nbodykit Lanczos windows do not perform well

Take a look at the cookbook recipe:

http://nbodykit.readthedocs.io/en/docs/cookbook/interpolation-windows.html

CIC + compensation, TSC + compensation, and then the DB kernels all perform reasonably well (accurate up to about 1/2 Nyquist, which is in line with my intuition from previous tests). Lanczos is quite bad though, maybe there is a compensation term we should add?

Aug 01 '17 20:08 nickhand

I added a bunch of higher order lanzcos and they still look bad.

Here is an interesting article (potentially) https://arxiv.org/abs/1401.2636

Aug 02 '17 00:08 rainwoodman

Huh, interesting. So maybe there is not a bug and they are just bad?

Aug 02 '17 04:08 nickhand

I guess it just means we will need to put in a compensation window for them (which probably doesn't come with a closed form in 3d).

They might be useful for other things. We'll see.

Aug 02 '17 05:08 rainwoodman

The sym20 kernel performs better than db20, which wasn't flat at large scale.

Symmetrizing the center of mass may have improved the large scale power?

tsc here was compensated.

Aug 05 '17 00:08 rainwoodman