1. science
2. physics

CMS CR -2014/351
Available on CMS information server
The Compact Muon Solenoid Experiment
Conference Report
Mailing address: CMS CERN, CH-1211 GENEVA 23, Switzerland
27 October 2014 (v4, 06 November 2014)
Future perspectives for jet substructure
techniques in LHC Run2 (ATLAS+CMS)
Matthias Ulrich Mozer for the ”Standard Model, QCD, W,Z,DIFF,FW”, ATLAS and CMS collaborations.
Abstract
The increased pile-up expected in the LHC Run 2 and High Luminosity LHC creates a challenging
environment for utilizing the jet-substructure techniques which were successfully demonstrated in
the LHC Run 1. The ATLAS and CMS experiments are studying a range of methods to improve
jet reconstruction to increase the resilience against high pile-up. Promising results are obtained in
simulation but await validation on the first Run 2 data.
Presented at ISMD2014 XLIV International Symposium on Multiparticle Dynamics
EPJ Web of Conferences will be set by the publisher
DOI: will be set by the publisher
c Owned by the authors, published by EDP Sciences, 2014
Future perspectives for jet substructure techniques in LHC Run2
Matthias Mozer1 , a
1
Institut für Experimentelle Kernphysik, KIT
Wolfgang-Gaede-Str. 1, 76131 Karlsruhe, Germany
Abstract. The increased pile-up expected in the LHC Run 2 and High Luminosity LHC creates a challenging
environment for utilizing the jet-substructure techniques which were successfully demonstrated in the LHC Run
1. The ATLAS and CMS experiments are studying a range of methods to improve jet reconstruction to increase
the resilience against high pile-up. Promising results are obtained in simulation but await validation on the first
Run 2 data.
1 Introduction
In the search for new physics at the LHC, jets play a dominant role. They are of particular importance in studies
involving electroweak bosons, top-quarks and the Higgs
boson, as these particles decay into hadron with large
branching ratios. At the high energies probed at the LHC
the decay hadrons will often form a single jet, so called
boosted topologies. The distribution of particles within
such jets differs substantially from jets originating from
the hadronization of single quarks or gluons and techniques that exploit these differences have been used successfully to suppress QCD induced backgrounds. The jet
characteristics used in the substructure methods originate
from the initial parton composition of the jets and the QCD
evolution of the parton shower. However, in the presence
of many simultaneous interactions (“pile-up”), the signatures can be diluted by the presence of additional jet constituents (calorimeter clusters, tracks...) not originating
from the primary interaction. In the upcoming LHC Run
2, the number of simultaneous interactions is expected to
rise to ∼ 40, compared to ∼ 25 in Run 1, and even more for
the future High Luminosity LHC (“HL-LHC”) upgrade.
Under these harsh conditions, the performance of jet substructure techniques, but also basic reconstruction, such as
the measurement of the jet transverse momentum (pT ) are
negatively impacted. The LHC experiments have started to
investigate various methods to reduce the effects of pile-up
on jet measurements in order to retain or even improve the
good performance observed in Run 1.
The LHC experiments are approaching this issue from
three directions:
• improving the basic event reconstruction of the experiments.
• removing pile-up particles before the event is interpreted
as a specific final state.
a e-mail: Matthias.Mozer@kit.edu
• refining physics interpretation tools such as b-tagging
and jet-substructure techniques to be more resilient
against pile-up.
2 Improved Reconstruction
Before data are reconstructed for analysis, events have
to pass the trigger selection in order to be permanently
stored. The ATLAS [1] and CMS [2] experiments both
use a multi-stage trigger design, where the first step of the
selection is performed by hardware designed for this purpose (referred to as L1 trigger) and the last step is implemented as a somewhat reduced version of the full reconstruction run on a dedicated but otherwise generic cluster
of computers (called HLT). The ATLAS experiment implements an intermediate step, while the CMS experiment
uses just the two steps discussed above. The whole range
of jet-substructure techniques can be relatively easily implemented in the HLT at the only cost of some additional
processing time. However, for triggers that solely rely on
boosted jet identification, the L1 trigger can impose a serious bottle neck. To improve on this situation, the ATLAS collaboration is currently implementing an upgrade
to their L1 trigger system that allows for wider jets [3].
