A study of heavy flavor quarks produced in association with top quark pairs at sqrt(s) = 7 TeV using the ATLAS detector

Using a sample of dilepton top-quark pair (ttbar) candidate events, a study is performed of the production of top-quark pairs together with heavy-flavor (HF) quarks, ttbar+b+X or ttbar+c+X, collectively referred to as ttbar+HF. The dataset used corresponds to an integrated luminosity of 4.7 fb-1 of proton--proton collisions at a center-of-mass energy of 7 TeV recorded by the ATLAS detector at the CERN Large Hadron Collider. The presence of additional HF quarks in the ttbar sample is inferred by looking for events with at least three b-tagged jets, where two are attributed to the b-quarks from the ttbar decays and the third to additional HF production. The dominant background to ttbar+HF in this sample is ttbar+jet events in which a light-flavor jet is misidentified as a heavy-flavor jet. To determine the heavy- and light-flavor content of the additional b-tagged jets, a fit to the vertex mass distribution of b-tagged jets in the sample is performed. The result of the fit shows that 79 +/- 14 (stat.) +/- 22 (syst.) of the 105 selected extra b-tagged jets originate from HF quarks, three standard deviations away from the hypothesis of zero ttbar+HF production. The result for extra HF production is quoted as a ratio (R_HF) of the cross section for ttbar+HF production to the cross section for ttbar production with at least one additional jet. Both cross sections are measured in a fiducial kinematic region within the ATLAS acceptance. R_HF is measured to be [6.2 +/- 1.1 (stat.) +/- 1.8 (syst.)]% for jets with pT>25 GeV and |eta|<2.5, in agreement with the expectations from Monte Carlo generators.


I. INTRODUCTION
In order to characterize the recently observed Higgslike particle (H) [1,2], quantities such as the Yukawa coupling of the top quark and the Higgs boson need to be measured with precision. For a Standard Model (SM) Higgs boson with a mass of 125 GeV, the decay mode with the largest branching ratio is H → bb. Thus, the channel with the largest yields for studying tt + H production is tt + H, H → bb. Production of topquark pair (tt) events featuring additional heavy-flavor (HF) b-and c-quarks, tt + b + X and tt + c + X, referred to as tt + HF, is the main irreducible background to tt + H, H → bb. A study of tt + HF production is useful to constrain models of heavy-flavor quark production at the scale of the top-quark mass. This analysis is also of interest because of the many potential phenomena beyond the SM, such as composite Higgs models [3] and processes leading to final states with four top quarks [4][5][6][7][8][9], that could produce additional heavy-flavor quarks in the tt candidate sample. This paper describes a study of tt + HF production. Within the SM, heavy-flavor quark pairs, cc and bb, are expected to be produced in association with tt mainly via gluon splitting from initial-and final-state radiation [10]. In addition, the heavy-flavor content of the proton could lead to tt final states with at least one additional HF quark, tt + c and tt + b. The data analyzed correspond to an integrated luminosity of 4.7 fb −1 at a center-of-mass energy of √ s = 7 TeV produced at the Large Hadron Collider (LHC) and recorded in 2011 with the ATLAS detector.
This analysis is performed on tt dilepton candidate events in which each top quark decays to a b-quark and a W boson, which subsequently decays to a neutrino and an isolated, charged lepton. The dilepton signature is selected for this measurement because it is relatively background free and precludes a third b-tagged jet from a hadronically decaying W boson, predominantly via W → sc. The tt + HF signal region is the subset of these events with three or more jets identified as containing HF quarks (b-tagged jets, or b-tags). However, jets without HF quarks may also be b-tagged, so that care must be taken to properly identify the flavor composition of the b-tagged jets in the sample. Two b-tagged jets from each event are presumed to originate from the b-quarks from top-quark decays, tt → W + bW −b . Therefore, all events in the signal region have at least one additional b-tag either from a b-or c-quark jet, or from a light-quark or gluon jet that was misidentified. The latter are referred to as light-flavor or LF jets.
Due to limited data statistics and discrimination between b-and c-jets, the sum of b-quark and c-quark jet rates is measured. Information about the composition of tt + b + X and tt + c+ X in tt+ HF is nevertheless required for the total correction due to acceptance, which is different for b-and c-quark jets. The composition is estimated with Monte Carlo simulation and tested in the data.
