16
John Marshall HZ Recoil Analysis 1 HZ Recoil Analysis: Selection & Fitting LCD WG6 Meeting, 03/04/2012 J.S. Marshall, University of Cambridge

HZ Recoil Analysis: Selection & Fitting

  • Upload
    jase

  • View
    46

  • Download
    0

Embed Size (px)

DESCRIPTION

HZ Recoil Analysis: Selection & Fitting. LCD WG6 Meeting, 03/04/2012 J.S. Marshall, University of Cambridge. HZ Recoil Analysis - Reminder. Relevant processes for this study are the recoil reaction e + e -  HZ  Hff , commonly called Higgsstrahlung. - PowerPoint PPT Presentation

Citation preview

Page 1: HZ Recoil Analysis: Selection & Fitting

John Marshall HZ Recoil Analysis 1

HZ Recoil Analysis:Selection & Fitting

LCD WG6 Meeting, 03/04/2012

J.S. Marshall, University of Cambridge

Page 2: HZ Recoil Analysis: Selection & Fitting

John Marshall HZ Recoil Analysis 2

HZ Recoil Analysis - Reminder

sEMsM ZZrecoil 222

s = 500GeVLint = 500fb-1

mH =120GeVNo polarization

• Relevant processes for this study are the recoil reaction e+e-HZHff, commonly called Higgsstrahlung.

• By detecting decay products of the Z, can search for Higgs signals without further assumption about Higgs decay modes: “model independent analysis”.

• For CDR V3, will search for decays Z and Zee: “X” and “eeX” channels.

• Signal is selected by identifying two well-measured leptons in final state, yielding Z mass. Can then compute recoil mass:

Page 3: HZ Recoil Analysis: Selection & Fitting

John Marshall HZ Recoil Analysis 3

Requests and Production Status

Process description

Whizard process ID

Cross-section

Cross-section

after pre-cuts

Events requested

X (signal) hzmumu 2.18 fb 2.18 fb 54,500

( e2e2nn 160.05 fb 160.05 fb 8.0 104

ff  e2e2ff 4592.55 fb 776.14 fb 3.9 105

eeX (signal) hzee 2.18 fb 2.18 fb 54,500

( ee eenunu 419.78 fb 419.78 fb 2.1 105

eeff  e1e1ff 6047.78 fb 1965.53 fb 9.8 105

Generator-level cuts for background samples ff and eeffpT l+l- > 10GeV, cut on transverse momentum, calculated from vector sum of two

leptons

|cos l+/l-| < 0.95, cut on angle of either of leptons

Status on 27/03/2012:

Events available

Percentage of total

59,381 100%81,337 100% )

295,000 76%

60,379 100%221,058 100% ) 993,000 100%

+ Two-fermion bkg test samples:

ee: 5,489 events : 9,990 events : 9,990 events

Page 4: HZ Recoil Analysis: Selection & Fitting

John Marshall HZ Recoil Analysis 4

Discriminant Variables• Previously, discussed procedure for finding

events with two leptons produced by Z-decay.

• Events passing di-lepton selection examined with aim of background rejection in mind.

• Know that selection will be rather difficult at s=500GeV and with radiative effects.

sEMsM ZZrecoil 222

Page 5: HZ Recoil Analysis: Selection & Fitting

John Marshall HZ Recoil Analysis 5

Discriminant Variables

TTdlTbal PPP

)/.( cosacop -12121 TTTT PPPP)/.( cosacol -1

2121 PPPP

Page 6: HZ Recoil Analysis: Selection & Fitting

John Marshall HZ Recoil Analysis 6

Selection - TMVA

X signal X bkg eeX signal eeX bkgEfficiency 87.6% 3.6% 79.3% 3.5%

• Use ROOT TMVA to perform selection. Allows simple cut optimization and comparison of multivariate techniques.

• Begin by making a few simple selection cuts to quickly veto background whilst leaving signal largely unharmed.

• Use TMVA to suggest a series of cuts, choosing to use those that approximate to 90% signal efficiency for X.

Pass Di-Lepton ID30 GeV < Mdl < 120 GeVMrecoil < 380 GeVPTdl > 65 GeVacopdl < 2.8PTBal > -100 GeV

Cuts for 90% signal efficiency

Page 7: HZ Recoil Analysis: Selection & Fitting

John Marshall HZ Recoil Analysis 7

Selection - BDT• Events passing selection cuts given to TMVA for

evaluation of different multivariate techniques.

• TMVA uses 50% of input events for training and 50% for testing, performing overtraining checks.