Figure 1 (left), shows how this upgraded trigger can increase the efficiency in identifying hadronic top quark decays.
The CMS collaboration has invested a significant effort
into improving the existing Particle Flow (PF) [6, 7] reconstruction to more accurately measure the properties of very
energetic jets. In previous versions of this reconstruction
techniques, the large multiplicity of tracks in a very energetic jet could overwhelm the tracking algorithm, leading
to a significant deficit in the fraction of charged particles
within a very energetic jet. This issue has been cured by
the addition of additional tracking steps, tuned specifically
to recover tracks aligned with the jet axis. Additionally,
1.0
0.8
ATLAS Preliminary
Simulation
0.6
tt s=14 TeV 〈µ 〉=80
0.4
anti- k T R=1.0 (5% trimmed)
≥1 k T D=0.3 subjet with p >20 GeV
T
jet
|η |<2.5; isolated by ∆ R>2
0.2
0.0
100
L1_J100 (Run 1 L1Calo sim.)
1 subjet
2 subjets
≥3 subjets
200
300
13 TeV
CMS
AK R=0.8
0.3
1.6 < p (W) < 2.4 TeV
T
|η|<2.4
Simulation Preliminary
merged PF neutrals
split PF photons
0.2
split PF photons+neutrals
0.1
L1_G140
1 subjet
2 subjets
≥3 subjets
400
uncalibrated p
Normalized Distribution
per-Jet Efficiency
EPJ Web of Conferences
500
jet
T
[GeV]
00
50
100
150
pruned jet mass (GeV)
Figure 1. Left: Simulated efficiency of new (L1_G140) and previous (L1_J100) L1 triggers of the ATLAS detector as function of jet pT
using boosted top quarks as case study. Large efficiency gains are observed in the case of boosted topologies with two or three subjets.
Right: Reconstructed jet mass for a hypothetical resonance with an invariant mass of 4 TeV decaying to two W bosons. The dotted line
shows the mass reconstructed with the pruning [4, 5] algorithm using the PF configuration as used in Run 1; the solid line adds higher
granularity treatment of deposits in the electromagnetic calorimeters and the points show the results when the hadronic calorimeters
are treated with finer granularity as well.
overlapping calorimeter deposits are treated with a finer
granularity, greatly improving the reconstruction of the jet
mass for very energetic jets. As an example, Figure 1
(right) from [8] shows the jet mass distribution for a hypothetical resonance with an invariant mass of 4 TeV decaying to two W bosons. Using the methods described above,
the reconstruction of the W mass improved substantially.
Additionally, algorithms that identify the hadronization of b-quarks (“b-tagging”) are being transferred from
their initial use in jets originating from a single b-quark
to the case of boosted hadronic decays with at least on bquark. Initial studies were driven by boosted top decays,
but with the discovery of the Higgs boson and its high
branching ratio to b-quark pairs, additional attention has
be focused on boosted hadronic Higgs decays.
The CMS collaboration has largely relied on reusing
the same algorithms tuned to the identification of jets
originating from single b-quarks also to identify b-quarks
within larger decay chains [9, 10]. While this approach
foregoes the opportunity for further optimization, it allows
for a very rapid use of b-tagging in boosted topologies and
focuses limited manpower resources. Even without specific optimizations, this approach leads to large gains in
the identification of boosted top quark (see Figure 2, left)
and Higgs boson decays. Conversely, the ATLAS collaboration is optimizing additional b-tagging algorithms
specifically for their performance within hadronic decays
of boosted heavy particles [11, 12]. The results are very
promising, as shown in Figure 2 (right), and excellent efficiencies and background rejection rates can be achieved.
3 Pile-Up Removal
Once reconstructed, particles from pile-up may be removed from further consideration by directly identifying them, or alternatively, by statistically subtracting their
contributions from other other variables, such as jet momenta or lepton isolation variables. Such corrections were
already necessary in the LHC Run 1 in order to meet expected resolutions and efficiencies. For Run 1, the predominant corrections were applied based on the jet- or
isolation-cone area and proportional to the pile-up activity measured in a given event [13, 14].
