From Bareiss' algorithm to the stable computation of partial correlations (Q1823641)

The paper is concerned with the problem of stable computation of sample partial correlation coefficients. A review of some known methods as well as a new algorithm are presented and their mutual relations are discussed. The authors start with the algorithm of \textit{E. H. Bareiss} [Numer. Math. 13, 404-424 (1969; Zbl 0174.204)] for the solution of linear systems with Toeplitz coefficient matrix. This algorithm computes a set of sample partial correlations of a given data matrix A, but it requires the formation of \(B=A^ TA\) and thus suffers from loss of numerical accuracy. The data flow graph of Bareiss' algorithm suggests a way of computing these partial correlations directly from A, method worked out by \textit{G. Cybenko} [Math. Comput. 40, 323-336 (1983; Zbl 0539.93089)]. However, the latter method has higher operation count and low potential for parallelism. The authors then describe an extension of Bareiss' algorithm to general coefficient matrices, whose normalized version reduces to the hyperbolic Cholesky algorithm when the initial matrix is positive definite. The authors propose a version of this method that is superior to Bareiss' and Cybenko's with regard to the combination of numerical accuracy, operation count, and inherent parallelism. Finally, this approach is extended to compute partial correlations with arbitrary sets of given variables.