In high-energy particle colliders, researchers look for heavy particles—such as $W$, $Z$, or Higgs bosons, and top quarks—by analyzing the sprays of subatomic particles they decay into. When these heavy particles are produced with immense kinetic energy, their decay products are highly collimated, merging into a single, wide "fat jet" (Precision Jet Substructure from Boosted Event Shapes).
To understand the physics of the original heavy particle, physicists must peer inside this jet and measure its internal properties, a field known as jet substructure (Boosted objects and jet substructure at the LiHC. Report of BOOST2012, held at IFIC Valencia, 23rd–27th of July 2012). However, high-energy collision environments are incredibly noisy. The primary signal is routinely obscured by a "haze" of extraneous, low-energy particles (Jet substructure at the Tevatron and LHC: new results, new tools, new benchmarks).
This noise stems from three main physical sources:
- Pileup: Multiple simultaneous proton-proton interactions occurring within the same bunch crossing (Exclusive γγ → µ^+µ^- production in proton-proton collisions at √s = 7 TeV).
- Initial-State Radiation (ISR): Gluons radiated from the incoming colliding partons before the hard scattering event occurs.
- The Underlying Event (UE): Soft interactions between the remaining "spectator" partons of the colliding protons.
To make precise physical measurements, physicists must find ways to cleanly separate the hard, primary signal from this soft radiation contamination.
The Geometry of Contamination
Why does this soft background radiation pose such a severe problem? The answer lies in the mathematics of how jet properties, particularly invariant mass, are calculated.
The invariant mass $m$ of a jet can be approximated by summing over the pairs of particles within it: $$m^2 \approx \sum_{i < j} E_i E_j \theta_{ij}^2$$ where $E_i$ and $E_j$ are the energies of the particles, and $\theta_{ij}$ is the opening angle between them.
Notice that the mass depends quadratically on the angle $\theta_{ij}$. The primary decay products of a boosted heavy particle are highly energetic and tightly collimated (small $\theta$). Conversely, pileup and underlying event radiation are diffuse, soft, and distributed at wide angles across the detector.
Even though an individual pileup particle has very low energy ($E$), its wide angle ($\theta$) relative to the jet axis gives it a disproportionate impact on the calculation. As a result, soft, wide-angle radiation artificially inflates the reconstructed mass of the jet, shifting and broadening the mass peaks that physicists use to identify heavy particles (Towards an understanding of the correlations in jet substructure).
This contamination problem is universal across high-energy physics experiments. For instance, in neutrino experiments like MINERvA, electromagnetic showers are similarly contaminated by stray energy deposits from unrelated neutrons or pions, degrading mass reconstruction (Neutral pion reconstruction using machine learning in the experiment at 〈E ν〉 6 GeV).
Algorithmic Pruning: How Jet Grooming Works
To combat this distortion, theorists and experimentalists developed jet grooming algorithms. These techniques exploit the geometric and energetic asymmetry between the signal (hard and collinear) and the noise (soft and wide-angle) to systematically strip away contamination.
Jet Trimming
One of the earliest widespread grooming techniques is jet trimming, introduced by David Krohn et al. (Jet trimming). Trimming works by taking an already formed jet and breaking it down into smaller, narrower "subjets" using a smaller clustering radius $R_{\text{trim}}$.
The algorithm inspects each subjet individually. If a subjet carries less than a specified fraction $f_{\text{trim}}$ of the total jet's transverse momentum ($p_T$), it is discarded. Because pileup and ISR form soft, isolated clusters, they fail this threshold and are scrubbed away, while the energetic cores of the heavy particle decay are preserved.
Soft Drop
A more dynamically adaptive approach is Soft Drop, developed by Andrew J. Larkoski et al. (Soft drop). Instead of creating subjets with a fixed radius, Soft Drop takes a jet and sequentially unwinds its clustering history.
Sequential clustering algorithms build a jet by pairing the closest particles first and working outward. Soft Drop reverses this process, starting from the final clustering step and splitting the jet into two subjets, $j_1$ and $j_2$. It then checks the following condition:
$$\frac{\min(p_{T1}, p_{T2})}{p_{T1} + p_{T2}} > z_{\text{cut}} \left(\frac{\Delta R_{12}}{R_0}\right)^\beta$$
where $z_{\text{cut}}$ is a sensitivity threshold, $\Delta R_{12}$ is the angular distance between the subjets, $R_0$ is the total jet radius, and $\beta$ is a tunable angular exponent.
