Non-Revelation Mechanism Design

We consider mechanism design and redesign for markets like Internet advertising where many frequent, small transactions are organized by a principal. Mechanisms for these markets rarely have truthtelling equilibria. We identify a family of winner-pays-bid mechanisms for such markets that exhibit three properties. First, equilibria in these mechanisms are simple. Second, the mechanisms’ parameters are easily reoptimized from the bid data that the mechanism generates. Third, the performance of mechanisms in the family is near the optimal performance possible by any mechanism (not necessarily within the family). Our mechanisms are based on batching across multiple iterations of an auction environment, and our approximation bound is asymptotically optimal, with loss inversely proportional to the cube root of the number of iterations batched. Our analysis methods are of broader interest in mechanism design and, for example, we also use them to give new sample complexity bounds for mechanism design in general single-dimensional agent environments.