Polynomial least squares fitting in the Bernstein basis (Q5962279): Difference between revisions

Based on the fact that in the least squares method used in the polynomial fitting, the rectangular Vandermonde matrix corresponding to the nodes, associated with the overdetermined system of the normal equations is usually an ill-conditioned matrix and the use of QR factorization doesn't solve too much of this problem, the paper is focused on the idea to take, instead of the classical monomial basis, a polynomial basis which leads to a matrix of the nodes with smaller condition number than the Vandermonde matrix. Thus, the Bernstein basis of polynomials, widely used in computer aided geometric design is considered. Moreover, the nodes are taken strictly increasing between 0 and 1. To solve this type of polynomial fitting, the authors use some results related to total positivity of the rectangular Bernstein-Vandermonde matrix, defined for stricly increasing nodes, between 0 and 1. The bidiagonal factorization is realized as starting point for the computation of QR decomposition for the Bernstein-Vandermonde matrix of the nodes, then a triangular system is solved by using the bidiagonal decomposition of the R factor. In this order, the authors propose an efficient algorithm with high accuracy dedicated to solve the problem of polynomial regression in the Bernstein basis. Finally, two numerical experiments illustrate the efficiency of the proposed algorithm showing that the accurcay is keeped when the condition number of the associated Bernstein-Vandermonde matrix increases.