Where Hadronization Information Lives: Multiplicity Scaling of Subgraph Retention in Jets

A jet is the spray of particles left behind when a quark or gluon fragments. Machine-learning models that read jets as graphs of particles are very accurate. What nobody can say is which information they use. Some of it comes from the hard, calculable part of the process: a few energetic sub-jets whose structure perturbation theory predicts. Some comes from hadronization, the poorly understood, non-perturbative stage where quarks bind into hadrons and scatter many soft particles. We propose a simple, quantitative way to tell these apart. Restrict a model to the best $k$ particles out of a jet's $N$, and measure how much task-relevant information survives as $N$ grows. If information sits in a few hard particles, the retained fraction depends on $k$ but not on $N$. If it is spread thinly over many soft hadrons, the retained fraction falls roughly like $k/N$.

The idea borrows its central question from wireless network simplification. That literature asks what fraction of a relay network's capacity survives when you route through only $k$ of $N$ relays. It found two qualitatively different answers. In full-duplex diamond networks the guaranteed fraction is $k/(k+1)$, independent of $N$ (Ozgur et al.; Nazaroglu et al.). In half-duplex networks the guarantee degrades as $N$ grows (Cardone et al.). We do not claim these theorems transfer to particle physics. We take the diagnostic, whether the retention curve depends on size, and turn it into a measurement of how non-perturbative a learned jet observable is. We then show that this one number predicts something practical: whether a jet model trained on low-multiplicity jets will keep working on high-multiplicity jets.

Core Claim

For a jet task with binary label $Y$ and a particle-selection rule that keeps $k$ of the jet's $N$ constituents, define the retention

$$R(k,N) = \frac{I(Y; X_{S_k})}{I(Y; X)},$$

where $I(Y;X)$ is the mutual information between the label and the full particle graph $X$, and $X_{S_k}$ is the subgraph on the selected particles. Both mutual informations are estimated from held-out cross-entropy as $H(Y)$ minus the classifier loss. We model retention as $R(k,N)\approx g(k)\,N^{-\gamma}$ within multiplicity bins.

The two limits follow from elementary arguments. If the likelihood ratio depends only on the $m$ hardest constituents, as for an IRC-safe prong-counting observable (infrared- and collinear-safe, meaning insensitive to soft or collinear splittings), then $R=1$ for every $k\ge m$, regardless of $N$, so $\gamma=0$. If the signal is a sum of $N$ roughly independent small per-particle contributions, divergence adds across particles and a fixed-$k$ subset captures about $k/N$ of it, so $\gamma=1$.

We call the fitted exponent $\gamma$ the diffuseness exponent of a task–model pair. We make three claims. First, $\gamma$ is near zero for top-quark tagging and near one for the parton-versus-hadron "calculability test". Second, $\gamma$ orders tasks consistently with existing robustness probes. Third, the generalization gap from low- to high-multiplicity jets grows with $\gamma$.

To our knowledge, no work in the file connects size-dependent retention bounds to hadronization. The two papers nearest in spirit each give a single snapshot rather than a size-resolved curve. Bright-Thonney, Nachman and Thaler show that an unconstrained energy-flow network separates parton- from hadron-level jets with AUC ~0.96, and that a Lipschitz-constrained version drops to ~0.55. Both numbers are global and say nothing about where in the jet the hadronization information sits. Mokhtar, Kansal and Duarte use edge-level relevance propagation, a method that assigns credit for a prediction back to individual graph edges. They find that about 40% of ParticleNet's most relevant edges in top jets connect distinct sub-jets, versus under 10% in QCD jets. That is a per-edge attribution, not a test of how much information a sub-graph can carry.

Research Program

The first component is the measurement apparatus. For each task, we train a message-passing graph network on the subgraph induced by $S_k$, using a fixed architecture: EdgeConv in the style of ParticleNet, at matched width. We do this for $k\in\{1,2,4,8,16,32\}$ and for multiplicity bins of $N$. We compare four selection rules:

  • Random: a lower envelope.
  • Hardest-$k$ by energy fraction: the IRC-motivated naive rule.
  • SoftDrop survivors: the particles kept by the standard grooming algorithm. When fewer than $k$ particles survive, we pad with the hardest removed particles.
  • Learned greedy: at each step we add the particle with the largest estimated marginal information gain, using a cheap scoring network amortized over jets.

The greedy rule comes from the submodular pruning analysis of Ezzeldin et al.. There, exploiting diminishing returns turns a loose $2^{-(N-k)}$ guarantee into tight statements. Whether jet information has diminishing returns at all is an empirical question. We test it directly by checking whether marginal gains shrink as the selected set grows. If they do not, the greedy rule loses its theoretical motivation, and we report that.

The second component fits $\gamma$ with bootstrap intervals per task, and checks that the power-law form is adequate against a two-parameter saturating alternative.

The third component is the topology-generalization test. Jets of different multiplicity are graphs of different size and connectivity. This is the same transfer problem that troubles graph policies trained on small graphs, as Gao, Eisen and Ribeiro illustrate in a wireless setting. We train each model on the lowest multiplicity tercile and evaluate on the highest. We then ask whether the size of the degradation is predicted by $\gamma$ measured on the training distribution alone.

