Approximation in Mechanism Design with Interdependent Values YUNAN LI, University of Pennsylvania The seminal work of [Myerson 1981] shows that the simple Vickrey-Clarke-Groves (VCG) mechanism with monopoly reserves is optimal in single-item auctions where agents have independent and identically distributed private values. [Hartline and Roughgarden 2009] and others prove approximate analogs of this statement in the more general one-dimensional type, independent private values setting. This paper studies the design of ex-post incentive compatible and approximately optimal mechanism in an environment with interdependent values. We show that the VCG mechanism is ex-post incentive compatible under a matroid feasibility constraint. We also give conditions under which the VCG mechanism with monopoly reserves could still generate near-optimal expected revenues. Categories and Subject Descriptors: F.0 [Theory of Computation]: General General Terms: Economics, Theory, Algorithms Additional Key Words and Phrases: Mechanism Design, Approximation, Interdependent Values, Revenue Maximization 1. INTRODUCTION In the independent private values (IPV) setting, it is well known that the VickreyClarke-Groves (VCG) mechanism is ex-post incentive compatible. In addition, when there is a single item for sale and the agents’ values are drawn from identical distributions, the seminal work of [Myerson 1981] shows that the VCG mechanism with monopoly reserves is revenue-maximizing. Recent works show that this idea is more general. [Hartline and Roughgarden 2009] show that in a variety of IPV settings monopoly reserves can be used to construct “approximately” optimal mechanisms. This paper considers the design of approximately optimal mechanisms for an important generalization of this setting. We relax the assumption that agents have independent private values, and consider the case of interdependent values. The case of interdependent values is pertinent for many practical applications, and has received much attention in the literature in economics on auctions, starting with [Milgrom and Weber 1982]. In an interdependent values setting, one agent’s valuation for winning can depend on other agents’ private information, and the private information of agents may be correlated. In this paper, we consider a setting where the agent’s private information can be summarized by a one-dimensional signal. As a motivating example, suppose the good for sale may be resold, and buyers may have different information about future states of the world, e.g. market conditions. Then the information possessed by other buyers, if known to a particular buyer, may affect his/her valuation of winning. For example a buyer’s value for an art piece depends not just on his/her private consumption value, but also on his/her beliefs about the art’s resale value, which depends on other buyers’ values. Furthermore, in many applications buyers may have different but correlated private information. A classical example is that of a tract of oil for sale, where each potential buyer surveys the tract and estimates the A full version of this paper is available at: https://economics.sas.upenn.edu/sites/economics.sas.upenn.edu/files/ec41li.pdf See the Acknowledgements section before REFERENCES. Author’s addresses: Yunan Li, Economics Department, University of Pennsylvania, Philadelphia, PA, USA, yunanli@sas.upenn.edu. Copyright is held by the author/owner(s). EC’13, June 16–20, 2013, Philadelphia, PA, USA. ACM 978-1-4503-1962-1/13/06. extractable oil — buyers’ estimates are therefore statistically correlated. Additionally a buyer’s private information will impact others’ valuations — for example if buyer 1 finds that buyer 2 has a lower estimate on the amount of extractable oil, he/she may revise his/her own valuation downwards. Several seminal applied papers on auctions use an interdependent values setting (see, e.g. [Hendricks and Porter 1988] and [Hendricks et al. 2003]). Instead of considering that there is a single item for sale, we consider the environment where there is a system of feasible sets specifying which subset of buyers can win simultaneously. For example, in the k-unit good auction with unit-demand buyers, the seller can sell to at most k buyers. The feasible sets are precisely those subsets which contain no more than k buyers. Another example is a combinatorial auction with single-minded buyers, where a feasible set corresponds to a subset of buyers seeking mutually disjoint bundles. In such environments, identifying the optimal mechanism remains an open question. This paper investigates a simpler question: can we find a simple mechanism that is ex-post incentive compatible and performs “reasonably” well? Specifically, we study the performance of the VCG mechanism with monopoly reserves, which is the optimal mechanism in the IPV, single unit environment. As is standard in the literature studying ex-post incentive compatibility for interdependent value problems, we assume that agents’ valuations satisfy a single-crossing condition. Under this condition, the VCG mechanism is ex-post incentive compatible in singleitem auctions (see, e.g. [Ausubel 2000]). Unfortunately, this need not be true once we leave the