Y1C1 Stellar and Galaxy Modeling Issues

-Eli Rykoff

Updated! See below.

The Tests

Previously, I looked at the consistency of the cluster red sequence across Y1C1 (specifically SPT-E) and found significant zero-point variations in the red galaxy colors across the field of view. I've now turned my attention to the stars and galaxies to start developing photometric null tests that may shed some light on what's going wrong.

All the analyses here are on the full Y1C1_COADD_CAL database, with color corrections from Bob Armstrong's tile-by-tile SLR (rearmstr.slr_cal in the database). I believe the SLR fits were run on MAG_AUTO stellar colors.

I first pixelize the SPT-E Y1C1 region with a mangle simple pixelization scheme of depth 8, which gives a pixel size of ~0.6 deg^2. In each pixel are 200-2000 galaxies (as determined with SPREAD_MODEL_I > 0.002) with 18<MAG_AUTO_I<21 and FLAGS_I = 0 (not deblended, saturated, etc). I only use galaxies with 0.0<MAG_AUTO_G-MAG_AUTO_R<2.0; 0.0<MAG_AUTO_R-MAG_AUTO_I<1.5; and 0.0<MAG_AUTO_I-MAG_AUTO_Z<1.5 to look at galaxies with "sane" colors. At these magnitudes there aren't many insane color outliers so this doesn't make a significant difference.

First, I have mapped out the homogenized PSF used in the coadd tiles:

PSF map, g band PSF map, r band
PSF map, i band PSF map, z band

There's a broad variation in the PSF across the fields. Furthermore, the g & r PSFs are correlated, because they were taken concurrently during dark time. Similarly, the i & z PSFs are correlated, because they were taken concurrently when the moon was up. Most importantly, the relative PSFs change over the field. Of course, since we know the PSF we should be able to take out any effect over the field, but let's see if we are...

Next, I look at galaxy color offsets between various magnitude measures. The most interesting are the comparisons between AUTO colors and DETMODEL colors:

AUTO color - DETMODEL color, g-r AUTO color - DETMODEL color, r-i
AUTO color - DETMODEL color, i-z

First, I note that the amplitude of the fluctuations is quite large. In r-i, it's almost 0.1 magnitudes from the south to the northeast. That is, the zero-points depend strongly on if you are using AUTO colors or DETMODEL colors. Next, I note that the largest excursions appear to be correlated with the worst PSF. The bad region in color space in the NE is correlated with the worst psf in r, and the bad region in color space in the S is correlated with the worst psf in i. (It is not surprising that large psfs in r and i have opposite signs in r-i color space.)

Finally, I look at stellar color offsets between various magnitude measures. The most interesting are the comparisons between AUTO colors (used for the SLR) and the PSF colors (which should be related to DETMODEL colors, as they use the PSF fits):

AUTO color - PSF color, stars, g-r AUTO color - PSF color, stars, r-i
AUTO color - PSF color, stars, i-z

And these have the same correlations with the PSF as the galaxies, and the same general pattern as the galaxies. However, and somewhat worrying, the magnitude of the bias appears to be worse for galaxies than stars.

Additional SLR tests

I ran Pat Kelly's BIGMACS code on the best and worst pixels as chosen by the r-i star residual map, running each with MAG_AUTO and MAG_PSF. The color fits for MAG_AUTO were roughly consistent with Bob's SLR runs, but using MAG_PSF yielded significantly different color zero-points, especially in r-i.


There are unacceptably large biases in the color zero-points that depend on the PSF, even with the PSF homogenization. These show up in both the stellar and galaxy colors, and may be the source of the red sequence spatial residuals I have observed, especially considering I used DETMODEL colors. Certainly, at the moment, a single zero-point solution is not appropriate for both AUTO and PSF-fit magnitudes. I have not yet confirmed if the same zero-points are appropriate for both stars and galaxies in the data (due to biases introduced by the psf homogenization, model fitting, etc.), though I think the red sequence tests can shed some light on that.


New on 5/6/13
-Eli Rykoff

More Tests

In the previous section I showed that there are significant differences in PSF and AUTO colors for stars and DETMODEL and AUTO colors for galaxies, and how we need to be extra careful as to which magnitudes we are using for calibration. But which is closer to the truth?

Let's go with the assumption that with a large enough aperture we can measure (close to) the total flux of a star or a galaxy, or at least large enough so that the PSF variations don't matter. This of course may be noisy, but should not be a big problem for brighter (i<20) stars and galaxies, provided the objects are suitably isolated. So we can compare AUTO based colors and PSF based colors of stars to large APER colors for these isolated objects.

Isolated Stars

In each coadd tile I select all stars that are 16<i<20 and do not have any nearby object brighter than i<21 within 10". I use APER8 (6" diameter) for my "large" aperture. I then measure the median offset between the AUTO/PSF/and APER8 based colors to look for a correlation 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.

Stars AUTO color minus PSF color Stars AUTO color minus APER color
Stars PSF color minus APER color

In these plots, each point represents a single coadd tile. I have not observed any significant trend in each field as a function of magnitude or color, so that's good. (Although the individual bands show a slope due to the size/flux relationship, this mostly cancels when measuring colors.)

The primary take-away here is that:
  1. The PSF color for stars is relatively insensitive to the difference in PSF size between the two bands (biases < 0.5%), and
  2. The AUTO color for stars is sensitive to differences in PSF size between the two bands, with biases as much as 4% when the two PSFs differ by 0.3".

I also note that you can see how the g/r and i/z psf sizes are correlated, while there is a large spread in r/i delta PSF as these fields are observed at different times in these two bands.

