Almost optimal estimates for best approximation by translates on a torus (Q2386419)

Let \({\mathbb T}^d\) be the \(d\)-dimensional torus, let \(\phi\) be some function from \(L_2=L_2({\mathbb T}^d)\), and let \(H_n(\varphi)\) denote the set of all functions of the form \[ \sum_{i=1}^n b_i \varphi (\cdot - a_i), \quad b_i \in {\mathbb R}, \quad a_i \in {\mathbb R}^d, \quad i=1, \ldots , n. \] The following lower estimate is established for the deviation in \(L_2\) of the Sobolev class \(W_2^r\) from \(H_n(\varphi)\): \[ \text{ dist}(W_2^r, H_n(\varphi))\geq C (n\log n)^{-r/d}, \] where \(C\) is a positive constant depending only on \(r\) and \(d\). A general method is developed for obtaining an upper estimate. In the case when the Fourier coefficients \({\widehat \varphi}_k\) decrease as \(| k| ^{-\alpha}\) with \(\alpha >\max\{r, d/2\}\) or as \(e^{-\alpha | k| }\) with \(\alpha >0\), this method gives \(\text{ dist}(W_2^r, H_n(\varphi))\leq C_1 n^{-r/d}\). These results are applied to obtain estimates for radial basis approximation.