Boosted Top Jet Tagging at CMS

Abstract

Multiple techniques for reconstructing highly Lorentz-boosted, hadronically-decaying top quarks are studied. These techniques, including the CMS Top Tagger, the HEP Top Tagger, and N-subjettiness, use jet substructure and jet mass observables to identify “top-jets”. The efficiency and misidentification rate of these algorithms are compared, both with and without identification of b-quark subjets. The efficiency is measured in data and in simulation, and the ratio of these measurements is presented as a data-simulation scale factor

Plots and tables from the paper JME-13-007 ( click on plot to get .pdf )

Simulation

Tagging variables

CMS Combined Tagger - tagging variables
	CA8 Jet mass for jets with pT > 500 GeV/c from a simulated tt POWHEG sample and from a simulated QCD PYTHIA 6 sample
	CA8 Jet mass after top tagging selections (Nsubjets ≥ 3, mmin > 50, τ3/τ2 < 0.55) for jets with pT > 500 GeV/c from a simulated tt POWHEG sample and from a simulated QCD PYTHIA 6 sample
	CMS Top Tagger minimum pairwise subjet mass for CA8 jets with pT > 500 GeV/c from a simulated tt POWHEG sample and from a simulated QCD PYTHIA 6 sample
	CMS Top Tagger minimum pairwise subjet mass after the N-subjettiness selection (τ3/τ2 < 0.55) for CA8 jets with pT > 500 GeV/c from a simulated tt POWHEG sample and from a simulated QCD PYTHIA 6 sample
	CA8 N-subjettiness ratio τ3/τ2 for jets with pT > 500 GeV/c from a simulated tt POWHEG sample and from a simulated QCD PYTHIA 6 sample
	CA8 N-subjettiness ratio τ3/τ2 after the CMS Top Tagger selection (Nsubjets ≥ 3, mmin > 50, 140 < mjet < 250 GeV/c2) for jets with pT > 500 GeV/c from a simulated tt POWHEG sample and from a simulated QCD PYTHIA 6 sample

HEP Combined Tagger - tagging variables
	HEP Top Tagger top jet mass (m123) for CA15 jets with pT > 200 GeV/c from a simulated tt MADGRAPH sample and from a simulated QCD PYTHIA 6 sample
	HEP Top Tagger top jet mass (m123) after the HEP Top Tagger W mass selection for CA15 jets with pT > 200 GeV/c from a simulated tt MADGRAPH sample and from a simulated QCD PYTHIA 6 sample
	CA15 N-subjettiness ratio τ3/τ2 for jets with pT > 200 GeV/c from a simulated tt POWHEG sample and from a simulated QCD PYTHIA 6 sample
	CA15 N-subjettiness ratio τ3/τ2 after the HEP Top Tagger top and W mass selections for jets with pT > 200 GeV/c from a simulated tt POWHEG sample and from a simulated QCD PYTHIA 6 sample
	Bi-dimensional distributions of m23/m123 vs. atan(m13/m12) for HEP Top Tagger subjets from CA15 jets with pT > 200 GeV/c for a simulated tt (MADGRAPH) sample
	Bi-dimensional distributions of m23/m123 vs. atan(m13/m12) for HEP Top Tagger subjets from CA15 jets with pT > 300 GeV/c for a simulated tt (MADGRAPH) sample
	Bi-dimensional distributions of m23/m123 vs. atan(m13/m12) for HEP Top Tagger subjets from CA15 jets with pT > 400 GeV/c for a simulated tt (MADGRAPH) sample
	Bi-dimensional distributions of m23/m123 vs. atan(m13/m12) for HEP Top Tagger subjets from CA15 jets with pT > 200 GeV/c for a simulated background sample (cross section weighted boson+jets, diboson, single-top, tt all-hadronic, and tt leptonic production)

