A dimension series for multivariate splines (Q2277525)

The study of piecewise polynomial functions over a polyhedral subdivision \(\Delta\) of a region in d-Euclidean space, involves not only algebra and geometry, but also analysis and combinatorics. The simplest case being a simplicial subdivision, one already meets difficulties when considering a planar simplicial subdivision, when global smoothness conditions are imposed on the functions. Problems also occur in the continuous case, with general polyhedral subdivisions in two dimensions. Piecewise polynomial functions (also called multivariate splines) turn up in a variety of contexts (differential equation approximation, interpolation of scattered multidimensional data, computer-aided geometric design, computer graphics, robotics). Let \(C^ r(\Delta)\) denote the set of all piecewise polynomial functions over \(\Delta\) which are globally smooth of order \(r\) and let \(C_ k^ r(\Delta)\) be the subset consisting of functions, only involving polynomials of degree at most \(k\). As \(C_ k^ r(\Delta)\) is a vector space over the reals, one can try to determine the dimension and to compute a basis for it. Problems arise from the fact that not only combinatorial but also geometric properties influence the result. Though this leaves no hope for certain individual invariance properties, there is some invariance as a whole for the \(C_ k^ r(\Delta)\). This paper is concerned with the generating function for the dimensions \(\sum_{k\geq 0}\dim_{{\mathbb{R}}}C_ k^ r(\Delta)\lambda^ k \). It is shown that under mild restrictions this series (considered as an element of the ring of formal power series in \(\lambda\) over \({\mathbb{Z}})\) always has the form \(P(\lambda)/(1-\lambda)^{d+1}\), where P(\(\lambda\)) is a polynomial in \(\lambda\) with integral coefficients and \(d=\dim (\Lambda)\). Given some other invariance results, one may conclude that, although the individual numbers \(\dim_{{\mathbb{R}}}C_ k^ r(\Delta)\) may change when the geometry of \(\Delta\) changes, the collection as a whole is rather more controlled. The calculation of P(\(\lambda\)) can be done, using Gröbner basis techniques.