You are on page 1of 6

Northwestern ECON 410-3: Spring 2015 Prof.

Jeff Ely
University TA Session 7 TAs: Arjada Bardhi
Egor Starkov

Incentive compatibility of a SCF


A suspect is in custody, accused of murder. If he goes to trial he will either be convicted or acquitted. If he is
convicted he will be sent to prison for life giving him a payoff of −1. If he is acquitted he goes free and has a
payoff of 0. The district attorney can offer plea bargains: allowing the defendant to plead guilty in return for a
lighter sentence. In particular, for any r ∈ (0, 1), the DA can offer a reduced sentence which, if accepted, would
give the defendant a payoff of −r.
The defendant is privately informed about his chances for acquittal at trial: θ ∈ [0, 1] is the defendant’s
privately known probability of acquittal. If the defendant does not enter into a plea bargain with the DA he will
go to trial and be convicted with probability 1 − θ.
Consider the mechanism design problem where the DA is the principal and the defendant is the agent. A
social choice function is a mapping f : [0, 1] → {trial} ∪ (0, 1) where f (θ) = trial means that type θ will go to
trial and f (θ) = r ∈ (0, 1) means that type θ accepts a plea bargain giving him a sentence with payoff −r.
1. Write down the inequalities that characterize whether a social choice function f is incentive-compatible for
the defendant.
2. What is the set of all incentive-compatible social choice functions?

Solution
By going to trial a defendant of type θ receives (expected) utility of −(1 − θ), while from accepting a plea bargain
his utility is −r. This implies that the IC constraint is f (θ) = r ⇒ −r ≥ −(1 − θ) & ∀θ0 6= θ f (θ0 ) ≥ f (θ)
We will characterize the set of IC social choice functions by a series of claims. For simplicity we will initially
assume that θ has full support on [0, 1], and then correct for other cases.
claim 1 f (·) has at most one value on the real line.
Proof: if f (θ1 ) < f (θ2 ) θ1 , θ2 ∈ [0, 1] then a defendant of type θ2 gains higher utility by declaring θ1 (as
−f (θ1 ) > −f (θ2 ). This implies the mechanism( is not IC for θ2 .
r if θ < θ̄
claim 2 f (·) has a cutoff at some θ̄. i.e. f (θ) = (value at θ̄ is not unique).
T if θ ≥ θ̄
Proof: assume θ0 > θ, f (θ) = T, f (θ0 ) = r. By IC for θ we know that −r ≤ −(1 − θ). However as
−(1 − θ0 ) > −(1 − θ) this implies that −(1 − θ0 ) > −r and we don’t have IC for θ0 .

claim 3 r = 1 − θ̄
Proof: consider IC for types close to θ̄: we have r ≤ 1 − (θ̄ − ) and r ≥ 1 − (θ̄ + ). As we push  to zero,
these imply the result.
These three claims imply that for any (r, θ̄) s.t. r = 1 − θ̄ the social choice function
(
r if θ < θ̄
f (θ) =
T if θ ≥ θ̄

is incentive compatible.
As long as the distribution of types has full support, the arguments above hold. However, if there is an
interval (a, b) with zero probability then when we set θ̄ = a any r ∈ [1 − b, 1 − a] can be used. This is the case
as claim 3 no longer holds (we don’t have types θ̄ + ), and we just need to get IC for type b. This multiplicity
of r occurs only if θ̄ is not in the interior of the support.

Page 1 of 6
Northwestern ECON 410-3: Spring 2015 Prof. Jeff Ely
University TA Session 7 TAs: Arjada Bardhi
Incentive compatibility of a SCF Egor Starkov

VCG mechanism: Bilateral trade


Consider the following buyer/seller framework. Let x ∈ [0, 100] = X denote the amount of a good that is
produced by a seller, which a buyer then consumes. Let θB , θS ∈ [1, 2] = ΘB = ΘS denote the types of the buyer
and seller, respectively. Types are independently distributed and privately known by each agent. Assume that
utility functions are quasilinear in money, and are given by

