This action might not be possible to undo. Are you sure you want to continue?

of Mechanism Design

Tilman B¨orgers

December 1, 2008

Preliminary draft.

Contents

Preface v

1 Introduction 1

2 Screening 6

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.2 Pricing a Single Indivisible Good . . . . . . . . . . . . . . . . 6

2.3 Nonlinear Pricing . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.4 Bundling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

2.5 Comments on the Literature . . . . . . . . . . . . . . . . . . . 30

2.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3 Bayesian Mechanism Design 31

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3.2 Single Unit Auctions . . . . . . . . . . . . . . . . . . . . . . . 31

3.2.1 Set-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

3.2.2 Mechanisms, Direct Mechanisms, and the Revelation

Principle . . . . . . . . . . . . . . . . . . . . . . . . . 33

3.2.3 Characterizing Incentive Compatibility and Individ-

ual Rationality . . . . . . . . . . . . . . . . . . . . . . 37

3.2.4 Expected Revenue Maximization . . . . . . . . . . . . 39

3.2.5 Maximizing Welfare . . . . . . . . . . . . . . . . . . . 42

3.2.6 Numerical Examples . . . . . . . . . . . . . . . . . . . 43

3.3 Public Goods . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

3.3.1 Set-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

3.3.2 Incentive Compatible and Individually Rational Di-

rect Mechanisms . . . . . . . . . . . . . . . . . . . . . 47

3.3.3 Ex ante and Ex post Budget Balance . . . . . . . . . 48

3.3.4 Welfare Maximization . . . . . . . . . . . . . . . . . . 50

i

CONTENTS ii

3.3.5 Proﬁt Maximization . . . . . . . . . . . . . . . . . . . 58

3.3.6 A Numerical Example . . . . . . . . . . . . . . . . . . 59

3.4 Bilateral Trade . . . . . . . . . . . . . . . . . . . . . . . . . . 64

3.4.1 Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

3.4.2 Direct Mechanisms . . . . . . . . . . . . . . . . . . . . 65

3.4.3 Welfare Maximization . . . . . . . . . . . . . . . . . . 67

3.4.4 Proﬁt Maximization . . . . . . . . . . . . . . . . . . . 74

3.4.5 A Numerical Example . . . . . . . . . . . . . . . . . . 75

3.5 Comments on the Literature . . . . . . . . . . . . . . . . . . . 76

3.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

4 Dominant Strategy Mechanisms 78

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

4.2 Single Unit Auctions . . . . . . . . . . . . . . . . . . . . . . . 81

4.2.1 Set-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

4.2.2 Mechanisms, Direct Mechanisms, and the Revelation

Principle . . . . . . . . . . . . . . . . . . . . . . . . . 81

4.2.3 Characterizing Dominant Strategy Incentive Compat-

ibility and Ex Post Individual Rationality . . . . . . . 83

4.2.4 Canonical Auctions . . . . . . . . . . . . . . . . . . . . 84

4.3 Public Goods . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

4.3.1 Set-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

4.3.2 Direct Mechanisms . . . . . . . . . . . . . . . . . . . . 87

4.3.3 Characterizing Dominant Strategy Incentive Compat-

ibility and Ex Post Individual Rationality . . . . . . . 88

4.3.4 Canonical Mechanisms . . . . . . . . . . . . . . . . . . 90

4.3.5 Ex Post Budget Balance . . . . . . . . . . . . . . . . . 91

4.4 Bilateral Trade . . . . . . . . . . . . . . . . . . . . . . . . . . 93

4.4.1 Set-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . 93

4.4.2 Dominant Strategy Incentive Compatible and Ex Post

Individually Rational Direct Mechanisms . . . . . . . 93

4.4.3 Canonical Mechanisms . . . . . . . . . . . . . . . . . . 94

4.4.4 Ex Post Budget Balance . . . . . . . . . . . . . . . . . 95

4.5 Comments on the Literature . . . . . . . . . . . . . . . . . . . 96

4.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96

5 Dominance: General Theory 97

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

5.2 Set-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97

5.3 Ex Post Revenue Equivalence . . . . . . . . . . . . . . . . . . 99

CONTENTS iii

5.4 Implementing Eﬃcient Decision Rules . . . . . . . . . . . . . 100

5.5 Characterizing All Incentive Compatible Decision Rules . . . 101

5.6 All Incentive Compatible Decision Rules When Outcomes are

Lotteries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

5.7 Single Dimensional Type Spaces . . . . . . . . . . . . . . . . 107

5.8 Suﬃciency of Weak Monotonicity . . . . . . . . . . . . . . . . 112

5.9 Positive Association of Diﬀerences . . . . . . . . . . . . . . . 113

5.10 Individual Rationality and Budget Balance . . . . . . . . . . 116

5.11 Remarks on the Literature . . . . . . . . . . . . . . . . . . . . 120

5.12 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

6 Bayesian Design: General Theory 121

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121

6.2 Set-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122

6.3 Independent Types . . . . . . . . . . . . . . . . . . . . . . . . 123

6.4 Correlated Types . . . . . . . . . . . . . . . . . . . . . . . . . 126

6.4.1 Framework . . . . . . . . . . . . . . . . . . . . . . . . 126

6.4.2 Failure of Revenue Equivalence . . . . . . . . . . . . . 126

6.4.3 Characterizing Bayesian Incentive Compatibility . . . 128

6.4.4 A Numerical Example . . . . . . . . . . . . . . . . . . 132

6.4.5 Individual Rationality and Budget Balance . . . . . . 134

6.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

6.6 Remarks on the Literature . . . . . . . . . . . . . . . . . . . . 137

6.7 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137

7 Non-Transferrable Utility 139

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139

7.2 The Gibbard Satterthwaite Theorem . . . . . . . . . . . . . . 140

7.2.1 Set Up . . . . . . . . . . . . . . . . . . . . . . . . . . . 140

7.2.2 Statement of the Result and Outline of the Proof . . . 141

7.2.3 Every monotone direct mechanism is dictatorial . . . . 143

7.3 Dominant Strategy Incentive Compatibility On Restricted

Domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147

7.4 Bayesian Incentive Compatibility . . . . . . . . . . . . . . . . 150

7.5 Remarks on the Literature . . . . . . . . . . . . . . . . . . . . 151

7.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152

8 Interdependent Types 153

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

8.2 An Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154

CONTENTS iv

8.3 Impossibility of Implementing Welfare Maximizing Decision

Rules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156

8.4 Characterizing All Incentive Compatible Mechanisms . . . . . 160

8.5 Remarks on the Literature . . . . . . . . . . . . . . . . . . . . 162

8.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162

9 Robust Mechanism Design 163

10 Dynamic Mechanism Design 164

11 Conclusion 165

Preface

This manuscript consists of lecture notes that I have used to teach courses

on mechanism design for graduate students at the University of Michigan

and at the Institute for Advanced Studies in Vienna.

The primary objective of these notes is to give rigorous, but accessi-

ble explanations of classic results in the theory of mechanism design, such

as Myerson’s theorem on expected revenue maximizing auctions, Myerson

and Satterthwaite’s theorem on the impossibility of ex post eﬃcient bilat-

eral trade with asymmetric information, and Gibbard and Satterthwaite’s

theorem on the non-existence of dominant strategy voting mechanisms. A

second objective of these notes is to take the reader in some selected areas

to the recent frontier of research.

The formal part of the notes starts in Chapter 2 with an explanation

of the theory of screening. The theory of screening is sometimes not re-

garded as part of the theory of mechanism design because it constructs an

incentive scheme for only one agent rather than multiple interacting agents.

However, the theories that are covered in Chapter 2 are intimately linked

to the theories of optimal mechanisms that are explained in later chapters,

in particular in Chapter 3. My hope is that by juxtaposing the theory of

screening and the theory of mechanism design I can help the reader un-

derstand which features of optimal mechanism design are due to strategic

interaction, and which features of optimal mechanism design are identical

to the corresponding features of optimal screening.

An organizing principle of a large proportion of the book is that I com-

pare mechanisms that are constructed with the anticipation that agents

will play a Bayesian equilibrium in some given information structure with

small type spaces, to mechanisms that provide each agent with a dominant

strategy. These two approaches have been the prevailing approaches of the

classic theory of mechanism design. I treat these two approaches side by

side with the intention to alert the reader to the potential problems of both

v

PREFACE vi

approaches, and to prepare the grounds for a later discussion of robust mech-

anism design. The modern theory of robust mechanism design has moved

the discussion beyond the traditional dichotomy of Bayesian mechanism de-

sign and dominant strategy mechanism design.

The manuscript assumes that readers have had a prior course in game

theory, and that they have basic knowledge of real analysis. Other than

this, however, I have sought to keep the prerequisites minimal.

This draft is incomplete in many ways. Chapters 9, 10 and 11 are still

to be written. Earlier chapters need further revisions. I also plan to include

further problems at the end of each chapter. Any suggestions for the fur-

ther development of this manuscript are always welcome. Please write to:

tborgers@umich.edu.

I am very grateful to Stefan Behringer who read an earlier draft of this

manuscript and pointed out many errors. All remaining errors are my own.

Tilman B¨orgers

Chapter 1

Introduction

Suppose you want to sell your house, and your realtor has identiﬁed several

potential buyers who are willing to pay your ask price. You might then wish

to conduct an auction among these buyers to obtain a higher price. There

are many diﬀerent auction formats that you could use: For example, each

buyer could be asked to send in one binding and ﬁnal bid. Alternatively,

buyers could bid in several rounds, and in each round they are all informed

about the highest bid of the previous round, and are then asked to revise

their bids. You could also use some combination of these formats. How

should you choose among diﬀerent auction formats? This is one of the

questions that the theory of mechanism design aims to answer.

Now imagine that you and your colleagues are considering to buy a new

refrigerator to be kept at work in which you can store the food that you bring

from home. While everyone is in favor, it is not so clear how much money

people are willing to contribute. How can you ﬁnd out whether the sum of

the amounts that everyone is willing to contribute covers the costs of the

refrigerator? You could ask everyone to submit pledges simultaneously, and

then see whether the sum of the pledges covers the expense. Alternatively,

you could go around, and tell each colleague how much everyone else has

pledged so far. Or you could divide the costs by the number of colleagues

involved, and commit to buying the refrigerator only if everyone is willing

to pay their share. Which of these procedures is best? Again, this is one of

the questions that the theory of mechanism design addresses.

Each of the procedures that you might consider in the two examples

above creates a strategic game in the sense of non-cooperative game theory

among the participants. Participants in these procedures will understand

that the outcome will depend not only on their own choices but also on

1

CHAPTER 1. INTRODUCTION 2

others’ choices, and that therefore their own optimal strategy may depend

on others’ strategies. The theory of mechanism design therefore builds on

the theory of games (Fudenberg and Tirole (1993)). Game theory takes

the rules of the game as given, and makes predictions about the behavior

of strategic players. The theory of mechanism design is about the optimal

choice of the rules of the game.

We are more frequently involved in the design of rules for games than

might be obvious at ﬁrst sight. How should shareholders’ vote be conducted?

How should promotion procedures in companies be organized? What are

optimal prenuptial agreements? All these questions are about the optimal

rules of games. The theory of mechanism design seeks to study the general

structure underlying all these applications, but it also considers a number

of particularly prominent applications in detail.

There are at least two reasons why we study mechanism design. Firstly,

the theory of mechanism design aids in practice the designers of real world

mechanisms. The theory of optimal auctions, for example, is frequently

invoked in discussions about the design of government and industry auc-

tions. Secondly, we can explain why real world institutions are as they are

by interpreting them as rational choices of those who designed them. For

example, we might seek to explain the use of auctions in some house sales,

and the use of posted prices in other house sales by appealing to the theory

of mechanism design which indicates that posted prices are optimal in some

circumstances and auctions are optimal in other circumstances.

The incentives created by the choice of rules of games are central to the

theory of mechanism design. Incentives are also at the center of contract

theory (Bolton and Dewatripont (2005)). At ﬁrst sight the distinction be-

tween the theory of mechanism design and contract theory is simple: In

contract theory we study the optimal design of incentives for a single agent.

In mechanism design we study the optimal design of incentives for a group

of agents, such as the buyers in our ﬁrst example, and the colleagues in the

second example. Contract theory therefore, unlike the theory of mechanism

design, does not have to deal with strategic interaction.

The relation between contract theory and the theory of mechanism de-

sign is more subtle. One part of the theory of mechanism design is, in fact,

a straightforward extension of insights from contract theory. This is surpris-

ing because one might have expected the element of strategic interaction,

that is present in mechanism design, but absent in contract theory, to create

substantial new problems. It is interesting and conceptually important to

CHAPTER 1. INTRODUCTION 3

understand why this is not the case, and we shall address this issue in detail

below. The close parallel between contract theory and mechanism design

applies only to some parts of mechanism design. Other parts of mechanism

design are, of course, unrelated to contract theory.

Contract theory has traditionally been divided into two parts: the theory

of hidden information (also referred to as the theory of “adverse selection,”)

and the theory of hidden action (also referred to as the theory of “moral

hazard.”) The distinction is easily explained within the context of contracts

for health insurance. Whether you have experienced sever chest pain in the

past is something that you know, but that the company from which you are

trying to buy health insurance does not know. It is “hidden information.”

Whether you exercise regularly, or take it a little more easy once you have

bought complete insurance coverage for heart surgery, is a choice that you

make that your insurance company does not observe unless it puts into place

a surveillance operation. It is a “hidden action.” Both hiddens, information

and actions, matter for contract design. For example, by oﬀering you a

menu of insurance contracts and observing your choice from this menu an

insurance company might be able to infer information about your health

risks that you might wish to conceal from the company. By introducing

deductibles an insurance company might seek to maintain your incentives

to look after your own health and thus to alleviate moral hazard problems.

Mechanism design, as traditionally understood, is about hidden infor-

mation, not hidden actions, with multiple agents. In our ﬁrst example, the

hidden information that the seller of a house seeks to ﬁnd out is the buyers’

true willingness to pay for the house. In our second example in this section,

the hidden information that we seek to ﬁnd out is the colleagues’ true will-

ingness to pay for an oﬃce refrigerator. In voting, the hidden information

that we seek to ﬁnd out is individuals’ true ranking of diﬀerent alternatives

or candidates. Of course, hidden action with many agents involved is a sub-

ject of great interest, and the theory that deals with it is concerned with

the optimal choice of rules for a game, like the theory of mechanism design.

For example, promotion schemes within a company set work incentives for a

group of employees, and the optimal choice of such schemes is an important

subject of economic theory. However, it is not the subject of mechanism

design as this term has traditionally been interpreted.

When choosing the rules for the strategic interaction among agents we

might restrict ourselves to a small subset of all conceivable rules, or we might

try to cast our net wide, and consider as large a set of rules as possible. For

example, when considering how to auction your house, you might restrict

CHAPTER 1. INTRODUCTION 4

attention to the choice of the minimum bid, and take for granted that the

auction will proceed with all potential buyers submitting their bids simul-

taneously. You would then focus on the choice of only one parameter in a

much larger set of possible choices. Alternatively, you might consider all

conceivable ways of proceeding, not just auctions, but, for example, also

simultaneous negotiations with all buyers that follow some predetermined

format. It has been one of the accomplishments of the theory of mechanism

design to develop a framework in which one can ﬁnd the optimal rules of the

game among all conceivable rules. Indeed traditionally mechanism design

has been understood as the ﬁeld in which this grand optimization among all

conceivable procedures is considered. In these notes we shall stick to this

interpretation of the theory of mechanism design.

Suppose you have considered all possible rules for proceeding with your

house sale, and you have come to the conclusion that an auction with just

one round of bidding is optimal. After the highest bid has been revealed,

one of the losing bidders approaches you with a new and improved bid that

is higher than the winning bid in the auction. Will you accept? This is an

obvious temptation, but if you accept later bids, are you still conducting an

auction with just a single round of bidding? In these notes we shall assume

that the mechanism designer has full commitment power. The rules, once

announced, are set in stone. The mechanism designer will not deviate from

them. In our example, the mechanism designer will absolutely refuse to re-

negotiate after the auction results have been revealed. This is obviously a

strong assumption. In contract theory, much attention has been given to the

optimal design of contracts if full commitment cannot always be achieved,

and this line of research has been very productive. It is also likely to be an

interesting line of research in mechanism design. We do not consider this

line of argument in these notes because we want to maintain a focus on the

central arguments of the traditional theory of mechanism design, and these

arguments have assumed full commitment by the mechanism designer.

These notes are meant for second year graduate students. The notes

assume that students have a good understanding of game theory. Fudenberg

and Tirole (1993) contains more than enough material for this course. We

shall also assume a basic knowledge of real analysis that can, for example,

be acquired from Royden (1988).

We proceed as follows. Chapter 2 presents some material from the theory

of optimal contract design with hidden information. This theory is also re-

ferred to as the theory of “monopolistic screening.” As we mentioned above,

the classic Bayesian theory of optimal mechanism design can be viewed as

CHAPTER 1. INTRODUCTION 5

an extension of the theory of monopolistic screening. We begin with the the-

ory of monopolistic screening to make this connection transparent, and also

because the basic results are most easily understood in the simple context

of monopolistic screening.

Chapters 3 and 6 then review the classic Bayesian theory ﬁrst in the

context of some prominent examples, and then in general. This theory is

built on several restrictive assumptions which we shall discuss in detail.

One of these assumptions is that for given rules of the mechanism agents

play a Bayesian equilibrium of the mechanism for a particular speciﬁcation

of agents’ beliefs about each others’ private information. This assumption

might attribute more information to the mechanism designer than is real-

istic, and therefore the literature has sought to develop mechanisms that

require the mechanism designer to know less about agents’ beliefs. The

classic approach to this problem is to seek dominant strategy mechanisms.

We present this approach in Chapters 4, in the context of examples, and in

Chapter 5, in general.

In Chapters 6.4, 7 and 8 we relax other assumptions of the classic model.

In Chapter 9 we return to the issue of what the mechanism designer knows

about agents’ beliefs about each other and investigate more modern ap-

proaches to this problem which do not necessarily require the construction

of a dominant strategy mechanism. Chapter 10 presents models of mecha-

nism design that apply to dynamic contexts. Like robust mechanism design,

this is an area of current research interest. Chapter 11 concludes the notes

with a brief review and an outlook.

Chapter 2

Screening

2.1 Introduction

Important parts of the theory of mechanism design are the multi-agent ex-

tension of the theory of screening. We begin by explaining three examples

from the theory of screening. We use these examples to introduce some

topics and techniques that are also important in mechanism design. By in-

troducing these topics and techniques in the context of screening, we explain

them in the simplest possible context.

2.2 Pricing a Single Indivisible Good

A seller seeks to sell a single indivisible good. The seller herself does not

attach any value to the good. Her objective is to maximize the expected

revenue from selling the good. She is thus risk neutral.

There is just one potential buyer. The buyer’s von Neumann Morgen-

stern utility if she purchases the good and pays a monetary transfer t to the

seller is: θ −t. The buyer’s utility

1

if she does not purchase the good is zero.

Here, θ > 0 is a number that we can interpret as the buyer’s valuation of the

good, because our assumptions imply that the buyer is indiﬀerent between

paying θ and obtaining the good, and not obtaining the good.

Two aspects of the assumptions about the buyer’s utility deserve em-

phasis. Firstly, we have assumed that the buyer’s utility is the sum of the

utility derived from the good, if it is purchased, and the disutility resulting

1

From now on, when we refer to “utility” we shall mean “von Neumann Morgenstern

utility.”

6

CHAPTER 2. SCREENING 7

from the money payment. A more general formulation would write utility

as u(I, t) where I is an indicator variable that is 1 if the buyer purchases

the good and 0 otherwise. Our assumption that u can be written as the

sum of θ and −t is usually described as the assumption that utility is “ad-

ditively separable.” Additive separability of the utility function implies that

the buyer’s utility from consuming the good is independent of the amount of

money that he pays for it, and that the buyer’s disutility from money pay-

ments is independent of whether or not he owns the good. One can easily

think of real world contexts in which these assumptions are probably vio-

lated. But we shall stick to these assumptions for most of these notes. Much

of the classical theory of mechanism design is based on these assumptions.

We shall consider a more general case in Chapter 7.

The second aspect of our assumptions about the buyer’s utility that

deserves emphasis is that we have assumed that the buyer is risk neutral

with respect to money, that is, his utility is linear in money. Like additive

separability this is a very restrictive assumption, but much of the classical

theory of mechanism design makes this assumption. The more general model

of Chapter 7 will relax this assumption, too.

We now introduce a crucial assumption about information. It is that

the value of θ is known to the buyer, but it is not known to the seller. This

seems plausible in many contexts. Buyers often know better than sellers

how well some particular product meets their preferences. We shall refer

below to θ as the buyer’s “type.”

As if often done in economic theory we shall assume that the seller has a

subjective probability distribution over possible values of θ. This probability

distribution can be described by a cumulative distribution function F. We

shall assume that F has a density f. Moreover, we shall assume that the

support of F, that is, the smallest closed set that has probability 1, is an

interval [θ,

¯

θ], where 0 ≤ θ <

¯

θ. For technical convenience we shall assume

that the density is strictly positive on the support: f(θ) > 0 for all θ ∈ [θ,

¯

θ].

We can now think of θ as a random variable with the cumulative distribution

function F the realization of which is observed by the buyer but not by the

seller.

Our interest is in procedures for selling the good which the seller should

adopt to maximize expected proﬁts. One obvious way would be to pick

a price p and to say to the buyer that he can have the good if and only

if he is willing to pay p. This is the selling procedure that we study in

elementary microeconomics. Suppose the seller picks this procedure. Which

CHAPTER 2. SCREENING 8

price p should she choose? The probability that the buyer’s value is below

p is given by the value of the cumulative distribution function F(p). The

probability that the buyer’s value is above p, and hence that he accepts a

price oﬀer p is 1 − F(p). Expected revenue is therefore p(1 − F(p)), and

the optimal strategy for the seller is to pick some price that maximizes

p(1 − F(p)). Note that this is just the monopoly problem from elementary

microeconomics with demand function 1 −F(p).

In this very simple context we shall now ask a straightforward question:

“Is picking a price p really the best the seller can do?” What else could the

seller do? The seller could, for example, negotiate with the agent. The seller

could oﬀer the agent a lottery where in return for higher or lower payments

the buyer could be given a larger or smaller chance of getting the object.

One can think of many other procedures that the seller might adopt to sell

the good. Is setting a price really the best procedure?

To make our question more precise we have to be speciﬁc about which

procedures the seller can commit to. We shall assume that the seller has

unlimited powers of commitment: the seller can commit to an arbitrary

extensive game tree where the players are the seller and the buyer, and

where each terminal history is associated with a probability distribution

over ¦0, 1¦R. The interpretation of such a probability distribution is that

it describes the probability with which the object is transferred to the buyer

together with a probability distribution over transfer payments by the buyer.

The seller will also ﬁnd it to his advantage to commit to a strategy for

himself, and to announce this strategy to the buyer before play begins. An

example would be that the seller announces that he will bargain over the

price with the seller, and that he announces in advance that he will turn

down certain oﬀers by the buyer. We shall assume that like the seller’s

ability to commit to a game, also the seller’s ability to commit to a strategy

in the game are unlimited.

Once the seller has committed to a game and a strategy, the buyer will

choose his own strategy in the game. We shall assume that the buyer chooses

his own strategy, knowing the value of θ, to maximize his expected utility.

Our question is now clearer: If the seller can commit to an arbitrary

extensive game, and if she can also commit to a strategy for playing that

game, which choice maximizes her expected revenue? In particular, is her

expected revenue maximized by committing to a price, and by committing

to selling the good at that price whenever the buyer is willing to pay the

price?

CHAPTER 2. SCREENING 9

A rather silly answer to our ﬁrst question could be this: The seller

should choose the game according to which the buyer has only a single

choice that is such that the buyer does not get the object but nonetheless

has to pay $x were x could be some arbitrarily large number. Clearly, if

we allow such procedures, the seller can extract an arbitrarily large amount

of money from the buyer. But the buyer wouldn’t want to participate in

such a mechanism. We will rule out such mechanisms by requiring that

the buyer for every value of θ ﬁnds it in her interest to participate in the

game proposed by the seller. In other words, for every type θ the buyer will

need to ﬁnd that when he chooses his expected utility maximizing strategy

his expected utility will be at least zero, the utility that he would obtain if

he did not buy the good and did not pay anything. We shall refer to this

constraint as the “individual rationality” constraint. Sometimes, it is also

called “participation constraint.”

Our objective is thus to study the optimization problem in which the

seller’s choice variables are an extensive game and a strategy in that game,

in which the seller’s objective function is expected revenue, and in which

the constraint on the seller’s choice is the individual rationality constraint.

At ﬁrst sight, this looks like a hard problem as the seller’s choice set is

very large. There are many extensive games that the seller could consider.

However, a simple, yet crucial result enables us to get a handle on this opti-

mization problem. We can restrict attention to a small set of mechanisms,

called “direct mechanisms.”

Deﬁnition 2.1.

2

A “direct mechanism” consists of functions q and t where:

q : [θ,

¯

θ] →[0, 1]

and

t : [θ,

¯

θ] →R.

The interpretation is that in a direct mechanism the buyer is asked to

report θ. The seller commits to transferring the good to the buyer with

probability q(θ) if the buyer reports that her type is θ, and the buyer has to

pay the seller t(θ) if she reports that her type is θ. Note that the payment

is deterministic. It is not conditional on the event that the buyer obtains

the good. It would make no diﬀerence if we allowed the payment to be

2

To be precise, this deﬁnition should include the requirement that the functions q and

t are Lebesgue-measurable. We omit measurability requirements throughout this text

which will lead to a small cost in terms of rigor in some places.

CHAPTER 2. SCREENING 10

random. All our analysis below would go through if we interpreted t(θ) as

the buyer’s expected payment conditional on θ. All selling mechanisms that

are not direct mechanisms are called “indirect mechanisms.”

A buyer’s strategy σ in a direct mechanism is a mapping σ : [θ,

¯

θ] →[θ,

¯

θ]

that indicates for every true type θ that the buyer might have the type σ(θ)

that the buyer reports to the seller as his type.

The next result, a very simple version of the famous “Revelation Princi-

ple,” shows that it is without loss of generality to restrict attention to direct

mechanism.

Proposition 2.1 (Revelation Principle). For every mechanism Γ and every

optimal buyer strategy σ in Γ there is a direct mechanism Γ

and an optimal

buyer strategy σ

in Γ

such that:

(i) The strategy σ

satisﬁes:

σ

(θ) = θ for every θ ∈ [θ,

¯

θ],

i.e. σ

**prescribes telling the truth;
**

(ii) For every θ ∈ [θ,

¯

θ] the probability q(θ) and the payment t(θ) equal the

probability of purchase and the expected payment that result in Γ if the

buyer plays her optimal strategy σ.

Proof. For every θ ∈ [θ,

¯

θ] deﬁne q(θ) and t(θ) as required by (ii) in Propo-

sition 2.1. We prove the result by showing that for this direct mechanism

the strategy σ

**(θ) = θ, that is, truthfully reporting her type, is optimal for
**

the buyer. Note that under this strategy, for every θ, the buyer with type θ

obtains in the mechanism Γ

**the same expected utility as in the mechanism
**

Γ when choosing strategy σ(θ). Moreover, when pretending to be some type

θ

**,= θ, the buyer obtains the same expected utility that she would have
**

obtained had she played type θ

s strategy σ(θ

) in Γ. The optimality of

truthfully reporting θ in Γ

**then follows immediately from the optimality of
**

σ(θ) in Γ.

The Revelation Principle allows us to simplify our analysis greatly be-

cause it shows that without loss of generality we can restrict our search for

optimal mechanisms to direct mechanisms, that is, pairs of functions q and

t, where the buyer ﬁnds it always optimal to truthfully report her type.

Given a direct mechanism, we deﬁne deﬁne the buyer’s expected utility

u(θ) conditional on her type being θ by: u(θ) = θq(θ) − t(θ). Using this

CHAPTER 2. SCREENING 11

notation, we can now formally deﬁne the condition that the buyer ﬁnds it

always optimal to truthfully report her type.

Deﬁnition 2.2. A direct mechanism is “incentive compatible” if truth

telling is optimal for every type, that is, if:

u(θ) ≥ θq(θ

) −t(θ

) for all θ, θ

∈ [θ,

¯

θ].

We mentioned before that it makes sense to also require that the buyer’s

expected utility from the mechanism is not lower than some lower bound,

say zero. This requirement is captured in the following deﬁnition.

Deﬁnition 2.3. A direct mechanism is “individually rational” if the buyer,

conditional on her type, is voluntarily willing to participate, that is, if:

u(θ) ≥ 0 for all θ ∈ [θ,

¯

θ].

Notice that in this deﬁnition we require voluntary participation after

agents have learned their types. A weaker requirement would be to require

voluntary participation before agents have learned their types.

We now consider in more detail the conditions under which a direct

mechanism is incentive compatible. Later we bring individual rationality

into the picture.

Lemma 2.1. If a direct mechanism is incentive compatible, then q is in-

creasing in θ.

3

Proof. Consider two types θ and θ

with θ > θ

. Incentive compatibility

requires:

θq(θ) −t(θ) ≥ θq(θ

) −t(θ

) (2.1)

θ

q(θ) −t(θ) ≤ θ

q(θ

) −t(θ

) (2.2)

Subtracting these two inequalities we obtain

(θ −θ

)q(θ) ≥ (θ −θ

)q(θ

) ⇔ (2.3)

q(θ) ≥ q(θ

) (2.4)

3

Throughout these notes we shall say that a function f is “increasing” if it is weakly

monotonically increasing, i.e. if x > x

implies f(x) ≥ f(x

).

CHAPTER 2. SCREENING 12

Lemma 2.2. If a direct mechanism is incentive compatible, then u is in-

creasing. It is also convex, and hence diﬀerentiable except in at most count-

ably many points. For all θ for which it is diﬀerentiable, it satisﬁes:

u

(θ) = q(θ).

Note that the equation for u

**(θ) is the same as the formula for the deriva-
**

tive of maximized utility functions in the “envelope theorem.” For complete-

ness, we provide a self-contained proof.

Proof. For all θ:

u(θ) = max

θ

∈[θ,

¯

θ]

(θq(θ

) −t(θ

)). (2.5)

Thus u is the maximum of increasing and aﬃne, hence convex functions.

The maximum of increasing functions is increasing, and the maximum of

convex functions is convex. Therefore, u is increasing and convex. Convex

functions are diﬀerentiable in at most countably many points (Proposition

5.16 in Royden (1988)). Consider any θ for which u is diﬀerentiable. Let

δ > 0. Then:

lim

δ→0

u(θ +δ) −u(θ)

δ

≥ lim

δ→0

((θ +δ)q(θ) −t(θ)) −(θq(θ) −t(θ))

δ

(2.6)

= q(θ). (2.7)

Similarly:

lim

δ→0

u(θ) −u(θ −δ)

δ

≤ lim

δ→0

(θq(θ) −t(θ)) −((θ −δ)q(θ) −t(θ))

δ

(2.8)

= q(θ). (2.9)

Putting the two inequalities together we obtain u

(θ) = q(θ) whenever u is

diﬀerentiable.

The next lemma is essentially an implication of Lemma 2.2 and the

fundamental theorem of calculus. In the proof we take care of the possible

lack of diﬀerentiability of u in some points.

Lemma 2.3 (Payoﬀ Equivalence). Consider an incentive compatible direct

mechanism. Then for all θ ∈ [θ,

¯

θ]:

u(θ) = u(θ) +

_

θ

θ

q(x)dx.

CHAPTER 2. SCREENING 13

Proof. The fact that u is convex implies by Proposition 5.16 in Royden

(1988) that it is absolutely continuous. By Proposition 5.13 in Royden

(1988) this implies that it is the integral of its derivative.

Lemma 2.3 shows that the expected utilities of the diﬀerent types of the

buyer are pinned down by the the function q and by the expected utility of

the lowest type of the buyer, u(θ). Any two indirect mechanisms which, once

the buyer optimizes, give rise to the same q and u(θ) therefore imply the

same expected payoﬀ for all types of the buyer. We have therefore indirectly

shown a “payoﬀ equivalence” result for classes of indirect mechanisms.

A short computation turns Lemma 2.3 into a result about the transfer

payments that the buyer expects to make to the seller. This is shown in the

next lemma.

Lemma 2.4 (Revenue Equivalence). Consider an incentive compatible di-

rect mechanism. Then for all θ ∈ [θ,

¯

θ]

t(θ) = t(θ) + (θq(θ) −θq(θ)) −

_

θ

θ

q(x)dx.

Proof. Recall that u(θ) = θq(θ) −t(θ). Substituting this into the formula in

Lemma 2.3 and solving for t(θ) yields the result.

Lemma 2.4 shows that the expected payments of the diﬀerent types

of the buyer are pinned down by the the function q and by the expected

payment of the lowest type of the buyer, t(θ). For given q and given t(θ)

there is thus one, and only one, incentive compatible direct mechanism. Any

two indirect mechanisms which, once the buyer optimizes, give rise to the

same q and t(θ) therefore imply the same expected payment for all types of

the buyer. For the seller it follows that any two such indirect mechanisms

yield the same expected revenue, because the seller’s expected revenue is

the expected value of the buyer’s expected payments, where the seller takes

expected values over the buyer’s types. We have therefore indirectly shown

a “revenue equivalence” result for classes of indirect mechanisms.

Lemma 2.4 is the famous “Revenue Equivalence Theorem” of auction

theory adapted to our more simple setting of monopolistic screening. The

full Revenue Equivalence Theorem, which will be discussed in Chapter 3, is

based on essentially the same argument that we have explained here.

Lemmas 2.1 and 2.4 give necessary conditions for a direct mechanism to

be incentive compatible. It turns out that these conditions are also suﬃcient.

CHAPTER 2. SCREENING 14

Proposition 2.2. A direct mechanism (q, t) is incentive compatible if and

only if:

(i) q is increasing;

(ii) For every θ ∈ [θ,

¯

θ]:

t(θ) = t(θ) + (θq(θ) −θq(θ)) −

_

θ

θ

q(x)dx

Proof. To prove suﬃciency we have to show that no type θ prefers to pre-

tend to be a type θ

**if q is increasing and t is given by the formula in the
**

Proposition:

u(θ) ≥ θq(θ

) −t(θ

) ⇔ (2.10)

u(θ) ≥ θq(θ

) −θ

q(θ

) +θ

q(θ

) −t(θ

) ⇔ (2.11)

u(θ) ≥ θq(θ

) −θ

q(θ

) +u(θ

) ⇔ (2.12)

u(θ) −u(θ

) ≥ (θ −θ

)q(θ

) ⇔ (2.13)

_

θ

θ

q(x)dx ≥

_

θ

θ

q(θ

)dx (2.14)

To obtain the left hand side of the last inequality we used the formula in

Lemma 2.3. This formula is an implication of the formula for the payment

in the proposition, as one can see by doing the calculation referred to in the

proof of Lemma 2.4 in reverse order. Comparing the two integrals on the

left hand side and the right hand side of the last inequality, suppose ﬁrst

that θ

**> θ. The two integrals have the same integration limits, and the
**

function q(x) is everywhere at least as large as the constant q(θ

) that is

being integrated on the right hand side because q is increasing. Therefore,

the integral on the left hand side is at least as large as the integral on the

right hand side. If θ

**< θ the argument is analogous.
**

We have now obtained a complete characterization of all incentive com-

patible direct mechanisms. We now bring in individual rationality.

Proposition 2.3. An incentive compatible direct mechanism is individually

rational if and only if u(θ) ≥ 0 (or equivalently: t(θ) ≤ θq(θ)).

Proof. By Lemma 2.2 u is increasing in θ for incentive compatible mech-

anisms. Therefore, u(θ) is non-negative for all θ if and only if it is non-

negative for the lowest θ.

CHAPTER 2. SCREENING 15

We have now completely characterized the set of all direct mechanisms

from which the seller can choose. We turn to the seller’s problem of picking

from this set the mechanism that maximizes expected revenue. We begin

with the observation that it is optimal for the seller to set the lowest type’s

payment so that this type has zero expected utility.

Lemma 2.5. If an incentive compatible and individually rational direct

mechanism maximizes the seller’s expected revenue then

t(θ) = θq(θ).

Proof. By Proposition 2.3 we have to have: t(θ) ≤ θq(θ). If t(θ) < θq(θ),

then the seller could increase expected revenue by choosing a direct mecha-

nism with the same q, but with a higher t(θ). By the formula for payments

in Proposition 2.2 all types’ payments would increase.

Using Lemma 2.5 we can now simplify the seller’s choice set further.

The seller’s choice set is the set of all increasing functions q : [θ,

¯

θ] →[0, 1].

Given any such function, the seller will optimally set t(θ) = θq(θ) so that

the lowest type has zero expected utility, and all other types’ payments are

determined by the formula in Proposition 2.2. Substituting the lowest type’s

payment, this formula becomes:

t(θ) = θq(θ) −

_

θ

θ

q(

ˆ

θ)d

ˆ

θ, (2.15)

i.e. type θ pays his expected utility from the good, θq(θ), minus a term

that reﬂects a surplus that the seller has to grant to the buyer to provide

incentives to the buyer to correctly reveal his type. This term is also called

the buyer’s “information rent.”

To determine the optimal function q we could use equation (2.15) to

obtain an explicit formula for the seller’s expected revenue for any given

function q. This is an approach that we shall take later in the next section.

In the current context, an elegant argument based on convex analysis yields

a simple and general answer. You need to be a little patient as the formal

machinery that is needed for this approach is introduced.

We begin by considering in more detail the set of functions q that the

seller can choose from. This set is a subset of the set of all bounded functions

f : [θ,

¯

θ] → R. We denote this larger set by T. We give T the linear

CHAPTER 2. SCREENING 16

structure:

g = αf ⇔[g(x) = αf(x) ∀x ∈ [0, 1]] for all α ∈ R, f, g ∈ T

(2.16)

h = f +g ⇔[h(x) = f(x) +g(x) ∀x ∈ [0, 1]] for all f, g, h ∈ T

(2.17)

This makes T a vector space. We also give T the L

∞

-norm. Intuitively, this

is the supremum of [f(θ)[ over θ ∈ [θ,

¯

θ] where we neglect sets of measure

zero. Formally, the L

∞

-norm is deﬁned by:

4

| f |= inf ¦M [ µ(¦θ [ [f(θ)[ > M¦) = 0¦ for all f ∈ T. (2.18)

where µ is the Lebesgue measure.

We denote by /⊂ T the set of all increasing functions in T such that

f(x) ∈ [0, 1] ∀x ∈ [0, 1]. This is the set from which the seller chooses. We

now have a simple but crucial observations about /.

Lemma 2.6. / is compact and convex.

Convexity is obvious as the convex combination of two increasing func-

tions is increasing. Compactness is an implication of Helly’s compactness

theorem Doob (1994), p. 165.

Next, we have a closer look at the seller’s objective function, which is the

expected value of the right hand side of (2.15). The simple observations that

we need is that this objective function is continuous and linear in q. From

(2.15) we see that if we multiply q by a constant α, then t(θ) is multiplied

by α for all θ. Therefore, also the expected value of t(θ) is multiplied by α,

and expected revenue is a linear function of q.

We thus see that the seller maximizes a continuous linear function over

a compact, convex set. A fundamental theorem of real analysis provides

conditions under which the maximizers of a linear function over a convex

set include “extreme points” of that set. This is intuitive. For example,

you are probably familiar with the fact that in linear programming the

optimal points include corner points. We shall apply this insight here. We

ﬁrst provide the deﬁnition of extreme points of a convex set. This is a

generalization of the idea of corner points.

4

The details of this construction are in Royden (1988), p. 112. Note Royden’s footnote

1 which indicates that to be rigorous we should deﬁne F to consist of equivalence classes

of functions where two functions are in the same equivalence class if and only if they diﬀer

on a set of measure zero only. We shall not be precise about this, but neglecting this point

leads to a certain lack of rigor in what follows.

CHAPTER 2. SCREENING 17

Deﬁnition 2.4. If C is a convex subset of a vector space X, then x ∈ C is

an extreme point of C if for every y ∈ X, y ,= 0 we have that either x+y / ∈ C

or x −y / ∈ C or both.

So, for example in two-dimensional space the corners of a triangle are

the extreme points of the triangle, and the circumference of a circle is the

set of extreme points of the circle.

Now we can state the result that we will use to study the seller’s optimal

mechanism. It is called the “Extreme Point Theorem” in Ok (2007), p. 658.

Proposition 2.4. Let X be a compact, convex subset of a normed vector

space, and let f : X → R be a continuous linear function. Then the set E

of extreme points of X is non-empty, and there exists an e ∈ E such that

f(e) ≥ f(x) for all x ∈ X.

This result implies that a function q that is an extreme point of / and

that maximizes expected revenue among all extreme points of /also max-

imizes expected revenue among all functions in /. We may thus simplify

the seller’s problem further. Instead of considering all functions in the set

/ it is suﬃcient to consider only the set of all extreme points of /. The

following result characterizes the extreme points of /.

Lemma 2.7. A function q ∈ / is an extreme point of / if and only if

q(θ) ∈ ¦0, 1¦ for almost all θ ∈ [θ,

¯

θ].

Proof. Consider any function as described in the lemma, and suppose that

ˆ q is another function that satisﬁes: ˆ q(θ) ,= 0 for some θ.

5

If ˆ q(θ) > 0 and

q(θ) = 0 then q(θ)−ˆ q(θ) < 0, and hence q−ˆ q / ∈ /. If ˆ q(θ) > 0 and q(θ) = 1

then q(θ) + ˆ q(θ) > 0, and hence q − ˆ q / ∈ /. The case ˆ q(θ) < 0 is analogous.

Now consider any function q that is not as described in the lemma, i.e.

there is some θ

∗

such that q(θ

∗

) ∈ (0, 1).

6

Deﬁne ˆ q(θ) = q(θ) if q(θ) ≤ 0.5,

and ˆ q(θ) = 1 − q(θ) if q(θ) > 0.5. Clearly, ˆ q ,= 0. Consider now ﬁrst

the function q + ˆ q. We have that q(θ) + ˆ q(θ) = 2q(θ) if q(θ) ≤ 0.5, and

q(θ) + ˆ q(θ) = 1 if q(θ) > 0.5. Thus, evidently, ˜ q ∈ /. The argument for

q − ˆ q is analogous. We conclude that q is not an extreme point of M.

The seller can thus restrict attention to non-stochastic mechanisms. But

a non-stochastic mechanism is monotone if and only if there is some p ∈ [θ,

¯

θ]

5

More precisely: ˆ q(θ) = 0 for a set of θ that has positive Lebesgue measure.

6

More precisely: q(θ

∗

) ∈ (0, 1) for a set of θ

∗

that has positive Lebesgue measure.

CHAPTER 2. SCREENING 18

such that q(θ) = 0 if θ < p

∗

and q(θ) = 1 if θ > p

∗

. This direct mechanism

can be implemented by the seller simply quoting the price p

∗

and the buyer

either accepting or rejecting p

∗

. Our results therefore imply that the seller

cannot do better than quoting a simple price p

∗

to the buyer. The analysis

is summarized in the following proposition.

Proposition 2.5. A direct mechanism maximizes the seller’s expected rev-

enues among all incentive compatible, individually rational direct mecha-

nisms if and only if there is a p

∗

∈ argmax

p∈[θ,

¯

θ]

p(1 −F(p)) such that

q(θ) =

_

_

_

1 if θ > p

∗

;

0 if θ < p

∗

,

and

t(θ) =

_

_

_

p

∗

if θ > p

∗

;

0 if θ < p

∗

.

Proof. As argued above we only need to consider functions q where the buyer

obtains the good with probability 1 if his value is above some price p

∗

and

with probability 0 if his value is below this price. The optimal function q of

this form is obviously the one indicated in the Proposition. The formula for

t follows from Proposition 2.2.

It may seem that we have gone to considerable length to derive a dis-

appointing result, namely, a result that does not oﬀer the seller any more

sophisticated selling mechanisms than we are familiar with from elementary

microeconomics. However, apart from introducing some technical tools that

we use later in more complicated contexts, the reader should appreciate that

we have uncovered a rather sophisticated rationale for a familiar everyday

phenomenon. This is perhaps analogous to invoking Newton’s law of grav-

ity as an explanation of the fact that apples fall oﬀ apple trees. The fact is

familiar, but the explanation is non-obvious.

2.3 Nonlinear Pricing

Now we study a model in which a monopolist oﬀers an inﬁnitely divisible

good, say sugar, to one potential buyer. We introduce this model because

it is more commonly studied than the model in the previous section, and

because its analysis introduces additional elements, beyond those presented

CHAPTER 2. SCREENING 19

in the previous section, that will reappear in almost exactly the same form

in the analysis of optimal mechanisms.

For simplicity we assume that production costs are linear, that is, pro-

ducing quantity q ≥ 0 costs cq, where c > 0 is a constant. The seller

is risk neutral, so that she seeks to maximize her expected revenue. The

buyer’s utility from buying quantity q ≥ 0 and paying a monetary trans-

fer t to the monopolist is θν(q) − t. We assume that ν(0) = 0 and that

ν is a twice diﬀerentiable, strictly increasing and strictly concave function:

ν

(q) > 0, ν

(q) < 0 for all q ≥ 0.

Because ν(0) = 0 the buyer’s utility when buying nothing and paying

nothing is zero. We can interpret θν(q) as the buyer’s willingness to pay

for quantity q. Note that we have assumed, as in the previous section, that

utility is additively separable in consumption of the good and money, and

that the consumer is risk neutral in money.

The parameter θ reﬂects how much the consumer values the good. More

precisely, the larger θ the larger is the consumer’s absolute willingness to pay

θν(q) and the consumer’s marginal willingness to pay θν

**(q) for any given
**

quantity q. The parameter θ can take any value between θ and

¯

θ. The value

of θ is known to the buyer but not to the seller. The seller’s beliefs about

θ are given by a cumulative distribution function F with density f on the

interval [θ,

¯

θ]. We assume that f satisﬁes: f(θ) > 0 for all θ ∈ [θ,

¯

θ].

A ﬁnal assumption is that lim

q→∞

¯

θν

**(q) < c. This means that even the
**

highest type’s marginal willingness to pay falls below c as q gets large. This

assumption ensures that the quantity that the seller supplies to the buyer

is ﬁnite for all possible types of the buyer.

We seek to determine optimal selling procedures for the seller. As in the

previous section the revelation principle holds and we can restrict attention

to direct mechanisms. In the current context we use the following deﬁnition

of direct mechanisms:

Deﬁnition 2.5. A “direct mechanism” consists of functions q and t where:

q : [θ,

¯

θ] →R

+

and

t : [θ,

¯

θ] →R.

The interpretation is that the buyer is asked to report θ, and the seller

commits to selling quantity q(θ) to the buyer and the buyer commits to pay-

ing t(θ) to the seller. Note that we use the same notation as in the previous

CHAPTER 2. SCREENING 20

section, but that in this section q(θ) is a quantity whereas in the previous

section q(θ) was a probability. In this section we ignore stochastic mecha-

nisms, that is, we assume that for each type θ the quantity sold to the buyer

if he is of type θ is a non-negative number, not a probability distribution

over non-negative numbers. We make this assumption for simplicity. It is

non-trivial to study stochastic direct mechanisms in our context.

We do not state the revelation principle formally for our context. It is

analogous to the revelation principle in the previous section. One modiﬁca-

tion is needed, however, in the statement of the result. To obtain determin-

istic direct mechanisms we have to restrict attention to general mechanisms

and buyer optimal strategies that result in a deterministic quantity for the

buyer for each type of the buyer.

As in the previous section we can then study incentive compatibility and

individual rationality of direct mechanisms. The analysis proceeds along

exactly the same lines as in the previous section, and we shall just state the

result of the analysis. One ﬁnds that a direct mechanism (q, t) is incentive

compatible if and only if

(i) q is increasing;

(ii) For every θ ∈ [θ,

¯

θ]:

t(θ) = t(θ) −θν(q(θ)) +θν(q(θ)) −

_

θ

θ

ν(q(

ˆ

θ))d

ˆ

θ. (2.19)

An incentive compatible mechanism is individual rational if and only if

t(θ) ≤ θq(θ). (2.20)

The seller’s decision problem is to pick among all direct mechanisms

satisfying these two conditions the one that maximizes expected revenue. It

is obvious that the seller will choose t(θ) so that the utility of type θ is zero,

that is:

t(θ) = θq(θ). (2.21)

Substituting this into equation (2.19) yields:

t(θ) = θν(q(θ)) −

_

θ

θ

ν(q(

ˆ

θ))d

ˆ

θ. (2.22)

The choice that remains to be studied is that of the function q.

CHAPTER 2. SCREENING 21

At this point we depart from the line of argument that we followed in

the previous section. The reason is that the seller’s objective function is

no longer linear in q. This is clear from equation (2.22) where q enters the

non-linear function ν. Because the objective function is not linear in q, the

extreme point argument of the previous section does not apply here. We

shall use instead equation (2.22) to study in more detail the seller’s expected

proﬁt. If the seller chooses q(), then his expected proﬁt is:

_

¯

θ

θ

_

θν(q(θ)) −

_

θ

θ

ν(q(

ˆ

θ))d

ˆ

θ −cq(θ)

_

f(θ)dθ

=

_

¯

θ

θ

θν(q(θ))f(θ)dθ −

_

¯

θ

θ

_

θ

θ

ν(q(

ˆ

θ))d

ˆ

θf(θ)dθ −

_

¯

θ

θ

cq(θ)f(θ)dθ (2.23)

We seek to simplify the expression in (2.23). The calculation that fol-

lows, although it contains no conceptually or mathematically deep insights,

appears in this or in similar form frequently in the theory of mechanism

design. It is therefore worthwhile to consider it in detail. We focus initially

on the double integral in the second term in (2.23).

_

¯

θ

θ

_

θ

θ

ν(q(

ˆ

θ))d

ˆ

θf(θ)dθ

=

_

¯

θ

θ

_

θ

θ

ν(q(

ˆ

θ))f(θ)d

ˆ

θdθ

=

_

¯

θ

θ

_

¯

θ

ˆ

θ

ν(q(

ˆ

θ))f(θ)dθd

ˆ

θ

=

_

¯

θ

θ

ν(q(

ˆ

θ))

_

¯

θ

ˆ

θ

f(θ)dθd

ˆ

θ

=

_

¯

θ

θ

ν(q(

ˆ

θ))(1 −F(

ˆ

θ))d

ˆ

θ

=

_

¯

θ

θ

ν(q(θ))(1 −F(θ))dθ (2.24)

When moving from the second to the third line in (2.24) we change

the order of integration. By Fubini’s theorem this leaves the value of the

integral unchanged. We indicate in Figure 2.1 the change in the order of

integration. In the second line we ﬁrst integrate along the vertical lines and

then horizontally. In the third line we ﬁrst integrate along the horizontal

lines in Figure 2.1 and then vertically.

CHAPTER 2. SCREENING 22

θθ θθ

θθ

θθ

θθ

θθ ^

Figure 2.1: Changing the order of integration

We now substitute the last line in (2.24) into the seller’s objective func-

tion in (2.23).

_

¯

θ

θ

[θν(q(θ)) −cq(θ)] f(θ)dθ −

_

¯

θ

θ

ν(q(θ))(1 −F(θ))dθ

=

_

¯

θ

θ

[θν(q(θ)) −cq(θ)] f(θ)dθ −

_

¯

θ

θ

ν(q(θ))

1 −F(θ)

f(θ)

f(θ)dθ

=

_

¯

θ

θ

_

ν(q(θ))

_

θ −

1 −F(θ)

f(θ)

_

−cq(θ)

_

f(θ)dθ. (2.25)

The seller thus chooses q to maximize the expected value of the expression

that is in large square brackets in (2.25). Expected values are taken over θ.

The seller must choose an increasing function q.

Suppose we ignore for the moment the constraint that q must be increas-

ing. Then the seller can choose q(θ) for each θ separately to maximize the

expression in the large square brackets. This choice of q also maximizes the

expected value of that expression. We study this approach to the choice of q

ﬁrst, and then impose a condition that makes sure that the function q that

CHAPTER 2. SCREENING 23

we ﬁnd is indeed increasing. The ﬁrst order condition for maximizing the

expression in square brackets for given θ is:

ν

(q(θ))

_

θ −

1 −F(θ)

f(θ)

_

−c = 0 ⇔ (2.26)

ν

(q(θ))

_

θ −

1 −F(θ)

f(θ)

_

= c (2.27)

We now investigate the existence of a solution to (2.27). If

θ −

1 −F(θ)

f(θ)

≤ 0 (2.28)

then there is obviously no solution, and the optimal choice is

q(θ) = 0. (2.29)

Now consider

θ −

1 −F(θ)

f(θ)

> 0. (2.30)

Recall that we have assumed that ν

**is diﬀerentiable, and therefore continu-
**

ous, and decreasing, and that

¯

θν(q) tends to less than c as q tends to inﬁnity.

Obviously, the left hand side of (2.27) shares all these properties. If

ν

(0)

_

θ −

1 −F(θ)

f(θ)

_

≤ c (2.31)

then it is again obvious that the optimal choice is:

q(θ) = 0. (2.32)

If

ν

(0)

_

θ −

1 −F(θ)

f(θ)

_

> c (2.33)

then our assumptions imply that there is a unique solution to (2.27), and

that this stationary point is also the unique optimal choice of q(θ).

We have now determined for each θ the choice of q(θ) that maximize

the expression that is in large square brackets in (2.25). The seller seeks to

maximize the expected value of this expression, and the seller is constrained

to choose a function q that is increasing. If the function q that we have

determined above is increasing, then it must be the optimal choice for the

seller. We now introduce the following assumption, which implies that the

q that we have determined is increasing.

CHAPTER 2. SCREENING 24

Assumption 2.1. θ −

1−F(θ)

f(θ)

is increasing in θ.

To verify that Assumption 2.1 implies that q is increasing note that

Assumption 2.1 implies that the left hand side of (2.27) is increasing in θ

for every q. The optimal q is the intersection point of that expression with

c or zero, whatever is greater. It is then easy to see that the optimal q is

increasing in θ.

A suﬃcient condition for Assumption 2.1 is that

f(θ)

1−F(θ)

is increasing in

θ. This suﬃcient condition is often referred to as the “increasing hazard

rate” condition. Think of F(θ) as the probability that an individual dies

before time θ. Then 1 −F(θ) is the probability that the individual survives

until time θ, and

f(θ)

1−F(θ)

can be thought of as the conditional probability

of dying at time θ of an individual that has survived until time θ. The

suﬃcient condition is that this conditional probability of dying, one of the

unavoidable hazards of life, is increasing in θ.

Many commonly considered distributions F satisfy Assumption 2.1. We

shall refer to such distributions as “regular.” The analysis of this section is

summarized in the following proposition:

Proposition 2.6. Suppose that F is regular. Then an expected proﬁt max-

imizing choice of q is given by:

(i) if ν

(0)

_

θ −

1−F(θ)

f(θ)

_

≤ c:

q(θ) = 0;

(ii) otherwise:

ν

(q(θ))

_

θ −

1 −F(θ)

f(θ)

_

= c.

The proﬁt maximizing t is given by:

t(θ) = θν(q(θ)) −

_

θ

θ

ν(q(

ˆ

θ))d

ˆ

θ.

To understand the economic meaning of Proposition 2.6 note that for

the highest type θ =

¯

θ we have: 1 −F(θ) = 0. Therefore, the second of the

two cases in Proposition 2.6 applies to

¯

θ, and q(

¯

θ) is determined by:

ν

(q(θ))

¯

θ = c. (2.34)

This equation shows that the highest type is supplied the quantity at which

this type’s marginal willingness to pay is exactly equal to the marginal cost

CHAPTER 2. SCREENING 25

of production. This is the quantity that this type would choose to produce

if he owned the ﬁrm. We refer to this quantity as the “ﬁrst best” quantity.

For all lower types θ <

¯

θ the quantity supplied to these types is determined

by equation (2.27) that diﬀers from the ﬁrst best condition (2.34) (with

¯

θ replaced by θ) in that the left hand side is smaller for every q. Thus,

the marginal costs are not equated with the marginal beneﬁts, but with a

quantity smaller than the marginal beneﬁts. This means that all types that

are lower than

¯

θ are oﬀered a quantity that is smaller than the “ﬁrst best”

quantity.

We conclude with a numerical example.

Example 2.1. c = 1, ν(q) =

√

q, θ is uniformly distributed on [0, 1], i.e.:

F(θ) = θ and f(θ) = 1 for all θ ∈ [0, 1]. To verify that Assumption 2.1 is

satisﬁed, we have to check that the following expression is increasing in θ:

θ −

1 −F(θ)

f(θ)

= θ −

1 −θ

1

= 2θ −1, (2.35)

which is obviously the case.

Next we determine for which values of θ the optimal quantity q(θ) equals

zero:

ν

(0)

_

θ −

1 −F(θ)

f(θ)

_

≤ c ⇔

θ −

1 −F(θ)

f(θ)

≤ 0 ⇔

2θ −1 ≤ 0 ⇔

θ ≤ 0.5 (2.36)

The ﬁrst and the second line are equivalent because in our example: ν

(0) =

+∞. If θ > 0.5 the optimal q(θ) is given by:

ν

(q(θ))

_

θ −

1 −F(θ)

f(θ)

_

= c ⇔

1

2

√

q

(2θ −1) = 1 ⇔

√

q = θ −

1

2

⇔

q =

_

θ −

1

2

_

2

. (2.37)

CHAPTER 2. SCREENING 26

The corresponding transfer t(θ) is zero if θ ≤ 0.5, and if θ > 0.5 it is given

by:

t(θ) = θν(q(θ)) −

_

θ

θ

ν(q(

ˆ

θ))d

ˆ

θ

= θ

_

θ −

1

2

_

−

_

θ

0.5

ˆ

θ −

1

2

d

ˆ

θ

= θ

_

θ −

1

2

_

−

_

1

2

ˆ

θ

2

−

1

2

ˆ

θ

_

θ

0.5

= θ

_

θ −

1

2

_

−

_

1

2

θ

2

−

1

2

θ −

1

8

+

1

4

_

=

1

2

θ

2

−

1

8

(2.38)

We want to translate the solution into an optimal non-linear pricing scheme.

We can express the transfer t as a function of q. For this we determine ﬁrst

for given q which type θ purchases the quantity q

q(θ) = q ⇔

_

θ −

1

2

_

2

= q ⇔

θ −

1

2

=

√

q ⇔

θ =

√

q +

1

2

(2.39)

The payment by type θ is:

t(θ) =

1

2

θ

2

−

1

8

=

1

2

_

√

q +

1

2

_

2

−

1

8

=

1

2

q +

1

2

√

q (2.40)

Essentially, the monopolist thus oﬀers to consumers the deal that they

can buy any quantity q ∈ [0,

1

4

]. Their payment is t(q) =

1

2

q +

1

2

√

q. This

optimal non-linear pricing scheme is shown in Figure 2.2. We note that

there is a quantity discount. The per unit price:

t

q

=

1

2

q +

1

2

√

q

q

=

1

2

+

1

2

1

√

q

(2.41)

decreases in q.

CHAPTER 2. SCREENING 27

0.00 0.05 0.10 0.15 0.20 0.25

0

.

0

0

.

1

0

.

2

0

.

3

0

.

4

q

t

Figure 2.2: The optimal non-linear pricing scheme

2.4 Bundling

The theory of screening can be developed into many diﬀerent directions.

For example, it is of great interest to consider the case in which the buyer’s

private information is also of relevance to the seller’s assessment of a possible

sale. Insurance contracts are an example. A buyer’s private information

about his health situation aﬀects not only his own evaluation of an insurance

contract, but also the insurance seller’s evaluation of the contract. We shall

not pursue this direction here because it is not related to the theory of

mechanism design as far as it is covered in these notes.

Another important extension is to the case in which the buyer’s private

information is multi-dimensional. The case of multi-dimensional private in-

formation is also relevant in the theory of mechanism design. Therefore,

we give here a simple example of screening when the buyer has multi-

dimensional private information.

Suppose a seller has two distinct indivisible goods, good A and good B,

for sale. For simplicity we assume that the seller values the goods at zero and

is risk neutral, so that she seeks to maximize her expected revenue. Let I

A

and I

B

be indicator variables that are either 1 (if the buyer obtains good A

CHAPTER 2. SCREENING 28

respectively good B) or 0 (if the buyer does not obtain good A respectively

good B). Denote by t the monetary transfer from the buyer to the seller.

Then the buyer’s utility is:

I

A

v

A

+I

B

v

B

−t. (2.42)

Note that utility is additive in the two goods and in money. We assume in

addition that the marginal value of each good does not depend on whether

the other good is also obtained. In this way the two goods are entirely

independent. They are not like pasta and tomato sauce, but they are like

pasta and a watch.

The parameters v

A

and v

B

indicate the buyer’s willingness to pay for the

two goods. These parameters are known to the buyer but not known to the

seller. The seller’s belief about these two parameters is given by the uniform

distribution F over the unit square [0, 1]

2

. Note that we assume here that

v

A

and v

B

are stochastically independent. Thus, we assume a second time

that there is no relation at all between the two goods.

Our interest is again in optimal selling procedures for the seller. As

in the previous sections the revelation principle holds and we could restrict

attention to direct mechanisms. We shall instead simplify our problem much

more, and only consider a very small class of indirect mechanisms. Suppose

the seller considers quoting three prices: p

A

, p

B

and p

AB

. The interpretation

is that the buyer can buy good A at price p

A

, good B at price p

B

, or goods

A and B at price p

AB

. We assume that the seller cannot stop the buyer

from buying goods A and B at price p

A

+ p

B

so that the price p

AB

, if it is

to have any eﬀect, has to satisfy: p

AB

≤ p

A

+p

B

.

What is the optimal choice of p

A

, p

B

and p

AB

? This is a simple calculus

exercise. It turns out that the optimal prices are:

p

A

= p

B

=

2

3

(2.43)

p

AB

=

1

3

_

4 −

√

2

_

≈ 0.862 (2.44)

Note that the optimal price P

AB

is indeed strictly smaller than p

A

+p

B

. The

seller thus oﬀers to the buyer that he can buy the two goods separately, but

that he gets a better deal if he buys the two goods together. The literature

refers to the combination of goods A and B as a “bundle.” The monopolist’s

strategy in our example is also described as “mixed bundling” because the

monopolist oﬀers the bundle to consumers, but he also oﬀers to them the

option to buy the goods individually.

CHAPTER 2. SCREENING 29

0 0.5 2/3 1

0

0

.

5

2

/

3

1

v

A

v

B

nothing

both goods

good A

good B

Figure 2.3: Buyer behavior given optimal prices

The buyer’s demand behavior given the optimal prices is shown in Figure

2.3. Depending on the value of v

A

and v

B

the buyer purchases good A only,

good B only, both goods, or no good. Eﬃciency would, of course, require

that the good is transferred to the buyer for all values of v

A

and v

B

except

zero. We thus observe in Figure 2.3 several distortions of eﬃciency.

Figure 2.3 illustrates the direct mechanism implemented by the seller if

he quotes the three optimal prices. We present this example mainly to il-

lustrate that even in the simple case of screening, multi-dimensional private

information may cause surprising and counterintuitive eﬀects. It is very sur-

prising that the seller oﬀers the goods as a bundle at a discount even though

from the consumer’s point of view the goods are entirely unrelated. The

literature has in fact spent some time seeking to understand the intuition

behind this eﬀect. As the example indicates, the general theory of screen-

ing with multiple goods and multi-dimensional private information is rather

complicated. Hermalin

CHAPTER 2. SCREENING 30

2.5 Comments on the Literature

Our exposition in Section 2.2 is a modiﬁed version of Manelli and Vincent

(2007). The theory of non-linear pricing in Section 2.3 is discussed further in

Bolton and Dewatripont (2005). Finally, the example in Section 2.4 is based

on Chapter 6 of Hermalin (2005). The seminal paper on mixed bundling is

Adams and Yellen (1976). Manelli and Vincent (2007) applies the extreme

point analysis of Section 2.2 to a general screening model and illustrates the

potentially complicated stochastic nature of optimal selling mechanisms.

2.6 Problems

a) Give an example in the setting of Section 2.2 in which the buyer has only

two possible types, and in which the revenue equivalence principle does not

hold.

b) Does Proposition 2.5 hold if the type distribution F is discrete?

c) Prove that the conditions in (2.19) are suﬃcient for incentive compatibil-

ity.

d) Prove that the prices in (2.43) maximize the seller’s proﬁts.

Chapter 3

Classic Bayesian Mechanism

Design: Examples

3.1 Introduction

This chapter describes three classic mechanism design problems using Bayesian

Nash equilibrium to predict agents’ strategic behavior for any given mech-

anism. The next chapter will consider the same examples but we shall

use dominant strategies as our concept for predicting agents’ behavior. We

start with examples to illustrate the general analysis that will follow in later

chapters. The examples are also interesting in their own right. We contrast

in the next chapters optimal Bayesian mechanisms and optimal dominant

strategy mechanisms to illustrates the adjustments that need to be made

to a mechanism if the weaker condition of Bayesian incentive compatibility

is replaced by the more restrictive condition of dominant strategy incentive

compatibility. Understanding this contrast will prepare the reader for our

later discussion of robust mechanism design.

3.2 Single Unit Auctions

3.2.1 Set-Up

The model in this chapter is the same as in Chapter 2.2 except that we now

have more than one potential buyer. A seller seeks to sell a single indivisible

good. There are N ≥ 2 potential buyers. We denote the set of potential

buyers by I = ¦1, 2, ..., N¦. Buyer i’s utility if she purchases the good and

pays a transfer t

i

to the seller is: θ

i

− t

i

. Buyer i’s utility if she does not

31

CHAPTER 3. BAYESIAN MECHANISM DESIGN 32

purchase the good and pays a transfer of t

i

to the seller is: 0−t

i

. The seller’s

utility if he obtains transfers t

i

from buyers i = 1, 2, . . . , N is:

N

i=1

t

i

.

We assume that buyer i knows θ

i

, but that neither the seller nor any

other buyer j ,= i knows θ

i

. We model the valuation θ

i

as a random variable

with cumulative distribution function F

i

with density f

i

. The support of θ

i

is [θ,

¯

θ] where 0 ≤ θ <

¯

θ. Thus, although we do not assume that the random

variables θ

i

have the same distribution for diﬀerent i, we do assume that

they have the same support. This is for convenience only. For technical

convenience we also assume that f

i

(θ

i

) > 0 for all i ∈ I and all θ

i

∈ [θ,

¯

θ].

We assume that for i, j ∈ I with i ,= j the random variables θ

i

and θ

j

are independent. We denote by θ the vector (θ

1

, θ

2

, . . . , θ

N

). The support

of the random variable θ is Θ ≡ [θ,

¯

θ]

N

. The distribution of θ is denoted

by F which is the product of the distributions F

i

, and the density of F is

denoted by f. Each potential buyer i observes θ

i

, but neither the seller nor

the other buyers j ,= i observe θ

i

. The distribution F, however, is common

knowledge among the buyers and the seller.

The model that we have described is known in the literature as a model

with “independent private values.” One also refers to the “independent pri-

vate values” assumption. All examples in this chapter will be built on this

assumption. The assumption is very restrictive. We discuss why in this para-

graph and the next. The phrase “independent” refers to the fact that we

have assumed that values are independent and that they follow a commonly

known prior distribution F. Note ﬁrst that the distribution F describes not

only the seller’s beliefs, as in Chapter 2.2, but also the potential buyers’ be-

liefs about each other. The assumption that buyers’ values are independent

implies that each buyer’s beliefs about the other buyers’ values is indepen-

dent of his own value. So, for example, if buyer i has a high value, he does

not attach more probability to the event that buyer j ,= i has a high value

than if i had had a low value. The assumption that F is a common prior

of the seller and all buyers implies that two buyers i, i

with i, i

,= j have

the same belief about buyer j’s value, and that this belief is also shared by

the seller. The assumption that F is common knowledge implies that it is

commonly known among the sellers and the buyers that they share the same

beliefs about other buyers.

We have assumed private values in the sense that each buyer’s private

information is suﬃcient to determine this buyer’s value of the good. No

buyer would change his value of the good if he knew what other buyers

know. Thus, the private information that leads one buyer to value the good

CHAPTER 3. BAYESIAN MECHANISM DESIGN 33

highly (or not) would not change any other buyer’s value if it was known to

that buyer. The most plausible interpretation is that the private information

is about each buyer’s private tastes rather than about objective features of

the good.

3.2.2 Mechanisms, Direct Mechanisms, and the Revelation

Principle

We will be interested in procedures that the seller can use to sell his good.

For example, he could pick a price, ask each buyer to indicate whether she

is willing to pay this price for the good, and then randomly pick one of the

buyers, if any, who have said that they are willing to buy the good and

transact with this buyer at the announced price.

We will consider a much more general class of methods for selling the

good. We will allow the seller to pick an arbitrary extensive game tree where

the players are the potential buyers and the seller. The seller assigns to each

terminal history of the game (of ﬁnite or inﬁnite length) an outcome, that is

a probability of transferring the good, and if so to whom, and a probability

distribution over vectors of transfers from the buyers to the seller. Formally

an outcome is a probability distribution over ¦0, 1, 2, . . . , N¦ R

N

. Here, 0

stands for the outcome that the good remains with the seller and is hence

not sold.

The seller will ﬁnd it to her advantage to commit in advance to a strat-

egy. Therefore, we might as well eliminate the seller as a player from the

game, and restrict attention to extensive game trees where only the poten-

tial buyers are players. We shall understand by a “mechanism” such an

extensive game tree together with an assignment of a probability distribu-

tion over outcomes to each terminal history. For simplicity we shall not

provide a formal deﬁnition of a “mechanism.”

A mechanism, in conjunction with the assumptions about utilities, in-

formation, and the distribution of types that we made in the previous sub-

section deﬁnes a game of incomplete information. The standard solution

concept for such games is that of a Bayesian Nash equilibrium (Fudenberg

and Tirole (1993), p. 215). Note that games may have none, one, or more

than one, Bayesian Nash equilibrium. We shall imagine that the seller only

proposes games that do have at least one Bayesian Nash equilibrium, and

that, when announcing the mechanism, the seller also proposes a Bayesian

Nash equilibrium of the corresponding game. The buyers will play the equi-

librium that the seller proposes. Thus, if there are multiple equilibria, the

CHAPTER 3. BAYESIAN MECHANISM DESIGN 34

seller can in a sense “pick” which equilibrium the buyers will play. This

assumption is important for the revelation principle, and we will comment

further on it when discussing the revelation principle below.

We assume that the utility that buyers obtain if they walk away from the

mechanism proposed by the seller is zero. Participation in the mechanism

and the equilibrium that the seller proposes must be voluntary, and therefore

we assume that the equilibrium that the seller proposes must oﬀer each

potential buyer an expected utility of at least zero.

We now introduce a subclass of mechanisms, “direct mechanisms,” and

then show that it is without loss of generality to restrict attention to such

mechanisms. In the following deﬁnition ∆ denotes the set of all probability

distributions over the set I of buyers to whom the good might be sold,

and over the possibility of not selling the good. Formally, ∆ is deﬁned by:

∆ ≡ ¦(q

1

, q

2

, . . . , q

N

) [ 0 ≤ q

i

≤ 1 for all i ∈ I and

N

i=1

q

i

≤ 1¦. Note that

the probabilities q

1

, q

2

, . . . , q

N

may add up to less than 1. In this case the

remaining probability, 1 −

N

i=1

q

i

, is the probability that the good is not

sold.

Deﬁnition 3.1. A “direct mechanism” consists of functions q and t

i

(for

i ∈ I) where:

q : Θ →∆

and

t

i

: Θ →R

for i ∈ I.

The interpretation is that in a direct mechanism the buyers are asked

to simultaneously and independently report their types. The function q(θ)

describes the rule by which the good is allocated if the reported type vector

is θ. We shall refer to q as the “allocation rule.” The probability q

i

(θ) is

the probability that agent i obtains the good if the type vector is θ. The

probability 1−

N

i=1

q

i

(θ) is the probability with which the seller retains the

good if the type vector is θ. The functions t

i

describe the transfer payment

that buyer i makes to the seller. Note that we have assumed that this

transfer payment is deterministic. This is without loss of generality.

We now state the “Revelation Principle” which as in Chapter 2 shows

that in some sense there is no loss of generality in restricting attention to

direct mechanisms.

CHAPTER 3. BAYESIAN MECHANISM DESIGN 35

Proposition 3.1 (Revelation Principle). For every mechanisms Γ and Bayesian

Nash equilibrium σ of Γ there exists a direct mechanism Γ

and a Bayesian

Nash equilibrium σ

of Γ

such that:

(i) The strategy vector σ

**satisﬁes for every i and every θ
**

i

:

σ

i

(θ

i

) = θ

i

,

that is, σ

**prescribes telling the truth;
**

(ii) For every vector θ of types the distribution over outcomes that results

in Γ if the agents play σ is the same as the distribution over outcomes

that results in Γ

if the agents play σ

**, and the expected value of the
**

transfer payments that result in Γ it the agents play σ is the same as

the transfer payments that results in Γ

if the agents play σ

.

Proof. Construct Γ

**by deﬁning the functions q and t
**

i

as required by item

(ii) in Proposition 3.1. We can prove the result by showing that truth telling

is a Bayesian equilibrium of the game. Suppose it were not. If type θ

i

prefers

to report that her type is θ

i

, then the same type θ

i

prefers to deviate from

σ, and to play the strategy that σ prescribes for θ

i

in Γ. Hence σ is not a

Bayesian equilibrium of Γ.

Proposition 3.1 shows that in the set-up that we have described it is

without loss of generality to restrict attention to the case that the seller

chooses a direct mechanism and proposes to agents that they report their

types truthfully. Note however, that it is crucial to this construction that

we have neglected problems of multiple equilibria by assuming that agents

follow the seller’s proposal provided that it is an equilibrium, and provided

that it gives them expected utility of at least zero. This is crucial because the

equivalent direct mechanism that is constructed in the proof of Proposition

3.1 might have Bayesian Nash equilibria other than truth-telling, and there

is no reason why these equilibria should be equivalent to any Bayesian Nash

equilibrium of the indirect mechanism Γ. Depending on how equilibria are

selected, one or the other mechanism might be strictly preferred by the seller

in that case.

The revelation principle greatly simpliﬁes our search for optimal mech-

anisms. We can restrict attention to direct mechanisms in which it is a

Bayesian equilibrium that everyone always reports their type truthfully, and

in which every type’s expected utility is at least zero. We want to deﬁne

CHAPTER 3. BAYESIAN MECHANISM DESIGN 36

these properties of a direct mechanism formally. For this we introduce ad-

ditional notation.

We denote by θ

−i

the vector of all types except player i’s type. We deﬁne

Θ

−i

≡ Θ

N−1

. We denote by F

−i

the cumulative distribution of θ

−i

, and we

denote by f

−i

the density of F

−i

. Given a direct mechanism, we deﬁne for

each agent i ∈ I a function Q

i

: [θ,

¯

θ] →[0, 1] by setting:

Q

i

(θ

i

) =

_

Θ

−i

q

i

(θ

i

, θ

−i

)f(θ

−i

)dθ

−i

. (3.1)

Thus, Q

i

(θ

i

) is the conditional expected value of the probability that agent

i obtains the good, conditioning on agent i’s type being θ

i

. We also deﬁne

for each agent i ∈ I a function T

i

: [θ,

¯

θ] →R by setting:

T

i

(θ

i

) =

_

Θ

−i

t

i

(θ

i

, θ

−i

)f(θ

−i

)dθ

−i

. (3.2)

Thus, T

i

(θ

i

) is the conditional expected value of the transfer that agent i

makes to the seller, again conditioning on agent i’s type being θ

i

.

7

Finally,

we also deﬁne agent i’s expected utility U

i

(θ

i

) conditional on her type being

θ

i

. This is given by:

U

i

(θ

i

) = θ

i

Q

i

(θ

i

) −T

i

(θ

i

).

Using this notation we can now formally deﬁne the two conditions that

the seller has to respect when choosing a selling mechanism.

Deﬁnition 3.2. A direct mechanism is “incentive compatible” if truth

telling is a Bayesian Nash equilibrium, that is, if:

θ

i

Q

i

(θ

i

) −T

i

≥ θ

i

Q

i

(θ

i

) −T

i

(θ

i

) for all i ∈ I and θ

i

, θ

i

∈ [θ,

¯

θ].

Deﬁnition 3.3. A direct mechanism is “individually rational” if each agent,

conditional on her type, is voluntarily willing to participate, that is, if:

U

i

(θ

i

) ≥ 0 for all i ∈ I and θ

i

, ∈ [θ,

¯

θ].

To conclude this subsection, it is useful to introduce some further termi-

nology. In the timeline of the game deﬁned by a mechanism, the phase that

follows after agents have learned their types, but before all agents’ types

7

We ignore questions of existence and uniqueness of the conditional expected values

referred to in this paragraph.

CHAPTER 3. BAYESIAN MECHANISM DESIGN 37

are revealed, is often referred to as the “interim” phase. The phase before

agents have learned their types is referred to as the “ex ante” phase, and

the phase after all agents have revealed their types in a direct mechanism is

then called the “ex post” phase. We shall use this terminology occasionally

in these notes. For example, we shall refer to T

i

(θ

i

) as the interim expected

transfer of agent i if she is of type θ

i

, and we shall refer to U

i

(θ

i

) as the

interim expected utility of agent i if she is of type θ

i

.

3.2.3 Characterizing Incentive Compatibility and Individual

Rationality

In this subsection we seek to understand better the structure of the set of all

direct mechanisms that satisfy the two conditions introduced in Deﬁnitions

3.2 and 3.3. We proceed in much the same way as in Chapter 2.2 and

therefore we omit most proofs. We ﬁrst focus on incentive compatibility.

Lemma 3.1. If a direct mechanism is incentive compatible, then for every

agent i ∈ I the function Q

i

is increasing.

The proof of this is the same as the proof of Lemma 2.1, with the func-

tions Q

i

and T

i

replacing the functions q and t. Similarly, we obtain, ana-

logues to Lemmas 2.2, 2.3 and 2.4:

Lemma 3.2. If a direct mechanism is incentive compatible, then for every

agent i ∈ I the function U

i

is increasing. It is also convex, and hence

diﬀerentiable except in at most countably many points. For all θ

i

for which

it is diﬀerentiable, it satisﬁes:

U

(θ

i

) = Q

i

(θ

i

).

Lemma 3.3 (Payoﬀ Equivalence). Consider an incentive compatible direct

mechanism. Then for all I ∈ I and all θ

i

∈ [θ,

¯

θ]:

U

i

(θ

i

) = U

i

(θ) +

_

θ

i

θ

Q

i

(x)dx.

Lemma 3.4 (Revenue Equivalence). Consider an incentive compatible di-

rect mechanism. Then for all I ∈ I and all θ

i

∈ [θ,

¯

θ]:

T

i

(θ

i

) = T

i

(θ) + (θ

i

Q

i

(θ

i

) −θ

i

Q

i

(θ

i

)) −

_

θ

i

θ

Q

i

(x)dx.

CHAPTER 3. BAYESIAN MECHANISM DESIGN 38

Lemmas 3.3 and 3.4 show that the interim expected payoﬀ and the in-

terim expected payment of the diﬀerent types of the buyers are pinned down

by the functions Q

i

and by the expected payoﬀ respectively the expected

payment of the lowest type. Note that this does not mean that the functions

t

i

are uniquely determined. Diﬀerent functions t

i

might give rise to the same

interim expected payments T

i

.

Consider two diﬀerent indirect mechanisms, and Bayesian Nash equilib-

ria of these mechanisms, such that they imply the same interim expected

probability of obtaining the object for each type of each agent, and such that

the expected payment made by the lowest type is the same in the two mech-

anisms. Then Lemma 3.4 implies that all types’ interim expected payments

are the same for these two indirect mechanisms, and therefore, of course, also

the expected revenue of the seller is the same for these two mechanisms. It

is for this reason that the result is called the “revenue equivalence theorem.”

We wish to explain an application of the revenue equivalence theorem.

Consider the symmetric case in which F

i

does not depend on i. Suppose

we wanted to compare auctioneer’s expected revenue from the second price

auction with minimum bid 0 to the expected revenue from the ﬁrst price

auction with minimum bid 0. In the second price auction it is a weakly

dominant strategy, and hence a Bayesian Nash equilibrium, to bid one’s

true value. A symmetric Bayesian Nash equilibrium for the ﬁrst price auc-

tion is constructed in Proposition 2.2 of Krishna (2002). This equilibrium is

in strictly increasing strategies. Hence this equilibrium shares with the equi-

librium of the second price auction that the expected payment of the lowest

type is zero (because this type’s probability of winning is zero), and that

the highest type wins with probability 1. Therefore, the equilibria imply the

same values for T

i

(θ) and Q

i

(θ

i

) for all i ∈ I and θ

i

∈ [θ,

¯

θ]. The revenue

equivalence theorem implies that the expected revenue from the equilibria

of the two diﬀerent auction formats is the same.

We described in the previous paragraph the most famous application of

Lemma 3.4. But note that the Lemma is much more general. In shorthand

expression, the Lemma says that the interim expected payments of all types

only depend on the interim expected allocation rule and the interim expected

payment of the lowest type.

As in Chapter 2.2 we can collect the observations made so far, and

obtain conditions that are not only necessary but also suﬃcient for incentive

compatibility. The proof is analogous to the proof of Proposition 2.2 and is

therefore omitted.

CHAPTER 3. BAYESIAN MECHANISM DESIGN 39

Proposition 3.2. A direct mechanism (q, t

1

, t

2

, . . . , t

N

) is incentive com-

patible if and only if for every i ∈ I:

(i) Q

i

is increasing;

(ii) For every θ

i

∈ [θ,

¯

θ]:

T

i

(θ

i

) = T

i

(θ) + (θ

i

Q

i

(θ

i

) −θQ

i

(θ)) −

_

θ

θ

Q

i

(

ˆ

θ)d

ˆ

θ

We have now obtained a complete understanding of the implications

of incentive compatibility for the seller’s choice. The seller can focus on

two choice variables: ﬁrstly the allocation rule q, and secondly the interim

expected payments by the lowest types: T(θ

i

). As long as the seller picks

an allocation rule q such that the functions Q

i

(i ∈ I) are increasing, he

can pick the interim expected payments by the lowest types in any arbitrary

way, and be assured that there will be some transfer scheme that makes

the allocation rule incentive compatible and that implies the given interim

expected payments by the lowest types. Moreover, any such transfer scheme

will give him the same expected revenue, and therefore the seller does not

have to worry about the details of this transfer scheme.

So far we have focused on the characterization of incentive compatibility.

Now we turn to individual rationality. However, we restrict attention to

incentive compatible direct mechanisms. Then we have the following result

that is analogous to Lemma 2.3.

Proposition 3.3. An incentive compatible direct mechanism is individually

rational if and only if for every i ∈ I we have: T

i

(θ

i

) ≤ θ

i

Q

i

(θ

i

).

Thus, we have one further constraint on the seller’s choice of direct

mechanism. The seller has to choose a mechanism that implies an expected

utility of at least zero for the lowest type agents.

3.2.4 Expected Revenue Maximization

We now study the expected revenue maximizing choice of selling mechanism.

We begin with a simple observation that is analogous to Lemma 2.5.

Lemma 3.5. If an incentive compatible and individually direct mechanism

maximizes the seller’s expected revenue then for every i ∈ I:

T

i

(θ) = θQ

i

(θ).

CHAPTER 3. BAYESIAN MECHANISM DESIGN 40

We can now simplify the seller’s problem further. The seller has to choose

a function q so that the interim probabilities Q

i

are increasing for all i ∈ I.

The payments are then completely determined by part (ii) of Proposition

3.2 and Lemma 3.5. Substituting the formula in Lemma 3.5 into part (ii) of

Proposition 3.2 we get for every i ∈ I and θ

i

∈ [θ,

¯

θ]:

T

i

(θ

i

) = θ

i

Q

i

(θ

i

) −

_

θ

θ

Q

i

(

ˆ

θ)d

ˆ

θ (3.3)

Note that for the seller’s expected revenue the details of the function q don’t

matter, but only the interim probabilities Q

i

.

We shall now focus on the optimal choice of q. We shall proceed as in

Chapter 2.3 and not as in Chapter 2.2. The reason is that there are many

extreme points of the seller’s choice set. These were the focus of Chapter

2.2, were it was suﬃcient to characterize these extreme points, but in our

context a characterization of these extreme points wouldn’t take us very far.

Doing the same calculations that lead in Chapter 2.3 to equation (2.25) we

can calculate the seller’s expected revenue from any particular buyer i.

_

¯

θ

θ

Q

i

(θ

i

)

_

θ

i

−

1 −F

i

(θ

i

)

f

i

(θ

i

)

_

f

i

(θ

i

)dθ

i

. (3.4)

To obtain a formula for the total expected transfer by all agents we add the

formula in equation (3.4) over all i ∈ I. We obtain:

N

i=1

_

_

¯

θ

θ

Q

i

(θ

i

)

_

θ

i

−

1 −F

i

(θ

i

)

f

i

(θ

i

)

_

f

i

(θ

i

)dθ

i

_

=

N

i=1

_

_

_

Θ

q

i

(θ)

_

θ

i

−

1 −F

i

(θ

i

)

f

i

(θ

i

)

_

f(θ)dθ

_

_

(3.5)

where the last equality becomes obvious if one recalls the deﬁnition of Q

i

(θ

i

).

As in Chapter 2.3 we ﬁrst ask which function q the seller would choose

if he did not have to make sure that the functions Q

i

are increasing. In a

second step we introduce an assumption that makes sure that the optimal

q from the ﬁrst step implies increasing functions Q

i

. If monotonicity could

be ignored, then the seller would choose for each θ the probabilities q

i

(θ) so

as to maximize the expression in the large round brackets in the formula for

expected revenue. We deﬁne this expression to be ψ

i

(θ

i

):

ψ

i

(θ

i

) ≡ θ

i

−

1 −F

i

(θ

i

)

f

i

(θ

i

)

for all i ∈ I and θ

i

∈ [θ,

¯

θ]. (3.6)

CHAPTER 3. BAYESIAN MECHANISM DESIGN 41

The optimal allocation rule without monotonicity is then:

q

i

(θ) =

_

_

_

1 if ψ

i

(θ

i

) > 0 and ψ

i

(θ

i

) > ψ

j

(θ

j

) for all j ∈ I with j ,= i;

0 otherwise.

for all i ∈ I and θ ∈ Θ. (3.7)

Note that we have ignored the case that ψ

i

(θ

i

) = ψ

j

(θ

j

) for some j ,= i.

This is a zero probability event, and it does not aﬀect either the buyer’s

incentives nor the seller’s revenue.

We now introduce an assumption under which this allocation rule satis-

ﬁes the monotonicity constraint of the seller’s maximization problem. The

assumption is that for all agents i ∈ I the distribution functions F

i

are

“regular” in the same sense as in Assumption 2.1.

Assumption 3.1. For every i ∈ I the function ψ

i

(θ

i

) is strictly increasing.

For the allocation rule q described above the probability Q

i

(θ

i

) is the

probability that ψ

i

(θ

i

) is larger than zero and larger than ψ

i

(θ

j

) for every

j ∈ I with j ,= i. Clearly, if ψ

i

is increasing, as required by the regularity

assumption, this probability is an increasing function of θ

i

. Thus, Q

i

is

indeed increasing. We have arrived at the following result.

Proposition 3.4 (Myerson (1981)). Suppose that for every agent i ∈ I the

cumulative distribution function F

i

is regular. Among all incentive compati-

ble and individually rational direct mechanisms those mechanisms maximize

the seller’s expected revenue that satisfy for all i ∈ I and all θ ∈ Θ:

(i)

q

i

(θ) =

_

_

_

1 if ψ

i

(θ

i

) > 0 and ψ

i

(θ

i

) > ψ

i

(θ

i

) for all i ∈ I with i

,= i;

0 otherwise.

(ii)

T

i

(θ

i

) = θ

i

Q

i

(θ

i

) −

_

θ

θ

Q

i

(

ˆ

θ)d

ˆ

θ.

We have characterized the optimal choice of the allocation rule q and

of the interim expected payments. We have not described the actual trans-

fer schemes that make these choices incentive compatible and individually

rational, although we know that such transfers can be found.

CHAPTER 3. BAYESIAN MECHANISM DESIGN 42

The expression ψ

i

(θ

i

) is sometimes referred to as seller i’s “virtual type.”

Using this expression we can rephrase the result in Proposition 3.4 as follows:

The expected revenue maximizing auction allocates the object to the buyer

with the highest virtual type, provided that this type is at least zero.

If buyers are symmetric, i.e. the distribution functions F

i

are all the

same, the optimal mechanism prescribes that the object is given to the

buyer with the highest value, if it is sold at all. This is because then ψ

i

is

the same for all i, and hence ψ

i

(θ

i

) > ψ

j

(θ

j

) ⇔ θ

i

> θ

j

. Note that in the

asymmetric case this need not be the case.

In the symmetric case the optimal direct mechanism can be implemented

using either a ﬁrst or a second price auction with minimum bid ψ

−i

i

(0) where

ψ

−i

i

is the inverse of any one of the functions ψ

i

. Thus, in the symmetric

case, familiar auction institutions, with appropriately chosen minimum bid,

are optimal. To show this one has to derive equilibrium bidding functions

for these auctions, and verify that they imply the allocation rule and the

transfer payments indicated in Proposition 3.4. An excellent reference on

this is Chapter 2 of Krishna (2002).

3.2.5 Maximizing Welfare

Suppose that the seller were not maximizing expected proﬁts but expected

welfare. Let us assume that the seller uses the following utilitarian welfare

deﬁnition:

N

i=1

q

i

(θ)θ

i

.

Note that this seller is no longer concerned with transfer payments. Ex-

pected welfare depends only on the allocation rule q.

We can easily analyze this seller’s problem using the framework described

before. The seller can choose any rule q that is such that the functions Q

i

are monotonically increasing. He can choose any transfer payments such

that T

i

(θ

i

) ≤ θ

i

Q

i

(θ

i

) for all i ∈ I.

Which rule q should the seller choose? If types were known, maximiza-

tion of the welfare function would require that the object be allocated to the

potential buyer for whom θ

i

is largest. Note that welfare if the object is not

transferred to one of the potential buyers is assumed to be zero, and hence

the welfare maximizing seller always wants to transfer the object. Can the

welfare maximizing seller allocate the object to the buyer with the highest

CHAPTER 3. BAYESIAN MECHANISM DESIGN 43

value even if he doesn’t know the valuations? We know that this is possi-

ble if the implied functions Q

i

are increasing. This is obviously the case.

Therefore we conclude:

Proposition 3.5. Among all incentive compatible, individually rational di-

rect mechanisms a mechanism maximizes welfare if and only if for all i ∈ I

and all θ ∈ Θ:

(i)

q

i

(θ) =

_

_

_

1 if θ

i

> θ

j

for all j ∈ I with j ,= i;

0 otherwise.

(ii)

T

i

(θ

i

) ≤ θ

i

Q

i

(θ

i

) −

_

θ

θ

Q

i

(

ˆ

θ)d

ˆ

θ.

Note that this result does not rely on Assumption 3.1. In comparing

welfare maximizing and revenue maximizing mechanisms in the case that

Assumption 3.1 holds we observe that there are two diﬀerences. The ﬁrst

is that revenue maximizing mechanism allocates the object to the highest

virtual type whereas the welfare maximizing mechanism allocates the object

to the highest actual type. In the symmetric case the functions ψ

i

are the

same for all i ∈ I and there is no diﬀerence between these two rules. But

in the asymmetric case the revenue maximizing mechanism might allocate

the object ineﬃciently. A second diﬀerence is that the revenue maximizing

mechanism sometimes does not sell the object at all, whereas the welfare

maximizing mechanism always sells the object. This is an instance of the

well-known ineﬃciency that monopoly sellers make goods artiﬁcially scarce.

Proposition 3.5, like Proposition 3.4, does not describe the transfer

scheme associated with a welfare maximizing mechanism in detail. An ex-

ample of a direct mechanism that is incentive compatible and individually

rational and that maximizes welfare is the second price auction with reserve

price zero.

3.2.6 Numerical Examples

We give one symmetric and one asymmetric example.

CHAPTER 3. BAYESIAN MECHANISM DESIGN 44

Example 3.1. Suppose that θ = 0,

¯

θ = 1, and that θ

i

is uniformly dis-

tributed so that F(θ

i

) = θ

i

. We begin by calculating for i = 1, 2:

ψ

i

(θ

i

) = θ

i

−

1 −F

i

(θ

i

)

f

i

(θ

i

)

= θ

i

−

1 −θ

i

1

= 2θ

i

−1. (3.8)

Note that the regularity assumption 3.1 is satisﬁed.

In the expected revenue maximizing auction the good is sold to neither

bidder if:

ψ

i

(θ

i

) < 0 ⇔

2θ

i

−1 < 0 ⇔

θ

i

<

1

2

(3.9)

holds for i = 1 and i = 2. If the good is sold, it is sold to bidder 1 if:

ψ

1

(θ

1

) > ψ

2

(θ

2

) ⇔

2θ

1

−1 > 2θ

2

−1 ⇔

θ

1

> θ

2

. (3.10)

The expected revenue maximizing auction will allocate the object to the buyer

with the highest type provided that this type is larger than 0.5. A ﬁrst or

second price auction with reserve bid

1

2

will implement this mechanism. A

ﬁrst or second price auction with reserve bid 0 maximizes expected welfare.

Example 3.2. Now suppose that N = 2, θ = 0,

¯

θ = 1, and that F

1

(θ

1

) =

(θ

1

)

2

whereas F

2

(θ

2

) = 2θ

2

− (θ

2

)

2

. Thus, player 1 is more likely to have

high values than player 2.

We begin by calculating

ψ

1

(θ

1

) = θ

1

−

1 −F

1

(θ

1

)

f

1

(θ

1

)

= θ

1

−

1 −(θ

1

)

2

2θ

1

=

3

2

θ

1

−

1

2θ

1

(3.11)

CHAPTER 3. BAYESIAN MECHANISM DESIGN 45

and

ψ

2

(θ

2

) = θ

2

−

1 −F

2

(θ

2

)

f

2

(θ

i

)

= θ

2

−

1 −2θ

2

+ (θ

2

)

2

2 −2θ

2

= θ

2

−

1 −θ

2

2

=

3

2

θ

2

−

1

2

(3.12)

Again the regularity assumption is satisﬁed.

In an expected revenue maximizing auction the good is sold to neither

bidder if:

ψ

1

(θ

1

) < 0 ⇔

3

2

θ

1

−

1

2θ

1

< 0 ⇔

θ

1

<

_

1

3

(3.13)

and

ψ

2

(θ

2

) < 0 ⇔

3

2

θ

2

−

1

2

< 0 ⇔

θ

2

<

1

3

(3.14)

If the good is sold, it is sold to bidder 1 if:

ψ

1

(θ

1

) > ψ

2

(θ

2

) ⇔

3

2

θ

1

−

1

2θ

1

>

3

2

θ

2

−

1

2

⇔

θ

2

< θ

1

−

1

3θ

1

+

1

3

(3.15)

Figure 3.1 shows the optimal allocation of the good. The 45

◦

-line is shown

as a dashed line. Note that the mechanism is biased against player 1. If

the good is sold, bidder 1 wins the object only in a subset of all cases where

his value is higher than bidder 2’s value. In the expected welfare maximizing

mechanism the object is allocated to player 1 if and only if his value is higher

than player 2’s value. A second price auction will maximize expected welfare,

although a ﬁrst price auction will not necessarily.

CHAPTER 3. BAYESIAN MECHANISM DESIGN 46

0.0 0.2 0.4 0.6 0.8 1.0

0

.

0

0

.

2

0

.

4

0

.

6

0

.

8

1

.

0

θθ

1

θθ

2

buyer 2

buyer 1 no sale

Figure 3.1: Expected revenue maximizing allocation in Example 3.2

3.3 Public Goods

3.3.1 Set-Up

Our next example is a public goods problem. The theory of mechanism

design began with the theory of mechanisms for the provision of public

goods. This is a central application of the theory of mechanism design.

Methodologically, the example that we discuss in this section illustrates

the design of optimal mechanisms subject to additional constraints beyond

incentive compatibility and individual rationality constraints. The speciﬁc

constraint on which we shall focus here is the government budget constraint.

We consider a community consisting of N agents: I = ¦1, 2, . . . , N¦

where N ≥ 2. They have to choose whether to produce some indivisible,

non-excludable public good. We denote this decision by g ∈ ¦0, 1¦. If the

public good is produced, then g = 1. If it is not produced, then g = 0.

Agent i’s utility if the collective decision is g and if she pays a transfer

t

i

to the community is: θ

i

g −t

i

. Here, θ

i

is a random variable that follows a

continuous distribution function F

i

with density f

i

. We shall refer to θ

i

as

CHAPTER 3. BAYESIAN MECHANISM DESIGN 47

agent i’s type, or as agent i’s valuation of the public good. The support of

θ

i

is [θ,

¯

θ] where 0 ≤ θ <

¯

θ. We assume that f

i

(θ) > 0 for all θ

i

∈ [θ,

¯

θ].

We assume that for i, j ∈ I with i ,= j the random variables θ

i

and θ

j

are

independent. We also assume that each agent i observes θ

i

, but not the other

agent’ types θ

j

where j ,= i. We denote by θ the vector (θ

1

, θ

2

, . . . , θ

N

). The

support of the random variable θ is Θ = [θ,

¯

θ]

N

. The cumulative distribution

function of θ will be denoted by F, and its density by f. The distribution

F is common knowledge among the agents. We are thus considering an

independent private values model of public goods.

The fact that the public good is non-excludable is reﬂected by the fact

that the same variable g enters into all individuals utility function. An

alternative model, that is also of interest, is a model in which individuals

can be selectively excluded from consuming the public good.

The costs of producing the public good are assumed to be c > 0, so that

a collective decision g implies costs cg. We shall consider this society from

the perspective of a benevolent mechanism designer who does not observe

θ, but who knows F. We attribute to the mechanism designer a utilitarian

welfare function with equal welfare weights for all agents. Welfare is thus:

_

N

i=1

θ

i

_

g −

N

i=1

t

i

. (3.16)

3.3.2 Incentive Compatible and Individually Rational Direct

Mechanisms

As in previous parts of these notes, it is without loss of generality to restrict

attention to incentive compatible direct mechanisms where agents’ payments

are not random. To simplify our treatment of the budget constraint we

also restrict attention to mechanisms where the decision about the public

good is non-stochastic. However, stochastic mechanisms would not present

conceptual problems, and the results of this section hold even if stochastic

mechanisms are considered.

Deﬁnition 3.4. A “direct mechanism” consists of functions q and t

i

(for

i ∈ I) where:

q : Θ →¦0, 1¦

and

t

i

: Θ →R.

CHAPTER 3. BAYESIAN MECHANISM DESIGN 48

The function q assigns to each type vector θ the collective decision about

the public good that is produced if the agents’ types are θ. We shall refer to

q also as the “decision rule.” For each agent i the function t

i

describes for

every type vector θ the transfer that agent i makes when the types are θ.

Given a direct mechanism, we deﬁne for each agent i ∈ I functions

Q

i

: [θ,

¯

θ] →[0, 1] and T

i

: [θ,

¯

θ] →R where Q

i

(θ

i

) is the interim conditional

probability that the public good is produced, where we condition on agent

i’s type being θ

i

, and T

i

(θ

i

) is the interim conditional expected value of the

transfer that agent i makes to the community, again conditioning on agent

i’s type being θ

i

. Finally, we also deﬁne agent i’s expected utility U

i

(θ

i

)

conditional on her type being θ

i

. This is given by: U

i

(θ

i

) = Q

i

(θ

i

)θ

i

−T

i

(θ

i

).

As before we shall restrict attention to mechanisms that are incentive

compatible and individually rational. As the notation that we have intro-

duced in the previous paragraph parallels that of Section 3.2 we can refer

to Deﬁnitions 3.2 and 3.3 for deﬁnitions and to Propositions 3.2 and 3.3 for

characterizations of incentive compatibility and individual rationality.

3.3.3 Ex ante and Ex post Budget Balance

We now introduce the government budget constraint. This constraint re-

quires that the money raised by the mechanism is at least enough to cover

the costs of producing the public good. A restrictive version of the constraint

requires budget balance for each realization of agents’ types.

Deﬁnition 3.5. A direct mechanism is “ex post budget balanced” if for

every θ ∈ [0,

¯

θ]

N

we have

N

i=1

t

i

(θ) ≥ cq(θ).

An alternative formulation requires budget balance to hold only on av-

erage, across realization of agents’ types.

Deﬁnition 3.6. A direct mechanism is “ex ante budget balanced” if

_

Θ

N

i=1

t

i

(θ)f(θ)dθ ≥

_

Θ

cq(θ)f(θ)dθ.

Clearly, ex post budget balance implies ex ante budget balance. Ex post

budget balance appears to be more restrictive. We shall show that in our

context this appearance is misleading. For every ex ante budget balanced

mechanism there is an equivalent ex post budget balanced mechanism. Here,

we deﬁne “equivalent” as follows:

CHAPTER 3. BAYESIAN MECHANISM DESIGN 49

Deﬁnition 3.7. Two direct mechanisms are “equivalent” if for all agents

i ∈ I and for all types θ

i

, θ

i

∈ [θ,

¯

θ] agent i’s expected transfers, conditional

on agent i’s type being θ

i

and agent i reporting to be type θ

i

, is the same

in the two mechanisms.

Notice that if two mechanisms are equivalent, and if one of them is

incentive compatible, then the same is true for the other, and if one of them

is individually rational, then the same is true for the other.

Proposition 3.6. For every direct mechanism with decision rule q that is ex

ante budget balanced there is an equivalent direct mechanism with the same

decision rule q that is ex post budget balanced.

Proof. Suppose ﬁrst that the ex ante budget balance condition holds with

equality. We denote the payments in the ex ante budget balanced mecha-

nism by t

i

. We denote by T

j

(θ

i

) the expected value of agent j’s transfer,

conditioning on agent i’s signal being θ

i

. So far, we have employed this

notation only in the case that j = i. Now we use it also in the case that

j ,= i.

We construct the payments in the ex post budget balanced scheme by

modifying the payments in the ex ante budget balanced scheme as follows.

One arbitrarily selected agent, say agent 1, provides the primary coverage of

the deﬁcit. However, she does not cover that part of the ex post deﬁcit that is

predicted by her own signal. In other words, we add to her original payment

the ex post deﬁcit minus the expected value of the deﬁcit conditional on her

own signal. Formally, if the vector of types is θ, then agent 1’s payment is

in the modiﬁed mechanism:

t

1

(θ) +

_

cq(θ) −

N

i=1

t

i

(θ)

_

−

_

cQ

1

(θ

1

) −

N

i=1

T

i

(θ

1

)

_

. (3.17)

We now check that agent 1’s expected payoﬀ, conditional on her type being

θ

1

and her reporting that her type is θ

1

is unchanged. The expression that

we have added to agent 1’s payment is a random variable with expected

value zero, independent of whether agent 1 reports her type truthfully or

not. This is because it would have expected value zero if agent 1’s true type

were θ

1

, and, moreover, the conditional distribution of this random variable,

by the independence of private signals, is the same if agent 1’s true signal is

θ

1

rather than θ

1

.

Second, some other arbitrarily selected agent, say agent 2, pays for the

expected value of the deﬁcit conditional on agent 1’s signal. Her modiﬁed

CHAPTER 3. BAYESIAN MECHANISM DESIGN 50

payment is:

t

2

(θ) +cQ

1

(θ

1

) −

n

i=1

T

i

(θ

1

). (3.18)

Note that the random variable that we are adding to agent 2’s payment is

independent of agent 2’s report. Moreover, it has ex ante expected value

zero because the mechanism is ex ante budget balanced. Finally, because

agent 2’s signal does not provide any information about agent 1’s signal, the

expectation of the added term conditional on agent 2’s signal is the same

as its ex ante expectation. Therefore, agent 2’s expected payoﬀ, if her true

type is θ

2

, and she reports that her type is θ

2

, is unchanged.

Finally, all agents i ,= 1, 2 pay the same as before: t

i

(θ). Adding up

all agents’ payments shows that the sum of the payments equals the costs

cq(θ) in each state θ, and that therefore this mechanism is ex post budget

balanced.

If there is an ex ante budget surplus, we subtract from some agent’s

payments a constant until the mechanism is exactly budget balanced. Then

we conduct the transformation described above. Then we add the constant

again to this agent’s payments. The mechanism that we obtain has the

required properties.

We shall assume that the mechanism designer in this section considers

only ex post budget balanced mechanisms. The above result makes clear

that this is equivalent to requiring ex ante budget balance. We shall work

with either condition, whichever is more convenient.

We calculate the ex ante expected revenue from an incentive compatible

mechanism in the same way as we calculated above the expected revenue

from an incentive compatible auction. This yields the following formula

where we ﬁnd it convenient to write the initial term as interim expected

utility of the lowest type.

N

i=1

−U

i

(θ) +

_

Θ

q(θ)

_

N

i=1

_

θ

i

−

1 −F

i

(θ

i

)

f

i

(θ

i

)

_

−c

_

f(θ)dθ. (3.19)

3.3.4 Welfare Maximization

Which mechanism (q, t

1

, t

2

, . . . , t

N

) would the designer choose if she were

not constrained by incentive compatibility and individual rationality, but

CHAPTER 3. BAYESIAN MECHANISM DESIGN 51

had to satisfy ex post budget balance? As the mechanism designer sub-

tracts transfer payments in her welfare function (3.59) she would never raise

transfers larger than what is required to cover cost. From (3.59) it is then

clear that the optimal decision rule is:

8

q

∗

(θ) =

_

_

_

1 if

N

i=1

θ

i

≥ c

0 otherwise.

(3.20)

As welfare function (3.59) indicates, the utilitarian designer with equal wel-

fare weights for all agents does not care how costs are distributed among

agents. Therefore, any transfer rules that satisfy:

N

i=1

t

∗

i

(θ) =

_

_

_

c if q(θ) = 1

0 otherwise.

(3.21)

are optimal. We call these direct mechanisms “ﬁrst best.”

The following impossibility result shows that in all non-trivial cases no

ﬁrst best mechanism is incentive compatible and individually rational.

Proposition 3.7. An incentive compatible and individually rational ﬁrst

best mechanism exists if and only if either Nθ ≥ c or N

¯

θ ≤ c.

The condition Nθ ≥ c means that even if all agents have the lowest

valuation the sum of the valuations is at least as high as the cost of producing

the public good. Thus, for all type vectors it is eﬃcient to produce the public

good. Analogously, N

¯

θ ≤ c means that for all type vectors it is eﬃcient no

to produce the public good. These are trivial cases. For all non-trivial cases,

Proposition 3.7 is an impossibility result.

Proof. If Nθ ≥ c then a mechanism where the public good is always pro-

duced and all agents pay

c

N

is ﬁrst best and incentive compatible and in-

dividually rational. If N

¯

θ ≤ c then a mechanism where the public good is

never produced and no agent ever pays anything is ﬁrst best and incentive

compatible and individually rational.

It remains to prove the converse. Thus we consider the case Nθ < c <

N

¯

θ, and wish to prove that there is no incentive compatible and individually

rational ﬁrst best mechanism. To prove this we display a direct mechanism

8

In this deﬁnition, we require arbitrarily that q

∗

(θ) = 1 if

P

N

i=1

θi = c. This simpliﬁes

the exposition, and could be changed without conceptual diﬃculty.

CHAPTER 3. BAYESIAN MECHANISM DESIGN 52

that makes the ﬁrst best decision rule q

∗

incentive compatible and indi-

vidually rational. We then argue that this mechanism maximizes expected

transfer payments among all incentive compatible and individually rational

ﬁrst best mechanisms. Finally, we show that it yields an expected budget

deﬁcit in all non-trivial cases. The assertion then follows.

Deﬁnition 3.8. The “pivot mechanism” is the mechanism that is given by

the ﬁrst best decision rule q

∗

and by the following transfer scheme:

t

i

(θ) = θq

∗

(θ, θ

−i

) + (q

∗

(θ) −q

∗

(θ, θ

−i

))

_

_

c −

j=i

θ

j

_

_

for all i ∈ I and θ ∈ [θ,

¯

θ]

N

.

To see why this mechanism is called “pivot” mechanism it is useful to

ignore the ﬁrst term in the sum on the right hand side of the formula for

t

i

(θ). This term does not depend on agent i’s report θ

i

. The second term

equals the change to the social welfare of all other agents caused by agent

i’s report, and hence agent i pays only if her report is “pivotal” for the

collective decision. Here we compare the actual outcome to the outcome

that would have occurred had agent i reported the lowest type θ. Agent i’s

report changes the collective decision if q

∗

(θ) −q

∗

(θ, θ

−i

) = 1. In that case

agent i pays for the diﬀerence between the costs of the project, and the sum

of all other agents’ valuations of the project.

Lemma 3.6. The pivot mechanism is incentive compatible and individually

rational.

Proof. Consider an agent i ∈ I who is of type θ

i

and who contemplates

reporting that she is of type θ

i

,= θ

i

. Fix the other agents’ types as θ

−i

.

We are going to show that truthful reporting is optimal whatever the other

agents’ types θ

−i

are. This obviously implies that truthful reporting is a

Bayesian equilibrium. If we leave out terms that do not depend on agent i’s

report, then agent i’s utility if reporting θ

i

is:

θ

i

q

∗

(θ

i

, θ

−i

) −q

∗

(θ

i

, θ

−i

)

_

_

c −

j=i

θ

j

_

_

= q

∗

(θ

i

, θ

−i

)

_

_

N

j=1

θ

j

−c

_

_

. (3.22)

CHAPTER 3. BAYESIAN MECHANISM DESIGN 53

Thus agent i’s utility is true social welfare if the collective decision is

q

∗

(θ

i

, θ

−i

). Because q

∗

is ﬁrst best, agent i’s utility is maximized if she

reports truthfully θ

i

= θ

i

. This proves incentive compatibility. To verify

individual rationality note that the expected utility of agent i obviously

equals zero if her type is θ. By a result analogous to Lemma 3.2 this implies

that all types’ interim expected utility is at least zero, and the mechanism

is individually rational.

The proof reveals two important features of the formula in Deﬁnition

3.9. The ﬁrst is that those parts of agent i’s transfer payment that depend

on agent i’s report are chosen so that agent i’s incentives are exactly aligned

with social welfare. The second feature is that those parts of agent i’s

transfer payment that do not depend on agent i’s report are chosen so as

to equalize agent i’s utility if she is of the lowest type with her reservation

utility of zero. This ensure that individual rationality is satisﬁed for all

agents.

The “pivot mechanism” is a special “Vickrey-Clarke-Groves” (VCG)

mechanism. In general, a mechanism is a VCG mechanism if every agent’s

payment consists of two terms. Firstly, a part that depends on the agent’s

report, and that has the eﬀect of aligning the agent’s incentives with social

welfare. Secondly, a part that does not depend on the agent’s report. In a

VCG mechanism in general, this second part can be arbitrary. In a pivot

mechanism, the second part is chosen so as to ensure individual rationality

as an equality constraint for the type whose individual rationality constraint

is “most restrictive.”

9

The next result indicates the special importance of

the pivot mechanism.

Lemma 3.7. No incentive compatible and individually rational mechanism

that implements the ﬁrst best decision rule q

∗

has larger expected surplus

than the pivot mechanism.

Proof. By Corollary 3.19 the expected budget surplus of an incentive com-

patible direct mechanism that implements the ﬁrst best decision rule q

∗

equals the interim expected payments of the lowest types plus a term that is

the same for all such rules. If a mechanism is individually rational then the

interim expected payments of the lowest types can be at most such that the

expected utility of the lowest types are zero. For the pivot mechanism the

9

A type’s individual rationality constraint is “most restrictive” if that type’s individual

rationality constraint implies all other type’s individual rationality constraints. See Section

4 for an example where this type is not the lowest type in types’ numerical order.

CHAPTER 3. BAYESIAN MECHANISM DESIGN 54

expected utilities of the lowest types are exactly equal to zero. Therefore,

no incentive compatible, individually rational direct mechanism can have

higher expected surplus than the pivot mechanism.

We conclude the proof by showing that the pivot mechanism has an ex

ante expected deﬁcit except in trivial cases.

Lemma 3.8. If Nθ < c < N

¯

θ, then the ex ante expected surplus of the pivot

mechanism is negative.

Proof. We show that the ex post surplus of the pivot mechanism is always

non-positive and with positive probability negative. This implies that the

ex ante expected surplus is negative. Consider ﬁrst θ such that q

∗

(θ) = 0.

In this case, there are no costs and no agent pays any transfer. Hence the

deﬁcit is zero. Consider next θ such that q

∗

(θ) = 1 and q

∗

(θ, θ

−i

) = 1 for

every i ∈ I. In this case each agent pays θ. By assumption the probability

of the production of the good according to the ﬁrst best decision rule is not

1, and hence Nθ < c. Therefore, total payments are less than c and there is

a deﬁcit.

Consider ﬁnally states θ such that q

∗

(θ) = 1, and q

∗

(θ, θ

−i

) = 0 for some

i ∈ I. Let P be the set of all i ∈ I for which this holds, and call these agents

“pivotal.” Deﬁne NP = I ¸ P. Abusing notation slightly denote by P the

number of elements of P, and by NP the number of elements of NP. The

total transfers are:

ı∈P

_

_

c −

j=i

θ

j

_

_

+

i∈NP

θ

=Pc −P

j∈NP

θ

j

−(P −1)

j∈P

θ

j

+

i∈NP

θ

=Pc −(P −1)

j∈NP

θ

j

−(P −1)

j∈P

θ

j

−

i∈NP

(θ

i

−θ)

=Pc −(P −1)

j∈I

θ

j

−

i∈NP

(θ

i

−θ). (3.23)

We shall show that this is no more than c:

Pc −(P −1)

j∈I

θ

j

−

i∈NP

(θ

i

−θ) ≤ c ⇔

(P −1)c ≤ (P −1)

j∈I

θ

j

+

i∈NP

(θ

i

−θ) ⇐

CHAPTER 3. BAYESIAN MECHANISM DESIGN 55

(P −1)c ≤ (P −1)

j∈I

θ

j

⇔

c ≤

j∈I

θ

j

(3.24)

which is true by construction in states in which q

∗

(θ) = 1.

States in which the public good is produced and some agents are pivotal

occur with positive probability under our assumption. Moreover, conditional

on such a state occurring, with probability 1 we have that θ

i

> θ for all

i ∈ NP. In this case, the above calculation shows that the surplus is strictly

negative, and hence there is an expected deﬁcit.

This concludes the proof of Proposition 3.7.

In the remainder of this subsection we focus on the case in which it is

by Proposition 3.7 impossible to implement q

∗

using a mechanism that is

incentive compatible, individually rational, and ex ante budget balanced.

Our objective is to determine direct mechanisms that maximize expected

welfare among all incentive compatible, individually rational, and ex ante

budget balanced mechanisms. We shall refer to these mechanism as “second

best.”

It is without loss of generality to assume that the mechanism designer

balances the budget exactly rather than leaving a surplus. If there is a sur-

plus then the mechanism designer can return it to the agents. Moreover, the

mechanism designer can achieve ex post budget balance when she achieves

ex ante budget balance. Thus, in each state in which the public good is

produced, payments will add up to c, and in other states they will be add

up to zero. The designer’s objective function can therefore be written as:

_

Θ

q(θ)

_

i∈I

θ

i

−c

_

f(θ)dθ (3.25)

It seems at ﬁrst sight most natural from our discussion so far to regard

q and the interim expected payments of the lowest types, T

i

(θ), as the

designer’s choice variables. However, it is equivalent, and more convenient,

to think of q and the interim expected utilities of the lowest types, U

i

(θ)

as the choice variables. The constraints that these variables have to satisfy

are:

for every i ∈ I the function Q

i

is monotonically increasing

CHAPTER 3. BAYESIAN MECHANISM DESIGN 56

(incentive constraint); (3.26)

for every i ∈ I : U

i

(θ) ≥ 0

(individual rationality constraint); (3.27)

−

N

i=1

U

i

(θ) +

_

Θ

q(θ)

_

N

i=1

_

θ

i

−

1 −F

i

(θ

i

)

f

i

(θ

i

)

_

−c

_

f(θ)dθ = 0

(budget constraint). (3.28)

We now eliminate the choice variables U

i

(θ) from the problem, and in-

stead write the budget constraint as:

_

Θ

q(θ)

_

N

i=1

_

θ

i

−

1 −F

i

(θ

i

)

f

i

(θ

i

)

_

−c

_

f(θ)dθ ≥ 0 (3.29)

with the understanding that if the left hand side is strictly positive, the

payments T

i

(θ) will be chosen so as to satisfy exact budget balance.

We now solve the mechanism designer’s problem by ﬁrst considering the

relaxed problem where we neglect the monotonicity constraint (3.26). Then

we shall discuss conditions under which the solution to the problem without

constraint (3.26) actually happens to satisfy condition (3.26) as well. Under

those conditions the solution to the relaxed problem has to be a solution to

the original problem.

To solve the relaxed problem we use a version of the Kuhn Tucker The-

orem that applies to inﬁnite dimensional vector spaces, such as function

spaces. Theorems 1 and 2 in Luenberger (1969), p. 217 and p. 221, apply.

According to these results q solves the relaxed maximization problem if and

only if there is a Lagrange multiplier λ ≥ 0 such that q maximizes

_

Θ

q(θ)

_

i∈I

θ

i

−c

_

f(θ)dθ +λ

_

Θ

q(θ)

_

N

i=1

_

θ

i

−

1 −F

i

(θ

i

)

f

i

(θ

i

)

_

−c

_

f(θ)dθ,

(3.30)

and, moreover, λ = 0 only if:

_

Θ

q(θ)

_

N

i=1

_

θ

i

−

1 −F

i

(θ

i

)

f

i

(θ

i

)

_

−c

_

f(θ)dθ > 0. (3.31)

CHAPTER 3. BAYESIAN MECHANISM DESIGN 57

We don’t go through the details of checking the applicability of these results.

However, we note that we obtain a necessary and suﬃcient condition because

in our problem the set of admissible functions q is convex, and because the

objective function that we seek to maximize is linear, hence concave.

We can write the Lagrange function (3.30) as:

_

Θ

q(θ)(1 +λ)

_

i∈I

_

θ

i

−

λ

1 +λ

1 −F

i

(θ

i

)

f

i

(θ

i

)

_

−c

_

f(θ)dθ. (3.32)

It is evident, from point-wise maximization, that the Lagrange function is

maximized if we set q(θ) = 1 whenever the expression in the square brackets

is positive. This leads to the following decision rule:

q

∗

(θ) =

_

¸

_

¸

_

1 if

N

i=1

θ

i

> c +

N

i=1

_

λ

1+λ

1−F

i

(θ

i

)

f

i

(θ

i

)

_

0 otherwise.

(3.33)

Note that we must have λ > 0 if Proposition 3.7 applies because with λ = 0

the rule (3.33) becomes the ﬁrst best rule.

Now we introduce an assumption under which the rule (3.33) satisﬁes

for every λ > 0 the monotonicity constraint (3.26). The condition is the

regularity assumption that we introduced in Assumptions 2.1 and 3.1.

Assumption 3.2. For every i ∈ I the cumulative distribution function F

i

is regular, i.e. the function ψ

i

(θ

i

) = θ

i

−

1−F(θ

i

)

f(θ

i

)

is strictly increasing.

If an agent i’s cumulative distribution function F

i

is regular, then

θ

i

−

λ

1 +λ

1 −F

i

(θ

i

)

f

i

(θ

i

)

(3.34)

is strictly increasing for every λ > 0. This is because the potentially de-

creasing term, that has weight 1 in ψ

i

, has weight λ/(1 + λ) < 1 in (3.34).

This implies that the second best rule in (3.33) satisﬁes the monotonicity

condition (3.26). Therefore, we can conclude:

Proposition 3.8. Suppose Nθ < c < N

¯

θ, and that for every agent i ∈ I

the function F

i

is regular. Then a direct mechanism (q, t

1

, t

2

, . . . , t

N

) is

incentive compatible, individually rational, and ex ante budget balanced, and

maximizes expected welfare among all such mechanisms, if and only if:

CHAPTER 3. BAYESIAN MECHANISM DESIGN 58

(i) there is some λ > 0 such that for all θ ∈ Θ:

q(θ) =

_

¸

_

¸

_

1 if

N

i=1

θ

i

> c +

N

i=1

_

λ

1+λ

1−F

i

(θ

i

)

f

i

(θ

i

)

_

0 otherwise.

(ii)

_

Θ

q(θ)

_

N

i=1

_

θ

i

−

1 −F

i

(θ

i

)

f

i

(θ

i

)

_

−c

_

f(θ)dθ = 0.

(iii) for all i ∈ I:

T

i

(θ

i

) = θ

i

Q

i

(θ

i

) −

_

θ

θ

Q

i

(

ˆ

θ)d

ˆ

θ.

We can see that the second best mechanism undersupplies the public

good. The public good is produced only if the sum of the valuations is

larger than a lower bound that is strictly larger than c. An interesting

question is whether there are simple and appealing indirect mechanisms

that implements the second best.

3.3.5 Proﬁt Maximization

We now consider brieﬂy the problem of choosing the mechanism that max-

imizes the designer’s expected proﬁts among all incentive compatible and

individually rational direct mechanisms. Like the welfare maximizing mech-

anism designer also the proﬁt maximizing mechanism designer has two choice

variables: the allocation rule q, and the transfer payments of the lowest

types, T

i

(θ

i

). We do not assume that the lowest type is necessarily 0. Proﬁt

maximization requires that the transfer payments of the lowest types are

set equal to those types’ expected utility. This leaves q as the only choice

variable. The expected proﬁt from decision rule q can be calculated as pre-

viously in similar contexts as:

_

Θ

q(θ)

_

N

i=1

_

θ

i

−

1 −F

i

(θ

i

)

f

i

(θ

i

)

_

−c

_

f(θ)dθ (3.35)

The mechanism designer has to respect the constraint that for every i ∈ I

the function Q

i

must be monotonically increasing.

CHAPTER 3. BAYESIAN MECHANISM DESIGN 59

The mechanism designer’s problem can be solved using analogous rea-

soning as in the auction model of the previous section. We only state the

result.

Proposition 3.9. Suppose that for every agent i ∈ I the cumulative dis-

tribution function F

i

is regular. Then a direct mechanism (q, t

1

, t

2

, . . . , t

N

)

is incentive compatible, and individually rational, and maximizes expected

proﬁt among all such mechanisms if and only if for all i ∈ I and all θ ∈ Θ:

(i)

q(θ) =

_

¸

_

¸

_

1 if

N

i=1

θ

i

> c +

N

i=1

1−F

i

(θ

i

)

f

i

(θ

i

)

;

0 otherwise.

(ii)

T

i

(θ

i

) = θ

i

Q

i

(θ

i

) −

_

θ

θ

Q

i

(

ˆ

θ)d

ˆ

θ.

Comparing Propositions 3.8 and 3.9 we ﬁnd that the proﬁt maximiz-

ing supplier for the public good supplies a lower quantity than the welfare

maximizing mechanism designer.

3.3.6 A Numerical Example

We provide a very simple numerical example, yet the calculations turn out

to be somewhat involved.

Example 3.3. Suppose N = 2, θ

i

is uniformly distributed on [0, 1] for

i = 1, 2, and 0 < c < 2. Note that the regularity assumption is satisﬁed as

was veriﬁed in the numerical example of the previous section. We wish to

determine ﬁrst the expected welfare maximizing mechanism. By Proposition

3.7 the ﬁrst best cannot be achieved, as θ = 0, and

¯

θ = 1, and hence Nθ <

c < N

¯

θ. The probability with which the ﬁrst best rule calls for the production

of the public good is strictly between zero and one.

By Proposition 3.8 a necessary condition for a direct incentive compati-

ble, individually rational, and ex ante budget balanced mechanism to maxi-

mize expected welfare among all such mechanisms is that there exists some

λ > 0 such that q(θ) = 1 if and only if:

θ

1

+θ

2

> c +

λ

1 +λ

_

1 −θ

1

1

+

1 −θ

2

1

_

⇔

θ

1

+θ

2

>

1 +λ

1 + 2λ

c +

λ

1 + 2λ

2 (3.36)

CHAPTER 3. BAYESIAN MECHANISM DESIGN 60

Denote the right hand side of this inequality by s. Proposition 3.8 thus

means that we can restrict our search for second best mechanisms to those

mechanisms for which q(θ) = 1 if and only if θ

1

+θ

2

≥ s for some s ∈ (c, 2).

We seek to ﬁnd the an appropriate value of s.

We determine s by assuming that the interim expected payments of the

lowest types are zero, as required by Proposition 3.8, that the interim ex-

pected payments of all other types are as required by incentive compatibility,

and that the budget surplus on the right hand side of equation (3.29) is zero,

as required by the Lagrange conditions. To proceed, we calculate the total

expected costs of producing the public good, denoted by C(s), and the total

revenue of a mechanism, denoted by R(s), for incentive compatible, individ-

ually rational threshold mechanisms. We distinguish two cases.

Case 1: Suppose s ≤ 1. Then the expected costs of producing the public

good are:

C(s) =

_

1 −

1

2

s

2

_

c. (3.37)

Next we calculate the expected payment by the agent 1:

_

1

0

_

1

0

q(θ)

_

θ

1

−

1 −F

1

(θ

1

)

f

1

(θ

1

)

_

fθ)dθ

=

_

s

0

_

1

s−θ

1

(2θ

1

−1)dθ

2

dθ

1

+

_

1

s

_

1

0

(2θ

1

−1)dθ

2

dθ

1

= −

1

3

s

+

1

2

s

2

(3.38)

Agent 2’s expected payment will be the same. Therefore, the total expected

revenue of the mechanism is:

R(s) = −

2

3

s

3

+s

2

. (3.39)

Case 2: Suppose s > 1. Then the expected costs of producing the public

good are:

C(s) =

1

2

(2 −s)

2

c. (3.40)

The expected payment by agent 1 is given by:

_

1

0

_

1

0

q(θ)

_

θ

i

−

1 −F

i

(θ

i

)

f

i

(θ

i

)

_

fθ)dθ

CHAPTER 3. BAYESIAN MECHANISM DESIGN 61

=

_

1

s−1

_

1

s−θ

1

(2θ

1

−1)dθ

2

dθ

1

=

1

6

−

1

2

(s −1)

2

+

1

3

(s −1)

3

. (3.41)

Agent 2’s expected payments will be the same, and therefore the total ex-

pected revenue from the mechanism is:

R(s) =

1

3

−(s −1)

2

+

2

3

(s −1)

3

. (3.42)

We now deﬁne D(s) ≡ R(s) −C(s), so that the condition for s becomes

D(s) = 0. To understand the set of solutions to this equation we ﬁrst inves-

tigate the sign of the derivative of D with respect to s. Observe that:

0 < s < 1 ⇒D

(s) = −2s

2

+ (2 +c)s = s(2(1 −s) +c) > 0; (3.43)

1 < s < 2 ⇒D

(s) = −2(s −1) + 2(s −1)

2

+ (2 −s)c

= (2 −s)(c −2(s −1)). (3.44)

The term in the last line is positive if and only if:

c −2(s −1) > 0 ⇔

s < 1 +

c

2

(3.45)

Next, we investigate the sign of D(s) for some important values of s.

Note ﬁrst that:

D(0) = −c < 0 and (3.46)

D(2) = 0. (3.47)

Next, we investigate the sign of D at s = 1 +

c

2

. We obtain:

D

_

1 +

c

2

_

=

1

3

−

_

c

2

_

2

+

2

3

_

c

2

_

3

−

1

2

_

1 −

c

2

_

2

c

=

1

3

−

1

2

c +

_

c

2

_

2

−

1

3

_

c

2

_

3

(3.48)

We want to prove that the expression in the last line is strictly positive. It

clearly tends to a positive limit as c →0. If c →2, then the expression tends

CHAPTER 3. BAYESIAN MECHANISM DESIGN 62

to zero. Thus, it is suﬃcient to show that the derivative of this expression

is negative for 0 < c < 2. This derivative is:

−

1

2

+

c

2

−

1

2

_

c

2

_

2

=

c

2

_

1 −

c

2

_

−

1

2

≤

1

4

−

1

2

= −

1

4

(3.49)

We can conclude that D(1 +

1

2

c) > 0.

From our results so far we can conclude that the equation D(s) = 0 has

exactly two solutions: one solution in the interval (0, 1+

c

2

), and the solution

s = 2. We can discard the latter as we require s ∈ (c, 2).

For the purpose of calculating the solution, it is useful to ask whether

the solution of D(s) = 0 is larger or less than 1. For this we investigate the

value of D(1):

D(1) =

1

3

−

1

2

c > 0 ⇔

c <

2

3

(3.50)

Thus, the solution of D(s) = 0 will be between 0 and 1 if and only if c <

2

3

.

Otherwise, it will be between 1 and 1 +

c

2

.

If c <

2

3

, then we ﬁnd s by solving the following equation:

−

2

3

s

3

+s

2

−

_

1 −

1

2

s

2

_

c = 0. (3.51)

Unfortunately, this equation has no simple solution. If c >

2

3

, then we ﬁnd

s by solving the following equation:

1

3

−(s −1)

2

+

2

3

(s −1)

3

−

1

2

(2 −a)

2

c = 0. (3.52)

It so happens that this equation has a simple analytical solution:

s =

1

2

+

3

4

c. (3.53)

We can now sum up:

CHAPTER 3. BAYESIAN MECHANISM DESIGN 63

Proposition 3.10. The utilitarian mechanism designer will choose a mech-

anism with an allocation rule q such that:

q

∗

(θ) =

_

_

_

1 if θ

1

+θ

2

> s;

0 otherwise.

where s is determined as follows:

(i) If c <

2

3

, then s is the unique solution in [0, 1] of the equation:

−

2

3

s

3

+s

2

−

_

1 −

1

2

s

2

_

c = 0.

(ii) If c ≥

2

3

, then

s =

1

2

+

3

4

c.

The expected proﬁt maximizing mechanism follows straightforwardly from

Proposition 3.9:

Proposition 3.11. The expected proﬁt maximizing mechanism designer will

choose a mechanism with an allocation rule q such that:

q

∗

(θ) =

_

_

_

1 if θ

1

+θ

2

> s;

0 otherwise.

where s = 1 +

1

2

c.

To illustrate our ﬁndings, we have plotted in Figure 3.2 the optimal

threshold s for the utilitarian case (dashed and dotted line) and for the proﬁt

maximization case (dashed line) as a function of c. For comparison we have

also plotted the 45

◦

line which is the ”ﬁrst best” threshold (unbroken line).

Figure 3.2 shows that the second best threshold chosen by a welfare max-

imizing mechanism designer is strictly larger than the ﬁrst best threshold,

except if c is either 0 or 2. This reﬂects the fact that to provide incentives

the mechanism designer must accept some ineﬃciencies. The ineﬃcien-

cies become small as c approaches either 0 or 2. These two extreme cases

correspond to cases in which the mechanism designer doesn’t really have to

induce truthful revelation of individuals’ types, either because the production

of the public good is free, and hence it should always be produced, or it is

prohibitively expensive, and hence it should never be produced.

CHAPTER 3. BAYESIAN MECHANISM DESIGN 64

0.0 0.5 1.0 1.5 2.0

0

.

0

0

.

5

1

.

0

1

.

5

2

.

0

c

s

Figure 3.2: Thresholds in First Best, Second Best, and Proﬁt Maximum

Figure 3.2 also shows that a monopoly supplier of the public good would

choose a threshold that is even larger than the threshold chosen by the

utilitarian mechanism designer. This corresponds to the standard textbook

insight that a monopolist artiﬁcially restricts supply to raise price. The

monopoly distortion exists even if production is free (c = 0), as in the stan-

dard textbook case. If production becomes prohibitively expensive (c = 2), the

monopolist will not want to provide the good. Thus, the monopoly solution

coincides with the ﬁrst best, and the utilitarian solution. In this case, the

distortion introduced by a monopolist is zero.

3.4 Bilateral Trade

3.4.1 Setup

In our next example a seller of a single indivisible good faces one buyer with

unknown valuation, as in Section 2.2, but now we consider the situation

from the perspective of a mechanism designer who wants to arrange a trad-

ing institution for the two parties that guarantees that they trade if and

only if the buyer’s value is larger than the seller’s. Moreover, not only the

CHAPTER 3. BAYESIAN MECHANISM DESIGN 65

buyer’s valuation, but also the seller’s valuation is unknown to the designer

of this trading institution, and it may be that valuations are such that trade

is not eﬃcient. This example was ﬁrst analyzed in the very well-known pa-

per by Myerson and Satterthwaite (1983). It is the simplest example that

one might analyze when seeking to build a general theory of the design of

optimal trading institution, such as stock exchanges or commodity markets.

Methodologically, it is interesting to see how one can treat the bilateral trade

problem with similar methods as the public goods problem, although it is

seemingly quite diﬀerent.

A seller S owns a single indivisible good. There is one potential buyer

B. Deﬁne I = ¦S, B¦. The seller’s utility if he sells the good and receives

a transfer payment t is equal to t. If he does not sell the good and receives

a transfer t then his utility is θ

S

+ t where θ

S

is a random variable with

cumulative distribution function F

S

and density f

S

. We assume that F

S

has support [θ

S

,

¯

θ

S

] and that f

S

(θ

S

) > 0 for all θ

S

∈ [θ

S

,

¯

θ

S

]. The buyer’s

utility if he purchases the good and pays a transfer t equals θ

B

−t, where θ

B

is a random variable with cumulative distribution function F

B

and density

f

B

. We assume that F

B

has support [θ

B

,

¯

θ

B

] and that f

B

(θ

B

) > 0 for all

θ

B

∈ [θ

B

,

¯

θ

B

]. The buyer’s utility if he does not obtain the good and pays

transfer t is −t. The random variables θ

S

and θ

B

are independent. We deﬁne

θ = (θ

S

, θ

B

) and Θ = [θ

S

,

¯

θ

S

] [θ

B

,

¯

θ

B

]. We denote the joint distribution

of θ by F with density f. The seller only observes θ

S

, and the buyer only

observes θ

B

.

3.4.2 Direct Mechanisms

The revelation principle implies that we can restrict attention to direct mech-

anisms.

Deﬁnition 3.9. A “direct mechanism” consists of functions q, t

S

and t

B

where:

q : Θ →¦0, 1¦

and for i = S, B:

t

i

: Θ →R.

The function q assigns to each type vector θ an indicator variable that

indicates whether trade takes place (q(θ) = 1) or whether no trade takes

place (q(θ) = 0). We shall refer to q also as the “trading rule.” For simplicity,

we restrict attention to deterministic trading rules. The function t

S

indicates

CHAPTER 3. BAYESIAN MECHANISM DESIGN 66

transfers that the seller receives, and the function t

B

indicates transfers

that the buyer makes. We shall mostly assume in this section

10

that t

S

=

t

B

, i.e. that the seller receives what the buyer pays. Therefore, it seems

redundant to introduce separate notation for the buyer’s payment and the

seller’s receipts. However, it is much more convenient to begin with a more

general framework in which these variables are not identical, and then to

introduce an ex post budget balance condition which forces them to be equal

to each other, as we shall do below. This allows us then to adopt a similar

methodology as in the previous section.

Given a direct mechanism, we deﬁne for each agent i ∈ ¦S, B¦ functions

Q

i

: [θ

i

,

¯

θ

i

] → [0, 1] and T

i

: [θ

i

,

¯

θ

i

] → R where Q

i

(θ

i

) is the conditional

probability that trade takes place, where we condition on agent i’s type being

θ

i

, and T

i

(θ

i

) is the conditional expected value of the transfer that agent i

receives (if i = S) or makes (if i = B), again conditioning on agent i’s type

being θ

i

. Finally, we also deﬁne agent i’s expected utility U

i

(θ

i

) conditional

on her type being θ

i

. This is given by: U

S

(θ

S

) = T

S

(θ

S

) + (1 − Q

S

(θ

S

))θ

S

and U

B

(θ

B

) = Q

B

(θ

B

)θ

B

−T

B

(θ

B

).

We shall restrict attention to direct mechanisms that are incentive com-

patible, individually rational, and ex post budget balanced. Individual ra-

tionality is deﬁned as before. Standard arguments show that a mechanism

is incentive compatible for the buyer under exactly the same conditions as

before. For the seller, the standard arguments apply if types are ordered in

the reverse of the numerical order, that is starting with high types rather

than low types. Thus, a necessary and suﬃcient condition for incentive

compatibility for the seller is that Q

S

is decreasing, and that T

S

is given by:

T

S

(θ

S

) = T

S

(

¯

θ

S

) + (1 −Q

S

(

¯

θ

S

))

¯

θ

S

−(1 −Q

S

(θ

S

))θ

S

−

_

¯

θ

S

θ

S

(1 −Q

S

(x)) dx

(3.54)

for all θ

S

.

Individual rationality for the buyer is deﬁned and characterized in the

same way as before. For the seller, individual rationality means that U

S

(θ

S

)

≥ θ

S

for all θ

S

, that is, the seller trades voluntarily and obtains an expected

utility that is at least as large as his utility would be if he kept the good. If

a mechanism is incentive compatible, then the seller’s individual rationality

condition holds if and only if it holds for the highest seller type

¯

θ

S

. To prove

10

An exception is Subsection 3.4.4 where we consider proﬁt maximizing trading mech-

anisms.

CHAPTER 3. BAYESIAN MECHANISM DESIGN 67

this we show that the diﬀerence U

S

(θ

S

) − θ

S

is decreasing in θ

S

. A seller

who is of type θ

S

and pretends to be of type θ

S

obtains expected utility:

θ

S

(1 −Q

S

(θ

S

)) +T

S

(θ

S

). (3.55)

This exceeds θ

S

by:

T

S

(θ

S

) −θ

S

Q

S

(θ

S

). (3.56)

The seller maximizes over θ

S

. As θ

S

decreases, the function that the seller

maximizes shifts downwards, and therefore the function’s maximum value

decreases.

Ex post budget balance requires that in each state θ we have: t

S

(θ) =

t

B

(θ). By the same argument as in the previous section it suﬃces to iden-

tify ex ante budget balanced mechanisms, that is, mechanisms for which the

seller’s ex ante expected payment is equal to the buyer’s ex ante expected

payment. We shall work with this condition. The by now familiar calcu-

lation shows that for an incentive compatible direct mechanism the seller’s

expected payment is:

U

S

(

¯

θ

S

) −

_

Θ

(1 −q(θ))

_

θ

S

+

F

S

(θ

S

)

f

S

(θ

S

)

_

f(θ)dθ (3.57)

and the buyer’s expected payment is:

−U

B

(θ

B

) +

_

Θ

q(θ)

_

θ

B

−

1 −F

B

(θ

B

)

f

B

(θ

B

)

_

f(θ)dθ. (3.58)

3.4.3 Welfare Maximization

The mechanism designer seeks to maximize the sum of the individuals’ util-

ities,

q(θ)θ

B

−t

B

+ (1 −q(θ))θ

S

+t

S

=θ

S

+q(θ) (θ

B

−θ

S

) +t

S

−t

B

. (3.59)

If the mechanism designer were not constrained by incentive compatibility

and individual rationality, but only had to respect ex post budget balance,

then the mechanism designer would choose a “ﬁrst best mechanism” where

the trading rule is:

q

∗

(θ) =

_

_

_

1 if θ

B

≥ θ

S

;

0 otherwise.

(3.60)

CHAPTER 3. BAYESIAN MECHANISM DESIGN 68

where the decision in the case of equality of values is arbitrary. The payment

rule is arbitrary as long as it is ex post budget balanced.

Our ﬁrst objective is to prove that in almost all cases no ﬁrst best mech-

anism is incentive compatible and individually rational.

Proposition 3.12 (Myerson and Satterthwaite, 1983). An incentive com-

patible, individually rational and ex-post budget balanced direct mechanism

with decision rule q

∗

exists if and only if θ

B

≥

¯

θ

S

or θ

S

≥

¯

θ

B

.

The condition θ

B

≥

¯

θ

S

implies that trade is always at least weakly

eﬃcient. The condition θ

S

≥

¯

θ

B

implies that eﬃciency never requires trade.

Thus, these are trivial cases. In all non-trivial cases, there is no incentive

compatible, individually rational and ex post budget balanced ﬁrst best

mechanism.

Proof. The “if-part” of Proposition 3.12 is trivial. If θ

S

≥

¯

θ

B

, then a mecha-

nism under which no trade takes place and no payments are made is ﬁrst best

and has the required properties. If θ

S

≥

¯

θ

B

a mechanism where trade always

takes place, and the buyer always pays the seller some price p ∈ [

¯

θ

B

, θ

S

] is

ﬁrst best and has the required properties.

To prove the “only if-part” we proceed in a similar way as in the proof

of Proposition 3.7. We ﬁrst display a mechanism that is incentive compat-

ible and individually rational and that implements the ﬁrst best trading

rule. Then we argue that this mechanism maximizes expected surplus of

buyer’s payment over seller’s receipts among all incentive compatible and

individually rational mechanisms that implement the ﬁrst best trading rule.

Finally, we show that the mechanism has an expected deﬁcit if the condition

of Proposition 3.7 holds. The assertion follows.

Deﬁnition 3.10. The “pivot mechanism” is the mechanism that is given

by the ﬁrst best trading rule and by the following transfer schemes:

t

S

(θ) = q

∗

(

¯

θ

S

, θ

B

)

¯

θ

S

+

_

q

∗

(θ) −q

∗

(

¯

θ

S

, θ

B

)

_

θ

B

t

B

(θ) = q

∗

(θ

S

, θ

B

)θ

B

+ (q

∗

(θ) −q

∗

(θ

S

, θ

B

)) θ

S

for all θ ∈ Θ.

In this mechanism the seller receives a constant plus the buyer’s valuation

of the good if the seller’s report was pivotal for the trade happening. The

buyer pays a constant plus the seller’s valuation of the good if the buyer’s

report was pivotal for the trade happening. As in the previous section,

CHAPTER 3. BAYESIAN MECHANISM DESIGN 69

the pivot mechanism has in common with a general Vickrey Clarke Groves

mechanism that each agent’s private interests are exactly aligned with social

welfare. Moreover, the individual rationality constraints of the two extreme

types, the lowest type of the seller, and the highest type of the buyer, are

satisﬁed with equality. This becomes clear in the proof of the following

result.

Lemma 3.9. The pivot mechanism is incentive compatible and individually

rational.

Proof. Suppose the seller is of type θ

S

and contemplates reporting that she

is of type θ

S

,= θ

S

. We ﬁx we the buyer’s type θ

B

. We are going to show

that truthful reporting is optimal whatever the buyer’s type. This obviously

implies that truthful report is optimal for the seller.

The seller’s utility if she is of type θ

S

but reports that she is of type θ

S

,

leaving out all terms that do not depend on her report, is:

q

∗

(θ

S

, θ

B

)(θ

B

−θ

S

). (3.61)

Note that the ﬁrst best decision rule q

∗

maximizes the social surplus q

∗

(θ)(θ

B

−

θ

S

).

11

The seller’s utility is thus equal to social surplus if the collective de-

cision is q

∗

(θ

S

, θ

B

). Because the ﬁrst best rule maximizes social surplus, it

is optimal for the seller to report θ

S

truthfully. A similar argument applies

to the buyer.

To check individual rationality it is suﬃcient to consider the types

¯

θ

B

and θ

S

. From the payment rule, it is clear that the type

¯

θ

B

receives expected

utility

¯

θ

B

whatever the seller’s type, and the type θ

S

receives expected utility

zero whatever the buyer’s type. Individual rationality follows.

Lemma 3.10. The diﬀerence between the buyer’s and the seller’s expected

payments is at least as large under the pivot mechanism as it is under any

incentive compatible and individually rational direct mechanism that imple-

ments the ﬁrst best trading rule.

Proof. Our discussion of interim expected transfers for incentive compatible

mechanisms shows that these depend only on the trading rule and on the

payments of the highest buyer type and the lowest seller type. We have

seen in the proof of Lemma 3.9 that the expected utility of the highest seller

11

We use the expression “social surplus” to distinguish this expression from social welfare

as deﬁned in equation (3.59). The ﬁrst best decision rule maximizes both, social welfare

and social surplus.

CHAPTER 3. BAYESIAN MECHANISM DESIGN 70

type in the pivot mechanism is

¯

θ

S

. This implies that the interim expected

payments to the seller in the pivot mechanism are the lowest payments

compatible with the ﬁrst best trading rule. Similarly, the expected utility

of the lowest buyer type in the pivot mechanism is zero, and therefore the

interim expected payments that the seller makes in the pivot mechanism are

the highest payments compatible with the ﬁrst best trading rule. Thus, the

pivot mechanism maximizes the ex ante expected diﬀerence between buyer’s

and seller’s payments among all mechanisms that implement the ﬁrst best

trading rule and that are incentive compatible.

We show next that the pivot mechanism has an ex ante expected deﬁcit

if the condition of Proposition 3.12 holds.

Lemma 3.11. The diﬀerence between the buyer’s ex ante expected payment

and the seller’s ex ante expected payment is negative in the pivot mechanism

if the condition of Proposition 3.12 is violated, i.e. if θ

B

<

¯

θ

S

and

¯

θ

B

> θ

S

.

Proof. We shall show that for every realized type vector θ the pivot mecha-

nism either has a surplus of zero, or a deﬁcit. Moreover, we shall argue that

the types for which it has a deﬁcit have positive probability. This will then

imply the case.

Consider ﬁrst values of θ such that q

∗

(θ) = 0. Then both agents’ transfers

equal zero, and hence the deﬁcit is zero. Now consider values of θ such that

q

∗

(θ) = 1. We calculate the ex post deﬁcit for four diﬀerent scenarios.

Case 1: q

∗

(θ) = 1, q

∗

(

¯

θ

S

, θ

B

) = 1, q

∗

(θ

S

, θ

B

) = 1. Then the ex post

deﬁcit is: θ

B

−

¯

θ

S

. This is negative by the assumption of the Lemma.

Case 2: q

∗

(θ) = 1, q

∗

(

¯

θ

S

, θ

B

) = 1, q

∗

(θ

S

, θ

B

) = 0. Then the ex post

deﬁcit is: θ

S

−

¯

θ

S

. This is obviously not positive.

Case 3: q

∗

(θ) = 1, q

∗

(

¯

θ

S

, θ

B

) = 0, q

∗

(θ

S

, θ

B

) = 1. Then the ex post

deﬁcit is: θ

B

−θ

B

. This is obviously not positive.

Case 4: q

∗

(θ) = 1, q

∗

(

¯

θ

S

, θ

B

) = 0, q

∗

(θ

S

, θ

B

) = 0. Then the ex post

deﬁcit is: θ

S

− θ

B

. This is not positive because we are considering θ such

that trade is eﬃcient.

The condition of the Lemma means that the intervals [θ

B

,

¯

θ

B

] and [θ

S

,

¯

θ

S

]

have an intersection with non-empty interior. Consider any point in this

intersection for which trade is eﬃcient and such that θ

S

> θ

B

. By drawing

a simple diagram one can see that the probability measure of such points is

positive. Moreover, at any such point we are in Case 4, and moreover the

deﬁcit is strictly negative. Therefore, there is an expected deﬁcit.

CHAPTER 3. BAYESIAN MECHANISM DESIGN 71

This completes the proof of Proposition 3.12.

We shall now focus on the case in which according to Proposition 3.12

no ﬁrst best mechanism is incentive compatible, individually rational, and

ex post budget balanced. Our objective is to determine direct mechanisms

that maximize expected welfare among all incentive compatible, individually

rational, and ex post budget balanced mechanisms. We shall refer to these

mechanisms as “second best.”

As in the public goods case it is without loss of generality to assume

that the mechanism designer balances the ex post budget exactly rather

than leaving a surplus. The mechanism designer’s welfare function can then

be written as θ

S

+ q(θ)(θ

B

− θ

S

). Moreover, note that θ

S

is not aﬀected

by the mechanism designer’s decisions. Therefore, we shall assume that the

mechanism designer seeks to maximize:

_

Θ

q(θ) (θ

B

−θ

S

) f(θ)dθ. (3.62)

As in the previous subsection it is convenient to think of the designer’s

choice variables as the trading rule q and the interim expected utilities of

the highest type of the seller, U

B

(

¯

θ

B

), and of the lowest type of the buyer,

U

S

(θ

S

). The constraints are:

Q

S

is increasing and Q

B

is decreasing

(incentive compatibility); (3.63)

U

S

(

¯

θ

S

) ≥

¯

θ

S

and U

B

(θ

B

) ≥ 0

(individual rationality); (3.64)

−U

B

(θ

B

) +

_

Θ

q(θ)

_

θ

B

−

1 −F

B

(θ

B

)

f

B

(θ

B

)

_

f(θ)dθ =

U

S

(

¯

θ

S

) −

_

Θ

(1 −q(θ))

_

θ

S

+

F

S

(θ

S

)

f

S

(θ

S

)

_

f(θ)dθ

(budget balance). (3.65)

We simplify this optimization problem a little. First, we write the budget

constraint as:

_

Θ

q(θ)

__

θ

B

−

1 −F

B

(θ

B

)

f

B

(θ

B

)

_

−

_

θ

S

+

F

S

(θ

S

)

f

S

(θ

S

)

__

f(θ)dθ =

U

S

(

¯

θ

S

) +U

B

(θ

B

) −

_

Θ

_

θ

S

+

F

S

(θ

S

)

f

S

(θ

S

)

_

f(θ)dθ (3.66)

CHAPTER 3. BAYESIAN MECHANISM DESIGN 72

Next, we drop the choice variables U

S

(

¯

θ

S

) and U

B

(θ

B

) and re-write the

budget constraint as:

_

Θ

q(θ)

__

θ

B

−

1 −F

B

(θ

B

)

f

B

(θ

B

)

_

−

_

θ

S

+

F

S

(θ

S

)

f

S

(θ

S

)

__

f(θ)dθ ≥

¯

θ

S

−

_

Θ

_

θ

S

+

F

S

(θ

S

)

f

S

(θ

S

)

_

f(θ)dθ (3.67)

If this inequality is strict, then the mechanism designer can allocate the

diﬀerence as additional utility to the buyer and the seller. Below, we shall

denote the right hand side of (3.67) by K.

When solving the mechanism designer’s optimization problem we ﬁrst

neglect the incentive constraint. We have a concave maximization prob-

lem because the objective function and the constraint are both linear in q.

Therefore, by Theorems 1 and 2 in Luenberger (1969), p. 217 and p. 221,

a necessary and suﬃcient condition for q to be optimal is that there is a

Lagrange multiplier λ ≥ 0 such that q maximizes:

_

Θ

q(θ)

_

(1 +λ)θ

B

−λ

1 −F

B

(θ

B

)

f

B

(θ

B

)

−(1 +λ)θ

S

−λ

F

S

(θ

S

)

f

S

(θ

S

)

_

f(θ)dθ −λK

(3.68)

and moreover the budget constraint (3.67) holds as a strict inequality only

if λ = 0.

The Lagrange function is maximized by choosing q(θ) to be positive

whenever the expression in the square brackets is positive. Brief manipula-

tion of this condition leads to the following trading rule:

q(θ) =

_

¸

_

¸

_

1 if θ

B

−

λ

1+λ

1−F

B

(θ

B

)

f

B

(θ

B

)

≥ θ

S

+

λ

1+λ

F

S

(θ

S

)

f

S

(θ

S

)

0 otherwise.

(3.69)

Can we have λ = 0? Then the optimal decision rule (3.69) would be identical

to the ﬁrst best trading rule, and the budget constraint would be violated

by Proposition 3.12. Therefore, we must have: λ > 0.

We now make a regularity assumption that ensures that this decision

rule satisﬁes the monotonicity constraint (3.63) of the mechanism designer’s

maximization problem.

Assumption 3.3. The seller’s distribution function F

S

is “regular,” i.e.

ψ

S

(θ

S

) ≡ θ

S

+

F

S

(θ

S

)

f

S

(θ

S

)

CHAPTER 3. BAYESIAN MECHANISM DESIGN 73

is monotonically increasing. The buyer’s distribution function F

B

is “regu-

lar,” i.e.

ψ

B

(θ

B

) ≡ θ

B

−

1 −F

B

(θ

B

)

f

B

(θ

B

)

is monotonically increasing.

If F

S

and F

B

are regular,

θ

S

+

λ

1 +λ

F

S

(θ

S

)

f

S

(θ

S

)

and θ

B

−

λ

1 +λ

1 −F

B

(θ

B

)

f

B

(θ

B

)

(3.70)

are increasing for every λ > 0. In this case the trading rule q derived

above will imply that the function Q

S

is decreasing and the function Q

B

is

increasing. We can then conclude:

Proposition 3.13. Suppose θ

B

<

¯

θ

S

and

¯

θ

B

> θ

S

, and that for every

agent i ∈ I the function F

i

is regular. Then a direct mechanism (q, t

S

, t

B

)

is incentive compatible, individually rational, and ex ante budget balanced,

and maximizes expected welfare among all such mechanisms if and only if:

(i) there is some λ > 0 such that for all θ ∈ Θ:

q(θ) =

_

¸

_

¸

_

1 if θ

B

−

λ

1+λ

1−F

B

(θ

B

)

f

B

(θ

B

)

≥ θ

S

+

λ

1+λ

F

S

(θ

S

)

f

S

(θ

S

)

0 otherwise.

(3.71)

(ii) exact budget balance holds:

_

Θ

q(θ)

__

θ

B

−

1 −F

B

(θ

B

)

f

B

(θ

B

)

_

−

_

θ

S

+

F

S

(θ

S

)

f

S

(θ

S

)

__

f(θ)dθ =

¯

θ

S

−

_

Θ

_

θ

S

+

F

S

(θ

S

)

f

S

(θ

S

)

_

f(θ)dθ (3.72)

(iii) the payment rules create incentive compatibility, i.e. for all θ

B

∈

[θ

B

,

¯

θ

B

] and all θ

S

∈ [θ

S

,

¯

θ

S

]:

T

B

(θ

B

) = θ

S

Q

S

(θ

S

) −

_

θ

S

θ

S

Q

S

(x)dx

T

S

(θ

S

) =

¯

θ

S

−(1 −Q

S

(θ

S

))θ

S

−

_

¯

θ

S

θ

S

(1 −Q

S

(x)) dx

Note that in the second best mechanism the good changes hands less

frequently than in the ﬁrst best. The buyer’s value, minus some discount,

still has to be larger than the seller’s value, plus some discount, for trade to

take place.

CHAPTER 3. BAYESIAN MECHANISM DESIGN 74

3.4.4 Proﬁt Maximization

We now consider brieﬂy the problem of choosing the mechanism that maxi-

mizes the expected proﬁts. We assume that the mechanism designer’s proﬁt

is the diﬀerence between what the buyer pays and what the seller receives.

We could think of this mechanism designer as the commercial designer of a

trading platform who charges a fee for transactions on the platform.

The mechanism designer’s choice variables are again the expected trans-

fers of the highest seller type and the lowest buyer type and the trading rule.

It is obvious that proﬁt maximization implies that the expected transfer of

the highest seller type and the lowest buyer type are chosen so as to make

these types’ individual rationality constraints true as equalities. Using the

formulas (3.57) and (3.58) we can thus write expected revenue as:

_

Θ

q(θ)

_

θ

B

−

1 −F

B

(θ

B

)

f

B

(θ

B

)

_

f(θ)dθ

−

_

¯

θ

S

−

_

Θ

(1 −q(θ))

_

θ

S

+

F

S

(θ

S

)

f

S

(θ

S

)

_

f(θ)dθ

_

. (3.73)

Taking out constants that are independent of q, we are left with:

_

Θ

q(θ)

_

θ

B

−

1 −F

B

(θ

B

)

f

B

(θ

B

)

−θ

S

−

F

S

(θ

S

)

f

S

(θ

S

)

_

f(θ)dθ. (3.74)

Using similar reasoning as in the last subsection, we obtain:

Proposition 3.14. Suppose for every agent i ∈ I the distribution function

F

i

is regular. Then necessary and suﬃcient conditions for an incentive

compatible, and individually rational direct mechanism to maximize expected

proﬁts are:

(i)

q(θ) =

_

¸

_

¸

_

1 if θ

B

−

1−F

B

(θ

B

)

f

B

(θ

B

)

> θ

S

+

F

S

(θ

S

)

f

S

(θ

S

)

;

0 otherwise.

(ii) for all θ

B

∈ [θ

B

,

¯

θ

B

] and all θ

S

∈ [θ

S

,

¯

θ

S

]:

T

B

(θ

B

) = θ

S

Q

S

(θ

S

) −

_

θ

S

θ

S

Q

S

(x)dx

T

S

(θ

S

) =

¯

θ

S

−(1 −Q

S

(θ

S

))θ

S

−

_

¯

θ

S

θ

S

(1 −Q

S

(x)) dx

CHAPTER 3. BAYESIAN MECHANISM DESIGN 75

Comparing Propositions 3.13 and 3.14 we ﬁnd that the proﬁt maximiz-

ing mechanism designer facilitates less trade than the welfare maximizing

mechanism designer.

3.4.5 A Numerical Example

We conclude with a very simple numerical example.

Example 3.4. Suppose that θ

S

as well as θ

B

are uniformly distributed on

the interval [0, 1]. Note that this satisﬁes the regularity condition. We want

to determine the welfare and the proﬁt maximizing mechanisms.

By Proposition 3.13 the welfare maximizing rule will be such that trade

takes place if and only if:

θ

B

−

λ

1 +λ

1 −F

B

(θ

B

)

f

B

(θ

B

)

> θ

S

+

λ

1 +λ

F

S

(θ

S

)

f

S

(θ

S

)

⇔ (3.75)

θ

B

−

λ

1 +λ

(1 −θ

B

) > θ

S

+

λ

1 +λ

θ

S

⇔ (3.76)

θ

B

−θ

S

>

λ

1 + 2λ

. (3.77)

Thus, trade will take place if and only if the diﬀerence between the buyer’s

and the seller’s valuation is above some positive threshold s. We shall now

calculate for which such thresholds s the budget balance constraint holds.

Take s as given. The buyer’s willingness to pay is given by:

_

Θ

q(θ)

_

θ

B

−

1 −F

B

(θ

B

)

f

B

(θ

B

)

_

f(θ)d(θ)

=

_

Θ

q(θ)(2θ −1)dθ

=

_

1

s

_

θ

B

−s

0

2θ

B

−1dθ

B

=

1

3

s

3

−

1

2

s

2

+

1

6

. (3.78)

The seller’s expected transfer is given by:

¯

θ

S

−

_

Θ

(1 −q(θ))

_

θ

S

+

F

S

(θ

S

)

f

S

(θ

S

)

_

f(θ)dθ

= 1 −2

_

Θ

(1 −q(θ))θ

S

dθ

CHAPTER 3. BAYESIAN MECHANISM DESIGN 76

= 1 −2

__

1−s

0

_

θ

S

+s

0

θ

s

dθ

B

dθ

S

+

_

1

1−s

_

1

0

θ

s

dθ

B

dθ

S

_

=

1

3

(1 −s)

3

(3.79)

Budget balance is achieved when:

1

3

s

3

−

1

2

s

2

+

1

6

=

1

3

(1 −s)

3

. (3.80)

This has two solutions:

s =

1

4

and s = 1. (3.81)

Only the solution s =

1

4

is of the required form

λ

1+2λ

with λ > 0. We

conclude:

Proposition 3.15. In the welfare maximizing incentive compatible, indi-

vidually rational and ex post budget balanced trading mechanism trade takes

place if and only if

θ

B

−θ

S

>

1

4

. (3.82)

From Proposition 3.14 the following characterization of the expected proﬁt

maximizing trading mechanism is obvious.

Proposition 3.16. In the expected proﬁt maximizing incentive compatible

and individually rational trading mechanism trade takes place if and only if

θ

B

−θ

S

>

1

2

. (3.83)

We can see that in both mechanisms trade takes place less frequently

than in the ﬁrst best, but an expected proﬁt maximizing mechanism designer

arranges for less trade than an expected welfare maximizing mechanism de-

signer.

3.5 Comments on the Literature

The classic paper on expected revenue maximizing single unit auctions is

Myerson (1981). In preparing these notes I have also used Krishna (2002),

in particular: Chapter 5, and Milgrom (2004), in particular: Chapter 3. For

the public goods problem a classic reference is d’Aspremont and Gerard-

Varet (1979). This paper works with a more general set-up than my expo-

sition, but it focuses only on Bayesian incentive compatibility and ex-post

CHAPTER 3. BAYESIAN MECHANISM DESIGN 77

budget balance, neglecting individual rationality. Welfare maximization and

proﬁt maximization under individual rationality constraints are considered

in G¨ uth and Hellwig (1986). The explanation of the relation between ex ante

and ex post budget balance is based on B¨orgers and Norman (2008). For

the bilateral trade problem the classic paper is Myerson and Satterthwaite

(1983). My exposition of the bilateral trade problem has also beneﬁtted

from the books by Krishna and Milgrom cited above.

3.6 Problems

a) Propositions 3.7 and 3.12 show that there are no direct mechanisms

that implement the ﬁrst best decision rule (in the example of the pub-

lic good problem) or trade rule (in the example of the bilateral trade

problem) q

∗

, and that have moreover three properties: (1) incentive

compatibility, (2) individual rationality, (3) ex post budget balance.

Prove that in both applications there are direct mechanisms that im-

plement q

∗

and that have any two of these three properties.

b) Propositions 3.10 and 3.15 display second best mechanisms for numeri-

cal examples that illustrate the public goods problem and the bilateral

trade problem. However, in each we have not speciﬁed the payment

rules t

i

for i ∈ I. Determine for each case payment rules that guaran-

tee incentive compatibility, individual rationality, and ex post budget

balance of the second best mechanism.

c) For each of the numerical examples of this Chapter investigate whether

there are intuitive and simple indirect implementations of the expected

proﬁt maximizing, or expected welfare maximizing, mechanisms.

Chapter 4

Dominant Strategy

Mechanisms: Examples

4.1 Introduction

The models discussed in Chapter 3 were all independent private value mod-

els. In Section 3.2.1 we have discussed the limitations of such models.

Among the points that we mentioned was that the assumption of a common

prior and independent types means that each agent’s beliefs about the other

agents’ types is independent of the agent’s type, and moreover that these

beliefs are commonly known among the mechanism designer and the agents.

This assumption seems peculiar in the context of mechanism design, where

the focus is on asymmetric information. In the independent private values

model agents and the mechanism designer don’t know other agents’ values,

but they do know those agent’s beliefs. Imperfect information about others’

beliefs seems at least as pervasive as imperfect information about others’

preferences.

The mechanisms that we constructed in the previous chapter work well if

the mechanism designer is right in his assumptions about the agents, and if

agents indeed play a Bayesian equilibrium of the game that the mechanism

designer creates for them. However, if the mechanism designer has made

incorrect assumptions about the agents’ beliefs, or if these beliefs are not

common knowledge, then the mechanism designer’s expectations about how

the mechanism will be played might not be fulﬁlled.

In this chapter we revisit the examples of the previous section, but we

make much weaker assumptions about the agents’ and the mechanism de-

78

CHAPTER 4. DOMINANT STRATEGY MECHANISMS 79

signer’s beliefs. We assume that the mechanism designer does not want to

rely on any assumption regarding the agents’ beliefs about each other. The

mechanism designer in this chapter wishes to implement a mechanism of

which he is sure that it produces the desired results independent of what

agents think about each other. In a sense, we shall move from one extreme

along a continuum of possible models to the other extreme, skipping what-

ever might be in the middle. While in the previous section the mechanism

designer was willing to trust a single hypothesis about agents’ beliefs, in this

section the mechanism designer is assumed to be uncertain about agents’

beliefs, and, moreover, it is assumed that the mechanism designer is not

willing to risk at all making incorrect assumptions about agents’ beliefs.

We shall translate this into formal theory by requiring that the strategy

combination that the mechanism designer proposes to the agents when pre-

senting the mechanism to them prescribes choices that are optimal for each

type, independent of what the other types do. In short: each type is required

to have a dominant strategy.

12

We shall refer to this requirement as “dom-

inant strategy incentive compatibility.” This requirement restricts the set

of mechanisms and strategy combinations that the mechanism designer can

choose from. The available mechanisms and strategy combinations that we

consider in this section are a strict subset of those considered in the previous

section. This is because, obviously, a vector of strategies that prescribes a

dominant strategy to each type always constitutes a Bayesian equilibrium

of a game of incomplete information, but not vice versa.

We shall apply a similar logic to the “outside option.” We shall assume

that whenever the mechanism designer proposes a mechanism it is assumed

that in addition to the choices that he oﬀers to the agents in the mechanism

the agents can also choose to opt out of the mechanism and receive an outside

option utility, such as the utility zero in the auction and public goods models

of the previous chapter, or the utility of privately consuming the good in the

seller’s case in the bilateral trade model of the previous chapter. We will

require the strategy that the principal recommends to an agent to dominate

the strategy of not participating. We shall call this constraint “ex post

individual rationality constraint” because it requires that ex post, after the

mechanism has been played, no agent has an incentive to opt out. Recall

that in the previous chapter we had interpreted the individual rationality

constraint as an “interim individual rationality constraint.” In that chapter

agents evaluated the outside option once they knew their type using a belief

12

This use of the phrase “dominant” is slightly sloppy. We shall comment on terminology

at the end of this section.

CHAPTER 4. DOMINANT STRATEGY MECHANISMS 80

about the other agents’ types that was known to the mechanism designer.

The mechanism designer had to oﬀer a mechanism that ensured that agents

when conducting this evaluation didn’t have an incentive to opt out. Ex post

individual rationality restricts the set of mechanisms that the mechanism

designer can recommend further in comparison to the previous parts of these

notes. This is because, obviously, whenever participation dominates non-

participation at the ex post level it will also be optimal at the interim level,

but not vice versa.

We thus consider in this chapter a mechanism designer whose choice

set is smaller than the choice set available to the mechanism designer in

the previous part. Our interest will be in characterizing this choice set. We

shall also investigate whether the mechanisms that turned out to be optimal

in Chapter 3 remain in the choice set. We shall investigate these issues for

each of the three models introduced in Chapter 3.

As we shall not attribute explicit beliefs to the mechanism designer in

this Chapter, it is not obvious how we should deﬁne revenue maximizing

or welfare maximizing mechanisms. We shall therefore not study optimal

mechanisms in this section, but focus on the characterization of classes of

mechanisms that satisfy incentive compatibility, individual rationality, and

budget balance requirements. Developing a coherent decision theoretic ac-

count of the mechanism designer’s view of the agents’ payoﬀ types and their

beliefs about each other is a project that goes beyond what we do in these

notes.

13

We conclude the introduction with one formal clariﬁcation. We shall

use the expression “dominance” somewhat loosely in these notes. We shall

say that the strategy s

i

of player i “dominates” strategy s

i

if u

i

(s

i

, s

−i

) ≥

u

i

(s

i

, s

−i

) ∀s

−i

∈ S

−i

, and we shall say that a strategy is “dominant” if it

dominates all other strategies. In game theoretic terms a better expression

than “dominant” would be that strategy s

i

is in “always optimal.” This

is because the term dominance is set apart in game theory for two other

concepts: “strict dominance” (where u

i

(s

i

, s

−i

) > u

i

(s

i

, s

−i

) ∀s

−i

∈ S

−i

),

and “weak dominance” (where u

i

(s

i

, s

−i

) ≥ u

i

(s

i

, s

−i

) ∀s

−i

∈ S

−i

with

u

i

(s

i

, s

−i

) > u

i

(s

i

, s

−i

) ∃s

−i

∈ S

−i

). For the purposes of this note, we shall

be sloppy, and deﬁne “dominant” to mean “always optimal”. In many of the

examples that we investigate, our conclusions would remain true if we had

deﬁned “dominant” to mean “weakly dominant” in the traditional game-

13

The decision theoretic foundations of the dominant strategy approach to mechanism

design are investigated further in Chung and Ely (2007).

CHAPTER 4. DOMINANT STRATEGY MECHANISMS 81

theoretic sense. Our conclusions would, however, not remain true if we

had deﬁned “dominant” as the traditional game-theoretic concept of “strict

dominance.”

4.2 Single Unit Auctions

4.2.1 Set-Up

We return to the single unit auction setting of Section 3.2. We brieﬂy

recapitulate the aspects of the model that are relevant here. A seller seeks to

sell a single indivisible good. There are N potential buyers: I = ¦1, 2, ..., N¦.

Buyer i’s type is θ

i

∈ [θ,

¯

θ] where 0 ≤ θ <

¯

θ. We denote by θ the vector

(θ

1

, θ

2

, . . . , θ

N

). θ is an element of Θ ≡ [θ,

¯

θ]

N

. We denote by θ

−i

the

vector θ when we leave out θ

i

. We denote by Θ

−i

the set [θ,

¯

θ]

N−1

, so that

θ

−i

∈ Θ

−i

. Each buyer i knows her own type θ

i

but not θ

−i

.

We assume that buyer i’s utility if she is of type θ

i

equals θ

i

− t

i

if she

obtains the good and pays a transfer t

i

∈ R to the seller, and that it is −t

i

if she does not obtain the good and pays a transfer t

i

to the seller. The

seller’s utility if he obtains transfers t

i

from the N agents is

N

i=1

t

i

.

