SRRM: Improving Recursive Transport Surrogates in the Small-Discrepancy Regime

A team of researchers including Yufei Zhang, Tao Wang, and Jingyi Zhang has published a paper introducing Selective Recursive Rank Matching (SRRM), a significant advancement in approximating the Wasserstein distance. The Wasserstein distance is a crucial metric in machine learning for comparing probability distributions, but existing recursive partitioning methods like Recursive Rank Matching (RRM) have struggled with accuracy in the 'small-discrepancy regime' where distributions are very similar. The researchers first established consistency and explicit convergence rates for RRM under quadratic cost, then identified the specific mismatch mechanisms causing resolution loss.

The new SRRM algorithm directly addresses these limitations by selectively suppressing the dominant mismatch patterns identified in their analysis. While maintaining computational efficiency characteristic of recursive methods, SRRM achieves substantially higher fidelity as a practical surrogate for the Wasserstein distance. The method represents a 30% improvement in accuracy for moderate additional computational cost, making it particularly valuable for applications requiring precise distribution comparisons, such as generative model evaluation, domain adaptation, and optimal transport problems in machine learning pipelines.