ctools

Something else that I noticed is that the model map shows some strange ring shapes. This seems to originate from the IRF convolution but doesn’t look physically sensible if I am not missing something.

In my test scenario, the fix in #2860 changes the number of counts in the model cube to 89149. (In the unbinned case: Nobs - Npred = 93267.000 - 93322.253 = -55.253)
The model map now looks much smoother:

Updated image.

I looked a bit in the problem and played around with the integration precisions in GCTAResponseIrf::irf_radial. Below a summary of the results, obtained using a modified simulation.py script where the cube energy range was limited to 0.5-90 TeV (to make sure to stay away from the threshold). I tried a variant where the maximum delta of the PSF was restricted to 99% of the PSF containment, but this led to some artefacts in the model cube at high energies. Setting the number of iterations in rho and phi to 6, as done in #2860, reduces the discrepancy between model and counts to about 3%. 1% is only reached when further increasing the integration precision in phi, but this brings the computation time to 10 min instead of 1 min when compared to the initial precision. In other words, the speed penalty would be a factor of 10, while with 6/6 it would be a factor of 4, which is already quite bad.

I prefer for the time being (and release 1.6) to fix the precision at 6/6, and defer it to the next release to think about a more precise but fast integration scheme. I keep this issue open for that purpose.