A Study of Electron Identification Jim Branson UCSD with collaborators from FNAL, UCSB & UCSD.

A Study of Electron Identification

Jim BransonUCSD

with collaborators from FNAL, UCSB & UCSD

2

Electron ID in CMS

• Need high efficiency, particularly in H4 lepton channel.• Need efficiency at low ET for for low mass H4 lepton and

HWW– Background increases at low ET– High background from fakes will complicate analyses

• CMS has some unusual features that impact electron ID.– High field– Thick tracker– Ecal with e/ near 2

• Need to use all the tools available to understand and optimize electron ID.– Understand physics of electron measurement now, – multivariate analyses… could come later, perhaps.

3

This is a Study of e ID• Attempt to learn about e ID in CMS

– No multivariate analysis for now

• Don’t feel constrained by current algorithms– But benefit from what was learned

• Aim for high selection efficiency, 97%• Try to compare to current cut-based e-ID

– But “the current ID” is not fully ready– And we’ve changed enough that comparison won’t be

fair.

• Look at very simple 5 cuts• And at what we can gain with some understanding

of the measurement.– Hopefully still “simple enough” (and more stable?)

4

Standard Electron ID

• Match in and • Loose cut on E/p (not very useful in CMS)

• Shower shape requirements (only in CMS)

• E/H cut (not very powerful in CMS (barrel))

We can use other “features” of CMS to help with electron ID.

5

E/p often Affected by p Measurement

E/p E/p

ESC −EMC

EMC

ptrack −pMC

pMC

Zee

6

Remove Some Producer Cuts

double maxEOverPBarrel = 3. double maxEOverPEndcaps = 5. double minEOverPBarrel = 0.35 double minEOverPEndcaps = 0.35 double maxHOverE = 0.2 double maxDeltaEta = 0.02 double maxDeltaPhi = 0.1 double ptCut = 5. 1.5

Keep these

Some producer cuts are too tight.

This increases the denominator for selection efficiency calculations and therefore decreases the efficiency calculated.

**See talk of Matteo Sani

7

Apply Simple Cuts on Five Variables

1. < cut

2. in < cut

3. Eseed/Pout > cut *can we replace this?

4. in < cut

5. H/E < cut

With straight cuts on these quantities, we get 97.3% selection efficiency and 4.1% of dijet(50-80) events

having a fake electron.

8

How much better can we do by using a bit more complex

cut-based analysis?

To describe the electron ID algorithm, we will show plots from

the Barrel only for simplicity.

Others in backup slides.

9

Use this 2D Plot to Study e ID

Electrons from Zee

Fakes from jets

E/p

fbrem

10

Color the PictureElectrons from Zee Fakes from jets

Electrons and jets overlap (Pyrite).

Mainly GOOD electrons Mainly fakes

We call a color in this picture a category.Are there selection differences beween the categories?

11

What is the Physics BehindElectrons from Zee

1. E/p is often well measured for electrons

2. Electrons usually radiate a good deal of energy in the tracker

3. E/p is not often measured to be less than 1 for electronsfbrem

E/p

12

What is the Physics Behind

1. Fakes from jets usually have fbrem around 0 (just charged pion tracks…)

2. Many fakes from jets have E/p<1 partly because of the low response of ECAL to charged pions...

Maybe some enhancement

due to interaction in tracker

13

Why Categories?

• Large differences in s/b in regions of this plot.– Looser cuts in high s/b regions.

• Try to use robust differences between electrons and jets.

Electrons Jetsfakes

14

Categories do two things

• Separate regions of high signal to background from low.

