[Written by Gene Magnier]

I'm now happy with how the flat-field correction analysis is working. I've measured the correction for the griy image sequences from 2008.09.20, and the results look very encouraging. To review: these datasets consist of ~36 images heavily dithered on a large range of scales and offsets across a dense star field. The analysis takes all of the stars which have been observed multiple times (> 3) in a filter, selecting out observations above a S/N cutoff and/or an instrumental magnitude cutoff. It then attempts to determine the average magnitude for each star, and a coarse grid of magnitude corrections as a function of position across the full mosaic to minimize the scatter of the residuals for all stars.

The analysis goes through a series of iterations, and there is a fair amount of outlier rejection during this process to reject a) stars which are very discrepant (their internal scatter is > 2x the median scatter for all stars); b) measurements which are very discrepant (> 3sigma from mean, where the sigma is measured from the inner 50% of the measurements for the star); c) images which are very discrepant (scatter > 2x the median scatter of all images or a zero point offset of more than 0.1 mags) -- no images were actually rejected in these particular analysis.

The first two plots below tell most of the story: the first is the star residuals vs the x-direction mosaic coordinate (note that i've shifted this to run from 0-40k, sorry) without the mosaic-position-dependent correction applied; the second shows the residuals after the the correction is applied (this is for r-band in this case). The effect of the correction is obvious. There are several hundred thousand detections, so the actual distribution after the correction is not very clear.

Grid.r.XdM.png Grid.r.XdMf.png

The third plot is the histogram of the scatter for each of the stars used in the analysis (note: histogram of sigmas, NOT a histogram of the residuals). Note that there are ~100k stars with scatter < 0.01 mags in this corrected dataset!


The next three plots show the same histogram for the other three filters (g,i,y).

Grid.g.dMhist.png Grid.i.dMhist.png Grid.y.dMhist.png

I have not yet applied this correction to any other datasets. So, I do not know if this is real or artificial: if the photometry errors that are being corrected are caused by anything equivalent to a flat-field error, and the effect is stable, then this is really good news. It would mean that we can obtain systematic relative photometry errors definitely better than 0.01 mags, with 0.005 mags looking like the typical value. On the other hand, if this correction is mostly fixing a random error in the photometry (say, it is totally driven by a poor psf-aperture magnitude correction) or if it is caused by a real instrumental error, but one that is variable with time, then things are not as rosy as these plots would suggest.

The last four images show the applied correction as a function of position in the mosaic (probably with the wrong parity relative to the camera coord system -- sorry). Each image is for a different filter: g,r,i,z. The grey-scale runs from -0.1 mag offset at white to +0.1 mag offset at black. The first obvious point is that most of the correction in g,r,i is a chip-to-chip error. Note that the analysis does not really know about the chips: they are subdivided into 4x4 subcells, and there is no difference in the analysis between a subset on one chip and a subcell on another chip. the fact that the chips stand out so strong suggests, to me, that this is a real effect of some kind. The y-band correction is very different: it looks more like the CFHT 'scattered light' correction. My guess is that the g,r,i results show that the dark inside of the dome really worked, but that the paint is not very black at 1um. There are also low-level circularly symmetric patterns in all four filters, but you need to flip back and forth to really notice them.

Grid.g.jpg Grid.r.jpg Grid.i.jpg Grid.y.jpg

At this point, I am trying to apply these corrections to the raw master flats, and testing out the related pipeline code at the same time.