3.5 Jet Substructure Tagging
75
The sensitivity to partonic radiation and non-perturbative effects makes q/g tagging particularly sensitive to modelling choices in simulation. Typically, Pythia
tends to describe quark jets better than Herwig, whereas the opposite is observed
for gluons [436]. For this reason, Pythia tends to overestimate the q/g tagging
performance with respect to data, while Herwig underestimates it [466]. The modelling of observables for q/g tagging is an active field of research, see for example
[467] for a recent study. A reduction of uncertainties related to the modelling of
quark and gluon jets can be achieved with dedicated measurements, sensitive to the
radiation pattern in jets. A very promising example is the measurement of the Lund
jet plane [468, 469], which has very recently been performed on dijet events by
ATLAS [470]. Since distinct regions in the Lund jet plane are dominated by contributions from different QCD processes, the measurement will be useful for tuning
non-perturbative models and parameters of event generators (see Sect. 4.1.4).
Numerous jet substructure observables have been studied for their q/g discrimination power. Some of the most useful ones are particle multiplicity, thrust [44–46],
broadening [471, 472], girth [473], integrated jet shapes [60] or the fragmentation
function p
D
T [444, 474]. The most powerful single variable has been found to be particle multiplicity which is not IRC safe, but can be modified using an iterative version
of soft drop grooming. The resulting soft drop multiplicity is IRC safe and can be
calculated perturbatively. It shows similar discrimination power as particle multiplicity [475]. In general, counting observables are sensitive to multiple emissions at LL
accuracy, resulting in better discrimination power than IRC-safe observables which
are dominated by a single emission at LL accuracy, giving rise to Casimir scaling
of the q/g discrimination power [223]. While particle multiplicity performs best as
single variable tagger, a gain in the q/g tagging performance can be obtained by
combining it with a Casimir scaling observable. Only marginal gains are obtained
by including more observables [476, 477].
Particle multiplicities are the most important input to the q/g taggers developed by
the ATLAS [466, 478–480] and CMS [436, 481, 482] Collaborations. While ATLAS
uses the number of tracks as an approximation for the number of jet constituents and
the jet width [480], CMS utilises the number of particle-flow constituents, p
D
T and
the angular opening of the minor jet axis σ 2 , computed from the p
2
T -weighted average
of the angular distance between the jet constituents [436]. The performance of the
CMS q/g tagger in terms of quark-jet efficiency versus gluon-jet rejection is shown
in Fig. 3.6. The gain by including p
D
T and σ 2 is an improvement of about 2% in
gluon-jet rejection at a quark-jet efficiency of 60%. The performance of q/g tagging
degrades considerably in the region outside of the tracker acceptance, where only
the granularity of the calorimeters can be used to infer the particle multiplicity. In
addition, this region has a higher susceptibility to pileup radiation than the central
region.
Since the distributions of variables used for q/g tagging depend on the jet kinematics and the pileup activity in the event, likelihood discriminators are built differential
in jet η, p T and the value ρ of the event. In order to mitigate the effect of modelling
75
The sensitivity to partonic radiation and non-perturbative effects makes q/g tagging particularly sensitive to modelling choices in simulation. Typically, Pythia
tends to describe quark jets better than Herwig, whereas the opposite is observed
for gluons [436]. For this reason, Pythia tends to overestimate the q/g tagging
performance with respect to data, while Herwig underestimates it [466]. The modelling of observables for q/g tagging is an active field of research, see for example
[467] for a recent study. A reduction of uncertainties related to the modelling of
quark and gluon jets can be achieved with dedicated measurements, sensitive to the
radiation pattern in jets. A very promising example is the measurement of the Lund
jet plane [468, 469], which has very recently been performed on dijet events by
ATLAS [470]. Since distinct regions in the Lund jet plane are dominated by contributions from different QCD processes, the measurement will be useful for tuning
non-perturbative models and parameters of event generators (see Sect. 4.1.4).
Numerous jet substructure observables have been studied for their q/g discrimination power. Some of the most useful ones are particle multiplicity, thrust [44–46],
broadening [471, 472], girth [473], integrated jet shapes [60] or the fragmentation
function p
D
T [444, 474]. The most powerful single variable has been found to be particle multiplicity which is not IRC safe, but can be modified using an iterative version
of soft drop grooming. The resulting soft drop multiplicity is IRC safe and can be
calculated perturbatively. It shows similar discrimination power as particle multiplicity [475]. In general, counting observables are sensitive to multiple emissions at LL
accuracy, resulting in better discrimination power than IRC-safe observables which
are dominated by a single emission at LL accuracy, giving rise to Casimir scaling
of the q/g discrimination power [223]. While particle multiplicity performs best as
single variable tagger, a gain in the q/g tagging performance can be obtained by
combining it with a Casimir scaling observable. Only marginal gains are obtained
by including more observables [476, 477].
Particle multiplicities are the most important input to the q/g taggers developed by
the ATLAS [466, 478–480] and CMS [436, 481, 482] Collaborations. While ATLAS
uses the number of tracks as an approximation for the number of jet constituents and
the jet width [480], CMS utilises the number of particle-flow constituents, p
D
T and
the angular opening of the minor jet axis σ 2 , computed from the p
2
T -weighted average
of the angular distance between the jet constituents [436]. The performance of the
CMS q/g tagger in terms of quark-jet efficiency versus gluon-jet rejection is shown
in Fig. 3.6. The gain by including p
D
T and σ 2 is an improvement of about 2% in
gluon-jet rejection at a quark-jet efficiency of 60%. The performance of q/g tagging
degrades considerably in the region outside of the tracker acceptance, where only
the granularity of the calorimeters can be used to infer the particle multiplicity. In
addition, this region has a higher susceptibility to pileup radiation than the central
region.
Since the distributions of variables used for q/g tagging depend on the jet kinematics and the pileup activity in the event, likelihood discriminators are built differential
in jet η, p T and the value ρ of the event. In order to mitigate the effect of modelling