ROC curves

ROC - low pT

Mistag rate vs. top-jet tagging efficiency for the HEP Top Tagger and HEP Com- bined Tagger for jets matched to generated partons with pT > 200 GeV/c2. The mistag rate is measured from QCD PYTHIA 6 Monte Carlo while the efficiency is measured with POWHEG tt Monte Carlo. R = 1.5 jets are used for all algorithms. A mass cut of 140 < m123 < 250 GeV/c2 is required. Signal jets are matched to simulated all-hadronic generated top quarks, while back- ground jets are matched to simulated partons from the hard scatter. The CMS Top Tagger and CMS Combined Tagger are not considered for this pT range because the decay products of tops with pT close to the 200 GeV/c are rarely merged within R=0.8 jets.

ROC - medium-high pT

Mistag rate vs. top-jet tagging efficiency as measured from QCD PYTHIA 6 Monte Carlo and POWHEG tt Monte Carlo, respectively. In the cases where a jet mass cut is applied, the cut is not varied and is fixed at 140 < mjet < 250GeV/c2. N-subjettiness is calculated using R = 0.8 jets except when used in combination with the HEP Top Tagger in which case R = 1.5 jets are used. Signal jets are matched to simulated all-hadronic generated top quarks, while background jets are matched to simulated partons from the hard scatter. Distributions are shown for three pT selections, where the pT cut is applied to the matched generated parton.

Working Point Definitions

	Working points for the HEP Combined Tagger
	Working points for the CMS Combined Tagger

Pileup

Slope of tagging efficiency and mistag rate (in percent) vs Nvtx for different tagging selections and pT > 500 GeV/c.

Data/MC comparison

CMS Combined Tagger

CMS Combined Tagger - semileptonic selection
	Number of subjets for full merged top jet candidate in the hadronic hemisphere after the semileptonic selection. tt is simulated with the MADGRAPH event generator. This distribution is used to evaluate the sequential top-tagging efficiency SF.
	Minimum subjet pairwise mass after the selection Nsubjets>=3 for full merged top jet candidates in the hadronic hemisphere after the semileptonic selection. tt is simulated with the MADGRAPH event generator. This distribution is used to evaluate the sequential top-tagging efficiency SF. The mmin distribution is not well modeled by the simulation. This effect may be due to mis-modeling of radiation within the top jet or merged subjets at very high jet momenta. The discrepancy is most evident at large jet pseudorapidity and therefore we choose to measure a pseudorapidity-dependent scale fac- tor.
	CA8 jet mass after the selections Nsubjets>=3 and mmin>50 for full merged top jet candidate in the hadronic hemisphere after the semileptonic selection. tt is simulated with the MADGRAPH event generator. This distribution is used to evaluate the sequential top-tagging efficiency SF.
	Nsubjettiness ratio τ3/τ2 after the selections Nsubjets>=3, mmin>50, 140<mjet<250 for full merged top jet candidate in the hadronic hemisphere after the semileptonic selection. tt is simulated with the MADGRAPH event generator. This distribution is used to evaluate the sequential top-tagging efficiency SF.
	Maximum subjet CSV b-discriminant after the selections Nsubjets>=3, mmin>50, 140<mjet<250, and τ3/τ2 < 0.55 for full merged top jet candidate in the hadronic hemisphere after the semileptonic selection. tt is simulated with the MADGRAPH event generator. This distribution is used to evaluate the sequential top-tagging efficiency SF.
	Jet mass for fully-merged top candidates from the semileptonic sample after selections Nsubjets>=3, mmin>50, 140<mjet<250, and τ3/τ2 < 0.55. tt is simulated with the MADGRAPH event generator.
	Jet mass for fully-merged top candidates from the semileptonic sample after selections Nsubjets>=3, mmin>50, 140<mjet<250, τ3/τ2 < 0.55 and CSV-medium b-tag. tt is simulated with the MADGRAPH event generator.

