Open dkirkby opened 4 years ago
The non-normalization of R is a "feature" of the spectroperfectionism math from Bolton & Schelgel 2009, roughly analogous to a weighted- vs. unweighted- fit and how they treat low-weight outlier points. I think you will find the largest Rsum outliers to be near noisy sky lines, which is why it is a bigger effect in Z than B.
The normalization comes from whether you sum over rows of columns in Bolton & Schelgel 2009 equations 11 and 12. If you swap the normalization axis (ex2d.py line 532) you can get something that does conserve flux in R with the somewhat surprising (but perhaps better) side effect of oscillating noise: neighboring bins of the extracted spectra can have radically different noise levels in a ringing-like oscillatory manner. But at least R is a real convolution matrix then. We explored that a bit years ago (see section 4.2.2 of DESI-0890) but didn't converge on a detailed study and specter currently implements the normalization as defined in the original paper.
All that being said, the effect you show here is larger than I expected, and perhaps more work could be done on
The FRAME resolution matrix documented here does not appear to be normalized in the SV0 reductions I have checked so far.
Specifically, I expected the sum of the 11 resolution function bins tabulated for each (fiber, wlen) to be close to 1, except possibly for some leakage beyond the 11-bin window.
The actual sums have a mean close to 1 but a lot of dispersion that depends mostly on the wavelength (rather than the fiber number). There might also be some correlation of R > 1 with cosmics. The spread in sums is largest in z and smallest in b.
To demonstrate the issue, I used:
which produces: