3.5 Jet Substructure Tagging
77
for low jet p T and higher for very high p T jets (see Sect. 2.4.1). Good separation
power between W and Z bosons is desirable in a number of analyses, most notably
searches for diboson resonances (see Sect. 5.1).
In experimental studies, simulated samples containing W jets rather than Z jets are
primarily used, as W jets are abundant in data and can be selected efficiently thanks
to the large quantity of tt events produced at the LHC. ATLAS and CMS performed
a broad range of studies, systematically identifying the influence of pileup mitigation
and grooming techniques on jet substructure observables used for V tagging [411,
435, 441, 452, 483, 484]. The discrimination power for a number of jet substructure
variables has been studied, including N -subjettiness [209, 210], Qjet volatility [255],
ratios of energy correlation functions C
(β)
2 [215], D
(β)
2 [218, 219] and N
(β)
2
[221],
angularities and planar flow [485], splitting scales [38, 486], the jet and subjet
quark/gluon likelihood, and the jet pull angle [226]. In addition to a good background
rejection at a given signal efficiency, the design of a V tagging algorithm also has to
take into account the accurate detector response to jet substructure observables.
Both ATLAS and CMS developed simple taggers that rely on the combination of
the jet mass with one other variable that improves the discriminating power between
the signal and background. The standard ATLAS V tagger for analyses of 13 TeV
data was chosen to be the trimmed jet mass and D
(β=1)
2
[483], while CMS decided
to use the soft drop jet mass and the N -subjettiness ratio τ 21 = τ 2 /τ 1 . Despite the
different choices of tagging observables and detector design, ATLAS and CMS reach
a very similar background rejection at a given tagging efficiency. An active field of
developments is the usage of multivariate techniques for boosted V identification
which have shown to be able to significantly improve the background rejection [435,
535].
In the ATLAS studies the variable C
(β=1)
2
in combination with the trimmed jet
mass has been shown to be as good a discriminator as τ 21
8 as shown in Fig. 3.7. This
is in contradiction to the study by CMS, where C 2 is one of the weaker observables;
however, a direct comparison is difficult, since in ATLAS groomed substructure
variables are used, calculated for trimmed jets, while in CMS ungroomed variables
are used. Also, the particulars of particle reconstruction have a large impact on
the performance of individual observables. While a study of the performance of
D 2 at CMS is still pending, the soft drop N 2 observable was found to give similar
performance to τ 21 in CMS [489]. The ATLAS measurements of signal efficiencies
versus background rejection power in Fig. 3.7 lie on the predicted performance
curves, giving confidence in performance studies using simulated events.
The CMS Collaboration studied the q/g likelihood discriminator for its potential
in V tagging applications [411], finding that a combination of the groomed jet mass
and the q/g likelihood achieved a similar discrimination power as the groomed jet
mass and τ 21 . When adding the q/g likelihood to a V tagger utilizing pruned jet
mass and τ 21 , the misidentification rate was reduced slightly from 2.6 to 2.3% at a
8 A different axis definition for the subjet axes is used in ATLAS when calculating τ N , known
as the-winner-takes-all axis [488], which is consistently found to perform slightly better than the
standard subjet axis definition in tagging bosons.
77
for low jet p T and higher for very high p T jets (see Sect. 2.4.1). Good separation
power between W and Z bosons is desirable in a number of analyses, most notably
searches for diboson resonances (see Sect. 5.1).
In experimental studies, simulated samples containing W jets rather than Z jets are
primarily used, as W jets are abundant in data and can be selected efficiently thanks
to the large quantity of tt events produced at the LHC. ATLAS and CMS performed
a broad range of studies, systematically identifying the influence of pileup mitigation
and grooming techniques on jet substructure observables used for V tagging [411,
435, 441, 452, 483, 484]. The discrimination power for a number of jet substructure
variables has been studied, including N -subjettiness [209, 210], Qjet volatility [255],
ratios of energy correlation functions C
(β)
2 [215], D
(β)
2 [218, 219] and N
(β)
2
[221],
angularities and planar flow [485], splitting scales [38, 486], the jet and subjet
quark/gluon likelihood, and the jet pull angle [226]. In addition to a good background
rejection at a given signal efficiency, the design of a V tagging algorithm also has to
take into account the accurate detector response to jet substructure observables.
Both ATLAS and CMS developed simple taggers that rely on the combination of
the jet mass with one other variable that improves the discriminating power between
the signal and background. The standard ATLAS V tagger for analyses of 13 TeV
data was chosen to be the trimmed jet mass and D
(β=1)
2
[483], while CMS decided
to use the soft drop jet mass and the N -subjettiness ratio τ 21 = τ 2 /τ 1 . Despite the
different choices of tagging observables and detector design, ATLAS and CMS reach
a very similar background rejection at a given tagging efficiency. An active field of
developments is the usage of multivariate techniques for boosted V identification
which have shown to be able to significantly improve the background rejection [435,
535].
In the ATLAS studies the variable C
(β=1)
2
in combination with the trimmed jet
mass has been shown to be as good a discriminator as τ 21
8 as shown in Fig. 3.7. This
is in contradiction to the study by CMS, where C 2 is one of the weaker observables;
however, a direct comparison is difficult, since in ATLAS groomed substructure
variables are used, calculated for trimmed jets, while in CMS ungroomed variables
are used. Also, the particulars of particle reconstruction have a large impact on
the performance of individual observables. While a study of the performance of
D 2 at CMS is still pending, the soft drop N 2 observable was found to give similar
performance to τ 21 in CMS [489]. The ATLAS measurements of signal efficiencies
versus background rejection power in Fig. 3.7 lie on the predicted performance
curves, giving confidence in performance studies using simulated events.
The CMS Collaboration studied the q/g likelihood discriminator for its potential
in V tagging applications [411], finding that a combination of the groomed jet mass
and the q/g likelihood achieved a similar discrimination power as the groomed jet
mass and τ 21 . When adding the q/g likelihood to a V tagger utilizing pruned jet
mass and τ 21 , the misidentification rate was reduced slightly from 2.6 to 2.3% at a
8 A different axis definition for the subjet axes is used in ATLAS when calculating τ N , known
as the-winner-takes-all axis [488], which is consistently found to perform slightly better than the
standard subjet axis definition in tagging bosons.