Вы находитесь на странице: 1из 65

Analysis of double

Higgs to bbZZ
Rami Kamalieddin, Ilya Kravchenko (UNL)
Michele de Gruttola (CERN)
Lesya Shchutska (ETH)

Big thanks for random help to


Benjamin Stieger, bb and bbWW teams

UNLHEP meeting, 15 May 2017


2

welcome to HH physics....

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


3

OUTLINE

HH production and analysis signature

Goal of the analysis

Complicated issue that we face

Possible solutions tried

Future plans

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


4

HH production,
HH bbZZ 2b 2l 2nu

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


5

Double Higgs production

Gluon fusion is the leading mechanism


* Pic from the talk by Marco Zaro
Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017
6

Higgs decay modes

Higgs BR + Total Uncert

LHC HIGGS XS WG 2013


1 WW
bb

gg ZZ
10-1

cc

10-2
Z

10-3

10-480 100 120 140 160 180 200


MH [GeV]

In our analysis we have HH decay through:


H to bb to give the highest BR
H to ZZ*, with Z to ll to give important handles,
which improve sensitivity
Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017
7

2b 2l 2 signature

MET from the off shell Z* boson

H ZZ*

2 leptons (eles, muons) from on shell Z

H bb
2 b-jets from the Higgs bb

* Pic based on the plot from bbWW team


Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017
8

Analysis strategy

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


9

Analysis strategy

Object Build Input to Combine Limits


selection Candidate

Requirements Simple common Simple shape: Final target


on leptons, jets sense cuts: HH trans.mass, is to derive
and met before Nb-jets, mbb, ZH mt2, etc the best
constructing a Nlep, mll, MET limits we
candidate OR can
OR

BDT with the BDT with all the


above variables possible
kinematic and
other variables

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


10

Analysis strategy

Object Build Input to Combine Limits


selection Candidate

Requirements Simple common Simple shape: Final target


on leptons, jets sense cuts: HH trans.mass, is to derive
and met before Nb-jets, mbb, ZH mt2, etc the best
constructing a Nlep, mll, MET limits we
candidate OR can
OR
BDT with all the
BDT with the
possible
above variables
kinematic and
other variables
Last two time we discussed object selection and building a
candidate, now lets focus on the Input to Combine
Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017
11

Goal of the analysis

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


12

Shape analysis

Our objective is:


Set 95% CL limit on
(pp X hh) BR(hh bbZZ bbllmet ) vs mass of X

The method we will use to achieve it:


Binned shape analysis using Higgs Combination Tool

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


13

Shape analysis, cont.


Questions that we need to address:

Which distribution to use?


Plenty of options.
Ideally use the one that gives the best limit.

How difficult is to construct this shape?


What could be possible complications at the
pre-approval/approval stage

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


14

Shape analysis, cont.


Questions that we need to address:

Which distribution to use?


Plenty of options.
Ideally use the one that gives the best limit.

How difficult is to construct this shape?


What could be possible complications at the
pre-approval/approval stage

Keeping in mind the latter,


lets explore in full depth the first question.
Any suggestions are highly appreciated.
Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017
15

HH transverse mass shape

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


16

Simplest shape
Possibilities:

1. HH_mt is the simplest


common sense option:

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


17

Simplest shape
Possibilities:

1. HH_mt is the simplest


common sense option:

+ Easy to construct this variable

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


18

Simplest shape
Possibilities:

1. HH_mt is the simplest


common sense option:

+ Easy to construct this variable

We cut on five different variables before


we build HH_mt. Thus, those five cuts
have to be optimized.

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


19

Simplest shape
Possibilities:

1. HH_mt is the simplest


common sense option:

+ Easy to construct this variable

We cut on five different variables before


we build HH_mt. Thus, those five cuts
have to be optimized.

