. Since µ3 < b2 , weak monopolist mixes between E and F .
2 If market 3 outcome is (E, A), then µ2 = 0, and entrant enters in market 2. If the outcome in market 3 is (E, F ), Note that the entrants stay out until market min{k : bk < ε} − 1,
then µ2 = b2 and entrant mixes. Since µ2 < b, weak monopolist mixes as well.
1 In market 1, weak monopolist accommodates if entry occurs. Entrant enters if the outcome of market 2 is O or
which is independent of the number of markets. Therefore, in the
(E, A), and mixes if it is (E, F ).
(E, A)
above equilibrium
hbi
O
UM (β, µ)
lim = a, for any ε > 0
N →∞ N
(E, F )
On the other hand, if ε = 0, in the unique SE,
UM (β, µ)
O (E, A)
=0
N
(E, F )
This shows that a small amount of uncertainty can lead to large
(E, F )
hεi O, (E, F ) O, (E, F ) O, (E, F ) O O

(E, A)
changes in equilibrium outcomes.
(E, A) (E, A) (E, A) bk (E, A) (E, A)
(E, A) (E, A)
h0i
K K−1 K−2 4 3 2 1
Reputation in Chain-Store Game
There are other Nash equilibria of this game that are not sequential
equilibria. For example, “always fight” for both types of the
monopolist and “always stay out” for the entrant is such an
equilibrium.
For some parameters, there are also other SE with different outcomes.
But Kreps and Wilson (1982, Journal of Economic Theory) show that
if ε 6= bk for k ≤ K and beliefs are such that
µ(τ |h, F ) ≥ µ(τ |h, A)
for any history h, then the equilibrium we described is unique on the

equilibrium path.
Levent Koçkesen (Koç University) Extensive Form Games III 57 / 57

Sequential Rationality and Subgame Perfect Equilibrium

Caricato da

Informazioni sul documento

Titolo originale

Copyright

Formati disponibili

Condividi questo documento

Condividi o incorpora il documento

Opzioni di condivisione

Hai trovato utile questo documento?

Questo contenuto è inappropriato?

Copyright:

Formati disponibili

Sequential Rationality and Subgame Perfect Equilibrium

Caricato da

Copyright:

Formati disponibili

Subgame Perfect Equilibrium

Strategic form of the game

2, 1 0, 0 0, 2 0, 1 SPE(Γ) = {(T, L), (O, R)}

(O, R) equilibrium is not plausible: R is strictly dominated for player 2

Sequential Rationality Beliefs and Assessments

Take any assessment (β, µ) ∈ B × M

Sequential Rationality Is Sequential Rationality Sufficient?

Definition (Sequential Rationality) 2

An assessment (β, µ) ∈ B × M is sequentially rational if for all I ∈ I

The following assessment is sequentially rational

Definition (Weak Consistency) F

An assessment (β, µ) ∈ B × M is weakly consistent if the belief system is

P β (h) The following assessment is a weak sequential equilibrium

Consistency Sequential Equilibrium

The intuition is that probability of an event conditional on zero

and the weak sequential equilibrium F A F A

β1 (O|{a0 }) = 1, β1 (A|{E}) = 1, β2 (F |{(E, F ), (E, A)}) = 1 −5, 0 8, −1 −1, 8 10, 10

µ((E, F )|{(E, F ), (E, A)}) = 1

Sequential Equilibrium Sequential Equilibrium

2 β2 (F |{(E, F ), (E, A)}) = 0 −5, 0 8, −1 −1, 8 10, 10

◮ sequential rationality implies β1 (A|{E}) = 1

Selten’s Horse Selten’s Horse

L R L R Note that αn → α, β n → β, γ n → γ, µn → 1/3 and

Properties of Sequential Equilibrium Properties of Sequential Equilibrium

Proposition (One-Deviation Property)

Games with Observed Actions Games with Observed Actions

3 Action determined beliefs: For all θi ∈ Θi

The following figure illustrates the application of the Bayes’ rule

ai conditional upon reaching history h, whereas the numerator is the L R L R

h0i h0i h1i

Unique SPE is (E, A)

An Application: Reputation in Chain-Store Game An Application: Reputation in Chain-Store Game

Formalization of this idea requires introducing incomplete information a, 0 a, 0

−1, b − 1 0, b 0, b − 1 −1, b Perfect Bayesian Equilibrium

⋆ UM ((F, β−M ), µ|ω, E) = −1 + 0 < 0 ≤ UM ((A, β−M ), µ|ω, E),

Therefore, the following is a PBE of the two-market game

Let’s verify that this is a PBE.

(1 − bk−1 )µ(τ |h)

hεi O, (E, F ) O, (E, F ) O, (E, F ) O O

µ(τ |h, F ) ≥ µ(τ |h, A)

for any history h, then the equilibrium we described is unique on the

Levent Koçkesen (Koç University) Extensive Form Games III 57 / 57

Potrebbero piacerti anche