Sparse data interpolation and smoothing on embedded submanifolds (Q777045)

Quasi-interpolation especially, and also interpolation, are popular tools in the approximation of functions and data in one or more dimensions for the purpose of smoothing (for data with noise) or extra- or interpolation. Several very good methods use approximations from spaces of piecewise polynomials, splines, or the related radial basis functions. Both of those have certain minimising properties especially when interpolation at data points in multivariable real spaces, spheres or manifolds is carried out (sometimes called kriging). A particular advantage of most of these algorithms is that they can be used when data are arbitrarily distributed (``scattered''). They minimise certain (semi-)norms (``energy norms'') directly related to the form of the splines (or otherwise) that span the linear approximation spaces. In many instances, these semi-norms can be identified via the generalised Fourier transforms of the kernels (B-splines, radial basis functions, thin-plate splines, multiquadrics etc). They are often Euclidean norms of certain partial derivatives of the approximants, so the approximations take place in Sobolev space. Especially when smoothing is wished for, these minimisation features render the described methods so attractive. In this article, these concepts are generalised to manifolds on which the approximations and approximants are defined. The simplest, quite well-studied cases are those of spheres, where the distances used in the arguments of the radial basis functions, are geodesics, but this paper is more general. The mentioned minimising properties are now guaranteed by minimising functions with constraints by penalty functions. The objective functions being minimised are the aforementioned energies defined by intrinsic differential operators on the manifolds, the constraints are the interpolation conditions. Some numerical examples using tensor-product splines as bases are presented to demonstrate the efficiency of the algorithms.