• Input variables: Mdl, PTdl, cosdl, acoldl, acopdl, PTBal

• Best performance (>20% purity): Boosted Decision Tree

Page 8: HZ Recoil Analysis: Selection & Fitting

John Marshall HZ Recoil Analysis 8

Selection - BDT• BDT also appropriate for signal selection in more difficult eeX channel (but could move to MLP).

• Can achieve small improvement with more aggressive initial cuts. However, this can lead to smaller samples for BDT training and hence overtraining. TMVA overtraining report OK:

• Application of full selection procedure removes all events from 2-fermion background test samples.

Page 9: HZ Recoil Analysis: Selection & Fitting

John Marshall HZ Recoil Analysis 9

Di-lepton Mass• Choice of BDT cut value currently a little arbitrary: simply try to maintain 30-40% signal efficiency.

X

BDT Cut > 0.17

Efficiency 41.7%

Purity 35.3%

Signal @ 500fb-1 454.9

Bkg @ 500fb-1 834.0

eeXBDT Cut > 0.15

Efficiency 33.1%

Purity 24.0%

Signal @ 500fb-1 360.7

Bkg @ 500fb-1 1139.1

Page 10: HZ Recoil Analysis: Selection & Fitting

John Marshall HZ Recoil Analysis 10

Recoil Mass

• Will try to fit these distributions to extract mH, nSig, nBkg

• Knew selection would be difficult. For comparison, scale results to ILD LoI cross-sections and luminosity.

• For ILD LoI analysis, reduced radiative effects also result in better Higgs mass peak.

Scale to ILD LoI , L

Page 11: HZ Recoil Analysis: Selection & Fitting

John Marshall HZ Recoil Analysis 11

Kernel Estimation• Need to model signal component of recoil mass distribution.

• Take half of selected signal and fit this “reference sample” using Simplified Kernel Estimation.

• Signal shape is approximated by sum of many Gaussians, one for each bin in the reference sample.

• Transformation x’ = x – mH allows sensitivity to Higgs mass; scaling distribution allows sensitivity to no. of signal events.

Page 12: HZ Recoil Analysis: Selection & Fitting

John Marshall HZ Recoil Analysis 12

Background Fit• Need to model background distribution for any given number of selected background events.

• Could just use available MC samples, but ideally want to remove statistical fluctuations.

• Choose to fit shape of selected background – 4th order polynomial seems to be OK.

• In final fit to recoil mass distribution, won’t vary background position or shape, just normalization.

Page 13: HZ Recoil Analysis: Selection & Fitting

John Marshall HZ Recoil Analysis 13

mH = 120 GeV nSig = 200 nBkg = 1000

mH = 50 GeV nSig = 500 nBkg = 1000

Predicted Distribution

mH = 200 GeV nSig = 500 nBkg = 1000

mH = 120 GeV nSig = 1000 nBkg = 200

X

• Using predicted signal and background, can create a prediction for any parameters mH, nSig, nBkg:

Page 14: HZ Recoil Analysis: Selection & Fitting

John Marshall HZ Recoil Analysis 14

Fit Distribution

bins

lnnll )(n-nn predobspred

• To produce input/fit distributions, use remaining half of signal sample, scaled back to 500fb-1. Add background by scaling polynomial fit and introducing Poisson fluctuations.

• Creates input distributions shown below. Error-bars reflect fact that both signal and background components derived from high statistics samples (L=25ab-1).

• Fit procedure uses MINUIT to vary mH, nSig and nBkg, producing a negative log likelihood value for each comparison of input and predicted distributions:

Page 15: HZ Recoil Analysis: Selection & Fitting

John Marshall HZ Recoil Analysis 15

Results• X sample best fit results:

• Trust MINUIT to provide error measurements for each parameter. In the case of mH, 1 shift from best-fit increases negative log likelihood by 0.5

• Will try to further optimize selection (in particular, will try varying BDT cut), but don’t expect to be able to extract parameters with much greater precision.

InputmH 120.0 GeV

nSig 454.9

nBkg 834.0

Fit ResultsmH 119.7 1.3

GeV

nSig 457.7 59.9

nBkg 831.3 62.9

Page 16: HZ Recoil Analysis: Selection & Fitting

John Marshall HZ Recoil Analysis 16

Results

• Approach for eeX sample not fully optimized, as have tried to maintain consistency with treatment of X sample. Could introduce differences in selection (currently only BDT training differs).

• eeX sample best fit results:

InputmH 120.0 GeV

nSig 360.7

nBkg 1139.1

Fit ResultsmH 120.5 2.7

GeV

nSig 356.1 78.8

nBkg 1143.7 83.7