LECTURE ON INCOMPLETE

INFORMATION GAMES
Swapnendu Banerjee
Jadavpur University
Kolkata, INDIA
 1.
Games of Incomplete
Information
 Bayesian Equilibrium
✘ Harsanyi, John C., 1967/1968. "Games with Incomplete Information Played by
Bayesian Players, I-III." Management Science 14 (3): 159-183 (Part I), 14 (5): 320-
334 (Part II), 14 (7): 486-502 (Part III).
✘ We consider a modified Prisoner’s Dilemma game

Assumption:
Prisoner 1 has an influential attorney who is her relative.
If she does not confess and so does prisoner 2 then her sentence is
completely waived.
3
 Modified Prisoner’s dilemma game
✘ Consider the modified prisoner’s dilemma game:
P2
Confess (C) Not Confess (NC)

P1 Confess (C) - 5, - 5 -1, -10

✘ Normal type
Not Confess (NC) -10, -1 0, - 2

✘ Prisoner 2 knows that if both play (NC, NC), then Prisoner 1’s sentence will be
completely waived.
✘ Now, the prisoner 2 can be of two types
a) normal type (exhibiting the payoffs as above)
b) good type -> hates to cheat his partner, pays a psychic penalty of 6
years more in jail if she confesses
4
 Modified prisoner’s dilemma
P2

Confess (C) Not Confess (NC)

P1 Confess (C) (−5, −11) (−1, −10) Good type

Not Confess (NC) (−10, −7) (0, - 2)

✘ In the “normal type” game prisoner 2’s optimal strategy is “C” and in
“good type” game optimal strategy is “ NC ”
✘ Prisoner 1 does not know the type of prisoner 2
 Incomplete information

5
 Harsanyi’s contribution
✘ Converted the incomplete information game into a game of complete
but imperfect information by introducing a new player ‘Nature”
This formalizes prior belief of Prisoner-1
 Prisoner 1 has the following belief:
With probability 𝜇 prisoner 2 is a normal type
With probability 1 − 𝜇 prisoner 2 is a good type

The above belief is common knowledge

6
 Extensive form representation of modified prisoners’ Dilemma
Nature

𝜇 1−𝜇

𝑃1
𝑁𝐶
C C
𝑁𝐶

𝑃2 (Normal) 𝑃2 (𝐺𝑜𝑜𝑑)
𝑁𝐶
C C
𝑁𝐶 𝑁𝐶 C 𝑁𝐶 C

(-1,-10) (-10,-1) (0,-2) (-5,-11) (-1,-10) (-10,-7) (0,-2)

(-5,-5)

7
 Extensive form representation of modified prisoners’ Dilemma
Nature

𝜇 1−𝜇

𝑃1
𝑁𝐶
C C
𝑁𝐶

𝑃2 (𝑁𝑜𝑟𝑚𝑎𝑙) 𝑃2 (Good)
𝑁𝐶
C C
𝑁𝐶 𝑁𝐶 C 𝑁𝐶 C

(-1,-10) (-10,-1) (0,-2) (-5,-11) (-1,-10) (-10,-7) (0,-2)

(-5,-5)

If P1 confess: −5𝜇 − 1(1 − 𝜇)= −4𝜇 − 1

8
 Extensive form representation of modified prisoners’ Dilemma
Nature

𝜇 1−𝜇

𝑃1
𝑁𝐶
C C
𝑁𝐶

𝑃2 (Normal) 𝑃2 (Good)
𝑁𝐶
C C
𝑁𝐶 𝑁𝐶 C 𝑁𝐶 C

(-1,-10) (-10,-1) (0,-2) (-5,-11) (-1,-10) (-10,-7) (0,-2)

(-5,-5)

If P1 not confess: −10𝜇 + 0(1 − 𝜇)

9
 Strategies
Prisoner 1 Prisoner 2

✘ Strategy set {Confess, Not ✘ Mapping from type to

Confess} action

✘ If C, expected payoff i. C if normal, C if good

✗ −4𝜇 − 1 ii. C if normal, NC if good
✘ If NC, expected payoff iii. NC if normal, C if good
✗ −10𝜇 iv. NC if normal, NC if good
✘ Optimal strategy: ✘ Optimal Strategy:
𝟏 𝟏
{C if 𝝁 > 𝟔, NC if 𝝁 < 𝟔} {C if normal, NC if good}
This is Bayesian Nash Equilibrium – an extension of NE for static game of incomplete information
Reference: MWG, Ch 8, Section 8 E, 8E.1

