On the consistent separation of scale and variance for Gaussian random fields (Q2380092)

A common situation in spatial statistics is when one has observations on a single realization of a random field \(Y\) at a large number of spatial points \({\mathbf t}_1\), \({\mathbf t}_2,\dots\) within some bounded region \(\Omega\subset {\mathbb R}^d\). One is then faced with the problem of predicting some quantity that depends on \(Y\) at unobserved points in \(\Omega\). Typically, a fully nonparametric estimation of the covariance structure of \(Y\) is difficult since the observations are from one realization of the random field. In this case, it is common to consider a class of covariance structures indexed by a finite number of parameters which are then estimated using the observations. Two common parameters found in many covariance models are an overall scale \(\alpha\) and an overall variance \(\sigma^2\). The simplest example of this model stipulates that the random field \(Y\) is a scale and amplitude chance by an unknown \(\alpha\) and \(\sigma\) of a known random field \(Z\): \[ \{Y({\mathbf t}):~{\mathbf t}\in \Omega\}\doteq \{\sigma Z(\alpha {\mathbf t}):~{\mathbf t}\in \Omega\}, \] where \(\doteq\) denotes equality of the finite-dimensional distributions. The author presents fixed domain asymptotic results that establish consistent estimates of the variance and scale parameters for a Gaussian random field with a geometric anisotropic Matérn autocovariance in dimension \(d>4\). When \(d<4\), this is impossible due to the mutual absolute continuity of Matérn Gaussian random fields with different scale and variance [see \textit{H. Zhang}, J. Amer. Statist. Assoc. 99, No. 465, 250--261 (2004; Zbl 1089.62538)]. Informally, when \(d>4\), the author shows that one can estimate the coefficient on the principle irregular term accurately enough to get a consistent estimate of the coefficient on the second irregular term. These two coefficients can then be used to separate scale and variance. The author extends his results to the general problem of estimating variance and geometric anisotropy for more general autocovariance functions. These results illustrate the interaction space and the number of increments used for the estimation. As a corrolary, these results establish the orthogonality of Matérn Gaussian random fields with different parameters when \(d>4\). The case \(d=4\) is still open.