Search for Standard Model Higgs Boson Production in Association with a W Boson using Neural Networks with 9.45fb-1 of CDF data
Timo Aaltonen1, Barbara Alvarez2, Adrian Buzatu3, Giorgio Chiarelli4, Jake Connors5, Jay Dittmann6,
Martin Frank6, John Freeman7, Craig Group8, Richard Hughes5, Nazim Hussain3, Tom Junk7,
Azeddine Kasmi6 Benjamin Kilminster7, Shinhong Kim9, Mike Kirby7, Sandra Leone4, Jan Lueck10,
Thomas Muller10, Yoshikazu Nagai9, Yuri Oksuzian8, Tom Phillips11, Elisabetta Pianori12, Manfredi Ronzani4,
Alberto Ruiz13, Federico Sforza4, Rick Snider7, Marco Trovato4, Rocio Vilar13, Jesus Vizan14,
Andreas Warburton3, Jon Wilson5, Brian Winer5, Homer Wolfe5, Zhenbin Wu6, Weiming Yao15

1University of Helsinki, 2Michigan State University, 3McGill University, 4INFN Pisa, 5The Ohio State University, 6Baylor University, 7FNAL,
8University of Virginia, 9University of Tsukuba, 10Karlsruhe Institute of Technology, 11Duke, 12University of Pennsylvania,
13IFCA (CSIC-Univ. Cantabria & Univ. Oviedo), 14UC Louvain, 15LBNL


 


- Abstract -
- Event Selection -
- NN b-jet energy correction -
- Bayesian Neural Network Inputs -
- Bayesian Neural Network Output -
- Systematics -
- Results -
- Diboson interpretation -
- Public Note -


- Link to the webpage of the previous version of the analysis -






Abstract:

We present a search for the standard model Higgs boson produced in association with a W± boson. The search uses data corresponding to an integrated luminosity of 9.45 fb-1. We select W→ lν bb candidate events with two or three jets, large missing transverse energy, and exactly one charged lepton from the central or forward regions of the detector. We further require that at least one jet be identified as originating from a b-quark. Discrimination between the Higgs boson signal and large Standard Model(SM) backgrounds is improved through the use of a Bayesian artificial neural network(BNN). The data agree with the SM background predictions within the systematic uncertainties. However, a modest(~2σ) broad excess for signal-like events is observed in the data. We set upper limits on the Higgs boson production cross section times the bb branching ratio, expressed in the units of SM Higgs production cross section:

σ(pp → WH)⋅BR(H→bb) < 3.13 (1.97)⋅SM - observed (expected) at mH = 115 GeV/c2





Event Selection:

We have analyzed 9.45 fb-1 of the full dataset recorded by CDF. The events were collected by triggers for high energy electrons, muons or large missing transverse energy. In the final event selection, we require events with high energy lepton candidates, large missing transverse energy, and two or three jets with at least one b-tag.

The main improvement to the analysis comparing to the previous result comes from the improved b-jet tagging algorithm. In the previous analysis tagging algorithm was based on combination of three b-jet taggers. In the current analysis we use the newly developed multivariate b-jet tagger,the Higgs-Optimized b-Identification Tagger (HOBIT), that has been optimized for the Higgs searches. Two HOBIT operational poins are used to define the tagging categories, corresponding to HOBIT Neural Net output values of 0.98, and 0.72. We denote them as Tight (T) and Loose (L) respectively. Based on these operational points, the following 5 orthogonal b-tagging categories are defined: TT, TL, LL, T, and L. We use TT, and TL only categories for events with 3 jets.

