A quadratically convergent class of modifications for Kovarik's method (Q1043698)

The author considers modifications of \textit{Z. Kovarik}'s method [SIAM J. Numer. Anal. 7, 386--389 (1970; Zbl 0217.21501)] that is given by the iteration \[ A_0 = A, \;K_k =(I-A_k A_k^T)(I+A_k A_k^T)^{-1}, \;A_{k+1} = (I+K_k)A_k, \;k \geq 0\text{ with }A \in \mathbb{R}^{m \times n},\;m \leq n. \] The iteration converges quadratically to \(A_\infty = [(AA^T)^{\frac{1}{2}}]^+A\), in which \(A^+\) is the Moore-Penrose pseudo inverse of \(A\), and the rows of \(A_\infty\) are approximatively orthogonal. \textit{D. Petcu} and \textit{C. Popa} [Int. J. Comput. Math. 82, No.~10, 1235--1246 (2005; Zbl 1085.65038)] approximated \((I+A_k A_k^T)^{-1}\) with \((I+0.5A_k A_k^T)\) and proved the linear convergence of this modification. Considering an extended parameter class, the author [Bull. Belg. Math. Soc. - Simon Stevin 15, No.~2, 377--384 (2008; Zbl 1151.65036)] proved that the modification of Petcu-Popa is quadratically convergent. In the present paper, the author studies a new quadratically convergent variant of Kovarik's method using the singular value sequence convergence of the iterates \(A_k\). The new algorithm contains the variant of Petcu-Popa, and numerical studies for various matrices of the moderate order \(n = 5, 10, 20, 50\) show that the number of iterations is considerably lower than that of others in the case of ill-conditioned matrices.