Simplicity in Auctions Revisited: The Primitive Complexity

In this paper we revisit the notion of simplicity in mechanisms. We consider a seller of m heterogeneous items, facing a single buyer with valuation v. We observe that previous attempts to define complexity measures often fail to classify mechanisms that are intuitively considered simple (e.g., the "selling separately"mechanism) as such. We suggest to view a menu as simple if a bundle that maximizes the buyer's profit can be found by conducting a few primitive operations that are considered simple. The primitive complexity of a menu is the number of primitive operations needed to (adaptively) find a profit-maximizing entry in the menu. In this paper, the primitive operation that we study is essentially computing the outcome of the "selling separately"mechanism.Does the primitive complexity capture the simplicity of other auctions that are intuitively simple? We consider bundle-size pricing, a common pricing method in which the price of a bundle depends only on its size. Our main technical contribution is determining the primitive complexity of bundle-size pricing menus in various settings. First, we connect the notion of primitive complexity to the vast literature on query complexity. We then show that for any distribution D over weighted matroid rank valuations, even distributions with arbitrary correlation among their values, there is always a bundle-size pricing menu with low primitive complexity that achieves almost the same revenue as the optimal bundle-size pricing menu. As part of this proof we provide a randomized algorithm that for any weighted matroid rank valuation v and integer k, finds the most valuable set of size k with only a poly-logarithmic number of demand and value queries. We show that this result is essentially tight in several aspects. For example, if the valuation v is submodular, then finding the most valuable set of size k requires exponentially many queries (this solves an open question of Badanidiyuru et al. [EC'12]). We also show that any deterministic algorithm that finds the most valuable set of size k requires [EQUATION] demand and value queries, even for additive valuations.