Angular sort in tufted cover impacts eigenvalue calculation

[svorderw]

Hello!

I have been looking at your project to calculate eigen information on pointclouds. During the construction of the tufted cover there is an angular sorting of the faces around the non-manifold edge. In the corresponding paper you mention that the Laplacian should be robust to the orderings, but I observe a significant difference in the eigenvalues if I compare the construction with and without angular sort. The difference is shown in the following picture.

I have produced a "minimal failing example" for reproducibility on a branch. The branch contains the addition of the Eigensolver and a control for the sorting. If you give me access rights, I can push the branch to your project, and you can try it yourself.

I was wondering if you had any insight why I observe these differences. Is it a floating-point issue? Is the tufted cover not unique and depends on ordering? @nmwsharp

Many thanks in advance,
and let me know if there is a better channel do discuss this,
Sophia Vorderwuelbecke

[svorderw]

@sitic