HEP Combined Tagger

CA15 jet pT
	Distribution of the pT of the Cambridge Aachen R=1.5 top candidate jet after all the selection cuts, but before applying the HEP Top Tagger. tt is simulated with the MADGRAPH event generator.

HEP Top Tagger top mass distribution (m123) - 3 pT regions
	HEP Top Tagger top mass distribution (m123) for CA15 jet pT > 200 GeV/c. tt is simulated with the MADGRAPH event generator. The m123 top mass peak is not well-modeled by simulation.
	HEP Top Tagger top mass distribution (m123) for CA15 jet pT > 300 GeV/c. tt is simulated with the MADGRAPH event generator. The m123 top mass peak is not well-modeled by simulation.
	HEP Top Tagger top mass distribution (m123) for CA15 jet pT > 400 GeV/c. tt is simulated with the MADGRAPH event generator. The m123 top mass peak is not well-modeled by simulation.

HEP Top Tagger top mass distribution (m123) after W mass selection - 3 pT regions
	HEP Top Tagger top mass distribution (m123) after the W mass selection but before the top mass selection for CA15 jet pT > 200 GeV/c. tt is simulated with the MADGRAPH event generator. The m123 top mass peak is not well-modeled by simulation.
	HEP Top Tagger top mass distribution (m123) after the W mass selection but before the top mass selection for CA15 jet pT > 300 GeV/c. tt is simulated with the MADGRAPH event generator. The m123 top mass peak is not well-modeled by simulation.
	HEP Top Tagger top mass distribution (m123) after the W mass selection but before the top mass selection for CA15 jet pT > 400 GeV/c. tt is simulated with the MADGRAPH event generator. The m123 top mass peak is not well-modeled by simulation.

HEP Top Tagger W mass selection
	Bi-dimensional distributions of m23/m123 vs. atan(m13/m12) for HEP Top Tagger jets in data (semi- leptonic selection). The area enclosed by the black lines denotes the region selected by the HEP Top Tagger W mass selection.

Type 2 tagger

	(a) Pruned jet mass of the leading jet in the hadronic hemisphere. This is the hadronic W boson candidate for partially merged “type 2” top quark candidates after the semileptonic selection. tt is simulated with the MADGRAPH event gen- erator. “NTMJ” represents non-top multijet backgrounds. These are measured in data by reversing the mass drop selection and normalizing through a fit to the HT distribution. The shaded regions represent the total uncertainty on the background model.
	Subjet Mass drop μ for the W boson candidate in the hadronic hemisphere for partially merged “type 2” top quark candidates after the semileptonic selection. tt is simulated with the MADGRAPH event gen- erator. “NTMJ” represents non-top multijet backgrounds. These are measured in data by reversing the mass drop selection and normalizing through a fit to the HT distribution. The shaded regions represent the total uncertainty on the background model.
	Pairwise mass of the W boson candidate and the closest jet in ∆R for partially merged “type 2” top quark candidates after the semileptonic selection. tt is simulated with the MADGRAPH event gen- erator. “NTMJ” represents non-top multijet backgrounds. These are measured in data by reversing the mass drop selection and normalizing through a fit to the HT distribution. The shaded regions represent the total uncertainty on the background model.

