Auction Design without Quasilinear Preferences
Brian Baisa ⇤†
August 6, 2013
Abstract Ianalyzeprivatevalueauctiondesignandassumeonlythatbiddersarerisk averse and have positive wealth effects (i.e. the good is normal). I show removing the standard quasilinearity restriction leads to qualitatively different solutions to the auction design problem with respect to both efficiency or revenue maximiza- tion. On efficiency, I show that probabilistic allocations of the good can Pareto dominate the second price auction; and there is no dominant-strategy mechanism that is both Pareto efficient and individually rationality. On revenue, I construct a probability demand mechanism with greater ex- pected revenues than standard auctions when there are sufficiently many bidders. In addition, I take a new approach to studying bid behavior when types are mul- tidimensional. Instead of characterizing bidder’s interim incentive constraints, Iplaceboundsontheirbids,andIshowthattheseboundsaresufficientfor obtaining revenue comparisons.
Keywords: Auctions; Multidimensional Mechanism Design; Wealth Effects. JEL Classification: C70, D44, D82.
⇤[email protected], Amherst College, Department of Economics. †I gratefully acknowledge financial support received from the Cowles Foundation Fellowship, the Yale School of Management, the International Center for Finance and the Whitebox Advisors Fellowship. I received excellent comments on this work from seminar and conference participants and faculty at Yale, Notre Dame, UNC, Johns Hopkins, Haverford College, Amherst College, Cambridge, the University of Michigan, Tel Aviv University, USC, Stony Brook and Lund. I am especially grateful to Dirk Bergemann, Benjamin Polak, Larry Samuelson and Johannes Hörner for numerous conversations and encouragement while advising me with this project. I would also like to thank Cihan Artunç, Lint Barrage, Yeon-Koo Che, Eduardo Faingold, Amanda Gregg, Yingni Guo, Vitor Farinha Luz, Drew Fudenberg, Adam Kapor, Phil Haile, Sofia Moroni, David Rappoport, Peter Troyan and Kieran Walsh for helpful comments and conversations.
1 1Introduction
In the auction design literature, it is standard to assume that bidders have quasilinear preferences. Yet there are many well-known environments in which this restriction is violated: bidders may be risk averse, have wealth effects, face financing constraints or be budget constrained. In this paper, I study the canonical independent private value auction setting for a single good and drop the quasilinearity restriction by assuming only that bidders are risk averse and have positive wealth effects (i.e., the indivisible good for sale is a normal good). I show that the auction design problem leads to qualitatively different prescriptions relative to those of the quasilinear benchmark. Instead of using standard auctions where the good is given to the highest bidder with probability one, the auctioneer prefers mechanisms where she can allocate the good to one of many different bidders, each with strictly positive probability. For auctioneers concerned with efficiency, such probabilistic allocations can Pareto dominate the second price auction. And for auctioneers concerned with maximizing expected revenues, Iconstructaprobabilitydemandmechanismthatgeneratesstrictlygreaterexpectedrevenues than standard auction formats when there are sufficiently many bidders. There are many examples of auctions where the quasilinearity restriction does not hold. One case is housing auctions. Housing auctions have developed into a multi-billion dollar industry in the United States. In Melbourne, Australia an estimated 25 50% of homes are sold via auction (Mayer (1998)). In a housing auction, a buyer’s bidding strategy is influenced by factors like how much wealth she has (wealth effects) and the terms of the mortgage offered by her bank (financing constraints). These factors are not included in the quasilinear model, where a bidders payofftype is described only by a single dimensional variable - her valuation for the good. For another example, consider firms bidding on spectrum rights or oil tracts. The cor- porate finance literature shows that many firms have an internal spending hierarchy (see Fazzari, Hubbard and Petersen (1988)). Specifically, it is more expensive for the firm to use external financing than internal financing, because firms pay higher interest on money borrowed from third parties. A firm may be able to place a relatively low bid in an auction without needing external financing, but in order to place a relatively high bid, the firm may need to obtain external financing and pay a higher interest rate on this debt. Increasing its bid by one dollar using external financing is more costly to the firm than increasing the bid by one dollar via internal financing. Even if the firm is risk neutral, this financing con- straint makes them behave as though they have declining marginal utility of money. Further evidence of risk aversion in auctions is discussed in the related literature section. I first study the problem of an auctioneer concerned with efficiency. In the single good en-
2 vironment with private values and quasilinear preferences, the second price auction is popular because it implements a Pareto efficient allocation in dominant strategies. I show that with- out quasilinearity, there are probabilistic allocations of the good that Pareto dominate the dominant strategy equilibrium outcome of the second price auction (Proposition 1). Is there another dominant strategy mechanism that is efficient in this more general environment? I show that the answer is no: there is no individually rational dominant strategy mechanism that satisfies budget balance and implements a Pareto efficient allocation (Proposition 3). While I obtain an impossibility result on efficiency, I show more positive results for revenue maximization. Specifically, I show that the auctioneer can use randomization to increase revenues. This may seem counterintuitive because bidders are risk averse, but the intuition follows directly from the assumption that the good is normal. Since the good is normal, a bidder’s willingness to pay for it increases as her wealth increases. Similarly, her willingness to pay for any giving probability of winning the good increases as her wealth increases. It follows that the bidder is willing to pay the highest price for her first marginal ‘unit’ of probability of winning, when she has still yet to spend any money and her wealth is the highest. Thus, the bidder is willing to buy a small probability of winning the good at a price per unit of probability that exceeds her willingness to pay for the entire good. Standard auctions that allocate the good to the highest ‘bidder’ do not exploit this property of bidder risk preferences. Iconstructaprobability demand mechanism that uses lotteries to better exploit this feature of bidder preferences. The mechanism sells probabilities of winning the good like a divisible good that is in net supply one. Bidders report a demand curve over probabilities of winning. The curve reports the probability of winning the bidder demands (Q) for a given price per unit price of probability (P ). The auctioneer uses an algorithm similar to that of the Vickrey auction for a divisible good to determine each bidder’s probability of winning and expected transfer. Without quasilinearity, it becomes more difficult to characterize bidder behavior. Now, a bidder’s private information is described by a utility function instead of a single dimensional valuation. This multidimensionality complicates mechanism design problems. Armstrong and Rochet (1999) show that even in the simplest of principal-agent models with multidi- mensional types, explicitly solving for equilibria can be difficult. I take a different approach to characterizing bid behavior in my probability demand mechanism. Instead of explicitly solving for equilibria, I show that we can bound what the bidder reports to the auctioneer by eliminating dominated strategies. In particular, I show that it is a dominated strategy for a bidder to underreport her type (Proposition 4). Iusethisboundonabidders’reportstoconstructalowerboundonexpectedrevenues
3 in the probability demand mechanism. With enough bidders, this lower bound on revenues strictly exceeds an analogously constructed upper bound on the revenues from any standard auction. That is, with enough bidders the probability demand mechanism has higher ex- pected revenues than any standard auction (Propositions 6 and 7). This class of standard auctions includes the first price, second price and all pay auctions, as well as modifications of these formats to allow for entry fees and/or reserve prices. The rest of the paper proceeds as follows. The remainder of the introduction relates my work to the current literature on auction design. Section 2 describes the model and specifies the assumptions I place on bidders’ preferences. Section 3 motivates the use of probabilistic allocations and provides an example in which the dominant strategy equilibrium outcome of the second price auction is Pareto dominated by a probabilistic allocation of the good. Section 4 shows there is no symmetric mechanism that respects individual rationality and implements a Pareto efficient allocation in dominant strategies. Section 5 outlines the con- struction of the probability demand mechanism. Section 6 focuses on revenue comparisons between the probability demand mechanism and standard auction formats. Section 7 pro- vides a numerical example illustrating the practical applicability of my results. Section 9 concludes.
