AUTO and DETMODEL Colors in X3¶
This is a follow-up on Testing Astrometry And Selection on Coadds, looking at the galaxy colors in X3.
It appears as though there are biases in redMaPPer richness values caused by shifts in detmodel colors caused by psf discontinuities (and corresponding underestimates of the detmodel color errors).
Is there a corresponding problem that could affect photo-zs derived from detmodel colors? Let's take a look!
I made a couple of redMaPPer test runs (after fixing a couple of cough bugs cough), using AUTO and DETMODEL colors. The training is done using spectra; for these runs I used all the spectra that we have available, using these as "seeds" to look for clusters around red spectroscopic galaxies. (See Rykoff+14 for details). At the end of the training run, I looked at a comparison of the richness values computed from AUTO and DETMODEL colors, and noticed something peculiar. There were a significant number of clusters where the AUTO color richness was more than a factor of 2 larger than the DETMODEL richness, which means something has gone seriously wrong. (typically the scatter is <~10%, for clusters with lambda>10). And then I noticed these were almost all concentrated in the SN X3 region (even though this isn't the only spectroscopic training region).
Here is one of the richest of these clusters, that drops from lambda~40 with AUTO colors to lambda~10 with DETMODEL colors (ouch!) at z=0.4:
A pretty cluster in a pretty blotchy region with variable depth.
I've zoomed in on a region of X3 with 36<RA<39 and -7<Dec<-4. And I'm looking at r-i colors since these are the most important for z~0.4 clusters (but presumably problems exist for other colors). First, on the left is a plot of the difference in color of all galaxies with i_auto<22. Red and blue lines denote the 10th and 90th percentile. On the right are the positions of these galaxies (red is low, white is high delta). And the cyan points show the problematic clusters.
And isn't that interesting. There is some very significant spatial structure here, and the problematic clusters are mostly in the same region (and note that there are small scale variations as well.)
This is a problem...
As an aside: I note that there is magnitude bias in the color shift between auto and detmodel colors, where the mean bias is larger at the bright end. (Similarly there are psf mag - mag_auto trends). Not sure what to do with this...
Which colors are correct, or are they both wrong? And what does this do to the spectroscopic training for photo-zs?
Finally, I've taken all the spectra in the test region, and plotted r-i vs z for both colors. Blue points are high delta-color outliers, red are low delta-color outliers:
And it's plainly obvious that in the right panel (detmodel colors) the blue points are significantly biased off the red sequence, while the corresponding points in the left panel (auto colors) are consistent with the rest of the galaxies. Thus, it appears as though the detmodel colors are biased in some regions, and this will presumably feed through to photo-z training and runs.
However, all is not all kittens and unicorns for the auto colors. In particular, there is a set of red points that with auto colors are crazy-red and these look incorrect. But it's a much smaller fraction of objects, but I'm not sure what to make of these...oh, I just took a look and I think these are in especially shallow regions. The error bars (not shown) are roughly consistent with the red sequence (if these are red dropout galaxies, consistent with them being the reddest objects at a given redshift). And MAG_DETMODEL, with a fixed profile, will certainly be better at forced photometry of dropouts than MAG_AUTO. Something to consider...