From the measurement of the fraction of jets with heavy flavor content, the cross section for tt production with at least one additional HF jet can be extracted. To reduce some systematic uncertainties, the result is quoted as a ratio, termed R HF , of the cross section for tt production with at least one additional HF jet to the cross section for tt production with at least one additional jet (tt + j), regardless of flavor. The measurement of tt + j production is performed in dilepton tt candidate events with at least three jets, at least two of which are b-tagged and assumed to come from topquark decays.
The paper is organized as follows. The ATLAS detector is briefly described in Sec. II. The data and Monte Carlo samples used in the analysis are described in Sec. III, followed by a description of the event selection in Sec. IV. The definition of the fiducial phase space used in the measurement of R HF , and the calculation of acceptances and efficiencies are presented in Sec. V. In Sec. VI, observed and expected numbers of events with ≥ 3 b-tagged jets are shown. Section VII describes a fit to the vertex mass distribution of b-tagged jets in these data events to extract the fraction of HF jets produced in association with tt. A discussion of the systematic uncertainties of the measurement is presented in Sec. VIII. Section IX shows the result of the measurements, followed by conclusions in Sec. X.

II. THE ATLAS DETECTOR
A detailed description of the ATLAS detector can be found elsewhere [11]. The innermost part of the detector is a tracking system that is immersed in a 2 T axial magnetic field and that measures the momentum of charged particles. The inner detector comprises a silicon pixel detector, a silicon microstrip detector, and a transition radiation tracker, providing tracking capability within the pseudorapidity range |η| < 2.5 [12]. The tracking system is also used to identify the displaced secondary vertex that is formed by hadrons containing a b-or c-quark. Calorimeter systems, which measure the electron, photon, and hadron energies, reside outside the inner detector and cover the region |η| < 4.9. Outside the calorimeters there is a muon spectrometer that is used to identify and measure the momentum of muons in an azimuthal magnetic field in the region |η| < 2.7. To reduce the data rate, a three-level trigger system selects the potentially interesting events that are recorded for offline analysis.

III. DATA AND MONTE CARLO SAMPLES
The total integrated luminosity for the analyzed data sample is 4.7 fb −1 at a center-of-mass energy of √ s = 7 TeV. During the 2011 data-taking period the instantaneous luminosity of the LHC increased, causing the average number of simultaneous inelastic pp interactions per beam crossing (pile-up) at the beginning of a pp fill to increase from about 6 to 17. Multiple pp interactions can occur either in the same bunch crossing as the primary vertex (termed 'in-time pile-up') or in an adjacent bunch crossing (termed 'out-of-time pile-up'). To account for these effects, all Monte Carlo simulated events are overlaid with additional inelastic events generated with the PYTHIA AMBT1 tune [13], and the distribution of the number of vertices in the simulation is reweighted to match the distribution of the number of additional interactions per bunch crossing measured in the data.
Monte Carlo simulation is used to study signal and background processes. Inclusive tt production and dedicated tt + HF samples are simulated using the multi-leg matrix-element generator ALPGEN v2.13 [14] with the CTEQ6L1 [15] parton distribution function (PDF) set. Parton showering and hadronization are performed by HERWIG v6.520 [16]. Effects due to the mass of the heavy-flavor quarks are included by default in ALPGEN. In these samples, additional jets (including heavy-flavor) can also be produced in the parton shower. The MLM [14] parton-jet matching scheme is applied to avoid double counting of configurations generated by both the parton shower and the leading-order (LO) matrix-element (ME) calculation. In addition, overlap between tt events with HF quarks that originate from ME production and those that originate from the parton shower is removed. This heavy flavor overlap removal (HFOR) is based on the ∆R qq [17] between simulated HF quarks. The event is taken from the ME calculation if it contains two well separated HF quarks (∆R qq > 0.4). The event is taken from the parton shower calculation if it contains two collinear HF quarks (∆R qq < 0.4).
