Geometry-Adaptive Clustering on High-Dimensional Tori

Abstract

Multivariate angular data lie on a torus, but clustering methods often use the flat torus. We introduce curved-torus energy (CTE), a position-dependent dissimilarity based on restricted Riemannian path energy and minimization over winding classes. CTE has a certified finite winding reduction, an exact flat endpoint, a geodesic upper bound, and stability with respect to the geometry parameter. For label-free geometry selection, we propose CEHS-HM, a held-out criterion for selecting the warping parameter at a specified clustering resolution. We establish a finite-grid population-score regret bound and also consider joint geometry-resolution selection over finite candidate grids. Controlled experiments on $\mathbb{T}^3$ show that CTE agrees with the flat method at zero warping and becomes more useful as geometric warping increases. Additional experiments show that the CEHS-selected geometry should be viewed as a data-adaptive clustering choice rather than as an estimator of a generative deformation parameter. Under a product-von-Mises mechanism, the selector tends toward the flat endpoint when curvature is unnecessary. On ILE protein torsion data, the proposed method identifies a stable five-mode representation of joint backbone-side-chain conformation.