| | 8 | I have been chasing down issues in the stack photometry analysis, and in the process have gotten side-tracked by concerns in the PSF matching code. The PSF-matched stacks were looking very poor (lots of ringing around bright stars). I identified some small bugs, but the main issue seemed to be a failure of the PSF matching code to generate well-matched images, despite reasonably easy inputs. The PSF matching code essentially employs the Alard-Lupton algorithm using in difference image analysis, so to explore this further, I worked with Mark Huber to compare ppSub and 'hotpants' (the photpipe implementation of Alard-Lupton) on some simple images. The result: even in some cases, ppSub was failing to get as good a subtraction as hotpants. (How does this jibe with reports from photfest that the two codes produce effectively equivalent subtractions? For a couple of reasons: (a) the differences are somewhat subtle and can be overwhelmed by the unmatchable differences in real image, (b) ppSub was able to give good results for part of the test images.). By running hotpants on the same images and looking at the code, I was able to identify an important difference. The convolution kernels need to be generated with zero net flux (due to the way the algorithm is defined). This require modification of the even-order polynomial kernels. In the case of ppSub, the zero flux was achieved by subtracting a delta-function from each of those kernel terms; for hotpants, the zero-order smallest kernel was being subtracted in full. These seem equivalent mathematically, but when I changed the ppSub code to use the same technique, the resulting subtractions were much better matches to the hotpants version. Although the issue was most obvious in the single-convolution difference, this fix also had the important side-benefit of substantially improving the dual convolution analysis. A comparison of the prior version of the code (from 9/18, just before photfest) and this most recent fix can be seen here: [wiki:ppSubModifications_201009]. |