To study the effect of different fixed-order calculations and matching schemes, samples of top-quark pair events are also generated using POWHEG v1.01 and showered with HERWIG. In this sample the tt process is described at next-to-leading order (NLO), while the extra jets are described at LO. For each sample showered with HERWIG, JIMMY v4.31 [18] and the AUET1 tune [19] are used to simulate the underlying event and to model various soft interactions. To assess the effect of different parton shower models, a sample is generated using ALPGEN v2.14 with the PYTHIA v6.425 [20] parton shower and hadronization, using the CTEQ5L PDF set [21]. The uncertainty associated with the CTEQ6L1 PDF set is evaluated with an envelope calculated using the uncertainty set from the NLO PDF MTSW2008nlo68cl [22], and an additional term to account for the difference between the central values of the LO and NLO calculations.
Initial-and final-state radiation (ISR/FSR) variations are studied using samples generated with AC-ERMC v2.0 [23] interfaced with PYTHIA v6.2. In these samples the parameters that control the amount of ISR/FSR are set to points consistent with the PE-RUGIA Hard/Soft tune [24] in a range constrained by current experimental data [25].
In all samples the top-quark mass is set to m t = 172.5 GeV. The cross section for Standard Model tt production at this mass is calculated using the approximate next-to-next-to-leading-order (NNLO) QCD calculation described in [26].
Background samples from the production of W and Z bosons are generated using the CTEQ6L1 PDFs with ALPGEN, which is interfaced to HERWIG for parton showering and hadronization; the ALPGEN matrix elements include diagrams with up to five additional partons. Separate samples of W + bb and Z + bb events are generated. The overlap between jets from the partonshower and the matrix-element in the n and n + 1 jet multiplicity samples is removed for the W +jets and Z+jets samples in the same manner as for the tt samples. Single top-quark production is modeled using ACERMC in the t-channel and MC@NLO v3.41 [27] for the W t-and s-channels. Diboson (WW , W Z, and ZZ) production is modeled using ALPGEN interfaced with HERWIG. Less than 0.5% of the expected yield in the tt + HF sample comes from the associated production of tt + W /Z and tt + H, and these processes are thus neglected in this analysis.
The resulting generated samples are passed through a GEANT4 simulation [28] of the ATLAS detector [29]. Events are then reconstructed in the same manner as the data.

IV. EVENT SELECTION
Events for the analysis are selected by at least one of the high-p T [12] single-electron or single-muon triggers, as described in Refs. [30] and [31]. The singleelectron triggers are based on calorimeter energy deposits, shower shape, and matching track quality constraints, while the single-muon triggers are based on a reconstructed track in the muon spectrometer that matches a track found in the inner detector. To ensure a final trigger rate that is compatible with the ATLAS data acquisition system, a minimum p T threshold for the electron and muon triggers is used. The p T threshold for the muon trigger is 18 GeV. For the electron trigger, the threshold is 20 GeV or 22 GeV, depending on the data-taking period due to varying LHC luminosity conditions.
The selected events are required to contain a reconstructed primary vertex with at least five associated tracks with p T > 0.4 GeV. Event reconstruction makes use of electrons (e), muons (µ), jets, and missing transverse momentum (E miss T ). Electrons are reconstructed by matching energy deposits in the electromagnetic calorimeter with tracks in the inner detector, and are required to have p T > 25 GeV and |η| < 2.47, excluding the transition region between the barrel and endcap calorimeters at 1.37< |η| <1.52 [32]. Muons are reconstructed by matching tracks in the inner detector with tracks measured in the muon spectrometer, and are required to have p T > 20 GeV and |η| < 2.5.
Tight isolation cuts are applied to both the electron and muon candidates to reduce the number of identified leptons (e, µ) that come from non-prompt (non-W/Z) sources and from misidentified hadrons. For electrons, the E T deposited in the calorimeter cells in a cone in η-φ space of radius ∆R = 0.2 around the electron position is summed, and the E T due to the electron is subtracted. The scalar sum of track transverse momenta in a cone of ∆R = 0.3, excluding the electron, is also measured. Cuts parametrized by the electron η and E T are made on these two isolation variables to ensure a constant efficiency over the entire (η, E T ) range. For muons, the corresponding calorimeter isolation energy in a cone of ∆R = 0.2 is required to be less than 4 GeV, and the scalar sum of track transverse momenta in a cone of ∆R = 0.3 is required to be less than 2.5 GeV after subtraction of the muon p T .
