Presentation is loading. Please wait.

Presentation is loading. Please wait.

H  in CMS Chris Seez Friday, 16 th March 2012 at LAL Orsay.

Similar presentations


Presentation on theme: "H  in CMS Chris Seez Friday, 16 th March 2012 at LAL Orsay."— Presentation transcript:

1 H  in CMS Chris Seez Friday, 16 th March 2012 at LAL Orsay

2 2 Muon endcaps (CSC+RPC) Muon barrel (DT+RPC) HCAL ECAL 3.8T Solenoid Preshower (Pb+Si) Silicon pixel + strip tracker 16 March 2012 Chris Seez

3 3 Search for SM Higgs boson decaying to two photons CMS analysis of 4.8fb -1 of data taken in 2011 Baseline result – four inclusive event classes based on conversion flag and barrel/endcap separation, plus dijet-tagged class (for VBF)  Accepted by PLB: arXiv:1202.1487arXiv:1202.1487 Multivariate analysis – presented at Moriond Not discussed: interpretation in fermiophobic model, with use of lepton-tag (for VH) and 2D fit of inclusive data 16 March 2012 Chris Seez

4 4 Key elements in search strategy:  Narrow mass peak  Separate events into classes based on resolution and inclusive S/B  Signal model from MC; background model from data Cut-based analysis exploits the fact that (in CMS) energy resolution and inclusive S/B vary in a similar way with conversion flag and pseudorapidity  Simple cuts used to classify events Multivariate analysis makes use of the available information in a more refined way to classify events 16 March 2012 Chris Seez Conversion flag: R 9 = E 3x3 /E supercluster

5 5 Primary vertex selection  Vertex BDT: Uses tracks in the event: underlying event and recoil tracks Tracks from converted photon  Vertex probability BDT: per-event probability of “better than 10mm” Photon selection  p T (  ) > m  /3, m  /4  Loose preselection followed by photon ID BDT Photon preselection efficiency  Data/MC scale factors from Z  ee (also trigger efficiency)  Electron veto data/MC scale factor from Z  Photon energy cluster corrections  From energy regression BDT; also BDT for per-photon resolution estimate 16 March 2012 Chris Seez

6 6 Energy scale, and resolution corrections to MC: from Z  ee Event classification:  Diphoton BDT trained on resolution and kinematics (“mass factorized”)  4 events classes based on diphoton BDT output – exploiting resolution and S/B differences – plus dijet-tagged (VBF) class Signal modelling: Monte Carlo (corrected by scale factors) Background modelling: fit to data (plus new method using mass sideband background model – powerful cross check) Statistical interpretation: limits and significance using maximum likelihood fit to m  in 5 classes  Plus limits and significance using complementary sideband background model (uses mass-window BDT combining mass and diphoton BDT output) 16 March 2012 Chris Seez

7 7 16 March 2012 Chris Seez

8

9 9 Loose preselection, but designed to be tighter than trigger  Also tighter than generator level enrichment filters used for MC simulation of background (used in BDT training) Isolation and shower shape variables designed to closely follow trigger definitions  Trigger uses R 9 and barrel/endcap classification (e.g. R 9 > 0.90 photons pass trigger with no isolation requirements) Includes “conversion-safe” electron veto 16 March 2012 Chris Seez

10 10 Efficiency of supercluster to pass preselection is measured with Z  ee tag and probe Data/MC scale factors are close to 1 Systematic uncertainties on scale factors examined: reweighting probe E T and R 9 distributions to match photons from Higgs – results in variations of < 0.8% (barrel), < 1.8% (endcap) 16 March 2012 Chris Seez

11 11 BDT selects hard interaction vertex rather than pileup vertex. Inputs:   p T 2 + two variables measuring tracks to diphoton-recoil balance   z conv /  conv (if there is a conversion track) 16 March 2012 Chris Seez Z  + jet H  simulation Inclusive efficiency ~80%

12 12 Another BDT trained to identify events where the incorrect vertex has been chosen  Probability of “correct to within 10mm” is linear function of BDT output 16 March 2012 Chris Seez p T (  ) N vtx N(conversion tracks): chosen vertex Vertex-choice-BDT value: for best 3 Distance between 2 nd and 3 rd Z   + jet

13 13 Photon ID BDT output intended as a prompt/fake discriminator Input variables ~same as used in cut-based analysis  Exception: remove p T scaling from isolation variables Shower shape:  R 9, H/E, cluster  width, cluster  width Isolation (pileup corrected):  Iso sum (selected vertex), Iso sum (“worst vertex”), track isolation (selected vertex), ECAL isolation, HCAL isolation Number of primary vertices in the event (to allow modulation of isolation tightness) 16 March 2012 Chris Seez

