Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Week 8:
Games of Incomplete Information

Dr Daniel Sgroi
Reading: Osborne chapter 9 (and 10).
With thanks to Peter J. Hammond.

EC202, University of Warwick, Term 2 1 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Introduction
So far we have always made an important assumption:
that the game played is common knowledge.
In particular, the players know:
1. who is playing;
2. what actions or “strategies” are possible for each player;
3. how outcomes translate into payoffs.
Furthermore, this knowledge of the game
is itself common knowledge.
Common knowledge underlies solution concepts such as:
• dominant strategy equilibrium;
• iterated elimination of dominated strategies;
• Nash equilibrium;
• even subgame perfect equilibrium in an extensive form.
EC202, University of Warwick, Term 2 2 of 45
Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Relaxing Common Knowledge

This common knowledge ideal excludes many interesting
and more realistic models of strategic interaction.
The Cournot and Bertrand duopoly models, for instance, each lead
to a clear, precise, easily understood outcome.
But is each firm’s cost function really known to the opponent,
let alone common knowledge?
Clearly, it is more convincing to allow some imprecision
into each firm’s beliefs about the opponent’s costs.
Similarly, a reasonable variant of prisoner’s dilemma allows
the possibility that each prisoner may be subject to omertà
(the Mafia code of silence or connivance).
Or whether in BoS, either player really prefers football to the opera.
How then does one model situations where players
have “incomplete information” about each other’s characteristics?

EC202, University of Warwick, Term 2 3 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Extended Beliefs
A key insight is that a similar issue
arises in any simultaneous move game:
players do not know each others’ actions,
but choose best responses to (probabilistic) beliefs.
Furthermore, equilibrium requires these beliefs,
and the appropriate best responses, to be consistent.
Realizing this similarity, John C. Harsanyi (1967) developed
a very elegant and practically useful way to model beliefs
not only over other players’ actions,
but also over these players’ other characteristics or types.
Harsanyi described a game as having incomplete information
when the players are uncertain about each other’s types.
As with games of complete information,
equilibrium analysis requires these beliefs to be consistent,
meaning there is common knowledge of players’ possible types,
and of the likelihood of each type profile.
EC202, University of Warwick, Term 2 4 of 45
Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Common Knowledge of Player and Action Spaces

Like a game of complete information,

a normal or strategic form game of incomplete information
retains the following two “physical” components
of an n-player simultaneous move game:
Players: the set of players N = {1, 2, ..., n};
Actions: the action space Ai of each player i ∈ N.
We still assume common knowledge of both these items.
But we allow uncertainty about players’ preferences.

EC202, University of Warwick, Term 2 5 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Complete vs. Incomplete Information

• In a game of complete information

each player has a single utility function
(more precisely, one equivalence class of utility functions)
that maps action or strategy profiles into payoffs.
• In a game of incomplete information
each player may have one of many possible utility functions.
Players’ preferences may not be common knowledge.
Player i’s type determines i’s preferences/payoffs.

EC202, University of Warwick, Term 2 6 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Randomized Types

Harsanyi (1967) captured this idea by assuming

that the game begins with a move by Chance
which selects the different players’ preferences
— or more generally, their types.
The type profile determines not only players’ payoffs
but also their beliefs about other players’ types.
Alternatively, one may think of a family of games,
and which one is played is determined by chance.

EC202, University of Warwick, Term 2 7 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

An Entry Game
1b
O @E
r @2r
@
0, 2 F @A
r @r
@
−1, −1 1, 1
Consider a simple “entry game”.
Firm 1 is a potential entrant to an industry,
which decides whether or to enter a market (E ) or stay out (O).
If player 1 does enter, the incumbent firm, player 2,
responds by either fighting (F ) or accommodating (A) entry.
The specified payoffs imply
that (E , A) is a unique subgame perfect equilibrium.
(Use a bimatrix game to convince yourself that (O, F ) is another
pure strategy Nash equilibrium that is subgame imperfect.)
EC202, University of Warwick, Term 2 8 of 45
Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

An Entry Game with Two Types of Incumbent

Chance b
p  HH 1 − p
 H
p 
r p p p p p p p p p 1p p p p p p p p H
pH
p r
 H

O @ E O @ E
@2rr 2cr
@ @
r r @
0, 2 F @A 0, 2 F @A
r @r
@ r @r
@
−1, −1 1, 1 −1, 2 1, 1

Now imagine that, whereas there is only one type of player 1,

there are two types of player 2.
One “rational” type labelled 2r has the payoffs originally specified;
the other “crazy” type labelled 2c enjoys fighting for its own sake,
so its payoff from (Entry, Fight) is 2 instead of −1.

