hands-on advanced tutorial session on jet substructure
TRANSCRIPT
Hands-on Advanced Tutorial Session on Jet Substructure
The Jet Substructure HATS team(Jake Anderson, Jim Dolen, Kalanand Mishra,
Ilya Osipenkov, Nhan Tran)
A pedagogical introduction
https://twiki.cern.ch/twiki/bin/view/CMS/SWGuideHATSJetSubstructureTwiki:
Indico: https://indico.cern.ch/conferenceDisplay.py?confId=251174
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 2
Outline
•Jet clustering: algorithms and calibration
•Jet substructure: introduction and motivation•grooming: filtering, pruning, trimming, mass drop
•Jet substructure distributions for garden-variety jets•comparisons of data & simulation, experimental effects
•Jet substructure observables and signal reconstruction•illustrate the techniques with W and top tagging
•Summary
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 3
Some bibliographyPhenomenology
(Introduced modern lingo and conventions)
‣CMS collaboration, “Jet mass in dijet and W/Z+jet events”, arXiv:1303.4811‣ATLAS collaboration, “Jet mass & substructure in incl. jets”, arXiv:1203.4606
Experimental results
(Exhaustive description of jet algorithms, “manifesto” of Fastjet)
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 4
Reality of event complexity at hadron colliders
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 5
Taming the reality: a reductionist approach
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 6
Zooming in to single objects
High energy partons unavoidably lead to collimated bunch of hadrons
parton
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 7
Our reductionist view of the same event
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 9
A fun exercise
1. which particles to put together ?2. how to combine their momenta?
Tell me how many jets you see here? 0, 1, 2, ..., 10, or more?
Two considerations:
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 10
So, what is a jet then? Formal definition
in modern lingo, i.e., post circa 2007
Most commonly used: direct 4-vector sums (E-scheme)
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 11
Two main classes of jet algorithms
Examples: kt, Cambridge-Aachen, anti-kt, ...
Examples: JetClu, MidPoint, SISCone,...
‣These algorithms are the most widely used nowadays, will discuss more about them in the following slides.
‣These are mostly historic, will NOT discuss them further.
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 12
Sequential recombination algorithms
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 13
Algorithmically, a jet is simply a collection of particles
Jet’s “reach”
For a number of reasons, it is however useful to consider its spatial extent, i.e., given the position of its axis, up to where does it collect particles? What is its shape?
These details are important for a number of corrections of various origin: perturbative, hadronization, pileup, detector-related etc.
Note that the intuitive picture of a jet being a cone (or radius R) is wrong. This is what kt jets look like.
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 14
From jet “reach” to jet areas ....
Cacciari, Salam, Soyez arXiv: 0802.1188
In the large number of particles limit all areas converge to the same value.
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 15
Jet areas for the most widely used algorithms
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 16
Jets in CMS
•Use charged hadron subtracted (CHS) Particle Flow jets-Remove all charged hadrons not from the primary vertex
•Neutrals are subtracted via jet area ρ correction where ρ is energy density computed with the KT6 algorithm
-Active area used in all cases including groomed
•Non-linearities in η and pT are corrected for using data & MC-AK5, AK7 jets have dedicated corrections while other large radii jets use AK7 corrections
•Jet finding & grooming algorithms implemented via FastJet3
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 17
Example of jets and jet observables in CMS
Jets extensively measured in hadronic collisions.Very good agreement with pQCD predictions over 10 orders of magnitude.
A dijet event from early run
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 18
Jet energy calibration: overview
Factorization facilitates the use of data-driven corrections- Breaking the correction into pieces that are naturally measured in collider data:
•Offset: pile-up and noise measured in zero-bias events.•MC: jet response vs. η, PT using MC truth.•Residual: jet response vs. η, PT using dijet balance and γ/Z+jet in data.
ReconstructedJets: after CHS
CalibratedJets
Offset: FastJet MC: η, pT Residual: η, pT
Required Corrections
In CMS the most widely used jet is anti-kT 0.5 (0.7 for QCD measurements). Jet substructure studies done with anti-kT 0.5, 0.7, 0.8 and CA 0.8 with various grooming techniques.
