On the Expressive Power of Mixed-Curvature Representations in Product Manifolds

Haitz Sáez de Ocáriz Borde

Track: tiny paper (up to 4 pages)

Keywords: mixed-curvature embeddings, product manifolds, product metrics, sectional curvature, local isometry, expressive power, geometric deep learning

Abstract: Mixed-curvature representations based on product manifolds of Euclidean, hyperbolic, and spherical spaces are widely used as latent geometries in representation learning and geometric deep learning. Their popularity stems from their analytical tractability and closed-form geodesic distances, which facilitate efficient backpropagation in deep learning. In the literature, they are often portrayed as richer representations than, for instance, hyperbolic embeddings. However, their expressive power is not well understood. In this work, we show that such mixed-curvature product manifolds impose rigid geometric constraints that fundamentally limit their expressive power. In particular, we prove that generic Riemannian manifolds cannot be locally represented isometrically by mixed-curvature product manifolds equipped with product metrics. The obstruction arises from curvature splitting: product metrics necessarily exhibit vanishing mixed sectional curvature (equivalently, flatness on mixed 2-planes), a rigidity that fails for a generic Riemannian metric.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Submission Number: 91