EC202, University of Warwick, Term 2 9 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Chance Types and Self Knowledge

We finish specifying the game by letting p denote
the exogenously specified and commonly known probability
that Chance chooses “rational” (r ) as player 2’s type.
This information allows us to depict the extensive form
of this entry game with incomplete information.
It is natural to suppose players know their own preferences,
and so can find best responses
given what they assume about other players’ behavior.
This explains the dashed information sets
in the extensive form entry game.
These show that firm 2, when making its decision,
knows its own preferences.
Player 1, however, remains unsure about player 2’s preferences.

EC202, University of Warwick, Term 2 10 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Type Probabilities . . .

Players know their own preferences,

but not other players’ preferences or types.
In order to use some kind of equilibrium analysis
we assume players form rational conjectures or correct beliefs
about these unknown preferences and types.
These beliefs allow players to predict other players’ behaviour.
More precisely, though each player
does not necessarily know other players’ actual preferences,
we assume he does know the precise way
in which Chance determines these preferences.
That is, each player knows the probability distribution over types,
and this itself is common knowledge among all the players.

EC202, University of Warwick, Term 2 11 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

. . . Are Commonly Known

In the extensive form of the entry game,

this is represented by player 1’s information set,
and by specifying that p is common knowledge.
COMMENT: Actually, only player 1’s estimate of p really matters;
player 2 has a unique dominant strategy at each information set,
so its behavior is independent of p.
An important feature of the extensive form
is that player 1 has one information set, whereas player 2 has two.
This is because player 2 knows its own type,
while player 1 only has a (correct) belief about player 2’s type.
This belief is specified by the probabilities p and 1 − p
that 2’s type is “rational” (r ) or “crazy” (c).

EC202, University of Warwick, Term 2 12 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

The Common Prior Assumption

Following the terminology of Bayesian statistics,

which Harsanyi himself adopted,
this is often referred to as the common prior assumption.
It means that all the players, before discovering their own types,
share a common prior probabilistic theory
of how Chance chooses the players’ random type profile.
The assumption is rather strong,
but essential if we are to discuss equilibrium behaviour
in any but very special models.

EC202, University of Warwick, Term 2 13 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Type Contingent Strategies

We can translate the entry game into its normal form.
In this form, player 1 must have four pure strategies:
at each of the two information sets
there are two actions to choose from.
A strategy of player 2 can be expressed
as a pair xy ∈ {Ar Ac , Ar Fc , Fr Ac , Fr Fc } = {Ar , Fr } × {Ac , Fc },
where x ∈ {Ar , Fr } describes
what player 2 does if its type is “rational” (r ),
and y ∈ {Ac , Fc } what it does if its type is “crazy” (c).
More generally: when information is incomplete, a player’s strategy
must prescribe what each type of that player should do.
(COMMENT:
Each player’s strategy is contingent on that player’s type.)
Of course, player 1’s strategy set is simply {E , O}.
EC202, University of Warwick, Term 2 14 of 45
Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Expected Payoffs
In this two-player game, each pair of pure strategies
gives rise to a path of play that starts with Nature’s choice,
and follows with the simultaneous actions of both players.
In this example, suppose player 2 plays Ar Fc
— i.e., A if “rational” (type r ), but F if “crazy” (type c).
Suppose player 1 plays E . The outcome will yield payoffs of:
• (1, 1) with probability p;
• (−1, 2) with probability 1 − p.
Expected payoffs from the strategy pair (E , Ar Fc ) are

u1 = p · 1 + (1 − p) · (−1) = 2p − 1
and u2 = p · 1 + (1 − p) · 2 = 2 − p.

EC202, University of Warwick, Term 2 15 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

The Entry Game in Normal Form

2
When p = 3 these payoffs will be (u1 , u2 ) = ( 31 , 43 ).
Similarly, the payoffs from the other seven pure strategy pairs
are as indicated in the table below:
P2
Ar Ac Ar Fc Fr Ac Fr Fc
O 0 0 0 0
2 2 2 2
1
P1 E 1 3 − 31 −1
4
1 3 − 31 0

This is like any other normal form game.