η-5 -4 -3 -2 -1 0 1 2 3 4 5
, GeV
〉 T,
offs
etp〈
0
0.2
0.4
0.6
0.8
1
1.2
1.4photonsem depositsneutral hadronshadronic depositscharged pile-upcharged hadrons
PVOffset/N
CMS preliminary = 8 TeVs
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 19
Pileup contribution to jet energy
✦Pileup (PU) measured with Zero Bias data •Most charged hadrons can be associated to pileup vertices and removed
•Part that can be removed is labeled “charged hadrons”•Part that remains as PU needs to be subtracted PU density x effective area
(FastJet-ρ)
https://twiki.cern.ch/twiki/bin/viewauth/CMS/SWGuideCMSDataAnalysisSchoolJetAnalysisMore details on jet calibration in CMSDAS short exercise
! Jet substructure is currently a very popular tool to analyze jets that are produced from heavy objects such as top quarks, Higgs bosons, W/Z bosons, or any beyond-Standard-Model hadronically-decaying object. ! An excellent overview of jet substructure techniques can be found here: Boosted objects: a probe of beyond the Standard Model physics by Abdesselam et al. ! Jet pruning algorithm: Designed to find heavy objects within a jet, and in the process, removes soft and wide-angle clusters from the clustering sequence.
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 20
Introduction to Jet Substructure & jet grooming! Jet substructure is currently a very popular tool to analyze jets that are produced from heavy objects such as top quarks, Higgs bosons, W/Z bosons, or any beyond-Standard-Model hadronically-decaying object. ! An excellent overview of jet substructure techniques can be found here: Boosted objects: a probe of beyond the Standard Model physics by Abdesselam et al. ! Jet pruning algorithm: Designed to find heavy objects within a jet, and in the process, removes soft and wide-angle clusters from the clustering sequence.
• Problem! Traditional techniques start to lose sensitivity (in part) due to jet merging at higher masses!
• Cannot rely on methods to assign partons to jets anymore
• Have to consider cases where partons merge into a single jet
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 21
Jet substructure: motivation
• Problem! Traditional techniques start to lose sensitivity (in part) due to jet merging at higher masses!
• Cannot rely on methods to assign partons to jets anymore
• Have to consider cases where partons merge into a single jet
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 22
Jet substructure: motivation
• Problem! Traditional techniques start to lose sensitivity (in part) due to jet merging at higher masses!
• Cannot rely on methods to assign partons to jets anymore
• Have to consider cases where partons merge into a single jet
To understand how to break up a jet, first we need to understand what it is!
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 23
Jet substructure: motivation
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 24
Jet grooming techniques I
✦Study jet mass properties with three grooming techniques • Filtering: arxiv: 0802.2470• Trimming: arxiv: 0912.1342• Pruning: arxiv: 0903.5081• This round of analysis uses default parameters from each
of the references.✦Filtering
• reclustering jet constituents with smaller radius, Rfilt, keeping nfilt hardest sub-jets
• default parameters: Rfilt = 0.3, nfilt = 3• sub-jet clustering algorithm:
Cambridge-Aachen
Steve Ellis, Chris Vermilion, Jon Walsh
D. Krohn, Jesse Thaler, LianTao Wang
Butterworth, Davison, Rubin, Salam
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 25
Jet grooming techniques II
✦Trimming• reclustering with smaller radius, Rfilt, keeping sub-jets with a fraction,
pTfrac,min, of original jet pT
• default parameters: Rfilt = 0.2, pTfrac,min = 0.03• sub-jet clustering algorithm: kT
