It appears that the convolution does handle variance plane correctly, and the variance is scaled appropriately back to (nearly) the theoretical value. This may be summarized in this IPython notebook and specifically this figure, which shows histograms of the per-pixel variances for an imageDifference run on a single DECam CCD, including the two original images (science and template), the "uncorrected" diffim ("orig"), the decorrelated diffim ("corr", for corrected) and the naive expected variances which is the sum of the raw variances ("expected"):

The mean values of the corrected and expected variances differ by about 0.5-sigma. The 1-sigma spread in variances is greater in the corrected diffim than one would naively expect. This is likely due to issues being investigated in ~~DM-6243~~ (the fact that we use a single decorrelation kernel based on matching kernel computed from the center of the image, and based on single (scalar) variances for the two input images). It is not clear that this will present an issue (again, see ~~DM-6243~~).

An additional analysis that is relevant to the variances is that of the resulting detections and their measured SNRs. This is to be part of ~~DM-6244~~.

This ticket is now considered Done.

Work underway. Additional test cases added to testImageDecorrelation.py - so this will have an IPython notebook as well as additional code in ip_diffim.