single-item auction setting. One contribution of this paper is to show that when the system of feasible sets is a matroid (described below), the VCG mechanism is still ex-post incentive compatible. We exhibit a novel example where the VCG mechanism is not ex-post incentive compatible when the system of feasible sets is not a matroid. The matroid feasibility constraint covers many interesting cases. Examples include single-item auctions, the allocation of homogeneous goods ([Ausubel 2004]) and digital good auctions ([Goldberg et al. 2001]), among others. If in addition the agents’ valuations and the distribution of their private information satisfy a generalized monotone hazard rate condition, we prove that the seller’s expected revenue by employing the VCG mechanism with monopoly reserves is at least 1/e of the full surplus, where e is the base of the natural logarithm. The proof uses the fact that distributions meeting the monotone hazard rate condition have tails no heavier than that of an exponential distribution (which has a constant hazard rate). The identified bound is tight for arbitrary number of buyers. It is attained in a single-item auction when there is only one “serious” buyer with an exponentially distributed valuation, and all the other buyers’ valuations of the item are negligible. This paper is broadly related to the large literature on revenue-maximizing mechanism design (see, e.g. [Krishna 2009]). The works most closely related to the current paper are those on approximation guarantees for simple mechanisms. The literature can be classified into two branches. One branch studies the general one-dimensional type environments where agents have independent private values. [Chawla et al. 2007], [Hartline and Roughgarden 2009] and [Dhangwatnotai et al. 2010] examine the extent to which simple mechanisms can achieve good approximations of the optimal one. This paper is closely related to [Hartline and Roughgarden 2009]. In their paper, they also study the general onedimensional type environments, but focus on the case where agents have independent private values. They consider two classes of environments, one in which the system of feasible sets is downward-closed and the valuation distributions satisfy the increasing hazard rate condition, and one in which the system of feasible sets is a matroid and the valuation distributions are regular. This paper only considers matroid environments since the ex-post incentive compatibility of the VCG mechanism fails in the more general downward-closed environments even if we assume the agents’ valuations satisfy the single-crossing condition. They prove that the seller’s expected revenue by employing the VCG mechanism with buyer specific reserves is at least 1/2 of the optimal revenue in both settings. See [Hartline 2012] for a thorough survey on this. The other branch focuses on the single-item auctions where agents have correlated private values. [Ronen 2001] proposes a mechanism that can achieve at least half of the optimal revenue. [Ronen and Saberi 2002] prove that no deterministic polynomial time ascending auction can achieve an approximation ratio better than 3/4. The paper most related to the current paper is [Neeman 2003], which compares the expected revenue generated by the English auction (with or without reserves) with the full surplus. He considers the environments where there are n agents and each agent’s expected value is as least α of his/her maximum possible value, and quantifies the fraction of surplus extracted as a function of both n and α. Contemporaneously with the present paper, [Roughgarden and Talgam-Cohen forthcoming] study single sample mechanisms in interdependent value settings, with positive results but slightly different assumptions. The reader is encouraged to see their paper for an alternative approach to approximately optimal mechanism design in these settings. The focus of the current paper is on the design of approximate revenue-maximizing mechanisms, but a number of papers have analyzed mechanisms that achieves social efficiency in the interdependent value cases. [Ausubel 2000] gives an elegant extension of the Vickrey auction that achieves efficiency when there are n buyers and k identical goods for sale. He does not restrict attention to unit-demand buyers, but assumes buyers’ marginal value from an additional unit of good is decreasing. The VCG mechanisms considered in this paper are a generalization of his to accommodate the more general feasibility constraint faced by the seller. [Dasgupta and Maskin 2000] constructs a detail-free mechanism achieves efficiency even if the goods for sale are heterogenous. In the homogeneous-goods downward sloping demand environment, [Perry and Reny 2002] proposes another simpler modification of Vickrey’s auction which consists of a collection of second-price single-unit auctions carried out over two rounds and can achieve an socially efficient outcome. [Bikhchandani et al. 2011] study ascending auctions that implement the Vickrey outcome when the seller is constrained to sell bases of a matroid. The rest of the paper is organized as follows. We present the model in Section 2. In Section 3, we show that the VCG mechanism is ex-post incentive