Schoenberg characterization of continuous non-stationary isotropic positive definite kernels

2025/06/27 by Felix Benning, Max David Schölpple, Benning, Felix +2
Computer Science · Mathematics · #Stochastic Gradient Optimization Techniques #Neural Networks and Applications #Tensor decomposition and applications

paper · pdf · doi:10.48550/arxiv.2506.22048

Abstract

We characterize the continuous isotropic positive definite kernels on ℝ^d, where isotropy refers to invariance under the orthogonal group O(d) but not necessarily stationarity. Furthermore, we characterize strict positive definiteness for such kernels. The class of isotropic kernels is fairly general as it unifies stationary isotropic and dot product kernels, and includes neural network kernels that arise from infinite-width limits of neural networks. As an application, we further characterize the continuous isotropic Gaussian random functions in terms of a series representation.

Citations

• Tensor Programs I: Wide Feedforward or Recurrent Neural Networks of Any Architecture are Gaussian Processes
• Handbook of Mathematical Functions With Formulas, Graphs and Mathematical Tables (National Bureau of Standards Applied Mathematics Series No. 55)
• Metric spaces and positive definite functions

Related

• Positive definite matrix-valued kernels and their scalar valued projections: counterexamples
• Entropy numbers of Reproducing Hilbert Space of zonal positive definite kernels on compact two-point homogeneous spaces
• A close look at the entropy numbers of the unit ball of the Reproducing Hilbert Space of isotropic positive definite kernels
• Strictly positive definite non-isotropic kernels on two-point homogeneous manifolds: The asymptotic approach
• Matrix valued positive definite kernels related to the generalized Aitken's integral for Gaussians