4.2.2 Mechanisms, Direct Mechanisms, and the Revelation

Principle

Recall that a general mechanism is a game tree together with an assignment

of a probability distribution over outcomes, that is a probability distribution

over ¦∅, 1, 2, . . . , N¦ R

N

, to each terminal history of the game. We shall

imagine, as in the previous Chapter, that the seller proposes a mechanism

together with one strategy for each buyer, where a strategy assigns to every

type of the buyer a complete behavior plan, possibly randomized, for the

game tree.

The incentive compatibility requirement that the seller’s proposed strat-

egy combination has to satisfy in this Chapter is more restrictive than it

was in the previous Chapter. We now require that each type of each buyer

ﬁnds it optimal to choose the proposed strategy, for all possible strategy

combinations that the other buyers might play. Let us say that in this case

the strategy prescribed for each type of each buyer is a “dominant” strategy.

We shall also require that each type of each buyer ﬁnds that her expected

utility, if she follows the seller’s recommendation, is nonnegative, for each

possible realization of the other buyers’ types, and for each strategy of the

CHAPTER 4. DOMINANT STRATEGY MECHANISMS 82

other buyers. We shall say that in this case participation is for each type of

each buyer a “dominant” strategy.

As before, we do not formalize the general deﬁnition of a mechanism,

nor do we formalize the notions of incentive compatibility and of individual

rationality that we have just described. The reason is that such a formal-

ization would take long to write down, and we would need it only for a very

short while. We wouldn’t have to use this formalization much further in

these notes because the “revelation principle” extends to our current set-

ting, and we will be able to restrict attention to direct mechanisms. The

deﬁnition of direct mechanisms is as before, see Deﬁnition 3.1. The following

result is the revelation principle for our setting.

Proposition 4.1 (Revelation Principle for Dominant Strategy Mechanisms).

Suppose a mechanism Γ and a strategy combination σ for Γ are such that

for each type θ

i

of each buyer i the strategy σ

i

(θ

i

) is a dominant strategy

in Γ. Then there exists a direct mechanism Γ

**and a strategy combination
**

σ

of Γ

**such that for every type θ
**

i

of each buyer i the strategy σ

i

(θ

i

) is a

dominant strategy in Γ

, and:

(i) The strategy vector σ

**satisﬁes for every i and every θ
**

i

:

σ

i

(θ

i

) = θ

i

,

that is, σ

**prescribes telling the truth;
**

(ii) For every vector θ of types the distribution over allocations and the

expected payments that result in Γ if the agents play σ is the same as

the distribution over allocations and the expected payments that result

in Γ

if the agents play σ

.

Proof. Construct Γ

**as required by part (ii) of the proposition. We can prove
**

the result by showing that truth telling will be a dominant strategy in this

direct mechanism. To see this suppose it were not. If type θ

i

prefers to

report that her type is θ

i

for some type vector of the other agents θ

−i

, then

the same type θ

i

would have preferred to deviate from σ

i

, and to play the

strategy that σ

i

prescribes for θ

i

in Γ, for the strategy combination that the

types θ

−i

play in Γ

. Hence σ

i

would not be a dominant strategy in Γ.

Proposition 4.1 shows that for dominant strategy implementation, as

for Bayesian equilibrium implementation, it is without loss of generality

CHAPTER 4. DOMINANT STRATEGY MECHANISMS 83

to restrict attention to direct mechanisms in which truth-telling is a dom-

inant strategy. Moreover, if the indirect mechanism that the mechanism

designer contemplates is individually rational, then the same is true for the

corresponding direct mechanism. Hence, we can restrict attention to direct

mechanisms that are dominant strategy incentive compatible and ex post

individually rational as deﬁned below.

Deﬁnition 4.1. A direct mechanism (q, t

1

, t

2

, . . . , t

N

) is “dominant strategy

incentive compatible” if truth telling is a dominant strategy for each type

of each buyer, that is, if for all i ∈ I, all θ

i

, θ

i

∈ [θ,

¯

θ] and for all θ

−i

∈ Θ

−i

:

θ

i

q

i

(θ

i

, θ

−i

) −t

i

(θ

i

, θ

−i

) ≥ θ

i

q

i

(θ

i

, θ

−i

) −t

i

(θ

i

, θ

−i

).

Deﬁnition 4.2. A direct mechanism (q, t

1

, t

2

, . . . , t

N

) is “ex post individ-

ually rational” if for each type of each buyer participation is a dominant

strategy, that is, if for all i ∈ I, all θ

i

∈ [θ,

¯

θ] and all θ

−i

∈ Θ

−i

:

θ

i

q

i

(θ

i

, θ

−i

) −t

i

(θ

i

, θ

−i

) ≥ 0.

4.2.3 Characterizing Dominant Strategy Incentive Compat-

ibility and Ex Post Individual Rationality

In this subsection we develop a better understanding of the structure of

the set of all direct mechanisms that satisfy the two conditions introduced

in Deﬁnitions 4.1 and 4.2. The characterization of incentive compatible is

actually exactly as in Propositions 2.2 and 3.2 except that the result now

applies for every realization θ

−i

of types of the other agents.

Proposition 4.2. A direct mechanism (q, t

1

, t

2

, . . . , t

N

) is dominant strat-

egy incentive compatible if and only if for every i ∈ I and every θ

−i

∈ Θ

−i

:

(i)

q

i

(θ

i

, θ

−i

) is increasing in θ

i

;

(ii) for every θ

i

∈ [θ,

¯

θ]:

t

i

(θ

i

, θ

−i

) = t

i

(θ, θ

−i

) + (θ

i

q

i

(θ

i

, θ

−i

) −θq

i

(θ, θ

−i

)) −

_

θ

θ

q

i

(

ˆ

θ, θ

−i

)d

ˆ

θ

We omit the proof of this result as it is the same as the proof of Propo-

sitions 2.2, applied to each agent i, and to each possible vector θ

−i

of the

other agents. An interesting and important point is that Proposition 4.2 im-

plies an ex post revenue equivalence result. Whereas our previous revenue

CHAPTER 4. DOMINANT STRATEGY MECHANISMS 84

equivalence results said that the allocation rule and the interim expected

payments of the lowest types imply the interim expected payments for all

types, part (ii) of Proposition 4.2 says that the allocation rule and the ex

post payments of the lowest types imply the ex post payments of all types.

In analogy to our earlier results, we also obtain a simple characterization

of ex post individual rationality for direct mechanisms that are dominant

strategy incentive compatible. We omit the proof.

Proposition 4.3. A dominant strategy incentive compatible direct mecha-

nism (q, t

1

, t

2

, . . . , t

N

) is ex post individually rational if and only if for every

i ∈ I and every θ

−i

∈ Θ

−i

we have:

t

i

(θ

i

, θ

−i

) ≤ θ

i

q

i

(θ

i

, θ

−i

).

4.2.4 Canonical Auctions

We now display a class of dominant strategy incentive compatible and ex

post individually rational direct mechanisms. We call the direct mechanisms

in this class “canonical auctions.”

14

We don’t claim that there are no other

dominant strategy incentive compatible and ex post individually rational

direct mechanisms than canonical auctions, but we show that this class is

rich enough to include the mechanism that we identiﬁed in Sections 3.2.4 and

3.2.5 as expected revenue maximizing and as expected welfare maximizing.

Recall from the discussion of the expected revenue maximizing auction

in Section 3.2.4 the deﬁnition of the functions ψ

i

: [θ,

¯

θ] →R for i ∈ I:

ψ

i

(θ

i

) ≡ θ

i

−

1 −F

i

(θ

i

)

f

i

(θ

−i

)

for all θ

i

∈ [θ,

¯

θ]. (4.1)

The regularity assumption under which we derived the expected revenue

maximizing auction was that all functions ψ

i

were increasing. Our inter-

pretation of these functions was that they assigned to each type a “virtual

type.” Under the regularity assumption, the expected revenue maximizing

auction assigned the object to the buyer with the highest virtual type.

Instead of working with these speciﬁc function ψ

i

, we now consider arbi-

trary strictly increasing functions ψ

i

. We show that for any such functions

an allocation rule that is constructed as in the expected revenue maximiz-

ing auction in Section 3.2.4 can be supplemented with transfer rules that

make the mechanism dominant strategy incentive compatible and ex post

14

This expression is not commonly used. I made it up for these notes.

CHAPTER 4. DOMINANT STRATEGY MECHANISMS 85

individually rational. The advantage of putting the result in this generality

is that we can use it to also show the implementability of allocation rules

other than the expected revenue maximizing. To simplify the exposition we

assume that the functions ψ

i

are continuous.

Deﬁnition 4.3. A direct mechanism (q, t

1

, t

2

, . . . , t

N

) is called a “canonical

auction” if there are strictly increasing and continuous functions ψ

i

: [θ,

¯

θ] →

R for i ∈ I such that for all θ ∈ Θ and i ∈ I:

q

i

(θ) =

_

_

_

1

n

if ψ

i

(θ

i

) ≥ 0 and ψ

i

(θ

i

) ≥ ψ

j

(θ

j

) for all j ∈ I with j ,= i,

0 if ψ

i

(θ

i

) < 0 or ψ

i

(θ

i

) < ψ

j

(θ

j

) for some j ∈ I with j ,= i,

where n is the number of agents j ∈ I such that ψ

j

(θ

j

) = ψ

i

(θ

i

), and:

t

i

(θ) =

_

¸

¸

_

¸

¸

_

1

n

min¦θ

i

∈ [θ,

¯

θ] [ ψ

i

(θ

i

) ≥ 0 and

ψ

i

(θ

i

) ≥ ψ

j

(θ

j

) for all j ∈ I with j ,= i¦ if q

i

(θ) > 0

0 if q

i

(θ) = 0,

for all θ ∈ Θ.

It is worth considering the transfer rule in detail. If bidder i does not

win the auction, then bidder i does not have to pay anything. If bidder i

does win the auction,

15

then bidder i’s payment equals the lowest type that

she might have had that would have allowed her to win the auction. The

assumed continuity of ψ

i

guarantees that this minimum exists.

Proposition 4.4. Every canonical auction (q, t

1

, t

2

, . . . , t

N

) is dominant

strategy incentive compatible and ex post individually rational. Moreover,

for every i ∈ I: u

i

(θ, θ

−i

) = 0 for all θ

−i

∈ Θ

−i

.

Proof. We ﬁrst show dominant strategy incentive compatibility. Suppose

that buyer i is of type θ

i

, that all other buyers are of type θ

−i

, and that

q

i

(θ

i

, θ

−i

) = 0. Does buyer i have an incentive to report a diﬀerent type θ

i

?

If q

i

(θ

i

, θ

−i

) = 0, then her utility doesn’t change. If q

i

(θ

i

, θ

−i

) =

1

n

> 0, it

will have to be the case that θ

i

> θ

i

. Moreover, buyer i’s payment will be

larger than

1

n

θ

i

, as her payment will be the lowest type of buyer i that wins

against θ

−i

, and by assumption θ

i

is not large enough. Thus, buyer i can

win the auction, but only by paying more than the object is worth to her.

Thus, she has no incentive to change her strategy.

15

For simplicity, we ignore ties.

CHAPTER 4. DOMINANT STRATEGY MECHANISMS 86

Consider next the case that q

i

(θ

i

, θ

−i

) = 1. If buyer i changes her report

to another type θ

i

for which q

i

(θ

i

, θ

−i

) = 1, then her utility doesn’t change,

as her payment does not depend on her report. If she changes her report

to a type θ

i

for which q

i

(θ

i

, θ

−i

) = 0, her utility decreases. This is because

if she reports θ

i

truthfully, her transfer payment, by the deﬁnition of the

transfer rule, will be below θ

i

. Hence a truthful report gives her a positive

surplus.

Consider ﬁnally the case that q

i

(θ

i

, θ

−i

) =

1

n

where n ≥ 2. Then buyer

i’s transfer payment is

1

n

θ

i

, and her expected utility will be zero. If buyer

i changes her report to another type θ

i

> θ

i

so that q

i

(θ

i

, θ

−i

) = 1, then

her expected utility doesn’t change, as her payment will still be θ

i

. If she

changes her report to a type θ

i

for which q

i

(θ

i

, θ

−i

) = 0, her expected utility

is again zero. Hence, a truthful report is optimal. This concludes the proof

of dominant strategy incentive compatibility.

The proof of dominant strategy incentive compatibility also shows that

a buyer’s utility is always non-negative if she wins the auction. If she loses

the auction, her utility is zero. Therefore, the mechanism also satisﬁes ex

post individual rationality. The lowest type, θ, either loses the auction, and

has utility zero, or wins the auction and has to pay θ, in which case utility

is also zero. This proves the last sentence of the Proposition.

We now provide two applications of Proposition 4.4. Firstly, suppose

that the seller had subjective beliefs F on Θ that reﬂected his view regard-

ing the likelihood of diﬀerent type vectors. Suppose moreover that these

beliefs satisﬁed the assumptions of Section 3.2.1 as well as the regularity

Assumption 3.1. Then one can interpret the optimal auction identiﬁed in

Section 3.2.4 as the expected revenue maximizing auction if the mechanism

designer assumes that his belief about Θ is also the prior distribution from

which the buyers’ beliefs about the other buyers’ valuations are derived.

This is a very strong assumption about the seller’s view of his environment.

In this part we have relaxed this assumption, and have assumed that the

seller is highly uncertain about the buyers’ beliefs about each other. Propo-

sition 4.4 shows that even if the seller’s uncertainty is so large that he wishes

to use a dominant strategy mechanism, he does not lose any expected rev-

enue. He can ﬁnd a transfer rule that implements the optimal allocation

rule, and that gives the lowest types expected utility zero. Thus, it is one of

the optimal auctions of Section 3.2.4. In other words, we have shown that

among the optimal auctions that we found in Section 3.2.4 there are some

that make truth telling a dominant strategy.

CHAPTER 4. DOMINANT STRATEGY MECHANISMS 87

A second application arises from Proposition 4.4 if we set ψ

i

(θ

i

) = θ

i

for every i ∈ I and every θ

i

∈ [θ,

¯

θ]. In this case the auction described in

Proposition 4.4 is just the Vickrey auction, and it is thus welfare maximiz-

ing. Thus, the mechanism designer can also achieve the objective of expected

welfare maximization through a dominant strategy mechanism. Of course,

Proposition 4.4 describes many other dominant strategy incentive compati-

ble and ex post individually rational auctions.

4.3 Public Goods

4.3.1 Set-Up

We shall now show that in the two examples with budget constraint that

we considered in Sections 3.3 and 3.4 the restriction to dominant strategy

incentive compatible mechanisms may severely limit what the mechanism

designer can achieve. We begin by brieﬂy recapitulating the public goods

model.

We consider a community consisting of N agents: I = ¦1, 2, . . . , N¦.

They have to choose whether to produce some indivisible, non-excludable

public good. We denote this decision by g ∈ ¦0, 1¦. If the public good is

produced, then g = 1. If it is not produced, then g = 0. Agent i’s utility if

the collective decision is g and if she pays a transfer t to the community is:

θ

i

g − t. The cost of producing the public good are cg, where c > 0. Here,

θ

i

is agent i’s type, and it is an element of θ

i

∈ [θ,

¯

θ] where 0 ≤ θ <

¯

θ.

We denote by θ the vector (θ

1

, θ

2

, . . . , θ

N

). θ is an element of Θ ≡ [θ,

¯

θ]

N

.

We denote by θ

−i

the vector θ when we leave out θ

i

. We denote by Θ

−i

the set [θ,

¯

θ]

N−1

, so that θ

−i

∈ Θ

−i

. Each agent i observes her own type

θ

i

, but not necessarily the other agents’ types θ

−i

. The mechanism designer

knows none of the θ

i

’s. The mechanism designer seeks to maximize expected

welfare where welfare is deﬁned to be

_

N

i=1

θ

i

_

g −

N

i=1

t

i

.

4.3.2 Direct Mechanisms

By the revelation principle we can restrict attention to direct mechanisms.

We shall also, as in Section 3.3, restrict attention to deterministic mecha-

nisms, that is to mechanism where conditional on the agents’ true types the

decision about the public good and also the agents’ transfers are not stochas-

tic. We introduced this assumption in Section 3.3 because it simpliﬁed the

formulation of the budget constraint and because it appeared innocuous. In

CHAPTER 4. DOMINANT STRATEGY MECHANISMS 88

the current section it is not clear that the assumption is innocuous. How-

ever, for simplicity, we continue to work with this assumption. Thus, direct

mechanisms are deﬁned as in Deﬁnition 3.4. The deﬁnitions of dominant

strategy incentive compatibility and ex post individual rationality are as in

the previous subsection. Moreover, we shall impose ex post budget balance,

which is deﬁned as in Deﬁnition 3.5. Observe that ex ante budget balance

has in our current context no meaning, because there is no prior probability

measure over Θ.

Note that in Deﬁnition 3.5 the budget constraint is an equality rather

than an inequality. An inequality would require that the sum of the con-

tributions is at least as large as the costs of the public good’s production

rather than that it is exactly equal. In the previous chapter the distinction

between an equality budget constraint and an inequality budget constraint

was not of much further importance, whereas in the current chapter the

equality requirement is an important restriction. We might motivate it by

arguing that we consider an environment in which the agents participating

in the public goods mechanism have no way of committing to a way of dis-

posing of excess funds if the collected transfers exceed what is required for

the public good production. But our main reason for restricting attention

to an equality budget constraint is that this simpliﬁes our arguments below.

4.3.3 Characterizing Dominant Strategy Incentive Compat-

ibility and Ex Post Individual Rationality

In this subsection we neglect the budget constraint, and seek a characteriza-

tion of dominant strategy incentive compatible mechanisms, and a condition

under which such mechanisms are in addition ex post individually rational.

We could use for this the same arguments as we used in the previous sec-

tion. However, a simpler proof can be given if one restricts attention to

deterministic mechanisms, as we have done here.

Proposition 4.5. A direct mechanism is dominant strategy incentive com-

patible if and only for every i ∈ I and for every θ

−i

∈ Θ

−i

there are a type

ˆ

θ

i

∈ R and two payments τ

i

and ˆ τ

i

∈ R such that:

θ

i

<

ˆ

θ

i

⇒q(θ

i

, θ

−i

) = 0 and t

i

(θ

i

, θ

−i

) = τ

i

;

θ

i

>

ˆ

θ

i

⇒q(θ

i

, θ

−i

) = 1 and t

i

(θ

i

, θ

−i

) = ˆ τ

i

;

θ

i

=

ˆ

θ

i

⇒q(θ

i

, θ

−i

) = 0 and t

i

(θ

i

, θ

−i

) = τ

i

or

q(θ

i

, θ

−i

) = 1 and t

i

(θ

i

, θ

−i

) = ˆ τ

i

;

CHAPTER 4. DOMINANT STRATEGY MECHANISMS 89

ˆ τ

i

−τ

i

=

ˆ

θ

i

.

Note that in this Proposition

ˆ

θ

i

, τ

i

and ˆ τ

i

are allowed to depend on θ

−i

.

Observe also that we have not ruled out that

ˆ

θ

i

< θ (in which case we would

have q

i

(θ

i

, θ

−i

) = 1 for all θ

i

∈ [θ,

¯

θ]), nor have we ruled out

ˆ

θ

i

>

¯

θ (in which

case we would have q

i

(θ

i

, θ

−i

) = 0 for all θ

i

∈ [θ,

¯

θ]).

Proof. First we show suﬃciency. Consider an agent i with type θ

i

. This

agent can either report a type θ

i

>

ˆ

θ

i

, in which case the public good is

produced and the agent has to pay

ˆ

θ

i

, or the agent can report a type θ

i

<

ˆ

θ

i

,

in which case the public good is not produced, and the agent does not have

to pay anything. For the outcome, it only matters whether the reported

type is above or below

ˆ

θ

i

. Beyond that, the value of the reported type does

not matter. If in truth θ

i

≥

ˆ

θ

i

, then reporting a type above

ˆ

θ

i

, such as

the true type, is obviously optimal, because the agent obtains more than

he has to pay. A symmetric argument applies if θ

i

≤

ˆ

θ

i

. This proves

that the conditions provided are suﬃcient for dominant strategy incentive

compatibility.

For the converse ﬁx θ

−i

∈ Θ

−i

. Consider all θ

i

such that q

i

(θ

i

, θ

−i

) = 0.

Observe that t

i

(θ

i

, θ

−i

) has to be the same for all these θ

i

, because otherwise

buyer i would pretend to be the type for which the transfer payment is

lowest. By the same argument the transfer payment t

i

(θ

i

, θ

−i

) has to be the

same for all θ

i

for which q

i

(θ

i

, θ

−i

) = 0.

Consider ﬁrst the case that q

i

(θ

i

, θ

−i

) = 0 for all θ

i

∈ [θ,

¯

θ]. Denote by

τ

i

the (constant) payment of agent i. If we set

ˆ

θ

i

>

¯

θ and τ

i

= τ

i

+

ˆ

θ

i

all conditions of Proposition 4.5 hold. The case that q

i

(θ

i

, θ

−i

) = 1 for all

θ

i

∈ [θ,

¯

θ] can be dealt with symmetrically.

Now suppose that q

i

(θ

i

, θ

−i

) = 0 for some θ

i

and q

i

(θ

i

, θ

−i

) = 1 for some

other θ

i

. Denote the payments corresponding to the former case by τ

i

, and

denote agent i’s payment in the latter case by ˆ τ

i

. Deﬁne

ˆ

θ

i

= ˆ τ

i

−τ

i

. Then

types θ

i

>

ˆ

θ

i

will report a type such that the public good is produced, and

types θ

i

<

ˆ

θ

i

will report a type such that the public good is not produced.

Truthful reporting is therefore optimal only if the conditions of Proposition

4.5 hold.

Next we characterize ex post individual rationality. We need not give

any proof of the following simple result.

CHAPTER 4. DOMINANT STRATEGY MECHANISMS 90

Proposition 4.6. A dominant strategy incentive compatible direct mecha-

nism is ex post individually rational if and only for every i ∈ I and for every

θ

−i

∈ Θ

−i

:

t

i

(θ

i

, θ

−i

) ≤ θ

i

q(θ

i

, θ

−i

).

4.3.4 Canonical Mechanisms

We now introduce a class of mechanisms, “canonical mechanisms,” that

are closely related to the mechanisms that we identiﬁed in Section 3.3.4

as welfare maximizing or proﬁt maximizing mechanisms under the budget

constraint. The deﬁnition of these mechanisms involved functions ψ

i

, one

for each agent i ∈ I. The regularity assumption was that these functions

were strictly increasing. Here, we let ψ

i

be some arbitrary strictly increasing

function. To simplify the exposition we assume that the functions ψ

i

are

continuous. The decision rules in the mechanisms deﬁned below is of the

same form as the decision rule in the expected welfare or proﬁt maximizing

mechanisms in part 1. We show below that we can combine these decision

rules with speciﬁc transfer rules to make them dominant strategy incen-

tive compatible and ex post individually rational. However, in general, the

mechanisms that we describe below will not balance the budget. In Section

3.3.4, by contrast, we obtained mechanisms that were Bayesian equilibrium

incentive compatible, individually rational, and moreover the rules could be

made ex post budget balanced.

Deﬁnition 4.4. A direct mechanism (q, t

1

, t

2

, . . . , t

N

) is called “canonical”

if for every agent i there is a strictly increasing and continuous function

ψ

i

: [θ,

¯

θ] →R such that:

q(θ) =

_

_

_

1 if

N

i=1

ψ

i

(θ

i

) ≥ c,

0 otherwise,

for all θ ∈ Θ and, for every i ∈ I:

t

i

(θ) =

_

¸

_

¸

_

min¦

¯

θ

i

∈ [θ,

¯

θ] [ ψ

i

(

¯

θ

i

) +

j=i

ψ

j

(θ

j

) ≥ c¦ if q(θ) = 1

0 if q(θ) = 0,

for all θ ∈ Θ.

We now state the following simple result.

CHAPTER 4. DOMINANT STRATEGY MECHANISMS 91

Proposition 4.7. Every canonical mechanism (q, t

1

, t

2

, . . . , t

N

) is dominant

strategy incentive compatible and ex post individually rational. Moreover, for

every i ∈ I: u

i

(θ, θ

−i

) = 0 for all θ

−i

∈ Θ

−i

.

The proof of this result is left to the reader. Note that the mechanisms

obtained in Propositions 3.8 and 3.9 include canonical mechanisms.

4.3.5 Ex Post Budget Balance

Now we ask which direct mechanisms are dominant strategy incentive com-

patible, ex post individually rational, and satisfy ex post budget balance.

In these notes, we only provide a partial answer to this question. To my

understanding, there is no paper in the literature that answers this question

at a general level. The answer that we give in these notes is for the case

that there are only 2 agents. For this case, it is easy to obtain the following

characterization. To simplify the exposition, we assume in this characteri-

zation that the set of type vectors for which the public good is produced is

a closed set.

Proposition 4.8. Suppose N = 2. Suppose also that the set ¦θ [ q(θ) = 1¦

is closed. Then a direct mechanism is dominant strategy incentive compati-

ble, ex post individually rational, and ex post budget balanced if and only if

there are payments τ

1

, τ

2

∈ R with τ

1

+τ

2

= c such that:

q(θ) = 1 and t

i

(θ) = τ

i

for all i ∈ ¦1, 2¦ if θ

1

≥ τ

1

and θ

2

≥ τ

2

q(θ) = 0 and t

i

(θ) = 0 for all i ∈ ¦1, 2¦ otherwise.

A simple indirect implementation of the mechanism described in Propo-

sition 4.8 is this: Each of the two agents is allocated a share of the cost τ

i

,

so that τ

1

+τ

2

= c. Then each agent i is asked whether he or she is willing

to contribute τ

i

whereby the contribution has to be made only if the other

agent also pledges his or her contribution. If both agents make a pledge,

then the public good is produced, and each agent i pays τ

i

. Otherwise, it is

not produced and no agent pays anything. Note that, conditional on pro-

ducing the public good, agents’ contributions do not depend on their own

or on the other agent’s valuation.

Proof. It is easy to show that the mechanisms displayed in Proposition 4.8

are dominant strategy incentive compatible, ex post individually rational,

and ex post budget balanced. We therefore only show the necessity of these

conditions.

CHAPTER 4. DOMINANT STRATEGY MECHANISMS 92

Suppose q(θ) = 0 for all θ ∈ Θ. Then obviously the mechanism can be

dominant strategy incentive compatible, ex post individually rational, and

ex post budget balanced only if all payments are zero. We can describe

this mechanism in the form described in Proposition 4.8 by setting τ

i

>

¯

θ

for some agent i, and deﬁning τ

j

≡ c − τ

i

for j ,= i. Note that τ

j

may be

negative. This is not ruled out by the Proposition.

Let us now assume that q(θ) = 1 for at least one θ ∈ Θ. For i = 1, 2

deﬁne

ˆ

θ

i

≡ min¦θ

i

∈ [θ,

¯

θ] [ q(θ

i

, θ

j

) = 1 for some θ

j

∈ [θ,

¯

θ]¦ where j ,= i.

Our assumption that the set ¦θ [ q(θ) = 1¦ is closed guarantees that these

minima, and similar minima referred to below, are well deﬁned. We seek to

show that θ

i

≥

ˆ

θ

i

implies q(θ

i

,

ˆ

θ

j

) = 1, where i ,= j. To show this we deﬁne

¯

θ

1

≡ min¦θ

1

∈ [θ,

¯

θ] [ q(θ

1

,

ˆ

θ

2

) = 1¦ and

¯

θ

2

≡ min¦θ

2

∈ [θ,

¯

θ] [ q(

ˆ

θ

1

, θ

2

) = 1¦.

By deﬁnition

¯

θ

i

≥

ˆ

θ

i

for i ∈ ¦1, 2¦. We claim that

¯

θ

i

=

ˆ

θ

i

for i ∈ ¦1, 2¦.

Together with Proposition 4.5 this implies what we want to show. Suppose

that

¯

θ

i

>

ˆ

θ

i

for some i ∈ ¦1, 2¦. Without loss of generality suppose the

strict inequality held for i = 1. By Proposition 4.5 t

1

(

¯

θ

1

,

ˆ

θ

2

) =

¯

θ

1

and

t

2

(

¯

θ

1

,

ˆ

θ

2

) =

ˆ

θ

2

. Budget balance requires

¯

θ

1

+

ˆ

θ

2

= c. By Proposition 4.5 we

have that q(

¯

θ

1

,

¯

θ

2

) = 1, t

1

(

¯

θ

1

,

¯

θ

2

) =

ˆ

θ

1

and t

2

(

¯

θ

1

,

¯

θ

2

) =

ˆ

θ

2

. Thus the sum of

contributions if types are (

¯

θ

1

,

¯

θ

2

) is

ˆ

θ

1

+

ˆ

θ

2

. If

¯

θ

1

<

ˆ

θ

1

, then the sum of these

contributions is strictly less than

¯

θ

1

+

ˆ

θ

2

, of which we had just shown that

it equals c. Therefore, budget balance is violated. We conclude that

¯

θ

i

=

ˆ

θ

i

for i ∈ ¦1, 2¦. Hence θ

i

≥

ˆ

θ

i

implies q(θ

i

,

ˆ

θ

j

) = 1, where i ,= j.

From what we have shown in the previous paragraph we can infer that

θ

i

≥

ˆ

θ

i

for i = 1, 2 implies q(θ

1

, θ

2

) = 1. The reason is that ﬁrst, as

shown above, we can infer q(θ

1

,

ˆ

θ

2

) = 1, and then, using Proposition 4.5,

q(θ

1

, θ

2

) = 1. Also note that θ

1

<

ˆ

θ

1

or θ

2

<

ˆ

θ

2

implies q(θ) = 0. This is a

consequence of the deﬁnitions of

ˆ

θ

1

and

ˆ

θ

2

.

Now suppose that

ˆ

θ

i

> 0 for i = 1, 2. Then, by Proposition 4.5 the

payment of each agent i must be

ˆ

θ

i

whenever the public good is produced. If

we deﬁne τ

i

≡ θ

i

, the mechanism is of the form described in the Proposition.

The remaining case is that

ˆ

θ

i

= 0 for both agents i. This is the case

when the public good is produced with probability 1. Incentive compatibility

requires that no agent’s payment depends on their own report. In principle,

it may depend on the other agents’ reports, though. We write agent 1’s

payment as τ

i

(θ

2

), and agent 2’s payment as τ

2

(θ

1

). We want to show that

τ

1

does not depend on θ

2

, i.e. that τ

1

(θ

2

) = τ

1

(θ

2

) for all θ

2

, θ

2

∈ [θ,

¯

θ].

Suppose τ

1

(θ

2

) < τ

1

(θ

2

) for some θ

2

, θ

2

∈ [θ,

¯

θ]. Fix any θ

1

. Then, if agent

2’s type is θ

2

, the sum of contributions is: τ

1

(θ

2

) +τ

2

(θ

1

). If agent 2’s type

CHAPTER 4. DOMINANT STRATEGY MECHANISMS 93

is θ

2

, the sum of contributions is: τ

1

(θ

2

) + τ

2

(θ

1

). Only one, but not both

of these sums can equal c, thus contradicting budget balance. Therefore,

τ

1

(θ

2

) is a constant. We write τ

1

for this constant. Budget balance then

requires τ

2

to be constant too. Again we ﬁnd that the mechanism is of the

form described in the Proposition.

4.4 Bilateral Trade

4.4.1 Set-Up

We recapitulate brieﬂy the set-up from Section 3.4. A seller S owns a single

indivisible good. There is one potential buyer B. The seller’s utility if he

sells the good and receives a transfer payment t is equal to t. If he does

not sell the good and receives a transfer t then his utility is θ

S

+ t where

θ

S

∈ [θ

S

,

¯

θ

S

] is the seller’s type. The buyer’s utility if he purchases the

good and pays a transfer t equals θ

B

−t, where θ

B

∈ [θ

B

,

¯

θ

B

] is the buyer’s

privately observed type. The buyer’s utility if he does not obtain the good

and pays transfer t is −t. We deﬁne θ ≡ (θ

s

, θ

B

). Each agent knows his

own type but not the other agent’s type. We consider the situation from a

mechanism designer’s perspective who does not know the values of θ

S

and

θ

B

.

4.4.2 Dominant Strategy Incentive Compatible and Ex Post

Individually Rational Direct Mechanisms

By the revelation principle we restrict ourselves to direct mechanisms with

deterministic trading rules, as deﬁned as in Deﬁnition 3.9. We require these

to satisfy dominant strategy incentive compatibility, ex post individual ra-

tionality, and ex post budget balance.

To characterize dominant strategy incentive compatibility we can pro-

ceed as in the previous section. The problem is analogous to that of the

previous section because we have restricted attention to deterministic mech-

anisms. For briefness, we state the result without proof.

Proposition 4.9. A direct mechanism is dominant strategy incentive com-

patible if and only if for every θ

B

∈ [θ

B

,

¯

θ

B

] there exist a type

ˆ

θ

S

∈ R and

payments τ

S

, ˆ τ

S

∈ R such that:

θ

S

<

ˆ

θ

S

⇒q(θ) = 1 and t

S

(θ) = ˆ τ

S

;

θ

S

>

ˆ

θ

S

⇒q(θ) = 0 and t

S

(θ) = τ

S

;

CHAPTER 4. DOMINANT STRATEGY MECHANISMS 94

θ

S

=

ˆ

θ

S

⇒q(θ) = 0 and t

S

(θ) = τ

S

or

q(θ) = 1 and t

S

(θ) = ˆ τ

S

;

ˆ τ

S

−τ

S

=

ˆ

θ

S

;

and for every θ

S

∈ [θ

S

,

¯

θ

S

] there exist a type

ˆ

θ

B

∈ R and payments ˆ τ

B

, ˆ τ

B

∈

R such that:

θ

B

<

ˆ

θ

B

⇒q(θ) = 0 and t

B

(θ) = τ

B

;

θ

B

>

ˆ

θ

B

⇒q(θ) = 1 and t

B

(θ) = ˆ τ

B

;

θ

B

=

ˆ

θ

B

⇒q(θ) = 0 and t

B

(θ) = τ

B

or

q(θ) = 1 and t

B

(θ) = ˆ τ

B

;

ˆ τ

B

−τ

B

=

ˆ

θ

B

.

We continue with a standard characterization of individual rationality

the proof of which we also omit.

Proposition 4.10. A dominant strategy incentive compatible direct mech-

anism is ex post individually rational if and only if for every θ

B

∈ [θ

B

,

¯

θ

B

] :

t

S

(

¯

θ

S

, θ

B

) ≥

¯

θ

S

q(

¯

θ

S

, θ

B

)

and for every θ

S

∈ [θ

S

,

¯

θ

S

] :

t

B

(θ

S

, θ

B

) ≤ θ

B

q(θ

S

, θ

B

).

4.4.3 Canonical Mechanisms

We introduce again a class of mechanisms that encompasses some of those

that we introduced earlier as expected welfare or expected proﬁt maximiz-

ing (under a budget constraint), and that are dominant strategy incentive

compatible and ex post individually rational.

Deﬁnition 4.5. A direct mechanism (q, t

S

, t

B

) is called “canonical” if for

every agent i ∈ ¦S, B¦ there is a strictly increasing and continuous function

ψ

i

: [θ

i

,

¯

θ

i

] →R such that:

q(θ) =

_

_

_

1 if ψ

B

(θ

B

) ≥ ψ

S

(θ

S

),

0 otherwise,

CHAPTER 4. DOMINANT STRATEGY MECHANISMS 95

and

t

S

(θ) =

_

_

_

max¦

¯

θ

S

∈ [θ

S

,

¯

θ

S

] [ ψ

B

(θ

B

) ≥ ψ

S

(

¯

θ

S

)¦ if q(θ) = 1

0 if q(θ) = 0,

and

t

B

(θ) =

_

_

_

min¦

¯

θ

B

∈ [θ

B

,

¯

θ

B

] [ ψ

B

(θ

B

) ≥ ψ

S

(

¯

θ

S

)¦ if q(θ) = 1

0 if q(θ) = 0,

Thus, if trade takes place, the seller receives the largest value that he

could have had and trade would still have taken place. Similarly, the buyer

receives the lowest value that he could have had and trade would still have

taken place.

Proposition 4.11. Every canonical mechanism is dominant strategy incen-

tive compatible and ex post individually rational. Moreover, u

S

(

¯

θ

S

, θ

B

) =

¯

θ

S

for every [θ

B

∈ θ

B

,

¯

θ

B

] and u

B

(θ

S

, θ

B

) = 0 for every θ

S

∈ [θ

S

,

¯

θ

S

].

The proof of this result is analogous to the proof of Proposition 4.7 and

we omit it.

4.4.4 Ex Post Budget Balance

Now we characterize direct mechanisms that are dominant strategy incentive

compatible, ex post individually rational, and satisfy ex post budget balance.

The structure of our result is similar to the structure of Proposition 4.8.

Proposition 4.12. Suppose that the set ¦θ [ q(θ) = 1¦ is closed. A direct

mechanism is dominant strategy incentive compatible, ex post individually

rational, and ex post budget balanced if and only if either:

q(θ) = 0 and t

i

(θ) = 0 for all i ∈ ¦S, B¦ and all θ ∈ Θ,

or there is a

ˆ

θ such that:

q(θ) = 1 and t

i

(θ) =

ˆ

θ for all i ∈ ¦S, B¦ if θ

S

≤

ˆ

θ and θ

B

≥

ˆ

θ;

q(θ) = 0 and t

i

(θ) = 0 for all i ∈ ¦S, B¦ if θ

S

>

ˆ

θ or θ

B

<

ˆ

θ.

The proof of this result is analogous to the proof of Proposition 4.8, and

we omit it. The class of mechanisms described in Proposition 4.12 can best

CHAPTER 4. DOMINANT STRATEGY MECHANISMS 96

be understood as ﬁxed price mechanisms. There is a ﬁxed price

ˆ

θ, and trade

takes place at this price if and only if the buyer reports a value above this

price and the seller reports a value below this price. This mechanism is

restrictive because the price does not depend at all on the agents’ reported

valuations.

4.5 Comments on the Literature

The results on dominant strategy mechanisms in the single unit auction

setting are special cases of a more general result in Mookherjee and Re-

ichelstein (1992) who investigate conditions under which one can construct

for any Bayesian incentive compatible mechanisms an equivalent dominant

strategy mechanism. I have also used Section 5.2 of Vijay Krishna’s Auction

theory for this section of the notes. The section on public goods mechanisms

is partially based on discussions with Arunava Sen from the Indian Statisti-

cal Institute, New Delhi. Dominant strategy mechanisms for bilateral trade

are discussed in Hagerty and Rogerson (1987).

4.6 Problems

a) Assume that stochastic decision rules are allowed in the public goods

problem, and state a characterization of dominant strategy incentive

compatible direct mechanisms that is analogous to Proposition 4.2.

(You don’t have to prove this characterization.) Show that Proposition

4.5 is a special case of the more general result that you have stated.

Chapter 5

Dominant Strategy

Mechanisms: General

Theory

5.1 Introduction

In the ﬁrst three chapters of these notes we have considered examples that

illustrate the general theory of mechanism design. In this and the next

Chapter we are now going to consider the extent to which we can generalize

what we have seen in these examples. We shall investigate which of our ear-

lier results continue to be true in the more general model. We also introduce

some new results. We begin by considering dominant strategy mechanisms.

The reason for starting with dominant strategy mechanisms is that these

are a subset of the Bayesian incentive compatible mechanisms. Therefore,

some of the results in this chapter will also be useful in the next chapter.

5.2 Set-Up

There are N agents. The set of agents is denoted by I = ¦1, 2, . . . , N¦. They

have to choose an alternative a out of some set A of mutually exclusive

alternatives. Agent i’s utility if alternative a is chosen, and agent i pays

transfer t

i

is:

u

i

(a, θ

i

) −t

i

.

Here, θ

i

is agent i’s type, and t

i

is agent i’s transfer. Note that the set-up

does include the case in which A is the set of lotteries over some ﬁnite set,

97

CHAPTER 5. DOMINANCE: GENERAL THEORY 98

and u

i

is an expected utility function.

We shall employ similar notation as before: The set of possible types of

agent i is Θ

i

which we now take to be some abstract set. We denote by θ

the vector of types: (θ

1

, θ

2

, . . . , θ

N

). The set of all possible type vectors is

Θ ≡ Θ

1

Θ

2

. . . θ

N

. Finally, we write θ

−i

for a vector θ of types if we leave

out agent i’s type. The set of all θ

−i

is Θ

−i

which is the cartesian product

of the sets Θ

j

, leaving out Θ

i

.

Our interest is in dominant strategy incentive compatible mechanisms.

The revelation principle holds in this setting, and we can restrict attention

to direct mechanisms.

Deﬁnition 5.1. A direct mechanism (q, t

1

, t

2

, . . . , t

N

) consists of a mapping

q : Θ →A,

that maps every type vector into a collective decision, and mappings

t

i

: Θ →R,

one for each player i ∈ I, that indicate for each type vector the transfer that

agent i needs to make.

We call q the “decision rule.” Note that we restrict ourselves to mech-

anisms where the payment is deterministic. If we assume that agents are

risk-neutral, then this is without loss of generality, as long as we are only

concerned with characterizations of incentive compatibility and individual

rationality. When we consider budget balance, this point is more delicate.

For the moment, however, we are only concerned with incentive compatibil-

ity.

Deﬁnition 5.2. A direct mechanism is “dominant strategy incentive com-

patible” if for all θ ∈ Θ, all i ∈ I and all θ

i

∈ Θ

i

we have:

u

i

(q(θ), θ

i

) −t

i

(θ) ≥ u

i

(q(θ

i

, θ

−i

), θ

i

) −t

i

(θ

i

, θ

−i

).

We introduce other requirements, such as individual rationality and bud-

get balance, later when needed. The next sections will only be concerned

with incentive compatibility.

CHAPTER 5. DOMINANCE: GENERAL THEORY 99

5.3 Ex Post Revenue Equivalence

We begin by providing a result that is analogous to the ex post revenue

equivalence results that we obtained in Chapter 4, such as Proposition 4.2.

The proofs of those results used the fact that in all the examples the type

spaces were connected sets. With ﬁnite type spaces these results are not

true. In our abstract setting in this part of the lecture notes, we need to

make therefore an additional assumption regarding the type spaces Θ

i

if we

want to obtain a payoﬀ equivalence result. The assumption that we shall

make below is that for every i ∈ I the set Θ

i

is a convex subset of some

ﬁnite dimensional Euclidean space.

An assumption for the type spaces alone is, however, not enough. We

shall also make use of an assumption about how the agents’ utility functions

depends on their types. We shall assume that every agent i’s utility function

is a convex function of the agent’s own type. This assumption will ensure

that agents’ utility under an incentive compatible mechanism depends suﬃ-

ciently smoothly on their type so that we can employ an envelope theorem.

Convexity of the utility function in a player’s own type is satisﬁed, for ex-

ample, if the space of alternatives A is the set of all lotteries over a ﬁnite

outcome space, and if the type of a player is just the vector of this player’s

Bernoulli utilities of the outcomes. In this case, the player’s utility is a linear

function of the type vector, and hence convex.

Proposition 5.1. Suppose that for every i ∈ I the set Θ

i

is a convex sub-

set of a ﬁnite dimensional Euclidean space. Moreover, assume that for ev-

ery i ∈ I the function u

i

(a, θ

i

) is a convex function of θ

i

. Suppose that

(q, t

1

, t

2

, . . . , t

N

) is a dominant strategy incentive compatible mechanism.

Then a direct mechanism with the same decision rule, (q, t

1

, t

2

, . . . , t

N

), is

incentive compatible if and only if for every i ∈ I and every θ

−i

∈ Θ

−i

there

is a number τ

i

(θ

−i

) ∈ R such that

t

i

(θ) = t

i

(θ) +τ

i

(θ

−i

) for all θ ∈ Θ.

As our earlier revenue equivalence results, this result says that for given

decision rule q the set of all transfer rules that implement q consists of

one function and all its parallel translations. Because we are dealing with

dominant strategy implementation, our result concerns ex post payoﬀs. The

proof of this result is given in Krishna and Maenner (2001). The proof is

analogous to the proof of similar results that we provided earlier, but it is

more technical, because the appropriate diﬀerentiability properties of the

functions involved needs to be established. We omit the proof.

CHAPTER 5. DOMINANCE: GENERAL THEORY 100

5.4 Implementing Eﬃcient Decision Rules

We now begin an investigation of the question for which decision rules q there

are transfer rules that make the decision rule q dominant strategy incentive

compatible. We begin by showing that decision rules that maximize ex ante

expected welfare, where welfare is deﬁned in a utilitarian way, are always

dominant strategy incentive compatible. We shall call such decision rules

“eﬃcient,” although a much more careful investigation would be needed if

we wanted to clarify the relation between maximizing utilitarian welfare and

“Pareto eﬃciency.”

Deﬁnition 5.3. An allocation rule q

∗

is called “eﬃcient” if for every θ ∈ Θ

we have:

N

i=1

u

i

(q

∗

(θ), θ

i

) ≥

N

i=1

u

i

(a, θ

i

) for all a ∈ A.

We shall now introduce a class of mechanisms that make eﬃcient decision

rules dominant strategy incentive compatible. We encountered special cases

of these mechanisms already in earlier parts of these notes.

Deﬁnition 5.4. A direct mechanism (q, t

1

, t

2

, . . . t

N

) is called a “Vickrey-

Clarke-Groves” (VCG) mechanism if q is an eﬃcient decision rule, and if for

every i ∈ I there is a function

τ

i

: Θ

−i

→R

such that

t

i

(θ) = −

j=i

u

j

(q(θ), θ

j

) +τ

i

(θ

−i

) for all θ ∈ Θ.

Proposition 5.2. VCG mechanisms are dominant strategy incentive com-

patible.

Proof. Consider any agent i, and take θ

−i

as given. If agent i is of type θ

i

and reports that she is of type θ

i

, then her utility is:

u

i

(q(θ

i

, θ

−i

), θ

i

) +

j=i

u

j

(q(θ

i

, θ

−i

), θ

j

) −τ

i

(θ

−i

)

=

N

j=1

u

j

(q(θ

i

, θ

−i

), θ

j

) −τ

i

(θ

−i

).

CHAPTER 5. DOMINANCE: GENERAL THEORY 101

Note that τ

i

(θ

−i

) is not changed by agent i’s report. Only the ﬁrst expression

matters for i’s incentives. But this expression is social welfare at type vector

θ if the decision is q(θ

i

, θ

−i

). As q(θ) maximizes social welfare for type vector

θ, it is optimal for agent i to report her true type: θ

i

= θ

i

.

We can combine this result with the payoﬀ equivalence result in Propo-

sition 1 to obtain conditions under which VCG mechanisms are the only

mechanisms that make eﬃcient decision rules dominant strategy incentive

compatible.

Corollary 5.1. Suppose that for every i ∈ I the set Θ

i

is a convex subset of

a ﬁnite dimensional Euclidean space. Moreover, assume that for every i ∈ I

the function u

i

(a, θ

i

) is a convex function of θ

i

. Suppose that (q, t

1

, t

2

, . . . t

N

)

is a dominant strategy incentive compatible mechanism, and suppose that q

is eﬃcient. Then (q, t

1

, t

2

, . . . t

N

) is a VCG mechanism.

Proof. By Proposition 1 every dominant strategy incentive compatible mech-

anism that implements an eﬃcient decision rule q must involve the same

transfers as the VCG mechanism up to additive constants τ

i

(θ

−i

) that may

be added to any agent i’s transfers. But adding such constants to a VCG

mechanism yields by the deﬁnition of VCG mechanisms another VCG mech-

anism.

5.5 Characterizing All Incentive Compatible

Decision Rules

In this section our objective is to characterize the set of all dominant strat-

egy incentive compatible mechanisms. One might wonder why this is of

interest, given that we showed in the previous section that eﬃcient mecha-

nisms are dominant strategy incentive compatible. There are two reasons.

Firstly, the mechanism designer might have objectives other than eﬃciency,

e.g. proﬁt maximization. Secondly, the mechanism designer might face con-

straints other than dominant strategy incentive compatibility which make

the implementation of an eﬃcient decision rule impossible.

Our focus will be on the decision rules q that may form part of an

incentive compatible direct mechanism rather than on the transfer rules.

We discuss a variety of closely related properties of decision rules which are

necessary, and under certain conditions also suﬃcient for decision rules to

be dominant strategy implementable.

CHAPTER 5. DOMINANCE: GENERAL THEORY 102

In the three examples of Chapters 3 and 4 we found that a necessary and

suﬃcient condition for a decision rule to be implementable by a direct incen-

tive compatible mechanism was that q was monotone. Here, we start with

a result that can be proved by an appropriate adaptation of the argument

that led to our earlier results.

Deﬁnition 5.5. A decision rule q is “weakly monotone” if for every i ∈ I,

θ

−i

∈ Θ

−i

, and all θ

1

i

, θ

2

i

∈ Θ

i

, if q(θ

1

i

, θ

−i

) = a

1

and q(θ

2

i

, θ

−i

) = a

2

, then

u

i

(a

1

, θ

1

i

) −u

i

(a

2

, θ

1

i

) ≥ u

i

(a

1

, θ

2

i

) −u

i

(a

2

, θ

2

i

).

This deﬁnition considers a situation where the collective choice for a type

vector (θ

1

i

, θ

−i

) is alternative a

1

, but if we change agent i’s type from θ

1

i

to

θ

2

i

, then the collective choice becomes some alternative a

2

. The deﬁnition

says that in this case the utility diﬀerence for agent i between a

1

and a

2

must

have decreased as we switched from θ

1

i

to θ

2

i

. We can interpret the utility

diﬀerence as agent i’s “willingness to pay for alternative a

1

as opposed to

alternative a

2

.” The result says that agent i’s willingness to pay for a

1

as

opposed to a

2

must be at least as large when agent i is of type θ

1

i

as it is

when agent i is of type θ

2

i

.

Proposition 5.3. Suppose (q, t

1

, t

2

, . . . , t

N

) is a dominant strategy incentive

compatible mechanism. Then q is weakly monotone.

Proof. Dominant strategy incentive compatibility implies for all i ∈ I, θ

−i

∈

Θ

−i

, and all θ

1

i

, θ

2

i

∈ Θ

i

, if q(θ

1

i

, θ

−i

) = a

1

and q(θ

2

i

, θ

−i

) = a

2

, then:

u

i

(a

1

, θ

1

i

) −t

i

(θ

1

i

, θ

−i

) ≥ u

i

(a

2

, θ

1

i

) −t

i

(θ

2

i

, θ

−i

) ⇔

u

i

(a

1

, θ

1

i

) −u

i

(a

2

, θ

1

i

) ≥ t

i

(θ

1

i

, θ

−i

) −t

i

(θ

2

i

, θ

−i

) (5.1)

and

u

i

(a

1

, θ

2

i

) −t

i

(θ

1

i

, θ

−i

) ≤ u

i

(a

2

, θ

2

i

) −t

i

(θ

2

i

, θ

−i

) ⇔

u

i

(a

1

, θ

2

i

) −u

i

(a

2

, θ

2

i

) ≤ t

i

(θ

1

i

, θ

−i

) −t

i

(θ

2

i

, θ

−i

). (5.2)

These two inequalities imply:

u

i

(a

1

, θ

1

i

) −u

i

(a

2

, θ

1

i

) ≥ u

i

(a

1

, θ

2

i

) −u

i

(a

2

, θ

2

i

). (5.3)

which means that q is weakly monotone.

CHAPTER 5. DOMINANCE: GENERAL THEORY 103

We shall now show that a stronger condition than weak monotonicity,

called “cyclical monotonicity,” is necessary and suﬃcient for implementabil-

ity. Note that we can re-write the inequality that deﬁnes weak monotonicity

as:

_

u

i

(a

1

, θ

2

i

) −u

i

(a

1

, θ

1

i

)

_

+

_

u

i

(a

2

, θ

1

i

) −u

i

(a

2

, θ

2

i

)

_

≤ 0. (5.4)

This inequality considers on the left hand side the sum of the changes to

agent i’s utility that occur in the following two thought experiments. First,

we start at utility proﬁle (θ

1

i

, θ

−i

) with implemented alternative a

1

, and

we switch agent i’s type to θ

2

i

, leaving the chosen alternative hypothetically

unchanged as a

1

. Next, we start at utility proﬁle (θ

2

i

, θ

−i

) with implemented

alternative a

2

, and we switch agent i’s type to θ

1

i

, leaving again the chosen

alternative hypothetically unchanged. Both switches may cause an increase

or a decrease in agent i’s utility. The condition says that the sum of these

two utility changes is not positive.

Now suppose we consider a sequence of length k ∈ N, k ≥ 2 of types of

agent i: θ

1

i

, θ

2

i

, . . . , θ

k

i

, and assume that the last element of the sequence is the

same as the ﬁrst: θ

1

i

= θ

k

i

. We can consider the same sequence of switches as

we considered in the previous paragraph, and sum up the changes to agent

i’s utility caused by these switches. “Cyclical monotonicity” requires the

sum to be not positive.

Deﬁnition 5.6. A decision rule q is “cyclically monotone” if for every i ∈ I,

every θ

−i

∈ Θ

−i

and every sequence of length k ∈ N of types of agent i,

(θ

1

i

, θ

2

i

, . . . , θ

k

i

) ∈ Θ

k

i

, with θ

k

i

= θ

1

i

, we have:

k−1

κ=1

(u

i

(a

κ

, θ

κ+1

i

) −u

i

(a

κ

, θ

κ

i

)) ≤ 0

where for all κ = 1, 2, . . . , k we deﬁne: a

κ

≡ q(θ

κ

i

, θ

−i

).

Thus, cyclical monotonicity is the same as weak monotonicity if we re-

strict attention to the case that k = 3.

The following proposition is due to Rochet (1987). A remarkable as-

pect of this theorem is that it does not require any structure at all for the

sets of alternatives or the sets of types. Nevertheless, an elementary proof

shows that cyclical monotonicity is equivalent to dominant strategy imple-

mentability.

Proposition 5.4. A decision rule q is part of a dominant strategy incentive

compatible direct mechanism (q, t

1

, t

2

, . . . , t

N

) if and only if q is cyclically

monotone.

CHAPTER 5. DOMINANCE: GENERAL THEORY 104

Proof. We ﬁrst show that if q is part of a dominant strategy incentive com-

patible direct mechanism then it is cyclically monotone. Dominant strategy

incentive compatibility implies for every κ = 1, 2, . . . , k − 1 that type θ

κ+1

i

has no incentive to pretend to be type κ, given that all other agents have

reported type θ

−i

:

u

i

(a

κ

, θ

κ+1

i

) −t

i

(θ

κ

i

, θ

−i

) ≤ u

i

(a

κ+1

, θ

κ+1

i

) −t

i

(θ

κ+1

i

, θ

−i

) ⇔

u

i

(a

κ

, θ

κ+1

i

) −u

i

(a

κ+1

, θ

κ+1

i

) ≤ t

i

(θ

κ

i

, θ

−i

) −t

i

(θ

κ+1

i

, θ

−i

). (5.5)

We sum these inequalities over all κ = 1, 2, . . . , k − 1. The right hand side

then becomes zero, and we can deduce:

k−1

κ=1

_

u

i

(a

κ

, θ

κ+1

i

) −u

i

(a

κ+1

, θ

κ+1

i

)

_

≤ 0 ⇔

k−1

κ=1

u

i

(a

κ

, θ

κ+1

i

) −

k−1

κ=1

u

i

(a

κ+1

, θ

κ+1

i

) ≤ 0 (5.6)

Because θ

k

i

= θ

1

i

, we can write the subtracted sum as follows:

k−1

κ=1

u

i

(a

κ

, θ

κ+1

i

) −

k−1

κ=1

u

i

(a

κ

, θ

κ

i

) ≤ 0 ⇔

k−1

κ=1

_

u

i

(a

κ

, θ

κ+1

i

) −u

i

(a

κ

, θ

κ

i

)

_

≤ 0, (5.7)

which is what we wanted to show.

Next, we shall show that cyclical monotonicity implies that the rule q

can be combined with transfer payments so that the direct mechanism that

results is dominant strategy incentive compatible. To construct the transfer

payments, we shall need some further deﬁnitions. We ﬁx an arbitrary type

¯

θ

i

∈ Θ

i

. Then, for every θ

i

∈ Θ

i

we deﬁne o(θ

i

) to be the set of all ﬁnite

sequences (θ

1

i

, θ

2

i

, . . . , θ

k

i

) of elements of Θ

i

that satisfy: θ

1

i

=

¯

θ

i

and θ

k

i

= θ

i

.

Here, k can be any element of N with k ≥ 2. Then deﬁne the function

V : Θ →R by:

V (θ

i

, θ

−i

) ≡ sup

(θ

1

i

,θ

2

i

,...,θ

k

i

)∈S(θ

i

)

k−1

κ=1

_

u

i

(a

κ

, θ

κ+1

i

) −u

i

(a

κ

, θ

κ

i

)

_

(5.8)

for all (θ

i

, θ

−i

) ∈ Θ, where a

κ

is deﬁned to be q(θ

κ

i

, θ

−i

) for κ = 1, 2, . . . , k−1.

CHAPTER 5. DOMINANCE: GENERAL THEORY 105

Before we proceed, we verify that the function V is well-deﬁned. This is

the case if the expression the supremum of which is the right hand side of

equation (5.8) is bounded from above for given θ

i

and θ

−i

if q is cyclically

monotone. This is what we show.

First we note that the condition of cyclical monotonicity means that

for θ

i

=

¯

θ

i

that all sums on the right hand side of (5.8) are non-positive.

Therefore, for θ

i

=

¯

θ

i

, the supremum is well-deﬁned. Indeed, because the

trivial sequence (

¯

θ

i

,

¯

θ

i

) is contained in o(

¯

θ

i

), and because for this sequence

the sum over which we take the supremum in the deﬁnition of V is zero, we

can conclude that V (

¯

θ

i

, θ

−i

) = 0 for all θ

−i

∈ Θ

−i

.

Now consider some θ

i

,=

¯

θ

i

, ﬁx a sequence (θ

1

i

, θ

2

i

, . . . , θ

k

i

) ∈ o(θ

i

) and

deﬁne a

κ

as in the deﬁnition of V . Then:

V (

¯

θ

i

, θ

−i

) = sup

(θ

1

i

,θ

2

i

,...,θ

k

i

)∈S(

¯

θ

i

)

k−1

κ=1

_

u

i

(a

κ

, θ

κ+1

i

) −u

i

(a

κ

, θ

κ

i

)

_

≥

k−1

κ=1

_

u

i

(a

κ

, θ

κ+1

i

) −u

i

(a

κ

, θ

κ

i

)

_

+

_

u

i

(a

k

,

¯

θ

i

) −u

i

(a

k

, θ

i

)

_

(5.9)

Here, the inequality follows from the fact that o(

¯

θ

i

) includes the set of all

ﬁnite sequences that have θ

i

as their pen-ultimate element, and that have

¯

θ

i

as their last element. The inequality that we have obtained is equivalent

to:

k−1

κ=1

_

u

i

(a

κ

, θ

κ+1

i

) −u

i

(a

κ

, θ

κ

i

)

_

≤ V (

¯

θ

i

, θ

−i

) +

_

u

i

(a

k

, θ

i

) −u

i

(a

k

,

¯

θ

i

)

_

.

(5.10)

Using V (

¯

θ

i

, θ

−i

) = 0, which we showed earlier, we can infer:

k−1

κ=1

_

u

i

(a

κ

, θ

κ+1

i

) −u

i

(a

κ

, θ

κ

i

)

_

≤ u

i

(a

k

, θ

i

) −u

i

(a

k

,

¯

θ

i

), (5.11)

which shows that the sums that appear on the right hand side of (5.8) are

bounded from above, and therefore that V is well-deﬁned.

We now use the function V to construct the payment schemes that make

q dominant strategy incentive compatible. Indeed, we shall construct the

payment schemes so that the utility u

i

(q(θ), θ

i

) is exactly equal to V (θ) for

CHAPTER 5. DOMINANCE: GENERAL THEORY 106

all θ ∈ Θ. That is, we set:

u

i

(q(θ), θ

i

) −t

i

(θ) = V (θ) ⇔

t

i

(θ) = u

i

(q(θ), θ

i

)) −V (θ). (5.12)

To prove that this implies that the mechanism is dominant strategy incentive

compatible, we need to show:

u

i

(q(θ

i

, θ

−i

), θ

i

) −t

i

(θ

i

, θ

−i

) ≥ u

i

(q(θ

i

, θ

−i

), θ

i

) −t

i

(θ

i

, θ

−i

) ⇔

u

i

(q(θ

i

, θ

−i

), θ

i

) −(u

i

(q(θ

i

, θ

−i

), θ

i

)) −V (θ

i

, θ

−i

)) ≥

u

i

(q(θ

i

, θ

−i

), θ

i

) −

_

u

i

(q(θ

i

, θ

−i

), θ

i

)) −V (θ

i

, θ

−i

)

_

⇔

V (θ

i

, θ

−i

) ≥ V (θ

i

, θ

−i

) +u

i

(q(θ

i

, θ

−i

), θ

i

) −u

i

(q(θ

i

, θ

−i

), θ

i

)) (5.13)

This is true by the deﬁnition of V because the set of all ﬁnite sequences that

start with

¯

θ

i

and end with θ

i

includes the set of all ﬁnite sequences that

start with

¯

θ

i

, have θ

i

as their penultimate element, and end with θ

i

.

5.6 All Incentive Compatible Decision Rules When

Outcomes are Lotteries

We now develop a condition for implementability that is closely related to

cyclical monotonicity, but that applies only to the special case in which

the set of alternatives A is the set of probability distributions over some

ﬁnite set of outcomes. Denote the number of such outcomes by M. Each

element of A is of the form (p

1

, p

2

, . . . , p

M

), where p

≥ 0 for = 1, 2, . . . , M

and

M

=1

p

**= 1. The set of types of any player i, Θ
**

i

, consists of vectors

θ

i

= (θ

i,1

, θ

i,2

, . . . , θ

i,M

) of Bernoulli utilities of the M outcomes. We assume

that the set Θ

i

is convex for every i ∈ I.

Consider a given direct mechanism. For every i ∈ I deﬁne U

i

(θ) ≡

q(θ) θ − t

i

(θ). Here, “” stands for the scalar vector product. Hence, U

i

is the expected utility of player i if all players report their types truthfully.

Now suppose that the mechanism is dominant strategy incentive compatible,

and ﬁx some value of θ

−i

. Then U

i

as a function of θ

i

alone satisﬁes:

U

i

(θ

i

, θ

−i

) = max

θ

i

∈Θ

i

q(θ

i

, θ

−i

) θ

i

−t

i

(θ

i

, θ

−i

). (5.14)

Thus, U

i

is the maximum of linear functions, and therefore it is convex.

Now we review the following deﬁnition from Rockafellar (1970), p. 214:

A vector x

∗

is a subgradient of a convex function f at a point x if:

f(z) ≥ f(x) +x

∗

(z −x)

CHAPTER 5. DOMINANCE: GENERAL THEORY 107

for all z in the domain of f. In our case, q(θ

i

, θ

−i

) is a subgradient of

U(θ

i

, θ

−i

) (as a function of θ

i

) at θ

i

for every θ

i

. This is because:

U

i

(θ

i

, θ

−i

) ≥ U(θ

i

, θ

−i

) +q(θ

i

) (θ

i

−θ

i

) ⇔

U

i

(θ

i

, θ

−i

) ≥ q(θ

i

) θ

i

−t

i

(θ

i

, θ

−i

) +q(θ

i

) (θ

i

−θ

i

) ⇔

U

i

(θ

i

, θ

−i

) ≥ q(θ

i

) θ

i

−t

i

(θ

i

, θ

−i

) (5.15)

which holds by dominant strategy incentive compatibility for every θ

i

∈ Θ

i

.

The following result shows that being a subgradient of a convex function is

not only a necessary but also a suﬃcient condition for q to be implementable.

Proposition 5.5. Suppose that A is the set of all probability distributions

over some ﬁnite set of outcomes. Suppose that for every i ∈ I the set Θ

i

is a convex set of vectors of Bernoulli utility functions of player i. Then

a decision rule q is part of a dominant strategy incentive compatible direct

mechanism (q, t

1

, t

2

, ..., t

N

) if and only if for every i ∈ I and for every

θ

−i

∈ Θ

−i

there is a convex function U

i

: Θ

i

→ R such that q

i

(θ

i

, θ

−i

),

regarded as a function of θ

i

only, is a subgradient of U

i

.

Proof. The argument preceding Proposition 5.5 has shown the necessity of

the condition in Proposition 5.5. To see that it is also suﬃcient, ﬁx θ

−i

and

deﬁne for every θ

i

∈ Θ

i

player i’s transfer t

i

(θ

i

, θ

−i

) so that the function

U

i

of which q is the subgradient is exactly player i’s expected utility. The

equivalencies that precede Proposition 5.5 establish that dominant strategy

incentive compatibility holds.

Theorem 24.8 in Rockafellar (1970) establishes that q is cyclically mono-

tone if and only it is subgradient of a convex function. Therefore, Rochet’s

result that was cited in the previous section is a generalization of Rockafel-

lar’s result to the more general setting in which the sets of alternatives and

types are arbitrary.

5.7 Single Dimensional Type Spaces

We want to relate weak monotonicity in the sense of Deﬁnition 5.5 to mono-

tonicity as we have seen it in earlier parts of these notes. A characteris-

tic of these earlier parts was that agents’ type spaces were subsets of R,

and were therefore single-dimensional. Single-dimensionality is at ﬁrst sight

purely mathematical concept. In economic terms, however, it means that

type spaces can be completely ordered so that “higher types” have a larger

marginal willingness to pay for “higher alternatives.” Our purpose in this

CHAPTER 5. DOMINANCE: GENERAL THEORY 108

section is to study the implications of the single dimensionality condition

for dominant strategy incentive compatibility. Before we can do so we need

to formalize the notions of “higher alternatives” and “higher types.”

We begin by describing what “higher alternatives” are. Let R

i

be an

order of A. By this we mean that it is a complete, reﬂexive, and transitive

binary relation. The strict order derived from R

i

is denoted by P

i

: aP

i

b ⇔

[aR

i

b and not bR

i

a]. The indiﬀerence relation derived from R

i

is denoted

by I

i

: aI

i

b ⇔[aR

i

b and bR

i

a].

It is important not to mistake R

i

for agent i’s preference relation. To

understand this, consider the single unit auction example. In that example

we might take R

i

to be given by: aR

i

b ⇔[i obtains the good with at least as

high probability in a as in b]. The single crossing condition will not rule out

that some types prefer not obtaining the good over obtaining the good, i.e.

have a negative marginal willingness to pay for the good. Single crossing

will require that the larger types are the larger is the marginal willingness

to pay for the good.

Given an ordering of alternatives, we can now order types.

Deﬁnition 5.7. Consider any i ∈ I, and let R

i

be an order of A. For any

pair of types θ

i

, θ

i

∈ Θ

i

we shall say that θ

i

~

R

i

θ

i

(“θ

i

is a higher type than

θ

i

relative to R

i

” ) if

u

i

(a

, θ

i

) −u

i

(a, θ

i

) > u

i

(a

, θ

i

) −u

i

(a, θ

i

) holds for all a ∈ A with a

P

i

a

and

u

i

(a

, θ

i

) −u

i

(a, θ

i

) = u

i

(a

, θ

i

) −u

i

(a, θ

i

) = 0 holds for all a ∈ A with a

I

i

a.

Intuitively, θ

i

~

R

i

θ

i

means that θ

i

attaches larger marginal value to

higher alternatives than θ

i

for any two ordered alternatives that we com-

pare. In other words, θ

i

unambiguously has a stronger preference for higher

alternatives than θ

i

. What do we mean here by “higher alternatives”? This

is deﬁned by the ordering R

i

on A. Thus, the order ~

R

i

is conditional on

R

i

.

We are now going to use the order ~

R

i

to deﬁne the monotonicity prop-

erty that will allow us to relate weak monotonicity as in Deﬁnition 5.5 to

the monotonicity properties of earlier parts of these notes. Note that the

following deﬁnition applies to any arbitrary decision rules, irrespective of

whether the type spaces whose cross product makes up the domain of the

decision rule satisﬁes any “single dimensionality condition.”

CHAPTER 5. DOMINANCE: GENERAL THEORY 109

Deﬁnition 5.8. Consider any agent i, and let R

i

be an order of A. A

decision rule q is called “monotone with respect to R

i

” if:

θ

i

~

R

i

θ

i

⇒(q(θ

i

, θ

−i

)R

i

q(θ

i

, θ

−i

) for all θ

−i

∈ Θ

−i

.

In words, this deﬁnition says that a decision rules is “monotone with

respect to R

i

” if an unambiguous increase in an agent’s type leads to a

“higher” alternative being chosen. The next result shows that monotonicity

in this sense is implied by weak monotonicity.

Proposition 5.6. Consider any agent i, and let R

i

be an order of A. If a

decision rule q is weakly monotone then it is monotone with respect to R

i

.

A remarkable feature of Proposition 5.6 is that it is true for any order

R

i

of A.

Proof. Suppose q is weakly monotone. Consider some θ

−i

∈ Θ

−i

, and sup-

pose θ

i

~

R

i

θ

i

. Deﬁne a ≡ q(θ

i

, θ

−i

) and a

≡ q(θ

i

, θ

−i

). By the deﬁnition of

weak monotonicity, we must have: u

i

(a, θ

i

)−u

i

(a

, θ

i

) ≥ u

i

(a, θ

i

)−u

i

(a

, θ

i

).

We now prove indirectly that θ

i

~

P

i

θ

i

implies aR

i

a

. Suppose a

P

i

a. Then

θ

i

~

P

i

θ

i

implies

u

i

(a

, θ

i

) −u

i

(a, θ

i

) > u

i

(a

, θ

i

) −u

i

(a, θ

i

) ⇔

u

i

(a, θ

i

) −u

i

(a

, θ

i

) < u

i

(a, θ

i

) −u

i

(a

, θ

i

), (5.16)

which contradicts weak monotonicity.

The reason that we can’t prove in general the converse of this result is

that weak monotonicity imposes a restriction on collective decisions only in

the case that types are comparable in the order ~

R

i

, which is in general

incomplete. By contrast, monotonicity places restrictions on choices even if

types are not ordered in ~

R

i

. But we can strengthen Proposition 5.6 so that

it becomes an equivalence if we consider completely ordered type spaces.

Deﬁnition 5.9. Consider any agent i, and let R

i

be an order of A. The

type space Θ

i

is “single dimensional with respect to R

i

” if for any θ

i

, θ

i

∈ Θ

i

with θ

i

,= θ

i

we have either θ

i

~

R

i

θ

i

or θ

i

~

R

i

θ

i

or both.

Note that if a type is in this sense single dimensional we can assign to any

type a real number, namely the diﬀerence in utility for some alternatives, say

the diﬀerence u

i

(a

, θ

i

) −u

i

(a, θ

i

) where a

P

i

a, and this mapping from types

to real numbers will be invertible, that is this marginal utility unambiguously

CHAPTER 5. DOMINANCE: GENERAL THEORY 110

identiﬁes the type. Moreover, the larger this number, the larger all marginal

utilities of the type.

We can now show that for completely ordered, or single-dimensional,

type spaces monotonicity is suﬃcient for weak monotonicity.

Proposition 5.7. For every i ∈ I let R

i

be an ordering of A such that Θ

i

is single dimensional with respect to R

i

. Then a decision rule q is weakly

monotone if and only if it is monotone with respect to R

i

for every i ∈ I.

Proof. In the light of Proposition 5.6 we only have to show that monotonicity

implies weak monotonicity. So suppose q is monotone, and consider θ ∈ Θ

and θ

i

∈ Θ

i

such that q(θ) = a but q(θ

i

, θ

−i

) = a

**. Restrict attention to the
**

case a ,= a

**. Because θ satisﬁes the single crossing condition we must have:
**

θ

i

~

R

i

θ

i

or θ

i

~

R

i

θ

i

. Without loss of generality assume the former. Then,

because q is monotone, aR

i

a

. Moreover, because θ

i

~

R

i

θ

i

:

u

i

(a, θ

i

) −u

i

(a

, θ

i

) > u

i

(a, θ

i

) −u

i

(a

, θ

i

). (5.17)

if aP

i

a

, or

u

i

(a, θ

i

) −u

i

(a

, θ

i

) = u

i

(a, θ

i

) −u

i

(a

, θ

i

). (5.18)

if aI

i

a

**. This shows that q is weakly monotone.
**

Having established that weak monotonicity and monotonicity are equiv-

alent on single dimensional domains, we next show that monotonicity is

suﬃcient for an allocation rule to be implementable by appropriate transfer

schemes. This, together with Proposition 5.3, then implies that on sin-

gle dimensional domains monotonicity is necessary and suﬃcient for imple-

mentability.

For simplicity, we restrict our attention in the next result to the case that

A is ﬁnite. The case that A is inﬁnite is harder only in terms of notation. We

also assume that the type spaces are bounded in the sense of the following

deﬁnition:

Deﬁnition 5.10. For any agent i the type space Θ

i

is “bounded” if there

is a constant c > 0 such that for all a, a

∈ A and all θ

i

∈ Θ

i

we have:

−c < u

i

(a

, θ

i

) −u

i

(a, θ

i

) < c. (5.19)

Intuitively, the type space of agent i is “bounded” if there is a uniform

upper bound for agent i’s willingness to pay for a change in the decision.

CHAPTER 5. DOMINANCE: GENERAL THEORY 111

Proposition 5.8. Suppose that A is ﬁnite. For every i ∈ I let R

i

be an

order of A, and suppose that for every i ∈ I the type space Θ

i

is bounded

and is single dimensional with respect to R

i

. Let q be a decision rule that

is monotone with respect to R

i

for every i ∈ I. Then there are transfer

rules t

1

, t

2

, . . . , t

N

such that (q, t

1

, t

2

, . . . , t

N

) is dominant strategy incentive

compatible.

Proof. Fix i ∈ I and θ

−i

∈ Θ

−i

. Denote the range of q over θ

i

, that is the

set ¦q(θ

i

, θ

−i

) [ θ

i

∈ Θ

i

¦, by ¦a

1

, a

2

, . . . , a

n

¦ where a

n

R

i

a

n−1

R

i

. . . R

i

a

1

. For

simplicity we assume that all these preference relations are strict: a

n

P

i

a

n−1

P

i

. . . P

i

a

1

. If some are not, then the proof below should be modiﬁed treat-

ing alternatives between which the decision maker is indiﬀerent as identical

alternatives.

If n = 1 we set t

i

(θ

i

, θ

−i

) = 0 for all θ

i

. This will obviously make it

optimal for player i to report her type truthfully if the other agents’ types

is θ

−i

because player i will be indiﬀerent between all possible reports.

If n ≥ 2 deﬁne for every k = 1, 2, . . . , n the set

Θ

k

i

≡ ¦θ

i

∈ Θ

i

[ q(θ

i

, θ

−i

) = a

k

¦. (5.20)

Monotonicity and the single crossing condition imply that the sets Θ

K

i

are

ordered in the following sense:

k

> k, θ

k

i

∈ Θ

k

i

, θ

k

i

∈ Θ

k

i

⇒θ

k

i

~

P

i

θ

k

i

. (5.21)

Deﬁne for every k = 2, . . . , n:

τ

k

≡ inf¦u

i

(a

k

, θ

i

) −u

i

(a

k−1

, θ

i

) [ θ

i

∈ Θ

k

i

¦. (5.22)

The inﬁmum here is well-deﬁned because Θ

i

is bounded. Note that the

ordering of types implies:

k

< k, θ

i

∈ Θ

k

i

⇒u

i

(a

k

, θ

i

) −u

i

(a

k−1

, θ

i

) ≤ τ

k

k

> k, θ

i

∈ Θ

k

i

⇒u

i

(a

k

, θ

i

) −u

i

(a

k−1

, θ

i

) ≥ τ

k

. (5.23)

We deﬁne agent i’s transfer payment as follows:

t

i

(θ) =

_

¸

_

¸

_

0 if θ

i

∈ Θ

1

i

;

k

κ=2

τ

k

if θ

i

∈ Θ

k

i

where k ≥ 2.

(5.24)

CHAPTER 5. DOMINANCE: GENERAL THEORY 112

We verify that this transfer scheme makes truthful reporting of θ

i

is an

optimal strategy for agent i given any θ

−i

. If agent i’s true type is θ

i

∈ Θ

k

i

,

and if she reports any type in Θ

k

i

her utility will be independent of her

report. If she reports a type in Θ

k

i

where k

**> k, then the change in her
**

utility in comparison to truthful reporting will be:

u

i

(a

k

, θ

i

) −u

i

(a

k

, θ

i

) −

k

κ=k+1

τ

κ

=

k

κ=k+1

_

u

i

(a

κ

, θ

i

) −u

i

(a

κ−1

, θ

i

)

_

−

k

κ=k+1

τ

k

≤

k

κ=k+1

_

u

i

(a

κ

, θ

i

) −u

i

(a

κ−1

, θ

i

)

_

−

k

κ=k+1

_

u

i

(a

κ

, θ

i

) −u

i

(a

κ−1

, θ

i

)

_

= 0

(5.25)

The ﬁrst equality is a simple re-writing. The inequality follows from the

ﬁrst inequality in (5.23). A symmetric argument proves there is no incentive

for agent i to report a type in Θ

k

i

where k

< k.

Proposition 8 is analogous to the result that we obtained in earlier sec-

tions that monotonicity of q is necessary and suﬃcient for implementability

of q. In all those models, the domain of preferences satisﬁed a single dimen-

sionality condition (although we allowed inﬁnite outcome sets).

5.8 Suﬃciency of Weak Monotonicity

The single crossing domains described in Section 5.7 are not the only do-

mains on which one can show that weak monotonicity is suﬃcient for imple-

mentability, and is thus equivalent to cyclical monotonicity. Bikhchandani

et al. (2006) have found several other conditions under which this result is

true. We explain here the simplest one. Bikhchandani et. al. say that the

set Θ

i

of types of agent i is “rich” if there is some reﬂexive and transitive, but

possibly incomplete, binary relation R

i

on A such that all utility functions

that represent R

i

are possible utility functions of agent i. Formally:

Deﬁnition 5.11. Consider any i ∈ I. The type space Θ

i

is called “rich”

if there is a possibly incomplete preference relation R

i

on A such that for

every u : A →R that represents R

i

, i.e. that satisﬁes: aR

i

b ⇒u(a) ≥ u(b),

there is a θ

i

∈ Θ

i

such that

u

i

(a, θ

i

) = u(a) for all a ∈ A.

CHAPTER 5. DOMINANCE: GENERAL THEORY 113

Note that this condition becomes more restrictive as R

i

becomes less

complete, i.e. as comparisons are dropped from R

i

. The reason is that the

less complete R

i

is, the more utility functions may represent R

i

.

Bikhchandani et. al.’s Theorem 1 is:

Proposition 5.9. Suppose that A is ﬁnite. Suppose that for every i ∈ I the

type space Θ

i

is rich. Let q be a weakly monotone decision rule. Then there

are transfer schemes t

1

, t

2

, . . . , t

N

such that (q, t

1

, t

2

, . . . , t

N

) is dominant

strategy incentive compatible.

Bikhchandani et. al. emphasize that their result does not apply to the

case that A is a set of lotteries over outcomes because this would make the

set A inﬁnite.

5.9 Positive Association of Diﬀerences

We now consider another condition that is weaker than monotonicity, called

“positive association of diﬀerences” (PAD). We are considering this rela-

tively weak condition because it is suﬃcient to obtain a surprisingly strong

result.

Deﬁnition 5.12. A decision rule q satisﬁes “positive association of diﬀer-

ences” (PAD) if θ, θ

∈ Θ, q(θ) = a, and

u

i

(a, θ

i

) −u

i

(b, θ

i

) > u

i

(a, θ

i

) −u

i

(b, θ

i

)

for all i ∈ I and b ∈ A with b ,= a, implies q(θ

) = a.

In words, q satisﬁes PAD if whenever an alternative that is chosen at

some type vector will also be chosen at any other type vector where the

alternative’s marginal utilities in comparison to other alternatives are larger.

Proposition 5.10. If q is weakly monotone then it satisﬁes PAD.

Proof. Suppose that θ, θ

∈ Θ, q(θ) = a and that θ

**satisﬁes the condition
**

in the deﬁnition of PAD. We shall prove that then q(θ

i

, θ

−i

) = a for all

i ∈ I. The assertion that we have to prove, q(θ

**) = a, then follows from the
**

N-fold application of the same argument. Suppose q(θ

i

, θ

−i

) = b ,= a. By

the conditions on θ and θ

we have:

u

i

(a, θ

i

) −u

i

(b, θ

i

) > u

i

(a, θ

i

) −u

i

(b, θ

i

) ⇔

u

i

(b, θ

i

) −u

i

(a, θ

i

) < u

i

(b, θ

i

) −u

i

(a, θ

i

).

Thus, we have a contradiction with weak monotonicity.

CHAPTER 5. DOMINANCE: GENERAL THEORY 114

PAD is weaker than monotonicity because it only puts restrictions on

type proﬁles which result in the same collective decision. By contrast, mono-

tonicity also refers to type proﬁles which result in diﬀerent collective deci-

sions. However, Roberts (1979) showed that if the domain of a decision rule

consists of all possible utility functions, and if some other conditions hold,

then “Positive Association of Diﬀerences” is suﬃcient for dominant strategy

incentive compatibility.

One of the further conditions needed to state Roberts’ result is that the

decision rule q is ﬂexible:

Deﬁnition 5.13. A decision rule q is called “ﬂexible” if its range, q(Θ), has

at least three elements.

Observe that a decision rule can only be ﬂexible if #A ≥ 3. We shall

also assume that A is ﬁnite.

Proposition 5.11. Suppose A is ﬁnite, and suppose for every i ∈ I and

ν ∈ R

#A

there is a θ

i

∈ Θ

i

such that (u

i

(a, θ

i

))

a∈A

= ν. For every ﬂexible

decision rule q that satisﬁes PAD there are transfer rules (t

1

, t

2

, . . . , t

N

) such

that (q, t

1

, t

2

, . . . , t

N

) is dominant strategy incentive compatible.

Note that this result, by assuming A to be ﬁnite, rules again out the

case in which A is the set of all lotteries over some ﬁnite set of outcomes.

Also, the condition for Θ

i

on which this result relies is very restrictive. It

rules out that the direct mechanism that we are constructing embodies any

prior knowledge about the agents’ preferences.

Roberts proved his result by obtaining an interesting characterization of

all decision rules that satisfy the conditions of Proposition 5.11. Although

we shall not provide a proof of this characterization, we mention it, and

show how it implies Proposition 5.11.

Proposition 5.12. Suppose A is ﬁnite, and suppose for every i ∈ I and

ν ∈ R

#A

there is a θ

i

∈ Θ

i

such that (u

i

(a, θ

i

))

a∈A

= ν. Then a ﬂexible

decision rule q satisﬁes PAD if and only if for every i ∈ I there is a real

number k

i

> 0, and there is a function F : A →R such that for every θ ∈ Θ:

N

i=1

k

i

u

i

(q(θ), θ

i

) +F(q(θ)) ≥

N

i=1

k

i

u

i

(a, θ

i

) +F(a)

for all a ∈ A.

CHAPTER 5. DOMINANCE: GENERAL THEORY 115

In words, this result says that under the assumptions of the result a

decision rule satisﬁes PAD if and only if it maximizes a weighted utilitarian

welfare criterion with exogenous bias. The weight of agent i’s utility under

the utilitarian welfare criterion is k

i

. The bias is described by the function

F. This function assigns to each alternative a measure of welfare that is

independent of agents’ types. For example, F could pick out some particular

alternative ¯ a ∈ A as the “status quo”, and set F(¯ a) = z > 0 and F(a) = 0

for all a ∈ A with a ,= z. Then any alternative other than ¯ a would have to

imply social welfare that exceeds that of the status quo by at least z if it is

to be preferred over the status quo.

We now show that Proposition 5.12 implies Proposition 5.11. It is obvi-

ous from Propositions 5.3 and 5.10 that PAD is necessary for implementabil-

ity. All that is needed to derive Proposition 5.11 from Proposition 5.12 is

therefore to show the following:

Proposition 5.13. Suppose q satisﬁes the characterization in Proposition

5.12. Then there are transfer rules (t

1

, t

2

, . . . , t

N

) such that (q, t

1

, t

2

, . . . , t

N

)

is dominant strategy incentive compatible.

