3.3 Pileup Mitigation
71
is usually combined with the area subtraction. First, the the pileup contribution from
charged particles is removed. In a second step, the remaining contributions from
neutral particles are removed with the area subtraction method.
While the above methods have been successfully deployed in ATLAS and CMS,
they each have some deficiencies for the reconstruction of jet substructure observables. Ideally, one would hope to remove pileup at the most granular level possible,
i.e. at the level of reconstructed particles or calorimeter clusters, in order to be as
generic as possible. For example, while area subtraction is very effective for correcting the jet p T , it is not capable to mitigate the pileup dependence of jet substructure
observables as it only removes pileup contributions on average. In fact, jet substructure variables are among the most difficult observables to correct for pileup because
they are so reliant on radiation profiles. A number of hybrid methods have been proposed operating at the event constituent level, such as jet cleansing [446], jets without
jets [447], SoftKiller [448], Constituent Subtraction [448–450] and PUMML [451].
An example of a method extensively used in CMS is the pileup per particle identification (PUPPI) algorithm [450]. This algorithm uses information related to charged
particle tracks, local particle densities and event pileup properties to determine if a
particle originates from pileup. A discriminator value α i is calculated event-by-event
for each particle i using p T,i and its angular distance to nearby particles. To translate
this value into a probability for a particle to originate from the LV, charged particles
assigned to pileup vertices are used to calculate the expected distribution in α per
event. The value of α i of each neutral particle is compared to the expected value
for pileup particles, assuming that charged and neutral particles result in the same
distribution in α. The difference between α i and the expected mean value is used to
calculate a weight for the particle four-vector, with large weights for parton showerlike radiation and small weights for pileup-like radiation. Almost all pileup particles
have values within a few standard deviations of the median of the α distribution and
are assigned small weights. Values that deviate significantly from the mean value
are indicative of a hard scatter, and these particles are assigned large weights. This
weighting method allows for experimental information, such as tracking, vertexing
and timing information, to be included.
3.3.2 Performance Studies
Pileup removal algorithms are commissioned for use in ATLAS and CMS via detailed
studies of jet observables in terms of the resolution, absolute scale and pileup dependence. Other observables include lepton identification efficiencies, missing transverse momentum resolution and the performance of jet substructure taggers. As an
example, pileup mitigation techniques can improve the lepton identification performance, where the energy in a small cone around the lepton is summed, to reduce the
susceptibility to pileup.
For observables like jet p T , dependencies on the number of reconstructed vertices
are observed with area subtraction methods for the pileup levels currently observed
71
is usually combined with the area subtraction. First, the the pileup contribution from
charged particles is removed. In a second step, the remaining contributions from
neutral particles are removed with the area subtraction method.
While the above methods have been successfully deployed in ATLAS and CMS,
they each have some deficiencies for the reconstruction of jet substructure observables. Ideally, one would hope to remove pileup at the most granular level possible,
i.e. at the level of reconstructed particles or calorimeter clusters, in order to be as
generic as possible. For example, while area subtraction is very effective for correcting the jet p T , it is not capable to mitigate the pileup dependence of jet substructure
observables as it only removes pileup contributions on average. In fact, jet substructure variables are among the most difficult observables to correct for pileup because
they are so reliant on radiation profiles. A number of hybrid methods have been proposed operating at the event constituent level, such as jet cleansing [446], jets without
jets [447], SoftKiller [448], Constituent Subtraction [448–450] and PUMML [451].
An example of a method extensively used in CMS is the pileup per particle identification (PUPPI) algorithm [450]. This algorithm uses information related to charged
particle tracks, local particle densities and event pileup properties to determine if a
particle originates from pileup. A discriminator value α i is calculated event-by-event
for each particle i using p T,i and its angular distance to nearby particles. To translate
this value into a probability for a particle to originate from the LV, charged particles
assigned to pileup vertices are used to calculate the expected distribution in α per
event. The value of α i of each neutral particle is compared to the expected value
for pileup particles, assuming that charged and neutral particles result in the same
distribution in α. The difference between α i and the expected mean value is used to
calculate a weight for the particle four-vector, with large weights for parton showerlike radiation and small weights for pileup-like radiation. Almost all pileup particles
have values within a few standard deviations of the median of the α distribution and
are assigned small weights. Values that deviate significantly from the mean value
are indicative of a hard scatter, and these particles are assigned large weights. This
weighting method allows for experimental information, such as tracking, vertexing
and timing information, to be included.
3.3.2 Performance Studies
Pileup removal algorithms are commissioned for use in ATLAS and CMS via detailed
studies of jet observables in terms of the resolution, absolute scale and pileup dependence. Other observables include lepton identification efficiencies, missing transverse momentum resolution and the performance of jet substructure taggers. As an
example, pileup mitigation techniques can improve the lepton identification performance, where the energy in a small cone around the lepton is summed, to reduce the
susceptibility to pileup.
For observables like jet p T , dependencies on the number of reconstructed vertices
are observed with area subtraction methods for the pileup levels currently observed