…or how we replaced the heuristics of a code by a rigorous backward error analysis.
We all know the procedure of developing a numerical method/algorithm:
- get it running,
- perform experiments,
- optimize the algorithm/code.
The emphasis is always on step 1. Unfortunately, step 3 is often not pursued because <<the PhD student left>, <the focus shifted>, <funding ended>,… >. In 2009 there existed a code for computing the action of the matrix exponential based on Leja method, see , , . However, the code was based on some heuristics and had room for optimization.
The main idea of the Leja method is to compute
Figure 1: 15 Leja points in the interval [-2,2].
In order to allow matrices with a large norm one takes advantage of the functional equation of the exponential to introduce scaling steps. In the end the interpolation essentially depends on three parameters, the number of scaling steps, the location/sequence of the interpolation points, and the degree of interpolation. The remaining question is: How do you select these parameters for a specific input?
The code from 2009 had the following heuristics included
% three magic numbers... maxm = 150; minm = 30; m = max (2.6 * ceil (gamma), minm); nsteps = ceil (m / maxm);
These magic numbers from above determined all of the necessary parameters to perform the interpolation. The
gamma in the formula describes an ellipse that encloses the field of values of the input and
nsteps corresponds to the number of scaling steps. The first attempt to get rid of the heuristics replaced the above selection procedure, see . The selection of the parameters was then based on the interpolation behaviour of a 2-by-2 matrix, with entries reflecting a bound of the extrema of the field of values of
In the preprint , the heuristics were finally eliminated. Yet again, the code was redesigned, based on a rigorous backward error analysis. The main idea is to interpret the result of the interpolation as the exact solution of a perturbed problem. The backward error analysis gives rise to a selection procedure of the essential parameters that can minimize the cost of the algorithm.
As a result, the code is simpler and has no longer any restrictions on the input matrix. Nevertheless, there is still room for optimization. Numerical experiments show that the method can have performance losses for nonnormal matrices. We are currently working to fix this glitch but that is for another blog entry.
The code is available at the Numerical Analysis Innsbruck project page.
 Caliari, M., Kandolf, P., Ostermann, A., Rainer, S., 2014. Comparison of methods for computing the action of the matrix exponential. BIT Numer. Math. 52 (1), 113-128.
 Caliari, M., Kandolf, P., Ostermann, A., Rainer, S., 2015. The Leja method revisited: backward error analysis for the matrix exponential. Preprint, arXiv:1506.08665.
 Caliari, M., Vianello, M., Bergamaschi, L., 2004. Interpolating discrete advection-diffusion propagators at Leja sequences. J. Comput. Appl. Math. 172 (1), 79-99.
 Huisinga, W., Pesce, L., Kosloff, R., Saalfrank, P., 1999. Faber and Newton polynomial integrators for open-system density matrix propagation. J. Chem. Phy., 110, 5538-5547.
 Reichel, L., 1990. Newton interpolation at Leja points. BIT Numer. Math. 30 (2), 332-346.
Author: Peter Kandolf, firstname.lastname@example.org