UB (x, t, θ) = θB x − tB
x2
US (x, t, θ) = − − tS
2θS

where t = (tB , tS ) denotes a vector of transfers that the agents pay. (If ti < 0, then agent i is receiving a
transfer.)
1. What is the efficient level of production x(θ)? What social surplus is attained by it?
2. Define a least charitable type of agent i to be any type θi ∈ Θi that minimizes the interim expected social
surplus, where the expectation is taken over others’ types. What are the least charitable types of each
agent?
3. Using the least charitable types as the default types, what are the VCG transfers? What are the buyer’s
and seller’s utilities in the mechanism? Does the mechanism run an ex post budget surplus for all θ? Is
the mechanism interim individually rational?
4. Show directly that, in the direct revelation game with the VCG transfer and allocation functions you
computed, it is weakly dominant for each player to truthfully reveal his type. That is, let US (θ̂S , θ̂B ; θS )
denote the seller’s utility when his type is θS , he announces θ̂S , and the buyer announces θ̂B . Show that
θS ∈ arg maxθ̂S US (θ̂S , θ̂B ; θS ) for all θ̂B , θS . Show the similar statement for the buyer. (Note: Every VCG
mechanism is dominant strategy incentive compatible, so this verifies in this special case something that
we know in general.)
5. Now instead of using the VCG mechanism to compute transfers, use the envelope theorem to compute
transfers that implement the efficient allocation. For what choices of UB (1, θS ) and US (1, θB ) do you get
the same transfers as computed with the VCG mechanism? What is the largest possible expected budget
surplus if interim individual rationality holds for all types?

Solution
2
x
1. It solves maxx θB x − 2θ S
2
, i.e. x(θ) = θB θS . The social surplus is θB θS − 21 θB
2
θS = 12 θB
2
θS .
2. θB = θS = 1.
3. If θB = 1, the seller sells x = θS and gets utility − 21 θS . For arbitrary θB , the seller’s utility is − 12 θB
2
θS .
1 1 2
 1 2
Therefore the transfer from the buyer is tB (θ) = − 2 θS − − 2 θB θS = 2 θS (θB − 1).
2
If θS = 1, the buyer gets x = θB and his utility is θB . For arbitrary θS , the buyer gets x(θ) = θB θS and
2 2 2 2
his utility is θB θS . Therefore the transfer from the seller is tS (θ) = θB − θB θS = θB (1 − θS ).
The budget surplus is tB (θ) + tS (θ) = 21 θS (θB
2 2
− 1) + θB 2
(1 − θS ) = θB (1 − 21 θS ) − 21 θS .
The buyer’s utility is UB (θB , θS ) = θB (θB θS ) − 12 θS (θB
2
− 1) = 1 2
2 θS (θB + 1) > 0. The seller’s utility is
2
US (θB , θS ) =− (θB2θθSS ) 2
− θB (1 − θS )
= 2 1
− 1) ≤ 0. So ex post IR is satisfied for the buyer only which
θB ( 2 θS
implies interim IR for the buyer as well. Interim IR for the seller is violated.
4. If a seller of type θS announces θ̂S , his payoff is

(θ̂B θ̂S )2 2
US (θ̂S , θ̂B ; θS ) = − − θ̂B (1 − θ̂S ).
2θS

VCG mechanism: Bilateral trade continued on next page. . . Page 2 of 6


Northwestern ECON 410-3: Spring 2015 Prof. Jeff Ely
University TA Session 7 TAs: Arjada Bardhi
VCG mechanism: Bilateral trade (continued) Egor Starkov

2
θ̂B θ̂S 2
First order condition with respect to θ̂S is − θS + θ̂B = 0. Solving it we obtain θ̂S = θS .
If a buyer of type θB announces θ̂B , his payoff is

1 2
UB (θ̂B , θ̂S ; θB ) = θB θ̂B θ̂S − θ̂S (θ̂B − 1).
2

First order condition with respect to θ̂B is θB θ̂S − θ̂S θ̂B = 0. Solving it we obtain θ̂B = θB .
5. Let θ = (θB , θS ). The envelope theorem says that
Z θB
∂vB (x(z, θS ), z)
tB (θ) = vB (x(θ), θB ) − UB (1, θS ) − dz.
1 ∂z
Z θB
1 2 1
= θB (θB θS ) − UB (1, θS ) − zθS dz = θ θS + θS − UB (1, θS )
1 2 B 2