Jets are reconstructed from clustered energy deposits in the calorimeters with the anti-k t [33] algorithm with a radius parameter R = 0.4 [34]. Jets selected for the analysis are required to have p T > 25 GeV and |η| < 2.5. In order to reduce the background from jets originating from pile-up interactions, additional selection criteria are applied to the fraction of the jet's p T (JV F) carried by tracks originating from the primary vertex, JV F > 0.75.
The transverse momentum of neutrinos produced in the top-quark decays, measured as E miss T , is inferred by balancing the vector sum of all visible transverse momenta. Specifically, the E miss T is constructed from the vector sum of all calorimeter cell energies contained in topological clusters [34] with |η| < 4.5, projected onto the transverse plane. Contributions to the E miss T from the calorimeter cells associated with jets are taken at the corrected energy scale that is used for jets, while the contribution from cells associated with electrons is substituted by the calibrated transverse momentum of the electron. The contribution to the E miss T from the p T of muons passing the selection requirements is also included.
The b-tagging algorithm [35,36] employed for this analysis uses impact parameter and vertex position measurements from tracks in the inner detector as inputs to a neural network. The b-tagging efficiency was calibrated in a multi-jet data sample where at least one jet contains a muon [36]. The c-tagging efficiency was calibrated in a data sample with reconstructed D * mesons [37]. For this analysis, b-tagged jets are required to satisfy a selection that is 75% efficient for b-quark jets, approximately 30% efficient for c-quark jets, and rejects light-flavor jets by a factor of approximately 35 in simulated tt events. In this paper, a 'b-tag' (or a 'b-tagged jet') refers to any jet passing this selection, regardless of flavor. A 'b-jet', by contrast, refers to a jet (which may or may not be b-tagged) which contains a b-quark. Similarly, 'c-jet' and 'HF jet' are statements of the flavor composition of the jet, not whether the jet is b-tagged. Three distinct subsets of the selected b-tagged jets with different b-jet purity are used in the measurement of σ fid (tt + HF), as described in Sec. VII.
Dilepton tt candidate events are selected by requir- ing exactly two opposite-sign leptons and at least two jets. To reduce the background from Z/γ * processes, events with like-flavor leptons are required to have E miss T above 60 GeV and a dilepton invariant mass satisfying |m ℓ + ℓ − − m Z | > 10 GeV. For events with one electron and one muon, the scalar sum of the lepton and jet transverse momenta is required to be above 130 GeV to reduce the backgrounds from Z/γ * → τ + τ − , as well as WW , W Z, and ZZ processes. This set of selection criteria is termed the 'nominal' tt selection criteria. The measurement of tt + HF production is carried out in the subset of these events that contain three or more b-tagged jets, whereas the measurement of tt production with at least one additional jet is performed in the subset with at least three jets, at least two of which are b-tagged.
Using the nominal selection criteria described above, data and Monte Carlo events are compared in three control regions: dilepton tt candidate events with zero, one, or two b-tagged jets. Data-to-simulation normalization corrections are applied to Monte Carlo simulation samples when calculating acceptances to account for observed differences in predicted and observed trigger and lepton reconstruction efficiencies, jet flavor tagging efficiencies and mistag rates, as well as jet and lepton energy scales and resolutions. In Fig. 1, the jet multiplicity distributions in the three regions are compared to Monte Carlo predictions. Agreement is observed within uncertainties.

V. DEFINITION OF THE FIDUCIAL PHASE SPACE AND CALCULATION OF CORRECTION FACTORS
To allow comparison of the analysis results to theoretical predictions, the measurement is made within a fiducial phase space. The fiducial volume is defined in Monte Carlo simulation by requiring two leptons (e, µ) from the t → W b → ℓνb decays (including electrons and muons coming from τ → ℓνν τ ) with p T > 25 (20) GeV for e (µ), and |η| < 2.5 as well as three or more jets with p T > 25 GeV and |η| < 2.5.
In the simulation, jets are formed by considering all particles with a lifetime longer than 10 ps, excluding muons and neutrinos. Particles arising from pile-up interactions are not considered. For the determination of the tt + HF fiducial cross section, σ fid (tt + HF), three or more jets are required to match a b-or c-quark, two of which must match a b-quark from top-quark decay. All simulated b-and c-quarks that were generated with p T > 5 GeV are considered for the matching, and are required to satisfy ∆R(quark, jet) < 0.25. Jets that match both a b-and a c-quark are considered as b-jets. For the calculation of σ fid (tt + j) three or more jets are required, two of which must contain a b-quark from top-quark decay.
