Y1C2 Stellar and Galaxy Modeling Issues

-Eli Rykoff

Following up on the Y1C1 Stellar and Galaxy Modeling Issues with Y1C2 data! In particular, "Detmodels gone bad".

Consistency of Stellar Zero-Points

Repeating on the Y1C1 analysis, I have gone to each coadd tile and selected all bright stars (16<i_auto<20) that are isolated (no objects brighter than i_auto<22 within 10"). I then use a big aperture (APER_8, 6") for a "large" aperture. I then measure the median color offset between AUTO/PSF/MODEL/DETMODEL colors and APER_8 colors to look for correlations with with delta PSF, where delta PSF is the difference between the homogenized PSF in the given tile with the two bands in question. Eg, if psf_g=1.2 and psf_r = 1.1, then delta psf=0.1.

The basic idea is that the very large aperture on isolated objects should give us a measure of the "true" color of the object (at least on average -- these large apertures will be noisy). Any systematic difference from the truth will lead to tile-to-tile color variations that will impact all our science.

Note also that applying the galaxy model fitting to the stars (as is done for MODEL and DETMODEL magnitudes) is a peculiar thing to do, but certainly something to check. In principle, we should be using PSF magnitudes to measure the zero-point, and these should be directly applicable to MODEL and DETMODEL colors which use the same psf model, only convolved with the appropriate galaxy profile.

For the following plots, I only use tiles with at least 50 isolated stars and 50 isolated galaxies (see below). This covers most of the SPT-E Y1C2 tiles, but excludes the LMC contaminated regions to the SE where there aren't enough isolated objects.

Stars: Auto - Aper Stars: PSF - Aper
Stars: Model - Aper Stars: Detmodel - Aper

The results are similar to Y1C1: MAG_AUTO has a systematic bias correlated with delta-psf, such that the stellar colors are biased more when the two images that make up a color have a large psf difference. However, MAG_PSF performs very well, with biases <0.5%. Interestingly the MODEL magnitudes work reasonably well for r-i and i-z, but show a systematic trend for g-r.

Finally, the DETMODEL zeropoints are all over the place, with biases as much as +/-20%, uncorrelated with delta-psf. While there appears to be a set of tiles where the DETMODEL zero-point makes sense, there are a lot more where it doesn't. I think this points to a bug somewhere in detmodel, but there could be issues elsewhere in the processing stack.

Consistency of Galaxy Zero-Points

I now repeat the analysis with galaxies. This time I use an isolation criteria such that no objects brighter than i_auto<22 are within 15". I use a larger aperture: APER_10 (8"), and a slightly different magnitude cut (17<i_auto<20). For the galaxies I do not use PSF magnitudes.

Again, for all tiles with at least 50 isolated stars and 50 isolated galaxies, we find:

Galaxies: Auto - Aper
Galaxies: Model - Aper Galaxies: Detmodel - Aper

For AUTO magnitudes we see a similar trend as for the stars (see below for a comparison). The MODEL magnitudes are a bit peculiar. Although there isn't a trend with delta-psf, there is a significant offset for g-r vs r-i vs i-z, which is largest for g-r. I'm not sure what's would be causing this. Finally, the DETMODEL magnitudes are once again all over the place, although with a smaller amplitude (+/-5%) as compared to the DETMODEL magnitudes when applied to stars.

Comparing Stars to Galaxies

Now, I look at the tile-by-tile relative zero-point comparison between stars and galaxies.

Galaxies vs Stars: Auto - Aper Galaxies vs Stars: Detmodel/psf - Aper
Galaxies vs Stars: Model - Aper Galaxies vs Stars: Detmodel - Aper

I have added a 1-1 line for reference. In addition to comparing AUTO, MODEL, and DETMODEL colors, I've also added galaxy DETMODEL vs stellar PSF, which should be the standard operating mode.

Here there are a number of troubling aspects.

  1. Although stars and galaxies have the same trend for AUTO magnitudes, the slope is significantly different. Thus, calibrating the zeropoint via AUTO magnitudes on stars will yield significant offsets (+/-2%) in the galaxy photometry.
  2. Things are slightly better using MODEL magnitudes, but there is the color offset issue that only shows up with galaxies and not with stars.
  3. Although the trends are the same for the DETMODEL colors on stars and galaxies (this is an improvement from Y1C1, I believe), the slope is not unity (similar to AUTO colors). Thus using DETMODEL magnitudes on the stars for zeropoint calibration will lead to significant biases for the galaxies.
  4. The default mode, of applying the stellar PSF derived zeropoint to the galaxy DETMODEL colors will lead to variations of up to 5%.


