Contract Theory

Contents

1 Hidden Information: Screening 6

1.1 A Simple Model of Price Discrimination . . . . . . . . . . . . . . . . . . . 8

1.1.1 Full Information Benchmark: First-Best Outcome or Perfect Price

Discrimination . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1.1.2 Asymmetric Information: Linear Pricing and Two-Part Tariffs . . . 9

1.1.3 Asymmetric Information: Second-Best Outcome or Optimal Nonlin-

ear Pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

1.2 Modeling Issues in Contracting Problem . . . . . . . . . . . . . . . . . . . 16

1.2.1 Ex-Ante Contracting . . . . . . . . . . . . . . . . . . . . . . . . . . 16

1.2.2 Countervailing Incentives . . . . . . . . . . . . . . . . . . . . . . . . 18

1.3 Other Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

1.3.1 Regulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

1.3.2 Collateral as a Screening Device in Loan Markets . . . . . . . . . . 23

1.3.3 Credit Rationing . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

1
 1.3.4 Optimal Income Taxation . . . . . . . . . . . . . . . . . . . . . . . 30

1.4 Appendix: Lagrangian Method . . . . . . . . . . . . . . . . . . . . . . . . 32

2 Hidden Action: Moral Hazard 34

2.1 A Simple 2 × 2 Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

2.1.1 First-Best: Complete Information Optimal Contract . . . . . . . . . 37

2.1.2 Risk-Neutral Agent and the First-Best Contract . . . . . . . . . . . 38

2.1.3 Limited Liability and Second-Best Contract . . . . . . . . . . . . . 39

2.1.4 Risk-Averse Agent and Second-Best Contract . . . . . . . . . . . . 40

2.2 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

2.2.1 More than 2 Outcomes . . . . . . . . . . . . . . . . . . . . . . . . . 42

2.2.2 Comparing Information Structures . . . . . . . . . . . . . . . . . . 43

2.2.3 Moral Hazard and Renegotiation . . . . . . . . . . . . . . . . . . . 45

2.3 Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

2.3.1 Moral Hazard in Insurance Markets . . . . . . . . . . . . . . . . . . 47

2.3.2 Moral Hazard in Teams . . . . . . . . . . . . . . . . . . . . . . . . 50

3 Incomplete Contracts 54

3.1 The Holdup Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

3.2 Organizational Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

3.3 Contractual Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

3.4 Cooperative Investment and Failure of Contracts . . . . . . . . . . . . . . 62

2
4 Auction Theory 64

4.1 Second-Price (or English) Auction . . . . . . . . . . . . . . . . . . . . . . . 67

4.2 First-Price Auction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

4.3 Revenue Equivalence and Optimal Auction . . . . . . . . . . . . . . . . . . 71

4.4 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

4.4.1 Risk Averse Bidders . . . . . . . . . . . . . . . . . . . . . . . . . . 76

4.4.2 Budget Constrained Bidders . . . . . . . . . . . . . . . . . . . . . . 78

4.4.3 Optimal Auction with Correlated Values . . . . . . . . . . . . . . . 79

4.5 Appendix: Envelope Theorem . . . . . . . . . . . . . . . . . . . . . . . . . 81

5 Matching Theory 83

5.1 Stable Matchings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

5.1.1 Gale-Shapley Algorithm . . . . . . . . . . . . . . . . . . . . . . . . 87

5.1.2 Incentive Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

5.1.3 Extension to Many-to-One Matching . . . . . . . . . . . . . . . . . 92

5.2 Efficient Matching . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

5.2.1 Top Trading Cycles Algorithm . . . . . . . . . . . . . . . . . . . . . 96

3
Introduction

• What is contract? - ‘A specification of actions that named parties are supposed to take

at various times, generally as a function of the conditions that hold’ (Shavell, 2004).

− An example is insurance contract under which (risk-averse) insureds pay premiums

and are covered against risk by an insurer.

− Some party might have incentive to behave opportunistically at the expense of others.

− In an ideal world, people can write a complete contingent contract that induces all

the parties to take the ‘right’ actions in every possible state of world, which leads to a

Pareto efficient outcome.

• Contract theory studies what will or should be the form of contracts in less than ideal

worlds, where there exist

− Hidden action (or moral hazard ): when the involved party’s behavior cannot be per-

fectly monitored by others.

− Hidden information (or adverse selection): when the involved party has private infor-

mation which is not known to others.

− Contractual incompleteness: when contracts do not deal with all relevant contingen-

cies.

• There are many applications of contract theory, among which the followings are impor-

tant:

4
 − Labor contracts

− Regulation

− Price discrimination

− Optimal taxation.

− Financial contracts

− Auctions

• In most applications, one party, called ‘principal ’, offers a contract to the other party,

called ‘agent’.

− One principal + one agent

− One principal + multiple agents, e.g. auction

− Multiple principals + one agent (common agency), e.g. lobby groups influencing a

government agency

− Multiple principals + multiple agents

5
Chapter 1

Hidden Information: Screening

• General setup

− An agent, informed party, who is privately informed about his type.

− A principal, uninformed party, who designs a contract in order to screen different types

of agent and maximize her payoff.

− This is a problem of hidden information, often referred to as screening problem.

• A list of of economic issues where this setup is relevant.

− Price discrimination: A monopolist selling to a buyer whose demand is unknown.

(often called ‘second-degree price discrimination’)

− Credit rationing in financial markets: A lender investing in a project whose profitability

is unknown.

− Optimal income taxation: A government designing the tax scheme for the people whose

income generating ability is unknown.

6
 − Implicit labor contract: An employer offering the wage contract to an employee whose

productivity is unknown.

In these examples, contracts correspond to the pricing scheme, investment decision, tax

scheme, and wage contract.

• To describe the contractual situations in general, we adopt the following time line:

− At date 0, the agent learns his type.

− At date 1, the principal offers a contract to the agent.

− At date 2, the agent accepts or rejects the contract. If rejects, then the principal and

agent get their outside utilities, which are normalized to zero. If accepts, then they go

to date 3.

− At date 3, the transaction (or allocation) occurs, following the contract

We then ask what is the optimal contract in the principal’s perspective, or the contract

which maximizes the principal’s payoff. Usually, however, there are so many contracts one

can conceive of that it is not feasible to try all possible contract one by one.

• Revelation principle

− To determine the optimal contract among all possible ones, it suffices to consider

the contracts which specify one allocation for each type of agent. (e.g. the insurance

example)

− Need to make sure that each type has an incentive to select only the allocation that is

destined to him.

7
1.1 A Simple Model of Price Discrimination

Consider a transaction between a buyer (agent) and a seller (principal), where the seller

does not know perfectly how much the buyer is willing to pay.

• The buyer has utility function is given by u(q, T, θ) = θv(q) − T

− q = the number of units purchased and T = the total amount paid to the seller.

− The buyer’s characteristic is represented by θ, which is only known to the buyer: θ = θL

with probability β ∈ (0, 1) and θ = θH > θL with (1 − β). Define ∆θ := θH − θL .

− We impose technical assumptions as follows: v(0) = 0, v 0 (q) > 0, and v 00 (q) < 0 for all

q.

− Outside utility of the buyer is fixed at some level ū = 0.

• The seller’s preference is given by π = T − cq, where c is the seller’s production cost per

unit.

1.1.1 Full Information Benchmark: First-Best Outcome or Per-

fect Price Discrimination

Suppose that the seller is perfectly informed about the buyer’s type. The seller can treat

each type of buyer separately and offer a type-dependent contract: (qi , Ti ) for type θi

(i = H, L).

• The seller solves

max Ti − cqi subject to θi v(qi ) − Ti ≥ 0.

qi ,Ti

8
 − The constraint is called participation constraint (or individual rationality, (IR), con-

straint): If this constraint is not satisfied, then the agent would not participate in the

contractual relationship.

− At the optimal solution, (IR) constraint must be satisfied as an equality or binding.

(Why?)

− A solution to this problem is (q̂i , T̂i ) given by

T̂i = θi v(q̂i ) (from binding (IR) constraint)

and

θi v 0 (q̂i ) = c. (from first order condition)

− This quantity is efficient since it maximizes the total surplus:

q̂i = arg max θi v(q) − cq for i = H, L.

q

− With this solution, the seller takes all the surplus equal to

β(θL v(q̂L ) − cq̂L ) + (1 − β)(θH v(q̂H ) − cq̂H ),

while the buyer gets no surplus. It is why this solution is called first-best or perfect

price discrimination.

1.1.2 Asymmetric Information: Linear Pricing and Two-Part

Tariffs

Suppose from now on that the seller cannot observe the type of the buyer, facing the adverse

selection problem. The first-best contract above is no longer feasible. (Why?) The contract

9
set is potentially large since the seller can offer any combination of quantity-payment pair

(q, T (q)). In other words, the seller is free to choose any function T (q). Let us first consider

two simple functions among others.

Linear Pricing: T (q) = P q

This is the simplest contract in which the buyer pays a uniform price P for each unit he

buys.

• Given this contract, the buyer of type θi chooses q to maximize

θi v(q) − P q, where i = L, H.

− From the first-order condition, we obtain θi v 0 (q) = P .

− We can derive the demand function: qi = Di (P ), i = L, H. Note that DH (P ) >

DL (P ).

− We can also calculate the buyer’s net surplus as Si (P ) := θi v(Di (P )) − P Di (P ),

i = L, H. Note that SH (P ) > SL (P ).

− Let us define

D(P ) := βDL (P ) + (1 − β)DH (P )

S(P ) := βSL (P ) + (1 − β)SH (P ).

• With linear pricing, the seller solves

max(P − c)D(P ).
P

− From the first-order condition, we obtain the optimal price P m

D(P m )
Pm = c − .
D0 (P m )

10
 − With this solution, the buyer obtains positive rents (why?) and consumes inefficiently

low quantities since

θi v 0 (qim ) = P m > c = θi v 0 (q̂i ).

Two-Part Tariff: T (q) = F + P q

With two-part tariff, the seller charges a fixed fee (F ) up-front, and a price P for each

unit purchased. For any given price P , the maximum fee the seller can charge up-front is

F = SL (P ) if he wants to serve both types. (∵ SL (P ) < SH (P ))

• The seller chooses P to maximize

β [SL (P ) + (P − c)DL (P )] + (1 − β) [SL (P ) + (P − c)DH (P )]

=SL (P ) + (P − c)D(P ).

− From the first order condition, we obtain the optimal unit price P t solving

D(P t ) + SL0 (P t )
Pt = c − .
D0 (P t )

− Since SL0 (P t ) = −DL (P t ) by the envelope theorem, we have D(P t ) + SL0 (P t ) > 0; in

addition, D0 (P t ) < 0, so P t > c. Again, the quantities are inefficiently low, or qit < q̂i

for i = L, H.

− But, the inefficiency reduces compared to the linear pricing, or qit > qim , since P t < P m .

To see it, we observe


d
= SL0 (P m ) = −DL (P m ) < 0

[(P − c)D(P ) + SL (P )]
dP P =P m

so that the seller is better off lowering the price from P m . (Intuition?)

11
• Let Bi := (qit , SL (P t ) + P t qit ) denote the bundle for type θi for i = L, H.

− The θH -type strictly prefers the bundle BH to BL .

− Perhaps, the seller could raise the payment of θH -type beyond SL (P t ) + P t qH

t
to offer
0 0
some other bundle BH so that the θH -type is indifferent between BL and BH .

1.1.3 Asymmetric Information: Second-Best Outcome or Opti-

mal Nonlinear Pricing

Here, we look for the best pricing scheme among all possible ones. In general, the pricing

scheme can be described as (q, T (q)), where the function T (q) specifies how much the buyer

has to pay for each quantity q. We do not restrict the function T (·) to be linear or affine

as before.

• The seller’s problem is to solve

max β(T (qL ) − cqL ) + (1 − β)(T (qH ) − cqH )

T (q)

subject to

θi v(qi ) − T (qi ) ≥ θi v(q) − T (q) for all q and i = L, H (IC)

and

θi v(qi ) − T (qi ) ≥ 0 for i = L, H. (IR)

− The first two constraints are called incentive compatibility constraint, which guarantees

that each type selects the bundle that is designed for him.

12
 − The next two constraints are called individual rationality or participation constraint,

which guarantees that each type is willing to participate in the seller’s contract.

− We solve this adverse selection problem step-by-step in the below.

• Step 1: Apply the revelation principle.

− We can use the revelation principle to restrict our attention to a couple of bundles,

(qL , T (qL )) for type θL and (qH , T (qH )) for type θH . Then, defining Ti := T (qi ) for

i = L, H, the above problem can be rewritten as

max β(TL − cqL ) + (1 − β)(TH − cqH ) (1.1)

(qL ,TL )
(qH ,TH )

subject to

θH v(qH ) − TH ≥ θH v(qL ) − TL (ICH )

θL v(qL ) − TL ≥ θL v(qH ) − TH (ICL )

θH v(qH ) − TH ≥ 0 (IRH )

θL v(qL ) − TL ≥ 0. (IRL )

− Note that the incentive compatibility constraint has been greatly simplified.

• Step 2: (IRL ) must be binding at the optimal solution.

− If not, then the seller can raise TL and TH by small  > 0, which does not violate any

constraint but raises the seller’s utility, a contradiction.

− Thus, we have

TL = θL v(qL ). (1.2)

13
• Step 3: (ICH ) must be binding at the optimal solution.

− If not, then the seller can raise TH by small  > 0, which does not violate any constraint

but raises the seller’s utility, a contradiction.

− Thus, we have

TH = θH v(qH ) − (θH v(qL ) − TL ) = θH v(qH ) − ∆θv(qL ) . (1.3)

| {z }
information rent

• Step 4: (IRH ) is automatically satisfied, provided that (ICH ) and (IRL ) are binding.

− Due to the binding (ICH ) and (IRL ),

θH v(qH ) − TH = θH v(qL ) − TL ≥ θL v(qL ) − TL = 0,

so (IRH ) is satisfied.

− We can thus ignore the (IRH ) constraint.

• Step 5: Given that (ICH ) is binding, (ICL ) is satisfied if and only if qH ≥ qL .

− Given that (ICH ) is binding, (ICL ) constraint can be written as

θL (v(qL ) − v(qH )) ≥ TL − TH = θH (v(qL ) − v(qH )),

which will be satisfied if and only if qH ≥ qL .

− Thus, we can replace (ICL ) constraint by the constraint qH ≥ qL .