Similarly
Z θS
∂vS (x(z, θB ), z)
tS (θ) = vS (x(θ), θS ) − US (1, θB ) − dz.
1 ∂z
Z θS
1 2 1 2 2 1 2
= − θB θS − US (1, θB ) − θB dz = −θB θS + θB − US (1, θB )
2 1 2 2

For tB (θ) = tVB CG (θ) and tS (θ) = tVS CG (θ), we need UB (1, θS ) = θS and US (θB , 1) = − 21 θB
2
.
Now we need to find transfers that maximize the expected budget surplus (with respect to θB , θS ) subject
R2 R2
to interim IR. So, we need 1 UB (1, θS )dθS = 1 US (θB , 1)dθB = 0. This gives us (assuming uniform
distribution for θS and θB ):
Z 2
7 1
tS (θB , θS )dθB = ( − θS )
1 3 2
Similarly, Z 2
3 1 2 1
tB (θB , θS )dθS = ( θ + )
1 2 2 B 2
Expected budget surplus (with respect to θB and θS ) is:
Z 2 Z 2 Z 2 Z 2
EθB ,θS (tS + tB ) = tS (θB , θS )dθB dθS + tB (θB , θS )dθS dθB
1 1 1 1

Z 2 Z 2
7 1 3 1 2 1 7 10 1
= ( − θS )dθS + ( θ + )dθB = − + =
1 3 2 1 2 2 B 2 3 4 6
As it was suggested in section, one could (mathematically) think of this interim IR constraint as an ex-post
IR constraint of the form: US (θB , 1) = UB (1, θS ) = 0; the maximum expected budget surplus from that
reasoning is 16 as well.

DSIC public good mechanisms


An economy consists of two individuals. An indivisible good can be produced at a cost c = 1. The good is a
public good so that both individuals will consume the good if it is produced. The consumption value to individual
i is denoted vi , for i = 1, 2. It is common-knowledge that the values (v1 , v2 ) are independently and uniformly
distributed on the unit interval.

DSIC public good mechanisms continued on next page. . . Page 3 of 6


Northwestern ECON 410-3: Spring 2015 Prof. Jeff Ely
University TA Session 7 TAs: Arjada Bardhi
DSIC public good mechanisms (continued) Egor Starkov

A mechanism is deterministic if the allocation rule is non-stochastic, i.e. Q : [0, 1]2 → {0, 1}. The good is
produced at profiles v for which Q(v) = 1 and the good is not produced when Q(v) = 0. A mechanism is ex post,
weakly budget balanced if t1 (v) + t2 (v) ≥ c for all v at which Q(v) = 1. Characterize all of the mechanisms that
are deterministic, dominant-strategy incentive-compatible, and ex-post weakly budget-balanced. (The analysis
is very similar to the analysis done in class of DSIC bilateral-trade mechanisms where we found that all such
mechanisms were equivalent to a fixed price.)

Solution
Given the problem setup, letting v = (v1 , v2 ), the constraints are:

No deficit (N D) : t1 (v) + t2 (v) ≥ cQ(v) ∀vi , v̂i , v−i


DSIC : vi Q(vi , v−i ) − ti (vi , v−i ) ≥ vi Q(v̂i , v−i ) − ti (v̂i , v−i ),
IR : vi Q(vi , v−i ) − ti (vi , v−i ) ≥ 0, ∀vi , v−i

1. ND + IR imply that Q(v) = 0 ⇒ t(v) = (0, 0).


Proof : if Q(v) = 0 then from ND t1 (v) + t2 (v) ≥ 0 while from IR −ti (v) ≥ 0. This is only possible if
ti (v) = 0.
2. ND + IR imply that Q(v) = 1 ⇒ v1 + v2 ≥ c.
Proof : If Q(v) = 1 then from IR vi ≥ ti (v). ND is t1 (v) + t2 (v) ≥ c. Combining the two yields the result.
3. DSIC implies that if vi , v−i and vi0 are such that Q(vi , v−i ) = Q(vi0 , v−i ) = 1 then it must be that
ti (vi , v−i ) = ti (vi0 , v−i ).
Proof : Suppose that vi and vi0 are such that Q(vi , v−i ) = Q(vi0 , v−i ) = 1. Then by DSIC vi − ti (vi , v−i ) ≥
vi − ti (vi0 , v−i ). However, DSIC for vi0 implies vi0 − ti (vi0 , v−i ) ≥ vi0 − ti (vi , v−i ). The two together imply
ti (vi , v−i ) = ti (vi0 , v−i ).
4. DSIC + IR imply that Q(v) = 1 ⇒ t1 (v) = v1 (v2 ) and t2 (v) = v2 (v1 ). Where vi (v−i ) = inf{vi : Q(vi , v−i ) =
1}.
Proof : Suppose not. Specifically suppose WLOG that t1 (v) 6= v1 (v2 ). Then there are two cases to consider:
• If t1 (v) < v1 (v2 ) then consider a type of P1 v1 ∈ (t1 (v), v1 (v2 )). Now note that u1 (v1 , v2 ) = 0 since
v1 < v1 (v2 ) and therefore trade will not occur and so the transfer will also be zero. However, consider
if player 1 lied and announced v1 (v2 ). Then trade will occur by definition and so u1 (v1 (v2 ), v2 ) =
v1 (v2 ) − t(v) > 0 using claim 3 and by assumption. Thus the agent would lie and so this violates
DSIC.
• If t1 (v) > v1 (v2 ) then suppose that v1 ∈ (v1 (v2 ), t1 (v)). Now note that u1 (v1 , v2 ) = v1 − t1 (v) < 0.
This, however violates IR.
Thus we have shown that Q(v) = 1 ⇒ ti (v) = vi (v−i )
5. DSIC implies that vi (v−i ) ≡ sup{vi : Q(vi , v−i ) = 1} = 1. (A.k.a., the regions where trade occurs must be
monotone.)
Proof : Suppose not. Suppose that vi (v−i ) < 1. Now consider a player such that vi > vi (v−i ). If this
player announces vi then ui (vi , v−i ) = 0 since trade does not occur at vi . However suppose if this player
instead announced his type as vi (v−i ). Then his payoff would be ui (vi (v−i ), v−i ) = vi − vi (v−i ) > 0 since
by definition of sup and inf vi (v−i ) ≥ vi (v−i ). Therefore this player could benefit by lying and so DSIC is
violated.
Combining all of our claims we see that the region in which trade occurs must be a monotone set, completely
contained in a rectangle that is completely contained above the line where v1 + v2 = c. The monotone set comes
from claim 5, the fact that the set must be above the line v1 + v2 = c comes from claim 2, and the fact that it

Page 4 of 6
Northwestern ECON 410-3: Spring 2015 Prof. Jeff Ely
University TA Session 7 TAs: Arjada Bardhi
DSIC public good mechanisms Egor Starkov

must be contained in a rectangle above this line comes from combining claim 4 with the ND constraint.

Gibbard-Satterthwaite Theorem
Definition: A decision rule α(θ) is DSIC if the social choice function f (θ) = (α(θ), t0 (θ)) is DSIC, where t0 (θ)
is a zero transfer function for all θ.

Definition: A decision rule α(θ) is dictatorial if there exists an agent i ∈ N such that α(θi , θ−i ) ∈ argmaxx∈X vi (x, θi )
for all θ = (θi , θ−i ).

Theorem: Suppose that X is finite and the type spaces include all possible strict orderings on X. A decision
rule α with at least three elements in its range is dominant strategy incentive compatible if and only if it is
dictatorial.

Prove the theorem for the case of N = 2 and X = {a, b, c}.

