When combining image frames scaled to the same exposure time, weighing the frames with weights proportional to their exposure time is optimum for photon-noise limited data. Weighing with weights proportional to the square of the exposure time is optimum for read-noise dominated data. In order to be systematic, an inverse variance weighting ***check this*** was used for all the final combination of the stacks. A single variance value was used for each chip, namely the expected variance for the sky level composed of the variances from the measured sky level, the dark current and the known read-noise. The sky levels were measured as 5-sigma clipped iterated means. The dark current is measured from the dark current exposures. The weights do not vary much between the three WFC chips, but it is slightly different for the PC chip, because of the differences in the dark current.
The expected variance was compared to the measured variance, and the latter typically turned out to be 10 % higher than the expected variance, which is entirely accountable by cosmic rays, since the sky was measured in pipelined frames with all the cosmic rays still in. Also a gradient in the image or scattered light can introduce a structure that entails a higher empirical variance.
GCOMBINE in principle could handle spatially variable weight maps, in order to combine each pixel with the weight appropriate for its value, but that has not been tried so far.
GCOMBINE outputs a variance map which is almost constant, except for the bad pixels. The variance variations caused by missing data due to cosmic rays introduce a small modulation ("ripple") on this otherwise flat surface. Thus when one feeds a constant weight (except for the bad pixels) into the subsequent DRIZZLE task one is presumably not doing too much wrong. In principle the variance map from GCOMBINE could be used for DRIZZLE, but in this processing round it has not been used.
Harry Ferguson ferguson@stsci.edu 1/14/95