Related literature
This paper builds on the private value auction literature pioneered by Vickrey (1961) and later by Myerson (1981) and Riley and Samuelson (1981). Each of these papers addresses an auction design problem under the restriction that bidders’ preferences are quasilinear. Myerson studies the problem of an expected revenue maximizing auctioneer, while Vickrey studies the problem of an auctioneer concerned with efficiency. Their results show the auction design problem is solved by a mechanism where the good is assigned to a bidder only if she has the highest valuation among all bidders.1 However, there is evidence that in many auction settings, bidders do not have quasilinear preferences. Prior work has argued that risk aversion can explain deviations in bidders’ behavior, relative to the predictions of the quasilinear benchmark. Bajari and Hortaçsu (2005) argue that risk aversion serves as a reasonable explanation of experimentally observed bid behavior. Specifically, they show that it serves as a better explanation than three other competing theories - one of which is the quasilinear benchmark. Similarly, Goeree, Holt and Palfrey (2002) show risk aversion provides a parsimonious explanation of experimentally observed overbidding in first price auctions. Buddish and Takeyama (2001) cite risk aversion
1In Myerson (1981), the good is assigned to the bidder with the highest ‘virtual’ valuation. Under the monotone likelihood ratio property, this corresponds to the bidder with the highest valuation.
4 as an explanation for why sellers use buy-it-now prices. Li and Tan (2000) cite risk aversion as an explanation for why sellers use secret reserve prices. In the theoretical auctions literature, while most work focuses on the quasilinear case, there is a smaller literature that considers auctions without quasilinearity. One of the first works in this literature is by Maskin and Riley (1984). Their paper characterizes certain properties of expected revenue maximizing auctions when a bidder’s type is a single dimen- sional variable, ✓ [0, 1]. They show that the exact construction depends on the common 2 prior over the distribution of types and the functional form of the bidders’ utility functions. Matthews (1983, 1987) and Hu, Matthews and Zhou (2010) also study auctions without quasilinearity. However, they focus on a more structured setting where all bidders have identical risk preferences and there are no wealth effects. Within this framework, they make comparisons between auction formats. Che and Gale (2006) consider a payoffenvironment that is closer to the one studied here. They assume that bidders have multidimensional types and use this to show that when bidders are risk averse, the first price auction generates greater revenues than the second price auction. This builds on their earlier work, which studies standard auction formats with more specific departures from the quasilinear environment (see Che and Gale (1996, 1998, 2000)). My work differs from this prior work in two respects: (1) I consider a very general setting that allows for heterogeneity in bidders across many dimensions; and (2) I study the problem from an auction design perspective. This contrasts with prior work that focuses on comparisons between standard auction formats. My solution to the auction design problem uses probabilistic allocations of the good. Randomization has been advocated in other auction design settings. Celis, Lewis, Mobius and Nazeradeh (2012) construct a randomized mechanism they call the buy-it-now or take- a-chance mechanism. They study a setting where bidder preferences are quasilinear. They show that using randomization can induce more aggressive bidding in ‘thinner’ markets. For the case of budget-constrained bidders, Pai and Vohra (2010), Maskin (2000) and Laffont and Robert (1996) show that the expected revenue-maximizing mechanism may employ probabilistic allocations. Baisa (2012) studies an auction design problem in which bidders use a non-linear probability weighting function to evaluate risks (as described by Kahneman and Tversky (1979, 1992)) and finds that the solution to the auction design problem uses probabilistic allocations of the good to exploit features of bidders’ probability weighting functions. In this paper, I construct a mechanism that uses randomization and is almost revenue maximizing with many bidders. My results can be seen as analogous to those shown of Armstrong (1999). Armstrong
5 studies a different problem, that of a multi-product monopolist selling to a representative consumer. He derives an almost optimal solution for the monopolists when there are many products. Outside of auction design, Morgan (2000) advocates using lotteries to raise money for public goods. The intuition for using lotteries in Morgan’s model is distinct from the intuition used here. Morgan shows that lottery tickets can help to overcome the free rider problem when people’s preferences are linear in money. No such free rider problem is present in my model.
2Themodel
2.1 The payoffenvironment
I consider a private value auction setting with a single risk neutral seller and N 2 buyers, indexed by i 1,...,N} . There is a single indivisible object for sale. Bidder i’s preferences 2{ are described by utility function ui, where
ui : 0, 1 R R. { }⇥ !
Iletui(1,wi) denote bidder i’s utility when she owns the object and has wealth wi.Similarly, ui(0,wi) denotes bidder i’s utility when she does not own the object and has wealth wi. The object is a “good” and not a “bad”; it is better to have the object than to not for any fixed wealth level, i.e. u (1,w) >u(0,w), w. i i 8 I additionally assume that bidders preferences are strictly increasing and twice continuously differentiable in wealth. This environment differs from the quasilinear case where a bidder’s preferences are com- pletely described by a one-dimensional signal, her ‘valuation’ of the object. This valuation is the most i is willing to pay for the object. In this setting, I define k(ui,wi) as bidder i’s willingness to pay for the good when she has an initial wealth wi and a utility function ui. This is the highest price bidder i will accept as a take-it-or-leave-it offer for the good.
Formally k(ui,wi) is defined as,
ui(1,wi k)=ui(0,wi). (2.1)
I place only two assumptions on bidder preferences. The first is that the good is a normal
6 good and the second is that bidders are risk averse. I assume that the good being sold is a normal good (i.e. positive wealth effects). My notion of positive wealth effects is analogous to the notion in the divisible goods case, where abidder’sdemandforthegoodincreasesasherwealthincreasesforaconstantpricelevel.
