If we take a stack of 100 images, calculate the total counts in each frame and tabulate this, and then take the list of totals and compare the mean of the list to its variance we find that variance is MUCH more than the mean. This should not be the case if the totals were Poisson-distributed.
We find that variance is 100s to 1000s times bigger than the mean of the list.
This could possibly be due to shutter variations – the shutter is … less than fantastic … after all.
Let us inspect the problem and see if it was less at the beginning of operations at MLO. Perhaps wear made it worse as time went on? We did collect 50000 images or so.
Perhaps atmospheric transmission changes on short time scales is this big? Each of the 100 images in a stack took about a second to capture. Does sky transparency change by a lot over that timescale, on a 1×1 degree field?
Repeat for subimages – perhaps the BS somehow dominates the total count in an unforeseen way?
Repeat for subimages of images that have been scaled to a realistic fixed number (e.g. The median of all images in the stack).