✦Pruning• reclustering with sequential recombination algorithm, veto soft and
large-angle recombinations between pseudojets i and j• veto: ΔRij > rcut×2m/pT; z = min(pTi,pTj)/pTi+j < zcut
• default parameters: zcut = 0.1, default rcut = 0.5• sub-jet clustering algorithm: Cambridge-Aachen
trimming
1 2
34
1 2
3
cluster with rfilt < R
keep nfilt
1 2
34
1 2
3
cluster with rfilt < R
keep pTi > pTfrac
4
veto soft or large angle
recombinations
min(pTi,pTj)/pTi+j < zcut or dij > rcut×2m/pT
filtering
trimming
pruning
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 26
Comparison of the three grooming techniques
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 27
Jet mass and substructure
•Jet substructure can be used to improve sensitivity of hadronic decays of boosted heavy particles such as Higgs, W/Z, and top
•Requires understanding of QCD radiation inside jet, structure of constituent particles
•In CMS, have analyzed jet mass properties-inclusive & systematic comparison of jet grooming algorithms
•Use di-jet and V+jets events as complimentary probes of gluon- and quark-enriched environments, respectively
Mass (GeV)0 20 40 60 80 100 120 140 160 180 20
-1Ev
ents
/ 8G
eV /
30fb
0
2
4
6
8
10
12
14
16
Mass (GeV)0 20 40 60 80 100 120 140 160 180 20
-1Ev
ents
/ 8G
eV /
30fb
0
2
4
6
8
10
12
14
16 qqV+jetsVVV+Higgs
= 2.1BS/in 112-128GeV
(a)
Mass (GeV)0 20 40 60 80 100 120 140 160 180 20
-1Ev
ents
/ 8G
eV /
30fb
0
20
40
60
80
100
Mass (GeV)0 20 40 60 80 100 120 140 160 180 20
-1Ev
ents
/ 8G
eV /
30fb
0
20
40
60
80
100 qqV+jetsVVV+Higgs
= 3.1BS/in 112-128GeV
(b)
Mass (GeV)0 20 40 60 80 100 120 140 160 180 20
-1Ev
ents
/ 8G
eV /
30fb
0
5
10
15
20
25
30
35
Mass (GeV)0 20 40 60 80 100 120 140 160 180 20
-1Ev
ents
/ 8G
eV /
30fb
0
5
10
15
20
25
30
35qq
V+jetsVVV+Higgs
= 2.9BS/in 112-128GeV
(c)
Mass (GeV)0 20 40 60 80 100 120 140 160 180 20
-1Ev
ents
/ 8G
eV /
30fb
0
20
40
60
80
100
120
140
Mass (GeV)0 20 40 60 80 100 120 140 160 180 20
-1Ev
ents
/ 8G
eV /
30fb
0
20
40
60
80
100
120
140 qqV+jetsVVV+Higgs
= 4.5BS/in 112-128GeV
(d)
arXiv: 0802.2470
Butterworth, Davison, Rubin, Salam
jet / mGroomjetm
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
Frac
tion
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14Filtered data RECOFiltered PYTHIA6, Z2 RECOFiltered PYTHIA6, Z2 GENTrimmed data RECOTrimmed PYTHIA6, Z2 RECOTrimmed PYTHIA6, Z2 GENPruned data RECOPruned PYTHIA6, Z2 RECOPruned PYTHIA6, Z2 GEN
Filtered data RECOFiltered PYTHIA6, Z2 RECOFiltered PYTHIA6, Z2 GENTrimmed data RECOTrimmed PYTHIA6, Z2 RECOTrimmed PYTHIA6, Z2 GENPruned data RECOPruned PYTHIA6, Z2 RECOPruned PYTHIA6, Z2 GEN
Filtered data RECOFiltered PYTHIA6, Z2 RECOFiltered PYTHIA6, Z2 GENTrimmed data RECOTrimmed PYTHIA6, Z2 RECOTrimmed PYTHIA6, Z2 GENPruned data RECOPruned PYTHIA6, Z2 RECOPruned PYTHIA6, Z2 GEN
= 7 TeV, AK7 Dijetss at -1CMS Preliminary, L = 5 fb
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 28
Jet mass for groomed jets
Comparison of grooming algorithms at particle level (GEN), reconstructed simulation (RECO) level, and in data
Pruning is the most aggressive, filtering is the least aggressive
arXiv:1303.4811
RECOη
-2 0 2
>G
ENun
groo
med
/pT
RECO
ungr
oom
ed<p
T
>G
ENgr
oom
ed/p
TRE
COgr
oom
ed<p
Tdo
uble
ratio
,
0.8
0.9
1
1.1
Ungroomed AK7Trimmed AK7Filtered AK7Pruned AK7
= 7 TeV, AK7 V+jetss at -1CMS Simulation, L = 5fb
[GeV]RECO
pT200 300 400 500
>G
ENun
groo
med
/pT
RECO
ungr
oom
ed<p
T
>G
ENgr
oom
ed/p
TRE
COgr
oom
ed<p
Tdo
uble
ratio
,
0.8
0.9
1
1.1
Ungroomed AK7Trimmed AK7Filtered AK7Pruned AK7
= 7 TeV, AK7 V+jetss at -1CMS Simulation, L = 5fb
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 29
Jet pT response for groomed jets
Groomed jet response within a few % of ungroomed case.