compatible, and approximately optimal when supplemented with monopoly reserves. We conclude in Section 4 with a discussion of future work. 2. THE MODEL There are a finite set N of buyers, where n = |N |, and a collection I ⊆ 2N of feasible sets of buyers, which are the subsets of buyers that can simultaneously win. For example, in a k−unit good auction with unit-demand buyers, a feasible set is any subset of N containing at most k buyers. We assume that (N , I) is a matroid, i.e.: Definition 2.1 ([Oxley 2006]). We say (N , I) is a matroid if (1) ∅ ∈ I; (2) For all S 0 ⊆ S ⊆ N , if S ∈ I then S 0 ∈ I; (3) For all S, S 0 ∈ I, if |S| > |S 0 |, then there exists i ∈ S\S 0 such that S 0 ∪ {i} ∈ I. A subset S of N is said to be independent if S ∈ I, and dependent if it is not independent. A set B ∈ I is called a base of the matroid if it becomes dependent upon adding any element of N . Examples of matroid environments include digital good auctions where I = 2N and k-unit good auctions with unit-demand buyers where I is the collection of all subsets of N which contains at most k buyers. The seller’s reservation value is 0. Each buyer i ∈ N receives a private signal (buyer i’s type) ti ∈ Ti ≡ [ti , t̄i ] ⊆ R+ . Let t ≡ (t1 , · · · , tn ) and T ≡ Πni=1 Ti . Let v : T → Rn+ , where vi (t) ∈ R+ denotes buyer i’s valuation of winning. We assume the payoff to buyer i when he/she does not win (and does not pay) is 0. We assume that vi (·) is non-negative, differentiable and increasing in all its elements, and strictly increasing in ti . Furthermore, we assume that v(·) satisfies the following single-crossing condition: Assumption 2.2. For all i, j (i 6= j), t−i and t0i > ti , we have vi (ti , t−i ) ≥ vj (ti , t−i ) =⇒ vi (t0i , t−i ) > vj (t0i , t−i ), and vi (t0i , t−i ) ≤ vj (t0i , t−i ) =⇒ vi (ti , t−i ) < vj (ti , t−i ). This single-crossing condition is typically assumed in most work that study ex-post incentive compatibility for interdependent value problems (e.g. [Crmer and McLean 1985] and [Ausubel 2000]). However, in the standard interpretation, vi is a reduced form utility function that defines agent i’s expected valuation of winning under some circumstances given the agents’ signals. Agent i’s valuation depends on others’ signals only to the extent that the signals provide information about the state of nature. [McLean and Postlewaite 2006] show that the conditions on the primitive utility functions that would ensure that the reduced form utility functions satisfy the singlecrossing condition are stringent. Let F : T → [0, 1] denote the cumulative distribution of t, with a joint density f such that f (t) > 0 for all t ∈ T . Let Fi (·|t−i ) denote the corresponding conditional cumulative distribution of ti given t−i , with a conditional density fi (·|t−i ). We assume that v(·) and F (·) satisfy the following generalized monotone hazard rate condition: Assumption 2.3. 1 − Fi (ti |t−i ) ∂vi (t) is decreasing in ti for all i and t−i . fi (ti |t−i ) ∂ti It is easy to see that the IPV settings where each buyer’s value is drawn from an increasing hazard rate distribution satisfy Assumptions 2.2 and 2.3. First, when buyers have private values, Assumption 2.2 is trivially satisfied since a buyer’s value only depends on his/her own type. Second, usually in the private values setting, we normalize vi (t) = ti . Then ∂vi (t)/∂ti = 1 and Assumption 2.3 is satisfied if ti follows a distribution with an increasing hazard rate. There are also other interesting cases that satisfy Assumptions 2.2 and 2.3. Note that one easy way to make sure Assumption 2.3 to be satisfied is to make sure that for all i and t−i : ∂vi (·, t−i ) is decreasing; and ∂ti (2) Fi (·|t−i ) has an increasing hazard rate. (1) We first give examples of v(·) that satisfy Assumption 2.2 and have strictly positive decreasing partial derivatives, ∂vi (t)/∂ti , and second we give examples of distributions F whose conditionals have increasing hazard rates. Example 2.4. P For all i ∈ N , buyer i’s valuation of winning takes the following form: vi (t) = ti + β j6=i tj with 0 < β < 1. Note that for any (v, F ), we can define (ṽ, F̃ ) such that ṽi (g(t)) = vi (t) and F̃ (g(t)) = F (t), where gi : Ti → Si ⊂ R+ is strictly increasing and g(t) = (g1 (t1 ), · · · , gn (tn )). Then (v, F ) and (ṽ, F̃ ) represent the same environment. Clearly, vi is non-negative, increasing and strictly increasing in ti if and only if ṽi is non-negative, increasing and strictly increasing in si = gi (ti ). Suppose further that gi is differentiable, then ∂ṽi (s) 0 ∂vi (t) g (tj ) = . ∂sj j ∂tj This implies that ṽ(·) satisfies the single-crossing condition if and only if v(·) does. Finally, suppose both v and g are twice differentiable, then we have −1 2 −1 2 ∂vi (t) ∂vi (t) ∂ ṽi (s) ∂ vi (t) gi0 0 (t) gi0 (ti )2 = − 0 . (1) 2 ∂ti ∂si ∂ti ∂t2i gi (t) h i0 Suppose for all i and ti , supt−i ln ∂v∂ti (t) < +∞, then we can define gi such that i 0 ∂vi (t) 0 [ln gi (ti )] ≥ sup ln . ∂ti t−i Then it follows from (1) that ∂ 2 ṽi (s)/∂s2i ≤ 0. Hence without much loss of generality we can assume that ∂vi (t)/∂ti is decreasing. We now turn to the condition that for all i and t−i the conditional distribution Fi (·|t−i ) has an increasing hazard rate. There