• Step 6: Eliminate TL and TH using (1.2) and (1.3) and solve the problem without any

constraint.

− For the moment, let us ignore the constraint qH ≥ qL , which will be verified later.

14
 − Substituting (1.2) and (1.3), the seller’s problem is turned into

max β(θL v(qL ) − cqL ) + (1 − β)(θH v(qH ) − cqH − ∆θv(qL )).

qL ,qH

∗ ∗
− From the first-order condition, we obtain the optimal quantities, qH and qH , solving

θH v 0 (qH
∗
)=c
c
θL v 0 (qL∗ ) = (1−β) ∆θ
> c, (1.4)
1− β θL

∗
which implies qH = q̂H > q̂L > qL∗ , as desired.

− If βθL < (1 − β)∆θ, then the RHS of (1.4) becomes negative so we would have a corner
∗
solution, qL∗ = 0 and qH = q̂H . (Try to interpret this)

• The second-best outcome exhibits the following properties.

∗
− No distortion at the top: qH = q̂H

− Downward distortion below the top: qL∗ < q̂L

− Why? → To reduce the information rent. (Refer to the equation (1.3))

15
1.2 Modeling Issues in Contracting Problem

1.2.1 Ex-Ante Contracting

There are situations in which the agent can learn his type only after he signs a contract:

For instance, an employee who is hired to work on some project, may not know whether

his expertise is suited to the project until he starts working on it. This kind of situation

can be modeled by modifying the timeline in page 7 and assuming that the agent privately

learns his type between date 2 and 3. So, it is at the ex-ante stage that the contracting

occurs. In the original model, by contrast, the contracting has occurred at the interim stage

in which the agent is already informed of his type. We analyze the problem of designing

optimal ex-ante contract in our basic model of price discrimination, though the results that

follow are much more general, applying to any adverse selection model. To put forward

the conclusion, it is possible for the principal to achieve the first-best outcome, despite the

asymmetric information arising after the contracting stage.

• The principal’s problem is now to solve

max β(TL − cqL ) + (1 − β)(TH − cqH )

(qL ,TL )
(qH ,TH )

subject to

β(θL v(qL ) − TL ) + (1 − β)(θH v(qH ) − TH ) ≥ 0 (IR)

θH v(qH ) − TH ≥ θH v(qL ) − TL (ICH )

θL v(qL ) − TL ≥ θL v(qH ) − TH (ICL )

16
 − Note that there is only one (IR) constraint that is for the agent to participate in the

contract without knowing his type.

− Note also that (ICL ) and (ICH ) constraints are intact since the information the agent

learns in the post-contracting stage remains to be private.

− As it turns out, the first-best outcome is achievable for the principal as follows:

(i) (IR) constraint must be binding (Why?).

(ii) Substitute the binding (IR) constraint into the object function to get

βθL v(qL ) + (1 − β)θH v(qH ) − βcqL − (1 − β)cqH ,

which is maximized at qH = q̂H and qL = q̂L , the first-best quantities.

(iii) Choose TL and TH to make (IR) binding and, at the same time, satisfy (ICH ) and

(ICL ). (How?)

There are some lessons to be learned from the above results: First, it is not the lack

of information itself but the asymmetry of information that causes the inefficiency in the

previous section. Also, what generates a rent for the agent is the asymmetric information at

the contracting stage. In the above setup where both parties symmetrically informed at the

contracting stage, the principal only needs to satisfy the ex-ante participation constraint,

which enables the principal to push down the low type’s payoff below outside utility while

guaranteeing the high type a payoff above outside utility. So, the agent without knowing

about his type breaks even on average and thus is willing to participate. This ex ante

incentive, however, would not be enough for inducing the agent’s participation if the agent

can opt out of the contractual relationship at any stage. In such case, both types should

17
be assured of their reservation payoff at least, which will revert the principal’s problem to

that in the previous section.

1.2.2 Countervailing Incentives

So far we have assume that the agent’s outside utility is type-independent, being uniformly

equal to zero. Often, this is not very realistic. In our price discrimination model, for

instance, the high-type consumer to pay may be able to find an outside opportunity that is

more lucrative than the low-type consumer does, thereby enjoying a higher outside utility

than the latter does. This situation can easily be modeled by modifying the (IRH ) as

follows: for some ū > 0,

0
θH v(qH ) − TH ≥ ū (IRH )

We maintain all other constraints. This small change in the contracting problem leads to

some dramatic change in the form of optimal contract.

• If ū ∈ [∆θv(q̂L ), ∆θv(q̂H )], then the seller as principal can achieve the first-best outcome.

− To see it, note that the first-best outcome requires (i) the first-best quantities or
0
qi = q̂i , i = H, L, and (ii) both (IRL ) and (IRH ) be binding so that TL = θL v(q̂L ) and

TH = θH v(q̂H ) − ū.

− We only need to verify that the quantities and transfers given above satisfy both (ICL )

and (ICH ) conditions, which can be written as

θL (v(qL ) − v(qH )) ≥ TL − TH ≥ θH (v(qL ) − v(qH )).

|{z} |{z}
ICL ICH

18
 By plugging the above numbers into this equation and rearranging, we obtain

∆θv(q̂H ) ≥ ū ≥ ∆θv(q̂L ),

which holds since ū ∈ [∆θv(q̂L ), ∆θv(q̂H )] as assumed.

What happens if ū > ∆θv(q̂H )?1 In this case, the first-best contract given above violates

(ICL ) only, which implies (ICL ) must be binding at the optimum. Also, analogously to

Step 5 in page 14, qH ≥ qL is necessary and sufficient for (ICH ) to be satisfied, given that

(ICL ) is binding. As before, we will ignore the constraint qH ≥ qL for the moment, which
0
can be verified later. Another observation is that (IRH ) also must be binding.2 We now

consider two cases depending on whether (IRL ) is binding or not.

0
• Let us first focus on the case in which (ICL ) and (IRH ) are binding while (IRL ) is not.

− From two binding constraints, one can derive

TH = θH v(qH ) − ū, TL = ∆θv(qH ) + θL v(qL ) − ū.

− By substitution, the principal’s problem can be turned into

max β(∆θv(qH ) + θL v(qL ) − ū − cqL ) + (1 − β)(θH v(qH ) − ū − cqH ).

qL ,qH

− The first-order condition yields the optimal quantities satisfying

c
θH v 0 (qH
∗∗
)= β∆θ
<c
1+ (1−β)θH

θL v 0 (qL∗∗ ) = c,
1 ∗
The same analysis and result as in the part 1.1.3 follow in case ū ≤ ∆θv(qL ), including the case ū = 0.
∗
If ū ∈ (∆θv(q̂L ), ∆θv(q̂L )), the analysis will be slightly different but yield the same qualitative result as in

1.1.3.
2
An argument to show this is a bit tedious and will be given in the class.

19
 which implies that qL∗∗ = q̂L and qH
∗∗
> q̂H . In contrast to the standard case, the quantity

here is distorted upward at the top.

− It remains to check that (IRL ) is satisfied. To do so, plug the above quantities and

transfers into the non-binding (IRL ) to obtain

∗∗
ū > ∆θv(qH ). (1.5)

So the above inequality is necessary (and sufficient) for having an optimal contract in
0
which only (ICL ) and (IRH ) are binding.

∗∗
• If (1.5) is violated, that is ū ∈ (∆θv(q̂H ), ∆θv(qH )], then, (IRL ) also must be binding.

0
− Plug (IRH ) and (IRL ) (as equality) into (ICL ) (as equality) to obtain the optimal

quantity for high type satisfying

∗∗∗
∆θv(qH ) = ū,

0
which then determines TH through the binding (IRH ) constraint.

− Then plug (IRL ) into the objective function, whose first-order condition results in the

optimal quantity for low type being equal to q̂L .

∗∗∗ ∗∗∗
− Note that since ∆θv(q̂H ) < ū = ∆θv(qH ), we must have q̂H < qH , an upward

distortion at the top again.

As seen above, if the high type entertains an attractive outside option, the participation

constraint countervails the incentive constraint so (ICH ) does not play a role in the design

of optimal contract. In other words, it is not much of a concern for the principal to save the

information rent needed to make the high type tell the truth. This leads to no downward

20
distortion at the bottom but rather an upward distortion at the top. This upward distortion

is the best way to satisfy the participation constraint for the high type while preventing

the low type from mimicking the high type.

1.3 Other Applications

1.3.1 Regulation

The public regulators are often subject to an informational disadvantage with respect to

the regulated utility or natural monopoly. Consider a regulator concerned with protecting

consumer welfare and attempting to force a natural monopoly to charge the competitive

price. The difficulty is that the regulator does not have full knowledge of the firm’s intrinsic

cost structure.

• Consider a natural monopoly with an exogenous cost parameter θ ∈ {θL , θH } with ∆θ =

θH − θL > 0, which is only observable to the monopolist.

− The cost parameter is θL with probability β and θH with 1 − β.

− The firm’s cost of producing good is observable (and contractible) and given by c =

θ − e, where e > 0 stands for the cost-reducing effort. Expending effort e has cost

ψ(e) = e2 /2.

− To avoid the distortionary tax, the regulator tries to minimize its payment P = c + s

to the firm, where s is a “subsidy” in excess of accounting cost c.

• As a benchmark, assume that there is no information asymmetry. For each type i, the

21
 regulator solves

min si + ci = si + θi − ei
(ei ,si )

subject to

si − e2i /2 ≥ 0. (IRi )

− The constraint (IRi ) is binding at the solution, which is then given by êi = 1 and

ŝi = 1/2 for both i = H and L.

− Under the asymmetric information, this contract is vulnerable to type L’s pretending

to be type H: To generate cost cH = θH −1, type L only need to exert effort e = 1−∆θ.

• Under the asymmetric information, letting ∆θ := θH − θL , the regulator solves

min β(sL − eL ) + (1 − β)(sH − eH )

(eL ,sL )
(eH ,sH )

subject to

sL − e2L /2 ≥ 0 (IRL )

sH − e2H /2 ≥ 0 (IRH )

sL − e2L /2 ≥ sH − (eH − ∆θ)2 /2 (ICL )

sH − e2H /2 ≥ sL − (eL + ∆θ)2 /2. (ICH )

− The binding constraints are (ICL ) and (IRH ). (Why?) Thus,

sL − e2L /2 = sH − (eH − ∆θ)2 /2

sH − e2H /2 = 0.

22
 − Substituting these equations, the optimization problem becomes

min β e2L /2 − eL + e2H /2 − (eH − ∆θ)2 /2 + (1 − β) e2H /2 − eH ,

eL ,eH

whose solution is given by

β
e∗L = 1 and e∗H = 1 − ∆θ.
1−β

− As before, the provision of effort is efficient at the top (θL ) while distorted downward

at the bottom (θH ).

− The downward distortion is severer as the cost differential is larger or the type is more

likely to be efficient.

1.3.2 Collateral as a Screening Device in Loan Markets

Consider a loan market in which the bank (lender) offers loan contracts to the liquidity-

constrained investors (borrowers). Investors are heterogeneous in terms of the probability

their investment projects succeed. Though liquidity-constrained, borrowers own some asset

that can be (only) used as collateral for a loan contract. We study how this collateral can

help the lender screen different types of borrowers, using the simple model as follows:

• Each investor has a project with a random return ỹ that can either fail (ỹ = 0) or succeed

(ỹ = y > 0).

− There are two types of investors who only differ in their failure probability: θH and

θL < θH .

− The proportion (or probability) of type k is given as βk , k = H, L.

23
 − Letting Uk denote the reservation utility for type k, we assume that

UL UH
> . (1.6)
1 − θL 1 − θH

• There is a monopolistic lender who offers a menu of loan contracts {(Ck , Rk )}k=L,H ,

where Rk and Ck denote the repayment (in case of success) and collateral for type k.

− If the project fails, then the lender can take the specified amount of collateral from

the borrower and liquidate it to obtain δCk with δ < 1. (Thus (1 − δ)Ck corresponds

to a liquidation cost.)

− Thus, the borrower’s payoff from type k is Rk if the project succeeds and δCk if fails.

So the expected payoff is given as

X
βk [(1 − θk )Rk + θk δCk ] (1.7)
k=H,L

− The payoff of type k borrower is (y − Rk ) if succeeds −Ck if fails. So the expected

payoff is given as

(1 − θk )(y − Rk ) − θk Ck .

− Note that this model has some fundamental difference from the ones we have seen so far

in the sense that the agent’s private information, θ, directly enters into the principal’s

utility function. In short, the agent’s information has an externality on the principal’s

payoff. A model with this kind of feature is generally referred to as ‘common value

model’.

• Consider as a benchmark that the lender is able to observe the borrower’s types.

24
 − At an optimal contract, the individual rationality constraints given by

(1 − θL )(y − RL ) − θL CL ≥ UL (IRL )

and (1 − θH )(y − RH ) − θH CH ≥ UH (IRH )

must be binding.

− Under the constraints (IRL ) and (IRH ), (1.7) is maximized by setting Ck = 0 and
Uk
Rk = y − 1−θk
for each type k: No collateral is required at the optimal contract since

the liquidation of collateral is costly.

− If the lender faces the asymmetric information problem, then the above contract would
UL UH
not work since RL = y − 1−θL
<y− 1−θH
= RH from (1.6) and thus both types will

prefer announcing θL .

• Consider now the problem of finding the optimal menu of contracts for the lender who

is uninformed about the borrower types.

− We need to maximize (1.7) under the (IRL ), (IRH ) and the (IC) constraints given as

(1 − θH )(y − RH ) − θH CH ≥ (1 − θH )(y − RL ) − θH CL (ICH )

(1 − θL )(y − RL ) − θL CL ≥ (1 − θL )(y − RH ) − θL CH . (ICL )

− As can be seen above, what matters is to get rid of the incentive of type θH to mimic

θL , which means that (ICH ) must be binding at the optimal contract.

− (ICH ) and (ICL ) constraints can be rewritten together as

1 − θL 1 − θH
(RL − RH ) ≤ CH − CL ≤ (RL − RH ), (1.8)
θL |{z} |{z} θH
(ICL ) (ICH )

1−θH 1−θL
which implies CH − CL ≤ 0 since θH
< θL
.

25
 − Given that (ICH ) is binding and CL ≥ CH , (1.8) implies that (ICL ) is automatically

satisfied and thus can be ignored.