groomed jet response ungroomed jet response
PV N0 5 10 15 20 25 30 35 40
num
ber
0
200
400
600
800
1000
1200
1400
1600
1800 , SimulationPVN
, Simulation reweightedPVN
, DataPVN
= 7 TeVs at -1CMS Preliminary, L = 5fb
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 30
Performance versus pileup
Pileup profile in 7 TeV data•It is of particular interest to understand the sensitivity of large size jets in presence of PU
-Grooming techniques may serve to mitigate pileup sensitivity by effectively reducing the jet area
•Understand performance of mean jet mass as a function of number of primary vertices
PVN
0 10 20 30 40
> (G
eV)
J<m
0
50
100
150data, Ungroomed AK5
Pythia6 Z2, Ungroomed AK5
data, Ungroomed AK7
Pythia6 Z2, Ungroomed AK7
data, Ungroomed AK8
Pythia6 Z2, Ungroomed AK8
Ungroomed AK5:
0.03 GeV± = 0.1PV dm/dN
Ungroomed AK7:
0.03 GeV± = 0.28PV dm/dN
Ungroomed AK8:
0.03 GeV± = 0.33PV dm/dN
= 7 TeV, AK7 W+jetss at -1CMS Preliminary, L = 5fb
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 31
Size matters here: larger jet size ⇒ larger effect
✦Ungroomed jet mass is very sensitive to PU
•<mJ> increases linearly as a function of the number of primary vertices
✦Effect becomes more pronounced as the jet size increases
•AK8 shows much worse effect than AK5
arXiv:1303.4811
PVN
0 10 20 30 40
> (G
eV)
J<m
0
50
100
150data, Ungroomed AK7
Pythia6 Z2, Ungroomed AK7
data, Trimmed AK7
Pythia6 Z2, Trimmed AK7
data, Filtered AK7
Pythia6 Z2, Filtered AK7
data, Pruned AK7
Pythia6 Z2, Pruned AK7
Ungroomed AK7:
0.03 GeV± = 0.28PV dm/dN
Trimmed AK7:
0.04 GeV± = 0.12PV dm/dN
Filtered AK7:
0.02 GeV± = 0.16PV dm/dN
Pruned AK7:
0.05 GeV± = 0.1PV dm/dN
= 7 TeV, AK7 W+jetss at -1CMS Preliminary, L = 5fb
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 32
Grooming helps to ameliorate the situation
✦Groomed jets are less sensitive to PU
•<mJ> vs NPV slope becomes flatter
✦Observe the expected behavior that <mJ> typically scales as R3
arXiv:1303.4811
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 33
• Perform an inclusive comparative study of jet mass for available jet grooming methods
• Make a comparison of data to simulation as a validation
of parton showering models
Understanding the “background”
How the jet mass distribution looks like for garden-variety boosted jets (either in multi-jet or W/Z+jets data)
Plots are presented in increasing aggressiveness of grooming [Ungroomed, Filtered, Trimmed, Pruned] and best digested by flipping through.
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 34
How the mJ looks like for QCD jets: Ungroomed
MC to data agreement is reasonable using both Pythia Z2 and Herwig++, though Herwig++ seems to have better agreement.
arXiv:1303.4811
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 35
Filtered
arXiv:1303.4811
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 37 24
As the grooming algorithm becomes more aggressive, Herwig++ agreement is better with data.
Turnover region moves to the left with harder grooming while tails remain similar in all cases − important feature in new physics searches.
Pruned arXiv:1303.4811
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 38
Jet substructure variables
•There are several jet substructure variables that can be used to discriminate between hadronic decays of boosted heavy particles (Higgs, W/Z, top,...) and the QCD jets
-Already discussed some of them, e.g., jet mass before and after grooming, jet pT, jet size, etc.-Others (# subjets, subjet mass/kinematics, ...) introduced in the following slides in the context of W & top tagging.
