The asymptotic behaviour of the maximum likelihood function of Kriging approximations using the Gaussian correlation function

Abstract

This study reports on the asymptotic behavior of the maximum likelihood function, encountered when constructing Kriging approximations using the Gaussian correlation function. Of specific interest is a maximum likelihood function that decreases continuously as the correlation function hyper-parameters approach zero. Since the global minimizer of the maximum likelihood function is an asymptote in this case, it is unclear if maximum likelihood estimation (MLE) remains valid. Numerical ill-conditioning of the correlation matrix also occurs in this case. Analytical and numerical examples are presented that demonstrates the validity of MLE, provided that arbitrary precision arithmetic is used. A recent result that claims the MLE function always approaches infinity as the hyper-parameters approach zero is also disproved.