Assumption 1. (Positive wealth effects) Bidder i has positive wealth effects:
@k(w, u ) i > 0, @w for all w.
Second, I assume that bidders have declining marginal utility from money.
Assumption 2. (Risk Averse) Bidder i has declining marginal utility of money:
2 @ ui(x, w) < 0 for x =0, 1. @w2
IdefineAnd, as the set of all utility functions which satisfy Assumptions 1 and 2.Notice U 2 @k(w,ui) @ ui(x,w) that quasilinear preferences are not included in since =0and 2 =0. U @w @w Note that this is a very general setting. The only two assumptions I place on bid- der preferences are positive wealth effects and risk aversion. I do not place any functional form restrictions on preferences. I also allow for multidimensional heterogeneity across bid- ders. This differs from most of the literature on risk aversion and auctions, where bidders preferences are described by specific functional forms and there is only single dimensional heterogeneity between bidders (see Maskin and Riley (1984), Mathews (1983, 1987)). There are many different functional forms of u that fit the risk aversion and normal good assumptions. Two examples are given below.
Example 1. If bidder i has preferences of the form ui(x, w)=gi(viIx=1 + fi(w)) where both fi and gi are concave functions and I is an indicator function, then ui .Asimpleexample 2U would be the case where gi is linear. We can then interpret this as saying bidder i’s gets a certain number of utils from owning the good, irrespective of her wealth level. She also has declining marginal utility of money, which is unaffected by whether or not she owns the good.
7 Example 2. The good being sold is a risky asset. It has a distribution of (monetary) returns g : R+ R+.Bidderi gets utility from wealth of fi.Shehasdecreasingabsolute ! risk aversion - for example, CRRA utility. Her expected utility of owning the asset when she starts with wealth w is ui(1,w) where
u (1,w)= f (w + x)g(x)dx. i ˆ i
Without the asset, she consumes her wealth,
ui(0,w)=fi(w).
In this example, bidder i could be a speculator bidding for a home. She knows this distribu- tion of resale prices for the house is given by g.Asshebecomesricher,herdegreeofabsolute risk aversion decreases. This implies that she demands a smaller risk premium when buying the home. Thus, as she becomes richer her willingness to pay for the home increases.
2.2 Allocations and mechanisms
By the revelation principle, I can limit attention to direct revelation mechanisms. A mecha- nism describes how the good is allocated and how transfers are made. I define A as the set of all feasible assignments, where
N N A := a a 0, 1 and ai 1 , { | 2{ } } i=1 X where ai =1if bidder i is given the object. A feasible outcome specifies both transfers and a feasible assignment: A RN . Idefine :=A RN as the set of feasible outcomes. 2 ⇥ ⇥ A (probabilistic) allocation is a distribution over feasible outcomes. Thus, an allocation ↵ is an element of ( ).
IusethenotationE↵ [ui,wi] to denote the expected utility of bidder i under allocation
↵ ( ) when she has preferences ui and initial wealth wi.Irefertotheauctioneeras 2 person 0 and use E↵ [u0] to denote the expected transfers she receives under allocation ↵. Note that the auctioneer does not value owning the good. In Section 4,IstudymechanismsthatimplementParetoefficientallocations.Inthe quasilinear benchmark, this problem simplifies to ensuring that the bidder with the highest valuation receives the good. Without quasilinearity, the description of a Pareto efficient allocation is more complicated. Icallanallocationposterior Pareto efficient if for a given profile of preferences, increasing
8 one person’s expected utility necessarily decreases another person’s expected utility (includ- ing the auctioneer). If a benevolent social planner knew each bidder’s private information, she would select a posterior Pareto efficient allocation.
Definition 1. (Posterior Pareto efficient allocations) N Assume bidders have preferences u =(u1,...,uN ) and initial wealth levels, w = 2U (w1,...,wN ).Afeasibleallocation↵ is posterior Pareto efficient if there does not exist a
↵0 ( ) s.t. 2 E↵ [ui,wi] E↵ [ui,wi] i =0, 1 ...N. 0 8 where the above inequality holds strictly for at least one i.
This is equivalent to the notion of an ex-post Pareto superior (efficient) allocation given in Holmström and Myerson (1983, 1991). I use the term posterior as my definition mirrors Laffont and Green’s (1987) notion of posterior implementability. In words, this means condi- tional upon knowing all of the bidders private information, we cannot increase one person’s expected payoffwithout decreasing another person’s. There are differences between posterior Pareto efficient allocations and the common no- tions of ex-ante, interim and ex-post Pareto efficiency. It is impossible to design a mechanism that implements an ex-ante or interim Pareto efficient allocation when bidders are risk averse. For example, if an allocation is ex-ante Pareto efficient (by the definition of Myerson (1991)), because the bidders are risk averse, their payments to the auctioneer can not depend on re- alized types. When the bidder’s payment is independent of her report, she has no incentive to tell the truth. At the same time, my posterior Pareto efficiency differs from the typical notion of ex- post Pareto efficiency. It allows the designer to randomly allocate the good to one of many different bidders. It does not necessarily ensure that a bidder with the highest willingness to pay for the good is given the good with certainty. Instead, using randomization can actually increase everyones payoff.
9 (ex-post) (interim) Good is allocated, Bidders report their payments are made, private information game is over.
(ex-ante) (posterior) Types are drawn Conditional on all bidders private information, auctioneer determines payments and the allocation of the good (possibly randomly).
Figure 2.1: Notions of efficiency AdirectrevelationmechanismM maps a profile of reported preferences and initial wealth levels to an allocation. That is,
M : N RN ( ). U ⇥ !
The direct revelation mechanism implements an posterior Pareto efficient allocation if for any reported preferences u =(u1 ...uN ) and initial wealth levels w =(w1 ...wN ),the allocation provided by M is Pareto efficient. That is, the mechanism specifies an allocation that would be Pareto efficient in the complete information environment where all bidders’ private information is known. Amechanismisimplementableindominantstrategiesifitisabestresponseforeach bidders to truthfully report her private information to the auctioneer, for any possible report of her opponent. Definition 2. (Dominant strategy incentive compatibility) AdirectrevelationmechanismM is dominant strategy incentive compatible if for any profile N N of preferences and initial wealth levels u =(ui,u i) and w =(wi,w i) R , 2U 2
EM(u,w) [ui,wi] EM((u0,u i),(w0,w i)) [ui,wi] , for any (u0,w0) R and i =1...N. 2U⇥
3Probabilisticallocations
I now show that auctioneers concerned with efficiency or revenue maximization should use mechanisms that allow for probabilistic allocations of the good. Such probabilistic allocations
10 give many bidders a positive probability of winning. This contrasts with most commonly studied mechanisms that assign the good to a bidder with the highest willingness to pay with certainty. The reason for using randomization is based on the assumption that the object is a normal good. When a good is a normal good, a bidder is willing to pay the most for her first ‘unit’ of probability of winning the good. As the bidder becomes poorer, she is willing to pay less for a marginal increase in her probability of winning. To illustrate this, I consider the following example. Suppose a bidder i has an initial wealth of wi and preferences described by the utility function u, where
u(x, w)=4Ix=1 + pw.