Each fiducial cross section is determined using measured quantities from the data, and a correction factor derived from the Monte Carlo simulation. The ratio of cross sections is defined as: The fiducial cross section for tt + HF production is determined from: where N HF is the number, after background subtraction, of b-tags from HF jets observed in the data, in addition to the two b-jets from top-quark decays. The integrated luminosity of the sample is denoted as L dt, and ε HF is a correction factor taken from Monte Carlo simulation that converts the number of observed b-tags from additional HF jets to the number of events in the signal fiducial volume. This correction factor includes the acceptance within the fiducial region, the reconstruction efficiency, and a factor to account for the multiplicity of extra b-tagged HF jets per tt+HF event in the signal region. This correction factor is different for tt + b + X (ε b ) and tt + c + X (ε c ), and thus ε HF is determined as a weighted sum of these two contributions. The weight used to form the sum is the fraction of tt + HF events in the fiducial volume which contain additional b-jets as opposed to c-jets. This fraction is termed F b/HF . The total correction factor (ε HF ) is calculated as: The denominator for R HF , σ fid (tt + j), is computed using a similar prescription: where N j is the yield of dilepton events in data with at least three jets, at least two of which are b-tagged, and ε j is the tt + j acceptance factor calculated from the Monte Carlo simulation. The acceptance calculation for each fiducial cross section assumes that all b-tagged jets are from real HF quarks. Events with b-tagged jets from LF quarks are treated as a background, and subtracted when computing both N HF and N j .
The prediction for R HF from the ALPGEN + HER-WIG Monte Carlo sample is 3.4%. The value obtained from the POWHEG v1.01 [38] generator showered with HERWIG [16] is R HF = 5.2%, with F b/HF = 0.34. While this R HF value is different to that from ALPGEN + HERWIG, the predicted F b/HF values are similar. Furthermore, a parton-level study using MAD-GRAPH5 v1.47 [39] gives F b/HF = 0.29. The value of F b/HF is also stable when different showering algorithms are used: the ALPGEN + PYTHIA Monte Carlo sample predicts a value of F b/HF = 0.32, in good agreement with the prediction when HERWIG is used. Based on comparison of these predictions for F b/HF , a symmetric 10% Monte Carlo systematic uncertainty is assigned, F b/HF = 0.31 ± 0.03. The prediction of F b/HF is also tested in data (see Sec. IX). Table I shows the number of events with ≥ 3 btagged jets expected in the Monte Carlo simulation from dilepton tt production and from various background sources. At this point, no distinction is made between events with a true additional HF jet and those containing a mistagged LF jet. The number of observed events is also shown. While Monte Carlo simulation is used to estimate tt + HF event rates and kinematic features, data-driven methods and Monte Carlo simulation are both used to estimate background processes, as detailed below.

VI. EXPECTED SIGNAL AND BACKGROUND YIELDS
Background processes containing real b-jets and leptons, such as single top-quark, Z/γ * + jets, and diboson (WW , W Z, and ZZ) production, are estimated using Monte Carlo simulation. Contributions from diboson production are found to be negligible.
A major source of background comes from tt events in which one or more of the b-tagged jets is from a mistagged LF jet. This background is estimated using Monte Carlo simulation for the measurement of σ fid (tt + j). However, in the measurement of σ fid (tt + HF), the final tt + LF background is determined by a fit to the vertex mass distribution of b-tagged jets in data, as explained in Sec. VII.
