Search papers, labs, and topics across Lattice.
This paper introduces regular Fourier features for approximating nonstationary Gaussian Processes (GPs) by directly discretizing the spectral representation, thereby avoiding the restrictive probability measure assumption required by traditional spectral methods. The method preserves the correlation structure among spectral weights, leading to a positive semi-definite low-rank approximation under a finite spectral support assumption. Experiments demonstrate the effectiveness of the approach on locally stationary kernels and harmonizable mixture kernels, including those with complex-valued spectral densities.
Bypass the limitations of probabilistic spectral methods for non-stationary Gaussian Processes with regular Fourier features that directly discretize the spectral representation, preserving correlation structure.
Simulating a Gaussian process requires sampling from a high-dimensional Gaussian distribution, which scales cubically with the number of sample locations. Spectral methods address this challenge by exploiting the Fourier representation, treating the spectral density as a probability distribution for Monte Carlo approximation. Although this probabilistic interpretation works for stationary processes, it is overly restrictive for the nonstationary case, where spectral densities are generally not probability measures. We propose regular Fourier features for harmonizable processes that avoid this limitation. Our method discretizes the spectral representation directly, preserving the correlation structure among spectral weights without requiring probability assumptions. Under a finite spectral support assumption, this yields an efficient low-rank approximation that is positive semi-definite by construction. When the spectral density is unknown, the framework extends naturally to kernel learning from data. We demonstrate the method on locally stationary kernels and on harmonizable mixture kernels with complex-valued spectral densities.