•We will exclusively focus on these in our second session today.
• If a top quark or other heavy object is produced with enough energy its decay products can be reconstructed within one jet
• Substructure can be used to identify these jets
- Top jets produce at least 3 subjets
- Two of the subjets should have the W mass and three should have the top mass
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 39
Example of boosted signal: Top jets
• Based on Kaplan et al. (arXiv:0806.0848) • CMS PAS JME-09-001 and CMS AN-2010/080 • Cluster a jet using a sequential recombination algorithm and
then decluster in order to find substructure • Subjets must satisfy two requirements
1. Momentum fraction criterion - Subjets should carry a significant fraction of the jets momentum ( pTsubjet > 0.05×pThard jet )
2. Adjacency criterion - subjets should be separated by some minimum distance ( ΔR(CA, CB) > 0.4 - 0.0004×pT )
• Iterative process - throw out subjets that fail momentum fraction cut and try to decluster again – Some jet grooming is built in
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 40
CMS top tagging algorithm
Define MinMass = minimum pairwise mass of subjets
Three Variables are used to tag tops:
1. Jet mass - top jets have mass within a top mass window
2. Number of subjets - top jets have 3 or 4 subjets
3. Minimum Pairwise Mass - Pairwise mass of two of the subjets should reconstruct the W mass
Generated top decay products: pairwise mass
Define Top Tag Cuts 140<mjet<250 Nsubjets ≥ 3 mmin>50
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 41
Top tagging variables
• A partially merged boosted top results in a W jet and a b jet
• W jets can be identified using jet and subjet properties:
- Jet mass = W mass
- Two subjets
- Subjet mass << jet mass
- Both subjets should have similar momentum
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 42
Example of boosted signal: W jets
13 June 2011
Simulation
Simulation
Conway, Erbacher, Dolen, Hu, Maksimovic, Rappoccio, Sierra-Vasquez
! Ellis et al. (arXiv:0903.5081) ! Improves mass resolution by
removing soft, large angle particles from the jet
! Recluster each jet, requiring that each recombination satisfy the following:
! For W tagging, require: Jet mass in 60-100 GeV/c2
Mass drop (mu) < 0.4
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 43
Example: W-tagging with pruning
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 44
Summary
The Jet substructure is coming up as a very popular tool to analyze jets that are produced from heavy objects
First systematic study of jet grooming techniques on jet mass performed at CMS•an important benchmark in understanding various algorithms for searches for new physics
With more data, probing deeper into boosted regime•e.g. boosted ttbar, VH, VV resonances (V = W or Z) •this tutorial should help you get started with boosted signals
This hands-on tutorial should get you started•Please find us during the breaks or email us in case you come up with a question later
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 46
Jet clustering in FastJet
Can be particle-flow particles or gen-particles or tracks or calo-towers
Defined above
Can be one of the algorithms described in the next slides
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 47
How to compute jet areas in FastJet?
Active areas for some of the most popular jet algorithms are shown on the next slide
RECOη
-2 0 2
GEN
/pT
RECO
,pT
σfit
ted
0
0.05
0.1
0.15
Ungroomed AK7Trimmed AK7Filtered AK7Pruned AK7
= 7 TeV, AK7 V+jetss at -1CMS Simulation, L = 5fb
[GeV]RECO
pT200 300 400 500
GEN
/pT
RECO
, pT
σfit
ted
0
0.05
0.1
0.15
Ungroomed AK7Trimmed AK7Filtered AK7Pruned AK7
= 7 TeV, AK7 V+jetss at -1CMS Simulation, L = 5fb
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 48
Jet pT resolution for groomed jets
Jet pT resolution for various grooming algorithms also shows good agreement to within a few percent.
•Groomed jet pT resolution degraded slightly.
resolution vs pTresolution vs η
Hands-on Advanced Tutorial Session on Jet Substructure / 44LPC, Fermilab 49
For visualization, present di-jet and V+jet distributions with one above the other.
pT bin:
300-450 GeV
N.B. not the exact same
observablesmJ vs <mJ> in bins of pT,
<pT>
Shows the difference in jet mass for di-jet and V+jet processes where the former (latter) radiates more (less) due to larger composition of
gluon (quark)-initiated jets
ungroomed pruned
Quark vs gluon differences