• Put events of similar characteristics together.– Well measured electrons– Electrons with track problems– Electrons with supercluster problems.– Fakes due to overlap– Fakes due to charge exchange…

15

Apply simple cuts in each category

1. < cut

2. in < cut

3. Eseed/Pout > cut

4. in < cut

5. H/E < cutLook in Seven categories (for STUDY).Look at barrel and endcap separately.

No Isolation Cuts Applied

16

Selection Cuts

E/p

Fbrem

Eseed/Pout

H/E

All electronsSurviving electronsET<15 electrons failing cutsAll fake candidatesSurviving fakes

17

Base Categories on This Plot

E/p

Fbrem=(Pin-Pout)/Pin

0

5

47

3

2

1

18

in Categories

0 1 2

3 4 5Electrons

Jets

“n-1” plotsLooser cuts in best category

19

in Categories

0 1 2 3

4 57

Electrons

Jets

Basically, we tighten the cut until it starts to cut more than a few electrons.

0 1 2

3 4 5

“n-1” plots

Tighter cuts in overlap category

20

in Categories

0 1 2 3

4 57

Electrons

Jets

0 1 2

3 4 5

“n-1” plots Looser cut for worst E/p category

21

Eseed/Pout in Categories

0 1 2 3

4 57

Electrons

Jets

0 1 2

3 4 5

“n-1” plots

22

H/E in Categories

0 1 2 3

4 57

Electrons

Jets

0 1 2

3 4 5

“n-1” plots

23

The Cuts: 22 valuesCat. 0 1 2 3 4 5 Ncut

< 0.014 0.0125 0.0125 0.0125 0.0125 0.0125 2

0.009 0.0085 0.0085 0.0035 0.0085 0.0035 3

0.06 0.06 0.10 0.02 0.06 0.06 3

E/Pout> 0.55 0.88 0.88 0.88 0.88 0.88 2

H/E 0.125 0.055 0.055 0.10 0.055 0.055 3

< 0.031 0.028 0.028 0.028 0.028 0.028 2

0.009 0.009 0.009 0.009 0.009 0.009 1

0.06 0.10 0.10 0.02 0.10 0.10 3

E/Pout> 0.85 0.85 0.85 0.85 0.85 0.85 1

H/E< 0.125 0.10 0.10 0.10 0.10 0.10 2

Use looser cuts for the best electrons.

Critical overlapCategory with

E/p=1

Barrel

Endcap

bad if E/p bad

For selection, we really only need 3-4 categories.

24

Why 22 Cut ParametersBarrel Endcap Sum

5 Cut-Variables 5 5 10

Looser Cuts on best electronsClaim looser cuts on best electrons make us less sensitive at startup!

4 3 7

Different Cuts in overlap region

and high E/p overlap 3 1 4

Open cut for worst E/p

1 0 1

Total 5+8 5+4 22

25

Simplification

• Reduce to 3 categories– Brem e with E/p1– Little brem e– Bad track E/p≠1 (brem)

• Open up cut a bit for E/p>>1• (Cut out low E/p && low fbrem region where there are almost

no e)

26

Why I Like this Selection

• Looser cuts for best electrons makes selection more robust.

• Tighter cuts in overlap region maintains robust rejection of fakes.

• Robust separation of regions since pion tracks should have fbrem0.– No dependence on clustering algorithms.

• (We must work with E/p for energy measurement anyway)

27

Nota Bene

• The 3 Categories can just be used for the selection.

• There is no reason we need to use them for later analysis.

28

Results on Standard Electron Reco

ETSC>5

-2.5<<2.5

THIS Selection

22 param.

THIS

Just 5 cuts,

Not tuned

Standard Selection

CrackGold

Selection Efficiency for e

from Zee97.0% 97.3% 91.8%**

Fakes per di-jet

event QCD_50_80

1.5%

(0.75%per jet)

4.7%

(2.3%per jet)

8.8%**

(4.4% per jet)

*Not fair: All crack electrons are accepted in the standard selection.

** Coded by us for comparison under same conditions. highly preliminary.

No Isolation Cuts Applied

29

Efficiency

Reco Eff.Reco*Sel. Eff. (THIS)Reco*Sel. Eff. ootb

All use same standard reco.Crack electrons treated as golden for ootb selec.

