Update GammaDiffuse cuts for IRFs

[vuillaut]

Cuts for full enclosure IRFs to be produced using GammaDiffuse need to be updated as current ones result in statistics that are too low in some bins.

Adrvice from experts is required.

My main concern is about the IRFs from the new diffuse test gamma - statistics is not great, so we should find a compromise for the off-axis bins.

Originally posted by @moralejo in #513

[mizorafa]

Just FYI, below is the config that I use for the GalCenter analysis (for lstchain v0.10.12 though).

Since the LST-1 response is more or less flat within the current view cone (2.5 deg), the offset angles can be widely split (0.5-deg width adopted here). I have not quantified how well this binning is good, but overall there seems no critical problem regarding the MC statistics (like when visualizing the energy migration at an usual energy), although of course some bins especially at high energies are not enough (but the essential solution is only producing MC more). For this, a slightly wider migration binning are also used here.
As the dynamic gammaness cut is not offset dependent but dependent only on the (reconstructed) energy, I guess the cut itself is not a problem also for the diffuse MC.

{
    "EventSelector": {
      "filters": {
        "intensity": [200, Infinity],
        "leakage_intensity_width_2": [0, 0.2],
        "event_type": [32, 32]
      }
    },
    "DL3Cuts": {
      "min_event_p_en_bin": 100,
      "gh_efficiency": 0.6,
      "min_gh_cut": 0.7,
      "max_gh_cut": 0.9,
      "allowed_tels": [1]
    },
    "DataBinning": {
      "true_energy_min": 0.001,
      "true_energy_max": 1000,
      "true_energy_n_bins": 30,
      "scale_true_energy": 1.00,
      "reco_energy_min": 0.001,
      "reco_energy_max": 1000,
      "reco_energy_n_bins": 30,
      "energy_migration_min": 0.1,
      "energy_migration_max": 10,
      "energy_migration_n_bins": 40,
      "fov_offset_min": 0.0,
      "fov_offset_max": 2.5,
      "fov_offset_n_edges": 6,
      "bkg_fov_offset_min": 0.0,
      "bkg_fov_offset_max": 2.5,
      "bkg_fov_offset_n_edges": 6,
      "source_offset_min": 0,
      "source_offset_max": 1.2,
      "source_offset_n_edges": 81
    },
    "__NotApplicable__": {
        "global_gh_cut": 0.7,
        "global_alpha_cut": 10,
        "global_theta_cut": 0.2,
        "theta_containment": 0.68,
        "alpha_containment": 0.68,
        "min_theta_cut": 0.1,
        "max_theta_cut": 0.32,
        "fill_theta_cut": 0.32,
        "min_alpha_cut": 1,
        "max_alpha_cut": 20,
        "fill_alpha_cut": 20
    }

[moralejo]

Hi @mizorafa , thanks a lot for your input!
I agree with your comment about the off-axis binning. It is a pity that there is no option to use variable binning, or binning in offset^2, which would be way more reasonable. Is that a gammapy limitation, not just of our DL2 to DL3 tool?

Did you use half of the (original) diffuse training MC as "test MC"? And can you share the Aeffs you get (for the cuts above)? Would be good to see in which E-range the Aeff is not too noisy.

[mdebony]

In the test I've run up to now, even with quite large bin in offset, the statistics at high energy is usually poor and you could see large fluctuation in the effective area. You could put quite large energy bin, which are likely OK for high energy, but the interpolated behavior at low energy might be a bit too simplistic.
I was thinking one day to try with non regular binning (thin at low energy and wide a high energy) but didn't yet took the time to do it. I don't know how gammapy would handle it, for applying this idea you might need to re-interpolate on regular bin interval.