Background from events in which at least one of the leptons is either non-prompt (originating from e.g. a photon conversion or b-quark decay) or is a misidentified hadron, is estimated using data and Monte Carlo simulation. For instance, W + jets, multi-jet, and tt events with one hadronically decaying W boson can contribute in this way. This contribution is determined by scaling the yield of events in the data with a pair of same-sign leptons by the ratio of oppositesign to same-sign yields (R OS/SS ) obtained in Monte Carlo simulation. The opposite-sign to same-sign ratio is determined separately for the three dilepton channels, and found to be 1.3 ± 0.1 (stat.) +1.8 −1.3 (syst.) for e + e − events, 1.2 ± 0.1 (stat.) ± 0.7 (syst.) for µ + µ − events, and 1.2 ± 0.1 (stat.) ± 0.5 (syst.) for events with one electron and one muon. The systematic uncertainty takes into account the unknown relative mixture of fake-lepton sources (photon conversions, band c-hadron decays, or misidentified hadrons) in the R OS/SS calculation. Since the central value of the prediction for this background is zero events, only variations in R OS/SS that lead to larger background predictions are considered in the systematic uncertainty calculation. This method for estimating the background due to events with fake leptons is validated in a control sample of dilepton events with less restrictive lepton identification requirements and no isolation criteria.
The dominant uncertainties on the total yield in Table I come from the jet energy scale, b-tagging efficiency, parton showering model, and initial-and final- state radiation.

VII. TEMPLATE FIT
For the measurement of σ fid (tt + HF), the fraction of heavy-flavor jets produced in association with tt is extracted by performing a binned maximum-likelihood fit on the displaced-vertex mass distribution using all b-tagged jets in the events with ≥ 3 b-tagged jets. Although the final result is for both flavors combined, the fit includes separate b-and c-quark components to improve the determination of the LF fraction, and to test the Monte Carlo prediction for F b/HF , which is used for the calculation of the correction factor described in Sec. V. This displaced-vertex mass is constructed from the inner detector tracks associated with the secondary vertex using the algorithm described in Ref. [40]. While the presence of a displaced vertex is an indication that a jet contains a b-quark, a jet may be b-tagged even if no vertex is reconstructed. In this case, the vertex mass is undefined. These jets are assigned a mass value of '−1 GeV' and they are included in the fit to the displaced-vertex mass distribution. Keeping the events without a reconstructed vertex improves the discrimination between heavy-flavor and light-flavor jets.
While the vertex mass is a powerful discriminant, Monte Carlo studies indicate that the sensitivity on the fitted fraction of LF jets increases when the jet p T is used as an additional discriminant. Considering only the statistical uncertainty, it is seen that a fit with both jet p T and vertex mass is approximately half a standard deviation more sensitive than a fit with only the vertex mass. It was thus decided to define a two-dimensional probability density function, termed a 'template,' for the fit using the vertex mass and jet p T .
The fit is performed simultaneously in three mutually exclusive bins of b-jet purity, defined by different ranges of the b-tagging neural network output value. Certain values of the neural network output, termed 'operating points', are defined by the average b-jet selection efficiency resulting from the applied selection. In this analysis, operating points of 60%, 70% and 75% efficiency are used to define the boundaries of the b-jet purity bins. The first bin uses only the tightest calibrated operating point (60%), contains the highest-purity sample of b-jets (referred to as 'high purity'), and has a btagging efficiency of 60% for b-jets. The second bin (referred to as 'medium purity') requires a b-tag selection between the tightest and second tightest (70%) operating points, and contains a larger fraction of LF jets and c-jets. The efficiency for this bin is 10% for b-jets, i.e. the difference between the 70% and 60% operating points. The final bin ('low purity') requires a b-tag selection between the second (70%) and third operating point (75%), and contains the largest fraction of LF jets. The efficiency for this bin is 5% for b-jets. The b-tagging efficiencies for b-jets, c-jets, and light-flavor jets for each selection are given in Table II.
All three classes of b-tag purity are used in the analysis so that a jet is considered 'b-tagged' if it satisfies any of these criteria. The discrimination power between LF and c-jets is greatly improved by using three (as opposed to one) classes of b-purity. The vertex mass distributions for all b-tagged jets in events passing the nominal tt selection criteria are shown in Fig. 2 to confirm that (a) the data are well described by the Monte Carlo simulation, (b) and the b-jet, c-jet and LFjet fractions are different in the three purity selections. For the purpose of illustration, the normalization of the b-jet, c-jet, and LF-jet components is taken from Monte Carlo simulation.