is a literature in mathematics about conditionally specified distributions, i.e. whether there exists a unique joint distribution compatible with given families of conditional distributions. [Arnold et al. 2001] provides a thorough survey on this topic. In the special case where the conditional distributions belong to exponential families, which includes many of the most common distributions including Poisson, binomial, exponential, gamma, multivariate normal, Weibull and many others, it is known that the joint distribution exists and must be of some specific form. In the literature about auctions with interdependent values, it is common to assume that t are affiliated and therefore positively correlated. For the bivariate case, [Arnold 1990] gives sufficient conditions on conditionals for the negative or positive dependence of the two variables, and his results can be extended to multivariate cases. Here are some examples of distributions F whose conditionals Fi (·|t−i ) have increasing hazard rates. Example 2.5. Suppose t follows a multivariate normal distribution, then the conditional distribution of ti given t−i is normal, which has an increasing hazard rate. Example 2.6. Suppose t is two-dimensional and has a joint density positive on the positive orthant and zero elsewhere. [Arnold and Strauss 1988] prove that the conditional distributions f1 (t1 |t2 ) and f1 (t2 |t1 ) are both exponential if and only if the joint distribution is proportional to exp(−λt1 − µt2 − νt1 t2 ) for some λ, µ > 0 and ν ≥ 0. It is well know that the exponential distribution has a constant hazard rate. Note that under this distribution t1 and t2 are either independent or negatively correlated. Without loss of generality we focus on direct mechanisms, in which each buyer reports his/her own signal directly. A mechanism M is a pair (q, p), where the allocation rule q maps every vector of reported signals t̃ to a probability mass function, q(t̃) ∈ ∆|I|−1 , and the payment rule p maps every t̃ to a payment vector (in Rn+ ). Let P xi (t̃) ≡ {S∈I:i∈S} q(t̃)(S) denote the probability with which buyer i is one of the winners. If the seller had complete information about the buyers’ valuations of winning and full bargaining power, the maximum total surplus he could extract from the buyers is: " # X E max vi (t) . (2) S∈I i∈S We are interested in obtaining a lower bound on the ratio between the expected revenue the seller can obtain by employing the VCG mechanism with monopoly reserves (described in detail later) and the full surplus given by (2). Obviously, the full surplus provides an upper bound for the expected revenue the seller can obtain by employing the optimal mechanism. It is attainable in some environments. For example, when the buyers’ information are correlated, [Crmer and McLean 1988] and [McAfee and Reny 1992] provide conditions under which the optimal mechanism succeeds in extracting the full surplus. Ideally, we want to compare the VCG mechanism with monopoly reserves with the optimal mechanism. However, for the general environment considered here, a characterization of the optimal mechanism is unavailable. To the extent that the optimal mechanism falls short of extracting the full surplus, our measurement provides a lower bound on the performance of the mechanism relative to the optimal one. 3. VCG MECHANISMS This section studies the incentive compatibility and approximate optimality of the VCG mechanisms. In Section 3.1 we define the mechanisms and show that they are ex-post incentive compatible if v(·) satisfies the single-crossing condition and the system of feasible sets is a matroid. We show in Section 3.2 that if in addition the generalized monotone hazard rate condition is satisfied, the VCG mechanism with monopoly reserves is approximately optimal. 3.1. Ex-Post Incentive Compatibility Definition 3.1. We say a mechanism M = (q, p) is ex-post incentive compatible if and only if for all i, t−i and ti vi (t)xi (t) − pi (t) ≥ vi (t)xi (t̃i , t−i ) − pi (t̃i , t−i ), ∀t̃i ∈ Ti . (3) We say a mechanism is Bayesian incentive compatible if E[vi (t)xi (t) − pi (t)|ti ] ≥ E[vi (t)xi (t̃i , t−i ) − pi (t̃i , t−i )|ti ] for all i, ti and t̃i . Note that if a mechanism is ex-post incentive compatible then it is Bayesian incentive compatible. The Bayesian incentive compatible mechanism has been very well characterized in the IPV settings by [Myerson 1981], and the same method can be easily extended to characterize ex-post incentive compatible mechanism in the IPV settings (see, e.g. [Archer and Tardos 2001]). The following lemma extended their results to the case where agents have interdependent values and correlated types: L EMMA 3.2. A mechanism M = (q, p) is ex-post incentive compatible if and only if xi (·, t−i ) is increasing for all i and t−i , and the payment function satisfies, for each t and each buyer i, Z ti ∂vi (s, ti ) xi (s, t−i )ds − ui (ti , t−i ), (4) pi (t) = vi (t)xi (t) − ∂s ti where ui (t) ≡ vi (t)xi (t) − pi (t). P ROOF. This proof follows [Myerson 1981]. Suppose M = (q, p) is ex-post incentive compatible, then, for all i, t−i and ti > t0i , we have ui (ti , t−i ) ≥ ui (t0i , t−i ) + [vi (ti , t−i ) − vi (t0i , t−i )]xi (t0i , t−i ), and ui (t0i , t−i ) ≥ ui (ti , t−i ) + [vi (t0i , t−i ) − vi (ti , t−i )]xi (ti , t−i ). Combine the above two inequalities and we get [vi (t) − vi (t0i , t−i )]xi (t) ≥ ui (t) − ui (t0i , t−i ) ≥ [vi (t) − vi (t0i , t−i )]xi (t0i , t−i ). (5) vi (t) − vi (t0i , t−i ) Since vi (·, t−i ) is strictly increasing by assumption, we have > 0. Then (5) implies that xi (ti , t−i ) ≥ xi (t0i , t−i ). That is, xi (·, t−i ) is increasing. Dividing (5) by ti −t0i and taking t0i → ti proves that for any t at which xi (·, t−i ) is continuous, ∂ui (t)/∂ti exists and ∂ui (t) ∂vi (t) = xi (t). ∂ti ∂ti Finally, ui (·, t−i ) is absolutely continuous, since vi is differentiable and |ui (t) − ui (t0i , t−i )| ≤ |vi (t) − vi (t0i , t−i )| max{xi (t0i , t−i ), xi (t)} ≤ |vi (t) − vi (t0i , t−i )|. Thus, Z ti ui (t) = ui (ti , t−i ) + ti ∂vi (s, ti ) xi (s, t−i )ds. ∂s Hence (4) holds by the definition of ui . It is straightforward to verify that when xi (·, t−i ) is increasing for all i and t−i and the payment function satisfies (4), the mechanism M = (q, p) is ex-post incentive compatible. A direct corollary of Lemma 3.2 is that for any ex-post incentive compatible mechanism, the expected payment of buyer i is E[pi (t)] = −E[ui (ti , t−i )] + E [xi (t)ϕi (t)] , where ϕi (t) ≡ vi (t) − 1 − Fi (ti |t−i ) ∂vi (t) . fi (ti |t−i ) ∂ti (6) Here ϕi is buyer i’s “virtual value” in the interdependent value case. Hence the seller gets the same expected revenues from any two ex-post incentive compatible mechanisms which have the properties that (1) given t, a buyer is one of the winners with the same probabilities, and (2) a buyer with the lowest signal gets the same expected utilities. Consider now the allocation rule of the VCG mechanism with no reserves. Given a vector of reportedP signals t̃, the mechanism chooses a feasible set S ∈ I that maximizes the total surplus i∈S vi (t̃). Let x∗i denote buyer i’s probability of winning implied by the mechanism, then x∗i (t) ∈ {0, 1} for all t since the mechanism is deterministic. In general, x∗i (·, t−i ) need not be increasing, and therefore the mechanism need not be ex-post incentive compatible: Example 3.3. Buyer 1 is “ big” and his/her valuation is v1 (t) = 5+t1 . Buyers 2, · · · , 11 are “small” and their valuations are vi (t) = ti + 0.5t1 . Then v(·) satisfies the singlecrossing condition. The feasible sets are precisely those that do not contain both the big buyer and a small buyer. Let ti = 0 for i ≥ 2. Suppose t1 = 1, then v1 (t) = 6 and P VCG mechanism will allocate to the big buyer. Suppose t1 = 2, i≥2 vi (t) = 5, and the P then v1 (t) = 7 and i≥2 vi (t) = 10, and the VCG mechanism will allocate to the set of small buyers. The probability with which the big buyer wins falls as his/her signal increases. x∗i (·, t−i ) is increasing when the system of feasible sets is a matroid: L EMMA 3.4. Suppose (N , I) is a matroid, and v(·) satisfies Assumption 2.2. Let x∗i denote the probability with which buyer i is one of the winners under the VCG mechanism with no reserves. Then for all i and t−i , x∗i (·, t−i ) is increasing. Furthermore, let t̂i (t−i ) ≡ inf{t0i |x∗i (t0i , t−i ) = 1}. (7) Then x∗i (t) = 1 if ti > t̂i (t−i ) and x∗i (t) = 0 if ti < t̂i (t−i ). The proof of Lemma 3.4 relies on the following well-known property of matroids: L EMMA 3.5 ([O XLEY 2006]). Let (N , I) be a matroid, and B denote the collection of its bases. For every A, B ∈ B and a ∈ A\B, there exists b ∈ B such that A\{a} ∪ {b} ∈ B and B\{b} ∪ {a} ∈ B. P ROOF OF L EMMA 3.4. Fix i and t−i . Suppose that there exist t0i > ti such that = 1 and x∗i (t0i , t−i ) = 0. Let W ≡ {j|x∗j (ti , t−i ) = 1} and W 0 ≡ {j|x∗j (t0i , t−i ) = 1} denote the set of winners under (ti , t−i ) and (t0i , t−i ), respectively. Then i ∈ W , but i ∈ / W 0 . Clearly, W and W 0 are bases of the matroid. By Lemma 3.5, there exits j ∈ W 0 \W 0 such that S1 ≡ W ∪ {j}\{i} ∈ I and P S2 ≡ W ∪ {i}\{j} P ∈ I. Since W is surplus maximizing given (ti , t−i ), we have i∈W vi (ti , t−i ) ≥ i∈S1 vi (ti , t−i ), which implies that x∗i (ti , t−i ) vi (ti , t−i ) ≥ vj (ti , t−i ). 0 (t0i , t−i ), (8) Similarly, W is Psurplus maximizing given P 0 0 v (t , 0 i i t−i ) ≥ i∈W i∈S2 vi (ti , t−i ), which implies that and therefore we have vj (t0i , t−i ) ≥ vi (t0i , t−i ). (9) However, by the single-crossing condition, (8) implies that vi (t0i , t−i ) > vj (t0i , t−i ), which contradicts with (9). Hence x∗i (·, t−i ) is increasing. It then follows directly that x∗i (t) = 1 if ti > t̂i (t−i ) and x∗i (t) = 0 if ti < t̂i (t−i ). Given Lemma 3.4, we can define the payment rule of the VCG mechanism with no reserves as following: a buyer pays if and only if he/she is one of the winners, and in case of winning he/she pays vi (t̂i (t−i ), t−i ). It is easy to verify that this is consistent with (4) and therefore by Lemma 3.2 the VCG mechanism with no reserves is ex-post incentive compatible. This is also intuitive. On the one hand, a buyer has no incentive to underreport since his/her winning probability is increasing in his/her own type, but his/her payment upon winning is independent of his/her own type. On the other hand, a buyer has no incentive to overreport since otherwise he/she might win, but pay an amount that is higher than his/her