Run Punzi significance or alike Use MVA methods (LD, BDT, DNN),
on 55 cuts (vary each cut at then simply cut on the response of
least twice down and up) the produced MVA distribution

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


20

BDT shape

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


21

BDT shape
Possibilities:

2. BDT response as an input shape


to Higgs Combination Tool

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


22

BDT shape
Possibilities:

2. BDT response as an input shape


to Higgs Combination Tool

+ Utilize discriminating power of


several promising variables

+ BDT does the work

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


23

BDT shape
Possibilities:

2. BDT response as an input shape to Higgs Combination Tool

Which variables to use?


We have 58 variables: pt, eta, phi, mt, mt2, dEta, dPhi, dR for b-jets,
leptons, Z, met, Hs and HH.
Discriminating power varies with the mva method.

How to address correlations?


Assume BDT will take care of it
Apply decorrelation
transformation prior to training

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


24

BDT shape
Possibilities:

2. BDT response as an input shape to Higgs Combination Tool

Which variables to use?


We have 58 variables: pt, eta, phi, mt, mt2, dEta, dPhi, dR for b-jets,
leptons, Z, met, Hs and HH.
Discriminating power varies with the mva method.

How to address correlations?


Assume BDT will take care of it
Apply decorrelation transformation prior to training

What to do with variables used to define control regions (CR):


Use for CRs (invert cut on variable) and exclude from BDT set of variables
Use in BDT and define new CRs

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


25

The scheme we chose

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


26

BDT shape for Combine

Optimize BDT with all 58 variables


Use BDT shape for Higgs Combination tool

+ Sounds like the simplest option

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


27

BDT shape for Combine


Optimize BDT with all 58 variables
Use BDT shape for Higgs Combination tool

+ Sounds like the simplest option

Type of training?
Binary classification (S vs total BG)
MultiClass classification (S vs TT vs DY)

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


28

BDT shape for Combine


Optimize BDT with all 58 variables
Use BDT shape for Higgs Combination tool

+ Sounds like the simplest option

Type of training?
Binary classification (S vs total BG)
MultiClass classification (S vs TT vs DY)

Low statistics for signal


Combine samples into four regions:
low mass (260 and 270 GeV)
at mass (@ the peak of DY and TT,
300-450 GeV)
mid mass (600 and 650 GeV)
high mass (900 and 1000 GeV)

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


29

BDT shape for Combine


Optimize BDT with all 58 variables
Use BDT shape for Higgs Combination tool

+ Sounds like the simplest option

Type of training?
Binary classification (S vs total BG)
MultiClass classification (S vs TT vs DY)

Low statistics for signal


Combine samples into four regions:
low mass (260 and 270 GeV)
at mass (@ the peak of DY and TT,
300-450 GeV)
mid mass (600 and 650 GeV)
high mass (900 and 1000 GeV)

Check what gives the best ROC integral


Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017
30

Binary classification:
S vs total BG

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


31

Binary classification
Binary classification: S vs total BG (TT+DY) of 58 vars
+ Fast
Information about correlations among BG samples is lost
260+270 GeV

ROC integral for LD = 0.547!


ROC integral for BDT = 0.918!

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


32

Binary classification
Binary classification: S vs total BG (TT+DY) of 58 vars
+ Fast
Information about correlations among BG samples is lost
260+270 GeV 300-450 GeV

ROC integral for LD = 0.547! ROC integral for LD =0.787!


ROC integral for BDT = 0.918!
ROC integral for BDT = 0.92!

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


33

Binary classification
Binary classification: S vs total BG (TT+DY) of 58 vars
+ Fast
Information about correlations among BG samples is lost
260+270 GeV 300-450 GeV 600+650 GeV

ROC integral for LD = 0.547! ROC integral for LD =0.787!


ROC integral for BDT = 0.918!
ROC integral for BDT = 0.92! ROC integral for LD = 0.981!
ROC integral for BDT = 0.992!

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


34

Binary classification
Binary classification: S vs total BG (TT+DY) of 58 vars
+ Fast
Information about correlations among BG samples is lost
260+270 GeV 300-450 GeV 600+650 GeV 900+1000 GeV

ROC integral for LD = 0.547! ROC integral for LD =0.787!


