heat icon indicating copy to clipboard operation
heat copied to clipboard

Published 20 hours ago • verified publisher icon helmholtz-analytics

Provide a matrix gallery for testing

Open d1saster opened this issue 5 years ago • 1 comments

Related This is mostly related to the current LA efforts.

Feature functionality We need some functions that can generate matrices of arbitrary size that have certain unique features that allow us to test the LA routines in extreme cases.

Additional context For example the Parter and Grcar matrices would be nice to test the SVD and eigenvalue functionality. Cf. also https://blogs.mathworks.com/cleve/2019/06/24/bohemian-matrices-in-the-matlab-gallery/ for a similar implementation in Matlab.

d1saster avatar Oct 01 '19 14:10 d1saster

@d1saster can we break this in smaller tasks?

ClaudiaComito avatar Apr 01 '22 09:04 ClaudiaComito

Bumping this up to the surface again and tagging @mrfh92 .

ClaudiaComito avatar Nov 10 '22 04:11 ClaudiaComito

Regarding the SVD i think the following test cases could be of interest:

  • matrices that are challenging for SVD-algorithms: I think the already implemented Parter matrix should belong to this class due to their clustered singular values (?)
  • matrices with known SVD and "low rank": in particular for testing truncated SVDs such as hSVD
  • at least one matrix that is definitively not "low" rank: for testing a future full SVD

mrfh92 avatar Nov 14 '22 08:11 mrfh92

I have added the second bullet point from above (randomly generate low-rank matrices with known SVD) to the hSVD-branch #1041

mrfh92 avatar Nov 14 '22 11:11 mrfh92

In the future, it will be easier to manage the topic of this issue as part of other issues (e.g. #1175) in which particular test cases have to be designed either for testing or benchmarking.

Therefore closed within #1109

mrfh92 avatar Aug 17 '23 11:08 mrfh92