We classify our events into the dedicated lepton categories: central tight leptons, forward electrons, loose muons and loose electron-like leptons. Loose muons category is formed from muons that fail the standard definition or point to a detector gap regions. Loose electron-like category is formed from W → eν/τν events, where an electron fails the standard identification or the τ lepton decays in a single charged hadron (one-prong). In addition to non-triggered electrons selected requiring an isolated track with significant deposits of energy in the calorimeter (loose isolated tracks), triggered electrons that fail the standard selection using a multivariate likelihood (loose electrons) are also included in this category. Recently the acceptance has been increased by selecting events by using additional triggers.
  • Extra central electrons and muons: The acceptance of central electrons and muons has been increased by using an inclusive set of triggers to select these kind of events instead of a single specific trigger. These kind of events are included in the central tight lepton category.
  • Gap Muons: Muons that do not fire the main trigger for central muons because they are not detected in both of the central muon CDF detectors (the Central Muon Detector, and the Central Muon Upgrade Detector), are recovered by employing specific triggers that only require that the muon is detected in one of the two detectors. A novel technique using NN regression is used to parameterize the trigger efficiencies. These kind of events are included in the central tight lepton category.
  • Extra loose muons: Loose muons that failed the standard identification and are reconstructed as tight isolated tracks. These kind of events were only selected using MET and jet based triggers before. Now a more inclusive set of triggers is considered. The trigger efficiency is measured again using NN regression. These kind of events are included in the loose muon lepton category.

A cut to reject QCD events based on a multivariate technique that makes use of a Support Vector Machine is applied. A different MET cut is applied depending on the reconstructed lepton in the event. The cut is set 20 (10) GeV for electrons (muons) from the central tight category. It is set to 25 GeV for plug electrons and loose electron-like leptons and it is set to 20 GeV for loose muons. For the latest category additional jet and MET requirements are set for events that are obtained from jet and MET based triggers.

We estimate the expected background contribution in each lepton type and b-tagging category. The following table shows the expected background and observed data events for the each b-tagging category for all the leptons:



The expected and observed number of background events for the different b-tagging and lepton categories is shown in the following tables:

Lepton Type 2 Jets, TT tags 2 Jets, TL tags 2 Jets, T tag 2 Jets, LL tags 2 Jets, L tag 3 Jets, TT tags 3 Jets, TL tags
Central Tight Leptons Event Yields Event Yields Event Yields Event Yields Event Yields Event Yields Event Yields
Plug Electrons Event Yields Event Yields Event Yields Event Yields Event Yields Event Yields Event Yields
Loose Muons Event Yields Event Yields Event Yields Event Yields Event Yields Event Yields Event Yields
Loose Electron-Like Leptons Event Yields Event Yields Event Yields Event Yields Event Yields




NN b-jet energy correction

The most sensitive variable for WH analysis is the dijet invariant mass. Improvement on dijet mass resolution directly results in improvement of the final sensitivity. To further improve the dijet mass resolution, b-jet energy corrections based on a multivariate technique are applied. More details can be found here




BNN Inputs

To take the full advantage of all the kinematical differences between WH events and SM background production processes we construct the final discriminant for WH analysis using multivariate algorithm by means of Bayesian Neural Network (BNN). Three sets of BNN discriminants are used for events with 2 jets: Double Tag Tight, Double Tag Loose , and Single Tag. For events with 3 jets only 2 sets of BNN discriminants are optimized for different background compositions (Double Tag Tight, and Double Tag Loose). In this case, BNN discriminant is optimized against the tt background. A cut on this variable is applied to subsequently to train an additional BNN against the W+bb background. The final discriminant separates the events in 2 regions according to the output value of the tt BNN. For each BNN function, we use the input variables defined HERE and the validation plots for these variables shown below

BNN input variables for events before b-tagging requirement(pretag) with central lepton and 2 jets.


More NN Input Plots
Lepton Type 2 Jets, Pretag 2 Jets, TT tags 2 Jets, TL tags 2 Jets, T tag 2 Jets, LL tags 2 Jets, L tag 3 Jets, TT tags 3 Jets, TL tags
Central Triggered Leptons Plots Plots Plots Plots Plots Plots Plots Plots
Plug Triggered Electrons Plots Plots Plots Plots Plots Plots Plots Plots
Non-Triggered Muons Plots Plots Plots Plots Plots Plots Plots Plots
Loose Electron-Like Leptons Plots Plots Plots Plots Plots Plots N/A N/A





BNN discriminants: pretag events

BNN output for pretag events with central lepton and 2 jets.


Bayesian Neural Network outputs before b-tagging, for both linear scale (Left) and log scale (Right).
From top to bottom, Central leptons, Plug electrons, Non-triggered muons, and Loose electron-like leptons, respectively.





BNN discriminants: b-tagged events

The following plots show the BNN output for all the b-tagging categories (signal region) combining all the considered lepton categories. Left plots show the normalized BNN output, middle (right) plots show the BNN output with linear (log) scale.

