BGS redshift efficiency under nominal (moon down!) conditions

[moustakas]

The notebook in this PR builds on lots of previous work by me, @akremin, @sbailey, and @julienguy.

Here, I attempt to quantify the redshift efficiency of BGS targets under "nominal" conditions (i.e., moon-down, 5 minute exposures). The results are just based on 1000 simulated spectra so far, so there's small-number statistics to contend with, but the results are nevertheless interesting.

This work is closely related to desihub/desisurvey#77, so input from @dkirkby would be welcome.

[dkirkby]

This looks good! I haven't gone through in detail, but it looks like you have simulated an ~unbiased sample of 1K BGS bright+faint targets and run them through redrock (which is taking ~3mins per spectrum on 4 cores?)

Is the ~linear drop in efficiency with r-mag > 19 expected? Efficiencies < 50% for the BGS-faint program are sobering.

I didn't expect the efficiency to be flat in redshift, due to the improvement in fiberloss with DA(z), but perhaps you aren't including this effect or there is a compensating effect?

Can you plot the efficiency vs median(SNR) in each camera, since that's closest to the "redshift success" proxy I am using for the exposure time model.

There is a canned plotting routine desiutil.plots.plot_slices that you could use for your qa_dz() function: http://desiutil.readthedocs.io/en/latest/api.html#desiutil.plots.plot_slices

[moustakas]

Thanks for the comments, @dkirkby. It's definitely a work in progress -- I'm building more QA that will help us understand what's going on. For example, regarding the significant drop in efficiency starting around r~19. It looks like nearly all (92%!) of the "lost" redshifts are at the correct redshift but have ZWARN==4, which is the "small_delta_chi2" flag. So I'm optimistic that we'll be able to recover these by staring at the chi2 vs redshift curves (volunteers are welcome).

Regarding the fiberloss: I'm using whatever is implemented in quickspectra, which looks like it's the fastsim option but with fixed half-light radii and Sersic indices (i.e., no redshift dependence). It would be good to get the redshift dependence you're incorporating into this piece of code (desisim.simexp.simulate_spectra). Or perhaps @julienguy has another suggestion?

As for the time -- there are 100 spectra per "chunk" of files so redrock is taking ~1.8 sec/spectrum with 4 cores (on my laptop). That's fast!

[moustakas]

Here's the current redshift efficiency curve alluded to, for reference.

[julienguy]

Just for confirmation, am I understanding correctly that the difference of efficiency you show with what I simulated is due to the moon?
My simulations had MOONALT=60 MOONFRAC=0.7 MOONSEP=50, and I got an efficiency dropping at lower magnitudes , with ~80% at rmag=19 for a 300s exposure.

[moustakas]

Right, my simulations are moon down / no moon, which is what @dkirkby needs (in desihub/desisurvey#77). I also have the "right" input distribution of redshifts and magnitudes because I tie the galaxy templates to the MXXL/BGS mock whereas I think you used a uniform input distribution of magnitudes and redshifts?

Related: In this notebook I should also derive the redshift efficiency with the set of BGS reference spectra generated in #285 and scaled to r=19.5 so we can tune the exposure time to achieve the redshift efficiency we need / want.

But suggestions for dealing with galaxies with ZWARN==4 would be helpful.

[moustakas]

This PR just adds a notebook, so I'd like to merge even though Travis tests have not yet passed.