She receives 4 utils from being given the object and pw utils from having final wealth w. This utility function satisfies the risk aversion and normal good assumptions. The bidder has declining marginal utility of money, and her willingness to pay for the good increases as her initial wealth increases.2 This is illustrated by Figure 3.1.
Utils u!1, w"
15 u!0, w"
10
Initial Wealth: 169; WTP: 88. 5 Wealth: 49; WTP: 40.
0 w 0 50 100 150 200
Figure 3.1: u(x, w)=4Ix=1 + pw
Assume that the bidder has an initial wealth of 100. The bidder is willing to pay 64 for the good, since u(1, 100 64) = u(0, 100) = 10. Suppose there are two identical lottery tickets that each give a 50 50 chance of winning the good. One of the two tickets is randomly selected, and that ticket holder wins the good.
2A bidder is risk averse and has positive wealth effects for any utility function which has this separable form. Specifically, any utility function of the form u(x, w)=vIx=1 + g(w) where v>0 and g is strictly increasing and concave satisfies Assumptions 1 and 2.
11 The bidder can purchase 0, 1 or both tickets. If she purchases both, she wins the good with certainty. The bidder’s willingness to pay for her first ticket is simple to calculate. It 1 is the price that makes the bidder indifferent between having a 2 probability of winning and remaining at her status quo initial wealth with no chance of winning the good. Her willingness to pay for the ticket is 36 as,
.5u(1, 100 36) + .5u(0, 100 36) = u(0, 100).
The bidders willingness to pay for the good is her willingness to pay for both tickets combined, which is 64. Thus, the bidder is willing to pay a higher price for her first ticket than her second. By the time she is deciding whether to purchase the second ticket, she is poorer and thus willing to pay less for the second ticket. This would not be true if bidder i had quasilinear preferences. If bidder i had quasilinear preferences, with a valuation of 64,she would be willing to pay an equal price for both tickets - in this case 32. This property of bidder preferences is important for an auctioneer concerned with effi- ciency. To see this, consider a case where there are two identical bidders who both have the same preferences as the bidder described above. If the auctioneer used a second price auction to sell the good, both bidders have a dominant strategy to bid their willingness to pay - in this case, 64. Since both bidders are identical, a winner is then randomly selected and pays 64 for the good. Thus, both bidders are indifferent between winning and losing. This is equivalent to bundling both lottery tickets together and selling them to one bidder for a price of 64. However, the auctioneer can use randomization to create an allocation that Pareto dom- inates this outcome. In this example, the reallocation is very simple. Recall that the loser is willing to pay more for her first ticket than she is willing to pay for her second. The loser is also ex-ante identical to the winner. Thus, the loser is willing to pay more for her first ticket than the winner is willing to pay for her second ticket. The winner could give one of her lottery tickets to the loser in exchange for 32. This is a Pareto improving reallocation. The revenue for the auctioneer is unchanged. The expected payoffof the loser increases - she is willing to pay up to 36 to obtain her first ticket, but here she only pays 32.Similarly,this trade increases the expected payofffor the winner as well - she is willing to sell her second ticket for as little as 28,buthereshegets32. This example illustrates the first qualitative difference from prescriptions of the quasilin- ear benchmark. When bidders’ preferences are restricted to be quasilinear, the second price auction implements a Pareto efficient allocation in dominant strategies. Without the restric- tion of quasilinearity, there are preferences where the auctioneer can construct a probabilistic
12 allocation of the good which Pareto dominates this allocation.
Proposition 1. a There exist probabilistic allocations of the good that Pareto dominate the dominant strategy equilibrium outcome of the second price auction.
The proposition is in contrast with the results of Sakai (2008) and Saitoh and Serizawa (2008). These authors claim the second price auction is Pareto efficient when bidders have non-quasilinear preferences. They obtain their results by restricting attention to determinis- tic mechanisms and using notions of envy-freeness in place of my notion Pareto efficiency. My example shows that when we use a notion of Pareto efficiency that allows for randomization, the second price auction does not guarantee Pareto efficiency. The proposition also relates to prior findings in the general equilibrium literature. Prescott and Townsend (1984) show that even when agents are risk averse, randomization is necessary for efficiency in a number of different general equilibrium settings. They write, “the key to this conclusion is that ex post full information inefficiencies and randomness are both con- sistent with ex ante private information optimal allocations.” Here we see that the same is true. While it is ex-post efficient to assign the good to the bidder with the highest willingness to pay, there are ex-ante gains in efficiency by using randomization, even when there is no private information. The conclusion of Proposition 1 leads to a natural question. Under the quasilinearity re- striction the second price auction is the unique dominant strategy mechanism that guarantees aParetoefficientallocationofthegood(Holmstrom,1979).Withoutthequasilinearityre- striction, Proposition 1 shows this is no longer true. This leads me to ask: is it possible to construct a mechanism that retains the desirable properties of the second price auction when we drop the quasilinearity restriction?IaddressthisquestioninSection4 and I show the answer is no. However, my main focus will be on studying the properties of revenue maximizing mecha- nisms. Here I obtain more positive results - I show that the auctioneer can use randomization to increase expected revenues. Using randomization allows the auctioneer to exploit the fact that a bidder is willing to pay a high price for her first marginal unit of probability of win- ning. In fact, she is willing to buy a positive probability of winning the good at a price per unit of probability that exceeds her willingness to pay for the entire good. I show that this property holds for any bidder who is risk averse and has positive wealth effects. In particular, I show that bidders’ have downward sloping demand curves for probabilities of winning the good. This probability demand curve states the probability of winning the good a bidder demands when she pays a price of p per unit of probability. The demand
13 curve captures the fact that bidders are willing to pay a higher price for their first units of probability. In fact, a bidder is willing to buy a positive probability of winning the good at prices (per unit) that exceed her willingness to pay for the entire good. Without the quasilinearity restriction, how a bidder’s payments are structured affects her payoff. Standard auctions do not equalize a bidder’s marginal utilities across different states of the world. In particular, a bidder may value a marginal dollar more when she wins than when she loses, or visa versa. By equalizing a bidder’s marginal utility across the win and lose state, the auctioneer can increase the bidder’s expected surplus. I will call such a payment scheme efficient. That is, given a bidder’s expected transfers and probability of winning the good, the efficient payment scheme maximizes her expected utility. I will show that with enough bidders, the auctioneer can extract the surplus created by efficient payments and increase revenues. I construct a bidder’s probability demand curve under the assumption that her payments are structured efficiently. To formalize this, consider a gamble where bidder i wins the good with probability q and pays x to the auctioneer in expectation. The bidder has preferences ui and initial wealth 2U wi. The efficient payment scheme maximizes bidder i’s expected utility given that she wins with probability q and pays x in expectation. Assume that she pays pw and pl conditional on winning or losing, respectively. In the efficient payment scheme, pw⇤ and pl⇤ are such that,
(pw⇤ ,pl⇤)=argmaxqui(1,wi pw)+(1 q)ui(0,wi pl). pw,pl
s.t. x = qp +(1 q)p . w l
The payments pw⇤ and pl⇤ maximize her expected utility under the constraint that she wins the good with probability q and pays x in expectation. This induces an indirect utility function Vi that is a function of her expected payments and the probability she wins the good. V (q, x):=max qu (1,w p⇤ )+(1 q)u (0,w p⇤). i pw,pl i i w i i l s.t. x = qp⇤ +(1 q)p⇤ w l The indirect utility function gives the maximal expected utility for bidder i conditional on winning the object with probability q and paying x in expectation. If bidder i makes an efficient payment, her expected utility from winning the good with probability q and paying a price of p per unit of probability is Vi(q, qp).Iformbidderi’s probability demand curve by assuming that she makes efficient payments.