The template fit has five components: b-jets from top-quark decays, non-tt background, extra b-tagged jets from b-quarks, extra b-tagged jets from c-quarks, and light-flavor b-tagged jets. The template for b-jets from top-quark decays is obtained from the data in tt dilepton events with exactly two b-tags. Monte Carlo simulation indicates that 97% of b-tagged jets in tt dilepton events with exactly two b-tags come from the decay of the top quark. To account for this in the shape of the data template, a template for b-tags not from the top-quark decays is derived from the tt Monte Carlo simulation, and subtracted with a 3% relative normalization from the data template. In the fit, the normalization for the template for b-jets from the top-quark decays is fixed assuming it contributes two of the three or more b-tags per observed event. Background events from non-dilepton tt processes are included using Monte Carlo simulation, and enter the fit with a fixed normalization. Monte Carlo simulation is used to obtain templates for additional (nont → W b) b-jets, c-jets, and LF jets.
In the fit to determine the number of b-tags from HF jets in addition to the two b-jets from top-quark decay, N HF , separate templates for each category of jet in each of the three purity classes (high, medium, and low) are used. The b-tagging efficiencies (Table II) for each flavor of jet are used to relate the number of jets in each purity bin. After the application of all constraints, the fit has two floating parameters: the fraction of LF jets and the fraction of additional b-jets. The fraction of additional c-jets makes up the remainder.
Monte Carlo pseudo-experiments show that the fitting method is unbiased in both best-fit values and estimated uncertainties. The fit strategy (including estimates of statistical and systematic uncertainties) was verified using 10% of the full data sample as well as with Monte Carlo pseudo-experiments before the fit was performed on the full data sample. These studies indicated that the fit could achieve only a 1σ separation of b-vs. c-jets based on the expected statistical uncertainty alone. Inclusion of the systematic uncertainty would further reduce the sensitivity. However, the LF-jet fraction is expected to be measured with sufficient precision to give a statistically significant measurement of the total HF content, defined as the fraction of additional b-tagged jets not coming from LF jets. In the fit, the individual fractions are not constrained to be positive or below unity.

VIII. SYSTEMATIC UNCERTAINTIES
Systematic uncertainties may affect the shape of the vertex mass and p T templates as well as the acceptance calculations. For the systematic uncertainties on the template shapes, the fit to the data is re-evaluated using new templates, derived by varying the relevant parameters by their systematic uncertainties, and a new fit to the data is performed. Major uncertainties that affect the fit are the jet energy scale and resolution, the tagging efficiencies for b-, c-and LF jets, the partonshower and hadronization models, and the Monte Carlo event generators.
The template for b-jets from top-quark decays is nominally taken from the data with exactly two btags. To account for kinematic biases due to additional heavy-flavor jets in the event, a systematic uncertainty on the shape of this template is assessed using b-jets from top-quark decays from Monte Carlo inclusive tt events with three or more b-tagged jets.
The vertex mass of additional b-and c-jets is sensitive to the number of HF quarks contained in a jet (for instance, for bb or cc produced via gluon splitting). The dominant uncertainty from this effect would manifest itself as a difference in the shape of the template for additional b-jets. To assess this uncertainty, the template for additional b-jets is replaced by the template for b-jets from top-quark decays.
By default, the normalization of the template for b-jets from top-quark decays is fixed to two per event. A systematic uncertainty on this normalization is assessed by using the predicted normalization from Monte Carlo simulation, which includes events with less than two b-tags from top-quark decays, due to btagging inefficiency. The total uncertainty due to specific template shape variations is referred to as 'additional fit uncertainties' for the rest of this paper.
Systematic uncertainties also affect the overall event reconstruction efficiency. Dominant sources of uncertainty for this category are: the tagging effiencies for b-, c-and LF jets, the jet energy scale and resolution, and the Monte Carlo event generator. Uncertainties on the lepton identification efficiency, E miss T reconstruction, and fragmentation modeling are negligible. In general, systematic uncertainties are evaluated on the full data sample, with each uncertainty being taken as the difference between the nominal and the varied resulting values of R HF .
An important uncertainty in this analysis comes from the flavor composition in the fiducial volume, namely in the value of F b/HF , the fraction of tt + HF events in the fiducial volume which contain b-jets, used to calculate the correction factor ε HF . As described in Sec. V, an uncertainty of 10% on F b/HF is estimated using different Monte Carlo generators. It is possible to evaluate F b/HF using the data, but with the present data set, significant discrimination between b-and cjets is not possible, making such a comparison of limited use. Nonetheless, the result of this study is presented as a point of comparison to the result obtained from the Monte Carlo.

