Hyperbolic Sliced-Wasserstein via Geodesic and Horospherical Projections

Clément Bonet, Laetitia Chapel, Lucas Drumetz, Nicolas Courty

Abstract

Embedding data in hyperbolic spaces has proven highly beneficial for capturing underlying hierarchical structures. However, comparing probability distributions in these curved spaces remains a challenge. While Optimal Transport (OT) distances are well-defined on Riemannian manifolds, they suffer from high computational costs. In Euclidean spaces, Sliced-Wasserstein (SW) distances provide a computationally efficient alternative by leveraging closed-form 1D solutions, but they are not directly applicable to hyperbolic geometry. This paper introduces two novel hyperbolic sliced-Wasserstein discrepancies—Geodesic (GHSW) and Horospherical (HHSW)—which use intrinsic projections to enable efficient distribution comparison. We demonstrate their effectiveness in tasks such as gradient flows and image classification.

Introduction

In machine learning, comparing probability distributions is a fundamental task for generative modeling, domain adaptation, and representation learning. The Wasserstein distance, derived from Optimal Transport theory, is a powerful metric but scales cubically ($O(n^3)$) with the number of samples. The Sliced-Wasserstein (SW) distance mitigates this by projecting high-dimensional distributions onto 1D lines, where the Wasserstein distance has a simple closed-form solution based on sorted samples.

However, many modern datasets—such as social networks, biological taxonomies, and linguistic hierarchies—are best represented in hyperbolic space, which acts as a continuous analog to trees. Standard SW fails here because hyperbolic space is curved; "straight lines" (geodesics) behave differently, and Euclidean projections ignore the intrinsic geometry. This work bridges that gap by defining slicing operators specifically for hyperbolic manifolds.

Proposed Methods: GHSW and HHSW

The authors propose two novel "slicing" strategies tailored for hyperbolic geometry, specifically for the Lorentz (hyperboloid) and Poincaré ball models:

  1. Geodesic Hyperbolic Sliced-Wasserstein (GHSW): This method projects points onto a geodesic (the shortest path between points) by finding the closest point on that path.
  2. Horospherical Hyperbolic Sliced-Wasserstein (HHSW): This method projects points along "horospheres"—surfaces that can be visualized as spheres with infinite radius. This approach is inspired by the Busemann function, which describes the "direction" of points at infinity.
Geodesic paths and projections in the Poincaré disk illustrate the intrinsic geometry used for slicing.

These projections allow the high-dimensional hyperbolic OT problem to be decomposed into multiple 1D problems on geodesics, maintaining the $O(Ln \log n)$ efficiency of Euclidean SW (where $L$ is the number of projections).

Key Insights and Results

The "Origin" Trap and Distance Preservation

A critical finding is that geodesic projections (GHSW) tend to "squeeze" points toward the origin, which can lead to numerical instability or slow convergence when data is far from the center. In contrast, horospherical projections (HHSW) tend to preserve relative distances between points more robustly. This makes HHSW particularly effective for gradient flows, where one distribution is iteratively moved to match another.

A comparison of projection mechanisms: Euclidean projections (straight lines) fail to capture the curvature inherent in hyperbolic data.

Performance and Scalability

  • Speed: The proposed methods are orders of magnitude faster than exact Wasserstein or Sinkhorn distances, matching the scalability of Euclidean SW.
  • Classification Accuracy: Using HHSW or GHSW as a global prior in hyperbolic image classifiers improved test accuracy on CIFAR-100. For instance, in 10 dimensions, HHSW achieved 62.80% accuracy, outperforming the previous state-of-the-art baseline (59.19%).
  • Dimension Independence: Like Euclidean SW, the sample complexity of these hyperbolic variants does not depend on the dimension of the space, making them ideal for high-dimensional embeddings.

Conclusion and Future Work

This work introduces the first intrinsic Sliced-Wasserstein discrepancies for hyperbolic spaces. By leveraging geodesic and horospherical projections, the authors provide a tool that is both theoretically grounded and computationally efficient for hierarchical data. While these measures are currently treated as "pseudo-distances," future work aims to prove they are true distances by investigating the injectivity of hyperbolic Radon transforms. Additionally, exploring other geometric objects for slicing, such as horocycles, may further improve performance for specific data geometries.

Summary updated

Abstract

It has been shown beneficial for many types of data which present an underlying hierarchical structure to be embedded in hyperbolic spaces. Consequently, many tools of machine learning were extended to such spaces, but only few discrepancies to compare probability distributions defined over those spaces exist. Among the possible candidates, optimal transport distances are well defined on such Riemannian manifolds and enjoy strong theoretical properties, but suffer from high computational cost. On Euclidean spaces, sliced-Wasserstein distances, which leverage a closed-form of the Wasserstein distance in one dimension, are more computationally efficient, but are not readily available on hyperbolic spaces. In this work, we propose to derive novel hyperbolic sliced-Wasserstein discrepancies. These constructions use projections on the underlying geodesics either along horospheres or geodesics. We study and compare them on different tasks where hyperbolic representations are relevant, such as sampling or image classification.
BibTeX
@inproceedings{bonet2023hyperbolic,
  author = {Clément Bonet and Laetitia Chapel and Lucas Drumetz and Nicolas Courty},
  title = {{Hyperbolic Sliced-Wasserstein via Geodesic and Horospherical Projections}},
  booktitle = {TAG-ML},
  year = {2023},
  url = {https://openreview.net/forum?id=g4XVO7gGOk},
}
Figure