Non-separable non-stationary random fields

We describe a framework for constructing nonsta- tionary nonseparable random fields that are based on an infinite mixture of convolved stochastic processes. When the mixing process is station- ary but the convolution function is nonstationary we arrive at nonseparable kernels with constant non-separability that are available in closed form. When the mixing is nonstationary and the convolu- tion function is stationary we arrive at nonsepara- ble random fields that have varying nonseparabil- ity and better preserve local structure. These fields have natural interpretations through the spectral representation of stochastic differential equations (SDEs) and are demonstrated on a range of syn- thetic benchmarks and spatio-temporal applica- tions in geostatistics and machine learning. We show how a single Gaussian process (GP) with these random fields can computationally and sta- tistically outperform both separable and existing nonstationary nonseparable approaches such as treed GPs and deep GP constructions