Tanaka-san found that LSF's widths becomes better when he divides the widths by 2.354. The black lines in the image below are pfsMergedLsf:

The black lines in the image below are pfsMergedLsf, whose widths are divided by 2.354:

The origin of pfsMergedLsf can be traced back to reducedExposure.py's config (gaussianLsfWidth). The values in this list are FWHM. They are input to GaussianLsf's constructor without being converted to standard deviations.

Could you review this PR?

I made Gaussian fits to multiple arc lines. See figs here. Although they can/should be cleaned up a bit, the running median shown as the red points suggest that the line width is roughly constant. The horizontal line is the global median after excluding bad fits (e.g., poitns at sigma=0.1, which is the initial guess). r3 and m3 look bad because these are from the Apr 2023 data.

The spectral resolution is dependent on fiberId. The dependency is different for different arms/grating, and the p-p variation is about 0.005-0.01nm. We need to account for this, but for now, I propose to adopt a constant LSF with sigma=0.081nm for b, 0.109nm for r, and 0.059nm for m. I will look at the n-arm and update this ticket.

There is no science arc for the n-arm in Run 11. There are some data in Run12 and I made an attempt to reduce them, but I am missing most of the lines at lambda>1000nm. I am not sure why, but I think I will wait for new calibs from this run and come back later.

I propose that we go with a specification resolution listed at https://pfs.ipmu.jp/research/performance.html for the n-arm for now, i.e., sigma=0.109 nm. I will measure the resolution once we have better calibs from the July 2023 run and file a separate ticket to update the number.

I revised the numbers in the ticket branch.

Merged. Thanks for reviewing.