Incomplete Information and Bayesian Games

Abhinash Borah

March 30, 2023

Incomplete Information

A situation in which some of the players do not know relevant

information about aspects of the strategic environment (while
others do)
Eg., a player’s payoff from different outcomes may not be known
to some or all of the players
That is, some players in the game hold private information that is
not common knowledge

Incomplete Information & Bayesian Games 2 / 22

Example: Incomplete Information
Consider an industry with two firms: an incumbent (player 1) and a
potential entrant (player 2). Player 1 decides whether to build a new
plant, and simultaneously player 2 decides whether to enter or not.
Player 2 is uncertain whether player 1’s cost of building is high or low,
while player 1 knows her own cost. Suppose payoffs when firm 1 has
high and low costs are as below:

Enter Don’t Enter Enter Don’t Enter

Build 0, −1 2, 0 1.5, −1 3.5, 0
Don’t Build 2, 1 3, 0 2, 1 3, 0

When 1 has high cost When 1 has low cost

Incomplete Information & Bayesian Games 3 / 22

Example: Incomplete Information

Enter Don’t Enter Enter Don’t Enter

Build 0, −1 2, 0 1.5, −1 3.5, 0
Don’t Build 2, 1 3, 0 2, 1 3, 0

When 1 has high cost When 1 has low cost

Don’t build is a dominant strategy for player 1 when her cost is

high
However, when 1’s cost is low, her optimal strategy depends on
whether 2 enters or not
Entering is profitable for player 2 if and only if player 1 does not
build
Incomplete Information & Bayesian Games 4 / 22
Example: Incomplete Information

Enter Don’t Enter Enter Don’t Enter

Build 0, −1 2, 0 1.5, −1 3.5, 0
Don’t Build 2, 1 3, 0 2, 1 3, 0

When 1 has high cost When 1 has low cost

In order to choose her action, player 1 (when her cost is low) must
try to predict player 2’s behavior, which, in turn depends on 2’s
prediction of 1’s behavior and how this behavior in influenced by
1’s private information about her cost.

Incomplete Information & Bayesian Games 5 / 22

Example: Incomplete Information

Enter Don’t Enter Enter Don’t Enter

Build 0, −1 2, 0 1.5, −1 3.5, 0
Don’t Build 2, 1 3, 0 2, 1 3, 0

When 1 has high cost When 1 has low cost

Harsanyi’s Proposal: Translate this situation of incomplete

information into an extensive game of imperfect information
Suppose our fictitious player Nature determines player 1’s
cost—her type. Suppose 1 is a high cost type with probability .5
and a low cost type with probability .5 and this probability
distribution is common knowledge.
Incomplete Information & Bayesian Games 6 / 22
Example: Incomplete Information
Representing the game with incomplete information as an
extensive game of imperfect information

High Cost [.5] [.5] Low Cost

1 1

Don’t Don’t
Build Build
Build Build

2 2 2 2
Don’t Don’t
Enter Enter
Enter Enter
Don’t Don’t
Enter Enter
Enter Enter

(0,-1) (2,0) (2,1) (3,0) (1.5,-1) (3.5,0) (2,1) (3,0)

Incomplete Information & Bayesian Games 7 / 22

Example: Incomplete Information
Strategies

Incomplete Information & Bayesian Games 8 / 22

Example: Incomplete Information
Strategies
Player 2 has just one information set in the game. So a strategy
for player 2 is just an action—enter (E) or don’t enter
(DE)—available at that information set.
Player 1 has two information sets—one after Nature chooses her
type as high cost, and one after Nature chooses her type as low
cost. So a strategy for 1 is a mapping from her set of types to an
action in the set {Build (B), Don’t Build (DB)}.
Accordingly, the set of strategies for player 1 is:
{(B, B), (B, DB), (DB, B), (DB, DB)}
E.g., (B, DB): Build if high cost; don’t build if low cost, etc.
Note: 1’s strategy specify her action for both her types even
though she knows what her type is
Incomplete Information & Bayesian Games 8 / 22
Example: Incomplete Information
Equilibrium

Normal-form representation of this extensive game:

E DE
DB,B .5(2, 1) + .5(1.5, −1) .5(3, 0) + .5(3.5, 0)
= (1.75, 0) = (3.25, 0)
DB,DB .5(2, 1) + .5(2, 1) .5(3, 0) + .5(3, 0)
= (2, 1) = (3, 0)
B,B .5(0, −1) + .5(1.5, −1) .5(2, 0) + .5(3.5, 0)
= (.75, −1) = (2.75, 0)
B,DB .5(0, −1) + .5(2, 1) .5(2, 0) + .5(3, 0)
= (1, 0) = (2.5, 0)

Incomplete Information & Bayesian Games 9 / 22

Example: Incomplete Information

Equilibrium
Normal-form representation and Nash Equilibrium

E DE
DB,B 1.75, 0 3.25, 0
DB,DB 2, 1 3, 0
B,B .75, −1 2.75, 0
B,DB 1, 0 2.5, 0

Two Nash Equilibria: ((DB, B), DE), ((DB, DB), E)

Since the extensive form has no proper subgames, both the Nash
equilibria and also subgame perfect

Incomplete Information & Bayesian Games 10 / 22

Summary: Modeling Strategy Adopted in the Example

How did we model situation of incomplete information in the example?

