EC941 - Game Theory

Lecture 6

Prof. Francesco Squintani

Email: [email protected]
 Structure of the Lecture

Extensive-Form Games with Imperfect Information

Spence Signalling Game
Crawford and Sobel Cheap Talk
In games of perfect information, the subgame perfect
equilibrium coincides with the backward-induction
solution.

But subgame-perfect equilibrium is a more general

concept, defined also for games without perfect
information.
Extensive Form Games with Imperfect
Information
In some strategic interactions, players make
decisions without knowing prior decisions of other
players.
This formulation includes the possibility of
simultaneous moves, and the possibility of
uncertainty about the opponents’ types.
Extensive form games with imperfect information
include also strategic form and Bayesian games.
 Example: Entry Game
Consider an entry game in which the entrant may or
not be ready for a fight, and this is private
information. 1
R U O
2
2,
S F S F 5

3, 1, 1 4, 0, 2
3 3
 1
R U O
The two decision subhistory of player 2 are linked,
to account for the2 fact that she does not know player
1’s prior action. 2,
S F S F 5

3, 1, 1 4, 0, 2
3 3
 Extensive Form Game
Definition An extensive game with imperfect information and chance
moves consists of
a set of players
a set of sequences (terminal histories) with the property that no
sequence is a proper subhistory of any other sequence
a function (the player function) that assigns either a player or
“chance” to every sequence that is a proper subhistory of some
terminal history
a probability distribution (the chance move) over the actions
available after any history to which the player function assigns
chance. The probability distributions are independent.
for each player, a partition (the player’s information partition) of
the set of histories assigned to that player by the player function
for each player, preferences over the set of terminal histories.

Definition A pure strategy of player i in an extensive game is a

function that assigns to each of player i’s information sets Ii an
action in A(Ii) (the set of actions available to player i at the
information set Ii).

With this definition we can transform any extensive form game

into a strategic form game, and find all its Nash Equilibria.
For example, in the entry game, we have:

The set of players is {1, 2}.

The terminal histories are {RF, RS, UF, US, O}.
The player function is P(∅) =1, P(R) = 2, P(U) =2.
There is no chance move,
Player 1’s information partition is {∅}, and player 2’s information
partition is {{R,U}}.
The preferences over the set of terminal histories are
u1(RF)=1, u1(RS)=3, u1(UF)=0, u1(US)=4, u1(O)=2
u2 (RF)=1, u2(RS)=3, u2(UF)=2, u2(US)=3, u2(O)=5
 The strategic form representation of the entry game is:
There are 2 pure
F
strategy S
Nash Equilibria (U, S) and
(O,F).
R
U 1, 1 3,Only (U, S) is reasonable.
3(O, F) contains a non-
0, threat.4,
credible
2 3
O equilibria are subgame perfect, because there are
Both 2, 5 2,
no
5
proper subgames in the extensive form representation.
 Perfect Bayesian Equilibrium
To account that only some Subgame Perfect
Equilibria are reasonable, we introduce the solution
concept of Perfect Bayesian Equilibrium.
We separate strategies from beliefs.
1. Strategies refer to the choice of a player at any
information set.
2. A belief is a probability over histories that belong to
the same information set.
We require beliefs to be `consistent’ with strategies.
Definition A behavioral strategy of player i in an extensive
game is a function that assigns to each of i’s information
sets Ii a probability distribution over the actions in A(I i),
the distributions are independent of each other.

Definition A belief system in an extensive game is a

collection of probability distributions, one for each
information set, over the histories in that information set.

Definition An assessment is a pair consisting of a profile of

behavioral strategies and a belief system.
 Definition An assessment is a Perfect Bayesian Equilibrium
if:
1. Each player’s strategy is optimal whenever she has to
move, given her belief and the other players’ strategies.
2. Each player’s belief is consistent with the strategy profile,
according to the Bayes rule.
3.

Consistency requires that the belief probability of

different histories in the same information set is
determined by the probability that the history is
reached according to the players’ strategies.
 1
R
The Nash Equilibrium (U,US) is a Perfect
O Bayesian
Equilibrium, with 2 the belief that assigns probability 1
to
the history U at the information set {U, R}.2,
S F S F 5

3, 1, 1 4, 0, 2
3 3
 1
R
The Nash Equilibrium U F) is notOa Perfect
(O,
Bayesian
2
Equilibrium, because the strategy F is not optimal,
2, {U,
regardless of 2’s beliefs at the information set
S
R}. F S F 5

3, 1, 1 4, 0, 2
3 3
 Spence Signalling Model
There is one student and two firms.