ROC integral for BDT = 0.918! ROC integral for LD = 0.981!
ROC integral for BDT = 0.92! ROC integral for BDT = 0.992! ROC integral for LD = 0.995!
ROC integral for BDT = 0.997!

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


35

Binary classification
Binary classification: S vs total BG (TT+DY) of 58 vars
+ Fast
Information about correlations among BG samples is lost
260+270 GeV 300-450 GeV 600+650 GeV 900+1000 GeV

Only btag of leading jet is present in all sets


low mass region has fewer variables in common
wrt to other sets
Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017
36

MultiClass:
S vs TT vs DY

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


37

MultiClass
MultiClass: S vs TT vs DY for all 58 vars

+ All the information available in the MC is preserved


+ May give increase in performance when enough statistics is available
Request/produce new big signal samples?

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


38

MultiClass
MultiClass: S vs TT vs DY for all 58 vars

+ All the information available in the MC is preserved


+ May give increase in performance when enough statistics is available
Request/produce new big signal samples?

Slow, easily scales from hours to days with more statistics, additional
different types of BG, increased # of BDT trees and # of layers
May be solved running on GPU with tf/theano backend

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


39

MultiClass
MultiClass: S vs TT vs DY for all 58 vars
260+270 GeV

BDT ROC integrals for:


TT DY Signal
0.340 0.056 0.997

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


40

MultiClass
MultiClass: S vs TT vs DY for all 58 vars
260+270 GeV 300-450 GeV

BDT ROC integrals for: BDT ROC integrals for:


TT DY Signal TT DY Signal
0.340 0.056 0.997 0.172 0.002 0.999

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


41

MultiClass
MultiClass: S vs TT vs DY for all 58 vars
260+270 GeV 300-450 GeV 600+650 GeV

BDT ROC integrals for: BDT ROC integrals for: BDT ROC integrals for:
TT DY Signal TT DY Signal TT DY Signal
0.340 0.056 0.997 0.172 0.002 0.999 0.328 0.289 0.999

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


42

MultiClass
MultiClass: S vs TT vs DY for all 58 vars
260+270 GeV 300-450 GeV 600+650 GeV 900+1000 GeV

BDT ROC integrals for: BDT ROC integrals for: BDT ROC integrals for: BDT ROC integrals for:
TT DY Signal TT DY Signal TT DY Signal TT DY Signal
0.340 0.056 0.997 0.172 0.002 0.999 0.328 0.289 0.999 0.190 0.783 1.000

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


43

MultiClass
MultiClass: S vs TT vs DY for all 58 vars
260+270 GeV 300-450 GeV 600+650 GeV 900+1000 GeV

None! of the variables appear in all sets

high mass region has fewer variables in common


wrt to other sets

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


44

Conclusions

Analysis strategy is clear binned shape analysis using Higgs Combine

Choose BDT over CnC for performance reasons/practicality

Multiclass seems worth using as it brings extra power

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


45

Plans
Out of almost fully exhaustive list of available options,
which were presented, make an educated guess and
pick few to get final limits

Optimize the option, which gave the best limit

Proceed with other items on the analysis list

* pic from the talk by Maxime Gouzevitch


Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017
46

Double Higgs is not easy,


but it is doable!
If it sounds complicated, you are not the only one

...the screenshot that I


took from the real HH talk!

Even Joe Incandela looks


double confused...

