rmandelb
commented
11 years ago This sadly neglected issue related to accuracy of real galaxy calculations needs an update. I'm going to post here a revised list based on the one originally posted here many months ago, plus updates that have occurred since then. Assume that the noise-padding stuff on #238 is done for the purpose of this discussion, since that's nearly true. Issues like optimization are not included, since right now the code is fairly fast and nobody is actively complaining. :) Comments on the plans and/or priority of various items are very welcome:
(1) tests of accuracy of shearing as a function of interpolation kernel: this is super important for the challenge or for any use of the code for the basic shear tests. It's also potentially quite difficult to test in a clear-cut way. Given how important this is, I think it would be nice to discuss how we can do this (pinging the usual suspects: @barnabytprowe , @rmjarvis , and also bothering @gbernstein since it's a natural extension of some of the SHERA vs. SBProfile tests we did before). I have a few ideas, but please let me know if you have more:
(a) Take some reasonably complicated real HST galaxy and maybe two target PSFs and pixel scales (one ground, one space). Simulate the HST galaxy with complications like shear, rotation, magnification, where for each choice of target PSF + shear + rotation + magnification we simulate the galaxy with a range of image padding and interpolant choices. Then, check how the observed images change for the different choices of padding and interpolants. Here we don't have ground truth but we do know which interpolants should do better than others (roughly) so we can at least look for convergence towards a particular answer with better interpolants / more appropriate choice of image padding.
(b) Take one of the HST images and apply to the observed image some operations like shearing, shifting, dilation, rotation. We know how some of the moments should change in each of those cases, so we can look for the correct change in the observed moments.
(c) Make some fake HST data using some analytic profiles / sums of profiles (e.g., bulge + disk) and an OpticalPSF. Use it to simulate sheared lower-resolution data for some target PSF and pixel scale, and compare the resulting images with those we get by directly shearing and convolving the original sums of profiles with the target PSF, then resampling to the target pixel scale.
What else? Is this enough? Depending on the results of these tests, we may end up changing default interpolants / image padding for the RealGalaxy objects, and should certainly update the documentation.
(2) Getting original image / PSF more easily (per Mike's request on this thread from months ago): this should be easy and can take advantage of the InterpolatedImage. It's not super important but it is quick and useful.
barnabytprowe
gbernstein
msimet
rmjarvis
Now that we have a way of simulating galaxies in HST training data, I'm opening this issue as a place to collect suggestions for changes to that part of the code. There are a number of obvious things to be done:
SBPseudoDeconvolveclass that carries out pseudo-deconvolution, so we can check whether those would be easier to photon-shoot.Other suggestions are welcome!