q(p, ui,wi):=argmaxVi(q, qp). (3.1) q [0,1] 2
14 Bidder i’s probability demand curve states the probability of winning she demands when she pays a price of p per unit of probability. I economize notation by writing bidder i’s probability demand curve as qi(p)=q(p, ui,wi). Abidder’sprobabilitydemandcurveiscontinuousandweaklydecreasinginthepricep. In addition, the bidder demands a strictly positive probability of winning the good at a per unit price that exceed her willingness to pay for the entire good.
Proposition 2. a
If bidder i has preferences ui ,initialwealthwi and willingness to pay ki,then 2U (1) qi(p) is continuous and weakly decreasing.
(2) qi(ki + ✏) > 0 for some ✏>0.
Figure 3.2 shows the probability demand curve for bidder i with initial wealth of 100 and preferences u(x, w)=4Ix=1 + pw,asintheaboveexample.
p 80 75 ki = 64 70 65 60 55 q p 0.2 0.4 0.6 0.8 1.0
Figure 3.2: A probability demand curveH L
As a basis for comparison, if a bidder has quasilinear preferences, with a valuation of v for the object, her probability demand curve is a step function where, q(p)=Ip v. Notice that this is discontinuous and q(p)=0for all prices above the bidders willingness to pay for the good v. It is the normal good assumption that implies that any bidder is willing to buy a strictly positive probability of winning the good at a per unit price that exceeds her willingness to pay for the good. This can be illustrated graphically. Consider a bidder i with an initial wealth of wi, and a willingness to pay for the good of ki.Supposethegoodwasbeing sold for exactly ki. Thus, bidder i is exactly indifferent between buying and not buying the good. Because the good is normal, increasing or decreasing bidder i’s wealth from wi would break this indifference. This means that bidder i would strictly prefers to (not) buy the good for the fixed price of ki if and only if her wealth is greater (less) than wi. Thus, ui(1,w ki) >ui(0,w) if and only if w>wi. This is illustrated in Figure 3.3. 15 Utils Utility if buy at price ki
Utility if do not buy
wealth wi
Figure 3.3: Bidder i’s indirect utility function
The upper envelope of the two curves is bidder i’s indirect utility as a function of her initial @ @ wealth when the object costs ki.Noticethat ui(1,wi ki) > ui(0,wi), which causes @w @w the curve ui(1,w ki) to cross the curve ui(0,w) from below to above at the point w = wi. This creates a local convexity in the indirect utility curve at the point wi.Byexploitingthis difference in bidders marginal utilities of money between the win state and the lose state, the auctioneer can construct a gamble where the bidder wins the good with positive probability and pays a price per unit of probability that strictly exceeds her willingness to pay for the entire good with certainty. This property of bidder preferences is similar to results shown by Kwang (1965). He shows indivisibilities in consumption can lead to risk loving behavior for the case where the utility of the good and money are separable.
4Paretoefficientmechanisms
When bidder preferences are restricted to be quasilinear, the second price auction is the unique format that implements an individually rational and Pareto efficient allocation in dominant strategies (Holmström, 1979). Proposition 1 shows that, without quasilinearity, the dominant strategy equilibrium outcome of the second price auction can be Pareto dominated by probabilistic allocations of the good. This leads to the question I address in this section: Is there a dominant strategy mechanism that respects individual rationality, generates a (weak) budget balance for the seller and is Pareto efficient? I show the answer to this question is no. Without the quasilinearity assumption, dominant strategy implementation imposes more restrictions on the design problem. With the quasilinearity restriction, a bidder can only
16 misreport her type by submitting another (single-dimensional) valuation. Without quasilin- earity, dominant strategy incentive compatibility requires that each bidder has no incentive to misreport her preferences for any set of reports by other bidders. She can misreport her preferences by reporting another (infinite-dimensional) utility function. Since each bidder has a much larger set of possible deviations, this makes dominant strategy implementation more difficult to achieve. Without the quasilinearity assumption, requiring Pareto efficiency also imposes more restrictions on the design problem. With the quasilinearity restriction, Pareto efficiency re- quires that the bidder with the highest valuation is given the object. Proposition 1 illustrates that this is not true when we drop the quasilinearity restriction. Without quasilinearity, re- quiring Pareto efficiency places restrictions both on how transfers are made and with what probability each bidder receives the good. With the quasilinearity restriction, there is a unique mechanism that is individual ratio- nal and implements a Pareto efficient allocation in dominant strategies. However, without quasilinearity, it becomes more difficult to design such a mechanism. This leads to the impossibility result I state below.
Proposition 3. a There is no dominant strategy direct revelation mechanism that is Pareto efficient and sat- isfies individual rationality and weak budget balance.
