Polynomial approximations for the matrix logarithm with computation graphs (Q6966756)

This interesting paper studies polynomial approximations for the matrix logarithm with computational graphs. A logarithm of a square matrix \(\mathbb C^{n\times n}\) say \(X\) satisfies the matrix equation \(A=\exp(X)\). As to the existence of the matrix \(X\), it is well known that if the matrix \(A\) has no eigenvalues on the negative real line, \(\log(A)\), called the principle logarithm, is the only solution of the matrix equation \(A=\exp(X)\) whose spectrum lies in the strip \(\left\{z\in \mathbb C:\,\pi<\mathrm{Im}(z)<\pi\right\}\). It is well known that \(\log(A)\) is used as a tool across a wide variety of fields for example quantum chemistry, mechanics, biomolecular dynamics, machine learning and many more. How to compute the matrix logarithm usually involves a combination of a Padé approximation, sometimes accompanied by a Schur decomposition in conjunction with an inverse scaling and squaring. In the paper under review, the authors show that the amount of work used in left matrix division and matrix-matrix multiplications can be substantially reduced, by using a graph based representation of an efficient polynomial evaluation scheme. A technique to analyze the rounding error is proposed, and a backward error analysis is adapted. The authors provide many simulations illustrating their results.\N\NThe paper is well written with a good set of references.