Scale Factors

	Efficiencies in data and simulation for each successive cut in the CMS Top Tagger, and for the addition of N-subjettiness and subjet b-tagging (CMS Combined Tagger WP3), as measured using the semileptonic selection. The denominator of each quoted efficiency is the number of events passing the previous selection. The efficiency is measured for three tt Monte Carlo generators (MADGRAPH, POWHEG, and MC@NLO).
	Data-simulation scale factors for three Monte Carlo generators for each successive cut in the CMS Top Tagger WP0, and for the addition of N-subjettiness and subjet b-tagging (CMS combined tagger WP3), as measured with the semileptonic selection. The scale factor measured in each row is with respect to the events passing the selection in the previous row. The scale factor is measured for three tt Monte Carlo generators (MADGRAPH, POWHEG, and MC@NLO). Only 21 events in the semileptonic selection in data survive the CMS and N-subjettiness selec- tions and therefore the subjet b-tagging scale factor measurement lacks statistics. If measured after the minmass selection or after the CMS Top Tagger selection the subjet b-tag scale factor is consistent with 1.
	Efficiencies in data and simulation for each successive cut in the CMS Top Tagger, and for the addition of N-subjettiness and subjet b-tagging (CMS Combined Tagger WP3), as measured using the semileptonic selection. The denominator of each quoted efficiency is the number of events passing the previous selection. The efficiency is measured for three tt Monte Carlo generators (MADGRAPH, POWHEG, and MC@NLO).
	Data-simulation scale factors after requiring all selections of the HEP Combined Tagger WP2 and after requiring all selections of the HEP Combined Tagger WP3. The scale factor is measured for three tt Monte Carlo generators in three pT regions and in two η regions. The scale factor is measured with the semileptonic selection with subjet b-tagging.
	Data-simulation scale factors for each sequential selection of the HEP Combined Tagger WP3. The scale factor measured in each row is with respect to the events passing the selection in the previous row. The scale factor is measured for three tt Monte Carlo generators in three pT regions and in two η regions. The scale factor is measured with the semileptonic selection with subjet b-tagging.

Supplementary plots ( click on plot to get .pdf )

Simulation plots

ROC curves

	Mistag rate vs. top-jet tagging efficiency for multiple combinations of subjet b-tagging and other top tagging algorithms, as measured from QCD PYTHIA 6 Monte Carlo and POWHEG tt Monte Carlo, respectively. In the cases where a jet mass cut is applied, the cut is not varied and is fixed at 140 < mjet < 250GeV/c2. Signal jets are matched to simulated all-hadronic generated top quarks, while background jets are matched to simulated partons from the hard scatter. Distributions are shown for three pT selections, where the pT cut is applied to the matched generated parton.
	Mistag rate vs. top-jet tagging efficiency for multiple combinations of N-subjettiness and other top tagging algorithms, as measured from QCD PYTHIA 6 Monte Carlo and POWHEG tt Monte Carlo, respectively. In the cases where a jet mass cut is applied, the cut is not varied and is fixed at 140 < mjet < 250GeV/c2. Signal jets are matched to simulated all-hadronic generated top quarks, while background jets are matched to simulated partons from the hard scatter. Distributions are shown for three pT selections, where the pT cut is applied to the matched generated parton.
	Mistag rate vs. top-jet tagging efficiency for multiple combinations of the HEP Top Tagger along with other top tagging algorithms, as measured from QCD PYTHIA 6 Monte Carlo and POWHEG tt Monte Carlo, respectively. In the cases where a jet mass cut is applied, the cut is not varied and is fixed at 140 < mjet < 250GeV/c2. Signal jets are matched to simulated all-hadronic generated top quarks, while background jets are matched to simulated partons from the hard scatter. Distributions are shown for three pT selections, where the pT cut is applied to the matched generated parton.

Mistag rate vs. top-jet tagging efficiency for the CMS Combined Tagger and HEP Combined Tagger, as measured from QCD PYTHIA 6 Monte Carlo and POWHEG tt Monte Carlo, respectively. Signal jets are matched to simulated all-hadronic generated top quarks, while background jets are matched to simulated partons from the hard scatter. Distributions are shown for four pT selections, where the pT cut is applied to the matched generated parton.

Data/MC comparison plots

CMS Combined Tagger

Minimum pairwise mass - Compare generators - 3 eta regions
	CMS Top Tagger subjet minimum pairwise mass for all eta (abs(eta)<2.4) for ttbar simulated with three event generators in addition to addional non-ttbar backgrounds.
	CMS Top Tagger subjet minimum pairwise mass for central eta (abs(eta)<1.0) for ttbar simulated with three event generators in addition to addional non-ttbar backgrounds.
	CMS Top Tagger subjet minimum pairwise mass for high eta (1.0<abs(eta)<2.4) for ttbar simulated with three event generators in addition to addional non-ttbar backgrounds.

