| | 10 | Another important cause of the elevated magic masking results from the post-8/23 modification of the PSF matching code. The PSF matching code improvements were motivated by the tendency of the PSF matching to leave behind donuts and dipoles. These were both examples of the PSF-matching not actually matching very well -- in these cases, the solutions tended to under fit the difference (in other words, the PSF-matched images were not convolved by enough of a kernel to modify the PSF enough). The work over the Fall (updates in 8/23 and 10/29) introduced modifications to the dual convolution to achieve better solutions, with good success to that goal. However, my investigations last week showed that these solutions could also over fit the PSF differences: they allowed potentially so many free parameters that the solutions could add excess noise to the images. At a suggestion from John Tonry, I have modified the code to require higher-order fits (from the dual convolution and spatial variations) only be used if they were justified by improvements in the chi-square (this was actually trickier than it seems -- larger kernels also smooth the image, reducing the effective chisq if the impact of that smoothing on the variance is not included). I tested this code over a wide range of simulated images which either should require or not require dual and/or spatial variations, and found that the code did a good job of choosing the expected set of fit parameters. |
