Lecture 1 - Mechanism Design

Focus of first two lectures: what is the best way to allocate an object? We look at situations
that are multi-agent extensions of the theory of screening.
This lecture looks at the most abstract setting, later we will focus on auction mechanisms. This
lecture is based on K, chapter 5. More specifically, the mechanism we look at in this lecture need
not be universal, that is, its rules can vary depending on the object for sale (via the distributions
of buyers’ values) nor it needs to be anonymous (we look at mechanisms that can treat different
buyers differently). Auctions are instead anonymous and universal.
The allocation of an indivisible object is only one example of situations in which the design
of mechanisms is a relevant tool. The example of the allocation of a private good captures a
general theme: it is hard to find mechanisms compatible with individual incentives that ensure
(a) efficient decisions, (b) voluntary participation of individuals, (c) balanced transfers.
Setting
A seller has an indivisible object to sell. The seller does not attach any value to the good. The
seller is risk-neutral and wants to maximize the revenue from selling the good. There are N
risk-neutral (potential) buyers, from set N = {1, .., N }. Buyers have no budget constraints.
Buyers have private and independently distributed values.1 Buyer i’s value Xi is distributed
over the interval X i := [0, ωi ] according to CDF Fi with PDF fi . We assume fi (xi ) > 0 over
the entire interval X i .
Let X := ×N
i=1 Xi and X−i := ×j6=i Xj . Let f (x) be the joint density of x := (x1 , .., xN ).

Independence ensures
f (x) = f1 (x1 ) × f2 (x2 ) × ... × fN (xN ).

Define f−i (x−i ) to be the joint density of x−i = (x1 , ..., xi−1 , xi+1 , .., xN ).
A selling mechanism (B, π, µ) is composed of

• a set of possible messages Bi for each buyer,

1Here “private” does not mean only privately known (although values ARE privately known). Private refers to
the fact that buyer i’s utility from the object depends only on her value Xi .
1
 2

• an allocation rule π : B → ∆, where ∆ is the set of probability distributions over N ,

(let πi (b) denote the probability that buyer i gets object (where b ∈ B))
• a payment rule µ : B → RN (let µi (b) denote the payment that i must make to the
seller).

Buyers have quasi-linear preferences: individual i’s utility is given by:

πi (b)xi − µi (b).

Every mechanism determines a game of incomplete information among the buyers. N strategies
βi : [0, ωi ] → Bi are part of an equilibrium of a mechanism if for all i and all xi , given β−i ,
βi (xi ) maximizes i’s expected payoff. We look for Bayes Nash Equilibria. An allocation rule π
is efficient if:

X X
πi (b)xi ≥ πi0 (b)xi ,
N N
0
for any allocation π .

The Revelation Principle.

Definition. A mechanism is called a direct mechanism if Bi = Xi . We refer to a direct mech-

anism as (Q, M) where Qi (x) is the probability that i will get the object and Mi (x) is the
expected payment by i.
In a direct mechanism, buyers are asked to simultaneously and independently report their types.

Proposition. (Revelation Principle) Given a mechanism and an equilibrium for that mech-
anism, there exists a direct mechanism in which (i) it is an equilibrium for each buyer to report
his or her value truthfully and (ii) the outcomes are the same as in the given equilibrium of the
original mechanism.

To see that the proposition holds, let Q : X → ∆ and M : X → RN be defined as: Q(x) =
π(β(x)) and M(x) = µ(β(x)) for all x.
 3

Incentive Compatibility. For a direct mechanism (Q, M) let:

ˆ
qi (zi ) = Qi (zi , x−i )f−i (x−i )dx−i ,
X−i
ˆ
mi (zi ) = Mi (zi , x−i )f−i (x−i )dx−i .
X−i

qi (zi ) is the conditional expected value of the probability that agent i obtains the good, condi-
tioning on buyer i’s reporting her type to be zi . mi (zi ) is the conditional expected value of the
transfer that agent i makes to the seller, conditioning on agent i reporting her type to be zi .

Definition. A direct mechanism is incentive compatible if ∀i ∈ N and ∀xi ∈ X i and ∀zi ∈ X i :

Ui (xi ) := qi (xi )xi − mi (xi ) ≥ qi (zi )xi − mi (zi ).

