A comparison of approaches for valid variogram achievement (Q2488430)

For a stationary isotropic spatial process (field) \(Z(s)\), \(s\in R^d\), the variogram is defined by \[ \gamma(\| t-s\| )=2^{-1} \text{Var}[Z(t)-Z(s)]. \] The authors consider three nonparametric estimates for \(\gamma\) by observations \(Z(s_i)\), \(i=1,\dots,N\): the empirical variogram, Cressie-Hawkins' robust estimator, and Nadaraya-Watson type kernel estimates of the form \[ \widehat\gamma(h)= \biggl(\sum_i\sum_j w_{ij}(Z(s_i)-Z(s_j))^2\biggr) \biggl/\sum_i\sum_j w_{ij}, \] where \(w_{ij}=K((h-\| s_i-s_j\| )/g)\), \(K\) is a symmetric kernel, and \(g\) is a bandwidth. These estimates are compared via simulations. Comparisons to parametric models fitted by generalized least squares are also presented.