[core] Trimmed performs badly

[Yey007]

Trimmed seems to perform way worse than it should. Are we doing something wrong?

[Yey007]

Jess found the issue! Or at least one of them. We are calculating the centroid for trimmed_b incorrectly. There can be duplicates in the array, skewing the result.

[Yey007]

Hmmm... I investigated a bit. I tried to fix this but the performance was way worse. It is possible that our original code is actually correct. If we treat the SVD method as something that simply optimizes the distance cost between pairs of points, fixing the centroid actually breaks the guarantees of the proof, it doesn't restore them. So maybe there is another cause?

[Yey007]

But if this is true, that would actually imply that vanilla is wrong. Could this be the case?