− At the optimal contract {(Ck , Rk )}k=L,H , we must have CH = 0: If CH > 0, then the
0 0 0
principal can offer an alternative contract {(CL , RL ), (CH , RH )}, where CH = CH − 
0 θH
and RH = RH + 1−θH
 for small  > 0. With this contract, (ICH ), (IRL ), and

(IRH ) all remain the same as with the original contract and thus are satisfied. Since
0
CH < CH ≤ CL and (ICH ) is binding,(ICL ) is also satisfied by the previous argument.

However, the principal’s profit increases by

θH
(1 − θH )  − δθH  = θH (1 − δ) > 0,
1 − θH

contradicting the optimality of the original contract.

− At the optimal contract, (IRL ) must be binding: Suppose it’s not binding or

(1 − θL )(y − RL ) − θL CL > UL . (1.9)

We can then consider an alternative contract, {(CL0 , RL0 ), (CH , RH )}, where CL0 = CL −
θH
and RL0 = RL + 1−θH
 for small  > 0. With this contract, (ICH ) and (IRH ) remain

the same as with the original contract and thus are satisfied. Also, (IRL ) is satisfied

since  is small. Since CL0 = CL −  > CH = 0 and (ICH ) is binding, (ICL ) is satisfied

for the same reason as before. However, the principal’s payoff increases by
 
θH θH (1 − θL )
(1 − θL )  − δθL  = θL − δ  > 0,
1 − θH θL (1 − θH )

contradicting the optimality of the original contract.

• Thus, the lender’s problem is simplified to

max βH (1 − θH )RH + βL [(1 − θL )RL + δθL CL ]

RH ,RL ,CL

26
 subject to

(1 − θH )(y − RH ) = (1 − θH )(y − RL ) − θH CL (1.10)

(1 − θL )(y − RL ) − θL CL = UL (1.11)

(1 − θH )(y − RH ) ≥ UH . (1.12)

− One can check that if RH is reduced by  > 0, then RL and CL must be changed
θL (1−θH )
by θH −θL
 and − (1−θθHL )(1−θ
−θL
H)
, respectively. Thus, the corresponding change of the

lender’s payoff is  times

 
βL (1 − δ)(1 − θL )θL
(1 − θH ) −βH + . (1.13)
θH − θL

− In the optimal contract, we have two cases: (i) if (1.13) is positive, then (CL , RL ) =
UL
(CH , RH ) = (0, y − 1−θL
) (and (1.12) is not binding); (ii) if (1.13) is negative, (1.12) is

binding and RH , RL , and CL are the solution of (1.10) to (1.12).

− As δ is small or βL big or θH close to θL , no collateral for type θL is optimal.

− We may conclude that

(a) High (low) risk type pays a high (low) interest rate but puts down no (some)

collateral as long as the optimal contract is separating;

(b) In case collateral is costly or types are likely to be low or two types are close, no

collateral is required and the optimal contract is pooling.

1.3.3 Credit Rationing

Adverse selection is typical of financial markets: A lender knows less than a borrower

about the quality of a project he invests in. Because of this informational asymmetry, good

27
quality projects are denied credit so there arises inefficiency in the allocation of investment

funds to project. This type of inefficiency is generally referred to as “credit rationing.”

• Consider a unit mass of borrowers who have no wealth but each own a project:

− Each project requires an initial investment of I = 1.

− There are two types of projects, safe and risky. A project i = s, r yields a random

return, denoted by Xi : Xi = Ri > 0 (success) with probability pi ∈ [0, 1] and Xi = 0

(failure) with 1 − pi .

− Assume that pi Ri = m > 1 while ps > pi and Rs < Rr . That is, two types of projects

have the same expected return, but type-s is safer than type-r.

− The proportion of “safe” borrowers is β and that of “risky” borrowers is 1 − β. The

bank has a total amount α of funds: α > max{β, 1 − β}

What type of lending contract should the bank offer a borrower? Under symmetric

information, the bank would be indifferent about financing either type of borrower: In

exchange for the initial investment Iˆ = 1, it requires a repayment of D̂i = Ri in case of

success from type-i borrowers. What if there is asymmetric information problem between

lender and borrowers?

• Let us first consider contracts where the bank specifies a fixed repayment, D.

− If the bank decides to accommodate type-r only, then setting D = Rr is optimal, which

gives the bank a profit of

(1 − β)(m − 1). (1.14)

28
 − If the bank decides to accommodate both types, setting D = Rs is optimal. Assuming

that each applicant has an equal chance of being financed, the bank’s profit is equal to

α [β(m − 1) + (1 − β)(pr − Rs − 1)] . (1.15)

− Ceteris paribus, (1.15) will be higher than (1.14) when 1 − β is small enough, or when

pr is close enough to ps , in which case some “risky” borrowers cannot get credit even

though they would be ready to accept a higher D.

Can’t bank do better by offering more sophisticated contracts?

• Consider a contract (xi , Di ) that offers financing with probability xi and repayment Di .

The bank’s problem is

max βxs (ps Ds − 1) + (1 − β)xr (pr Dr − 1)

(xs ,Ds )
(xr ,Dr )

subject to

0 ≤ xi ≤ 1 for all i = s, r

Di ≤ Ri for all i = s, r (IRi )

xi pi (Ri − Di ) ≥ xj pi (Ri − Dj ) for all i = s, r (ICi )

βxs + (1 − β)xr ≤ α.

− The same line of reasoning as in the previous model shows (how?) that (IRs ) and

(ICr ) constraints are binding, which implies that

xs
Ds = Rs and Dr = Rr − (Rr − Rs ).
xr

− As before, we can ignore (ICs ) and (IRr ) constraints.

29
• So, the proceeding problem becomes

max βxs (ps Rs − 1) + (1 − β) [xr (pr Rr − 1) − xs pr (Rr − Rs )]

xs ,xr

subject to

βxs + (1 − β)xr ≤ α.

− The solution of this problem is given by: x∗r = 1 and


 0 if β(ps Rs − 1) − (1 − β)pr (Rr − Rs ) < 0

∗
xs =
 α−(1−β) otherwise.

β

− The risky types are not rationed any more.

− The safe types are not fully funded but are indifferent about being funded

− The informational asymmetry does not necessarily give rise to credit rationing.

− We restricted attention to a fixed repayment schedule, so called debt contract. Why

not use a return-contingent repayment schedule, which is to set repayments equal to

realized returns and is much more efficient than debt contracts?

1.3.4 Optimal Income Taxation

In this part, we study the trade-off between allocative efficiency and redistributive taxation

in the presence of adverse selection. A lump-sum taxation does entail no distortion in

general, but may not be feasible since its size is limited by the lowest income level. To

be able to raise higher tax revenues, the tax need be based on an individual’s income-

generating-ability, which is not observable, however so adverse selection arises.

30
• Consider an economy where each individual from a unit mass of population is endowed

with production function q = θe, where q, e, and θ denote income, effort, and ability,

respectively.

− A proportion β of individuals has low productivity θL while a proportion (1 − β) has

high productivity θH > θL .

− All individuals have the same utility function u(q − t − ψ(e)), where t is the net tax

paid to (or received from) the government and ψ(e) is an increasing and convex cost

function.

− The government’s budget constraint is given by

0 ≤ βtL + (1 − β)tH , (1.16)

where ti is the net tax from a type-i individual.

• Under symmetric information, a utilitarian government maximizing the sum of individual

utilities, solves

max βu(θL eL − tL − ψ(eL )) + (1 − β)u(θH eH − tH − ψ(eH )),

(eL ,tL )
(eH ,tH )

subject to (1.16).

− At the optimum, (1.16) is binding and the first order conditions yield

u0L := u0 (θL êL − tL − ψ(êL )) = u0 (θH êH − tH − ψ(êH )) =: u0H (1.17)

ψ 0 (êL ) = θL

ψ 0 (êH ) = θH .

31
• Under asymmetric information, the government has extra conditions to satisfy:
 
θH eH
θL eL − tL − ψ(eL ) ≥ θH eL − tH − ψ (ICL )
θL
 
θL eL
θH eH − tH − ψ(eH ) ≥ θL eL − tL − ψ . (ICH )
θH

− Note that the allocation satisfying (1.17) violates (ICH ) since

 
θL êL
θH êH − tH − ψ(êH ) = θL êL − tL − ψ(êL ) > θL êL − tL − ψ .
θH

Thus, in a second-best optimum, (ICH ) must be binding together with (1.16).

− Substituting these two binding constraints and applying the first order condition with

respect to eH and eL yields

ψ 0 (e∗H ) = θH
θL e∗L
  
0 θL 0
ψ (e∗L ) = θL − (1 − β)γ ψ 0
(e∗L )− ψ ,
θH θH
| {z }
Q
where γ := (u0L − u0H )/[βu0L + (1 − β)u0H ]. Thus, e∗H = êH but e∗L < êL , since Q > 0

(why?).

− We can interpret this result in terms of the income tax code: The marginal tax rate is
Q
equal to 0 at output qH = θH e∗H while it is equal to θL
at output qL = θL e∗L .

1.4 Appendix: Lagrangian Method

Let f and hi , i = 1, · · · m be concave C 1 functions defined on the open and convex set

U ⊂ Rn . Consider the following maximization problem:

Maximize f (x) subject to x ∈ D = {z ∈ U | hi (z) ≥ 0, i = 1, · · · , m}.

32
 Set up the Lagrangian for this problem, which is defined on U × Rm
+ , as follow:

m
X
L(x, λ) = f (x) + λi hi (x).
i=1

We call a pair (x∗ , λ∗ ) ∈ U × Rm

+ saddle point of L if it satisfies the following conditions:

Df (x∗ ) + λ∗i Dhi (x∗ ) = 0 (1.18)

hi (x∗ ) ≥ 0 and λ∗i hi (x∗ ) = 0, ∀i = 1, · · · , m. (1.19)

Here, (1.18) implies that given λ∗ , x∗ maximizes L while (1.19) implies that given x∗ ,

λ∗ minimizes L, which is why we call (x∗ , λ∗ ) saddle point. Condition (1.19) is called

complementary slackness condition.

Theorem 1.1 (Kuhn-Tucker). x∗ ∈ U solves the above problem if and only if there is

λ∗ ∈ Rm ∗ ∗
+ such that (x , λ ) is a saddle point of L.

Remark 1. If functions f and hi are quasi-concave, then x∗ maximizes f over D provided

at least one of the following conditions holds: (a) Df (x∗ ) 6= 0; (b) f is concave.

Remark 2. If some of the constraints, say constraints k to m, are not binding that is

hi (x∗ ) 6= 0 for i = k, · · · , m, then x∗ must be a solution of the following relaxed problem

also:

Maximize f (x) subject to x ∈ D0 = {z ∈ U | hi (z) ≥ 0, i = 1, · · · , k − 1}.

33
Chapter 2

Hidden Action: Moral Hazard

• Basic setup:

− Agent takes an action that is not observable or verifiable and thus cannot be contracted

upon.

− The agent’s action is costly to himself but benefits the principal.

− The agent’s action generates a random outcome, which is both observable and verifi-

able.

− Principal designs a contract based on the observable outcome.

We refer this kind of setup as moral hazard problem.

• We are interested in

− How to design a contract which gives agent an incentive to participate (participation

constraint) and choose the ‘right’ action (incentive constraint).

34
 − When and how the unobservability of agent’s action distorts the contract away from

the first-best.

• The usual time line is as follows:

− At date 0, principal offers a contract.

− At date 1, agent accepts or rejects the contract. If rejects, then both principal and

agent get their outside utilities, zero. If accepts, then they proceed.

− At date 2, agent chooses action.

− At date 3, outcome is realized.

− At date 4, the contract is executed.

• A list of examples includes1

− Employer and employee: Induce employee to take a profit-enhancing action ← Base

employee’s compensation on employer’s profits

− Plaintiff and attorney: Induce attorney to exert effort to increase plaintiff’s chance of

prevailing at trial ← Make attorney’s fee contingent on damages awarded plaintiff.

− Landlord and tenant: Induce tenant to make investment that preserve property’s value

to the landlord ← Make tenant post deposit to be forfeited if value declines too much.

1
These examples draw on Hermalin’s note.

35
2.1 A Simple 2 × 2 Model

• Suppose that there is an employer (principal) and an employee (agent)

− Agent could be idle (e = 0 or low effort) or work hard (e = 1 or high effort), which is

not observable to the principal.

− The level of production, which is observable and verifiable, is stochastic, taking two

values q and q with q > q: Letting πe denote the probability that q realizes when agent

exerts e, we assume π0 < π1 .

− The agent’s utility when exerting e and paid t: u(t) − ce with u(·) increasing and

concave, and u(0) = 0. (h := u−1 , inverse function of u(·))

− The principal’s utility when q realizes and t is paid to agent: S(q) − t.

• The principal can only offer a contract based on the level of q: t(q) with t := t(q) and

t := t(q). We also denote S := S(q) and S := S(q).

− If agent exerts e, then principal receives the expected utility

Ve = πe (S − t) + (1 − πe )(S − t).

− To induce a high effort, the (moral hazard) incentive constraint is needed:

π1 u(t) + (1 − π1 )u(t) − c ≥ π0 u(t) + (1 − π0 )u(t).

− To ensure that agent participates, the participation constraint is needed:

π1 u(t) + (1 − π1 )u(t) − c ≥ 0.

36
2.1.1 First-Best: Complete Information Optimal Contract

In this benchmark case, we assume that the effort level can contracted upon since it is

observable and verifiable.

• If principal wants to induce effort level e, his problem becomes

max πe (S − t) + (1 − πe )(S − t)
t,t

s.t. πe u(t) + (1 − πe )u(t) − ce ≥ 0 (2.1)

− Letting λ denote the Lagrangean multiplier for the participation constraint (2.1), the

first-order conditions w.r.t. t and t yield

f
−πe + λπe u0 (t ) = 0

−(1 − πe ) + λ(1 − πe )u0 (tf ) = 0

1 1 f
From this, we immediately derive λ = f = u0 (tf )
, or t = tf = tf for some tf ≥ 0.
u0 (t )

− Since (2.1) must be binding at the optimum, tf = h(ce).

− So, the risk-neutral principal offers a full insurance to the risk-averse agent and then

extracts the full surplus.

− Principal prefers e = 1 if

π1 S + (1 − π1 )S − h(c) ≥ π0 S + (1 − π0 )S

or

∆π∆S
| {z } ≥ h(c) (2.2)
|{z}
expected gain first-best cost
of of
increasing effort increasing effort

and otherwise prefers e = 0.