The student’s skill is H or L<H, private information.

First, the student chooses her amount e of education.

Then, the firms observe e, and offer wages w1 and w2.

Finally, the worker chooses one of the wage offers,

w*.
 The firms believe that the student skill is H with
probability p.
The cost of studying is e/K for a student of skill K.
(Studying is less costly for a high skill student).
The student’s payoff is w* – e/K.
Each firm i’s payoff is K - wi if hiring the student,
and zero otherwise.
 Perfect Bayesian Equilibrium
There are many Perfect Bayesian Equilibria.
In one PBE, the high skill type chooses a positive
amount of education, e*.
Consider the following assessment.
Student’s strategy. Type H chooses e = e* and type L
chooses e = 0. After observing the firms’ wage
offers, both types choose one of the highest offer.
Firms’ belief. Each firm believes that a student is
type H if she chooses e* and type L otherwise.
Firms’ strategies. Each firm offers the wage H to a
student who chooses e* and the wage L to a student
who chooses any other value of e.
The firms’ beliefs are consistent with the student’s
strategy. (No student chooses an education level e
other than e* and 0, so consistency imposes no
restriction on the firms’ beliefs after observing any
such e.)
Now, consider sequential rationality.
Student. The strategy of accepting the highest offer
at the end of the game is clearly optimal.
Now consider the student’s education choice.
A type H student obtains the payoff H - e*/H if she
follows her strategy, choosing the education level e*,
and the payoff L - e/H if deviating to any other level
e.
So, for the assessment to be a PBE, ones need
H - e*/H > L, or e* < H(H - L).
A type L student obtains the payoff L, if she follows
her strategy and chooses the education level 0.
If deviating to any e other than e*, she obtains payoff
L - e/L.
If deviating to e = e*, then the firms believe her
ability to be H, and she obtains the payoff H - e*/L.
So, for the assessment to be a PBE, one needs
L > H - e*/L, i.e., e* > L(H - L).
Each firm’s payoff is 0, given its belief and its
strategy.
If it raises the wage it offers in response to any value
of e, its expected profit is negative, given its belief.
If it lowers the wage its expected profit remains zero
(its offer is not accepted).
In summary, the assessment is a Perfect Bayesian
Equilibrium if and only if
L (H - L) < e* < H (H - L).
 There are other Perfect Bayesian Equilibria.
For example, consider the following assessment.
Student’s strategy. Type H and L chooses e = e’.
After observing the firms’ wage offers, both types
choose one of the highest offer.
Firms’ belief. Each firm believes that a student is
type H with probability p if she chooses e = e ’.
Otherwise, she believes that the student is of type L.
Firms’ strategies. Each firm offers the wage L to a
student choosing e = e’, and w = pH+(1-p)L to a
student who chooses e = e’.
The firms’ beliefs are consistent with the student’s
strategies.
The student’s wage acceptance strategy is optimal.
The type H student’s strategy of choosing e = e’ is
also optimal, because it gives her the highest wage,
as long as pH+(1-p)L – e’/H > L.
 The type L student’s strategy of choosing e = e’ is
also optimal, because it gives her the highest wage,
as long as pH+(1-p)L – e’/L > L.
In sum, it is required that e’ < pL(H – L).
It can be argued that these “pooling” PBE are less
plausible than the “separating” PBE seen before.
The pooling PBE is fragile because there are
deviations e* = e’, that only high-skill students
would choose, if the firms believed that only high-
skill students chose e*.
Crawford and Sobel Cheap Talk
An expert privately observes a state t in [0, 1].

The expert sends a report r (a real number) to a

decision maker.

After receiving the report, the decision maker

chooses an action y (also, a real number).

The DM knows that t is uniformly distributed in

[0,1], i.e. the p.d.f. f is such that f(t) = 1 for t in [0, 1].
 The players’ payoffs are independent of r.
(I.e., the report is cheap talk).
Both players care that the action matches the state.
But the expert is biased relatively to the DM.
The expert’s payoff is uE(y, t) = -(y-(t + b))2.
(Her optimal action given the state t is y = t + b).
The DM payoff is uDM(y, t) = -(y-t)2.
(Her optimal action given the state t is y = t ).
 Perfect Communication
We first show that there cannot be any PBE in which
the expert accurately reports the state.
(I.e., an invertible strategy ρ cannot be part of a PBE.)

