Advanced Microeconomics

HKU Business School

Masaki Miyashita

Chapter 5:
Games with Incomplete Information
 Incomplete Information

• So far: Games with complete information

• The game structure is common knowledge across all players.

• Now: Games with incomplete information

• Players may not understand the partial/full details about the game in play.

• Each player may have private information that is not known to the others.
 Example 1: Cournot Duopoly

• Two rms i ∈ {1,2} simultaneously select their production levels qi ≥ 0.

• Firm i’s pro t is

ui = qi (a − (qi + qj)) − θiqi

where θi is private marginal cost, random & unknown to the opponent.

• The rm’s optimal production level qi may depend not only on θi but also on their
expectation about θj .

• This is because θj can affect the opponent’s choice of qj .

fi
fi
fi
 Example 2: First-price Auction

• n bidders compete for an indivisible good.

• Each bidder has a private valuation θi for the good, drawn independently and
uniformly from [0,1] .

• The realization of θi is known to i but not to others (including the auctioneer).

• Bidders simultaneously submit their bids bi ≥ 0.

• The highest bidder wins the good and pays own bid bi .

θi − bi if bi > maxj≠i bj,

{0
ui =
if bi < maxj≠i bj .

• The bidder’s optimal strategy may depend not only on θi but also on their
expectation about θj .
Incomplete Information Game
 Incomplete Information Game

• A random move by the nature at the beginning of the game.

• The nature determines a type of each player, which corresponds to a different

payoff function that the player might have.

• All players commonly know the ex-ante distribution of types.

• However, each individual type is private information; the player does not know
the realization of the opponent’s type.
 Incomplete Information Game

• N = {1,…, n} is a nite set of players.

• For each player i …

• Ai is a set of actions.

• Θi is a set of of types.

• ui is a payoff function,

ui : A × Θi → ℝ where A = A1 × ⋯ × An .

• μ ∈ Δ(Θ) is a probability distribution over type pro les.

• An incomplete information game is summarized as the pro le

G = ⟨N, (Ai, Θi, ui)i∈N, μ⟩ .

fi
fi
fi
 Incomplete Information Game

• Timeline:

• The nature randomly draws a type pro le θ = (θ1, …, θn) according to the
prior distribution μ .

• Each player is privately informed of her own type θi .

• Each player simultaneously chooses action ai from Ai .

Remark. Complete information games correspond to the special case of this model,
where | Θi | = 1 for all players.
fi
 Incomplete Information Game

• A player i’s (pure) strategy is any function si : Θi → Ai .

• Si denotes the set of i’s pure strategies.

• A strategy speci es a contingent plan detailing the player’s actions for each
type realization.

- We can also de ne a mixed strategy (though not used further): A mixed strategy is
any function σi : Θi → Δ(Ai) that assigns a lottery σi(θi) over actions to each type θi .

Remark: In incomplete information games, actions and pure strategies are distinct
concepts.
fi
fi
 Incomplete Information Game

• Given any strategy pro le s = (s1, …, sn), the player i ’s (ex-ante) expected
payoff is de ned as follows:

μ(θ)ui (s(θ), θi) .

∑
ui(s) =
θ∈Θ

• μ(θ) is the ex-ante probability that type pro le θ = (θ1, … . θn) realizes.

• s(θ) = (s1(θ1), …, sn(θn)) is the action pro le that results when type pro le θ
occurs, given each player j adopts strategy sj .
fi
fi
fi
fi
fi
 Bayesian Nash Equilibrium

• The notion of NE is extended to incomplete information games.

De nition.

• A strategy pro le s* = (s*

1
, …, s*
n ) forms a Bayesian Nash equilibrium

in incomplete information game G = ⟨N, (Ai, Θi, ui)i∈N, μ⟩ if

ui(s*
i
, s*
−i
) ≥ ui(si, s*
−i
), ∀i ∈ N, ∀si ∈ Si .
fi
fi
 Bayesian Nash Equilibrium

• In BNE, each player’s strategy maximizes their ex-ante expected payoff,

where a strategy in an incomplete information game is a contingent plan that
speci es the player’s action for each type realization.

• The following are equivalent:

• si maximizes the ex-ante expected payoff.