14 14 Shower shape rescaling in MC  To account for slightly too wide showers in GEANT version used Performance: (comparison with previous cut-based ID)  Choosing a cut on the photon ID BDT output value have the same efficiency as the previous cut-based ID, gives a 13% (30%) reduction of fakes in barrel (endcap) Main benefit is to provide a continuous variable to pass to the diphoton BDT Only a very loose cut on photon ID BDT is imposed on top of the preselection (almost 100% efficient) 16 March 2012 Chris Seez

15 15 16 March 2012 Chris Seez Plots show data/MC comparison for lead photons in events with m  > 160 GeV (relatively pure) Also validated with Z  ee  (reconstructed as “Z  ”) Systematic uncertainty on the output value taken as ±0.025 fully correlated shift, which is propagated through the analysis (red bands in plots) barrel endcap

16 16 ECAL cluster corrections based on multivariate regression Input variables: supercluster and subcluster positions, shower shapes, local coordinates, N vtx Another BDT, with same input variables, provides per-photon energy resolution estimate 16 March 2012 Chris Seez

17 17 Stable energy scale achieved throughout 2011 after applying transparency corrections Average signal loss of 2.5% in the ECAL barrel is corrected with an RMS stability of 0.14%  Loss in the endcaps is larger 16 March 2012 Chris Seez

18 18 Energy resolution and energy scale determined using Z  ee  Separate by R 9 and by ECAL region 16 March 2012 Chris Seez Scale uncertainty Very little additional smearing of MC needed in central barrel

19

20 20 Aim is to encode all relevant information on signal/background discrimination (apart from m  itself) in a single variable Diphoton BDT is trained on signal and background MC samples Input variables largely mass independent The output of this BDT is then used to classify events  Four classes + additional dijet-tagged class Alternative method combines the diphoton BDT output with  m/m together with “mass sideband background model” 16 March 2012 Chris Seez

21 21 16 March 2012 Chris Seez 1p T (1)/m  Leading photon p T scaled by diphoton mass 2p T (2)/m  Trailing photon p T scaled by diphoton mass 3 11 Leading photon  4 22 Trailing photon  5 cos(   )Azimuthal angle between photons – highly correlated with p T (  ) 6  m /m  (right vtx) Mass resolution – from per-photon energy resolution 7  m /m  (wrong vtx)Mass resolution – wrong vertex: with  z contribution 8p vtx Probability of correct vertex 9ID 1 Photon ID BDT output – lead photon 10ID 2 Photon ID BDT output – trailing photon

22 22 Since inclusive S/B does not change with resolution, the BDT is insensitive to the benefit of good resolution This is changed in training by weighting signal events by 1/(resolution)  Right and wrong primary vertex hypotheses resolutions weighted by per- event probability 16 March 2012 Chris Seez

23 23 Data – MC agreement rather good, but correctness of result does not depend on correct prediction of BDT output for background: any discrepancies simply make result less optimal 16 March 2012 Chris Seez

24 24 Z  ee events selected with full diphoton selection (but inverted electron veto) Reweight p T ee spectrum in MC to match data 16 March 2012 Chris Seez AllBarrel-barrelBarrel-endcapEndcap-endcap

25 25 Events classified according to diphoton BDT output Class boundaries optimized by iteratively scanning boundary values – to obtain best median expected limit using MC background Little improvement adding more than five BDT classes Then dropping the fifth class causes little loss of perfromance  This is like a cut on diphoton BDT output value  Events used in analysis: preselection (including loose photon ID BDT cut), followed by cut on diphoton BDT output value 16 March 2012 Chris Seez

26 26 After preselection, including the loose cut on photon ID BDT, and cut on value of diphoton BDT  Prompt-prompt ~70% 16 March 2012 Chris Seez

27 27 Highest score class is almost exclusively p T (  ) > 40 GeV Quite strong correlation with event classes used in cut-based analysis 16 March 2012 Chris Seez Signal (m H = 120 GeV) MC p T (  ) > 40 GeVp T (  ) < 40 GeV

28 28 Similar plot for events in data passing the cut-based analysis selection cuts 16 March 2012 Chris Seez

29 29 Vector boson fusion production process gives two forward jet – in opposite pseudorapidity hemispheres Identifying this distinct pattern disfavours background and selects a subset of events with high S/B Events passing dijet-tag are removed from the other classes Selection (after cut on diphoton BDT):  Harder cut on lead photon  Jets separated in   Small separation in  of diphoton and dijet systems  Large dijet mass  Diphoton and dijet balance in  16 March 2012 Chris Seez p T /m  (lead photon)> 55/120 GeV p T (jet 1 )> 30 GeV p T (jet 2 )> 20 GeV  (jet 1 -jet 2 )| > 3.5  (  ) – (  (j 1 ) +  (j 2 ))/2| < 2.5 m jj > 350 GeV  (jj,  ) > 2.6 (150º)