The type space will be: Θ = Θ1 × Θ2 = {abc, cab, bac, bca, cab, cba} × {abc, cab, bac, bca, cab, cba}. By assump-
tion Range(α) = {a, b, c}. We will first prove by construction that if α is DSIC, then it must be dictatorial. In
the following table we assign an outcome value α for every type profile:

abc acb bac bca cab cba


abc
acb
bac
bca
cba
cab

The rows denote the type of P1 and the columns denote the type of P2. First, notice that in order for α to
be DSIC, it must be that whenever both players agree on their top alternative, it assigns that alternative as the
outcome. If not, then both players can agree to misreport in such a way that they get their top alternative. Then,
α(abc, abc) = α(abc, acb) = α(acb, abc) = α(acb, acb) = a, α(bac, bac) = α(bac, bca) = α(bca, bac) = α(bca, bca) =
b, α(cab, cab) = α(cab, cba) = α(cba, cab) = α(cba, cba) = c.
Second, let us consider α(abc, bca). Notice that α(abc, bca) 6= c; otherwise, P1 has an incentive to misreport
as bac and get b instead of c. So, α(abc, bca) ∈ {a, b}. For the rest of this analysis, let α(abc, bca) = a and we
will show that P1 is the dictator in this case. The case α(abc, bca) = b ensures that P2 is the dictator, by a very
similar argument.
If α(abc, bca) = a, then α(abc, bac) = a = α(abc, bac) (by the IC for P2) and α(acb, bca) = a (by the IC for
P1). The IC for P2 ensures also that α(abc, cab) = α(abc, cba) = α(acb, cab) = α(acb, cba) = a. So far, the table
should look like:

abc acb bac bca cab cba


abc a a a a a a
acb a a a a a a
bac b b
bca b b
cba c c
cab c c

Gibbard-Satterthwaite Theorem continued on next page. . . Page 5 of 6


Northwestern ECON 410-3: Spring 2015 Prof. Jeff Ely
University TA Session 7 TAs: Arjada Bardhi
Gibbard-Satterthwaite Theorem (continued) Egor Starkov

Consider now α(bac, cba). If it is c, then P1 with bac has incentives to misreport acb. If it is a, then P2 with
type cba has incentives to misreport as bca. So α(bac, cba) = b. But α(bca, cba) = b also so that P1 with type bca
has no incentives to misreport as bac. It must also be that α(bac, cab) = α(bca, cab) = b so that P1 of type abc
or acb or cba has no incentives to misreport as cab. For P2 of type cab not to deviate and misreport abc or acb,
we need also α(bac, abc) = α(bac, acb) = α(bca, abc) = α(bca, acb) = b. This leaves us with:

abc acb bac bca cab cba


abc a a a a a a
acb a a a a a a
bac b b b b b b
bca b b b b b b
cba c c
cab c c

Now consider α(cab, bca). It cannot be a because P2 of type bca would misreport as cab and it cannot be
b because then P1 of type cab would misreport as abc or acb. So α(cab, bca) = c. For P2 of type cba not to
misreport as cab, we need α(cba, bca) = c as well. Notice also that α(cba, bac) 6= a so that P1 of type cba
does not misreport as bca and α(cba, bac) 6= b so that P2 of type bca does not misreport as bac. Similarly
α(cba, abc) = α(cba, acb) = c. By the IC for P1, it must also be that α(cab, abc) = α(cab, acb) = α(cab, bac) = c.
So the table becomes:

abc acb bac bca cab cba


abc a a a a a a
acb a a a a a a
bac b b b b b b
bca b b b b b b
cba c c c c c c
cab c c c c c c

For the α we constructed so that DSIC is satisfied for any type profile, P1 is the dictator. By a symmetric
argument, had we let α(abc, bca) = b, P2 would have been the dictator.
We have shown that if α is DSIC, then it is dictatorial. Let us now show that if α is dictatorial, then it is
DSIC. Let i be the dictator in α. Then α(θi , θ−i ) always assigns as the outcome the top alternative in θi , so i
has no incentives to misreport θi . On the other hand, if player −i misreport θ−i as θ̂−i , α(θi , θ−i ) = α(θi , θ̂−i .
So misreporting of −i does not alter the outcome assigned by α as that outcome depends only on θi . Hence, −i
does not have incentives to misreport either. So, α is DSIC.

Page 6 of 6

You might also like