[ Jet Split ]
/ \
/ \
Subjet 1 Subjet 2 (Softer: min(pT1, pT2))
| |
[Hard Core] [Soft & Wide-Angle?]
|
---> YES: Drop Subjet 2, Recurse on Subjet 1
---> NO: Stop. Jet is clean.
If the split fails this condition—meaning one branch is highly asymmetric (too soft) and located at a wide angle—the softer branch is completely dropped from the jet. The algorithm then recurses, unwinding the clustering history of the harder branch and repeating the test.
By tuning $\beta$, physicists can control how aggressively wide-angle radiation is penalized. When $\beta = 0$, the angular dependence drops out, acting as a pure energy cutoff. This mathematical cleanliness allows Soft Drop to strip away non-perturbative noise so effectively that researchers can directly expose fundamental quantum chromodynamics (QCD) parameters, such as the underlying splitting function that governs how quarks and gluons radiate (Exposing the QCD Splitting Function with CMS Open Data).
As colliders push to higher energies and luminosities, the density of pileup increases dramatically. To handle these extreme environments, Frédéric A. Dreyer et al. introduced Recursive Soft Drop (Recursive Soft Drop). Rather than stopping at the first branch that passes the grooming condition, the recursive variant continues down both branches of the clustering tree, purging soft radiation that might be hidden deeper within the jet structure.
Alternative Paths to Noise Robustness
While grooming algorithms physically remove particles from a jet, other research directions achieve noise resilience by changing how jets are measured or defined.
Track-Based Observables
One elegant way to bypass pileup entirely is to ignore neutral particles. Detectors can trace charged particles back to their precise point of origin (vertex) along the beamline. Hsi-Ming Chang et al. developed a theoretical framework called "track functions" (Calculating Track-Based Observables for the LHC). By calculating jet observables using only these vertex-validated charged tracks, pileup contamination from secondary collisions can be almost completely filtered out at the hardware level.
Recoil-Free Observables
Instead of filtering particles, one can design mathematical functions that are intrinsically insensitive to soft noise. Andrew J. Larkoski et al. introduced Energy Correlation Functions (Energy correlation functions for jet substructure), which measure the geometric configurations of energy flow within a jet without ever defining subjets. By using specific ratios of these functions ($C_N$), the observables remain robust against the soft "recoil" effects induced by background noise. Similarly, using "winner-take-all" recombination schemes allows the defined axis of a jet to remain stable even when bombarded by diffuse pileup (Aspects of jets at 100 TeV).
Fuzzy and Machine Learning Approaches
Traditional jet algorithms draw rigid, binary boundaries: a particle is either entirely inside a jet or entirely outside it. This makes them highly sensitive to edge-case pileup particles. Benjamin Nachman et al. proposed "fuzzy jets" (Fuzzy jets), which utilize probabilistic mixture models via the Expectation-Maximization algorithm. Particles are assigned a continuous weight belonging to a jet, allowing the jet to dynamically adapt its shape and naturally down-weight diffuse background noise.
In a similar vein, modern machine learning architectures bypass manual algorithmic tuning by learning the complex correlations between signal and noise. For example, T. Q. Nguyen et al. demonstrated that Graph Neural Networks (specifically Interaction Networks) can map the relationships between particles and secondary vertices (Interaction networks for the identification of boosted H → bb¯ decays). These networks inherently learn to isolate the hard scattering core, maintaining high particle identification accuracy even when operating in high-luminosity environments where traditional methods begin to degrade (Exposing the dead cone effect with jet substructure techniques, The first calculation of fractional jets).
Key Foundational Work
Jet substructure at the Tevatron and LHC: new results, new tools, new benchmarks (2012) — contributed 1 concept elements to this direction.
Jet trimming (2010) — contributed 1 concept elements to this direction.
Precision Jet Substructure from Boosted Event Shapes (2012) — contributed 1 concept elements to this direction.