Its origins as a game of incomplete information are irrelevant.
Standard tools can find best responses and Nash equilibria.

EC202, University of Warwick, Term 2 16 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

2
Nash Equilibria in the Entry Game when p = 3

P2
Ar Ac Ar Fc Fr Ac Fr Fc
O 0 0 0∗ 0∗
2∗ 2∗ 2∗ 2∗
1∗
P1 E 1∗ 3 − 31 −1
4
1 3∗ − 13 0

There are three Nash equilibria in pure strategies:

(O, Fr Ac ), (O, Fr Fc ), and (E , Ar Fc ).
In the extensive form, only (E , Ar Fc ) is subgame perfect.
Also, Ar Fc weakly dominates all player 2’s other strategies.
The only mixed strategy equilibria involve 1 playing O,
and 2 mixing Fr Ac and Fr Fc arbitrarily.
(To deter entry, firm 2 must fight if “rational”!)

EC202, University of Warwick, Term 2 17 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Nash Equilibria in the Entry Game for General p

P2
Ar Ac Ar Fc Fr Ac Fr Fc
O 0 0 0 0
2 2 2 2
P1 E 1 2p − 1 1 − 2p −1
1 2−p 1 − 2p 2−p

In case p > 12 , the best responses are as for p = 32 .

The Nash equilibria are (O, Fr Ac ), (O, Fr Fc ), and (E , Ar Fc ),
of which only (E , Ar Fc ) is subgame perfect in the extensive form.
But in case p < 12 , the Nash equilibria are (O, Ar Fc ) and
(O, Fr Fc ),
of which only (O, Ar Fc ) is subgame perfect in the extensive form.

EC202, University of Warwick, Term 2 18 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

“Meta” Players . . .

Each type of any player i ∈ N,

given specific beliefs about other players’ types and actions,
can calculate expected utility separately.
But we averaged the payoffs of all i’s types
using the likelihood of each type as its weight.
This is the average payoff across the different types
of some artificial “meta” player.
Will this “meta” version of player i have optimal responses
that combine all the optimal responses of i’s different types?
This is precisely where
the “type contingent” definition of strategies is so useful.

EC202, University of Warwick, Term 2 19 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

. . . and Their Strategies

For example, suppose player 1 chooses E in the entry game.

When the “meta” player 2 responds with Ar Ac ,
player 2’s “rational” type is playing a best response A,
while the “crazy” type plays the dominated strategy A.
This corresponds precisely to the fact
that Ar Ac is not a best response to E , but Ar Fc is.
Note that Ar Fc gives one (“crazy”) type a higher payoff,
and the other (“rational”) type the same payoff.
Thus, the type contingent strategy Ar Fc
gives the “meta” player a higher expected payoff than Ar Ac .
Once again, an elegant part of Harsanyi’s solution.

EC202, University of Warwick, Term 2 20 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Specifying Games of Incomplete Information

Like a game of complete information,

a game of incomplete information
retains the usual physical components
— namely, the set of players and their action spaces.
These are complemented with preferences
and information components as follows:
• Chance/Nature: a probability distribution over players’ types,
which determine their preferences.
• Each player knows his own type,
but not the other players’ types.
• The probability distribution over types is common knowledge.

EC202, University of Warwick, Term 2 21 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Bayesian Games

A normal form game of complete information is specified

as G = hN, (Si )ni=1 , (ui (·))ni=1 i.
Recall that N = {1, 2, . . . , n} is the set of players,
that Si is player i’s action or strategy space,
and ui : S → R is player i’s utility (payoff) function,
where S := S1 × S2 × · · · × Sn .
In a “Bayesian” game of incomplete information,
players know their own payoffs from different strategy profiles,
but may not know other players’ payoffs.
To allow for this, we have introduced three new ideas.

EC202, University of Warwick, Term 2 22 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Three New Features

1. Before the game is actually played,

Chance or Nature chooses the different players types.
2. Each type ti can represent information
about player i’s own payoffs, or more generally,
about other relevant attributes of the game.
Thus, there is a type space Ti for each player i ∈ N,
representing the range from which Nature (or Chance)
chooses i’s type.
3. We introduce a commonly known
common prior probability distribution p(·) on ni=1 Ti
Q
to describe how Chance/Nature chooses a type profile (ti )ni=1 .
That is, every player i knows his own type ti ∈ Ti ,
and uses this prior to form posterior beliefs
over other profiles t−i of agents’ types.