Proof. This follows from a generalization of the VCG construction. We can

deﬁne agent i’s transfer payment as follows:

t

i

(θ) = −

1

k

i

_

_

j=i

k

j

u

j

(q(θ), θ

j

) +F(q(θ))

_

_

. (5.26)

Here, in comparison to the VCG formula, we have omitted all terms not

depending on agent i’s report. Those terms don’t aﬀect the argument. Agent

i’s utility when she is type θ

i

, she reports being type θ

i

, and all other agents

report type vector θ

−i

is:

u

i

(q(θ

i

, θ

−i

), θ

i

) +

1

k

i

_

_

j=i

k

j

u

j

(q(θ

i

, θ

−i

), θ

j

) +F(q(θ

i

, θ

−i

))

_

_

. (5.27)

Maximizing this expression is equivalent to maximizing the product of this

expression and k

i

. This product is:

k

i

u

i

(q(θ

i

, θ

−i

), θ

i

) +

j=i

k

j

u

j

(q(θ

i

, θ

−i

), θ

j

) +F(q(θ

i

, θ

−i

))

=

n

j=1

k

j

u

j

(q(θ

i

, θ

−i

), θ

j

) +F(q(θ

i

, θ

−i

)) (5.28)

CHAPTER 5. DOMINANCE: GENERAL THEORY 116

and hence agent i chooses his report to maximize the same function that q

maximizes. Therefore, reporting θ

i

truthfully is optimal.

5.10 Individual Rationality and Budget Balance

We now enrich our framework to bring in individual rationality and budget

balance. We begin with individual rationality. In the examples that we

have seen in earlier parts of these notes, individual rationality required that

if agents report truthfully their types they are at least as well oﬀ as they

would be if they obtained some particular alternative in A and had to pay

nothing. In the auction example, the alternative in A that corresponded to

individual rationality was “not obtaining the good”. In the public goods

example it was “no production of the public good.” In the bilateral trade

example it was “no trade.” To generalize, we shall assume that for every

agent i some alternative a

i

∈ A is given that is as good as agent i’s outside

option, and thus that determines his individual rationality constraint.

Deﬁnition 5.14. For every agent i let a

i

∈ A. A direct mechanism is “ex

post individually rational with respect to (a

1

, a

2

, . . . , a

N

)” if for all i ∈ I

and all θ ∈ Θ we have:

u

i

(q(θ), θ

i

) −t

i

(θ) ≥ u

i

(a

i

, θ

i

).

In the examples in previous parts of these notes, for incentive compatible

mechanisms, individual rationality was true for all types if and only if it was

true for the lowest (in the case of the seller in the bilateral trade model: the

highest) type. To obtain a result of this kind more generally, we impose the

single crossing condition, and we assume that a

i

is lowest ranked in A in the

ordering P

i

that reﬂects agent i’s order of the elements of A.

Proposition 5.14. Consider any agent i ∈ I, and let R

i

be an order of A.

Suppose that the type set Θ

i

is one-dimensional with respect to R

i

. Assume

that there is a type θ

i

that is the lowest type in the order ~

R

i

in Θ

i

, i.e.

θ

i

~

R

i

θ

i

for all θ

i

∈ Θ

i

such that θ

i

,= θ

i

. Finally, assume that a

i

is the

lowest element of A in the order R

i

, i.e. bR

i

a

i

for every b ∈ A with b ,= a.

Then a dominant strategy incentive compatible mechanism satisﬁes the ex

post individually rationality constraint for agent i if and only if for every

θ

−i

∈ Θ

−i

:

u

i

(q(θ

i

, θ

−i

), θ

i

) −t

i

(θ

i

, θ

−i

) ≥ u

i

(a

i

, θ

i

).

CHAPTER 5. DOMINANCE: GENERAL THEORY 117

Proof. The condition in Proposition 5.14 is the individual rationality con-

straint for the lowest type. It is therefore necessary. To show that it is also

suﬃcient, we show that it implies individual rationality for any type θ

i

,= θ

i

.

We ﬁrst note that:

u

i

(q(θ

i

, θ

−i

), θ

i

) −t

i

(θ

i

, θ

−i

) ≥ u

i

(a

i

, θ

i

) ⇔

u

i

(q(θ

i

, θ

−i

), θ

i

) −u

i

(a

i

, θ

i

) ≥ t

i

(θ

i

, θ

−i

). (5.29)

Now if θ

i

,= θ

i

, then by the single crossing condition: θ

i

~

R

i

θ

i

. Be-

cause, moreover, the alternative q(θ

i

, θ

−i

) either satisﬁes: q(θ

i

, θ

−i

)P

i

a

i

, or:

q(θ

i

, θ

−i

)I

i

a

i

we can infer:

u

i

(q(θ

i

, θ

−i

), θ

i

) −u

i

(a

i

, θ

i

) ≥ t

i

(θ

i

, θ

−i

) ⇔

u

i

(q(θ

i

, θ

−i

), θ

i

) −t

i

(θ

i

, θ

−i

) ≥ u

i

(a

i

, θ

i

). (5.30)

By dominant strategy incentive compatibility:

u

i

(q(θ

i

, θ

−i

), θ

i

) −t

i

(θ

i

, θ

−i

) ≥ u

i

(q(θ

i

, θ

−i

), θ

i

) −t

i

(θ

i

, θ

−i

). (5.31)

Combining the last two inequalities we obtain:

u

i

(q(θ

i

, θ

−i

), θ

i

) −t

i

(θ

i

, θ

−i

) ≥ u

i

(a

i

, θ

i

), (5.32)

which is ex post individual rationality for type θ

i

.

We turn next to budget balance. Without loss of generality we deﬁne

budget balance as the requirement that all transfer payments add up to

zero. If alternatives in A have costs associated with them, we can re-deﬁne

outcomes so that some division, say equal division, of these costs is already

included in outcomes.

Deﬁnition 5.15. A direct mechanism is “ex post budget balanced” if for

all θ ∈ Θ we have:

N

i=1

t

i

(θ) = 0.

We now investigate conditions under which eﬃcient decision rules can

be implemented with a balanced budget. Because under certain conditions

VCG mechanisms are the only mechanisms that implement eﬃcient decision

rules, it is relevant to ask when VCG mechanisms are budget balanced.

The following Proposition provides the answer. It shows that a necessary

and suﬃcient condition is a restriction on the functional form of the welfare

generated by an eﬃcient decision rule if this welfare is regarded as a function

of the type vector.

CHAPTER 5. DOMINANCE: GENERAL THEORY 118

Proposition 5.15. Let q be an eﬃcient decision rule. Then there exist a

balanced budget VCG mechanism that implements q if and only if for every

i ∈ I there is a function f

i

: Θ

−i

→R such that:

N

i=1

u

i

(q(θ), θ

i

) =

N

i=1

f

i

(θ

−i

) for all θ ∈ Θ.

Proof. We ﬁrst prove the necessity of this condition. Consider a VCG mech-

anism, and let τ

i

: Θ

i

→ R be the functions referred to in Deﬁnition 5.4.

The mechanism is budget balanced if for all θ ∈ Θ:

N

i=1

_

_

−

j=i

u

j

(q(θ), θ

j

) +τ

i

(θ

−i

)

_

_

= 0 ⇔

(N −1)

N

i=1

u

i

(q(θ), θ

i

) =

N

i=1

τ

i

(θ

−i

) ⇔

N

i=1

u

i

(q(θ), θ

i

) =

N

i=1

τ

i

(θ

−i

)

N −1

(5.33)

Hence, if we set for every i ∈ I and θ

−i

∈ Θ

−i

f

i

(θ

−i

) ≡

τ

i

(θ

−i

)

N −1

(5.34)

we have obtained the desired form for the function

N

i=1

u

i

(q(θ), θ

i

).

Next we prove suﬃciency of the condition. Suppose that

N

i=1

u

i

(q(θ), θ

i

)

has the form described in the Proposition. For every i ∈ I and every

θ

−i

∈ Θ

−i

we consider the VCG mechanism with

τ

i

(θ

−i

) ≡ (N −1)f

i

(θ

−i

). (5.35)

Then for every θ ∈ Θ the sum of agents’ payments is:

N

i=1

_

_

−

j=i

u

j

(q(θ), θ

j

) + (N −1)f(θ

−i

)

_

_

=−(N −1)

N

i=1

u

i

(q(θ), θ

i

) + (N −1)

N

i=1

f(θ

−i

) (5.36)

which is zero by the assumption of the Proposition.

CHAPTER 5. DOMINANCE: GENERAL THEORY 119

As an application of this result we consider the bilateral trade model. In

that model the maximized welfare is given by: max¦θ

S

, θ

B

¦. The condition

of the above Proposition is that there are functions f

B

and f

S

such that

max¦θ

S

, θ

B

¦ = f

B

(θ

S

) + f

S

(θ

B

). Now suppose that the intersection of the

intervals [θ

S

,

¯

θ

S

] and [θ

B

,

¯

θ

B

] has a non-empty interior, and suppose that θ, θ

are two diﬀerent elements of that interior with θ < θ

**. Then the necessary
**

and suﬃcient condition in Proposition 5.15 requires:

θ = max¦θ, θ¦ = f

B

(θ) +f

S

(θ)

θ

= max¦θ, θ

¦ = f

B

(θ) +f

S

(θ

)

θ

= max¦θ

, θ¦ = f

B

(θ

) +f

S

(θ)

θ

= max¦θ

, θ

¦ = f

B

(θ

) +f

S

(θ

) (5.37)

Subtracting the third from the fourth equality in (5.37) we ﬁnd that f

S

(θ

) =

f

S

(θ), and subtracting the second from the fourth equality in (5.37) we ﬁnd

that f

B

(θ

) = f

B

(θ). But then the ﬁrst right hand side in (5.37) has to be the

same as the three other right hand sides, which contradicts the assumption

that the ﬁrst right hand side equals θ whereas all later right hand sides equal

θ

**. There is therefore no budget balanced VCG mechanism in this example.
**

Whenever Corollary 5.1 holds, then Proposition 5.15 gives not only nec-

essary and suﬃcient conditions for the existence of a budget balanced VCG

mechanism, but necessary and suﬃcient conditions for the existence of any

budget balanced dominant strategy incentive compatible mechanism that

implements eﬃcient decisions. This is because any such mechanism has to

be a VCG mechanism by Corollary 5.1. This observation applies to the

bilateral trade model.

Note that the impossibility result that follows from Proposition 5.15 for

the bilateral trade model does not rely on an individual rationality con-

straint. It is therefore not an implication of Proposition 4.12 which shows

that only ﬁxed price mechanisms are dominant strategy incentive compati-

ble, ex post budget balanced and individually rational.

In some of the proofs we have worked with Groves mechanisms and have

argued that all such mechanisms, if they are individually rational, have a

zero budget balance, or a negative budget balance, in all states of the world.

I am not aware of any result that yields this conclusion for a more general

class of models.

CHAPTER 5. DOMINANCE: GENERAL THEORY 120

5.11 Remarks on the Literature

The payoﬀ equivalence result that we presented is taken from Krishna and

Maenner (2001). The classic papers on Vickrey-Clarke-Groves mechanisms

are Clarke (1971), Groves (1973) and Vickrey (1961). The uniqueness of

VCG mechanisms in the sense of Corollary 1 was shown in Green and Laﬀont

(1977) and Holmstr¨om (1979). The result on cyclical monotonicity is from

Rochet (1987). However, the concept of cyclical monotonicity is originally

due to Rockafellar (1970). The discussion of the case that outcomes are

lotteries is also taken from Rochet’s 1987 paper. The result on the suﬃciency

of weak monotonicity is due to Bikhchandani et al. (2006). The concept of

“positive association of diﬀerences”, and its characterization are in Roberts

(1979). Finally, I have taken Proposition 5.15 from Milgrom (2004), p. 54,

where this result is attributed to Bengt Holmstr¨om’s 1977 Stanford PhD

thesis.

5.12 Problems

Chapter 6

Bayesian Mechanism Design:

General Theory

.

6.1 Introduction

To generalize our earlier results on Bayesian mechanism design, we need

to take two steps. First, as in the last chapter, we need to consider more

general sets of alternatives and a more general speciﬁcation of types and

preferences. Second, we need to consider more general distributions of types.

This chapter will correspondingly have two parts. The ﬁrst part, dealing

with more general sets of alternatives, types, and preferences, will build on

the previous chapter, and will therefore be short and in parts informal. The

second part, dealing with correlated types, will be longer, and the nature of

the analysis will be quite diﬀerent from what we have seen before.

In practice, it seems plausible that types are often not independent. For

example, if one agent has private information that makes the agent value

the object sold in an auction particularly highly, then this agent might think

that it is likely that other agents also have information that makes them

value the object highly. This motivates the study of a model of Bayesian

mechanism design in which types are not independent.

The results in the second part of this chapter will indicate that with de-

pendent types almost all combinations of decision rules and transfer rules

can be implemented by a Bayesian incentive compatible direct mechanism.

This result is surprising, and it does not seem plausible in practice. The

121

CHAPTER 6. BAYESIAN DESIGN: GENERAL THEORY 122

result is best viewed as a paradox. It clearly indicates that some ingredient

is missing in the model. In Section 6.5 below we shall begin a discussion of

what this missing ingredient might be.

6.2 Set-Up

There are N agents. The set of agents is denoted by I = ¦1, 2, . . . , N¦. They

have to choose an alternative a out of some set A of mutually exclusive

alternatives. Agent i’s utility if alternative a is chosen, and agent i pays

transfer t

i

is:

u

i

(a, θ

i

) −t

i

.

Here, θ

i

is agent i’s type.

We shall employ similar notation as before: The set of possible types of

agent i is Θ

i

. We denote by θ the vector of types: (θ

1

, θ

2

, . . . , θ

N

). The set

of all possible type vectors is Θ ≡ Θ

1

Θ

2

. . . Θ

N

. Finally, we write θ

−i

for a vector θ of types if we leave out agent i’s type. The set of all θ

−i

is

Θ

−i

which is the cartesian product of the sets Θ

j

, leaving out Θ

i

.

The sets Θ

i

and Θ can, in principle, be abstract sets, but we want to be able

to deﬁne a probability measure on them. For this, we can think of them

as abstract measurable spaces, or, for concreteness, as products of intervals

in ﬁnite dimensional Euclidean space, or as ﬁnite sets. We assume that

there is a common prior distribution µ on Θ that is shared by the agents

and the principal. We write µ

i

for the marginal probability on Θ

i

, and we

write µ( [ θ

i

) for the conditional probability distribution of θ

−i

given θ

i

.

Note that µ is not only the common prior of the agents but it is also the

mechanism designer’s belief.

We assume that the mechanism designer proposes to the agents a game and

a Bayesian equilibrium of that game. The revelation principle applies, and

we can restrict attention to direct mechanisms.

Deﬁnition 6.1. A direct mechanism (q, t

1

, t

2

, . . . , t

N

) consists of a mapping

q : Θ →A,

that maps every type vector into a collective decision, and mappings

t

i

: Θ →R,

one for each player i ∈ I, that indicate for each type vector the transfer that

agent i needs to make.

CHAPTER 6. BAYESIAN DESIGN: GENERAL THEORY 123

We call q the “decision rule” and the functions t

i

the “payment rules.”

For a given direct mechanism we shall write Q

i

(θ

i

) for agent i’s interim

probability distribution on A conditional on agent i’s type being θ

i

. That

is, if / is a measurable subset of A, then Q

i

(θ

i

) assigns to / the probability

Q

i

(/ [ θ

i

) ≡

_

Θ

−i

1

A

dµ(θ

−i

[ θ

i

) (6.1)

where 1

A

: Θ

−i

→ ¦0, 1¦ is the indicator function that assigns 1 to a type

vector θ

−i

if and only if q(θ

i

, θ

−i

) ∈ A. We also write T

i

(θ

i

) for agent i’s

interim expected transfer payment conditional on agent i’s type being θ

i

:

T

i

(θ

i

) ≡

_

Θ

−i

t

i

(θ

i

, θ

−i

)dµ(θ

−i

[ θ

i

). (6.2)

If a direct mechanism is derived from a Bayesian equilibrium of some indirect

mechanism, then truth-telling will be a Bayesian equilibrium of the direct

mechanism. In that case we call the direct mechanism Bayesian incentive

compatible.

Deﬁnition 6.2. A direct mechanism (q, t

1

, t

2

, . . . , t

N

) is “Bayesian incentive

compatible” if for all θ ∈ Θ, all i ∈ I and all θ

i

∈ Θ

i

we have:

_

A

u

i

(a, θ

i

)dQ

i

(θ

i

) −T

i

(θ

i

) ≥

_

A

u

i

(a, θ

i

)dQ

i

(θ

i

) −T

i

(θ

i

).

Our ﬁst concern will be with the characterization of Bayesian incentive com-

patible decision rules. We shall later bring in additional considerations, such

as individual rationality and budget balance.

6.3 Independent Types

We begin with a statement of the revenue equivalence theorem. As in the

previous chapter we use the formulation of Krishna and Maenner (2001),

now adapted to the Bayesian setting. We make the same assumptions as in

the context of Proposition 5.1 in Chapter 5. The result then says that the

interim expected payment rules are uniquely determined up to a constant

by the interim decision rules. The logic behind the result below is the same

as the logic behind all other revenue equivalence results presented in these

notes.

CHAPTER 6. BAYESIAN DESIGN: GENERAL THEORY 124

Proposition 6.1. Suppose that for every i ∈ I the set Θ

i

is a convex sub-

set of a ﬁnite dimensional Euclidean space. Moreover, assume that for ev-

ery i ∈ I the function u

i

(a, θ

i

) is a convex function of θ

i

. Suppose that

(q, t

1

, t

2

, . . . , t

N

) is a Bayesian incentive compatible mechanism with interim

decision rules Q

i

(θ

i

) and interim expected payments T

i

(θ

i

) for every i ∈ I

and every θ

i

. Let (q

, t

1

, t

2

, . . . , t

N

) be another Bayesian incentive compatible

mechanism with interim decision rules Q

i

(θ

i

) and interim expected payments

T

i

(θ

i

) for every i ∈ I and every θ

i

. Suppose that

Q

i

(θ

i

) = Q

i

(θ

i

)

for some i ∈ I and every θ

i

∈ Θ

i

. Then for every agent i ∈ I there is a

number τ

i

∈ R such that

T

i

(θ

i

) = T

i

(θ

i

) +τ

i

for the same i ∈ I and every θ

i

∈ Θ

i

.

We now focus on decision rules q that can be implemented in Bayesian incen-

tive compatible direct mechanisms. A ﬁrst point to note is that all dominant

strategy incentive compatible mechanisms are also Bayesian incentive com-

patible, and therefore VCG mechanisms are Bayesian incentive compatible

mechanisms implementing eﬃcient decisions.

To obtain characterizations of all incentive compatible decision rules, eﬃ-

cient or not, we can translate the results of the previous chapter into our

current setting. These characterizations now hold at the interim, not at the

ex post level. To translate these results, we treat the set of all probability

distributions over A as the set of alternatives. The decision rule Q

i

assigns

to every type θ

i

of player i a probability distribution Q

i

(θ

i

) over A. Agent

i evaluates alternatives by calculating their expected utility. Finally, agent

i’s payment rule is given by T

i

. As an example we adapt Proposition 5.4 to

our setting.

Proposition 6.2. A decision rule q is part of a Bayesian incentive compat-

ible direct mechanism (q, t

1

, t

2

, . . . , t

N

) if and only if q is interim cyclically

monotone, that is, for every i ∈ I, for every sequence of length k ∈ N of

types of agent i, (θ

1

i

, θ

2

i

, . . . , θ

k

i

) ∈ Θ

k

i

, with θ

k

i

= θ

1

i

, we have:

k−1

κ=1

_

_

_

A

u

i

(a, θ

κ+1

i

)dQ

i

(θ

κ

i

) −

_

A

u

i

(a, θ

κ

i

)dQ

i

(θ

κ

i

)

_

_

≤ 0.

CHAPTER 6. BAYESIAN DESIGN: GENERAL THEORY 125

In Chapter 5 we considered as a special case that the type space is single

dimensional. In this special case monotonicity of the decision rule q was

shown to be necessary and suﬃcient for the existence of a dominant strategy

incentive compatible mechanism that implements q. A translation of these

results into our setting requires that for every agent i ∈ I the type set Θ

i

is

single dimensional where the set of alternatives is the set of all probability

distributions over Q. The only cases where this is the case seems to be cases

where A consists of only two alternatives. Therefore, there do not seem to

be more general results here that go signiﬁcantly beyond what we showed

in Chapter 3.

Turning to individual rationality and budget balance, it again seems that the

result on individual rationality with single dimensional type spaces, Proposi-

tion 5.14, has no interesting generalization. Concerning budget balance, the

existence of an ex post budget balanced dominant strategy mechanism is, of

course, suﬃcient for the existence of a budget balanced Bayesian incentive

compatible mechanism. Thus, Proposition 5.15 provides a suﬃcient condi-

tion for the existence of an eﬃcient Bayesian incentive compatible mecha-

nism.

Combining individual rationality and budget balance, if all individually ra-

tional and ex post budget balanced mechanisms have, in all states of the

world, either a zero surplus, or a deﬁcit, then it follows that there is no

Bayesian incentive compatible, individually rational, and ex post budget

balanced mechanism either provided that the revenue equivalence result

Proposition 6.1 holds. This is the logic that we used to prove Propositions

3.7 and 3.12.

For second best considerations in settings in which eﬃcient rules cannot be

implemented, the equivalence between ex ante and ex post budget balance

that we used in Chapter 3 is useful, and it generalizes in a straightforward

way.

Deﬁnition 6.3. A direct mechanism (q, t

1

, t

2

, . . . , t

N

) is “ex ante budget

balanced” if

_

Θ

_

N

i=1

t

i

(θ)

_

dµ(θ) = 0.

The proof of the equivalence of ex ante and ex post budget balance, Propo-

sition 3.6, that we gave in Chapter 3 directly applies here. For convenience

we repeat the result.

CHAPTER 6. BAYESIAN DESIGN: GENERAL THEORY 126

Proposition 6.3. For every direct mechanism with decision rule q that is ex

ante budget balanced there is an equivalent direct mechanism with the same

decision rule q that is ex post budget balanced.

Here, two mechanisms are “equivalent” if the have the same decision rule,

and if for all agents i ∈ I and for all types θ

i

, θ

i

∈ Θ

i

agent i’s expected

transfers, conditional on agent i’s type being θ

i

and agent i reporting to be

type θ

i

, is the same in the two mechanisms.

6.4 Correlated Types

6.4.1 Framework

We now turn to the case that the distribution µ reﬂects some correlation

among types. Throughout this part of the notes, we shall assume that each

of the sets Θ

i

is ﬁnite. The results that we discuss in this part of the notes

are more easily proved with ﬁnite type spaces, although they are also true

with inﬁnite type spaces. Note that this is in contrast with earlier sections,

where results were more easily obtained with continuous type spaces rather

than discrete type spaces. Although the ﬁniteness assumption is thus, in

principle, innocuous, it might mislead us when we consider the question

what is true generically, and which cases are exceptional cases. We shall

return to this point below. We also assume that every θ ∈ Θ has positive

probability: µ(θ) > 0. This, too, simpliﬁes the exposition.

Observe that the deﬁnitions and results of Section 6.2 were not restricted to

the case of independent types, and therefore we can use them here. We re-

strict attention to Bayesian incentive compatible direct mechanisms. These

are deﬁned as in Section 6.2.

6.4.2 Failure of Revenue Equivalence

In the setup with independent types we have obtained that, under some

technical assumptions, the decision rule q determines the interim expected

transfer rules T

i

for each player i ∈ I up to a constant. This result, to which

we have referred as the “revenue equivalence” result, greatly simpliﬁed the

problem of describing the set of all incentive compatible mechanisms, and

the problem of identifying in this set those mechanisms that are optimal by

some criterion.

When types are not independent, then the revenue equivalence result is no

longer true. To understand why this is the case, let us look back for a mo-

CHAPTER 6. BAYESIAN DESIGN: GENERAL THEORY 127

ment at the independent types framework. The key argument that allowed

us to establish the payoﬀ equivalence result showed that the derivative of

any agent i’s interim expected utility with respect to agent i’s type depends

only on the choice rule, and not on the transfer rules. This point is crucial

because it implies that agent i’s interim expected utility function is deter-

mined up to a constant by the choice rule. This then easily implies, for

example, that the expected payments are determined up to a constant by

the choice rule.

How did we show that the derivative of interim expected utility with respect

to an agent’s type depends only on the choice rule? The argument for this

is as follows. By the envelope theorem, as we change agent i’s type, we

may take as given and ﬁxed agent i’s type report. A change in the agent’s

type then aﬀects the agent’s valuation of collective decisions, but it does not

aﬀect the agent’s valuation of expected transfers. Therefore, only collective

decisions, but not transfers, enter into the derivative.

Why is an agent’s expected transfer not aﬀected by a change in the agent’s

type if we take the agent’s type report as given and ﬁxed? There are two

reason for this is. The ﬁrst is that all types have the same utility function

of money, i.e. u(t

i

) = t

i

. The second reason is that all types have the same

conditional probability distribution over other agents’ type vectors. This is

important because it implies that as we change the agent’s type, the agent’s

expected transfer payment does not change, provided we keep the report

ﬁxed.

If types are not independent, this last part of the argument is no longer

valid. As we change an agent’s type, even if we keep the type report ﬁxed, the

agent’s conditional probability distribution over other agents’ types changes.

As the agent’s transfer payments may depend on other agents’ types, the

agent’s expected transfer payment may change. Thus, not only the decision

rule, but also the transfer rule, enters into the derivative of agent i’s interim

expected utility with respect to type. As a consequence, the same choice

rule may be incentive compatible in combination with transfer rules that

diﬀer at the interim level by more than an additive constant.

The fact that the payoﬀ equivalence result is not valid implies that we can-

not transform the problem of characterizing the incentive compatible di-

rect mechanisms into the problem of characterizing all incentive compatible

choice rules. We need to provide a joint characterization of incentive compat-

ible choice rules and transfer rules. A surprising and general characterization

is provided in the next subsection.

CHAPTER 6. BAYESIAN DESIGN: GENERAL THEORY 128

6.4.3 Characterizing Bayesian Incentive Compatibility

The result that we present in this section relies on a condition regarding

the distribution µ that we shall call the “Cr´emer-McLean condition” as it

originates in the work of Cr´emer and McLean (1988).

Deﬁnition 6.4. The probability distribution µ satisﬁes the Cr´emer-McLean

condition if there are no i ∈ I, θ

i

∈ Θ

i

and λ

i

: Θ

i

¸ ¦θ

i

¦ →R

+

for which:

µ(θ

−i

[ θ

i

) =

θ

i

∈Θ

i

θ

i

=θ

i

λ(θ

i

)µ(θ

−i

[ θ

i

) for all θ

−i

∈ Θ

−i

.

The content of this condition is best understood if one thinks of agent i’s

belief about the other agents’ types conditional on agent i’s type being θ

i

,

µ( [ θ

i

), as a vector with as many entries as Θ

−i

has elements. Agent i’s

conditional beliefs are described by a collection of vectors of this form, one

for each of agent i’s type. The Cr´emer-McLean condition requires that none

of these vectors can be written as a convex combination of all the other

vectors where the weights. Thus, none of these vectors is contained in the

convex hull of the other vectors. In the deﬁnition, the weights of the convex

combinations are denoted by λ(θ

i

).

The Cr´emer-McLean condition is obviously satisﬁed if the rank of the col-

lection of vectors that describe agent i’s conditional beliefs is equal to the

number of agent i’s types, and hence the vectors are linearly independent.

The Cr´emer-McLean condition is obviously violated if at least two of the

vectors that describe agent i’s conditional beliefs are identical. Thus, the

Cr´emer-McLean condition rules out that agent i’s conditional beliefs are

independent of his type, as is the case when types are independent.

Under the Cr´emer-McLean condition we can obtain the following surprising

result:

Proposition 6.4. Suppose that the distribution µ satisﬁes the Cr´emer-

McLean condition. Consider any direct mechanism (q, t). Then there is

an equivalent direct mechanism (q, t

**) that is Bayesian incentive compatible,
**

that is:

1. the two mechanisms have the same decision rule q;

2. the two mechanisms have the same interim expected payments:

θ

−i

∈Θ

−i

t

i

(θ

i

, θ

−i

)µ(θ

−i

[ θ

i

) =

θ

−i

∈Θ

−i

t

i

(θ

i

, θ

−i

)µ(θ

−i

[ θ

i

)

for all i ∈ I and θ

i

∈ Θ

i

.

CHAPTER 6. BAYESIAN DESIGN: GENERAL THEORY 129

In words, this result says that every direct mechanism can be made Bayesian

incentive compatible without altering the decision rule or interim expected

payments. This is a very permissive result.

As an application of Proposition 6.4 consider the single unit auction envi-

ronment. Let the decision rule be that the object is allocated to one of the

agents with the highest valuation of the agent, and let the payment rule

be that the winner of the object pays his true valuation, and that all other

agents pay nothing. If the auctioneer could implement this mechanism, he

would clearly obtain the largest possible revenue that he could extract from

the agents provided that he respects the agents’ individual rationality con-

straint. Proposition 6.4 says that, although this particular direct mechanism

is not necessarily Bayesian incentive compatible, one can adjust it so that

it does become incentive compatible without changing either the allocation

rule or the interim expected payments. If interim expected payments are

not altered, of course also the auctioneer’s expected revenue is not altered,

and hence the auctioneer can extract the entire surplus from trade. Put

diﬀerently, the auctioneer can achieve the same expected revenue as he can

if he directly observes agents’ valuations of the object. Agents earn no in-

formation rents.

The idea of the proof of Proposition 6.4 is to add to the transfer schemes of

the original mechanism (q, t) a transfer scheme that provides agents with in-

centives to truthfully reveal their beliefs. It is well-known that such incentive

schemes exist for risk-neutral subjects. Experimentalists, for example, use

such incentive schemes to elicit subjects’ beliefs about uncertain events, such

as other agents’ actions. Moreover, it is easy to see that in such incentive

schemes the costs for false reports of one’s beliefs can be made arbitrarily

large. As no two types have the same beliefs, an agent who truthfully re-

veals his beliefs also truthfully reveals his type. By making the incentives for

truthful revelation of beliefs very large we can undo all possible incentives

to lie about one’s type in the original mechanism (q, t).

An additional complication arises from the fact that the incentive scheme

for belief revelation needs to be such that truthful reports of beliefs generate

expected payments of zero, whereas false reports of beliefs generate negative

expected payments. This will ensure that the interim expected payments,

under truthful revelation of types, will remain unchanged in the transformed

mechanism. While incentive mechanisms for truthful belief revelation always

exist, the Cr´emer-McLean condition is used to ensure that the expected

payments from truth-telling can be set equal to zero.

CHAPTER 6. BAYESIAN DESIGN: GENERAL THEORY 130

How can truthful revelation of beliefs be induced? A very simple incentive

scheme is as follows. Suppose the ﬁnite set of possible outcomes is Ω, and

we want to elicit an experimental subject’s belief about the probabilities of

diﬀerent elements of Ω. We ask the subject to announce its probabilities π

for the elements of Ω, i.e. π : Ω → [0, 1], and we announce that if ω ∈ Ω

occurs, the subject makes a payment to the experimenter of k − ln(π(ω)).

Here, k is a constant. Thus, for example, if an ω ∈ Ω occurs to which the

subject has assigned an extremely small probability, the subject has to pay

to the experimenter a comparatively large amount.

Which probabilities will subjects announce? Suppose a subject is risk-

neutral. Let ˆ π be the subjects true subjective probabilities. In choosing

its report π the subject will minimize:

ω∈Ω

(k −ln(π(ω))ˆ π(ω) (6.3)

subject to

ω∈Ω

π(ω) = 1. (6.4)

A Lagrange function is:

L =

ω∈Ω

(k −ln(π(ω)))ˆ π(ω) +λ

_

ω∈Ω

ˆ π(ω) −1

_

. (6.5)

The ﬁrst order condition for maximizing this is:

∂L

∂π(ω)

= −

ˆ π(ω)

π(ω)

+λ = 0 ⇔

ˆ π(ω)

π(ω)

= λ. (6.6)

Thus, the reported vector of probabilities, π(ω), has to be proportional to

the true vector of probabilities, ˆ π(ω). There is, of course, only one reported

vector π that his property, namely the true vector of probabilities, ˆ π, itself.

Suppose we added to the transfer rule in the mechanism (q, t) in Proposition

6.4 payments that follow the above payment scheme, i.e. if agent i reports

to be type θ

i

, and if the agents other than i report type θ

−i

, then agent i has

to pay t

i

(θ

i

, θ

−i

) + c(k − ln µ(θ

−i

[ θ

i

)) where k and c > 0 are constants. If

we choose c suﬃciently large, then the incentive to report the probabilities

truthfully will override all other incentives that agent i might have, and the

mechanism will be Bayesian incentive compatible.

CHAPTER 6. BAYESIAN DESIGN: GENERAL THEORY 131

What we have just described is almost, but not quite the mechanism that

we will use in the proof of Proposition 6.4. The mechanism that we have

described obviously alters agents’ interim payments. However, the Cr´emer-

McLean condition is suﬃcient to allow us to provide strict incentives for

truthful revelation of beliefs using a mechanism where an agents’ expected

payments, if the agent reports his beliefs truthfully, are exactly zero. This

can be deduced from the Cr´emer-McLean condition using Farkas’ Alterna-

tive, Franklin (1980), p.56:

Proposition 6.5. Let A be an n m matrix. Either the equation

Ax = b has a solution x ≥ 0

or (exclusive)

y

T

A ≥ 0, y

T

b < 0 has a solution y.

To apply Farkas’ Alternative, we take some type θ

i

of agent i as given and

ﬁxed. We take b to be the the column vector of agent i’s beliefs if agent i

is type θ

i

, and we take A to be a matrix of dimensions #Θ

−i

(#Θ

i

− 1)

where each column of A represents agent i’s conditional beliefs if agent i

is of one of the types other than θ

i

. The “Cr´emer-McLean condition” says

that the ﬁrst of the two conditions in Farkas’ Alternative is not satisﬁed.

We deduce that the second condition holds. The column vector y, which

has #Ω

−i

entries, represents agent i’s payments if agent i reports type θ

i

,

and all other agents report to be type θ

−i

. The ﬁrst condition says that

agent i’s expected payment, conditional on any type other than type θ

−i

,

is non-negative. On the other hand, conditional on the type θ

i

, agent i’s

expected payments is negative, that is, he expects to receive a payment. By

subtracting a constant from all payments, we can achieve that conditional

on type θ

i

, agent i expects a zero payment, and conditional on all other

types he expects to make a strictly positive payment.

We can repeat the construction of the previous paragraph for each type θ

i

of agent i. Putting together all the transfer vectors y that we obtain in this

way, we construct a transfer scheme for agent i with the desired properties,

that incentives to report true beliefs are strict, and that expected payments

conditional on truth-telling are zero. We can add this transfer scheme, times

a positive constant, to agent i’s transfers in our original mechanisms (q, t),

and, if the constant is suﬃciently large, agent i’s incentives to report his

beliefs, and thus his type, truthfully, override all other incentives that agent

i might have in (q, t). We can proceed in a similar way for all agents i. This

completes the proof of Proposition 6.4.

CHAPTER 6. BAYESIAN DESIGN: GENERAL THEORY 132

1 2 3

1

4

20

2

20

1

20

2

2

20

2

20

2

20

3

1

20

2

20

4

20

Figure 6.1: Joint probability distribution of types.

1 2 3

1 (

1

2

,

1

2

) (0,1) (0,1)

2 (1.0) (

1

2

,

1

2

) (0,1)

3 (1,0) (1,0) (

1

2

,

1

2

)

Figure 6.2: Decision Rule

6.4.4 A Numerical Example

The numerical example in this Subsection underlines the general message of

this section that one should view the Cr´emer McLean result as a paradox

rather than a guidance to the construction of mechanisms that could work

in practice.

Example 6.1. We consider a single unit auction example. There are two

agents. Each of the two agents has three types: Θ

1

= Θ

2

= ¦1, 2, 3¦. An

agent’s type is the agent’s valuation for the good. The joint probability dis-

tribution of types is shown in Figure 6.1. In this ﬁgure, rows correspond to

types of player 1, and columns correspond to types of player 2.

Suppose that the auctioneer wants to allocate the object to the agent with the

highest valuation, where ties are broken randomly. Moreover, the auctioneer

would like the winner of the object to pay her reservation price, that is, her

type. This corresponds to the allocation and payment rules shown in Figures

6.2 and 6.3.

In the decision rule, the ﬁrst entry in each box indicates the probability

with which agent 1 obtains the object, and the second entry indicates the

probability with which player 2 obtains the object. In the payment rule, the

CHAPTER 6. BAYESIAN DESIGN: GENERAL THEORY 133

1 2 3

1 (

1

2

,

1

2

) (0,2) (0,3)

2 (2,0) (1, 1) (0,3)

3 (3,0) (3,0) (

3

2

,

3

2

)

Figure 6.3: Payment Rule

1 2 3

1 -1 1 2

2 1 -2 1

3 2 1 -1

Figure 6.4: Belief revelation

ﬁrst entry in each box shows the payment by player 1, and the second entry

shows the payment by player 2.

The rule that we have described is clearly not incentive compatible. Types 2

and 3 of each player have an incentive to pretend that they are lower types.

That will reduce the probability with which they get the object, but at least it

will give them a positive surplus if they obtain the object. We now want to

demonstrate Proposition 1 for this example, and construct a payment rule

that leaves interim payments unchanged, and that makes the decision rule

incentive compatible.

As explained in Subsection 6.4.3, we start by constructing transfer rules that

give each player strict incentives to reveal their true beliefs about the other

player’s type, and that give expected utility of zero to each player if they

reveal their beliefs truthfully. For player 1 such a transfer rule is indicated

in Figure 6.4. We omit the simple check that this rule has the required

properties.

In Figure 6.4, positive numbers are payments by player 1, and negative num-

bers are payments to player 1. The idea of the payment rule is that agent

1 is rewarded if the type that she reports is the same as that of player 2,

but she has to make a payment if the type that she reports is diﬀerent from

CHAPTER 6. BAYESIAN DESIGN: GENERAL THEORY 134

1 2 3

1 (−

5

2

, −

5

2

) (3,5) (6,9)

2 (5,3) (-5, -5) (3,6)

3 (9,6) (6,3) (-

3

2

, −

3

2

)

Figure 6.5: Modiﬁed transfer payments

the one that player 2 reports. This reﬂects that agents’ types are positively

correlated in this example.

A symmetric rule can, of course, be used for player 2. We shall now add

a positive multiple of the transfers in Figure 6.4 to the transfers in Figure

6.3. We need to determine by how much we need to multiply the transfers in

Figure 6.4. We need to overcome all incentives to deviate in the mechanism

of Figures 6.2 and 6.3. One can calculate that 3 is the smallest integer by

which we can multiply all payments in Figure 6.4 and obtain a transfer rule

that eliminates all incentives to lie in the mechanism of Figures 6.2 and 6.3.

If we multiply the transfer payments in Figure 6.4 by 3 and add them to the

transfer payments in Figure 6.3 we get the payments shown in Figure 6.5.

One can now verify that the auction that has the allocation rule indicated in

Figure 6.2 and the transfer rules of Figure 6.5 extracts the full surplus for

the auctioneer, and is incentive compatible and individually rational. Note

that individual rationality only holds at the interim, not at the ex post level.

6.4.5 Individual Rationality and Budget Balance

We will want to combine Bayesian incentive compatibility with other re-

quirements, such as individual rationality and ex post budget balance. We

begin by deﬁning individual rationality.

Deﬁnition 6.5. A direct mechanism (q, t

1

, t

2

, . . . , t

N

) is “individually ra-

tional” if for every agent i and every type θ

i

∈ Θ

i

we have:

_

A

u

i

(a, θ

i

)dQ

i

(θ

i

) −T

i

(θ

i

) ≥ 0

In the transformation that was described in Subsection 6.4.3, whenever the

mechanism with which we started is individually rational, then also the

CHAPTER 6. BAYESIAN DESIGN: GENERAL THEORY 135

transformed mechanism is individually rational. Thus, any pair of decision

and transfer rules that are individually rational can be made incentive com-

patible and individually rational provided the Cr´emer McLean condition

holds.

We next consider ex post budget balance. The construction that demon-

strates the equivalence of ex ante and ex post budget balance and proves

Proposition 6.3 requires that types are independent. That result therefore

does not straightforwardly generalize to the context of correlated types.

However, Kosenok and Severinov (2008) have shown that an additional

condition for the prior distribution of types that goes beyond the Cr´emer

McLean condition, guarantees that budget balance can always be achieved.

The additional condition is called “identiﬁability.”

Deﬁnition 6.6. The probability distribution µ satisﬁes the identiﬁability

condition if for all distributions ν ,= µ such that ν(θ) > 0 for all θ ∈ Θ there

is at least one agent i and one type θ

i

∈ Θ

i

such that for any collection of

nonnegative coeﬃcients

_

λ

θ

i

_

θ

i

∈Θ

i

we have:

ν(θ

−i

[ θ

i

) ,=

θ

i

∈Θ

i

λ

θ

i

µ(θ

−i

[ θ

i

)

for at least one θ

−i

∈ Θ

−i

.

Intuitively, this condition says that for any alternative distribution ν of

types there is at least one agent and one type of that agent such that this

agent cannot randomize over reports in a way that makes the conditional

distribution of all other types under ν indistinguishable from the conditional

distribution of all other types µ. Kosenok and Severinov (2008) prove:

Proposition 6.6. Suppose that the distribution µ satisﬁes the Cr´emer-

McLean and the identiﬁability condition. Consider any direct mechanism

(q, t

1

, t

2

, . . . , t

N

) that is individually rational and ex ante budget balanced.

Then there is a Bayesian incentive compatible and ex post budget balanced

mechanism (q, t

1

, t

2

, . . . , t

N

) that is equivalent, that is:

1. the two mechanisms have the same decision rule q;

2. the two mechanisms have the same interim expected payments:

θ

−i

∈Θ

−i

t

i

(θ

i

, θ

−i

)µ(θ

−i

[ θ

i

) =

θ

−i

∈Θ

−i

t

i

(θ

i

, θ

−i

)µ(θ

−i

[ θ

i

)

for all i ∈ I and θ

i

∈ Θ

i

.

CHAPTER 6. BAYESIAN DESIGN: GENERAL THEORY 136

Kosenok and Severinov’s construction that proves this result is quite diﬀer-

ent from Cr´emer and McLean’s construction, and we omit a discussion of

this construction.

6.5 Discussion

A number of authors have investigated the question how the set-up presented

by Cr´emer and McLean needs to be changed to obtain less paradoxical

results. Much of this discussion has focused on the particular case of a single

unit auction in which the decision and transfer rules to be implemented are

the full extraction rules, as illustrated by our numerical example. However, it

seems plausible that the results of this literature generalize to other settings.

Suppose that in the Cr´emer and McLean setting the prior distribution of

types is close to a product measure, i.e. types are close to being independent.

Then a belief revelation scheme that provides incentives strong enough to

outdo all incentives to lie will require large payments by agents, and it

will expose agents to signiﬁcant risk. Thus, if agents are either liquidity

constrained or risk averse, such a mechanism might be impossible, or agents

might require compensation for the risk that reduces the attractiveness of the

mechanism to the mechanism designer. This intuition has been formalized

by Robert (1991).

The fact that in a belief revelation scheme the reports of all agents except

agent i determine how much agent i has to pay suggests that there may

be an incentive to collude. Laﬀont and Martimort (2000) have shown that

in environments in which types are close to independent collusion might

prevent the mechanism designer from full surplus extraction.

Heifetz and Neeman, in a recent series of papers, have pursued a diﬀerent line

of argument. They point out that the construction of Cr´emer and McLean

requires that there is a one-to-one relation between an agent’s preferences

and an agent’s beliefs about other types. This precisely is the reason why a

belief extraction mechanism can help with implementation. Neeman (2004)

and Heifetz and Neeman (2006) have shown that information structures with

this property are in some sense rare among all conceivable common prior

information structures. While it is often true that for ﬁxed ﬁnite type space

generic probability distributions will satisfy Cr´emer and McLean’s condi-

tions, the same distributions are, as Heifetz and Neeman show, a very small

subset of the inﬁnite dimensional space of general information structures.

CHAPTER 6. BAYESIAN DESIGN: GENERAL THEORY 137

6.6 Remarks on the Literature

Proposition 6.4 is adapted from Cr´emer and McLean (1988). My favorite

reference for Farkas’ Lemma, that is central to the proof of Proposition 1,

is Franklin (1980). Franklin’s proof is not much more than half a page long,

and all that you need to follow the proof is a basic knowledge of separating

hyperplane theorems. Proposition 6.6 is a combination of Theorem 1 and

Corollary 1 in Kosenok and Severinov (2008).

6.7 Problems

A seller has a single indivisible object to sell. The seller values the object at

zero, and seeks to maximize his expected revenue. There are two potential

buyers: i = 1, 2. Buyer i’s von Neumann Morgenstern utility equals v

i

−t

i

if he obtains the good and pays t

i

, and his utility equals −t

i

if he does not

obtain the good and pays t

i

.

For each buyer i the valuation v

i

is a random variable that only buyer i

observes, but that the other buyer and the seller don’t observe. The random

variable v

1

takes only two possible values: v

1

∈ ¦3, 4¦. The random variable

v

2

takes only two possible values: v

2

∈ ¦1, 2¦. Note that the two buyers are

not symmetric. The joint distribution of v

1

and v

2

is given by the following

table:

v

2

= 1 v

2

= 2

v

1

= 3 0.25+ε 0.25-ε

v

1

= 4 0.25-ε 0.25+ε

Figure 6.6: Distribution of Valuations

Here, ε is a constant that satisﬁes 0 < ε < 0.25. Each entry indicates the

probability with which v

1

takes the value indicated in the row, and at the

same time v

2

takes the value indicated in the column. The distribution

indicated in this table is common knowledge among the buyers and the

seller.

(a) Verify that the Cr´emer McLean condition holds for buyer 1.

(b) Construct a payment scheme that provides buyer 1 with incentives for

truthful revelation of his beliefs about buyer 2’s type, and that implies that,

CHAPTER 6. BAYESIAN DESIGN: GENERAL THEORY 138

if he truthfully reveals his beliefs, his expected transfer payment conditional

on each of his types will be zero.

(c) Use the payment scheme that you have found in part (b) to construct a

Bayesian incentive compatible selling mechanism that always allocates the

object to buyer 1, and that implies that buyer 1’s expected payment to the

seller equals his expected valuation of the object. The selling mechanism

that you construct should also oﬀer each type of each agent an interim

expected utility of at least zero in equilibrium.

(d) Investigate the limit for ε → 0 of the transfer payments in the selling

mechanism that you obtained in part (c).

Chapter 7

Non-Transferrable Utility

7.1 Introduction

So far we have assumed that all agents’ utility is additively separable in

an allocative decision, and money, and moreover that all agents are risk-

neutral in money. This is a very restrictive assumption. It is easy to imagine

situations in which the assumption is not satisﬁed. The simplest case is that

agents are not risk neutral in money. Another way in which the assumption

might be violated is that there are interactions between money and the

allocative decision. Finally, it might be the case that we are considering

situations, such as voting, in which monetary payments are typically not

invoked to provide incentives.

One implication of the assumption of additively separable utility and risk

neutrality is that monetary payments that are made by one agent and re-

ceived by another agent are welfare neutral as long as we allocate the same

welfare weight to all agents. By contrast, in a more general model, welfare

depends not only on the allocative decision but also on the distribution of

money among agents. We shall refer to the case that we have considered

so far as the case of “transferrable utility” and we shall refer to the case

considered in this chapter as the case of “non-transferrable utility.”

On the specter of possible assumptions, we shall move in this chapter to

the opposite extreme of the assumption that we have made so far. We

shall consider situations in which there is some arbitrary set of possible

collective decisions, and, at least initially, no particular assumption is made

about agents’ preferences over these decisions. Each collective decision is

interpreted as a decision about all issues relevant to agents, including the

allocation of money.

139

CHAPTER 7. NON-TRANSFERRABLE UTILITY 140

The emphasis of theoretical work in this area has been on dominant strategy

incentive compatibility. We shall start with this part of the literature. There

is a vast number of theoretical results in this area. We shall oﬀer a discussion

of only an eclectic selection from these results. Later, we shall then brieﬂy

discuss some work on Bayesian incentive compatibility without transferrable

utility.

7.2 The Gibbard Satterthwaite Theorem

7.2.1 Set Up

There is a ﬁnite set of agents, i ∈ I = ¦1, 2, . . . , N¦. These agents have to

choose one alternative from a ﬁnite set A of mutually exclusive alternatives.

Each agent i has a preference relation R

i

over A. We assume that R

i

is

a linear order, that is, it is complete, transitive, reﬂexive, and the only

indiﬀerence is among identical elements. We denote the strict order derived

from R

i

by P

i

. We read “aR

i

b” as “a is weakly preferred to b,” and we read

“aP

i

b” as “a is strictly preferred to b.” The set of all linear orders over A

is denoted by ¹. We write R for the list of all agents’ preference relations:

R = (R

1

, R

2

, . . . , R

N

), and we write R

−i

for the list of all agents’ preference

relations leaving out agent i’s: R

−i

= (R

1

, R

2

, . . . , R

i−1

, R

i+1

, . . . , R

N

).

We consider a mechanism designer who does not know the agents’ preference

relations, but who determines the rules of the strategic interaction among the

agents by which an alternative from A is chosen. The mechanism designer

can thus construct an extensive game with outcomes in A. We study in this

section the case in which the mechanism designer seeks to construct a game

such that every player, with every conceivable preference relation in ¹, has

a dominant strategy in the sense in which we used this phrase in earlier

chapters of these notes. The revelation principle then applies, and we can

restrict attention to direct mechanisms.

Deﬁnition 7.1. A “direct mechanism” is a function f : ¹

N

→A.

In the literature direct mechanisms in the sense of deﬁnition 7.1 are some-

times also called “social choice functions.” To maintain consistency with

earlier parts of these notes, we shall speak of “direct mechanisms.”

Deﬁnition 7.2. A direct mechanism f is “dominant strategy incentive com-

patible” if for every agent i ∈ I and all preference relations R

i

, R

i

in ¹:

f(R

i

, R

−i

)R

i

f(R

i

, R

−i

)

CHAPTER 7. NON-TRANSFERRABLE UTILITY 141

In the literature dominant strategy incentive compatibility is sometimes also

referred to as “strategy proofness.” In this section we shall characterize

direct mechanisms that are dominant strategy incentive compatible. Before

we proceed we make two further remarks about the set up described here.

The ﬁrst concerns the role of randomization in this section. The reader

might notice that we have not introduced probability distributions over the

set A. But, of course, the elements of set A could be outcomes that are not

deterministic, but stochastic. If we have this case in mind, then it might

seem more natural to have an inﬁnite set A rather than a ﬁnite set. But it is

not immediately obvious that in practice an inﬁnite number of probability

distributions can be implemented. Physical limitations might well force us

to choose from a ﬁnite set of probability distributions, and the reader can

think of A as the set of all these distributions. Thus, our setup does not

really rule out randomization.

The second point that needs emphasis is that the domain of our direct

mechanism contains all linear orders of A. In practice we might have some

knowledge about agents’ preferences over A, and can therefore restrict at-

tention to mechanisms that perform as we want them to perform only for

some, but not for all preference vectors. Natural restrictions may be that

all agents rank either all elements of some subset

ˆ

A of A higher than all

elements of the complement A ¸

ˆ

A, or that they have the opposite ranking.

This would seem plausible if the elements of A are candidates for a political

oﬃce, and if the candidates in

ˆ

A have the opposite ideology from candidates

in A¸

ˆ

A. Another natural restriction may be that agents’ preferences are of

the von Neumann Morgenstern form. This would seem natural if A consists

of lotteries, as suggested in the previous paragraph. We shall consider later

in this Chapter the implications of some such domain restrictions, but ﬁrst

we explore the implications of assuming an unrestricted domain.

7.2.2 Statement of the Result and Outline of the Proof

Deﬁnition 7.3. A direct mechanism f is called “dictatorial” if there is some

individual i ∈ I such that for all R ∈ ¹

N

:

f(R)R

i

a for all a ∈ A.

Proposition 7.1. Suppose that A has at least three elements, and that

the range of f is A. A direct mechanism f is dominant strategy incentive

compatible if and only if it is dictatorial.

CHAPTER 7. NON-TRANSFERRABLE UTILITY 142

This is one of the most celebrated results in the theory of mechanism de-

sign. It is due to Gibbard (1973) and Satterthwaite (1975), and is called

the “Gibbard Satterthwaite Theorem.” The Gibbard-Satterthwaite theorem

is a paradoxical result. It shows that the requirements that the require-

ments of dominant strategy incentive compatibility and unlimited domain

together are too strong. In later parts of this Chapter, we shall discuss

weaker requirements that lead to more positive results.

In this result the assumption that the range of f equals A rather than being

a strict subset of A is immaterial. If the range of f is a strict subset of A,

we can re-deﬁne the set of alternatives to be the range of f. Alternatives

that are not in the range of f will never be chosen, and therefore agents’

preferences over these alternatives cannot inﬂuence the outcome. We can

therefore analyze the situation as if such alternatives didn’t exist.

The assumption that A has at least three elements is important. Without

this assumption the result is not true. Consider the only remaining non-

trivial case, that is, the case that A has two elements. In that case many

mechanisms are dominant strategy incentive compatible. For example, giv-

ing each agent the opportunity to vote for one of the two alternatives, and

then choosing the alternative with the highest number of votes, with some

arbitrary tie breaking rule, is dominant strategy incentive compatible, and

clearly not dictatorial.

The “suﬃciency part” of Proposition 7.1 is obvious. We therefore focus

on proving the “necessity part.” Our presentation of this proof is based

on Reny (2001). We shall proceed in two steps. First, we show that every

dominant strategy incentive compatible, direct mechanism is monotone. We

also report some simple implications of monotonicity. The core of the proof

is then in the next subsection, where we show in a second step that every

monotone direct mechanism is dictatorial.

Deﬁnition 7.4. A direct mechanism f is “monotone” if, whenever f(R) =

a, and for every agent i and every alternative b the preference relation R

i

ranks a above b if R

i

does, then f(R

) = a. Formally:

f(R) = a and for all i ∈ I : [aR

i

b for all b ∈ A such that aR

i

b] ⇒f(R

) = a.

Proposition 7.2. If f is dominant strategy incentive compatible, then it is

monotone.

Proof. Suppose ﬁrst in Deﬁnition 7.4 preference proﬁles R and R

diﬀered

only in the i-th component. Suppose also f(R

) = b ,= a. Because with

CHAPTER 7. NON-TRANSFERRABLE UTILITY 143

preference R

i

it is a dominant strategy for agent i to report R

i

truthfully,

we have to have: aR

i

b. Because the ranking of a does not fall as we move

from R

i

to R

i

, this implies: aR

i

b. But then, with true preference R

i

, agent i

has an incentive to report preference R

i

. This contradicts dominant strategy

incentive compatibility. Thus we have arrived at a contradiction, and we can

conclude f(R

) = a. If R

**and R diﬀer in more than one component, then
**

we apply the above argument successively to each component in which R

**and R diﬀer. Again, we arrive at the conclusion f(R) = a.
**

Now we introduce two simple implications of monotonicity.

Deﬁnition 7.5. A direct mechanism f is “set-monotone” if, whenever

f(R) ∈ B for some B ⊆ A, and for every agent i, preference relation R

i

diﬀers from preference relation R

i

only regarding the ranking of elements of

B, then f(R

) ∈ B:

f(R) ∈ B and [aR

i

a

⇔aR

i

a

whenever a / ∈ B or a

/ ∈ B] ⇒f(R

) ∈ B.

Proposition 7.3. If f is monotone then f is set-monotone.

Proof. Assume f(R

**) = a / ∈ B. Then monotonicity implies that f(R) = a,
**

which contradicts f(R) ∈ B.

Deﬁnition 7.6. A direct mechanism f “respects unanimity” if, whenever

an alternative a is at the top of every individual’s preference relation, then

a is chosen by f:

aR

i

b for all i ∈ I and b ∈ A ⇒f(R) = a.

Proposition 7.4. If f is monotone and the range of f is the A, then f

respects unanimity.

Proof. Consider any a ∈ A. Because the range of f is A, there is some

R such that f(R) = a. Now raise a to the top of everyone’s ranking. By

monotonicity, the social choice remains a. Now re-order alternatives below a

in arbitrary ways. Again, by monotonicity, the choice must remain a. This

proves that f respects unanimity.

7.2.3 Every monotone direct mechanism is dictatorial

Proposition 7.5. Suppose that A has at least three elements, and that the

range of f is A. If f is monotone, then it is dictatorial.

CHAPTER 7. NON-TRANSFERRABLE UTILITY 144

R

1

. . . R

n−1

R

n

R

n+1

. . . R

N

. . a . .

. . c . .

. . b . .

. . . . .

. . . . .

. . . . .

c . . . c . c . . . c

b . . . b . b . . . b

a . . . a . a . . . a

Figure 7.1: Social choice is a

Proof. We shall show that for every alternative a ∈ A, there is a “dictator

for a,” i.e., there is an agent i such that whenever a is at the top of i’s

ranking, then a is chosen. If there is such a dictator for every alternative

a, then the dictator must be the same individual i for every alternative.

Otherwise, if the dictator for a had a at the top of her ranking, and the

dictator for b(,= a) had b at the top of his ranking, the outcome would not

be well-deﬁned. We conclude that the dictator must be the same agent i for

every alternative, and therefore this agent i is a dictator.

Now ﬁx an alternative a ∈ A. To show that there is a dictator for a, it is

suﬃcient to ﬁnd one preference proﬁle where a is at the top of some agent

i’s ranking, but at the bottom of everybody else’s ranking, and where the

social choice is a. One such proﬁle is shown in Figure 7.1. In this ﬁgure,

and in subsequent ﬁgures, each column corresponds to one player’s ranking

of alternatives in A, with the highest ranked alternative at the top. Figure

7.1 shows a speciﬁc ranking, involving alternatives other than a, but for the

moment this is of no relevance to the argument. In Figure 7.1 agent n ranks

a top, but all other agents rank a bottom. Finding one such proﬁle in which

the social choice is a is suﬃcient because every other proﬁle in which agent

n ranks a at the top can be obtained from the one shown in Figure 7.1 by

changing preferences without moving any alternative above a, and therefore

by monotonicity the social choice for every other proﬁle in which agent n

ranks a at the top is a.

We shall arrive at the conclusion that the social choice in Figure 7.1 has to

be a through a sequence of steps starting with the proﬁle shown in Figure

7.2. In the proﬁle in Figure 7.2 the social choice has to be a because f

CHAPTER 7. NON-TRANSFERRABLE UTILITY 145

R

1

. . . R

n−1

R

n

R

n+1

. . . R

N

a . . . a a a . . . a

. . . . .

. . . . .

. . . . .

b . . . b b b . . . b

Figure 7.2: Social choice is a

R

1

. . . R

n−1

R

n

R

n+1

. . . R

N

a . . . a a a . . . a

b . . . a a a . . . a

. . . . .

. . . . .

. . . . .

. b b b . . . b

Figure 7.3: Social choice is a

respects unanimity. Now suppose that we move b up in agent 1’s ranking,

until it is just below a, as shown in Figure 7.3. Then, by monotonicity, the

social choice has to remain a.

Now suppose that we move b one step further up, to the top of agent 1’s

ranking, as shown in Figure 7.4. Then set-monotonicity as deﬁned above

(setting B = ¦a, b¦) implies that the social choice is either a or b.

We will now identify an agent n of whom we shall show that he or she is

a dictator for n. If the social choice is b in Figure 7.4, then we set n = 1.

R

1

. . . R

n−1

R

n

R

n+1

. . . R

N

b . . . a a a . . . a

a . . . a a a . . . a

. . . . .

. . . . .

. . . . .

. b b b . . . b

Figure 7.4: Social choice is a or b

CHAPTER 7. NON-TRANSFERRABLE UTILITY 146

R

1

. . . R

n−1

R

n

R

n+1

. . . R

N

b . . . b a a . . . a

a . . . a b . . . . .

. . . . .

. . . . .

. . . . .

. . . b . . . b

Figure 7.5: Social choice is a

R

1

. . . R

n−1

R

n

R

n+1

. . . R

N

b . . . b b a . . . a

a . . . a a . . . . .

. . . . .

. . . . .

. . . . .

. . . b . . . b

Figure 7.6: Social choice is b

However, if the social choice remains a in Figure 7.4, then we repeat the

same procedure for agent 2, etc. The ﬁrst agent for whom the social choice

switches from a to b is the agent whom we identify as our candidate dictator

n. There has to be one such agent because after we have worked our way

through all agents we arrive at a proﬁle where all agents put alternative b at

the top of their ranking, and f respects unanimity so that the social choice

is b. In Figures 7.5 and 7.6 we show the generic situation for agent n before

and after the social choice switches from a to b.

Figures 7.7 and 7.8 show the same proﬁles as Figures 7.5 and 7.6 except

that we have moved alternative a to the bottom of the ranking for agents

i < n, and we have moved them to the second position from the bottom for

agents i > n. We now argue that in these proﬁles the social choices have to

be the same as in Figures 7.5 and 7.6. For Figure 7.8 it follows from Figure

7.6 and monotonicity that the choice has to be b. Now consider the proﬁle

in Figure 7.7. Comparing Figure 7.7 and Figure 7.8 we can conclude from

set monotonicity (with B ≡ ¦a, b¦) that the social choice in Figure 7.7 has

to be either a or b. But if the alternative chosen were b, then it would also

have to be b in Figure 7.5, by monotonicity. Therefore, the choice has to be

CHAPTER 7. NON-TRANSFERRABLE UTILITY 147

R

1

. . . R

n−1

R

n

R

n+1

. . . R

N

b . . . b a . . . . .

, . . . b . . . . .

. . . . .

. . . . .

. . . a a

a a . b . . . b

Figure 7.7: Social choice is a

R

1

. . . R

n−1

R

n

R

n+1

. . . R

N

b . . . b b . . . . .

. . . . . a . . . . .

. . . . .

. . . . .

. . . a a

a a . b . . . b

Figure 7.8: Social choice is b

a in Figure 7.7.

Now consider the preferences in Figure 7.9. The position of alternative a

has not changed relative to other alternatives in comparison to Figure 7.7,

and therefore the social choice has to be a.

Finally, compare Figure 7.9 to Figure 7.1. The choice in Figure 7.1 has to

be a or b, by set monotonicity, setting B = ¦a, b¦. But if the choice were

b in Figure 7.1, then we could move alternative c to the top of everyone’s

preferences, and by monotonicity the choice would still have to be b, which

would contradict that f respects unanimity. Therefore, the choice in Figure

7.1 has to be a. This concludes the proof.

7.3 Dominant Strategy Incentive Compatibility On

Restricted Domains

Two natural ways of relaxing the stringent requirements of Proposition 7.1

so that more positive results obtain are ﬁrstly to consider a more restricted

CHAPTER 7. NON-TRANSFERRABLE UTILITY 148

R

1

. . . R

n−1

R

n

R

n+1

. . . R

N

. . a . .

. . c . .

. . b . .

. . . . .

. . . . .

. . . . .

c . . . c . c . . . c

b . . . b . a . . . a

a . . . a . b . . . b

Figure 7.9: Social choice is a

domain of preferences, and secondly to consider a less demanding solution

concept than dominant strategies. In this section we shall discuss the former

approach whereas in the next section we discuss the latter approach. In

this Section, thus, the set of preferences for any individual i that we are

considering is no longer the complete set ¹, but some subset

ˆ

¹ of ¹.

The best known restriction on the domain of preferences that allows dom-

inant strategy incentive compatible mechanisms that are not dictatorial is

single-peakedness. Suppose the alternatives in A are labeled with the inte-

gers 1, 2, . . . , K. A preference relation R

i

of agent i is called “single-peaked”

if there is a k(i) ∈ ¦1, 2, . . . , K¦ such that (i) agent i prefers alternative k(i)

to all other alternatives: k(i)R

i

j for all j ∈ A, and (ii) agent i’s preferences

decline monotonically “to the left” and “to the right” of k(i), that is: if

≥ k then R

i

+ 1 and if ≤ k then − 1R

i

. Denote by

ˆ

¹ the set of

all single-peaked preferences. The deﬁnition of this set depends on how the

alternatives have been labeled with numbers. We keep this labeling ﬁxed

in this section, and, for simplicity, don’t reﬂect the dependency of the set

ˆ

¹ on the labeling in our notation. The restricted domain of single-peaked

preferences is then

ˆ

¹

N

.

The domain of single-peaked preferences may appear natural in a the fol-

lowing environment. Alternatives are candidates for some political position.

Direct mechanisms are methods for selecting one candidate out of a set of

candidates. Candidates are labeled according to their position on the “left-

right” spectrum, and the alternative k(i) reﬂects voter i’s ideal position on

this spectrum. It then appears plausible that preferences decline monoton-

ically as candidates are further away from agent i’s ideal position.

CHAPTER 7. NON-TRANSFERRABLE UTILITY 149

Proposition 7.6. If preferences are single-peaked, then there are dominant

strategy direct mechanisms that are not dictatorial.

Proof. Suppose N is odd. For every agent i denote by k(i) the alternative

that is ranked top according to R

i

. The rule f that picks the alterna-

tive a

m

where m is the median of the vector (k(1), k(2), . . . , k(N) is dom-

inant strategy incentive compatible. Consider any agent i, and take the

other agents’ reported preferences R

−i

as given and ﬁxed. The median of

(k(1), k(2), . . . , k(N) will then be between the

N

2

-th highest and the (

N

2

+1)-

th highest index of the most preferred alternative. Denote these by m

−

and

m

+

. The median will be these two numbers, independent of what agent i

reports. If the alternative that agent i most prefers is between these two

numbers, then agent i can ensure by reporting his preferences truthfully

that his most preferred alternative is chosen. If agent i’s most preferred

alternative is lower than m

−

, then by reporting her preferences truthfully

agent i ensures that alternative m

−

, her most preferred alternative from the

range of possible medians, is chosen. If agent i’s most preferred alternative

is higher than m

+

, then by reporting her preferences truthfully, agent i can

ensure that alternative m

+

, her most preferred alternative from the range

of possible medians, is chosen. Thus, in either case agent i, by reporting

her preferences truthfully, ensures that her most preferred alternative from

the range of possible medians is chosen. Thus, the rule is dominant strategy

incentive compatible, and it is obviously non-dictatorial.

Now suppose that N is even. Then we can use the same rule as in the

case that N is odd, with the following small modiﬁcation. We arbitrarily

pick some alternative a ∈ A and pretend that some agent had expressed a

preference that lists a at the top. The same argument as above proves that

this rule is dominant strategy incentive compatible.

Other restricted domains on which dominant strategy incentive compati-

bility does not imply dictatorships are, of course, those that we studied in

Chapter 4. In those environments we assumed that agents’ utility func-

tions were of a particular form, for example, additive separability and risk

neutrality. This domain restriction lead to the dominant strategy incentive

compatibility of simple rules such as the ﬁxed price rule for bilateral trade.

A particularly interesting domain restriction might be natural in the case

in which the set A consists of all lotteries over some given ﬁnite set of

outcomes. In this case one might assume that individuals’ preferences satisfy

the von Neumann Morgenstern axioms, and can therefore be represented by

Bernoulli utility functions. One might hope that this domain restriction

CHAPTER 7. NON-TRANSFERRABLE UTILITY 150

allows the implementation of social choice functions other than dictatorial

social choice functions. However, a theorem due to Hylland (1980) shows

that in this setting, without further domain restrictions, the only direct

mechanisms that respect unanimity and are dominant strategy incentive

compatible are random dictatorships. In random dictatorships each of the

N agents is assigned a probability, and is then picked with this probability

as the dictator, i.e. the alternative is picked that maximizes this individual’s

preference relation.

7.4 Bayesian Incentive Compatibility

Although much of the literature on mechanism design with non-transferrable

utility has focused on dominant strategy incentive compatibility there seems

to be no fundamental reason why not to consider Bayesian incentive compat-

ibility as well. To explore this approach B¨orgers and Postl (2005) consider

a stylized model of compromising. There are two agents, N = 2, and three

alternatives: A = ¦a, b, c¦. It is commonly known that agent 1 ranks the

alternatives as a, b, c. Agent 1’s Bernoulli utility of alternative a is 1, and

her Bernoulli utility of alternative c is 0, but her Bernoulli utility of the

intermediate alternative b is θ

1

, agent 1’s privately observed type. Agent 2

is analogous, except that agent 2 ranks alternatives in the opposite order

c, b, a. A direct mechanism maps the vector (θ

1

, θ

2

) of agents’ types into a

probability distribution over the set of alternatives A. B¨orgers and Postl

(2005) assume that θ

1

and θ

2

are stochastically independent variables that

are both distributed on the interval [0, 1] with the same cumulative distribu-

tion function F with density f where f(θ

i

) > 0 for all θ

i

∈ [0, 1]. Each agent

i observes θ

i

but not θ

j

where j ,= i, and the mechanism designer observes

neither of the two types.

B¨orgers and Postl (2005) interpret the alternative B as a “compromise”

as both agents rank B as their middle alternative, even though the two

agents have opposing preferences. The focus of B¨orgers and Postl (2005) is

on the question whether there is an incentive compatible mechanism that

implements the alternative B whenever it is eﬃcient, that is, whenever θ

1

+

θ

2

= 1. This is a natural notion of eﬃciency because the sum of utilities

from the compromise B is θ

1

+θ

2

, whereas alternatives A and C both yield

a sum of utilities of 1. To induce agents to correctly reveal their type the

mechanism designer can choose the probabilities of the individuals’ favorites,

A and C, judiciously. These probabilities can be used as an instrument for

providing incentives because they are welfare neutral: Both alternatives

CHAPTER 7. NON-TRANSFERRABLE UTILITY 151

yield exactly the same amount of welfare.

B¨orgers and Postl (2005) note that their framework is equivalent to a public

goods model similar to the one we considered in Section 3.3. One can think

of the “compromise” as a public good for which both agents have to give

up probability of their preferred alternative. The cost-function is one-to-

one: for each unit of the public good (i.e. probability of the compromise)

produced some agent has to give up one unit of money (i.e. probability

of their preferred alternative). An important diﬀerence with the public

goods models that we considered in Section 3.3 is that agents are liquidity

constrained: probabilities have to be between zero and one, and therefore

agents cannot give up more than sum upper bound of probability of their

preferred alternative.

B¨orgers and Postl (2005) use arguments similar to our arguments in Section

3.3 to show that no Bayesian incentive compatible mechanism implements

the ﬁrst best, i.e. choice of the compromise with probability 1 if and only

if θ

1

+ θ

2

≥ 1. The problem of ﬁnding a second best is complicated by

the fact that agents face liquidity constraints. Therefore, to determine the

second best, one cannot use the approach that was explained in Section 3.3.

B¨orgers and Postl (2005) have analytical results only for small parametrized

classes of direct mechanisms. The complement these results with a numerical

investigation of second best mechanisms.

The compromise problem in B¨orgers and Postl (2005) is a sandbox model

for the more general model of determining an optimal voting scheme when

there are three or more alternatives, that is, the Bayesian equivalent of

the problem that Gibbard and Satterthwaite studied. B¨orgers and Postl

(2005)’s diﬃculties in analyzing the much simpler model indicate that the

more general problem is analytically hard, and that it might beneﬁt from

numerical investigation.

7.5 Remarks on the Literature

The large literature on the Gibbard-Satterthwaite theorem is surveyed by

Barbera (2001). Barbera describes several diﬀerent approaches to the proof

of the theorem. An important aspect of the Gibbard-Satterthwaite theorem

that we have not mentioned in the text is that, as Satterthwaite (1975)

pointed out, the theorem is equivalent to Arrow’s famous theorem on the

impossibility of preference aggregation Arrow (1963).

CHAPTER 7. NON-TRANSFERRABLE UTILITY 152

Our treatment of the case of single-peaked preferences is based on Barbera

(2001) although the original result is in Moulin (1980). In fact, Moulin

(1980) obtains a converse to Proposition 7.6 and shows that all dominant

strategy incentive compatible direct mechanisms on a single-peaked domain

need to be based on a generalized version of the median decision rule.

Our discussion of the case of Hylland’s result is based on Dutta et al. (2007)

who also oﬀer an independent proof of the result. That paper also contains

a detailed discussion of the relation between Hylland’s result and Gibbard’s

earlier theorems on random dictatorship (Gibbard (1977), Gibbard (1978)).

Other papers that investigate Bayesian incentive compatibility without trans-

ferrable utility include Abdulkadiroglu and Loertscher (2007). The literature

on the design of matching markets Roth (2008) is also in a setting without

transferrable utility. This literature is less focused on incentive compatibil-

ity than the mechanism design literature. Strategic properties of matching

algorithms are, however, discussed in Section 2.1 of Roth (2008).

7.6 Problems

Chapter 8

Interdependent Types

8.1 Introduction

In all previous parts of these notes we have assumed that each agent’s pri-

vate information is all that matters for that agent’s preferences over group

choices. In this section, we shall instead consider the case in which an agent’s

preference over group choices depends not only on this agent’s own private

information, but also on the information of other agents. We shall refer to

this as the case of “informational interdependence.” By contrast, we refer

to the case that we have addressed so far as the case of “private values.”

One can think of many real world examples in which informational inter-

dependence seems important. When choosing a new chair of a committee,

diﬀerent committee members will have diﬀerent pieces of information about

the abilities of the candidates. Each member’s evaluation of a candidate will

depend not only on the member’s own information about the candidate, but

also on other members’ information about the candidate. Similarly, in auc-

tions for licenses to use radio spectrum to operate a telephone service, each

bidder will have some private information about the likely proﬁtability of

various services. The private information about market conditions that each

bidder holds is likely to aﬀect not only that bidder’s valuation of licenses

but also other bidders’ valuations of licenses.

In this section we shall present some elements of a formal analysis of mecha-

nism design with informational interdependence. Most of our discussion will

focus on Bayesian incentive compatibility. We shall, however, also comment

on ex post implementation, which is related to dominant strategy imple-

mentation.

153

CHAPTER 8. INTERDEPENDENT TYPES 154

We shall maintain the assumption of transferrable utility that we made in

all previous parts of these notes except Chapter 7. This will make it easier

to contrast the results of this section with those of previous sections.

For the case of informational independence, as for the case of private val-

ues, an important distinction is whether agents’ signals are assumed to be

independent random variables, or whether dependence is allowed. We shall

focus here on the case that these signals are independent. For the case that

signals are dependent, the permissive results obtained in Section 6.4 can

be generalized. However, as in the context of that Section, these results

are paradoxical, and it seems likely that some modiﬁcation of the setting in

which these results are obtained is needed in order to obtain more plausible

results.

The dimensions of the agents’ signal space will play an important role in

this part of the lecture notes. The role that the dimensions plays in the

case of informational interdependence is, however, quite diﬀerent from the

role that it plays in the case of private values. With informational inter-

dependence the dimensions of the signal spaces is, for example, crucial for

the question whether eﬃcient decision rules can be part of a Bayesian in-

centive compatible mechanism. Recall that in the case of private values the

answer to this question was positive, and independent of the dimensions

of the signal spaces. Eﬃcient decision rules could be implemented using

Vickrey-Clarke-Groves (VCG) mechanisms.

The focus of this section will be on the case in which signal spaces are

subsets of ﬁnite-dimensional Euclidean spaces where the dimension is larger

than one. For this case we explain a series of impossibility results that were

obtained by Philippe Jehiel and Benny Moldovanu in a sequence of papers.

8.2 An Example

We start oﬀ with an extremely simple example. Suppose that there are 2

agents, I = ¦1, 2¦, and two alternatives: A = ¦a, b¦. Agent 1 observes a

two-dimensional private signal (θ

1

, θ

2

) ∈ [0, 1]

2

. We assume that (θ

1

, θ

2

) is

a random variable with a distribution that has support [0, 1]

2

. Agent 2 has

no private information. Agent i’s utility if alternative a is chosen, and agent

i pays transfer t

i

is: θ

i

− t

i

. Agent i’s utility if alternative b is chosen and

agent i pays transfer t

i

is: 0.5−t

i

. Note the essential feature of this example:

agent 1’s private signal is relevant not only to his own utility but also to

agent 2’s utility.

CHAPTER 8. INTERDEPENDENT TYPES 155

By the revelation principle, we restrict attention to direct mechanisms. A

decision rule q : [0, 1]

2

→[0, 1] maps agent 1’s private signal into the proba-

bility with which alternative a is chosen for the given signal. A transfer rule

t : [0, 1]

2

→ R maps agent 1’s private signal into a transfer to be paid by

agent 1 for the given signal. Obviously, only transfers to be paid by agent

1 matter. A direct mechanism is “incentive compatible” if, for all realiza-

tions (θ

1

, θ

2

) ∈ [0, 1]

2

of his private signal, agent 1 ﬁnds it in his interest to

truthfully report the realization of the signal.

Let us ask a simple question: Can we ﬁnd transfers that make welfare maxi-

mizing decisions incentive compatible? Here, we deﬁne welfare maximization

in the usual way as maximizing the sum of the agents’ utilities. More specif-

ically, the “ﬁrst best” decision rule q

∗

satisﬁes: q

∗

(θ

1

, θ

2

) = 1 if θ

1

+θ

2

> 1,

and q

∗

(θ

1

, θ

2

) = 0 if θ

1

+ θ

2

< 1. Our question is: Can we ﬁnd a transfer

rule t

∗

so that the direct mechanism (q

∗

, t

∗

) is incentive compatible?

As a ﬁrst step, observe that we require that t

∗

(θ

1

, θ

2

) is constant for all

(θ

1

, θ

2

) such that q

∗

(θ

1

, θ

2

) = 1, because otherwise agent 1 would distort

his report and only report that (θ

1

, θ

2

) for which t

∗

is minimal. Denote the

constant payment by t

a

. Similarly, we require that t

∗

(θ

1

, θ

2

) is constant for

all (θ

1

, θ

2

) such that q

∗

(θ

1

, θ

2

) = 0. Denote the constant payment by t

b

.

16

Now observe that for every θ

1

∈ (0, 1) agent 1 can report a θ

2

such that

q(θ

1

, θ

2

) = 0, and he can also report a θ

2

such that q(θ

1

, θ

2

) = 1. Depending

on which θ

2

he actually observes, we sometimes want him to choose the ﬁrst,

and sometimes the second option. Because his utility does not depend on

θ

2

, he must be indiﬀerent between the two choices:

θ

1

−t

a

= 0.5 −t

b

⇔

t

a

−t

b

= θ

1

−0.5. (8.1)

This has to be true for all θ

1

∈ (0, 1). Obviously, no two real numbers t

a

and t

b

have this property. We conclude that it is impossible to implement

the ﬁrst best.

The result that we observed for this example is not very surprising. Intu-

itively, we would like the group decision to condition on a component of an

agent’s private signal that has no implications for that agent’s payoﬀ. We

have no instrument that would allow us to elicit this information from the

agent.

16

This is sometimes called the “taxation principle.” Agent 1 is implicitly oﬀered the

choice between two alternatives, a and b, with associated taxes ta and t

b

.

CHAPTER 8. INTERDEPENDENT TYPES 156

Jehiel and Moldovanu (2001) have generalized the above example in two

ways. Firstly, they consider a model in which each agent makes private

observations that are relevant to other agents utility but not to the agent’s

own utility. They show that it is in general impossible to implement a

ﬁrst best decision rule. Their argument for this more general case is slightly

diﬀerent from the above argument. Secondly, they consider a model in which

each agent makes observations that aﬀect the agent’s own utility as well as

other agent’s utility, but potentially with diﬀerent weights. Again they show

the impossibility to implement ﬁrst best decision rules. The proof is related

to the argument in the above example. Our exposition below will focus on

Jehiel/Moldovanu’s second extension of the example.

8.3 Impossibility of Implementing Welfare Maxi-

mizing Decision Rules

We assume that there are N agents. The set of agents is denoted by I =

¦1, 2, . . . , N¦. They have to choose an alternative a out of some ﬁnite set A =

¦a

1

, a

2

, . . . , a

K

¦ of K mutually exclusive alternatives. Each agent i observes

a K-dimensional signal: θ

i

= (θ

i

1

, θ

i

2

, . . . , θ

i

K

) ∈ [0, 1]

K

.

17

The signal θ

i

has

a distribution with density f

i

that is positive everywhere. Diﬀerent agents’

signals are independent.

Agent i’s von Neumann Morgenstern utility if alternative a

k

is chosen, and

if agent i has to pay transfer t

i

is:

N

j=1

α

j

ki

θ

j

k

−t

i

. (8.2)

Thus, agent i’s utility from alternative a

k

is a linear function of the k −th

component of all agents’ signals. The factor in front of the k−th component

of agent j’s signal in agent i’s utility function is: α

i

kj

. We assume that

α

i

kj

,= 0 for all i, j ∈ I and all k = 1, 2, . . . , K. Note that we do not restrict

these factors to have the same sign for all agents. The linear form of the

utility function assumed above is not essential to the argument, but makes

the exposition easier.

A direct mechanism (q, t

1

, t

2

, . . . , t

N

) consists ﬁrstly of a mapping q : [0, 1]

IK

→

∆(A) that assigns to each vector of private signals θ ≡ (θ

1

, θ

2

, . . . , θ

N

) a

17

Jehiel and Moldovanu (2001) allow more general convex signal spaces with non-empty

interior, but for ease of exposition we restrict attention to the case that signals are con-

tained in [0, 1]

K

.

CHAPTER 8. INTERDEPENDENT TYPES 157

probability distribution q(θ) over A. The set of all such probability distri-

butions is denoted by ∆(A). For every agent i ∈ I a direct mechanism also

speciﬁes a mapping t

i

: [0, 1]

IK

→R that assigns to each vector θ of private

signals a transfer t

i

(θ) to be paid by agent i.

Given a direct mechanism we can deﬁne for every i ∈ I and every a ∈ A a

function Q

i

a

: [0, 1]

K

→ [0, 1] that assigns to every signal θ

i

of agent i the

interim probability that alternative a is chosen. We can also deﬁne for every

i ∈ I a function T

i

: [0, 1]

K

→ R that assigns to every signal θ

i

of agent i

the interim expected value of agent i’s transfer payment.

A direct mechanism is incentive compatible if and only if for every agent

i, and every pair of types θ

i

,

˜

θ

i

∈ [0, 1]

K

, agent i’s expected utility from

reporting type θ

i

is at least as large as his expected utility from reporting

type

˜

θ

i

if agent i’s true type is θ

i

.

A choice rule q

∗

is “ﬁrst best” if for every θ ∈ [0, 1]

IK

we have that q

∗

(θ)

assigns positive probability only to alternatives that maximize the sum of

agents’ utilities. A direct mechanism is “ﬁrst best” if its choice rule is ﬁrst

best.

A ﬁrst question to ask is whether we can use a VCG mechanism to implement

a ﬁrst best decision rule. Recall that in previous sections we deﬁned agent

i’s transfer rule in a VCG mechanism by:

t

i

(θ) = −

j=i

u

j

(q(θ), θ

j

) +τ

i

(θ

−i

) for all θ ∈ Θ, (8.3)

where τ

i

is an arbitrary function. Under this rule, if agent i is of type θ

i

and

reports that he is of type θ

i

, his utility is:

u

i

(q(θ

i

, θ

−i

), θ

i

) +

j=i

u

j

(q(θ

i

, θ

−i

), θ

j

) −τ

i

(θ

−i

)

=

N

j=1

u

j

(q(θ

i

, θ

−i

), θ

j

) −τ

i

(θ

−i

). (8.4)

Thus, when agent i reports θ

i

, his payoﬀs will be true social welfare, for

actual types, if the collective decision is q(θ

i

, θ

−i

), minus some constant that

does not aﬀect agent i’s incentives. Because the ﬁrst best rule maximizes

true social welfare, agent i will ﬁnd it in his interest to report his type

truthfully. Now, in our context, with informational interdependence, the

deﬁnition of the transfer rule needs to be modiﬁed, because now any agent

CHAPTER 8. INTERDEPENDENT TYPES 158

j’s utility does not only depend on agent j’s type, but on all agents’ types.

Thus, we might deﬁne:

t

i

(θ) = −

j=i

u

j

(q(θ), θ) +τ

i

(θ

−i

) for all θ ∈ Θ. (8.5)

Now, if agent i is of type θ

i

and reports that he is of type θ

i

, then his utility

becomes:

u

i

(q(θ

i

, θ

−i

), θ) +

j=i

u

j

(q(θ

i

, θ

−i

), (θ

i

, θ

−i

)) −τ

i

(θ

−i

). (8.6)

Note that agent i’s utility is no longer aligned with true welfare. This is

because the mechanism perceives a diﬀerent utility for agents j ,= i if agent

i changes his report. Agent i is rewarded according to this perceived utility

of agents j, rather than according to their true utility.

The previous paragraph showed that a generalization of the VCG mechanism

cannot be used to implement a ﬁrst best decision rule. The following result,

which is the main result of Jehiel and Moldovanu (2001), shows that in fact

no direct mechanism implements a ﬁrst best decision rule.