Isolated Galaxies

I now repeat this procedure using galaxies that are 18<i<20 and do not have any nearby object brighter than i<21 within 20". I used APER_10 (8" diameter) for my "large" aperture, although APER_8 shows basically the same "pattern". I have replaced PSF colors of stars with DETMODEL colors of galaxies. We now have:

Galaxies AUTO color minus DETMODEL color Galaxies AUTO color minus APER color
Galaxies DETMODEL color minus APER color
The primary take-away here is that:
  1. The AUTO colors show very similar biases that are seen in the AUTO magnitudes for stars, though the biases are slightly smaller, at 2% bias when the the PSFs differ by 0.3".
  2. The DETMODEL colors are simply wrong. All over the place, when compared to aperture colors. And not necessarily correlated with delta PSF.

At the moment, using AUTO colors will be less problematic than DETMODEL colors, but there are the issues of the few percent biases in the stellar colors when using AUTO magnitudes. But there must be a bug in DETMODEL or in the pipeline implementation of DETMODEL, to create such scatter.

Re-Coadding Images

One of the most striking things about the galaxy DETMODEL - APER10 color plot above is the large scatter in DETMODEL colors that does not appear to be correlated with delta-PSF. In particular, there appear to be three clumps in the i-z color even though the psfs are well matched. What's up with this?

The Test Data

I have decided to focus on the i-z color for 3 fields, one from each clump. These are DES0426-6039, DES0427-5831, and DES0428-5914. Quite alarmingly, these are all from the same general area of SPT-E.

I have looked at these 3 fields in 4 different ways:

  1. The catalogs as output from Y1C1 (run with SExtractor v2.18.?, exp+dev model, using the _det.fits detection i-band image.)
  2. A rerun of SExtractor 2.18.10 on the Y1C1 coadds (exp + dev, using _det.fits) to confirm I'm running equivalent code.
  3. A rerun of SExtractor 2.18.10 on the Y1C1 coadds (sersic, using _det.fits) -- is this related to the type of model?
  4. A non-homogenized restack of the finalcut images. I ran my own scamp + swarp (mean stacks) + psfex (default parameters) + SExtractor 2.18.10 (sersic, using i band image as detection image).

The zero-points for each of these images was set from Bob Armstrong's field-by-field SLR matching stars via the PSF magnitudes.


Here is a table of the color offsets for the three fields for stars/galaxies for all these tests:

                    DES0426-6039 DES0427-5831  DES0428-5914 
Regular Y1C1:
  auto_m_aper    :   -0.00153351 -0.000865936  -0.00452995
  psf_m_aper     :   0.000257492  0.000144958  0.000797272
  model_m_aper   :   0.000885010  0.000469208  0.000564575
  detmodel_m_aper:     -0.212461 -0.000562668   -0.0942459

  auto_m_aper    :   -0.00108337 -0.000698090  -0.00233650
  model_m_aper   :    0.00167274   0.00141907   0.00245857
  detmodel_m_aper:    -0.0618095   0.00916100   -0.0226688

Y1C1 resex (exp+dev):

  auto_m_aper    :    -0.0152969  -0.00131798   -0.0185299
  psf_m_aper     :   0.000400543  0.000370026   0.00106239
  model_m_aper   :   0.000829697  0.000576019  0.000654221
  detmodel_m_aper:     -0.211813 -0.000318527   -0.0945835

  auto_m_aper    :   -0.00135231 -0.000947952  -0.00423050
  model_m_aper   :    0.00165176 -0.000848770  -0.00131989
  detmodel_m_aper:    -0.0632267   0.00942993   -0.0238132

Y1C1 resex (sersic):

  auto_m_aper    :    -0.0152969  -0.00131798   -0.0185299
  psf_m_aper     :   0.000400543  0.000370026   0.00106239
  model_m_aper   :   -0.00330544  0.000425339  -0.00251961
  detmodel_m_aper:     -0.211407 -0.000347137   -0.0945072

  auto_m_aper    :   -0.00135231 -0.000947952  -0.00423050
  model_m_aper   :    0.00171280   0.00454903   0.00526237
  detmodel_m_aper:    -0.0637398   0.00925446   -0.0255566

re-coadd (sersic):

  auto_m_aper    :  -0.000627518  -0.00845909  -0.00376701
  psf_m_aper     :   0.000747681   0.00263596  -0.00184441
  model_m_aper   :    0.00146294   0.00387001 -2.09808e-05
  detmodel_m_aper:  -0.000619888   0.00480652  -0.00190163

  auto_m_aper    :   -0.00112915  -0.00170708  -0.00135040
  model_m_aper   :    0.00106430   0.00341606   0.00199890
  detmodel_m_aper:     0.0103569    0.0112457    0.0143986

The takeaway:

  1. My run of SExtractor is virtually identical to the official run of SExtractor
  2. The problem is not the exp+dev vs sersic model fitting
  3. The MODEL colors are noisy but unbiased for stars and galaxies.
  4. For DETMODEL, stars and galaxies are doing different things.
  5. The DETMODEL magnitudes on stars (using a galaxy model, mind you!) can be waaaay off: 0.2 mag, and this is entirely due to some issue with the z band DETMODEL mag on stars that are systematically offset by 0.2 mag for DES0426-6039.
  6. All the homogenized runs have similar problems.
  7. The non-homogenized run looks fine to first order (though there may be a systematic offset of 1% in the galaxies).

Something is going wrong in the DETMODEL colors for the homogenized coadds in particular, and it does not depend on the model fit. One thing to note is that when DETMODEL is run the model is fit on the detection image while when MODEL is run the model is fit on the measurement image (even though these are based on the same input stacks). Is there some characteristic of the detection image that is screwing up the model fitting as applied to DETMODEL?