We call Ui the equilibrium payoff function.

Note that qi (zi )xi − mi (zi ) is an affine (hence convex) function of xi .
Hence, as incentive compatibility requires

Ui (xi ) = max {qi (zi )xi − mi (zi )}

zi ∈Xi

then:
1) incentive compatibility implies convexity of Ui (xi ) (the maximum over a set of convex
functions is a convex function);
2) convex functions are not differentiable in at most countably many points. Consider any
xi for which Ui (·) is differentiable. Let δ > 0. By incentive compatibility we have:

Ui (xi + δ) − Ui (xi ) (xi + δ)Ui (xi ) − mi (xi ) − (xi Ui (xi ) − mi (xi ))

lim ≥ lim = qi (xi ),
δ→0 δ δ→0 δ
Ui (xi ) − Ui (xi − δ) (xi )Ui (xi ) − mi (xi ) − ((xi − δ) Ui (xi ) − mi (xi ))
lim ≤ lim = qi (xi ),
δ→0 δ δ→0 δ
 4

thus
Ui0 (xi ) = qi (xi );

3) as any continuous function is the definite integral of its derivative, we have:

ˆ xi
(1) Ui (xi ) = Ui (0) + qi (ti )dti .
0

Using the definition of Ui (xi ), (1) can be used to characterize the conditional expected transfer
mi (xi ):

ˆ xi
Ui (xi ) = qi (xi )xi − mi (xi ) = −mi (0) + qi (ti )dti
0

↔
ˆ xi
(2) mi (xi ) = qi (xi )xi + mi (0) − qi (ti )dti
0

As the shape of the payoff function is entirely determined by the allocation rule Q, the payment
rule is determined by the IC constraint.
4) If a direct mechanism is incentive compatible, then qi (·) is non-decreasing.
To see this, consider 2 types xi , zi such that xi > zi . Incentive compatibility requires:

qi (xi )xi − mi (xi ) ≥ qi (zi )xi − mi (zi ),

qi (zi )zi − mi (zi ) ≥ qi (xi )zi − mi (xi ).

Subtracting the two inequalities: qi (xi )xi − mi (xi ) − (qi (xi )zi − mi (xi )) ≥ qi (zi )xi − mi (zi ) −
(qi (zi )zi − mi (zi )) ↔qi (xi ) (xi − zi ) ≥ qi (zi ) (xi − zi ) ↔ qi (xi ) ≥ qi (zi ).

Individual Rationality. In many (but not all) applications, it makes sense to assume that
potential buyers participate in the mechanism after learning their types, so individual rationality
requires that for all i and xi we have Ui (xi ) ≥ 0.

Definition. A direct mechanism is individually rational if ∀i ∈ N and ∀xi ∈ X i :

 5

Ui (xi ) ≥ 0.

By equation (1) we see that it is sufficient to check that Ui (0) ≥ 0 (and since Ui (0) = −mi (0),
this is equivalent to mi (0) ≤ 0).

Optimal Mechanisms
Optimal mechanisms = mechanisms that maximize the seller expected revenue, subject to IC
and IR constraints. We focus on direct mechanisms.
The expected revenue of seller is:

X
E [R] = E [mi (Xi )]
i∈N

ex-ante expected payment of buyer i:

ˆ ωi
E [mi (Xi )] = mi (xi )fi (xi )dxi
0

Using (2):

ˆ ωi ˆ ωi ˆ xi
E [mi (Xi )] = mi (0) + qi (xi )xi fi (xi )dxi − qi (ti )fi (xi )dti dxi .
0 0 0

Switching the order of integration, the last term can be written:

ˆ ωi ˆ xi ˆ ωi ˆ ωi
qi (ti )fi (xi )dti dxi = qi (ti )fi (xi )dxi dti
0 0 0 ti
ˆ ωi ˆ ωi
= qi (ti ) fi (xi )dxi dti
0 ti
ˆ ωi
= qi (ti )(1 − Fi (ti ))dti
0

