Kriging, cokriging, radial basis functions and the role of positive definiteness (Q1207565)

The author studies two interpolators: a) \(z^*(x)=\sum_{i=1}^ n b_ i g(\| x_ i- x\|)+\sum_{k=0}^ p c_ k f_ k(x)\), \(x_ i\in \mathbb{R}^ k\), \(1\leq i\leq n\), where \(f_ k(x)\), \(0\leq k\leq p\) are linearly independent monomials in the position coordinates of \(x\). b) \(z^*(x)=\sum_{i=1}^ n \lambda_ i(x) z(x_ i)\), where \(z(x_ i)\), \(1\leq i\leq n\), are observed data. These two interpolators represent two different approaches to the same problem and in certain special cases they are equivalent. It is noted that positive definiteness plays a central role with respect to the uniqueness of the coefficients in the interpolators a) and b). There are, at least, three developments for interpolators that lead to the same functional form for the interpolators: the thin plate spline, radial basis functions and the regression method known as kriging. All of them are studied.