Asymptotic coefficients for Gaussian radial basis function interpolants (Q983990): Difference between revisions

Radial basis functions (RBFs) are used in multidimensional interpolation on irregular or scattered data. The form of the approximation is \[ f(\vec{x})\approx \sum_{j=1}^{N} \lambda_j\varphi(\| \vec{x}-\vec{x}_j\|_2), \quad x\in\mathbb R^d \] \((d\geq1)\) for some function \(\varphi(r)\) and some set of \(N\) points \(\vec{x}_j\), which are called the ``centers''. The coefficients \(\lambda_j\), \(j=\overline{1,N}\) are found from the interpolation conditions at a given set of points \(\vec{y}_k\), \(k=\overline{1,N}\). In this paper the authors consider the Gaussian RBFs, i.e., \(\varphi(x)=\exp(-[\alpha^2/h^2]x^2)\) on an infinite interval with uniform grid spacing \(h\), where the parameter \(\alpha\) is the so called inverse width of RBF functions relative to \(h\), For general RBF species, one obtains the RBF coefficients of the trivial function \(f(x)\equiv1\), and, in particular, for Gaussian RBFs. For \(\alpha\) fixed and \(h\rightarrow0\) one proves that the error for a general \(f(x)\) is proportional to that for the trivial function \(f(x)\equiv1\). Also, for \(\alpha\rightarrow0\) and \(h\) fixed one obtains \(\lambda_{j}\approx\exp (\pi^2/(4\alpha^2))\). For \(\alpha\approx1/2\) the size of the Gaussian RBFs coefficients and the condition number of the interpolation matrix are both no larger than \(O(10^{4})\). The representative example of the ``Witch of Agnesi'' is considered.