Higher density MC production for declination line 6166

[FrancaCassol]

The present density in the line is with a cos(0.04) spacing (see issue https://github.com/cta-observatory/lst-sim-config/issues/50). Unfortunately, this is not enough to avoid discontinuities in the reconstruction. See for example the following plots that show the constructed energy as function of the telescope zenith in the case of two data runs:

In the case of MAGIC VLZA (https://arxiv.org/pdf/2001.09566.pdf) the MC nodes have been produced on 100 bins in cos(zd) for zd=0-90:

We can try to improve the RF targets in order to make the results less dependent on cos(zd), however it is better to compare all this with a reference MC production with a high node density. For that, we ask to complete the declination line 6166 with the nodes in the following file, with a step of cos(zd)=0.01, both for test and for training.

dec_6166_zenith_azimuth_grid_step_cos_zd_eq_0.01_new_points.txt

This production will be used for the LZA study of src4, for which we have also several hours at low zd.

[maxnoe]

I will note here what I proposed on slack: to not have a hard cut in zd / cos(zd) in the random forest, which will always happen when using discrete values, we could try to randomize the zenith value, so that the random forest does not make hard splits in zenith.

[FrancaCassol]

Do you mean instead of giving the real "alt_tel" to give to give a randomized "alt_tel" value (adding an artificial uncertainty)? Instead, what about to make the RF in sequence and add to the energy RF the estimated photon direction (instead of alt_tel), this would automatically smear the zenith information over the camera field with a real physical meaning (but I am expert of RF, so there are perhaps some draw back in making a RF dependent on the result of the other?)

[moralejo]

I will note here what I proposed on slack: to not have a hard cut in zd / cos(zd) in the random forest, which will always happen when using discrete values, we could try to randomize the zenith value, so that the random forest does not make hard splits in zenith.

As we discussed elsewhere, this would indeed "iron out" the steps, and would be equivalent to "knowing" the telescope pointing with an accuracy just equal to the MC steps. Hence the RF would perform worse than it could, if trained with finer pointing steps.

Relative to what we do now, we would turn a (ZD-dependent) systematic error into a statistical one - not clear if that is better or not.