The proof of the proposition is developed in the appendix. I show that the necessary conditions for Pareto efficiency and dominant strategy incentive compatibility form a contra- diction. I prove this result by considering only a subset of the incentive constraints required for dominant strategy implementation. I show that there is no mechanism that both satisfies this subset of incentive constraints and implements a Pareto efficient allocation. Thus, there is no mechanism that satisfies all dominant strategy incentive constraints and implements a Pareto efficient outcome. In particular, I consider a subset that includes a continuum of types that are almost quasilinear. Additionally, I consider a single type with non-linear utility of money. I use the necessary conditions for Pareto efficiency and dominant strategy incentive compatibility to show that if all players report types that are almost quasilinear, the mechanism must specify an allocation that is close to the allocation specified by the second price auction. In particular, the bidder with the highest type wins the object with certainty and pays an amount approximately equal to the second highest bidder’s willingness to pay for the good. However, when one bidder is of the non-linear type, the necessary conditions for Pareto efficiency place strict restrictions on the structure of the allocation. Under these necessary
17 conditions, if a bidder has almost quasilinear preferences, she can gain by underreporting when another bidder reports her type to be non-linear. The reason I study mechanisms that satisfy individual rationality and weak budget bal- ance is to avoid trivializing the problem. If I removed the individual rationality condition, I could form a mechanism which is both Pareto efficient and one such mechanism would randomly select a ‘winner’. It would then take a large sum of money from all other bidders and also give this money to the winner. Now that the winner is very rich and the losers are very poor, it is efficient to assign the winner the good (recall the positive wealth effects assumption). Similarly, if I removed the budget balance requirement, I could construct a mechanism in which the auctioneer does the same thing. She randomly selects a winner and gives her enough money such that it is efficient that she also owns the object. The relatively benign assumptions of weak budget balance and individual rationality avoid these trivializations. My proof strategy is similar to that used by Jehiel, Moldavanu, Meyer-Ter-Vehn and Zame (2006). They study a related multidimensional single good allocation problem when bidders have quasilinear preferences and interdependent values. They show that ex-post im- plementation of non-trivial social choice functions is impossible in generic multi-dimensional type spaces. To show this, they study a two bidder setting where i’s type is described by 2 (✓i1,✓i2) [0, 1] .Hervaluationisgivenby 2
I
vi(✓)=✓i1 + ✏ ✓j2. j=1 Y They show that in this setting, the only ex post incentive compatible social choice functions are trivial. My proof is analogous in that I also show an impossibility result by showing that the necessary conditions for dominant strategy implementation and Pareto efficiency form a contradiction, even when studying a limited type space. The results of Jehiel et. al and Proposition 3 illustrate similar points. The former il- lustrates the difficulties of implementing social choice functions when bidders preferences are restricted to be quasilinear, but are unrestricted in their interdependence. The latter illustrates the difficulties of implementing efficient allocations when bidders are restricted to have private values, but have unrestricted risk preferences. In both cases, multidimen- sional heterogeneity between bidders makes robust implementation of specific social choice functions more difficult than in the benchmark setting.
18 5Theprobabilitydemandmechanism
The previous section illustrated the importance of probabilistic allocations for an auctioneer who is concerned with efficiency. Proposition 1 shows that probabilistic allocations of the good can improve on efficiency relative to the prescriptions of the benchmark model, though Proposition 3 provides a more negative result. Now I shift focus to the second question, the problem of an auctioneer concerned with maximizing expected revenues. I obtain more positive results and show that using ran- domization increases revenues. I construct a probability demand mechanism that generates strictly greater expected revenues than standard auctions when there are sufficiently many bidders. Recall that Proposition 2 shows that if a bidder is indifferent between accepting or re- jecting a take-it-or-leave-it offer for the good at a price of k,therearegamblesshestrictly prefers to enter where she wins the good with positive probability q and pays strictly greater than qk in expectation. This is not true when bidders’ preferences are restricted to be quasilinear. Any such gamble violates a bidder’s individual rationality constraint. Auction formats that are commonly studied in the benchmark setting do not exploit this feature of bidders’ preferences. For example, in the first and second price auctions, it is a dominated strategy for a bidder to submit a bid above her willingness to pay for the good. However, my probability demand mechanism uses this property of bidder preferences to generate greater revenues than standard auctions. The probability demand mechanism sells bidders probabilities of winning the good like a divisible good that is in net supply one. Instead of submitting single dimensional bids, bidders reports their probability demand curves to the auctioneer. Recall a bidder’s probability demand curve states how much probability of winning she demands when she pays a price of p per unit. The construction exploits bidders preferences towards risk - it allows the auctioneer to sell bidders probabilities of winning the good at prices per unit that exceed their willingness to pay for the good with certainty. The auctioneer uses the reported probability demand curves to calculate each bidder’s expected transfer and probability of winning. This is done by using an algorithm that resembles the Vickrey auction for a divisible good. In the Vickrey auction for a divisible good, bidders report demand curves for quantities of the good. The auctioneer uses the reported demand curves to calculate each bidder’s transfer and quantity allocation. In the probability demand mechanism, I use bidders probability demand curves to analogously determine a bidder’s expected transfers and probability of winning the good. Though my probability demand mechanism and the Vickrey auction for a divisible good
19 have similar constructions, there are important differences in bidder behavior between the two. In the latter, it is a dominant strategy for bidders to truthfully report their preferences. But removing the quasilinearity restriction poses challenges not seen in the benchmark divis- ible goods setting. The standard approach of solving for an equilibrium by studying bidders’ interim incentive constraints proves to be intractable in multidimensional settings. Thus, I take a different approach to characterizing bid behavior and generating revenue comparisons. Instead, I partially characterize bidder behavior by placing bounds on their reports to the auctioneer. Specifically, in my probability demand mechanism, there is a lower bound on what a bidder will want to report to the auctioneer. I show it is a dominated strategy for a bidder to underreport her demand curve to the auctioneer. I can then use a bidder’s truthful report as a lower bound on what she actually will report to the auctioneer. I use this bound on bidder behavior to obtain revenue comparisons between my probability demand mechanism and standard auctions. In this section I discuss the construction of the probability demand mechanism. Section 6 discusses revenue comparisons between the probability demand mechanism and standard auction formats.
