discrimination with converted photons Nicolas Chanon, Zhang Zhen Guoming Chen, Suzanne Gascon-Shotkin, Morgan Lethuillier 22/05/2009 – IHEP CMS weekly meeting I – Converted photons selection II – TMVA results
discrimination with converted photons I – Converted photons selection Samples : mH=120GeV H->, QCD Preselection at generator level : Et1>37.5 GeV, Et2>22.5 GeV (background only) Selection at reconstructed level : - Use of converted photon collection in CMSSW_1_6_12 and ask isConverted=1. - ||<2.5. - Events are selected if there is at least one photon with Et>40 GeV. TMVA is then applied to all photons which have Et>25 GeV - Tracker ISO : No tracks with pt>1.5 GeV inside ΔR<0.3 around the direction of the photon candidate. We consider tracks with hits in at least two layers of the silicon pixel detector. - Ecal ISO : Sum of Et of the ECAL basic clusters within 0.06<ΔR<0.35 around the direction of the photon candidate <6 GeV in barrel, <3 GeV in endcap. If one of the candidates is in endcap the other has to satisfy : Sum of Et of the ECAL<3 - Hcal ISO : Sum of Et of the HCAL towers within ΔR<0.3 around the direction of the photon candidate<6 GeV (5 GeV) in barrel (endcap)
discrimination with converted photons I – Converted photons selection Contents of the converted photon collection : Event N, converted photon collection : - convphot 1 : SC 1 : track 1 - convphot 2 : SC 1 : track 2 - convphot 3 : SC 1 : track 3 - convphot 4 : SC 2 : track 4 and track 5 - convphot 5 : SC 2 : track 4 and track 6 - convphot 6 : SC 2 : track 5 and track 6 - convphot 7 : SC 3 : nTracks=0, isConverted=0 => For each converted photon supercluster, how to select the best tracks ? candidate per SC candidate per SC 1trk Case 2trk Case Ntrk>=2 : Sg : 23% Bg : 29% Ntrk>=3 : Sg : 64% Bg : 62%
discrimination with converted photons I – Converted photons selection Some informations about the converted photon collection - The converted photon collection has dedicated tracks, which are not belonging to the usual tracks collection - In the case where many converted photons have only 1 track associated with the same SC, all tracks candidates have the same sign. - For the 2 tracks case, each pair has opposite charge.
discrimination with converted photons I – Converted photons selection Selection of the best tracks among all converted photon tracks candidates : - Previously, I was using EoverP closest to 1. - In last presentation and this one, I used Nhits minimum : improves slightly the 2 tracks case (but the 1track stay almost the same). Official way in CMSSW_2_X_X and CMSSW_3_X_X : - 1 track case : EoverP closest to 1 (which can be improved, because 1 track is lost so there is no reason for EoverP to be close to 1) - 2 track case : Use of a likelihood which takes as imput EoverP, abs(cot_theta), abs(delta_phi), chi2_max_pt, chi2_min_pt => No more available in CMSSW_1_6_12 ! I should move soon to 2_2_9.
discrimination with converted photons II – TMVA results Input variables Variables tried since last time : - curvature of the tracks - chi2/n.d.f. of the fit - dz, d0 : impact parameters - sBT : asymmetry top-bottom in the seed cluster - distance between the ecal impact position of the tracks - Z ecal impact position Removing variables in the 2 track case : - paireta : gives slightly better result than cotTheta => results shown use it. - Hcal isolation energy of the track => dR=0.1 gives the best result. Can replace trk34overEt (2 track case) with the same performance
discrimination with converted photons II – TMVA results 1 track case Results MLP Ranking result (top variable is best ranked) Rank : Variable : Importance 1 : conpho_cPP : 1.636e+06 2 : conpho_cEP : 6.603e+03 3 : conpho_s9/(s9-s1-s2) : 4.476e+00 4 : conpho_ptoverjetpt : 1.540e+00 5 : conpho_r9 : 7.695e-01 6 : conpho_closestSC_dR : 4.042e-01 7 : conpho_dR_SCtrkclosest : 2.999e-03
discrimination with converted photons II – TMVA results 1 track case Results : Pt bins - Divide in pt bin with almost the same number of photons - Try adding variable Et in the TMVA input => Outside [50,70] conpho_closestSC_dR falls at the last place Et added as input variable
discrimination with converted photons II – TMVA results 1 track case Results : Application For 90% signal efficiency, Gam+Jets rejection = 55% -0.34
discrimination with converted photons II – TMVA results 2 track case Results MLP Ranking result (top variable is best ranked) Rank : Variable : Importance 1 : conpho_cPP : 1.745e+06 2 : conpho_cEP : 9.233e+03 3 : conpho_SigmaPhi : 2.679e+01 4 :s9/(s9-s1-s2) : 5.172e+00 5 : conpho_ptoverjetpt : 4.247e+00 6 : conpho_paireta : 8.635e-01 7 : conpho_trk34overEt : 7.935e-01 8 : conpho_closestSC_dR : 2.951e-01 9 : conpho_dR_SCtrkclosest : 9.017e-02
discrimination with converted photons II – TMVA results 2 track case Results : Pt Bin Et added as input variable
discrimination with converted photons II – TMVA results 2 track case Results : Application For 90% signal efficiency, Gam+Jets rejection = 34% -0.32
discrimination with converted photons II – TMVA results Converted photon et 1 trk case 2 trk case - 50<Et<70 : more Higgs photons than QCD photons - GammaJet : very different pt shape. Should try the application depending on the pt (more photons in the tail, less photons with 25<Et<70)
discrimination with converted photons Conclusions : - 1 trk is still at 63% of background rejction, and 2 trk case at 60%. Some variables that increase the performance with VBF don't work for gluon fusion. - Still too many input variables in the 2 track case - Dividing sample in different pt bin has been done. The discrimination has the best performance for 50<Et<70. Elsewhere, conpho_closestSC_dR is less efficient and the performance is 10% less. - Application of the TMVA to Gamma+Jets samples has been done. The rejection rate is too low for the 2 track case. Perspectives : - Use Ted's likelihood to select the best converted photon per SC in the 2 tracks case. => migrate soon to CMSSW_2_2_9 (Hgg and Jets samples seems to be available). - Investigate TMVA application on 2 trck case on Gamma+Jet samples (try pt bins, non-prompt photons only, different MVA) - Combinate Pt bins to know the total background rejection - Check with MC truth info the proportion of real Pi0 in the background samples (later, with Nancy's implementation of conversion MC truth). - Investigate the asymmetry variable pointed out by Anagnostou