Source-linked AI summary
Quark-Gluon Tagging: Machine Learning vs Detector
Gregor Kasieczka, Nicholas Kiefer, Tilman Plehn, Jennifer M. Thompson
TL;DR
Separating quark from gluon jets using low-level detector information is challenging, and the paper tests whether LoLa tagging helps in two LHC search benchmarks. Detector effects degrade idealized tagging, while the benchmark studies report potential for mono-jet searches and improved reach for relatively light di-jet resonances.
Problem
Quark–gluon tagging from low-level detector output is a challenging LHC task, and its value in realistic search analyses needs evaluation.
Source: §1 · p. 2, §2.2 · p. 7Method
The paper extends the LoLa tagger to quark–gluon discrimination using particle-flow jet constituents and studies mono-jet and di-jet resonance benchmarks.
Source: §1 · p. 3, §2.2 · p. 7, §3 · p. 12Results
Detector effects degrade idealized LoLa results until a classic BDT becomes competitive; mono-jet tagging may improve reach, while di-jet tagging can help for hadronic Z′ bosons below the TeV range.
Source: §5 · p. 18Takeaways & Limitations
Quark–gluon tagging has applications with both gluon-rich and quark-rich signals, but its effect on LHC reach merits evaluation in properly described experimental setups.
Source: §5 · p. 18Neither mono-jet nor di-jet resonance searches is an obvious benchmark for estimating quark–gluon tagging’s impact on LHC reach, and mono-jet analysis yields no simple significance improvement.
Source: §3 · p. 15, §5 · p. 18
Abstract
from arXiv · showhide
Distinguishing quarks from gluons based on low-level detector output is one of the most challenging applications of multi-variate and machine learning techniques at the LHC. We first show the performance of our 4-vector-based LoLa tagger without and after considering detector effects. We then discuss two benchmark applications, mono-jet searches with a gluon-rich signal and di-jet resonances with a quark-rich signal. In both cases an immediate benefit compared to the standard event-level analysis exists.
1 Introduction
Jet substructure analysis is shifting from hand-designed constituent observables toward machine learning on low-level detector outputs. The paper motivates LoLa tagging and two LHC benchmark channels while emphasizing theoretical, detector, and analysis challenges.
1 Introduction
- Modern machine learning can analyze low-level detector outputs such as measured 4-vectors instead of relying only on high-level constituent observables.The paper describes this shift as a change in how jet substructure is analyzed.Source: §1 · p. 2
- Quark and gluon jets pose theoretical and simulation challenges because they are not defined in QCD beyond tree level.The paper identifies well-defined LHC-signature hypotheses involving mostly gluons in signal and background as one alternative.Source: §1 · p. 2
- The paper proposes using LoLa to extract quark–gluon information from measured particle-flow objects and evaluates detector effects, tracking sensitivity, and correlations with jet momentum.It compares the tagger with traditional quark–gluon variables and studies mono-jet and di-jet benchmarks.Source: §1 · pp. 2–3
- Traditional quark–gluon observables rely on tracking, are not infrared-safe, and are not easily interpreted in perturbative QCD.Low-level approaches must combine calorimeter information with tracking's better angular resolution while accounting for detector effects.Source: §1 · p. 2
- The benchmark program covers mono-jet searches with a gluon-dominated signal and di-jet resonance searches with a quark-dominated signal.For both cases, the paper motivates tagging, demonstrates LoLa's role in signal extraction, and discusses realistic-analysis limitations.Source: §1 · pp. 2–3
2 Ideal world
The section compares standard subjet observables with a low-level LoLa tagger, including detector effects and jet-momentum dependence. LoLa improves overall on the six-observable BDT, while the gap narrows after detector simulation.
2 Ideal world
- The study begins with tree-level, Monte Carlo-truth quark and gluon jets; switching from Pythia to Sherpa makes discrimination somewhat harder.The authors note that the parton-level definition becomes ambiguous beyond leading-order QCD and that higher-order effects remain to be tested.Source: §2 · p. 3
2.1 Standard observables
- nPF and pT D are detector-sensitive, whereas pT-weighted wPF and C0.2 are more stable under detector effects.Figure 1 compares normalized quark- and gluon-jet distributions with and without detector effects for jets with pT = 200–220 GeV.Source: §2.1 · p. 4, Figure 1 · p. 4
- The detector changes observable distributions, but its effect stays within 10% of the ideal ROC across most of the curves; nPF and N95 are the most stable single observables.The 10% comparison excludes the numerically fluctuating region with quark efficiency below 0.01.Source: §2.1 · p. 5, Figure 2 · p. 5
- In the BDT, wPF and C0.2 provide the leading information: nPF and N95 are strong single variables but are correlated with wPF, while C0.2 adds information.The two leading observables, wPF and C0.2, are also identified as infrared-safe.Source: §2.1 · pp. 6–7, Figure 3 · p. 6
2.2 Charging LoLa
- LoLa takes pT-sorted constituent 4-vectors through trainable recombination and Lorentz-feature layers, with charge information encoded in its inputs.For most phase space considered, tagging performance hardly improves with charge information; the authors caution against generalizing that result.Source: §2.2 · pp. 7–8
- LoLa improves overall on the six-observable BDT, with a smaller performance gap after detector effects; without simulation, it reproduces the enhanced-images setup of Ref. [31].Figure 4 compares ROC and significance-improvement curves for LoLa and BDT variants, with and without detector simulation.Source: Figure 4 · p. 8, §2.2 · p. 10
- The added LoLa observables have limited impact on quark–gluon separation, independent of the options used for the final entry.Source: §2.2 · p. 8
2.3 Jet kinematics
- LoLa performance is stable and rises slightly toward higher jet momenta, but transfer from 200–350 GeV training to pT > 600 GeV seriously suffers; charge leaves performance unchanged below 600 GeV.Source: §2.3 · p. 11
3 Mono-jets
The mono-jet benchmark contrasts a gluon-dominated signal with quark-dominated backgrounds, but its tagging performance depends on jet momentum. The study finds the discrimination weakest in the highest-momentum range and does not establish a simple significance improvement for a realistic detector-level analysis.
