Fork me on GitHub

Opened 11 years ago

Last modified 11 years ago

#208 new How to

Possible problem with tau tagging efficiency

Reported by: simone gennai Owned by:
Priority: major Milestone:
Component: Delphes miscellaneous Version: Delphes 3
Keywords: Tau Efficiency Cc:

Description

Hi,
I am running some jobs with Delphes_3.0.10.pre7 with official CMS tcl cards and, independently of the PU scenario I am using, I cannot reproduce a flat tau efficiency as a function of eta and pT.
I am using the following formula:
EfficiencyFormula {15} {(pt <= 10.0) * (0.000) + \

(abs(eta)<4.0) * (pt > 10) * (0.65) +\
(abs(eta)>4.0) * (0.00)}

}

The efficiency in eta is farily flat but with a lower efficiency (~ 30 GeV) while the efficiency vs pT is growing linearly with the pT.
I have checked several times the code I am using to measure efficiency (# of tagged jets matched to gen level tau divided by the # of gen level taus) and I was not able to find any problem with that.
Any suggestion on other control plots to be made in order to dig further the problem would be welcome.

Best,

Simone

Change History (7)

comment:1 by Michele Selvaggi, 11 years ago

Hi Simone,

I have checked your formula on a Z->tautau sample and I get a flat efficiency.
My 2 cents are that maybe what you see is related to the way you are estimating the efficiency. The way the TauTagging module works in Delphes is the following: for each reconstructed jet we look if we can find a hadronically decayed tau-parton within a cone DR w.r.t the center of jet.
If it is found then we apply the efficiency (by shooting a random number).

The fact that you see a non-flat efficiency makes me think that you are actually incorporating in your calculation both the jet-parton matching efficiency (which is for sure increasing with pt) and the flat efficiency defined in your config card. In other words:

e = #(tau-tagged jets) / #(jets matched with gentau)
e' = #(tau-tagged jets) / # (gentaus)

(e) will give you what you have in your delphes card, (e') will probably give you something increasing vs pt.

Does this solve your issue?

Cheers,
Michele

Last edited 11 years ago by Michele Selvaggi (previous) (diff)

comment:2 by simone gennai, 11 years ago

Hi,
this indeed solves the issue.
I have tested and taking as a reference the matched jet to a gen tau the efficiency is indeed flat Vs eta and pT.

Unfortunately this opens another issue: why the matching efficiency has such a large turn on in pT and increases quite a lot at large eta?
This is not what is being seen in CMS and for sure with such a low matching efficiency, the actual implementation of the tau tagging cannot represent CMS performance.
I would suggest that (for CMS case) a smearing and efficiency are applied directly on the gen Tau, without passing through the jets.
A further cleaning of jets removing those matched to identified taus would be needed, though.

Thanx again for the help,
S.

comment:3 by Michele Selvaggi, 11 years ago

Hi again,

we have looked a bit into this and it seems that the "turn on" on the matching efficiency curve is largely due to the jet pt threshold. Due to neutrinos, reconstructed jets originating from taus have a much lower energy compared to their relative gen tau.

In the Delphes card you have a threshold which is set by default to 20 GeV. If you bring it down to a very low value (e.g. 1 GeV) you will see that the matching efficiency curve becomes flat again (and close to 100%).

Probably the proper way to do this would be to determine the correct jet energy scale, and apply it to reco jets. There is a module called EnergyScale that allows you to apply a (pt,eta) dependent energy scale factor to reconstructed jets.

Besides this, we can always, of course, write the module you have requested.

Michele

comment:4 by simone gennai, 11 years ago

Hi,
indeed lowering the jet pT helps.
In fact already at 10 GeV the turn on is much steeper.
Anyway digging deeper I have noticed that the real problem was how the reference generator level tau was defined.
In CMS we define the reference MC tau as the visible part of the tau decay, i.e. (genLevel P4 - neutrino P4).
When you take this visible decay product even with a jet pT the turn on is quite fast (the energy scale of the jet is infact much nearer to the one of the visible tau products than to the original tau).
Applying a pT cut of 20 GeV (the same applied to the offline jet) on the visible tau products gives a flat matching efficiency wrt pT and eta.
If I am not wrong in the code you are using the full genlevel tau four vector. This may give some mis-matching specially with high density of jets. Do you think you could change the code to use the visible tau decay products when requesting the matching? This would give a matching efficiency of very near 100% regardless the jet pT cut.
In this way no other changes (to the jet energy scale as an example would be needed).

Best,

S.

comment:5 by Pavel Demin, 11 years ago

Hi Simone,

Thank you very much for your comments and suggestions.

I've just modified the TauTagging module to use the visible tau decay products as you suggest.

Here is a preliminary version with this modification:

http://cp3.irmp.ucl.ac.be/downloads/Delphes-3.0.10.pre9.tar.gz

Regards,

Pavel

comment:6 by simone gennai, 11 years ago

Hi,
I have checked and for low PU the results looks fine and as expected (the matching efficiency is quite high).
I am now testing with large PU (140). For that case a tuning of the dR matching between taus and jets could be needed, but this can happen on the config. file level.

Best,

S.

comment:7 by Pavel Demin, 11 years ago

The modified the TauTagging module that calculates tau momentum as a sum of the momenta of the visible tau decay products is available in new release 3.0.10.

Note: See TracTickets for help on using tickets.