In the CMS experiment, the PF reconstruction additionally allowed individually reconstructed charged particles to be classified as pile-up according to the association
of their measured tracks to the vertices in the event. While
this method (called charged hadron subtraction, CHS) has
the advantage that the contributions from charged pileup particles is subtracted exactly, neutral pile-up contributions can only be subtracted statistically.
In a recent study [15], the CMS has investigated more
sophisticated methods to reduce the pile-up distribution.
The so called constituent subtraction method [16] works
similarly to the statistical jet correction described above,
but instead of defining a jet area, an effective area and
corresponding correction is assigned to each reconstructed
particle. Additionally the Pileup Per Particle Identification (PUPPI) algorithm [17] was studied, which uses interparticle correlations to assign a pile-up probability to each
reconstructed particle. The probabilities are one (zero) for
charged particles from the primary (pile-up) vertices, respectively, similar to the CHS method. However, for neutral particles the probability is evaluated from the correlations with surrounding particles, leading to weighting fac-
Mistag Rate
CMS Simulation, s = 8 TeV
HEP Top Tagger
HEP + τ3/τ2
HEP + sub. b-tag
HEP + τ3/τ2 + sub. b-tag
HEP WP0
HEP Comb. WP1
HEP Comb. WP2
HEP Comb. WP3
-2
10
-3
10
0.4
0.35
0.3
ATLAS Simulation Preliminary
s = 8 TeV, k/Mplank = 1.0
Track jets (R=0.2)
Track jets (R=0.3)
anti-k t track jets
kt calo subjets
Calo subjets (R=0.3)
Track jets (R=0.4)
0.25
0.2
0.15
0.1
10-4
Matched parton
p > 200 GeV/c
T
10-5
0
Efficiency
XLIV International Symposium on Multiparticle Dynamics
0.1
0.2
0.05
0
0.3
0.4
Top Tag Efficiency
1000
1500
2000
2500
Graviton Mass [GeV]
Figure 2. B-tagging performance in boosted objects.
Left: CMS top tagging performance with (brown lines) and without (purple lines) using subjet b-tags to select boosted top-quark
decays.
Right: Efficiency to select a resonance decaying to two hadronically decaying Higgs bosons using several different working points for
subjet b-tagging with the ATLAS detector.
tors between zero and one. Figure 3 compares the jet mass
spectra and resolutions for QCD jets and boosted hadronic
W decays using the different algorithm. The best performance is achieved with the PUPPI algorithm, closely followed by the constituent subtraction. CHS and statistical
corrections as used in Run 1 perform much worse.
In Run 1, the so called Jet Veto Fraction (JVF) was
widely used in the ATLAS collaboration to suppress jets
largely consisting of pile-up particles. The JVF is defined
as the sum of the pT of the tracks within a jets area coming
from the primary vertex, divided by the sum of the pT of
all tracks within the jets area. A requirement for the JVF to
exceed a given threshold was then employed to select jets
originating from the primary vertex. However, the fixed
cut requirement leads to a noticeable dependence of the
selection efficiency on the number of simultaneous collisions in the event, as shown in Figure 4, left. This issue
is solved by two improvements: For future use, the JVF
computation is rescaled to explicitly take into account the
dependence on the number of vertices in the event. Additionally the JVF variable is combined with a variable related to the fraction of charged particles within a jet to
further reduce the pile-up dependence of the variable and
increase overall tagging efficiency [18].
4 Jet-Substructure Techniques
In the LHC Run 1, a number of different jet-substructure
techniques have been employed. In the CMS collaboration, the so called pruning algorithm [19] has seen
widespread use in order to reconstruct the invariant mass
of hadronic W- and Z-boson. The algorithm is designed to
clean the jet from soft and wide-angle particles. Jets are
reclustered with a modified Cambrige-Aachen (CA) algorithm [20], where combinations only proceed if the to constituents pass criteria on their relative angle and transverse
momentum.
