[vtkusers] PCA statistics change
david.thompson at kitware.com
Fri Feb 16 12:12:06 EST 2018
> We are seeing this too when running our company's own tests built against VTK master.
> Is this change considered a breakage of vtkPCAStatistics? Will you be changing things to make vtkPCAStatistics behave as it used to?
No. Eigenvectors reported by singular value decomposition are not totally constrained:
1. Vectors may have their direction flipped.
2. In the case of eigenvalues with identical or nearly identical values (such that differences in floating point calculations may change their ordering) the reported order of vectors may vary.
We have switched the implementation of SVD we use (from an old alglib to a new Eigen), so these differences are expected and in some cases preferable. You should adjust baselines appropriately.
However, if you see a case where the SVD is failing (i.e., Nan/Inf values where there were none before), that is a bug we should fix.
More information about the vtkusers