Approximation in \(L_p (\mathbb{R}^d)\) from a space spanned by the scattered shifts of a radial basis function (Q5937039)

Let \(\lambda\in Z_+\). Approximation schemes of the form \[ s(x):= \sum_{\xi\in X}c_\xi\varphi (x-\xi)\quad (x\in\mathbb{R}^d) \] with the ``shifted'' surface spline \[ \varphi(x): =\begin{cases} \bigl(|x|^2+c^2\bigr)/2 \quad & \text{if }\lambda, d\text{ odd},\\ \bigl(|x|^2+c^2 \bigr)/2 \log \bigl( |x|^2+ c^2)^{1/2}\quad &\text{if }\lambda,d\text{ even}\end{cases} \] are known to be effective for approximation to scattered data. Here \(X\subset \mathbb{R}^d\) is the set of nonuniformly distributed centers and \(c\geq 0\) is a ``tension'' parameter. The aim of this interesting paper is to construct an approximation scheme with following properties: (i) every function of a Sobolev space should be well approximated; (ii) every coefficient \(c_\xi\) should be determined by a few values of the data; (iii) the scheme should have a ``smoothing'' effect. The presented approximation schema satisfies all the above properties and provides spectral \(L_p\)-approximation orders with \(1\leq p\leq\infty\), i.e., approximation orders that depend on the smoothness of the approximants. In a numerical test with \(d=2\), the new approximation schema gives better results than the known surface spline interpolation method.