First-Price Auctions with Budget Constraints*† Maciej H. Kotowski‡ March 23, 2013 Abstract Consider a first-price sealed-bid auction where participants have affiliated valuations and private budget constraints; that is, bidders have private multidimensional types. This article investigates the existence and nature of monotone equilibria in this setting. Hard budget constraints introduce two competing effects on bidding. The direct effect depresses bids as participants hit their spending limit. The strategic effect encourages more aggressive bidding by participants with large budgets. Together these effects can yield discontinuous equilibrium strategies stratifying competition along the budget dimension. The strategic consequences of private budget constraints can be a serious confound in interpreting bidding behavior in auctions. Keywords: First-Price Auctions, Budget Constraints JEL: D44 *This draft is made available to stimulate discussion and to gather feedback. It has not benefited from a formal peer review. Please do not (re)post this article. Please contact the author for the latest draft before citing. All comments and suggestions are greatly appreciated. †I thank Benjamin Hermalin, John Morgan, Chris Shannon, Adam Szeidl, Steven Tadelis, and especially Shachar Kariv. I am grateful for the feedback of numerous seminar audiences. This paper supersedes and corrects results and discussion originally communicated in Kotowski (2010)and included in Kotowski(2011). I accept responsibility for any remaining errors or omissions. ‡Kennedy School of Government, Harvard University. 79 JFK Street, Cambridge, MA 02138. E-mail: maciej_kotowski@hks.harvard.edu. 1 Budget constraints are a fundamental feature of many auctions. For example, bidders may be buying real estate. A house’s attractiveness is partly idiosyncratic and all potential buyers face a private spending limit determined by a financial institution. Problems securing a large loan may prevent a buyer from placing a competitive bid on a desirable property. Oncebiddershave multidimensional privateinformationcomposedofavalue-signalwhich informs their preferences, and a budget, which defines their feasible strategy set, even simple auctions feature nontrivial equilibrium behavior. For concreteness, consider a first-price, sealed-bid auction for one item. Bidders simultaneously submit their best offer to the seller and the highest bidder wins. Only the winner pays her own bid. In this straightforward setting, private budgets introduce two effects on equilibrium bidding. Budgets have a direct effect constraining bidders; however, they also introduce a strategic effect whereby highbudget bidders deliberately outbid high-valuation opponents who have the misfortune of being budget-constrained. The strategic effect emerges as a discontinuity in the equilibrium strategy and it endogenously stratifies competition within the auction along the budget dimension. These effects change the strategic interaction of the first-price auction with consequences for its efficiency, revenue potential, and bidders’ perception of the winner’s curse. Even the existence of a monotone equilibrium becomes a delicate matter. 1Inthis studywe investigate equilibrium bidding infirst-price auctionswith privatebudget constraints. Our analysis considers both the existence of a well-behaved equilibrium and the strategic implications of equilibrium bidding. While our general model accommodates interdependent and affiliated values and continuously distributed types, we begin with two examples highlighting the novelties in our environment. Examples 1 and 2 extend the classic Vickrey (1961) model.As we argue below, the issues they raise apply generally. Example 1. There are two risk-neutral bidders with private values. Each bidder receives a signal Si i.i.d.~ U[0,1] regarding her own value for the item. If a bidder with value-signal Si=siwins by bidding bi, her payoff is si-bi; otherwise, it is zero. Ties are resolved with a fair coin flip. Suppressing bidder subscripts, the function a(s) = s/2 is the symmetric Bayesian-Nash equilibrium bidding strategy. Suppose, additionally, that each bidder has a private budget wi, which is independent of her value-signal. With probability 1/2, a bidder has a budget of w= 1/4 and with probability 1/2it is wi=3/4. We assume that a bidder of type (si,wii)cannot bid above 1Textbook treatments and other extensions to Vickrey’s model are found in Krishna (2002) and Menezes & Monteiro (2005). Milgrom (2004, p. 132) studies an extreme variant of Example 1 where all bidders have a low budget with certainty. 2 ß(si, 3 )4 ) bi 12 ß(si, 14 14 0 '1 ˆsˆs si Figure 1: The strategy equilibrium in Example 1. High-budget bidders increase their bid discontinuously to 1/4at ˆs. Low-budget bidders must bid below 1/4. irrespective of si. Hence, a(si)=si Figure 1 presents a sketch of the new symmetric equilibrium ß(si,wi). Its definition is si i =3/4 if si ∈[0,ˆs] , 5+9s2 i ]and w =1/4 18+18si 1+9s '∈(ˆs,ˆs if si 6+18si ß(si,wi)= if si i ' 14 3 ' ∈ ( ˆ s wi /2can no longer be an equilibrium strategy since a low-budget bidder cannot afford to bid above 1/4. 2i if si ∈(ˆs,1]and w =1/4 i ,1]and w 2 =11/27. w The proof that ßis an equilibrium is very simple, but long. It is presented in here ˆs=1/3and ˆs online appendix. Example 1 conveys two main points concerning equilibrium behavior. First, the equilibrium strategy features a prominent discontinuity at ˆs. The intuition for this discontinuity rests on the strategic tradeoff made by a high-budget bidder in equilib A high-budget bidder always has the option of placing a bid above 1/4. Exercising option, which is worthwhile only for a bidder with a valuation greater than ˆs, substantially increases the probability of a win as she is guaranteed to outbid a low-budget competitor. Private budget constraints may therefore encourage more aggressive bidding among certain auction participants than would occur in a world without budget limits. The intuition for the discontinuity at ˆs'is similar; however, such bidders may tie with others rather than win outright. Second, although the discontinuities imply higher bids by some types of bidders, on ithe margin bids are often less brazen than if budget constraints were absent. In Figure 1, the slope of ß(·,w) declines moving past ˆs. The intuition here hinges on the endogenous stratification of competition along the budget dimension. For instance, bidders placing a bid above 1/4are competing on the margin only against other high-budget opponents. The decline in competition reduces the incentive to bid aggressively at distinct budget levels. A similar intuition holds for low-budget bidders when si'∈(ˆs,ˆs]. Although the above lessons appear to be specific to Example 1, or are perhaps an artifact of the discrete budget distribution, this is not the case. The strategic effects, and the associated discontinuities, can arise even when budgets assume a continuum of values. While the construction is more elaborate, and we detail it below, the intuition is fully captured by the reasoning in Example 1. The strategic effects seen in Example 1 are just one of the new issues raised by budget constraints. A more subtle matter concerns how the private wealth-effects associated with budgets may interact with preferences, a topic previously unexplored in the auction literature. Seemingly unintuitive effects may arise, with implications for equilibrium existence. For example, consider risk-aversion. Suppose that increasing a bidder’s wealth decreases her aversion to risk. In a first-price auction, decreasing aversion to risk encourages a bidder to bid less (Krishna, 2002, p. 38). Intuitively, the insurance with respect to the auction outcome from a higher bid is less valuable. Thus, a high-budget bidder may justifiably bid less than a counterpart with the same value-signal but a lower budget. This observation conflicts with the usual understanding of a monotone, nondecreasing strategy, which is at the core of the vast majority of auction studies. The following example, adapting the model of Cox et al. (1988), highlights this phenomenon. Example 2. Consider the model of Example 1. Again, Si i.i.d.~ U[0,1]and a bidder may have a budget of 1/4or 3/4with equal probability. Unlike Example 1, suppose a winning bidder of type (si,wi) receives a payoff of (si-bi)wi+1 4. Payoffs following a loss are zero. While contrived for tractability, these preferences capture the idea that bidders with a large budget are less risk-averse. A high budget implies risk-neutrality while a low budget implies strict 4 ˆs'' bi ˆs' 1 ˆs ß(si, 3 )4 ) 12 ß(si, 14 14 si 0 Figure 2: Equilibrium strategy in Example 2. risk-aversion. The following strategy, sketched in Figure 2, is a symmetric equilibrium: v 11-3)s24(si2 i+33(19 v if si ∈[0,ˆs] ) = 2si +24( if si ''∈[0,ˆs] 3 11-63)+ v 11-3)2 ,ˆs] ß( si,3 4) = ,1] si22s2 i+13 v11-42 4(si+1) 16s3 i if si ∈(ˆs,1] , ß( si, 14 14 if si ''∈(ˆs if si '∈(ˆs ' and ˆsis presented in the online appendix. When si 5 ˜0.298. A verification that ßis an equilibrium and the precise definition of ˆs'is low, the w is dominant and a high-budget bidder bids strictly less budget’s effect on risk preferences here ˆs= v than a low-budget opponent. When siis large, the strategic effects noted already in Example 1 dominate. Consequently, competition becomes stratified along the budget dimension. The strategies exhibited in Example 2 question the existence of an equilibrium in monotone, nondecreasing strategies in auctions with private budget constraints. Investigating this question, albeit in an environment admitting more natural preferences and interdependent values, forms the starting point of our analysis in Section 2. Since monotone equilibria form the core of most applied auction studies, it is critical to investigate their existence in this arguably prosaic setting. We address this question with political deftness by redefining the notion of a monotone, nondecreasing strategy in a manner appropriate for our setting. Our definition accommodates the countervailing incentives associated with budget constraints. Using a recent equilibrium existence result from Reny (2011), our relaxed notion of mono''11-3, ˆs = 3 (4 v '11-3), tonicity allows us to conclude that an equilibrium with economically-meaningful properties exists. Roughly, bidders with larger value-signals always bid more in equilibrium. Those with larger budgets tend to—but need not necessarily—bid more in equilibrium. The equilibrium in Example 2 exhibits these features and they apply more generally. Except for the private-values case, passing the equilibrium to a continuum action space depends on an endogenous tie-breaking rule (which we discuss below), a common feature of auctions with multidimensional types and interdependent values. Whereas the existence of a “monotone” equilibrium is reassuring, it is alimited conclusion regarding equilibrium behavior. Thus, in Section 3 we specialize to a symmetric model with interdependent and affiliated valuations, as in Milgrom & Weber (1982), and we augment it with continuously distributed private budget constraints. Fang & Parreiras (2002) analyze the second-price auction in this setting and it extends the analysis of Che & Gale (1998). Example 1’s main features, such as discontinuous equilibrium strategies and endogenously stratified competition, can appear in this benchmark environment. For example, equilibrium bid distributions can be bimodal even when the underlying type-space is uniform and expost valuations have a unimodal distribution. The documented strategic effects seriously confound inference regarding model primitives as bidders often react non-monotonically to the presence of private budgets. A discussion of applications and a brief comparison with the second-price auction concludes our analysis. An appendix collects proofs and an online appendix presents extensions and discussions of a more narrow interest. 1 Literature 2Researchers have recognized the possibility of budget constraints in auctions for over thirty years;however, few studies treat both valuations and budgets as private information. Indeed, many studies examine budget constraints in multi-unit or sequential settings where a budget’s strategic importance is more transparent a priori (Pitchik & Schotter, 1988; Benoît & Krishna, 2001;Pitchik, 2009;Brusco & Lopomo,2008,2009). In such models, assumptions simplifying the information structure or the auction format are a practical necessity. Budgets feature in many internet-motivated applications, such as selling advertisements (Ashlagi et al. , 2010), and play a prominent role in spectrum auctions (Cramton, 1995; Salant, 1997; Bulow et al. , 2009). Thus, budget constraints are an economically-meaningful element of 2Rothkopf 6 (1977) and Palfrey (1980) are early discussions of multi-unit auctions with budgets. many auction markets. We contend that complex or multi-unit environments, like spectrum auctions, are not necessary to appreciate the strategic effects posed by budgets—“classic,” single item settings are sufficient, although surprisingly research here has been more limited. Early research by Che & Gale (1996a,b) considers first-price, second-price, and all-pay auctions in a commonvalue setting where only budgets are private information. In an important paper, Che & Gale (1998) discuss first-price auctions allowing for private budget constraints and explore some of the revenue implications. Their analysis was limited to the private values case with risk-neutral bidders. In studying auction revenues when bidders have multidimensional, independent types, Che & Gale (2006) develop techniques that can be applied to the case of budgetconstraints. Fang&Parreiras (2002,2003)study equilibrium bidding inatwo-bidder, second-price auction with private budget constraints allowing for affiliated and interdependent valuations. Their setting forms the starting point of our analysis in section 3. Kotowski & Li (2012) consider the all-pay auction and the war of attrition with private budget constraints in the case of N-bidders. Textbook discussions of basic models of auctions with budget constraints are included in Krishna (2002) and Milgrom (2004). 3The optimal auction design problem in the spirit of Myerson (1981) incorporating budget constraints is particularly challenging and has been considered under various guises (Laffont & Robert, 1996; Monteiro & Page, 1998; Che & Gale, 1999, 2000; Maskin, 2000; Malakhov & Vohra, 2008; Richter, 2011). Pai & Vohra (2011) offer the most comprehensive description of optimal mechanisms but restrict attention to discrete types. Auction design accounting for budget limits has also spurred a growing literature among computer scientists. Whereasmostoftheabovemodelsassume“hard” budgetconstraints, somepapersexplore expostdefaultorasymmetries inbidfinancingability(Zheng,2001;Jaramillo,2004;RhodesKropf & Viswanathan, 2005). While appropriate for some applications, such models may mask the strategic effects of budget constraints alone. Therefore, we treat budgets as a rigid bound for bids. Cho et al. (2002) suggest that overcoming budget constraints can encourage collusion or the formation of bidding consortia. We assume non-collusive behavior. 