Figure out why PSF photometry is wrong in simple gaussian case

[rknop]

On my branch that will merge "soon", see test_pipeline.py::test_pipeline_run_simple_gauss1.

That makes a plot. You can see the PSF is way off.

If you look at the output files with the decorrelated difference and the decorellated PSF, it's clear the PSF is not the right thing for the. The decoreelated difference has a PSF that's the same size as in the input images, but the decorrelated PSF is a delta function with some ringing.

[rknop]

Some things that may be relevant : this is the decorrelation kernel that we get (real part of inverse fourier transform of variable FKDECO_GPU), which is the thing that is convolved with the PSF and images. It's a delta function -- which I think you'd expect from the science and template images having the same seeing, but I don't really understand what's going on -- but with a lot of ringing.

[rknop]

This image is a plot of the truth, aperture sum, and PSF photometry using σ=2.0 pixel Gaussians. I think the aperture sum is sytematically low because it's in an aperture of radius 4 pixels, which only includes ~86% of the flux (if I did my integral right). But, obviously, something's wrong with PSF photometry, and I think it's the PSF. (It's not the zeropoint, because aperture sum and psf plots use the same zeropoint.)

test_pipeline_run_simple_gauss1.pdf

[rknop]

...I think the erorrbars are not plotted right, because they are not consistent on the PSF points in the bottom and to plots. That's almost certainly just a plotting error.