GammaLib

May it be a problem, that the rmf is not re-normalized to unity after the cuts are applied?

For testing purposes I calculated a sum in the way shown below. The sums for different locations in the code are shown in the plot.

for (int itrue=0; itrue<ntrue; ++itrue) {
    double sum = 0;
    for (int ireco=0; ireco<nreco; ++ireco) {
        sum += m_rmf(itrue, ireco);
    }
}

If I apply a quick renormalization via dividing per etrue bin by the just calculated sum the fit result enhances a bit:

      | Reference (hess tool, gammapy) | joint, activated cut on rmf, renormalized
------+--------------------------------+-------------------------------------------
Norm  |      1.802 +- 0.212 e-11       |         1.957 +- 0.210 e-12 (+9%)
Index |      3.429 +- 0.269            |         3.568 +- 0.256.     (+4%)

I think the underlying problem is something different and disabling the cut on the rmf and/or renormalizing the energy dispersion are just something that patched the actual issue (maybe by chance to reasonable fit results):

We discussed in our group here in Erlangen and we think that the effective area should not be zero beyond the energy threshold (at least for the case where energy dispersion is enabled!). As this seems to be a different major topic I opened a new issue: https://cta-redmine.irap.omp.eu/issues/2674

A zero effective area beyond the threshold prevents the model flux with true energies below the threshold from contributing to the predicted counts above and in the vicinity of the threshold although their probability of being reconstructed to energies above the threshold may be >0 (depends on specific edisp!).

This would mean that the predicted counts in the vicinity of the threshold are systematically underestimated.

why did renormalization improve results?¶

Hence, indeed renormalization of the energy dispersion after the cut on the rmf was applied improves the fit results as the predicted counts in the vicinity of the thresholds, which are underestimated, are “manually” scaled up.

why did disabling the cuts on the rmf improve results?¶

This is pretty sure due to the full contribution of the energy dispersion to the predicted flux also beyond thresholds. See the following plot showing the predicted flux expanding beyond thresholds when the cut on the rmf is not done (here, effective area is zero beyond thresholds, how it is implemented in ctools).

For comparison the following plot shows the predicted flux when the cut on the rmf is done but the effective area is not set to zero beyond thresholds.

I fixed the application of the energy threshold in #2674. Does this also fix this issue?