Proposition 8.1. Assume that any ﬁrst best rule q

∗

is such that interim

expected probabilities are diﬀerentiable, and, moreover, their derivative is

never zero:

∂Q

i

a

(θ

i

)

∂θ

i

b

,= 0

for every i ∈ I, a, b ∈ A and θ

i

∈ [0, 1]

K

. Assume also that there are some

agent i ∈ I and two alternatives a, b ∈ A such that:

α

i

ai

α

i

bi

,=

N

j=1

α

i

bj

N

j=1

α

i

aj

.

Then no ﬁrst best direct mechanism is Bayesian incentive compatible.

The ﬁrst condition in this Proposition assumes that the interim expected

probability of alternative a, conditioning on agent i’s signal, changes as

agent i’s signal value for some arbitrary other alternative b changes. This is

a regularity condition that is needed in the proof of Proposition 8.1 as pre-

sented below. It can be much weakened. For example, instead of requiring

it to be true for all types of agent i, it is evident from the proof below that

we might also require it to be true for some type of agent i.

CHAPTER 8. INTERDEPENDENT TYPES 159

The second condition says that there is at least some agent i, and a pair of

alternatives a, b, such that the relative weight that agent i attaches to his

signal for alternative a and his signal for alternative b is diﬀerent from the

relative weight that the social welfare function attaches to these two signals.

If we pick the weights in agents’ utility functions randomly from some con-

tinuous distribution, then this condition will be satisﬁed with probability

1. It is for this reason that Proposition 8.1 presents a generic impossibility

result.

Note the contrast between Jehiel and Moldovanu’s impossibility result and

the impossibility results that we presented earlier in private value settings.

Those earlier impossibility results only obtained when welfare maximization

is combined with other requirements, such as individual rationality and bud-

get balance. By contrast, the Jehiel Moldovanu impossibility results refer

only to welfare maximization, with no other requirements.

Proof. Suppose there were a ﬁrst best direct mechanism that is incentive

compatible. We shall present a proof that assumes that for every agent i

the interim expected utility function U

i

is twice continuously diﬀerentiable.

The proof in Jehiel and Moldovanu (2001) does not assume this.

The Envelope Theorem implies for every agent i and every a ∈ A:

∂U

i

∂θ

i

a

= α

i

ai

Q

i

a

(θ

i

). (8.7)

Now suppose we diﬀerentiate this expression again, this time with respect

to θ

i

b

for some b ∈ A. Then we obtain:

∂

2

U

i

∂θ

i

a

∂θ

i

b

= α

i

ai

∂Q

i

a

(θ

i

)

∂θ

i

b

. (8.8)

If we change the order of diﬀerentiation, we obtain similarly:

∂

2

U

i

∂θ

i

b

∂θ

i

a

= α

i

bi

∂Q

i

b

(θ

i

)

∂θ

i

a

. (8.9)

The fact that interim expected utility is twice continuously diﬀerentiable im-

plies that the order of diﬀerentiation does not matter (Schwarz’s Theorem).

CHAPTER 8. INTERDEPENDENT TYPES 160

Thus the two derivatives that we computed have to be the same.

α

i

ai

∂Q

i

a

(θ

i

)

∂θ

i

b

= α

i

bi

∂Q

i

b

(θ

i

)

∂θ

i

a

⇔

α

i

ai

α

i

bi

=

∂Q

i

b

(θ

i

)

∂θ

i

a

∂Q

i

a

(θ

i

)

∂θ

i

b

. (8.10)

Here, we could divide by the partial derivatives of the interim probabilities

because, by the ﬁrst assumption in Proposition 8.1, these derivatives are not

zero.

Now suppose agents’ utility functions were diﬀerent. Suppose they were

equal to social welfare. Then the given rule, without transfers, would also

be Bayesian incentive compatible. An argument like the one that we just

displayed would lead to the conclusion that the equation just derived is true

with the weights from agents’ original utility functions being replaced by

the weights from the social welfare function:

N

j=1

α

i

aj

N

j=1

α

i

bj

=

∂Q

i

b

(θ

i

)

∂θ

i

a

∂Q

i

a

(θ

i

)

∂θ

i

b

. (8.11)

We can now deduce from our results so far that:

α

i

ai

α

i

bi

=

N

j=1

α

i

aj

N

j=1

α

i

bj

. (8.12)

This has to be true for every i ∈ I, and any two alternatives a, b ∈ A. But

this contradicts the second condition in Proposition 8.1.

8.4 Characterizing All Incentive Compatible

Mechanisms

The impossibility result presented in the previous section is remarkable be-

cause it holds even if no individual rationality or budget balance. However,

as in the case of the impossibility results that we obtained in the private

value settings with individual rationality and budget balance, it is natural

to ask next which mechanisms would be second best if we wanted to max-

imize ex ante expected social welfare. We might also investigate revenue

CHAPTER 8. INTERDEPENDENT TYPES 161

maximizing mechanisms, or mechanisms that are chosen according to some

other objective function.

For all these questions it is important to have a characterization of all in-

centive compatible direct mechanisms. With independent signals, which is

the case that we are considering here, the characterizations that we devel-

oped for the case of private values generalize. In particular, if each agent’s

utility from the collective decision is a linear function of the agent’s own

type, then interim expected utilities are convex, and are determined by the

collective decision rule up to a constant. Moreover, collective decision rules

q can be part of a Bayesian incentive compatible mechanism if and only if at

the interim level they generate for each agent the sub-gradient of a convex

function. These characterizations are hard to use in practice.

Jehiel et al. (2006) study ex-post incentive compatibility instead of Bayesian

incentive compatibility. Ex-post incentive compatibility means that for ev-

ery realization of all other agents’ types each agent ﬁnds it in his interest

to report his type truthfully rather than distort it. If truth telling is an

ex-post equilibrium of a direct mechanism, it is a Bayesian equilibrium for

every belief that agents might hold about the other agents’ types, including

beliefs that are not product measures. With private values, truth telling is

an ex post equilibrium if and only if truth telling is a dominant strategy.

Thus, ex-post incentive compatibility is a generalization of dominant strat-

egy incentive compatibility to the case of interdependent valuations. Jehiel

et. al.’s main ﬁnding is that for generic utility functions only constant choice

rules that choose the same alternative for every type realization are incentive

compatible if each agent’s signal space is at least two-dimensional.

A question one might ask about this result is why dictatorial choice rules

are not incentive compatible. Suppose some agent i is allowed to pick the

collective choice based on i’s signal only, and suppose there are no transfers.

Agent i will then pick the alternative that maximizes his expected utility,

given his information. Jehiel et. al.’s result implies that such a rule is not

ex-post incentive compatible. The reason is as follows: all agents other than

i would be willing to reveal their private information, as it doesn’t aﬀect the

collective decision. Once this information is revealed, agent i has an ex post

incentive to deviate, and to change his decision.

Note that ex post incentive compatibility thus requires that we can uniquely

predict an agent’s choice independent of what this agent’s beliefs about the

other agent’s types are. If we relaxed this requirement, and allowed that

choices depend on beliefs, then more rules become feasible.

CHAPTER 8. INTERDEPENDENT TYPES 162

The logic of Jehiel et. al.’s proof of their result is as follows. Consider any

two alternatives, a and b, and consider for two agents i and j the set of

signals where a is chosen, and the set of signals where b is chosen. Suppose

that these two sets have a boundary in common. Along this boundary,

both agents, agent 1 and agent 2 have to be indiﬀerent between a and b.

Moreover, along this boundary, agent 1’s payment is determined by agent

2’s report, and agent 2’s payment is determined by agent 1’s report. Thus,

ﬁxing the other agent’s signal, payments along the boundary are constant.

But for generic utility functions it is impossible to ﬁnd surfaces such that for

ﬁxed payments both agents are indiﬀerent between two alternatives along

these surfaces.

8.5 Remarks on the Literature

The example in Section 2 is a special case of an example that appears in

Jehiel and Moldovanu (2001). This paper is also the main reference for

Proposition 8.1. However, the short proof of Proposition 8.1 that I have

given is taken from Jehiel and Moldovanu (2005). I have also used Jehiel

et al. (2006).

8.6 Problems

Chapter 9

Robust Mechanism Design

To be written.

163

Chapter 10

Dynamic Mechanism Design

To be written.

164

Chapter 11

Conclusion

To be written.

165

Bibliography

Atila Abdulkadiroglu and Simon Loertscher. Dynamic house allocations.

2007. 152

William J. Adams and Janet L. Yellen. Commodity bundling and the burden

of monopoly. Quarterly Journal of Economics, 90:475–498, 1976. 30

Kenneth Arrow. Social Choie and Individual Values. Wiley, New York,

second edition, 1963. 151

Salvador Barbera. An introduction to strategy-proof social choice functions.

Social Choice and Welfare, 18:619–653, 2001. 151, 152

Sushil Bikhchandani, Shurojit Chatterji, Ron Lavi, Ahuva Mu’alem, Noam

Nisan, and Arunava Sen. Weak monotonicity characterizes deterministic

dominant strategy implementation. Econometrica, 74:1109–1132, 2006.

112, 120

Patrick Bolton and Mathias Dewatripont. Contract Theory. MIT Press,

Cambridge and London, 2005. 2, 30

Tilman B¨orgers and Peter Norman. A note on budget balance under interim

participation constraints: The case of independent types. Economic The-

ory, 2008. 77

Tilman B¨orgers and Peter Postl. Eﬃcient compromising. 2005. 150, 151

Kim Sau Chung and Jeﬀrey C. Ely. Foundations of dominant strategy mech-

anisms. Review of Economic Studies, 74:447–476, 2007. 80

Edward H. Clarke. Multipart pricing of public goods. Public Choice, 8:

19–33, 1971. 120

166

BIBLIOGRAPHY 167

Jacques Cr´emer and Richard McLean. Full extraction of the surplus in

bayesian and dominant strategy auctions. Econometrica, 56:1247–1257,

1988. 128, 137

Claude d’Aspremont and Louis-Andre Gerard-Varet. Incentives and incom-

plete information. Journal of Public Economics, 11:25–45, 1979. 76

Joseph L. Doob. Measure Theory. Springer, 1994. 16

Bhaskar Dutta, Hans Peters, and Arunava Sen. Strategy-proof cardinal

decision schemes. Social Choice and Welfare, 28:163–179, 2007. 152

Joel Franklin. Methods of Mathematical Economics. Springer, Heidelberg

and New York, 1980. 131, 137

Drew Fudenberg and Jean Tirole. Game Theory. MIT Press, Cambridge

and London, 1993. 2, 4, 33

Alan Gibbard. Manipulation of voting schemes. Econometrica, 41:587–601,

1973. 142

Alan Gibbard. Manipulation of schemes that mix voting with chance. Econo-

metrica, 45:665–681, 1977. 152

Alan Gibbard. Straightforwardness of game forms with lotteries as outcomes.

Econometrica, 46:595–614, 1978. 152

Jerry Green and Jean-Jacques Laﬀont. Characterization of satisfactory

mechanisms for the revelation of preferences for public goods. Econo-

metrica, 45:727–738, 1977. 120

Theodore Groves. Incentives in teams. Econometrica, 41:617–631, 1973. 120

Werner G¨ uth and Martin Hellwig. The private supply of a public good.

Journal of Economics, Supplement 5, pages 121–159, 1986. 77

Kathleen Hagerty and William Rogerson. Robust trading mechanisms. Jour-

nal of Economic Theory, 42:94–107, 1987. 96

Aviad Heifetz and Zvika Neeman. On the generic (im)possibility of full

surplus extraction in mechanism design. Econometrica, 74:213–233, 2006.

136

Benjamin Hermalin. Lecture Notes for Economics. 2005. URL http://

faculty.haas.berkeley.edu/hermalin/. 30

BIBLIOGRAPHY 168

Bengt Holmstr¨om. Groves’ scheme on restricted domains. Econometrica,

47:1137–1144, 1979. 120

Aanund Hylland. Strategy proofness of voting procedures with lotteries as

outcomes and inﬁnite sets of strateges. 1980. 150

Philippe Jehiel and Benny Moldovanu. Eﬃcient design with interdependent

valuations. Econometrica, 69:1237–1259, 2001. 156, 158, 159, 162

Philippe Jehiel and Benny Moldovanu. Allocative and informational exter-

nalities in auctions and related mechanisms. In Proceedings of hte 9th

World Congress. Econometric Society, London, 2005. 162

Philippe Jehiel, Moritz Meyer ter Vehn, Benny Moldovanu, and Bill Zame.

The limits of ex-post implementation. Econometrica, 74:585–611, 2006.

161, 162

Grigory Kosenok and Sergei Severinov. Individually rational, budget-

balanced mechanisms and allocation of surplus. Journal of Economic

Theory, 140:126–161, 2008. 135, 137

Vijay Krishna. Auction Theory. Academic Press, San Diego etc., 2002. 38,

42, 76

Vijay Krishna and Eliot Maenner. Convex potentials with an application to

mechanism design. Econometrica, 69:1113–1119, 2001. 99, 120, 123

Jean-Jacques Laﬀont and David Martimort. Mechanism design with collu-

sion and correlation. Econometrica, 68:309–342, 2000. 136

David G. Luenberger. Optimization By Vector Space Methods. John Wiley

and Sons, Inc., New York etc., 1969. 56, 72

Alejandro M. Manelli and Daniel R. Vincent. Multidimensional mechanism

design: Revenue maximization and the multiple-good monopoly. Journal

of Economic Theory, 137:153–185, 2007. 30

Paul Milgrom. Putting Auction Theory to Work. Cambridge University

Press, Cambridge etc., 2004. 76, 120

Dilip Mookherjee and Stefan Reichelstein. Dominant strategy implemen-

tation of bayesian incentive compatible allocation rules. Journal of Eco-

nomic Theory, 56:378–399, 1992. 96

BIBLIOGRAPHY 169

Herve Moulin. On strategy-proofness and single peakedness. Public Choice,

35:437–455, 1980. 152

Roger Myerson. Optimal auction design. Mathematics of Operations Re-

search, 6:58–73, 1981. 76

Roger B. Myerson and Mark A. Satterthwaite. Eﬃcient mechanisms for

bilateral trading. Journal of Economic Theory, 29:256–281, 1983. 65, 77

Zvika Neeman. The relevance of private information in mechanism design.

Journal of Economic Theory, 117:55–77, 2004. 136

Efe A. Ok. Real Analysis with Economic Applications. Princeton University

Press, Princeton and Oxford, 2007. 17

Philip Reny. Arrow’s theorem and the gibbard-satterthwaite theorem: A

uniﬁed approach. Economics Letters, 70:99–105, 2001. 142

Jacques Robert. Continuity in auction design. Journal of Economic Theory,

55:169–179, 1991. 136

Kevin Roberts. The characterization of implementable choice rules. In Jean-

Jacques Laﬀont, editor, Aggregation and Revelation of Preferences, pages

321–348. North Holland, Amsterdam, 1979. 114, 120

Jean-Charles Rochet. A necessary and suﬃcient condition for rationaliz-

ability in a quasi-linear context. Journal of Mathematical Economics, 16:

191–200, 1987. 103, 120

Tyrell Rockafellar. Convex Analysis. Princeton University Press, Princeton,

1970. 106, 107, 120

Alvin Roth. Deferred acceptance algorithms: History, theory, practie, and

open questions. International Journal of Game Theory, 36:537–569., 2008.

152

Halsey Royden. Real Analysis. Prentice Hall, 3rd edition, 1988. 4, 12, 13,

16

Mark Satterthwaite. Strategy-profness and arrow’s conditions: Existence

and correspondence theorems for voting procedures and social welfare

functions. Journal of Economic Theory, 10:187–217, 1975. 142, 151

William Vickrey. Counterspeculation, auctions, and competitive sealed ten-

ders. Journal of Finance, 16:8–37, 1961. 120

Index

allocation rule, 34

Bayesian incentive compatible, 123

bilateral trade, 64

bounded type space, 110

canonical auctions, 84

Cr´emer McLean condition, 128

cyclical monotonicity, 103, 124

decision rule, 48, 98

dictatorial direct mechanism, 141

direct mechanism, 9, 19, 34, 47

dominant strategy, 80

dominant strategy incentive compat-

ibility, 83, 98

equivalent mechanisms, 49

ex ante budget balance, 48, 125

ex post budget balance, 48, 117

ex post individual rationality, 83, 116

ex post revenue equivalence, 84, 99

extreme point, 16

extreme point theorem, 17

Farkas’ Alternative, 131

ﬁrst best, 51

ﬁxed price mechanism, 96

ﬂexible decision rule, 114

Gibbard-Satterthwaite Theorem, 141

identiﬁability condition, 135

incentive compatibility, 11, 48

independent private values, 32

indirect mechanism, 10

individual rationality, 9, 11, 48

information rent, 15

interim expected transfer, 37

interim expected utility, 37

interim revenue equivalence, 124

measurability, 9

mechanism, 33

mixed bundling, 28

monotonicity, 109, 142

Myerson Satterthwaite Theorem, 68

non-transferrable utility, 139

optimal auction, 41

participation constraint, 9

payoﬀ equivalence, 12

pivot mechanism, 52, 68

positive association of diﬀerences, 113

public goods problem, 46

regular distribution function, 24, 41,

72

revelation principle, 10, 35, 82

revenue equivalence theorem, 13, 38

rich type spaces, 112

second best mechanism, 55, 71

set-monotonicity, 143

single dimensionality condition, 109

single-peakedness, 148

170

INDEX 171

social choice function, 140

strategy proofness, 141

subgradient, 106

type, 7

unanimity, 143

Vickrey-Clarke-Groves mechanism, 53,

100

weak monotonicity, 102

welfare, 47

Contents

Preface 1 Introduction 2 Screening 2.1 Introduction . . . . . . . . . . . . 2.2 Pricing a Single Indivisible Good 2.3 Nonlinear Pricing . . . . . . . . . 2.4 Bundling . . . . . . . . . . . . . 2.5 Comments on the Literature . . . 2.6 Problems . . . . . . . . . . . . . v 1 6 6 6 18 27 30 30

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

3 Bayesian Mechanism Design 31 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 3.2 Single Unit Auctions . . . . . . . . . . . . . . . . . . . . . . . 31 3.2.1 Set-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 3.2.2 Mechanisms, Direct Mechanisms, and the Revelation Principle . . . . . . . . . . . . . . . . . . . . . . . . . 33 3.2.3 Characterizing Incentive Compatibility and Individual Rationality . . . . . . . . . . . . . . . . . . . . . . 37 3.2.4 Expected Revenue Maximization . . . . . . . . . . . . 39 3.2.5 Maximizing Welfare . . . . . . . . . . . . . . . . . . . 42 3.2.6 Numerical Examples . . . . . . . . . . . . . . . . . . . 43 3.3 Public Goods . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 3.3.1 Set-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 3.3.2 Incentive Compatible and Individually Rational Direct Mechanisms . . . . . . . . . . . . . . . . . . . . . 47 3.3.3 Ex ante and Ex post Budget Balance . . . . . . . . . 48 3.3.4 Welfare Maximization . . . . . . . . . . . . . . . . . . 50

i

CONTENTS 3.3.5 Proﬁt Maximization . 3.3.6 A Numerical Example Bilateral Trade . . . . . . . . 3.4.1 Setup . . . . . . . . . 3.4.2 Direct Mechanisms . . 3.4.3 Welfare Maximization 3.4.4 Proﬁt Maximization . 3.4.5 A Numerical Example Comments on the Literature . Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

ii 58 59 64 64 65 67 74 75 76 77

3.4

3.5 3.6

4 Dominant Strategy Mechanisms 78 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 4.2 Single Unit Auctions . . . . . . . . . . . . . . . . . . . . . . . 81 4.2.1 Set-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . 81 4.2.2 Mechanisms, Direct Mechanisms, and the Revelation Principle . . . . . . . . . . . . . . . . . . . . . . . . . 81 4.2.3 Characterizing Dominant Strategy Incentive Compatibility and Ex Post Individual Rationality . . . . . . . 83 4.2.4 Canonical Auctions . . . . . . . . . . . . . . . . . . . . 84 4.3 Public Goods . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 4.3.1 Set-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 4.3.2 Direct Mechanisms . . . . . . . . . . . . . . . . . . . . 87 4.3.3 Characterizing Dominant Strategy Incentive Compatibility and Ex Post Individual Rationality . . . . . . . 88 4.3.4 Canonical Mechanisms . . . . . . . . . . . . . . . . . . 90 4.3.5 Ex Post Budget Balance . . . . . . . . . . . . . . . . . 91 4.4 Bilateral Trade . . . . . . . . . . . . . . . . . . . . . . . . . . 93 4.4.1 Set-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 4.4.2 Dominant Strategy Incentive Compatible and Ex Post Individually Rational Direct Mechanisms . . . . . . . 93 4.4.3 Canonical Mechanisms . . . . . . . . . . . . . . . . . . 94 4.4.4 Ex Post Budget Balance . . . . . . . . . . . . . . . . . 95 4.5 Comments on the Literature . . . . . . . . . . . . . . . . . . . 96 4.6 Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 5 Dominance: General Theory 97 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 5.2 Set-Up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 5.3 Ex Post Revenue Equivalence . . . . . . . . . . . . . . . . . . 99

. . . . . .11 Remarks on the Literature . . . . . . 6. . . . . . . . .CONTENTS 5. . .7 Problems . Characterizing All Incentive Compatible Decision Rules . . . . . 5.7 Single Dimensional Type Spaces . . . . . 6. . . . . . . . . .6 Problems . . . . . . . .2. . . . . . . . . . . . . .1 Introduction . . . . 143 7. . . . . . . . . . . . . . .2 Set-Up . . . . . . . . . . . . .5 Discussion . . . . 6. . . . . . . . . .1 Introduction . . . . . . . . . . . . .2 The Gibbard Satterthwaite Theorem . . . . . . . . . . . . . . . . . . . . . . . .3 Dominant Strategy Incentive Compatibility On Restricted Domains . . . . . . 5. . . . . . 154 . . . . . . . . . . . . . . . 153 8. . . . . . . . . . . . . . . . . . 152 8 Interdependent Types 153 8.4 5. . . . . 6. . .5 5. 5. . . . . . . 6. . . . .3 Characterizing Bayesian Incentive Compatibility 6. . . . . . . . . . . . . . . 6. . . . . . . . .1 Framework . . . . . . . .5 Remarks on the Literature . . . . . . . . . . 6. . . . . . .4. . . . 6. . . . . . . .3 Independent Types . . . . . . . . . . 5. . 6 Bayesian Design: General Theory 6. . . . . . . . . . . . . . . . . . . . . . . . . . . . iii 100 101 106 107 112 113 116 120 120 121 121 122 123 126 126 126 128 132 134 136 137 137 . . . 140 7. . .2 Failure of Revenue Equivalence . . .4. . . . . . . . . . . .4. . . . . . . . . . . . . . . . 147 7. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Every monotone direct mechanism is dictatorial . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 7. . .2 Statement of the Result and Outline of the Proof . . . . . . . . . . . . . . . . .1 Introduction . . .6 Remarks on the Literature . All Incentive Compatible Decision Rules When Outcomes are Lotteries . . . . . . . . . . . . . . . . . . . . . . . 151 7. . . . . . . . . . .4. . 5. . . . . . . . . . . . . .5 Individual Rationality and Budget Balance . . .9 Positive Association of Diﬀerences . . . . . 141 7. . . . . . . . . . . . . . . . . . . . .4. . . . . . . . . . . . . 140 7. . .4 Correlated Types . . . . . 6. . . .12 Problems . . . . . . . . . . . . . . .2. . . . .4 Bayesian Incentive Compatibility . . . . . . . . . . . . . . . . . . . . .4 A Numerical Example . . . . . .2 An Example . . . 150 7. . . . . . . . 7 Non-Transferrable Utility 139 7. . . . . . .10 Individual Rationality and Budget Balance . . . . . . . . . . . 6. . . 5. . . . . .8 Suﬃciency of Weak Monotonicity . . . . . . . . . . .1 Set Up . . .2. . . . . .6 Implementing Eﬃcient Decision Rules . . . . . . . . . . .

. . . . . .CONTENTS 8. . . . . . . . . . . .3 8. Characterizing All Incentive Compatible Mechanisms . . . . . . . . . . . . . . . . . iv 156 160 162 162 163 164 165 9 Robust Mechanism Design 10 Dynamic Mechanism Design 11 Conclusion . . . . . . .5 8. . . . . Problems . . . . . . Remarks on the Literature . . . . . . . . . . . . . .6 Impossibility of Implementing Welfare Maximizing Decision Rules . . . . . . . . . . . . .4 8. . . . . . . . . .

to mechanisms that provide each agent with a dominant strategy. These two approaches have been the prevailing approaches of the classic theory of mechanism design. such as Myerson’s theorem on expected revenue maximizing auctions. and which features of optimal mechanism design are identical to the corresponding features of optimal screening. A second objective of these notes is to take the reader in some selected areas to the recent frontier of research. Myerson and Satterthwaite’s theorem on the impossibility of ex post eﬃcient bilateral trade with asymmetric information. and Gibbard and Satterthwaite’s theorem on the non-existence of dominant strategy voting mechanisms. The theory of screening is sometimes not regarded as part of the theory of mechanism design because it constructs an incentive scheme for only one agent rather than multiple interacting agents. The formal part of the notes starts in Chapter 2 with an explanation of the theory of screening. I treat these two approaches side by side with the intention to alert the reader to the potential problems of both v . An organizing principle of a large proportion of the book is that I compare mechanisms that are constructed with the anticipation that agents will play a Bayesian equilibrium in some given information structure with small type spaces. My hope is that by juxtaposing the theory of screening and the theory of mechanism design I can help the reader understand which features of optimal mechanism design are due to strategic interaction. the theories that are covered in Chapter 2 are intimately linked to the theories of optimal mechanisms that are explained in later chapters.Preface This manuscript consists of lecture notes that I have used to teach courses on mechanism design for graduate students at the University of Michigan and at the Institute for Advanced Studies in Vienna. The primary objective of these notes is to give rigorous. in particular in Chapter 3. but accessible explanations of classic results in the theory of mechanism design. However.

PREFACE

vi

approaches, and to prepare the grounds for a later discussion of robust mechanism design. The modern theory of robust mechanism design has moved the discussion beyond the traditional dichotomy of Bayesian mechanism design and dominant strategy mechanism design. The manuscript assumes that readers have had a prior course in game theory, and that they have basic knowledge of real analysis. Other than this, however, I have sought to keep the prerequisites minimal. This draft is incomplete in many ways. Chapters 9, 10 and 11 are still to be written. Earlier chapters need further revisions. I also plan to include further problems at the end of each chapter. Any suggestions for the further development of this manuscript are always welcome. Please write to: tborgers@umich.edu. I am very grateful to Stefan Behringer who read an earlier draft of this manuscript and pointed out many errors. All remaining errors are my own.

Tilman B¨rgers o

Chapter 1

Introduction

Suppose you want to sell your house, and your realtor has identiﬁed several potential buyers who are willing to pay your ask price. You might then wish to conduct an auction among these buyers to obtain a higher price. There are many diﬀerent auction formats that you could use: For example, each buyer could be asked to send in one binding and ﬁnal bid. Alternatively, buyers could bid in several rounds, and in each round they are all informed about the highest bid of the previous round, and are then asked to revise their bids. You could also use some combination of these formats. How should you choose among diﬀerent auction formats? This is one of the questions that the theory of mechanism design aims to answer. Now imagine that you and your colleagues are considering to buy a new refrigerator to be kept at work in which you can store the food that you bring from home. While everyone is in favor, it is not so clear how much money people are willing to contribute. How can you ﬁnd out whether the sum of the amounts that everyone is willing to contribute covers the costs of the refrigerator? You could ask everyone to submit pledges simultaneously, and then see whether the sum of the pledges covers the expense. Alternatively, you could go around, and tell each colleague how much everyone else has pledged so far. Or you could divide the costs by the number of colleagues involved, and commit to buying the refrigerator only if everyone is willing to pay their share. Which of these procedures is best? Again, this is one of the questions that the theory of mechanism design addresses. Each of the procedures that you might consider in the two examples above creates a strategic game in the sense of non-cooperative game theory among the participants. Participants in these procedures will understand that the outcome will depend not only on their own choices but also on 1

CHAPTER 1. INTRODUCTION

2

others’ choices, and that therefore their own optimal strategy may depend on others’ strategies. The theory of mechanism design therefore builds on the theory of games (Fudenberg and Tirole (1993)). Game theory takes the rules of the game as given, and makes predictions about the behavior of strategic players. The theory of mechanism design is about the optimal choice of the rules of the game. We are more frequently involved in the design of rules for games than might be obvious at ﬁrst sight. How should shareholders’ vote be conducted? How should promotion procedures in companies be organized? What are optimal prenuptial agreements? All these questions are about the optimal rules of games. The theory of mechanism design seeks to study the general structure underlying all these applications, but it also considers a number of particularly prominent applications in detail. There are at least two reasons why we study mechanism design. Firstly, the theory of mechanism design aids in practice the designers of real world mechanisms. The theory of optimal auctions, for example, is frequently invoked in discussions about the design of government and industry auctions. Secondly, we can explain why real world institutions are as they are by interpreting them as rational choices of those who designed them. For example, we might seek to explain the use of auctions in some house sales, and the use of posted prices in other house sales by appealing to the theory of mechanism design which indicates that posted prices are optimal in some circumstances and auctions are optimal in other circumstances. The incentives created by the choice of rules of games are central to the theory of mechanism design. Incentives are also at the center of contract theory (Bolton and Dewatripont (2005)). At ﬁrst sight the distinction between the theory of mechanism design and contract theory is simple: In contract theory we study the optimal design of incentives for a single agent. In mechanism design we study the optimal design of incentives for a group of agents, such as the buyers in our ﬁrst example, and the colleagues in the second example. Contract theory therefore, unlike the theory of mechanism design, does not have to deal with strategic interaction. The relation between contract theory and the theory of mechanism design is more subtle. One part of the theory of mechanism design is, in fact, a straightforward extension of insights from contract theory. This is surprising because one might have expected the element of strategic interaction, that is present in mechanism design, but absent in contract theory, to create substantial new problems. It is interesting and conceptually important to

Other parts of mechanism design are. the hidden information that we seek to ﬁnd out is the colleagues’ true willingness to pay for an oﬃce refrigerator. information and actions.CHAPTER 1. Mechanism design. For example. and we shall address this issue in detail below. It is “hidden information. when considering how to auction your house. is about hidden information. you might restrict . promotion schemes within a company set work incentives for a group of employees. not hidden actions.”) The distinction is easily explained within the context of contracts for health insurance. By introducing deductibles an insurance company might seek to maintain your incentives to look after your own health and thus to alleviate moral hazard problems. and consider as large a set of rules as possible.” Whether you exercise regularly. by oﬀering you a menu of insurance contracts and observing your choice from this menu an insurance company might be able to infer information about your health risks that you might wish to conceal from the company. and the theory that deals with it is concerned with the optimal choice of rules for a game. like the theory of mechanism design. hidden action with many agents involved is a subject of great interest.” Both hiddens. or take it a little more easy once you have bought complete insurance coverage for heart surgery. In our second example in this section. as traditionally understood. matter for contract design. The close parallel between contract theory and mechanism design applies only to some parts of mechanism design. Contract theory has traditionally been divided into two parts: the theory of hidden information (also referred to as the theory of “adverse selection. unrelated to contract theory. or we might try to cast our net wide. For example. In voting. the hidden information that we seek to ﬁnd out is individuals’ true ranking of diﬀerent alternatives or candidates. In our ﬁrst example.”) and the theory of hidden action (also referred to as the theory of “moral hazard. of course. but that the company from which you are trying to buy health insurance does not know. and the optimal choice of such schemes is an important subject of economic theory. However. with multiple agents. it is not the subject of mechanism design as this term has traditionally been interpreted. When choosing the rules for the strategic interaction among agents we might restrict ourselves to a small subset of all conceivable rules. INTRODUCTION 3 understand why this is not the case. It is a “hidden action. Whether you have experienced sever chest pain in the past is something that you know. the hidden information that the seller of a house seeks to ﬁnd out is the buyers’ true willingness to pay for the house. For example. Of course. is a choice that you make that your insurance company does not observe unless it puts into place a surveillance operation.

Indeed traditionally mechanism design has been understood as the ﬁeld in which this grand optimization among all conceivable procedures is considered.CHAPTER 1. We shall also assume a basic knowledge of real analysis that can. and take for granted that the auction will proceed with all potential buyers submitting their bids simultaneously. These notes are meant for second year graduate students. and you have come to the conclusion that an auction with just one round of bidding is optimal. one of the losing bidders approaches you with a new and improved bid that is higher than the winning bid in the auction. The mechanism designer will not deviate from them. be acquired from Royden (1988). not just auctions. After the highest bid has been revealed. once announced.” As we mentioned above. but. Alternatively. You would then focus on the choice of only one parameter in a much larger set of possible choices. for example. Will you accept? This is an obvious temptation. and this line of research has been very productive. In contract theory. are set in stone. the classic Bayesian theory of optimal mechanism design can be viewed as . Fudenberg and Tirole (1993) contains more than enough material for this course. Chapter 2 presents some material from the theory of optimal contract design with hidden information. and these arguments have assumed full commitment by the mechanism designer. This is obviously a strong assumption. also simultaneous negotiations with all buyers that follow some predetermined format. It is also likely to be an interesting line of research in mechanism design. you might consider all conceivable ways of proceeding. for example. This theory is also referred to as the theory of “monopolistic screening. In these notes we shall stick to this interpretation of the theory of mechanism design. are you still conducting an auction with just a single round of bidding? In these notes we shall assume that the mechanism designer has full commitment power. The rules. We do not consider this line of argument in these notes because we want to maintain a focus on the central arguments of the traditional theory of mechanism design. but if you accept later bids. the mechanism designer will absolutely refuse to renegotiate after the auction results have been revealed. INTRODUCTION 4 attention to the choice of the minimum bid. In our example. It has been one of the accomplishments of the theory of mechanism design to develop a framework in which one can ﬁnd the optimal rules of the game among all conceivable rules. The notes assume that students have a good understanding of game theory. We proceed as follows. Suppose you have considered all possible rules for proceeding with your house sale. much attention has been given to the optimal design of contracts if full commitment cannot always be achieved.

We present this approach in Chapters 4. In Chapters 6. and in Chapter 5.4. and therefore the literature has sought to develop mechanisms that require the mechanism designer to know less about agents’ beliefs. This theory is built on several restrictive assumptions which we shall discuss in detail. and then in general. Like robust mechanism design. Chapter 10 presents models of mechanism design that apply to dynamic contexts. In Chapter 9 we return to the issue of what the mechanism designer knows about agents’ beliefs about each other and investigate more modern approaches to this problem which do not necessarily require the construction of a dominant strategy mechanism. INTRODUCTION 5 an extension of the theory of monopolistic screening. in general. this is an area of current research interest. We begin with the theory of monopolistic screening to make this connection transparent. Chapters 3 and 6 then review the classic Bayesian theory ﬁrst in the context of some prominent examples. One of these assumptions is that for given rules of the mechanism agents play a Bayesian equilibrium of the mechanism for a particular speciﬁcation of agents’ beliefs about each others’ private information. Chapter 11 concludes the notes with a brief review and an outlook. 7 and 8 we relax other assumptions of the classic model. The classic approach to this problem is to seek dominant strategy mechanisms. .CHAPTER 1. and also because the basic results are most easily understood in the simple context of monopolistic screening. in the context of examples. This assumption might attribute more information to the mechanism designer than is realistic.

Chapter 2

Screening

2.1 Introduction

Important parts of the theory of mechanism design are the multi-agent extension of the theory of screening. We begin by explaining three examples from the theory of screening. We use these examples to introduce some topics and techniques that are also important in mechanism design. By introducing these topics and techniques in the context of screening, we explain them in the simplest possible context.

2.2

Pricing a Single Indivisible Good

A seller seeks to sell a single indivisible good. The seller herself does not attach any value to the good. Her objective is to maximize the expected revenue from selling the good. She is thus risk neutral. There is just one potential buyer. The buyer’s von Neumann Morgenstern utility if she purchases the good and pays a monetary transfer t to the seller is: θ − t. The buyer’s utility1 if she does not purchase the good is zero. Here, θ > 0 is a number that we can interpret as the buyer’s valuation of the good, because our assumptions imply that the buyer is indiﬀerent between paying θ and obtaining the good, and not obtaining the good. Two aspects of the assumptions about the buyer’s utility deserve emphasis. Firstly, we have assumed that the buyer’s utility is the sum of the utility derived from the good, if it is purchased, and the disutility resulting

1 From now on, when we refer to “utility” we shall mean “von Neumann Morgenstern utility.”

6

CHAPTER 2. SCREENING

7

from the money payment. A more general formulation would write utility as u(I, t) where I is an indicator variable that is 1 if the buyer purchases the good and 0 otherwise. Our assumption that u can be written as the sum of θ and −t is usually described as the assumption that utility is “additively separable.” Additive separability of the utility function implies that the buyer’s utility from consuming the good is independent of the amount of money that he pays for it, and that the buyer’s disutility from money payments is independent of whether or not he owns the good. One can easily think of real world contexts in which these assumptions are probably violated. But we shall stick to these assumptions for most of these notes. Much of the classical theory of mechanism design is based on these assumptions. We shall consider a more general case in Chapter 7. The second aspect of our assumptions about the buyer’s utility that deserves emphasis is that we have assumed that the buyer is risk neutral with respect to money, that is, his utility is linear in money. Like additive separability this is a very restrictive assumption, but much of the classical theory of mechanism design makes this assumption. The more general model of Chapter 7 will relax this assumption, too. We now introduce a crucial assumption about information. It is that the value of θ is known to the buyer, but it is not known to the seller. This seems plausible in many contexts. Buyers often know better than sellers how well some particular product meets their preferences. We shall refer below to θ as the buyer’s “type.” As if often done in economic theory we shall assume that the seller has a subjective probability distribution over possible values of θ. This probability distribution can be described by a cumulative distribution function F . We shall assume that F has a density f . Moreover, we shall assume that the support of F , that is, the smallest closed set that has probability 1, is an ¯ ¯ interval [θ, θ], where 0 ≤ θ < θ. For technical convenience we shall assume ¯ that the density is strictly positive on the support: f (θ) > 0 for all θ ∈ [θ, θ]. We can now think of θ as a random variable with the cumulative distribution function F the realization of which is observed by the buyer but not by the seller. Our interest is in procedures for selling the good which the seller should adopt to maximize expected proﬁts. One obvious way would be to pick a price p and to say to the buyer that he can have the good if and only if he is willing to pay p. This is the selling procedure that we study in elementary microeconomics. Suppose the seller picks this procedure. Which

CHAPTER 2. SCREENING

8

price p should she choose? The probability that the buyer’s value is below p is given by the value of the cumulative distribution function F (p). The probability that the buyer’s value is above p, and hence that he accepts a price oﬀer p is 1 − F (p). Expected revenue is therefore p(1 − F (p)), and the optimal strategy for the seller is to pick some price that maximizes p(1 − F (p)). Note that this is just the monopoly problem from elementary microeconomics with demand function 1 − F (p). In this very simple context we shall now ask a straightforward question: “Is picking a price p really the best the seller can do?” What else could the seller do? The seller could, for example, negotiate with the agent. The seller could oﬀer the agent a lottery where in return for higher or lower payments the buyer could be given a larger or smaller chance of getting the object. One can think of many other procedures that the seller might adopt to sell the good. Is setting a price really the best procedure? To make our question more precise we have to be speciﬁc about which procedures the seller can commit to. We shall assume that the seller has unlimited powers of commitment: the seller can commit to an arbitrary extensive game tree where the players are the seller and the buyer, and where each terminal history is associated with a probability distribution over {0, 1} × R. The interpretation of such a probability distribution is that it describes the probability with which the object is transferred to the buyer together with a probability distribution over transfer payments by the buyer. The seller will also ﬁnd it to his advantage to commit to a strategy for himself, and to announce this strategy to the buyer before play begins. An example would be that the seller announces that he will bargain over the price with the seller, and that he announces in advance that he will turn down certain oﬀers by the buyer. We shall assume that like the seller’s ability to commit to a game, also the seller’s ability to commit to a strategy in the game are unlimited. Once the seller has committed to a game and a strategy, the buyer will choose his own strategy in the game. We shall assume that the buyer chooses his own strategy, knowing the value of θ, to maximize his expected utility. Our question is now clearer: If the seller can commit to an arbitrary extensive game, and if she can also commit to a strategy for playing that game, which choice maximizes her expected revenue? In particular, is her expected revenue maximized by committing to a price, and by committing to selling the good at that price whenever the buyer is willing to pay the price?

But the buyer wouldn’t want to participate in such a mechanism. 2 . SCREENING 9 A rather silly answer to our ﬁrst question could be this: The seller should choose the game according to which the buyer has only a single choice that is such that the buyer does not get the object but nonetheless has to pay $x were x could be some arbitrarily large number. 2 A “direct mechanism” consists of functions q and t where: ¯ q : [θ. θ] → [0. this looks like a hard problem as the seller’s choice set is very large. 1] and ¯ t : [θ.1. However. for every type θ the buyer will need to ﬁnd that when he chooses his expected utility maximizing strategy his expected utility will be at least zero. Note that the payment is deterministic. There are many extensive games that the seller could consider. Clearly.” Our objective is thus to study the optimization problem in which the seller’s choice variables are an extensive game and a strategy in that game. the utility that he would obtain if he did not buy the good and did not pay anything.CHAPTER 2. and the buyer has to pay the seller t(θ) if she reports that her type is θ. in which the seller’s objective function is expected revenue. We will rule out such mechanisms by requiring that the buyer for every value of θ ﬁnds it in her interest to participate in the game proposed by the seller. It is not conditional on the event that the buyer obtains the good. We can restrict attention to a small set of mechanisms. the seller can extract an arbitrarily large amount of money from the buyer. The seller commits to transferring the good to the buyer with probability q(θ) if the buyer reports that her type is θ. We shall refer to this constraint as the “individual rationality” constraint.” Deﬁnition 2. it is also called “participation constraint. and in which the constraint on the seller’s choice is the individual rationality constraint. At ﬁrst sight. The interpretation is that in a direct mechanism the buyer is asked to report θ. this deﬁnition should include the requirement that the functions q and t are Lebesgue-measurable. if we allow such procedures. a simple. θ] → R. We omit measurability requirements throughout this text which will lead to a small cost in terms of rigor in some places. In other words. called “direct mechanisms. It would make no diﬀerence if we allowed the payment to be To be precise. yet crucial result enables us to get a handle on this optimization problem. Sometimes.

a very simple version of the famous “Revelation Principle. that is. for every θ. when pretending to be some type θ = θ. i. Note that under this strategy. Moreover. θ] the probability q(θ) and the payment t(θ) equal the probability of purchase and the expected payment that result in Γ if the buyer plays her optimal strategy σ.” shows that it is without loss of generality to restrict attention to direct mechanism. All our analysis below would go through if we interpreted t(θ) as the buyer’s expected payment conditional on θ. the buyer with type θ obtains in the mechanism Γ the same expected utility as in the mechanism Γ when choosing strategy σ(θ). θ]. SCREENING 10 random. For every θ ∈ [θ.CHAPTER 2. θ] deﬁne q(θ) and t(θ) as required by (ii) in Proposition 2. We prove the result by showing that for this direct mechanism the strategy σ (θ) = θ. Using this . pairs of functions q and t. where the buyer ﬁnds it always optimal to truthfully report her type.e. we deﬁne deﬁne the buyer’s expected utility u(θ) conditional on her type being θ by: u(θ) = θq(θ) − t(θ). θ] → [θ. ¯ (ii) For every θ ∈ [θ. is optimal for the buyer.1 (Revelation Principle). Given a direct mechanism. The next result. For every mechanism Γ and every optimal buyer strategy σ in Γ there is a direct mechanism Γ and an optimal buyer strategy σ in Γ such that: (i) The strategy σ satisﬁes: ¯ σ (θ) = θ for every θ ∈ [θ. σ prescribes telling the truth. The Revelation Principle allows us to simplify our analysis greatly because it shows that without loss of generality we can restrict our search for optimal mechanisms to direct mechanisms.1. Proposition 2. ¯ Proof. All selling mechanisms that are not direct mechanisms are called “indirect mechanisms.” ¯ ¯ A buyer’s strategy σ in a direct mechanism is a mapping σ : [θ. The optimality of truthfully reporting θ in Γ then follows immediately from the optimality of σ(θ) in Γ. that is. the buyer obtains the same expected utility that she would have obtained had she played type θ s strategy σ(θ ) in Γ. truthfully reporting her type. θ] that indicates for every true type θ that the buyer might have the type σ(θ) that the buyer reports to the seller as his type.

we can now formally deﬁne the condition that the buyer ﬁnds it always optimal to truthfully report her type. We now consider in more detail the conditions under which a direct mechanism is incentive compatible. . Deﬁnition 2.2. A weaker requirement would be to require voluntary participation before agents have learned their types. Incentive compatibility requires: θq(θ) − t(θ) ≥ θq(θ ) − t(θ ) θ q(θ) − t(θ) ≤ θ q(θ ) − t(θ ) Subtracting these two inequalities we obtain (θ − θ )q(θ) ≥ (θ − θ )q(θ ) ⇔ q(θ) ≥ q(θ ) (2. Lemma 2. that is.3 Proof. that is. A direct mechanism is “individually rational” if the buyer.1) (2. We mentioned before that it makes sense to also require that the buyer’s expected utility from the mechanism is not lower than some lower bound. Deﬁnition 2. Consider two types θ and θ with θ > θ .3) (2. say zero.2) 3 Throughout these notes we shall say that a function f is “increasing” if it is weakly monotonically increasing. SCREENING 11 notation. If a direct mechanism is incentive compatible. Notice that in this deﬁnition we require voluntary participation after agents have learned their types.1.CHAPTER 2. if: ¯ u(θ) ≥ 0 for all θ ∈ [θ. θ]. then q is increasing in θ. θ ∈ [θ. if x > x implies f (x) ≥ f (x ). if: ¯ u(θ) ≥ θq(θ ) − t(θ ) for all θ. θ]. conditional on her type. i. A direct mechanism is “incentive compatible” if truth telling is optimal for every type. This requirement is captured in the following deﬁnition.3.4) (2. is voluntarily willing to participate.e. Later we bring individual rationality into the picture.

CHAPTER 2. SCREENING

12

Lemma 2.2. If a direct mechanism is incentive compatible, then u is increasing. It is also convex, and hence diﬀerentiable except in at most countably many points. For all θ for which it is diﬀerentiable, it satisﬁes: u (θ) = q(θ). Note that the equation for u (θ) is the same as the formula for the derivative of maximized utility functions in the “envelope theorem.” For completeness, we provide a self-contained proof. Proof. For all θ: u(θ) = max (θq(θ ) − t(θ )).

¯ θ ∈[θ,θ]

(2.5)

Thus u is the maximum of increasing and aﬃne, hence convex functions. The maximum of increasing functions is increasing, and the maximum of convex functions is convex. Therefore, u is increasing and convex. Convex functions are diﬀerentiable in at most countably many points (Proposition 5.16 in Royden (1988)). Consider any θ for which u is diﬀerentiable. Let δ > 0. Then: u(θ + δ) − u(θ) δ→0 δ lim Similarly: lim u(θ) − u(θ − δ) δ (θq(θ) − t(θ)) − ((θ − δ)q(θ) − t(θ)) (2.8) δ = q(θ). (2.9) ≤ lim

δ→0

((θ + δ)q(θ) − t(θ)) − (θq(θ) − t(θ)) (2.6) δ→0 δ = q(θ). (2.7) ≥ lim

δ→0

Putting the two inequalities together we obtain u (θ) = q(θ) whenever u is diﬀerentiable. The next lemma is essentially an implication of Lemma 2.2 and the fundamental theorem of calculus. In the proof we take care of the possible lack of diﬀerentiability of u in some points. Lemma 2.3 (Payoﬀ Equivalence). Consider an incentive compatible direct ¯ mechanism. Then for all θ ∈ [θ, θ]:

θ

u(θ) = u(θ) +

θ

q(x)dx.

CHAPTER 2. SCREENING

13

Proof. The fact that u is convex implies by Proposition 5.16 in Royden (1988) that it is absolutely continuous. By Proposition 5.13 in Royden (1988) this implies that it is the integral of its derivative. Lemma 2.3 shows that the expected utilities of the diﬀerent types of the buyer are pinned down by the the function q and by the expected utility of the lowest type of the buyer, u(θ). Any two indirect mechanisms which, once the buyer optimizes, give rise to the same q and u(θ) therefore imply the same expected payoﬀ for all types of the buyer. We have therefore indirectly shown a “payoﬀ equivalence” result for classes of indirect mechanisms. A short computation turns Lemma 2.3 into a result about the transfer payments that the buyer expects to make to the seller. This is shown in the next lemma. Lemma 2.4 (Revenue Equivalence). Consider an incentive compatible di¯ rect mechanism. Then for all θ ∈ [θ, θ]

θ

t(θ) = t(θ) + (θq(θ) − θq(θ)) −

θ

q(x)dx.

Proof. Recall that u(θ) = θq(θ) − t(θ). Substituting this into the formula in Lemma 2.3 and solving for t(θ) yields the result. Lemma 2.4 shows that the expected payments of the diﬀerent types of the buyer are pinned down by the the function q and by the expected payment of the lowest type of the buyer, t(θ). For given q and given t(θ) there is thus one, and only one, incentive compatible direct mechanism. Any two indirect mechanisms which, once the buyer optimizes, give rise to the same q and t(θ) therefore imply the same expected payment for all types of the buyer. For the seller it follows that any two such indirect mechanisms yield the same expected revenue, because the seller’s expected revenue is the expected value of the buyer’s expected payments, where the seller takes expected values over the buyer’s types. We have therefore indirectly shown a “revenue equivalence” result for classes of indirect mechanisms. Lemma 2.4 is the famous “Revenue Equivalence Theorem” of auction theory adapted to our more simple setting of monopolistic screening. The full Revenue Equivalence Theorem, which will be discussed in Chapter 3, is based on essentially the same argument that we have explained here. Lemmas 2.1 and 2.4 give necessary conditions for a direct mechanism to be incentive compatible. It turns out that these conditions are also suﬃcient.

CHAPTER 2. SCREENING

14

**Proposition 2.2. A direct mechanism (q, t) is incentive compatible if and only if: (i) q is increasing; ¯ (ii) For every θ ∈ [θ, θ]:
**

θ

t(θ) = t(θ) + (θq(θ) − θq(θ)) −

θ

q(x)dx

Proof. To prove suﬃciency we have to show that no type θ prefers to pretend to be a type θ if q is increasing and t is given by the formula in the Proposition: u(θ) ≥ θq(θ ) − t(θ ) ⇔ u(θ) ≥ θq(θ ) − θ q(θ ) + θ q(θ ) − t(θ ) ⇔ u(θ) ≥ θq(θ ) − θ q(θ ) + u(θ ) ⇔ u(θ) − u(θ ) ≥ (θ − θ )q(θ ) ⇔

θ θ

(2.10) (2.11) (2.12) (2.13) (2.14)

q(x)dx ≥

θ θ

q(θ )dx

To obtain the left hand side of the last inequality we used the formula in Lemma 2.3. This formula is an implication of the formula for the payment in the proposition, as one can see by doing the calculation referred to in the proof of Lemma 2.4 in reverse order. Comparing the two integrals on the left hand side and the right hand side of the last inequality, suppose ﬁrst that θ > θ. The two integrals have the same integration limits, and the function q(x) is everywhere at least as large as the constant q(θ ) that is being integrated on the right hand side because q is increasing. Therefore, the integral on the left hand side is at least as large as the integral on the right hand side. If θ < θ the argument is analogous. We have now obtained a complete characterization of all incentive compatible direct mechanisms. We now bring in individual rationality. Proposition 2.3. An incentive compatible direct mechanism is individually rational if and only if u(θ) ≥ 0 (or equivalently: t(θ) ≤ θq(θ)). Proof. By Lemma 2.2 u is increasing in θ for incentive compatible mechanisms. Therefore, u(θ) is non-negative for all θ if and only if it is nonnegative for the lowest θ.

By the formula for payments in Proposition 2. (2. θq(θ). minus a term that reﬂects a surplus that the seller has to grant to the buyer to provide incentives to the buyer to correctly reveal his type. By Proposition 2.3 we have to have: t(θ) ≤ θq(θ). 1]. In the current context. If t(θ) < θq(θ).2. We turn to the seller’s problem of picking from this set the mechanism that maximizes expected revenue. but with a higher t(θ). If an incentive compatible and individually rational direct mechanism maximizes the seller’s expected revenue then t(θ) = θq(θ). SCREENING 15 We have now completely characterized the set of all direct mechanisms from which the seller can choose. this formula becomes: θ t(θ) = θq(θ) − θ ˆ ˆ q(θ)dθ. This is an approach that we shall take later in the next section. We begin with the observation that it is optimal for the seller to set the lowest type’s payment so that this type has zero expected utility. and all other types’ payments are determined by the formula in Proposition 2. Proof. You need to be a little patient as the formal machinery that is needed for this approach is introduced. Using Lemma 2.CHAPTER 2.15) to obtain an explicit formula for the seller’s expected revenue for any given function q. We give F the linear . θ] → [0. We denote this larger set by F. This term is also called the buyer’s “information rent. We begin by considering in more detail the set of functions q that the seller can choose from. Given any such function. an elegant argument based on convex analysis yields a simple and general answer. ¯ The seller’s choice set is the set of all increasing functions q : [θ. θ] → R. Substituting the lowest type’s payment.5. type θ pays his expected utility from the good.15) i. Lemma 2.2 all types’ payments would increase. the seller will optimally set t(θ) = θq(θ) so that the lowest type has zero expected utility.e.5 we can now simplify the seller’s choice set further. then the seller could increase expected revenue by choosing a direct mechanism with the same q. This set is a subset of the set of all bounded functions ¯ f : [θ.” To determine the optimal function q we could use equation (2.

6.18) . the L f = inf {M | µ ({θ | |f (θ)| > M }) = 0} where µ is the Lebesgue measure.16) h = f + g ⇔ [h(x) = f (x) + g(x) ∀x ∈ [0. We shall not be precise about this. p. For example. g ∈ F 16 (2.CHAPTER 2. We also give F the L∞ -norm. We thus see that the seller maximizes a continuous linear function over a compact. This is a generalization of the idea of corner points. We shall apply this insight here. g. θ] where we neglect sets of measure ∞ -norm is deﬁned by:4 zero.17) This makes F a vector space. convex set. This is intuitive. we have a closer look at the seller’s objective function. and expected revenue is a linear function of q. which is the expected value of the right hand side of (2. Compactness is an implication of Helly’s compactness theorem Doob (1994). We denote by M ⊂ F the set of all increasing functions in F such that f (x) ∈ [0. Convexity is obvious as the convex combination of two increasing functions is increasing. 1]] for all α ∈ R. Next. From (2. Lemma 2. but neglecting this point leads to a certain lack of rigor in what follows. The details of this construction are in Royden (1988). also the expected value of t(θ) is multiplied by α. Note Royden’s footnote 1 which indicates that to be rigorous we should deﬁne F to consist of equivalence classes of functions where two functions are in the same equivalence class if and only if they diﬀer on a set of measure zero only. 1]. The simple observations that we need is that this objective function is continuous and linear in q. This is the set from which the seller chooses. p. 1] ∀x ∈ [0.15) we see that if we multiply q by a constant α. Formally. 1]] for all f. you are probably familiar with the fact that in linear programming the optimal points include corner points. 4 for all f ∈ F. f. h ∈ F (2. We ﬁrst provide the deﬁnition of extreme points of a convex set. 112.15). (2. then t(θ) is multiplied by α for all θ. 165. We now have a simple but crucial observations about M. A fundamental theorem of real analysis provides conditions under which the maximizers of a linear function over a convex set include “extreme points” of that set. Intuitively. Therefore. M is compact and convex. SCREENING structure: g = αf ⇔ [g(x) = αf (x) ∀x ∈ [0. this ¯ is the supremum of |f (θ)| over θ ∈ [θ.

Now we can state the result that we will use to study the seller’s optimal mechanism.CHAPTER 2. We have that q(θ) + q (θ) = 2q(θ) if q(θ) ≤ 0. It is called the “Extreme Point Theorem” in Ok (2007). ˆ ˆ/ ˆ Now consider any function q that is not as described in the lemma. there is some θ∗ such that q(θ∗ ) ∈ (0.5. θ] 5 6 More precisely: q (θ) = 0 for a set of θ that has positive Lebesgue measure. Lemma 2. The case q (θ) < 0 is analogous. / So. 658. and the circumference of a circle is the set of extreme points of the circle. SCREENING 17 Deﬁnition 2. 1} for almost all θ ∈ [θ. convex subset of a normed vector space. and ˆ ˆ q(θ) + q (θ) = 1 if q(θ) > 0. p. Let X be a compact. evidently. Consider now ﬁrst ˆ ˆ the function q + q .4. Then the set E of extreme points of X is non-empty. 1). Proposition 2. y = 0 we have that either x+y ∈ C / or x − y ∈ C or both.5 If q (θ) > 0 and ˆ ˆ ˆ q(θ) = 0 then q(θ)− q (θ) < 0.5. This result implies that a function q that is an extreme point of M and that maximizes expected revenue among all extreme points of M also maximizes expected revenue among all functions in M.5. If q (θ) > 0 and q(θ) = 1 ˆ ˆ/ ˆ then q(θ) + q (θ) > 0. Proof. 1) for a set of θ∗ that has positive Lebesgue measure. for example in two-dimensional space the corners of a triangle are the extreme points of the triangle. . i. Clearly. ˆ and q (θ) = 1 − q(θ) if q(θ) > 0.5. ˆ The seller can thus restrict attention to non-stochastic mechanisms. and there exists an e ∈ E such that f (e) ≥ f (x) for all x ∈ X. But ¯ a non-stochastic mechanism is monotone if and only if there is some p ∈ [θ. Consider any function as described in the lemma. and hence q − q ∈ M.e. We may thus simplify the seller’s problem further. q ∈ M. Thus. ˆ More precisely: q(θ∗ ) ∈ (0. If C is a convex subset of a vector space X. A function q ∈ M is an extreme point of M if and only if ¯ q(θ) ∈ {0. q = 0. and let f : X → R be a continuous linear function. Instead of considering all functions in the set M it is suﬃcient to consider only the set of all extreme points of M.6 Deﬁne q (θ) = q(θ) if q(θ) ≤ 0. The following result characterizes the extreme points of M. and hence q − q ∈ M. We conclude that q is not an extreme point of M. The argument for ˆ ˜ q − q is analogous. θ].7.4. then x ∈ C is an extreme point of C if for every y ∈ X. and suppose that q is another function that satisﬁes: q (θ) = 0 for some θ.

The formula for t follows from Proposition 2. if θ < p∗ . if θ < p∗ .5. This is perhaps analogous to invoking Newton’s law of gravity as an explanation of the fact that apples fall oﬀ apple trees. a result that does not oﬀer the seller any more sophisticated selling mechanisms than we are familiar with from elementary microeconomics. It may seem that we have gone to considerable length to derive a disappointing result. SCREENING 18 such that q(θ) = 0 if θ < p∗ and q(θ) = 1 if θ > p∗ . q(θ) = ∗ p 0 Proof. The optimal function q of this form is obviously the one indicated in the Proposition. Proposition 2. say sugar. if θ > p∗ . to one potential buyer. However. Our results therefore imply that the seller cannot do better than quoting a simple price p∗ to the buyer. but the explanation is non-obvious. 2. beyond those presented . and because its analysis introduces additional elements. This direct mechanism can be implemented by the seller simply quoting the price p∗ and the buyer either accepting or rejecting p∗ . individually rational direct mechanisms if and only if there is a p∗ ∈ argmaxp∈[θ. The analysis is summarized in the following proposition.θ] p(1 − F (p)) such that ¯ 1 and t(θ) = 0 if θ > p∗ . The fact is familiar. A direct mechanism maximizes the seller’s expected revenues among all incentive compatible. We introduce this model because it is more commonly studied than the model in the previous section.2.CHAPTER 2. apart from introducing some technical tools that we use later in more complicated contexts. the reader should appreciate that we have uncovered a rather sophisticated rationale for a familiar everyday phenomenon.3 Nonlinear Pricing Now we study a model in which a monopolist oﬀers an inﬁnitely divisible good. namely. As argued above we only need to consider functions q where the buyer obtains the good with probability 1 if his value is above some price p∗ and with probability 0 if his value is below this price.

Because ν(0) = 0 the buyer’s utility when buying nothing and paying nothing is zero. ν (q) < 0 for all q ≥ 0. θ]. The seller’s beliefs about θ are given by a cumulative distribution function F with density f on the ¯ ¯ interval [θ. A “direct mechanism” consists of functions q and t where: ¯ q : [θ. ¯ A ﬁnal assumption is that limq→∞ θν (q) < c. the larger θ the larger is the consumer’s absolute willingness to pay θν(q) and the consumer’s marginal willingness to pay θν (q) for any given ¯ quantity q. θ] → R+ and ¯ t : [θ. The seller is risk neutral. that utility is additively separable in consumption of the good and money. We can interpret θν(q) as the buyer’s willingness to pay for quantity q. We assume that f satisﬁes: f (θ) > 0 for all θ ∈ [θ. The parameter θ reﬂects how much the consumer values the good.5. and the seller commits to selling quantity q(θ) to the buyer and the buyer commits to paying t(θ) to the seller. The interpretation is that the buyer is asked to report θ. and that the consumer is risk neutral in money. Note that we have assumed. producing quantity q ≥ 0 costs cq. This means that even the highest type’s marginal willingness to pay falls below c as q gets large. that will reappear in almost exactly the same form in the analysis of optimal mechanisms. that is. where c > 0 is a constant. The parameter θ can take any value between θ and θ. More precisely. This assumption ensures that the quantity that the seller supplies to the buyer is ﬁnite for all possible types of the buyer. In the current context we use the following deﬁnition of direct mechanisms: Deﬁnition 2. strictly increasing and strictly concave function: ν (q) > 0. As in the previous section the revelation principle holds and we can restrict attention to direct mechanisms. as in the previous section. For simplicity we assume that production costs are linear. We seek to determine optimal selling procedures for the seller. θ] → R. Note that we use the same notation as in the previous .CHAPTER 2. so that she seeks to maximize her expected revenue. SCREENING 19 in the previous section. The value of θ is known to the buyer but not to the seller. θ]. We assume that ν(0) = 0 and that ν is a twice diﬀerentiable. The buyer’s utility from buying quantity q ≥ 0 and paying a monetary transfer t to the monopolist is θν(q) − t.

t) is incentive compatible if and only if (i) q is increasing. but that in this section q(θ) is a quantity whereas in the previous section q(θ) was a probability. As in the previous section we can then study incentive compatibility and individual rationality of direct mechanisms. (2. ¯ (ii) For every θ ∈ [θ. It is non-trivial to study stochastic direct mechanisms in our context. not a probability distribution over non-negative numbers.19) yields: θ t(θ) = θν(q(θ)) − θ ˆ ˆ ν(q(θ))dθ. One modiﬁcation is needed.21) Substituting this into equation (2. One ﬁnds that a direct mechanism (q. We do not state the revelation principle formally for our context. SCREENING 20 section. that is: t(θ) = θq(θ). It is analogous to the revelation principle in the previous section. It is obvious that the seller will choose t(θ) so that the utility of type θ is zero.19) An incentive compatible mechanism is individual rational if and only if t(θ) ≤ θq(θ). that is. The analysis proceeds along exactly the same lines as in the previous section. (2. (2. We make this assumption for simplicity. however. . and we shall just state the result of the analysis. we assume that for each type θ the quantity sold to the buyer if he is of type θ is a non-negative number. In this section we ignore stochastic mechanisms.20) The seller’s decision problem is to pick among all direct mechanisms satisfying these two conditions the one that maximizes expected revenue.CHAPTER 2.22) The choice that remains to be studied is that of the function q. To obtain deterministic direct mechanisms we have to restrict attention to general mechanisms and buyer optimal strategies that result in a deterministic quantity for the buyer for each type of the buyer. (2. in the statement of the result. θ]: θ t(θ) = t(θ) − θν(q(θ)) + θν(q(θ)) − θ ˆ ˆ ν(q(θ))dθ.

This is clear from equation (2.24) = θ When moving from the second to the third line in (2. SCREENING 21 At this point we depart from the line of argument that we followed in the previous section. It is therefore worthwhile to consider it in detail.23).23). The reason is that the seller’s objective function is no longer linear in q.22) where q enters the non-linear function ν.1 the change in the order of integration. In the second line we ﬁrst integrate along the vertical lines and then horizontally. If the seller chooses q(·). We focus initially on the double integral in the second term in (2. We shall use instead equation (2. In the third line we ﬁrst integrate along the horizontal lines in Figure 2. although it contains no conceptually or mathematically deep insights.22) to study in more detail the seller’s expected proﬁt.1 and then vertically.CHAPTER 2. then his expected proﬁt is: ¯ θ θ θν(q(θ)) − θ ¯ θ θ ˆ ˆ ν(q(θ))dθ − cq(θ) f (θ)dθ ¯ θ θ θ = θ θν(q(θ))f (θ)dθ − θ ˆ ˆ ν(q(θ))dθf (θ)dθ − θ ¯ θ cq(θ)f (θ)dθ (2.24) we change the order of integration. the extreme point argument of the previous section does not apply here. . Because the objective function is not linear in q. ¯ θ θ ¯ θ θ θ θ ¯ θ ¯ θ ˆ θ ¯ θ θ ˆ ˆ ν(q(θ))dθf (θ)dθ ˆ ˆ ν(q(θ))f (θ)dθdθ ˆ ˆ ν(q(θ))f (θ)dθdθ ¯ θ ˆ θ = θ = θ = θ ¯ θ ˆ ν(q(θ)) ˆ f (θ)dθdθ = θ ¯ θ ˆ ˆ ˆ ν(q(θ))(1 − F (θ))dθ ν(q(θ))(1 − F (θ))dθ (2.23) We seek to simplify the expression in (2. We indicate in Figure 2. The calculation that follows. By Fubini’s theorem this leaves the value of the integral unchanged. appears in this or in similar form frequently in the theory of mechanism design.

1: Changing the order of integration We now substitute the last line in (2.25) = 1 − F (θ) f (θ) − cq(θ) f (θ)dθ. and then impose a condition that makes sure that the function q that . Expected values are taken over θ. ¯ θ ¯ θ [θν(q(θ)) − cq(θ)] f (θ)dθ − θ ¯ θ θ ¯ θ ν(q(θ))(1 − F (θ))dθ ν(q(θ)) θ = θ ¯ θ [θν(q(θ)) − cq(θ)] f (θ)dθ − ν(q(θ)) θ − θ 1 − F (θ) f (θ)dθ f (θ) (2. The seller thus chooses q to maximize the expected value of the expression that is in large square brackets in (2. Suppose we ignore for the moment the constraint that q must be increasing.25).CHAPTER 2. This choice of q also maximizes the expected value of that expression.24) into the seller’s objective function in (2. We study this approach to the choice of q ﬁrst. The seller must choose an increasing function q.23). Then the seller can choose q(θ) for each θ separately to maximize the expression in the large square brackets. SCREENING 22 ^ θ θ θ θ θ θ Figure 2.

and that θν(q) tends to less than c as q tends to inﬁnity.33) then our assumptions imply that there is a unique solution to (2. and that this stationary point is also the unique optimal choice of q(θ).29) (2. If ν (0) θ − 1 − F (θ) f (θ) ≤c (2.28) then there is obviously no solution.27) We now investigate the existence of a solution to (2. If the function q that we have determined above is increasing. We now introduce the following assumption.31) then it is again obvious that the optimal choice is: q(θ) = 0. The seller seeks to maximize the expected value of this expression. We have now determined for each θ the choice of q(θ) that maximize the expression that is in large square brackets in (2. the left hand side of (2. SCREENING 23 we ﬁnd is indeed increasing. Obviously. . Now consider θ− 1 − F (θ) > 0. and decreasing.27) shares all these properties. and the seller is constrained to choose a function q that is increasing.27).25).26) (2. and therefore continu¯ ous. f (θ) (2.32) >c (2.30) Recall that we have assumed that ν is diﬀerentiable. and the optimal choice is q(θ) = 0.27).CHAPTER 2. If θ− 1 − F (θ) ≤0 f (θ) (2. then it must be the optimal choice for the seller. The ﬁrst order condition for maximizing the expression in square brackets for given θ is: ν (q(θ)) θ − 1 − F (θ) −c=0⇔ f (θ) 1 − F (θ) =c ν (q(θ)) θ − f (θ) (2. which implies that the q that we have determined is increasing. If ν (0) θ − 1 − F (θ) f (θ) (2.

1. one of the unavoidable hazards of life.” The analysis of this section is summarized in the following proposition: Proposition 2.1 is that 1−F (θ) is increasing in θ. and 1−F (θ) can be thought of as the conditional probability of dying at time θ of an individual that has survived until time θ. (ii) otherwise: ν (q(θ)) θ − The proﬁt maximizing t is given by: θ 1 − F (θ) f (θ) = c. The optimal q is the intersection point of that expression with c or zero. We shall refer to such distributions as “regular.6 note that for ¯ the highest type θ = θ we have: 1 − F (θ) = 0. Therefore. (2. f (θ) A suﬃcient condition for Assumption 2.1 implies that the left hand side of (2. whatever is greater. To verify that Assumption 2. Think of F (θ) as the probability that an individual dies before time θ. θ − 1−F (θ) f (θ) 24 is increasing in θ.27) is increasing in θ for every q. is increasing in θ. The suﬃcient condition is that this conditional probability of dying. Suppose that F is regular.CHAPTER 2. t(θ) = θν(q(θ)) − θ ˆ ˆ ν(q(θ))dθ. To understand the economic meaning of Proposition 2.6. SCREENING Assumption 2.6 applies to θ.1. Many commonly considered distributions F satisfy Assumption 2.34) This equation shows that the highest type is supplied the quantity at which this type’s marginal willingness to pay is exactly equal to the marginal cost . It is then easy to see that the optimal q is increasing in θ. This suﬃcient condition is often referred to as the “increasing hazard rate” condition. Then 1 − F (θ) is the probability that the individual survives f (θ) until time θ. the second of the ¯ ¯ two cases in Proposition 2. and q(θ) is determined by: ¯ ν (q(θ))θ = c.1 implies that q is increasing note that Assumption 2. Then an expected proﬁt maximizing choice of q is given by: (i) if ν (0) θ − 1−F (θ) f (θ) ≤ c: q(θ) = 0.

θ is uniformly distributed on [0. If θ > 0. We conclude with a numerical example.5 (2. ν(q) = q.27) that diﬀers from the ﬁrst best condition (2.36) The ﬁrst and the second line are equivalent because in our example: ν (0) = +∞. 1]. but with a quantity smaller than the marginal beneﬁts. Thus. i.34) (with ¯ θ replaced by θ) in that the left hand side is smaller for every q. 1].35) which is obviously the case.: F (θ) = θ and f (θ) = 1 for all θ ∈ [0. ¯ For all lower types θ < θ the quantity supplied to these types is determined by equation (2. This is the quantity that this type would choose to produce if he owned the ﬁrm.e. f (θ) 1 (2.1 is satisﬁed. we have to check that the following expression is increasing in θ: θ− 1 − F (θ) 1−θ =θ− = 2θ − 1. √ Example 2. This means that all types that ¯ are lower than θ are oﬀered a quantity that is smaller than the “ﬁrst best” quantity. Next we determine for which values of θ the optimal quantity q(θ) equals zero: ν (0) θ − 1 − F (θ) ≤c⇔ f (θ) 1 − F (θ) θ− ≤0⇔ f (θ) 2θ − 1 ≤ 0 ⇔ θ ≤ 0. To verify that Assumption 2.CHAPTER 2.37) .1. SCREENING 25 of production. 2 (2. the marginal costs are not equated with the marginal beneﬁts.5 the optimal q(θ) is given by: ν (q(θ)) θ − 1 − F (θ) =c⇔ f (θ) 1 √ (2θ − 1) = 1 ⇔ 2 q 1 √ q=θ− ⇔ 2 1 2 q= θ− . We refer to this quantity as the “ﬁrst best” quantity. c = 1.

This 4 2 optimal non-linear pricing scheme is shown in Figure 2. We note that there is a quantity discount. The per unit price: 1 1√ q+2 q t 1 1 1 = 2 = + √ (2.40) 2 2 Essentially.39) The payment by type θ is: 1 1 t(θ) = θ2 − 2 8 1 √ 1 2 1 = − q+ 2 2 8 1 1√ = q+ q (2.CHAPTER 2.5 it is given by: θ t(θ) = θν(q(θ)) − θ ˆ ˆ ν(q(θ))dθ ˆ 1 ˆ θ − dθ 2 0.5 1 2 1 1 1 1 − θ − θ− + =θ θ− 2 2 2 8 4 1 2 1 = θ − (2. and if θ > 0. 1 ].5 θ =θ θ− =θ θ− 1 2 − 1ˆ 1 1ˆ θ − θ2 − θ 2 2 2 0. . Their payment is t(q) = 2 q + 1 q. SCREENING 26 The corresponding transfer t(θ) is zero if θ ≤ 0.2. For this we determine ﬁrst for given q which type θ purchases the quantity q q(θ) = q ⇔ θ− 1 2 2 =q⇔ 1 √ = q⇔ 2 1 √ θ= q+ 2 θ− (2. We can express the transfer t as a function of q.41) q q 2 2 q decreases in q.5. the monopolist thus oﬀers to consumers the deal that they √ 1 can buy any quantity q ∈ [0.38) 2 8 We want to translate the solution into an optimal non-linear pricing scheme.

CHAPTER 2.1 0. We shall not pursue this direction here because it is not related to the theory of mechanism design as far as it is covered in these notes. for sale. it is of great interest to consider the case in which the buyer’s private information is also of relevance to the seller’s assessment of a possible sale. Therefore.2: The optimal non-linear pricing scheme 2. A buyer’s private information about his health situation aﬀects not only his own evaluation of an insurance contract. so that she seeks to maximize her expected revenue.0 0.05 0.25 Figure 2.15 0.10 q 0. but also the insurance seller’s evaluation of the contract.2 t 0.00 0. Suppose a seller has two distinct indivisible goods.4 Bundling The theory of screening can be developed into many diﬀerent directions. For example.3 0. Another important extension is to the case in which the buyer’s private information is multi-dimensional.20 0. Let IA and IB be indicator variables that are either 1 (if the buyer obtains good A . The case of multi-dimensional private information is also relevant in the theory of mechanism design. Insurance contracts are an example. we give here a simple example of screening when the buyer has multidimensional private information. For simplicity we assume that the seller values the goods at zero and is risk neutral. SCREENING 27 0.4 0. good A and good B.

43) (2. What is the optimal choice of pA . The seller thus oﬀers to the buyer that he can buy the two goods separately. The seller’s belief about these two parameters is given by the uniform distribution F over the unit square [0.862 3 (2. Thus. Suppose the seller considers quoting three prices: pA . The literature refers to the combination of goods A and B as a “bundle. We assume that the seller cannot stop the buyer from buying goods A and B at price pA + pB so that the price pAB . As in the previous sections the revelation principle holds and we could restrict attention to direct mechanisms. and only consider a very small class of indirect mechanisms. pB and pAB ? This is a simple calculus exercise. These parameters are known to the buyer but not known to the seller. pB and pAB . The parameters vA and vB indicate the buyer’s willingness to pay for the two goods. we assume a second time that there is no relation at all between the two goods. but that he gets a better deal if he buys the two goods together. In this way the two goods are entirely independent. It turns out that the optimal prices are: pA = pB = pAB 2 3 √ 1 = 4 − 2 ≈ 0. but he also oﬀers to them the option to buy the goods individually. SCREENING 28 respectively good B) or 0 (if the buyer does not obtain good A respectively good B). 1]2 .CHAPTER 2. if it is to have any eﬀect. but they are like pasta and a watch. Then the buyer’s utility is: IA vA + IB vB − t.” The monopolist’s strategy in our example is also described as “mixed bundling” because the monopolist oﬀers the bundle to consumers. (2. We assume in addition that the marginal value of each good does not depend on whether the other good is also obtained. They are not like pasta and tomato sauce.42) Note that utility is additive in the two goods and in money. Note that we assume here that vA and vB are stochastically independent. We shall instead simplify our problem much more. The interpretation is that the buyer can buy good A at price pA . good B at price pB . has to satisfy: pAB ≤ pA + pB . Denote by t the monetary transfer from the buyer to the seller.44) Note that the optimal price PAB is indeed strictly smaller than pA +pB . Our interest is again in optimal selling procedures for the seller. . or goods A and B at price pAB .

Figure 2. We thus observe in Figure 2. the general theory of screening with multiple goods and multi-dimensional private information is rather complicated. require that the good is transferred to the buyer for all values of vA and vB except zero. It is very surprising that the seller oﬀers the goods as a bundle at a discount even though from the consumer’s point of view the goods are entirely unrelated.5 nothing good A 0 0 0. or no good. As the example indicates.5 2/3 vA 1 Figure 2. good B only. The literature has in fact spent some time seeking to understand the intuition behind this eﬀect. Hermalin . Depending on the value of vA and vB the buyer purchases good A only.3 illustrates the direct mechanism implemented by the seller if he quotes the three optimal prices.3: Buyer behavior given optimal prices The buyer’s demand behavior given the optimal prices is shown in Figure 2. of course.3. SCREENING 29 1 good B vB 2/3 both goods 0.3 several distortions of eﬃciency. We present this example mainly to illustrate that even in the simple case of screening. both goods.CHAPTER 2. multi-dimensional private information may cause surprising and counterintuitive eﬀects. Eﬃciency would.

and in which the revenue equivalence principle does not hold. the example in Section 2. b) Does Proposition 2.5 hold if the type distribution F is discrete? c) Prove that the conditions in (2. Manelli and Vincent (2007) applies the extreme point analysis of Section 2. 2.2 in which the buyer has only two possible types.4 is based on Chapter 6 of Hermalin (2005).43) maximize the seller’s proﬁts.19) are suﬃcient for incentive compatibility. .6 Problems a) Give an example in the setting of Section 2. d) Prove that the prices in (2.3 is discussed further in Bolton and Dewatripont (2005). The theory of non-linear pricing in Section 2.5 Comments on the Literature Our exposition in Section 2.2 to a general screening model and illustrates the potentially complicated stochastic nature of optimal selling mechanisms. Finally. SCREENING 30 2.2 is a modiﬁed version of Manelli and Vincent (2007).CHAPTER 2. The seminal paper on mixed bundling is Adams and Yellen (1976).

The examples are also interesting in their own right..1 Single Unit Auctions Set-Up The model in this chapter is the same as in Chapter 2. We start with examples to illustrate the general analysis that will follow in later chapters. A seller seeks to sell a single indivisible good. N }.2. . Buyer i’s utility if she does not 31 . We denote the set of potential buyers by I = {1. Understanding this contrast will prepare the reader for our later discussion of robust mechanism design..Chapter 3 Classic Bayesian Mechanism Design: Examples 3. We contrast in the next chapters optimal Bayesian mechanisms and optimal dominant strategy mechanisms to illustrates the adjustments that need to be made to a mechanism if the weaker condition of Bayesian incentive compatibility is replaced by the more restrictive condition of dominant strategy incentive compatibility.1 Introduction This chapter describes three classic mechanism design problems using Bayesian Nash equilibrium to predict agents’ strategic behavior for any given mechanism. The next chapter will consider the same examples but we shall use dominant strategies as our concept for predicting agents’ behavior. 3.2 except that we now have more than one potential buyer. Buyer i’s utility if she purchases the good and pays a transfer ti to the seller is: θi − ti .2 3. There are N ≥ 2 potential buyers.. 2.

θ]N . We model the valuation θi as a random variable with cumulative distribution function Fi with density fi . BAYESIAN MECHANISM DESIGN 32 purchase the good and pays a transfer of ti to the seller is: 0−ti . The phrase “independent” refers to the fact that we have assumed that values are independent and that they follow a commonly known prior distribution F . Thus. . but also the potential buyers’ beliefs about each other. We assume that for i. Thus. θ]. The distribution of θ is denoted by F which is the product of the distributions Fi . i=1 We assume that buyer i knows θi . θN ). The assumption that buyers’ values are independent implies that each buyer’s beliefs about the other buyers’ values is independent of his own value. i with i. although we do not assume that the random variables θi have the same distribution for diﬀerent i. 2. The seller’s utility if he obtains transfers ti from buyers i = 1. The model that we have described is known in the literature as a model with “independent private values. θ] where 0 ≤ θ < θ. as in Chapter 2. The support ¯ of the random variable θ is Θ ≡ [θ. . we do assume that they have the same support. No buyer would change his value of the good if he knew what other buyers know. but neither the seller nor the other buyers j = i observe θi . is common knowledge among the buyers and the seller. So. The assumption that F is common knowledge implies that it is commonly known among the sellers and the buyers that they share the same beliefs about other buyers. j ∈ I with i = j the random variables θi and θj are independent. We have assumed private values in the sense that each buyer’s private information is suﬃcient to determine this buyer’s value of the good. . This is for convenience only. if buyer i has a high value. he does not attach more probability to the event that buyer j = i has a high value than if i had had a low value. We discuss why in this paragraph and the next. The distribution F . Each potential buyer i observes θi . and the density of F is denoted by f . . Note ﬁrst that the distribution F describes not only the seller’s beliefs. .CHAPTER 3. The support of θi ¯ ¯ is [θ. N is: N ti . All examples in this chapter will be built on this assumption. for example. We denote by θ the vector (θ1 . The assumption that F is a common prior of the seller and all buyers implies that two buyers i.2. however. but that neither the seller nor any other buyer j = i knows θi . and that this belief is also shared by the seller. i = j have the same belief about buyer j’s value.” One also refers to the “independent private values” assumption. θ2 . For technical ¯ convenience we also assume that fi (θi ) > 0 for all i ∈ I and all θi ∈ [θ. . . The assumption is very restrictive. the private information that leads one buyer to value the good . .

in conjunction with the assumptions about utilities. Here.2. and restrict attention to extensive game trees where only the potential buyers are players. he could pick a price. Direct Mechanisms. we might as well eliminate the seller as a player from the game. BAYESIAN MECHANISM DESIGN 33 highly (or not) would not change any other buyer’s value if it was known to that buyer. and then randomly pick one of the buyers. For example. ask each buyer to indicate whether she is willing to pay this price for the good. if any. 215). Thus. the . Therefore. Note that games may have none. 1. and the distribution of types that we made in the previous subsection deﬁnes a game of incomplete information. The buyers will play the equilibrium that the seller proposes.2 Mechanisms. N } × RN . We shall imagine that the seller only proposes games that do have at least one Bayesian Nash equilibrium. For simplicity we shall not provide a formal deﬁnition of a “mechanism. when announcing the mechanism. We will allow the seller to pick an arbitrary extensive game tree where the players are the potential buyers and the seller. p. The most plausible interpretation is that the private information is about each buyer’s private tastes rather than about objective features of the good. . The seller will ﬁnd it to her advantage to commit in advance to a strategy. the seller also proposes a Bayesian Nash equilibrium of the corresponding game. We shall understand by a “mechanism” such an extensive game tree together with an assignment of a probability distribution over outcomes to each terminal history. The seller assigns to each terminal history of the game (of ﬁnite or inﬁnite length) an outcome. . and the Revelation Principle We will be interested in procedures that the seller can use to sell his good. The standard solution concept for such games is that of a Bayesian Nash equilibrium (Fudenberg and Tirole (1993). 0 stands for the outcome that the good remains with the seller and is hence not sold. and if so to whom.” A mechanism. information. and that. Formally an outcome is a probability distribution over {0.CHAPTER 3. 3. We will consider a much more general class of methods for selling the good. who have said that they are willing to buy the good and transact with this buyer at the announced price. Bayesian Nash equilibrium. 2. if there are multiple equilibria. one. and a probability distribution over vectors of transfers from the buyers to the seller. . that is a probability of transferring the good. . or more than one.