3Borgs et al. (2005) and Dobzinski et al. (2012) are representative contributions of the vast literature in this tradition. 7 2 A General Model and Equilibrium Existence 4Recalling the challenge posed by Example 2, we begin by investigating the existence of a “monotone” equilibrium.Wealsointroducenotationthatwillbeusedthroughoutthesequel. Our model rests on three assumptions. Assumption A-1 considers bidders’ preferences. Assumptions A-2 and A-3 introduce two complementary information structures. Each places some limit on the correlation in bidders’ private information, which is a practical necessity in this environment. We discuss both model components in turn. Preferences and Strategies There is a set of bidders N = {1,...,N} participating in a first-price, sealed-bid auctionfor one item. Each player has a private type θi=(si,wi)∈Si×W i=Ti. si=[si,¯siiis a bidder’s value-signal and it takes values in S]. A value-signal is a bidder’s information about the item for sale. wiis a player’s budget and W i=[wi,¯wi]⊂Ri. The budget defines an absolute spending limit above which a bidder cannot bid. Each bidder has available a set of valid competitive bids B⊂[ri+,8)and a non-competitive null bid l<mini{ri}, which guarantees a loss in the auction. Bids are placed simultaneously and the bidder submitting greatest competitive bid wins the item. Only the winner makes a payment to the auctioneer. We adopt standard notation. Signal profiles are in boldface, as in s=(s1), and θ=(θ1,...,θNN i=1)is a profile of types. We let S =×Si,...,sNand we define W and Tsimilarly. As usual, s-i=(s1,...,si-1,si+1,...,sN)and S-iis defined analogously. Our first assumption concerns a bidder’s utility function. Assumption A-1. Whenever ri =bi =wi, the payoffs of bidder isatisfy the following: bounded, continuous and decreasing in bi, nondecreasing in1(s,wi), and differentiable. Moreover, ∂¯ui/∂si=κi>0. . ¯ u i ( s , w i , b i ) : S × W i →Ris bidder i’s payoff when she does not win the auction. ui i i > ¯ B, ¯ui(s,wi,b ) <ui(wi)for all (s,wi). 2. ui(wi):W i is bounded, nondecreasing, and differentiable. i 3. There exists ¯ B=r such that for all b =0, then for all (s,wi), ¯ui(s,wi,bi)=ui(wi). 4. If bi Establishing the existence of an equilibrium in auction games is a research program spanning several 4 decades. deCastro&Karney(2012)presentacomprehensivesurveyofthis(now)large,specializedliterature. 8 5. For all bi >b' i =ri, ¯ui(s,wi,bi)-¯ui(s,wi ∂¯ui =sup dui <8. 6. inf dwi ,b' i)is nondecreasing in (s,wi). ∂wi A-1 adapts standard conditions from Maskin & Riley (2000), Reny & Zamir (2004), or McAdams (2007), among others, to our environment. While the differentiability requirements are stronger than typically encountered, they greatly simplify much of the following argument. The lower bound on ∂¯uin A-1(1) is similar in our setting to Assumption A2(v) from Athey (2001). A-1(3) says that bidders do not wish to spend an unbounded amount and A-1(4) says a bidder would happily take the item for free. A-1(5) is an increasing-differences condition common to the literature. A-1(6) is a technical condition which is not restrictive. Its role will be apparent shortly. To simplify our notation, we write ui(s,wi,bi)=¯ui(s,wi,bi)-ui(wii/∂si). As highlighted by Example 2, when bidders’ preferences are sufficiently general, equilibrium behavior may not conform with traditional conceptions of monotonicity. To accommodate the behavior suggested by Example 2, we will amend the standard notion of a monotone strategy. Reny (2011) pioneered the weakening of the definition of a monotone strategy in games of incomplete information and our analysis closely parallels his examination of the multi-unit auction with risk-averse bidders. The analogy between our model of a single-item auction with private budget constraints and a multi-unit auction is quite clear. An auction with private budget constraints can be interpreted as an unusual multi-unit auction: the highest bidder wins the item for sale and all bidders “win” their budget irrespective of their feasible bid. As a first step in (re)defining a monotone strategy, we introduce an alternative partial order on the type-space. This ordering relates changes in a bidder’s information to changes in her utility. ∂¯ui . = sup dui -inf∂¯u∂w i Definition 1. Let ξi inf∂s dwi i i on Si ×W i as (si,wi)=θi 9 (s' i,w' i) ⇐⇒ wi i(wi i =w' i and si i -s' i i i,bi Define the partial order =θi = ξ-w' i). Figure 3 sketches the greater- and less-than sets for type. An increase in si ) = v(s)+wi ˆθ increases, si alone implies a higher type. If wmust increase commensurably to imply a higher type. If bidders have prototypical quasilinear preferences, i.e. ¯u(s,w -b an d wi θ' =θi ˆ θi i ¯wi ˆ θ i wi si Figure 3: The =θi ordering . =0 and =θi reduces to the usual coordinate-wise ordering of R2. We can now defin u (wi)=wi, then ξi strategies of interest. Definition 2. An admissible (pure) strategy for bidder iis a measurable f B∪{l}suchthatßi(si,wi)=wi. Astrategyisnondecreasingifθi=θiθ' i=⇒ ßi(θii:T) 2weakens the usual definition ofamonotone, nondecreasing strategy. A budget and a sufficiently higher value-signal imply a higher bid. Althou outside of our general model, its equilibrium reflects this intuition. The Information Structure Our analysis studies in parallel two distinct, information structures. Both have natural motivations and particular applications will dictate th specification. The first posits independent types across bidders but allo between a bidder’s value-signal and budget. The second allows for som dependence among value-signals, but demands budgets be determine Our first environment considers a setting where bidders’ types are i However, a bidder’s own value-signal and budget may be correlated. C conduct their analysis in a special case of this environment. Assumption A-2. The joint density of players’ types is f(θiand θjare inde f(θ) is strictly positive, bounded, and for all i=j, θ 10 5While a tractable formulation, the independence of bidders’ private information is often untenable in applications. Therefore, we propose a complementary second environment that allows for some statistical dependence among bidders’ value-signals in the form of affiliation.Even though affiliation is already a restrictive property (de Castro, 2010), it is often insufficient to guarantee the existence of an equilibrium in nondecreasing strategies. Therefore, in Assumption A-3 we introduce an additional limit on the “relative degree” of affiliation. Similar limits have been fruitfully exploited by Fang & Parreiras (2002), Krishna & Morgan (1997), and Lizzeri & Persico (2000), among others. To state Assumption A-3, we first introduce some notation. If h(s)is a density function, we write H(s-i|si) = t-i=s-ih(t-i|si)dt-iand ¯ H(s-i|si) = t-i=s-ih(t-i|s. Also, define λ(s-i|si)=h(s-i|s) ¯H(s-ii|si)and s(s-i|si)=h(s-i|s) H(s-ii|si)i)dt-i. λand sare (conditional) hazard-rate and reverse hazard rate functions. Assumption A-3. The joint density of players’ types, f(θ1,...,θN)=h(s)g(w), is strictly positive, bounded, and it satisfies the following conditions: 1. Value-signals are affiliated, but independent of budgets. Their joint density is h(s). Moreover, if (si,wi)=θi(s' i,w' i)and ri=bi=min{ ¯ B,w' i}, then for a.e. s-i (a) ui(si,s-i,wi,bi)λ(s-i|si)=ui(s' i,s-i,w' i,bi)λ(s-i|s' i), and (b) ui(si,s-i,wi,bi)s(s-i|si)=ui(s' i,s-i,w' i,bi)s(s-i|s' i) with strict inequalities if si >s' i.6 2. Budgets are mutually independentand independentof value-signals. Theirjointdensity is g(w). Remark 1. Assumption A-3 is satisfied if value-signals are independent. It can also be satisfied when value-signals are strictly correlated. For example, consider N = 2, r= 0, ui(s,wi,bi)=1 2(si+sj)-bi, and h(s1,s2)=4 5(1+s1s2iθi). By continuity, it will also hold in sufficiently nearby models where the =ordering has bite. While assumptions akin to A-3 have appeared before, our formulation is slightly more restrictive owing to A-3(1b). Intuitively, A-3(1a) and A-3(1b) ensure that the direct influence of sion preferences dominates its informational content concerning s7-i. Consequently, interim expected utility is nondecreasing in a bidder’s own value-signal. 5Weusethepropertiesofaffiliatedrandomvariablesfreely. Milgrom&Weber(1982)offeranintroduction. A-3(1b) was absent from this study’s previous versions and we introduce it here. 7A-3(1a) and A-3(1b) facilitate verification of a single crossing condition when value-signals are affiliated 6Condition 11 Equilibrium Existence In sufficiently general auction models, the existence of an equilibrium depends intimately ion the prevailing tie-breaking procedures (Jackson et al. , 2002). As Araujo et al. (2008) discuss, this is particularly true when bidders’ private information is multidimensional— as in the case of private budget constraints. We therefore offer two statements regarding equilibrium existence. The first statement concerns a setting where the set of each bidder’s valid bids, B, is set so as to preclude relevant ties. Subsequently, we discuss equilibrium existence when ties can occur and tie -breaking procedures matter. Let ϕi(bi,b-i)be the probability that bidder iwins the auction with a bid of bgiven that other bidders bid b-i. Following tradition, we suppose ϕiiconforms to uniform tie-breaking and ϕi(l,b-i)=0for all b-i. Thus, bidder i’s (interim) expected payoff at the bid biis Ui(bi,ß-i|θi)=ui(wi)+ T-i ϕi(bi,ß-i)ui(s,wi,bi)f(θ-i|θi)dθ-i. By following the strategy , bidder i’s (ex-ante) payoff is Ui(ßi,ß-i)=E[Ui(ßi(θi),ß-i)|θ)]. The strategy profile ßi -i) is a Nash equilibrium if for every i, Ui(ß* i,ß* -ii) = Ui(ßi,ß* -i) for all admissible ßi. We can now offer our first result concerning equilibrium bidding. Theorem 1. Suppose that for each i, Bi is finite, non-empty, and ∀i=j, Bi nBj =Ø. 1. If A-1 and A-2 are satisfied, there exists an equilibrium in nondecreasing strategies. 82. If A-1 and A-3 are satisfied, there exists an equilibrium in nondecreasing strategies. The proof of Theorem 1 rests on ensuring that our model satisfies the conditions outlined in Reny (2011). We already mentioned the analogy between our model of an auction with budget constraints and his analysis of the multiunit auction. Readers familiar with that argument will recognize the key lemmas in our analysis. When bidders place bids from a continuum, relevant ties are a serious technical concern andthetie-breakingprotocolmatters. Regrettably, theuniformtie-breakingruleisnoteasily accommodating of an equilibrium. We explain why this is the case below. The following theorem outlines some of the possibilities when bidders place bids from a continuum. while computing an expectation conditional on T⊂T. Without these conditions, Tadditionally needs to be a sublattice of T-i-i, which is not the typical case in our application. 8In the online appendix we show how Theorem 1(2) obtains without Assumption A-3(1b) when there are two bidders with quasilinear preferences. 12 Theorem 2. Suppose that for all i, Bi =[ri,8). Theorem 1 continues to hold if either: 1. Bidders have private values and ties among high bidders are resolved with a uniform randomization; or, 2. The standard allocation rule with uniform tie-breaking, ϕi, is replaced by an “endogenous” tie-breaking rule, which depends on an announcement of bidder’s private information to resolve ties. 9Endogenous tie-breaking rules are the central focus of Simon & Zame (1990) and Jackson et al. (2002). For convenience, the tie-breaking rule in part 2 calls upon the construction of Araujo & de Castro (2009), which they originally propose for a double-auction setting. Their construction carries over naturally to our environment as well. Roughly, a tie-breaking rule is defined with respect to a sequence of monotone equilibria from near-by games where the allocation rule is smoothed with slight perturbations. The rule ensures continuity of payoffs facilitating the existence of an equilibrium. de Castro (2011) presents a recent discussion of approximating games in this manner and the implications for equilibrium existence. He also places such methods in the wider context of equilibrium existence in discontinuous games. 10Jackson (2009) discusses special tie-breaking rules in auctions and we refer the reader to his thorough assessment. In our environment, the need for special tie-breaking protocols stems from the indecision experienced by bidders should they tie for the winning bid. Standard arguments regarding equilibrium existence in first-price auctions depend critically on a bidder’s desire to bid infinitesimally more to break a hypothetical tie in her own favor. Although obvious in a private-value setting, winner’s curse effects make this desire not obvious once valuations are interdependent.Absent strong assumptions, it is not generally true in a setting with budget constraints, even with two bidders, since winner’s curse effects are more intricate. To sketch some of the complications, suppose bidder ibids band bidders j=iare each following a nondecreasing strategy ßj. Bidder i’s bid defeats two kinds of opponents. First, it defeats all opponents who have a budget less than b. This event is good news concerning the item’s true value. It is a “winner’s blessing” to defeat budget-constrained opponents since their value-signals will be of average quality and not negatively selected. Second, the bid defeats a subset of opponents who may have a budget greater than bbut a relatively 9Important contributions in this literature include Dasgupta & Maskin (1986) and Reny (1999), among others. 10Reny & Zamir (2004) prove this as part of their analysis. 13 low value-signal. With regard to this group, winner’s curse effects are more acute and a tie will disproportionally implicate this second class of opponents. Depending on the situation’s details, bidder imay prefer to win or lose such a tie. The “relatively-low” value-signals may or may not be sufficiently good to warrant a slightly higher bid. The concerns highlighted above aretypical ofmultidimensional auctionenvironments and are mainly due to value interdependence. However, value interdependence alone does not preclude equilibrium existence under standard tie-breaking rules. The environment’s asymmetries alsoplayarole. Aswe documentbelow, inabenchmark symmetric environment with continuously-distributed types and interdependent values, we can identify an equilibrium in nondecreasing strategies where ties are zero-probability events. Thus, the tie-breaking rule is superfluous. More interestingly, however, the equilibrium strategies we identify exhibit the qualitative features of Example 1—such as discontinuities and an endogenous stratification of competition—suggesting that these economic phenomena are a recurring feature of bidding in budget-constrained environments and deserve careful scrutiny in applied studies. 