30 30 Best event classes have excellent resolution Class 1 Class 0 16 March 2012 Chris Seez

31 31 16 March 2012 Chris Seez Class 2Class 3

32 32 16 March 2012 Chris Seez Overall resolution better than in cut-based analysis  Because events with poor resolution have been discarded All classes dijet-tagged class

33 33 Considerable variation of inclusive S/B Best event classes have excellent resolution Overall resolution significantly better than in cut-based analysis  Because events with poor resolution have been discarded 16 March 2012 Chris Seez

34

35 35 Major concern has been possible bias in the background fit Fit in each event class tested individually to find background model that has bias less than 1/5 of the statistical uncertainty on the estimated background  Tested with pseudo-data toys thrown from wide range of functions Verify with combined toys that the overall bias is at a similar level 16 March 2012 Chris Seez

36 36 Best event classes have good resolution and high S/B 16 March 2012 Chris Seez Class 1 Class 0 Analysis range

37 37 16 March 2012 Chris Seez Class 2Class 3

38 38 16 March 2012 Chris Seez All classesdijet-tagged class

39 39 16 March 2012 Chris Seez

40 40 16 March 2012 Chris Seez Shape systematics on photon ID BDT, and  E estimation are propagated to diphoton BDT, where they result in “event class migration” Shown here as applied to background MC which is not used in the analysis, and on which there are large k-factor uncertainties (not shown)

41 41 20% improvement in expected limit compared to cut-based reference analysis Observed 95% CL exclusion:  110.0–111.0, 117.5–120.5, 128.5–132.0,139.0–140.0 and 146.0–147.0 GeV 16 March 2012 Chris Seez

42 42 Most significant excess, at 125 GeV, has 2.9  local significance Global significance 1.6  (trials factor of about 28) 16 March 2012 Chris Seez

43 43 16 March 2012 Chris Seez

44 44 16 March 2012 Chris Seez

45 45 Best fit signal strength 1.6±0.7 x SM cross section 16 March 2012 Chris Seez For mass hypothesis of 125 GeV

46

47 47 The information needed to extract the signal is contained in two variables (plus the dijet tag):  diphoton BDT output value and m  Cross check employs alternative to fitting the mass spectrum in event classes “Sideband background model”:  Less sensitive to shape of m  distribution  Has sliding window exclusion of signal region  Allows, in a simple and natural way, the inclusion of a systematic uncertainty on the background normalization 16 March 2012 Chris Seez

48 48 Use for signal extraction a ±2% sliding window centred on the mass hypothesis, m H Train BDT with two inputs: diphoton BDT,  m/m H  Use only events in the four event classes 16 March 2012 Chris Seez m H = 120 GeV

49 49 Mass-window BDT output is then optimally binned (6 or 7 bins – depending on mass) Dijet-tagged class added as an additional bin   7 or 8 bin histogram 16 March 2012 Chris Seez

50 50 9 x ±2% bands constructed about mass hypothesis 6 used to build background model  Used to predict the fraction of (background) events in each mass-window BDT bin 16 March 2012 Chris Seez

51 51 Inputs to mass-window BDT are almost, but not entirely, mass independent  Main mass dependence comes from effect of the varying fake composition on the diphoton BDT Final background fraction per BDT bin is obtained from linear fit to sidebands  Systematic uncertainty from choice of linear fit included in final analysis 16 March 2012 Chris Seez

52 52 Sideband procedure gives shape of BDT output for background Overall normalization obtained from fit excluding signal region Fit uses double power law: f(m) = N(m -a + bm -c ) Possible bias assessed and 20% systematic uncertainty added to background normalization 16 March 2012 Chris Seez

53 53 The statistical and systematic uncertainties are combined Final limits and p-values then extracted from information represented in this plot (this one for m H = 124 GeV) 16 March 2012 Chris Seez

54 54 Very close agreement with the mass-fit result Discrete signal window results in a more spikey appearance 16 March 2012 Chris Seez

55 55 p-values also agree closely with mass-fit result 16 March 2012 Chris Seez

56 56 16 March 2012 Chris Seez Mass-fitSideband background model Cut-based arXiv:1202.1487

57 57 Multivariate techniques used to improve sensitivity by ~20% in terms of cross section The observed exclusion region, at 95% CL, is extended  110.0–111.0, 117.5–120.5, 128.5–132.0,139.0–140.0 and 146.0–147.0 GeV Excesses at 125 and 137 GeV remain  The small shifts in position of the excesses have no significance An alternative background model, with different assumptions about the mass spectrum, has comparable sensitivity and gives compatible results 16 March 2012 Chris Seez


Download ppt "H  in CMS Chris Seez Friday, 16 th March 2012 at LAL Orsay."

Similar presentations


Ads by Google