In addition the jet pruning, results of the ATLAS collaboration have been prepared with the trimming [21] and
mass-drop [23] algorithms (see Ref. [24] for corresponding performance studies). In the trimming algorithm a jet
with wide radius parameter is reclustered with a narrower
radius parameter. Of the resulting subjets, only the ones
carrying a certain fraction of the wide jets transverse momentum are retained. The mass drop algorithm is applied
to jets reconstructed with the CA algorithm. In each step,
the most recent combination of the CA algorithm is undone and it is checked whether the more massive of the
two resulting subjets carries more than a certain fraction of
the mass of the parent jet, in which case the lighter subjet
is discarded and the procedure continues on the remaining
subjet.
In preparation for the HL-LHC and the large amount
of pile-up expected, the ATLAS collaboration has revisited
the topic of jet substructure algorithms. The results show
(see Figure 5, left) that even in the extreme case of 200 simultaneous interactions, the mass of hadronically decaying top-quarks can be well reconstructed.
In preparation for Run 2, the CMS collaboration has
performed a comprehensive study of several jet cleaning
algorithms in combination with methods that reduce pileup at the reconstruction level [15]. Together with the
pruning and trimming algorithms, discussed above, the so
called soft drop algorithm [22] (an evolution of the mass
drop algorithm described above) was investigated in combination with the CHS and PUPPI techniques described
above. Figure 6 shows the mass resolution for the different combinations of algorithms. Similar to the ATLAS
results, the trimming algorithm emerges as particularly advantageous. In addition to good stability against pile-up,
the trimming algorithm has less pronounced tails in it’s
resolution as estimated by comparing a Gaussian fit of the
resolution to its RMS.
EPJ Web of Conferences
25000
CMS
13 TeV
Simulation Preliminary
Pythia QCD
Anti-kT (R=0.8)
<nPU> = 40
200 GeV < p < 600 GeV
T
|η| < 2.5
arbitrary units
arbitrary units
CMS
30000
GEN
PF+PUPPI
PF
10000
8000
PF+CHS
PF(Cleansing)
20000
13 TeV
Simulation Preliminary
Pythia RS Graviton → WW
Anti-kT (R=0.8)
<nPU> = 40
200 GeV < p < 600 GeV
T
|η| < 2.5
GEN
PF+PUPPI
PF
PF+CHS
PF(Cleansing)
PF+CHS(Const.Sub.)
PF+CHS(Const.Sub.)
6000
15000
4000
10000
2000
5000
20
arbitrary units
CMS
35000
30000
40
60
80
100
120
140
160
0
0
180
200
m (GeV)
PF+PUPPI
<∆m>=-0.6 GeV
RMS=10.9 GeV
PF
<∆m>=13.7 GeV
RMS=17.9 GeV
25000
4500
4000
3500
PF+CHS
<∆m>=-6.2 GeV
RMS=14.6 GeV
20000
60
80
100
120
140
160
180
200
m (GeV)
13 TeV
Simulation Preliminary
Pythia RS Graviton → WW
Anti-kT (R=0.8)
<nPU> = 40
200 GeV < p < 600 GeV
T
|η| < 2.5
PF+PUPPI
<∆m>=-0.5 GeV
RMS=10.0 GeV
PF
<∆m>=12.1 GeV
RMS=15.4 GeV
PF+CHS
<∆m>=-5.0 GeV
RMS=12.4 GeV
2500
PF(Cleansing)
<∆m>=-1.9 GeV
RMS=12.1 GeV
2000
PF+CHS(Const.Sub.)
<∆m>=0.5 GeV
RMS=13.7 GeV
10000
40
3000
PF(Cleansing)
<∆m>=-0.8 GeV
RMS=12.7 GeV
15000
20
CMS
13 TeV
Simulation Preliminary
Pythia QCD
Anti-kT (R=0.8)
<nPU> = 40
200 GeV < p < 600 GeV
T
|η| < 2.5
arbitrary units
0
0
PF+CHS(Const.Sub.)
