Closed-form Hilbert metric for the SPD bicone, with extended Gaussians

FrnkNlsn · x · 2026-11-19

arXiv paper 2508.14369 (21 pp) by Jacek Karwowski and Frank Nielsen (Sony CSL) studies the Hilbert geometry of the symmetric positive-definite bicone — the parameter space of the extended Gaussian family, obtained by closing the Gaussian family under degenerate covariance and/or precision. The authors derive a closed-form formula for the Hilbert metric distance, exhaustively characterize its invariance properties, classify all isometries of the variance-precision model (generated by orthogonal conjugation and inversion), compare against the Affine-Invariant Riemannian metric (AIRM), and extend results to the boundary and to non-centered Gaussians via the Calvo-Oller embedding. Subjects: computational geometry, machine learning, probability.

Original post →

More from Research

Research channel →