Regular Fourier Features for Nonstationary Gaussian Processes
Researchers have developed regular Fourier features to address challenges in simulating nonstationary Gaussian processes. This new method discretizes the spectral representation directly, avoiding the need for probability assumptions that limit stationary processes. The approach yields an efficient, low-rank approximation that maintains correlation structure and positive semi-definiteness, and it can be extended to kernel learning from data when the spectral density is unknown. AI
IMPACT This research offers a more efficient method for simulating complex Gaussian processes, potentially improving machine learning models that rely on these statistical techniques.