Niklas Z

@WarrenWeckesser You're welcome! **Edit** **⚠️ The following solution is wrong, but it's fixed in the next comment! ⚠️** I did some research and found the publication Turley R. S., *Polynomial...

Maybe I should not have fried my brain 🍳🧠 trying to do banded QR-decomposition before writing the fix. **I completely forgot about `pos` and an actual `assert` in my previous...

@WarrenWeckesser Very nice! `delta` indeed slipped through. I'm happy that this already does the trick 🎉

I actually found a tiny little improvement that can be made, nothing really dramatic and more a matter of preference I guess. The computation of the polynomials could be streamlined...

@WarrenWeckesser will you implement and make a PR for this? For me, the hurdle of making a fork of SciPy and building SciPy on Windows is right now a bit...

Hi @WarrenWeckesser I think I already found how we can align this. I forked your repo and will see how we can improve this 👍

Hi @WarrenWeckesser I finally got it to work and now I actually obtain maximum floating point accuracy (at the cost of some performance however, maybe you have an idea how...

I hope I wrote everything correctly in terms of equations and text. Sorry, but after writing and parsing all of this, everything blurred when I read it at some point....

I checked and it seems that `fsum` does not make that much (if any) of a difference. Therefore, the implementation of the correction can be simplified: ```python def _correct_pinv_row_normalized_polyvander( pinv_row_idx:...

No worries! I tried to tackle the hilarious runtime in the meanwhile. The simple NumPy-polynomial and the super stabilised version with pseudoinverse correction both take ~1 ms for `polyorder=10, window_length=51,...