| | 8 | I spent much of the week looking in detail at the interactions between ppSub and psphot. This has led me to explore in detail the behavior of the PSF matching analysis and its impact on the resulting PSF. I have been working with simulated images with known PSFs, and have discovered some issues with the PSF matching code. For example, I was having a lot of trouble getting a pair of images with the same PSF to subtract well. I had not fully appreciated that the PSF matching algorithm is attempting to fit the difference of the two images with zero-flux basis functions (A - norm*B - \sum_i a_i C_i where C_i = k_i x B). Realizing this, I noticed that the raw difference (A - norm * B) had significant residuals -- this meant that the normalization was not measured correctly, partly due to an error in the kernel basis function normalizations (see Paul's report) and partly due to sensitivity of the analysis to outlier pixels. I added a windowing function to limit the normalization calculation to the dominant flux from the matched stars. A related issue is that the coefficients of the basis functions (which are strongly coupled -- this is not an orthogonal basis set) are sensitive to ringing. I have been looking into options for re-conditioning the matrix equation using SVD. |