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


47

BACK UP

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


48

Material for HH meeting

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


49

TWO different BDTs

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


50

TWO different BDTs


Possibilities:

3. Two different (by purpose) BDTs:


One BDT with common sense variables to construct
the best promising variables
Another BDT is to utilize these best promising variables in order to
create the shape for Higgs Combination Tool

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


51

TWO different BDTs


Possibilities:

3. Two different (by purpose) BDTs:


One BDT with common sense variables to construct
the best promising variables
Another BDT is to utilize these best promising variables in order to
create the shape for Higgs Combination Tool

+ May be the best option performance-wise

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


52

TWO different BDTs


Possibilities:

3. Two different (by purpose) BDTs:


One BDT with common sense variables to construct
the best promising variables
Another BDT is to utilize these best promising variables in order to
create the shape for Higgs Combination Tool

+ May be the best option performance-wise

Difficulty validating them.

Which goes where:


common sense variables vs variables to define CRs.
variables to define CRs vs promising variables variables

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


53

N-dimensional BDT

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


54

N-dimensional BDT
Possibilities:

4. N-dimensional BDT shapes mapped into 1D hist for Higgs Combination Tool

+ Similar procedure is used in tHq/ttH (2D into 1D) and other analyses

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


55

N-dimensional BDT
Possibilities:

4. N-dimensional BDT shapes mapped into 1D hist for Higgs Combination Tool

+ Similar procedure is used in tHq/ttH (2D into 1D) and other analyses

1D shape has to be optimized too

Complexity of problem how many BDTs we need taking into account:


Two dominant backgrounds (TT and DY)
Signal like BG (SM ZH)
Signal MC samples cover 10 points from 260 to 1000 GeV

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


56

N-dimensional BDT
4. N-dimensional BDT shapes mapped into 1D hist for Higgs Combination Tool

+ Similar procedure is used in tHq/ttH (2D into 1D)

1D shape has to be optimized too

Complexity of problem how many BDTs we need?


Two dominant backgrounds (TT and DY)
Signal like BG (SM ZH)
Signal MC samples cover 10 points from 260 to 1000 GeV

Optimize 10 signal Optimize:


mass points low mass (260 and 270 GeV)
at mass (@ the peak of DY and TT,
300-450 GeV)
Optimize low One parametrized
mid mass (600 and 650 GeV)
and high mass learning training
high mass (900 and 1000 GeV)
regions 4D BDT shape can be constructed,
(see it as 3D shape wrapped into heat map)
then translated into 1D histogram
Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017
57

Physics objects

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


58

Building HH candidate

At least 2 b-jets

90 GeV < Hbb mass < 150 GeV

Exactly 2 leptons

76 GeV < Dilepton mass < 106 GeV,


to be separated from bbWW

Missing ET > 20 GeV

Transverse mass of HH > 250 GeV

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


59

b-jets
Jet selection:
PF CHS jets with charged particles
from non-primary vertices (PU)
removed before clustering
Anti-kT jet algorithm to cluster jets
withing the cone of R=0.4
Good jets with pt > 30 GeV, || < 2.4
Loose jet ID and PU-ID are applied

b-jet selection:
CSV algorithm, WP Medium
at least 2 b-jets
Higgs mass (HCSV/Hbb) is built
from the best btag pair

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


60

Leptons (muons)

Use only events with 2


opposite sign muons of
>20/15 GeV Pt and || < 2.4

Apply for muon WP Loose


and mu_pfRelIso04 < 0.25

Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017


61

Missing transverse energy

PF MET type 1 corrected


Cut on MET > 20 GeV mostly
affects only Drell-Yan
Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017
62

On shell Z ll

As expected Drell-Yan is the main background


Hard to make judgments about shown Z pT of the
300 GeV signal sample, but higher mass samples
will have evident large boost
Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017
63

Higgs bb

Two b-jets with highest CSV score are


used to form H bb
Small statistics of signal MC sample
limits resolution of H bb mass
Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017
64

Higgs ZZ*

Z ll plus MET are used to form Higgs ZZ*


Reconstructed mass of Higgs ZZ*
has a clear positive skew and high mass mean
Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017
65

HH transverse mass

m T2 = E 2 - p 2z

One of our main handles for the analysis


Rami Kamalieddin (UNL) HHbbZZ2b2l2 UNLHEP 15 May 2017

Вам также может понравиться