5.1 The probability demand mechanism
In the probability demand mechanism, a bidder reports her probability demand curve q(p, ui,wi), preferences ui and initial wealth level wi. The auctioneer uses the reported demand curves to calculate each bidder’s probability of winning and expected payment. Given a bidder’s probability of winning and expected payments, her payments are then structured efficiently. It is equivalent to assume that each bidder reports only her preferences and initial wealth levels and that the auctioneer derives her probability demand curve directly. I work with the indirect mechanism, where bidders also report probability demand curves, to illustrate the connection to the divisible goods setting. For economy of notation I write qi(p)=q(p, ui,wi). The probability that bidder i wins the object is calculated using the reported probability demand curves. Given the reported demand curves, the auctioneer calculates the (lowest) price for probabilities of winning the good that ‘clears the market.’ That is, she finds the
(lowest) price p⇤ where the total reported demand for probabilities of winning the good equals one. N
p⇤ := inf p :1= qi(p) . (5.1) p { } i=1 X Given p⇤,theprobabilitybidderi wins the good is qi(p⇤). This is the (reported) probability of winning that she demands when she pays a price of p⇤ per unit of probability and payments
20 are made efficiently.
The market clearing price p⇤ is not the per unit price bidders pay for probabilities of winning the good. Instead each bidder faces a probability supply curve that represents her marginal price curve for probabilities of winning the good. The supply curve curve is the residual demand for probabilities of winning. It is analogous to the residual demand curve in Vickrey auction for a divisible goods auction. Thus, the price a bidder pays for a unit of probability is determined by her opponent’s actions.
Given a price p,abidder’sprobabilitysupplycurveSi(p) states the amount of probability of winning the good that is not demanded by the N 1 other bidders.
1 j=i qj(p) if 1 j=i qj(p) > 0 Si(p)= 6 6 (5.2) 80 otherwise < P P Bidder i’s (reported) probability: demand curve equals her probability supply curve at the price p⇤. Thus, the market clearing price sets each bidder’s probability demand curve equal to her probability supply curve. Bidder i’s expected payments are determined by treating her probability supply curve as her (expected) marginal price curve. Her expected payment to the auctioneer is Xi, where,
p ⇤ Xi = tdSi(t). (5.3) ˆ0 This payment rule can be viewed as the continuous analog to the payment rule in the second price auction. Note that I suppress notation in writing Xi;itisafunctionofthecomplete profile of reported demand curves (q1 ...qN ).Figure5.1 illustrates this graphically.
p 8 qi p Si p 6
4 H L H L
2 i's expected payment 0 q 0.0 0.2 0.4 0.6 0.8 1.0
Figure 5.1: Expected payment in the probability demand mechanism.
21 Thus, the reported probability demand curves determine each bidder’s expected transfers and probability of winning the good. Notice that bidder i pays a marginal price for a unit of probability that is lower than the market clearing price p⇤ . Thus, Xi p⇤qi(p⇤). To complete the description of the mechanism, I specify precisely how payments are structured. I assume that each bidder’s payment to the auctioneer is structured efficiently given the reports. Thus, if the profile of reported demand curves (q1 ...qN ) is such that bidder i wins with probability qi(p⇤) and pays the auctioneer Xi in expectation, she makes an efficient payment. If bidder i reports her preferences to be ui and her initial wealth to be wi,shethenpayspi,w⇤ when she wins and pi,l⇤ when she loses where,
(pi,w⇤ ,pi,l⇤ )=argmaxqi(p⇤)ui(1,wi pw)+(1 qi(p⇤)) ui(0,wi pl). (5.4) pw,pl
s.t. X = q (p⇤)p +(1 q (p⇤))p . i i w i l Definition 3. (The probability demand mechanism)
The probability demand mechanism maps reported demand curves (q1,...,qN ),preferences
(u1 ...uN ),andinitialwealthlevels(w1 ...wN ) to a probabilistic allocation described by (5.1) (5.4). The probability demand mechanism can be modified so bidders only submit a probability demand curve to the auctioneer. This is done using a two-step procedure where bidders first submit probability demand curves and each bidder’s probability of winning and ex- pected payments are calculated as discussed above. Then, the auctioneer tells each bidder her probability of winning qi(p⇤) and the expected payment she will make Xi.Beforethe actual winner is selected, each bidder decides how to structure her payments, given that her probability of winning is qi(p⇤) and her expected payment is Xi.Alternatively,theauc- tioneer can use an ascending-bid clinching mechanism, as described in Ausubel (2004), to determine bidders probabilities of winnings and expected payments. This would not even require bidders to submit a probability demand curve.
5.2 Behavior in the probability demand mechanism
The construction of the probability demand mechanism illustrates its similarities to the Vickrey auction for a divisible good. Yet, in the benchmark divisible good setting, bidders’ preferences are restricted to be quasilinear. Under this restriction on preferences, the Vick- rey auction for a divisible good has a dominant strategy equilibrium (see Krishna (2001)). However, in the indivisible good setting that I study, I drop the quasilinearity restriction
22 on bidders’ preferences and I find that it is no longer a dominant strategy for a bidder to truthfully report her preferences. Removing the quasilinearity restriction leads to multidimensional heterogeneity between bidders. This multidimensional heterogeneity makes characterizing bidders interim incentive constraints intractable. Thus, in order to describe bidder behavior, I take a new approach. In particular, I show that we can obtain a partial characterization of bid behavior by placing bounds on a bidder’s report to the auctioneer. I show that this partial characterization of bidder behavior will already be sufficient for making revenue comparisons between my probability demand mechanism and standard auctions. In this section, I derive the bound on a bidder’s report to the auctioneer. Specifically, I show it is a dominated strategy for a bidder to report a probability demand curve which underreports her demand for winning probabilities. The intuition for this result can be understood graphically. First, I consider a hypothetical case where it is a best response for a bidder to truthfully reveal her demand curve. Suppose bidder i faces a perfectly elastic probability supply curve, and she pays a constant marginal price for a unit of probability of winning the good. Assume this price is pE.Herbestresponse is to maximize her expected utility given she pays a price pE per unit of probability. The solution to this maximization problem is qi(pE) by definition - qi(pE) is the probability of winning the good that she desires when she pays price of pE per unit of probability. Thus, truthful reporting is a best response. By truthful reporting, she wins the good with probability qi(pE) and pays pEqi(pE) in expectation.
p 8 qi p 6 S p p* i 4 H L H L 2 i's expected payment q 0.0 0.2 0.4 0.6 0.8 1.0
Figure 5.2: Expected payment when facing a perfectly elastic supply curve.