value. Note also that the VCG mechanism with no reserves is ex-post efficient and ex-post individually rational. In the rest of the paper, we focus on the following class of VCG mechanisms with reserves (VCGr): Definition 3.6. The VCG mechanism with reserves (VCGr), r̃i (·), is a direct mechanism in which buyers simultaneously report their signals, and given the vector of reported signals t̃, the set of winners and their payments are decided in the following three steps: P (1) Choose a set W ∗ ∈ arg maxS∈I i∈S vi (t̃), and let x∗i denote the probability with ∗ which buyer i is in W . (2) Delete all the buyers i in W ∗ with t̃i such that t̃i < r̃i (t̃−i ), where r̃i : T−i → Ti . Then the set of winners is W = {i ∈ W ∗ |t̃i ≥ r̃i (t̃−i )}. (3) Charge each winner the larger of vi (r̃i (t̃−i ), t̃−i ) and vi (t̂i (t̃−i ), t̃−i ), where t̂i (·) is given by (7) using x∗i in step 1. The VCG mechanism with no reserves is a special VCGr mechanism with r̃i (t−i ) = ti for all i and all t−i . The VCGr mechanisms are ex-post incentive compatible under the matroid feasibility constraint. To see this let x denote the allocation rule of a VCGr, then xi (t) = 1 if and only if x∗i (t) = 1 and ti ≥ r̃i (t−i ). Suppose xi (t) = 1 and consider t0i > ti . First, x∗i (t0i , t−i ) = 1 by Lemma 3.4. Second, t0i > ti ≥ r̃i (t−i ). Hence xi (t0i , t−i ) = 1. That is, xi (·, t−i ) is increasing. Finally, buyer i’s payment function in the VCGr satisfies: 0 if ti < max{t̂i (t−i ), r̃i (t−i )}, pi (t) = vi (max{t̂i (t−i ), r̃i (t−i )}, t−i ) if ti > max{t̂i (t−i ), r̃i (t−i )}. One can easily verify that the payment scheme satisfies (4). Hence by Lemma 3.2 the VCGr mechanisms are ex-post incentive compatible. 3.2. Approximate Optimality The monopoly reserves, ri (·), are defined such that ϕi (ri (t−i ), t−i ) = 0. They are well defined since under Assumption 2.3 buyer i’s virtual value, ϕi (t), is strictly increasing in ti . The VCG mechanism with monopoly reserves is the VCGr mechanism in which the reserves are given by the monopoly reserves: r̃i (·) = ri (·). It generates the highest expected revenue among all the VCGr mechanisms. To see this, consider the seller’s expected revenue from a VCGr mechanism: X X X E[pi (t)] = − E[ui (ti , t−i )] + E x∗i (t)1{ti ≥r̃i (t−i )} (t)ϕi (t) , i i i where 1{ti ≥r̃i (t−i )} (·) is an indicator function which equals 1 if and only if ti ≥ r̃i (t−i ). The expected revenue is maximized when a buyer in W ∗ wins if and only if his/her virtual value is positive, i.e. when r̃i (·) = ri (·). The main result of the paper is the following theorem: T HEOREM 3.7. For the class of v(·) and F (·) satisfying Assumption 2.3, the seller’s expected revenue by employing the VCG mechanism with monopoly reserves is at least 1/e of the full surplus. The proof of Theorem 3.7 relies on the following lemma: L EMMA 3.8 (L EMMA 3.9 IN [D HANGWATNOTAI ET AL . 2010]). Suppose a nonnegative random variable v follows a cumulative distribution F with a density function f (v) ∗ f , and 1−F (v) is increasing. Then for all v̂ ≥ 0 and s = max{v̂, v } Z 1 s[1 − F (s)] ≥ vdF (v), e v≥v̂ where v ∗ is defined by v ∗ − [1 − F (v ∗ )]/f (v ∗ ) = 0. P ROOF OF T HEOREM 3.7. Let x∗i denote buyer i’s winning probability under the VCG mechanism with no reserves, then the full surplus is given by X X X E[ x∗i (t)vi (t)] = E[x∗i (t)vi (t)] = E[E[x∗i (t)vi (t)|t−i ]]. (10) i i i Similarly, the seller’s expected revenue by employing the VCG mechanism with monopoly reserves is given by X X X E[ pi (t)] = E[pi (t)] = E[E[pi (t)|t−i ]]. (11) i i i We now prove that for all i and t−i , 1 E[x∗i (t)vi (t)|t−i ]. (12) e Once we prove (12), the theorem follows immediately from taking expectations over t−i and summing over all buyers. Fix i and t−i . Let Gi (vi (t)|t−i ) ≡ Fi (ti |t−i ), i.e. Gi (·|t−i ) is the cumulative distribution function of vi conditional on t−i . Let gi (·|t−i ) denote the corresponding density function. Then −1 fi (ti |t−i ) ∂vi (t) gi (vi |t−i ) = , 1 − Gi (vi |t−i ) 1 − Fi (ti |t−i ) ∂ti E[pi (t)|t−i ] ≥ which is increasing by Assumption 2.3 and by the fact that vi (·, t−i ) is increasing. Let v̂i ≡ vi (t̂i (t−i ), t−i ) and vi∗ ≡ vi (ri (t−i ), t−i ), then vi∗ − gi (vi∗ |t−i ) = 0. 1 − Gi (vi∗ |t−i ) Let s = max{v̂i , vi∗ }. Then the LHS of (12) can be rewritten as E[pi (t)|t−i ] = s[1 − Gi (s|t−i )], The RHS of (12) multiplied by e is E[x∗i (t)vi (t)|t−i ] Z t̄i vi (s, t−i )f (s|t−i )ds, = Zt̂i vi dGi (vi |t−i ). = (13) vi ≥v̂i By Lemma 3.8, s[1 − Gi (s|t−i )] ≥ 1 e Z vi dGi (vi |t−i ). vi ≥v̂i (12) follows immediately. Considering the single-item auction where there is a single buyer with an exponentially distributed valuation shows that the bound in Theorem 3.7 is tight. Without further restrictions on F , this bound is also tight for an arbitrary number of buyers. Consider the single-item auction with n buyers where one buyer’s valuation is exponentially distributed and the other n − 1 buyers have valuation 0. This is essentially an auction with a single buyer. This result is similar to [Neeman 2003]. He finds that in the general private value environment, the worst-case ratio of the expected revenue from an English auction to full surplus stays bounded away from 1 as the number of buyers tends to infinity even when the seller sets reserve prices optimally. This is because in the worst case there is only one “serious” buyer, it is not surprising that increasing in the number of buyers has a negligible effect on the revenue. This example also shows that reserve prices are important for revenue guarantee, since in this case the expected revenue from the VCG mechanism with no reserves is 0. 