By contradiction, say that ρ is invertible and PBE.

Then, consistency implies that the DM believes

(correctly) that the state is t, when the report is r =
ρ(t).
Hence, upon receiving report r = ρ(t), the DM
chooses the action y = t.
Anticipating this, when the state is t, the expert
deviates from the strategy ρ.
Instead of sending the report r = ρ(t), she sends the
report r’= ρ(t+b), which induces the action y = t+b.
Hence, ρ is not sequentially rational, and cannot be
part of a PBE.
Perfect Bayesian Equilibrium
By the above result, every Perfect Bayesian
Equilibrium is described by a partition of the state
space [0, 1].

Each partition is composed of K intervals, where the

admissible set of K depends on the bias level b.

The K intervals are [t0, t1), [t1, t2), …, [tK-1, tK], where
t0=0 and tK=1.
 For any k = 1, …, K-1, all expert types t in [tk-1, tk)
send the same report rk.
Hence, upon receiving report rk, the DM knows that t
is in [tk-1, tk), and chooses y so as to maximize
tk
- (y-t)2 dt
tk-1
tk
Taking the first-order condition, 0 = - 2 (y-t) dt
= - 2[ y - (tk-1+tk)/2 ]. tk-1
So, upon receiving report rk, the DM chooses the
action y = (tk-1+ tk)/2.
In a PBE, the expert anticipates that if reporting r =
rk, she induces the action y = (tk-1+ tk)/2.
For any k = 1, …, K-1, an expert of type tk must be
indifferent between sending report rk and rk+1.
I.e., - ((tk-1+ tk)/2 -(tk + b))2 = - ((tk+ tk+1)/2 -(tk + b))2.
Or, tk+1 - tk = tk – tk-1 +4b.
If these conditions hold, then, for any k = 1, …, K,
every type of expert t in [tk-1, tk) prefers to send
report r = rk than any other report.
The conditions give a second-order difference
equation, solved with boundary conditions t0 = 0 and
tK = 1.
The interval [tk-1, tk) has length tk – tk-1 = t1 + (k-1)4b.
The sum of the intervals [tk-1, tk), for k = 1, …, K,
must be equal to 1:
t1 + (t1 +4b) + … + [t1 +(K-1)4b] = 1.
 Or, Kt1 + 4b [1+ … + (K-1)b] = 1.
I.e., Kt1 + 2bK(K-1) = 1.
So, if b is small enough that 2bK(K-1) < 1, then
there is a value of t1 that satisfies the condition, and
hence a PBE with K intervals.
The largest is b, the smallest is the maximum number
of intervals K* that can be a PBE.
For b> 1/4, K* = 1. For 1/4 < b < 1/12, K* =2, etc.
 We conclude by ranking PBE in ex-ante sense
(i.e., before the state t is realized).
The sender’s ex-ante equilibrium payoff is:
- Σ Pr{t in [tk-1, tk)}
[ (tk-1+tk)/2 - (t+b)]2dt /Pr[tk-1, tk).
The decision maker’s
K tk ex-ante equilibrium payoff is:
- Σ Pr{t in [tk-1, tk)} [ (tk-1+tk)/2 - t]2dt /Pr[tk-1, tk).
k=1 tk-1

K tk

k=1 tk-1
 Expanding the sender’s ex-ante equilibrium payoff,

KΣ tk {[(tk-1+tk)/2 – t] 2 + b2 – 2b [(tk-1+tk)/2 – t]}dt

- Σ t [(tk-1+tk)/2 – t]2 dt - b2,
=k=1
k-1
because, for any k,
[(tk-1
K+tk)/2
tk – t]dt = (tk-1+tk)/2 - (tk-1+tk)/2 = 0.

k=1 tk-1

tk

tk-1
In sum, the sender’s and the decision maker’s ex-
ante equilibrium payoff differ only by the constant
b2.
The PBE maximizing one, maximizes also the other.
It can be shown that, for any b, the PBE maximizing
ex-ante payoff is the one with K* intervals.
This is because the sum of the loss terms in the ex-
ante payoff expression is larger, the coarser is the
equilibrium partition.
 Summary of the Lecture

Extensive-Form Games with Imperfect Information

Spence Signalling Game
Crawford and Sobel Cheap Talk
 Preview of the Next Lecture

Infinitely Repeated Games

Nash and Subgame-Perfect Equilibrium
Finitely Repeated Games