PardisoSimplicialLDLT vs Eigen::SimplicialLDLT about maks HOT 3 CLOSED

Ok I found the culprit when using Eigen solver

from maks.

Hi,
Thanks for the feedback. The choice between MKL Pardiso and Eigen::SimplicalLDLT was purely based on performance since the code was initially implemented on an i7 and there was a considerable difference between the two. Regarding accuracy, it works best in sparse graphs with a moderate noise level in the rotations (smallest 3 eigenvalues of the Rotations Laplacian close to zero).

You may want to try different eigenvalue targets (sigma) of the symKrylovSchurLDL function, which is set at -1e-6. Tweaking this may lead to instability in the LDL decomposition in some cases, however.

from maks.

Hi, thanks for your answer it is very appreciated!
I made a cmake file for this project, taking advantage of "modern and robust" detection of recent version of MKL (the versions with the "oneAPI" branding) that fallback to Eigen::SimplicialLDLT if MKL is not found. I think it could be a good way for scientists and practitioners to test MAKS if they do not want to mess with MKL. Are you willing to integrate this to the repository?

from maks.