Modeling of multivariate random fields through Gaussian processes calls for the construction of valid cross-covariance functions describing the dependence between any two component processes at different spatial locations. The required validity conditions often present challenges that lead to complicated restrictions on the parameter space. The purpose of this work is to present techniques using multivariate mixtures for establishing validity that are simultaneously simplified and comprehensive. This is accomplished using results on conditionally negative semidefinite matrices and the Schur product theorem. For illustration, we use the recently-introduced Confluent Hypergeometric (CH) class of covariance functions. In addition, we establish the spectral density of the Confluent Hypergeometric covariance and use this to construct valid multivariate models as well as propose new cross-covariances. Our approach leads to valid multivariate cross-covariance models that inherit the desired marginal properties of the Confluent Hypergeometric model and outperform the multivariate Mat\'ern model in out-of-sample prediction under slowly-decaying correlation of the underlying multivariate random field. We also establish properties of the new models, including results on equivalence of Gaussian measures. We demonstrate the new model's use for multivariate oceanography dataset consisting of temperature, salinity and oxygen, as measured by autonomous floats in the Southern Ocean.
翻译:暂无翻译