• si(θi) maximizes the conditional expected payoff for each type θi .

fi
 Bayesian Nash Equilibrium

• Bayes’ rule:
μ(θi, θ−i)
∑
μ(θ−i ∣ θi) = where μ(θi) = μ(θi, θ′−i) .
μ(θi) θ′−i∈Θ−i

Lemma (Equivalence between ex-ante & interim).

• A strategy pro le s* = (s*

1
, …, s*
n ) forms a Bayesian Nash equilibrium if

and only if for all i and θi (with μ(θi) > 0), s*

i
(θi) solves

μ(θ−i ∣ θi)ui(ai, s* (θ ), θi) .

∑
max −i −i
ai∈Ai
θ−i∈Θ−i


fi
 Bayesian Nash Equilibrium

Proof.

• The equivalence is con rmed by Bayes’ rule:

Ex-ante expected payo

μ(θi, θ−i) ⋅ ui (si(θi), s−i(θ−i), θi)

∑
(θi,θ−i)

μ(θi)μ(θ−i ∣ θi) ⋅ ui (si(θi), s−i(θ−i), θi)

∑
=
(θi,θ−i)

μ(θ−i ∣ θi) ⋅ ui (si(θi), s−i(θ−i), θi) .

∑ ∑
= μ(θi)
θi θ−i

Conditional expected payo

f
fi
f
Example 1: Cournot Duopoly
 Example 1: Cournot Duopoly

• The pro t function is given as

qi (a − (qi + qj)) − θiqi

where θi is the rm i’s marginal cost of production.

• Suppose θ1 = c for sure.

• Suppose θ2 takes cH = c + ϵ and cL = c − ϵ with equal probabilities.

Exercise. Derive the BNE in this Cournot duopoly under asymmetric information.
Compare the equilibrium output levels to the symmetric equilibrium output under
complete information. How does each rm’s expected pro t vary across ϵ?
fi
fi
fi
fi
 Example 1: Cournot Duopoly

• We can derive BNE by solving the interim expected payoff maximization for each
possible type.

• Take θi (known to i) and qi (i’s choice) out from the expectation:

max qi (a − (qi + [q*

j ))
] − θiqi .
q ≥0
i

• From the rm i’s point of view, [q*] is a constant.

• By taking FOC, we get the following best response:

a − θi − [q*
j ]
q*
i
(θi) = .
2
𝔼
𝔼
𝔼
fi
 Example 1: Cournot Duopoly

• Firm 1’s output level is deterministic:

a − c − [q*
2]
q*
1
= .
2
• Firm 2’s output level depends on the marginal cost:

a − cH − q*
1 a − cL − q*
1
q* (c ) =
2 H
and q*(c ) =
2 L
.
2 2

• By the assumption that c = 0.5cL + 0.5cH, we have q*

1
= [q*
2
].

• Consequently,
a−c ϵ
− if cH = c + ϵ,
{ a −3 c
a−c 3 2
q*
1
= and q*
2
= ϵ
3 + if cL = c − ϵ .
2
𝔼
𝔼
 Example 1: Cournot Duopoly
• The expected pro t is higher for Firm 2 than for Firm 1 —- Why?

• The uncertainty in θ2 impacts the pro ts of both rms through the market price,
while θ2 is the private information of Firm 2.

• Firm 2 can adjust the production more ef ciently than Firm 1 in response to θ2.
fi
fi
fi
fi
Example 2: Auction
 First-price Auction

• There are n bidders.

• Each bidder’s private valuation θi (= type) is independently and uniformly

distributed over [0,1] .

• By Lemma 1, a strategy pro le (b*

1
, …, b*
n ) constitutes a BNE if for every i and θi,

b*
i
(θi) solves
Gain on winning

( )
max (θi − bi) ⋅ Pr bi > max b*
j
(θj) .
bi≥0 j≠i

Winning probability

Exercise. Derive the BNE of the rst-price auction. Also, calculate the equilibrium
expected revenue to the seller and compare it to that in the second-price auction
(recall that truth-telling is weakly dominant in 2PA).
fi
fi
 First-price Auction

• Let us focus on a “symmetric” BNE, where every bidder adopts the same bidding
strategy f .

• While the actual bid b*

i
= f(θi) varies with each bidder’s type realization, the
dependency of b*
i
on type θi is common across all bidders.

• To derive the equilibrium strategy f, we use a guess-and-verify approach.

Guess: f is a linear function of θi .