<∆m>=0.6 GeV
RMS=11.4 GeV
1500
1000
5000
0
-100
500
-80
-60
-40
-20
0
20
40
0
-100
60
80
100
mreco - mgen(GeV)
-80
-60
-40
-20
0
20
40
60
80
100
mreco - mgen(GeV)
0.1
Fake Rate
Fake Rate
Figure 3. Mass distributions (top) and resolutions (bottom) for QCD jets (left) and boosted hadronic W decays (right) using several
different pile-up mitigation techniques.
ATLAS Simulation Preliminary
0.08
Pythia8 dijets
Anti-k t LCW+JES R=0.4
20 < pT < 30 GeV, |η| < 2.4
Target signal efficiency = 0.9
JVF>0.5
JVT>0.6
ATLAS Simulation Preliminary
10-1
0.06
Pythia8 dijets
Anti-k t LCW+JES R=0.4
|η| < 2.4
20 < p < 50 GeV
T
JVF = 0.5
JVF
corrJVF
RpT
JVT
JVF = 0.25
0.04
10-2
0.02
0
5
10
15
20
25
NVtx
0.8
0.85
0.9
0.95
Efficiency
Figure 4. Left: Comparison of the pile-up-dependence of the JVF used in Run 1 to the improved combined variable (JVT).
Right: Fake rate as function of tagging efficiency for the JVF used in Run 1 compared to the scaled JVF (corrJVF), the additional
charge-fraction variable (R pT ) and the combination (JVT).
5 Summary and Outlook
The LHC experiments are demonstrating with simulation
studies, that the previously used jet reconstruction and
substructure techniques are affected by the increased pileup expected in Run 2 and the HL-LHC. However, improved reconstruction and substructure methods have been
devised to minimize the impact of the additional pile-up.
Nevertheless, these methods will have to be validated and
possibly revised when high pile-up data-taking starts with
the LHC Run 2 in 2015.
References
[1] ATLAS Collaboration, JINST 3 S08003 (2008)
[2] CMS Collaboration, JINST 3 S08004 (2008)
XLIV International Symposium on Multiparticle Dynamics
0.2
0.18
0.16
0.22
ATLAS Simulation Preliminary
anti-k t LCW jets with R=1.0, 0.0< |η| < 1.2
<µ >=0, σ
No jet grooming, no jet pileup correction
(µ =30)
noise
pileup
<µ >=40, σ
s = 14 TeV, 25 ns bunch spacing
<µ >=80, σ
jet
500 < pT < 750 GeV
(µ =40)
noise
pileup
Z’
(µ =80)
noise
pileup
<µ >=140, σ
Pythia8 Z’ → tt (m =2 TeV)
0.14
pileup
<µ >=200, σ
(µ =140)
noise
pileup
noise
(µ =200)
0.12
Normalized Entries
Normalized Entries
0.22
0.2
0.18
0.16
<µ >=0, σ
Trimmed, jet 4-vector pileup correction
s = 14 TeV, 25 ns bunch spacing
(µ =30)
noise
pileup
<µ >=80, σ
jet
500 < pT < 750 GeV
(µ =40)
noise
pileup
(µ =80)
noise
pileup
<µ >=140, σ
Z’
0.14
pileup
<µ >=40, σ
Pythia8 Z’ → tt (m =2 TeV)
<µ >=200, σ
(µ =140)
noise
pileup
noise
(µ =200)
0.12
0.1
0.1
0.08
0.08
0.06
0.06
0.04
0.04
0.02
0.02
0
ATLAS Simulation Preliminary
anti-k t LCW jets with R=1.0, 0.0< |η| < 1.2
-100
0
100
200
300
400 500
mjet [GeV]
0
-100
0
100
200
300
400 500
mjet [GeV]
Resolution (GeV)
Figure 5. Reconstructed mass for boosted top quarks in the ATLAS detector with different amounts of pile-up without (left) and with
(right) jet-grooming techniques and pile-up corrections applied.
