bug in b-tagging module?

[Fabian Bach]

When I try to extend the default b-tagging module by another parametrization setup with results stored in BitNumber 1 (e. g., as in the two detector charts attached), I observe that the result of the original setup in BitNumber 0 is affected, which does not make sense to me. For example, in the single top s-channel sample attached (100 events, partonic t+b final state showered and fragmented with Pythia6), counting just the positive flags in BitNumber 0 of the Jet_BTag container entries, I find 72 b-tags after running with delphes_card_ATLAS.tcl and 66 with delphes_card_ATLAS2.tcl.

[Alexandre Mertens]

Hello,

The b-tagging algorithm is using random numbers, so what you observe is probably due to that.
We just throw a number n between 0 and 1 and the jet is b-tagged if the number is higher than the efficiency.

Can you check with more event and see if the differences get statistically smaller?

Alexandre

[Fabian Bach]

Hi Alexandre,

I checked with larger samples, and the difference is significant. I just reduced the size so that the hep file doesn't get too large. Moreover, when I just redefine the setup in BitNumber 0 to "perfect tagging" (i.e. btag probability =1, mistagging prob. =0), I find 168 flags after running delphes_card_ATLAS.tcl, and 172 on delphes_card_ATLAS2.tcl, which shouldn't depend on any randum number any more, right?

[Alexandre Mertens]

Hello,

The b-tagging is applied on each reconstructed jet. And the jet reconstruction depend on random numbers (for smearing and track efficiencies).
So you can check that the two cards will probably also provide you a different number of jet.

Does that answer you question?

Alexandre

[Fabian Bach]

Hi Alexandre,

thanks for the hint, the jet number is indeed different on different cards. However, whenever I rerun the same card I get identical results, so I assume that internally the random seed is somehow fixed, but either the seed or the random sequence changes when going to another card, so that results from the same setup differ between different cards. Is it possible to vary the random seed in order to check whether the results from the two cards are consistent within statistical fluctuations?

[Fabian Bach]

Just as a follow-up: since I didn't know how to vary the random seed (unfortunately, I discovered ticket #202 just today, after the announcement of v3.0.10), I simply divided my entire 2M events sample into 40 sub-samples of 50k events each and looked at the statistics of b-tagging results over these samples. It turns out that indeed the results in BitNumber 0 from both detector cards coincide within statistical fluctuations. Alternatively, using the hint of #202 I will check this to be true upon rerunning a small fixed sample several times with varying random seed, just for completeness.