Verify: Show that there is a constant c > 0 such that f(θi) = cθi constitutes BNE.
 First-price Auction

• Since bidders’ types are independently distributed,

( )
Pr bi > max f(θj) = Pr (bi > f(θj), ∀j ≠ i)
j≠i

Pr (bi > f(θj)) = Pr (bi > f(θj))

n−1

∏
= .
j≠i

• Since we guess f(θj) = cθj for each j,

Pr (bi > f(θj)) = Pr (bi > cθj) = Pr (θj < bi /c) = bi /c .

• Hence, the bidder i’s winning probability is

( ) (c)
bi n−1
Pr bi > max f(θj) = .
j≠i
 First-price Auction
• The bidder i’s payoff maximization problem:

(c)
n−1
bi
max (θi − bi) ⋅ .
bi≥0

• Taking FOC,

(c) (c)
n−1 n−2
bi (n − 1)(θi − bi) bi
− + = 0.
c

• Solving this equation, the optimal b*

i
is given by

n−1
b*
i
= ⋅ θi .
n
• Setting c = (n − 1)/n, our initial guess is veri ed!

• Thus, in the symmetric BNE, each bidder reports (n − 1)/n times the true type.
fi
 First-price Auction

• The seller’s expected revenue is

[ i=1,…,n ] [ i=1,…,n ]
n−1
max f(θi) = ⋅ max θi .
n
⋆

n
• Since θi ∼i.i.d. Unif[0, 1], we can calculate that ⋆ = n+1
.
n−1
• Therefore, the seller’s revenue is n+1
.

• The revenue is increasing in n . In other words, the seller gains from having
additional participants in the auction.
n−1
• In fact, the expected revenue from 2PA is also n+1
. More generally, the
revenue equivalence theorem holds: under certain conditions, all
standard auction formats yield the same expected revenue.
𝔼
𝔼
 Expected Value of Highest Type

Calculating ⋆.

• Let X = maxi=1,…,n θi .

• For each x ∈ [0,1], we have Pr (X < x) = x n .

• So, the cumulative distribution function (CDF) of X is FX(x) = x n .

• Moreover, the probability density function (PDF) of X is fX(x) = nx n−1 .

• Now, the expected value of X is calculated as

1 1 1
n+1
x n
∫0 ∫0
n
[X] = x ⋅ fX(x)dx = n x dx = n ⋅ = .
n+1 n+1
0

• Consequently, the seller’s revenue in 1PA is

n−1
.
n+1
𝔼
 Second-price Auction

• In 2PA, we have con rmed that the truth-telling strategy is optimal for each bidder
regardless of the behavior of the opponents.

• In particular, this implies f(θi) = θi forms the BNE in 2PA.

• Let us calculate the seller’s expected revenue in 2PA.

• Let Y denote the second highest type among n bidders.

• Since truth-telling constitutes an equilibrium, the seller’s revenue is equal to the

expected value of Y .
fi
 Second-price Auction

Step 1: Calculate the CDF of Y .

• Given any value y ∈ [0,1], observe that Y < y occurs if and only if either of the
following occurs:

1. All θi fall below y, which occurs with probability y n .

2. Exactly one θi exceeds y, which occurs with probability n(1 − y)y n−1 .

• Consequently,

FY (y) = Pr(Y < y) = y n + n(1 − y)y n−1 = ny n−1 + (n − 1)y n .

 Second-price Auction

Step 2: Calculate the expected value of Y .

• Differentiating FY , we get

fY (y) = n(n − 1)(y n−2 − y n−1) .

• Hence, it follows that

1 1

∫0 ∫0 ( )
n−1 n
[Y ] = y ⋅ fY (y)dy = n(n − 1) y − y dy

[n n + 1] (n n + 1) n + 1
yn y n+1 1 1 n−1
= n(n − 1) − = n(n − 1) − = .
0

• Therefore, the revenue equivalence holds between 1PA and 2PA.

𝔼
 2PA with Reserve Price

• Now, assume that the seller can set a reserve price r ∈ (0,1) .

• If all bids fall below r, then the good remains unsold.

• Otherwise, if at least one bidder bids above r , the highest bidder wins the
good and pays the higher of r or the highest bid among the losing bidders.

Exercise. Show that truth-telling is still weakly dominant. Then, calculate the
equilibrium expected revenue to the seller and discuss how it depends on r .
Speci cally, derive the seller-optimal reserver price r* for revenue maximization.
fi