Spectral Mixture Kernels for Multi-Output Gaussian Processes

Initially, multiple-output Gaussian processes models (MOGPs) were constructed as linear combinations of independent, latent, single-output Gaussian processes (GPs). This resulted in cross-covariance functions with limited parametric interpretation, thus conflicting with single-output GPs and their intuitive understanding of lengthscales, frequencies and magnitudes to name but a few. On the contrary, current approaches to MOGP are able to better interpret the relationship between different channels by directly modelling the cross-covariances as a spectral mixture kernel with a phase shift. We propose a parametric family of complex-valued crossspectral densities and then build on Cramer's Theorem, the multivariate version of Bochner's Theorem, to provide a principled approach to design multivariate covariance functions. The so-constructed kernels are able to model delays among channels in addition to phase differences and are thus more expressive than previous methods, while also providing full parametric interpretation of the relationship across channels. The proposed method is first validated on synthetic data and then compared to existing MOGP methods on two real-world examples.