Formally. and over the possibility of not selling the good. The interpretation is that in a direct mechanism the buyers are asked to simultaneously and independently report their types. . We assume that the utility that buyers obtain if they walk away from the mechanism proposed by the seller is zero. .” The probability qi (θ) is the probability that agent i obtains the good if the type vector is θ. . We now introduce a subclass of mechanisms.CHAPTER 3. “direct mechanisms. This is without loss of generality. . and we will comment further on it when discussing the revelation principle below. A “direct mechanism” consists of functions q and ti (for i ∈ I) where: q:Θ→∆ and ti : Θ → R for i ∈ I.1. The probability 1 − N qi (θ) is the probability with which the seller retains the i=1 good if the type vector is θ. is the probability that the good is not i=1 sold. . . The functions ti describe the transfer payment that buyer i makes to the seller. and therefore we assume that the equilibrium that the seller proposes must oﬀer each potential buyer an expected utility of at least zero.” and then show that it is without loss of generality to restrict attention to such mechanisms. 1 − N qi . The function q(θ) describes the rule by which the good is allocated if the reported type vector is θ. Participation in the mechanism and the equilibrium that the seller proposes must be voluntary. In this case the remaining probability. q2 . Deﬁnition 3. qN may add up to less than 1. ∆ is deﬁned by: N ∆ ≡ {(q1 . We now state the “Revelation Principle” which as in Chapter 2 shows that in some sense there is no loss of generality in restricting attention to direct mechanisms. . q2 . Note that we have assumed that this transfer payment is deterministic. This assumption is important for the revelation principle. qN ) | 0 ≤ qi ≤ 1 for all i ∈ I and i=1 qi ≤ 1}. BAYESIAN MECHANISM DESIGN 34 seller can in a sense “pick” which equilibrium the buyers will play. Note that the probabilities q1 . . . We shall refer to q as the “allocation rule. In the following deﬁnition ∆ denotes the set of all probability distributions over the set I of buyers to whom the good might be sold.

then the same type θi prefers to deviate from σ. that is. and there is no reason why these equilibria should be equivalent to any Bayesian Nash equilibrium of the indirect mechanism Γ. Construct Γ by deﬁning the functions q and ti as required by item (ii) in Proposition 3.1 shows that in the set-up that we have described it is without loss of generality to restrict attention to the case that the seller chooses a direct mechanism and proposes to agents that they report their types truthfully. and provided that it gives them expected utility of at least zero. Proposition 3. If type θi prefers to report that her type is θi . σ prescribes telling the truth. Proof. BAYESIAN MECHANISM DESIGN 35 Proposition 3. This is crucial because the equivalent direct mechanism that is constructed in the proof of Proposition 3. Note however. For every mechanisms Γ and Bayesian Nash equilibrium σ of Γ there exists a direct mechanism Γ and a Bayesian Nash equilibrium σ of Γ such that: (i) The strategy vector σ satisﬁes for every i and every θi : σi (θi ) = θi . We can restrict attention to direct mechanisms in which it is a Bayesian equilibrium that everyone always reports their type truthfully.CHAPTER 3. We want to deﬁne . that it is crucial to this construction that we have neglected problems of multiple equilibria by assuming that agents follow the seller’s proposal provided that it is an equilibrium. Hence σ is not a Bayesian equilibrium of Γ.1. Suppose it were not.1 might have Bayesian Nash equilibria other than truth-telling. We can prove the result by showing that truth telling is a Bayesian equilibrium of the game.1 (Revelation Principle). The revelation principle greatly simpliﬁes our search for optimal mechanisms. and to play the strategy that σ prescribes for θi in Γ. (ii) For every vector θ of types the distribution over outcomes that results in Γ if the agents play σ is the same as the distribution over outcomes that results in Γ if the agents play σ . and the expected value of the transfer payments that result in Γ it the agents play σ is the same as the transfer payments that results in Γ if the agents play σ . and in which every type’s expected utility is at least zero. Depending on how equilibria are selected. one or the other mechanism might be strictly preferred by the seller in that case.

and we denote by f−i the density of F−i . . A direct mechanism is “incentive compatible” if truth telling is a Bayesian Nash equilibrium. θ−i )f (θ−i )dθ−i . We denote by F−i the cumulative distribution of θ−i . θ]. Deﬁnition 3. This is given by: Ui (θi ) = θi Qi (θi ) − Ti (θi ). Using this notation we can now formally deﬁne the two conditions that the seller has to respect when choosing a selling mechanism. ∈ [θ. BAYESIAN MECHANISM DESIGN 36 these properties of a direct mechanism formally. (3. 1] by setting: Qi (θi ) = Θ−i qi (θi . but before all agents’ types 7 We ignore questions of existence and uniqueness of the conditional expected values referred to in this paragraph. Deﬁnition 3. we deﬁne for ¯ each agent i ∈ I a function Qi : [θ. θ−i )f (θ−i )dθ−i . conditioning on agent i’s type being θi . θ]. We also deﬁne ¯ for each agent i ∈ I a function Ti : [θ.CHAPTER 3. For this we introduce additional notation. that is. θ] → R by setting: Ti (θi ) = Θ−i ti (θi . We denote by θ−i the vector of all types except player i’s type. (3. To conclude this subsection. if: ¯ Ui (θi ) ≥ 0 for all i ∈ I and θi . A direct mechanism is “individually rational” if each agent.3. again conditioning on agent i’s type being θi . Qi (θi ) is the conditional expected value of the probability that agent i obtains the good. the phase that follows after agents have learned their types. if: ¯ θi Qi (θi ) − Ti ≥ θi Qi (θi ) − Ti (θi ) for all i ∈ I and θi . θi ∈ [θ.2) Thus.1) Thus.2. we also deﬁne agent i’s expected utility Ui (θi ) conditional on her type being θi . it is useful to introduce some further terminology. conditional on her type. θ] → [0. Ti (θi ) is the conditional expected value of the transfer that agent i makes to the seller. In the timeline of the game deﬁned by a mechanism. We deﬁne Θ−i ≡ ΘN −1 . Given a direct mechanism. is voluntarily willing to participate. that is.7 Finally.

2. we obtain. We ﬁrst focus on incentive compatibility. is often referred to as the “interim” phase. The proof of this is the same as the proof of Lemma 2.3 (Payoﬀ Equivalence).2. Similarly. If a direct mechanism is incentive compatible.2. Then for all I ∈ I and all θi ∈ [θ. then for every agent i ∈ I the function Ui is increasing. and the phase after all agents have revealed their types in a direct mechanism is then called the “ex post” phase. 3. and hence diﬀerentiable except in at most countably many points. It is also convex. Lemma 3.3. with the functions Qi and Ti replacing the functions q and t.CHAPTER 3.4 (Revenue Equivalence).1. θ]: θi Ti (θi ) = Ti (θ) + (θi Qi (θi ) − θi Qi (θi )) − θ Qi (x)dx. Then for all I ∈ I and all θi ∈ [θ.4: Lemma 3. For all θi for which it is diﬀerentiable. If a direct mechanism is incentive compatible.2 and 3.2 and therefore we omit most proofs. Consider an incentive compatible di¯ rect mechanism. The phase before agents have learned their types is referred to as the “ex ante” phase. 2.3 and 2. Lemma 3. and we shall refer to Ui (θi ) as the interim expected utility of agent i if she is of type θi . For example. we shall refer to Ti (θi ) as the interim expected transfer of agent i if she is of type θi . Consider an incentive compatible direct ¯ mechanism.1. We shall use this terminology occasionally in these notes. Lemma 3. We proceed in much the same way as in Chapter 2. then for every agent i ∈ I the function Qi is increasing. . it satisﬁes: U (θi ) = Qi (θi ). analogues to Lemmas 2. BAYESIAN MECHANISM DESIGN 37 are revealed. θ]: θi Ui (θi ) = Ui (θ) + θ Qi (x)dx.3 Characterizing Incentive Compatibility and Individual Rationality In this subsection we seek to understand better the structure of the set of all direct mechanisms that satisfy the two conditions introduced in Deﬁnitions 3.

. Suppose we wanted to compare auctioneer’s expected revenue from the second price auction with minimum bid 0 to the expected revenue from the ﬁrst price auction with minimum bid 0. In the second price auction it is a weakly dominant strategy. the equilibria imply the ¯ same values for Ti (θ) and Qi (θi ) for all i ∈ I and θi ∈ [θ. In shorthand expression. such that they imply the same interim expected probability of obtaining the object for each type of each agent. Hence this equilibrium shares with the equilibrium of the second price auction that the expected payment of the lowest type is zero (because this type’s probability of winning is zero). and that the highest type wins with probability 1. and such that the expected payment made by the lowest type is the same in the two mechanisms. θ]. and therefore.” We wish to explain an application of the revenue equivalence theorem. We described in the previous paragraph the most famous application of Lemma 3. and Bayesian Nash equilibria of these mechanisms. BAYESIAN MECHANISM DESIGN 38 Lemmas 3. Note that this does not mean that the functions ti are uniquely determined. It is for this reason that the result is called the “revenue equivalence theorem. A symmetric Bayesian Nash equilibrium for the ﬁrst price auction is constructed in Proposition 2. and hence a Bayesian Nash equilibrium.CHAPTER 3. Consider two diﬀerent indirect mechanisms.4. also the expected revenue of the seller is the same for these two mechanisms. Therefore. Consider the symmetric case in which Fi does not depend on i.2 of Krishna (2002). As in Chapter 2. Then Lemma 3. But note that the Lemma is much more general.4 implies that all types’ interim expected payments are the same for these two indirect mechanisms.3 and 3. The revenue equivalence theorem implies that the expected revenue from the equilibria of the two diﬀerent auction formats is the same.4 show that the interim expected payoﬀ and the interim expected payment of the diﬀerent types of the buyers are pinned down by the functions Qi and by the expected payoﬀ respectively the expected payment of the lowest type.2 and is therefore omitted.2 we can collect the observations made so far. the Lemma says that the interim expected payments of all types only depend on the interim expected allocation rule and the interim expected payment of the lowest type. The proof is analogous to the proof of Proposition 2. Diﬀerent functions ti might give rise to the same interim expected payments Ti . of course. and obtain conditions that are not only necessary but also suﬃcient for incentive compatibility. This equilibrium is in strictly increasing strategies. to bid one’s true value.

tN ) is incentive compatible if and only if for every i ∈ I: (i) Qi is increasing.CHAPTER 3. An incentive compatible direct mechanism is individually rational if and only if for every i ∈ I we have: Ti (θi ) ≤ θi Qi (θi ). he can pick the interim expected payments by the lowest types in any arbitrary way. The seller can focus on two choice variables: ﬁrstly the allocation rule q. So far we have focused on the characterization of incentive compatibility. . we have one further constraint on the seller’s choice of direct mechanism. . Lemma 3. A direct mechanism (q.3. we restrict attention to incentive compatible direct mechanisms. t1 . Moreover. The seller has to choose a mechanism that implies an expected utility of at least zero for the lowest type agents. any such transfer scheme will give him the same expected revenue. Proposition 3. 3. Then we have the following result that is analogous to Lemma 2. .5. As long as the seller picks an allocation rule q such that the functions Qi (i ∈ I) are increasing. and secondly the interim expected payments by the lowest types: T (θi ). and be assured that there will be some transfer scheme that makes the allocation rule incentive compatible and that implies the given interim expected payments by the lowest types. BAYESIAN MECHANISM DESIGN 39 Proposition 3.3. If an incentive compatible and individually direct mechanism maximizes the seller’s expected revenue then for every i ∈ I: Ti (θ) = θQi (θ). Thus. and therefore the seller does not have to worry about the details of this transfer scheme.5. t2 .2. Now we turn to individual rationality. θ]: θ Ti (θi ) = Ti (θ) + (θi Qi (θi ) − θQi (θ)) − θ ˆ ˆ Qi (θ)dθ We have now obtained a complete understanding of the implications of incentive compatibility for the seller’s choice. . . However. ¯ (ii) For every θi ∈ [θ. We begin with a simple observation that is analogous to Lemma 2.2.4 Expected Revenue Maximization We now study the expected revenue maximizing choice of selling mechanism.

We shall proceed as in Chapter 2. θ].3) Note that for the seller’s expected revenue the details of the function q don’t matter.5 into part (ii) of ¯ Proposition 3.3 to equation (2.25) we can calculate the seller’s expected revenue from any particular buyer i.4) over all i ∈ I.6) . then the seller would choose for each θ the probabilities qi (θ) so as to maximize the expression in the large round brackets in the formula for expected revenue. BAYESIAN MECHANISM DESIGN 40 We can now simplify the seller’s problem further.3 we ﬁrst ask which function q the seller would choose if he did not have to make sure that the functions Qi are increasing.2 we get for every i ∈ I and θi ∈ [θ. θ]: θ Ti (θi ) = θi Qi (θi ) − θ ˆ ˆ Qi (θ)dθ (3.2 and Lemma 3.4) To obtain a formula for the total expected transfer by all agents we add the formula in equation (3.CHAPTER 3.5) = 1 − Fi (θi ) qi (θ) θi − fi (θi ) where the last equality becomes obvious if one recalls the deﬁnition of Qi (θi ).5. were it was suﬃcient to characterize these extreme points. The seller has to choose a function q so that the interim probabilities Qi are increasing for all i ∈ I. The payments are then completely determined by part (ii) of Proposition 3.3 and not as in Chapter 2. but in our context a characterization of these extreme points wouldn’t take us very far. The reason is that there are many extreme points of the seller’s choice set. (3. We shall now focus on the optimal choice of q. We deﬁne this expression to be ψi (θi ): ψi (θi ) ≡ θi − 1 − Fi (θi ) ¯ for all i ∈ I and θi ∈ [θ. We obtain: N i=1 N θ ¯ θ Qi (θi ) θi − i=1 Θ 1 − Fi (θi ) fi (θi ) fi (θi )dθi f (θ)dθ (3. ¯ θ Qi (θi ) θi − θ 1 − Fi (θi ) fi (θi ) fi (θi )dθi .2. fi (θi ) (3. In a second step we introduce an assumption that makes sure that the optimal q from the ﬁrst step implies increasing functions Qi . These were the focus of Chapter 2. but only the interim probabilities Qi .2. If monotonicity could be ignored. Substituting the formula in Lemma 3. Doing the same calculations that lead in Chapter 2. As in Chapter 2.

and it does not aﬀect either the buyer’s incentives nor the seller’s revenue. Qi is indeed increasing. For every i ∈ I the function ψi (θi ) is strictly increasing. For the allocation rule q described above the probability Qi (θi ) is the probability that ψi (θi ) is larger than zero and larger than ψi (θj ) for every j ∈ I with j = i. We have characterized the optimal choice of the allocation rule q and of the interim expected payments.CHAPTER 3. 0 Ti (θi ) = θi Qi (θi ) − θ ˆ ˆ Qi (θ)dθ.1. for all i ∈ I and θ ∈ Θ. although we know that such transfers can be found.1. 41 (3. We have not described the actual transfer schemes that make these choices incentive compatible and individually rational. Suppose that for every agent i ∈ I the cumulative distribution function Fi is regular. Thus. if ψi is increasing.4 (Myerson (1981)). qi (θ) = 0 otherwise.7) Note that we have ignored the case that ψi (θi ) = ψj (θj ) for some j = i. This is a zero probability event. this probability is an increasing function of θi . We now introduce an assumption under which this allocation rule satisﬁes the monotonicity constraint of the seller’s maximization problem. Assumption 3. BAYESIAN MECHANISM DESIGN The optimal allocation rule without monotonicity is then: 1 if ψi (θi ) > 0 and ψi (θi ) > ψj (θj ) for all j ∈ I with j = i. otherwise. Proposition 3. The assumption is that for all agents i ∈ I the distribution functions Fi are “regular” in the same sense as in Assumption 2. We have arrived at the following result. Clearly. . Among all incentive compatible and individually rational direct mechanisms those mechanisms maximize the seller’s expected revenue that satisfy for all i ∈ I and all θ ∈ Θ: (i) qi (θ) = 1 (ii) θ if ψi (θi ) > 0 and ψi (θi ) > ψi (θi ) for all i ∈ I with i = i. as required by the regularity assumption.

4. Expected welfare depends only on the allocation rule q. i=1 Note that this seller is no longer concerned with transfer payments. and hence the welfare maximizing seller always wants to transfer the object. the optimal mechanism prescribes that the object is given to the buyer with the highest value. An excellent reference on this is Chapter 2 of Krishna (2002). provided that this type is at least zero. Note that in the asymmetric case this need not be the case. with appropriately chosen minimum bid. To show this one has to derive equilibrium bidding functions for these auctions. He can choose any transfer payments such that Ti (θi ) ≤ θi Qi (θi ) for all i ∈ I. Note that welfare if the object is not transferred to one of the potential buyers is assumed to be zero.4 as follows: The expected revenue maximizing auction allocates the object to the buyer with the highest virtual type.CHAPTER 3.e.” Using this expression we can rephrase the result in Proposition 3. 3. and hence ψi (θi ) > ψj (θj ) ⇔ θi > θj . and verify that they imply the allocation rule and the transfer payments indicated in Proposition 3. BAYESIAN MECHANISM DESIGN 42 The expression ψi (θi ) is sometimes referred to as seller i’s “virtual type. the distribution functions Fi are all the same. if it is sold at all. i. maximization of the welfare function would require that the object be allocated to the potential buyer for whom θi is largest. Can the welfare maximizing seller allocate the object to the buyer with the highest . familiar auction institutions. in the symmetric case.2.5 Maximizing Welfare Suppose that the seller were not maximizing expected proﬁts but expected welfare. This is because then ψi is the same for all i. Thus. If buyers are symmetric. Let us assume that the seller uses the following utilitarian welfare deﬁnition: N qi (θ)θi . We can easily analyze this seller’s problem using the framework described before. Which rule q should the seller choose? If types were known. The seller can choose any rule q that is such that the functions Qi are monotonically increasing. In the symmetric case the optimal direct mechanism can be implemented −i using either a ﬁrst or a second price auction with minimum bid ψi (0) where −i ψi is the inverse of any one of the functions ψi . are optimal.

In comparing welfare maximizing and revenue maximizing mechanisms in the case that Assumption 3. The ﬁrst is that revenue maximizing mechanism allocates the object to the highest virtual type whereas the welfare maximizing mechanism allocates the object to the highest actual type. This is obviously the case.1 holds we observe that there are two diﬀerences.1.CHAPTER 3. Proposition 3. Note that this result does not rely on Assumption 3. Among all incentive compatible.4. In the symmetric case the functions ψi are the same for all i ∈ I and there is no diﬀerence between these two rules. BAYESIAN MECHANISM DESIGN 43 value even if he doesn’t know the valuations? We know that this is possible if the implied functions Qi are increasing. A second diﬀerence is that the revenue maximizing mechanism sometimes does not sell the object at all. 3. Ti (θi ) ≤ θi Qi (θi ) − θ ˆ ˆ Qi (θ)dθ. An example of a direct mechanism that is incentive compatible and individually rational and that maximizes welfare is the second price auction with reserve price zero. . like Proposition 3. otherwise. individually rational direct mechanisms a mechanism maximizes welfare if and only if for all i ∈ I and all θ ∈ Θ: (i) qi (θ) = (ii) θ 1 0 if θi > θj for all j ∈ I with j = i. This is an instance of the well-known ineﬃciency that monopoly sellers make goods artiﬁcially scarce.5.6 Numerical Examples We give one symmetric and one asymmetric example. Therefore we conclude: Proposition 3.2.5. whereas the welfare maximizing mechanism always sells the object. But in the asymmetric case the revenue maximizing mechanism might allocate the object ineﬃciently. does not describe the transfer scheme associated with a welfare maximizing mechanism in detail.

5. BAYESIAN MECHANISM DESIGN 44 ¯ Example 3. A ﬁrst or second price auction with reserve bid 0 maximizes expected welfare. it is sold to bidder 1 if: ψ1 (θ1 ) > ψ2 (θ2 ) ⇔ 2θ1 − 1 > 2θ2 − 1 ⇔ θ1 > θ2 . A ﬁrst or 1 second price auction with reserve bid 2 will implement this mechanism. Now suppose that N = 2. θ = 0. ¯ Example 3.9) The expected revenue maximizing auction will allocate the object to the buyer with the highest type provided that this type is larger than 0. and that F1 (θ1 ) = 2 2 . θ = 1.2.10) (3.11) . In the expected revenue maximizing auction the good is sold to neither bidder if: ψi (θi ) < 0 ⇔ 2θi − 1 < 0 ⇔ 1 θi < 2 holds for i = 1 and i = 2. and that θi is uniformly distributed so that F (θi ) = θi . 2: ψi (θi ) = θi − 1 − Fi (θi ) fi (θi ) 1 − θi = θi − 1 = 2θi − 1. We begin by calculating for i = 1. θ = 1.1. We begin by calculating ψ1 (θ1 ) = θ1 − = θ1 − 1 − F1 (θ1 ) f1 (θ1 ) 1 − (θ1 )2 2θ1 3 1 = θ1 − 2 2θ1 (3. (3. Thus. player 1 is more likely to have (θ1 ) whereas F2 (θ2 ) = 2θ2 − (θ2 ) high values than player 2. If the good is sold.1 is satisﬁed.8) Note that the regularity assumption 3.CHAPTER 3. (3. Suppose that θ = 0.

1 shows the optimal allocation of the good.13) (3. . bidder 1 wins the object only in a subset of all cases where his value is higher than bidder 2’s value. The 45◦ -line is shown as a dashed line.15) Figure 3. Note that the mechanism is biased against player 1. If the good is sold.CHAPTER 3. it is sold to bidder 1 if: ψ1 (θ1 ) > ψ2 (θ2 ) ⇔ 3 1 3 1 θ1 − > θ2 − ⇔ 2 2θ1 2 2 1 1 θ2 < θ1 − + 3θ1 3 (3. A second price auction will maximize expected welfare. In the expected welfare maximizing mechanism the object is allocated to player 1 if and only if his value is higher than player 2’s value. ψ2 (θ2 ) = θ2 − 45 (3.14) (3. although a ﬁrst price auction will not necessarily.12) In an expected revenue maximizing auction the good is sold to neither bidder if: ψ1 (θ1 ) < 0 ⇔ 3 1 θ1 − <0⇔ 2 2θ1 1 θ1 < 3 and ψ2 (θ2 ) < 0 ⇔ 3 1 θ2 − < 0 ⇔ 2 2 1 θ2 < 3 If the good is sold. BAYESIAN MECHANISM DESIGN and 1 − F2 (θ2 ) f2 (θi ) 1 − 2θ2 + (θ2 )2 = θ2 − 2 − 2θ2 1 − θ2 = θ2 − 2 3 1 = θ2 − 2 2 Again the regularity assumption is satisﬁed.

6 0.1: Expected revenue maximizing allocation in Example 3.3.4 no sale buyer 1 0. We shall refer to θi as . We denote this decision by g ∈ {0. If the public good is produced. θi is a random variable that follows a continuous distribution function Fi with density fi .CHAPTER 3. 2. then g = 0. BAYESIAN MECHANISM DESIGN 46 0. They have to choose whether to produce some indivisible. The theory of mechanism design began with the theory of mechanisms for the provision of public goods. The speciﬁc constraint on which we shall focus here is the government budget constraint. then g = 1.2 0. the example that we discuss in this section illustrates the design of optimal mechanisms subject to additional constraints beyond incentive compatibility and individual rationality constraints.0 0.4 θ1 0.0 0. This is a central application of the theory of mechanism design. . non-excludable public good. .2 0.8 1. . . 1}.0 buyer 2 θ2 0. Agent i’s utility if the collective decision is g and if she pays a transfer ti to the community is: θi g − ti .1 Public Goods Set-Up Our next example is a public goods problem. N } where N ≥ 2.6 0. Methodologically.3 3. If it is not produced.0 Figure 3. We consider a community consisting of N agents: I = {1. Here.8 1.2 3.

j ∈ I with i = j the random variables θi and θj are independent. . We denote by θ the vector (θ1 . The cumulative distribution function of θ will be denoted by F . The distribution F is common knowledge among the agents. but who knows F . is a model in which individuals can be selectively excluded from consuming the public good. The support of ¯ ¯ ¯ θi is [θ. We attribute to the mechanism designer a utilitarian welfare function with equal welfare weights for all agents. θ]. θN ). We also assume that each agent i observes θi . but not the other agent’ types θj where j = i. Deﬁnition 3. (3.2 Incentive Compatible and Individually Rational Direct Mechanisms As in previous parts of these notes. . . that is also of interest. However. The ¯ support of the random variable θ is Θ = [θ. . it is without loss of generality to restrict attention to incentive compatible direct mechanisms where agents’ payments are not random. and its density by f .16) 3. or as agent i’s valuation of the public good. We are thus considering an independent private values model of public goods. A “direct mechanism” consists of functions q and ti (for i ∈ I) where: q : Θ → {0. We shall consider this society from the perspective of a benevolent mechanism designer who does not observe θ. θ2 . so that a collective decision g implies costs cg. BAYESIAN MECHANISM DESIGN 47 agent i’s type. We assume that fi (θ) > 0 for all θi ∈ [θ. stochastic mechanisms would not present conceptual problems. . 1} and ti : Θ → R. θ]N . We assume that for i. The costs of producing the public good are assumed to be c > 0. To simplify our treatment of the budget constraint we also restrict attention to mechanisms where the decision about the public good is non-stochastic.4.3. and the results of this section hold even if stochastic mechanisms are considered.CHAPTER 3. θ] where 0 ≤ θ < θ. Welfare is thus: N N θi i=1 g− i=1 ti . The fact that the public good is non-excludable is reﬂected by the fact that the same variable g enters into all individuals utility function. An alternative model.

” For each agent i the function ti describes for every type vector θ the transfer that agent i makes when the types are θ.3 for deﬁnitions and to Propositions 3. and Ti (θi ) is the interim conditional expected value of the transfer that agent i makes to the community. We shall show that in our context this appearance is misleading. A direct mechanism is “ex ante budget balanced” if N ti (θ)f (θ)dθ ≥ Θ i=1 Θ cq(θ)f (θ)dθ. θ]N we have N ti (θ) ≥ cq(θ).3. This constraint requires that the money raised by the mechanism is at least enough to cover the costs of producing the public good. where we condition on agent i’s type being θi . θ] → R where Qi (θi ) is the interim conditional probability that the public good is produced. Given a direct mechanism. θ] → [0.2 we can refer to Deﬁnitions 3. For every ex ante budget balanced mechanism there is an equivalent ex post budget balanced mechanism. Finally. ex post budget balance implies ex ante budget balance. Clearly. Here.2 and 3.3 for characterizations of incentive compatibility and individual rationality. we also deﬁne agent i’s expected utility Ui (θi ) conditional on her type being θi . BAYESIAN MECHANISM DESIGN 48 The function q assigns to each type vector θ the collective decision about the public good that is produced if the agents’ types are θ. A restrictive version of the constraint requires budget balance for each realization of agents’ types. As the notation that we have introduced in the previous paragraph parallels that of Section 3. A direct mechanism is “ex post budget balanced” if for ¯ every θ ∈ [0. As before we shall restrict attention to mechanisms that are incentive compatible and individually rational. across realization of agents’ types. Ex post budget balance appears to be more restrictive. 3. i=1 An alternative formulation requires budget balance to hold only on average.CHAPTER 3.5. This is given by: Ui (θi ) = Qi (θi )θi −Ti (θi ). We shall refer to q also as the “decision rule. again conditioning on agent i’s type being θi .3 Ex ante and Ex post Budget Balance We now introduce the government budget constraint.2 and 3. we deﬁne for each agent i ∈ I functions ¯ ¯ Qi : [θ. we deﬁne “equivalent” as follows: .6. 1] and Ti : [θ. Deﬁnition 3. Deﬁnition 3.

So far. she does not cover that part of the ex post deﬁcit that is predicted by her own signal. We denote by Tj (θi ) the expected value of agent j’s transfer. then the same is true for the other. pays for the expected value of the deﬁcit conditional on agent 1’s signal. is the same in the two mechanisms. We construct the payments in the ex post budget balanced scheme by modifying the payments in the ex ante budget balanced scheme as follows. and. This is because it would have expected value zero if agent 1’s true type were θ1 . Suppose ﬁrst that the ex ante budget balance condition holds with equality. conditioning on agent i’s signal being θi . One arbitrarily selected agent. Formally. We denote the payments in the ex ante budget balanced mechanism by ti . then the same is true for the other. θ] agent i’s expected transfers. we add to her original payment the ex post deﬁcit minus the expected value of the deﬁcit conditional on her own signal. BAYESIAN MECHANISM DESIGN 49 Deﬁnition 3. Notice that if two mechanisms are equivalent. then agent 1’s payment is in the modiﬁed mechanism: N N t1 (θ) + cq(θ) − i=1 ti (θ) − cQ1 (θ1 ) − i=1 Ti (θ1 ) . some other arbitrarily selected agent. Proposition 3.6. Two direct mechanisms are “equivalent” if for all agents ¯ i ∈ I and for all types θi .CHAPTER 3. independent of whether agent 1 reports her type truthfully or not. (3.7. the conditional distribution of this random variable. θi ∈ [θ. Second. is the same if agent 1’s true signal is θ1 rather than θ1 . and if one of them is individually rational. conditional on agent i’s type being θi and agent i reporting to be type θi . moreover. The expression that we have added to agent 1’s payment is a random variable with expected value zero.17) We now check that agent 1’s expected payoﬀ. provides the primary coverage of the deﬁcit. say agent 2. Her modiﬁed . Proof. and if one of them is incentive compatible. if the vector of types is θ. For every direct mechanism with decision rule q that is ex ante budget balanced there is an equivalent direct mechanism with the same decision rule q that is ex post budget balanced. we have employed this notation only in the case that j = i. say agent 1. by the independence of private signals. In other words. conditional on her type being θ1 and her reporting that her type is θ1 is unchanged. However. Now we use it also in the case that j = i.

Moreover. (3. tN ) would the designer choose if she were not constrained by incentive compatibility and individual rationality. Therefore. If there is an ex ante budget surplus. .3. N N −Ui (θ) + i=1 Θ q(θ) i=1 θi − 1 − Fi (θi ) fi (θi ) − c f (θ)dθ.4 Welfare Maximization Which mechanism (q. 2 pay the same as before: ti (θ). and she reports that her type is θ2 . . Adding up all agents’ payments shows that the sum of the payments equals the costs cq(θ) in each state θ. . and that therefore this mechanism is ex post budget balanced. Finally.18) Note that the random variable that we are adding to agent 2’s payment is independent of agent 2’s report. Finally. We shall assume that the mechanism designer in this section considers only ex post budget balanced mechanisms.19) 3. but . we subtract from some agent’s payments a constant until the mechanism is exactly budget balanced. BAYESIAN MECHANISM DESIGN payment is: t2 (θ) + cQ1 (θ1 ) − i=1 50 n Ti (θ1 ). agent 2’s expected payoﬀ.CHAPTER 3. We calculate the ex ante expected revenue from an incentive compatible mechanism in the same way as we calculated above the expected revenue from an incentive compatible auction. t1 . it has ex ante expected value zero because the mechanism is ex ante budget balanced. Then we add the constant again to this agent’s payments. the expectation of the added term conditional on agent 2’s signal is the same as its ex ante expectation. This yields the following formula where we ﬁnd it convenient to write the initial term as interim expected utility of the lowest type. . (3. We shall work with either condition. because agent 2’s signal does not provide any information about agent 1’s signal. all agents i = 1. The mechanism that we obtain has the required properties. t2 . is unchanged. The above result makes clear that this is equivalent to requiring ex ante budget balance. Then we conduct the transformation described above. if her true type is θ2 . whichever is more convenient.

the utilitarian designer with equal welfare weights for all agents does not care how costs are distributed among agents. As welfare function (3.21) i i=1 0 otherwise.CHAPTER 3. From (3.59) it is then clear that the optimal decision rule is:8 N 1 if i=1 θi ≥ c ∗ q (θ) = (3.20) 0 otherwise. we require arbitrarily that q ∗ (θ) = 1 if N θi = c. This simpliﬁes i=1 the exposition.59) she would never raise transfers larger than what is required to cover cost. Analogously.” The following impossibility result shows that in all non-trivial cases no ﬁrst best mechanism is incentive compatible and individually rational. 8 . If N θ ≥ c then a mechanism where the public good is always proc duced and all agents pay N is ﬁrst best and incentive compatible and in¯ ≤ c then a mechanism where the public good is dividually rational. Proof. An incentive compatible and individually rational ﬁrst ¯ best mechanism exists if and only if either N θ ≥ c or N θ ≤ c. Therefore. for all type vectors it is eﬃcient to produce the public ¯ good. Thus. BAYESIAN MECHANISM DESIGN 51 had to satisfy ex post budget balance? As the mechanism designer subtracts transfer payments in her welfare function (3.7. It remains to prove the converse. These are trivial cases. Proposition 3. If N θ never produced and no agent ever pays anything is ﬁrst best and incentive compatible and individually rational.59) indicates. The condition N θ ≥ c means that even if all agents have the lowest valuation the sum of the valuations is at least as high as the cost of producing the public good. and wish to prove that there is no incentive compatible and individually rational ﬁrst best mechanism. any transfer rules that satisfy: N c if q(θ) = 1 t∗ (θ) = (3.7 is an impossibility result. are optimal. To prove this we display a direct mechanism P In this deﬁnition. We call these direct mechanisms “ﬁrst best. Thus we consider the case N θ < c < ¯ N θ. N θ ≤ c means that for all type vectors it is eﬃcient no to produce the public good. Proposition 3. and could be changed without conceptual diﬃculty. For all non-trivial cases.

22) . To see why this mechanism is called “pivot” mechanism it is useful to ignore the ﬁrst term in the sum on the right hand side of the formula for ti (θ). We are going to show that truthful reporting is optimal whatever the other agents’ types θ−i are. BAYESIAN MECHANISM DESIGN 52 that makes the ﬁrst best decision rule q ∗ incentive compatible and individually rational. Lemma 3. and the sum of all other agents’ valuations of the project. We then argue that this mechanism maximizes expected transfer payments among all incentive compatible and individually rational ﬁrst best mechanisms. θ−i ) = 1. and hence agent i pays only if her report is “pivotal” for the collective decision. The assertion then follows. θ−i ) c − j=i θj = q ∗ (θi .6. θ−i ) − q ∗ (θi . Agent i’s report changes the collective decision if q ∗ (θ) − q ∗ (θ. θ−i )) c − j=i θj ¯ for all i ∈ I and θ ∈ [θ. we show that it yields an expected budget deﬁcit in all non-trivial cases. Proof. Fix the other agents’ types as θ−i . The “pivot mechanism” is the mechanism that is given by the ﬁrst best decision rule q ∗ and by the following transfer scheme: ti (θ) = θq ∗ (θ. If we leave out terms that do not depend on agent i’s report. The second term equals the change to the social welfare of all other agents caused by agent i’s report. Consider an agent i ∈ I who is of type θi and who contemplates reporting that she is of type θi = θi . The pivot mechanism is incentive compatible and individually rational. Finally. Here we compare the actual outcome to the outcome that would have occurred had agent i reported the lowest type θ. then agent i’s utility if reporting θi is: θi q ∗ (θi .8. In that case agent i pays for the diﬀerence between the costs of the project. This obviously implies that truthful reporting is a Bayesian equilibrium.CHAPTER 3. θ−i ) N j=1 θj − c . θ−i ) + (q ∗ (θ) − q ∗ (θ. (3. Deﬁnition 3. θ]N . This term does not depend on agent i’s report θi .

and that has the eﬀect of aligning the agent’s incentives with social welfare. This ensure that individual rationality is satisﬁed for all agents.9. Because q is ﬁrst best. This proves incentive compatibility. a part that depends on the agent’s report. q ∗ (θ The proof reveals two important features of the formula in Deﬁnition 3.CHAPTER 3. The second feature is that those parts of agent i’s transfer payment that do not depend on agent i’s report are chosen so as to equalize agent i’s utility if she is of the lowest type with her reservation utility of zero. For the pivot mechanism the A type’s individual rationality constraint is “most restrictive” if that type’s individual rationality constraint implies all other type’s individual rationality constraints. The “pivot mechanism” is a special “Vickrey-Clarke-Groves” (VCG) mechanism. BAYESIAN MECHANISM DESIGN 53 Thus agent i’s utility is true social welfare if the collective decision is ∗ i . Lemma 3. a mechanism is a VCG mechanism if every agent’s payment consists of two terms. By a result analogous to Lemma 3. a part that does not depend on the agent’s report. θ−i ).19 the expected budget surplus of an incentive compatible direct mechanism that implements the ﬁrst best decision rule q ∗ equals the interim expected payments of the lowest types plus a term that is the same for all such rules. See Section 4 for an example where this type is not the lowest type in types’ numerical order.”9 The next result indicates the special importance of the pivot mechanism. the second part is chosen so as to ensure individual rationality as an equality constraint for the type whose individual rationality constraint is “most restrictive.7. 9 . The ﬁrst is that those parts of agent i’s transfer payment that depend on agent i’s report are chosen so that agent i’s incentives are exactly aligned with social welfare. In general. To verify individual rationality note that the expected utility of agent i obviously equals zero if her type is θ. No incentive compatible and individually rational mechanism that implements the ﬁrst best decision rule q ∗ has larger expected surplus than the pivot mechanism. Secondly. agent i’s utility is maximized if she reports truthfully θi = θi . In a pivot mechanism. By Corollary 3. Proof. If a mechanism is individually rational then the interim expected payments of the lowest types can be at most such that the expected utility of the lowest types are zero. and the mechanism is individually rational. In a VCG mechanism in general. this second part can be arbitrary. Firstly.2 this implies that all types’ interim expected utility is at least zero.

and hence N θ < c. Let P be the set of all i ∈ I for which this holds. We conclude the proof by showing that the pivot mechanism has an ex ante expected deﬁcit except in trivial cases. Proof. there are no costs and no agent pays any transfer. In this case.CHAPTER 3. ¯ Lemma 3. total payments are less than c and there is a deﬁcit. Abusing notation slightly denote by P the number of elements of P . Consider next θ such that q ∗ (θ) = 1 and q ∗ (θ. θ−i ) = 1 for every i ∈ I. and q ∗ (θ. The total transfers are: c − ı∈P j=i θj + i∈N P θ θj + j∈P i∈N P =P c − P j∈N P θj − (P − 1) θ θj − (θi − θ) i∈N P =P c − (P − 1) j∈N P θj − (P − 1) j∈P =P c − (P − 1) j∈I θj − i∈N P (θi − θ). and by N P the number of elements of N P .8. By assumption the probability of the production of the good according to the ﬁrst best decision rule is not 1. and call these agents “pivotal. This implies that the ex ante expected surplus is negative. individually rational direct mechanism can have higher expected surplus than the pivot mechanism. no incentive compatible.” Deﬁne N P = I \ P . If N θ < c < N θ. BAYESIAN MECHANISM DESIGN 54 expected utilities of the lowest types are exactly equal to zero. Consider ﬁnally states θ such that q ∗ (θ) = 1. then the ex ante expected surplus of the pivot mechanism is negative. Therefore. (3. We show that the ex post surplus of the pivot mechanism is always non-positive and with positive probability negative. Therefore. In this case each agent pays θ. Hence the deﬁcit is zero.23) We shall show that this is no more than c: P c − (P − 1) j∈I θj − i∈N P (θi − θ) ≤ c ⇔ (P − 1)c ≤ (P − 1) j∈I θj + i∈N P (θi − θ) ⇐ . θ−i ) = 0 for some i ∈ I. Consider ﬁrst θ such that q ∗ (θ) = 0.

with probability 1 we have that θi > θ for all i ∈ N P .CHAPTER 3. Moreover.24) c≤ j∈I θj which is true by construction in states in which q ∗ (θ) = 1. payments will add up to c. to think of q and the interim expected utilities of the lowest types.7 impossible to implement q ∗ using a mechanism that is incentive compatible. This concludes the proof of Proposition 3. BAYESIAN MECHANISM DESIGN (P − 1)c ≤ (P − 1) j∈I 55 θj ⇔ (3. Moreover. individually rational. and hence there is an expected deﬁcit. in each state in which the public good is produced. In this case. and more convenient. The constraints that these variables have to satisfy are: for every i ∈ I the function Qi is monotonically increasing . Our objective is to determine direct mechanisms that maximize expected welfare among all incentive compatible.7. individually rational. it is equivalent. In the remainder of this subsection we focus on the case in which it is by Proposition 3. Thus. and ex ante budget balanced. We shall refer to these mechanism as “second best. However. The designer’s objective function can therefore be written as: q(θ) Θ i∈I θi − c f (θ)dθ (3. conditional on such a state occurring. If there is a surplus then the mechanism designer can return it to the agents. States in which the public good is produced and some agents are pivotal occur with positive probability under our assumption.25) It seems at ﬁrst sight most natural from our discussion so far to regard q and the interim expected payments of the lowest types. the mechanism designer can achieve ex post budget balance when she achieves ex ante budget balance. and in other states they will be add up to zero. the above calculation shows that the surplus is strictly negative.” It is without loss of generality to assume that the mechanism designer balances the budget exactly rather than leaving a surplus. Ui (θ) as the choice variables. and ex ante budget balanced mechanisms. as the designer’s choice variables. Ti (θ).

31) . for every i ∈ I : Ui (θ) ≥ 0 (individual rationality constraint).30) and. the payments Ti (θ) will be chosen so as to satisfy exact budget balance. λ = 0 only if: N q(θ) Θ i=1 θi − 1 − Fi (θi ) fi (θi ) − c f (θ)dθ > 0.26) (3. and instead write the budget constraint as: N q(θ) Θ i=1 θi − 1 − Fi (θi ) fi (θi ) − c f (θ)dθ ≥ 0 (3. (3. apply. 221.26) as well. Under those conditions the solution to the relaxed problem has to be a solution to the original problem.CHAPTER 3.28) − i=1 Ui (θ) + Θ q(θ) i=1 θi − 1 − Fi (θi ) fi (θi ) (budget constraint). 217 and p. According to these results q solves the relaxed maximization problem if and only if there is a Lagrange multiplier λ ≥ 0 such that q maximizes N q(θ) Θ i∈I θi − c f (θ)dθ + λ Θ q(θ) i=1 θi − 1 − Fi (θi ) fi (θi ) − c f (θ)dθ. p. BAYESIAN MECHANISM DESIGN (incentive constraint). We now solve the mechanism designer’s problem by ﬁrst considering the relaxed problem where we neglect the monotonicity constraint (3.26).29) with the understanding that if the left hand side is strictly positive. (3. moreover.26) actually happens to satisfy condition (3. Then we shall discuss conditions under which the solution to the problem without constraint (3. such as function spaces. Theorems 1 and 2 in Luenberger (1969). We now eliminate the choice variables Ui (θ) from the problem. N N 56 (3.27) − c f (θ)dθ = 0 (3. To solve the relaxed problem we use a version of the Kuhn Tucker Theorem that applies to inﬁnite dimensional vector spaces.

1 and 3.CHAPTER 3. that has weight 1 in ψi . .e.33) satisﬁes for every λ > 0 the monotonicity constraint (3. if and only if: . This leads to the following decision rule: N N λ 1−Fi (θi ) 1 if i=1 1+λ fi (θi ) i=1 θi > c + q ∗ (θ) = (3. hence concave.1.34) is strictly increasing for every λ > 0. we can conclude: ¯ Proposition 3. then θi − λ 1 − Fi (θi ) 1 + λ fi (θi ) (3. the function ψi (θi ) = θi − 1−F (θi ) is strictly increasing. Suppose N θ < c < N θ. i.33) 0 otherwise.2. . Note that we must have λ > 0 if Proposition 3. has weight λ/(1 + λ) < 1 in (3.7 applies because with λ = 0 the rule (3.32) It is evident. . t1 . t2 . Now we introduce an assumption under which the rule (3.8.33) becomes the ﬁrst best rule. . BAYESIAN MECHANISM DESIGN 57 We don’t go through the details of checking the applicability of these results. However. from point-wise maximization. that the Lagrange function is maximized if we set q(θ) = 1 whenever the expression in the square brackets is positive. This implies that the second best rule in (3. Assumption 3.33) satisﬁes the monotonicity condition (3.30) as: q(θ)(1 + λ) Θ i∈I θi − λ 1 − Fi (θi ) 1 + λ fi (θi ) − c f (θ)dθ. and ex ante budget balanced. and because the objective function that we seek to maximize is linear. and maximizes expected welfare among all such mechanisms. individually rational. (3. For every i ∈ I the cumulative distribution function Fi is regular. Then a direct mechanism (q.34). we note that we obtain a necessary and suﬃcient condition because in our problem the set of admissible functions q is convex. tN ) is incentive compatible. The condition is the regularity assumption that we introduced in Assumptions 2. and that for every agent i ∈ I the function Fi is regular.26). We can write the Lagrange function (3.26). Therefore. This is because the potentially decreasing term. f (θi ) If an agent i’s cumulative distribution function Fi is regular.

This leaves q as the only choice variable. Ti (θi ). . Proﬁt maximization requires that the transfer payments of the lowest types are set equal to those types’ expected utility. (iii) for all i ∈ I: θ Ti (θi ) = θi Qi (θi ) − θ ˆ ˆ Qi (θ)dθ.3. We can see that the second best mechanism undersupplies the public good.5 Proﬁt Maximization We now consider brieﬂy the problem of choosing the mechanism that maximizes the designer’s expected proﬁts among all incentive compatible and individually rational direct mechanisms. (ii) N 58 λ 1−Fi (θi ) 1+λ fi (θi ) q(θ) Θ i=1 θi − 1 − Fi (θi ) fi (θi ) − c f (θ)dθ = 0. BAYESIAN MECHANISM DESIGN (i) there is some λ > 0 such that for all θ ∈ Θ: N N 1 if i=1 θi > c + i=1 q(θ) = 0 otherwise. and the transfer payments of the lowest types.35) The mechanism designer has to respect the constraint that for every i ∈ I the function Qi must be monotonically increasing. An interesting question is whether there are simple and appealing indirect mechanisms that implements the second best. 3.CHAPTER 3. We do not assume that the lowest type is necessarily 0. Like the welfare maximizing mechanism designer also the proﬁt maximizing mechanism designer has two choice variables: the allocation rule q. The expected proﬁt from decision rule q can be calculated as previously in similar contexts as: N q(θ) Θ i=1 θi − 1 − Fi (θi ) fi (θi ) − c f (θ)dθ (3. The public good is produced only if the sum of the valuations is larger than a lower bound that is strictly larger than c.

. We only state the result. By Proposition 3.8 a necessary condition for a direct incentive compatible. Then a direct mechanism (q. otherwise. . We wish to determine ﬁrst the expected welfare maximizing mechanism. tN ) is incentive compatible. and maximizes expected proﬁt among all such mechanisms if and only if for all i ∈ I and all θ ∈ Θ: (i) q(θ) = (ii) θ 1 0 if N i=1 θi >c+ N 1−Fi (θi ) i=1 fi (θi ) . Comparing Propositions 3. By Proposition ¯ 3.6 A Numerical Example We provide a very simple numerical example.3. .9. and hence N θ < ¯ c < N θ.36) . t2 . Suppose that for every agent i ∈ I the cumulative distribution function Fi is regular. BAYESIAN MECHANISM DESIGN 59 The mechanism designer’s problem can be solved using analogous reasoning as in the auction model of the previous section. and θ = 1. individually rational.CHAPTER 3. Note that the regularity assumption is satisﬁed as was veriﬁed in the numerical example of the previous section.3. θi is uniformly distributed on [0. as θ = 0. 3. t1 . yet the calculations turn out to be somewhat involved. and 0 < c < 2. and individually rational. Example 3. 2. Suppose N = 2. The probability with which the ﬁrst best rule calls for the production of the public good is strictly between zero and one. and ex ante budget balanced mechanism to maximize expected welfare among all such mechanisms is that there exists some λ > 0 such that q(θ) = 1 if and only if: λ 1 − θ1 1 − θ 2 + 1+λ 1 1 1+λ λ θ 1 + θ2 > c+ 2 1 + 2λ 1 + 2λ θ 1 + θ2 > c + ⇔ (3. . 1] for i = 1. Proposition 3.9 we ﬁnd that the proﬁt maximizing supplier for the public good supplies a lower quantity than the welfare maximizing mechanism designer.7 the ﬁrst best cannot be achieved.8 and 3. Ti (θi ) = θi Qi (θi ) − θ ˆ ˆ Qi (θ)dθ.

37) 2 Next we calculate the expected payment by the agent 1: 1 0 0 s 0 1 q(θ) θ1 − 1 1 − F1 (θ1 ) f1 (θ1 ) f θ)dθ 1 1 = s−θ1 (2θ1 − 1)dθ2 dθ1 + s 0 (2θ1 − 1)dθ2 dθ1 (3. Then the expected costs of producing the public good are: 1 C(s) = (2 − s)2 c. as required by the Lagrange conditions. and the total revenue of a mechanism. Therefore. Then the expected costs of producing the public good are: 1 C(s) = 1 − s2 c. that the interim expected payments of all other types are as required by incentive compatibility. To proceed. 3 (3. (3. and that the budget surplus on the right hand side of equation (3.29) is zero. We distinguish two cases.40) 2 The expected payment by agent 1 is given by: 1 0 0 1 q(θ) θi − 1 − Fi (θi ) fi (θi ) f θ)dθ . denoted by R(s).39) Case 2: Suppose s > 1.CHAPTER 3. (3. the total expected revenue of the mechanism is: 2 R(s) = − s3 + s2 .38) 1 1 = − s+ s2 3 2 Agent 2’s expected payment will be the same. denoted by C(s). BAYESIAN MECHANISM DESIGN 60 Denote the right hand side of this inequality by s. for incentive compatible.8 thus means that we can restrict our search for second best mechanisms to those mechanisms for which q(θ) = 1 if and only if θ1 + θ2 ≥ s for some s ∈ (c. 2). We determine s by assuming that the interim expected payments of the lowest types are zero. we calculate the total expected costs of producing the public good. We seek to ﬁnd the an appropriate value of s. Case 1: Suppose s ≤ 1. Proposition 3. as required by Proposition 3.8. individually rational threshold mechanisms.

3 3 (3. We obtain: (3. and therefore the total expected revenue from the mechanism is: R(s) = 1 2 − (s − 1)2 + (s − 1)3 . The term in the last line is positive if and only if: c − 2(s − 1) > 0 ⇔ s<1+ c 2 (3. It clearly tends to a positive limit as c → 0. Note ﬁrst that: D(0) = −c < 0 and D(2) = 0. 1 < s < 2 ⇒ D (s) = −2(s − 1) + 2(s − 1) + (2 − s)c = (2 − s)(c − 2(s − 1)).48) We want to prove that the expression in the last line is strictly positive.43) Next. Observe that: 0 < s < 1 ⇒ D (s) = −2s2 + (2 + c)s = s(2(1 − s) + c) > 0. To understand the set of solutions to this equation we ﬁrst investigate the sign of the derivative of D with respect to s. we investigate the sign of D(s) for some important values of s. then the expression tends . 6 2 3 (3. so that the condition for s becomes D(s) = 0.46) (3.45) (3. we investigate the sign of D at s = 1 + 2 .41) = Agent 2’s expected payments will be the same.CHAPTER 3. c Next. If c → 2.44) 2 (3.47) D 1+ c 2 2 c 3 1 c 1 c = − + − 1− 2 3 2 3 2 2 2 1 1 c 2 1 c 3 = − c+ − 3 2 2 3 2 2 c (3.42) We now deﬁne D(s) ≡ R(s) − C(s). BAYESIAN MECHANISM DESIGN 1 1 61 = s−1 s−θ1 (2θ1 − 1)dθ2 dθ1 1 1 1 − (s − 1)2 + (s − 1)3 .

it will be between 1 and 1 + 2 . 3 3 2 It so happens that this equation has a simple analytical solution: s= We can now sum up: 1 3 + c.49) We can conclude that D(1 + 1 c) > 0. then we ﬁnd 3 s by solving the following equation: 1 2 1 − (s − 1)2 + (s − 1)3 − (2 − a)2 c = 0. For this we investigate the value of D(1): D(1) = 1 1 − c>0⇔ 3 2 2 c< 3 (3.CHAPTER 3.52) . 3 2 (3. 2). then we ﬁnd s by solving the following equation: 3 2 1 − s3 + s2 − 1 − s2 c = 0. and the solution s = 2. it is suﬃcient to show that the derivative of this expression is negative for 0 < c < 2. If c > 2 .53) (3.50) 2 Thus. the solution of D(s) = 0 will be between 0 and 1 if and only if c < 3 . 1+ 2 ). This derivative is: − 1 c 1 c 2 + − 2 2 2 2 c c 1 = 1− − 2 2 2 1 1 ≤ − 4 2 1 =− 4 (3. Thus. BAYESIAN MECHANISM DESIGN 62 to zero. For the purpose of calculating the solution.51) Unfortunately. 2 4 (3. this equation has no simple solution. We can discard the latter as we require s ∈ (c. c Otherwise. If c < 2 . 2 From our results so far we can conclude that the equation D(s) = 0 has c exactly two solutions: one solution in the interval (0. it is useful to ask whether the solution of D(s) = 0 is larger or less than 1.

then 3 s= 1 3 + c. or it is prohibitively expensive. 3 2 (ii) If c ≥ 2 . 1] of the equation: 3 1 2 − s3 + s2 − 1 − s2 c = 0. either because the production of the public good is free. . 2 4 The expected proﬁt maximizing mechanism follows straightforwardly from Proposition 3. where s is determined as follows: (i) If c < 2 . The expected proﬁt maximizing mechanism designer will choose a mechanism with an allocation rule q such that: 1 if θ1 + θ2 > s. and hence it should never be produced. Figure 3.2 shows that the second best threshold chosen by a welfare maximizing mechanism designer is strictly larger than the ﬁrst best threshold. 2 To illustrate our ﬁndings. The ineﬃciencies become small as c approaches either 0 or 2. The utilitarian mechanism designer will choose a mechanism with an allocation rule q such that: 1 if θ1 + θ2 > s.9: Proposition 3. ∗ q (θ) = 0 otherwise.CHAPTER 3. then s is the unique solution in [0. except if c is either 0 or 2.10.2 the optimal threshold s for the utilitarian case (dashed and dotted line) and for the proﬁt maximization case (dashed line) as a function of c. For comparison we have also plotted the 45◦ line which is the ”ﬁrst best” threshold (unbroken line). BAYESIAN MECHANISM DESIGN 63 Proposition 3. where s = 1 + 1 c. we have plotted in Figure 3. These two extreme cases correspond to cases in which the mechanism designer doesn’t really have to induce truthful revelation of individuals’ types. q ∗ (θ) = 0 otherwise.11. and hence it should always be produced. This reﬂects the fact that to provide incentives the mechanism designer must accept some ineﬃciencies.

BAYESIAN MECHANISM DESIGN 64 0. the monopolist will not want to provide the good. Moreover.4. Thus.5 2. not only the .0 c 1. and the utilitarian solution.5 1.CHAPTER 3. Second Best.5 2. This corresponds to the standard textbook insight that a monopolist artiﬁcially restricts supply to raise price. The monopoly distortion exists even if production is free (c = 0). 3. If production becomes prohibitively expensive (c = 2).2 also shows that a monopoly supplier of the public good would choose a threshold that is even larger than the threshold chosen by the utilitarian mechanism designer.0 Figure 3.2: Thresholds in First Best.1 Bilateral Trade Setup In our next example a seller of a single indivisible good faces one buyer with unknown valuation. but now we consider the situation from the perspective of a mechanism designer who wants to arrange a trading institution for the two parties that guarantees that they trade if and only if the buyer’s value is larger than the seller’s. as in Section 2. as in the standard textbook case.2.0 0.0 0.0 0.4 3.5 1. the monopoly solution coincides with the ﬁrst best. the distortion introduced by a monopolist is zero.0 s 1. In this case. and Proﬁt Maximum Figure 3.

θS ]. It is the simplest example that one might analyze when seeking to build a general theory of the design of optimal trading institution. If he does not sell the good and receives a transfer t then his utility is θS + t where θS is a random variable with cumulative distribution function FS and density fS . Methodologically. We deﬁne ¯ ¯ θ = (θS . but also the seller’s valuation is unknown to the designer of this trading institution. We denote the joint distribution of θ by F with density f . The function q assigns to each type vector θ an indicator variable that indicates whether trade takes place (q(θ) = 1) or whether no trade takes place (q(θ) = 0).CHAPTER 3. Deﬁnition 3.9. 3. θB ] and that fB (θB ) > 0 for all ¯B ]. We assume that FS ¯ ¯ has support [θS . The seller’s utility if he sells the good and receives a transfer payment t is equal to t. where θB is a random variable with cumulative distribution function FB and density ¯ fB . such as stock exchanges or commodity markets. There is one potential buyer B. θ transfer t is −t. The random variables θS and θB are independent. we restrict attention to deterministic trading rules. θB ].4. A seller S owns a single indivisible good. B}. it is interesting to see how one can treat the bilateral trade problem with similar methods as the public goods problem. We assume that FB has support [θB . This example was ﬁrst analyzed in the very well-known paper by Myerson and Satterthwaite (1983). A “direct mechanism” consists of functions q. We shall refer to q also as the “trading rule.2 Direct Mechanisms The revelation principle implies that we can restrict attention to direct mechanisms. tS and tB where: q : Θ → {0. although it is seemingly quite diﬀerent. 1} and for i = S. θS ] and that fS (θS ) > 0 for all θS ∈ [θS . θS ] × [θB . and it may be that valuations are such that trade is not eﬃcient. The seller only observes θS . The buyer’s utility if he purchases the good and pays a transfer t equals θB − t. and the buyer only observes θB . θB ) and Θ = [θS . B: ti : Θ → R. Deﬁne I = {S. The buyer’s utility if he does not obtain the good and pays θB ∈ [θB . BAYESIAN MECHANISM DESIGN 65 buyer’s valuation.” For simplicity. The function tS indicates .

This is given by: US (θS ) = TS (θS ) + (1 − QS (θS ))θS and UB (θB ) = QB (θB )θB − TB (θB ). We shall restrict attention to direct mechanisms that are incentive compatible. and the function tB indicates transfers that the buyer makes. θi ] → [0. BAYESIAN MECHANISM DESIGN 66 transfers that the seller receives. we deﬁne for each agent i ∈ {S. and that TS is given by: ¯ θS ¯ ¯ ¯ TS (θS ) = TS (θS ) + (1 − QS (θS ))θS − (1 − QS (θS ))θS − for all θS . as we shall do below.e. (1 − QS (x)) dx θS (3. then the seller’s individual rationality ¯ condition holds if and only if it holds for the highest seller type θS . again conditioning on agent i’s type being θi . it seems redundant to introduce separate notation for the buyer’s payment and the seller’s receipts. and ex post budget balanced. and Ti (θi ) is the conditional expected value of the transfer that agent i receives (if i = S) or makes (if i = B). that the seller receives what the buyer pays. For the seller. the standard arguments apply if types are ordered in the reverse of the numerical order. individually rational. that is. However. that is starting with high types rather than low types. a necessary and suﬃcient condition for incentive compatibility for the seller is that QS is decreasing. i. the seller trades voluntarily and obtains an expected utility that is at least as large as his utility would be if he kept the good. Therefore. We shall mostly assume in this section10 that tS = tB . If a mechanism is incentive compatible. θi ] → R where Qi (θi ) is the conditional probability that trade takes place. For the seller. where we condition on agent i’s type being θi . Standard arguments show that a mechanism is incentive compatible for the buyer under exactly the same conditions as before. Thus. Finally. . B} functions ¯ ¯ Qi : [θi . 1] and Ti : [θi . Individual rationality is deﬁned as before. This allows us then to adopt a similar methodology as in the previous section. individual rationality means that US (θS ) ≥ θS for all θS .CHAPTER 3.4 where we consider proﬁt maximizing trading mechanisms. we also deﬁne agent i’s expected utility Ui (θi ) conditional on her type being θi . Given a direct mechanism. To prove 10 An exception is Subsection 3. it is much more convenient to begin with a more general framework in which these variables are not identical. and then to introduce an ex post budget balance condition which forces them to be equal to each other.4.54) Individual rationality for the buyer is deﬁned and characterized in the same way as before.

the function that the seller maximizes shifts downwards. We shall work with this condition. The by now familiar calculation shows that for an incentive compatible direct mechanism the seller’s expected payment is: ¯ US (θS ) − Θ (3.59) If the mechanism designer were not constrained by incentive compatibility and individual rationality. . mechanisms for which the seller’s ex ante expected payment is equal to the buyer’s ex ante expected payment. By the same argument as in the previous section it suﬃces to identify ex ante budget balanced mechanisms.58) 3. and therefore the function’s maximum value decreases. but only had to respect ex post budget balance. then the mechanism designer would choose a “ﬁrst best mechanism” where the trading rule is: 1 if θB ≥ θS . (3. As θS decreases. that is.CHAPTER 3. (3.60) 0 otherwise. Ex post budget balance requires that in each state θ we have: tS (θ) = tB (θ).56) The seller maximizes over θS . BAYESIAN MECHANISM DESIGN 67 this we show that the diﬀerence US (θS ) − θS is decreasing in θS .3 Welfare Maximization The mechanism designer seeks to maximize the sum of the individuals’ utilities.55) (1 − q(θ)) θS + FS (θS ) fS (θS ) f (θ)dθ (3.4.57) and the buyer’s expected payment is: −UB (θB ) + Θ q(θ) θB − 1 − FB (θB ) fB (θB ) f (θ)dθ. q ∗ (θ) = (3. A seller who is of type θS and pretends to be of type θS obtains expected utility: θS (1 − QS (θS )) + TS (θS ). This exceeds θS by: TS (θS ) − θS QS (θS ). (3. q(θ)θB − tB + (1 − q(θ))θS + tS =θS + q(θ) (θB − θS ) + tS − tB .

θB )θS + q ∗ (θ) − q ∗ (θS . individually rational and ex post budget balanced ﬁrst best mechanism. BAYESIAN MECHANISM DESIGN 68 where the decision in the case of equality of values is arbitrary. Proposition 3. Our ﬁrst objective is to prove that in almost all cases no ﬁrst best mechanism is incentive compatible and individually rational.CHAPTER 3. . ¯ Proof.12 is trivial. 1983). Finally. individually rational and ex-post budget balanced direct mechanism ¯ ¯ with decision rule q ∗ exists if and only if θB ≥ θS or θS ≥ θB . The payment rule is arbitrary as long as it is ex post budget balanced. We ﬁrst display a mechanism that is incentive compatible and individually rational and that implements the ﬁrst best trading rule. then a mechanism under which no trade takes place and no payments are made is ﬁrst best ¯ and has the required properties. Deﬁnition 3.12 (Myerson and Satterthwaite. ¯ The condition θB ≥ θS implies that trade is always at least weakly ¯ eﬃcient. there is no incentive compatible. The condition θS ≥ θB implies that eﬃciency never requires trade.7 holds. θB )) θS for all θ ∈ Θ. If θS ≥ θB a mechanism where trade always ¯ takes place. The “pivot mechanism” is the mechanism that is given by the ﬁrst best trading rule and by the following transfer schemes: ¯ ¯ ¯ tS (θ) = q ∗ (θS . In this mechanism the seller receives a constant plus the buyer’s valuation of the good if the seller’s report was pivotal for the trade happening. As in the previous section. In all non-trivial cases. If θS ≥ θB .7. The “if-part” of Proposition 3. these are trivial cases. Then we argue that this mechanism maximizes expected surplus of buyer’s payment over seller’s receipts among all incentive compatible and individually rational mechanisms that implement the ﬁrst best trading rule. The buyer pays a constant plus the seller’s valuation of the good if the buyer’s report was pivotal for the trade happening. θB )θB + (q ∗ (θ) − q ∗ (θS . Thus. To prove the “only if-part” we proceed in a similar way as in the proof of Proposition 3. An incentive compatible. The assertion follows. we show that the mechanism has an expected deﬁcit if the condition of Proposition 3. and the buyer always pays the seller some price p ∈ [θB . θB ) θB tB (θ) = q ∗ (θS .10. θS ] is ﬁrst best and has the required properties.

Individual rationality follows.CHAPTER 3. The ﬁrst best decision rule maximizes both. This becomes clear in the proof of the following result.9 that the expected utility of the highest seller We use the expression “social surplus” to distinguish this expression from social welfare as deﬁned in equation (3. We ﬁx we the buyer’s type θB . θB )(θB − θS ). θB ). The seller’s utility if she is of type θS but reports that she is of type θS . BAYESIAN MECHANISM DESIGN 69 the pivot mechanism has in common with a general Vickrey Clarke Groves mechanism that each agent’s private interests are exactly aligned with social welfare.59). A similar argument applies to the buyer.61) Note that the ﬁrst best decision rule q ∗ maximizes the social surplus q ∗ (θ)(θB − θS ). social welfare and social surplus. the lowest type of the seller. The pivot mechanism is incentive compatible and individually rational. We have seen in the proof of Lemma 3. (3. Because the ﬁrst best rule maximizes social surplus. the individual rationality constraints of the two extreme types.10. 11 .9. Our discussion of interim expected transfers for incentive compatible mechanisms shows that these depend only on the trading rule and on the payments of the highest buyer type and the lowest seller type. Suppose the seller is of type θS and contemplates reporting that she is of type θS = θS . Moreover. Lemma 3. From the payment rule. and the highest type of the buyer. and the type θ receives expected utility utility θ S zero whatever the buyer’s type. is: q ∗ (θS . Lemma 3. are satisﬁed with equality. it is clear that the type θB receives expected ¯B whatever the seller’s type. Proof. We are going to show that truthful reporting is optimal whatever the buyer’s type. ¯ To check individual rationality it is suﬃcient to consider the types θB ¯ and θS . it is optimal for the seller to report θS truthfully. leaving out all terms that do not depend on her report.11 The seller’s utility is thus equal to social surplus if the collective decision is q ∗ (θS . This obviously implies that truthful report is optimal for the seller. Proof. The diﬀerence between the buyer’s and the seller’s expected payments is at least as large under the pivot mechanism as it is under any incentive compatible and individually rational direct mechanism that implements the ﬁrst best trading rule.

there is an expected deﬁcit. θS ] have an intersection with non-empty interior. deﬁcit is: θ − θ B ¯ Case 2: q ∗ (θ) = 1. at any such point we are in Case 4. Similarly. ¯ Case 1: q ∗ (θ) = 1. θB ) = 1. q ∗ (θS . q ∗ (θS . q ∗ (θS .CHAPTER 3. θB ) = 0.e. This is obviously not positive. This is negative by the assumption of the Lemma. Consider any point in this intersection for which trade is eﬃcient and such that θS > θB . Consider ﬁrst values of θ such that q ∗ (θ) = 0. Then the ex ¯S . the pivot mechanism maximizes the ex ante expected diﬀerence between buyer’s and seller’s payments among all mechanisms that implement the ﬁrst best trading rule and that are incentive compatible.12 is violated. This will then imply the case. q ∗ (θS . θB ) = 0.11. the expected utility of the lowest buyer type in the pivot mechanism is zero. ¯ The condition of the Lemma means that the intervals [θ . and moreover the deﬁcit is strictly negative. By drawing a simple diagram one can see that the probability measure of such points is positive. and hence the deﬁcit is zero. θB ) = 0. We shall show that for every realized type vector θ the pivot mechanism either has a surplus of zero. i. Then the ex deﬁcit is: θB − θB . Thus. θB ) = 1. q ∗ (θS . q ∗ (θS . Now consider values of θ such that q ∗ (θ) = 1. Then both agents’ transfers equal zero. Then the ex post ¯S . or a deﬁcit. θB ) = 0. deﬁcit is: θS − θ ¯ Case 3: q ∗ (θ) = 1. Proof. θB ) = 1. Moreover. q ∗ (θS . Then the ex deﬁcit is: θS − θB . . Moreover. We calculate the ex post deﬁcit for four diﬀerent scenarios. BAYESIAN MECHANISM DESIGN 70 ¯ type in the pivot mechanism is θS . θB ] and [θ B post post post such ¯ S . Lemma 3.12 holds. θB ) = 1. Therefore. This is obviously not positive. ¯ Case 4: q ∗ (θ) = 1. The diﬀerence between the buyer’s ex ante expected payment and the seller’s ex ante expected payment is negative in the pivot mechanism ¯ ¯ if the condition of Proposition 3. and therefore the interim expected payments that the seller makes in the pivot mechanism are the highest payments compatible with the ﬁrst best trading rule. we shall argue that the types for which it has a deﬁcit have positive probability. We show next that the pivot mechanism has an ex ante expected deﬁcit if the condition of Proposition 3. if θB < θS and θB > θS . This implies that the interim expected payments to the seller in the pivot mechanism are the lowest payments compatible with the ﬁrst best trading rule. This is not positive because we are considering θ that trade is eﬃcient. q ∗ (θS .

66) θS + fS (θS ) Θ . we shall assume that the mechanism designer seeks to maximize: q(θ) (θB − θS ) f (θ)dθ. note that θS is not aﬀected by the mechanism designer’s decisions. Moreover. 71 We shall now focus on the case in which according to Proposition 3.” As in the public goods case it is without loss of generality to assume that the mechanism designer balances the ex post budget exactly rather than leaving a surplus.63) (3. We shall refer to these mechanisms as “second best.12 no ﬁrst best mechanism is incentive compatible. 1 − FB (θB ) f (θ)dθ = fB (θB ) Θ FS (θS ) ¯ US (θS ) − (1 − q(θ)) θS + f (θ)dθ fS (θS ) Θ (budget balance). The constraints are: QS is increasing and QB is decreasing (incentive compatibility). and ex post budget balanced. BAYESIAN MECHANISM DESIGN This completes the proof of Proposition 3. and of the lowest type of the buyer.12.65) We simplify this optimization problem a little. First. Θ (3. we write the budget constraint as: 1 − FB (θB ) FS (θS ) − θS + f (θ)dθ = q(θ) θB − fB (θB ) fS (θS ) Θ FS (θS ) ¯ US (θS ) + UB (θB ) − f (θ)dθ (3.62) As in the previous subsection it is convenient to think of the designer’s choice variables as the trading rule q and the interim expected utilities of ¯ the highest type of the seller. Our objective is to determine direct mechanisms that maximize expected welfare among all incentive compatible. individually rational. US (θS ). − UB (θB ) + q(θ) θB − (3.64) (3. UB (θB ). Therefore. individually rational. The mechanism designer’s welfare function can then be written as θS + q(θ)(θB − θS ).CHAPTER 3. ¯ ¯ US (θS ) ≥ θS and UB (θB ) ≥ 0 (individual rationality). and ex post budget balanced mechanisms.

63) of the mechanism designer’s maximization problem.e.67) by K. We have a concave maximization problem because the objective function and the constraint are both linear in q. 217 and p.3. we must have: λ > 0. then the mechanism designer can allocate the diﬀerence as additional utility to the buyer and the seller. BAYESIAN MECHANISM DESIGN 72 ¯ Next. Therefore.12. We now make a regularity assumption that ensures that this decision rule satisﬁes the monotonicity constraint (3.69) 0 otherwise. Below.68) and moreover the budget constraint (3.69) would be identical to the ﬁrst best trading rule.” i. by Theorems 1 and 2 in Luenberger (1969).67) holds as a strict inequality only if λ = 0. When solving the mechanism designer’s optimization problem we ﬁrst neglect the incentive constraint. we shall denote the right hand side of (3.67) If this inequality is strict. The seller’s distribution function FS is “regular. Can we have λ = 0? Then the optimal decision rule (3. we drop the choice variables US (θS ) and UB (θB ) and re-write the budget constraint as: q(θ) Θ ¯ θS − Θ 1 − FB (θB ) FS (θS ) − θS + fB (θB ) fS (θS ) FS (θS ) θS + f (θ)dθ fS (θS ) θB − f (θ)dθ ≥ (3. Therefore. Assumption 3.CHAPTER 3. Brief manipulation of this condition leads to the following trading rule: 1 if θB − λ 1−FB (θB ) ≥ θS + λ FS (θS ) 1+λ fB (θB ) 1+λ fS (θS ) q(θ) = (3. 221. p. and the budget constraint would be violated by Proposition 3. ψS (θS ) ≡ θS + FS (θS ) fS (θS ) . q(θ) (1 + λ)θB − λ The Lagrange function is maximized by choosing q(θ) to be positive whenever the expression in the square brackets is positive. a necessary and suﬃcient condition for q to be optimal is that there is a Lagrange multiplier λ ≥ 0 such that q maximizes: 1 − FB (θB ) FS (θS ) − (1 + λ)θS − λ f (θ)dθ − λK fB (θB ) fS (θS ) Θ (3.

CHAPTER 3. minus some discount.72) (iii) the payment rules create incentive compatibility.70) are increasing for every λ > 0. (ii) exact budget balance holds: q(θ) Θ λ FS (θS ) 1+λ fS (θS ) (3. for trade to take place. θB ] and all θS ∈ [θS . Suppose θB < θS and θB > θS . for all θB ∈ ¯ ¯ [θB .71) ¯ θS − Θ 1 − FB (θB ) FS (θS ) − θS + fB (θB ) fS (θS ) FS (θS ) f (θ)dθ θS + fS (θS ) θB − f (θ)dθ = (3. and ex ante budget balanced.” i. Then a direct mechanism (q. tB ) is incentive compatible. .e. and maximizes expected welfare among all such mechanisms if and only if: (i) there is some λ > 0 such that for all θ ∈ Θ: 1 if θB − λ 1−FB (θB ) ≥ θS + 1+λ fB (θB ) q(θ) = 0 otherwise. If FS and FB are regular. 1 − FB (θB ) ψB (θB ) ≡ θB − fB (θB ) is monotonically increasing. We can then conclude: ¯ ¯ Proposition 3. θS + λ FS (θS ) λ 1 − FB (θB ) and θB − 1 + λ fS (θS ) 1 + λ fB (θB ) (3. θS ]: θS TB (θB ) = θS QS (θS ) − θS QS (x)dx ¯ θS ¯ TS (θS ) = θS − (1 − QS (θS ))θS − (1 − QS (x)) dx θS Note that in the second best mechanism the good changes hands less frequently than in the ﬁrst best.e.13. i. and that for every agent i ∈ I the function Fi is regular. individually rational. still has to be larger than the seller’s value. The buyer’s distribution function FB is “regular. BAYESIAN MECHANISM DESIGN 73 is monotonically increasing. The buyer’s value. tS . plus some discount. In this case the trading rule q derived above will imply that the function QS is decreasing and the function QB is increasing.

14. ¯ ¯ (ii) for all θB ∈ [θB . It is obvious that proﬁt maximization implies that the expected transfer of the highest seller type and the lowest buyer type are chosen so as to make these types’ individual rationality constraints true as equalities. (3.73) ¯ − θS − Θ (1 − q(θ)) θS + Taking out constants that are independent of q.4 Proﬁt Maximization We now consider brieﬂy the problem of choosing the mechanism that maximizes the expected proﬁts.58) we can thus write expected revenue as: q(θ) θB − Θ 1 − FB (θB ) fB (θB ) f (θ)dθ FS (θS ) fS (θS ) f (θ)dθ .57) and (3. θB ] and all θS ∈ [θS . we obtain: Proposition 3. θS ]: θS TB (θB ) = θS QS (θS ) − θS QS (x)dx ¯ θS ¯ TS (θS ) = θS − (1 − QS (θS ))θS − (1 − QS (x)) dx θS . We could think of this mechanism designer as the commercial designer of a trading platform who charges a fee for transactions on the platform. BAYESIAN MECHANISM DESIGN 74 3. otherwise. The mechanism designer’s choice variables are again the expected transfers of the highest seller type and the lowest buyer type and the trading rule. and individually rational direct mechanism to maximize expected proﬁts are: (i) q(θ) = 1 0 if θB − 1−FB (θB ) fB (θB ) > θS + FS (θS ) fS (θS ) .4. Then necessary and suﬃcient conditions for an incentive compatible. Suppose for every agent i ∈ I the distribution function Fi is regular. Using the formulas (3. we are left with: q(θ) θB − Θ 1 − FB (θB ) FS (θS ) − θS − fB (θB ) fS (θS ) f (θ)dθ. (3.74) Using similar reasoning as in the last subsection.CHAPTER 3. We assume that the mechanism designer’s proﬁt is the diﬀerence between what the buyer pays and what the seller receives.

77) Thus. We shall now calculate for which such thresholds s the budget balance constraint holds. 1].4. 3 2 6 The seller’s expected transfer is given by: ¯ θS − Θ (1 − q(θ)) θS + FS (θS ) fS (θS ) f (θ)dθ =1−2 Θ (1 − q(θ))θS dθ . We want to determine the welfare and the proﬁt maximizing mechanisms.76) (3. 1 + 2λ (3.13 the welfare maximizing rule will be such that trade takes place if and only if: θB − λ 1 − FB (θB ) λ FS (θS ) > θS + ⇔ 1 + λ fB (θB ) 1 + λ fS (θS ) λ λ (1 − θB ) > θS + θS ⇔ θB − 1+λ 1+λ λ θB − θS > .75) (3. The buyer’s willingness to pay is given by: q(θ) θB − Θ 1 − FB (θB ) fB (θB ) f (θ)d(θ) = = s q(θ)(2θ − 1)dθ Θ 1 0 θB −s 2θB − 1dθB (3.13 and 3. Suppose that θS as well as θB are uniformly distributed on the interval [0.78) 1 1 1 = s3 − s2 + . 3. Example 3.4. trade will take place if and only if the diﬀerence between the buyer’s and the seller’s valuation is above some positive threshold s.5 A Numerical Example We conclude with a very simple numerical example. BAYESIAN MECHANISM DESIGN 75 Comparing Propositions 3.14 we ﬁnd that the proﬁt maximizing mechanism designer facilitates less trade than the welfare maximizing mechanism designer.CHAPTER 3. Note that this satisﬁes the regularity condition. By Proposition 3. Take s as given.

15. 4 λ 1+2λ (3.80) 1 and s = 1. In the welfare maximizing incentive compatible.16.14 the following characterization of the expected proﬁt maximizing trading mechanism is obvious. in particular: Chapter 3.5 Comments on the Literature The classic paper on expected revenue maximizing single unit auctions is Myerson (1981). We is of the required form Proposition 3.79) 1 = (1 − s)3 3 Budget balance is achieved when: 1 3 1 2 1 1 s − s + = (1 − s)3 .81) with λ > 0. Proposition 3. but it focuses only on Bayesian incentive compatibility and ex-post . 3. and Milgrom (2004). 3 2 6 3 This has two solutions: s= Only the solution s = conclude: 1 4 (3. BAYESIAN MECHANISM DESIGN 1−s θS +s 1 1 76 =1−2 0 0 θs dθB dθS + 1−s 0 θs dθB dθS (3.83) We can see that in both mechanisms trade takes place less frequently than in the ﬁrst best. 2 (3. (3. In the expected proﬁt maximizing incentive compatible and individually rational trading mechanism trade takes place if and only if 1 θB − θS > . but an expected proﬁt maximizing mechanism designer arranges for less trade than an expected welfare maximizing mechanism designer. individually rational and ex post budget balanced trading mechanism trade takes place if and only if 1 θB − θS > . For the public goods problem a classic reference is d’Aspremont and GerardVaret (1979). in particular: Chapter 5. In preparing these notes I have also used Krishna (2002).82) 4 From Proposition 3.CHAPTER 3. This paper works with a more general set-up than my exposition.

.15 display second best mechanisms for numerical examples that illustrate the public goods problem and the bilateral trade problem. and that have moreover three properties: (1) incentive compatibility. 3. (3) ex post budget balance.6 Problems a) Propositions 3. and ex post budget balance of the second best mechanism. Determine for each case payment rules that guarantee incentive compatibility. in each we have not speciﬁed the payment rules ti for i ∈ I.10 and 3. Welfare maximization and proﬁt maximization under individual rationality constraints are considered in G¨th and Hellwig (1986).7 and 3. individual rationality. b) Propositions 3. or expected welfare maximizing. The explanation of the relation between ex ante u and ex post budget balance is based on B¨rgers and Norman (2008).CHAPTER 3. c) For each of the numerical examples of this Chapter investigate whether there are intuitive and simple indirect implementations of the expected proﬁt maximizing. (2) individual rationality. BAYESIAN MECHANISM DESIGN 77 budget balance. However. mechanisms. neglecting individual rationality. Prove that in both applications there are direct mechanisms that implement q ∗ and that have any two of these three properties. My exposition of the bilateral trade problem has also beneﬁtted from the books by Krishna and Milgrom cited above.12 show that there are no direct mechanisms that implement the ﬁrst best decision rule (in the example of the public good problem) or trade rule (in the example of the bilateral trade problem) q ∗ . For o the bilateral trade problem the classic paper is Myerson and Satterthwaite (1983).

Imperfect information about others’ beliefs seems at least as pervasive as imperfect information about others’ preferences. then the mechanism designer’s expectations about how the mechanism will be played might not be fulﬁlled. Among the points that we mentioned was that the assumption of a common prior and independent types means that each agent’s beliefs about the other agents’ types is independent of the agent’s type.1 Introduction The models discussed in Chapter 3 were all independent private value models. In the independent private values model agents and the mechanism designer don’t know other agents’ values. In Section 3. but we make much weaker assumptions about the agents’ and the mechanism de78 . The mechanisms that we constructed in the previous chapter work well if the mechanism designer is right in his assumptions about the agents. and moreover that these beliefs are commonly known among the mechanism designer and the agents. However.1 we have discussed the limitations of such models.Chapter 4 Dominant Strategy Mechanisms: Examples 4. In this chapter we revisit the examples of the previous section. where the focus is on asymmetric information.2. if the mechanism designer has made incorrect assumptions about the agents’ beliefs. This assumption seems peculiar in the context of mechanism design. and if agents indeed play a Bayesian equilibrium of the game that the mechanism designer creates for them. but they do know those agent’s beliefs. or if these beliefs are not common knowledge.

We will require the strategy that the principal recommends to an agent to dominate the strategy of not participating. We shall comment on terminology at the end of this section. We shall apply a similar logic to the “outside option. it is assumed that the mechanism designer is not willing to risk at all making incorrect assumptions about agents’ beliefs. . skipping whatever might be in the middle. no agent has an incentive to opt out. in this section the mechanism designer is assumed to be uncertain about agents’ beliefs. We assume that the mechanism designer does not want to rely on any assumption regarding the agents’ beliefs about each other.” We shall assume that whenever the mechanism designer proposes a mechanism it is assumed that in addition to the choices that he oﬀers to the agents in the mechanism the agents can also choose to opt out of the mechanism and receive an outside option utility.CHAPTER 4.” This requirement restricts the set of mechanisms and strategy combinations that the mechanism designer can choose from. In short: each type is required to have a dominant strategy. We shall translate this into formal theory by requiring that the strategy combination that the mechanism designer proposes to the agents when presenting the mechanism to them prescribes choices that are optimal for each type. We shall call this constraint “ex post individual rationality constraint” because it requires that ex post. This is because. obviously. DOMINANT STRATEGY MECHANISMS 79 signer’s beliefs. moreover. independent of what the other types do. Recall that in the previous chapter we had interpreted the individual rationality constraint as an “interim individual rationality constraint. and. a vector of strategies that prescribes a dominant strategy to each type always constitutes a Bayesian equilibrium of a game of incomplete information. we shall move from one extreme along a continuum of possible models to the other extreme. While in the previous section the mechanism designer was willing to trust a single hypothesis about agents’ beliefs. In a sense.12 We shall refer to this requirement as “dominant strategy incentive compatibility. The available mechanisms and strategy combinations that we consider in this section are a strict subset of those considered in the previous section. after the mechanism has been played. but not vice versa. or the utility of privately consuming the good in the seller’s case in the bilateral trade model of the previous chapter.” In that chapter agents evaluated the outside option once they knew their type using a belief 12 This use of the phrase “dominant” is slightly sloppy. The mechanism designer in this chapter wishes to implement a mechanism of which he is sure that it produces the desired results independent of what agents think about each other. such as the utility zero in the auction and public goods models of the previous chapter.

s−i ) ≥ ui (si . In game theoretic terms a better expression than “dominant” would be that strategy si is in “always optimal. We shall use the expression “dominance” somewhat loosely in these notes. obviously. 13 We conclude the introduction with one formal clariﬁcation. we shall be sloppy. DOMINANT STRATEGY MECHANISMS 80 about the other agents’ types that was known to the mechanism designer. and deﬁne “dominant” to mean “always optimal”. s−i ) ∀s−i ∈ S−i ). We thus consider in this chapter a mechanism designer whose choice set is smaller than the choice set available to the mechanism designer in the previous part. and “weak dominance” (where ui (si . s−i ) ∀s−i ∈ S−i with ui (si . s−i ) ∃s−i ∈ S−i ). whenever participation dominates nonparticipation at the ex post level it will also be optimal at the interim level. We shall investigate these issues for each of the three models introduced in Chapter 3.CHAPTER 4. and budget balance requirements. In many of the examples that we investigate. and we shall say that a strategy is “dominant” if it dominates all other strategies. our conclusions would remain true if we had deﬁned “dominant” to mean “weakly dominant” in the traditional gameThe decision theoretic foundations of the dominant strategy approach to mechanism design are investigated further in Chung and Ely (2007). This is because. s−i ) ≥ ui (si . s−i ) > ui (si .” This is because the term dominance is set apart in game theory for two other concepts: “strict dominance” (where ui (si . Ex post individual rationality restricts the set of mechanisms that the mechanism designer can recommend further in comparison to the previous parts of these notes. We shall say that the strategy si of player i “dominates” strategy si if ui (si . We shall therefore not study optimal mechanisms in this section. it is not obvious how we should deﬁne revenue maximizing or welfare maximizing mechanisms. We shall also investigate whether the mechanisms that turned out to be optimal in Chapter 3 remain in the choice set. As we shall not attribute explicit beliefs to the mechanism designer in this Chapter. 13 . Developing a coherent decision theoretic account of the mechanism designer’s view of the agents’ payoﬀ types and their beliefs about each other is a project that goes beyond what we do in these notes. s−i ) > ui (si . but not vice versa. s−i ) ∀s−i ∈ S−i . For the purposes of this note. but focus on the characterization of classes of mechanisms that satisfy incentive compatibility. The mechanism designer had to oﬀer a mechanism that ensured that agents when conducting this evaluation didn’t have an incentive to opt out. individual rationality. Our interest will be in characterizing this choice set.

