Verified computation of the matrix exponential (Q2631982)

There are many traps to numerically computing the exponential of an $n \times n$ complex matrix $A$ (see, for example, [\textit{C. Moler} and \textit{C. Van Loan}, SIAM Rev. 45, No. 1, 3--49 (2003; Zbl 1030.65029)]). This paper analyzes two methods to compute interval matrices for $\exp (A)$. If $A$ can be diagonalized, and we have an approximate relation $A \approx X^{ -1}\Lambda X$ for some diagonal matrix $\Lambda $, then an approximation to $\exp (A)$ can be easily calculated since $\exp (\Lambda )$ is a diagonal matrix. In this case the author shows how to obtain an interval matrix for $\exp (A)$ from floating point estimates of $X$ and $\Lambda $. On the other hand, in the case that $A$ is not diagonalizable, the author proposes using a Jordan decomposition of $A$ [\textit{B. Kagström} and \textit{A. Ruhe}, ACM Trans. Math. Softw. 6, 398--419 (1980; Zbl 0434.65020)] in a similar way. The two methods are compared with two other computational methods for accurately calculating $\exp (A)$ [\textit{P. Bochev} and \textit{S. Markov}, Computing 43, No. 1, 59--72 (1989; Zbl 0685.65035)] and [\textit{J. Rohn}, ``VERSOFT: verification software in MATLAB/INTLAB'', \url{http://uivtx.cs.cas.cz/rohn/matlab}] but the comparisons seem inconclusive.