Stepwise Randomized Combinatorial Auctions Achieve Revenue Monotonicity

Abstract

In combinatorial auctions that use VCG, a seller can sometimes increase revenue by dropping bidders (see e.g. [5]). In our previous work [26], we showed that such failures of "revenue monotonicity" occur under an extremely broad range of deterministic strategyproof combinatorial auction mechanisms, even when bidders have "known single-minded" valuations. In this work we consider the question of whether revenue monotonic, strategyproof mechanisms for such bidders can be found in the broader class of randomized mechanisms. We demonstrate that---surprisingly- such mechanisms do exist, show how they can be constructed, and consider algorithmic techniques for implementing them in polynomial time.More formally, we characterize a class of randomized mechanisms defined for known single-minded bidders that are strategyproof and revenue monotonic, and furthermore satisfy some other desirable properties, namely participation, consumer sovereignty and maximality, representing the mechanism as a solution to a quadratically constrained linear program (QCLP). We prove that the QCLP is always feasible (i.e., for all bidder valuations) and give its solution analytically. Furthermore, we give an algorithm for running such a mechanism in time polynomial in the number of bidders and goods; this is interesting because constructing an instance of such mechanisms from our QCLP formulation in a naive way can require exponential time.