Local Kernels and the Geometric Structure of Data

TL;DR

This paper develops a theory of local kernels that connect kernel-based data analysis methods to the underlying geometric structure of data, enabling the design of invariant embeddings and geometric reconstructions.

Contribution

It introduces a unified geometric framework for local kernels, linking them to the generator of Markov processes and enabling the creation of geometry-aware data analysis tools.

Findings

Local kernels approximate the generator of a continuous Markov process.

Symmetric kernels relate to the Laplace-Beltrami operator and geometry.

Framework enables invariant data embeddings and diffeomorphism reconstructions.

Abstract

We introduce a theory of local kernels, which generalize the kernels used in the standard diffusion maps construction of nonparametric modeling. We prove that evaluating a local kernel on a data set gives a discrete representation of the generator of a continuous Markov process, which converges in the limit of large data. We explicitly connect the drift and diffusion coefficients of the process to the moments of the kernel. Moreover, when the kernel is symmetric, the generator is the Laplace-Beltrami operator with respect to a geometry which is influenced by the embedding geometry and the properties of the kernel. In particular, this allows us to generate any Riemannian geometry by an appropriate choice of local kernel. In this way, we continue a program of Belkin, Niyogi, Coifman and others to reinterpret the current diverse collection of kernel-based data analysis methods and place them in a geometric framework. We show how to use this framework to design local kernels invariant to various features of data. These data-driven local kernels can be used to construct conformally invariant embeddings and reconstruct global diffeomorphisms.