EC202, University of Warwick, Term 2 23 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Definition of Static Bayesian Game, I

Definition
The normal form representation
of an n-player static Bayesian game is

G ∗ = hN, (Ai )ni=1 , (Ti )ni=1 , ((ui (·; ti ))ti ∈Ti )ni=1 , (pi )ni=1 i.

As before, N = {1, 2, . . . , n} is the set of players,

but now Ai is player i’s action set.
Also, Ti = {ti1 , ti2 , ..., tiki } is player i’s type space,
and ui : A × Ti → R is player i’s type Qn dependent utility function,
where A := A1 × A2 × · · · × An = i=1 Ai .

EC202, University of Warwick, Term 2 24 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Definition of Static Bayesian Game, II

Consider again the general static Bayesian game

G ∗ = hN, (Ai )ni=1 , (Ti )ni=1 , ((ui (·; ti ))ti ∈Ti )ni=1 , (pi )ni=1 i.

The notation ui (a; ti ) indicates

that ui is a utility function of the action profile a,
but it also depends on ti as a parameter.
Finally, pi describes player i’s belief about other players’ types
— that is, given that i knows his type is ti ,
the map t−i 7→ pi (t−i |ti ) is the (posterior) conditional distribution
on t−i (profiles of all other players’ types, excluding i).

EC202, University of Warwick, Term 2 25 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Timing of Static Bayesian Game

We assume the timing of the static Bayesian game is:

1. Nature chooses a type profile (ti )ni=1 ∈ ni=1 Ti .
Q

2. Each player i ∈ N learns his own type ti ∈ Ti ,

which is his private information,
and uses his prior pi to form beliefs pi (t−i |ti )
over the other players’ types t−i .
3. Players simultaneously (as in a static game)
choose actions ai ∈ Ai (i ∈ N).
4. Given the players’ choices a = (a1 , a2 , . . . , an ),
the payoffs ui (a; ti ) of each player i ∈ N are realized.

EC202, University of Warwick, Term 2 26 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Private Values

Note that this setup has a player’s utility ui (a; ti )

depend only on all the players’ actions a ∈ A
and on i’s own type ti ,
but it does not depend on other players’ types t−i .
We call this the private values case,
since each type’s payoff depends only on private information.
For example, how much are you willing to pay
for a minor artist’s oil painting,
when the resale value is negligible?

EC202, University of Warwick, Term 2 27 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Common Values

Later, some interesting examples will require us

to discuss the common values case.
This allows utility functions ui (a1 , a2 , . . . , an ; t1 , t2 , . . . , tn )
that depend on all players’ types.
For example, how much are you willing to pay
for one of Renoir’s oil paintings?
It should last longer than you will,
so the resale value is important,
and depends on other connoisseur’s tastes.

EC202, University of Warwick, Term 2 28 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Bayes’ Rule
Suppose two of many possible states S (sunny weather)
and H (high waves) can occur exclusively or together
according to some prior probability distribution p(·).
Thus, p(S) denotes the prior probability of sunshine,
p(H) the prior probability of high waves,
and p(S ∩ H) the prior probability
that it will be sunny with high waves.
When you wake up, you see it is sunny;
what can you infer about the probability of high waves
(without seeing the sea)?
Bayes’ Rule: Conditional on state S occurring,
the probability that state H occurs as well is

p(S ∩ H)
Pr{H|S} = .
p(S)

EC202, University of Warwick, Term 2 29 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Intuitive Justification

Here is an intuitive justification.