2 Jets, "TT" b-tagging category
2 Jets, "TL" b-tagging category
2 Jets, "T" b-tagging category
2 Jets, "LL" b-tagging category
2 Jets, "L" b-tagging category
3 Jets, "TT" b-tagging category
3 Jets, "TL" b-tagging category


BNN discriminants for various lepton types

Central Triggered Leptons Plots
Plug Triggered Electrons Plots
Non-Triggered Muons Plots
Loose Electron-Like Leptons Plots





Systematics:

We address systematic uncertainty on the signal acceptance from several different sources:
  • Lepton reconstruction: The reconstruction efficiency uncertainty is ≈ 2% for Central Tight Leptons and for Plug Electrons. The uncertainty for loose muons and loose electron-like leptons is ≈ 4.5%.
  • Trigger: The trigger efficiency uncertainty is ≈ 1% for Central Tight Leptons and for Plug Electrons. The uncertainty for loose muons and loose electron-like leptons is ≈ 3%.
  • Initial and final state radiation: The uncertainties due to initial- and final-state radiation are estimated by changing the parameters related to ISR and FSR, halving and doubling the default values. This uncertainty ranges from ≈ 5 - 10 % for most of the considered categories.
  • Jet energy scale: The effect of the uncertainty in the jet energy scale is evaluated by applying jet-energy corrections that describe ± 1σ variations to the default correction factor. This uncertainty ranges from ≈ 2 - 10 % for most of the considered categories.
  • Luminosity : We consider a further 6.0% uncertainty in the estimation of the integrated luminosity for all the lepton and b-tagging categories.
  • b-tagging and mistag uncertainties: The uncertainties in the b-tagging and mistag efficiencies are considered independently for the Tight and Loose operational points of the b-tagging algorithm. The following table shows the corresponding rate uncertainties due to these effects for the different tagging categories.

  • Uncertainty b-tag Tight b-tag Loose Mistag Tight Mistag Loose
    T+T 7.8% 0 40% 0
    T+L 3.9% 3.2% 19% 10%
    T 3.9% 0 19% 0
    L+L 0 6.3% 0 20%
    L 0 3.2% 0 10%



    The uncertainty in the shape of the BNN discriminant due to the JES is also taken into account. Also shape and rate systematics are considered for the uncertainty in the renormalization scale used to generate the W + jets MC samples by halving and doubling the default value.




    Results:

    We perform a binned likelihood fit of the BNN discriminant where we constrain the backgrounds to their estimated rates within uncertainties. For optimal sensitivity, we perform a separate simultaneous search in each tag channel and lepton category.
    The plot and the table show the observed and expected 95% confidence level upper limits on the WH production cross section times the bb branching ratio. The result is obtained using all the lepton types, 2 and 3 jets, and all the b-tagging categories.

    Observed and expected upper limits for WH analysis with the full dataset.


    mH [GeV/c2] Observed Expected
    90 1.38 1.36
    95 2.07 1.53
    100 1.92 1.44
    105 2.36 1.58
    110 3.03 1.76
    115 3.13 1.97
    120 4.33 2.30
    125 4.93 2.79
    130 6.47 3.59
    135 8.51 4.85
    140 10.9 6.59
    145 14.4 9.91
    150 21.7 15.9





    The evolution of improvements for WH sensitivity:









    Diboson interpretation:

    The production rate of WZ boson pairs is significantly higher than a low-mass Higgs boson.The measurement of this process using the tools designed for the Higgs boson search could provide a powerful validation of the WH analysis. We perform the diboson analysis using exactly the same event selection and the tools as for WH search. The dijet mass distribution shown below is clearly sensitive to the diboson production. However, in order to confirm the WH framework, we construct the BNN discriminant similar to the procedure used for WH. The fit for the total WZ cross section distributions yields , which is a little bit higher than the SM prediction of 3.2 ± 0.2pb, but the result is still consistent with the NLO SM prediction at within 1.5 standard deviations.



    Invariant dijet mass for TT+TL b-tagging combined categories BNN for WZ for TT+TL b-tagging combined categories




    Public Note:

    More details about the analysis can be found in the public note.
    - Back to Top page -






    -- Contact