TarekHC
commented
1 year ago I just wonder about the choice to not use true classification for background. I understand why the majority of events would go to the "worse" event type. However, I would expect that once the gamma/hadron cuts are applied, this bias would be "resolved" and the partition would be somewhat realistic. Is it not the case?
The difference between the true and reconstructed direction for a proton does not really make sense, no? Proton showers may look like gamma ones when there is a pion sub-shower that very likely will have a very different direction to the impinging proton. So ranking cosmic rays (even the gamma-like ones!) w.r.t. this difference we found does not make much sense...
Even for electrons, I'm not sure how strong is the effect, but the initial electron bending due to magnetic field will worsen the "true to reconstructed" distance. This is probably a very small effect... But again something its probably more realistic to evade.
orelgueta
These changes have been made with the purpose to test the calculation of true event types and try to make it more realistic. While doing this, 2 things have been noticed: 1) The partition of true event types should be independent from the actual partition, as we want to use true event types to test the room for improvement of the whole process. 2) True event types are not correctly defined for the background. We know that the direction is never well reconstructed for the background, as the previous analysis only expects it to be good for gammas. This means that partitioning true event types for background (by true misdirection) always ends up sending most of the background to the worst type, which affects heavily to the event-type-wise sensitivity computation, among others.
Both things have been fixed. The first one by having two partition matrices, one for reco and one for true. The second one by assuming the reconstructed event types for the background are good enough to use them directly as true event types (it seems not optimal, but without a definition, this is the best we can do to be able to compute realistic IRFs with "almost" true event types).