Super-smooth cubic Powell-Sabin splines on three-directional triangulations: B-spline representation and subdivision (Q2222184)

In this paper, the famous Powell-Sabin splines are considered; they are on a three-directional triangulation, becoming 6-directional after refinement. Much like box-splines for example they can be viewed as shadows of multidimensional polytopes in \(\mathbb{R}^2\), but the splines considered here do not generate shift-invariant spaces as the box-splines usually do. (But here is a regular, symmetric, 3-directional triangulation, too, and there are other relations with box-splines mentioned in the article.) In practice, because they form convex partitions of \(1\) like the classical, univariate B-splines, they could be used for quasi-interpolation with splines, a most useful method for the approximation of multivariable functions when no interpolation (collocation) properties are needed, and that is also available for other kernel functions such as radial basis functions. They are extended here into so-called super-splines, i.e., splines that are piecewise polynomials but not stitched together to be once continuously differentiable but twice. They are cubic piecewise two-dimensional polynomials. Their relation to simplex splines is pointed out, too (the super-splines of the kind derived now from the Powell-Sabin splines are multiples of simplex splines). Both B-spline representations of the ordinary Powell-Sabin splines as well as those of the super-splines are considered. The authors derive conditions for twice differentiability over the whole domain as well, and explicit, detailed computations for all coefficients for the various splines are provided, and so in particular the dimensions of the splines spaces are given. The B-spline representations are particularly attractive because they are compressed through making use of the particular symmetries of the triangulation. A particular important aspect is proved in Proposition 3 of this article, namely the stability of the B-spline representations, furthermore linear independence (following of course from stability) and local support are guaranteed. Also, all approximation orders known before for these splines are kept. In fact they are improved upon because the splines on triangulations generated by the cubic B-splines are only \(C^1\) and not \(C^2\) (giving cubic precision instead of quadratic precision).