3.5 Jet Substructure Tagging
83
The commissioning of H → bb tagging algorithms is challenging since efficiency
measurements are not possible on jets from boosted H → bb and Z → bb production
due to the smallness of the production cross sections. Instead, jets originating from
g → bb splittings are being used. ATLAS and CMS select jets with one or two
muons associated associated to them, where the muons originate from semi-leptonic
B hadron decays. While jets associated with two muons are more similar to the
topology of jets from H → bb decays, a selection of jets with one muon results in a
larger dataset for the performance measurement. It should be noted that the jet mass
depends on the number of muons and a large fraction of the signal will not contain
two muons.
A measurement of the double-b tagging efficiency in ATLAS is shown in Fig. 3.9
(right). One of the two ghost-associated track subjets is required to contain a muon
track. About 6% of the jets originate from g → bb splittings and another 6% from
g → cc before applying the ATLAS double-b tagger. The flavour fractions in simulation have been adjusted to the ones in data by a template fit to the distribution of
the mean signed impact parameter significance, calculated from the three leading- p T
tracks. The flavour fraction corrections are found to vary between 0.7 and 1.7 depending on the jet p T , with a statistical uncertainty below 10%. The rate of double-b tagged
jets is measured to be 3.5% at p T = 500 GeV and 1.6% at p T = 1000 GeV, consistent with the expectation of a 70% efficiency for H → bb tagging. The decrease in
efficiency with increasing p T is a result of the increasing collimation and eventual
merging of the two track jets, as well as the degradation of the tracking performance
inside high p T jets. The uncertainty in the measurement is dominated by uncertainties
in b tagging variables and modelling of the g → bb signal. The background rejection
of the double-b tagger for jets from boosted top quarks is about an order of magnitude
smaller than the one for light quark and gluon jets [514, 516], caused by the presence
of a b quark from the t → bW decay. This misidentification rate has been measured
in a sample enriched in high- p T tt production [504]. The misidentification rate in
data and simulation agree within the uncertainties, which are of the order of 10%,
dominated by the statistical precision of the data.
In contrast to H → bb, the decay H → W W
∗
→ 4q has not been studied experimentally so far, except for a search for V H resonances in all-hadronic final states by
CMS [518]. In this analysis, H → W W
∗
→ 4q jets are identified using a selection
in the pruned jet mass and the N -subjettiness ratio τ 42 = τ 4 /τ 2 . Since the distribution
in τ 42 is p T dependent, the signal efficiencies have been observed to decrease from
approximately 30 to 20% with increasing p T from 500 to 1000 GeV, for a selection
of τ 42 < 0.55. A direct measurement of the signal efficiency proves difficult because
of the inability to enrich a sample with jets from four prong decays without applying
a jet substructure selection. Instead, the efficiency correction of a W tagger based
on τ 21 measured in a tt sample has been extrapolated using the difference between
Pythia and Herwig in modelling the H → W W
∗
→ 4q decay as systematic uncertainty. Further studies are clearly needed for the applicability of this decay channel
in future analyses, where advanced machine learning concepts offer the opportunity
83
The commissioning of H → bb tagging algorithms is challenging since efficiency
measurements are not possible on jets from boosted H → bb and Z → bb production
due to the smallness of the production cross sections. Instead, jets originating from
g → bb splittings are being used. ATLAS and CMS select jets with one or two
muons associated associated to them, where the muons originate from semi-leptonic
B hadron decays. While jets associated with two muons are more similar to the
topology of jets from H → bb decays, a selection of jets with one muon results in a
larger dataset for the performance measurement. It should be noted that the jet mass
depends on the number of muons and a large fraction of the signal will not contain
two muons.
A measurement of the double-b tagging efficiency in ATLAS is shown in Fig. 3.9
(right). One of the two ghost-associated track subjets is required to contain a muon
track. About 6% of the jets originate from g → bb splittings and another 6% from
g → cc before applying the ATLAS double-b tagger. The flavour fractions in simulation have been adjusted to the ones in data by a template fit to the distribution of
the mean signed impact parameter significance, calculated from the three leading- p T
tracks. The flavour fraction corrections are found to vary between 0.7 and 1.7 depending on the jet p T , with a statistical uncertainty below 10%. The rate of double-b tagged
jets is measured to be 3.5% at p T = 500 GeV and 1.6% at p T = 1000 GeV, consistent with the expectation of a 70% efficiency for H → bb tagging. The decrease in
efficiency with increasing p T is a result of the increasing collimation and eventual
merging of the two track jets, as well as the degradation of the tracking performance
inside high p T jets. The uncertainty in the measurement is dominated by uncertainties
in b tagging variables and modelling of the g → bb signal. The background rejection
of the double-b tagger for jets from boosted top quarks is about an order of magnitude
smaller than the one for light quark and gluon jets [514, 516], caused by the presence
of a b quark from the t → bW decay. This misidentification rate has been measured
in a sample enriched in high- p T tt production [504]. The misidentification rate in
data and simulation agree within the uncertainties, which are of the order of 10%,
dominated by the statistical precision of the data.
In contrast to H → bb, the decay H → W W
∗
→ 4q has not been studied experimentally so far, except for a search for V H resonances in all-hadronic final states by
CMS [518]. In this analysis, H → W W
∗
→ 4q jets are identified using a selection
in the pruned jet mass and the N -subjettiness ratio τ 42 = τ 4 /τ 2 . Since the distribution
in τ 42 is p T dependent, the signal efficiencies have been observed to decrease from
approximately 30 to 20% with increasing p T from 500 to 1000 GeV, for a selection
of τ 42 < 0.55. A direct measurement of the signal efficiency proves difficult because
of the inability to enrich a sample with jets from four prong decays without applying
a jet substructure selection. Instead, the efficiency correction of a W tagger based
on τ 21 measured in a tt sample has been extrapolated using the difference between
Pythia and Herwig in modelling the H → W W
∗
→ 4q decay as systematic uncertainty. Further studies are clearly needed for the applicability of this decay channel
in future analyses, where advanced machine learning concepts offer the opportunity