3 A Symmetric, Two-Bidder Model Although the existence of a well-behaved equilibrium is reassuring, a question remains to what degree Example 1’s economic features carry over to a more general environment. In this section we provide one extension by allowing bidders’ types to assume a continuum of values. While our discussion is necessarily incremental, in case 3 below we construct a discontinuous, monotone, pure strategy, symmetric equilibrium in a symmetric first-price auction. Although unusual, its intuition is fully reflective of the equilibrium from Example 1. Such equilibria can introduce serious confounds in empirical applications. For example, even when the type-space is uniform and valuations are unimodal, bimodal equilibrium bid distributions can arise. iWe maintain assumptions A-1 and A-3. (However, A-3(1b) is not necessary for our subsequent analysis). Thus, our environment is one of affiliated, interdependent values. We specialize the model further by supposing that there are two ex ante symmetric bidders, value-signals take on values in S=[0,1], budgets are in W i=[w ,¯w], and there is no reserve price. We briefly address reserve prices before the conclusion. Assumption A-4 summarizes the key additional restrictions. Assumption A-4. Suppose the environment satisfies the following conditions: 1. For all i, ¯ui(s,wi,bi)=v(si,sj)+wi-bi, ui(wi)=wi, v(0,0)=0, and v(1,1)=¯v. 14 2. The density h(s1,s2)is permutation symmetric and continuously differentiable. 3. Bidders’ budgets are distributed i.i.d. according to the c.d.f. G(·). G(·)is C 211, concave, and its support satisfies w =¯v= ¯w. A-4imposesadditionalsymmetryonpreferencesandontheinformationstructure. Therefore, it builds on Milgrom & Weber (1982)’s classic setting. Like Fang & Parreiras (2002) we restrict attention to two bidders for expositional clarity. A-4(3) is stronger than considered in Fang & Parreiras (2002) and the concavity condition fulfills the same role as Assumption 5 from Che & Gale (1998). It is a sufficient condition for our analysis, which can be weakened in specific cases. The discussion below is organized into three cases depending on w . In case 1, wis low and the equilibrium is defined primarily by a differential equation. However, identifying the appropriate boundary condition requires care. This case also introduces concepts that help identify the remaining two scenarios. Second, if w is large, the equilibrium is a constrained extension of the equilibrium strategy without budget constraints. Finally, when w is intermediate, equilibrium bidding features a discontinuity in the spirit of Example 1. In equilibrium, bidders endogenously separate into two groups—those with relatively large budgets and others with small budgets. Unlike Example 1, the equilibrium bid distribution has no mass points and has connected support. This case highlights the tension between the direct and strategic effects introduced by private budget constraints. Bids become depressed by budget limits; however, the chance to exploit others’ (possible) spending limit is a strategic option encouraging more aggressive bidding. Although our cases are defined in reference to w , we hope that our discussion makes clear that the qualitative effects we identify can easily emerge in more general circumstances. While nontrivial, the equilibria we study are likely among the tamest that may emerge in further generalizations of our model. Since we focus onsymmetric equilibria, we alsoeconomize onsubscripts by presenting our model from bidder 1’s view-point. We let (S,W)be her value-signal and budget respectively. We denote bidder 2’s value-signal and budget pair with (Y,Z). Finally, since it appears frequently, recall from Milgrom & Weber (1982) that in the corresponding model without budget constraints the equilibrium bidding strategy a(s)is the solution to a'(s)=(v(s,s)-a(s)) h(s|s)H(s|s) , a(0)=0. (1) 11To limit subscripts, hereafter section we write g(·)for the density of the budget distribution of a single bidder. In Section 2, g(·)was the joint density of all bidders’ budgets. No confusion should result. 15 Cas e 1: Lo w Min ima l Bud get s (w˜ 0) In a bro ad sen se, the foll owi ng ana lysi s app lies wh ene ver w is “lo w” (de fine d bel ow) , but for con cret ene ss ass um e w= 0. To con stru ct an equ ilibri um, our app roa ch is to “gu ess , ana lyze , and veri fy.” Spe cific ally, sup pos ea sy mm etri c equ ilibri um stra teg y has the for m ß ( s wh ere ¯ b(s) —t heb idpl ace dby abi dde roft ype (s,¯ w) —is diff ere ntia ble and stri ctlyi ncr eas ing. Thi s stra , w ) = m i n { ¯ b ( s ) , w } ( 2 ) teg y imp lies Leo ntie f “iso bid” cur ves in (s, w)spa ce with ¯ b(s) defi nin g the loc us of ki n kp oi nt s. Si n c e C h e & G al e (1 9 9 8) , (2 ) h a s b e e n th e ty pi c al st ra te g y c o n si d er e d in th e lit er at ur e. It is m o n ot o n e in th e s e n s e of d e fi ni ti o n 2. If th e o p p o si n g bi d d er is fo ll o wi n g (2 ), th e e x p e ct e d p a y o ff of a bi d d er w h o bi d s ß( x, ¯ w )= ¯ b( x) is (v(s,y)¯ b(x))h(y|s)g(z)dydz+ Ui0(ß(x,¯w)|s,¯w)= ¯ b(x) 0 (v(s,y)¯ b(x))h(y|s)g(z)dydz. 0 1 ¯ w ¯b(x) x ¯ b' The first term is the contribution to expected utility from defeating all opponents with a budget z=¯ b(x). The second term is the contribution to expected utility from defeatingall opponents with a budget z> ¯ b(x)but who have a value-signal y=x, and are thereforebidding less than ¯ b(x). Computing the necessary first-order condition,∂Ui(ß(x,¯w)|s,¯w) ∂x =0, gives: d(b,x|s)=b+ G(b)g(b) + γ(z)= g(z) dy, h(y|s) 1-H(y|s) . (3) v(s,y)h(y|s) , η(x|s)= x1 1-H(x|s) H(x|s) g(b)(1-H 1-G(z) , λ(y|s)= ¯ 1 ¯ b(s)-v(s,s))(s)= λ(s|s)(γ( ¯ b(s))(η(s|s)-d( b(s),s|s)) 6 and simplifying notation gives ¯ b' , We make two observations. First, it is not clear that (3) has a solution satisfying the desiderata asserted above. Notably,¯ b(s)must be strictly increasing for all swhile the sign of (3) is unclear. Second, (3) gives little guidance concerning boundary conditions. With interdependent values,¯ b(0)>0may be a reasonable situation. Conditional on s=0 and on winning the auction with a small bid, the item may have a positive expected value if the losing opponent was budget constrained with high probability. Toaddresstheseissuesweproceedtodevelopanunderstandingofthequalitativebehavior of solutions to (3). Specifically, rather than analyzing (3) directly, we will study instead the closely-related plane-autonomous system ˙s =γ(b)(η(s|s)-d(b,s|s)) ˙b =λ(s|s)(b-v(s,s)) . (4) Following Birkhoff & Rota (1978, Chapter 5) we can recover solutions of (3) by noting that d = dbdt d = . b s ˙b(s,b) d dt ˙s(s,b) s Interpreting the problem in this manner allows for the use of elementary techniques from phase-plane analysis to discern the qualitative behavior of candidate solutions and has the advantage of working around the inherent singularities associated with a direct analysis of (3). We emphasize that the direction of motion as a function of time in (4) is immaterial for our purposes. Rather, we are interested in the entire paths taken by solution orbits, which (locally) are integral curves of (3). Our goal is to identify solutions to (4) that can be represented by a function meeting the desiderata imposed above on¯ b(s). This section’s 12remainder undertakes this task. To describe the behavior of a plane-autonomous system, we are initially interested in two related objects: the nullclines and the critical points. The nullclines of (4) are loci in (s,b)spacewhereorbitsarehorizontalorvertical. Respectively, thesesetsareν={(s,b):˙ b(s,b)=0}and ψ={(s,b): ˙s(s,b)=0}. It will prove convenient to express νand ψas functions of sin the following manner: ν(s)={b:(s,b)∈ν}and ψ(s)={b:(s,b)∈ψ}. By inspection, when ν(s)=Ø, ν(s)=v(s,s). Similarly, when ψ(s)=Ø, ψ(s)is single-valued and is defined 12We note that ˙sand ˙ bare notdefined atthe boundary whereb= ¯wand s=1, respectively. This technical point is not a concern since we will be able to extend the identified solution to the boundary via (3) using standard results in the theory of ordinary differential equations. For expositional clarity we suppress this technical detail. 17 13implicitly by η(s|s)=d(ψ(s),s|s). Moreover, if Sψ ={s:ψ(s)=Ø}, the function ψ(s) if s∈Sψw if s/∈Sψ * *) is 15 a critical point if ˙s(s* * b(s*,b 2. If (s*,b 1 4 i s ˜ ψ(s)= and the function ˜ ψ(s)will play an im ˙*)=0. When w =0, there exists at least one critical point.The follo the behavior of the (˙s,˙ b)system around the nullclines and the T Lemma h 1. Consider the system (˙s, ˙ b)defined in (4). 1. If eb>(<)ν(s), then ˙ b(s,b)>(<)0. If b>(<)ψ(s), then ˙s(s,b)<(>)0 c o n t s i e n t u o S u ψ s . * 16)is a critical point, then it is (generically) either a node or a saddle point. Critical points occur where nullclines meet. (s ,b ,b Aconsequence ofLemma1isthatthereexist solutionsof (4)thattendtowardoremanate from each critical point. In the case when the critical point is a saddle point, these solutions would correspond to the point’s stable and unstable manifolds. In the case of a stable or unstable node, there would be infinitely many such solutions. Going forward, we introduce the following assumption to simplify our presentation. Assumption A-5. There exists at most one critical point. The online appendix extends ouranalysis to the case ofmultiple critical points. However, the substantive conclusions we reach carry over to the more general case. A sufficient condition ensuring Assumption A-5 holds is that ψ(s) is non-increasing. We are unaware of any economic interpretation to this sufficient condition. Leveraging Assumption A-5, we can construct a qualitatively accurate phase portrait for (4). Figure 4 sketches a generic situation. The nullclines ψand νpartition the space into four regions—R1,...,R4—that meet at the critical point (s*,b*). We plot several sample solutions and the thick arrows indicate the direction of the system’s solutions in each region. When there is one critical point, it is a saddle point. (In the case of multiple critical points, they alternate between saddles and nodes as sincreases.) G(·)is not concave, ψ(s)may not be single-valued. Much of the analysis we develop carries over to this situation, but many more cases need to be considered. 14See Lemma B9 in the appendix. 15See Lemma B10 in the appendix. 16See Remark B3 in the appendix. 13When 18 b* ¯ b0 ν R4 b ¯ b1 s1 R1 R2 R3 s* 0 ψ Figure 4: Phase-portait of the (˙s, ˙ b) system defined in (4). There is one critical point at ( ,b*). It is a saddle point. Together, the (thick, black) stable manifolds define ¯ s b(s). * Figure 4 also informs our understanding of ¯ b(s). This function was hypothesized to be continuous, strictly increasing, and satisfying (3). If such a function exists, it must begin in region R1, “pass through” the critical point (s*,b*)and continue in R. Only in regions R1and R33is ˙ b/˙s>0. Our preceding analysis also confirms that the only possible paths consistent with these requirements correspond to the stable manifolds of (4), illustrated in Figure 4 as paths tending toward (s*,b*) from the points (0, ¯ b0) and (1, ¯ b1). Let ¯ bbe the ¯ b}. (5) closure of the union of these two paths and let ¯ b(s)={b:(s,b)∈ Defined in this manner, ¯ b(s)is continuous, strictly increasing and an integral curve of (3). It has an endogenously determined initial condition ¯ ¯ and terminal condition ¯ b b(1)= b(0)= 0 ¯ b 1 is routine. symmetric equilibrium. . Having identified a function¯ b(s)that meets our requirements, proving the following theorem Theorem 3. Suppose w =0, then ß(s,w)=min{ ¯ b(s),w}where ¯ b(s)is defined by (5) is a The following example illustrates our discussion thus far. The example corresponds (via scaling) to one used by Fang & Parreiras (2002) to illustrate their model of the second-price auction with budget constraints. 19 Υ 1b b Α 0.5 Ψ 0 0.5 1 s Figure 5: Equilibrium sketch in Example 3. There is a single critical point at the intersection of ν(s)and ψ(s). Example 3. Suppose v(si,sj)=si +sj , W i i.i.d.~ U[0,2], and Si i.i.d. ¯ . ¯ b(s)) (1+s)(3s-1)-4(s-1) ¯ b(s) b ' + 3s 8 v 5and ν(s) = 2s. The critical point is a 1 5We can show that ψ(s) 5 4+ 1 s-1 0 4 =(5-2 v-2 5( v ( v ~ U[0,1]. Some algebra gives (s)= 2(2¯ b(s))(2s- 5) ˜ 0.105. The eigenvalues of the Jacobian of (˙s, ˙ b) at the critical point are5+ v205) ˜ -6.62 and2 5205v5) ˜ 4.83. Therefore, the critical point is a saddle point and the stable manifolds approach it at a slope of˜0.85. Numerical approximations allow us to conclude that¯ b˜0.1202while ¯ b 13 v 5+v 205˜0.878. Figure 5 plots¯ b(s)and a(s)=s, which is the equilibrium in the same model but with no budget constraints. The nullclines and representative solutions from the associated planeautonomous system are included for context. This example highlights the attenuation of the winner’s curse: low value-signal bidders with relatively large budgets bid more than in the same environment absent budget constraints. When w >0, two factors may complicate the preceding analysis. First, the expected payoff from the bid¯ b(0) may be negative—a fact inconsistent with equilibrium behavior. 0,w )for some >0. ,b* s0 20 Second, the manifold tending to (s ) from below may not originate on the vertical axis as it does in Figur its initial condition could be (s In either case, Theorem 3 cannot apply; otherwise, we have the following definition and corollary. Definition 3. w is low if a critical point exists, s0=0, and η(0|0)-w=0. Corollary 1. If w is low, Theorem 3 continues to hold. Proof. The preceding analysis continues to apply. As¯ b(s)depends on G(·), the equilibriumchanges with w , but it is still of the form ß(s,w)=min{ ¯ b(s),w}. Analysis of the subsequent two cases builds on the preceding results. Notably, the function¯ b(s):[s0,1]→[w ,¯w]defined as the closure of the union of the upward-sloping (stable) manifolds passing through the critical point (s*,b*)will reappear. ¯ b(s)is reserved as notation for this function. Case 2: High Minimal Budgets (w »0) If budgets are likely to be large relative to most valuations, a strong intuition suggests that they should matter strategically only for high-valuation bidders. In Example 1, for instance, low-valuation bidders (s<ˆs) adopt the strategy a(s)bidding as if budgets are strategically irrelevant. Extending this observation to the general model is relatively straightforward. When w is sufficiently large all bidders with a value-signal below some threshold ˜swill bid according to the strategy a(s), defined in (1). Bidders with a value-signal s>˜s, will bid min{˜ b(s),w}where˜ b(s)is an appropriate solution to (3). Definition 4. w >0is sufficiently large if a-1(w) /∈Sψ. The nomenclature “sufficiently large” is motivated by the limiting case w = a(1) which ensures definition 4 is satisfied in practice. However, noting Example 4 below, w can be sufficiently large in less-trivial instances as well. The following theorem describes equilibrium bidding in this case. Theorem 4. Suppose w is sufficiently large and let ˜s=a-1(w). Let ˜ b(s)be the solution to a(s) if s=˜s min{˜ (7) b(s),w} if s>˜s ˜ ˜ b(s)-v(s,s))(s)= λ(s|s)(γ( ˜ , ˜ b(˜s)=w . (6) b b(s))(η(s|s)-d(˜ b(s),s|s)) ' Then , ß(s,w)= 2 1 is a symmetric equilibrium. Unlike case 1, the desired solution to the main differential equation (6) bypasses the critical point (if such a point exists at all). Abusing terminology, we can say that˜ b(s) extends a(s) into the range of bids above w and as a result ß(s,w) is continuous. ˜ b(s) accounts for the change in the marginal effectiveness of bidding due to budget limits. Bids above w defeat opponents who have a low value-signal and a relatively low budget. Locally, this encourages more aggressive bidding by bidders with relatively large budgets. The next corollary and example make this effect precise. i such that for all bidders of type (s,w), with s∈(˜s,s'˜ b(s)=a(s). Corollary 2. There exists s' i i.i.d. ˜ b' suppose w>0, a-1(w)∈Sψ -1 )and w sufficiently close to ¯w, ß(s,w)= Example 4. Again, suppose ,sj)=si +sj and Si i.i.d. v(s ~ U[0,1]. Unlike Example 3, suppose w =1/2and thus W~ U[1/2,2]. The differential equation describing ˜ b(s)reduces to ˜ b(s))(2- ˜ b(s)) . (s)= 2(2s' of Corollary 2 holds with strict inequality. 