Mixed-Curvature Tree-Sliced Wasserstein Distance

ICLR 2026 Conference Submission5099 Authors

Keywords: mixed curvature space, sliced optimal transport

TL;DR: We propose MC-TSW, a valid and efficient metric for comparing distributions in mixed-curvature spaces, showing superior performance over product-space and constant-curvature baselines.

Abstract: Mixed-curvature spaces have emerged as a powerful alternative to their Euclidean counterpart, enabling data representations better aligned with the intrinsic structure of complex datasets. However, comparing probability distributions in such spaces remains underdeveloped: existing measures such as KL divergence and Wasserstein either rely on strong assumptions on distributions or incur high computational costs. The Sliced-Wasserstein (SW) framework provides an alternative approach for constructing distributional distances; however, its reliance on one-dimensional projections limits its ability to capture the geometry of the ambient space. To address this limitation, the Tree-Sliced Wasserstein (TSW) framework employs tree structures as a richer projected space. Motivated by the intuition that such a space is particularly suitable for representing the geometric properties of mixed-curvature manifolds, we introduce the Mixed-Curvature Tree-Sliced Wasserstein (MC-TSW), a novel discrepancy measure that is computationally efficient while faithfully capturing both the topological and geometric structures of mixed-curvature spaces. Specifically, we introduce an adaptation of tree systems and Radon transform to mixed-curvature spaces, which yields a closed-form solution for the optimal transport problem on the tree system. We further provide theoretical analysis on the properties of the Radon transform and the MC-TSW distance. Experimental results demonstrate that MC-TSW improves distributional comparisons over product-space-based distance and line-based baselines, and that mixed-curvature representations consistently outperform constant-curvature alternatives, highlighting their importance for modeling complex datasets.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 5099