Vanilla-SPDE Exchange cuts Gaussian process costs from cubic to near-linear

Gaussian processes (GPs) are a cornerstone of probabilistic machine learning, but their O(n³) complexity limits scalability—especially in spatio-temporal settings where dense grids and disjoint observation/prediction locations balloon costs. Traditional state-space SPDE formulations bring linear time complexity but still suffer from cubic spatial cost, and the problem worsens when inference points don't align with training data. The Vanilla-SPDE Exchange, introduced by Zhang and colleagues, tackles this by exploiting a formal equivalence between the standard GP kernel and SPDE-based representations, allowing a hybrid inference scheme that avoids the worst-case cubic blowup.

The method essentially chooses the cheaper representation at each step: using the standard GP when spatial points are few, and switching to the SPDE formulation when the spatial grid grows dense. This hybrid exchange reduces asymptotic complexity, especially for problems where predictions are needed on a fine grid but observations are sparse or scattered. Numerical experiments on synthetic and real-world data show orders-of-magnitude speedups without sacrificing accuracy. The paper provides rigorous complexity analysis and demonstrates practical gains for dynamic, large-scale GP inference—a key bottleneck in weather forecasting, sensor networks, and spatial statistics.