CMS Search for hgγγ

Yuri Gershtein

…or… squeezing blood from stone

Yuri Gershtein

Hgγγ Search

!  How come this is hard?!

4/6/12 Yuri Gershtein 3

M = 2E1E2 (1! cos!)

!  Main idea: measure energy of the two photons and their opening angle

The Challenge !  Huge “irreducible” background from QCD di-photon

production (plus instrumental backgrounds)

4/6/12 Yuri Gershtein 4

Tracker Material

4/6/12 Yuri Gershtein 5

!   Reasonably well described by simulation !   Degrades energy resolution

4/6/12 Yuri Gershtein 6

What Material Does to Photons

!   Generate back-to-back photons and scan event displays

Black: hits in tracker

Green: ECAL outline

Red: ET of ECAL cells

Several bad effects at once

•  hard to identify (showers not narrow in azimuth)

•  energy resolution has tails due to not clustered energy

•  some energy never reaches ECAL (B = 4 Tesla!)

4/6/12 Yuri Gershtein 7

10-4

10-3

10-2

10-1

0 5 10 15 20 25 30 35Reconstructed Et

Arb

itrar

y Un

its

Single electrons 30 GeV pt

Putting Humpty Dumpty Together Again

!   Some of the spray may be recovered by making clusters wide in azimuth !   at a cost of picking up energy from underlying

event

!   How to make optimal correction given the detector region and all reconstructed pieces of the photon?

Super-Cluster window

The First Squeeze ! Improve photon energy resolution !   Energy resolution is affected by

!   Impact point in the calorimeter (containment) !   Whether the photon converted !   Radius of conversion !   The amount of material that electrons from conversion have to

traverse before impacting calorimeter

!   Very non-trivial correlations between measured photon properties and the corrections that one needs to apply

4/6/12 Yuri Gershtein 8

The First Squeeze ! Improve photon energy resolution !   Energy resolution is affected by

!   Impact point in the calorimeter (containment) !   Whether the photon converted !   Radius of conversion !   The amount of material that electrons from conversion have to

traverse before impacting calorimeter

!   Very non-trivial correlations between measured photon properties and the corrections that one needs to apply

4/6/12 Yuri Gershtein 9

“trivial” in one dimension

Multivatiate Analyses ! Classification

4/6/12 Yuri Gershtein 10

! Regression

!   The math that goes into either is quite similar – multi-dimensional minimization problem

Boosted Decision Trees !   BDT’s have been used for classification problems in high

energy physics for more then a decade !   i.e. miniBOONE particle identification

!   Robust against noisy variables !   Robust against outlier events !   Robust against discrete variables !   Fairly intuitive

!   Boosting !   Instead of one tree use a weighted

ensemble of many trees

4/6/12 Yuri Gershtein 11

root node

2.1

0.35 0.63 0.17

1.5

regression tree

The First Squeeze ! Improve photon energy resolution !   Energy resolution is affected by

!   Impact point in the calorimeter (containment) !   Whether the photon converted !   Radius of conversion !   The amount of material that electrons from conversion have to

traverse before impacting calorimeter

!   Depending on the region in the detector, can achieve up to 20% improvement in resolution compared to “standard” factorized correction method !   And can check that the used variables and their correlations are

described by simulation by comparing electrons from Z decays in data and MC

4/6/12 Yuri Gershtein 12

The Second Squeeze !   Some photons are measured well, some are not – instead of

mixing all events together, can we separate them into classes and combine the searches in individual classes?

!   First 5/fb result – just 2 variables: Barrel – Endcap and converted-unconverted based on shower narrowness (fraction of cluster energy in the 3x3 crystal matrix)

4/6/12 Yuri Gershtein 13

The Second Squeeze !   Some photons are measured well, some are not – instead of

mixing all events together, can we separate them into classes and combine the searches in individual classes?

!   First 5/fb result – just 2 variables: Barrel – Endcap and converted-unconverted based on shower narrowness (fraction of cluster energy in the 3x3 crystal matrix)

!   Can we be smarter?

4/6/12 Yuri Gershtein 14

The Second Squeeze !   Some photons are measured well, some are not – instead of

mixing all events together, can we separate them into classes and combine the searches in individual classes?