37
 − From now on, we assume (2.2) holds so pirncipal would prefer e = 1 without moral

hazard problem.

2.1.2 Risk-Neutral Agent and the First-Best Contract

Suppose that there is moral hazard problem and suppose also that agent is risk-neutral,

i.e. u(t) = t. In this case, the principal can achieve the first-best outcome.

• The principal’s optimal contract to induce e = 1 must solve the following problem:

max π1 (S − t) + (1 − π1 )(S − t)
t,t

s.t. π1 t + (1 − π1 )t − c ≥ π0 t + (1 − π0 )t (2.3)

π1 t + (1 − π1 )t − c ≥ 0 (2.4)

− One contract (t, t) that achieves the first-best can be found by making both (2.4) and

(2.3) binding, which yields

π0 ∗ 1 − π0
t∗ = − c < 0 and t = c > 0.
∆π ∆π
So, agent is rewarded if output is high while punished if output is low. The agent’s
∗
expected gain of exerting a high effort is ∆π(t − t∗ ) = c.

− The principal obtains the first-best surplus, π1 S + (1 − π1 )S − c.

• In fact, there are many other contracts that achieve the first-best. Among them, the

“selling the firm” contract is often observed in the real world.

− The idea is to let the agent buy out the principal’s firm at a fixed price T ∗ by setting
∗0
t = S − T ∗ and t∗ 0 = S − T ∗ with T ∗ = π1 S + (1 − π1 )S − c: Easy to see that this

achieves the first-best surplus for the principal.

38
 − Note that the agent’s incentive constraint is satisfied since

∗0
∆π(t − t∗ 0 ) = ∆π∆S > h(c) = c,

which holds due to (2.2).

2.1.3 Limited Liability and Second-Best Contract

Let us assume that the agent has no wealth and is protected by limited liability constraint

that the transfer received by the agent should not be lower than zero:

t≥0 (2.5)

t ≥ 0. (2.6)

• Then, the principal’s problem is written as

max π1 (S − t) + (1 − π1 )(S − t)
t,t

subject to (2.3) to (2.6).

− One can argue that (2.3) and (2.6) must be binding, from which we obtain t∗ = 0 and
∗ c
t = ∆π
. Then, it can be checked that other constraints are automatically satisfied.

− The agent’s rent is positive:

∗ π0
π1 t + (1 − π1 )t∗ − c = c > 0.
∆π

− So, the principal’s payoff is equal to

∗ π1
V1∗ := π1 (S − t ) + (1 − π1 )(S − t∗ ) = π1 S + (1 − π1 )S − c,
∆π

which is lower than her first-best payoff by as much as the agent’s rent.

39
 − The principal would then like to induce e = 1 if V1∗ ≥ V0 or

π1 c π0 c
|∆π∆S ≥ =c+
{z } ∆π ∆π
expected gain |{z} |{z}
of second-best cost agent’s
increasing effort of rent
increasing effort

and otherwise induce e = 0.

2.1.4 Risk-Averse Agent and Second-Best Contract

Let us now turn to the second source of inefficiency in a moral hazard context: Agent’s

risk aversion.

• The principal’s problem is written as

(P ) max π1 (S − t) + (1 − π1 )(S − t)
t,t

subject to

π1 u(t) + (1 − π1 )u(t) − c ≥ π0 u(t) + (1 − π0 )u(t) (2.7)

π1 u(t) + (1 − π1 )u(t) − c ≥ 0. (2.8)

− It is not clear that (P ) is a ‘concave program’ for which the Kuhn-Tucker condition is

necessary and sufficient.

− Making the change of variables u := u(t) and u := u(t) or equivalently t = h(u) and

t = h(u), the problem (P ) turns into

(P 0 ) max π1 (S − h(u)) + (1 − π1 )(S − h(u))

u,u

40
 subject to

π1 u + (1 − π1 )u − c ≥ π0 u + (1 − π0 )u (2.9)

π1 u + (1 − π1 )u − c ≥ 0, (2.10)

which is a concave program.

• Set up the Lagrangian function for (P 0 ) as

L = π1 (S − h(u)) + (1 − π1 )(S − h(u)) + λ(∆π(u − u) − c) + µ(π1 u + (1 − π1 )u − c).

− Differentiating L with u and u yields

− π1 h0 (u∗ ) + λ∆π + µπ1 = 0

− (1 − π1 )h0 (u∗ ) − λ∆π + µ(1 − π1 ) = 0.

Rearranging, we obtain

∆π
h0 (u∗ ) = µ + λ (2.11)
π1
∆π
h0 (u∗ ) = µ − λ . (2.12)
(1 − π1 )

− One can argue that both λ and µ are positive so that (2.9) and (2.10) are both binding:

(i) If λ = 0, then (2.11) and (2.12) imply u∗ = u∗ , which violates (2.9), a contradiction.

(ii) If µ = 0, then (2.12) implies h0 (u∗ ) < 0, which is not possible since h(·) = u−1 (·) is

an inceasing function.

− From (2.9) and (2.10) as equality, we obtain

1 − π0
u∗ = c
∆π
π0
u∗ = − c.
∆π

41
 − Using this, one can show that the second-best cost of inducing a high effort is higher

than the first-best cost:

∗
C ∗ := π1 t + (1 − π1 )t∗ =π1 h(u∗ ) + (1 − π1 )h(u∗ )

>h(π1 u∗ + (1 − π1 )u∗ ) = h(c),

where the inequality holds due to the convexity of h(·).

− So, the principal would like to induce e = 1 if

∆π∆S ≥ C ∗

and otherwise induce e = 0.

2.2 Extensions

2.2.1 More than 2 Outcomes

Suppose that there are n possible output levels, {q1 , · · · , qn } with q1 < q2 < · · · < qn . Each

qi is realized with probability πie given the effort level, e, with πe = (π1e , · · · , πne ).

• Now, the optimal contract that induces e = 1 must solve

n
X
max πi1 (Si − ti ),
t1 ,··· ,tn
i=1

subject to
n
X n
X
πi1 u(ti ) − c ≥ πi0 u(ti ) (2.13)
i=1 i=1
Xn
πi1 u(ti ) − c ≥ 0. (2.14)
i=1

42
 − Let λ and µ denote the Lagrangian multipliers for (2.13) and (2.14), respectively.

Then, the first-order condition with respect to ti is given as

πi1
= µπi1 + λ (πi1 − πi0 ) . (2.15)
u0 (t∗i )

− Analoguously to the two-outcome case, both λ and µ can be shown to be positive: An

arugment to show µ > 0 is very similar and thus is omitted. To show µ > 0, sum up

equations (2.15) across i = 1, · · · , n and obtain

n n n
X πi1 X X
0< 0 (t∗ )
=µ πi1 + λ (πi1 − πi0 ) = µ.
i=1
u i i=1 i=1

− One may ask under what condition t∗i is increasing with i or wage is increasing with
πi1
output level. The condition for this is that the likelihood ratio is monotone, or πi0
is

increasing with i. To see it, rewrite (2.15) as

 
1 πi0
=µ+λ 1− .
u0 (t∗i ) πi1

2.2.2 Comparing Information Structures

The performance of optimal contract in moral hazard setup is in part affected by how

informative a contractible variable is about an agent’s hidden action. In the extreme case

the contractible varaible pefectly reveals the agent’s action, for instance, the principal will

be able to achieve the first-best outcome. In this part, we will investigate how the principal’s

payoff is affected by the informativeness of contractible variables in less than extreme cases.

• Consider two information structures, π and π̄, satisfying

n
X
π̄ej = πei pij for all j = 1, · · · , n
i=1
Pn
for some n × n transition matrix P = (pij ) with j=1 pij = 1 for each i = 1, · · · , n.

43
 − We say that the information structure π is Blackwell-sufficient for the information

structure π̄. In other words, the information structure π̄ is a garbling of the original

information structure π.

Let us define C ∗ (π) and C ∗ (π̄) as the second-best costs of inducing e = 1 under the

information structures π and π̄, that is

n
X n
X
C ∗ (π) = π1i t∗i and C ∗ (π̄) = π̄1i t̄∗i
i=1 i=1

• The information structure π is more efficient than π̄ if C ∗ (π) ≤ C ∗ (π̄), which is indeed

true if π is Blackwell-sufficient for π̄, as shown in the following proof.

Proof. Consider a different wage schedule (t01 , · · · , t0n ) defined by

n
X
u(t0i ) = pij u(t̄∗j ).
j=1

This wage schedule satisfies the incentive compatibility and individual rationality condi-

tions under the original information structure π:

n n n
! n n
!
X X X X X
c= (π̄1j − π̄0j )u(t̄∗j ) = (π1i − π0i ) pij u(t̄∗j ) = (π1i − π0i ) pij u(t̄∗j )
j=1 j=1 i=1 i=1 j=1

and ! !
n
X n
X n
X n
X n
X
c= π̄1j u(t̄∗j ) = π1i pij u(t̄∗j ) = pij u(t̄∗j ) .
j=1 j=1 i=1 i=1 j=1

Thus, we must have

n
X
πi t01i ≥ C ∗ (π).
i=1

44
 We can also obtain by the Jensen’s inequality
n n n
! n n
!
X X X X X
π1i t0i = π1i h pij u(t̄∗j ) ≤ π1i pij h(u(t̄∗j ))
i=1 i=1 j=1 i=1 j=1
n n
! n n
! n
X X X X X
= π1i pij t̄∗j = π1i pij t̄∗j = π̄1j t̄∗j .
i=1 j=1 j=1 i=1 j=1
Pn Pn
Thus, we conclude that C ∗ (π) ≤ i=1 π1i t0i ≤ j=1 π̄1j t̄∗j = C ∗ (π̄).

2.2.3 Moral Hazard and Renegotiation

Let us alternatively assume that the agent’s effort is not verifiable but can be observed

by the principal. Also, after observing the effort choice by agent but before the output

is realized, the principal can propose to renegotiate the initial contract. We investigate

whether this change in the contractual environment affects the implementability of the

first-best and then how.

It is more convenient for our analysis to generalize the setup as follows. The effort e,

which can be any nonnegative value, costs g(e). The principal’s profit, denoted S, is drawn

from [0, S] according to distribution p(S|e). The contract is now a function, t : [0, S] → R.

• Without renegotiation, the second-best contract t∗ (·) for inducing a target effort level e

will satisfy (IC) and (IR) as follows:

Z S
e = arg max
0
u(t∗ (S))p(S|e0 )dS − g(e0 ) (IC)
e ≥0 0

and
Z S
u(t∗ (S))p(S|e)dS − g(e) = 0. (IR)
0

45
 Suppose now that the principal is allowed to renegotiate after observing the agent’s

effort but before the profit is realized.

• After e0 has been chosen, the principal offers t̃(·|e0 ) in renegotiation that solves the

following program:
Z S
max
0
(S − t(S|e0 ))p(S|e0 )
t(·|e ) 0

subject to
Z S Z S
0 0
u(t(S|e ))p(S|e )dS ≥ u(t∗ (S))p(S|e0 )dS. (2.16)
0 0

− Clearly, t̃(·|e0 ) must be a full insurance contract that binds (2.16), that is
!
Z S
t̃(S|e0 ) = t̃(e0 ) := h u(t∗ (S))p(S|e0 )dS for all S ∈ [0, S].
0

− Note that t̃(e) corresponds to the minimum cost of inducing e:

!
Z S
t̃(e) = h u(t∗ (S))p(S|e)dS = g(e).
0

− It is important to ensure that the renegotiation does not change the agent’s incentive

to choose e rather than e0 6= e. For this, it suffices to see that

Z S Z S
0 0 0 0 0
u(t̃(S|e ))p(S|e )dS − g(e ) = u(t̃(e )) − g(e ) = u(t∗ (S))p(S|e0 )dS − g(e0 ).
0 0

− We conclude that the principal can achieve the first-best outcome.

− What happens if the renegotiation occurs without principal observing the agent’s ef-

fort?

46
2.3 Applications

2.3.1 Moral Hazard in Insurance Markets

Moral hazard is pervasive in insurance markets. Let us consider a risk-averse agent with

utility function u(·) and initial wealth w. Effort e ∈ {0, 1} is a level of safety care while

ce is the cost of choosing e. Given the effort level e, an accident occurs with probability

1 − πe with the damage worth d being incurred and otherwise no damage incurrred. We

assume that without insurance, the agent would choose e = 1 (or due care), that is

∆π(u(w) − u(w − d)) ≥ c,

where ∆π ≡ π1 − π0 > 0. Thus, the agent’s outside utility, denoted û, is given as

û = π1 u(w) + (1 − π1 )u(w − d) − c.

Let us assume that there is a competitive insurance market. In other words, there are

many insurance companies offering insurance contracts to the agent. An insurance contract

can be described by (u, u), where u (u, resp.) is the agent’s utility from his final wealth if no

accident (accident, resp.) occurs.2 Because of the competition among insurance companies,

the contract should maximize the agent’s expected utility subject to the standard incentive

compatibility constraint

c
u−u≥ , (2.17)
∆π

and the constraint that the profit of the insurance company be non-negative

π1 (w − h(u)) + (1 − π1 )(w − d − h(u)) ≥ 0. (2.18)

2
Note that since the total wealth is divided between agent and insurance company, the latter’s profit

is equal to w − h(u) if accident occurs and equal to w − d − h(u) if not.

47
• Without moral hazard, the incentive compatibility constraint can be ignored so the

equilibrium contract offered by insurance company must solve the following problem:

max π1 u + (1 − π1 )u − c
(u,u)

subject to (2.18).

− As in Part 2.1.1, the solution of the above problem yields u = u, which is denoted as

uf . Letting U f = uf − c denote the agent’s utility at the first-best, we have

h(U f + c) = w − d(1 − π1 ),

since (2.18) must be binding.

− The market would not break down if the agent is (weakly) better off with insurance

than without, or U f ≥ û, which is equivalent to

u(w − d(1 − π1 )) − c ≥ π1 u(w) + (1 − π1 )u(w − d) − c. (2.19)

This always holds since u(·) is a concave function.

• Under moral hazard, assuming that e = 1 is chosen in the equilibrium,3 the equilibrium

contract must solve the following problem:

(P ) max(u,u) π1 u + (1 − π1 )u − c

subject to (2.17) and (2.18).

3
Note that we need to check that the agent is willing to participate, and also that the insurance firms

cannot do better by inducing the agent to choose e = 0. We will later see whether and when they hold.