Modeled what a player does not know about the other player in
terms of probabilistic beliefs
Assumed that players’ probabilistic beliefs are common knowledge
Constructed an extensive game of imperfect information with
corresponding normal form representation and used standard
equilibrium analysis
In the process, constructed “meta player(s)”

Incomplete Information & Bayesian Games 11 / 22

Types
The type of a player embodies any private information that is
relevant to the player’s decision making.
Private information refers to information that a player knows but
is not common knowledge amongst all players.
This may include, in addition to the players’ payoff function, her
belief about other player’s payoff function, her beliefs about
others’ beliefs about her beliefs, her beliefs about others’ beliefs
about her beliefs about others’s beliefs, and so on.
Models with higher order beliefs are analytically challenging. So,
in most applications a player’s type is restricted to certain
“parameters” of her payoff function and the probabilistic process
generating these parameters is assumed to be common knowledge
Incomplete Information & Bayesian Games 12 / 22
Bayesian Game

A static Bayesian game consists of:

1 A set of player
2 A set of types for each player
3 A probability distribution over the set of type profiles
4 A set of actions for each player
5 Players’ payoffs—we assume they are expected utility maximizers

Incomplete Information & Bayesian Games 13 / 22

Bayesian Game

A static Bayesian game may be thought of as proceeding through the

following steps:
1 Nature chooses a profile of types
2 Each player learns her own type, which is her private information.
3 Players update beliefs about types of other players in accordance
with Bayes rule
4 Players simultaneously choose actions
5 Given the players’ choices, each player’s payoff is realized

Incomplete Information & Bayesian Games 14 / 22

Bayesian Game: Strategies

Observe that in a Bayesian game, each type of a player can be

identified with an information set. Therefore, a player’s strategy is
nothing but a mapping from a player’s set of types to her actions.

Incomplete Information & Bayesian Games 15 / 22

Bayesian Nash Equilibrium

A Bayesian Nash equilibrium is a Nash equilibrium of the

extensive form representation of the Bayesian game. This is the
same thing as saying it is a Nash equilibrium of the normal form
representation1 of this extensive form game.

Why not demand subgame perfection?

1
This normal form is referred to as the Bayesian normal form in Watson
Incomplete Information & Bayesian Games 16 / 22
Example (Watson, figures 24.3, 24.4)

Two players, 1 and 2, engage in a game captured by the following

matrix.

C D
A x, 9 3, 6
B 6, 0 6, 9

Player 1’s payoff number x is private information. It is common

knowledge, though, that x = 12 with probability 2/3 and x = 0 with
probability 1/3

Incomplete Information & Bayesian Games 17 / 22

Example: Cournot Duopoly

Cournot Duopoly with Incomplete Information: 2 firms

have to simultaneously decide how much quantity of a
homogenous product to produce in a market with demand
function P = 100 − Q. Both firms have constant unit cost given
by ci (qi ) = θi qi , i = 1, 2. It is common knowledge that θ1 = 60.
Firm 2, on the other hand, has private information about its unit
cost. It is common knowledge, though, that θ2 = 70 with
probability .5 and θ2 = 50 with probability .5.
This situation can be modeled as a Bayesian game where player 1
has a single type θ1 = 60, whereas player 2’s type θ2 is either 70 or
1
50 with probability 2 each.

Incomplete Information & Bayesian Games 18 / 22

Example: Cournot Duopoly
For any type θ2 , firm 2’s equilibrium choice q2 (θ2 ) must satisfy:

q2 (θ2 ) ∈ arg max{(100 − q1 − q2 )q2 − θ2 q2 }

q2

First order condition:

100 − q1 − θ2 − 2q2 = 0
1 q1 θ 2
q2 (θ2 ) = [100 − q1 − θ2 ] = 50 − −
2 2 2
That is,
q1
q2 (70) = 15 −
2
q1
q2 (50) = 25 −
2

Incomplete Information & Bayesian Games 19 / 22

Example: Cournot Duopoly
Firm 1 does not know which type of firm 2 it faces, so its payoff is the
expected value over firm 2’s types:
1
q1 ∈ arg max{ [(100 − q1 − q2 (70))q1 − 60q1 ] +
q1 2
1
[(100 − q1 − q2 (50))q1 − 60q1 ]}
2
1
That is, q1 ∈ arg max{40q1 − q12 − q1 [q2 (70) + q2 (50)]}
q1 2
First order condition:
1
40 − 2q1 − [q2 (70) + q2 (50)] = 0
2
1
That is, q1 = 20 − [q2 (70) + q2 (50)]
4

Incomplete Information & Bayesian Games 20 / 22

Example: Cournot Duopoly
Accordingly,
1
q1 = 20 − [q2 (70) + q2 (50)]
4
1 q1 q1 1
q1 = 20 − [15 − + 25 − ] = 20 − [40 − q1 ]
4 2 2 4
q1
q1 = 10 +
4
40
q1 =
3
and,
q1 25
q2 (70) = 15 − =
2 3
q1 55
q2 (50) = 25 − =
2 3

Incomplete Information & Bayesian Games 21 / 22

Useful Insight

For type of applications we will deal with in this course, there are two
equivalent ways of computing equilibrium in Bayesian games.
1 Compute Nash equilibrium using the normal form. This is
recommended for any game that can be represented in matrix
form.
2 The second method entails treating the types of each player as
separate players. Treating types separately often simplifies the
analysis of infinite games, where calculus may be required

Incomplete Information & Bayesian Games 22 / 22