Mathematicians unveil Wasserstein Mahalanobis distance for AI geometry

A team of mathematicians—Chuxiangbo Wang, Shiying Li, and Caroline Moosmüller—has published a groundbreaking preprint introducing the Wasserstein Mahalanobis distance, a new metric that bridges classical Mahalanobis distance with optimal transport theory. Published on arXiv (arXiv:2608.06560), this 62-page paper redefines how geometric structure can be recovered from nonlinear observations by replacing Euclidean displacement vectors with optimal transport displacement fields and covariance matrices with covariance operators defined on Wasserstein tangent spaces.

The theoretical contribution is substantial: the authors prove that the Wasserstein Mahalanobis distance approximates classical Mahalanobis geometry even under smooth nonlinear transformations, with exact correspondence for affine transformations and controlled higher-order error for general cases. Numerical experiments confirm these predictions, demonstrating accurate recovery of latent geometric structure in high-dimensional data. This work establishes a distribution-valued analog of classical Mahalanobis geometry, offering a robust foundation for covariance-adapted learning directly in Wasserstein space—potentially transforming fields like nonlinear independent component analysis and generative modeling.