48
− Let λ and µ denote the Langrangian multipliers for (2.17) and (2.18). Then, the

first-order condition is given by

π1 + λ = µπ1 h0 (u∗ )

and

1 − π1 − λ = µ(1 − π1 )h0 (u∗ ).

− By the same argument as in Part 2.1.4, one can show that both λ and µ are positive

so that (2.17) and (2.18) are binding.

− Letting U ∗ = π1 u∗ +(1−π1 )u∗ −c denote the agent’s expected utility at the second-best,

the binding constraint (2.17) yields

π0
u∗ = U ∗ − c (2.20)
∆π
1 − π0
u∗ = U ∗ + c (2.21)
∆π

− Plugging (2.20) and (2.21) into the binding constraint (2.18) yields
 
∗ 1 − π0 
∗ π0 
π1 h U + + (1 − π1 )h U − c = w − d(1 − π1 ). (2.22)
∆π ∆π

− We need to check that (i) the agent has an incentive to participate or U ∗ ≥ û and (ii)

an insurance company can do better by offering a contract that induces the agent to

choose e = 0.

(1) To see (i), it is enough to note that the pair of utility (u(w), u(w − d)) satisfies

both (2.17) and (2.18) and also yields the expected utility û, which means that U ∗ ,

the expected utility obtained from the problem (P ) above, cannot be smaller than

û, as desired.

49
 (2) For (ii), an insurance company should not be able to make the agent better off by

offering the optimal contract inducing e = 0, which is obtained from solving

max π0 u + (1 − π0 )u − c
(u,u)

subject to

π0 (w − h(u)) + (1 − π0 )(w − d − h(u)) ≥ 0.

The solution of this problem is u = u = u(w − d(1 − π0 )) (Why?). Thus, (ii) will

holds if

U ∗ ≥ u(w − d(1 − π0 )) − c.

Otherwise, e = 0 will be chosen in equilibrium and the equilibrium contract will

be u = u = u(w − d(1 − π0 )).

2.3.2 Moral Hazard in Teams

Consider n agents working in a partnership to produce aggregate output

Q = Q(a1 , a2 , · · · , an ),

where ai ∈ [0, ∞) is the agent i’s action. Assume that the function Q(·) satisfies

∂Q ∂ 2Q ∂ 2Q
> 0, < 0, and ≥ 0.
∂ai ∂a2i ∂ai ∂aj

Suppose that each agent is risk-neutral: With wage wi and effort ai , the agent i’s utility

is wi − gi (ai ) with gi (·) strictly increasing and being convex. Since each agent’s effort

or individual output is not observable, a contract in the partnership can only depend

on the aggregate output: w(Q) = (w1 (Q), w2 (Q), · · · , wn (Q)). We require each wi (·) to

50
be differentiable (almost everywhere). All agents in the partnership share the aggregate

output in the following sense:

n
X
wi (Q) = Q for each Q. (2.23)
i=1

A key observation in this model is the externality among agents: If agents are rewarded for

rasing output, then one agent working hard to raise aggregate output will benefit others

also. This provides a possible source for an inefficient outcome.

• The first-best profile of actions â = (â1 , · · · ân ) solves the following problem
n
X n
X
max (wi (Q) − gi (ai )) = Q(a1 , · · · , an ) − gi (ai ).
(a1 ,··· ,an )
i=1 i=1

− The corresponding first-order condition is given by

∂Q(â)
= gi0 (âi ) for each i. (2.24)
∂ai

When actions are not contractible, the first-best is no longer achievable.

• Given other agents’ actions a−i = (a1 , · · · , ai−1 , ai+1 , · · · , an ), agent i will choose an effort

ai that satisfes the first-order condition:

dwi (Q(ai , a−i )) ∂Q(ai , a−i )

= gi0 (ai ) for each i.
dQ ∂ai

− If the first-best action profile â were a Nash equilibrium, then (2.24) would imply

dwi (Q(âi , â−i ))

= 1 for each i.. (2.25)
dQ

− However, (2.23) tells us that for all a,

n
X dwj (Q(a)) ∂Q(a) ∂Q(a)
= ,
j=1
dQ ∂ai ∂ai

which contradicts with (2.25).

51
Here, the failure to achieve the first-best is partly due to the balanced-budget constraint

(2.23). Indeed, if there is a third party who can act as a budget breaker, then it is possible

to design a contract which achieves the first-best outcome.

• Let a third party pay wi (Q) = Q to each of agents, who in return hand over the entire

output Q and make an up-front payment zi to the third party.

− Given this contract, each agent i optimally chooses ai = âi , the first-best level.

− Two participation constraints need to be satisfied: the third party not losing money

requires
n
X
Q(â) + zi ≥ nQ(â), (2.26)
i=1

while each agent not losing money requires

zi ≤ Q(â) − gi (âi ). (2.27)

− There exist zi ’s that satisfy both (2.26) and (2.27) if

n
X n
X
nQ(â) − gi (âi ) ≥ (n − 1)Q(â) or Q(â) − gi (âi ) ≥ 0,
i=1 i=1

which has to be satisfied at the first-best.

There is some problem with the above contract: If the team performance is better than

the first-best, then the third party will be forced to overpay the agents and lose money.

Then, the agents might be able to exploit this oppotunity by deciding collusively on an

excessively high level of actions. There is an alternative contract, called Mirrlees contract,

that avoids such problem and achieves the first-best.

52
• Each agent is rewarded with bonus bi if output level Q(â) is realized and punished with

a penalty k to be paid to the budget breaker if any other output is realized.

− Under this contract, it is a Nash equilibrium for each agent to choose âi : given that

other agents choose â−i , each agent i will choose âi if bi and k satisfy

bi − gi (âi ) ≥ −k.

− However, this game might have other equilibria in which either only a few agents work

or all shirk.

53
Chapter 3

Incomplete Contracts

So far, the parties could write a complete contract that identifies every contingency which

will arise at a later stage.

• However, the parties’ ability to write such contract might well be limited for the following

reasons

− Unforseen contingencies: “Parties cannot define ex ante the contingencies that may

occur later on.”

− Cost of writing contracts: “Even if one could foresee all contingencies, they might be

so numerous that it would be too costly to describe them in a contract.”

− Cost of enforcing contracts: “Courts must understand the terms of the contract and

verify the contracted upon contingencies and actions in order to enforce the contract.”

• For the above reasons, the contracts could be incomplete in the sense that they leave out

what to do in some contingencies, which has the following implications:

54
− The parties have to rely on ex post bargaining or renegotiation in those contingencies

− The parties’ opportunism at the ex post bargaining stage might deter each other from

taking an efficient action at an earlier stage (e.g. GM-Fisher Body)

− The organizational arrangement such as allocation of ownership rights, authority, or

financial control becomes important

55
3.1 The Holdup Problem

• Consider the following setup with one seller (S) and one buyer (B) where two parties

are not integrated:

− At date 0, the parties sign a contract

− At date 1 (ex ante), each party decides on how much to invest, i ∈ [0, 1] at cost φ(i)

by S, and j ∈ [0, 1] at cost ψ(j) by B. (Assume that φ(·) and ψ(·) are increasing and

convex functions.)

− At date 2 (ex post), the state of nature θ = (v, c), where v is B’s value and c S’s cost,

is realized as follows: v ∈ {vL , vH } with vL < vH and Prob(vH ) = j, and c ∈ {cL , cH }

with cL < cH and Prob(cL ) = i.

− At date 3, two parties trade a quantity q ∈ [0, 1] at price P .

− The resulting payoff levels are vq − P − ψ(j) for B and P − cq − φ(i) for S.

− Assuming that cH > vH > cL > vL , the ex-post efficient level of trade is q = 1 if

θ = (vH , cL ) and q = 0 otherwise. Given this, the ex-ante efficient investment solves

max ij(vH − cL ) − ψ(j) − φ(i),

i,j

which yields the optimal investment levels i∗ and j ∗ as follows:

i∗ (vH − cL ) = ψ 0 (j ∗ ) and j ∗ (vH − cL ) = φ0 (i∗ ). (3.1)

• What happens if i, j, and θ are not contractible, that is the contracts are incomplete?

Take an extreme case where there is no contract signed at date 0.

56
 − Two parties bargain over q and P once θ is realized. According to the Nash bargaining

solution, two parties evenly split the (extra) surplus so the payoffs become

1
ij[ (vH − cL )] − ψ(j)
2

for B and
1
ij[ (vH − cL )] − φ(i)
2
for S, which lead to the first-order condition

1 1
î(vH − cL ) = ψ 0 (ĵ) and ĵ(vH − cL ) = φ0 (î) (3.2)
2 2

− Comparing (3.1) and (3.2), one can see that ĵ < j ∗ and î < i∗ , underinvestment.

3.2 Organizational Solution

One solution to the underinvestment problem due to the incomplete contract is organiza-

tional arrangement such as the integration of firms by (re)defining ownership rights. To

illustrate the point, we modify the above model and assume that only buyer makes an

investment while seller’s cost is fixed at c with vH > c > vL . Also, B and S can find

another, less efficient and competitive, trading partner: seller S 0 with cost c0 ∈ (c, vH ) and

buyer B 0 with v 0 ∈ (c, vH ). Defining χ = vH − c, χS = v 0 − c and χB = vH − c0 , assume that

χ > χS + χB (3.3)

so B and S can generate a greater surplus by trading with each other than with outside

partners. Also, (3.3) implies that the marginal effect of B’s investment is greater within

the relationship or relationship-specific. We consider two other regimes in addition to the

57
no-integration regime: B-integration and S-integration. Under i-integration, the party i

has the ownership of both B and S’s assets. The ownership title gives the owner the right

to all revenues generated by the asset.

• Under no integration, B and S each hold their own assets so that if B and S fail to trade

ex post, then they obtain the outside payoffs χB and χS , respectively.

− Thus, B’s ex ante payoff is

1
j[χB + (χ − χB − χS )] − ψ(j),
2

which leads to the first-order condition

1 1
χB + (χ − χB − χS ) = (χ + χB − χS ) = ψ 0 (jN
∗
). (3.4)
2 2

• Under the B-integration, if B and S fail to trade ex post, B obtains the outside payoff

χB while S obtains zero since he has no asset.

− Thus, B’s ex ante payoff is given by

1
j[χB + (χ − χB )] − ψ(j),
2

which leads to the first-order condition

1 1
χB + (χ − χB ) = (χ + χB ) = ψ 0 (jB∗ ). (3.5)
2 2

• Under the S-integration, if B and S fail to trade ex post, then S obtains the outside

payoff χS while B obtains zero.

− Thus, B’s ex ante payoff is given by

1
j[ (χ − χS )] − ψ(j),
2
58
 which leads to the first-order condition

1
(χ − χS ) = ψ 0 (jS∗ ). (3.6)
2

Which regime is better in terms of generating a higher investment from B? Comparing

(3.4),(3.5), and (3.6), it is straightforward to see that jS∗ < jN

∗
< jB∗ . The lessen here is

that

(1) The ownership protects the investing party against the ex post opportunism of other

parties.

(2) The equilibrium allocation of ownership rights will be determined by the relative

value of each party’s ex ante specific investments.

The above model has many limitations. Among others, the ranking of investment may

depend on the particular bargaining solution assumed. More importantly, the organization

intervention may not be necessary for resolving the holdup problem, in particular when a

simple contract suffices to do so.

3.3 Contractual Solution

In this section, we show that there exists a simple contract that can induce the ex-ante

investment and resolve the holdup problem. Let us keep the model from the previous section

and assume for simplicity that there is no outside partner. We consider two contracts:

specific performance contract (SPC) and option contract (OC).

• The SPC specifies the following action to be taken unless renegotiation occurs: ‘B and

59
S trade a quantity q̂ ∈ (0, 1) at price P̂ .’ We ask if there is some q̂ that induces B to

choose j ∗ . There are two cases to be considered.

− If θ = (vH , c), then the renegotiation will result in q = 1 since it is more efficient than

q̂ < 1. In this case, the B’s payoff will be

1
vH q̂ − P̂ + (1 − q̂)(vH − c).
2

− If θ = (vL , c), then the renegotiation will result in q = 0, which is more efficient than

q̂ > 0. In this case, the B’s payoff will be

1
vL q̂ − P̂ + q̂(c − vL ).
2

− Thus, B’s ex ante payoff from investing j is

1 1
j[vH q̂ − P̂ + (1 − q̂)(vH − c)] + (1 − j)[vL q̂ − P̂ + q̂(c − vL )] − ψ(j),
2 2

which leads to the first order condition

1 1 1
vH (1 + q̂) − q̂vL − c = ψ 0 (j). (3.7)
2 2 2

− If some q̂ can be set to make the LHS of (3.7) equal to (vH − c), then j ∗ , efficient level,

will solve (3.7). Note that with q̂ = 0,

1
LHS of (3.7) = (vH − c) < vH − c
2

while with q̂ = 1,
1 1
LHS of (3.7) = vH − vL − c > vH − c
2 2
since vL < c. Thus, by continuity, there exists q̂ ∈ (0, 1) such that the LHS of (3.7) is

equal to (vH − c), as desired.

60
 To consider the option contract, let us now be more specific about the bargaining

protocol. After θ has been realized, the parties can simultaneously send to each other

letters containing trading offers, which the receiving party can later present as evidence to

the court. We make the simplifying assumption that there are only two trading possibilities,

trade (q = 1) or no trade (q = 0).

• Option contracts are such that ‘S receives P0 in case of no trade, and has the option to

trade at P1 = P0 + K, where P0 and K are set to satisfy K > c and vH > P1 = P0 + K’.

− Note first that according to the contract, S would always want to trade (if there is no

renegotiation) since ‘trade’ will yield P1 − c to S while ‘no trade’ P0 < P1 − c.

− Consider first the case θ = (vL , c). In this case, it is socially efficient to trade, which

will indeed take place since (i) S prefers ‘trade’ over ‘no trade’ and (ii) B cannot make

any renegotiation offer that induces S to choose ‘no trade’ (that is makes S better with

‘no trade’ than with ‘trade at P1 ’) and at the same time makes himself better off.

− Even in the case θ = (vL , c), S would like to trade for the above reason. Then, B

will send a letter offering P00 = P1 − c(+) for ‘no trade’ so that his payoff can be

−P00 = −P1 + c instead vL − P1 (< −P1 + c). Also, B can hide whatever offers has been

sent by S.

