Approximate radical for clusters: A global approach using Gaussian elimination or SVD (Q2468363)

This paper is about approximation of roots of a polynomial ideal \(I\subset {\mathbb C}[{\mathbf x}]\) in \(m\) variables \({\mathbf x}=[x_1,\dots,x_m]\). The authors introduce a concept ``matrix of traces'' to a zero dimensional ideal \(\tilde I\) which is considered as a numerical variant of \(I\). The concept is based on Dickson's lemma about the Jacobian radical of a finite dimensional associative algebra, and extended to an approximate radical of \(\tilde I\) with zero clusters, i.e., the approximate radical ideal has exactly one root in each cluster for sufficiently small clusters. Instead of locations of the roots in the clusters, the coefficients of the system of polynomials defining \(\tilde I\) are used in the proposed method so that it works simultaneously for all clusters globally and reduces the problem to the computation of the numerical nullspace of the matrix of traces, which itself can be computed efficiently from the generating polynomials of \(\tilde I\). Gaussian elimination and singular value decomposition are proposed for computing the numerical nullspace of the matrix of traces. It is demonstrated that this problem can be reduced further to finding eigenvalues of smaller matrices with separated eigenvalues. In the univariate case it leads to a simpler approximate square-free factorization algorithms.