4(1-s) ˜ b(s)+s(3s+2)-2 Since a(s)=s, ˜s=1/2. Figure 6 depicts the functions ˜ b(s),a(s), and ν(s). The conclusion Case 3: Intermediate Minimal Budgets The intermediate case is defined in opposition to the low- and high-w cases. Therefore, , and that neither of the above cases applies. In this situation, the preceding theory is inadequate to describe equilibrium behavior. For instance, at (˜s,w ), ˜s= a(w ), the differential equation (6) evaluates to ˜ b<0 precluding an increasing extension as in the high-wcase. Such features prompt consideration of discontinuous bidding strategies similar to Example 1. Although the equilibrium strategy we introduce in this section is the natural analogue of the strategies we defined in the preceding cases, its definition is comparatively intricate and we must take a brief detour to provide motivation. By analogy to Example 1, we posit that bidders will bid a(s)when their value-signal is sufficiently low. At some critical value, ˆs, high-budget bidders will increase their bid discontinuously. To identify the location and 22 1b Υ b w 0.5 Α 0 0.5 1 s Figure 6: Equilibrium sketch in Example 4. Given the example’s parameters, there is no critical point. the magnitude of this discontinuity we begin by defining the function µ:[0,1]→[w ,¯w]as µ(s)= w if s<smin{ ¯ b(s),0˜ ψ(s)} if s=s0if there is a critical point. If there is no critical point, ¯ b(s)does not exists and we simply let µ(s)= ˜ ψ(s). The technical motivation for µ(s) is that when µ(s) >w the differential equation ˆ'b ˆ b(s)-v(s,s))(s)= λ(s|s)(γ( ˆ b(s))(η(s|s)-d(ˆ b(s),s|s)) will admit a strictly increasing solution corresponding to the initial condition (s,µ(s)). As beforehand, this follows by examining the system (4) and extending a solution through the critical point if necessary. Hence, {(s,µ(s)):µ(s) >w } is a natural set of candidate “landing-points” for the discontinuous bid increase. To entertain a discontinuous bid increase from a(s)to µ(s), requires a bidder to be just 23 indifferent between the two options. To identify such cases consider the inequality = sUtility from bid of a(s) (8) (v(s,y)-µ(s))h(y|s)dy+G(µ(s)) s1 Utility from bid of µ(s) -1 a(s), if s=ˆs b (s) if . (9) s>ˆs,w<f(s) min{ 1ˆ b(s),w} if s>ˆs,w=f(s) γ( ˆ b(s))(η(s|s)-d(ˆ b(s),s|s)) ß(s,w)= ' subject to ˆ b(ˆs)=µ(ˆs). 1 (v(s,y)-µ(s))h(y|s)dy and define the set Z ={s=a s 0 0(v(s,y)-a(s))h(y|s)dy (w ):(8) holds}. In the above inequality, the utility from the bid µ(s) is defined under the conjecture that the other bidder also increases her bid discontinuously to µ(y) when her value-signal is y= sand that higher types bid strictly more. Of course, in a symmetric equilibrium this conjecture will prove correct. Lemma B11 in the appendix confirms that when w is intermediate, Z =Ø. We define ˆs=supZ. At this point, (8) holds with equality. Like in Example 1, ˆsis the value-signal of the lowest type who increases her bid discontinuously. With ˆsdefined, we can introduce the strategy of interest, which we sketch in Figure 7. We provide intuition motivating this strategy after its definition. Let The value ˆsis defined as above, a(s)is defined in (1) and: • ˆ b(s)is the solution to ˆ bˆ b(s)-v(s,s))(s)= λ(s|s)( • b 17 (s)is the solution to (s)= (v(s,s)-b b' 1 subject to b1(ˆs)=a(ˆs). G(f(y))h(y|s)dy (11) (10 ) 1(s))h(s|s)G(f(s)) H( From the arguments in the low-wcase, ˆ b(s) exists and is strictly increasing and continuous. We can 1 employ the (˙s, ˙ b)system to define the appropriate solution as beforehand. When there exists a critical 7 point ˆ b(ˆs)=µ(ˆs)and extends it to values s>s* 24 and ˆs=s*, ¯ b(s)solves (10) with the initial condition . The online appendix discusses constructingˆ b(s)when there are multi b min{ ˆ b(s),w} sw¯w ˆs1 µ(ˆs)ˆs w' Figure 7: ß(s,w)when wis intermediate. • f(s):[ˆs,1]→[w ,µ(ˆs)]is a non-increasing, continuous function such that µ(ˆs) b1(s ) w a(s) = G(f(s))(1-H(s|s)) f(s)-b1(s) G(f(y))h(y|s)dy , (12) η(s|s)-f(s) 1 (ˆs' f ( ˆ s ) = µ ( ˆ s ) , a n d f ( ˆ s ' ) = b b i d w i n s w it h g r e a t e r p r o b a b il it y . S e c o n d , b e 'Thedefinedstrategyhasadis c continuityalongtheset{(ˆs,w) a :w=µ(ˆs)}∪{(s,f(s)):s∈[ˆs,ˆs+u) }. Importantly, however, s ß(·,w ):[0,1]→Ris e continuous and ß’s range is a an interval. The intuition b underlying this strategy i echos Example 1. Biddersd who increase their bid o discontinuously at ˆsare f indifferent between the ˆ higher bid µ(ˆs)and the b bid a(ˆs). The larger bid ( ˆ is attractive for two reasons. First, a higher s ) bests H many bidders with low budgets, it has the added benefit of ( attenuating the winner’s curse. These reasons encourage a bidder with a value-signal ˆ than ˆsto favor the much higher bid. greater s Ifbidders | with abudgetofw= ˆ b(ˆs)bidstrictly above w , bidders with lower budgets mustsomehow respond. Consider the problem now facing a bidder with a value-signal s ˆs=ˆs+o and a budget ˆwoo= ˆ b(ˆs)-o. Such a bidder is tempted to bid her entire ) endowment and so asto benefit from the two positive effects noted above. However, sh + does not compare bidding a(ˆso)versus ˆwoto see if the high bid is indeed worthwhile. Th fact that ) all opposing types with a budget in excess ofˆ b(ˆs) are already outbidding her changes = her incentives—she is facingwless effective competition on the margin if she bids low. The same effect is seen 25 f o r s o m e ˆ s ' s ˆ s ∈ ( ˆ s , 1 ] . 1 for bidders with a low budget and a value-signal of s∈ (ˆs,ˆs). Therefore, she compares bidding ˆwoversus b1(ˆso)which accounts for this change in marginal incentives. The subtle point, however, is that b1(·) is not independent of the rate at which bidders prefer to bid their entire endowment versus b1(s). As sincreases and more types discontinuously increase their bid above w , bidders with even lower budgets but higher value-signals have even less incentive to bid aggressively on the margin. 'Example Theorem 5. Suppose w is intermediate and value-signals are independent. Then (9) is a symmetric equilibrium. Proof. See Appendix B. Verifying that this is an equilibrium is routine, but the discontinuities necessitate a case-by-case argument. Preparatory work involves showing that the functionsfandb1existwiththestatedproperties. (Theyareintroducedaboveinaself-referential manner.) The independence assumption greatly simplifies the overall argument. The strategy ß(s,w)generates a bimodal distribution of equilibrium bids with few bids near w. After observing such a bid distribution, an observer not accounting for the strategic implications of budget constraints may infer erroneously that there are two distinct classes of bidders: those with high values and those with low values. Even with a uniform type distribution, such an endogenous segregation of bids can arise. * 3 v 2˜ +sj , W i i.i.d.~ U[1/5,2], and Si i.i.d. Example 5. Suppose v(si,sj)=si 5 1 ˆ b' ˆ b(s)-2s) 26 ~ U[0,1]. This is the setting of Examples 3 and 4 except now w =1/5. There is a critical point at s =10.151. The differential equation describing ˆ b(s)is ˆ (s)= 10(2b(s))( Figure 8 depicts the functions a,b 20 ˆ b(s)(s-1)-5s(3s+2)+7 and a(s)=s. The initial discontinuity is at ˆs˜0.181where some bidders shift to a bid of µ(ˆs)˜0.286—an increase in bid of approximately 58 percent. Therefore, the discontinuity can be quantitatively large. , ˆ b,ν,ψ, and f. Several representative paths from the associated plane-autonomous system are included to relate the example with previously derived results. Figure 9 presents a histogram of the resulting bid distribution. Only 24 percent of types actually place a bid equal to their budget; nevertheless, the strategic consequences of budget constraints are pronounced. Without budget constraints, the equilibrium bid distribution would be uniform. 1b Υ b 0.5 ΨΑ w F b 1 0 0 . 5 1 s F i g u r e 8 : E q u i l i b r i u m s k e t Dis cus sio n and App licat ion s We con clu de by pla cin g the pre ced ing ana lysi s in con text by con sid erin g sev eral c h i n E x a m p l e 5 . app licat ion s. Inp arti cul ar, we pro vid e aco mp aris on with the sec ond -pri ce auc tion ’ssy mm etri c equ ilibri um stra teg y. We als o hig hlig ht a diff ere nce in the co mp arat ive stat ic beh avi or of equ ilibri um stra tegi es bet we en the two for mat s. A disc ussi on of res erv e pric es con clu des . Wh en bud get con stra ints are pre sen t, eve n sm all res erv e pric es ma y dec rea se exp ect ed rev enu es. The Sec ond -Pri ce Auc tion Fan g& Par reir as (20 02) stu dy the sec ond -pri ce auc tion ’s sym met ric equ ilibri um. Ref erri ng to thei r deri vati on but ado ptin g our not atio n, the equ ilibri um stra teg y in the sec ond -pri ce auc tion is ß2 (s ,w )= v( s, s) if s <ˆ s m in {b 22 (s ), w } if s =ˆ s2 where (s)is a solution to the boundary-value b2 problem (s)) (s)-v(s,s)] 2(s)] , b2(1)=v(1,1). [b2 [η(s|s)-b 27 b' 2(s)= λ(s|s)γ(b2 4000 3000 2000 1000 0.2 0.4 0.6 0.8 Bid Figure 9: Histogram of equilibrium bids in Example 5. The figure was generated by evaluating ß(s,w)on a fine uniform grid. When >0, it is a point of discontinuity in the bidding strategy and v(ˆs2,ˆs2 ˆs2 b2 + 2s→ˆs 2,ˆs2)up to w at s=ˆs2, all types with value-signals s>ˆs2 If ß1 (s,w)and ß2 1 8 1(s,w)=ß2 1 1(s,w) = ß2 18 28 ) <w = lim(s). As in the first-price auction, the discontinuity’s presence depends on model parameters. Although discontinuities may exist, budget constraints in a second-price auction do not introduce the same strategic implications as in the first-price auction. There is no endogenous stratification of competition along the budget dimension. If bids increase discontinuously from v(ˆswill bid at least w. (s,w)are symmetric equilibrium strategies in the first- and second-price auctions, respectively, and for simplicity w =0, it is straightforward to confirm that bidders in a first-price auction shade their bids more: ß(s,w)(see the online appendix). When ß(s,w)<w, the preceding weak inequality is strict. On the other hand, noting instances where ß(s,w), we can exploit the invariance of bids to the auction format to test for the presence of budget constraints or to draw some inferences concerning their nature. Such tests are unusual insofar as they demand observing bidding across auction formats while maintaining the environment otherwise constant.However, they do in principle enable some identification of private budget constraints, as distinct from valuations, based on bidding behavior alone. Needless to say, For example, bidders could submit format-contingent bids. extending existing theories of empirical identification in auctions to accommodate privateG1' (w)=w(4-w)/4. G0 budget constraints will require novel forms of variation and strong assumptions. Otherwise, learning about two unobserved primitives (s,w)from a single bid is a difficult proposition. Changing Financial Constraints Budget constraints offer a new dimension along which the auction environment may change. Our preceding discussion focused on changing the budget distribution’s support parameterized by w . Here we consider changing the budget distribution G(·)holding [w ,¯w]fixed. Specifically, consider two distributions of budgets, Gand G1, with the same support, such that G0likelihood ratio dominates G10. Intuitively, financial constraints are more severe when budgets are distributed according to Grather than G01. For example, such differences may reflect changing macroeconomic conditions. In the second-price auction, changing the budget distribution from G0to G1increases unconstrained types’ bids (Fang & Parreiras, 2002). In the first-price auction, in contrast, increasing the severity of budget constraints analogously has ambiguous effects on bidding: an unconstrained bidder may bid more or less. As likelihood ratio dominance is already a restrictive stochastic ordering, this ambiguity can be interpreted as a negative conclusion from an empirical perspective. To confirm an ambiguous relationship, we revisit the model of Example 3 but with three alternative distributions of budgets on [0,2]: G0(w) = w/2, G1 li screen bidders along multiple dimensions, its introduction may reduce the auction’s expected revenue. Corollary 3 follows from Theorem 3 and we state it without proof. Corollary 3. Suppose w is low and let ß(s,w)=min{ ¯ b(s),w} be the equilibrium strategyabsent any reserve prices. Let r∈(w, ¯ b(0)]be a common reserve price. Then ßr(s,w)= l if w<r min{¯ b(s),w} if w=ris an equilibrium when the reserve price is r. The auction’s expected revenue has decreased. When r< ¯ b(0) the bids of all types with a budget w= rare unchanged. Expectedrevenue declines since the reserve price reduces the probability of sale by screening only low-budget bidders. Without budget constraints, in contrast, a small reserve price screens out low-valuation bidders and increases expected revenue through its uplift of the lowestvaluation participating bidder’s equilibrium bid. When private budget constraints exist, setting the revenue-maximizing reserve price is a delicate exercise. For instance, even in the introductory model of Example 1 the revenuemaximizing reserve price varies discontinuously with the model’s parameters (the budget values and their distribution). If high budgets are very likely, a reserve price of r *= 1/2 is optimal. This is also the revenue-maximizing reserve price when budget constraints are absent from the model. If bidders are often meaningfully budget-constrained, the optimal reserve price will be strictly lower. 4 Concluding Remarks Private budget constraints are a key feature of many economic environments and, as argued above, have a deep strategic effect which organizes interaction in the first-price auction. Although monotone equilibria (in a qualified sense) exist, equilibrium strategies demand a marked amendment of existing intuition. Private budgets have both direct and strategic effects on equilibrium bids and they can subtly interact with preferences changing bidder’s incentives. Increasing the wealth of a bidder may not necessarily imply a greater bid. Budget constraintsalsochangethenatureofcompetitionleadingtoitsstratificationalongthebudget dimension. The option to exploit other bidders’ budget constraints introduces a strategic effect that breaks the natural relationship between valuations and bids. 30 Our results suggest many possibilities for further research. For instance, the degree to which the inefficiencies introduced by budget limits can be ameliorated by resale is largely unexplored. Similarly, extending the analysis to multi-unit settings or addressing deeper technical matters within the standard environment, such as equilibrium uniqueness or empirical identification, are open questions. Untangling the more subtle relationship between bids, valuations, and budgets stands as a new, and in many cases continuing, challenge for theoretical and empirical researchers alike. 