!   First 5/fb result – just 2 variables: Barrel – Endcap and converted-unconverted based on shower narrowness (fraction of cluster energy in the 3x3 crystal matrix)

!   Train the second regression – this time do not try to improve the resolution by asking BDT to guess the energy, ask the BDT to guess how well this particular photon is measured !   Same variables as for energy regression !   Different target – instead of Etrue/Ereco regress to |(Ereco-Etrue)/Etrue|

4/6/12 Yuri Gershtein 15

Regression to resolution !  Regress to |Ereco-Etrue|/Etrue

!   average value of that is 0.7985 of gaussian sigma

16 4/6/12 Yuri Gershtein

Fit resolution in bins of BDT

17 4/6/12 Yuri Gershtein

simulation

Fit resolution in bins of BDTG

18

Just as with energy regression, we can check the resolution regression using observed width of Zgee

4/6/12 Yuri Gershtein

simulation

Hgγγ resolution classification

4/6/12 Yuri Gershtein 19

M = 2E1E2 (1! cos!)

!  Mass resolution depends on the energy resolution of the photons and on the precision of the opening angle measurement

!  If the vertex is reconstructed correctly (within ~1cm) the contribution of angle in mass resolution is negligible

!  If not, angular resolution dominates

Choosing the right vertex

!  Track activity in the best events – with no photon conversion – look very similar to a minimum bias event

4/6/12 Yuri Gershtein 20

The Third Squeeze !  Using BDT to pick the correct vertex !  Simulation result on the Higgs signal sample

4/6/12 Yuri Gershtein 21

no conversions with conversions

The Third Squeeze !   Total efficiency to get the right vertex ~80%

!   Validated with Zgµµ and γ+jet events in data

!   Higgs pT is a good predictor of whether the vertex is correct

4/6/12 Yuri Gershtein 22

The Fourth Squeeze !   Instead of using one variable to predict the probability that

the vertex is picked correctly, use an MVA (BDT) ! pT of the di-photon system !   number of vertices !   per-vertex BDT values for the best three vertices ! Δz between first and second and third vertices !   “pulls” of the reconstructed conversions

4/6/12 Yuri Gershtein 23

z vtx!zconv! conv

Zgµµ

!   In the hypothesis that the vertex is unknown, the resolution can be calculated analytically

as a function of photon’s η, φ, and the distance from (0,0,0) to ECAL cluster r

!   This formula simplifies if r1=r2

4/6/12 Yuri Gershtein 24

vertex impact on mass resolution

The story so far !  Minimized mass resolution

!   regression to energy !   categorization of the vertices

!  Evaluated mass resolution precision !   regression to energy resolution !   regression to vertex selection probability !   analytical formula for resolution due to incorrect vertex

4/6/12 Yuri Gershtein 25

The story so far !  Minimized mass resolution

!   regression to energy !   categorization of the vertices

!  Evaluated mass resolution precision !   regression to energy resolution !   regression to vertex selection probability !   analytical formula for resolution due to incorrect vertex

! Left to do: !  Suppress instrumental backgrounds (photon ID MVA) !   Identify differences in kinematical features of signal and

background and make a final discriminant optimizing signal-to-background ratio (per-event categorization MVA)

4/6/12 Yuri Gershtein 26

Final Classification !   Demonstration of what final classifier is sensitive to:

!   separate Higgs MC events into high/low di-photon pT

4/6/12 Yuri Gershtein 27

!   Vertical lines correspond to optimized classifier bins !   Mass distributions are fit in classifier bins

!   Colored histograms are event classes as in non-MVA analysis !   Note overlaps between the old categories!

high S/B low S/B high S/B low S/B

Higgs simulation

Higgs simulation

Final Classification !  Demonstration of what final classifier is sensitive to:

!  plot mass resolution for each classifier bin

4/6/12 Yuri Gershtein 28

Crunching the Numbers !  Five exclusive event categories

!  VBF tagged events ! pT

γ > mγγ/2.18, mγγ/4, |η|<1.44 or 1.57<|η|<2.5 !  2 jets, ET>30, 20 GeV, |η|<4.7 !  Δη> 3.5, mJJ>350 GeV, |Z|<2.5, Δφ(jj,γγ)>2.6

!  Four BDT categories ! pT

γ > mγγ/3, mγγ/4, |η|<1.44 or 1.57<|η|<2.5 !  VBF tagged events are removed

4/6/12 Yuri Gershtein 29

Crunching the Numbers

4/6/12 Yuri Gershtein 30

!  New analysis methods result in effectively almost 40% larger data sample

4/6/12 Yuri Gershtein 31

Expected sensitivity@125 GeV is ~1.2σSM (compared to ~1.6σSM for ATLAS)

The result

The result

4/6/12 Yuri Gershtein 32

2.9σ maximum local significance. After LLE in 110-150 GeV window the significance is 1.6σ (non-MVA numbers were 3.1σ and 1.8σ)