3 Mono-jets
- The benchmark is invisible SM-like Higgs production in mono-jets, where the signal is gluon-dominated and the Z+jets background is mostly quark-initiated.The relative gluon purity of the Higgs signal gradually falls as pT,j increases.Source: §3 · p. 12
- The signal-to-background ratio is at the per-mille level even for a formally 100% invisible Higgs branching ratio, while leading-jet kinematics are essentially identical to background.The second jet is softer for signal; a stringent missing-energy cut-and-count strategy is not optimal because the Higgs and Z masses are similar.Source: §3 · p. 13, Figure 8 · p. 14
- AUC performance is best for networks trained and tested in the same pT,j range, then gradually decreases as pT,j rises in the mono-jet samples.The mono-jet AUCs are compared with those for pure samples in Table 1.Source: §3 · p. 14, Table 2 · p. 15
- The ROC and significance-improvement curves show a performance drop for 600–650 GeV jets, making discrimination least efficient where missing energy is largest.The analysis instead uses the full pT,j distribution and a shape analysis with large systematic uncertainties, so no simple significance improvement is quoted.Source: §3 · pp. 14–15, Figure 9 · p. 15
4 Di-jet resonances
The di-jet study uses quark-dominated Z′ decays against multigluon backgrounds as a resonance-search benchmark. A dedicated network combining both jets boosts tagging performance, though the available passage cautions that correlated event-level observables may limit the improvement in LHC reach.
4 Di-jet resonances
- The benchmark pairs quark-initiated resonance decay jets with a multigluon background, using an axial-vector Z′ with democratic quark couplings.The study considers resonance masses of 450 and 750 GeV and coupling gZ′ = 0.1.Source: §4 · p. 16
- Tagging works slightly better for lower-mass resonances or lower typical pT,j because of the QCD-background purity in that phase-space region.The second jet from the lighter resonance is comparably soft, making it hard to separate from QCD radiation without strongly shaping the background.Source: §4 · p. 16
- For a 750 GeV Z′, a dedicated two-jet LoLa network combining both jet outputs receives a significant efficiency boost.The harder jet is more sensitive at high signal efficiencies, while the second jet is more sensitive at lower efficiencies.Source: §4 · p. 16, Figure 10 · p. 17
- Event-level observables sensitive to jet flavor are correlated with di-jet tagging, so tagging is unlikely to significantly improve LHC reach once those observables are considered.The authors also note that fast neural networks could enrich quark-containing samples at trigger level, suppressing multijet backgrounds in difficult phase-space regions.Source: §4 · pp. 16–17
5 Summary
The study tests low-level quark–gluon tagging in realistic benchmark channels, finding detector effects substantially affect tagging performance. Mono-jet and di-jet applications show potential benefits, but neither is a clear benchmark for estimating LHC reach.
5 Summary
- Detector effects degrade LoLa’s separation of pure quark and gluon jets until a classic BDT becomes competitive, although its standard observables have theoretical and experimental drawbacks and non-trivial correlations.Including charge information in LoLa can be useful for hard jets, and training and testing on pT,j slices gives surprisingly stable results.Source: §5 · p. 18
- In mono-jet searches with a gluon-rich signal, subjet information can supplement limited event-level observables and may improve LHC reach, especially when controlling the pT,j distribution.The impact of pT-dependent training on systematic uncertainties is expected to be controllable.Source: §5 · p. 18
- For hadronically decaying Z′ bosons below the TeV range, LoLa quark–gluon discrimination can be useful and improves reach for relatively light Z′ bosons coupled to quarks.The study applies a network trained on pure samples to di-jet resonances with a quark-rich signal.Source: §5 · p. 18
- Neither mono-jet nor di-jet resonance searches is an obvious benchmark for estimating quark–gluon tagging’s impact on LHC reach; realistic experimental setups remain necessary.The conclusion identifies possible studies of triggering, background systematics, and signal extraction.Source: §5 · p. 18