Spatial autoregression and related spatio-temporal models. (Q1421867): Difference between revisions

Let \(\alpha_1, \dots, \alpha_p>0\) be distinct constants, \(A_k= \left[ \alpha_k \prod_{j\neq k} (\alpha_j^2-\alpha_k^2) \right]^{-1}\), \(k=1,\dots,p\), \(p>1\). Define \(C({\mathbf s}) = \sum_{k=1}^p A_k \exp(-\alpha_k \| {\mathbf s} \|)\), \({\mathbf s}\in {\mathbb R}^d\). The author proves that \(C({\mathbf s})\) is a covariance function on \({\mathbb R}^d\) if and only if \(\| \cdot \|\) is the Euclidean norm. This result is surprising since for \(p=1\) and \(d=2\), \(\exp(-\| {\mathbf s} \|)\) is a covariance function on \({\mathbb R}^2\) for every norm \(\| \cdot \|\). The function \(C({\mathbf s})\) is a \(d\)-dimensional generalization of the one-dimensional covariance function \(C_0(t) = \sum_{k=1}^p A_k \exp(-\alpha_k| t |)\), \(t \in {\mathbb R}\). The author shows that not all continuous-time covariance functions on the real line with the rational spectral density can be extended to a higher dimensional domain. A class of semiparametric spatio- temporal models is introduced that are stationary in space and have spatial autoregressive margins.