So:
 6

ˆ ωi ˆ ωi
E [mi (Xi )] = mi (0) + qi (xi )xi fi (xi )dxi − qi (xi )(1 − Fi (xi ))dxi
0 0
ˆ ωi  
1 − Fi (xi )
= mi (0) + qi (xi )fi (xi ) xi − dxi
0 fi (xi )
´
Using qi (xi ) = X−i
Qi (x)f−i (x−i )dx−i , and (f (x) = f1 (x1 ) × f2 (x2 ) × ... × fN (xN )), we have:

ˆ ωi ˆ  
1 − Fi (xi )
E [mi (Xi )] = mi (0) + Qi (x)f (x)dx−i xi − dxi
0 X−i fi (xi )
ˆ  
1 − Fi (xi )
= mi (0) + Qi (x)f (x) xi − dx
X fi (xi )

Where we are using:

ˆ ˆ ˆ ˆ
g(x)dx = .. g(x)dxn ..dx2 dx1 .
X x1 x2 xn

The objective of the seller is to find a mechanism that maximizes

X X Xˆ  1 − Fi (xi )

E [mi (Xi )] = mi (0) + xi − Qi (x)f (x)dx.
i∈N i∈N i∈N X f i (xi )

subject to IC and IR constraint.

1−Fi (xi )
Definition. Let ψi (xi ) := xi − fi (xi )
be called the virtual valuation of a buyer with value xi .
A design problem is regular if for all i function ψi is increasing in the true value xi .

1 fi
Note that since ψi (xi ) := xi − λi (xi )
,where λi := 1−Fi
is the hazard rate of Fi , a sufficient
condition for regularity is that λi is non-decreasing (λi is non-decreasing for the uniform, normal,
exponential distributions, among others). From now on, we assume regularity.
The function that the seller maximizes can be re-written as:

X ˆ X
(3) mi (0) + (ψi (Xi )Qi (x)) f (x)dx.
i∈N X i∈N
 7

Hence we observe:

(4) Qi (x) > 0 ↔ ψi (xi ) = max ψj (xj ) ≥ 0

j∈N

to the object is assigned with probability 1 to one of the buyers with the highest virtual valuation,
as long as the virtual valuation is non-negative.

ˆ xi
(5) Mi (x) = Qi (x)xi − Qi (zi , x−i )dzi
0

Conditions (4) and (5) together define an optimal mechanism. To get a more intuitive interpre-
tation:

yi (x−i ) = inf {zi : ψi (zi ) ≥ 0 and ∀j 6= i, ψi (zi ) ≥ ψj (xj )}

thus (4) is equivalent to:



1

if zi > yi (x−i )
Qi (zi , x−i ) =

0
 if zi < yi (x−i )

thus


ˆ xi


 xi − yi (x−i ) if xi > yi (x−i )
Qi (zi , x−i )dzi =
0 
0
 if xi < yi (x−i )

and therefore:



yi (x−i ) if Qi (x) = 1


Mi (x) =

0
 if Qi (x) = 0

This mechanism is not efficient:

 8

1) the mechanism requires the seller to keep the good if all virtual valuations are negative, but
as all buyers have positive valuations and seller has valuation 0, efficiency requires that the
seller assigns the good to the seller with the highest valuation rather than keeping it.
2) the object is allocated to the buyer with the highest virtual valuation, and this needs not be
the agent with the highest value (though the two correspond in case of identical distributions
of valuations).
In the setup considered so far, a second-price auction is efficient: in a second-price auction it is
optimal for every buyer to report her type truthfully, and the object is allocated to one of the
buyers with the highest report, so with the highest valuation.

Symmetric Example
Suppose N = 2, ωi = 1, Xi is distributed uniformly, for i = 1, 2.
1−F (xi )
Thus: ψi (xi ) = xi − xi
= 2xi − 1. Note that the design problem is regular.
The revenue maximizing mechanism does not sell to any buyer if for both i = 1 and i = 2:
ψi (xi ) < 0 ↔ xi < 21 .
If the good is sold, it is sold to the buyer with the highest virtual valuation, that is, with the
buyer with the highest valuation.
1
It can be checked that a first or second price auction with reverve price 2
will implement this
mechanism (not that in this case the efficient mechanism is a first or a second price auction
with reserve price 0).
Asymmetric Example
Suppose N = 2, ω1 = ω2 = 1, and F1 (x1 ) = x21 , F2 (x2 ) = 2x2 − x22 .
Thus ψ1 (x1 ) = 32 x1 − 1
2x1
and ψ2 (x2 ) = 32 x2 − 21 .
The revenue maximizing mechanism does not sell to any buyer if
q
ψ1 (x1 ) < 0 ↔ x1 < 13 ; ψ2 (x2 ) < 0 ↔ x1 < 31 .
If the good is sold, it is sold to buyer 1 if ψ1 (x1 ) > ψ2 (x2 ) ↔ x2 < x1 + 13 − 1
3x1

