Average Case tractability of multivariate approximation with Gaussian kernels
read the original abstract
We study the problem of approximating functions of $d$ variables in the average case setting for the $L_2$ space $L_{2,d}$ with the standard Gaussian weight equipped with a zero-mean Gaussian measure. The covariance kernel of this Gaussian measure takes the form of a Gaussian kernel with non-increasing positive shape parameters $\gamma_j^2$ for $j = 1, 2, \dots, d$. The error of approximation is defined in the norm of $L_{2,d}$. We study the average case error of algorithms that use at most $n$ arbitrary continuous linear functionals. The information complexity $n(\varepsilon, d)$ is defined as the minimal number of linear functionals which are needed to find an algorithm whose average case error is at most $\varepsilon$. We study different notions of tractability or exponentially-convergent tractability (EC-tractability) which the information complexity $n(\varepsilon, d)$ describe how behaves as a function of $d$ and $\varepsilon^{-1}$ or as one of $d$ and $(1+\ln\varepsilon^{-1})$. We find necessary and sufficient conditions on various notions of tractability and EC-tractability in terms of shape parameters. In particular, for any positive $s>0$ and $t\in(0,1)$ we obtain that the sufficient and necessary condition on $\gamma^2_ j$ for which $$\lim_{d+\varepsilon^{-1}\to\infty}\frac{n(\varepsilon,d)}{\varepsilon^{-s}+d^t}=0$$ holds is $$ \lim_{j\to \infty}j^{1-t}\gamma_j^2\,\ln^+ \gamma_j^{-2}=0,$$where $\ln^+ x=\max(1,\ln x)$.
This paper has not been read by Pith yet.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.