4. CONCLUSION In the general one-dimensional type setting where agents have interdependent values and correlated types, this paper shows that (1) when the system of feasible sets is a matroid and the agents’ valuations satisfy the single-crossing condition, the VCG mechanisms are ex-post incentive compatible; (2) if in addition the valuation distribution satisfies the generalized monotone hazard rate conditions, the VCG mechanism with monopoly reserves is approximately optimal, i.e. the expected revenue from the VCG mechanism with monopoly reserves is at least 1/e of the full surplus. Looking toward future work, one limitation of the present paper is that the main result heavily depends on the increasing hazard rate assumption. In the IPV settings [Hartline and Roughgarden 2009] proves approximation result under matroid feasibility constraints and regular valuation distributions. Their analysis is facilitated by the fact that in the IPV settings the optimal revenue is given by the expected total virtual surplus. However, this is no longer true when we allow buyers’ types to be correlated (see, e.g. [Crmer and McLean 1988]). Therefore in this paper we compare the expected revenue of the VCG mechanism with monopoly reserves with the full surplus. It is not clear in this case how to connect the full surplus with virtual surplus under the weaker regularity assumption. Another disadvantage of the VCG mechanism with monopoly reserves studied in this paper is that it is not anonymous. [Hartline and Roughgarden 2009] show that there is an anonymous random reserve price such that the expected revenue of VCG with that reserve is at least 1/4 of the expected revenue of the optimal auction. Their result is based on [Bulow and Klemperer 1996] type result. [Bulow and Klemperer 1996] shows that in a single-item auction where buyers have interdependent values but independent types, the VCG mechanism with no reserves is more profitable in expectation than the optimal mechanism with one less buyer. [Hartline and Roughgarden 2009] proves an approximate analog of this statement in the general one-dimensional type environments where agents have independent private values. Specifically, they consider the following duplication of the original environment: Each buyer i is replaced by a pair i, n + i of buyers. For ease of notation let i0 = n + i. Let N d ≡ {1, 2, · · · , 2n} and I d denote the collection of the feasible sets of the duplicated environment. Then for each S d ⊂ N d , S d ∈ I d if and only if (1) for each i ∈ N , at most one buyer from the pair {i, i0 } is selected; (2) S d , when naturally interpreted as a set of buyers from the original environment, is a feasible set in that environment. It is easy to verify that (N d , I d ) is a matroid if (N , I) is a matroid. They show that ratio of the expected revenue from the VCG mechanism with no reserves in the duplicated environment to the optimal revenue in the original environment is at least 1/3 when the system of feasible sets is downward-closed and the the valuation distributions have increasing hazard rates, and 1/2 when the system of feasible sets is a matroid and the the valuation distributions are regular. A direct extension of their results to the case where agents have interdependent values and correlated types is difficult for several reasons. First, how to define a buyer’s valuation of winning properly when the total number of buyers changes? Let t and t0 denote the types of the original buyers and their duplicates, respectively. One way to deal with this question is to assume v̄i (t) = E[vi (t, t0 )|t] as buyer i’s valuation when there are only n buyers (see [Bulow and Klemperer 1996]). The other is to assume that a buyer’s value only depends on his/her own type and the state of the world, and the other buyers’ types affect his/her expected valuation only to the extent that they are correlated with the state of the world (see, e.g. [McLean and Postlewaite 2002] and [McLean and Postlewaite 2004]). Second, consider a buyer i and his/her duplicate i0 , we need to make assumptions so that not only their valuations depend on the other buyers’ types in a similar way, but also their types affect the other buyers’ valuations in a similar way. Finally, the correlation between i and i0 ’s types and how their valuations depend on each other’s type also matters for the approximation result. ACKNOWLEDGMENTS I am grateful to Steven Matthews, Mallesh Pai, Andrew Postlewaite and Tim Roughgarden for extremely helpful discussions. All remaining errors are my sole responsibility. REFERENCES A RCHER , A. AND T ARDOS, É. 2001. Truthful mechanisms for one-parameter agents. In Foundations of Computer Science, 2001. Proceedings. 