The probability that S occurs
must be the sum p(S) = p(S ∩ H) + p(S \ H)
because S ∩ H and S \ H are mutually exclusive events.
Once S has occurred, then conditional on this knowledge,
the likelihood of H occurring
is the relative likelihood of both S and H occurring,
among all the states in which S occurs. Hence,

p(S ∩ H) p(S ∩ H)
Pr{H|S} = = .
p(S ∩ H) + p(S \ H) p(S)

EC202, University of Warwick, Term 2 30 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Bayesian Game Example

Consider an example of a Bayesian game with two players,
each having two possible types (say, a and b).
So tik will indicate that player i is of type k.
Nature chooses pairs of players’ types
according to a prior joint distribution over
the four possible type combinations (t1k , t2l ), where k, l ∈ {a, b}.
This prior can be described by a joint distribution matrix:

t2a t2b
1 1
t1a 6 3
1 1
t1b 3 6

Table: Prior Joint Distribution

The “common prior” assumption requires everybody

to take this joint distribution as given.
EC202, University of Warwick, Term 2 31 of 45
Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Posterior Beliefs
t2a t2b
1 1
t1a 6 3
1 1
t1b 3 6

Table: Prior Joint Distribution

Suppose player 1 learns that he is of type a.

What will be his posterior belief about player 2’s type?
Using Bayes’ rule we have,
1
Pr{t1a ∩ t2a } 1
p1 (t2a |t1a ) = = 6
= ,
Pr{t1a } 1
6 + 1
3
3
and similarly,
1
Pr{t1a ∩ t2b } 2
p1 (t2b |t1a ) = = 3
= .
Pr{t1a } 1
6 + 1
3
3
EC202, University of Warwick, Term 2 32 of 45
Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Correlated Types

t2a t2b
1 1
t1a 6 3
1 1
t1b 3 6

Table: Prior Joint Distribution

Note that each player has a 12 chance

of being either type a or type b.
But the types are not independent;
rather, they are negatively correlated.
It is only half as likely that they will be the same
as that they will differ.

EC202, University of Warwick, Term 2 33 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Strategies and Bayesian Nash Equilibrium

In static normal form games of complete information,

actions and strategies need not be distinguished,
since choices are made once and for all.
For games of incomplete information, however,
each player i ∈ N can be of several different types ti ∈ Ti ,
each of which may choose a different action from the set Ai .
Thus, a pure strategy for player i must specify
what each of player i’s types ti ∈ Ti will choose
when Nature calls upon this type to act.
Similarly, a mixed strategy is a probability distribution
over a player’s pure strategies.

EC202, University of Warwick, Term 2 34 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Strategies Are . . .

Definition
Consider a static Bayesian game with private values

G ∗ = hN, (Ai )ni=1 , (Ti )ni=1 , ((ui (·; ti )ti ∈Ti )ni=1 , (pi )ni=1 i.

A pure strategy for player i is a function si : Ti → Ai

that specifies a pure action si (ti ),
which is what i will choose when his type is ti .
Similarly, a mixed behavioral strategy
is a function σi : Ti → ∆Ai that specifies a lottery σi (ti )
for each of i’s possible types ti ∈ Ti .

EC202, University of Warwick, Term 2 35 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

. . . Type Contingent

Thus, at the first stage,

even before Chance determines their types,
each player chooses a type-contingent strategy.
As a result of that Chance move,
players finally learn their types,
and then play according to their chosen strategy.
This offers a convenient way to model
players’ beliefs over other players’ strategies
when these depend on their different types.

EC202, University of Warwick, Term 2 36 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Strategies in the Entry Game

In the Entry Game example, the incumbent firm 2
has two possible types r and c (for “rational” and “crazy”).
There are two information sets, labelled 2r and 2c,
so 22 = 4 type-dependent strategies.
Player 1 has beliefs about the type distribution
that are given by Nature’s probability distribution.
Together with a specific strategy of player 2,
these probabilities determine player 1’s beliefs
over different continuation paths in the game,
and the resultant outcomes.
So the formulation allows players to evaluate
their expected utility from their alternative choices.
In the Entry Game example, let t2k denote player 2’s type,
with k ∈ {r , c}.

EC202, University of Warwick, Term 2 37 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Expected Utility in the Entry Game

Suppose player 1 believes that player 2 is using the pure strategy
(
A if t2 = t2r
s2 (t2 ) =
F if t2 = t2c

This simple example has only one type of player 1,

but we still denote it by t1 .
From player 1’s own perspective (with some redundant notation),
the expected utility from playing E will be

Eu1 (E , s2 (·); t1 ) = p1 (t2r |t1 ) u1 (E , s2 (t2r ); t1 )

+ p1 (t2c |t1 ) u1 (E , s2 (t2c ); t1 )
= p · 1 + (1 − p) · (−1) = 2p − 1

When p = 32 this yields Eu1 (E , s2 (·); t1 ) = 13 ,

which is the payoff entry from the pure strategy pair (E , Ar Fc )
in the bimatrix that represents this game.
EC202, University of Warwick, Term 2 38 of 45
Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

First observation:

If player i uses a pure (type-dependent) strategy,

while Chance chooses player i’s type,
for other players j 6= i it is as if player i uses a mixed strategy.
Again, in the Entry Game example,
suppose player 2 uses the strategy Ar Fc .
For player 1 it is as if player 2
chooses A with probability p
and F with probability 1 − p.