31 A Appendix: Proofs of Theorems 1 and 2 The proofs of Theorems 1 and 2 follow a standard argument, similar to those employed by Athey (2001) and Reny & Zamir (2004). Following some preliminary lemmas, we appeal to the equilibrium existence results in Reny (2011) to complete the argument. =bi =w' i. If (si,wi)=θi (s' i,w' i), then ui(si,s-i,wi,bi Lemma A1. Fix (s-i,w-i)and ri i(s' i,s-i,w' i,bi). ' i)= inf ∂ -s' i)+ inf i∂¯u∂ ∂ (wi -w' i) u ¯u i (s ¯ ' i) i ui ¯ui(si,s-i,wi,bi)-¯ui(s' i,s-i )= u Proof. We can establish the following inequalities: i i ∂si ∂wi i ∂wi -w' i) i ,w ,b ∂ i ∂wi -w' i)+ inf (w -w ξ i(w si ∂¯ (w = inf= ui sup i =ui(w )-ui (w' i) Rearranging the final inequality gives the desired result. i >s' i. (1) θi h(s-i|s' i) i i -i ¯H(s-i|s' i) = h(s-i|s) (2) H(s-i|s' i)=H(s-i|si). (3) ¯ H(s-i|s' i 'i ¯H(s-ii|si) -i θ H(s-i|si) i i(bi,ß-i|θ 1. Ui(bi,ß-i|θ' i 2. Ui(bi,ß-i|θi i )= ¯ L emma A2. Let si )-Ui Proof. These facts about affiliation follow from Milgrom & Weber (1982). Lemma A3. Suppose A-1 and A-3 hold. Let ßbe a profile of nondecreasing strategie bidders N \{i}. Let θ=' i. Suppose b>bare feasible bids for a bidder of type θwhich tie probability zero with the bids submitted by bidders j=i. Moreover, suppose U ' i)=Ui(l,ß i|θ). Then, (b' i,ß-i ' i|θ)=(>)0 =⇒ U (bi,ß |θi)-Ui (b' i,ß-i|θi)=(>)0. )-Ui (b' i,ß-i |θi)=(<)0 =⇒ U ( ,ß-i ' i|θ)-Ui (b' i,ß-i ' i|θ)=(<)0. b Proof. We prove part 1. Part 2 follows similarly. It is sufficient to show that Ui(bi,ß-i i (b' ' i|θ)=Ui(bi,ß-i|θi)-Ui (b' i,ß-i|θi). i,ß-i 32 |θ)U ' i Suppose that b' i'= land define, = {(s-i,w-i): maxj=i' ßj (sj,wj ' i) = b '. Then, maxj=i Aßj(sj,wj)=bi -i A' i,ß-i ' i|θ)-Ui (b' i,ß-i ' i|θ ,w' i ' i)]h(s ds-i |s' i)g(w-i)dw-ids -i [¯ui(s' i -i,w' i i)-¯u (s' i,s -i (1) -i,w-i): u (s' i,s-i ˜A ' -i )g(w -i)dw-i ,s-i ' i,bi ,b ,b -i|s (2) i(s i ,w-i ' ' -i,w -i . -i ' -i )∈A } and A = {(s}. Suppose that Pr[A]>0and let ˜ A=A\A -i -i 0=Ui(b ,s,b) =i ,b -i -i -i )∈A -i , if (s-i,w [¯ ' i,s-i,w' i,bi)-¯ui(s' i,s-i,w' i ui ,b' i)]h(s-i|s' i)g(w-i)dw-ids-i ,b -i -i -i =w-i,s' i] ] ')|W -i 'i =E [¯u=E =E = A' i -ids →[¯ui(s' i,s ,w' )-¯ui(s' i,s ,w' ' i)]1((s-i,w-i ' 19)is (s i i,b nondecreasing. + i,w)h(s' i We consider term (1) first. From Assumption A-1, ¯u' i,w i A' E[[ ¯u ) is non-decreasing in san sthen for all s= s, (s-i') ∈ A (s' i,S-i,w' i,bi)-¯ui(s' i,S-i ' ,b' i)]1((S-i,W i ' ,W -i )]1((S i ,w i(si,S-i,wi,bi)-¯ui(si,S-i E[[ ¯u -i [ ,wi,b i(si,s-i,wi,bi)-¯ui(si,s-i )]h(s | )g(w-i)dw ¯ s u Next, consider term (2). Given ˜ A⊂T-i , define the following subsets: ˜ )=0}, ˜-A={(s-i,w-i)∈ ˜ ={(s-i,w-i)∈ ˜ A:u A+ A:uii(s(sii,s-i,s-i,w,wii,bi,bi)<0}. 19 1(x∈X)is 33 the indicator function. ˜ A+Using A-3 we note the following inequalities: = u˜ A= i(s+˜ A= +˜ A+' h(s' i)g(w-i|s' i) ¯H(s) h(s-i-i|s|s' ii-i) ¯) ¯H(s)h(s-i-i|s|sii)dwH(s) ¯-iH(s)g(w-i-i-i|sds' i|s' i-i)g(w-i)g(w)dw-ids-i-i)dw-i)dw-ids-ids-i ¯ H(s-i|s' i) The second inequality follows from ui(si,s-i,wi,bi)= 0 and ¯H(s-i|si) iuiuu,sii(s-i' i(s(sii,w,s-i,s-i,s-i' i,bi,w' i,w,wii)h(s,bi,bi,bi-i|s) i(si,s-i,wi,bi)<0and H(s-i -i|si) -i|s' i)g(w-i)dw-ids-i ) h(s) h(s-i|s' i) H(s-i-i|s|s' ii) H(s) H(s-i|si-i) H(s-i|s' i|s' i)g(w-i)g(w-i)dw-i)dw-ids-ids-i ui ˜ A- i)h(s-i|si)g(w-i)dw-ids-i = ' i,s-i,w' i,bi)h(s ui(s' i,s-i,w' i,bi ui(si,s-i,wi,bi ui(si,s-i,wi,b ˜ A = = - ˜ A - ˜ A - ' i,s-i,w' i i)h(s -i |θ|s' i)g(w-i -i i ui(si,s i = ,wi,b )h(s .|si)g(w-i)dw-ids i 'i ui(s i ' i,ß i i -i i i ˜A |s' i= 1. Similarly, since u) H(s(s=1, ,b-iCombining the previous results gives )dw Therefore, 0=Ui(bi,ß-i' ' i)-U ( b d s -i|θ )=U ˜ A -i ( ,ß | )-U b θ i (b' i,ß -i -i | ). If instead b' θ i ⊂ ¯[ B<8. 34r i,wi i , ¯ i B ] f o r s o m e = l or Pr[A]=0, then the above analysis applies but only considering term (2). Remark A1. If A-1 and A-2 are satisfied, Lemma A3 continues to obtain. The proof is analogous, but simpler. Proof of Theorem 1 We introduce two formal modifications to our model. First suppose that there is no upper bound on the bid that a bidder of type θcan place, i.e. there is no budget constraint, but bidders continue to have a two-dimensional type (s). Noting A-1, we can assume that for each i, B Second, toincorporatethebudgetlimit, modifyeachbidder’sutilityfunctionasfollows. If abidder exceeds her budget, bi>wi, then bidder ireceives apenalty of-F utils irrespective of the auction’s outcome in addition to their auction payoff. Set F such that all bids bi>wi are strictly dominated by the bid lirrespective of the bids of other bidders. Since and ui ¯ui are bounded, such a constant exists. Examining players’ behavior in such modified games where agents are penalized for deviating “inappropriately” (from the analyst’s perspective) is often seen in the auction literature (for example see Jackson & Swinkels (2005)). Also Che & Gale (1998) note this possibility in the context of auctions with budget constraints. let ρi(θi;ß-i Noting this modified auction, for a profile of nondecreasing strategies ß-i i: UiUi(bi(bi,ß-i,ß-i|θ-i|θ-i)=wi . if bi)-F if bi >wi ρi(θi;ß-i)=arg maxbi∈Bi∪{l} i θi i,ß )=Ø. Since a bid above wi -i. For all θi, ρi(θi;ß-i i -i, supρi(θi,ß-i )=w Proof. Since the set Bi θ ) be the set of bidder i’s best-response bids when her type is θ Lemma A4. Consider the modified auction game described above. Fix a nondecreasing strategy profile ß)is nonempty, bounded, and increasing in the strong set order. -i ∪{l}is nonempty and discrete, ρis strictly dominated by ( θ lirrespective of ß ' i(θi;ß-i) is nondecreasing, let θ = i 'i and let b ∈ ρ ( ;ß i θ 'i (b' i,ß-i|θi)=Ui(bi,ß-i ' i|θ)-U (b ,ß |θ ∈ρ (θ ;ß ), -i |θ i 35 (b' ,ß-i |θ i |θi i(θ = w' i = b i i ( θ i ' i,ß ' i,ß i)= |θ i i -i i) i i ) . = ⇒ b U -i i i ' i)-U ( b -i|θ i(b -i i )-U i ' i) =⇒ b ∈ρ ' i;ß (b' i,ß-i )-U i(bi,ß-i i i i . To show that ρ -i) and b ' ' i;ß-i). If bi for all such bids, we are done. Otherwise, suppose b' >bi i i i an 0=Ui ' i= >b , both bids are feasible for bidders of type θ and θ' i d b ∈ ρ. Since w. By Lemma A3 or Remark A1, 0=U )is increasing in the strong set order. As the interim best-reply correspondence is nondecreasing, player i has at least one nondecreasing best reply when others are following nondecreasing strategies (Topkis, 1998, Theorem 2.8.3). Hence, ρ i Lemma A5. Consider the modified auction game described above. There exists an equilibrium in nondecreasing strategies. Proof. The conclusion follows from Reny (2011) and specifically Proposition 4.4, Theorems 4.1 and 4.3, and Remark 1. His conditions G.1–G.6 are satisfied by our game and are easy to verify. Remark A2. An equilibrium in the modified auction game is an equilibrium in an auction with budget constraints where each bidder’s set of available (competitive) bids is finite. Proof of Theorem 2 To prove Theorem 2(1) we start with the setup of Theorem 1 and we consider a sequence auction games, indexed by k, where the set of a bidder’s feasible bids becomes dense in [rik i,8). In auction k, the set of feasible bids for bidder iis B=[ri, ¯ B]nPk ik iwhere Pi= {r-νi+km 2:m=1,...,Mk ik i}and M=⌈2k( ¯ B-(ri-νi))⌉. {νk ik j}are small fixed constants such that PnP=Øfor all k. We let Bi=[ri, ¯ B]. Let ßkibe the corresponding sequence of equilibria. LemmaA6(Theorem2(1)). Assumeprivatevalues, i.e. ui(si,s-i,wi,bi)=ui(si ,s' -i,wi,bi i =[ri k + i 1 k -i i:T→Bi i i ˜ ßk i(θi)= l wk imax{b∈B∪{l}:b<min(ßi(θi)+As k→8, the (ex-ante) loss to bidder ifrom using ˜ ßk i1 2k,wi)} wversus ßiii , ¯ B].). Then there exists an equilibrium in nondecreasing strategies when ∀i, B Proof. This is a consequence of better-reply security (Reny, 1999); see also the proof of Corollary 5.2 in Reny (2011) and Jackson & Swinkels (2005). Briefly, for kvery large ßis an o-equilibrium of the first-price auction where all bidders can place bids from B. Given any monotone strategy profile ß, bidder ican approximate the payoff she would receive from the (feasible) strategy ß∪{l}. Specifically, suppose bidder ibids + 1k . <r =r becomes arbitrarily small and this conclusion holds versus all strategies that the other bidders may follow. To prove Theorem 2(2), we adopt the argument of Araujo & de Castro (2009, Theorem 2). Since the construction is very similar, forbrevity we only sketch the relevant steps. Again take a sequence of auctions indexed by k; however, each auction has the following features: 36 k i1. i In auction k, B =[ri, ¯ B]. The bid lis always available and l<mini ri -2. =wi 2. The penalty for exceeding a budget constraint is modified as follows. If , it is zero. If w<bi=wi+1/k, it bi +ok 1,...,bN +ok N k 1,...o k. Considering the sequence of equilibria, ßk a.e. andfora.e. (si,wi), ˆ ß (si,wi)=wi i such thatß ˆ ßi i i i˜ = ˆ ßj(tj i(θ ϕ (Bj : ( × k i j i ( ß k ( t ) ) i f b ˜ϕi(b;t)= limki . The allocation rule in auction kfor bidder iis ˜ϕk i(b)=E[ϕi(b1 k i → i )]where (ok N)is 3 a vector of i.i.d. uni [-1/k,1/k]. The modified allocation and penalty r a bidder has a nondecreasing best respo strategies. For each k, an auction with th nondecreasing strategies, ß , and passing to a subsequence as need Lemma A.10), there exists a nondecreas i ∪{l}))×T→[0,1]as ˆ ß j)∀j ϕ(b) otherwise . ), and an announcement of her type, ˜s ˆ , a bidder’s strategy is a bid, ˜ . There exists aspecial tie-breaking rule, which depends on pla ßi ß announcement of their private information, for which the biddin equilibrium bidding strategy. Define this allocation rule ˜ϕ Under ˜ϕ ). ˜ ϕ i(θi allocate s the item to the highest bidder (if unambi guous) and announ cement s of private types are used to resolve ties among high bidders. Lemma A7 (Theorem 2(2)). Suppose Bstrategy profile where for all bidders i, ˜ ß i i)and ˜s ( )=θ is an θ equilibrium. i i i ( )= ˆ ( θ ß θ i =[ri, ¯ B]and the allocation rule ˜ϕis in effect. The Proof. The argument mirrors the proof of Lemma 4 in Araujo & de Castro (2009). B Appendix: Proofs from Section 3 Lemma B8. d(b,x|s) is increasing in band xand decreasing in s. η(x|s)is increasing in both arguments. 37 Proof. Since g'(b)=0, dis increasing in b. dis increasing in xas H(x|s)is increasing in x. Finally from affiliation, if s'>sthen H(y|s')=H(y|s)and thus,H(y|s') 1-H(y|s')=H(y|s) 1-H(y|s)'. Thus, d(b,x|s)=d(b,x|s). The conclusions concerning ηfollow easily from its definition. exists s* >0 such that ˜ ψ(s* Lemma B9. ˜ ψ(s):[0,1)→[w,8)is continuous. Proof. Suppose ˜ ψ(s) is not continuous at some s1. If ˜ ψ(s1) >w , then ˜ ψ(s) and η(s1|s1) = d(ψ(s1),s1|s11) =ψ(s1) by definition. Hence ψ(s) = Ø for all ssufficiently close to s, ˜ψ(s)=ψ(s). Since ψ(s)is continuous when it is defined ˜ ψ(s)must be continuous at s11— contradiction. Therefore,˜ ψ(s1) = w . Since ˜ ψ(s) is not continuous at s, we can find a sequence tk→ s1such that for each k ˜ ψ(tk1) >w+ofor o>0 sufficiently small. Since ˜ψ(tk)=¯v, we can find a further subsequence, tk' such that { ˜ ψ(t' )}converges to some value ˜ψ1= ¯w+o. By continuity, however, we conclude that lim kk' η(tk' |tk' )-d( ˜ ψ(tk' ),tk' |t' ) = η(s1|s1)=d( ˜ ψ1,s1|s1). But then either ˜ ψ1∈ψ(s1)or ˜ ψ1k=w . Both of these conclusions are contradictions. Hence˜ ψ(s)is continuous. Lemma B10. If w=0, then there exists a critical point. Proof. Since 0=d(0,0|0)=η(0|0)=¯v= ¯w=d(¯w,0|0), ψ(0)=Ø. If ψ(0)=0, then (0,0) is a critical point. If ψ(0)>0, then since ) is n o n- n e g at iv e. T h e n, *) ∂b (s*,b ) =-γ(b ∂b (s*,b <0 ) *)-d(b,s*|s* =0 ∂ ˙ 2b ∂˙s ∂s- are real-valued. The eigenvalues will be real if ∂˙s ∂b (s*,b*) +4 =γ' *(b)(η(s*|s* )) ∂˙s ∂ ˙ bwhen ∂b ∂s (s* ∂b +γ(b * 3 8 ∂ d evaluated at *) ∂d ,b* * w h er e th e c o n cl u si o n fo ll o w s fr o m d st ri ctl y in cr e a si n g in b a n d γ( b* )= 0. Al s o, *(b -v(s*,s* = ∂λ∂s ∂s∂ ˙ b (s*,b*) (s*,b*)Therefor e, 4 R e m ar k B 3. Si n c e th e ei g e n v al u e s of (B 1) ar e re al -v al u e d, th e o ut st a n di n g c o +λ(s*|s*) ∂v ∂s *,b ) =-λ(s*|s* ∂s (s*,b <0 ) ∂ s ∂ ∂ ˙ s ˙ ∂ b b =0)) >0as neede d. * ( s ) ∂v * n c er n is th at o n e of th e ei g e n v al u e s is z er o. T h at is, w h e n e v al u at e d at (s *, b* ), ∂˙s ∂s S u c + ∂ ˙ b+ ∂b ∂˙s ∂s ∂ ˙ 2b ∂b +4 ∂˙s ∂ ˙ b=0, ∂b ∂s or ∂˙s + ∂ ˙ b ∂s ∂b ∂˙s ∂s ∂ ˙ 2b ∂b +4 ∂˙s ∂b ∂˙ b∂s =0. h c a s e s ar e n o ng e n er ic a n d a s m al l p er tu rb at io n to th e m o d el re s ol v e s th e sit u at io n. A ty pi c al in st a n c e of a n o ng e n er ic c a s e w o ul d o cc ur if, fo r e x a m pl e, ψ (s )a n d ν( s) w er e e x a ctl y ta n g e nt at a cr iti c al p oi nt (s *, b* ). Pr o of of T h e or e m 3. S u p p o s e th e ot h er bi d d er is fo ll o wi n g th e pr e sc ri b e d st ra te g y ß(y,z) = min{ ¯ b(y),z}. The expected utility of a bidder of type (s,w), w =bids ¯ ¯ b(0), b(x)=wis who 0¯ b(x) Di ff er e nt ia ti n g wi th re s p e ct to x a n d c ol le cti n ¯ 01 Ui( ¯ b(x) |s,w )= v(s,y)- 0 ¯ b(x) h(y|s)g(z)dydz+ ¯ w x v(s,y)- ¯ b(x) b ( x ) h(y|s)g(z)dydz. g te r m s, ∂Ui( ¯ b(x)|s,w) ∂x =g( ¯ b(x))(1-H(x|s)) v(s,x)λ(x|s)γ( ¯ b(x)) To show that ¯ b(s) is optimal among all bids¯ b(x), x∈ ¯ ∂Ui( ¯ b-1 [0,∂Ui b(x)|s,w)∂x ∂x ∂Ui( ¯ b(x)|s,w)∂x = By a similar argument, if x <s, = 0 a n d x < s = ⇒ (¯ b(x )|s, w) ( w )], it is s u ffi ci e nt to s h o w th at x > s = ⇒ = i ∂Ui( ¯ b(x)|s,w)∂x 39 ¯ b(·) . ¯ b(x) ¯ b'(x) η(x|s)-d(¯ b(x),x|s) . + 0. S u p p o s e x > s. B y a ffi li at io n, η( x| x) = η( x| s) a n d d( ¯ b( x) ,x |x )= d( ¯ b( x) ,x |s ). H o w e v er , al s o d u e to a ffi li at io n λ( x| x) = λ( x| s) a n d b y A ss u m pt io n A3( 1 a) , v( s, x) λ( x| s) = v( x, x) λ( x| x) . T h u s, ∂U ( ¯ b(x)|x,w)∂x =0. (B2) = 0. T h u s, a bi d d er of ty p e (s ,w ) h a s n o in c e nt iv e to d e vi at e to a n y fe a si bl e bi d in th e ra n g e of Since bids greater than ¯ b(1)are inferior to ¯ b(1), it remains to verify that the bid b< ¯ b(0 ) is n ot a pr o fit a bl e d e vi at io n. T h e re s ul ti n g e x p e ct e d ut ilit y is 0 1 (v(s,y)-b)h(y|s)g(w)dydw=G(b)(η(0|s)-b). Ui(b|s,w)= 0b This expression is concave in band achieves a maximum at ˇ bwhich solves η(0|s) ˇ =G( ˇ b)/g(ˇ b). Recall that ψ(0) satisfies η(0|0) = ψ(0)+G(ψ(0)) g(ψ(0)) b+ henceUi(b|s,w)=Ui( ¯ consi for ssufficiently close to zero, ¯ b(s) = ψ(s), b(0)|s,w)=Ui(ß(s,w)|s,w). Finally der a ¯ b(0) = ψ(0). Therefore, ∀b = ¯ b(0) , . Consequently, by ˜ b(x), x= ˜s. The expected p this affiliation, η(0|s) = η(0|0) =⇒ˇ b = ψ(0) because b+G(b)/g(b) is increasing. However, bidder with a budget of w< other bidder, she solves maxb=w ¯ b(0). Given the strategy followed by the G(b)(η(0|s)-b). By the argument above, ß(s,w)= wis their optimal bid. Proof of Theorem 4. As in Theorem 3, we need only verify that no type of bidder wishes to deviate to an alternative bid in the range of ß(s,w). There are two cases: 1. Consider a bidder with a value-signal s<˜sand budget w. Then ß(s,w)=a(s). Since a(s)is the equilibrium bid in a first-price auction without budget constraints, this type of bidder will not have any profitable deviations to any bids a(x), x∈[0,˜s]. Therefore, we need only rule out deviation to a bid 0 1 0 x bid is (v(s,y)˜ b(x))h(y|s)dy. Ui( ˜ b(x)|s,w)=G(˜ b(x)) (v(s,y)˜ b(x))h(y|s)dy+(1-G(˜ b(x)) ∂Ui( ˜ b(x)|s,w)∂x i( ˜ b(x)|s,w)=Ui(w |s,w)=Ui(a(˜s)|s,w)=Ui( ˜ bf 40 y an argument parallel to (B2) we know that if x>˜s= s, = 0. This implies B that U(s)|s,w). 