Suppose instead bidder i faces a more inelastic (relative to perfectly elastic) probability supply curve. Suppose her residual probability demand curve still passes through the point
(pE,qi(pE)). If bidder i truthfully reports her probability demand curve, she wins the good with the probability qi(pE). Thus, her probability of winning the good is the same as it is
23 when she faces the perfectly elastic supply curve. However, she pays less when she faces the more inelastic supply curve. The marginal price she pays for all but the final unit of probability she acquires is less than pE. Thus, she pays Xi which is less than pEqi(pE) for a qi(pE) probability of winning. It is as though she faced the same supply curve that was perfectly elastic, with constant price pE,andthenshewasgivenarefundofpEqi(pE) Xi 0. This increases her demand of the good relative to the case where she pays the price of pE per unit of probability.
p 8 qi p
Si p 6 H L p* H L 4
2 i's expected payment q 0.0 0.2 0.4 0.6 0.8 1.0
Figure 5.3: Expected payment when facing a relative inelastic supply curve. Thus, while it was a best response for bidder i to truthfully report her demand curve when she faced a constant marginal price curve, she actually has an incentive to over report her demand curve when she faces an upward sloping marginal price curve. Since bidders are restricted to reporting downward sloping demand curves, bidder i will always face an upward sloping marginal price curve. The precise amount that bidder i wants to overreport her demand curve depends on the elasticity of the supply curve she expects to face. Her incentive to misreport is greater when facing a more inelastic supply curve - she is given a larger ‘refund’ relative to the case where she faces a perfectly elastic demand curve. Thus her demand increases more relative to the perfectly elastic case. What is clear is that it is never a best reply for bidder i to ever underreport her demand curve. This observation allows us to use bidder i’s truthful report as a lower bound on her actual report. Proposition 4. a
Assume bidder i has probability demand curve qi(p).Itisaweaklydominatedstrategyto report a probability demand curve q˜i where q˜i(p) 24 Proposition 4 shows that truthful reporting can serve as a lower bound on a bidder’s possible report. This lower bound on a bidder’s report enables revenue comparisons between the probability demand mechanism and other auctions. 6Revenuecomparisons In this section, I show that with sufficiently many bidders, the expected revenues from the probability demand mechanism exceeds the expected revenues of a large class of standard auctions. In addition, the probability demand mechanism achieves approximate revenue maximization while most standard auctions do not. Included within the class of standard auctions is the first price, second price and all pay auctions. Also included are any of these auctions modified to include reserve prices or entry fees. Ishowthisinthreesteps.First,Ideriveanupperboundontheexpectedrevenuesfrom any mechanism that respects bidders’ interim individual rationality constraints. Next, I use Proposition 4 to place a lower bound on bidders’ reports to the auctioneer in the probability demand mechanism. Using the lower bound on bidders’ reports, I obtain a lower bound on the expected revenues. As N ,thislowerboundontheexpectedrevenuefromthe !1 probability demand mechanism approaches the revenue upper bound for any individually rational mechanism. Lastly, I define a broad class of ‘highest-bid’ wins mechanisms. This class of mechanisms includes most commonly studied auctions. I show that even when there are many bidders, the expected revenues from any highest-bid wins mechanism does not approach the revenue upper bound. Thus, it attains a strictly lower bound than the one attained by my probability demand mechanism. 6.1 A revenue upper bound for all mechanisms Iassumethatbidders’preferencesareindependentlyandidenticallydistributed.Inpartic- ular, I assume there is a distribution over possible ‘types’ of bidder preferences. I describe M bidder i’s preferences by a type ti, where ti R and M is finite. I let the first element 2 represent bidder i’s initial wealth level wi. A bidder with type ti has preferences described by the utility function u(x, w, ti) when her type is ti.Iassumethatallbidderspreferenceshave M declining marginal utility of money and positive wealth effects. That is, for any ti R , 2 u(x, w, ti) . At the same time, this setup allows for heterogeneity across multiple dimen- 2U sions. I do not place restrictions on the functional forms of bidders’ utility functions. While the standard quasilinear model allows for heterogeneity in ‘valuations’, this setup allows for heterogeneity across risk preferences, initial wealth levels and financing constraints. 25 M If two types t, t0 R are relatively close, they have similar preferences. Specifically, I 2 assume u(x, w, t) is continuous in t. Each bidder’s preference is an i.i.d. draw of a random variable with density f. The distribution f has full support over T ,acompactsubsetofRM . The density maps types to anon-negativenumber,f : T R+. ! With this notation, I will briefly talk about the existence of equilibria in the probability demand mechanism. While I do not explicitly solve for an equilibrium, I can use a recent result by Reny (2011) to show that there exists a symmetric Bayes Nash equilibrium in pure strategies. However to use Reny’s result, I must first make two more assumptions. First, I assume that f is atomless and second I assume monotonicity on types which means that if ti tj for all coordinates, then q(p, ti) q(p, tj) for all p 0.Notethatthisassumption is only necessary to use Reny’s existence result. I do not impose these two assumptions anywhere else in the paper. Proposition 5. a Under the monotonicity and atomless distribution assumptions, there exists a unique sym- metric pure strategy Bayes Nash equilibrium of the probability demand mechanism. With this, I return to the problem of deriving a revenue upper bound. For a given distribution over types, I can define an upper bound on revenues from any mechanism that is interim individually rational. To do so, I first consider the highest per unit price where a bidder demands a positive probability of winning the good. Let pi be defined as the highest per unit price where bidder i demands a positive probability of winning. Note that pi is a function of bidder i’s type, ti. Thus I write it as p(ti), where p(ti):=sup p : q(p, ti) > 0 . p { } Since preferences u are continuous in t, it follows that p(ti) is continuous in ti.Giventhe distribution of preferences f,IletP denote the highest possible per unit price where a bidder demands a positive probability of winning the good. P := sup p(t):t supp(f) . { 2 } IassumethatP< . Thus, for a sufficiently high per unit price, no bidder demands any 1 probability of winning the good. The price P serves as an upper bound on the expected revenues from any interim individually rational mechanism. Since no bidder is ever willing to pay a per unit price that exceeds P for any positive probability of winning the good, a 26 mechanism with expected revenues that exceed P necessarily violates some bidder’s interim individual rationality constraint. Remark 1. a All mechanisms which are interim individually rational generate expected revenues less than or equal to P . This is a generous upper bound on revenues. It only assumes that a mechanism respects interim individual rationality constraints and holds for any number of bidders. 6.2 Revenues from the probability demand mechanism Proposition 4 states that in the probability demand mechanism, it is a weakly dominated strategy for a bidder to submit a demand curve where she underreports her demand. By assuming that bidders play only undominated strategies, I show that as N ,alower !1 bound on expected revenue from the probability demand mechanism approaches the upper bound on expected revenues from any interim individually rational mechanism, P . Proposition 6. a As N ,ifallbiddersplayundominatedstrategies,theexpectedrevenuefromtheproba- !1 bility demand mechanism approaches P . The formal proof of the above result is in the appendix, though I discuss the intuition here. Using the definition of P and the continuity of p in t,thereisapositiveprobability that a randomly drawn bidder has a demand curve where qi(p) > 0 for any p