Adaptive Lanczos methods for recursive condition estimation (Q1186619): Difference between revisions

The authors consider adaptive estimation of the condition number of a matrix varying in discrete time like in recursive least squares and signal processing computations. For this aim, the Cholesky factor of the covariance matrix is updated or ``downdated'' (rather than up- or downdating the covariance matrix itself) to avoid numerical difficulties like loss of symmetry or positive definiteness of the matrix. Connected with the up- and downdating procedures is a Lanczos based estimation of the extreme singular values and vectors -- using the known fast convergence of these extreme singular values during the Lanczos method. The proposed method is compared with estimators by \textit{C. H. Bischof}, \textit{J. G. Lewis}, and \textit{D. J. Pierce} [SIAM J. Mat. Anal. Appl. 11, No. 4, 644-659 (1990; Zbl 0726.65038)] and by \textit{D. J. Pierce} and \textit{R. J. Plemmons} [ibid. 13, No. 1, 274-291 (1992; Zbl 0747.65029)] (where the latter is cheaper and faster for \(n\geq 7\)). The new estimator is shown by numerical experiments on several data sets to be more reliable (if not counting a small initial time interval).