Ambient approximation on embedded submanifolds (Q2216644): Difference between revisions

Approximation on manifolds is a currently much researched-into area and there are several approaches to it for a collection of special types of manifolds. An example is the sphere where in particular splines and radial basis functions offer a good class of choices for approximants. It is quite a different matter to make approximations on general (sub-)manifolds that are embedded into higher dimensional ambient spaces. Hypersurfaces of a variety of co-dimensions have been studied already, but in this paper, a more general approach is taken. It uses (tensor-product) B-splines as a tool but the other tensor-product B-splines that were used before as a main scheme are now only employed as an intermediate approximation. The splines provide the well-known approximation orders within the ambient space, \(h^m\) say, for \(C^{m-1}\) functions. They provide the least number of operations on a computer and are available in three (but not four) dimensions for example (remember the hairy-ball theorem in this context). Other approaches that would be an alternative to the algorithms of this article and are available for scattered data (e.g. for solving partial differential equations numerically) are spherical splines and spherical harmonics on spheres or approximations simply on the global (ambient) space by radial basis functions, then restricted to the manifold. On the ambient space, an approximation with the very useful method of quasi-interpolation could be used. Since the manifolds can be defined by so-called charts, also approximations individually on the charts are possible, but they lead to non-linear difficult equations to be solved and to blending problems. Therefore the authors do not use this method either. Here an ambient approximation is carried out by extending the approximant constantly along normals within tubes about the manifold, then use approximation with a local method, e.g. again, the extremely efficient quasi-interpolation (for instance with radial basis functions) that consists of evaluating the now transformed approximant point-wise and making semi-discrete convolutions with shifts of the so-called quasi-Lagrange functions. This is done for any dimension, not only for hypersurfaces as before; after the intermediate approximation within the tube, the approximant is restricted to the manifold again. These steps together form a relatively simple, efficient method for approximation in very general circumstances. The author provides also interesting numerical examples.