CMS
50
13 TeV
Simulation Preliminary
RS Graviton → WW, Anti-kT (R=0.8)
PF
<nPU>=40
PF + CHS
p > 300 GeV
T
fitted σ
RMS
PF + PUPPI
|η| < 2.5
40
Trimming
Pruning
Soft Drop
30
20
10
0
m
mra
w
rs =
r
r
r
z
z
z
z
β=
β=0
ub 0
.2 psub =0.1 psub =0.2 psub =0.3 pcut =0.1 Rcut =0.05 cut =0.05 cut =0.1 R 1
T
T
T
T
R
R
=
cut =0
cut =0 cut =0.7
frac =0.0 frac =0.0 frac =0.0 frac =0.0cut 0.5
.5
.75
5
5
3
3
3
β=2
Figure 6. Comparison of the jet mass resolution of a variety of jet grooming algorithms in combination with pile-up reduction techniques. The trimming, pruning and soft drop algorithms are combined with CHS and PUPPI pile-up reduction techniques. Shown are
mass resolutions from a Gaussian fit as well as the RMS values of the distributions, indicating the size of non-Gaussian tails in the
resolution.
[3] ATLAS Collaboration, Global Feature Extraction
(gFEX) Performance Plots, ATLAS-COM-DAQ-2014087 (2014), https://cds.cern.ch/record/1749167
[4] Ellis, S. D. et. al., Phys. Rev. D 80, 094023 (2010),
[5] Ellis, S. D. et. al., Phys. Rev. D 81, 094023 (2010)
[6] CMS Collaboration, Particle–Flow Event Reconstruction in CMS and Performance for Jets, Taus, and
ETmiss , CMS-PAS-PFT-09-001 (2009), https://cds.cern.
ch/record/1194487
[7] CMS Collaboration, Commissioning of the Particleflow Event Reconstruction with the first LHC collisions
recorded in the CMS detector, CMS-PAS-PFT-10-001
(2010), https://cds.cern.ch/record/1247373
[8] CMS Collaboration, V Tagging Observables and Correlations, CMS-PAS-JME-14-002 (2014), https://cds.
cern.ch/record/1754913
[9] CMS Collaboration, Boosted Top Jet Tagging at
CMS, CMS-PAS-JME-13-007 (2014), https://cds.cern.
ch/record/1647419
[10] CMS
Collaboration, Performance of b tagging
√
at
s=8 TeV in multijet, tt and boosted topology
events, CMS-PAS-BTV-13-001 (2013), https://cds.
cern.ch/record/1581306
[11] ATLAS Collaboration, Flavor Tagging with Track
Jets in Boosted Topologies with the ATLAS Detector, ATL-PHYS-PUB-2014-013 (2014), https://cds.
cern.ch/record/1750681
EPJ Web of Conferences
[12] ATLAS Collaboration, b-tagging in dense environments, ATL-PHYS-PUB-2014-014 (2014), https://cds.
cern.ch/record/1750682
[13] M. Cacciari, G. P. Salam and G. Soyez, JHEP 0804,
005 (2008)
[14] M. Cacciari and G. P. Salam, Phys. Lett. B 659, 119
(2008)
[15] CMS Collaboration, Study of Pileup Removal Algorithms for Jets, CMS-PAS-JME-2014-001 (2014),
https://cds.cern.ch/record/1751454
[16] P. Berta, M. Spousta, D. W. Miller and R. Leitner,
JHEP 1406 092 (2014)
[17] D. Bertolini, P. Harris, M. Low and N. Tran, JHEP
1410 59 (2014)
[18] ATLAS Collaboration, Search for high-mass states
with one lepton plus missing transverse momentum in
√
pp collisions at s = 8 TeV with the ATLAS detector, ATLAS-CONF-2014-018 (2014), https://cds.cern.
ch/record/1700870
[19] CMS Collaboration, arXiv:1410.4227 [hep-ex].
[20] M. Wobisch and T. Wengler, In *Hamburg
1998/1999, Monte Carlo generators for HERA
physics* 270-279 [hep-ph/9907280].
[21] D. Krohn, J. Thaler and L. T. Wang, JHEP 1002
(2010) 084
[22] A. J. Larkoski, S. Marzani, G. Soyez and J. Thaler,
JHEP 1405 (2014) 146
[23] J. M. Butterworth, A. R. Davison, M. Rubin and
G. P. Salam, Phys. Rev. Lett. 100 (2008) 242001
[24] ATLAS Collaboration, JHEP 1309 (2013) 076