Implement BRIGHT program exposure time model

[dkirkby]

This issue is to discuss and converge on a first implementation of a BRIGHT program exposure time model. As a reminder, we currently use the DARK/GRAY model for BRIGHT exposures.

I am following the same approach already used for the DARK/GRAY model here, which has two components:

1. Determine what exposure time is necessary under nominal conditions to achieve an acceptable redshift success rate for the targeted BGS sample. Note that the nominal conditions are an arbitrary but unambiguous reference for normalization purposes and not intended to be typical. They are defined in machine-readable format here:

   nominal_conditions:
   # Moon below the horizon
   seeing: 1.1 arcsec
   airmass: 1.0
   transparency: 1.0
   EBV: 0.0

   I believe @moustakas is working now on this component and I hope he can post updates here. We currently assume the answer is 300s, but if the answer is much different we need to know soon.

2. Determine the actual exposure time necessary to achieve equivalent redshift success under different lunar and twilight observing conditions. For this study, we use the SNR predicted by specsim simulations as a proxy for "equivalent redshift success", rather than running redshift fits, in order to study a wide range of observing conditions efficiently. I am working on this component, using this earlier DARK/GRAY program study as a template. The areas where I need input are:

   • Is the new twilight model in specsim close enough to reality to base this study on? (@segasi has already made key contributions here).
   • What are the threshold BGS targets that the exposure time should be calibrated for? Specifically, I need a sample of SEDs to simulate and will then determine what exposure time achieves a target median SNR for this sample, as a function of lunar + twilight observing conditions. I guess these should sample the expected n(z) and have r = 19.5, or perhaps r=20? @moustakas is working on this.
   • How should the SNR of the simulated spectra be calculated so that it is a good proxy of "equivalent redshift success". For the DARK/GRAY program, we use the integrated SNR of the [OII] doublet. By default, I will use an r-band weighted SNR, but better ideas are welcome (perhaps SNR near the 4000A break?)

[julienguy]

For the last point, one could use one of the redrock simulation runs. For several definitions of SNR (r-band , around 4000A restframe ...):

• compute SNR on spectra of cframe files
• compute correlation coefficient of SNR with the redshift error reported by redrock

[moustakas]

I've generated a set of reference spectra (No 2, bullet two in @dkirkby's original post) for the BGS sample. See https://github.com/desihub/desisim/pull/285 or just the notebook -- https://github.com/desihub/desisim/blob/bgs-refspec/doc/nb/bgs-reference-spectra.ipynb

The spectra themselves are at NERSC in /global/project/projectdirs/desi/spectro/sim/bgs-refspec/bgs-refspec-v1.0.fits

Comments are welcome.

[dkirkby]

Here is a first look at the reference spectra provided by @moustakas (details in desihub/desisim#285).

The redshift distribution is reasonable (with some spikes due to realistic clustering). The magnitudes cover a range, so I am rescaling them all to DECam-r = 19.5 to define the "threshold" sample.

Here are the fiberloss fractions calculated for each target in the sample:

These are generally low compared with the DARK program targets because of the larger galaxy sizes. The strong correlation with redshift is because I scale disk and bulge sizes with DA(z), so closer galaxies have less overlap with the fiber:

disk half-light radius = 4.7" DA(0.1) / DA(z)
bulge half-light radius = 1.3" DA(0.1) / DA(z)

[moustakas]

I've begun characterizing the redshift efficiency of BGS targets under "nominal" observing conditions -- see https://github.com/desihub/desisim/pull/286. This work is still preliminary because it's only based on 1000 spectra and also because the fiberloss fractions are not totally up to date with the ongoing work by @dkirkby (although I'm not sure in what sense). But it nevertheless gives a rough idea of how we should expect to perform with 5-minute exposures.

Comments are welcome here or at https://github.com/desihub/desisim/pull/286.

[dkirkby]

The code and notebook used to develop this model is now checked into desihub/surveysim#56

[schlafly]

We've defined effective exposure times for the BRIGHT program, which is the extent to which desisurvey cares about this issue. I'm inclined to close. There's a PR to implement a better sky model in surveysim https://github.com/desihub/surveysim/pull/79 but that doesn't touch desisurvey directly.