Baselines

The do-nothing reference is the full-graph ParticleNet-style model trained on all constituents. It defines the denominator $I(Y;X)$ and the in-distribution ceiling. On top tagging, Mokhtar et al. report an overall AUC of 98.23%, which our reimplementation must approach before any retention number is trusted. On the calculability task, the reference is the unconstrained energy-flow network of Bright-Thonney et al. at AUC ~0.96. For Herwig nominal-versus-varied hadronization, the reference is the Deep Sets discriminator of Chan et al. at AUC 0.77, against roughly 0.5 for traditional observables.

The strongest existing method for isolating hadronization sensitivity is the Lipschitz energy-flow network of Bright-Thonney et al. Its spectral normalization bounds how much the output can move per unit of Energy Mover's Distance. If our exponent carries no information beyond "this model's AUC collapses under a Lipschitz constraint", the proposal adds nothing. We run it on identical samples with the same hyperparameter search budget as our subgraph models. We record its calculability AUC and its low-to-high multiplicity transfer gap.

The proposal counts as beating it if $\gamma$ explains variance in the transfer gap across tasks and model variants after controlling for the Lipschitz-constrained AUC drop. We test this with a held-out-task regression. Success is $\gamma$ remaining a significant predictor and improving rank correlation with the gap.

Hardest-$k$ and SoftDrop selection serve both as naive baselines for the selection rule and as physics-standard grooming. Learned greedy selection beats them if it reaches $R\ge 0.9$ at strictly smaller $k$, averaged over multiplicity bins, under the same training budget.

The key ablation removes the size resolution. We compute retention pooled over all $N$, which is the ordinary "AUC versus number of constituents" plot. We then ask whether the pooled curve predicts the transfer gap as well as $\gamma$ does. If it does, the scaling framing is unnecessary.

Why not just…?

Why not just plot accuracy against the number of kept constituents, as jet physicists already do? This is the most serious threat, because our apparatus is exactly that plot sliced by multiplicity. The claim stands or falls on the pooled-versus-resolved ablation above. If the slicing adds no predictive power for transfer, we will report the exponent as a reparameterization of a known plot and drop the stronger claims.

Why not just use the Lipschitz calculability test? It answers whether a model can be made insensitive to hadronization, at the cost of accuracy. It does not say whether the original model's information is concentrated or diffuse, and it requires retraining a constrained architecture that, as built, applies only to Deep Sets (Bright-Thonney et al.). The exponent applies unchanged to message-passing graphs. The head-to-head regression above decides whether it adds anything.

Why not just groom with SoftDrop and train on what survives? SoftDrop is a fixed kinematic rule with no statement about what it discards. Chen et al. show that even after SoftDrop, every major generator misdescribes ALEPH substructure by several percent in soft regions. That suggests the groomed-away region matters for hadronization modelling. SoftDrop is included as a selection rule, so its retention curve is measured directly rather than assumed.

Why not just improve size generalization with multiplicity augmentation or size-normalized pooling? These may well shrink the transfer gap. They answer a different question, though: how to fix transfer, not why a given model fails it. We include a size-augmented full-graph model. The diagnostic claim holds only if $\gamma$ still predicts which tasks benefit most from augmentation.

Proposed Experiments

We use three datasets.

  • Top tagging: the top-quark versus QCD jet benchmark used by Mokhtar et al..
  • Calculability: parton- versus hadron-level jets generated as in Bright-Thonney et al..
  • Hadronization variants: Herwig 7.2.1 $e^+e^-$ events at 91.2 GeV, with nominal versus varied cluster-hadronization parameters, following Chan et al.. We add the flavor-resolved variant of Chan et al. to check whether discrete species labels on particles change $\gamma$.

The metrics are:

  • the retention $R(k,N)$ from cross-entropy mutual-information estimates, with AUC and background rejection at fixed signal efficiency reported alongside;
  • the fitted $\gamma$ with 95% bootstrap intervals;
  • the marginal-gain monotonicity rate, which tests diminishing returns;
  • the low-to-high multiplicity drop in mutual information and in AUC.

There are four decisive checks.

  1. Separation of known cases. Top tagging should give $\gamma$ consistent with zero for $k\ge 4$, and the calculability task should give $\gamma$ near one. If both tasks give similar exponents, the diagnostic fails.
  2. Ordering of hadronization tasks. The Herwig parameter-variation task should sit between the two. Its exponent should rise when parameters that mainly affect soft multiplicity are varied, compared with those affecting cluster kinematics.
  3. Prediction of transfer. Across tasks, architectures, and Lipschitz-constrained variants, $\gamma$ measured on the training multiplicity range should rank-correlate with the transfer gap, and it should outperform both the pooled curve and the Lipschitz AUC drop as a predictor.
  4. Selection efficiency. Learned greedy selection should reach 90% retention at smaller $k$ than hardest-$k$ and SoftDrop, with the margin largest on high-$\gamma$ tasks. Large-margin cases are where a grooming rule with a stated information budget would be useful in practice.

A failure of check 3 with a pass on check 1 would still yield a clean, model-agnostic measurement of where hadronization information lives in a jet, but not the stronger claim about generalization.