Probably not going to play in this sandbox for the first half of this week - Real Life chores and obligations press.
- As @weightt_an points out, there are (probable) distortions in how I’m rendering a eigen_vector compared to how his script renders it out. My suspicions run toward the initial calculation and rounding of N that I do. That, and the padding out of the offset diagonals when it comes to boundary policy. Currently I use periodic. That actually adds coefficients to the matrix not in the original unrolled image. Plausible, then, that in effect I make a slightly different seed well than what @weightt_an does. Solution to that is to switch to Dirichlet Boundary policy - i.e. consider off-image pixels as black. That adds no coefficients not already present in the original unrolled image.
- My diagonalizer can put up to five huge images on the image stack at one time. These are related to consumption of up to a 16 GB bulge. I like @Reptorian’s diagonalizer because it writes to one large image from smaller sources. Still think it is worth some time and energy to track down what’s going on with that approach because it is the more memory friendly one.
- Even with a defective diagonalizer, it is not that far away to hook together a G’MIC only pipeline. If somebody else doesn’t beat me to it, that would be my next destination on my a periodic bus trips to Schrodinger’s Well. Have fun!