Combinatorial Stationary Prophet Inequalities
Neel Patel, David Wajc
TL;DR
This work studies the stationary prophet inequality (SPI) in infinite-horizon markets with combinatorial buyer demands, establishing a fundamental link between SPI performance and offline contention resolution schemes (CRS). By reducing SPI to the best offline CRS balance, the authors prove that the ex-ante competitive ratio is within a constant factor of this equilibrium, specifically achieving an ex-ante $\frac{c}{2}$-competitive algorithm for constraint family $\mathcal{F}$ with CRS balance $c$ (and matching hardness within $\varepsilon$). The framework extends beyond linear valuations to monotone submodular objectives, matroid constraints, and multi-good settings, yielding concrete ratios such as a $0.393$-approximation for multi-good SPI and a $\tfrac{1}{2}(1-1/e)\approx 0.316$-competitive matroid-SPI (with a tighter polytime approximation of $\approx 0.318$). The approach hinges on offline CRS as a black-box tool, leverages PASTA to relate arrivals to stationary distributions, and uses “saved” items and stochastic-dominance to manage correlations, offering a versatile blueprint for broader valuation classes and online-optimality considerations.
Abstract
Numerous recent papers have studied the tension between thickening and clearing a market in (uncertain, online) long-time horizon Markovian settings. In particular, (Aouad and Sarita{ç} EC'20, Collina et al. WINE'20, Kessel et al. EC'22) studied what the latter referred to as the Stationary Prophet Inequality Problem, due to its similarity to the classic finite-time horizon prophet inequality problem. These works all consider unit-demand buyers. Mirroring the long line of work on the classic prophet inequality problem subject to combinatorial constraints, we initiate the study of the stationary prophet inequality problem subject to combinatorially-constrained buyers. Our results can be summarized succinctly as unearthing an algorithmic connection between contention resolution schemes (CRS) and stationary prophet inequalities. While the classic prophet inequality problem has a tight connection to online CRS (Feldman et al. SODA'16, Lee and Singla ESA'18), we show that for the stationary prophet inequality problem, offline CRS play a similarly central role. We show that, up to small constant factors, the best (ex-ante) competitive ratio achievable for the combinatorial prophet inequality equals the best possible balancedness achievable by offline CRS for the same combinatorial constraints.