42nd IEEE Symposium on. IEEE, 482–491. A RNOLD, B. C. 1990. Dependence in conditionally specified distributions. Lecture Notes-Monograph Series 16, pp. 13–18. A RNOLD, B. C., C ASTILLO, E., AND S ARABIA , J. M. 2001. Conditionally specified distributions: An introduction. Statistical Science 16, 3, pp. 249–265. A RNOLD, B. C. AND S TRAUSS, D. 1988. Bivariate distributions with exponential conditionals. Journal of the American Statistical Association 83, 402, pp. 522–527. AUSUBEL , L. 2000. A generalized vickrey auction. In Econometric Society World Congress 2000 Contributed Papers. Econometric Society. AUSUBEL , L. M. 2004. An efficient ascending-bid auction for multiple objects. The American Economic Review 94, 5, pp. 1452–1475. B IKHCHANDANI , S., D E V RIES, S., S CHUMMER , J., AND V OHRA , R. 2011. An ascending vickrey auction for selling bases of a matroid. Operations research 59, 2, 400–413. B ULOW, J. AND K LEMPERER , P. 1996. Auctions versus negotiations. The American Economic Review 86, 1, pp. 180–194. C HAWLA , S., H ARTLINE , J., AND K LEINBERG, R. 2007. Algorithmic pricing via virtual valuations. In Proceedings of the 8th ACM conference on Electronic commerce. ACM, 243–251. C HAWLA , S., H ARTLINE , J., M ALEC, D., AND S IVAN, B. 2010. Multi-parameter mechanism design and sequential posted pricing. In Proceedings of the 42nd ACM symposium on Theory of computing. ACM, 311–320. C RMER , J. AND M C L EAN, R. P. 1985. Optimal selling strategies under uncertainty for a discriminating monopolist when demands are interdependent. Econometrica 53, 2, pp. 345–361. C RMER , J. AND M C L EAN, R. P. 1988. Full extraction of the surplus in bayesian and dominant strategy auctions. Econometrica 56, 6, pp. 1247–1257. D ASGUPTA , P. AND M ASKIN, E. 2000. Efficient auctions. The Quarterly Journal of Economics 115, 2, 341–388. D HANGWATNOTAI , P., R OUGHGARDEN, T., AND YAN, Q. 2010. Revenue maximization with a single sample. In Proceedings of the 11th ACM conference on Electronic commerce. ACM, 129–138. D OBZINSKI , S., F U, H., AND K LEINBERG, R. 2011. Optimal auctions with correlated bidders are easy. In Proceedings of the 43rd annual ACM symposium on Theory of computing. ACM, 129–138. G OLDBERG, A., H ARTLINE , J., AND W RIGHT, A. 2001. Competitive auctions and digital goods. In Proceedings of the twelfth annual ACM-SIAM symposium on Discrete algorithms. Society for Industrial and Applied Mathematics, 735–744. H ARTLINE , J. AND R OUGHGARDEN, T. 2008. Optimal mechanism design and money burning. In Proceedings of the 40th annual ACM symposium on Theory of computing. ACM, 75–84. H ARTLINE , J. D. 2012. Approximation in mechanism design. The American Economic Review 102, 3, 330–336. H ARTLINE , J. D. AND R OUGHGARDEN, T. 2009. Simple versus optimal mechanisms. In Proceedings of the 10th ACM conference on Electronic commerce. ACM, 225–234. H ENDRICKS, K., P INKSE , J., AND P ORTER , R. H. 2003. Empirical implications of equilibrium bidding in first-price, symmetric, common value auctions. The Review of Economic Studies 70, 1, pp. 115–145. H ENDRICKS, K. AND P ORTER , R. H. 1988. An empirical study of an auction with asymmetric information. The American Economic Review 78, 5, pp. 865–883. K RISHNA , V. 2009. Auction theory. Academic press. M C A FEE , R. P., M C M ILLAN, J., AND R ENY, P. J. 1989. Extracting the surplus in the common-value auction. Econometrica 57, 6, pp. 1451–1459. M C A FEE , R. P. AND R ENY, P. J. 1992. Correlated information and mecanism design. Econometrica 60, 2, pp. 395–421. M C L EAN, R. AND P OSTLEWAITE , A. 2002. Informational size and incentive compatibility. Econometrica 70, 6, pp. 2421–2453. M C L EAN, R. AND P OSTLEWAITE , A. 2004. Informational size and efficient auctions. The Review of Economic Studies 71, 3, pp. 809–827. M C L EAN, R. AND P OSTLEWAITE , A. 2006. Implementation with interdependent valuations. M ILGROM , P. R. AND W EBER , R. J. 1982. A theory of auctions and competitive bidding. Econometrica 50, 5, pp. 1089–1122. M YERSON, R. B. 1981. Optimal auction design. Mathematics of Operations Research 6, 1, pp. 58–73. N EEMAN, Z. 2003. The effectiveness of english auctions. Games and Economic Behavior 43, 2, 214 – 238. O XLEY, J. 2006. Matroid theory. Vol. 3. Oxford University Press, USA. PAPADIMITRIOU, C. AND P IERRAKOS, G. 2011. On optimal single-item auctions. In Proceedings of the 43rd annual ACM symposium on Theory of computing. ACM, 119– 128. P ERRY, M. AND R ENY, P. J. 2002. An efficient auction. Econometrica 70, 3, pp. 1199– 1212. R ONEN, A. 2001. On approximating optimal auctions. In Proceedings of the 3rd ACM conference on Electronic Commerce. ACM, 11–17. R ONEN, A. AND S ABERI , A. 2002. On the hardness of optimal auctions. In Foundations of Computer Science, 2002. Proceedings. The 43rd Annual IEEE Symposium on. IEEE, 396–405. R OUGHGARDEN, T. AND T ALGAM -C OHEN, I. forthcoming. Optimal and near-optimal mechanism design with interdependent values. In Proceedings of the 14th ACM conference on Electronic commerce. ACM. R USTICHINI , A., S ATTERTHWAITE , M. A., AND W ILLIAMS, S. R. 1994. Convergence to efficiency in a simple market with incomplete information. Econometrica 62, 5, pp. 1041–1063. S ATTERTHWAITE , M. AND W ILLIAMS, S. 2004. The optimality of a simple market mechanism. Econometrica 70, 5, 1841–1863.