10
 Application: Research and development game
✘ The Model:
✘ R&D research consortium
✘ Two members: Firm 1 and 2
✘ Rule: Independent invention by one firm has to be shared fully
with the other
✘ New invention can come from any of the two firms
✘ To develop a new idea, there is cost 𝑐 ∈ 0,1

11
 Research and development game
✘ The ability to internalize the knowledge from the invention is
each firm’s “type”
✘ For firm 𝑖, type is 𝜃𝑖 , 𝑖 = 1,2
✘ Type is private information
✘ Benefit from R&D is 𝜃𝑖 2 , 𝑖 = 1,2 , also private information
✘ Both firms believe that 𝜃𝑖′ s are drawn independently and
uniformly from [0,1] ⇒ Find pdf…??? f(x)=??...prior belief
✘ Firms simultaneously decide whether to innovate or not
✘ 𝑆𝑖 𝜃𝑖 = 1 if 𝑖th firm innovates, = 0 otherwise.

12
✘ If firm 1 decides to innovate, then net payoff is 𝜃1 2 − 𝑐
✘ If firm 1 decides not to innovate, then expected payoff is 𝜃1 2 ×
𝑃𝑟𝑜𝑏 (𝑆2 𝜃2 = 1)
𝑐
✘ Firm 1 will innovate iff 𝜃1 ≥ = 𝜃1 (Say)
1−𝑃𝑟𝑜𝑏 (𝑆2 𝜃2 =1)

✘ If 𝜃1 ≥ 𝜃1 then 𝑆1 𝜃1 = 1, otherwise 0
✘ Therefore, only a sufficiently high type firm will go for R&D.

13
✘ In isolation (without any other firm), firm 1 innovates if 𝜃1 2 − 𝑐 ≥
0 ⇒ 𝜃1 ≥ 𝑐 = 𝜃1 (say)
✘ Evidently, 𝜃1 > 𝜃1
R&D range (with
Free-riding range another firm)

0 𝜃1 𝜃1 1

No R &D range (with

another firm)

14
 Research and development game
✘ From uniform distribution we get,
P𝑟𝑜𝑏 [𝑆1 𝜃1 = 1]=1−𝜃1 (How?? Use the pdf)
Similarly, 𝑃𝑟𝑜𝑏 [𝑆2 𝜃2 = 1]=1−𝜃2
𝑐
From = 𝜃1
1−𝑃𝑟𝑜𝑏 (𝑆2 𝜃2 =1)

2 𝑐 𝑐
We get, 𝜃1 = = 1−1+𝜃
1−𝑃𝑟𝑜𝑏 [𝑆2 𝜃2 =1] 2
2
=>𝜃1 𝜃2 = c …(1)
Similarly, from firm 2’s condition, we will get
2
𝜃2 𝜃1 = c …(2)
From (1) and (2) we get 𝜃1 = 𝜃2 …(3)
15
 Research and development game
✘ Substituing (3) in (1) we get,
𝜃 ∗ = 𝑐 1/3
✘ Optimal threshold types depends on c, the parameter of the model.
✘ Since 𝑐 ∈ 0,1 , 𝜃 ∈ 0,1
✘ Optimal strategies of both players:

Player 1 Player 2
∀𝜃1 ∈ [0, 𝜃 ∗ ), 𝑆1 𝜃1 = 0 ∀𝜃2 ∈ [0, 𝜃 ∗ ), 𝑆2 𝜃2 = 0

Bayesian Equilibrium ∀𝜃1 ∈ [ 𝜃 ∗ ,1], 𝑆1 𝜃1 = 1 ∀𝜃2 ∈ [ 𝜃 ∗ ,1], 𝑆2 𝜃2 = 1

16
 What else can we say?
✘ Question: What is the equilibrium probability that both firms go for
R&D?
2
✘ Answer: (1−𝜃 ∗ ).(1−𝜃 ∗ )=(1−𝜃 ∗ )
✘ Question: What is the equilibrium probability that only one firm will
opt to innovate?
✘ Answer: 2𝜃 ∗ (1−𝜃 ∗ )
✘ Question: What is the equilibrium probability that no firm will opt to
innovate?
2
✘ Answer: 𝜃 ∗
✘ Reference: MWG, Ch 8, Section 8 E, 8E.2

17
Application 2:Provision of Public good Under Incomplete Info
✘ Simplified model of Palfrey and Rosenthal (1981)
The Model:
✘ There are 2 players, i=1,2
✘ Players decide simultaneously whether to contribute towards a
public good or not
✗ 1-> contribute
✗ 0-> don’t contribute