The DETMODEL issues exhibited in the Y1C1 coadds are still present, at a level that makes extra-galactic science not possible. Furthermore, there appears to be something odd with the MODEL magnitudes, especially for g-r. Looking back at my plots for Y1C1 (not all on the wiki) I see this problem was there before, but it may be slightly stronger now.

Looking at Individual Images

One possibility that is the source of our problems is that something in the processing is screwing up the zeropoints when going from single-images (where GCM is run on the finalcut individual frames using stellar PSF magnitudes) to coadd images.

In this section I investigate whether there are color zeropoint shifts when moving from the individual finalcut frames to the coadd frames.

The Data

I am focusing on 6 test fields: DES0410-4957, DES0414-5748, DES0432-5540, DES0441-5957, DES0502-5248, and DES0522-5705. These were chosen because they were also reprocessed as non-homogenized (nh) coadds in run number 20130523000011.


For each field I started with the coadd catalog, taking all objects with i<22. For each individual image/ccd I matched all objects within 1" (via ALPHAWIN_J2000/DELTAWIN_J2000) to the coadd catalog, and recomputed all magnitudes using zeropoints from the ZEROPOINT database table where SOURCE='GCM'. For all objects with at least 2 individual good observations (FLAGS<=3; MAG<90) I calculated the weighted mean magnitude (using MAG_AUTO, MAG_PSF, MAG_MODEL, MAG_APER_8, MAG_APER_10). We don't have DETMODEL magnitudes for the finalcut images because they are not processed with a deep detection image.

I note that the magnitude errors in the finalcut images that I looked at are wrong: too low by a factor of at least 100. This bug had been seen previously in the firstcut images.

Results: Homogenized Coadds


I first select stars from the coadd catalog, cutting on abs(SPREAD_MODEL_I) < 0.002, 16.0 < MAG_AUTO_I < 20.0, 0< color < 1.5. There are no isolation criteria. I then calculate the median color offset between the color computed with the coadd image and that computed with the weighted mean of individual catalogs. Note that the x axis just shows the index for the 6 fields checked and has no other relevant meaning.

First, the PSF magnitude based color zeropoints do fluctuate, but <~1%. Similarly, the large aperture magnitudes are consistent within 1%, with perhaps a bit smaller scatter than PSF magnitudes. However, the MODEL zeropoints (wrong for stars!) are all over the place (amplitude of 5%) as well as the AUTO zeropoints (amplitude of 6%), which are strongly dependent on the relative PSFs of the two images.

Is this helpful for triangulating the problem? The PSF magnitude color zeropoints seem reasonably well behaved. But the MODEL zeropoints are all over the place, which may be why things aren't working.


I select galaxies from the coadd catalog, cutting on SPREAD_MODEL_I > 0.003, 16.0 < MAG_AUTO_I < 20.0, 0 <color < 1.5. I then calculate the median color offset between the color computed with the coadd image and that computed with the weighted mean of individual catalogs. Note that the x axis just shows the index for the 6 fields checked and has no other relevant meaning.

As with the stars, the large aperture-based color zeropoints are very stable. But the AUTO and MODEL zeropoints are all over the place! This may be coming close to giving us a clue: even if the PSF and APER based zeropoints are consistent when moving from individual frames to coadds, the galaxy MODEL and AUTO zeropoints certainly are not. Somewhere, something is going wrong... this the problem with the PSF homogenization?

Results: Non-Homogenized Coadds


All analysis the same, now with the DESDM processed non-homogenized coadds.

This looks very similar to the homogenized case, except that the aperture color zeropoints are much better behaved. Interestingly, the zeropoint scatter for PSF magnitudes is not significantly larger in this non-homogenized case.


These look much the same as for the homogenized case, with large scatter in the model and auto color zeropoints, but reasonable performance with the large apertures.


Something in the coaddition procedure in DESDM is adding a large amount of scatter in the color-based zeropoints derived from the individual images. It does not appear that the psf homogenization is the source of this scatter.