2. Consider a bidder with a value-signal s>˜sand budge Theorem 3 establishes that this bidder has no profitable Consider a deviation to a bid a(x), x= ˜s. The expected Ui(a(x)|s,w)= ∂Uix 0(v(s,y)-a(x))h(y|s)dy. Differentiating this expression gives(a(x)|s,w) ∂x=H(x|s) (v(s,x)-a(x))h(x|s) H(x|s)=H(x|s) (v(x,x)-a(x))h(x|x) H(x|x)-a'-a(x) '(x) =0. Theinequalityfollowsfromaffiliation. Hence, Ui(a(x)|s,w)=Ui(w |s,w)= Ui(ß(s,w)|s,w).(a(˜s)|s,w)=Ui +s→˜ ˜ b'(s). We can Proof of Corollary 2. It is sufficient to show that a'(˜s)<lim s write H(s|s) H(s|s) (1-G( + (v(s,s)˜ h(s|s) 1-H(s|s) ˜ b'(s)= -γ( ˜ b(s) H(s|s) . 1-G( ˜ b(s)) b(s))(η(s|s)) G( ˜ b(s)) ˜ ˜ b(s))(1-H(s|s) b ( s ) ) + The first bracketed term is positive for all s= ˜s. Moreover, since η(s|s)> ˜ b(s), g(w )=0,and lim+s→˜s˜ b(s)=w , the limit as s→˜s+of the first bracketed term is strictly greater than 1. But, lims→˜s+ (v(s,s)˜ b(s))h(s|s) H(s|s)=a'(˜s). Thus, lim+s→˜s˜ b(s)>a'(˜s). Lemma B11. When wis intermediate, Z=Øand at ˆs=supZ, (8) holds with equality. Proof. We first show that Z=Ø. There are two cases. First, suppose µ(s)=w for some s= a -1(w ). Forsuchans, (8)reducesto s 0(v(s,y)-w )h(y|s)dy= s 0(v(s,y)-a(s))h(y|s)dy, which is true since w >a(s). Suppose instead µ(s)>wfor all s<a-1(w). In this case, there is a critical point, s0=0, and ¯ b(s0)=w. Hence, (8) reduces to G( ¯ b(s0)) 01(v(0,y)¯ b(0))h(y|0)dy=G(¯ b(s0))(η(0|0)¯ b(0))=0. The inequality obtains since η(0|0)¯ b(0)=η(0|0)-w =0.To confirm that there is a value at which (8) holds with equality, it is sufficient to show that at ˜s=a-1(w ), the inequality in (8) is reversed. Consider the following maximization problem: maxb=w˜s 0(v(˜s,y)-b)h(y|˜s)dy+G(b) 1 ˜s(v(˜s,y)-b)h(y|˜s)dy. The first-order 41 condition leads to 0=η(˜s|˜s)-d(b,˜s|˜s)or b=ψ(˜s). Since ˜s∈S0˜s, ψ(˜s)=µ(˜s)=w. Thus, (v(˜s,y)-µ(˜s))h(y|˜s)dy+G(b) ˜s1ψ(v(˜s,y)-µ(˜s))h(y|˜s)dy= 0˜s(v(˜s,y)-a(˜s))h(y|˜s)dy. That (8) holds with equality at ˆsfollows from continuity. Proof of Theorem 5 The proof has two parts. The first set of lemmas concerns the functions fand b. The final part verifies that the proposed strategy is an equilibrium. We specialize our notation to the case of independent value-signals as stated in the theorem; thus, H(y|s)=H(y), etc.We start by deriving some necessary relationships between b 11and f, which we use to construct a map Λ. A fixed point of Λwill be used to define b1and f. b1and fneed to be consistent with two facts: f(s)describes an indifference condition. For ˆs=s=ˆs, types (s,f(s))are indifferent between bids of b1(s)and f(s). The expected utility from bidding f(x), x=ˆs, is '1. Ui(f(x)|s,f(s))= 0ˆs(v(s,y)-f(x))h(y)dy (B3)+ ˆsx(v(s,y)-f(x))G(f(y))h(y)dy (B4)+G(f(x)) x1(v(s,y)-f(x))h(y)dy. (B5) (B3) is the contribution to utility from defeating all bidders of type y<ˆs. (B4) is the contribution to utility from defeating bidders of type y∈[ˆs,x)but who have w<f(y). Such bidders are bidding according to b1(·). (B5) is the contribution to utility from defeating all bidders of type y=xwho have a budget less than f(x). In an analogous manner, the expected utility of bidding b1(x)is Ui(b1(x)|s,w)= 0ˆs(v(s,y)-b1(x))h(y)tdy+ ˆsx(v(s,y)-b1(x))G(f(y))h(y)dy.(B6) Both expressions for expected payoffs assume b1is increasing and continuous and that fis decreasing and continuous. In equilibrium, we will require that U(s)|s,f(s))= Uii(b1(f(s)|s,f(s)). Rearranging the terms in this equality gives (12) which is reproduced 42 here: f(s)-b1(s) η(s|s)-f(s) (·) must be optimal for all types who do not bid above w . The usua 2. b1 condition yields the differential equation (11) which is reproduced 1,c2)=sups∈[ˆs,1] 1 '=0 =⇒ b(s)= [v(s 1(x)|s,w) ∂x x=s ˆsG(f(y))h(y)dy (B8 1 2 |c1 ∂ U i ( b We wish to show that there exist band fthat meet (B7) and (B8), along with the collection of boundary and monotonicity conditions from (9). Let C[ˆs,1] be the set of continuous functions on the interval [ˆs,1]equipped with the usual metric, d(c(s)c(s)|, and define the following subsets: ∈ [ˆs,1], then |b(s)-b(s' supH(ˆs)s∈[ˆs,1] v(s,s)h(s) ')|= sups∈[ˆs,1] F(s|b,ζ)=min arg minˆ f∈[w ' ζ(s) ˆ f-b(s)η(s| s)ˆ f = 0, ,η(s|s)] • B ⊂ C[ˆs,1]. If b∈ B then b(ˆs) = a(ˆs) and if s,s' )| = |s-s'|.• Z ⊂C[ˆs,1]. If ζ∈Z then ζ(ˆs)=H(ˆs), H |ζ(s)-ζ(sh(s) |s-s'|.B and Z are both equicontinuo With (b,ζ)∈B ×Z given, consider the following fu G( ˆ f)(1-H(s)) . (B9) max 0, ˆf-b(s) η(s|s)-ˆ fG( ˆ f)(1-H(s))ζ(s) Given F, define the map Λas follows: Λ(b,ζ)=( ˇ b, ˇ b(s)=a(ˆs)+ ˆss max{v(x,x)-b(x),0}G(F(x|b,ζ))h(x) ζ(x) dx (B10) To develop an intuition for F, consider Figure B1, which sketches the value outputted by Ffor a fixed s, b, and ζ. Ideally, Fwould like to output a solution to max . When there are two solutions, it selects the smaller one (Figures B1b and B1c). (There can be at most two solutions.) When the desired solution does not exist, Fminimizes the distance between the two sides of the equation (Figure B1a). (This minimum exists and will be unique.) Clearly, Fis continuous in s, b, and ζ. ˇ ζ)where ˇ ζ(s)=H(ˆs)+ ˆss G(F(x|b,ζ))h(x)dx (B11) 43 L LR F LR fb R Min Distance w η Ff b w ηFf b (a) No Solution (b) Two Solutions Figure B1: Definition of F. Notation: L(f)=max 0, wη (c) Solution at w f-b(s) η(s|s)-f , R(f)= G(f)(1-H(s)) ζ(s) . Λis continuous, Λ(B×Z)⊂B×Z, and by Schauder’s fixed point theorem, Λhas a fixed point. Lemma B12. Let (b,ζ) be a fixed point of Λ and let F(s) = F(s|b,ζ). Then, (1) b(s) satisfies the integral equation b(s)=a(ˆs)+s ˆs[v(x,x)-b(x)]G(F(x|b,ζ))h(x) ζ(x)dx, (2) b(s) is strictly increasing if and only if F(s)>w , and (3) b(s)=w. (1) Suppose there is an interval [ˇs,ˇs] ⊂ [ˆs,1] such that for all x∈ [ˇs,ˇs'], v(x,x)b(x) =0. Then b(x) =b(ˇs) for all x∈ [ˇs,ˇs']. Since b(s) is continuous and b(ˆs)=a(ˆs)< v(ˆs,ˆs), without loss of generality, we may assume that at ˇs, v(ˇs,ˇs) = b(ˇs). However, for all x> ˇs, v(x,x) >v(ˇs,ˇs). Hence, x∈ (ˇs,ˇs) =⇒ v(x,x) >b(ˇs) = b(x), which is a contradiction. (2) By the preceding part, v(s,s)-b(s)>0a.e. Thus, the conclusion follows. (3) Suppose b(s)>w for some s. Since b(s)is nondecreasing there exists ˇs>ˆssuch that for all s>ˇs, b(s) >w . But from the definition of F, if b(s) = w, F(s) = w; thus, the integrand above is zero for all x> ˇs. Coupled with b(·)’s continuity, however, this is a contradiction. Lemma B13. Let (b,ζ)be a fixed point of Λand let F(s)=F(s|b,ζ). Then for all s=ˆs, F(s)-b(s) = G(F(s))(1-H(s))ζ(s) . (B12) η(s|s)-F(s) 'Proof. Proof. If b(s) = w , then (B12) holds since b(s) = F(s) = w ensures equality. Suppose b(s) < w. We first establish that (B12) holds at s = ˆs. When s = ˆs, (B12) becomes 44 F(ˆs)-a(ˆs) η(ˆs|ˆs)-F(ˆs) d df = d d s = g ( f ) ( η f ) H G(F(ˆs))(1-H(ˆs)) H(ˆs) g(f) - = g ( f ) 1 H . From Lemma B11, if F(ˆs)=µ(ˆs), then the preceding equation is satisfied. Thus, F(ˆs) = µ(ˆs). Suppose ∃f < µ(ˆs) such that. Rearranging terms and suppressing the dependence on ˆs, -a=(η-f)G(f)(1-H)/H-f (B13) Differentiating the right-hand side of (B13) with respect to fgives (η-f)G(f)(1-H)/H-f 1-H -G(f) 1-HH -1 H =g(f) 1-H η-fG(f) H g(f)(1-H) H (η(ˆs|ˆs)-d(f,ˆs|ˆs)). When f<µ(ˆs), η(ˆs|ˆs)-d(f,ˆs|ˆs)>0. So the right-hand side of (B13) is strictly increasing as f↑µ(ˆs). Indeed, the above derivative is zero only at f=ψ(ˆs)=µ(ˆs). Therefore, there is no f<µ(ˆs)that satisfies (B13). Thus, F(ˆs)=µ(ˆs). As F(s), b(s), and ζ(s)are continuous, if (B12) fails to be satisfied, it must fail on some interval, say (˜s,˜s''). For all such s∈(˜s,˜s),F(s)-b(s) η(s|s)-F(s)=G(F(s))(1-H(s)) ζ(s). Equivalently, (F(s)-b(s))ζ(s)=(η(s|s)-F(s))G(F(s))(1-H(s)). (B14) 'As the inequality in (B14) becomes strict for s∈(˜s,˜s), at s=˜s, the derivative of the lefthand side of (B14) with respect to s(holding F(˜s)>w fixed) must exceed the derivative of the right-hand side of (B14) with respect to s. Intuitively, the right-hand side escapes downward from the left-hand side locally at ˜s. Therefore, the following must be true, [(F(˜s)-b(s))ζ(s)]|s=˜s = dds [(η(s|s)-F(˜s))G(F(˜s))(1-H(s))]| .s=˜ s (˜s) = Some algebra and substituting b' ∂ ∂s[v(˜s,˜s)-b(˜s)]G(F(˜s))h(˜s)and ζ(˜s)v(s,y) s=˜s ζ '' 45 (˜s) = G(F(˜s))h(˜s) gives 0=G(F(˜s))1 ˜s∂ ∂sv(s,y) s=˜sh(y)dy. However,>0. Since ˜s=1and F(˜s)>w , this is a contradiction. Lemma B14. Let (b,ζ)be a fixed point of Λand let F(s)=F(s|b,ζ). There exists ˆs', b(s)=F(s)=w .∈(ˆs,1] such that for all s=ˆs Proof. As (B12) holds, at s=1, we have F(1)=b(1). Moreover, as b(s)= w =F(s) we must have b(1)=F(1)=w. The lemma’s conclusion follows since whenever b(s)=w, then F(s)=w as well, and b(s)is nondecreasing. Lemma B15. Let (b,ζ)be a fixed point of Λand let F(s)=F(s|b,ζ). Then, (1-H(s))[g(F(s))(η(s|s)-F(s))-G(F(s))]=ζ(s). Proof. Suppressing sinthenotation, F(s)satisfies F-b=(η-F)G(F)(1-H)/ζ. ForF>w , the right-hand side of this expression is concave in F, therefore at the minimal solution to this equation, the slope of the right-hand side must exceed the slope of the left-hand side. Differentiation gives [g(F)(η-F)-G(F)](1-H)/ζ=1, as desired. Lemma B16. Let (b,ζ)be a fixed point of Λ. F(s)=F(s|b,ζ)is non-increasing on [ˆs,1]. Proof. Noting Lemma B13, Ui(F(s)|s,F(s))=Ui(b(s)|s,F(s))for all s>ˆs. Moreover, since band ζare differentiable a.e., F(s)is differentiable a.e. Suppressing a bidder’s budget-type in the following notation, differentiating the preceding expression with respect to sgives ∂Ui(F(s)|s) dF(s) + ∂Ui(F(s)|s) = ∂Ui(b(s)|s) db(s + ∂Ui(b(s)|s) (B15 ∂F(s) ds ∂s ∂b(s) ) ds ∂s ) = d = (1-H(s))(g(F(s))(η(s|s)-F(s))-G(F(s)))-ζ(s) F'= 0ˆsd dF(s) dsv(s,y)h(y) dsv(s,y)h(y)dy+ +G(F(s)) s1d dsˆssd dsv(s,y)h(y)G(F(y))dy 0 dy+ ˆss ˆ ds d dsv(s,y)h(y)G(F(y))dy (B19) s i d b ( s ) d s i i (s) (B16) ∂U(F(s)|s) ∂s v(s,y)h(y)dy (B17) ∂U(b(s)|s) ∂b(s)=0 (B18) ∂U(b(s)|s) ∂s w here ∂Ui(F(s)|s) ∂F(s) (B18)iszero by theenvelope theorem. The firstterms of (B16), collectively +G(F(s)) 1 s d ds * ∂Ui(F(s)|s) dF(s) ds ∂F(s) 1 * ∂Ui(F(s)|s) ∂F(s) ' *,ζ *,ζ* 46 , isnonnegative by Lemma B15. Simplifying (B15),v(s,y)h(y)dy=0. When F(s)>w , the second term on the left-hand side is positive. So, F(s)=0. Choose a fixed point of Λ, (b), and define b(s) = b(s) and f(s) = F(s|b). Therefore, the strategy ß(s,w)exists with the stated properties. Lemma B17. Suppose w is intermediate. Then ß(s,w)is a symmetric equilibrium. Proof. We divide the argument into cases. As usual, a bidder will not have a profitable deviation to a bid outside the range of ß(s,w). 1. Consider a bidder with a value-signal of s<ˆs. This bidder’s expected payoff when following the strategy ß(s,w)is Ui(a(s)|s,w)= s 0(v(s,y)-a(s))h(y)dy.(a) By the usual argument confirming an equilibrium in a first-price auction, this type of bidder will not have a profitable deviation to any bid a(x), x∈[0,ˆs]. (b) Suppose this bidder bids '(x), x∈[ˆs,ˆs]. The expected payoff from this bid is b1 Ui(b1(x)|s,w)= 0ˆs(v(s,y)-b 1(x))h(y)dy+ ˆsx(v(s,y)-b1(x))G(f(y))h(y)dy. ˆsx G(f(y))h(y)dy . ' 1(x)]G(f(x))h(x)-b(x) H(ˆs)+ 1Hence, d dxUi(b1(x)|s,w)=[v(s,x)-b d dxThus, s→Ui(b1 d dxUi(b1(x)|s,w) = 0 and x<s =⇒d dxUis=x=0. Therefore, s<x =⇒i(b1(x)|s,w) = 0. Ui( Hence, when s = ˆs = x, U(x)|s,w) = Ui(b1(b1(ˆs)|s,w) = U(a(ˆs)|s,w) = Ui'(a(s)|s,w). Therefore, this is not a profitable deviation. (c) Suppose this bidder bids f(x), x∈[ˆs,ˆsi]. The expected payoff from this bid is (v(s,y)-f(x))h(y)dy. x ˆs(v(s,y)-f(x))h(y)dy+ (v(s,y)-f(x))G(f(y))h(y)dy U i ( +G(f(x)) x1f ( x ) | s , w ) = ˆs 0 Consider the difference Ui(f(x)|s,w)-U= 0= 0ˆs[bˆs1[b1i(b1(x)|s,w)(x)-f(x)]h(y)dy+ ˆs(x)-f(x)]h(y)dy+ ˆsxx[b1[b1(x)-f(x)]h(y)G(f(y))dy+ x(x)-f(x)]h(y)G(f(y))dy+ x11[v(s,y)-f(x)]h(y)dy[v(x,y)-f(x)]h(y)dy =Ui(f(x)|x,w)-Ui(b1(x)|x,w)=0 By case 1b, Ui(f(x)|s,w)=Ui(b1(x)|s,w)=Ui(a(s)|s,w). 47 (d) Suppose this bidder bids ˆ b(x), x∈[ˆs,1]. Analogously to the proof of Theorem 4and noting case 1c, Ui( ˆ b(x)|s,w)=Ui( ˆ b(ˆs)|s,w)=Ui(f(ˆs)|s,w)=Ui(a(s)|s,w). 2. Consider a bidder with a value-signal s=ˆsand a budget w=µ(ˆs). Thus, ß(s,w)= min{ˆ b(s),w}. Hence, ∃sw∈[ˆs,s]such that ß(s,w)= ˆ b(sw). (a) Following Theorem 3, we can rule out all deviations to bids above µ(ˆs). (b) Suppose this bidder bids a(x), x∈[0,ˆs]. Analogously to the proof of Theorem 4,x=s =⇒ Ui(a(x)|s,w)=Ui(a(ˆs)|s,w). Noting that ˆ b(ˆs)=f(ˆs)=µ(ˆs), ˆs [ ˆ b(ˆs)-a(ˆs)]h(y)dy-G(f(ˆs)) Ui(a(ˆs)|s,w)-Ui( ˆ b(ˆs)|s,w)= (a(ˆs)|ˆs,w)-Uii=U= 0ˆs[ ˆ Thus, Ui(a(x)|s,w)=Ui b(ˆs)-a(ˆs)]h(y)dy-G(f(ˆs))( ˆ b(ˆs)|ˆs,w)=0. (a(ˆs)|s,w)=Ui( ˆ b(ˆs)|s,w)=Ui( ˆ b(sw )|s,w). (c) Suppose this bidder bids f(x), x=s. A calculation gives d Ui(f(x)|s,w)=f'(x) g(f(x)) x1[v(s,y)-f(x)]h(y)dy-G(f(x))(1-H(x)) dx dx 1 dx 0 ˆ b b v(x,y)h(y)dy=0 =-G(f(x)) =f'(x) H(ˆs)+ H(ˆs)+ ˆsx G(f(y))h(y)dy g(f(x)) x1 [v(x,y)-f(x)]h(y)dy-G(f(x))(1-H(x)) x G(f(y))h(y)dy ˆs (f(x)|s,w)= U Therefore, Ui (f(ˆs)|s,w)= Ui( ˆ b(ˆs)|s,w)= Ui( ˆ b(s)|s,w). If instead x > s, by the reason part 1c, Ui(f(x)|s,w) = Uw(x)|s,w) = Ui(b1(s)|s,w)=Ui(f(s)|s,w)=Ui( ˆ b(s)|s,w Suppose this bidder bids b1w'(x), x∈ [ˆs,ˆsi(b1']. Suppose first that s= ˆs. B reasoning in case 1b, Ui(b1(x)|s,w) = Ui(b1(s)|s,w). However, Ui(b1(s)|s,w 48 i ')|s,w). If s>ˆs, then Ui(b1(x)|s,w)= Ui(b1)|s,w). i i(f(ˆs (ˆs' ' 1(s). i(a(ˆs)|s,w)=Ui(b1 i(f(x)|s,w)=Ui(b1 (a) Bids of f(x), x∈[ˆs,sw i i(f(sw i(f(sw)|s,w)=Ui '' i(b1(x)|s,w)=Ui(b1 49 1 )|s,w)= U')|s,w)=Ui( ˆ b(sw U 3. Consider a bidder with a value-signal s= ˆsand a budget w= f(s). In this cas i(f(s)|s,w)=Ui( ˆ ß(s,w)=b b(sw (b1(s)|s,w)=U (a) Standard arguments show that U i(b1 (ˆs)|s,w)=U i(b1(s)|s,w). (b) Regarding bids f(x), x ∈ [sw (x), x∈[ˆs,ˆs w i(f(s)|s,w)=U (b1(x)|s,w)for all x∈[ˆs,ˆs i )=w. w ]. (b) Since s=ˆs, by a standard argument we can see that for all x=ˆs, U(a(x)|s,w)= U(s)|s,w). (c) Deviations to a bid above f(s) are not feasible. Thus, suppose this bidder bids f(x) for some x = s. Replicating the argument in case 1c shows that U(x)|s,w)=U 4. Consider a bidder with a value-signal s= ˆsand a budget f(s)=w=µ(ˆs). In this case, ß(s,w)=wand there exists some s=ssuch that f(s ), or ˆ b(s), x=ˆs, are not feasible deviations. ,s], then the reasoning of case 2c confirms that U(f(x)|s,w). (c) Regardingbidsf(x),x=s,thenthereasoningofcase1cshowsthatU(f(x)|s,w)= U(s)|s,w)=U)|s,w). (d) Bids of b], and a(y), y=ˆs, can be ruled out by reasoning similar to the arguments already presented above. The discussion above has studied all cases; thus, ß(s,w)is a symmetric equilibrium. References Araujo, Aloisio, & de Castro, Luciano I. 2009. Pure Strategy Equilibria of Single and Double Auctions with Interdependent Values. Games and Economic Behavior, 65(1), 25–48. Araujo, Aloisio, de Castro, Luciano I., & Moreira, Humberto. 2008. Non-Monotinicities and the All-Pay Auction Tie-Breaking Rule. Economic Theory, 35(3), 407–440. Ashlagi, Itai, Braverman, Mark, Hassidim, Avinatan, Lavi, Ron, & Tennenholtz, Moshe. 2010. Position Auctions with Budgets: Existence and Uniqueness. The B.E. Journal of Theoretical Economics: Advances, 10(1), Article 20. Athey, Susan. 2001. Single Crossing Properties and the Existence of Pure Strategy Equilibria in Games of Incomplete Information. Econometrica, 69(4), 861–889. Benoît, Jean-Pierre, & Krishna, Vijay. 