Interpolation by radial functions on Sobolev space (Q5958883): Difference between revisions

The purpose of this paper is to study the behavior of interpolation by smooth radial basis functions on larger spaces, especially on the homogeneous Sobolev spaces. The author considers approximation of real-valued functions \(f\) which are known only at a discrete set \(X:=\{x_1,\dots,x_N\}\) in \(\mathbb{R}^d\), \(d\geq 1\). Given data \((x_j,f(x_j))\), \(j=1,\dots,N\), the radial basis function approach is to choose a function \(\varphi\): \(\mathbb{R}^d\to\mathbb{R}\) and to difine an approximant \(s_{f,X}\) by \[ s_{f,X}(x):=p(x)+\sum_{j=1}^N\alpha_j\varphi_h(x-x_j), \quad p\in \Pi_m , \] where \(\varphi_h:=\varphi(\cdot/h)\) and \(\alpha_j\) are chosen so that \(\sum_{j=1}^N\alpha_j q(x_j)=0\) for all \(q\in \Pi_m\). Here \(\Pi_m\) denotes the class of all algebraic polynomials of degree less than \(m\) on \(\mathbb{R}^d\). The Fourier transform of \(f\in L_1(\mathbb{R}^d)\) is defined as \[ \widehat{f}(\theta):=\int_{\mathbb{R}^d}f(t)e^{-i\theta\cdot t} dt . \] The author gets the following result: Let \(\widehat{\varphi}\) satisfy the condition \(|\cdot|^{2m}\widehat{\varphi}=F>0\) with \(m>d/2\) and \(F\in L_\infty(\mathbb{R}^d)\). Assume that there exists a constant \(\rho>0\) such that \(h/q\leq\rho\). Let \(\varphi\) be a Gaussian function \(\varphi(x):=e^{-|x|^2}\), and assume that \(\widehat{\varphi}\) decays slower than \(\widehat{\varphi}\) around \(\infty\) such that \(\widehat{\varphi}_q/\widehat{\varphi}_h\) is uniformly bounded. Then, for every \(f\in W_2^m(\mathbb{R}^d)\cap W_\infty^m(\mathbb{R}^d)\), we have an error bound of the form \[ |f(x)-s_{f,X}(x)|\leq ch^{m-d/2}(1+P_{\varphi,X/h}(x/h) , \] where the constant \(c>0\) depends on \(|f|_{m,2}\) and \(|f|_{m,\infty}\). Here \(W_p^m(\mathbb{R}^d)\) denotes the homogeneous Sobolev space, i.e., \[ W_p^m(\mathbb{R}^d):=\Biggl\{f:|f|_{m,p}:=\biggl(\sum_{|\alpha|= m}||D^\alpha f||^p_{L_p(\mathbb{R}^d)}\biggr)^{1/p}<\infty\Biggr\} \] with \(1\leq p\leq\infty\), and \(P_{phi,X}(x)\) is the norm of the error functional.