IX. RESULTS
In the 106 events in the signal sample (with ≥ 3 btagged jets), there are 325 b-tagged jets. After subtracting the non-tt background component, and the contribution from the tagged jets from the t → W b decay, the number of additional b-tags is found to be 105. As described in Sec. VII, a template fit to all b-tagged jets is performed to determine the flavor composition of these additional b-tagged jets. The result of the fit to all 325 b-tagged jets is shown in Fig. 3. The weighted sums of all fit templates are shown, with contributions for extra   Uncertainties are quoted separately for the number of HF jets measured in the fit (N HF ), the portion of the calculation affecting only the correction factors (ε HF ), and the full calculation. As the fit prefers 100% charm for additional heavyflavor jets, it is sensitive to differences in the extra b-tagged jets from the c-quark template shape. the Monte Carlo simulation for tt + HF production, σ fid (tt + HF) is found to be 0.16 ± 0.03 (stat.) pb. ALPGEN interfaced with HERWIG predicts a value of 0.10 pb. The uncertainty on the fitted fraction of light-flavor jets is significantly smaller than the uncertainty on the fitted fraction of additional b-jets. This is understood as an effect of fitting in multiple b-purity bins: the lowpurity bin is dominated by light-flavor jets and thus gives improved discrimination. The data resolve the total observed HF production rate with a significance of about 3σ .
In the data, 1656 tt dilepton candidate events are observed with at least three jets, at least two of which are b-tagged. The total background estimate, which is dominated by LF jets misidentified as b-jets from topquark decay, is found to be 112 ± 4 (stat.), leading to a background subtracted yield of 1544 ± 41 (stat.). Using Eq. 2, and the quoted acceptance factor for tt + j production, σ fid (tt + j) is found to be 2.55 ± 0.07 (stat.) pb, compared to 2.83 pb predicted by ALPGEN and HERWIG. Taking into account the total uncertainty, it is found that R HF = [6.2 ± 1.1 (stat.) ± 1.8 (syst.)]%. A full breakdown of the systematic uncertainties contributing to R HF is given in Table IV.
The extracted value of σ fid (tt + HF) is very sensitive to the value of F b/HF . As indicated in Sec. V, the efficiency for tt + b + X events is approximately a factor of three higher than the corresponding efficiency for tt + c + X events, implying a potential change in σ fid (tt + HF) by a factor of three if F b/HF is allowed to vary Using the fitted fraction of additional b-jets in data results in F b/HF = −0.02, with one and two sigma statistical upper bounds of F b/HF = 0.09 and 0.27, respectively. This value is found to be compatible with the predicted value to within 1σ when systematic uncertainties are included. Figure 4 shows R HF as a function of F b/HF . The predicted and data-driven ranges of F b/HF are overlaid. With F b/HF = −0.02 the central value for R HF is determined as 10.7%.

X. CONCLUSIONS
A 4.7 fb −1 sample of 7 TeV proton-proton collisions recorded by the ATLAS detector at the LHC was used to measure the ratio R HF of the fiducial cross section for the production of tt events with at least one additional HF quark jet (tt + b + X or tt + c + X) to that for the production of tt events with at least one additional jet, regardless of flavor, each with p T > 25 GeV and |η| < 2.5. A fit to the vertex mass distribution for b-tagged jets in tt candidate events with three or more b-tagged jets is performed to determine the heavy-and light-flavor content of the additional b-tagged jets. The result of the fit shows that 79 ± 14 (stat.) ± 22 (syst.) of the 105 selected b-tagged jets originate from HF quarks, three standard deviations away from the hypothesis of zero tt + HF production. A value of R HF = [6.2±1.1 (stat.)±1.8 (syst.)]% is extracted. This value of R HF is consistent with the leading order predictions of 3.4% obtained from the ALPGEN Monte Carlo generator interfaced with HERWIG and 5.2% from a calcu-lation using POWHEG interfaced with HERWIG.

Acknowledgments
We thank CERN for the very successful operation of the LHC, as well as the support staff from our institutions without whom ATLAS could not be operated efficiently. We