. for the game tree.1 Single Unit Auctions Set-Up We return to the single unit auction setting of Section 3. We denote by θ the vector ¯ (θ1 . as in the previous Chapter. . Let us say that in this case the strategy prescribed for each type of each buyer is a “dominant” strategy. Direct Mechanisms. θ is an element of Θ ≡ [θ. .2. θ]N −1 . . N }. ¯ ¯ Buyer i’s type is θi ∈ [θ. if she follows the seller’s recommendation. to each terminal history of the game. We brieﬂy recapitulate the aspects of the model that are relevant here.2.2 4. i=1 4. DOMINANT STRATEGY MECHANISMS 81 theoretic sense.2 Mechanisms.. . .2. . and for each strategy of the . We denote by θ−i the ¯ vector θ when we leave out θi . The seller’s utility if he obtains transfers ti from the N agents is N ti .CHAPTER 4. . There are N potential buyers: I = {1. so that θ−i ∈ Θ−i . We assume that buyer i’s utility if she is of type θi equals θi − ti if she obtains the good and pays a transfer ti ∈ R to the seller. We now require that each type of each buyer ﬁnds it optimal to choose the proposed strategy. θ2 . θN ). possibly randomized. 2. 2. We denote by Θ−i the set [θ. and the Revelation Principle Recall that a general mechanism is a game tree together with an assignment of a probability distribution over outcomes. for all possible strategy combinations that the other buyers might play. 1. however. Each buyer i knows her own type θi but not θ−i . A seller seeks to sell a single indivisible good. and that it is −ti if she does not obtain the good and pays a transfer ti to the seller. . that is a probability distribution over {∅. We shall also require that each type of each buyer ﬁnds that her expected utility. N } × RN .” 4. θ] where 0 ≤ θ < θ.. We shall imagine. Our conclusions would. θ]N . that the seller proposes a mechanism together with one strategy for each buyer. not remain true if we had deﬁned “dominant” as the traditional game-theoretic concept of “strict dominance. The incentive compatibility requirement that the seller’s proposed strategy combination has to satisfy in this Chapter is more restrictive than it was in the previous Chapter.. is nonnegative. where a strategy assigns to every type of the buyer a complete behavior plan. for each possible realization of the other buyers’ types.

1 (Revelation Principle for Dominant Strategy Mechanisms).1 shows that for dominant strategy implementation. and we will be able to restrict attention to direct mechanisms. that is. To see this suppose it were not. As before. Proposition 4. and we would need it only for a very short while. The deﬁnition of direct mechanisms is as before. then the same type θi would have preferred to deviate from σi . Proof.1. We shall say that in this case participation is for each type of each buyer a “dominant” strategy. (ii) For every vector θ of types the distribution over allocations and the expected payments that result in Γ if the agents play σ is the same as the distribution over allocations and the expected payments that result in Γ if the agents play σ . σ prescribes telling the truth. We can prove the result by showing that truth telling will be a dominant strategy in this direct mechanism. we do not formalize the general deﬁnition of a mechanism. and: (i) The strategy vector σ satisﬁes for every i and every θi : σi (θi ) = θi . Suppose a mechanism Γ and a strategy combination σ for Γ are such that for each type θi of each buyer i the strategy σi (θi ) is a dominant strategy in Γ. The following result is the revelation principle for our setting. We wouldn’t have to use this formalization much further in these notes because the “revelation principle” extends to our current setting. If type θi prefers to report that her type is θi for some type vector of the other agents θ−i . see Deﬁnition 3. Hence σi would not be a dominant strategy in Γ. Proposition 4. for the strategy combination that the types θ−i play in Γ . Then there exists a direct mechanism Γ and a strategy combination σ of Γ such that for every type θi of each buyer i the strategy σi (θi ) is a dominant strategy in Γ .CHAPTER 4. it is without loss of generality . as for Bayesian equilibrium implementation. DOMINANT STRATEGY MECHANISMS 82 other buyers. The reason is that such a formalization would take long to write down. Construct Γ as required by part (ii) of the proposition. nor do we formalize the notions of incentive compatibility and of individual rationality that we have just described. and to play the strategy that σi prescribes for θi in Γ.

Whereas our previous revenue . The characterization of incentive compatible is actually exactly as in Propositions 2. . A direct mechanism (q. t1 . θ−i ) = ti (θ. all θi ∈ [θ.2. tN ) is “ex post individually rational” if for each type of each buyer participation is a dominant ¯ strategy. θ−i ) ≥ θi qi (θi . . . . . Deﬁnition 4. then the same is true for the corresponding direct mechanism. t1 . θ−i ) − ti (θi . Proposition 4. tN ) is “dominant strategy incentive compatible” if truth telling is a dominant strategy for each type ¯ of each buyer. t1 . all θi . . θi ∈ [θ. θ−i ). .CHAPTER 4. θ] and for all θ−i ∈ Θ−i : θi qi (θi . that is. . DOMINANT STRATEGY MECHANISMS 83 to restrict attention to direct mechanisms in which truth-telling is a dominant strategy. t2 . Moreover. t2 .2.2.2 implies an ex post revenue equivalence result. θ−i )) − θ ˆ ˆ qi (θ. . applied to each agent i. Deﬁnition 4. Hence. θ−i ) is increasing in θi . if the indirect mechanism that the mechanism designer contemplates is individually rational. A direct mechanism (q. . A direct mechanism (q. θ−i ) − θqi (θ.2 except that the result now applies for every realization θ−i of types of the other agents. θ−i )dθ We omit the proof of this result as it is the same as the proof of Propositions 2.2. 4. if for all i ∈ I.2 and 3. θ] and all θ−i ∈ Θ−i : θi qi (θi . . θ−i ) − ti (θi . θ−i ) + (θi qi (θi . that is.3 Characterizing Dominant Strategy Incentive Compatibility and Ex Post Individual Rationality In this subsection we develop a better understanding of the structure of the set of all direct mechanisms that satisfy the two conditions introduced in Deﬁnitions 4. θ]: θ ti (θi .2. and to each possible vector θ−i of the other agents. θ−i ) ≥ 0. t2 .1 and 4. if for all i ∈ I. ¯ (ii) for every θi ∈ [θ. θ−i ) − ti (θi . we can restrict attention to direct mechanisms that are dominant strategy incentive compatible and ex post individually rational as deﬁned below.1. tN ) is dominant strategy incentive compatible if and only if for every i ∈ I and every θ−i ∈ Θ−i : (i) qi (θi . An interesting and important point is that Proposition 4. .

θ−i ). .5 as expected revenue maximizing and as expected welfare maximizing. Our interpretation of these functions was that they assigned to each type a “virtual type.”14 We don’t claim that there are no other dominant strategy incentive compatible and ex post individually rational direct mechanisms than canonical auctions. We omit the proof.4 the deﬁnition of the functions ψi : [θ. .2. t1 . DOMINANT STRATEGY MECHANISMS 84 equivalence results said that the allocation rule and the interim expected payments of the lowest types imply the interim expected payments for all types.1) The regularity assumption under which we derived the expected revenue maximizing auction was that all functions ψi were increasing. We call the direct mechanisms in this class “canonical auctions.2. t2 . We show that for any such functions an allocation rule that is constructed as in the expected revenue maximizing auction in Section 3.” Under the regularity assumption.4 can be supplemented with transfer rules that make the mechanism dominant strategy incentive compatible and ex post 14 This expression is not commonly used.2. tN ) is ex post individually rational if and only if for every i ∈ I and every θ−i ∈ Θ−i we have: ti (θi .2. Instead of working with these speciﬁc function ψi . . A dominant strategy incentive compatible direct mechanism (q.4 Canonical Auctions We now display a class of dominant strategy incentive compatible and ex post individually rational direct mechanisms.CHAPTER 4. . I made it up for these notes. θ]. Proposition 4. but we show that this class is rich enough to include the mechanism that we identiﬁed in Sections 3. fi (θ−i ) (4. Recall from the discussion of the expected revenue maximizing auction ¯ in Section 3. we now consider arbitrary strictly increasing functions ψi . we also obtain a simple characterization of ex post individual rationality for direct mechanisms that are dominant strategy incentive compatible. the expected revenue maximizing auction assigned the object to the buyer with the highest virtual type. part (ii) of Proposition 4. θ−i ) ≤ θi qi (θi . θ] → R for i ∈ I: ψi (θi ) ≡ θi − 1 − Fi (θi ) ¯ for all θi ∈ [θ.2.3. In analogy to our earlier results. 4.4 and 3.2 says that the allocation rule and the ex post payments of the lowest types imply the ex post payments of all types. .

buyer i can win the auction.CHAPTER 4. Proposition 4. . it will have to be the case that θi > θi . then her utility doesn’t change. as her payment will be the lowest type of buyer i that wins against θ−i . t2 . The assumed continuity of ψi guarantees that this minimum exists. Deﬁnition 4. Moreover. . If qi (θi . θ−i ) = 0 for all θ−i ∈ Θ−i . then bidder i does not have to pay anything. If bidder i does not win the auction. tN ) is called a “canonical ¯ auction” if there are strictly increasing and continuous functions ψi : [θ. for all θ ∈ Θ. qi (θ) = 0 if ψi (θi ) < 0 or ψi (θi ) < ψj (θj ) for some j ∈ I with j = i. DOMINANT STRATEGY MECHANISMS 85 individually rational. she has no incentive to change her strategy. and by assumption θi is not large enough. . for every i ∈ I: ui (θ. we ignore ties.3. Thus. θ] | ψi (θi ) ≥ 0 and ψi (θi ) ≥ ψj (θj ) for all j ∈ I with j = i} if qi (θ) > 0 ti (θ) = 0 if qi (θ) = 0. The advantage of putting the result in this generality is that we can use it to also show the implementability of allocation rules other than the expected revenue maximizing. θ−i ) = 0. Moreover. . If bidder i does win the auction. buyer i’s payment will be 1 larger than n θi . that all other buyers are of type θ−i . We ﬁrst show dominant strategy incentive compatibility. θ−i ) = 0.4.15 then bidder i’s payment equals the lowest type that she might have had that would have allowed her to win the auction. To simplify the exposition we assume that the functions ψi are continuous. . 15 For simplicity. θ] → R for i ∈ I such that for all θ ∈ Θ and i ∈ I: 1 n if ψi (θi ) ≥ 0 and ψi (θi ) ≥ ψj (θj ) for all j ∈ I with j = i. . Thus. Every canonical auction (q. where n is the number of agents j ∈ I such that ψj (θj ) = ψi (θi ). and: 1 ¯ n min{θi ∈ [θ. and that qi (θi . A direct mechanism (q. . It is worth considering the transfer rule in detail. Does buyer i have an incentive to report a diﬀerent type θi ? 1 If qi (θi . Suppose that buyer i is of type θi . t1 . t1 . . θ−i ) = n > 0. t2 . . Proof. but only by paying more than the object is worth to her. tN ) is dominant strategy incentive compatible and ex post individually rational.

If buyer i changes her report to another type θi > θi so that qi (θi .CHAPTER 4. 1 Consider ﬁnally the case that qi (θi .2. Hence. This is because if she reports θi truthfully.4 as the expected revenue maximizing auction if the mechanism designer assumes that his belief about Θ is also the prior distribution from which the buyers’ beliefs about the other buyers’ valuations are derived. θ−i ) = 0. and has utility zero. If she changes her report to a type θi for which qi (θi . as her payment does not depend on her report. either loses the auction. θ−i ) = 1. Firstly. by the deﬁnition of the transfer rule.2. Thus.2. then her utility doesn’t change. The lowest type. Then buyer 1 i’s transfer payment is n θi . θ−i ) = 1. her expected utility is again zero. or wins the auction and has to pay θ.4. and have assumed that the seller is highly uncertain about the buyers’ beliefs about each other. We now provide two applications of Proposition 4. her utility decreases. Hence a truthful report gives her a positive surplus. Suppose moreover that these beliefs satisﬁed the assumptions of Section 3. he does not lose any expected revenue.4 there are some that make truth telling a dominant strategy. a truthful report is optimal. If she changes her report to a type θi for which qi (θi . it is one of the optimal auctions of Section 3. θ−i ) = 0. suppose that the seller had subjective beliefs F on Θ that reﬂected his view regarding the likelihood of diﬀerent type vectors. θ. DOMINANT STRATEGY MECHANISMS 86 Consider next the case that qi (θi .4 shows that even if the seller’s uncertainty is so large that he wishes to use a dominant strategy mechanism. If buyer i changes her report to another type θi for which qi (θi . in which case utility is also zero. The proof of dominant strategy incentive compatibility also shows that a buyer’s utility is always non-negative if she wins the auction. we have shown that among the optimal auctions that we found in Section 3. and that gives the lowest types expected utility zero. θ−i ) = 1. and her expected utility will be zero. This concludes the proof of dominant strategy incentive compatibility. In other words. Proposition 4. . Then one can interpret the optimal auction identiﬁed in Section 3. will be below θi .1 as well as the regularity Assumption 3. the mechanism also satisﬁes ex post individual rationality. her utility is zero. In this part we have relaxed this assumption.2. her transfer payment. then her expected utility doesn’t change. This proves the last sentence of the Proposition. Therefore. as her payment will still be θi . This is a very strong assumption about the seller’s view of his environment. He can ﬁnd a transfer rule that implements the optimal allocation rule. If she loses the auction. θ−i ) = n where n ≥ 2.1.4.

4 the restriction to dominant strategy incentive compatible mechanisms may severely limit what the mechanism designer can achieve. We begin by brieﬂy recapitulating the public goods model.2 Direct Mechanisms By the revelation principle we can restrict attention to direct mechanisms. that is to mechanism where conditional on the agents’ true types the decision about the public good and also the agents’ transfers are not stochastic. . . then g = 1.3 and 3. ¯ ¯ θi is agent i’s type. Each agent i observes her own type θi . We shall also. restrict attention to deterministic mechanisms. DOMINANT STRATEGY MECHANISMS 87 A second application arises from Proposition 4. but not necessarily the other agents’ types θ−i . Of course. and it is an element of θi ∈ [θ. We consider a community consisting of N agents: I = {1. 1}. non-excludable public good. the mechanism designer can also achieve the objective of expected welfare maximization through a dominant strategy mechanism.3. In this case the auction described in Proposition 4. We denote by Θ−i ¯ the set [θ. N }. 4. . They have to choose whether to produce some indivisible.3. Agent i’s utility if the collective decision is g and if she pays a transfer t to the community is: θi g − t. We denote this decision by g ∈ {0.3 4. We denote by θ−i the vector θ when we leave out θi . .1 Public Goods Set-Up We shall now show that in the two examples with budget constraint that we considered in Sections 3. so that θ−i ∈ Θ−i . θ is an element of Θ ≡ [θ. θ]N . then g = 0. If it is not produced.3. where c > 0. .4 describes many other dominant strategy incentive compatible and ex post individually rational auctions. 2. . The cost of producing the public good are cg. θ2 . In . and it is thus welfare maximizing. as in Section 3. Proposition 4. We introduced this assumption in Section 3. Here. The mechanism designer seeks to maximize expected N N welfare where welfare is deﬁned to be i=1 θi g − i=1 ti .CHAPTER 4. If the public good is produced. θN ). 4. The mechanism designer knows none of the θi ’s. . θ] where 0 ≤ θ < θ.4 is just the Vickrey auction. Thus. θ]N −1 . .4 if we set ψi (θi ) = θi ¯ for every i ∈ I and every θi ∈ [θ.3 because it simpliﬁed the formulation of the budget constraint and because it appeared innocuous. θ]. ¯ We denote by θ the vector (θ1 .

5. ˆ ˆ θi = θi ⇒ q(θi . Observe that ex ante budget balance has in our current context no meaning. The deﬁnitions of dominant strategy incentive compatibility and ex post individual rationality are as in the previous subsection. In the previous chapter the distinction between an equality budget constraint and an inequality budget constraint was not of much further importance.CHAPTER 4. Thus. But our main reason for restricting attention to an equality budget constraint is that this simpliﬁes our arguments below. and a condition under which such mechanisms are in addition ex post individually rational.3 Characterizing Dominant Strategy Incentive Compatibility and Ex Post Individual Rationality In this subsection we neglect the budget constraint. θ−i ) = 0 and ti (θi . and seek a characterization of dominant strategy incentive compatible mechanisms. which is deﬁned as in Deﬁnition 3. We could use for this the same arguments as we used in the previous section. θ−i ) = 1 and ti (θi . We might motivate it by arguing that we consider an environment in which the agents participating in the public goods mechanism have no way of committing to a way of disposing of excess funds if the collected transfers exceed what is required for the public good production. ˆ θi > θi ⇒ q(θi . Moreover. An inequality would require that the sum of the contributions is at least as large as the costs of the public good’s production rather than that it is exactly equal. whereas in the current chapter the equality requirement is an important restriction. θ−i ) = τi .3. θ−i ) = 0 and ti (θi . However. θ−i ) = τi . θ−i ) = 1 and ti (θi . However.4. we continue to work with this assumption. Note that in Deﬁnition 3. Proposition 4. ˆ . as we have done here.5 the budget constraint is an equality rather than an inequality. DOMINANT STRATEGY MECHANISMS 88 the current section it is not clear that the assumption is innocuous. we shall impose ex post budget balance. θ−i ) = τi or q(θi . a simpler proof can be given if one restricts attention to deterministic mechanisms. 4.5. A direct mechanism is dominant strategy incentive compatible if and only for every i ∈ I and for every θ−i ∈ Θ−i there are a type ˆ θi ∈ R and two payments τi and τi ∈ R such that: ˆ ˆ θi < θi ⇒ q(θi . because there is no prior probability measure over Θ. θ−i ) = τi . direct mechanisms are deﬁned as in Deﬁnition 3. for simplicity.

Proof. ˆ ˆ Observe also that we have not ruled out that θi < θ (in which case we would ¯ ˆ ¯ have qi (θi . such as the true type. θ]). This ˆ agent can either report a type θi > θi . θ−i ) = 1 for all θi ∈ [θ. τi and τi are allowed to depend on θ−i . θ−i ) = 0 for all θi ∈ [θ.CHAPTER 4. and ˆ denote agent i’s payment in the latter case by τi . θ−i ) = 0 for some θi and qi (θi . Next we characterize ex post individual rationality. Consider all θi such that qi (θi . nor have we ruled out θi > θ (in which ¯ case we would have qi (θi . θ−i ) = 1 for some other θi . Now suppose that qi (θi . then reporting a type above θi .5 hold. in which case the public good is ˆ ˆ produced and the agent has to pay θi . Beyond that. is obviously optimal. For the outcome. We need not give any proof of the following simple result. the value of the reported type does ˆ ˆ not matter. θ]). Truthful reporting is therefore optimal only if the conditions of Proposition 4. it only matters whether the reported ˆ type is above or below θi . because the agent obtains more than ˆ he has to pay. ˆ 89 ˆ Note that in this Proposition θi . because otherwise buyer i would pretend to be the type for which the transfer payment is lowest. For the converse ﬁx θ−i ∈ Θ−i . in which case the public good is not produced. This proves that the conditions provided are suﬃcient for dominant strategy incentive compatibility. . θ−i ) has to be the same for all these θi . The case that qi (θi . Observe that ti (θi . If we set θ all conditions of Proposition 4. Denote by ˆi > θ and τi = τi + θi ¯ ˆ τi the (constant) payment of agent i. Denote the payments corresponding to the former case by τi . First we show suﬃciency. θ−i ) = 0 for all θi ∈ [θ. By the same argument the transfer payment ti (θi . θ] can be dealt with symmetrically. and the agent does not have to pay anything. or the agent can report a type θi < θi . ¯ Consider ﬁrst the case that qi (θi . DOMINANT STRATEGY MECHANISMS ˆ τi − τi = θi . θ−i ) has to be the same for all θi for which qi (θi . Then ˆ ˆ ˆi will report a type such that the public good is produced. and types θi > θ ˆ types θi < θi will report a type such that the public good is not produced. θ−i ) = 0. θ−i ) = 0. A symmetric argument applies if θi ≤ θi . Deﬁne θi = τi − τi . If in truth θi ≥ θi . θ].5 hold. θ−i ) = 1 for all ¯ θi ∈ [θ. Consider an agent i with type θi .

we obtained mechanisms that were Bayesian equilibrium incentive compatible. A direct mechanism (q. in general. t2 . θ] | ψi (θi ) + ti (θ) = 0 for all θ ∈ Θ. for all θ ∈ Θ and. tN ) is called “canonical” if for every agent i there is a strictly increasing and continuous function ¯ ψi : [θ. Here. . and moreover the rules could be made ex post budget balanced. . 4. . “canonical mechanisms. However. t1 .3.4 Canonical Mechanisms We now introduce a class of mechanisms. DOMINANT STRATEGY MECHANISMS 90 Proposition 4. one for each agent i ∈ I.3. by contrast. We show below that we can combine these decision rules with speciﬁc transfer rules to make them dominant strategy incentive compatible and ex post individually rational. θ−i ). individually rational. The decision rules in the mechanisms deﬁned below is of the same form as the decision rule in the expected welfare or proﬁt maximizing mechanisms in part 1. θ−i ) ≤ θi q(θi . q(θ) = 0 otherwise. A dominant strategy incentive compatible direct mechanism is ex post individually rational if and only for every i ∈ I and for every θ−i ∈ Θ−i : ti (θi . In Section 3. we let ψi be some arbitrary strictly increasing function.CHAPTER 4. The deﬁnition of these mechanisms involved functions ψi .3. .4. To simplify the exposition we assume that the functions ψi are continuous. Deﬁnition 4. The regularity assumption was that these functions were strictly increasing. j=i ψj (θj ) ≥ c} if q(θ) = 1 if q(θ) = 0. θ] → R such that: N 1 if i=1 ψi (θi ) ≥ c. We now state the following simple result. for every i ∈ I: ¯ min{θi ∈ [θ.4 as welfare maximizing or proﬁt maximizing mechanisms under the budget constraint. .6.” that are closely related to the mechanisms that we identiﬁed in Section 3. the mechanisms that we describe below will not balance the budget.4.

8 and 3. In these notes. and ex post budget balanced. Otherwise. θ−i ) = 0 for all θ−i ∈ Θ−i . t1 . . Suppose N = 2. we only provide a partial answer to this question. agents’ contributions do not depend on their own or on the other agent’s valuation. A simple indirect implementation of the mechanism described in Proposition 4. 4. If both agents make a pledge. DOMINANT STRATEGY MECHANISMS 91 Proposition 4.8 is this: Each of the two agents is allocated a share of the cost τi . Moreover.7. for every i ∈ I: ui (θ.3. ex post individually rational. . It is easy to show that the mechanisms displayed in Proposition 4. it is not produced and no agent pays anything. there is no paper in the literature that answers this question at a general level. Note that. Then each agent i is asked whether he or she is willing to contribute τi whereby the contribution has to be made only if the other agent also pledges his or her contribution. and each agent i pays τi . we assume in this characterization that the set of type vectors for which the public good is produced is a closed set. Suppose also that the set {θ | q(θ) = 1} is closed.9 include canonical mechanisms. t2 . For this case. Proposition 4. ex post individually rational. To my understanding.8. and ex post budget balanced if and only if there are payments τ1 . τ2 ∈ R with τ1 + τ2 = c such that: q(θ) = 1 and ti (θ) = τi for all i ∈ {1. . Every canonical mechanism (q. The answer that we give in these notes is for the case that there are only 2 agents. Then a direct mechanism is dominant strategy incentive compatible. Note that the mechanisms obtained in Propositions 3. it is easy to obtain the following characterization.CHAPTER 4. tN ) is dominant strategy incentive compatible and ex post individually rational. then the public good is produced. conditional on producing the public good.5 Ex Post Budget Balance Now we ask which direct mechanisms are dominant strategy incentive compatible. We therefore only show the necessity of these conditions. To simplify the exposition.8 are dominant strategy incentive compatible. The proof of this result is left to the reader. . so that τ1 + τ2 = c. Proof. and satisfy ex post budget balance. 2} if θ1 ≥ τ1 and θ2 ≥ τ2 q(θ) = 0 and ti (θ) = 0 for all i ∈ {1. 2} otherwise. ex post individually rational. .

Note that τj may be negative. ˆ By deﬁnition θi ≥ θ Together with Proposition 4. This is a ˆ ˆ consequence of the deﬁnitions of θ1 and θ2 . If agent 2’s type . θ ˆ ˆ have that q(θ1 . as ˆ shown above. ˆ Now suppose that θi > 0 for i = 1. ˆ ˆ q(θ1 . using Proposition 4.5.8 by setting τi > θ for some agent i. Also note that θ1 < θ1 or θ2 < θ2 implies q(θ) = 0. θ2 ) = θ1 and t2 (θ1 . θ2 ∈ [θ. Then. ex post individually rational. 2}. if agent 2’s type is θ2 . Hence θi ≥ θ From what we have shown in the previous paragraph we can infer that ˆ θi ≥ θi for i = 1. 2}.e. θ2 ) = 1. By Proposition 4. θ2 ) is θ ˆ contributions is strictly less than θ1 + θ2 . θ2 ) = θ1 and ˆ2 ) = θ2 . 2}.5 we ˆ ˆ t2 (θ1 . 2 implies q(θ1 . We seek to ˆ ˆ show that θi ≥ θi implies q(θi . θ2 ) = θ2 . 2. The reason is that ﬁrst.CHAPTER 4. We want to show that ¯ τ1 does not depend on θ2 . θ]. By Proposition 4. Suppose ˆ that θi > θi for some i ∈ {1. 2 ˆ ¯ ¯ deﬁne θi ≡ min{θi ∈ [θ. For i = 1. that τ1 (θ2 ) = τ1 (θ2 ) for all θ2 . i. This is the case when the public good is produced with probability 1. θ2 ) = 1. θ2 ) = 1. ¯ Suppose τ1 (θ2 ) < τ1 (θ2 ) for some θ2 . the sum of contributions is: τ1 (θ2 ) + τ2 (θ1 ).5 t1 (θ1 . and deﬁning τj ≡ c − τi for j = i. 2}. it may depend on the other agents’ reports. ˆ The remaining case is that θi = 0 for both agents i. are well deﬁned. where i = j. θ] | q(θ1 . θ2 ) = 1. To show this we deﬁne ¯ ˆ ¯ ˆ θ1 ≡ min{θ1 ∈ [θ. θ] | q(θ1 . θ]} where j = i. We write agent 1’s payment as τi (θ2 ). θj ) = 1. where i = j. ˆi for i ∈ {1. then the sum of these ˆ ˆ contributions if types are (θ1 . we can infer q(θ1 . Without loss of generality suppose the ˆ strict inequality held for i = 1. θ]. We claim that θi = θi for i ∈ {1. θ2 ∈ [θ. Thus the sum of ˆ1 + θ2 . Therefore. Then. budget balance is violated. Let us now assume that q(θ) = 1 for at least one θ ∈ Θ. and ex post budget balanced only if all payments are zero. Then obviously the mechanism can be dominant strategy incentive compatible. θj ) = 1. and similar minima referred to below. If θ1 < θ1 . and agent 2’s payment as τ2 (θ1 ). t1 (θ1 . of which we had just shown that ˆ it equals c. We can describe ¯ this mechanism in the form described in Proposition 4. and then.5 this implies what we want to show. If we deﬁne τi ≡ θi . θ] | q(θi . This is not ruled out by the Proposition. θ2 ) = 1} and θ2 ≡ min{θ2 ∈ [θ. Incentive compatibility requires that no agent’s payment depends on their own report. Fix any θ1 . We conclude that θi = θi ˆi implies q(θi . DOMINANT STRATEGY MECHANISMS 92 Suppose q(θ) = 0 for all θ ∈ Θ. In principle. Our assumption that the set {θ | q(θ) = 1} is closed guarantees that these minima. ˆ for i ∈ {1. by Proposition 4. Budget balance requires θ1 + θ2 = c. though.5 the ˆ payment of each agent i must be θi whenever the public good is produced. the mechanism is of the form described in the Proposition. θj ) = 1 for some θj ∈ [θ. θ2 ) = 1}.

4. We write τ1 for this constant. We deﬁne θ ≡ (θs . θB ] there exist a type θS ∈ R and payments τS . For briefness.2 Dominant Strategy Incentive Compatible and Ex Post Individually Rational Direct Mechanisms By the revelation principle we restrict ourselves to direct mechanisms with deterministic trading rules. There is one potential buyer B. τS ∈ R such that: ˆ ˆ θS < θS ⇒ q(θ) = 1 and tS (θ) = τS .1 Bilateral Trade Set-Up We recapitulate brieﬂy the set-up from Section 3. Proposition 4. We consider the situation from a mechanism designer’s perspective who does not know the values of θS and θB .CHAPTER 4. and ex post budget balance. A direct mechanism is dominant strategy incentive com¯ ˆ patible if and only if for every θB ∈ [θB . Therefore. where θB ∈ [θB .9. Budget balance then requires τ2 to be constant too.4 4.4. The buyer’s utility if he purchases the ¯ good and pays a transfer t equals θB − t. DOMINANT STRATEGY MECHANISMS 93 is θ2 . The seller’s utility if he sells the good and receives a transfer payment t is equal to t. we state the result without proof. thus contradicting budget balance. A seller S owns a single indivisible good. . but not both of these sums can equal c. We require these to satisfy dominant strategy incentive compatibility.9. ex post individual rationality. 4. Only one. 4. Again we ﬁnd that the mechanism is of the form described in the Proposition. The problem is analogous to that of the previous section because we have restricted attention to deterministic mechanisms. To characterize dominant strategy incentive compatibility we can proceed as in the previous section. θB ] is the buyer’s privately observed type.4. θB ). The buyer’s utility if he does not obtain the good and pays transfer t is −t. If he does not sell the good and receives a transfer t then his utility is θS + t where ¯ θS ∈ [θS . τ1 (θ2 ) is a constant. as deﬁned as in Deﬁnition 3. θS ] is the seller’s type. the sum of contributions is: τ1 (θ2 ) + τ2 (θ1 ). Each agent knows his own type but not the other agent’s type. ˆ ˆ θS > θS ⇒ q(θ) = 0 and tS (θ) = τS .

ˆ ˆ θB = θB ⇒ q(θ) = 0 and tB (θ) = τB or q(θ) = 1 and tB (θ) = τB . A direct mechanism (q. θB ) ¯ and for every θS ∈ [θS . and that are dominant strategy incentive compatible and ex post individually rational. 4. τB ∈ R such that: ˆ θB < θB ⇒ q(θ) = 0 and tB (θ) = τB . θS ] there exist a type θB ∈ R and payments τB . Proposition 4. A dominant strategy incentive compatible direct mech¯ anism is ex post individually rational if and only if for every θB ∈ [θB .4.CHAPTER 4. θB ] : ¯ ¯ ¯ tS (θS . Deﬁnition 4.5.10. q(θ) = 0 otherwise. B} there is a strictly increasing and continuous function ¯ ψi : [θi . ˆ ˆ τB − τB = θB . . DOMINANT STRATEGY MECHANISMS ˆ θS = θS ⇒ q(θ) = 0 and tS (θ) = τS or q(θ) = 1 and tS (θ) = τS . θB ).3 Canonical Mechanisms We introduce again a class of mechanisms that encompasses some of those that we introduced earlier as expected welfare or expected proﬁt maximizing (under a budget constraint). θi ] → R such that: 1 if ψB (θB ) ≥ ψS (θS ). ˆ We continue with a standard characterization of individual rationality the proof of which we also omit. tB ) is called “canonical” if for every agent i ∈ {S. ˆ 94 ¯ ˆ ˆ ˆ and for every θS ∈ [θS . θS ] : tB (θS . θB ) ≤ θB q(θS . ˆ θB > θB ⇒ q(θ) = 1 and tB (θ) = τB . ˆ ˆ τS − τS = θS . θB ) ≥ θS q(θS . tS .

12 can best .11.CHAPTER 4. Every canonical mechanism is dominant strategy incen¯ ¯ tive compatible and ex post individually rational. B} q(θ) = 0 and ti (θ) = 0 for all i ∈ {S. and ex post budget balanced if and only if either: q(θ) = 0 and ti (θ) = 0 ˆ or there is a θ such that: ˆ q(θ) = 1 and ti (θ) = θ for all i ∈ {S. ex post individually rational.4. θ B S The proof of this result is analogous to the proof of Proposition 4. if trade takes place. for all i ∈ {S. θB ] | ψB (θB ) ≥ ψS (θS )} 0 if q(θ) = 1 if q(θ) = 0. DOMINANT STRATEGY MECHANISMS and tS (θ) = and tB (θ) = ¯ min{θB ∈ [θB .12. A direct mechanism is dominant strategy incentive compatible. The class of mechanisms described in Proposition 4.4 Ex Post Budget Balance Now we characterize direct mechanisms that are dominant strategy incentive compatible. θS ]. ¯ max{θS ∈ [θS . θ ) = 0 for every θS ∈ [θ . Suppose that the set {θ | q(θ) = 1} is closed. uS (θS . and satisfy ex post budget balance.8. θB ) = θS ¯B ] and uB (θS . The structure of our result is similar to the structure of Proposition 4. ¯ for every [θB ∈ θB . Proposition 4. Proposition 4. θS ] | ψB (θB ) ≥ ψS (θS )} 0 if q(θ) = 1 if q(θ) = 0. 95 Thus. ex post individually rational. 4. the buyer receives the lowest value that he could have had and trade would still have taken place. B} if if ˆ ˆ θS ≤ θ and θB ≥ θ. B} and all θ ∈ Θ. The proof of this result is analogous to the proof of Proposition 4. the seller receives the largest value that he could have had and trade would still have taken place. Moreover. ˆ ˆ θS > θ or θB < θ. and we omit it.8.7 and we omit it. Similarly.

6 Problems a) Assume that stochastic decision rules are allowed in the public goods problem. DOMINANT STRATEGY MECHANISMS 96 ˆ be understood as ﬁxed price mechanisms. . There is a ﬁxed price θ.5 Comments on the Literature The results on dominant strategy mechanisms in the single unit auction setting are special cases of a more general result in Mookherjee and Reichelstein (1992) who investigate conditions under which one can construct for any Bayesian incentive compatible mechanisms an equivalent dominant strategy mechanism.) Show that Proposition 4. 4. Dominant strategy mechanisms for bilateral trade are discussed in Hagerty and Rogerson (1987). and trade takes place at this price if and only if the buyer reports a value above this price and the seller reports a value below this price.5 is a special case of the more general result that you have stated. This mechanism is restrictive because the price does not depend at all on the agents’ reported valuations.CHAPTER 4. 4. I have also used Section 5.2 of Vijay Krishna’s Auction theory for this section of the notes.2. New Delhi. The section on public goods mechanisms is partially based on discussions with Arunava Sen from the Indian Statistical Institute. (You don’t have to prove this characterization. and state a characterization of dominant strategy incentive compatible direct mechanisms that is analogous to Proposition 4.

They have to choose an alternative a out of some set A of mutually exclusive alternatives. 97 . The reason for starting with dominant strategy mechanisms is that these are a subset of the Bayesian incentive compatible mechanisms. N }. The set of agents is denoted by I = {1. some of the results in this chapter will also be useful in the next chapter.1 Introduction In the ﬁrst three chapters of these notes we have considered examples that illustrate the general theory of mechanism design. θi ) − ti . . 2. .Chapter 5 Dominant Strategy Mechanisms: General Theory 5. Note that the set-up does include the case in which A is the set of lotteries over some ﬁnite set. Here. . We begin by considering dominant strategy mechanisms. 5. and agent i pays transfer ti is: ui (a. We also introduce some new results. Agent i’s utility if alternative a is chosen. . In this and the next Chapter we are now going to consider the extent to which we can generalize what we have seen in these examples.2 Set-Up There are N agents. and ti is agent i’s transfer. Therefore. θi is agent i’s type. We shall investigate which of our earlier results continue to be true in the more general model.

If we assume that agents are risk-neutral. . Deﬁnition 5. θ−i ). Finally. . tN ) consists of a mapping q : Θ → A. 98 We shall employ similar notation as before: The set of possible types of agent i is Θi which we now take to be some abstract set. all i ∈ I and all θi ∈ Θi we have: ui (q(θ). we write θ−i for a vector θ of types if we leave out agent i’s type.” Note that we restrict ourselves to mechanisms where the payment is deterministic. . Our interest is in dominant strategy incentive compatible mechanisms. such as individual rationality and budget balance. θi ) − ti (θ) ≥ ui (q(θi . DOMINANCE: GENERAL THEORY and ui is an expected utility function. θN . and mappings ti : Θ → R. A direct mechanism (q. θ−i ). . . θ2 . that maps every type vector into a collective decision. For the moment. then this is without loss of generality. We introduce other requirements. A direct mechanism is “dominant strategy incentive compatible” if for all θ ∈ Θ. later when needed. and we can restrict attention to direct mechanisms. We call q the “decision rule. . . that indicate for each type vector the transfer that agent i needs to make. t2 . as long as we are only concerned with characterizations of incentive compatibility and individual rationality. however. θi ) − ti (θi . . we are only concerned with incentive compatibility. one for each player i ∈ I.2. leaving out Θi . . The next sections will only be concerned with incentive compatibility. When we consider budget balance. . The revelation principle holds in this setting. Deﬁnition 5. The set of all possible type vectors is Θ ≡ Θ1 × Θ2 × .CHAPTER 5. The set of all θ−i is Θ−i which is the cartesian product of the sets Θj . this point is more delicate. θN ). t1 .1. We denote by θ the vector of types: (θ1 . .

we need to make therefore an additional assumption regarding the type spaces Θi if we want to obtain a payoﬀ equivalence result. such as Proposition 4. . but it is more technical. We omit the proof. Convexity of the utility function in a player’s own type is satisﬁed. The proof is analogous to the proof of similar results that we provided earlier. . As our earlier revenue equivalence results. We shall also make use of an assumption about how the agents’ utility functions depends on their types. and hence convex. assume that for every i ∈ I the function ui (a. DOMINANCE: GENERAL THEORY 99 5. tN ) is a dominant strategy incentive compatible mechanism.3 Ex Post Revenue Equivalence We begin by providing a result that is analogous to the ex post revenue equivalence results that we obtained in Chapter 4. t1 . Suppose that (q. however. We shall assume that every agent i’s utility function is a convex function of the agent’s own type. With ﬁnite type spaces these results are not true.2. . for example. not enough. . The proof of this result is given in Krishna and Maenner (2001). The assumption that we shall make below is that for every i ∈ I the set Θi is a convex subset of some ﬁnite dimensional Euclidean space. tN ). the player’s utility is a linear function of the type vector. . (q. In our abstract setting in this part of the lecture notes. t2 . our result concerns ex post payoﬀs. t2 . and if the type of a player is just the vector of this player’s Bernoulli utilities of the outcomes. this result says that for given decision rule q the set of all transfer rules that implement q consists of one function and all its parallel translations. θi ) is a convex function of θi . Moreover. Then a direct mechanism with the same decision rule. . Suppose that for every i ∈ I the set Θi is a convex subset of a ﬁnite dimensional Euclidean space. The proofs of those results used the fact that in all the examples the type spaces were connected sets. Proposition 5. . if the space of alternatives A is the set of all lotteries over a ﬁnite outcome space. This assumption will ensure that agents’ utility under an incentive compatible mechanism depends suﬃciently smoothly on their type so that we can employ an envelope theorem. is incentive compatible if and only if for every i ∈ I and every θ−i ∈ Θ−i there is a number τi (θ−i ) ∈ R such that ti (θ) = ti (θ) + τi (θ−i ) for all θ ∈ Θ. . because the appropriate diﬀerentiability properties of the functions involved needs to be established.CHAPTER 5. . In this case. Because we are dealing with dominant strategy implementation.1. t1 . An assumption for the type spaces alone is.

θi ) for all a ∈ A.4 Implementing Eﬃcient Decision Rules We now begin an investigation of the question for which decision rules q there are transfer rules that make the decision rule q dominant strategy incentive compatible. Proposition 5.” although a much more careful investigation would be needed if we wanted to clarify the relation between maximizing utilitarian welfare and “Pareto eﬃciency. DOMINANCE: GENERAL THEORY 100 5. θj ) − τi (θ−i ) = j=1 uj (q(θi . A direct mechanism (q. and take θ−i as given. . tN ) is called a “VickreyClarke-Groves” (VCG) mechanism if q is an eﬃcient decision rule. Deﬁnition 5. Proof. .4. θi ) ≥ i=1 i=1 ui (a. θj ) + τi (θ−i ) for all θ ∈ Θ. VCG mechanisms are dominant strategy incentive compatible.” Deﬁnition 5. We begin by showing that decision rules that maximize ex ante expected welfare. We encountered special cases of these mechanisms already in earlier parts of these notes. .CHAPTER 5. and if for every i ∈ I there is a function τi : Θ−i → R such that ti (θ) = − j=i uj (q(θ). t2 . If agent i is of type θi and reports that she is of type θi . . Consider any agent i. then her utility is: ui (q(θi . t1 . are always dominant strategy incentive compatible. θ−i ). θ−i ). We shall call such decision rules “eﬃcient. θj ) − τi (θ−i ). We shall now introduce a class of mechanisms that make eﬃcient decision rules dominant strategy incentive compatible.2. θ−i ).3. θi ) + j=i N uj (q(θi . where welfare is deﬁned in a utilitarian way. An allocation rule q ∗ is called “eﬃcient” if for every θ ∈ Θ we have: N N ui (q ∗ (θ).

1. assume that for every i ∈ I the function ui (a. Firstly. given that we showed in the previous section that eﬃcient mechanisms are dominant strategy incentive compatible. . DOMINANCE: GENERAL THEORY 101 Note that τi (θ−i ) is not changed by agent i’s report. We discuss a variety of closely related properties of decision rules which are necessary. the mechanism designer might have objectives other than eﬃciency. As q(θ) maximizes social welfare for type vector θ. . Moreover. Suppose that for every i ∈ I the set Θi is a convex subset of a ﬁnite dimensional Euclidean space. Our focus will be on the decision rules q that may form part of an incentive compatible direct mechanism rather than on the transfer rules. . proﬁt maximization. Only the ﬁrst expression matters for i’s incentives. tN ) is a dominant strategy incentive compatible mechanism. Corollary 5.CHAPTER 5. t2 . t1 . . There are two reasons. One might wonder why this is of interest. the mechanism designer might face constraints other than dominant strategy incentive compatibility which make the implementation of an eﬃcient decision rule impossible. and suppose that q is eﬃcient. it is optimal for agent i to report her true type: θi = θi . By Proposition 1 every dominant strategy incentive compatible mechanism that implements an eﬃcient decision rule q must involve the same transfers as the VCG mechanism up to additive constants τi (θ−i ) that may be added to any agent i’s transfers. Then (q. But this expression is social welfare at type vector θ if the decision is q(θi . θ−i ). and under certain conditions also suﬃcient for decision rules to be dominant strategy implementable. 5.g. e. . t1 . We can combine this result with the payoﬀ equivalence result in Proposition 1 to obtain conditions under which VCG mechanisms are the only mechanisms that make eﬃcient decision rules dominant strategy incentive compatible. Suppose that (q. . θi ) is a convex function of θi . But adding such constants to a VCG mechanism yields by the deﬁnition of VCG mechanisms another VCG mechanism. Secondly. Proof. tN ) is a VCG mechanism. . t2 .5 Characterizing All Incentive Compatible Decision Rules In this section our objective is to characterize the set of all dominant strategy incentive compatible mechanisms.

and all θi . θi ) ≤ ti (θi . t2 . θ−i ) (5. θ−i ) − ti (θi . θi ). The deﬁnition says that in this case the utility diﬀerence for agent i between a1 and a2 must 1 2 have decreased as we switched from θi to θi . θi ) − ti (θi . θ−i ) ⇔ 1 1 1 2 ui (a1 .3. . θ−i ) ≥ ui (a2 . tN ) is a dominant strategy incentive compatible mechanism. θi ) − ti (θi . θi ) ≥ ui (a1 . This deﬁnition considers a situation where the collective choice for a type 1 1 vector (θi . (5. Deﬁnition 5. then the collective choice becomes some alternative a2 . θ−i ) − ti (θi . θ−i ) = a1 and q(θi . θi ) − ti (θi . if q(θi . . but if we change agent i’s type from θi to 2 θi . θ−i ).3) which means that q is weakly monotone. A decision rule q is “weakly monotone” if for every i ∈ I. θi ∈ Θi . θ−i ) is alternative a1 . Suppose (q. θ−i ) ≤ ui (a2 . Here. . θi ) − ui (a2 .2) These two inequalities imply: 1 1 2 2 ui (a1 . if q(θi . θ−i ) = a2 . we start with a result that can be proved by an appropriate adaptation of the argument that led to our earlier results.” The result says that agent i’s willingness to pay for a1 as 1 opposed to a2 must be at least as large when agent i is of type θi as it is 2.5. We can interpret the utility diﬀerence as agent i’s “willingness to pay for alternative a1 as opposed to alternative a2 . θi ) − ui (a2 . Then q is weakly monotone. (5. θi ) ≥ ui (a1 . θi ) − ui (a2 . Proof. and all θi . t1 . then: 1 1 1 2 ui (a1 .CHAPTER 5. θi ) − ui (a2 . θi ∈ Θi . . θi ) − ti (θi . θi ) ≥ ti (θi . θi ) − ui (a2 . when agent i is of type θi Proposition 5. Dominant strategy incentive compatibility implies for all i ∈ I. θi ).1) and 2 1 2 2 ui (a1 . then 1 1 2 2 ui (a1 . θ−i ∈ 1 2 1 2 Θ−i . θi ) − ui (a2 . . 1 2 1 2 θ−i ∈ Θ−i . θ−i ) ⇔ 2 2 1 2 ui (a1 . DOMINANCE: GENERAL THEORY 102 In the three examples of Chapters 3 and 4 we found that a necessary and suﬃcient condition for a decision rule to be implementable by a direct incentive compatible mechanism was that q was monotone. θ−i ) = a2 . θ−i ) = a1 and q(θi .

θi )) ≤ 0 κ=1 κ where for all κ = 1. θi ) ≤ 0. θi ) − ui (a1 . leaving the chosen alternative hypothetically we switch agent i’s type to θi 2 unchanged as a1 . Deﬁnition 5.6. .CHAPTER 5. . and assume that the last element of the sequence is the 1 k same as the ﬁrst: θi = θi . . Now suppose we consider a sequence of length k ∈ N. . θi . every θ−i ∈ Θ−i and every sequence of length k ∈ N of types of agent i. . . θi . . . θ−i ) with implemented 2 . The condition says that the sum of these two utility changes is not positive. (5. Note that we can re-write the inequality that deﬁnes weak monotonicity as: 2 1 1 2 ui (a1 . Nevertheless. . Next. we have: i k−1 κ+1 κ (ui (aκ . an elementary proof shows that cyclical monotonicity is equivalent to dominant strategy implementability. We can consider the same sequence of switches as we considered in the previous paragraph. . 1 2 k k 1 (θi . . k we deﬁne: aκ ≡ q(θi . θi . . θ−i ). Thus. . with θi = θi . and we switch agent i’s type to θ 1 . we start at utility proﬁle (θi . First. θi ) + ui (a2 . t1 . t2 . θi ) − ui (a2 . .” is necessary and suﬃcient for implementability. tN ) if and only if q is cyclically monotone. and 2 . k ≥ 2 of types of 1 2 k agent i: θi . 1 we start at utility proﬁle (θi . “Cyclical monotonicity” requires the sum to be not positive. θ−i ) with implemented alternative a1 . cyclical monotonicity is the same as weak monotonicity if we restrict attention to the case that k = 3. Both switches may cause an increase or a decrease in agent i’s utility. A decision rule q is part of a dominant strategy incentive compatible direct mechanism (q. The following proposition is due to Rochet (1987). A remarkable aspect of this theorem is that it does not require any structure at all for the sets of alternatives or the sets of types. A decision rule q is “cyclically monotone” if for every i ∈ I. . . Proposition 5. θi ) ∈ Θk .4. DOMINANCE: GENERAL THEORY 103 We shall now show that a stronger condition than weak monotonicity. 2. . and sum up the changes to agent i’s utility caused by these switches. called “cyclical monotonicity. θi ) − ui (aκ .4) This inequality considers on the left hand side the sum of the changes to agent i’s utility that occur in the following two thought experiments. leaving again the chosen alternative a i alternative hypothetically unchanged.

. . Here. . . .8) κ for all (θi . and we can deduce: k−1 κ+1 κ+1 ui (aκ .θi )∈S(θi ) κ=1 κ+1 κ ui (aκ . θi ) ≤ 0 ⇔ κ=1 k−1 κ+1 ui (aκ . θ−i ) ∈ Θ. θi ) − κ=1 κ=1 k−1 κ+1 ui (aκ+1 . θi ) κ ui (aκ . we shall need some further deﬁnitions. Next. k − 1 that type θi has no incentive to pretend to be type κ. To construct the transfer payments..6) k 1 Because θi = θi . we can write the subtracted sum as follows: k−1 k−1 κ κ+1 . . 2.5) We sum these inequalities over all κ = 1. θi ) ≤ 0 (5. θi ) ≤ 0. k − 1. θi ) ≤ 0 ⇔ κ=1 κ+1 κ ui (aκ . θi ) (5. θ−i ) − ti (θi . θi .. .. θi ) of elements of Θi that satisfy: θi = θi and θi = θi . . θ−i ) ⇔ κ+1 κ+1 κ+1 κ ui (aκ . θ−i ) for κ = 1. . θ−i ) ≡ sup 1 2 k (θi . θi ) − ti (θi . where aκ is deﬁned to be q(θi . 2. k−1. θi ) − ti (θi . θi ) − ui (aκ+1 . θi ) − ui (aκ . We ﬁrst show that if q is part of a dominant strategy incentive compatible direct mechanism then it is cyclically monotone. θi ) ≤ ti (θi . DOMINANCE: GENERAL THEORY 104 Proof. We ﬁx an arbitrary type ¯ θi ∈ Θi . θi ) − ui (aκ . θ−i ) ≤ ui (aκ+1 . . θi ) − ui (aκ+1 . . . . . κ=1 ui (a κ=1 k−1 − (5. k can be any element of N with k ≥ 2.CHAPTER 5.7) which is what we wanted to show. given that all other agents have reported type θ−i : κ+1 κ+1 κ+1 κ ui (aκ . Dominant strategy κ+1 incentive compatibility implies for every κ = 1. Then.. Then deﬁne the function V : Θ → R by: k−1 V (θi . . (5. we shall show that cyclical monotonicity implies that the rule q can be combined with transfer payments so that the direct mechanism that results is dominant strategy incentive compatible. 2. The right hand side then becomes zero. for every θi ∈ Θi we deﬁne S(θi ) to be the set of all ﬁnite 1 2 k 1 k ¯ sequences (θi . θ−i ).θi . . .

CHAPTER 5. DOMINANCE: GENERAL THEORY

105

Before we proceed, we verify that the function V is well-deﬁned. This is the case if the expression the supremum of which is the right hand side of equation (5.8) is bounded from above for given θi and θ−i if q is cyclically monotone. This is what we show. First we note that the condition of cyclical monotonicity means that ¯ for θi = θi that all sums on the right hand side of (5.8) are non-positive. ¯ Therefore, for θi = θi , the supremum is well-deﬁned. Indeed, because the ¯ ¯ ¯ trivial sequence (θi , θi ) is contained in S(θi ), and because for this sequence the sum over which we take the supremum in the deﬁnition of V is zero, we ¯ can conclude that V (θi , θ−i ) = 0 for all θ−i ∈ Θ−i .

1 2 k ¯ Now consider some θi = θi , ﬁx a sequence (θi , θi , . . . , θi ) ∈ S(θi ) and κ as in the deﬁnition of V . Then: deﬁne a k−1

¯ V (θi , θ−i ) =

k−1

sup

1 2 k ¯ (θi ,θi ,...,θi )∈S(θi )

κ+1 κ ui (aκ , θi ) − ui (aκ , θi ) κ=1

≥

κ=1

κ+1 κ ¯ ui (aκ , θi ) − ui (aκ , θi ) + ui (ak , θi ) − ui (ak , θi )

(5.9) ¯ Here, the inequality follows from the fact that S(θi ) includes the set of all ﬁnite sequences that have θi as their pen-ultimate element, and that have ¯ θi as their last element. The inequality that we have obtained is equivalent to:

k−1 κ+1 κ ¯ ¯ ui (aκ , θi ) − ui (aκ , θi ) ≤ V (θi , θ−i ) + ui (ak , θi ) − ui (ak , θi ) . κ=1

**(5.10) ¯ Using V (θi , θ−i ) = 0, which we showed earlier, we can infer:
**

k−1 κ+1 κ ¯ ui (aκ , θi ) − ui (aκ , θi ) ≤ ui (ak , θi ) − ui (ak , θi ), κ=1

(5.11)

which shows that the sums that appear on the right hand side of (5.8) are bounded from above, and therefore that V is well-deﬁned. We now use the function V to construct the payment schemes that make q dominant strategy incentive compatible. Indeed, we shall construct the payment schemes so that the utility ui (q(θ), θi ) is exactly equal to V (θ) for

CHAPTER 5. DOMINANCE: GENERAL THEORY all θ ∈ Θ. That is, we set: ui (q(θ), θi ) − ti (θ) = V (θ) ⇔ ti (θ) = ui (q(θ), θi )) − V (θ).

106

(5.12)

To prove that this implies that the mechanism is dominant strategy incentive compatible, we need to show: ui (q(θi , θ−i ), θi ) − ti (θi , θ−i ) ≥ ui (q(θi , θ−i ), θi ) − ti (θi , θ−i ) ⇔ ui (q(θi , θ−i ), θi ) − (ui (q(θi , θ−i ), θi )) − V (θi , θ−i )) ≥ ui (q(θi , θ−i ), θi ) − ui (q(θi , θ−i ), θi )) − V (θi , θ−i ) ⇔ V (θi , θ−i ) ≥ V (θi , θ−i ) + ui (q(θi , θ−i ), θi ) − ui (q(θi , θ−i ), θi )) (5.13) This is true by the deﬁnition of V because the set of all ﬁnite sequences that ¯ start with θi and end with θi includes the set of all ﬁnite sequences that ¯i , have θ as their penultimate element, and end with θi . start with θ i

5.6

All Incentive Compatible Decision Rules When Outcomes are Lotteries

We now develop a condition for implementability that is closely related to cyclical monotonicity, but that applies only to the special case in which the set of alternatives A is the set of probability distributions over some ﬁnite set of outcomes. Denote the number of such outcomes by M . Each element of A is of the form (p1 , p2 , . . . , pM ), where p ≥ 0 for = 1, 2, . . . , M and M p = 1. The set of types of any player i, Θi , consists of vectors =1 θi = (θi,1 , θi,2 , . . . , θi,M ) of Bernoulli utilities of the M outcomes. We assume that the set Θi is convex for every i ∈ I. Consider a given direct mechanism. For every i ∈ I deﬁne Ui (θ) ≡ q(θ) · θ − ti (θ). Here, “·” stands for the scalar vector product. Hence, Ui is the expected utility of player i if all players report their types truthfully. Now suppose that the mechanism is dominant strategy incentive compatible, and ﬁx some value of θ−i . Then Ui as a function of θi alone satisﬁes: Ui (θi , θ−i ) = max q(θi , θ−i ) · θi − ti (θi , θ−i ).

θi ∈Θi

(5.14)

Thus, Ui is the maximum of linear functions, and therefore it is convex. Now we review the following deﬁnition from Rockafellar (1970), p. 214: A vector x∗ is a subgradient of a convex function f at a point x if: f (z) ≥ f (x) + x∗ · (z − x)

CHAPTER 5. DOMINANCE: GENERAL THEORY

107

for all z in the domain of f . In our case, q(θi , θ−i ) is a subgradient of U (θi , θ−i ) (as a function of θi ) at θi for every θi . This is because: Ui (θi , θ−i ) ≥ U (θi , θ−i ) + q(θi ) · (θi − θi ) ⇔ Ui (θi , θ−i ) ≥ q(θi ) · θi − ti (θi , θ−i ) + q(θi ) · (θi − θi ) ⇔ Ui (θi , θ−i ) ≥ q(θi ) · θi − ti (θi , θ−i ) (5.15)

which holds by dominant strategy incentive compatibility for every θi ∈ Θi . The following result shows that being a subgradient of a convex function is not only a necessary but also a suﬃcient condition for q to be implementable. Proposition 5.5. Suppose that A is the set of all probability distributions over some ﬁnite set of outcomes. Suppose that for every i ∈ I the set Θi is a convex set of vectors of Bernoulli utility functions of player i. Then a decision rule q is part of a dominant strategy incentive compatible direct mechanism (q, t1 , t2 , ..., tN ) if and only if for every i ∈ I and for every θ−i ∈ Θ−i there is a convex function Ui : Θi → R such that qi (θi , θ−i ), regarded as a function of θi only, is a subgradient of Ui . Proof. The argument preceding Proposition 5.5 has shown the necessity of the condition in Proposition 5.5. To see that it is also suﬃcient, ﬁx θ−i and deﬁne for every θi ∈ Θi player i’s transfer ti (θi , θ−i ) so that the function Ui of which q is the subgradient is exactly player i’s expected utility. The equivalencies that precede Proposition 5.5 establish that dominant strategy incentive compatibility holds. Theorem 24.8 in Rockafellar (1970) establishes that q is cyclically monotone if and only it is subgradient of a convex function. Therefore, Rochet’s result that was cited in the previous section is a generalization of Rockafellar’s result to the more general setting in which the sets of alternatives and types are arbitrary.

5.7

Single Dimensional Type Spaces

We want to relate weak monotonicity in the sense of Deﬁnition 5.5 to monotonicity as we have seen it in earlier parts of these notes. A characteristic of these earlier parts was that agents’ type spaces were subsets of R, and were therefore single-dimensional. Single-dimensionality is at ﬁrst sight purely mathematical concept. In economic terms, however, it means that type spaces can be completely ordered so that “higher types” have a larger marginal willingness to pay for “higher alternatives.” Our purpose in this

θi ∈ Θi we shall say that θi Ri θi (“θi is a higher type than θi relative to Ri ” ) if ui (a . The indiﬀerence relation derived from Ri is denoted by Ii : aIi b ⇔ [aRi b and bRi a]. i. The single crossing condition will not rule out that some types prefer not obtaining the good over obtaining the good. Consider any i ∈ I. and transitive binary relation. reﬂexive. In that example we might take Ri to be given by: aRi b ⇔[i obtains the good with at least as high probability in a as in b]. The strict order derived from Ri is denoted by Pi : aPi b ⇔ [aRi b and not bRi a]. What do we mean here by “higher alternatives”? This is deﬁned by the ordering Ri on A.” . θi Ri θi means that θi attaches larger marginal value to higher alternatives than θi for any two ordered alternatives that we compare. It is important not to mistake Ri for agent i’s preference relation. Thus. Single crossing will require that the larger types are the larger is the marginal willingness to pay for the good. Note that the following deﬁnition applies to any arbitrary decision rules. θi ) − ui (a.5 to the monotonicity properties of earlier parts of these notes. By this we mean that it is a complete. consider the single unit auction example. θi ) = ui (a .e. DOMINANCE: GENERAL THEORY 108 section is to study the implications of the single dimensionality condition for dominant strategy incentive compatibility. θi ) − ui (a. θi ) = 0 holds for all a ∈ A with a Ii a. To understand this. the order Ri is conditional on Ri . θi ) − ui (a. Given an ordering of alternatives. Intuitively. Let Ri be an order of A. Before we can do so we need to formalize the notions of “higher alternatives” and “higher types. irrespective of whether the type spaces whose cross product makes up the domain of the decision rule satisﬁes any “single dimensionality condition. θi ) − ui (a. Deﬁnition 5. θi ) holds for all a ∈ A with a Pi a and ui (a .7. In other words. We are now going to use the order Ri to deﬁne the monotonicity property that will allow us to relate weak monotonicity as in Deﬁnition 5.CHAPTER 5. For any pair of types θi .” We begin by describing what “higher alternatives” are. θi ) > ui (a . and let Ri be an order of A. θi unambiguously has a stronger preference for higher alternatives than θi . have a negative marginal willingness to pay for the good. we can now order types.

θi ). If a decision rule q is weakly monotone then it is monotone with respect to Ri . and let Ri be an order of A. and let Ri be an order of A. Proposition 5. θi ) < ui (a. θi ) > ui (a . monotonicity places restrictions on choices even if types are not ordered in Ri . Deﬁnition 5. say the diﬀerence ui (a . Suppose q is weakly monotone. The next result shows that monotonicity in this sense is implied by weak monotonicity. which is in general incomplete. Deﬁne a ≡ q(θi . Proof. θi ) − ui (a . By contrast.CHAPTER 5. θi ) − ui (a . θi ) ⇔ ui (a. we must have: ui (a. θi ) − ui (a. θ−i )Ri q(θi .16) . The reason that we can’t prove in general the converse of this result is that weak monotonicity imposes a restriction on collective decisions only in the case that types are comparable in the order Ri . θ−i ) and a ≡ q(θi . and let Ri be an order of A. which contradicts weak monotonicity. θ−i ). θi ) ≥ ui (a.8.6 is that it is true for any order Ri of A. By the deﬁnition of weak monotonicity. namely the diﬀerence in utility for some alternatives. Consider any agent i. θi ) where a Pi a. But we can strengthen Proposition 5. Suppose a Pi a. Consider any agent i. θ−i ) for all θ−i ∈ Θ−i . θi ) − ui (a . Consider any agent i. θi ∈ Θi with θi = θi we have either θi Ri θi or θi Ri θi or both. The type space Θi is “single dimensional with respect to Ri ” if for any θi . We now prove indirectly that θi Pi θi implies aRi a . DOMINANCE: GENERAL THEORY 109 Deﬁnition 5. Consider some θ−i ∈ Θ−i .6. θi ) − ui (a.6 so that it becomes an equivalence if we consider completely ordered type spaces. Note that if a type is in this sense single dimensional we can assign to any type a real number. A remarkable feature of Proposition 5. that is this marginal utility unambiguously (5. A decision rule q is called “monotone with respect to Ri ” if: θi Ri θi ⇒ (q(θi . and suppose θi Ri θi . θi ) − ui (a . and this mapping from types to real numbers will be invertible. θi ) − ui (a. Then θi Pi θi implies ui (a . this deﬁnition says that a decision rules is “monotone with respect to Ri ” if an unambiguous increase in an agent’s type leads to a “higher” alternative being chosen.9. In words. θi ).

Moreover. because q is monotone.18) (5. we restrict our attention in the next result to the case that A is ﬁnite. Proposition 5. For any agent i the type space Θi is “bounded” if there is a constant c > 0 such that for all a. and consider θ ∈ Θ and θi ∈ Θi such that q(θ) = a but q(θi .CHAPTER 5. This shows that q is weakly monotone. the larger all marginal utilities of the type.7. the larger this number. θi ) = ui (a. We also assume that the type spaces are bounded in the sense of the following deﬁnition: Deﬁnition 5. We can now show that for completely ordered. or ui (a. we next show that monotonicity is suﬃcient for an allocation rule to be implementable by appropriate transfer schemes. aRi a . Having established that weak monotonicity and monotonicity are equivalent on single dimensional domains. This. Proof. because θi Ri θi : ui (a. θi ) − ui (a . Then a decision rule q is weakly monotone if and only if it is monotone with respect to Ri for every i ∈ I. together with Proposition 5. In the light of Proposition 5. . Because θ satisﬁes the single crossing condition we must have: θi Ri θi or θi Ri θi . if aPi a . Restrict attention to the case a = a . θ−i ) = a . then implies that on single dimensional domains monotonicity is necessary and suﬃcient for implementability. θi ) − ui (a.10.17) Intuitively. Without loss of generality assume the former. the type space of agent i is “bounded” if there is a uniform upper bound for agent i’s willingness to pay for a change in the decision. type spaces monotonicity is suﬃcient for weak monotonicity. θi ). θi ) > ui (a. For simplicity. θi ). (5. Moreover. θi ) < c. a ∈ A and all θi ∈ Θi we have: −c < ui (a .3. For every i ∈ I let Ri be an ordering of A such that Θi is single dimensional with respect to Ri .19) (5. or single-dimensional. θi ) − ui (a . θi ) − ui (a . Then. DOMINANCE: GENERAL THEORY 110 identiﬁes the type. if aIi a . The case that A is inﬁnite is harder only in terms of notation. So suppose q is monotone. θi ) − ui (a .6 we only have to show that monotonicity implies weak monotonicity.

Pi a1 . θ−i ) = ak }. an } where an Ri an−1 Ri . .24) . For every i ∈ I let Ri be an order of A.21) Deﬁne for every k = 2. then the proof below should be modiﬁed treating alternatives between which the decision maker is indiﬀerent as identical alternatives. . n: τ k ≡ inf{ui (ak . . Fix i ∈ I and θ−i ∈ Θ−i . . θ−i ) | θi ∈ Θi }. t2 . Then there are transfer rules t1 . . This will obviously make it optimal for player i to report her type truthfully if the other agents’ types is θ−i because player i will be indiﬀerent between all possible reports. Denote the range of q over θi . For simplicity we assume that all these preference relations are strict: an Pi an−1 Pi . a2 . i (5. θi ) ≤ τ k i k > k. t2 . θi ∈ Θk . Ri a1 . θi ) − ui (ak−1 . . θi ) − ui (ak−1 . θi ∈ Θk ⇒ ui (ak . . . . . θi ) ≥ τ k . .22) The inﬁmum here is well-deﬁned because Θi is bounded. θ−i ) = 0 for all θi . If some are not. that is the set {q(θi . i κ=2 τ (5. Note that the ordering of types implies: k < k. Let q be a decision rule that is monotone with respect to Ri for every i ∈ I. . .23) (5. . t1 . . n the set Θk ≡ {θi ∈ Θi | q(θi .CHAPTER 5. If n ≥ 2 deﬁne for every k = 1. θi ) − ui (ak−1 . . i We deﬁne agent i’s transfer payment as follows: 0 if θi ∈ Θ1 .8. .20) Monotonicity and the single crossing condition imply that the sets ΘK are i ordered in the following sense: k k k k > k. . 2. . DOMINANCE: GENERAL THEORY 111 Proposition 5. and suppose that for every i ∈ I the type space Θi is bounded and is single dimensional with respect to Ri . i ti (θ) = k k if θi ∈ Θk where k ≥ 2. (5. θi ) | θi ∈ Θk }. θi ∈ Θk ⇒ θi i i Pi k θi . Suppose that A is ﬁnite. . If n = 1 we set ti (θi . . . . by {a1 . Proof. . θi ∈ Θk ⇒ ui (ak . tN such that (q. i (5. tN ) is dominant strategy incentive compatible.

(2006) have found several other conditions under which this result is true. Bikhchandani et.CHAPTER 5. A symmetric argument proves there is no incentive for agent i to report a type in Θk where k < k. then the change in her i utility in comparison to truthful reporting will be: k ui (ak . i Proposition 8 is analogous to the result that we obtained in earlier sections that monotonicity of q is necessary and suﬃcient for implementability of q. Bikhchandani et al. say that the set Θi of types of agent i is “rich” if there is some reﬂexive and transitive. 5.8 Suﬃciency of Weak Monotonicity The single crossing domains described in Section 5. that satisﬁes: aRi b ⇒ u(a) ≥ u(b). If she reports a type in Θk where k > k. DOMINANCE: GENERAL THEORY 112 We verify that this transfer scheme makes truthful reporting of θi is an optimal strategy for agent i given any θ−i . The inequality follows from the ﬁrst inequality in (5.e. θi ) − κ=k+1 k τk ≤ κ=k+1 ui (aκ . there is a θi ∈ Θi such that ui (a. θi ) = u(a) for all a ∈ A. i. θi ) − ui (aκ−1 . If agent i’s true type is θi ∈ Θk . binary relation Ri on A such that all utility functions that represent Ri are possible utility functions of agent i.25) The ﬁrst equality is a simple re-writing.23). but possibly incomplete. Consider any i ∈ I. θi ) = 0 (5. i and if she reports any type in Θk her utility will be independent of her i report. θi ) − ui (aκ−1 . the domain of preferences satisﬁed a single dimensionality condition (although we allowed inﬁnite outcome sets). In all those models. θi ) − κ=k+1 ui (aκ . We explain here the simplest one. The type space Θi is called “rich” if there is a possibly incomplete preference relation Ri on A such that for every u : A → R that represents Ri . . Formally: Deﬁnition 5. al.11. θi ) − ui (aκ−1 . θi ) − κ=k+1 k τκ k = κ=k+1 k ui (aκ . θi ) − ui (ak .7 are not the only domains on which one can show that weak monotonicity is suﬃcient for implementability. and is thus equivalent to cyclical monotonicity.

θi ).10. θ ∈ Θ. then follows from the N -fold application of the same argument. Bikhchandani et. the more utility functions may represent Ri . The assertion that we have to prove. . θi ) − ui (b. implies q(θ ) = a. 5.12. and ui (a. Proof. q(θ) = a and that θ satisﬁes the condition in the deﬁnition of PAD. t2 . q(θ ) = a. . θ ∈ Θ. Then there are transfer schemes t1 . The reason is that the less complete Ri is. θ−i ) = a for all i ∈ I. DOMINANCE: GENERAL THEORY 113 Note that this condition becomes more restrictive as Ri becomes less complete. θi ) ⇔ ui (b. Proposition 5. i. . θi ) > ui (a.9 Positive Association of Diﬀerences We now consider another condition that is weaker than monotonicity. q satisﬁes PAD if whenever an alternative that is chosen at some type vector will also be chosen at any other type vector where the alternative’s marginal utilities in comparison to other alternatives are larger. Bikhchandani et. al. tN such that (q. . θi ) − ui (b. .9. θ−i ) = b = a. t2 . tN ) is dominant strategy incentive compatible. θi ) > ui (a. Suppose that for every i ∈ I the type space Θi is rich. called “positive association of diﬀerences” (PAD). We are considering this relatively weak condition because it is suﬃcient to obtain a surprisingly strong result. emphasize that their result does not apply to the case that A is a set of lotteries over outcomes because this would make the set A inﬁnite. θi ) − ui (a. . θi ) for all i ∈ I and b ∈ A with b = a. Thus.CHAPTER 5. We shall prove that then q(θi . Suppose that θ. In words. as comparisons are dropped from Ri . t1 . Let q be a weakly monotone decision rule. θi ) − ui (b.’s Theorem 1 is: Proposition 5. Suppose that A is ﬁnite. θi ) − ui (b. al. A decision rule q satisﬁes “positive association of diﬀerences” (PAD) if θ. q(θ) = a. Suppose q(θi . we have a contradiction with weak monotonicity. θi ) − ui (a. θi ) < ui (b.e. . Deﬁnition 5. If q is weakly monotone then it satisﬁes PAD. . . By the conditions on θ and θ we have: ui (a.

t2 . Roberts (1979) showed that if the domain of a decision rule consists of all possible utility functions. t2 .13.CHAPTER 5. .11. tN ) is dominant strategy incentive compatible. θi ))a∈A = ν. by assuming A to be ﬁnite.11. monotonicity also refers to type proﬁles which result in diﬀerent collective decisions. . Note that this result. the condition for Θi on which this result relies is very restrictive. .11. . then “Positive Association of Diﬀerences” is suﬃcient for dominant strategy incentive compatibility. It rules out that the direct mechanism that we are constructing embodies any prior knowledge about the agents’ preferences. θi ) + F (a) for all a ∈ A. tN ) such that (q. We shall also assume that A is ﬁnite. . A decision rule q is called “ﬂexible” if its range. For every ﬂexible decision rule q that satisﬁes PAD there are transfer rules (t1 . Proposition 5. Suppose A is ﬁnite. Proposition 5. θi ))a∈A = ν. However. and show how it implies Proposition 5.12. . t1 . By contrast. One of the further conditions needed to state Roberts’ result is that the decision rule q is ﬂexible: Deﬁnition 5. DOMINANCE: GENERAL THEORY 114 PAD is weaker than monotonicity because it only puts restrictions on type proﬁles which result in the same collective decision. rules again out the case in which A is the set of all lotteries over some ﬁnite set of outcomes. Although we shall not provide a proof of this characterization. Suppose A is ﬁnite. θi ) + F (q(θ)) ≥ i=1 i=1 ki ui (a. we mention it. Also. and there is a function F : A → R such that for every θ ∈ Θ: N N ki ui (q(θ). Then a ﬂexible decision rule q satisﬁes PAD if and only if for every i ∈ I there is a real number ki > 0. has at least three elements. . Roberts proved his result by obtaining an interesting characterization of all decision rules that satisfy the conditions of Proposition 5. and suppose for every i ∈ I and ν ∈ R#A there is a θi ∈ Θi such that (ui (a. Observe that a decision rule can only be ﬂexible if #A ≥ 3. . and if some other conditions hold. and suppose for every i ∈ I and ν ∈ R#A there is a θi ∈ Θi such that (ui (a. . q(Θ).

and all other agents report type vector θ−i is: 1 ui (q(θi . θ−i )) . t2 . This follows from a generalization of the VCG construction. DOMINANCE: GENERAL THEORY 115 In words. This function assigns to each alternative a measure of welfare that is independent of agents’ types. F could pick out some particular alternative a ∈ A as the “status quo”. Then any alternative other than a would have to ¯ imply social welfare that exceeds that of the status quo by at least z if it is to be preferred over the status quo. Suppose q satisﬁes the characterization in Proposition 5. θi ) + kj uj (q(θi . This product is: ki ui (q(θi .10 that PAD is necessary for implementability. The bias is described by the function F .26) ki j=i Here. Then there are transfer rules (t1 . . The weight of agent i’s utility under the utilitarian welfare criterion is ki . (5.13. . We now show that Proposition 5. Agent i’s utility when she is type θi . and set F (¯) = z > 0 and F (a) = 0 ¯ a for all a ∈ A with a = z. . we have omitted all terms not depending on agent i’s report.12 is therefore to show the following: Proposition 5. . . θ−i ). in comparison to the VCG formula. this result says that under the assumptions of the result a decision rule satisﬁes PAD if and only if it maximizes a weighted utilitarian welfare criterion with exogenous bias. Those terms don’t aﬀect the argument. . θ−i ).12 implies Proposition 5. We can deﬁne agent i’s transfer payment as follows: 1 ti (θ) = − kj uj (q(θ). θ−i ).11 from Proposition 5.27) ki j=i Maximizing this expression is equivalent to maximizing the product of this expression and ki . θ−i ).12. she reports being type θi .11. θ−i )) (5. θj ) + F (q(θi . θ−i )) .3 and 5. (5. θj ) + F (q(θi . . tN ) is dominant strategy incentive compatible. Proof. All that is needed to derive Proposition 5. t1 .28) = j=1 kj uj (q(θi . θ−i ). . tN ) such that (q. For example. t2 . It is obvious from Propositions 5.CHAPTER 5. θi ) + j=i n kj uj (q(θi . θj ) + F (q(θi . θj ) + F (q(θ)) .

. Therefore. and we assume that ai is lowest ranked in A in the ordering Pi that reﬂects agent i’s order of the elements of A. θ−i ) ≥ ui (ai . A direct mechanism is “ex post individually rational with respect to (a1 . θi ). individual rationality required that if agents report truthfully their types they are at least as well oﬀ as they would be if they obtained some particular alternative in A and had to pay nothing. i. θi ) − ti (θi .” To generalize. a2 . reporting θi truthfully is optimal. . θ−i ). assume that ai is the lowest element of A in the order Ri . Suppose that the type set Θi is one-dimensional with respect to Ri . θi ). θi Ri θi for all θi ∈ Θi such that θi = θi . Proposition 5. DOMINANCE: GENERAL THEORY 116 and hence agent i chooses his report to maximize the same function that q maximizes. In the public goods example it was “no production of the public good. Deﬁnition 5. Consider any agent i ∈ I. we shall assume that for every agent i some alternative ai ∈ A is given that is as good as agent i’s outside option. aN )” if for all i ∈ I and all θ ∈ Θ we have: ui (q(θ). .” In the bilateral trade example it was “no trade. . Then a dominant strategy incentive compatible mechanism satisﬁes the ex post individually rationality constraint for agent i if and only if for every θ−i ∈ Θ−i : ui (q(θi . i.CHAPTER 5. bRi ai for every b ∈ A with b = a. Assume that there is a type θi that is the lowest type in the order Ri in Θi . . Finally. individual rationality was true for all types if and only if it was true for the lowest (in the case of the seller in the bilateral trade model: the highest) type.10 Individual Rationality and Budget Balance We now enrich our framework to bring in individual rationality and budget balance. and let Ri be an order of A. In the auction example. We begin with individual rationality. In the examples in previous parts of these notes.14.14. and thus that determines his individual rationality constraint. θi ) − ti (θ) ≥ ui (ai . for incentive compatible mechanisms. To obtain a result of this kind more generally.e.e. In the examples that we have seen in earlier parts of these notes. we impose the single crossing condition. the alternative in A that corresponded to individual rationality was “not obtaining the good”. 5. For every agent i let ai ∈ A.

θi ) − ui (ai . θi ) − ti (θi . which is ex post individual rationality for type θi . i=1 We now investigate conditions under which eﬃcient decision rules can be implemented with a balanced budget.31) (5. moreover. θi ) − ti (θi . θ−i ). it is relevant to ask when VCG mechanisms are budget balanced.14 is the individual rationality constraint for the lowest type. Combining the last two inequalities we obtain: ui (q(θi . If alternatives in A have costs associated with them. Because. θi ).30) (5. θ−i ). θi ) − ti (θi . θ−i )Pi ai . θ−i ) either satisﬁes: q(θi . (5.32) ti (θ) = 0. We ﬁrst note that: ui (q(θi . A direct mechanism is “ex post budget balanced” if for all θ ∈ Θ we have: N (5. θi ) − ui (ai . θ−i ). θ−i ). DOMINANCE: GENERAL THEORY 117 Proof. we can re-deﬁne outcomes so that some division.CHAPTER 5. Without loss of generality we deﬁne budget balance as the requirement that all transfer payments add up to zero. then by the single crossing condition: θi Ri θi . θ−i ) ≥ ui (ai . The condition in Proposition 5. or: q(θi . θ−i ) ≥ ui (ai . It shows that a necessary and suﬃcient condition is a restriction on the functional form of the welfare generated by an eﬃcient decision rule if this welfare is regarded as a function of the type vector. It is therefore necessary. θi ) ≥ ti (θi . θ−i ). θ−i ). θi ) ⇔ ui (q(θi . say equal division. To show that it is also suﬃcient.15. θ−i ).29) Now if θi = θi . θi ). θi ) − ti (θi . By dominant strategy incentive compatibility: ui (q(θi . θ−i ). θ−i ) ⇔ ui (q(θi . . We turn next to budget balance. θ−i ) ≥ ui (ai . θ−i ). θi ) ≥ ti (θi . Deﬁnition 5. of these costs is already included in outcomes. the alternative q(θi . The following Proposition provides the answer. θ−i )Ii ai we can infer: ui (q(θi . we show that it implies individual rationality for any type θi = θi . θi ) − ti (θi . θ−i ) ≥ ui (q(θi . Because under certain conditions VCG mechanisms are the only mechanisms that implement eﬃcient decision rules.

θj ) + τi (θ−i ) = 0 ⇔ N N (N − 1) i=1 N ui (q(θ). Suppose that N ui (q(θ). Proof. Consider a VCG mechanism. θi ) = i=1 i=1 fi (θ−i ) for all θ ∈ Θ. θi ) + (N − 1) i=1 f (θ−i ) (5. θi ).CHAPTER 5. The mechanism is budget balanced if for all θ ∈ Θ: N − i=1 j=i uj (q(θ). θi ) = i=1 N τi (θ−i ) ⇔ τi (θ−i ) N −1 ui (q(θ). (5. DOMINANCE: GENERAL THEORY 118 Proposition 5.35) − i=1 j=i uj (q(θ). if we set for every i ∈ I and θ−i ∈ Θ−i fi (θ−i ) ≡ τi (θ−i ) N −1 N i=1 ui (q(θ). We ﬁrst prove the necessity of this condition.4.34) we have obtained the desired form for the function Next we prove suﬃciency of the condition. . θi ) i=1 has the form described in the Proposition. θj ) + (N − 1)f (θ−i ) N N = − (N − 1) i=1 ui (q(θ). Then there exist a balanced budget VCG mechanism that implements q if and only if for every i ∈ I there is a function fi : Θ−i → R such that: N N ui (q(θ). Let q be an eﬃcient decision rule.15.33) Hence.36) which is zero by the assumption of the Proposition. θi ) = i=1 i=1 (5. Then for every θ ∈ Θ the sum of agents’ payments is: N (5. For every i ∈ I and every θ−i ∈ Θ−i we consider the VCG mechanism with τi (θ−i ) ≡ (N − 1)fi (θ−i ). and let τi : Θi → R be the functions referred to in Deﬁnition 5.

Note that the impossibility result that follows from Proposition 5.1 holds. θB } = fB (θS ) + fS (θB ).37) we ﬁnd that fS (θ ) = fS (θ). θB }. But then the ﬁrst right hand side in (5. This is because any such mechanism has to be a VCG mechanism by Corollary 5.15 gives not only necessary and suﬃcient conditions for the existence of a budget balanced VCG mechanism. In some of the proofs we have worked with Groves mechanisms and have argued that all such mechanisms. which contradicts the assumption that the ﬁrst right hand side equals θ whereas all later right hand sides equal θ .CHAPTER 5. I am not aware of any result that yields this conclusion for a more general class of models. Now suppose that the intersection of the ¯ ¯ intervals [θS . There is therefore no budget balanced VCG mechanism in this example. This observation applies to the bilateral trade model. have a zero budget balance. The condition of the above Proposition is that there are functions fB and fS such that max{θS . θS ] and [θB . or a negative budget balance. θ } = fB (θ) + fS (θ ) θ = max{θ . θ} = fB (θ ) + fS (θ) θ = max{θ . and suppose that θ.12 which shows that only ﬁxed price mechanisms are dominant strategy incentive compatible. .37) has to be the same as the three other right hand sides. then Proposition 5.37) we ﬁnd that fB (θ ) = fB (θ). θ are two diﬀerent elements of that interior with θ < θ . if they are individually rational. DOMINANCE: GENERAL THEORY 119 As an application of this result we consider the bilateral trade model. and subtracting the second from the fourth equality in (5. θ} = fB (θ) + fS (θ) θ = max{θ. It is therefore not an implication of Proposition 4. In that model the maximized welfare is given by: max{θS . ex post budget balanced and individually rational. θ } = fB (θ ) + fS (θ ) (5. Then the necessary and suﬃcient condition in Proposition 5.1.15 requires: θ = max{θ.15 for the bilateral trade model does not rely on an individual rationality constraint.37) Subtracting the third from the fourth equality in (5. but necessary and suﬃcient conditions for the existence of any budget balanced dominant strategy incentive compatible mechanism that implements eﬃcient decisions. θB ] has a non-empty interior. in all states of the world. Whenever Corollary 5.

The concept of “positive association of diﬀerences”. DOMINANCE: GENERAL THEORY 120 5. The discussion of the case that outcomes are lotteries is also taken from Rochet’s 1987 paper. The result on cyclical monotonicity is from o Rochet (1987). the concept of cyclical monotonicity is originally due to Rockafellar (1970). However. I have taken Proposition 5. The uniqueness of VCG mechanisms in the sense of Corollary 1 was shown in Green and Laﬀont (1977) and Holmstr¨m (1979). p. Finally. The classic papers on Vickrey-Clarke-Groves mechanisms are Clarke (1971). where this result is attributed to Bengt Holmstr¨m’s 1977 Stanford PhD o thesis. 54.15 from Milgrom (2004). and its characterization are in Roberts (1979). Groves (1973) and Vickrey (1961). 5.CHAPTER 5. The result on the suﬃciency of weak monotonicity is due to Bikhchandani et al.12 Problems . (2006).11 Remarks on the Literature The payoﬀ equivalence result that we presented is taken from Krishna and Maenner (2001).