EC202, University of Warwick, Term 2 39 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Second observation:
As in an extensive form game,
we effectively specify player i’s action
at all its information sets, one for each type.
Thus, even though player i knows its realized type,
we specify “counterfactually” what i would have done
even for those types that have not been realized
(and never will be).
This is necessary so that players other than i
can form well defined beliefs over player i’s behavior.
Namely, players j 6= i need to combine
their posterior beliefs over i’s types
with their beliefs over
what each type ti of player i would want to do.

EC202, University of Warwick, Term 2 40 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Bayesian Nash Equilibrium Defined

In the static Bayesian game
G ∗ = hN, (Ai )ni=1 , (Ti )ni=1 , ((ui (·; ti )ti ∈Ti )ni=1 , (pi )ni=1 i,
a strategy profile s ∗ = (s1∗ (·), s2∗ (·), ..., sn∗ (·))
is a pure strategy Bayesian Nash Equilibrium (BNE)
if for every player i ∈ N,
and for each realization ti ∈ Ti of player i’s type,
the action ai = si∗ (ti ) is a best response because it solves
X
∗
max pi (t−i |ti ) ui (ai , s−i (t−i ); ti ).
ai ∈Ai t−i ∈T−i
∗ (t ) = (s ∗ (t ))
Note that s−i −i j j j6=i
is the other players’ equilibrium strategy profile
Q
when their type profile is t−i = (tj )j6=i ∈ j6=i Tj .
Note too that in the maximand we could replace ai
with any type-contingent deviation s̃i (·)
without affecting the result.
EC202, University of Warwick, Term 2 41 of 45
Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Conditional Expected Utility

Remark
In the definition of Bayesian Nash equilibrium, we can write
the best response condition more generally and concisely
as requiring each si∗ (ti ) to solve
∗
max Et−i [ui (ai , s−i (t−i ); ti )|ti ].
ai ∈Ai

We take player i’s conditional expectations Et−i [·|ti ]

over the random realizations of t−i
given that player i knows his own type ti .
The expectation is of ui (a; ti )
when player i’s component ai of strategy profile a is chosen freely,
but the other players j 6= i
choose their equilibrium type dependent strategies sj∗ (tj ).

EC202, University of Warwick, Term 2 42 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Multiple Integration

The notation Et−i [·|ti ] is more general

in the sense that it applies even when one or more players
have a continuum of possible types.
For instance, each player may have
an interval Ti = [t i , t̄i ] ⊆ R as the type space,
with the random type drawn from Ti
with cumulative distribution Fi (ti )
and density function fi (ti ) = Fi0 (ti ) (if Fi is differentiable).
In this case player i’s expected utility will be an integral
(more precisely, a multiple integral in n − 1 dimensions)
over the realizations of the other players’ types,
and the corresponding actions specified by their strategies.

EC202, University of Warwick, Term 2 43 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Reflections on Equilibrium

Hence, a Bayesian Nash equilibrium has each player

choose a type-contingent strategy si∗ (·) so that,
given any one of his types ti ∈ Ti ,
∗ (·),
and his beliefs about other players’ strategies s−i
∗
his conditional expected utility from si (ti )
is no less than from any other action ai ∈ Ai .
The conditional probability distribution is determined
from the other players’ type contingent strategies,
and by the Chance moves that each player faces
through the conditional beliefs pi (t−i |ti ).

EC202, University of Warwick, Term 2 44 of 45

Introduction Incomplete Information Strategies Bayesian Games Posterior Beliefs Bayesian Equilibrium

Zero Probability Events

Remark
Bayes’ Rule gives an ill defined answer
if there is a zero probability event
that appears in the denominator
of the formula for a conditional probability.
This matters little for now,
but is crucial when one considers sequential rationality
in games of incomplete information.

EC202, University of Warwick, Term 2 45 of 45