moustakas
commented
1 year ago Thanks for reporting this issue @VFawcett. I drilled down on the first object in your list and I think I understand what's going on.
In the code, there are two rounds of fitting (neglecting the broad-line fitting):
- An initial round where the line-widths (
sigma) and velocity centers (vshift) of the forbidden and narrow Balmer+He lines are all tied together; - A final round where
sigmaandvshiftof the Balmer+He lines are tied together separately fromsigmaof the forbidden lines (thevshiftconstraint on the forbidden lines is removed to allow for wavelength calibration adjustments).- Moreover, in this final round, the
sigmaandvshiftparameters are constrained to be within +/-20% of their initial values, to prevent the optimizer from falling off a cliff given the greater number of degrees of freedom.
- Moreover, in this final round, the
For the narrow lines, the prior on sigma is [1,750] kms and the prior on vshift is [-500,+500] km/s. Initial values are set based on a quick estimate of the line-widths (if possible); in the case of the first object in your list, the initial line-width could not be determined and so it was set to a default of 200 km/s (I know this is large, but convergence isn't too sensitive to this value and this value is also used for masking during continuum-fitting, so it's intentionally conservative / large).
After the initial fitting round, if any parameter remains at its initial value it is dropped. (I played around with other choices like: "drop the line if it's within X% of its lower/upper bounds" but found that real lines were being dropped this way.)
In your first example, sigma for [OIII] 5007 is 749.7799 km/s, i.e., barely below the max allowed value of 750 km/s. Then going into the final fitting round, the prior becomes [599.8239, 899.7359] km/s (+/-20% of its initial value) and the final optimization ends with a value of 599.8310 km/s.
So: this logic explains not only your spike at ~600 km/s but also the one at ~900 km/s (which surprised me at first because I knew the prior was 750 km/s!). (In fact, I think this means that all line-widths >750 km/s should be suspect).
The good news is that, looking at the QA figure below, [OIII] is not significant, so we should be able to remove these warts by demanding well-detected lines, e.g., OIII_5007_AMP*sqrt(OIII_5007_IVAR)>2 (or 1.5 or 3, whatever), but keep in mind that there are likely to be real lines which have these values which we'll want to try to recover.
Going forward, I would be grateful for any thoughts you (or anyone else) may have for "cleaning up" issues like these. And if you could also report (in this ticket) how you're going to decide to proceed, that'd be helpful.
To generate the QA below:
source /global/cfs/cdirs/desi/software/desi_environment.sh 23.1
module swap desispec/0.57.0
module load fastspecfit/2.1.2
fastspecfit-qa $DESI_ROOT/spectro/fastspecfit/fuji/v2.0/healpix/special/dark/254/25463/fastspec-special-dark-25463.fits.gz \
--targetids 39627631057962664
Exploring the sigma distribution for some of the narrow lines there seems to be a strange peak at ~600. This is especially prominent in [OIII] and [OII].
Some example sources: [39627631057962664, 'special', 'dark', 25463] [39627637076787376, 'special', 'dark', 26146] [39628239173321503, 'special', 'dark', 27987] [39627601714610456, 'special', 'dark', 44996] [39628492240849767, 'special', 'dark', 9328] [39628512960711724, 'special', 'dark', 9336] [39628512956520409, 'special', 'dark', 9336] [39633179044218807, 'sv2', 'dark', 10955] [39633170546557400, 'sv2', 'dark', 10976] [39633304214834237, 'sv2', 'dark', 11007] [39633274842120987, 'sv2', 'dark', 11002] [39633349047747989, 'sv2', 'dark', 11193] [39633321927378204, 'sv2', 'dark', 11175] [39633335667920174, 'sv2', 'dark', 11186] [39633318404163394, 'sv2', 'dark', 11177] [39633345738442330, 'sv2', 'dark', 11187] [39633325408652474, 'sv2', 'dark', 11184] [39633304231611292, 'sv2', 'dark', 11176] [39633345776192464, 'sv2', 'dark', 11167] [39633365351008432, 'sv2', 'dark', 11214] [39633345813939941, 'sv2', 'dark', 11203]