Cat 0-3 went from ~1.3×10-2 to ~3.0×10-2

Fermiophobic interpretation

A little different analysis !   VBF tag is essentially the

same (except no cut on classification MVA)

!   add lepton tag from VH !   for events not tagged by

the two analyses above do a 2-D unbinned likelihood fit (mass vs higgs pT) in “old” categories

4/6/12 Yuri Gershtein 33

Fermiophobic interpretation

4/6/12 Yuri Gershtein 34

FP interpretation: with WW/ZZ

“van

illa”

FP

is n

ot f

avor

ed

4/6/12 Yuri Gershtein 35

Challenge for 2012 !   Preliminary estimates on MC

!   As-is analysis degrades by ~20%

!   Pile-up !   Degrades ID !   Degrades energy resolution !   Increases chance of misidentifying the primary vertex !   Degrades VBF tag

!   Particle flow techniques seem to help quite a bit !   Possible to recover and may be even improve

4/6/12 Yuri Gershtein 36

Summary !   New MVA methods have been developed for Higgs

search in di-photons and give a substantial improvement in the expected statistical precision

!  The fundamental conclusion of the re-analysis of the 2011 data is however unchanged

Need more data

4/6/12 Yuri Gershtein 37

Summary !   New MVA methods have been developed for Higgs

search in di-photons and give a substantial improvement in the expected statistical precision

!  The fundamental conclusion of the re-analysis of the 2011 data is however unchanged

Need more data

(but less then we’d needed without MVA’s)

4/6/12 Yuri Gershtein 38

4/6/12 Yuri Gershtein 39

Systematic errors

4/6/12 Yuri Gershtein 40

Boosted Decision Trees !   A toy example

4/6/12 Yuri Gershtein 41

z = {1+ ay, x < 0!1+ by, x > 0

Boosted Decision Trees !   Ensemble of 1000 trees

!   Tree parameters: depth=1, number of splits/variable = 4

4/6/12 Yuri Gershtein 42

Boosted Decision Trees !   Ensemble of 1000 trees

!   Tree parameters: depth=2, number of splits/variable = 4

4/6/12 Yuri Gershtein 43

Boosted Decision Trees !   Ensemble of 1000 trees

!   Tree parameters: depth=3, number of splits/variable = 4

4/6/12 Yuri Gershtein 44

Boosted Decision Trees !   Ensemble of 1000 trees

!   Tree parameters: depth=1, number of splits/variable = 100

4/6/12 Yuri Gershtein 45

The Third Squeeze !  Use BDT to choose the “signal” vertex !   For all events

!   “intensity” of the vertex

!   tracks should follow direction of higgs recoil

!   track sum should be similar to higgs recoil

!   For events with at least one reconstructed conversion

!   “pull”

4/6/12 Yuri Gershtein 46

!pT!2

!pT! "!pT!!

!pT!!

!pT! "!pT!!

!pT! +!pT!!

z vtx!zconv! conv

vertex impact on mass resolution !   Toy MC: take ideal energy resolution, and *always* pick a

wrong vertex !   The error is on 1-cosα, so back-to-back configuration is less affected

4/6/12 Yuri Gershtein 47

vertex impact on mass resolution !   Resolution also depends on the polar angles of the photons

4/6/12 Yuri Gershtein 48

more affected less affected

Instrumental Backgrounds !   From previous (non-MVA) analysis:

4/6/12 Yuri Gershtein 49

Photon ID MVA !   Garden variety classification BDT

!   Shower shape variables !   Isolation variables !   Underlying event activity

4/6/12 Yuri Gershtein 50

!   Cluster rapidity

Barrel, Mγγ>160 GeV

Endcap, Mγγ>160 GeV

Final Squeeze: putting everything together into the classification MVA !   Kinematics

! pTγ/mγγ for both photons

! pesudo-rapidities of both photons !   cosine of opening angle in azimuthal plane

!   Instrumental background !   Photon ID BDT values for both photons

!   Mass resolution !   for correct vertex choice !   for incorrect vertex choice !   probability of the vertex to be chosen correctly

!   Events need to be weighted so that best resolution events get higher classifier values

4/6/12 Yuri Gershtein 51

Final Classification !   Although data-MC agreement is very good, strictly speaking it

is not necessary for background. It just shows that training is close to optimal

4/6/12 Yuri Gershtein 52

HIG-11-033 results

4/6/12 Yuri Gershtein 53

Combination

4/6/12 Yuri Gershtein 54

Combination

4/6/12 Yuri Gershtein 55

Combination

4/6/12 Yuri Gershtein 56