Minimum pairwise mass in 2 eta regions - MADGRAPH
	CMS Top Tagger subjet minimum pairwise mass for central eta (abs(eta)<1.0) with the MADGRAPH event generator.
	CMS Top Tagger subjet minimum pairwise mass for high eta (1.0<abs(eta)<2.4) with the MADGRAPH event generator.

Minimum pairwise mass in 2 eta regions - POWHEG
	CMS Top Tagger subjet minimum pairwise mass for central eta (abs(eta)<1.0) with the POWHEG event generator.
	CMS Top Tagger subjet minimum pairwise mass for high eta (1.0<abs(eta)<2.4) with the POWHEG event generator.

Minimum pairwise mass in 2 eta regions - MC@NLO
	CMS Top Tagger subjet minimum pairwise mass for central eta (abs(eta)<1.0) with the MC@NLO event generator.
	CMS Top Tagger subjet minimum pairwise mass for high eta (1.0<abs(eta)<2.4) with the MC@NLO event generator.

Subjet pairwise mass - all eta - MADGRAPH
	CMS Top Tagger subjet pairwise mass m12 for all eta (abs(eta)<2.4). Subjets are numbered based on subjet pT, with subjet 1 being the leading subjet. Subjets are found by the CMS Top Tagger using CA8 jets with pT>400 and abs(eta)<2.4. tt is simulated with the MADGRAPH event generator.
	CMS Top Tagger subjet pairwise mass m13 for all eta (abs(eta)<2.4). Subjets are numbered based on subjet pT, with subjet 1 being the leading subjet. Subjets are found by the CMS Top Tagger using CA8 jets with pT>400 and abs(eta)<2.4. tt is simulated with the MADGRAPH event generator.
	CMS Top Tagger subjet pairwise mass m23 for all eta (abs(eta)<2.4). Subjets are numbered based on subjet pT, with subjet 1 being the leading subjet. Subjets are found by the CMS Top Tagger using CA8 jets with pT>400 and abs(eta)<2.4. tt is simulated with the MADGRAPH event generator.

Subjet pairwise mass - central eta - MADGRAPH
	CMS Top Tagger subjet pairwise mass m12 for central eta (abs(eta)<1.0). Subjets are numbered based on subjet pT, with subjet 1 being the leading subjet. Subjets are found by the CMS Top Tagger using CA8 jets with pT>400 and abs(eta)<1.0. tt is simulated with the MADGRAPH event generator.
	CMS Top Tagger subjet pairwise mass m13 for central eta (abs(eta)<1.0). Subjets are numbered based on subjet pT, with subjet 1 being the leading subjet. Subjets are found by the CMS Top Tagger using CA8 jets with pT>400 and abs(eta)<1.0. tt is simulated with the MADGRAPH event generator.
	CMS Top Tagger subjet pairwise mass m23 for central eta (abs(eta)<1.0). Subjets are numbered based on subjet pT, with subjet 1 being the leading subjet. Subjets are found by the CMS Top Tagger using CA8 jets with pT>400 and abs(eta)<1.0. tt is simulated with the MADGRAPH event generator.

