Nonuniform sampling and recovery of multidimensional bandlimited functions by Gaussian radial-basis functions

TL;DR

This paper investigates the convergence of Gaussian-based interpolation of multidimensional bandlimited functions, showing that under certain conditions, the interpolant converges to the original function as the Gaussian variance increases.

Contribution

It establishes new convergence results for Gaussian radial-basis function interpolation of bandlimited functions in multiple dimensions, extending known univariate theorems.

Findings

Interpolant converges in L2 and uniformly as Gaussian variance tends to infinity.

Convergence holds for functions bandlimited to scaled Euclidean balls under specific geometric conditions.

Existence of Riesz-basis sequences depends on the shape of the domain in higher dimensions.

Abstract

Let $S\subset\R^d$ be a bounded subset with positive Lebesgue measure. The Paley-Wiener space associated to $S$, $PW_S$, is defined to be the set of all square-integrable functions on $\R^d$ whose Fourier transforms vanish outside $S$. A sequence $(x_j:j\kin\N)$ in $\R^d$ is said to be a Riesz-basis sequence for $L_2(S)$ (equivalently, a complete interpolating sequence for $PW_S$) if the sequence $(e^{-i\la x_j,\cdot\ra}:j\kin\N)$ of exponential functions forms a Riesz basis for $L_2(S)$. Let $(x_j:j\kin\N)$ be a Riesz-basis sequence for $L_2(S)$. Given $\lambda>0$ and $f\in PW_S$, there is a unique sequence $(a_j)$ in $\ell_2$ such that the function $$ I_\lambda(f)(x):=\sum_{j\in\N}a_je^{-\lambda \|x-x_j\|_2^2}, \qquad x\kin\R^d, $$ is continuous and square integrable on $\R^d$, and satisfies the condition $I_\lambda(f)(x_n)=f(x_n)$ for every $n\kin\N$. This paper studies the convergence of the interpolant $I_\lambda(f)$ as $\lambda$ tends to zero, {\it i.e.,\} as the variance of the underlying Gaussian tends to infinity. The following result is obtained: Let $\delta\in(\sqrt{2/3},1]$ and $0<\beta<\sqrt{3\delta^2 -2}$. Suppose that $\delta B_2\subset Z\subset B_2$, and let $(x_j:j\in\N)$ be a Riesz basis sequence for $L_2(Z)$. If $f\in PW_{\beta B_2}$, then $f=\lim_{\lambda\to 0^+} I_\lambda(f)$ in $L_2(\R^d)$ and uniformly on $\R^d$. If $\delta=1$, then one may take $\beta$ to be 1 as well, and this reduces to a known theorem in the univariate case. However, if $d\ge2$, it is not known whether $L_2(B_2)$ admits a Riesz-basis sequence. On the other hand, in the case when $\delta<1$, there do exist bodies $Z$ satisfying the hypotheses of the theorem (in any space dimension).