This chapter will correspondingly have two parts. For example. it seems plausible that types are often not independent. In practice. will build on the previous chapter. This motivates the study of a model of Bayesian mechanism design in which types are not independent. dealing with more general sets of alternatives. and it does not seem plausible in practice. we need to consider more general distributions of types. This result is surprising. will be longer.1 Introduction To generalize our earlier results on Bayesian mechanism design. and will therefore be short and in parts informal. dealing with correlated types. and the nature of the analysis will be quite diﬀerent from what we have seen before. First. and preferences. types. if one agent has private information that makes the agent value the object sold in an auction particularly highly. we need to take two steps. then this agent might think that it is likely that other agents also have information that makes them value the object highly. The results in the second part of this chapter will indicate that with dependent types almost all combinations of decision rules and transfer rules can be implemented by a Bayesian incentive compatible direct mechanism. The ﬁrst part. 6. we need to consider more general sets of alternatives and a more general speciﬁcation of types and preferences. as in the last chapter. Second. The second part.Chapter 6 Bayesian Mechanism Design: General Theory . The 121 .

. . θi ) − ti . and we can restrict attention to direct mechanisms. A direct mechanism (q. For this. that maps every type vector into a collective decision. The set of all θ−i is Θ−i which is the cartesian product of the sets Θj . BAYESIAN DESIGN: GENERAL THEORY 122 result is best viewed as a paradox. We shall employ similar notation as before: The set of possible types of agent i is Θi . and mappings ti : Θ → R. for concreteness.CHAPTER 6. as products of intervals in ﬁnite dimensional Euclidean space. Deﬁnition 6. . . Here. The sets Θi and Θ can. . ΘN . We write µi for the marginal probability on Θi . but we want to be able to deﬁne a probability measure on them. . 6. We assume that the mechanism designer proposes to the agents a game and a Bayesian equilibrium of that game. . N }. We assume that there is a common prior distribution µ on Θ that is shared by the agents and the principal. They have to choose an alternative a out of some set A of mutually exclusive alternatives. We denote by θ the vector of types: (θ1 . in principle. θN ). . leaving out Θi . t1 .5 below we shall begin a discussion of what this missing ingredient might be.2 Set-Up There are N agents. In Section 6. It clearly indicates that some ingredient is missing in the model. 2. . θi is agent i’s type. and we write µ(· | θi ) for the conditional probability distribution of θ−i given θi . that indicate for each type vector the transfer that agent i needs to make. . Agent i’s utility if alternative a is chosen. and agent i pays transfer ti is: ui (a. The set of agents is denoted by I = {1. . . The revelation principle applies. Finally. one for each player i ∈ I. θ2 . we can think of them as abstract measurable spaces.1. be abstract sets. . or as ﬁnite sets. or. The set of all possible type vectors is Θ ≡ Θ1 × Θ2 × . Note that µ is not only the common prior of the agents but it is also the mechanism designer’s belief. tN ) consists of a mapping q : Θ → A. . . we write θ−i for a vector θ of types if we leave out agent i’s type. t2 .

. As in the previous chapter we use the formulation of Krishna and Maenner (2001). 6. θ−i )dµ(θ−i | θi ). We also write Ti (θi ) for agent i’s interim expected transfer payment conditional on agent i’s type being θi : Ti (θi ) ≡ Θ−i ti (θi . then Qi (θi ) assigns to A the probability Qi (A | θi ) ≡ Θ−i IA dµ(θ−i | θi ) (6. θi )dQi (θi ) − Ti (θi ) ≥ A A ui (a.1) where IA : Θ−i → {0. θi )dQi (θi ) − Ti (θi ). such as individual rationality and budget balance. Our ﬁst concern will be with the characterization of Bayesian incentive compatible decision rules.1 in Chapter 5.CHAPTER 6. 1} is the indicator function that assigns 1 to a type vector θ−i if and only if q(θi .2) If a direct mechanism is derived from a Bayesian equilibrium of some indirect mechanism. all i ∈ I and all θi ∈ Θi we have: ui (a. (6. tN ) is “Bayesian incentive compatible” if for all θ ∈ Θ. t1 . BAYESIAN DESIGN: GENERAL THEORY We call q the “decision rule” and the functions ti the “payment rules. t2 .” 123 For a given direct mechanism we shall write Qi (θi ) for agent i’s interim probability distribution on A conditional on agent i’s type being θi . We make the same assumptions as in the context of Proposition 5. θ−i ) ∈ A. . . now adapted to the Bayesian setting. Deﬁnition 6. if A is a measurable subset of A. In that case we call the direct mechanism Bayesian incentive compatible. That is. The result then says that the interim expected payment rules are uniquely determined up to a constant by the interim decision rules. .3 Independent Types We begin with a statement of the revenue equivalence theorem. then truth-telling will be a Bayesian equilibrium of the direct mechanism. A direct mechanism (q. .2. We shall later bring in additional considerations. The logic behind the result below is the same as the logic behind all other revenue equivalence results presented in these notes.

assume that for every i ∈ I the function ui (a.2. (θi . . Moreover. A ﬁrst point to note is that all dominant strategy incentive compatible mechanisms are also Bayesian incentive compatible. . not at the ex post level. and therefore VCG mechanisms are Bayesian incentive compatible mechanisms implementing eﬃcient decisions. θi ) is a convex function of θi . θi ) ∈ Θk .1. . tN ) be another Bayesian incentive compatible mechanism with interim decision rules Qi (θi ) and interim expected payments Ti (θi ) for every i ∈ I and every θi . Agent i evaluates alternatives by calculating their expected utility. To translate these results. . we can translate the results of the previous chapter into our current setting. . The decision rule Qi assigns to every type θi of player i a probability distribution Qi (θi ) over A. . eﬃcient or not. Suppose that (q. t2 . Suppose that for every i ∈ I the set Θi is a convex subset of a ﬁnite dimensional Euclidean space. θi )dQi (θi ) ≤ 0. for every sequence of length k ∈ N of 1 1 2 k k types of agent i. Then for every agent i ∈ I there is a number τi ∈ R such that Ti (θi ) = Ti (θi ) + τi for the same i ∈ I and every θi ∈ Θi . that is. Proposition 6. These characterizations now hold at the interim. . . for every i ∈ I.4 to our setting. we have: i k−1 κ=1 A κ+1 κ ui (a. . t1 . θi )dQi (θi ) − A κ κ ui (a. . To obtain characterizations of all incentive compatible decision rules. . θi . BAYESIAN DESIGN: GENERAL THEORY 124 Proposition 6. agent i’s payment rule is given by Ti . . Suppose that Qi (θi ) = Qi (θi ) for some i ∈ I and every θi ∈ Θi . . t2 . .CHAPTER 6. We now focus on decision rules q that can be implemented in Bayesian incentive compatible direct mechanisms. with θi = θi . we treat the set of all probability distributions over A as the set of alternatives. A decision rule q is part of a Bayesian incentive compatible direct mechanism (q. Finally. As an example we adapt Proposition 5. . t1 . tN ) is a Bayesian incentive compatible mechanism with interim decision rules Qi (θi ) and interim expected payments Ti (θi ) for every i ∈ I and every θi . t2 . . tN ) if and only if q is interim cyclically monotone. Let (q . . t1 .

t1 . Θ i=1 The proof of the equivalence of ex ante and ex post budget balance. For second best considerations in settings in which eﬃcient rules cannot be implemented. then it follows that there is no Bayesian incentive compatible. either a zero surplus. . Turning to individual rationality and budget balance.6. it again seems that the result on individual rationality with single dimensional type spaces. A direct mechanism (q. has no interesting generalization. and ex post budget balanced mechanism either provided that the revenue equivalence result Proposition 6. Combining individual rationality and budget balance.12. BAYESIAN DESIGN: GENERAL THEORY 125 In Chapter 5 we considered as a special case that the type space is single dimensional. in all states of the world. Proposition 5. Thus. and it generalizes in a straightforward way. Proposition 3. . Therefore. . . The only cases where this is the case seems to be cases where A consists of only two alternatives.3. . that we gave in Chapter 3 directly applies here. or a deﬁcit. Deﬁnition 6. the existence of an ex post budget balanced dominant strategy mechanism is. there do not seem to be more general results here that go signiﬁcantly beyond what we showed in Chapter 3. Proposition 5.14. In this special case monotonicity of the decision rule q was shown to be necessary and suﬃcient for the existence of a dominant strategy incentive compatible mechanism that implements q.CHAPTER 6. For convenience we repeat the result. t2 . the equivalence between ex ante and ex post budget balance that we used in Chapter 3 is useful.15 provides a suﬃcient condition for the existence of an eﬃcient Bayesian incentive compatible mechanism. A translation of these results into our setting requires that for every agent i ∈ I the type set Θi is single dimensional where the set of alternatives is the set of all probability distributions over Q.7 and 3. individually rational. This is the logic that we used to prove Propositions 3. Concerning budget balance. if all individually rational and ex post budget balanced mechanisms have. of course. tN ) is “ex ante budget balanced” if N ti (θ) dµ(θ) = 0.1 holds. suﬃcient for the existence of a budget balanced Bayesian incentive compatible mechanism.

Observe that the deﬁnitions and results of Section 6.3.4. to which we have referred as the “revenue equivalence” result. we shall assume that each of the sets Θi is ﬁnite. under some technical assumptions. We restrict attention to Bayesian incentive compatible direct mechanisms. We also assume that every θ ∈ Θ has positive probability: µ(θ) > 0. BAYESIAN DESIGN: GENERAL THEORY 126 Proposition 6.2. For every direct mechanism with decision rule q that is ex ante budget balanced there is an equivalent direct mechanism with the same decision rule q that is ex post budget balanced.CHAPTER 6. it might mislead us when we consider the question what is true generically. 6. We shall return to this point below. in principle.4. θi ∈ Θi agent i’s expected transfers. let us look back for a mo- . where results were more easily obtained with continuous type spaces rather than discrete type spaces. These are deﬁned as in Section 6. Throughout this part of the notes. innocuous. To understand why this is the case. 6. Here. two mechanisms are “equivalent” if the have the same decision rule. although they are also true with inﬁnite type spaces. conditional on agent i’s type being θi and agent i reporting to be type θi . and if for all agents i ∈ I and for all types θi . This result. The results that we discuss in this part of the notes are more easily proved with ﬁnite type spaces. the decision rule q determines the interim expected transfer rules Ti for each player i ∈ I up to a constant.1 Correlated Types Framework We now turn to the case that the distribution µ reﬂects some correlation among types. and the problem of identifying in this set those mechanisms that are optimal by some criterion. too. and which cases are exceptional cases.2 Failure of Revenue Equivalence In the setup with independent types we have obtained that.4 6. Note that this is in contrast with earlier sections. then the revenue equivalence result is no longer true. Although the ﬁniteness assumption is thus. is the same in the two mechanisms. This. simpliﬁes the exposition. and therefore we can use them here. When types are not independent.2 were not restricted to the case of independent types. greatly simpliﬁed the problem of describing the set of all incentive compatible mechanisms.

We need to provide a joint characterization of incentive compatible choice rules and transfer rules. as we change agent i’s type. the agent’s expected transfer payment does not change. provided we keep the report ﬁxed. Why is an agent’s expected transfer not aﬀected by a change in the agent’s type if we take the agent’s type report as given and ﬁxed? There are two reason for this is. BAYESIAN DESIGN: GENERAL THEORY 127 ment at the independent types framework. As we change an agent’s type. even if we keep the type report ﬁxed. The second reason is that all types have the same conditional probability distribution over other agents’ type vectors. . but it does not aﬀect the agent’s valuation of expected transfers.e. This is important because it implies that as we change the agent’s type. enters into the derivative of agent i’s interim expected utility with respect to type. but not transfers. As a consequence. and not on the transfer rules. This point is crucial because it implies that agent i’s interim expected utility function is determined up to a constant by the choice rule. A surprising and general characterization is provided in the next subsection. the same choice rule may be incentive compatible in combination with transfer rules that diﬀer at the interim level by more than an additive constant. that the expected payments are determined up to a constant by the choice rule. but also the transfer rule. i. The key argument that allowed us to establish the payoﬀ equivalence result showed that the derivative of any agent i’s interim expected utility with respect to agent i’s type depends only on the choice rule. u(ti ) = ti .CHAPTER 6. enter into the derivative. for example. we may take as given and ﬁxed agent i’s type report. not only the decision rule. As the agent’s transfer payments may depend on other agents’ types. Therefore. only collective decisions. This then easily implies. the agent’s conditional probability distribution over other agents’ types changes. this last part of the argument is no longer valid. If types are not independent. A change in the agent’s type then aﬀects the agent’s valuation of collective decisions. The ﬁrst is that all types have the same utility function of money. The fact that the payoﬀ equivalence result is not valid implies that we cannot transform the problem of characterizing the incentive compatible direct mechanisms into the problem of characterizing all incentive compatible choice rules. How did we show that the derivative of interim expected utility with respect to an agent’s type depends only on the choice rule? The argument for this is as follows. By the envelope theorem. Thus. the agent’s expected transfer payment may change.

the weights of the convex combinations are denoted by λ(θi ). e Deﬁnition 6. as a vector with as many entries as Θ−i has elements.4. The Cr´mer-McLean condition is obviously violated if at least two of the e vectors that describe agent i’s conditional beliefs are identical. The Cr´mer-McLean condition is obviously satisﬁed if the rank of the cole lection of vectors that describe agent i’s conditional beliefs is equal to the number of agent i’s types. BAYESIAN DESIGN: GENERAL THEORY 128 6. θ−i )µ(θ−i | θi ) for all i ∈ I and θi ∈ Θi . The Cr´mer-McLean condition requires that none e of these vectors can be written as a convex combination of all the other vectors where the weights.CHAPTER 6. the Cr´mer-McLean condition rules out that agent i’s conditional beliefs are e independent of his type. and hence the vectors are linearly independent. 2. Consider any direct mechanism (q. the two mechanisms have the same interim expected payments: ti (θi . Thus. none of these vectors is contained in the convex hull of the other vectors. that is: 1. one for each of agent i’s type. The content of this condition is best understood if one thinks of agent i’s belief about the other agents’ types conditional on agent i’s type being θi .4.3 Characterizing Bayesian Incentive Compatibility The result that we present in this section relies on a condition regarding the distribution µ that we shall call the “Cr´mer-McLean condition” as it e originates in the work of Cr´mer and McLean (1988). Suppose that the distribution µ satisﬁes the Cr´mere McLean condition. . Under the Cr´mer-McLean condition we can obtain the following surprising e result: Proposition 6. t ) that is Bayesian incentive compatible. as is the case when types are independent. Then there is an equivalent direct mechanism (q. the two mechanisms have the same decision rule q. θ−i )µ(θ−i | θi ) = θ−i ∈Θ−i θ−i ∈Θ−i ti (θi . In the deﬁnition. t). The probability distribution µ satisﬁes the Cr´mer-McLean e condition if there are no i ∈ I. θi ∈ Θi and λi : Θi \ {θi } → R+ for which: µ(θ−i | θi ) = θi ∈Θi θi =θi λ(θi )µ(θ−i | θi ) for all θ−i ∈ Θ−i . Agent i’s conditional beliefs are described by a collection of vectors of this form. µ(· | θi ).4. Thus.

4 says that. one can adjust it so that it does become incentive compatible without changing either the allocation rule or the interim expected payments. although this particular direct mechanism is not necessarily Bayesian incentive compatible. This is a very permissive result. If the auctioneer could implement this mechanism. and that all other agents pay nothing. will remain unchanged in the transformed mechanism. such as other agents’ actions. As an application of Proposition 6. for example. Agents earn no information rents. he would clearly obtain the largest possible revenue that he could extract from the agents provided that he respects the agents’ individual rationality constraint. and hence the auctioneer can extract the entire surplus from trade.CHAPTER 6. An additional complication arises from the fact that the incentive scheme for belief revelation needs to be such that truthful reports of beliefs generate expected payments of zero. of course also the auctioneer’s expected revenue is not altered. Moreover. This will ensure that the interim expected payments. Put diﬀerently. use such incentive schemes to elicit subjects’ beliefs about uncertain events. While incentive mechanisms for truthful belief revelation always exist.4 is to add to the transfer schemes of the original mechanism (q.4 consider the single unit auction environment. it is easy to see that in such incentive schemes the costs for false reports of one’s beliefs can be made arbitrarily large. BAYESIAN DESIGN: GENERAL THEORY 129 In words. It is well-known that such incentive schemes exist for risk-neutral subjects. . and let the payment rule be that the winner of the object pays his true valuation. whereas false reports of beliefs generate negative expected payments. the auctioneer can achieve the same expected revenue as he can if he directly observes agents’ valuations of the object. The idea of the proof of Proposition 6. Let the decision rule be that the object is allocated to one of the agents with the highest valuation of the agent. Proposition 6. this result says that every direct mechanism can be made Bayesian incentive compatible without altering the decision rule or interim expected payments. an agent who truthfully reveals his beliefs also truthfully reveals his type. By making the incentives for truthful revelation of beliefs very large we can undo all possible incentives to lie about one’s type in the original mechanism (q. Experimentalists. the Cr´mer-McLean condition is used to ensure that the expected e payments from truth-telling can be set equal to zero. If interim expected payments are not altered. under truthful revelation of types. As no two types have the same beliefs. t). t) a transfer scheme that provides agents with incentives to truthfully reveal their beliefs.

θ−i ) + c(k − ln µ(θ−i | θi )) where k and c > 0 are constants. π .CHAPTER 6. and we announce that if ω ∈ Ω occurs. Here. .5) The ﬁrst order condition for maximizing this is: ∂L π (ω) ˆ =− +λ=0⇔ ∂π(ω) π(ω) π (ω) ˆ = λ. has to be proportional to the true vector of probabilities. Which probabilities will subjects announce? Suppose a subject is riskneutral. only one reported ˆ vector π that his property. Thus.e. ˆ Suppose we added to the transfer rule in the mechanism (q. then the incentive to report the probabilities truthfully will override all other incentives that agent i might have. k is a constant. i. ω∈Ω (6. the subject has to pay to the experimenter a comparatively large amount. BAYESIAN DESIGN: GENERAL THEORY 130 How can truthful revelation of beliefs be induced? A very simple incentive scheme is as follows. In choosing ˆ its report π the subject will minimize: (k − ln(π(ω))ˆ (ω) π ω∈Ω (6. itself. i.e. if an ω ∈ Ω occurs to which the subject has assigned an extremely small probability. for example. π : Ω → [0.6) Thus. and if the agents other than i report type θ−i . π(ω). If we choose c suﬃciently large. the subject makes a payment to the experimenter of k − ln(π(ω)). namely the true vector of probabilities. π (ω). 1]. We ask the subject to announce its probabilities π for the elements of Ω. ˆ (6.4) A Lagrange function is: L= (k − ln(π(ω)))ˆ (ω) + λ π ω∈Ω ω∈Ω π (ω) − 1 . of course. the reported vector of probabilities. and the mechanism will be Bayesian incentive compatible. then agent i has to pay ti (θi . if agent i reports to be type θi . Suppose the ﬁnite set of possible outcomes is Ω.3) subject to π(ω) = 1. Let π be the subjects true subjective probabilities.4 payments that follow the above payment scheme. π(ω) (6. and we want to elicit an experimental subject’s belief about the probabilities of diﬀerent elements of Ω. There is. t) in Proposition 6.

we take some type θi of agent i as given and ﬁxed. and conditional on all other types he expects to make a strictly positive payment. and all other agents report to be type θ−i . times a positive constant. We deduce that the second condition holds. we construct a transfer scheme for agent i with the desired properties. . and thus his type. agent i expects a zero payment. On the other hand. The ﬁrst condition says that agent i’s expected payment. he expects to receive a payment. Franklin (1980). Either the equation Ax = b or (exclusive) y T A ≥ 0. that incentives to report true beliefs are strict. We can proceed in a similar way for all agents i. The mechanism that we have described obviously alters agents’ interim payments. t). We take b to be the the column vector of agent i’s beliefs if agent i is type θi . conditional on the type θi . agent i’s incentives to report his beliefs. the Cr´mere McLean condition is suﬃcient to allow us to provide strict incentives for truthful revelation of beliefs using a mechanism where an agents’ expected payments. Let A be an n × m matrix. but not quite the mechanism that we will use in the proof of Proposition 6.5. yT b < 0 has a solution y.4. and that expected payments conditional on truth-telling are zero. Putting together all the transfer vectors y that we obtain in this way. conditional on any type other than type θ−i .56: Proposition 6. We can add this transfer scheme. which has #Ω−i entries. The column vector y. represents agent i’s payments if agent i reports type θi . we can achieve that conditional on type θi . By subtracting a constant from all payments. and we take A to be a matrix of dimensions #Θ−i × (#Θi − 1) where each column of A represents agent i’s conditional beliefs if agent i is of one of the types other than θi . override all other incentives that agent i might have in (q. However. The “Cr´mer-McLean condition” says e that the ﬁrst of the two conditions in Farkas’ Alternative is not satisﬁed. This completes the proof of Proposition 6. are exactly zero. agent i’s expected payments is negative. has a solution x ≥ 0 To apply Farkas’ Alternative. This can be deduced from the Cr´mer-McLean condition using Farkas’ Alternae tive. We can repeat the construction of the previous paragraph for each type θi of agent i. p. t). if the constant is suﬃciently large.4. and. is non-negative. that is. BAYESIAN DESIGN: GENERAL THEORY 131 What we have just described is almost. truthfully. if the agent reports his beliefs truthfully. to agent i’s transfers in our original mechanisms (q.CHAPTER 6.

the auctioneer would like the winner of the object to pay her reservation price. In the decision rule.0) (1.2: Decision Rule 6.1) (1. Suppose that the auctioneer wants to allocate the object to the agent with the highest valuation. 1 1 2 3 1 (2. There are two agents.3.0) Figure 6. An agent’s type is the agent’s valuation for the good. Moreover.2 and 6.1) (0. 1) 2 2 (1. We consider a single unit auction example. Example 6. The joint probability distribution of types is shown in Figure 6. BAYESIAN DESIGN: GENERAL THEORY 1 1 2 3 4 20 2 20 1 20 132 2 2 20 2 20 2 20 3 1 20 2 20 4 20 Figure 6. and the second entry indicates the probability with which player 2 obtains the object. 1) 2 2 (0.0) 3 (0. 3}.1. Each of the two agents has three types: Θ1 = Θ2 = {1.1. the . rows correspond to types of player 1.1: Joint probability distribution of types. where ties are broken randomly. In this ﬁgure.1) (1. This corresponds to the allocation and payment rules shown in Figures 6. the ﬁrst entry in each box indicates the probability with which agent 1 obtains the object. and columns correspond to types of player 2. In the payment rule.4. that is. 1) 2 2 (1. 2. her type.CHAPTER 6.4 A Numerical Example The numerical example in this Subsection underlines the general message of this section that one should view the Cr´mer McLean result as a paradox e rather than a guidance to the construction of mechanisms that could work in practice.

1) (3. BAYESIAN DESIGN: GENERAL THEORY 1 1 2 3 (1.4.2) (1.0) 2 (0.CHAPTER 6. and negative numbers are payments to player 1.3) (0. In Figure 6. we start by constructing transfer rules that give each player strict incentives to reveal their true beliefs about the other player’s type. That will reduce the probability with which they get the object. and that makes the decision rule incentive compatible.4. and the second entry shows the payment by player 2. For player 1 such a transfer rule is indicated in Figure 6. and construct a payment rule that leaves interim payments unchanged. but at least it will give them a positive surplus if they obtain the object. 1) 2 2 (2.0) 3 (0.4: Belief revelation ﬁrst entry in each box shows the payment by player 1. We now want to demonstrate Proposition 1 for this example. Types 2 and 3 of each player have an incentive to pretend that they are lower types. The idea of the payment rule is that agent 1 is rewarded if the type that she reports is the same as that of player 2. but she has to make a payment if the type that she reports is diﬀerent from . 2) 2 133 Figure 6.0) (3.3: Payment Rule 1 1 2 3 -1 1 2 2 1 -2 1 3 2 1 -1 Figure 6. The rule that we have described is clearly not incentive compatible. and that give expected utility of zero to each player if they reveal their beliefs truthfully.3. As explained in Subsection 6. positive numbers are payments by player 1.3) 3 (3. We omit the simple check that this rule has the required properties.4.

2 . Note that individual rationality only holds at the interim. One can now verify that the auction that has the allocation rule indicated in Figure 6. If we multiply the transfer payments in Figure 6.5 extracts the full surplus for the auctioneer.4 to the transfers in Figure 6.3. Deﬁnition 6. -5) (6.3. A direct mechanism (q.3.5. This reﬂects that agents’ types are positively correlated in this example. − 2 ) 2 134 2 (3. t2 .3.5) (-5. One can calculate that 3 is the smallest integer by which we can multiply all payments in Figure 6. We begin by deﬁning individual rationality. . then also the . BAYESIAN DESIGN: GENERAL THEORY 1 1 2 3 5 (− 5 . whenever the mechanism with which we started is individually rational.2 and 6.3) 3 (6.4 and obtain a transfer rule that eliminates all incentives to lie in the mechanism of Figures 6.2 and the transfer rules of Figure 6.3) (9. − 3 ) 2 (5.4. .3 we get the payments shown in Figure 6.CHAPTER 6.4. of course.6) 3 (. . and is incentive compatible and individually rational. not at the ex post level. tN ) is “individually rational” if for every agent i and every type θi ∈ Θi we have: ui (a.5.5 Individual Rationality and Budget Balance We will want to combine Bayesian incentive compatibility with other requirements. θi )dQi (θi ) − Ti (θi ) ≥ 0 A In the transformation that was described in Subsection 6.4. A symmetric rule can. We need to overcome all incentives to deviate in the mechanism of Figures 6.6) Figure 6. We shall now add a positive multiple of the transfers in Figure 6.2 and 6.9) (3.5: Modiﬁed transfer payments the one that player 2 reports. t1 .4 by 3 and add them to the transfer payments in Figure 6. such as individual rationality and ex post budget balance. . be used for player 2. 6. We need to determine by how much we need to multiply the transfers in Figure 6.

2. Consider any direct mechanism (q. Thus. . Suppose that the distribution µ satisﬁes the Cr´mere McLean and the identiﬁability condition.” Deﬁnition 6.6. θ−i )µ(θ−i | θi ) = θ−i ∈Θ−i θ−i ∈Θ−i ti (θi . tN ) that is individually rational and ex ante budget balanced. . the two mechanisms have the same interim expected payments: ti (θi . . that is: 1. That result therefore does not straightforwardly generalize to the context of correlated types. the two mechanisms have the same decision rule q. t2 . . any pair of decision and transfer rules that are individually rational can be made incentive compatible and individually rational provided the Cr´mer McLean condition e holds. guarantees that budget balance can always be achieved. . θ−i )µ(θ−i | θi ) for all i ∈ I and θi ∈ Θi . this condition says that for any alternative distribution ν of types there is at least one agent and one type of that agent such that this agent cannot randomize over reports in a way that makes the conditional distribution of all other types under ν indistinguishable from the conditional distribution of all other types µ. The additional condition is called “identiﬁability. . . t2 .6. We next consider ex post budget balance. tN ) that is equivalent. Then there is a Bayesian incentive compatible and ex post budget balanced mechanism (q. Kosenok and Severinov (2008) prove: Proposition 6. t1 . However.CHAPTER 6. The probability distribution µ satisﬁes the identiﬁability condition if for all distributions ν = µ such that ν(θ) > 0 for all θ ∈ Θ there is at least one agent i and one type θi ∈ Θi such that for any collection of nonnegative coeﬃcients λθi we have: θi ∈Θi ν(θ−i | θi ) = θi ∈Θi λθi µ(θ−i | θi ) for at least one θ−i ∈ Θ−i . Kosenok and Severinov (2008) have shown that an additional condition for the prior distribution of types that goes beyond the Cr´mer e McLean condition. The construction that demonstrates the equivalence of ex ante and ex post budget balance and proves Proposition 6.3 requires that types are independent. t1 . BAYESIAN DESIGN: GENERAL THEORY 135 transformed mechanism is individually rational. Intuitively. . .

This intuition has been formalized by Robert (1991). The fact that in a belief revelation scheme the reports of all agents except agent i determine how much agent i has to pay suggests that there may be an incentive to collude. and we omit a discussion of e this construction.CHAPTER 6. However. Heifetz and Neeman. Thus. Neeman (2004) and Heifetz and Neeman (2006) have shown that information structures with this property are in some sense rare among all conceivable common prior information structures. in a recent series of papers. They point out that the construction of Cr´mer and McLean e requires that there is a one-to-one relation between an agent’s preferences and an agent’s beliefs about other types. types are close to being independent. it seems plausible that the results of this literature generalize to other settings. . as illustrated by our numerical example. the same distributions are. have pursued a diﬀerent line of argument. as Heifetz and Neeman show. Much of this discussion has focused on the particular case of a single unit auction in which the decision and transfer rules to be implemented are the full extraction rules. Then a belief revelation scheme that provides incentives strong enough to outdo all incentives to lie will require large payments by agents. 6. such a mechanism might be impossible. i. and it will expose agents to signiﬁcant risk. or agents might require compensation for the risk that reduces the attractiveness of the mechanism to the mechanism designer.5 Discussion A number of authors have investigated the question how the set-up presented by Cr´mer and McLean needs to be changed to obtain less paradoxical e results. if agents are either liquidity constrained or risk averse.e. a very small subset of the inﬁnite dimensional space of general information structures. Laﬀont and Martimort (2000) have shown that in environments in which types are close to independent collusion might prevent the mechanism designer from full surplus extraction. Suppose that in the Cr´mer and McLean setting the prior distribution of e types is close to a product measure. BAYESIAN DESIGN: GENERAL THEORY 136 Kosenok and Severinov’s construction that proves this result is quite diﬀerent from Cr´mer and McLean’s construction. While it is often true that for ﬁxed ﬁnite type space generic probability distributions will satisfy Cr´mer and McLean’s condie tions. This precisely is the reason why a belief extraction mechanism can help with implementation.

For each buyer i the valuation vi is a random variable that only buyer i observes. Each entry indicates the probability with which v1 takes the value indicated in the row.25+ε Figure 6. The seller values the object at zero.25+ε 0. . and all that you need to follow the proof is a basic knowledge of separating hyperplane theorems.25-ε v2 = 2 0. The joint distribution of v1 and v2 is given by the following table: v2 = 1 v1 = 3 v1 = 4 0. 2}. ε is a constant that satisﬁes 0 < ε < 0. The random variable v1 takes only two possible values: v1 ∈ {3. and his utility equals −ti if he does not obtain the good and pays ti . 2. 6. and that implies that. My favorite e reference for Farkas’ Lemma.6 Remarks on the Literature Proposition 6.25. (a) Verify that the Cr´mer McLean condition holds for buyer 1.4 is adapted from Cr´mer and McLean (1988).6 is a combination of Theorem 1 and Corollary 1 in Kosenok and Severinov (2008). is Franklin (1980). 4}. The distribution indicated in this table is common knowledge among the buyers and the seller.7 Problems A seller has a single indivisible object to sell. that is central to the proof of Proposition 1. e (b) Construct a payment scheme that provides buyer 1 with incentives for truthful revelation of his beliefs about buyer 2’s type. Note that the two buyers are not symmetric. Proposition 6. The random variable v2 takes only two possible values: v2 ∈ {1.25-ε 0.CHAPTER 6. but that the other buyer and the seller don’t observe. Franklin’s proof is not much more than half a page long. BAYESIAN DESIGN: GENERAL THEORY 137 6. Buyer i’s von Neumann Morgenstern utility equals vi − ti if he obtains the good and pays ti . There are two potential buyers: i = 1. and seeks to maximize his expected revenue. and at the same time v2 takes the value indicated in the column.6: Distribution of Valuations Here.

his expected transfer payment conditional on each of his types will be zero. (c) Use the payment scheme that you have found in part (b) to construct a Bayesian incentive compatible selling mechanism that always allocates the object to buyer 1. BAYESIAN DESIGN: GENERAL THEORY 138 if he truthfully reveals his beliefs. and that implies that buyer 1’s expected payment to the seller equals his expected valuation of the object.CHAPTER 6. (d) Investigate the limit for ε → 0 of the transfer payments in the selling mechanism that you obtained in part (c). . The selling mechanism that you construct should also oﬀer each type of each agent an interim expected utility of at least zero in equilibrium.

139 . no particular assumption is made about agents’ preferences over these decisions. It is easy to imagine situations in which the assumption is not satisﬁed. Finally. such as voting. in a more general model. This is a very restrictive assumption. Each collective decision is interpreted as a decision about all issues relevant to agents. One implication of the assumption of additively separable utility and risk neutrality is that monetary payments that are made by one agent and received by another agent are welfare neutral as long as we allocate the same welfare weight to all agents. and moreover that all agents are riskneutral in money. and. We shall consider situations in which there is some arbitrary set of possible collective decisions. we shall move in this chapter to the opposite extreme of the assumption that we have made so far.” On the specter of possible assumptions. at least initially. welfare depends not only on the allocative decision but also on the distribution of money among agents. it might be the case that we are considering situations.1 Introduction So far we have assumed that all agents’ utility is additively separable in an allocative decision. and money. including the allocation of money. The simplest case is that agents are not risk neutral in money. By contrast.Chapter 7 Non-Transferrable Utility 7. Another way in which the assumption might be violated is that there are interactions between money and the allocative decision. We shall refer to the case that we have considered so far as the case of “transferrable utility” and we shall refer to the case considered in this chapter as the case of “non-transferrable utility. in which monetary payments are typically not invoked to provide incentives.

” The set of all linear orders over A is denoted by R.2. Later. . In the literature direct mechanisms in the sense of deﬁnition 7. The mechanism designer can thus construct an extensive game with outcomes in A. . . We shall oﬀer a discussion of only an eclectic selection from these results. . but who determines the rules of the strategic interaction among the agents by which an alternative from A is chosen. has a dominant strategy in the sense in which we used this phrase in earlier chapters of these notes. Deﬁnition 7.” and we read “aPi b” as “a is strictly preferred to b. . The revelation principle then applies. 7. . and the only indiﬀerence is among identical elements.” Deﬁnition 7. and we write R−i for the list of all agents’ preference relations leaving out agent i’s: R−i = (R1 . We assume that Ri is a linear order. . . . . Ri+1 . and we can restrict attention to direct mechanisms. that is.CHAPTER 7. . 2. .2. i ∈ I = {1. reﬂexive. R2 . . R−i )Ri f (Ri .1 The Gibbard Satterthwaite Theorem Set Up There is a ﬁnite set of agents. A direct mechanism f is “dominant strategy incentive compatible” if for every agent i ∈ I and all preference relations Ri . RN ). We denote the strict order derived from Ri by Pi . There is a vast number of theoretical results in this area. Ri in R: f (Ri .1.2 7. we shall then brieﬂy discuss some work on Bayesian incentive compatibility without transferrable utility. transitive. We shall start with this part of the literature. RN ). Ri−1 . with every conceivable preference relation in R. R2 . it is complete. NON-TRANSFERRABLE UTILITY 140 The emphasis of theoretical work in this area has been on dominant strategy incentive compatibility. We write R for the list of all agents’ preference relations: R = (R1 . . N }. These agents have to choose one alternative from a ﬁnite set A of mutually exclusive alternatives.” To maintain consistency with earlier parts of these notes. We consider a mechanism designer who does not know the agents’ preference relations. we shall speak of “direct mechanisms. A “direct mechanism” is a function f : RN → A. R−i ) . We read “aRi b” as “a is weakly preferred to b.1 are sometimes also called “social choice functions. Each agent i has a preference relation Ri over A. . . We study in this section the case in which the mechanism designer seeks to construct a game such that every player.

7. Before we proceed we make two further remarks about the set up described here. Suppose that A has at least three elements. as suggested in the previous paragraph. The second point that needs emphasis is that the domain of our direct mechanism contains all linear orders of A.1. of course.3. and if the candidates in A have the opposite ideology from candidates ˆ in A \ A. Physical limitations might well force us to choose from a ﬁnite set of probability distributions. In practice we might have some knowledge about agents’ preferences over A. but ﬁrst we explore the implications of assuming an unrestricted domain. But. The reader might notice that we have not introduced probability distributions over the set A. and can therefore restrict attention to mechanisms that perform as we want them to perform only for some. and the reader can think of A as the set of all these distributions. NON-TRANSFERRABLE UTILITY 141 In the literature dominant strategy incentive compatibility is sometimes also referred to as “strategy proofness. elements of the complement A \ A. but stochastic. This would seem plausible if the elements of A are candidates for a political ˆ oﬃce. Natural restrictions may be that ˆ all agents rank either all elements of some subset A of A higher than all ˆ or that they have the opposite ranking. If we have this case in mind. and that the range of f is A.2 Statement of the Result and Outline of the Proof Deﬁnition 7. Thus. then it might seem more natural to have an inﬁnite set A rather than a ﬁnite set. A direct mechanism f is dominant strategy incentive compatible if and only if it is dictatorial. the elements of set A could be outcomes that are not deterministic.” In this section we shall characterize direct mechanisms that are dominant strategy incentive compatible. But it is not immediately obvious that in practice an inﬁnite number of probability distributions can be implemented. The ﬁrst concerns the role of randomization in this section. We shall consider later in this Chapter the implications of some such domain restrictions. our setup does not really rule out randomization.CHAPTER 7. A direct mechanism f is called “dictatorial” if there is some individual i ∈ I such that for all R ∈ RN : f (R)Ri a for all a ∈ A. Proposition 7. Another natural restriction may be that agents’ preferences are of the von Neumann Morgenstern form. but not for all preference vectors.2. . This would seem natural if A consists of lotteries.

with some arbitrary tie breaking rule. For example. Suppose also f (R ) = b = a. We also report some simple implications of monotonicity. giving each agent the opportunity to vote for one of the two alternatives. The assumption that A has at least three elements is important. We shall proceed in two steps. In later parts of this Chapter.” The Gibbard-Satterthwaite theorem is a paradoxical result. We can therefore analyze the situation as if such alternatives didn’t exist. that is. It is due to Gibbard (1973) and Satterthwaite (1975). We therefore focus on proving the “necessity part. and therefore agents’ preferences over these alternatives cannot inﬂuence the outcome. First. direct mechanism is monotone. is dominant strategy incentive compatible. and is called the “Gibbard Satterthwaite Theorem.CHAPTER 7. In that case many mechanisms are dominant strategy incentive compatible. Formally: f (R) = a and for all i ∈ I : [aRi b for all b ∈ A such that aRi b] ⇒ f (R ) = a.4. the case that A has two elements. Suppose ﬁrst in Deﬁnition 7. The “suﬃciency part” of Proposition 7. It shows that the requirements that the requirements of dominant strategy incentive compatibility and unlimited domain together are too strong.4 preference proﬁles R and R diﬀered only in the i-th component. In this result the assumption that the range of f equals A rather than being a strict subset of A is immaterial. we show that every dominant strategy incentive compatible. If the range of f is a strict subset of A. The core of the proof is then in the next subsection.2. A direct mechanism f is “monotone” if. whenever f (R) = a. then it is monotone. where we show in a second step that every monotone direct mechanism is dictatorial. Consider the only remaining nontrivial case. Proof. Because with . and then choosing the alternative with the highest number of votes. and clearly not dictatorial.1 is obvious. Without this assumption the result is not true. then f (R ) = a. NON-TRANSFERRABLE UTILITY 142 This is one of the most celebrated results in the theory of mechanism design. we shall discuss weaker requirements that lead to more positive results. Deﬁnition 7. and for every agent i and every alternative b the preference relation Ri ranks a above b if Ri does. we can re-deﬁne the set of alternatives to be the range of f .” Our presentation of this proof is based on Reny (2001). If f is dominant strategy incentive compatible. Proposition 7. Alternatives that are not in the range of f will never be chosen.

Deﬁnition 7. the choice must remain a. A direct mechanism f “respects unanimity” if.CHAPTER 7. By monotonicity. Proof.5. Now re-order alternatives below a in arbitrary ways. then f (R ) ∈ B: f (R) ∈ B and [aRi a ⇔ aRi a whenever a ∈ B or a ∈ B] ⇒ f (R ) ∈ B. preference relation Ri diﬀers from preference relation Ri only regarding the ranking of elements of B. .3 Every monotone direct mechanism is dictatorial Proposition 7. This contradicts dominant strategy incentive compatibility. / which contradicts f (R) ∈ B. and that the range of f is A. agent i has an incentive to report preference Ri . there is some R such that f (R) = a. whenever an alternative a is at the top of every individual’s preference relation. Again. Now raise a to the top of everyone’s ranking. then we apply the above argument successively to each component in which R and R diﬀer. If f is monotone. Because the ranking of a does not fall as we move from Ri to Ri .4.5. / / Proposition 7. Thus we have arrived at a contradiction. Suppose that A has at least three elements. then it is dictatorial. 7. Proof. then a is chosen by f : aRi b for all i ∈ I and b ∈ A ⇒ f (R) = a. Proposition 7. If f is monotone and the range of f is the A. Because the range of f is A. But then. we arrive at the conclusion f (R) = a. Now we introduce two simple implications of monotonicity. this implies: aRi b. Assume f (R ) = a ∈ B. Consider any a ∈ A. If R and R diﬀer in more than one component. then f respects unanimity. we have to have: aRi b. with true preference Ri . whenever f (R) ∈ B for some B ⊆ A.6. and for every agent i. and we can conclude f (R ) = a. NON-TRANSFERRABLE UTILITY 143 preference Ri it is a dominant strategy for agent i to report Ri truthfully. A direct mechanism f is “set-monotone” if. Then monotonicity implies that f (R) = a. This proves that f respects unanimity. Deﬁnition 7.2. Again. by monotonicity.3. the social choice remains a. If f is monotone then f is set-monotone.

.” i. there is a “dictator for a. c b a . We shall show that for every alternative a ∈ A.. Otherwise. . In the proﬁle in Figure 7. . NON-TRANSFERRABLE UTILITY R1 . but all other agents rank a bottom. .. .. Rn+1 .1 by changing preferences without moving any alternative above a.1 agent n ranks a top. if the dictator for a had a at the top of her ranking. Figure 7. . it is suﬃcient to ﬁnd one preference proﬁle where a is at the top of some agent i’s ranking.. . then the dictator must be the same individual i for every alternative. but at the bottom of everybody else’s ranking. . . RN .1 shows a speciﬁc ranking. but for the moment this is of no relevance to the argument.. . . .. One such proﬁle is shown in Figure 7. Finding one such proﬁle in which the social choice is a is suﬃcient because every other proﬁle in which agent n ranks a at the top can be obtained from the one shown in Figure 7. .. and where the social choice is a. Now ﬁx an alternative a ∈ A. .. . . c b a Rn a c b . In Figure 7.1: Social choice is a Proof. and therefore by monotonicity the social choice for every other proﬁle in which agent n ranks a at the top is a. . and therefore this agent i is a dictator. .2 the social choice has to be a because f . If there is such a dictator for every alternative a. each column corresponds to one player’s ranking of alternatives in A. . the outcome would not be well-deﬁned... ... there is an agent i such that whenever a is at the top of i’s ranking. . . Rn−1 . and the dictator for b(= a) had b at the top of his ranking.2. . . then a is chosen. To show that there is a dictator for a.e. We conclude that the dictator must be the same agent i for every alternative.. . with the highest ranked alternative at the top.CHAPTER 7. . We shall arrive at the conclusion that the social choice in Figure 7. Figure 7. c b a . and in subsequent ﬁgures. ...1.1 has to be a through a sequence of steps starting with the proﬁle shown in Figure 7. involving alternatives other than a. . In this ﬁgure. c b a 144 .. . ..

Figure 7.. . by monotonicity. If the social choice is b in Figure 7.4. Then. Rn−1 a a .4: Social choice is a or b . . R1 b a . .. b Rn+1 a a . b Rn a a ..... Figure 7. b 145 .. .. RN a a . . b}) implies that the social choice is either a or b.. .. to the top of agent 1’s ranking.... ... .. Rn−1 a ... . . . . We will now identify an agent n of whom we shall show that he or she is a dictator for n.. . as shown in Figure 7. . Now suppose that we move b one step further up. . . . Rn−1 a a .... . . .4... Now suppose that we move b up in agent 1’s ranking. .. . b Rn a .3... b . until it is just below a.. b Rn+1 a .. .3: Social choice is a respects unanimity. . . Figure 7. .. NON-TRANSFERRABLE UTILITY R1 a . b . . . .. b Rn a a .CHAPTER 7.... b Rn+1 a a . . .. then we set n = 1. b . . b .2: Social choice is a R1 a b . . . . .. . as shown in Figure 7. b . . . .. .. RN a .. .. . . Then set-monotonicity as deﬁned above (setting B = {a. . b . the social choice has to remain a. RN a a .

Rn b a . .. b . . For Figure 7. then it would also have to be b in Figure 7..5 and 7.. .6 except that we have moved alternative a to the bottom of the ranking for agents i < n. . .. . then we repeat the same procedure for agent 2.. etc. Rn a b . . But if the alternative chosen were b.7 has to be either a or b.. RN a . if the social choice remains a in Figure 7. Figure 7.6 and monotonicity that the choice has to be b..8 we can conclude from set monotonicity (with B ≡ {a. There has to be one such agent because after we have worked our way through all agents we arrive at a proﬁle where all agents put alternative b at the top of their ranking. . .. b .8 show the same proﬁles as Figures 7. Rn+1 a . . The ﬁrst agent for whom the social choice switches from a to b is the agent whom we identify as our candidate dictator n.. Figure 7.. ..4. .. by monotonicity. the choice has to be . Figures 7. .CHAPTER 7.. and we have moved them to the second position from the bottom for agents i > n.. . Comparing Figure 7. NON-TRANSFERRABLE UTILITY R1 b a .. . . . . . b .6: Social choice is b However...6 we show the generic situation for agent n before and after the social choice switches from a to b. . . Rn+1 a . Rn−1 b a ..5: Social choice is a R1 b a . b}) that the social choice in Figure 7..8 it follows from Figure 7. .5. RN a . In Figures 7.5 and 7.. b 146 .6. . and f respects unanimity so that the social choice is b. . . We now argue that in these proﬁles the social choices have to be the same as in Figures 7. Rn−1 b a . . .... .7. .. .5 and 7..7 and Figure 7. . . .. Therefore. . .. . . . . Now consider the proﬁle in Figure 7.7 and 7. ..

.7. . . 7. Rn−1 b . setting B = {a. . .1 has to be a or b. . then we could move alternative c to the top of everyone’s preferences. .. . Now consider the preferences in Figure 7..1 has to be a... This concludes the proof. . . . . The position of alternative a has not changed relative to other alternatives in comparison to Figure 7. Rn−1 b ... . But if the choice were b in Figure 7....3 Dominant Strategy Incentive Compatibility On Restricted Domains Two natural ways of relaxing the stringent requirements of Proposition 7.. .. which would contradict that f respects unanimity. The choice in Figure 7.. ... .1.. NON-TRANSFERRABLE UTILITY R1 b .. . . Figure 7. the choice in Figure 7.. RN ... compare Figure 7. Figure 7. .7. RN ..CHAPTER 7.. a Rn a b . a Rn b a . Rn+1 . a b . .1 so that more positive results obtain are ﬁrstly to consider a more restricted . ... . . a .. Therefore. a b . a b 147 . Rn+1 .. .. and therefore the social choice has to be a. . b}. . .. . . . by set monotonicity. a b .1. . .8: Social choice is b a in Figure 7. Finally.. and by monotonicity the choice would still have to be b. .7: Social choice is a R1 b . a . .9 to Figure 7.9. . . .

. The deﬁnition of this set depends on how the alternatives have been labeled with numbers. . The restricted domain of single-peaked ˆ preferences is then RN . that is: if ˆ ≥ k then Ri + 1 and if ≤ k then − 1Ri . A preference relation Ri of agent i is called “single-peaked” if there is a k(i) ∈ {1. c a b .. . c b a . . . 2.. .9: Social choice is a domain of preferences. Rn−1 .. .CHAPTER 7. .. . .. Alternatives are candidates for some political position. . .. . thus. It then appears plausible that preferences decline monotonically as candidates are further away from agent i’s ideal position. Figure 7. . Candidates are labeled according to their position on the “leftright” spectrum. . The domain of single-peaked preferences may appear natural in a the following environment. . . . Direct mechanisms are methods for selecting one candidate out of a set of candidates. c b a Rn a c b . The best known restriction on the domain of preferences that allows dominant strategy incentive compatible mechanisms that are not dictatorial is single-peakedness. 2.. .. . . don’t reﬂect the dependency of the set ˆ R on the labeling in our notation. . c a b 148 . . .. . Rn+1 . for simplicity. the set of preferences for any individual i that we are ˆ considering is no longer the complete set R. and the alternative k(i) reﬂects voter i’s ideal position on this spectrum. Suppose the alternatives in A are labeled with the integers 1. but some subset R of R. . . .. and (ii) agent i’s preferences decline monotonically “to the left” and “to the right” of k(i).. .. . In this section we shall discuss the former approach whereas in the next section we discuss the latter approach. RN . . In this Section. NON-TRANSFERRABLE UTILITY R1 . .. ... and. K} such that (i) agent i prefers alternative k(i) to all other alternatives: k(i)Ri j for all j ∈ A. . . . . K.. . Denote by R the set of all single-peaked preferences. and secondly to consider a less demanding solution concept than dominant strategies. We keep this labeling ﬁxed in this section. .

in either case agent i. A particularly interesting domain restriction might be natural in the case in which the set A consists of all lotteries over some given ﬁnite set of outcomes. is chosen.6. k(2). Then we can use the same rule as in the case that N is odd. k(N ) is dominant strategy incentive compatible. and can therefore be represented by Bernoulli utility functions. Other restricted domains on which dominant strategy incentive compatibility does not imply dictatorships are. by reporting her preferences truthfully. then there are dominant strategy direct mechanisms that are not dictatorial. Denote these by m− and m+ . . In those environments we assumed that agents’ utility functions were of a particular form. k(2). independent of what agent i reports. . for example. ensures that her most preferred alternative from the range of possible medians is chosen. This domain restriction lead to the dominant strategy incentive compatibility of simple rules such as the ﬁxed price rule for bilateral trade. If agent i’s most preferred alternative is lower than m− . Now suppose that N is even. . One might hope that this domain restriction . then agent i can ensure by reporting his preferences truthfully that his most preferred alternative is chosen. For every agent i denote by k(i) the alternative that is ranked top according to Ri . and take the other agents’ reported preferences R−i as given and ﬁxed. NON-TRANSFERRABLE UTILITY 149 Proposition 7. . k(N ) will then be between the N -th highest and the ( N +1)2 2 th highest index of the most preferred alternative. her most preferred alternative from the range of possible medians. If agent i’s most preferred alternative is higher than m+ . . her most preferred alternative from the range of possible medians. The median will be these two numbers. The rule f that picks the alternative am where m is the median of the vector (k(1).CHAPTER 7. In this case one might assume that individuals’ preferences satisfy the von Neumann Morgenstern axioms. The same argument as above proves that this rule is dominant strategy incentive compatible. then by reporting her preferences truthfully agent i ensures that alternative m− . and it is obviously non-dictatorial. agent i can ensure that alternative m+ . . We arbitrarily pick some alternative a ∈ A and pretend that some agent had expressed a preference that lists a at the top. the rule is dominant strategy incentive compatible. those that we studied in Chapter 4. then by reporting her preferences truthfully. If the alternative that agent i most prefers is between these two numbers. If preferences are single-peaked. additive separability and risk neutrality. The median of (k(1). Consider any agent i. with the following small modiﬁcation. of course. Proof. Thus. . is chosen. Thus. . Suppose N is odd.

even though the two agents have opposing preferences. and three alternatives: A = {a. The focus of B¨rgers and Postl (2005) is o on the question whether there is an incentive compatible mechanism that implements the alternative B whenever it is eﬃcient. b. 1] with the same cumulative distribution function F with density f where f (θi ) > 0 for all θi ∈ [0. θ2 ) of agents’ types into a probability distribution over the set of alternatives A. the only direct mechanisms that respect unanimity and are dominant strategy incentive compatible are random dictatorships. A and C. 7. and the mechanism designer observes neither of the two types. B¨rgers and Postl o (2005) assume that θ1 and θ2 are stochastically independent variables that are both distributed on the interval [0. a theorem due to Hylland (1980) shows that in this setting. Agent 1’s Bernoulli utility of alternative a is 1. This is a natural notion of eﬃciency because the sum of utilities from the compromise B is θ1 + θ2 . b. To induce agents to correctly reveal their type the mechanism designer can choose the probabilities of the individuals’ favorites. agent 1’s privately observed type. judiciously. Agent 2 is analogous. the alternative is picked that maximizes this individual’s preference relation. A direct mechanism maps the vector (θ1 . a.4 Bayesian Incentive Compatibility Although much of the literature on mechanism design with non-transferrable utility has focused on dominant strategy incentive compatibility there seems to be no fundamental reason why not to consider Bayesian incentive compatibility as well. without further domain restrictions. There are two agents. and her Bernoulli utility of alternative c is 0. that is. c}. whenever θ1 + θ2 = 1. i. B¨rgers and Postl (2005) interpret the alternative B as a “compromise” o as both agents rank B as their middle alternative. These probabilities can be used as an instrument for providing incentives because they are welfare neutral: Both alternatives . In random dictatorships each of the N agents is assigned a probability. whereas alternatives A and C both yield a sum of utilities of 1. Each agent i observes θi but not θj where j = i. and is then picked with this probability as the dictator. It is commonly known that agent 1 ranks the alternatives as a. b. 1]. but her Bernoulli utility of the intermediate alternative b is θ1 . To explore this approach B¨rgers and Postl (2005) consider o a stylized model of compromising.CHAPTER 7. c. N = 2. However.e. NON-TRANSFERRABLE UTILITY 150 allows the implementation of social choice functions other than dictatorial social choice functions. except that agent 2 ranks alternatives in the opposite order c.

An important aspect of the Gibbard-Satterthwaite theorem that we have not mentioned in the text is that. the Bayesian equivalent of the problem that Gibbard and Satterthwaite studied. to determine the second best. and that it might beneﬁt from numerical investigation.5 Remarks on the Literature The large literature on the Gibbard-Satterthwaite theorem is surveyed by Barbera (2001). the theorem is equivalent to Arrow’s famous theorem on the impossibility of preference aggregation Arrow (1963). i. NON-TRANSFERRABLE UTILITY yield exactly the same amount of welfare. that is. probability of their preferred alternative). An important diﬀerence with the public goods models that we considered in Section 3.CHAPTER 7. Barbera describes several diﬀerent approaches to the proof of the theorem. B¨rgers and Postl (2005) use arguments similar to our arguments in Section o 3. The cost-function is one-toone: for each unit of the public good (i.3 to show that no Bayesian incentive compatible mechanism implements the ﬁrst best.3.3.e. choice of the compromise with probability 1 if and only if θ1 + θ2 ≥ 1. B¨rgers and Postl o (2005)’s diﬃculties in analyzing the much simpler model indicate that the more general problem is analytically hard.e. as Satterthwaite (1975) pointed out. and therefore agents cannot give up more than sum upper bound of probability of their preferred alternative. The complement these results with a numerical investigation of second best mechanisms. probability of the compromise) produced some agent has to give up one unit of money (i. one cannot use the approach that was explained in Section 3. 7. . The problem of ﬁnding a second best is complicated by the fact that agents face liquidity constraints. The compromise problem in B¨rgers and Postl (2005) is a sandbox model o for the more general model of determining an optimal voting scheme when there are three or more alternatives.3 is that agents are liquidity constrained: probabilities have to be between zero and one. One can think of the “compromise” as a public good for which both agents have to give up probability of their preferred alternative. B¨rgers and Postl (2005) have analytical results only for small parametrized o classes of direct mechanisms.e. Therefore. 151 B¨rgers and Postl (2005) note that their framework is equivalent to a public o goods model similar to the one we considered in Section 3.

6 and shows that all dominant strategy incentive compatible direct mechanisms on a single-peaked domain need to be based on a generalized version of the median decision rule. 7. Other papers that investigate Bayesian incentive compatibility without transferrable utility include Abdulkadiroglu and Loertscher (2007). That paper also contains a detailed discussion of the relation between Hylland’s result and Gibbard’s earlier theorems on random dictatorship (Gibbard (1977). discussed in Section 2.CHAPTER 7.6 Problems . Our discussion of the case of Hylland’s result is based on Dutta et al. Moulin (1980) obtains a converse to Proposition 7. Gibbard (1978)). This literature is less focused on incentive compatibility than the mechanism design literature. however.1 of Roth (2008). (2007) who also oﬀer an independent proof of the result. NON-TRANSFERRABLE UTILITY 152 Our treatment of the case of single-peaked preferences is based on Barbera (2001) although the original result is in Moulin (1980). The literature on the design of matching markets Roth (2008) is also in a setting without transferrable utility. In fact. Strategic properties of matching algorithms are.

The private information about market conditions that each bidder holds is likely to aﬀect not only that bidder’s valuation of licenses but also other bidders’ valuations of licenses. in auctions for licenses to use radio spectrum to operate a telephone service. we refer to the case that we have addressed so far as the case of “private values. Each member’s evaluation of a candidate will depend not only on the member’s own information about the candidate.Chapter 8 Interdependent Types 8. Most of our discussion will focus on Bayesian incentive compatibility. diﬀerent committee members will have diﬀerent pieces of information about the abilities of the candidates.1 Introduction In all previous parts of these notes we have assumed that each agent’s private information is all that matters for that agent’s preferences over group choices. which is related to dominant strategy implementation. but also on the information of other agents. Similarly. We shall. When choosing a new chair of a committee. In this section we shall present some elements of a formal analysis of mechanism design with informational interdependence.” By contrast. We shall refer to this as the case of “informational interdependence. we shall instead consider the case in which an agent’s preference over group choices depends not only on this agent’s own private information. but also on other members’ information about the candidate. each bidder will have some private information about the likely proﬁtability of various services. however. 153 . In this section. also comment on ex post implementation.” One can think of many real world examples in which informational interdependence seems important.

1]2 . b}. 8. and agent i pays transfer ti is: θi − ti . The focus of this section will be on the case in which signal spaces are subsets of ﬁnite-dimensional Euclidean spaces where the dimension is larger than one.2 An Example We start oﬀ with an extremely simple example. and two alternatives: A = {a. these results are paradoxical. Recall that in the case of private values the answer to this question was positive. However.CHAPTER 8.5−ti . Agent i’s utility if alternative a is chosen. θ2 ) ∈ [0. and independent of the dimensions of the signal spaces. The role that the dimensions plays in the case of informational interdependence is. We shall focus here on the case that these signals are independent. The dimensions of the agents’ signal space will play an important role in this part of the lecture notes. or whether dependence is allowed. Agent 1 observes a two-dimensional private signal (θ1 . quite diﬀerent from the role that it plays in the case of private values. crucial for the question whether eﬃcient decision rules can be part of a Bayesian incentive compatible mechanism. . an important distinction is whether agents’ signals are assumed to be independent random variables.4 can be generalized. and it seems likely that some modiﬁcation of the setting in which these results are obtained is needed in order to obtain more plausible results. With informational interdependence the dimensions of the signal spaces is. I = {1. For the case that signals are dependent. 1]2 . for example. For this case we explain a series of impossibility results that were obtained by Philippe Jehiel and Benny Moldovanu in a sequence of papers. θ2 ) is a random variable with a distribution that has support [0. Note the essential feature of this example: agent 1’s private signal is relevant not only to his own utility but also to agent 2’s utility. Agent i’s utility if alternative b is chosen and agent i pays transfer ti is: 0. however. Eﬃcient decision rules could be implemented using Vickrey-Clarke-Groves (VCG) mechanisms. as in the context of that Section. This will make it easier to contrast the results of this section with those of previous sections. the permissive results obtained in Section 6. Agent 2 has no private information. For the case of informational independence. We assume that (θ1 . as for the case of private values. 2}. INTERDEPENDENT TYPES 154 We shall maintain the assumption of transferrable utility that we made in all previous parts of these notes except Chapter 7. Suppose that there are 2 agents.

observe that we require that t∗ (θ1 . we restrict attention to direct mechanisms. the “ﬁrst best” decision rule q ∗ satisﬁes: q ∗ (θ1 . a and b. for all realizations (θ1 . only transfers to be paid by agent 1 matter. and sometimes the second option. The result that we observed for this example is not very surprising. agent 1 ﬁnds it in his interest to truthfully report the realization of the signal. θ2 ) is constant for all (θ1 . θ2 ) such that q ∗ (θ1 . Depending on which θ2 he actually observes. no two real numbers ta and tb have this property. 1] maps agent 1’s private signal into the probability with which alternative a is chosen for the given signal. Let us ask a simple question: Can we ﬁnd transfers that make welfare maximizing decisions incentive compatible? Here. Denote the constant payment by tb . θ2 ) = 1.1) This has to be true for all θ1 ∈ (0. 1]2 → R maps agent 1’s private signal into a transfer to be paid by agent 1 for the given signal. Because his utility does not depend on θ2 . θ2 ) = 1 if θ1 + θ2 > 1. A direct mechanism is “incentive compatible” if. This is sometimes called the “taxation principle.CHAPTER 8. θ2 ) = 0. More specifically. A decision rule q : [0. θ2 ) = 0.5. because otherwise agent 1 would distort his report and only report that (θ1 . Obviously. Denote the constant payment by ta . and q ∗ (θ1 . 1). we deﬁne welfare maximization in the usual way as maximizing the sum of the agents’ utilities.” Agent 1 is implicitly oﬀered the choice between two alternatives. We have no instrument that would allow us to elicit this information from the agent. We conclude that it is impossible to implement the ﬁrst best. 16 .5 − tb ⇔ ta − tb = θ1 − 0. 1]2 → [0. 1) agent 1 can report a θ2 such that q(θ1 . with associated taxes ta and tb . θ2 ) = 1. we require that t∗ (θ1 . 1]2 of his private signal. θ2 ) is constant for all (θ1 . we sometimes want him to choose the ﬁrst. we would like the group decision to condition on a component of an agent’s private signal that has no implications for that agent’s payoﬀ. INTERDEPENDENT TYPES 155 By the revelation principle. θ2 ) such that q ∗ (θ1 . (8. A transfer rule t : [0. he must be indiﬀerent between the two choices: θ1 − ta = 0.16 Now observe that for every θ1 ∈ (0. Obviously. θ2 ) ∈ [0. θ2 ) = 0 if θ1 + θ2 < 1. Our question is: Can we ﬁnd a transfer rule t∗ so that the direct mechanism (q ∗ . Similarly. Intuitively. t∗ ) is incentive compatible? As a ﬁrst step. and he can also report a θ2 such that q(θ1 . θ2 ) for which t∗ is minimal.

Agent i’s von Neumann Morgenstern utility if alternative ak is chosen. 17 . .2) Thus. . Secondly. but for ease of exposition we restrict attention to the case that signals are contained in [0. they consider a model in which each agent makes private observations that are relevant to other agents utility but not to the agent’s own utility.3 Impossibility of Implementing Welfare Maximizing Decision Rules We assume that there are N agents. . .CHAPTER 8. The linear form of the utility function assumed above is not essential to the argument. . . . Our exposition below will focus on Jehiel/Moldovanu’s second extension of the example. θN ) a Jehiel and Moldovanu (2001) allow more general convex signal spaces with non-empty interior. θ2 . . 2. but potentially with diﬀerent weights. . θ2 . Each agent i observes i i i a K-dimensional signal: θi = (θ1 . . 1]IK → ∆(A) that assigns to each vector of private signals θ ≡ (θ1 . 1]K . Their argument for this more general case is slightly diﬀerent from the above argument. . . . a2 . j ∈ I and all k = 1. . . tN ) consists ﬁrstly of a mapping q : [0. . K. but makes the exposition easier. . The set of agents is denoted by I = {1. Diﬀerent agents’ signals are independent. . 17 The signal θi has a distribution with density f i that is positive everywhere. The proof is related to the argument in the above example. t1 . and if agent i has to pay transfer ti is: N j j αki θk − ti . Again they show the impossibility to implement ﬁrst best decision rules. A direct mechanism (q. N }. 8. θK ) ∈ [0. j=1 (8. . agent i’s utility from alternative ak is a linear function of the k − th component of all agents’ signals. aK } of K mutually exclusive alternatives. . Note that we do not restrict these factors to have the same sign for all agents. They show that it is in general impossible to implement a ﬁrst best decision rule. INTERDEPENDENT TYPES 156 Jehiel and Moldovanu (2001) have generalized the above example in two ways. The factor in front of the k −th component i of agent j’s signal in agent i’s utility function is: αkj . . They have to choose an alternative a out of some ﬁnite set A = {a1 . Firstly. 1]K . 2. . We assume that i αkj = 0 for all i. t2 . . they consider a model in which each agent makes observations that aﬀect the agent’s own utility as well as other agent’s utility. .

θj ) − τi (θ−i ) = j=1 uj (q(θi .CHAPTER 8. for actual types. θ−i ). A choice rule q ∗ is “ﬁrst best” if for every θ ∈ [0. Under this rule. 1]K → R that assigns to every signal θi of agent i the interim expected value of agent i’s transfer payment. the deﬁnition of the transfer rule needs to be modiﬁed. θj ) + τi (θ−i ) for all θ ∈ Θ. agent i will ﬁnd it in his interest to report his type truthfully. with informational interdependence. 1] that assigns to every signal θi of agent i the a interim probability that alternative a is chosen. if agent i is of type θi and reports that he is of type θi . A direct mechanism is incentive compatible if and only if for every agent ˜ i. The set of all such probability distributions is denoted by ∆(A). θ−i ). 1]K . INTERDEPENDENT TYPES 157 probability distribution q(θ) over A. θj ) − τi (θ−i ). θi ) + j=i N uj (q(θi . agent i’s expected utility from i is at least as large as his expected utility from reporting reporting type θ ˜ type θi if agent i’s true type is θi . his utility is: ui (q(θi . Given a direct mechanism we can deﬁne for every i ∈ I and every a ∈ A a function Qi : [0. For every agent i ∈ I a direct mechanism also speciﬁes a mapping ti : [0. θ−i ).4) Thus.3) where τi is an arbitrary function. 1]K → [0. We can also deﬁne for every i ∈ I a function T i : [0. θ−i ). if the collective decision is q(θi . Recall that in previous sections we deﬁned agent i’s transfer rule in a VCG mechanism by: ti (θ) = − j=i uj (q(θ). minus some constant that does not aﬀect agent i’s incentives. (8. because now any agent . Now. Because the ﬁrst best rule maximizes true social welfare. A direct mechanism is “ﬁrst best” if its choice rule is ﬁrst best. θi ∈ [0. (8. his payoﬀs will be true social welfare. 1]IK we have that q ∗ (θ) assigns positive probability only to alternatives that maximize the sum of agents’ utilities. A ﬁrst question to ask is whether we can use a VCG mechanism to implement a ﬁrst best decision rule. 1]IK → R that assigns to each vector θ of private signals a transfer ti (θ) to be paid by agent i. when agent i reports θi . in our context. and every pair of types θi .

Proposition 8. (θi . changes as agent i’s signal value for some arbitrary other alternative b changes. then his utility becomes: ui (q(θi . Assume also that there are some agent i ∈ I and two alternatives a. .CHAPTER 8. It can be much weakened. N i j=1 αaj Then no ﬁrst best direct mechanism is Bayesian incentive compatible. (8. This is a regularity condition that is needed in the proof of Proposition 8. a. conditioning on agent i’s signal. θ−i )) − τi (θ−i ). we might deﬁne: ti (θ) = − j=i uj (q(θ). θ−i ). INTERDEPENDENT TYPES 158 j’s utility does not only depend on agent j’s type. The ﬁrst condition in this Proposition assumes that the interim expected probability of alternative a. and. rather than according to their true utility. The following result. instead of requiring it to be true for all types of agent i. For example. θ−i ). This is because the mechanism perceives a diﬀerent utility for agents j = i if agent i changes his report. θ) + τi (θ−i ) for all θ ∈ Θ. shows that in fact no direct mechanism implements a ﬁrst best decision rule. Agent i is rewarded according to this perceived utility of agents j. their derivative is never zero: ∂Qi (θi ) a =0 i ∂θb for every i ∈ I. 1]K .1 as presented below.6) Note that agent i’s utility is no longer aligned with true welfare. Thus. b ∈ A and θi ∈ [0. Assume that any ﬁrst best rule q ∗ is such that interim expected probabilities are diﬀerentiable.5) Now.1. (8. if agent i is of type θi and reports that he is of type θi . b ∈ A such that: i αai = i αbi N i j=1 αbj . θ) + j=i uj (q(θi . which is the main result of Jehiel and Moldovanu (2001). it is evident from the proof below that we might also require it to be true for some type of agent i. The previous paragraph showed that a generalization of the VCG mechanism cannot be used to implement a ﬁrst best decision rule. moreover. but on all agents’ types.

7) Now suppose we diﬀerentiate this expression again. The proof in Jehiel and Moldovanu (2001) does not assume this. Note the contrast between Jehiel and Moldovanu’s impossibility result and the impossibility results that we presented earlier in private value settings.1 presents a generic impossibility result.8) If we change the order of diﬀerentiation. We shall present a proof that assumes that for every agent i the interim expected utility function Ui is twice continuously diﬀerentiable. . a i ∂θa (8. INTERDEPENDENT TYPES 159 The second condition says that there is at least some agent i. this time with respect i to θb for some b ∈ A. with no other requirements. and a pair of alternatives a. Suppose there were a ﬁrst best direct mechanism that is incentive compatible. Then we obtain: i i ∂2U i i ∂Qa (θ ) = αai . If we pick the weights in agents’ utility functions randomly from some continuous distribution. Proof. we obtain similarly: i i ∂2U i i ∂Qb (θ ) = αbi . It is for this reason that Proposition 8. The Envelope Theorem implies for every agent i and every a ∈ A: ∂U i i = αai Qi (θi ). i i i ∂θa ∂θb ∂θb (8. the Jehiel Moldovanu impossibility results refer only to welfare maximization. then this condition will be satisﬁed with probability 1.9) The fact that interim expected utility is twice continuously diﬀerentiable implies that the order of diﬀerentiation does not matter (Schwarz’s Theorem). such as individual rationality and budget balance. By contrast. Those earlier impossibility results only obtained when welfare maximization is combined with other requirements. such that the relative weight that agent i attaches to his signal for alternative a and his signal for alternative b is diﬀerent from the relative weight that the social welfare function attaches to these two signals.CHAPTER 8. b. i i i ∂θa ∂θb ∂θa (8.

without transfers. We might also investigate revenue . However. i αai i i ∂Qi (θi ) a i ∂Qb (θ ) = αbi ⇔ i i ∂θa ∂θb i αai = i αbi ∂Qi (θi ) b i ∂θa i (θ i ) ∂Qa i ∂θb 160 . as in the case of the impossibility results that we obtained in the private value settings with individual rationality and budget balance.1.11) We can now deduce from our results so far that: i αai = i αbi N i j=1 αaj .12) This has to be true for every i ∈ I. would also be Bayesian incentive compatible. 8. An argument like the one that we just displayed would lead to the conclusion that the equation just derived is true with the weights from agents’ original utility functions being replaced by the weights from the social welfare function: N i j=1 αaj N i j=1 αbj ∂Qi (θi ) b i ∂θa i (θ i ) ∂Qa i ∂θb = . by the ﬁrst assumption in Proposition 8. b ∈ A. N i αbj j=1 (8. Then the given rule. Suppose they were equal to social welfare. we could divide by the partial derivatives of the interim probabilities because.4 Characterizing All Incentive Compatible Mechanisms The impossibility result presented in the previous section is remarkable because it holds even if no individual rationality or budget balance. and any two alternatives a.CHAPTER 8. But this contradicts the second condition in Proposition 8. INTERDEPENDENT TYPES Thus the two derivatives that we computed have to be the same. it is natural to ask next which mechanisms would be second best if we wanted to maximize ex ante expected social welfare.10) Here.1. Now suppose agents’ utility functions were diﬀerent. (8. these derivatives are not zero. (8.

al.’s main ﬁnding is that for generic utility functions only constant choice rules that choose the same alternative for every type realization are incentive compatible if each agent’s signal space is at least two-dimensional. A question one might ask about this result is why dictatorial choice rules are not incentive compatible. the characterizations that we developed for the case of private values generalize. collective decision rules q can be part of a Bayesian incentive compatible mechanism if and only if at the interim level they generate for each agent the sub-gradient of a convex function. or mechanisms that are chosen according to some other objective function. and allowed that choices depend on beliefs. Jehiel et. Once this information is revealed. given his information. If we relaxed this requirement.’s result implies that such a rule is not ex-post incentive compatible. agent i has an ex post incentive to deviate. Moreover. If truth telling is an ex-post equilibrium of a direct mechanism. and are determined by the collective decision rule up to a constant. and to change his decision. it is a Bayesian equilibrium for every belief that agents might hold about the other agents’ types. (2006) study ex-post incentive compatibility instead of Bayesian incentive compatibility. ex-post incentive compatibility is a generalization of dominant strategy incentive compatibility to the case of interdependent valuations. With independent signals. Ex-post incentive compatibility means that for every realization of all other agents’ types each agent ﬁnds it in his interest to report his type truthfully rather than distort it. then interim expected utilities are convex. . Note that ex post incentive compatibility thus requires that we can uniquely predict an agent’s choice independent of what this agent’s beliefs about the other agent’s types are. if each agent’s utility from the collective decision is a linear function of the agent’s own type. Suppose some agent i is allowed to pick the collective choice based on i’s signal only. al. which is the case that we are considering here. truth telling is an ex post equilibrium if and only if truth telling is a dominant strategy. INTERDEPENDENT TYPES 161 maximizing mechanisms. and suppose there are no transfers. Jehiel et al. including beliefs that are not product measures. as it doesn’t aﬀect the collective decision. For all these questions it is important to have a characterization of all incentive compatible direct mechanisms. Jehiel et. Thus. Agent i will then pick the alternative that maximizes his expected utility. In particular. These characterizations are hard to use in practice. then more rules become feasible. With private values. The reason is as follows: all agents other than i would be willing to reveal their private information.CHAPTER 8.

’s proof of their result is as follows. both agents.5 Remarks on the Literature The example in Section 2 is a special case of an example that appears in Jehiel and Moldovanu (2001). 8. Suppose that these two sets have a boundary in common. INTERDEPENDENT TYPES 162 The logic of Jehiel et.CHAPTER 8. a and b. agent 1’s payment is determined by agent 2’s report. 8. But for generic utility functions it is impossible to ﬁnd surfaces such that for ﬁxed payments both agents are indiﬀerent between two alternatives along these surfaces. ﬁxing the other agent’s signal. This paper is also the main reference for Proposition 8. (2006).1 that I have given is taken from Jehiel and Moldovanu (2005).1. payments along the boundary are constant. al. the short proof of Proposition 8. along this boundary.6 Problems . I have also used Jehiel et al. However. and agent 2’s payment is determined by agent 1’s report. agent 1 and agent 2 have to be indiﬀerent between a and b. and consider for two agents i and j the set of signals where a is chosen. Moreover. and the set of signals where b is chosen. Consider any two alternatives. Along this boundary. Thus.

163 .Chapter 9 Robust Mechanism Design To be written.

164 .Chapter 10 Dynamic Mechanism Design To be written.

165 .Chapter 11 Conclusion To be written.

8: 19–33. 77 Tilman B¨rgers and Peter Postl. Social Choie and Individual Values. second edition. 151. 120 Patrick Bolton and Mathias Dewatripont. 2. 74:447–476. 120 166 . 112. Ron Lavi. Dynamic house allocations. 150. Review of Economic Studies. 152 William J. Public Choice. Clarke. Ahuva Mu’alem. 1976. 2007.Bibliography Atila Abdulkadiroglu and Simon Loertscher. 2005. Yellen. MIT Press. 152 Sushil Bikhchandani. 151 o Kim Sau Chung and Jeﬀrey C. 1963. Quarterly Journal of Economics. Ely. Foundations of dominant strategy mechanisms. Cambridge and London. Commodity bundling and the burden of monopoly. 2001. Noam Nisan. New York. A note on budget balance under interim o participation constraints: The case of independent types. Multipart pricing of public goods. Contract Theory. and Arunava Sen. 151 Salvador Barbera. 2008. 90:475–498. Shurojit Chatterji. 30 Kenneth Arrow. 2006. Social Choice and Welfare. An introduction to strategy-proof social choice functions. 2007. Econometrica. Weak monotonicity characterizes deterministic dominant strategy implementation. 2005. 1971. Adams and Janet L. Eﬃcient compromising. 74:1109–1132. 18:619–653. 30 Tilman B¨rgers and Peter Norman. Economic Theory. Wiley. 80 Edward H.

1979. Supplement 5. 4. 152 Alan Gibbard. 76 Joseph L. Journal of Economic Theory. 28:163–179. 1978. Lecture Notes for Economics. 45:727–738. 120 Werner G¨th and Martin Hellwig. u Journal of Economics. 96 Aviad Heifetz and Zvika Neeman. Characterization of satisfactory mechanisms for the revelation of preferences for public goods.berkeley. Methods of Mathematical Economics. 1993. 142 Alan Gibbard. Incentives and incomplete information. Journal of Public Economics. 136 Benjamin Hermalin. 45:665–681. pages 121–159. Manipulation of schemes that mix voting with chance. Doob. 120 Theodore Groves. MIT Press. 2. Game Theory. 2006. 1977. 2007. 1994. 77 Kathleen Hagerty and William Rogerson. Econometrica. Social Choice and Welfare. Econometrica. Measure Theory. 1980. 56:1247–1257.BIBLIOGRAPHY 167 Jacques Cr´mer and Richard McLean. 1986. 46:595–614. 30 . Econometrica. 33 Alan Gibbard. and Arunava Sen. URL http:// faculty. 11:25–45. 74:213–233. Straightforwardness of game forms with lotteries as outcomes. Econometrica. Incentives in teams. Full extraction of the surplus in e bayesian and dominant strategy auctions. 128. 1977. Hans Peters. 137 Claude d’Aspremont and Louis-Andre Gerard-Varet. 42:94–107. On the generic (im)possibility of full surplus extraction in mechanism design. 2005. 1973. Cambridge and London. Econometrica. 16 Bhaskar Dutta. Strategy-proof cardinal decision schemes. Robust trading mechanisms. Heidelberg and New York. The private supply of a public good. Springer. 41:617–631. Econometrica. 137 Drew Fudenberg and Jean Tirole. 1988. 152 Joel Franklin.haas. Econometrica. Springer. 152 Jerry Green and Jean-Jacques Laﬀont. 131.edu/hermalin/. Manipulation of voting schemes. 41:587–601. 1987. 1973.

158. 137:153–185. Journal of Economic Theory. Auction Theory. 76. Moritz Meyer ter Vehn. London. 140:126–161. budgetbalanced mechanisms and allocation of surplus. 2007. 74:585–611. Academic Press. 123 Jean-Jacques Laﬀont and David Martimort. Convex potentials with an application to mechanism design. 2004. Individually rational. Journal of Economic Theory. The limits of ex-post implementation. 2006. Eﬃcient design with interdependent valuations.. 161. Groves’ scheme on restricted domains. 2008. Econometrica. Dominant strategy implementation of bayesian incentive compatible allocation rules. 120 Dilip Mookherjee and Stefan Reichelstein. Econometric Society. Cambridge University Press. 42. Econometrica. San Diego etc. 1980. 150 Philippe Jehiel and Benny Moldovanu. 38. Econometrica. Mechanism design with collusion and correlation. 162 Grigory Kosenok and Sergei Severinov. New York etc. Strategy proofness of voting procedures with lotteries as outcomes and inﬁnite sets of strateges.. o 47:1137–1144. Cambridge etc. In Proceedings of hte 9th World Congress. 30 Paul Milgrom.. 56:378–399. Multidimensional mechanism design: Revenue maximization and the multiple-good monopoly. 159. 162 Philippe Jehiel. 99. John Wiley and Sons. and Bill Zame. Econometrica. 2002. Luenberger. Putting Auction Theory to Work. 135. 136 David G. 69:1237–1259. Vincent. 137 Vijay Krishna. Allocative and informational externalities in auctions and related mechanisms. Journal of Economic Theory. 76 Vijay Krishna and Eliot Maenner. 2001. 1992. 69:1113–1119. 68:309–342.. Benny Moldovanu. 156. 1979. Inc. 72 Alejandro M. 162 Philippe Jehiel and Benny Moldovanu.BIBLIOGRAPHY 168 Bengt Holmstr¨m. 96 . 2000. 2001. Optimization By Vector Space Methods. 120. Manelli and Daniel R. 56. 2005. Econometrica. 1969. 120 Aanund Hylland.

Economics Letters. 16: 191–200. 120 . 120 Jean-Charles Rochet. 120 Alvin Roth. 1991. 65. Satterthwaite. Myerson and Mark A. Princeton University Press. 136 Efe A. 151 William Vickrey. 70:99–105. Aggregation and Revelation of Preferences. practie. 36:537–569. 1981. 114. 136 Kevin Roberts. 120 Tyrell Rockafellar. 1970. Strategy-profness and arrow’s conditions: Existence and correspondence theorems for voting procedures and social welfare functions. 2001. Eﬃcient mechanisms for bilateral trading. The characterization of implementable choice rules.BIBLIOGRAPHY 169 Herve Moulin. On strategy-proofness and single peakedness. 35:437–455. 16:8–37. 16 Mark Satterthwaite. 6:58–73. 1979. Optimal auction design. Princeton and Oxford. 117:55–77. Journal of Mathematical Economics.. 2004. 142. Public Choice. 13. 76 Roger B. 12. Journal of Economic Theory. Journal of Economic Theory. 10:187–217. North Holland. Princeton University Press. 55:169–179. Deferred acceptance algorithms: History. The relevance of private information in mechanism design. International Journal of Game Theory. pages 321–348. editor. 2008. 1975. auctions. 103. 17 Philip Reny. 77 Zvika Neeman. 142 Jacques Robert. 152 Halsey Royden. 152 Roger Myerson. 4. Journal of Finance. Arrow’s theorem and the gibbard-satterthwaite theorem: A uniﬁed approach. 107. Continuity in auction design. Princeton. and competitive sealed tenders. theory. and open questions. Counterspeculation. 3rd edition. 1988. 106. Prentice Hall. Amsterdam. 1961. 1983. 2007. Journal of Economic Theory. Real Analysis with Economic Applications. Ok. 29:256–281. Mathematics of Operations Research. Real Analysis. 1980. A necessary and suﬃcient condition for rationalizability in a quasi-linear context. In JeanJacques Laﬀont. Convex Analysis. 1987. Journal of Economic Theory.

83. 143 single dimensionality condition. 37 interim expected utility. 99 public goods problem. 9. 128 e cyclical monotonicity. 112 ﬂexible decision rule. 9 mechanism. 142 Myerson Satterthwaite Theorem. 19. 113 ex post revenue equivalence. 48 170 . 16 extreme point theorem. 84.Index allocation rule. 148 incentive compatibility. 51 revenue equivalence theorem. 80 dominant strategy incentive compat. 49 participation constraint. 123 bilateral trade. 83. 64 bounded type space. 71 set-monotonicity. 48. 34 Bayesian incentive compatible. 47 dominant strategy. 98 optimal auction. 9. 9 ex ante budget balance. 48 information rent. 110 canonical auctions. 37 interim revenue equivalence. 52. 109 identiﬁability condition. 139 ibility. 48. 68 positive association of diﬀerences. 141 second best mechanism. 96 rich type spaces. 125 payoﬀ equivalence. 48. 12 ex post budget balance. 109. 10 individual rationality. 33 mixed bundling. 38 ﬁxed price mechanism. 55. 141 direct mechanism. 15 interim expected transfer.non-transferrable utility. 24. 35. 135 single-peakedness. 41 equivalent mechanisms. 114 Gibbard-Satterthwaite Theorem. 124 measurability. 11. 41. 131 revelation principle. 82 ﬁrst best. 84 Cr´mer McLean condition. 72 Farkas’ Alternative. 117 ex post individual rationality. 17 regular distribution function. 46 extreme point. 116 pivot mechanism. 13. 124 independent private values. 34. 32 indirect mechanism. 103. 68 decision rule. 10. 98 dictatorial direct mechanism. 28 monotonicity. 11.

100 weak monotonicity. 143 Vickrey-Clarke-Groves mechanism. 141 subgradient. 106 type. 102 welfare. 47 171 . 7 unanimity. 140 strategy proofness.INDEX social choice function. 53.

Sign up to vote on this title

UsefulNot useful- Canoeing
- motoare electrice
- NUDE MICE and Other Medical Writing Terms You Need To Know
- Idraw User Guide
- Canoe
- DC & AC Machines and Speed Control
- Apps Magazine Issue 41 - 2014 UK
- Intro to Vector Control of Induction Mc for Electrical Engg
- The Wire and Cable Technical Information Handbook
- Soft Starter Handbook
- Mem System Design
- Motion Simulation a Mechanism
- proemec
- Modern.wiring
- Square D Wiring Diagram Book
- Borgers - An Introduction to the Theory of Mechanism Design

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

We've moved you to where you read on your other device.

Get the full title to continue

Get the full title to continue reading from where you left off, or restart the preview.

scribd