A study of Schröder's method for the matrix \(p\)th root using power series expansions (Q2287864)

Fix an integer \(p\geq2\) and for each \(m>0\) define \(T_{m}(t)\) to be the Taylor polynomial of degree \(m\) approximating \((1-t)^{1/p}\). The family of Schröder's methods for computing the \(p\)-th root \(a^{1/p}\) of a scalar \(a\) consists of the iterative schemes \(\mathcal{S}_{m}:x_{0}=1\) and \(x_{k+1}:=x_{k}T_{m}(1-ax_{k}^{-p})\) for \(k\geq0\). Both Newton's method (\(m=1\)) and Chebyshev's method (\(m=2\)) lie in this family. Applying Schröder's method with \(1-z\) in place of \(a\) we obtain a power series in \(z\) for \(x_{k}\) for \(k=1,2,\dots\). In the first half of the paper it is shown that the coefficients of these power series are all positive for all numerical values of \(p\), \(m\) and \(k\) (Theorem 6). This confirms an earlier conjecture (see [\textit{K. Ziȩtak}, J. Comput. Appl. Math. 272, 468--486 (2014; Zbl 1294.65050)]). The second half of the paper considers the use of Schröder's method for the computation of the \(p\)-th root of a real or complex matrix \(A\) and the rate of convergence of the method. The scheme \(\mathcal{S}_{m}\) converges to \(A^{1/p}\) whenever no eigenvalues lie on the negative real axis and all eigenvalues of \(A\) lie in the unit disc \(\mathcal{D}\): \(|z| <1\). More generally \(\mathcal{S}_{m}\) converges if the eigenvalues of \(A\) all lie in the set \(\mathcal{F}:=\left\{z\mid |T_{m}(z)| <1\right\}\); note that \(\mathcal{D\subseteq F}\). It is proved (Theorem 9) that \(\mathcal{S}_{m}\) converges to \(A^{1/p}\) if all eigenvalues of \(A\) lie in the connected component \(\mathcal{E}\) of \(\mathcal{F}\) containing \(\mathcal{D}\). There are also more specific theorems on the convergence for \(Z\)-matrices and \(M\)-matrices.