Interpolating splines on graphs for data science applications (Q778025): Difference between revisions

Interpolation in many variables is an established tool in numerical analysis and approximation theory. In this context, it is especially important to identify properties of Lagrange functions that, given some data points, are each 1 at just one point and vanish at all the others. (The idea comes, of course, from univariate Lagrange interpolation by polynomials.) This is because (i) they allow a particularly useful, easy formulation of the approximant , and (ii) the study of their properties give important insights into the qualities of our method -- localness (reflected by decay of Lagrange functions), stability (estimating Lebesgue numbers), convergence properties (reproduction of some finite dimensional subset, usually polynomials of some degree). Good examples are based on Lagrange functions based on kernels, radial basis functions, polyharmonic splines (note that here, for example, localness is guaranteed by the Lagrange functions' exponential decay). In this papers the authors pursue exactly that track, but the Lagrange functions and their 0/1 values are defined on \textit{graphs} (defined by edges and vertices) rather than points. All other features of the \textit{Ansatz} described can be carried through in the same way as mentioned above: easy formulation of interpolants on \textbf{graphs} instead of points, localness etc. This is established in much detail, compared with the classical approaches and results from radial basis function approximation, and applications in useful, modern fields like machine learning are given.