Abstract:We introduce a new family of distances, relative-translation invariant Wasserstein distances ($RW_p$), for measuring the similarity of two probability distributions under distribution shift. Generalizing it from the classical optimal transport model, we show that $RW_p$ distances are also real distance metrics defined on the quotient set $\mathcal{P}_p(\mathbb{R}^n)/\sim$ and invariant to distribution translations. When $p=2$, the $RW_2$ distance enjoys more exciting properties, including decomposability of the optimal transport model, translation-invariance of the $RW_2$ distance, and a Pythagorean relationship between $RW_2$ and the classical quadratic Wasserstein distance ($W_2$). Based on these properties, we show that a distribution shift, measured by $W_2$ distance, can be explained in the bias-variance perspective. In addition, we propose a variant of the Sinkhorn algorithm, named $RW_2$ Sinkhorn algorithm, for efficiently calculating $RW_2$ distance, coupling solutions, as well as $W_2$ distance. We also provide the analysis of numerical stability and time complexity for the proposed algorithm. Finally, we validate the $RW_2$ distance metric and the algorithm performance with three experiments. We conduct one numerical validation for the $RW_2$ Sinkhorn algorithm and show two real-world applications demonstrating the effectiveness of using $RW_2$ under distribution shift: digits recognition and similar thunderstorm detection. The experimental results report that our proposed algorithm significantly improves the computational efficiency of Sinkhorn in certain practical applications, and the $RW_2$ distance is robust to distribution translations compared with baselines.

Efficient algorithms computing distances between Radon measures on R

A new class of transport distances

Analysis of particle methods for structured population models with nonlocal boundary term in the framework of bounded Lipschitz distance

Computing the Distance between unbalanced Distributions -- The flat Metric

Metrics on fractals and sub-gaussian heat kernel estimates

Radon measures and Lipschitz graphs

The Riemannian geometry of Sinkhorn divergences

Relative-Translation Invariant Wasserstein Distance

On the Wasserstein Median of Probability Measures

Computing distances on Riemann surfaces

Empirical measures and random walks on compact spaces in the quadratic Wasserstein metric

The boundary method for semi-discrete optimal transport partitions and Wasserstein distance computation

Near-linear time approximation algorithms for optimal transport via Sinkhorn iteration

Sinkhorn Distances: Lightspeed Computation of Optimal Transport

Rate estimates for total variation distance with applications

Stereographic Spherical Sliced Wasserstein Distances

Alpha Procrustes metrics between positive definite operators: A unifying formulation for the Bures-Wasserstein and Log-Euclidean/Log-Hilbert-Schmidt metrics

On the equivalence between Fourier-based and Wasserstein distances for probability measures on $\mathbb N$

Unbalanced Optimal Transport and Maximum Mean Discrepancies: Interconnections and Rapid Evaluation

On the Equivalence of Statistical Distances for Isotropic Convex Measures