2001. Multiple-Object Auctions with Budget Constrained Bidders. Review of Economic Studies, 68(1), 155–179. Birkhoff, Garrett, & Rota, Gian-Carlo. 1978. Ordinary Differential Equations. New York: John Wiley & Sons. Borgs, Christian, Chayes, Jennifer, Immorlica, Nicole, Mahdian, Mohammad, & Saberi, Amin. 2005. Multi-unit Auctions with Budget-Constrained Bidders. Pages 44–51 of: EC ’05: Proceedings of the 6th ACM Conference on Electronic Commerce. New York: ACM. Brusco, Sandro, & Lopomo, Giuseppe. 2008. Budget Constraints and Demand Reduction in Simultaneous Ascending-Bid Auctions. The Journal of Industrial Economics, 56(1), 113–142. Brusco, Sandro, & Lopomo, Giuseppe. 2009. Simultaneous Ascending Auctions with Complementarities and Known Budget Constraints. Economic Theory, 38(1), 105–124. Bulow, Jeremy, Levin, Jonathan, &Milgrom, Paul. 2009(March). Winning Play in Spectrum Auctions. Working Paper 14765. NBER. Che, Yeon-Koo, & Gale, Ian. 1996a. Expected Revenue of All-Pay Auctions and First Price Sealed-Bid Auctions with Budget Constraints. Economics Letters, 50(3), 373–379. Che, Yeon-Koo, & Gale, Ian. 1996b. Financial Constraints in Auctions: Effects and Antidotes. Pages 97–120 of: Baye, Michael R. (ed), Advances in Applied Microeconomics, vol. 6. Greenwich, CT: Jia Press. Che, Yeon-Koo, &Gale, Ian. 1998. StandardAuctions with Financially Constrained Bidders. Review of Economic Studies, 65(1), 1–21. Che, Yeon-Koo, & Gale, Ian. 1999. Mechanism Design with a Liquidy Constrained Buyer: The 2×2 Case. European Economic Review, 43(4–6), 947–957. 50 Che, Yeon-Koo, & Gale, Ian. 2000. The Optimal Mechanism for Selling to a BudgetConstrained Buyer. Journal of Economic Theory, 92(2), 198–233. Che, Yeon-Koo, & Gale, Ian. 2006. Revenue Comparisons for Auctions when Bidders have Arbitary Types. Theoretical Economics, 1(1), 95–118. Cho, In-Koo, Jewell, Kevin, & Vohra, Rajiv. 2002. A Simple Model of Coalitional Bidding. Economic Theory, 19(3), 435–457. Cox, James C., Smith, VernonL., &Walker, James M.1988. Theory andIndividual Behavior of First-Price Auctions. Journal of Risk and Uncertainty, 1(1), 61–99. Cramton, Peter. 1995. Money out of Thin Air: The Nationwide Narrowband PCS Auction. Journal of Economics & Management Strategy, 4(2), 267–343. Dasgupta, Partha S., & Maskin, Eric S. 1986. The Existence of Equilibrium in Discontinuous Economic Games, I: Theory. Review of Economic Studies, 53(1), 1–26. de Castro, Luciano I. 2010 (March). Affiliation, Equilibrium Existence and Revenue Ranking of Auctions. Working Paper. de Castro, Luciano I. 2011. Equilibrium Existence and Approximation of Regular Discontinuous Games. Economic Theory, 48(1), 67–85. de Castro, Luciano I., & Karney, Daniel H. 2012. Equilibria Existence and Characterization in Auctions: Achiements and Open Questions. Journal of Economic Surveys, 26(5), 911– 932. Dobzinski, Shahar, Lavi, Ron, & Nisan, Noam. 2012. Multi-unit Auctions with Budget Limits. Games and Economic Behavior, 74(2), 486–503. Fang, Hanming, & Parreiras, Sérgio. 2002. Equilibrium of Affiliated Value Second Price Auctions with Financially Constrained Bidders: The Two-Bidder Case. Games and Economic Behavior, 39(2), 215–236. Fang, Hanming, & Parreiras, Sérgio. 2003. On the Failure of the Linkage Principle with Financially Constrained Bidders. Journal of Economic Theory, 110(2), 374–392. Jackson, Matthew O. 2009. Non-existence of Equilibrium in Vickrey, Second-Price, and English Auctions. Review of Economic Design, 13(1–2), 137–145. Jackson, Matthew O., & Swinkels, Jeroen M. 2005. Existence of Equilibrium in Single and Double Private Value Auctions. Econometrica, 73(1), 93–139. Jackson, Matthew O., Simon, Leo K., Swinkels, Jeroen M., & Zame, William R. 2002. Communication and Equilibrium in Discontinuous Games of Incomplete Information. Econometrica, 70(5), 1711–1740. 51 Jaramillo, Jose E. Quintero. 2004 (January). Liquidity Constraints and Credit Subsidies in Auctions. Working Paper 04-06. Universidad Carlos III de Madrid. Kotowski, Maciej H. 2010 (November). First-Price Auctions with Budget Constraints. Working Paper, University of California, Berkeley. Kotowski, Maciej H. 2011. Studies of Auction Bidding with Budget-Constrained Participants. Ph.D. thesis, University of California, Berkeley. Kotowski, Maciej H., & Li, Fei. 2012 (January). On the Continuous Equilibria of AffiliatedValue, All-Pay Auctions with Private Budget Constraints. Working Paper, Harvard Kennedy School and University of Pennsylvania. Krishna, Vijay. 2002. Auction Theory. San Diego, CA: Academic Press. Krishna, Vijay, & Morgan, John. 1997. An Analysis of the War of Attrition and the All-Pay Auction. Journal of Economic Theory, 72(2), 343–362. Laffont, Jean-Jacques, & Robert, Jacques. 1996. Optimal Auction with Financially Constrained Buyers. Economics Letters, 52(2), 181–186. Lizzeri, Alessandro, & Persico, Nicola. 2000. Uniqueness and Existence of Equilibrium in Auctions with a Reserve Price. Games and Economic Behavior, 30(1), 83–114. Malakhov, Alexey, & Vohra, Rakesh. 2008. Optimal Auctions for Asymmetrically Budget Constrained Bidders. Review of Economic Design, 12(4), 245–257. Maskin, Eric S. 2000. Auctions, Development, and Privatization: Efficient Auctions with Liquidity-Constrained Buyers. European Economic Review, 44(4–6), 667–681. Maskin, Eric S., & Riley, John G. 2000. Equilibrium in Sealed High Bid Auctions. Review of Economic Studies, 67(3), 439–454. McAdams, David. 2007. Uniqueness in Symmetric First-Price Auctions with Affiliation. Journal of Economic Theory, 136(1), 144–166. Menezes, Flavio M., & Monteiro, Paulo K. 2005. An Introduction to Auction Theory. New York: Oxford University Press. Milgrom, Paul. 2004. Putting Auction Theory to Work. Cambridge: Cambridge University Press. Milgrom, Paul, & Weber, Robert J. 1982. A Theory of Auctions and Competitive Bidding. Econometrica, 50(5), 1089–1122. Monteiro, Paulo K., & Page, Frank H. 1998. Optimal Selling Mechanisms for Multiproduct Monopolists: Incentive Compatibility in the Presence of Budget Constraints. Journal of Mathematical Economics, 30(4), 473–502. 52 Myerson, Roger B. 1981. Optimal Auction Design. Mathematics of Operations Research, 6(1), 58–73. Pai, Mallesh M., & Vohra, Rakesh. 2011 (March). Optimal Auctions with Financially Constrained Bidders. Working Paper, University of Pennsylvania. Palfrey, Thomas R. 1980. Multiple-Object, Discriminatory Auctions with Bidding Constraints: A Game-Theoretic Analysis. Management Science, 26(9), 925–946. Pitchik, Carolyn. 2009. Budget-Constrained Sequential Auctions with Incomplete Information. Games and Economic Behavior, 66(2), 928–949. Pitchik, Carolyn, & Schotter, Andrew. 1988. Perfect Equilibria in Budget-Constrained Sequential Auctions: An Experimental Study. RAND Journal of Economics, 19(3), 363–388. Reny, Philip J. 1999. On the Existence of Pure and Mixed Strategy Nash Equilibria in Discontinuous Games. Econometrica, 67(5), 1029–1056. Reny, Philip J. 2011. On the Existence of Monotone Pure Strategy Equilibria in Bayesian Games. Econometrica, 79(2), 499–553. Reny, Philip J., & Zamir, Shmuel. 2004. On the Existence of Pure Strategy Monotone Equilibria in Asymmetric First-Price Auctions. Econometrica, 72(4), 1105–1124. Rhodes-Kropf, Matthew, & Viswanathan, S. 2005. Financing Auction Bids. RAND Journal of Economics, 36(4), 789–815. Richter, Michael. 2011 (October). Mechanism Design With Budget Constraints and a Continuum of Agents. Working Paper, New York University. Rothkopf, Michael H. 1977. Bidding in Simultaneous Auctions with a Constraint on Exposure. Operations Research, 25(4), 620–629. Salant, David J. 1997. Up in the Air: GTE’s Experience in the MTA Auction for Personal Communication Services Licenses. Journal of Economics & Management Strategy, 6(3), 549–572. Simon, Leo K., & Zame, William R. 1990. Discontinuous Games and Endogenous Sharing Rules. Econometrica, 58(4), 861–872. Topkis, Donald M. 1998. Supermodularity and Complementarity. Princeton, NJ: Princeton University Press. Vickrey, William. 1961. Counterspeculation, Auctions, and Competitive Sealed Tenders. Journal of Finance, 16(1), 8–37. Zheng, Charles Z.2001. High Bids and Broke Winners. Journal of Economic Theory, 100(1), 129–171. 53 X Online Appendix X.1 Example 1 In this section we prove that the strategy proposed in Example 1 is an equilibrium. The proof is in a more general environment rendering the construction more transparent. Suppose bidders have independent private values which are distributed according to the c.d.f. H:[0,¯s]→[0,1]. H(s)admits a density, h(s), which is strictly positive for s∈(0,¯s). With probability pa bidder has abudget w which we assume satisfies 0<w<¯s-¯s 0H(z)dz. With probability 1-pa bidder has a budget of ¯w= ¯s. (Example 1 does not meet this requirement but the argument is unchanged since a bidder with a high budget bids less than 3/4 in equilibrium anyway.) Ties are resolved with a fair coin flip. We begin with two preliminary lemmas which define ˆsand ˆs'1 H(˜s). Let ˜sbe the solution to w =˜s-˜s 0H(z)dz.Lemma X1. There exists a unique ˆs∈(w ,˜s)such that Proof. Let t(s)=[p+(1-p)H(s)](s-w )s 0ˆs 0H(z)dz=[p+(1-p)H(ˆs)](ˆs-w ).H(z)dz. It suffices to show that t(s)crosses zero exactly one time. When s=w , t(w)<0. If instead s=˜s, t(˜s)=[p+(1-p)H(˜s)](˜sw)-H(˜s)(˜s-w)>0. Thus, there exists ˆs∈(w,˜s)such that t(ˆs)=0.To show that ˆsis unique, note that t'(s)=p(1-H(s))+(1-p)(s-w )h(s)>0when w =s<¯s. Thus, t(s)is strictly increasing and therefore ˆsis unique. Lemma X2. Let ˆsbe as defined in Lemma X1. There exists a unique ˆs'0ˆs>ˆssuch that H(z)dz+ ˆs'ˆs[(1-p)H(ˆs)+pH(z)]dz= (1-p)H(ˆs)+p1+H(ˆs') 2 (ˆs'-w ). Proof. Let t(s)= χ(s)= 0ˆsH(z)dz+ ˆss[(1-p)H(ˆs)+pH(z)]dz, and(1-p)H(ˆs)+p 1+H(s)2 (s-w ). X1 t(ˆs)= 0 At s=ˆs, Instead, if s=¯s, 1+H(s) 2 ˆs H(z)dz =[p+(1-p)H(ˆs)](ˆs-w )> (1-p)H(ˆs)+p 1+H(ˆs)2 (ˆs-w )=χ(ˆs). [(1-p)H(ˆs)+pH(z)]dz <[p+(1-p)H(ˆs)](ˆs-w )+[p+(1-p)H(ˆs)](¯s-ˆs) t(¯s)=[ =[p+(1-p)H(ˆs)](¯s-w)=χ(¯s). p+(1p)H(ˆs )](ˆs-w )+ ˆs¯s ' ' )=χ(ˆs' '∈(ˆs,¯s)such that t(ˆsis '(s) <χ unique, note that t ' p 2h(s)s. Since p1+H(s) 2 ' > p H ( s ) , ' t Thus, there exists ). To show that ˆs(s) = (1-p)H(ˆs)+pH(s) but χ+(s) = (1p)H(ˆs)+p'(s) and s→t(s)-χ(s) ˆs' strictly decreasing. Thus, ˆsis unique. Theorem X1. Let ˆsand ˆsbe as defined in Lemmas X1 and X2. The strategy profile where all bidders follow the strategy 1 H(s) s 0 s ˆs[p+(1-p)H(z)]dz p+(1-p)H(s)s ˆs[pH(z)+(1-p)H(ˆs)]dz pH(s)+(1-p)H(ˆs) ' i f ˆs 0 X2 ˆs 0 s ∈ ( ˆ s , ˆ s ' i , ¯ s ] , ( w X = 1) w s0 H(z)dz+ H(z)dz if s∈[0,ˆs],w∈{w ,¯w} s-H(z)dz+ if s∈(ˆs,¯s],w= ¯w s-],w=w w if s∈(ˆs ß ( s s is a symmetric Bayesian-Nash equilibrium. Proof. To simplify notation, let¯ , w ß(s)=ß(s,¯w)and ß (s)=ß(s,w). We will also hold fixedthe strategy of player j=iat the ) = strategy ßand suppress it from our notation. If any type has a profitable prescribed deviation from ß(s,w), it will be to different feasible bid in the range of ß; thus, we need only verify that such deviations are not worthwhile. In tota there are five cases to consider. 1. Consider bidder iof type s∈ [0,ˆs] with budget ¯w. If this bidder follows the prescribed strategy, her expected payoff is U( ¯ ß(s)|s,¯w)=H(z)dz. There are 4 kinds of alternative bids to consider. (a) A deviation to a bid ¯ ß(x), x∈ [0,ˆs], is unprofitable. The standard argument'], will give this bidder a payoff ofconfirming an equilibrium in a symmetric, first-price sealed-bid auction applies. (b) A deviation to a bid ß (x), x∈(ˆs,ˆs Ui(ß (x)|s,¯w)=[(1-p)H(ˆs)+pH(x)](s-x)+ 0ˆsH(z)dz+ ˆsx[(1-p)H(ˆs)+pH(z)]dz. The net gain from this alternative bid is Ui(ß (x)|s,¯w)-Ui( ¯ ß(s)|s,¯w)=[(1-p)H(ˆs)+pH(x)](s-x)+ = sˆssˆsH(z)dz+ H(z)dz+[(1-p)H(ˆs)+pH(x)](s-ˆs)+pH(x)(ˆs-x)+ ˆsxpH(z)dzˆsx[(1-p)H(ˆs)+pH(z)]dz =H(ˆs)(ˆs-s)+[(1-p)H(ˆs)+pH(ˆs)](s-ˆs)+pH(x)(ˆs-x)+pH(x)(x-ˆs)=0 Thus, Ui(ß (x)|s,¯w)=Ui( ¯ ß(s)|s,¯w). (c) Since the bid w is placed with strictly positive probability by bidder j=iwhen she is following the prescribed strategy, bidder iwill never wish to deviate to this bid since a bid slightly higher will guarantee a strictly higher payoff. A slightly higher bid is feasible for bidder isince she has a budget of ¯w. (d) Finally, consider a deviation to a bid ¯ ß(x), x∈(ˆs,¯s]. This bid will give bidder ian expected payoff of 0 ˆ H(z)dz+ ˆsx[p+(1-p)H(z)]dz. s Ui( ¯ ß(x)|s,w)=[p+(1-p)H(x)](s-x)+ X3 The net gain from this alternative bid is Ui( ¯ ß(x)|s,¯w)-U= sˆsi( ¯ ß(s)|s,¯w)H(z)dz+[p+(1-p)H(x)](s-ˆs)+[p+(1-p)H(x)](ˆs-x)+ ˆsx[p+(1-p)H(z)]dz=H(ˆs)(ˆs-s)+[p+(1-p)H(x)](s-ˆs)+ ˆsx(1-p)[H(z)-H(x)]dz=0 The final inequality follows from x=ˆs=s. Thus, Ui( ¯ ß(x)|s,¯w)=Ui( ¯ ß(s)|s,¯w). 2. Consider bidder iof type s∈ (ˆs,¯s] with budget w= ¯w. If this bidder follows the prescribed strategy, her expected payoff is Ui( ¯ ß(s)|s,¯w) =ˆs 0H(z)dz+ s ˆs[p+(1p)H(z)]dz. There are four kinds of alternative bids she may consider. (a) A deviation to a bid ¯ ß(x), x∈[0,ˆs], will give this bidder a payoff of Ui( ¯ ß(x)|s,¯w)=H(x)(s-x)+ 0 x H(z)dz. The net gain from this alternative bid is Ui( ¯ ß( x)| s,¯ w) -Ui (¯ ß( s)| s,¯ w) =H (x) (sx) + 0= H( x)( ˆsx) + = xˆsˆ sxx H( z)dz0ˆs H( z) dzˆ sH (z) dz +H (x) (sˆs) sˆs[ p+ (1p) H( z)] dz s[p +( 1p) H( z)] dz [H(x)-H(z)]dz+H(x)(s-ˆs) -[p+(1-p)H(ˆs)](s-ˆs)=0 'Therefore, this deviation is not profitable. (b) A deviation to a bid ß (x), x∈[ˆs,ˆs], will give this bidder a payoff of Ui(ß (x)|s,¯w)=[(1-p)H(ˆs)+pH(x)](s-x)+ 0ˆsH(z)dz+ ˆsx[(1-p)H(ˆs)+pH(z)]dz. X4 The net gain from this alternative bid is Ui(ß (x)|s,¯w)-Ui( ¯ ß(s)|s,¯w)=[(1-p)H(ˆs)+pH(x)](s-x)+ ˆss[p+(1-p)H(z)]dzˆs=[(1-p)H(ˆs)+pH(x)](ˆs-x)+ ˆsxx[(1-p)H(ˆs)+pH(z)]dz[(1-p)H(ˆs)+pH(z)]dzs ˆ [p+(1-p)H(z)]dz s ˆ s p[H(z)-H(x)]dz = +[(1-p)H(ˆs)+pH(x)](s-ˆs)x +[(1-p)H(ˆs)+pH(x)](s-ˆs)-[p+(1-p)H(ˆs)](s-ˆs) =ˆsxp[H(z)-H(x)]dz+p(s-ˆs)(H(x)-1)=0 Therefore, this deviation is not profitable. (c) As in the previous case, a bidder can strictly improve upon the bid w by bidding omore. (d) A deviation ¯ ß(x), x∈(ˆs,¯s], will give this bidder a payoff to a bid of 0 ˆ H(z)dz+ ˆsx[p+(1-p)H(z)]dz. Ui ( ¯ ß(s)|s,¯w)=[p+(1-p)H(x)](s-x)+ s (1-p)[H(z)-H(x)]dz=0 Therefore, this deviation is not profitable. 