Higgs Branching RatioH→bb, cc, gg

for new detector models at ILDMasakazu Kurata, Ryo Yonamine, Hiroaki Ono

ILD meeting

02/27/2019

Introduction • Start to re-optimize new detector models using physics

benchmarks

• Higgs analyses are one of the most important benchmark processes for detector optimization at ILD

• H→bb, cc, gg measurement:• Flavor tagging is indispensable for jet flavor separation

• Try@500GeV, focus on ννH process

• Based on Hiroaki’s 1TeV analysis in DBD era

• So far, do not consider separation of ZH and VBF• But, considering it for future plan

2

LCFIPlus summary: Comparison among any situation• Ryo made great efforts to resolve internal problems in

LCFIPlus for new verson of iLCsoft

• Use 6f flavor tag samples: bbbbbb, cccccc, qqqqqq

• Tune MVA training torecover performance tothose of DBD era

• Almost same performanceas DBD era

3

Comparison between large and small• MVAoutput between large and small

• Very similar…

4

but

5

Small has excess hereAll the jet flavors have excess

After selectionLargesmall

Check samples• Check nnH→nncc sample

• Compare MVAoutput between Large and Small

• Same tendency can be seen• This excess directly reflects on measurement precision

• No bias on c-jet likeliness from:• Event selection

• Beam background rejection

6

Status of analysis

Iso

late

d L

epto

n ID Look for

electron and muon

Jet

Clu

ster

ing Durham with beam background rejection

2 jet clustering

Even

t se

lect

ion Reject

backgrounds

Preselection

MVA

Polarization (e-, e+)=(-,+) (+,-) (-,-) (+,+)

Luminosity(fb-1) 1600 1600 400 400

• ECM=500GeV• Luminosity: 4ab-1

• Analysis flow

• Signal and backgrounds• Signal: use nnH→nnbb, nncc, nngg• Backgrounds: 2f, 4f, 5f, 6f, aa, ZH, nnh→nn(no bb,cc,gg) 7

Beam background rejection

• 𝑦𝑖𝑗 =2min(𝐸𝑖

2, 𝐸𝑗2)(1−cos 𝜃)

𝐸𝑣𝑖𝑠2 , 𝑦𝑏𝑒𝑎𝑚 =

2𝐸𝑖2α2(1−cos 𝜃)

𝐸𝑣𝑖𝑠2

α: beam rejection parameter

smaller→ beam rejection becomes stronger• Particle i with yij>ybeam is discarded

8

ννZ@500GeV(DBD) • 2 jet clustering• Parameters are tuned

for better result

w/o beam b.g. rejectionKtDurhamValencia

Apply to nnH events• Apply Durham beam b.g. rejection to nnH→nnbb

events

• Parameter scan to make Higgs mass distribution better• 5.5 seems best for both detector models

large small

9

Preselection

10

• Just reject backgrounds which are trivial ones• Signal tail part of each variable is rejected

• No Isolated lepton

• Njets=2

• Npfo≧5

• Evis≦300

• 40≦m2jets≦200

Multivariate Analysis

• Expect better background rejection efficiency than cut based

• Need to reject other Higgs processes as much as SM backgrounds

• Important to separate both SM backgrounds and other Higgs processes for backgrounds

• Use Binary Classification• Due to the phase space difference of each background

component

• Signal vs.: other Higgs, 2f&4f, 5f&6f&aa

• 3kinds of classifier trained

11

Input variables• Need to suppress bias between H→bb,cc and H→gg

• Reduce difference by combining variables• m2jets, cosθH, cosθjj, Ej1, cosθj1, Ej2, cosθj2, mmiss, Pt,

Principal Thrust, Major Thrust, Log(y23)*npfo, Log(y34)*npfo

• Example: signal vs. 2f&4f

12

Stacked

13

vs. other Higgs vs. 2f&4f

vs. 5f&6f&aa

• Determine operation point:MVA_1>x.xx && MVA_2>y.yy && MVA_3>z.zz

Cut table• Large, (e-, e+)=(-,+) polarization,

L=1600fb-1

Signal efficiency: typically ～60% for all the polarization & detector models

Significance: 275.3

14

Template fit• Toy MC to extract the measurement precision of H→bb,

cc, gg• From LCFIPlus output, calculate x-likeliness in each event:

x1, x2: LCFIPlus output(b, c, bc=c/(b+c))

• Create 3-D template with those b, c, bc likeliness

• Fitting is performed according to Poisson statistics• 3 scale factors of signal events are parameters, other fixed:

• Do Toy MC• 10000 pseudo-experiments performed

15

Template examples• Templates are 3-D, so project into 2-D space

16

Results• Large detector, all the polarization

• Small detector, all the polarization

• Comparison with DBD for check• (-0.8, +0.3) 500fb-1

17

Process(-,+) H→bb H→cc H→gg

Precision(%) 0.69 6.79 2.89

Process(-,+) H→bb H→cc H→gg

Precision(%) 0.66 6.2 4.1

IDR-L DBD(125GeV scaled)

Plot• Results of branching ratio measurement

18

Plot• Combined only

19

• We construct pseudo jet flavor taggers for comparison

• We do the pseudo experiments with those (artificially) perfect & pessimistic case

• Artificial fraction• Perfect Pessimistic

• According to those fraction, jet flavor is determined• All the samples are processed, and measurement precision is

estimated

Impact of flavor tagging on precision measurement

20

• (-,+) polarization, L=1600fb-1

• Better flavor tagging is very important for better precision• b-tagging is excellent• Better c-tagging is important

Comparison between pseudo experiments

21

Plot• Impact of flavor tagging performance on measurement

22

Summary & Prospects• We can get the results of both detectors & all the

polarizations• B-tagging is very excellent already

• Small detector has better performance than large in H→cc measurement

• Coming from c-tag output distribution

• Thanks to stronger B-field(4.0T>3.5T)?

• We will investigate that point

• Template fitting is being checked now• We cannot see huge bias on binning so far(backup)

• Checking and discussing fitting stability with reviewers now

• Note & Going to IDR input23

Backups

24

MVA output

25

vs. other Higgs vs. 2f&4f

vs. 5f&6f&aa

• large

26

• small

27

Check template fit• Need to check

• Num. of empty bins: must be as small as possible

• Bins with small statistics: need to avoid

Should be checked

• Num. of binning is important• To include the shape effect

• This is the advantage of template fit

• The tendency of template binning:• b-likeliness is very robust: num. of bin is 2 enough

• H→gg is relatively robust: don’t need much care about (b, c, bc)=(0.0,0.0,0.0) value

• Shape of c & bc-likeliness is very important

⇒H→cc precision varies very much 28

One case• Num. of bins: (b-, c-, bc-)=(2,3,3) (LCWS: (10,10,10))

• Variable bin size is used(LCWS: 0.1 interval from 0 to 1)

• Check the bin stat. (all the events accumulated)

29

b-likeliness<0.4Num. of empty bins: 2Smallest stat: 1072

b-likeliness>0.4Num. of empty bins: 7Smallest stat: 904.8

More aggressive case• Num. of bins: (b-, c-, bc-)=(2,5,5) to include shape info more

• Variable bin size is used

• Check the bin stat. (all the events accumulated)• Typically, num. of events in a bin is >～100

30

b-likeliness<0.4Num. of empty bins: 10Smallest stat: 98.1

b-likeliness>0.4Num. of empty bins: 19Smallest stat: 11.0

Check different binning effect• (e-,e+)=(-,+),L=1600fb-1

• Large:

• Small:

• Small is better for c-tagging• Results does not change so drastically, but H→cc varies if shape info includes more(have to be

careful of statistics)• Zero bins do not affect well• We use (2,5,5) binning for result

31

Large(10,10,10) H→bb H→cc H→gg

Precision(%) 0.43 3.80 1.68

Large(2,3,3) H→bb H→cc H→gg

Precision(%) 0.43±0.01 3.98±0.02 1.76±0.01

Small(2,3,3) H→bb H→cc H→gg

Precision(%) 0.43±0.01 3.69±0.02 1.81±0.01

Small(10,10,10) H→bb H→cc H→gg

Precision(%) 0.42 3.55 1.73

Large(2,5,5) H→bb H→cc H→gg

Precision(%) 0.43±0.01 3.88±0.02 1.70±0.01

Small(2,5,5) H→bb H→cc H→gg

Precision(%) 0.44±0.01 3.56±0.02 1.77±0.01

Optimization with 2 detector models

• Re-optimize ILD detector• Revisit optimization of cost and detector performance

Detector models ILD-L ILD-S

B-field 3.5T 4T

VTX inner radius 1.6cm 1.6cm

TPC inner radius 33cm 33cm

TPC outer radius 180cm 146cm

TPC length (z/2) 235cm 235cm

Inner ECAL radius 184cm 150cm

Outer ECAL radius 202.5cm 168.5cm

Inner HCAL radius 206cm 172cm

Outer HCAL radius 335cm 301cm

Coil inner radius 344cm 310cm

ILD-L ILD-S

32

Check samples• Beam background rejection?

• No bias from beam background rejection

33

Before beam b.g. rejectionLargeSmall