Efficient Mechanisms
Consider a more general setup in which Xi = [αi , ωi ] ⊂ R for each agent.
 9

An allocation rule Q∗ : X → ∆ is efficient if for all x ∈ X ,

X
Q∗ (x) ∈ arg max Q j xj .
Q∈∆
j∈N

That is, an efficient allocation rule allocates the object to one of the agents with the highest
valuation.
The maximized value of the social welfare is defined as:

X
W (x) = Q∗j (x)xj .
j∈N

The VCG Vickrey-Clarke-Groves Mechanism. The VCG mechanism is an efficient mech-

anism (Q∗ , M V ) where:

MiV (x) = W (αi , x−i ) − W−i (x)

for

X
W−i (x) = Q∗j (x)xj .
j6=i

With αi = 0, the VCG mechanism corresponds to a second price auction. Note that the
mechanism can be thought of as a “pivotal mechanism”:

• Say for reports x the object is not assigned to agent i. Then W (αi , x−i ) = W−i (x), thus
MiV (x) = 0
• Say for reports x the object is assigned to agent i with some positive probability. Then
W (αi , x−i ) < W−i (x), thus MiV (x) < 0.

Thus an agent pays only if she is pivotal, that is only if her presence changes the utility of the
other agents. In this case, the payment is exactly equal to the loss of value imposed by the
agent on the other agents.
The VCG is incentive compatible. As long as others report x−i , by reporting zi buyer i0 s payoff
is:
 10

X
Q∗i (zi , x−i )xi − M V (zi , x−i ) = Q∗j (zi , x−i )xj − W (αi , x−i )
j∈N

Q∗j (zi , x−i )xj is maximized by reporting zi = xi , while the term

P
Note that the term j∈N

−W (αi , x−i ) does not depend on the report, so reporting truthfully is optimal.
Using what we learned, as the mechanism is incentive compatible, we know that the equilibrium
payoff:

UiV (xi ) = E [W (xi , X−i ) − W (αi , X−i )]

is convex and increasing. Hence UiV (αi ) = 0 implies that IR holds.

Two observations:
1) Any mechanism that satisfies IC and IR and is efficient must have an M function with the
same shape as the VCG mechanism. Moreover UiV (αi ) = 0 ensures that the VCG mechanism
maximizes expected payments among the efficient mechanisms that satisfy IC and IR.
2) A mechanism balances the budget (ex-post) if for all x:

X
Mi (x) = 0.
N

VCG is not budget balanced ex post.

An example of efficient mechanism that satisfies ex post budget balance is the Arrow- d’Aspremont-
Gerard-Varet (or AGV) mechanism (Q∗ , M A ), defined by:

1 X
MiA (x) =


Ex−j [W−j (xj , X−j )] − Ex−i [W−i (xi , X−i )] .
N − 1 j6=i

So:
 11

!
X X 1 X
MiA (x)


= Ex−j [W−j (xj , X−j )] − Ex−i [W−i (xi , X−i )] .
N i
N − 1 j6=i
 
P 1

= i (N − 1) Ex−i [W−i (xi , X−i )] − Ex−i [W−i (xi , X−i )] = 0.
N −1

The AGV mechanism is incentive compatible. To see this, suppose all other agents are reporting
x−i truthfully. The expected payoff to i from reporting zi when the true value is xi is equal to:

" #
1 X
Ex−i [Q∗i (zi , X−i )xi + W−i (zi , X−i )] − Ex−i Ex [W−j (xj , X−j )] .
N − 1 j6=i −j

The second term is independent of zi , while the first term is maximized by zi = xi . On the
other hand, the AGV mechanism may not satisfy the IR constraint.