18
✘ Each player derives a benefit of 1 unit if at least one of them
contributes to the public good provision.
✘ Each player gets ‘0’ if no one contributes.
✘ The benefit of the players are common knowledge.
✘ Cost of contribution is 𝑐𝑖 for player 𝑖.
✘ Cost is private knowledge, 𝑐𝑖 is the type of player 𝑖.
✘ Both players believe that 𝑐𝑖 𝑠 are drawn independently from
the same continuous and strictly increasing cumulative
distribution function P(. ) on [𝑐, 𝑐].
✘ This belief is again common knowledge, where P 𝑐 = 0 and
P(𝑐)=1, => Common Priors.
19
✘ We assume 𝑐 < 1 < 𝑐
✘ Pure strategy of this game is a function 𝑆𝑖 𝑐𝑖 , where 𝑐𝑖 is the type
of the player 𝑖 P2

Contribute (C) Not contribute

(NC)

Contribute (C) (1 − 𝑐1 ), (1 − 𝑐2 ) (1 − 𝑐1 ), 1
P1
Not contribute 1, (1 − 𝑐2 ) 0, 0
(NC)

✘ Characterize the Bayesian Nash equilibrium of this game?

20
✘ Let 𝑍𝑗 = 𝑃𝑟𝑜𝑏 (𝑆𝑗∗ 𝑐𝑗 = 1) be the equilibrium probability that player 𝑗 (where 𝑗 =
1,2) contributes .
✘ Expected payoff of Player-1 if she contributes-
(1 − 𝑐1 )𝑍2 + (1 − 𝑐1 )(1 − 𝑍2 ) = (1 − 𝑐1 )
✘ Expected payoff of Player-1 if she doesn’t contribute-
1. 𝑍2 + 0.(1 − 𝑍2 ) = 𝑍2
So player 1 will contribute iff
(1 − 𝑐1 )≥ 𝑍2
⇒ 𝑐1 ≤ 1 − 𝑍2 = 𝑐1∗ (can be interpreted as threshold type) …(1)
Similarly, for Player-2, 𝑐2 ≤ 1 − 𝑍1 = 𝑐2∗ …(2)
21
Note that 𝑍2 = 𝑃𝑟𝑜𝑏 𝑐 ≤ 𝑐2 ≤ 𝑐2∗ = P(𝑐2∗ )
From (1) 𝑐1 ≤ 1 − P(𝑐2∗ ) = 𝑐1∗
From (2) 𝑐2 ≤ 1 − P(𝑐1∗ ) = 𝑐2∗

✘ Therefore, Player-1 will contribute, i.e., 𝑆1 (𝑐1 )=1 , if 𝑐1 ∈ [𝑐, 𝑐1∗ ].

Player-1 will not contribute, i.e., 𝑆1 (𝑐1 )=0 , if 𝑐1 ∈ (𝑐1∗ , 𝑐 ].
Similarly
For Player-2, 𝑆2 (𝑐2 )=1 , if 𝑐2 ∈ [𝑐, 𝑐2∗ ].
𝑆2 (𝑐2 )=0, if 𝑐2 ∈ (𝑐2∗ , 𝑐 ].

22
✘ Specific Example:
Let the distribution of 𝑐𝑖 be uniform on [0,2].
Therefore, P 𝑐 = 𝑐/2.
Now one can get specific values of 𝑐1∗ and 𝑐2∗ .
From 1 − P(𝑐2∗ ) = 𝑐1∗ and 1 − P(𝑐1∗ ) = 𝑐2∗
we get 1 − 𝑐2∗ /2 = 𝑐1∗ and 1 − 𝑐1∗ /2 = 𝑐2∗ ,
Solving
we get 𝑐1∗ = 𝑐2∗ =2/3.
Thus the thresholds are same ⇒ players are symmetric (obvious)

23
Therefore,
Player-1’s Strategy:
𝑆1 (𝑐1 )=1 , if 𝑐1 ∈ [0, 2/3 ].
𝑆1 (𝑐1 )=0 , if 𝑐1 ∈ (2/3, 2 ].
Similarly
For Player-2, 𝑆2 (𝑐2 )=1 , if 𝑐2 ∈ [0, 2/3 ].
𝑆2 (𝑐2 )=0, if 𝑐2 ∈ (2/3, 2 ].

(2/3, 1 ] is the free riding range, where the benefit is greater than cost but
still the players prefer not to contribute, tries to free ride on other’s
contribution.
Reference: Fudenberg and Tirole Ch 6, Section 6.2
24