− Thus, B’s ex ante payoff is

j[vH − P1 ] + (1 − j)[−P00 ] − ψ(j) = jvH + (1 − j)c − P1 − ψ(j),

which leads to the first-order condition

vH − c = ψ 0 (j),

resulting in j = j ∗ .

61
3.4 Cooperative Investment and Failure of Contracts

It has been assumed so far that investments are selfish in that one party’s investment has

no direct impact on the other’s payoff. Often, however, investments are cooperative. For

example, the Fisher Body’s decision to build a plant near GM may benefit both parties

by reducing shipping costs and so on. This externality turns out to have an important

consequence on the efficiency properties of contracts.

• For a simple model of cooperative investment, assume that only S invests i, which only

benefits B so that Prob(vH ) = i while S’s cost is fixed at c ∈ (vL , vH ).

− The first-best level of investment is defined as before to satisfy

vH − c = φ0 (i∗ ).

− Without any contract, S’s ex ante payoff would be

1
i[ (vH − c)] − φ0 (i)
2

so S chooses î satisfying

1
(vH − c) − φ(î) = 0. (3.8)
2

• Now consider the specific-performance contract (P̂ , q̂). Once θ is realized, the renegoti-

ation proceeds as follows:

− If θ = (vH , c) realizes, then S’s renegotiated payoff is

1
P̂ − cq̂ + (1 − q̂)(vH − c).
2

62
− If θ = (vL , c) realizes, then S’s renegotiated payoff is

1
P̂ − cq̂ + q̂(c − vL ).
2

− Thus, S’s ex ante payoff is given as

1 1
i[P̂ − cq̂ + (1 − q̂)(vH − c)] + (1 − i)[P̂ − cq̂ + q̂(c − vL )] − φ(i),
2 2

whose differentiation with i yields

1 1
[(1 − q̂)vH + q̂vL − c] − φ0 (i) ≤ (vH − c) − φ0 (i). (3.9)
2 2

− Comparing (3.8) and (3.9) shows that S’s investment is no higher with contracts than

without.

63
Chapter 4

Auction Theory

• Why do we study auctions?

− Practical significance; commonly used in many areas

− Good welfare and revenue features

− Serves as a “behavioral” foundation of markets

− Empirical testability

• Different Environments

− Identity of the auctioneer: Seller auction, buyer auction, or double auction

− Number of objects: Single Unit or Multiunit

− Information: Bidders have signals which convey information about the value of auc-

tioned object

(1) Private vs interdependent values (c.f., Mineral rights auction)

64
 (2) Independent vs correlated signals

(3) Single-dimensional vs multidimensional signals

• Different Auction Rules

− Single Unit: First-price, Second-price (Vickrey), Open-oral ascending, Dutch, All-pay,

and many others

− Multi Unit:

(1) Discriminatory price auction (pay your bid)

(2) Uniform price auction (pay highest losing bid)

• Our study focuses on the single-unit IPV model, whose basic setup is given as follows:

− 1 unit on sale

− n ≥ 2 risk-neutral bidders. Let N = {1, · · · , n} denote the set of all bidders.

− Bidder i’s value: θi ∈ Θi := [θi , θi ] drawn according to distribution Fi (·) with density

fi (·)

− Each bidder i knows his value θi while others only know its distribution Fi (·).

− Bidder i’s utility: θi xi − ti , where xi is the probability of bidder i winning the object

and ti the bidder i’s payment. Each bidder’s reservation utility is zero.
Pn
− Seller’s utility: i=1 ti

• Two criteria to evaluate an auction outcome

− Efficiency: Whether the auctioned objects go to the bidders who value them most

highly.

65
− Revenue: How much revenue the seller raise in expectation.

66
4.1 Second-Price (or English) Auction

The auction rule is as follows: The bidder who submits the highest bidder wins the object

and pays the second highest bid. If there is a tie, then the winner is randomly determined

among the highest bidders.

• It is a (weakly) dominant strategy for each bidder i to submit a bid equal to his value θi .

Proof. Consider bidder 1, say, and suppose that p1 := maxj6=i bj is the highest competing

bid. By bidding b1 , bidder 1 will win if b1 > p1 and not win if b1 < p1 . It is straightforward

that bidding θ1 is always (weakly) better than bidding some b1 < θ1 : If p1 ≤ b1 < θ1 or

p1 < b1 < θ1 , then bidder 1 is indifferent between θ1 and b1 . If b1 ≤ p1 ≤ θ1 , then θ1 is

strictly better than b1 . A similar argument shows that bidding θ1 is always better than

bidding some b1 > θ1 . Thus, it is weakly dominant to bid θ1 .

− By the same logic, it is weakly dominant for a bidder in English auction to drop out

when the price reaches his value.

− As a result, the object is allocated to the highest value bidder, who pays the second

highest value. So, the equilibrium allocation is efficient.

− The seller’s revenue is

RSP A := E[θ(2) ],

where θ(2) is the second order statistic.

− Assuming that bidder are symmetric, that is Fi (·) = F (·) for some F : [θ, θ] → [0, 1],

each bidder i of type θi pays in expectation

Z θi
s(n − 1)f (s)F n−2 (s)ds, (4.1)
θ

67
 which will prove useful for comparing the revenues between SPA and FPA.

− It doesn’t matter whether bidders know the others’ values or not since they play the

weakly dominant strategy.

− There are other equilibria which are less reasonable: Bidder 1, say, always bids ∞

while others bid 0 → These equilibria might be utilized by the collusive bidders.

4.2 First-Price Auction

The rule is the same as in the second-price auction except that the winner pays his own

bid. It is very difficult to fully characterize the equilibrium bidding strategy if bidders are

asymmetric, that is Fi (·) 6= Fj (·) for i 6= j. So, we assume that bidders are symmetric. We

focus on (Bayesian) Nash equilibrium in which bidders use the symmetric and increasing

bidding strategy, β : [θ, θ] → R+ .1

• In equilibrium, the optimal bid of bidder i with value θi is

β(θi ) = arg max(θi − b)F n−1 (β −1 (b)).

b∈R+

This problem can be translated into

θi = arg max (θi − β(θ̂))F n−1 (θ̂), (4.2)

θ̂∈[θ,θ]

which must be maximized at θ̂ = θi .

− To differentiate (4.2) with θ̂ and set it equal to zero at θ̂ = θi yields

− β 0 (θi )F n−1 (θi ) + (θi − β(θi ))(F n−1 (θi ))0 = 0, (4.3)
1
It can be shown that the equilibrium bidding strategy we will discuss is indeed unique.

68
 which can be rearranged to yield

0
β(θi )F n−1 (θi ) = θi (F n−1 (θi ))0 ,

and then
Z θi
β(θi )F n−1
(θi ) = s(F n−1 (s))0 ds + K.
θ

Clearly, K = 0. Rearranging once more, we obtain

R θi Rθ R θi n−1
θ
s(F n−1 (s))0 ds θi F n−1 (θi ) − θ i F n−1 (s)ds θ
F (s)ds
β(θi ) = = = θi − , (4.4)
F n−1 (θi ) F n−1 (θi ) F n−1 (θi )

where the second equality follows from the integration by parts.

− The equilibrium bidding function in (4.4) has resulted from only considering the first-

order condition. To check that β(θ) is indeed a global optimum (or second-order con-

dition) for each type θi ∈ [θ, θ], substitute (4.4) into (4.2) to obtain
Z θ̂
n−1
(θi − θ̂)F (θ̂) + F n−1 (s)ds.
θ

Differentiate this with θ̂ to verify

(θi − θ̂)(F n−1 (θ̂))0 T 0 if θ̂ S θi ,

implying that θ̂ = θi achieves the global maximum.

− Applying the formula in (4.4) to the case of uniform distribution F (θi ) = θi on [0, 1]

yields
R θi
0
sn−1 ds n−1
β(θi ) = θi − n−1 = θi .
θi n

− Note that the equilibrium allocation is efficient as in the second-price auction.

69
− The seller’s revenue is equal to
Z θ
RF P A := β(θ)(F n (θ))0 dθ.
θ

In case of the uniform distribution,

θ
n−1 n 0 n−1
Z
RF P A = θ(θ ) dθ = .
θ n n+1

− The seller’s revenue is the same across SPA and FPA since in FPA, the amount each

bidder i of type θi pays in expectation is given as, using the second expression in (4.4),
Z θi
F n−1
(θi )β(θi ) = s(F n−1 (s))0 ds,
θ

which is equal to (4.1).

70
4.3 Revenue Equivalence and Optimal Auction

Here, we ask two questions: (1) Can we generalize the observation that the first-price

and second-price auctions generate the same revenue? (2) Does there exist the auction

mechanism that maximizes the seller’s revenue among all possible auction mechanisms? It

turns out that the answer to both questions are positive. To do so, we allow for bidders to

be asymmetric.

• Let us consider an auction in its most general format: Given a type realization θ =

(θ1 , · · · , θn ), xi (θ) is the probability bidder i wins the object and ti (θ) is his payment.

− For instance, at the equilibrium of the first-price auction with the uniform distribution,

 1, n−1 θi if θi > maxj6=i θj


n
(xi (θ), ti (θ)) =
 (0, 0)

otherwise.

− Call Xi (θi ) := Eθ−i [xi (θi , θ−i )] and Ti (θi ) := Eθ−i [ti (θi , θ−i )] interim wining probability

and interim payment, respectively.

− If bidder i of type θi pretends to be of type θi0 (or uses the equilibrium strategy of

type θi0 ), then his expected payoff is given by ui (θi0 , θi ) := θi Xi (θi0 ) − Ti (θi0 ). Thus, his

equilibrium payoff is given by Ui (θi ) := ui (θi , θi ).

• There are two constraints for the above auction mechanism to work well:

Ui (θi ) ≥ ui (θi0 , θi ) for all θi ∈ [θi , θi ] and θi0 ∈ [θi , θi ] (IC)

Ui (θi ) ≥ 0 for all θi ∈ [θi , θi ]. (IR)

− The following result helps us express the above constraints in a simple form, which

will be instrumental to the ensuing analysis.

71
 Theorem 4.1. The constraint (IC) holds if and only if for all i,

Xi (·) is non-decreasing (M )
Z θi
Ui (θi ) = Ui (θi ) + Xi (s)ds for all θi . (Env)
θi

Proof. (→) To first show that (IC) implies (M ), observe that

θi Xi (θi ) − Ti (θi ) = Ui (θi ) ≥ ui (θi0 , θi ) = θi Xi (θi0 ) − Ti (θi0 )

θi0 Xi (θi0 ) − Ti (θi0 ) = Ui (θi0 ) ≥ ui (θi , θi0 ) = θi0 Xi (θi ) − Ti (θi ),

which yields

θi (Xi (θi0 ) − Xi (θi )) ≤ Ti (θi0 ) − Ti (θi ) ≤ θi0 (Xi (θi0 ) − Xi (θi )),

or

(θi0 − θi )(Xi (θi0 ) − Xi (θi )) ≥ 0,

implying (M ). To next show that (IC) implies (Env), observe that

θi = arg max ui (θi0 , θi ) and Ui (θi ) = max ui (θi0 , θi ). (4.5)

θi0 ∈[θi ,θi ] θi0 ∈[θi ,θi ]

Thus, applying the envelope theorem to (4.5), we obtain

∂ui (θi0 , θi ) 

dUi (θi )
= = Xi (θi ),
dθi ∂θi 0
θ =θi i

which implies (Env).

(←) Do this for yourself.

• The seller’s revenue is given by

" n
# n
X X
max E ti (θ) = E [Ti (θi )]
(xi (·),ti (·))n
i=1
i=1 i=1

72
 − Using (Env), we can make the following substitution
Z θi
Ti (θi ) = θi Xi (θi ) − Xi (s)ds − Ui (θi ). (4.6)
θ

− Then,
n n Z
!
X X θi Z θi
E [Ti (θi )] = θi Xi (θi ) − Xi (s)ds − Ui (θi ) fi (θi )dθi ,
i=1 i=1 θi θi

which results in the following theorem:

Theorem 4.2 (Revenue Equivalence). If two auction mechanisms generate the

same allocation and the same utility to the lowest type of each bidder, then they yield

the same expected revenue to the seller.

Let us now search for the optimal auction which maximizes the seller’s revenue.

• The seller’s problem is to solve

" n
# n
X X
max E ti (θ) = E [Ti (θi )]
(xi (·),ti (·))n
i=1
i=1 i=1

subject to (IC) and (IR), or equivalently (M ), (Env), and (IR).

− From (4.6), we have

!
Z θi Z θi
E[Ti (θi )] = θi Xi (θi ) − Xi (s)ds − Ui (θi ) fi (θi )dθi
θi θi
 θi 
1 − Fi (θi )
Z
= θi − Xi (θi )fi (θi )dθi − Ui (θi )
θi fi (θi )
  
1 − Fi (θi )
= E θi − xi (θ) − Ui (θi )
fi (θi )

− The seller’s revenue can be expressed as

" n # " n # n
X X X
E ti (θ) = E Ji (θi )xi (θ) − Ui (θi ),
i=1 i=1 i=1

73
 where
1 − Fi (θi )
Ji (θi ) =: θi − ,
fi (θi )
called “virtual valuation.” We assume that Ji (·) is non-decreasing, as is satisfied by

most distribution functions.

− Let us temporarily ignore (M ). Then, the optimal auction mechanism requires that (i)

the (IR) constraint for the lowest type be binding or Ui (θi ) = 0 and (ii) the allocation

rule be given as

 1

if Ji (θi ) > max{maxj6=i Jj (θj ), 0},
x∗i (θ) =
 0

otherwise.
Note that this allocation rule satisfies (M ) since Ji (·) is nondecreasing.

− Let us assume that bidders are symmetric or Fi (·) = F (·), which implies that Ji (·) =

J(·). Then, 
 1

if θi > maxj6=i θj and θi > θ̂,
x∗i (θ) =
 0

otherwise,

where θ̂ := J −1 (0).

− In case of uniform distribution, J(θ) = 2θ − 1 and thus θ̂ = 1/2. In the optimal

auction, the object is sold to the highest value bidder only if his value is greater than
n−1 1 1 n


1/2. The optimal revenue is n+1 + n+1 2
.

− How can we practically implement the optimal auction mechanism above? → Use the

first-price or the second-price auction with a reserve price set at θ̂. For instance, in the

first-price auction with a reserve price θ̂, the equilibrium bidding strategy is given as
 R θi n−1
 F (s)ds
 θi − θ̂

if θi ≥ θ̂
F n−1 (θi )
β(θi ) =

 0

otherwise.

