1. Top Tagging at CLIC 1.4TeV Using Jet Substructure
Alasdair Winter
University of Birmingham
21/02/2017

2. Overview Aims: identify TTBar decays in boosted events
select events without introducing bias to the top mass
examine prospects for CPT violation measurements in the top sector
Boosted topology makes conventional top tagging techniques a challenge- btagging alone no longer viable!
Approach will be to use the concept of fat jets and look at jet substructure
Only focusing on event selection- NOT reconstruction! (yet…)

3. Samples used 𝑠 =1.4TeV P(e-) = -80% for signal y=d,s,b x=u,c l=mu,tau
Process ID
CrossSection (fb)
N Generated Events
eeyyveyx
6589
31.9
790,000
eeyyxyev
6592
29.0
800,000
eeyyvlyx
6634
40.7
1,140,000
eeyyxylv
6637
1,130,000
eeqqlv
3249
4309.7
980,000
eeqqqq
2163
1328.1
300,000
eeqq
2091
4009.5
460,000
eeqqqqlv
2169
115.3
180,000
eeqqqqll
2166
71.7
210,000
eeqqqqvv
2152
24.7
230,000
Signal!

4. MC Sample Selection- Stage 1
Samples are 6 fermion final states consistent with TTBar final state- need to select only those that are truly from TTBar decays
Check truth tables contain:
2 b quarks
2 light quarks
Combination of light quarks and single b quark consistent with top mass GeV
Remove events where the isolated lepton is a Tau (will add these as background in future)
Process
Sample CrossSection (fb)
Pass Rate(%)
New CrossSection (fb)
eeyyveyx
31.9
19.8
15.7
eeyyxyev
29.0
20.3
15.1
eeyyvlyx
40.7
12.5
13.7
eeyyxylv
eeqqqqlv
115.3
39.4
45.4
Also need to remove signal from background sample

5. MC Selection- stage 2 Broad tail in energy spectrum
Split analysis into boosted and low E regions, design separate selection for both
Work presented here is mainly focused on the boosted region
1350GeV

6. MC Selection- stage 2 Process Whole Sample CrossSection (fb)
Boosted Events CrossSection (fb)
LowE CrossSection (fb)
eeyyveyx
15.7
6.32
9.41
eeyyxyev
15.1
5.90
9.16
eeyyvlyx
13.7
5.10
8.61
eeyyxylv
5.08
8.59
High E BDT Discriminator
BDT Score 1
Event Selection
Final
Discriminating Variable
Event Variables
Low Energy BDT Discriminator
BDT Score 2

7. Analysis Strategy Find Isolated Lepton in TightSelectedPFOs1
Find Isolated Lepton in TightSelectedPFOs 2
Cluster remaining PFOs into 2 Fat Jets 3
Determine FatJet structure- multiplicity, nSubJettiness, angular distribution of subjets 4
Apply 2 BDTs using kinematic and jet substructure variables. One looking for boosted top events, one for low energy tops. 5
Use scores from each BDT as discriminating variables for a second tier classifier that makes the final selection 6
Check top mass dependence on final selection before looking at physics studies

8. 1) Lepton Identification
Cluster all TightSelectedPFOs into 5 jets using ee_kt_algorithm 2
Select TightPFOs with Pandora PID=11 or 13, E>10GeV as isolated lepton candidates 3
Calculate 𝐸 𝑃𝐹𝑂 𝐸 𝐽𝑒𝑡 (where EJet is the energy of the jet the PFO was clustered into) for each candidate- provides an estimate for how isolated the candidate is 4
Select the PFO with the highest 𝐸 𝑃𝐹𝑂 𝐸 𝐽𝑒𝑡 ratio to be the isolated lepton 5
If no candidate is selected, repeat process without the inclusion of step 2
By construction, 1 lepton identified per event
Purity= 85%
High purity essential for many applications

9. 2) Fat Jet Finding Combine remaining PFOs into two jets-
ee_kt_algorithm and Valencia (1.5,1,1) tested
ExclusiveNJets=2
Declare jet with the highest energy to be the hadronically decaying top
Narrow peak at 174 GeV but peaks at low mass too...
Broad peak at 184GeV
ee kt algortihm
Valencia
Preferred for now
Need to investigate distribution post selection

10. 3) Jet Substructure
MicroJet Counting- recluster fat jets into microjets with very small ΔR and count them to find jet multiplicity
R increased from to 0.5, supress possible effects of systematics arising from hadronization modelling
NSubjettiness- estimate number of prongs by reclustering fat jet into N subjets and measuring distance of microjets relative to the jet axes
Jet topology- recluster fat jet into three subjets and measure the relative angles between subjets
Redefinition of microjets has an impact on many jet related variables:
Jet variables no longer as highly ranked by BDT
Nsubjettiness removed completely from selection process
Need to examine systematics in detail to understand if this change is necessary
Currently looking into using Herwig for hadronization rather than Pythia

