Reproduce “Matérn Gaussian processes on Riemannian manifolds” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Viacheslav Borovitskiy, Alexander Terenin, Peter Mostowsky, Marc Peter Deisenroth

This research develops practical methods for defining and computing Matérn and squared exponential Gaussian process kernels on compact Riemannian manifolds. The approach utilizes the spectral theory of the Laplace–Beltrami operator to derive explicit kernel formulas and spectral measures, enabling scalable and geometrically consistent modeling on non-Euclidean data.

Key insights

  • SPDE-defined Matérn and squared exponential GPs on compact Riemannian manifolds are demonstrably equivalent to the reproducing kernels of specific Sobolev and diffusion Hilbert spaces, respectively.
  • For the d-dimensional torus, the general framework reduces to kernels obtained by periodic summation of Euclidean Matérn kernels, providing insight into why naive geodesic distance kernels fail.
  • The constructive derivation of kernels and spectral measures effectively transforms a theoretically sound but practically intractable SPDE formulation into a computationally amenable one.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.