74
 − What will the optimal auction look like if bidders are asymmetric? Consider the

following example with two bidders: F1 (θ1 ) = θ and F2 (θ) = θ2 , so bidder 2 is stronger

(why?). Note that

1 − F1 (θ) 1 − θ2 1 − F2 (θ)
=1−θ < = .
f1 (θ) 2θ f2 (θ)

Thus, the optimal auction favors the weaker bidder, or handicaps the stronger bidder.

The idea behind the optimal auction is not much different from that of the monopoly

pricing. Consider a monopolist who faces n markets and assume that the market i is

populated by consumers whose value distribution follows Fi (·). So, the demand function

and the total revenue in the market i are q = 1 − Fi (p) and qFi−1 (1 − q), respectively. The

marginal revenue is thus

d 1 1 − Fi (p)
(qFi−1 (1 − q)) = Fi−1 (1 − q) − q −1 =p− .
dq fi (Fi (1 − q)) fi (p)

Thus, the virtual valuation of a bidder can be interpreted as a marginal revenue. The

allocation rule of the optimal auction is to sell the good to the buyer associated with the

highest marginal revenue as long as it is nonnegative.

4.4 Extensions

In this section, we relax three important assumptions we have adopted so far in order to

make the model more realistic: (1) risk-neutrality of bidders, (2) bidders with unlimited

budget, and (3) independent value distributions. We will see that relaxing the above

assumptions leads us to quite different conclusion than before. It turns out that if (1)

or (2) is relaxed, then two auctions are no longer equivalent with the seller favoring the

75
first-price auction. If (3) is relaxed, then the seller may be able to extract the entire surplus

from bidders.

4.4.1 Risk Averse Bidders

So far, we have assumed that bidders are risk neutral in the sense that each bidders’ utilities

are linear in net surplus (= value minus payment). We ask here what happens to the seller’s

revenue if bidders are risk averse so that each bidder’s utility is represented by a strictly

concave function u : R+ → R satisfying u(0) = 0, u0 > 0, and u00 < 0. Then, if bidder with

value θ wins and pays b, his utility is given as u(θ − b). We maintain the assumption that

bidders’ values are identically and independently distributed in the interval [0, 1].

Proposition 4.1. With risk averse bidders, the expected revenue is higher in the first-price

auction than in the second-price auction.

Proof. In the second-price auction, it is still a dominant strategy for each bidder to bid his

or her value.

Let us examine the equilibrium strategy for the first-price auction. Letting γ : [0, 1] →

R+ denote the (symmetric) equilibrium bidding strategy, it must be that γ(0) = 0 (why?).

Also, each bidder i of type θi > 0 must solve

max
0
G(θi0 )u(θi − γ(θi0 )),
θi

where G(θi0 ) = F n−1 (θi0 ). The first-order condition for this problem is

g(θi )u(θi − γ(θi0 )) − G(θi0 )γ 0 (θi )u0 (θi − γ(θi0 )) = 0.

76
Setting θi0 = θi and rearranging yield

u(θi − γ(θi )) g(θi )

γ 0 (θi ) = . (4.7)
u0 (θi − γ(θi )) G(θi )

If bidders were risk-neutral or u(x) = x, then (4.7) would become

g(θi )
β 0 (θi ) = (θi − β(θi )) ,
G(θi )

which is the same expression as (4.3) after a rearrangement. Since u being strictly concave

with u(0) = 0 implies u(x)/x > u0 (x) or u(x)/u0 (x) > x, we obtain the following: Whenever

γ(θi ) = β(θi ),

u(θi − γ(θi )) g(θi ) g(θi ) g(θi )

γ 0 (θi ) = 0
> (θi − γ(θi )) = (θi − β(θi )) = β 0 (θi ).
u (θi − γ(θi )) G(θi ) G(θi ) G(θi )

Now, we make use of the following lemma:

Lemma 4.1. Suppose f : R → R is a continuously differentiable such that f (x) ≥ 0.

Suppose that for all x ≥ x, f 0 (x) > 0 whenever f (x) = 0. Then, f (x) > 0 for all x > x.

Defining f (θi ) := γ(θi ) − β(θi ), the function f (·) satisfies the condition in Lemma 4.1

so f (θi ) = γ(θi ) − β(θi ) > 0 for all θi ∈ (0, 1]. To conclude, in the first-price auction, the

risk aversion causes an increase in equilibrium bids while it does not in the second-price

auction.

Why does risk aversion lead to higher bid in the first-price auction? Raising one’s bid

slightly in the first-price auction is analogous to buying partial insurance: it reduces the

probability of a zero payoff and increases the chance of winning at a lower profit margin.

Risk averse bidders value this insurance, so they bid more than they would if they were

risk neutral.

77
4.4.2 Budget Constrained Bidders

Let us relax the assumption that bidders are not constrained in budgeting their bids. An

alternative assumption, among others, would be that bidders face a fixed budget equal

to B, so they cannot make any bid higher than B. Letting β(·) denote the equilibrium

strategy in the first-price auction with no budget constraint, we assume B < β(1) to avoid

a trivial case.

• In the second-price auction, each player has a dominant strategy:

βS (θ) = min{θ, B}.

− The proof follows the same line as when there is no budget constraint. (Try for

yourself.)

• In the first-price auction, a candidate equilibrium is the following:


 β(θ) if θ ≤ θ̂

βF (θ) :=
 B

if θ > θ̂,

where θ̂ satisfies
n−1  
n−1
X 1 n−1
(θ̂ − β(θ̂))F (θ̂) = (θ̂ − B) (1 − F (θ̂))k F n−k−1 (θ̂).
k=0
k+1 k

− Here, θ > θ̂ bids B while θ < θ̂ bids β(θ) as if the budget were absent. And θ̂ is

indifferent between B and β(θ̂).

− Note that B > β(θ̂) so the equilibrium strategy jumps at θ̂.

− Given all other bidders employ βF (·), it is optimal for bidder i of type θi to bid βF (θi ).

(Why?)

78
• We have the following proposition.

Proposition 4.2. With the same budget B, the expected revenue is higher in the first-

price auction than in the second-price auction.

Proof. The equilibrium allocation in the first-price auction is the same as the second-

price auction with a larger budget equal to θ̂. Thus, by the revenue equivalence theorem,

the first-price auction with budget B and the second-price auction with budget θ̂ > B

generate the same revenue, which is higher than the revenue generated by the second-

price auction with budget B.

− A rough intuition behind this result is as follows: Without budget constraint, any given

type of bidder makes the same interim (or average) payment in two auction formats

but the payment in the SPA is random and thus sometimes exceeds the average. This

makes the budget constraint more likely to bind in the SPA, which translates into a

higher revenue of the FPA.

4.4.3 Optimal Auction with Correlated Values

We now relax the assumption that values are independently distributed. Relaxing this

assumption leads to a striking result in terms of the design of optimal auction: The seller

can achieve the first-best allocation and extract all the surplus from the bidders. Here, this

result is shown in a simple 2 × 2 model, though it is much more general.

• Assume that there are 2 bidders and 2 possible valuations, θH and θL < θH , for each

bidder.

79
 − Let pij denote the probability that bidder 1 has θi and bidder 2 has θj .
pHH pLL
− Suppose that valuations are correlated: pLL pHH − pLH pHL 6= 0: Letting ψ := ,
pHL pLH
if ψ > 1 (< 1), then values are positively (negatively) correlated.

• Consider the general auction rule: xij = the probability that bidder 1 gets the object

and tij = bidder 1’s payment, when his value is i and bidder 2’s value is j.

− The full extraction means that (i) the allocation must be efficient

1
xHH = xLL = , xHL = 1, and xLH = 0.
2

(ii) the individual rationality for both types must be biding:

 
θL
pLL − tLL − pLH tLH = 0
2
 
θH
pHH − tHH + pHL (θH − tHL ) = 0,
2

from which we have

 
pLL θL
tLH = − tLL (4.8)
pLH 2
 
pHH θH
tHL = θH + − tHH . (4.9)
pHL 2

− Also, we need to satisfy the incentive compatibility condition:

   
θL θL
0 = pLL − tLL − pLH tLH ≥ pLL (θL − tHL ) + pLH − tHH (ICL )
2 2
 
θH θH
0 = pHH − tHH + pHL (θH − tHL ) ≥ −pHH tLH + pHL ( − tLL ). (ICH )
2 2

Substituting (4.8) and (4.9) into this yields

pLH
0≥ (θL − ψθH ) − pLL (θH − θL ) + pLH (ψ − 1)tHH (4.10)
2
pLH
0≥ (θH − ψθL ) + pHL (ψ − 1)tLL . (4.11)
2

80
 − Thus, we are done by choosing tHH and tLL that satisfy (4.10) and (4.11).

• Some remarks are in order.

− If ψ = 1 or values are independent, then it is impossible to satisfy the above inequali-

ties, which is what we already know from the analysis of independent types.

− As ψ gets larger beyond 1 or values become more positively correlated, (4.10) and

(4.11) can be satisfied by setting tHH and tLL lower while setting tHL and tLH relatively

higher: This is effective in giving bidders an incentive to tell truthfully whey types are

highly positively correlated.

− If ψ > 1 but ψ ' 1, then tHH and tLL have to be very large, which may cause some

problem for a budget-constrained bidder.

4.5 Appendix: Envelope Theorem

Let X be the choice set and [t, t] the parameter set. Consider the following problem:

max f (x, t).

x∈X

Let V (t) := maxx∈X f (x, t), called value function, and x(t) := arg maxx∈X f (x, t), called

the set of maximizers. Assume for simplicity that x(t) is a singleton (though it need not

be). We have the following theorem.

Theorem 4.3 (Envelope Theorem). Take t ∈ [t, t] and x ∈ x(t) and suppose that
∂f
f2 (x, t) exists, where f2 (x, t) = ∂t
. Then, V 0 (t) = f2 (x(t), t) whenever V 0 (t) exists.

81
Example: Consider f (x, t) = (3t − 1)x − 21 x2 and X = [0, 1] and t ∈ R. Then,
 
1
if t ≤ 3 if t ≤ 13
 


 0 

 0

 

x(t) = 1
3t − 1 if 3 < t < 3 2 and V (t) = 1 2 1 2

  2 (3t − 1) if 3 < t <
 3
 
2  3t − 2 if t ≥ 23
 
 1 if t ≥ 3
 
3

One can verify 

if t ≤ 13



 0


V 0 (t) = f2 (x(t), t) = 3(3t − 1) if 13 < t < 2
 3


if t ≥ 23 ,

 3


as desired by Theorem 4.3.

In addition to the assumptions in Theorem 4.3, let us assume that V 0 (t) exists almost

everywhere.2 Then, we have the envelope theorem in the integral form:

Z t
V (t) = V (t) + f2 (x(s), s)ds. (4.12)
t

2
This is equivalent to assuming that the function V (·) is absolute continuous.

82
Chapter 5

Matching Theory

In this part, we will study the very basics of matching theory, which is becoming more

popular among economists both theoretically and practically. A typical matching problem

is described by a situation in which there are agents on two sides of a market who each

wish to match with agent(s) on the other side.

• The examples include matchings between

− Men and women in a ‘marriage market’

− Firms and workers in a labor market

− Medical interns and Hospitals

− Students and colleges/high schools

− Tenants and houses

− Donators and patients in organ exchanges

83
 A matching mechanism is a rule that specifies a matching for each preference profile of

all involved agents.

• A desirable matching rule is expected to satisfy the following properties:

− Pareto efficiency: Ex-ante or ex-post

− Stability: No incentive for agents to break up their matches

− Strategy-proofness: No incentive for agents to misreport their preferences

A matching rule that satisfies the above properties is hard to come by since they need

to be satisfied by many matchings that can arise from all possible preference profiles of

agents on two sides. It is an important task of the matching theory to find such a rule.

84
5.1 Stable Matchings

Our benchmark model is the marriage problem in which there are two finite and disjoint

sets M and W : M = {m1 , m2 , · · · , mn } is the set of men, and W = {w1 , w2 , · · · , wp } is the

set of women. The preference of each man m is represented by an ordered list, P (m), on

the set W ∪ {m}: For example, P (m) = w3 , w1 , m, w2 , · · · , wp indicates that w3 is m’s first

choice, w1 is his second choice, remaining single is his third choice, and so on. Sometimes,

we just write P (m) = w3 , w1 , ignoring the part of list that is worse than remaining single.

For another example, P (m0 ) = w2 , [w1 , w3 ], m0 , · · · , wk indicates that for man m0 , w2 is his

first choice, w1 and w3 are his second choice and indifferent, and so on. We write w >m w0

to mean that m prefers w to w0 , and w ≥m w0 to mean that m likes w at least as well as

w0 . We say that w is acceptable to m if m likes w at least as well as remaining single, that

is w ≥m m, and otherwise w is unacceptable to m. The same notions apply to women as

well.

• A matching µ is a one-to-one mapping from the set M ∪ W onto itself such that for each

m ∈ M and w ∈ W , µ(m) ∈ W ∪ {m}, µ(w) ∈ M ∪ {w}, and µ(m) = w if and only if

µ(w) = m.

− We say that the matching µ is blocked by an individual, say m, if m > µ(m), or blocked

by a pair, say (m, w), if w >m µ(m) and m >w µ(w).

− A matching µ is individually rational if it is not blocked by any individual.

− A matching is stable if it is not blocked by any individual or any pair of agents.

− What is the justification about requiring matching to be stable? → An unstable

matching would not sustain if agents are well informed of one another’s preference and

85
 can easily access each other.

− One can show that any stable matching is Pareto-optimal in the sense that there is no

other matching that makes some agents better off without making anyone worse off.

(Try to prove this for yourself.)

• Let us consider the following example with three men and three women:

P (m1 ) = w2 , w1 , w3 P (w1 ) = m1 , m3 , m2

P (m2 ) = w1 , w3 , w2 P (w2 ) = m3 , m1 , m2

P (m3 ) = w1 , w2 , w3 P (w3 ) = m1 , m3 , m2 .

− The matching µ given by

w1 w2 w3
µ=
m1 m2 m3

is unstable since (m1 , w2 ) is a blocking pair

− The matching given by

w1 w2 w3
µ0 =
m1 m3 m2
is stable.

