Regular Fourier Features for Nonstationary Gaussian Processes

A team of researchers including Arsalan Jawaid, Abdullah Karatas, and Jörg Seewig has published a paper titled 'Regular Fourier Features for Nonstationary Gaussian Processes,' introducing a breakthrough method for simulating complex data patterns. Traditional Gaussian process simulation requires sampling from high-dimensional distributions with cubic scaling complexity (O(n³)), making it computationally prohibitive for large datasets. While spectral methods using Fourier representations have helped stationary processes by treating spectral density as a probability distribution, this approach fails for nonstationary processes where spectral densities aren't valid probability measures. The new method directly addresses this limitation for harmonizable processes.

The proposed 'Regular Fourier Features' method discretizes the spectral representation directly, preserving the correlation structure among spectral weights without requiring probability assumptions. Under finite spectral support assumptions, this yields an efficient low-rank approximation that's positive semi-definite by construction. When spectral density is unknown, the framework extends naturally to kernel learning from data. The researchers demonstrated their method on locally stationary kernels and harmonizable mixture kernels with complex-valued spectral densities, showing practical applications in spatial-temporal modeling, climate prediction, and financial time series analysis where data patterns change over time or space. This represents a significant advancement in scalable machine learning for complex real-world datasets.