Subjet pairwise mass - high eta - MADGRAPH
	CMS Top Tagger subjet pairwise mass m12 for high eta (1.0<abs(eta)<2.4). Subjets are numbered based on subjet pT, with subjet 1 being the leading subjet. Subjets are found by the CMS Top Tagger using CA8 jets with pT>400 and 1.0<abs(eta)<2.4. tt is simulated with the MADGRAPH event generator.
	CMS Top Tagger subjet pairwise mass m13 for high eta (1.0<abs(eta)<2.4). Subjets are numbered based on subjet pT, with subjet 1 being the leading subjet. Subjets are found by the CMS Top Tagger using CA8 jets with pT>400 and 1.0<abs(eta)<2.4. tt is simulated with the MADGRAPH event generator.
	CMS Top Tagger subjet pairwise mass m23 for high eta (1.0<abs(eta)<2.4). Subjets are numbered based on subjet pT, with subjet 1 being the leading subjet. Subjets are found by the CMS Top Tagger using CA8 jets with pT>400 and 1.0<abs(eta)<2.4. tt is simulated with the MADGRAPH event generator.

HEP Combined Tagger

HEP subjet pairwise mass - MADGRAPH
	Invariant mass of the combination of the two highest pT subjets in the CA15 Jet after a semileptonic selection
	Invariant mass of the combination of the pT-ordered subjets in position 1 and 3 in the CA15 jet after a semileptonic selection
	Invariant mass of the combination of the pT-ordered subjets in position 2 and 3 in the CA15 jet after a semileptonic selection

HEP W-tag variables - before W mass selection - MADGRAPH
	Distribution of the arctangent of the ratio of m_13 and m_12 for subjets of CA15 jets after a semileptonic selection, where m_ij is the invariant mass of the combination of the pT-oredered subjets in position i and j.
	Distribution of the ratio of m_23 and m_123 for subjets of CA15 jets after a semileptonic selection, where m_ij is the invariant mass of the combination of the pT-oredered subjets in position i and j, and m_123 is the jet mass obtained from the sum of the three subjets.
	Distribution of the invariant mass of the two subjets whose mass is closest to the W mass from CA15 Jets after a semileptonic selection.

HEP W-tag variables - after W mass selection - MADGRAPH
	Distribution of the arctangent of the ratio of m_13 and m_12 for subjets of CA15 jets after a semileptonic selection, where m_ij is the invariant mass of the combination of the pT-oredered subjets in position i and j. The HEP Top Tagger W mass selection has been applied.
	Distribution of the ratio of m_23 and m_123 for subjets of CA15 jets after a semileptonic selection, where m_ij is the invariant mass of the combination of the pT-oredered subjets in position i and j, and m_123 is the jet mass obtained from the sum of the three subjets. The HEP Top Tagger W mass selection has been applied.
	Distribution of the invariant mass of the two subjets whose mass is closest to the W mass from CA15 Jets after a semileptonic selection. The HEP Top Tagger W mass selection has been applied.

Diagrams and flowcharts

CMS Top Tagger
	Cartoon demonstrating the two stage application of the CMS Top Tagger decomposition procedure. The jet is declustered into A and B. A and B are then individually declustered into A' A'' and B' B''
	Flowchart of the CMS Top Tagger decomposition procedure
	An example decomposition using the CMS Top Tagger

HEP Top Tagger
	HEP Top Tagger algorithm description
	Labeled HEP Top Tagger W mass selection in bi-dimensional plane
	Flowchart for HEP Top Tagger mass drop decomposition procedure
	HEP Top Tagger cartoon 1 - mass drop decomposition. The circles represent PF constituents. The circle radius is proportional to the constituent momentum. The colors represent subjets found by the decomposition. Some constituents are discarded during the decomposition process.
	HEP Top Tagger cartoon 2 - consider all combinations of 3 mass drop subjets.
	HEP Top Tagger cartoon 3 - recluster each combination of three mass drop subjets. Distance parameter is determined by the separation between the two closest mass drop subjets in each combination. The hollow blue circles represent subjets found by the reclustering.
	HEP Top Tagger cartoon 4 - jet filtering. Keep only the 5 leading subjets.
	HEP Top Tagger cartoon 5 - filter all combinations of 3 mass drop subjets
	HEP Top Tagger cartoon 6 - keep the combination with filtered mass closest to the top mass. Recluster to force 3 subjets.

-- JamesDolen - 14 Jan 2014