3.5 Jet Substructure Tagging
81
secondary vertex or charged particle tracks originating from the B hadron decay. For
13 TeV analyses, CMS uses the CSVv2 algorithm [504] and ATLAS uses the MV2c10
algorithm [505, 506]. Typically, efficiencies of around 70% with misidentification
rates of 0.3–1% for light quark and gluon jets and 10–20% for charm jets are achieved
with these algorithms [507, 508].
The challenge of identifying boosted H → bb decays lies in the busy hadronic
environment, where secondary vertices are difficult to identify and properties of the
B decay vertices calculated relative to the jet axis have less discrimination power
than for isolated b production. To meet this challenge, various algorithms have been
developed. For large-R jets, algorithms in use by ATLAS and CMS are subjet b
tagging and double-b tagging. Once the b content of the large-R jet has been identified, adding other substructure variables to the H → bb tagger can give additional
improvements. An advantage of the a tagger using only b tagging is that other jet
substructure observables like the jet mass are unaffected by an H → bb selection
and can be used as discriminating variables in H analyses or searches.
Subjet b tagging [509–514] is a technique where the standard b tagging algorithm
is applied to each of the subjets of the large-R jet. In ATLAS, the subjets are track
jets with a radius of 0.2, matched to the large-R jet using the ghost-association
technique [433]. In CMS, either subjets from the N -subjettiness minimisation or
from the soft drop algorithm are used. At p T = 800 GeV and higher, the subjets start
to overlap, causing the subjet b tagging techniques to break down due to doublecounting of tracks and secondary vertices. In this regime, b tagging on the large-R jet
gives better performance when requiring the presence of two B hadrons. The ATLAS
Collaboration has also performed studies to improve the efficiency at high p T , using
variable-radius track jets, exclusive k T subjets and calorimeter subjets reconstructed
in the centre-of-mass frame of the Higgs jet candidate [515]. For highly boosted
Higgs bosons, these reconstruction techniques promise to improve the performance
compared to the usage of fixed-radius track jets.
Double-b tagging in ATLAS means that the two leading p T track-jets must pass
the same b tagging requirement [514]. In CMS, the double-b tagger [504, 516] is
a dedicated algorithm combining several discriminating variables using a boosted
decision tree. The double-b tagger has been designed for jets with a pruned mass
of 50 < m jet < 200 GeV and p T > 300 GeV. The algorithm exploits not only the
presence of two B hadrons inside the large-R jet, but also the correlation between
the directions of the momenta of the two B hadrons and the energy flows of the
two subjets. It employs the same variables as used in the CSVv2 algorithm. It is
constructed such that its performance is largely uncorrelated to the jet mass and
p T . A recent improvement of double-b tagging in ATLAS uses VR track-jets [515]
for subjet b tagging. The reason for this development is a decreasing H → bb
identification efficiency with increasing large-R jet p T . For larger H boson boosts,
track-jets with R = 0.2 start to merge, making it difficult to identify two B hadron
decays within the H jet. The shrinking size of VR track-jets with increasing p T helps
to reduce the inefficiency at high p T . These are obtained by clustering tracks with the
VR algorithm with the parameters ρ = 30 GeV, R min = 0.02 and R max = 0.4. The
VR track-jets are ghost-associated to large-R jets for double-b tagging. For large-R
81
secondary vertex or charged particle tracks originating from the B hadron decay. For
13 TeV analyses, CMS uses the CSVv2 algorithm [504] and ATLAS uses the MV2c10
algorithm [505, 506]. Typically, efficiencies of around 70% with misidentification
rates of 0.3–1% for light quark and gluon jets and 10–20% for charm jets are achieved
with these algorithms [507, 508].
The challenge of identifying boosted H → bb decays lies in the busy hadronic
environment, where secondary vertices are difficult to identify and properties of the
B decay vertices calculated relative to the jet axis have less discrimination power
than for isolated b production. To meet this challenge, various algorithms have been
developed. For large-R jets, algorithms in use by ATLAS and CMS are subjet b
tagging and double-b tagging. Once the b content of the large-R jet has been identified, adding other substructure variables to the H → bb tagger can give additional
improvements. An advantage of the a tagger using only b tagging is that other jet
substructure observables like the jet mass are unaffected by an H → bb selection
and can be used as discriminating variables in H analyses or searches.
Subjet b tagging [509–514] is a technique where the standard b tagging algorithm
is applied to each of the subjets of the large-R jet. In ATLAS, the subjets are track
jets with a radius of 0.2, matched to the large-R jet using the ghost-association
technique [433]. In CMS, either subjets from the N -subjettiness minimisation or
from the soft drop algorithm are used. At p T = 800 GeV and higher, the subjets start
to overlap, causing the subjet b tagging techniques to break down due to doublecounting of tracks and secondary vertices. In this regime, b tagging on the large-R jet
gives better performance when requiring the presence of two B hadrons. The ATLAS
Collaboration has also performed studies to improve the efficiency at high p T , using
variable-radius track jets, exclusive k T subjets and calorimeter subjets reconstructed
in the centre-of-mass frame of the Higgs jet candidate [515]. For highly boosted
Higgs bosons, these reconstruction techniques promise to improve the performance
compared to the usage of fixed-radius track jets.
Double-b tagging in ATLAS means that the two leading p T track-jets must pass
the same b tagging requirement [514]. In CMS, the double-b tagger [504, 516] is
a dedicated algorithm combining several discriminating variables using a boosted
decision tree. The double-b tagger has been designed for jets with a pruned mass
of 50 < m jet < 200 GeV and p T > 300 GeV. The algorithm exploits not only the
presence of two B hadrons inside the large-R jet, but also the correlation between
the directions of the momenta of the two B hadrons and the energy flows of the
two subjets. It employs the same variables as used in the CSVv2 algorithm. It is
constructed such that its performance is largely uncorrelated to the jet mass and
p T . A recent improvement of double-b tagging in ATLAS uses VR track-jets [515]
for subjet b tagging. The reason for this development is a decreasing H → bb
identification efficiency with increasing large-R jet p T . For larger H boson boosts,
track-jets with R = 0.2 start to merge, making it difficult to identify two B hadron
decays within the H jet. The shrinking size of VR track-jets with increasing p T helps
to reduce the inefficiency at high p T . These are obtained by clustering tracks with the
VR algorithm with the parameters ρ = 30 GeV, R min = 0.02 and R max = 0.4. The
VR track-jets are ghost-associated to large-R jets for double-b tagging. For large-R