An error analysis of the modified scaling and squaring method (Q2468911)

The paper develops \textit{N. J. Higham}'s scaling and squaring method for the matrix exponential [SIAM J. Matrix Anal. Appl. 26, No. 4, 1179--1193 (2005; Zbl 1081.65037)] of linearly stiff ordinary differential equations \( y'=\Lambda y + f(y)\), where \(\Lambda\) is an \(N\times N\) constant matrix. The author investigates a modified scaling and squaring method for the computation of \(\varphi\)-functions \[ \varphi_n (\Lambda h) =(\Lambda h)^{-n}\left( e^{\Lambda h} - \sum_{k=o}^{n-1}\frac{(\Lambda h)^k}{k!}\right). \] The main operations of the algorithm are: dividing the matrix \(\Lambda h\) by some \(s\)-power of two (scaling), the diagonal Padé approximation of \(\exp(\Lambda h/2^s)\) and \(\varphi_n(\Lambda h/2^s)\), and squaring the obtained approximants in \(2^s\) degree. The propagation of the rounding error in the squaring process is investigated. Numerical realizations of the proposed algorithm for famous test matrices are performed.