11. 4) Variables Used for BDT-22
Leptonic Top Properties
Jet Mass
Jet Energy
Jet Multiplicity
Jet Pt
Lep Charge
Lep Energy
Lep Pz
Lep Pt
Lep PID
Hadronic Top Properties
Fat Jet Energy
Fat Jet Multiplicity
Fat Jet Pt
Angle between fat jet and beam axis
W Mass
W Energy
W Pt
Angle between highest and mid E subjets
General Event Properties
Visible Energy
Visible Pt
Visible Pz
Y45 (switch from 4-5 jet topology)
Angular separation of isolep and hadronic top

12. BDT Response

13. Significance Plot from TMVA

14. Expected number of events after selection
Taking L = 1.5ab-1, BDT cut= 0.19
Significance = 104 ( 𝑆 𝑆+𝐵 )
Signal Efficiency = 52%
Background efficiency = 0.07%
Purity after selection = 62%
Process
CrossSection (fb)
N Generated Events
Selection Efficiency (%)
Expected Events Passing for 1.5ab-1
eeyyveyx
6.32
790,000 42
4000
eeyyxyev
5.90
800,000 44
eeyyvlyx
5.10
1,140,000 64
4900
eeyyxylv
5.08
1,130,000 63
4800
eeqqlv
4309.7
980,000
0.02
1000
eeqqqq
1328.1
300,000
0.04
900
eeqq
4009.5
460,000
0.03
1600
eeqqqqlv
45.4
180,000
9.6
6500
eeqqqqll
71.7
210,000
0.78
800
eeqqqqvv
24.7
230,000
0.16
100
Need to examine this in more detail- is it really all background??

15. Very Preliminary Results….

16. Low E Selection
Naively using same BDT variables as for boosted events, but train with low E events
Maximum Significance = 144
Signal Efficiency = 67 %
Background efficiency = 0.18 %
Purity after selection = 57%
Process
CrossSection (fb)
N Generated Events
Selection Efficiency (%)
Expected Events Passing for 1.5ab-1
eeyyveyx
6.32
790,000 60
8500
eeyyxyev
5.90
800,000
8200
eeyyvlyx
5.10
1,140,000 73
9500
eeyyxylv
5.08
1,130,000 76
9800
eeqqlv
4309.7
980,000
0.05
3200
eeqqqq
1328.1
300,000
0.18
3600
eeqq
4009.5
460,000
0.07
4400
eeqqqqlv
45.4
180,000 18
11800
eeqqqqll
71.7
210,000
2.8
3000
eeqqqqvv
24.7
230,000
1.9
700

17. Combined Selection
Accept any events that pass either of the first tier BDT cuts- not optimal way to combine BDTs but quickest!
Significance = 195
Signal Efficiency = 66 %
Background efficiency = 0.2 %
Purity after selection = 65%
Process
CrossSection (fb)
N Generated Events
Selection Efficiency (%)
Expected Events Passing for 1.5ab-1
Signal
58.2
3,800,000 66
58,000
eeqqlv
4309.7
980,000
0.06
3,800
eeqqqq
1328.1
300,000
0.02
4,100
eeqq
4009.5
460,000
0.09
5,100
eeqqqqlv
45.4
180,000 20
13,300
eeqqqqll
71.7
210,000
3.2
3,400
eeqqqqvv
24.7
230,000
1.9
700
Improved performance appears to arise from boosted signal events that are missed in the boosted BDT selection being picked up in the low E selection

18. Still to do….
Implement more sophisticated second tier of selection (secondary BDT?)
Examine systematics from jet related variables
Ongoing effort to produce events in Whizard then switch to Herwig for hadronization
Look at dependence on the MC mass window we choose for the top
Top mass reconstruction after event selection
Reinvestigate use of Valencia algorithm for top reconstruction based on work by Lars, Martin…
Determine prospects for CPT measurements
Extend approach to look at 3TeV stage

19. Summary
Switching to new analysis approach- two tiered selection looking for both boosted and low energy tops
New lepton finder implemented based on jet isolation
ee_kt_algorithm still preferred for fat jet finding
Additional backgrounds added to BDT, though necessary to further examine the effect of MC selection to make sure there is no mixing between signal and background samples
Microjet parameters have been adjusted to reduce possible systematics, however this reduced the discriminating power of many variables previously used for event selection

20. Overtraining checks