− There is another stable matching

w1 w2 w3
µ00 = .
m3 m1 m2

− The above example shows that there can be multiple stable matchings. Some inspec-

tion reveals that every man likes µ00 as well as µ0 while every woman likes µ0 as well as

µ00 .

86
 − A matching is man-optimal if it is stable and every man likes it as well as any other

stable matching. Also, a matching is woman-optimal if it is stable and every woman

likes it as well as any other stable matching.

5.1.1 Gale-Shapley Algorithm

We ask if there is a matching rule or algorithm that always yields a stable matching whatever

preferences agents have.

• Gale and Shapley provided one such algorithm, called deferred acceptance procedure.

− The man-proposing deferred acceptance procedure consists of the following steps:

Step 1: Each man proposes to his first choice among the acceptable women on his

preference list; Each woman only holds her most preferred among the acceptable men

who have proposed to her, rejecting all others;

In general, at

Step k: Each man who was rejected in the previous step proposes to his next choice

among the acceptable women on his list; Each woman only holds her most preferred

among the group consisting of the new acceptable proposers and any man she has held

from the previous step, rejecting all others;

Termination: The process terminates when a step is reached in which no new man is

rejected; After termination, each woman is married to the man she is holding.

− Let µM denote the matching that results from this procedure.

− There is also the woman-proposing deferred acceptance procedure in which the roles

of man and woman are just switched.

87
− Consider the following example with 5 men and 4 women:

P (m1 ) = w1 , w2 , w3 , w4 P (w1 ) = m2 , m3 , m1 , m4 , m5

P (m2 ) = w4 , w2 , w3 , w1 P (w2 ) = m3 , m1 , m2 , m4 , m5

P (m3 ) = w4 , w3 , w1 , w2 P (w3 ) = m5 , m4 , m1 , m2 , m3

P (m4 ) = w1 , w4 , w3 , w2 P (w4 ) = m1 , m4 , m5 , m2 , m3

P (m5 ) = w1 , w2 , w4

Step 1: m1 , m4 , and m5 propose to w1 , who only holds m1 ; m2 and m3 propose to w4 ,

who only holds m2 . This yields a tentative matching

w1 w2 w3 w4
;
m1 m2

Step 2: m3 , m4 , and m5 propose to w3 , w4 , and w2 , respectively; w4 rejects m2 and

holds m4 . The tentative matching is thus revised to

w1 w2 w3 w4
;
m1 m5 m3 m4

Step 3: m2 proposes to w2 , who rejects m5 and holds m2 . The tentative matching is

thus revised to
w1 w2 w3 w4
;
m1 m2 m3 m4

Step 4: m5 proposes to w4 , who rejects m5 and holds m4 , which stops the procedure

and results in
w1 w2 w3 w4 (m5 )
µM = .
m1 m2 m3 m4 m5

88
 − In our previous example with three men and three women, the man-proposing algo-

rithm yields µ00 while the woman-proposing algorithm yields µ0 .

• A matching arising from the man-proposing (resp. woman-proposing) deferred accep-

tance procedure is man-optimal (resp. woman-optimal):

Theorem 5.1. Suppose that every agent has a strict preference list. Then, the matching

µM is man-optimal.

− Proof of stability: Suppose that man m and woman w are matched in µM but

m prefers another woman w0 . Then, it must be that at some step of the algorithm, m

proposes to w0 and she rejects m in favor of a man, say m0 , whom she prefers. This,

given how the algorithm works, means that w0 is matched in µM with a man as least

as high-ranked as m0 on her list, which implies that (m, w0 ) is not a blocking pair.

− Proof of man-optimality: Let us say that w is possible for m if there is some

stable match at which these two are matched. We show by induction that there is no

round at which a women rejects a man for whom she is possible.

Induction hypothesis: Up to some step t, no woman has rejected a man for whom

she is possible.

We now show that if at step t + 1, some woman w rejects some man m in favor of some

other man m0 , then w is not possible for m. To this end, we consider any matching µ

such that µ(m) = w and show that it is not stable. Suppose to the contrary that µ is

stable. Then, w0 := µ(m0 ) is possible for m0 so, by the above hypothesis, m0 has yet to

propose to w0 through step t (why?), which implies that m0 prefers w to w0 . Thus, a

pair (m0 , w) blocks µ, a contradiction.

89
 − Symmetrically, the woman-proposing deferred acceptance algorithm yields a stable

matching that is woman-optimal.

− One can also show that when all agents have strict preferences, the interests of two

sides are opposed on the set of stable matchings: If µ and µ0 are stable, then all men

like µ at least as well as µ0 if and only if all women like µ0 at least as well as µ.

Proof. Let µ and µ0 stable and µ(m) ≥m µ0 (m) for all m ∈ M . Suppose to the contrary

that for some w ∈ W , µ(w) >w µ0 (w). We show that µ0 cannot be stable. Let m = µ(w).

Then, m 6= µ0 (w), so µ0 (m) 6= w. By strict preference, m prefers w to µ0 (w). Hence,

(m, w) blocks µ0 .

− Moreover, the set of stable matchings has an algebraic structure, called lattice.

5.1.2 Incentive Problem

In a situation where agents’ preferences are unknown, a stable matching can be implemented

by asking each agent to report his/her preference list and then running, for instance, the

deferred acceptance procedure according to the reported preference profile. With no mea-

sures to check the truthfulness of each agent’s report, however, there arises an incentive

problem, i.e. whether each agent is willing to report his/her true preference, given the

specified matching algorithm. To investigate this problem, we do not confine ourselves to

Gale-Shapley algorithm but consider all possible matching mechanisms:

Definition 5.1. A matching mechanism is a mechanism in which agents report their

(possibly false) preferences and which maps the reported preference profile to a matching

outcome.

90
A matching mechanism defined this way will be without an incentive problem if it is

strategy-proof in the following sense:

Definition 5.2. A matching mechanism is strategy-proof if it is a dominant strategy for

each agent to report his/her true preference.

Then, we obtain an impossibility result as follows.

Proposition 5.1. There does exist no stable matching mechanism that is strategy-proof.

Proof. The proof is done with an example with two men and two women:

P (m1 ) = w1 , w2 P (w1 ) = m2 , m1

P (m2 ) = w2 , w1 P (w2 ) = m1 , m2 .

There are only two stable matchings, µ and ν: µ(mi ) = wi for i = 1, 2 and ν(mi ) =

wj for i = 1, 2 and j 6= i. Consider any stable mechanism that chooses µ. Observe

that if w2 reports a preference Q(w2 ) = m1 while all others report their true preferences,

then ν is the only stable matching with respect to the reported preference profile P 0 =

(P (m1 ), P (m2 ), P (w1 ), Q(w2 )) so any stable mechanism must select ν when P 0 is reported.

Thus, w2 is better off being matching with m1 after reporting Q(w2 ) instead of her true

preference.

Despite the negative result in Proposition 5.1, Gale-Shapley algorithm works well with

respect to the incentives of proposers:

• Given the man-proposing algorithm, it is a dominant strategy for each man to report his

true preference.

91
 − For a proof, refer to Roth and Sotomayor (1992).

− Combined with this observation, Proposition 5.1 implies that women do have incentive

to misrepresent their preference, as was showcased in the proof of Proposition 5.1.

− Let us go back to the previous example with three men and three women. If the

man-proposing algorithm is run so that men are reporting truthfully, then women can

report false preferences as follows:

Q(w1 ) = m1 Q(w2 ) = m3 Q(w3 ) = m2 ,

which will result in µ0 , woman-optimal matching. Thus, it is a Nash equilibrium for

agents to report P 0 = (P (m1 ), P (m2 ), P (m3 ), Q(w1 ), Q(w2 ), Q(w3 )).(Why?)

5.1.3 Extension to Many-to-One Matching

In many matching situations such as college admission problem or internship assignment

problem, an agent (college or hospital) on one side is matched to more than one agent (stu-

dent or intern) on the other side. Our analysis in the previous section can be easily extended

to this many-to-one matching problem. A prime example of the many-to-one matching is

the college admission problem in which there are a set of colleges, C = {C1 , · · · , Cn }, and

a set of students, S = {s1 , · · · , sm }. Each college C ∈ C has a quota of qC ≥ 1. and

has preference over individual students, denoted P (C) while each student can only attend

one school and has preference over C. Each college also has preference over 2S , i.e. sets

of students. We assume that the preferences of colleges over 2S are responsive in the fol-

lowing sense: For any college C and two sets of students A and A0 that are identical with

|A| = |A0 | ≤ qC except that s ∈ A\A0 and s0 ∈ A0 \A for some s, s0 ∈ S, we have A >C A0

92
if and only if s >C s0 .

• A matching is a mapping µ that satisfies the followings: (i) µ(s) ∈ C ∪ {s}; (ii) µ(C) ⊂ S

with |µ(C)| ≤ qC ; (iii) µ(s) = C if and only if s ∈ µ(C).

− As before, a matching µ is (pairwise) stable if it is not blocked by any individual agent

or any college-student pair, that is if there is no pair (C, s) for which C >s µ(s) and

s >C s0 for some s0 ∈ µ(C).

− There is a stronger notion of stability, called group stability, that requires a matching

to be immune to blocking by a coalition of students and colleges. The group stability,

however, is equivalent to the pairwise stability if colleges have responsive preferences.

There is an easy way to translate the college admission model into a related marriage

problem. For a college C, treated as a men, create qC “men”, c1 , c2 · · · , cqC , each of whom

has the same preference over students as C. The preference list of each student, treated as

a woman, is then modified by replacing each entry C, wherever it appear on her list, by

the string c1 , c2 , · · · , cqC .

• Some, but not all, results from the marriage problem generalize to college admission

problem.

− A matching of the college admission problem is stable if and only if the corresponding

matchings of the related marriage problem are stable.

− When the colleges’ preferences over individual students are strict, the matching result-

ing from the student-proposing (college-proposing, resp.) mechanism is student-optimal

(college-optimal, resp.).

93
− In a matching mechanism that yields the student-optimal stable matchings, it is a

dominant strategy for students to report their true preferences.

− However, there does not exist any matching mechanism in which it is a dominant

strategy for colleges to report their true preferences. This is proved with the following

example in which there are four students and three colleges with q1 = 2 and q2 = q3 = 1:

P (s1 ) = C3 , C1 , C2 P (C1 ) = s1 , s2 , s3 , s4

P (s2 ) = C2 , C1 , C3 P (C2 ) = s1 , s2 , s3 , s4

P (s3 ) = C1 , C3 , C2 P (C3 ) = s3 , s1 , s2 , s4

P (s4 ) = C1 , C2 , C3 .

There is a unique stable matching,

C1 C2 C3
µ= .
{s3 , s4 } {s2 } {s1 }

Suppose now that C1 reports a false preference P 0 (C1 ) = s1 , s4 , C1 while all other

agents report their true preferences, so that the reported preference profile is P 0 =

(P 0 (C1 ), P (C2 ), P (C3 ), P (s1 ), · · · , P (s4 )). Given P 0 , a unique stable matching is

C1 C2 C3
µ0 = ,
{s1 , s4 } {s2 } {s3 }

which is preferred by C1 to µ(C3 ).

− The above matching µ has another problem: There is a (unstable) matching that is

preferred by every college,

C1 C2 C3
µ00 = .
{s2 , s4 } {s1 } {s3 }

94
 This kind of problem does not arise in the one-to-one matching.

− Consider another example in which there are three colleges with qi = 1, i = 1, 2, 3 and

three students:

P (s1 ) = C2 , C1 , C3 P (C1 ) = s1 , s3 , s2

P (s2 ) = C1 , C2 , C3 P (C2 ) = s2 , s1 , s3

P (s3 ) = C1 , C2 , C3 P (C3 ) = s2 , s1 , s3 .

It is easy to check there is only one stable matching,

C1 C2 C3
.
{s1 } {s2 } {s3 }

This matching, however, is Pareto-dominated for the students by a (unstable) matching,

C1 C2 C3
.
{s2 } {s1 } {s3 }

The last example shows that if one considers that only students’ preferences are important

in evaluating the efficiency of matchings, then the stability may conflict with the Pareto

efficiency.

5.2 Efficient Matching

In this section, we study the matching problem in which the preferences of agents on one

side are not as important as those of agents on the other side. As shown in the last example

of Subsection 5.1.3, the stability and Pareto efficiency may not stand together as far as the

preferences of students are concerned. Here, we ask whether by forgoing the stability

95
requirement, we can find a matching mechanism that always yields the Pareto efficiency

for the students while maintaining the strategy-proofness.

5.2.1 Top Trading Cycles Algorithm

We adopt the same setup as in Subsection 5.1.3 except that colleges are called (high) schools

and colleges’ preferences are treated as schools’ priority rankings. For instance, a priority

ranking of a student for a given school can be determined according to: (i) whether (s)he

has a sibling in that school; (ii) how close the school is to the student’s residence; and so

on. Students in the same priority group are randomly ordered by a lottery. We assume

that the prior rankings of schools are publicly known before a matching is implemented

while the preferences of students are only known to themselves.

• The following matching mechanism, called top trading cycles, yields a Pareto efficient

matching from the students’ perspectives

− Given the reported preference profile, the top trading cycles mechanism is run in the

following steps:

Step 1: Each student points to his/her first choice among the schools; Each school

points to the student who has the highest priority for the school; There must be at

least one circle (why?), say (s1 , C1 , s2 , · · · , sk , Ck ), in which s1 points to C1 , C1 points

to s2 , ... , sk points to Ck , and Ck points to s1 . Every student in a circle is assigned a

seat in the school (s)he points to and is removed. Also, each school removed as soon

as its quota is filled.

In general, at

96
 Step k: Each remaining student points to his/her first choice among the remaining

schools; Each remaining school points to the student who has the highest priority for

the school; Whatever circle is found, every student in that circle is assigned a seat in

the school she points to and is removed. Also, each school is removed as soon as its

quota is filled.

Termination: The algorithm terminates when all students are assigned a seat or all

schools fill their quotas.

− Consider the last example in subsection 5.1.3 and apply the top trading cycle. In the

fist step, there forms a circle (s1 , C2 , s2 , C1 ). Once these are removed, in the second

step, there forms a circle (s3 , C3 ). The resulting matching is

C1 C2 C3
,
{s2 } {s1 } {s3 }

which is Pareto-efficient for the students.

− The top trade cycles mechanism is Pareto efficient for the students.

− The top trading cycles mechanism is strategy-proof for the students.

97