3. Consider a bidder of type s∈ [0,ˆs] with budget w= w . By identical arguments to those presented above fora high-budgetbidder with avalue-signal less than ˆs, this type of bidder will not have a profitable deviation to any bid, except possibly ¯w= ß (x), X5 x The net gain from this alternative bid is =[p+(1-p)H(x)](s-x)+ = sxs [p+(1-p)H(z)]dz Ui( ¯ ß(x)|s,¯w)-Ui( ¯ ß(s)|s,¯w) x>ˆs'. Noting the uniform tie breaking rule, this deviation will yield a payoff of 1+H(ˆs') (s-w ). Ui(w|s,w)= (1-p)H(ˆs)+p 2 The net gain from this alternative bid is ' [(1-p)H(ˆs)+pH(z)]dz=0 ˆs (1-p)H(ˆs)+p 1+H(ˆs) 2 [(1-p)H(ˆs)+pH(z)]dz '(1-p)H(ˆs)+p 1+H(ˆs) 2 + (1-p)H(ˆs)+p 1+H(ˆs) 2 Ui(w |s,w)-U= = = i(ß(s)|s,w) '(1-p)H(ˆs)+p (s-ˆs' (s-ˆs')+ 1+H(ˆs') -w )0 (1-p)H(ˆs)+p)+ (s-ˆs)+' 2 (ˆs' (ˆs-ˆsssˆsˆsH(z)dz+ H(z)dz)+ ' 1+H(ˆs) 2 'ˆsˆsˆs' Therefore, this deviation is not profitable. Consider a bidder of type s∈ (ˆs,ˆs] with budget w= w . If this bidder follows the prescribed strategy, her expected payoff is Ui(ß (s)|s,w)= ˆs 0H(z)dz+ s ˆs[pH(z)+(1p)H(ˆs)]dz. (a) A deviation to a bid ß (x), x∈[0,ˆs], will give this bidder a payoff of Ui(ß (x)|s,w)=H(x)(s-x)+ 0x H(z)dz '4. The net gain from this alternative bid is U i ( ß ( x ) | s , X6 [ ( 1 p ) H ( ˆ s ) + p H ( z ) ] d z s [ ( 1 p ) H ( ˆ s ) + p H ( z ) ] d z =0+H(x)(s-ˆs)-[(1-p)H(ˆs) +pH(ˆs)](s-ˆs) =[H(x)-H(ˆs)](s-ˆs)=0 Therefore, this deviation is not profitable. '(b) A deviation to a bid ß(x), x∈[ˆs,ˆs], will give this bidder a payoff of Ui(ß (x)|s,w)=[(1-p)H(ˆs)+pH(x)](s-x)+ 0 ˆ H(z)dz+ s [(1-p)H(ˆs)+pH(z)]dz The net gain from this alternative bid is Ui(ß (x)|s,w)-Ui(ß(s)|s,w) =[(1-p)H(ˆs)+pH(x)](s-x)+ x [(1-p)H(ˆs)+pH(z)]dz. ˆs sx = sx p[H(z)-H(x)]dz=0 Therefore, this deviation is not profitable. (c) A deviation to a bid of w can be ruled out by an argument fully analogous to that presented in case 3 above. Consider a bidder of type s∈(ˆs,¯s]with budget w=w . If this bidder follows the prescribed strategy, her expected payoff is Ui'1+H(ˆs) 2 (a) A deviation to a bid ß (x), x∈[0,ˆs], will give this bidder a payoff of H(z)dz. Ui(ß (x)|s,w)=H(x)(s-x)+ 0x '5. X7 w). [(1-p)H(ˆs)+pH(z)]dz The net gain from this alternative bid is 1+H(ˆs') Ui(ß (x)|s,w)-Ui(ß(s)|s,w)=H(x)(s-x)+ =H(x)(s-x)+ 2 00xxH(z)dz (1-p)H(ˆs)+pH(z)dz'(1-p)H(ˆs)+p 1+H(ˆs=H(x)(ˆs-x)+ ˆsx0) 2H(z)dz'ˆsˆsH(z)dz (s-ˆs')ˆsˆs' '+H(x)(ˆs -ˆs)ˆs [(1-p)H(ˆs)+pH(z)]dz +H(x)(s-ˆs ' ) (1-p)H(ˆs)+p 1+H(ˆs') (s-ˆs')=0 2 (s-w ) 'Therefore, this deviation is not profitable. (b) A deviation to the bid ß (x), x∈[ˆs,ˆs], can be ruled out with reasoning similar to the preceding case. The above discussion has exhausted all possible cases; therefore, (X1) is a symmetric equilibrium strategy profile. Remark X1. This examplecanbegeneralized tothecaseofN>2bidders. The equilibrium’s qualitative features remain unchanged. X.2 Example 2 The reasoning below echos the proof of Theorem X1. We specialize to the parameters of interest to simplify the overall argument. X8 16 s 11-3)s224(s+ v 11-3)2 s2 2s2+13 v11-42 4(s+1) 3+24( v ''if s∈(ˆs ,ˆs' ] ''if s∈[0,ˆs] if s∈[0,ˆs] +33(19 v 11-63) 2s 3 Theorem X2. The following is a symmetric equilibrium in Example 2. ß(s,1/4)= if s∈(ˆs,1] if s∈(ˆs ß(s,3/4)= 1 2 ',1] (ˆs-1/4) 2 ˆs' +24( v11-3)s2 wh er e ˆs =v where q(s)= Pro of. As in the pro of of The ore m X1 ther e ' ' = 1 1 3 , 11-3), and ˆs+ˆs 2solves the equation (ˆs-q(ˆs))= 1 2'· ˆs+1 2+ 4 v 12 ˆs 3( ˆ s 16s 3 +33(19 v11-63) 24(s+v 11-3) . (Approximately, ˆs' ˜0.298.) 14 ' 1 2 are mul tipl e cas es. Cle arly , we nee d onl y con sid er pos sibl e dev iati ons into the ran ge of ß(s, w). To sim plify not atio n, we will writ e ¯ ß ( s ) = ß ( s , 3 / 4 ) a n d ß ( s ) = ß ( s , 1 / 4 ) . 1. C o n si d er a bi d d er wi th a b u d g et of w = 3/ 4 a n d a v al u esi g n al s ∈ [0 ,ˆ s] . If th is bi d d er pl a c e s a bi d ¯ ß( x) , x ∈ [0 ,ˆ s] , s h e wi ll d ef e at al l hi g hb u d g et o p p o n e nt s wi th a v al u esi g n al le ss th a n x a n d al l lo w -b u d g et o p p o n e nt s wi th a v al u esi g n al le ss th a n 3 x/ 4. T h u s, th e pr o b a bi lit y th at s h e wi n s th e a u cti o n is 7 x/ 8. T h e e x p e ct e d p a y o ff fr o m th is bi d is th er ef or e Ui( ¯ ß(x)|s,¯w)= M a xi m izi n g th e a b o v e e x pr e ss 7x8 s- x . 2 io n wi th re s p e ct to x =ˆ s gi v e s a s ol ut io n of x = s. T h u s, ¯ ß( s) is th e o pt i m al bi d w h e n c o n st ra in e d to a bi d in th e ra n g e [0 ,ˆ s/ 2] . S u p p o s e in st e a d th is bi d d er bi d s ¯ ß( x) a n d x ∈ (ˆ s, 1] . T hi s bi d wi ll d ef e at al l lo w -b u d g et o p p o n e nt s si n c e it is a b o v e 1/ 4 a n d it wi ll d ef e at al l hi g hb u d g et o p p o n e nt s wi th a v al u esi g n al le ss th a n x. T h er ef or e, th e e x p e ct e d p a y o ff fr o m th is bi d is Ui( ¯ ß(x)|s,¯w)= X 9 1+x2 (s¯ ß(x)) It is easy to calculate that definition of ˆs, we see that 1 1+ˆs 2 s- d Ui( dx ¯ ß(x)|s,¯w) = (s-x)/2 = 0. However, given our 4 2 = ( v 11-2)(4s-1)7s8 =16 ∀s∈[0,ˆs] Since Ui( ¯ ß(s)|s,¯w)= Ui( ¯ ß(x)|s,¯w)= lim+x→ˆs= ' Ui( ¯ ß(x)|s,¯w) ''). '), For x∈[ˆs 7s2 16 = 7s∆(x,s)=Ui216 - '' i (s-ß (x)). ,ˆs, this proposed deviation is not profitable. The remaining candidate deviation is to a bid ß (x), x∈[ˆs (¯ (ß ß(s)|s,¯w)-U (x)|s,¯w) x+ˆs2 ' ,ˆs' ' It is sufficient to establish that ∆(x,s) = 0 for all x∈ '' [ˆs''(noting that ß (s)is continuous at ˆs'' = 78 (s-ˆs)=0. ∂∆(ˆs'',s) ∂s ''Hence, for all s=ˆs, ∆(ˆs 8(v 11-3-3s)+32x49(19 v 11-63)(x+ v =0 = 148 11-3) 2 8(v 11-3-3ˆs)+32ˆs'' = (ˆs + v 11-3)2 '' 148 ) and s= ˆs. From the definition of ˆsand ˆs), ∆(ˆs,ˆs)=0. Next, note that ,ˆ s let ,s)=0. Thus, it is sufficient to confirm that 49(19v 11-63) ∂∆(x,s) ∂x ∂∆(x,s) ∂x =0for all s∈[0,ˆs]. A direct computation gives where we used the fact that 19 v11-63>0. Therefore, a bidder of type s<ˆshas no incentive to deviate into this range of bids. 2. Consider a bidder with a budget of w = 3/4 and a value-signal s∈ (ˆs,1]. By an argument fully analogous to the preceding case, all deviations to bids¯ ß(x), x∈[0,1]can be ruled out. We therefore consider a possible deviation to a bid ß (x) where X10 1+s 2 ' ' ¯ ß(s)),ˆs'). Let (s-ß (x)). ( ¯ ß(s)|s,¯w)-U (s-i(ß (x)|s,¯w) x+ˆs2 x ∈ ∆(x,s)=Ui [ = ˆ s '',ˆs)=0. Note that for any x∈[ˆs '' ,ˆ s' ∂∆(x,s) = 4- v 11+s-x2 = 4v11+ˆs-x 2 >0. ∂s ]. As before, ∆(ˆs'',ˆs'], Again, it is sufficient to verify that ∆(x,s)=0for s∈(ˆs,1]and x∈[ˆs ∂∆(x,ˆs) =0. A direct calculation gives Thus, it is sufficient to confirm that ∂x ''= 2ˆs3 11 +1=0. 3 ∂∆(x,ˆs) = 2x3 - 49(19 v11-63) 48(x+v +1 ∂x 11-3)49(19 v211-63) 48(ˆs''+ v11-3)v211 3v Thus, a bidder with a value-signal of s>ˆshas no profitable deviation from ¯ ß(s). ''3. Consider a bidder with a budget of w= 1/4 and a value-signal s∈ [0,ˆs'']. If this bidder places a bid of ß (x), x∈[0,ˆs], she will defeat all low-budget opponents with a value-signal less than xand all high-budget opponents with a value-signal less than 4x/3. Therefore, the expected payoff from this bid is 1/2 Ui (ß (x)|s,w)= 7x6 s2x 3 . ''Maximizing the above expression with respect to x= ˆs Ui(ß(x)|s,w)= x+ˆs 2 (s-ß (x)) X11 1 2 . gives a solution of x= s. Thus, ß(s)is the the range [0,'',ˆs'3]. If instead this bidder plac payoff would be ''2ˆs =0. Differentiating this expression with respect to xgives 2 49(19 v 11-63)(x+ v 11-3) d dxUi(ß (x)|s,w)= v8(3s+ v6(s-x) 11-3)-16xTherefore, since ß is continuous at ˆs'', Ui(ß(x)|s,w)=Ui (ß(ˆ '' )|s,w)=U(ß(s)|s,w). Finally, since ˆs''=3 4( vi11-3)˜0.23, a s negative payoff. Therefore, this type of bidder does not (ß (s)|s,w)= 8s-49( 19 v 11-63) ( s+ = v 11-33 2)2(ˆs+ ,ˆ v+8( v 11-3)8ˆs''11-3) = (-49(19 v 11-63)(ˆs''+ v 11-3)v s' 11-3) 382v 11-1262+8( v 11-3)˜0.563 ]. By an argument fully analogous to th ''4. Consider a bidder with a budget of w = 1/4 and a value-signal s ∈ (ˆs (x), x∈ [0,ˆs] can be ruled out. We there the bid of w = 1/4. Define ∆(s) = U(w|s, '' (s) = 0; equivalently, that i(ß(s)|s,w)-Ui) = 0. Therefore, it is sufficient to , ∆(ˆs verify that ∆ = 11-5 4v 4ˆs-1= 11-29 16 ˜0.901and ˆs ' d dsU v 1 1 3 U i i( ß(s)|s, w )=11-5 4v 4s-1 3 s+ 2 5. Considerabidderwithabudgetofw=1/4andavalue-signals∈(ˆs,1]. Byarguments analogous to those presented in the previous cases, it is simple to verify that ß(ˆs ')is the utility-maximizing, feasible bid other than w . Thus, we verify that U(w|s,w)= Ui(ß (ˆs'')|s,w). By definition of ˆs, Ui'(w|ˆs,w)=Ui(ß(ˆs'')|ˆsi,w). Thus, it is sufficient to X12 ' verify that for all s>ˆs', d dsUi d d s U i = ( ˆ s 11+ ' ' + 2 v d d s ( ß ( w ( ˆ s | s , w ) Ui (ß (ˆs')|s,w)= ) | s , w ) Ui(w |s,w). Consider therefore, ( +2 v11-5) ˆ 8(3s+ v2 ' s v24s-6(ˆs - '11-3)-16ˆs+ v11-3) ' = ' 11-5) 2 v 11)s-86 v8 v 8(3s+ v 11-3)-16ˆs- 49(19 v 11-63)(ˆs''+ v 11-3)2 = ( 24s-6(ˆs+ v11-3) 3 4s-1 This final expression is decreasing in sand at s=1we can 3 3 conclude that + 6 8 v 50 23 d (ß (ˆs 'i U i which is the desired d ( s result. w d ds The pre ced ing disc ussi on has exh aus ted all pos d d s | s , w ) U )|s,w) = 1 1501-54v 11˜1.51=1 2 4 49(19 v 11-63)(ˆs'+ v 11-3)2 sibl e cas es; ther efor e, ß(s, w)is a sy mm etri c equ ilibri um. X .3 R el a xi n g A s s u m pt io n A 3 ( 1 b ) As not ed in foot not e 8, in the cas e of two bid der s with qua silin ear pref ere nce s Ass um ptio n A-3 (1b) is not nec ess ary. In this sec tion we ela bor ate on this rem ark. ˜ A, d e fi n e d b el o w, m a y n ot b e a s u bl at tic e. 20 T h e o nl yr ol e of A ss u m pt io n s A3( 1 a) a n d A3( 1 b) isi nt h e v er ifi c at io n of a si n gl e cr o ss in g c o n di ti o n. T h er ef or e, it s u ffi c e s to s h o w th at L e m m a A 3 o bt ai n s wi th o ut A ss u m pt io n A 3( 1 b) . L e m m a X 4 d e m o n st ra te s th is c o n cl u si o n. G e n er al ly, th is ar g u m e nt d o e s n ot a p pl y if N = 3 si n c e th e s et Lemma X3. Let s' i =si and suppose =sj =¯sj, then s' j |s' i Proof. We note that for all sj, ¯ H(sj|si)= ¯ H(sjj' i=siand zj=tj 20 ; thu s, the upp er bou nd foll ows . Sin ce h(·) is logsup erm odu ) ¯ H(sj|s' i) ¯H(sj|si) = ¯ H(s' j|s' i) ¯H(s' j|si) =1. |si)h(tj |s' i)= lar, for s, the foll owi ng ine qua lity hol ds: h(z T h i s d i s c u s s i o n c o r r e c t s a n d c l a r i fi e s a r e s u l X 1 3 t o r i g i n a l l y i n c l u d e d i n t h e w o r k i n g p a p e r K o t o w s k i ( 2 0 1 0 ) . i)dzj sj |s' i)h(tj|si). Therefore, j |s' i)=h(zj |si)h(tj j=s' j h(tj|s' i h ( z 'j t j j j=s ) =⇒=zh(zj|stj=sj)dt= sj=zj=sj h(zj ¯ H(s' ¯ H(s|sj ' i)= ¯ H(s' ' i|s' i)¯ ) ¯ j|s' i)= ¯ H(s' H(sj|s) ¯ j|s H(sj j H(sj|si) (si,s-i,wi,b j j ) ¯ H(sj|si H(sj H(sj )=v ' i'i i) -i = ¯ H(s|s|si) H(s (wi)=wi i)gj(wj)dwjdsj. i , u 2 0= ˜ A i, ˜A =⇒ ¯ H(s' j|si)¯ H(sj|si |s' i)h(tj |si h(z |s' i)dz i) =⇒=⇒ ¯|s) ¯|s|s' j' i) ¯' j j 'i i(si,s )+wi -b i Lemma X4. Suppose ¯ui ') i i ,b )h(sj|s θi and w i , and N=2. Then Lemma A3 continues to hold without Assumption A-3(1b). Proof. We first note that the =ordering reduces to the coordinate-wise partial order of R. We specialize our notation to the case of two bidders. Noting the proof of Lemma A3, it is sufficient to show that for s=s=w ui(s' i,sj,w' i,bi)h(s |s' (wj)dwjdsj = ui(si,sj ,wi i i)gj is independent of , the random variables (Si,Sj,W Sj h( j i i ×Sj ×W j and wj where the ˜ A={(sSince W jj,wj follows: (si,sj,wj) ˜ =(s, (Si' i,S,sj' j,w,W j' j)if and =s' i, sj =s' j only if si j jas j ) < ' (sj,wj)=bi}and Pr[ ˜ A]>0. j : ß i j b )are affiliated. Redefine the ordering of S=w' j. Again by the independence of W)are affiliated when “affiliationness” is defined with the˜ = order. S×S×Wremains a lattice under this alternative ordering. j ×W under the ˜ = ordering. To see this, suppose= and w j The set ˜ A is a sublattice w' j. Then, since ßj of S) ∈ ˜ Abut s= s' jj (sj,wj),(s' j,w' j is nondecreasing in the usual coordinate-wise partial order, =ßj(s' j ,w' j)=ßj (sj,w' j)=ßj (sj˜ ∨s',w j j˜ ∨w' j)=ßj (sj,wj)=bi (sj,wj)=bi. (s j˜ ∧s' j b' i ,wj˜ ∧w' j)=ßj =ßj(s' j ,w' j)=ßj (s' j,wj)=ßj 'i , and b X14 Thus, (sj ˜ ∧s' j˜ ∧w' j),(sj˜ ∨s' j,wj˜ )∈ ˜ A. We can now establish the j,w following: ∨w' j ui(s' i,sj,w' i,bi)h(sj |s' (wj)dw jdsj i)gj j ' i) |s' i)gj(wj)dwjdsj j|s ¯ ˜A = ˜ A= u˜ i(s' i,sj,w' i,bi) h(sj|s' i) H(sj A ¯H(s |s' i)gj(wj)dwjdsj ) ¯H(s|s ui(si,sj,wi,bi) h(s |s H(sj j i ) ¯i =Pr[ ˜ =si ,wi,bi ) ¯H(Sj|s' i) ˜ =si A|Si ¯H(Sj|si) A,Si i ii ¯H(Sj|s' i) i)| ˜ E ¯|s) ˜ A,S i =si j ]E ]E uu]E ,wi,buiiii(s(s(siii ,S,S,S)h(sjjjj|si = =Pr[ ˜ si A|Si =si =Pr[ ˜ A|S= ˜ ,sj Auii(si ,wi, b ,w ,b )| ˜ A,S = si H( Sj A,S i i =si dsj )g (wj)dwj The first inequality follows from Assumption A-3(1a). The second inequality follows from the fact that ui(si,·,wi,bi):Sj×W j→ R is nondecreasing,¯ H(·|s' i) ¯H(·|si):Sj×W ¯ H(sj|s' i→ R is nonincreasing (Lemma X3), and Milgrom & Weber (1982, Theorem 23). The final inequality follows from) ¯H(sj|si)being positive and bounded above by 1.j X.4 Multiple Critical Points Multiple critical points can arise if ψ(s) is not monotone. Our conclusions carry over to this case as well. The analysis’ essence is to identify a family of connected, upward sloping solutions from (4) and to link appropriate solutions at critical points to define the bidding strategy. Here we sketch how these definitions may be carried out. Definition of ¯ b(s)(Low w ). Figure X1 presents a typical case of three critical points and w=0. The central critical point is an unstable node while the others are saddle points. The thick arrows indicate the motion of (4) in the various regions. Confining attention to the regions between ν(s) and ψ(s), we are able to link adjacent nodes constructing a path for ¯ b(s)which satisfies the desired conditions. Definition of ˆ b(s) (Intermediate w ). The preceding procedure can be generalized to define ˆ b(s)whenthere aremultiple critical pointsandw isintermediate. Inthiscase however, X15 b b ν ¯b(s) ψ 0 s Figure X1: Multiple critical points and the definition of ¯ b(s). explicitconsiderationneedstobegiventoµ(s). Consider, forexample, thesituationinFigure X2 with three critical points. Suppose the equilibrium strategy is discontinuous at ˆs. µ(s) is indicated by the thick dotted curve. Since solutions to (4) radiate from the critical point at s* 2, the initial segment of ˆ b(s)approaches this critical point from below. Beyond s * 2, ˆ b(s) equals ¯ b(s)which ensures continuous passage through the final critical point at s* 3. X.5 Comparing First- and Second-Price Auctions In this section we prove the following standard result. LemmaX5. Supposew =0andletß1(s,w)andß2(s,w)besymmetric equilibriumstrategies for the first-price and second-price auctions respectively. Then, ß1(s,w)=ß(s,w)and the inequality is strict if ß1(s,w)<w.2 Proof. If there are no critical points, the result is immediate as ß1(s,w)=v(s,s)=ß(s,w). Suppose instead that there is a critical point (s*,b*2)and let ¯ b(s)be the solution (3) passingthrough (s* ,b*). It is sufficient to show that ¯ b(s)<b2(s). From their definitions, it is easy X16 b ψ ν s ¯ b(s) s * 2 s ˆs a ˆ b(s) s* 3 * 1 w s0 Figure X2: Construction of ˆ b(s) (orange) when w is intermediate and there are multiple critical points. µ(s)is the dotted curve. For s>s* 2, ˆ b(s)= ¯ b(s) . . As both functions are continuous, theremustex Notingthatd(b,s|s)>b, b' 2(ˇs)< ¯ b'(ˇs). Therefore when s<s*. This contradicts b2(s*)>v(s*,s*). Thus t o see that for all s∈[s*,1], bSuppose instead that b22(s)>v(s,s)= ¯ b(s).(s)= ¯ b(s)for some 0<s<s* G0(w)= w2 , G1(w)= w 2, G X17 ' (w)= w(4-w)14 . X.6 Change in the Budget Distribution In this section we detail our comparative static c constraints. Our argument is a simple example. with three alternative distributions of budgets on ψ1(s)= 9s likelihood ratio dominates both G1 and G k 9s5 4 3' . The critical p 2 implies a distinct ψk(s): ψ0(s)= 3s2 -7s2 v9s4 1 0(s* 0). +3s ) occurs at G the intersection of ν(s)and each distribution G +3s -2s+1) ψ' (s)= 9-6s-3s12-36s3+48s2-84s+57 6-6s 1 1 1 Table 1 summarizes approximate values for the critical point in each case. The critica s 8 (points are strictly ordered: s* 1<s* 0<s* 1'. In a neighborhood of s* 0, unconstrained bidd s bid less when budgets are distributed according to Gversus G 0: ¯ b1(s* 0)< ¯ b0(s* 0). T converse is true when budgets follow G' : ¯ b1' (s1* 0)> ¯ b Table 1: Critical Points for Different Budget Distributions b +2s-1 4s-43+3s+3-4 v2 0 Distribution s* k *k 1 0.0864G 0.1728 G00.1056 0.2111 G' 0.1264 0.25281 X18