Power series kernels (Q734058)

Let \(\alpha\in\mathbb{N}_{0}^{d}\) and let \(w_{\alpha}\) be a positive real number such that \(\sum_{\alpha\in\mathbb{N}_{0}^{d}}\) \(\frac{w_{\alpha}}{(\alpha!)^{2}}\) \(<\infty\) holds. The expression \(K(x,y)=\sum_{\alpha\in\mathbb{N}_{0}^{d} }w_{\alpha}\frac{x^{\alpha}y^{a}}{(\alpha!)^{2}}\) where \(x,y\in\Omega \subset\mathbb{R}^{d}\) is called a power series kernel (PSK for short). The author considers the PSK and under weak additional conditions on the weights \(w_{\alpha},\) she determines the ``native'' Hilbert space \(\mathcal{N}_{K}\) of multivariate functions in which the kernels \(K(x,y)\) are reproducing. For a function \(f\in\mathcal{N}_{K}\) and \(X=\{x_{1},x_{2},\dots ,x_{N}\}\subset\Omega\) the author considers the uniquely determined interpolant of the form \(s_{f,X}(x)=\sum_{j=1}^{N}\alpha_{j}K(x,x_{j})\) satisfying \(s_{f,X} (x_{k})=f(x_{k})\) for \(k=1,2,\dots ,N\) and the error between \(f\) and its interpolant \(s_{f,X}\) are expressed in terms of the fill distance \(h:=h_{X,\Omega}:=\sup_{x\in\Omega}\min_{x_{j}\in X}\| x-x_{j} \|_{2}\). She shows that the convergence occurs for the function and all derivatives, when \(h\rightarrow0\). Finally, by truncation of the interpolant one obtains a multivariate polynomial \(T^{(k)}s_{f,X}\) and the approximation error between \(f\) and \(T^{(k)}s_{f,X}\) is studied. So the classical Bernstein theorem concerning polynomial approximation of analytic functions is extended to the multivariate case. An application to machine learning algorithm is presented.