Efficient reconstruction of functions on the sphere from scattered data (Q2384707): Difference between revisions

In this interesting paper, the authors consider the problem of reconstructing a spherical polynomial of degree \(N\) from \(M\) sample values at scattered nodes on the unit sphere. A main result is that for given sampling nodes the polynomial degree \(N\) can either be chosen small enough with respect to the inverse mesh norm or large enough with respect to the inverse separation distance of the sampling set to ensure a well conditioned spherical Fourier matrix. For the overdetermined case \(M>(N+1)^2\), the authors present a fast iterative solution of the least squares approximation problem. Using results of \textit{H. G. Feichtinger, K. Gröchenig} and \textit{T. Strohmer} [Numer. Math. 69, No. 4, 423--440 (1995; Zbl 0837.65148)], the idea of adaptive weights is generalized to the sphere and an \(L^2\)-Marcinkiewicz-Zygmund inequality for dense sampling sets is proven. In the underdetermined case \(M<(N+1)^2\), the authors show stable interpolation for well separated sampling sets. In both cases, the rate of convergence of the fast iterative solvers depends only on the mesh norm and the separation distance of the nodes. For both iterative methods, the total number of floating point operations is \({\mathcal O}(N^2\, \log^2 N + M)\) by the use of spherical fast Fourier transforms [see \textit{S. Kunis} and \textit{D. Potts}, J. Comput. Appl. Math. 161, No.~1, 75--98 (2003; Zbl 1033.65123)]. Finally, several numerical experiments are presented.