From Fairness to Efficiency in Repeated Bilateral Trade

Abstract

Recently, Bachoc et al. (2026) proposed an axiomatic approach to fairness for repeated bilateral trade. Their resulting objective is a H\''older mean of seller and buyer utilities with non-positive order $λ$, interpolating between Rawlsian fairness at $λ= -\infty$ and Nash fairness at $λ= 0$. Here we study this family of H\''older means in the complementary setting of orders $λ\in (0,1]$, which constitutes a tradeoff between fairness and efficiency, with the weight of efficiency increasing with $λ$, and where $λ=1$ yields the existing gain from trade criterion, focusing solely on efficiency. We obtain an axiomatic characterization of these H\''older means, relying on a Pigou-Dalton principle of transferring from rich to poor. From an online learning perspective, we show that, contrary to the case $λ\le 0$, the problem is unlearnable in full generality, even under seller's and buyer's valuation independence. We focus on seller's and buyer's valuations with distributions having bounded densities, in which case we obtain matching lower and upper bounds on the cumulative regret, of order $\widetildeΘ(T^{2/3})$ with respect to the time horizon $T$. The upper bounds rely on a reconstruction identity where the expected objective is written as a gain-from-trade term plus a two-dimensional integral involving the seller and buyer distribution and survival functions, and a kernel function. For $0 < λ< 1/2$, the strong singularity of the kernel is addressed by two main ingredients. First, we propose a new zooming algorithm that sequentially rejects or refines layers of the set of prices. Second, we propose a tailored decomposition of the two-dimensional integral, enabling us to truncate it with a controlled impact on the truncation error. For $1/2 \le λ\le 1$, the singularity of the kernel is mild enough, so that the existing scouting bandit algorithm, already proposed for the case $λ=1$, can be extended.