Jet Cleansing: Pileup Removal at High Luminosity

TL;DR

Pileup at the LHC obscures jet observables, especially jet mass, at high luminosity. The paper introduces jet cleansing, a subjet-level technique that combines tracking and calorimeter information to reconstruct the pileup-free leading-vertex momentum, with three implementations (linear, Gaussian, and JVF cleansing) that leverage a per-subjet constraint on charged-to-total momentum. Cleansing demonstrates superior correlation to truth and higher $S/\sqrt{B}$ relative to area subtraction, CHS, and jet-vertex methods, for both dijet invariant mass and jet mass, and remains effective with or without grooming. The approach scales favorably with pileup and has been validated in full detector simulations, offering a practical path to improved precision QCD comparisons and enhanced sensitivity to new physics in high-luminosity runs.

Abstract

One of the greatest impediments to extracting useful information from high luminosity hadron-collider data is radiation from secondary collisions (i.e. pileup) which can overlap with that of the primary interaction. In this paper we introduce a simple jet-substructure technique termed cleansing which can consistently correct for large amounts of pileup in an observable independent way. Cleansing works at the subjet level, combining tracker and calorimeter-based data to reconstruct the pileup-free primary interaction. The technique can be used on its own, with various degrees of sophistication, or in concert with jet grooming. We apply cleansing to both kinematic and jet shape reconstruction, finding in all cases a marked improvement over previous methods both in the correlation of the cleansed data with uncontaminated results and in measures like S/rt(B). Cleansing should improve the sensitivity of new-physics searches at high luminosity and could also aid in the comparison of precision QCD calculations to collider data.

Figures (6)

• Figure 1: Dijet mass distributions for various methods with 20 and 140 pileup vertices. Results shown are without grooming, groomed results can be seen in Table \ref{['table:resultsD']}.
• Figure 2: Jet mass distributions for various methods with 20 and 140 pileup vertices. Results shown are without grooming, groomed results can be seen in Table \ref{['table:resultsD']}.
• Figure 3: Correlations for dijet mass, a kinematic variable, are shown between between events with 140 pileup interactions, corrected via subtraction or cleansing, and the truth version of the same events, with pileup explicitly removed. The top row shows the uncorrected correlations, the middle row demonstrates the performance of Cacciari:2007fd, and the bottom row shows the performance of the linear cleansing method described here.
• Figure 4: Correlations for jet mass, a substructure variable, are shown between between events with 140 pileup interactions, corrected via subtraction or cleansing, and the truth version of the same events, with pileup explicitly removed. The top row shows the uncorrected correlations, the middle row demonstrates the performance of Soyez:2012hv, and the bottom row shows the performance of the linear cleansing method described here.
• Figure 5: Top: the distribution of $\gamma_0$, the charged to total $p_T$ ratio in pileup, for various average number of pileup interactions. Bottom: the correlation between the true value of $\gamma_1$, the charged to total $p_T$ ratio coming from the leading vertex, with its approximation using Eq. \ref{['eq:solveg1']}.