A contribution to perturbation analysis for total least squares problems (Q2360675): Difference between revisions

The total least squares (TLS) solution \(x\) of an overdetermined system \(Ax\approx b\) solves \(x\) from \((A+E)x=b+f\) with \(\|A~f\|_F\) minimal to make the system solvable. The condition number says how sensitive \(x\) is to perturbations in the data. Three condition numbers for TLS that appeared in the literature [\textit{L. Zhou} et al., Numer. Algorithms 51, No. 3, 381--399 (2009; Zbl 1171.65031)], [\textit{M. Baboulin} and \textit{S. Gratton}, SIAM J. Matrix Anal. Appl. 32, No. 3, 685--699 (2011; Zbl 1242.65087)], and [\textit{B. Li} and \textit{Z. Jia}, Linear Algebra Appl. 435, No. 3, 674--686 (2011; Zbl 1228.65058)] are reviewed and shown to be identical. Here alternative computable and more accurate bounds of the form \[ \frac{\|\delta x\|_2}{\|x\|_2} \leq \kappa_b \frac{\|\delta b\|_2}{\|b\|_2} + \kappa_A \frac{\|\delta A\|_2}{\|A\|_2} \] are obtained and tested numerically. It is assumed that the problem is generic, i.e.\ \(\tilde{\sigma}_n>\sigma_{n+1}\) where \(\tilde\sigma_n\) and \(\sigma_{n+1}\) are the smallest singular values of \(A\) and \([A~b]\), respectively.