I've modified the luptitude softening b parameter to reflect the 1-sigma error in the coadded images as a function of filter. The error as a function of magnitude now makes sense. The single-pass errors cross 20% at r=22.65, the 26-frame coadd at r=24.4. This difference is 1.75 mags, where 1.77 mags is expected from sqrt(N).
Having said that, the 20% crossings in the coadded u,g,r,i,z are
23.50, 24.75, 24.40, 23.85, 22.45
which differ from the depth of single-pass data+1.75 mags by
0.05, -0.2, 0, -0.2, -0.1. That is, the g and i bands are two tenths
less deep than I would expect. This is perhaps understandable
as we select images for inclusion in r and take what comes with
that selection in the other bandpasses.
Michael mentions the increased scatter in the psf-model vs model magnitude diagram in the g-band (upper right plot, page 13 in the coadd_color_color-rerun5 plots) This seems to be because the psf is more poorly measured there- remarkably there means 17-20 in g-band. The following plot has on the x-axis the ratio of the adaptive second moment size of the psf as measured on the image (M_rr_cc) and as evaluated at the object position (M_rr_cc_psf) for clean stars between 18 and 19 (in the model mag of the relevant band); this should be 1 if the psf was perfectly modeled. On the y-axis is the psg-model mag for the band pass of interest. u-band is blue, r-band is red, i-band is orange, z-band purple, and g-band is black. The g-band has lots of data points outside the locus of the other filters and it is these are the increased scatter.
Why do these g-band objects have poorly modeled psf?
For that matter, why is there a correlation between measured psf size and
the ratio of measured to reconstructed psf size? As I write that, it becomes
clear to me that this is because the reconstructed psf has constant size. Can
that be right? Then perhaps the question is not the perfect line seen
in the z-band below but why the scatter plot seen in the g-band...
Here is another plot showing the star galaxy separation problem. The data are all the col 2 data, about 1.1 sq-degrees. The star galaxy separator is just that of photo. The lines are the Bahcall-Sonaira model for stars and the Metcalfe R-band galaxy number count slope (with arbitrary zeropoint) for the galaxies. Stars start masquerading as galaxies starting around r=22. Since galaxies outnumber stars (do they really at r= 20?) this isn't really a problem untill r~23.75 - 24, where the galaxy counts roll over while the star counts keep rising.
It is tempting to identify the spray of objects to the right (psf gt model) in the r-band psf-model vs. mag plot, the middle left plot on page 13 of Michael's plot. The "stars"- which I take to be the objects on the locus at psf-model = 0- start becoming sparse at r~22. The spray to psf-model ~0.03 kick in there and this is about where s/g separation becomes a problem.
In other words, there is no evidence the photo s/g separator works well
below the single-pass magnitude limits. Changing to a LRG-like S/G at
psf-model=0.05 moves the start of misclassification to model=23, but the ridge
of data at r~24 allows no further progress with that simple classifier.
At r~23, the S/N = 20, so the statement is that probabilistic methods are necessary
at S/N < 20.
