Local inversion-free estimation of spatial Gaussian processes (Q2008608)

The purpose in this article is to extend and investigate the applicability of the surrogate loss based method for covariance estimation proposed in the article [\textit{M. Anitescu} et al., ``An inversion-free estimating equations approach for Gaussian process models'', J. Comput. Graph. Stat. 26, No. 1, 98--107 (2007; \url{doi:10.1080/10618600.2016.1164056})]. It has been shown that this method is computationally more efficient than the standard MLE. In the present paper one defines a surrogate loss function which will be referred as LIF (local inversion-free) loss. The preconditioned samples are divided into ``bins'', and the LIF loss will be a weighted average of the IF loss functions over these bins. The covariance estimates are obtained by optimizing with respect to the LIF loss function. In the second section of the article the preconditioning technique and IF algorithm are introduced. A large class of scalable covariance estimation algorithms based on this approach are shown. Further, the authors examine the fixed-domain asymptotic properties of the defined IF estimator. A consequence of one of the results is the \(\sqrt{n}\)-consistency of the maximizers of the LIF objective function. The authors establishe the asymptotic distribution of all the stationary points of the LIF loss function. The fifth section of the article is devoted to an extended presentation on simulation results, the sixth section to conclusions and further research. The seventh section and appendices A and B contain proofs and additional technical results related to the paper.