ET

30

Cracks

• Selection efficiency dips by about 5% in the cracks in the ECAL.– The crack events have distributions similar to

background– They are removed by all of the cuts.– Special selection in cracks would clearly allow

background in cracks.

• We don’t think there is much to be gained in the cracks.

31

Accepted Crack Electrons: SC OK

ESC/E

All electronsCrack electrons

32

Low ET Electrons

• Our selection efficiency also dips by about 5% at low ET.

• Some events with low E/p populate background regions.

• Probably best to look at SuperCluster reco to see if some energy can be recovered.

• Would also help reconstruction efficiency.

33

More Cuts? No.

Isolation

34

Summary of Loose Selection

• We can have a loose selection for electrons with about 97% efficiency and low fake rate.

• A hopefully “robust” categorization uses E/p and fbrem.– Electrons and fakes separate to a great extent

in this categorization.– Separate different types of measured events.

• Simple cuts are used to work on understanding of electrons.

06/08/2007 - EGamma POG meeting

35

New Electron Sequence• Try to match each reconstructed Super Cluster to a standard

Combinatorial Track Finder (CTF) seed:– using the GlobalMixedSeed collection:

• we expect an efficiency improvement at least at large eta since the number of pixel layers is lower and the CTF seeder uses the Silicon Strip hits in addition.

– at the moment the matching is done with all the seeds inside a fixed size cone around the Super Cluster direction.

• For each matched seed we propagate the track using the standard Gaussian Sum Filter (GSF) builder:

• Feed the matched pair (Super Cluster & track) to the standard electron producer adapted to return a new collection of identical objects labelled as “GlobalGsfElectrons”.

From Matteo Sani

GGE See also talk in Higgs meeting by Boris Mangano.

06/08/2007 - EGamma POG meeting

36

Comparison in Zee events• The following table reports the number of candidate

electrons reconstructed by either our custom or the standard algorithm.

• A candidate is defined by a Super Cluster and a CTF track matching the MC electron direction.

OOTB152 – OK OOTB152 – FAILS

GGE – OK 5296 734

GGE - FAILS 28 84

GGE reconstructs electron GGE reconstructs electron

but OOTB152 fails.but OOTB152 fails.

OOTB152 reconstructsOOTB152 reconstructs

electron but GGE fails.electron but GGE fails.From Matteo Sani

37

Preliminary e-ID PerformanceOn GGE reco sequence

ETSC>5

-2.5<<2.5

THIS Selection

Standard Selection

CrackGold

Selection Efficiency for e from

Zee

96.8% 90.7%

Fakes per dijet event

QCD_50_80

1.2%

(0.6% per jet)

5.6%

(2.8% per jet)

This reco is new and a

rapidly moving target.

Results are therefore

very preliminary.

38

A Look at Tighter e-ID

-2.5<<2.5

PTMC>5

Selection Efficiency for e

from Zee

Fakes per dijet event

QCD_50_80

ETSC>5

Fakes per dijet event

QCD_50_80

ETSC>10

Loose Selection

5 cut params.

TkIso<7

96.9%

96.9%

3.1%

1.5%

2.2%

0.99%

Loose Selection

(6 cat 60 param.)

TkIso<7

97.2%

97.4%

1.3%

0.60%

0.95%

0.42%

Loose Selection

(22 cut param.)

TkIso<7

96.8%

96.6%

1.2%

0.56

0.94%

0.40%

Tight Selection

TkIso<7

94.1%

94.4%

0.53%

0.22%

0.40%

0.16%

39

Summary

• New electron loose electron selection with high efficiency and low background.– Even good with very simple cuts

• Study of simple vs. “less simple” selection– Factor of 2.5 in fake rate

• Study of tighter selection– 2.7% lower eff. For more than factor of 2 in fake rate

• Isolation seems to commute with selection• New reconstruction sequence using standard

seeder from tracker, also shows similar high selection efficiency and low background.– With higher reco efficiency