Compact Representation of Particle-Collision Events for Physics-Informed Machine Learning

This paper introduces RMM-C46, a compact and physics-driven representation that compresses the high-dimensional rapidity mass matrix into a low-dimensional, interpretable feature set, achieving performance comparable to the full matrix in machine learning tasks while enabling compatibility with near-term quantum computing applications.

The Gist

Imagine you are a detective trying to solve a crime, but instead of a few witnesses, you have thousands of security cameras recording every single movement in a massive city. Each camera captures a tiny detail: a shoe print, a shadow, a specific color of a car.

In the world of particle physics (specifically at the Large Hadron Collider, or LHC), scientists face a similar problem. When protons smash together, they create a chaotic explosion of particles. To understand what happened, physicists use a "map" called the Rapidity-Mass Matrix (RMM).

The Problem: The "Over-Engineered" Map

The original RMM is like a giant spreadsheet with 2,600 cells for every single collision.

• The Good News: It records everything. It knows exactly how every particle relates to every other particle.
• The Bad News: Most of those 2,600 cells are empty (zeros) because a single collision doesn't produce 2,600 different types of particles. It's like filling out a 2,600-page form where 90% of the boxes are blank.
• The Consequence: This makes it incredibly slow for computers (and especially for new "Quantum Computers") to process. It's like trying to drive a Ferrari through a traffic jam of empty boxes.

The Solution: The "RMM-C46"

The authors of this paper, W. Islam and S. Chekanov, invented a new way to look at the data. They call it RMM-C46.

Think of the original 2,600-cell spreadsheet as a giant, messy warehouse full of boxes.

• The Old Way: You try to count every single item in every single box, even the empty ones.
• The New Way (RMM-C46): You group the boxes into 46 specific categories (zones). Instead of counting every item, you just measure the "total weight" or "total energy" of each category.

How it works in simple terms:

1. Grouping: They take all the "Jet" particles and group them together. They take all the "Electron" particles and group them.
2. Summarizing: Instead of listing every single electron, they calculate a single number that represents the "total punch" (energy) of all the electrons combined.
3. The Result: They shrink that massive 2,600-cell map down to a neat, tidy list of just 46 numbers.

Why is this a Big Deal?

1. It's Smarter, Not Just Smaller
You might think, "If you throw away 98% of the data, you must lose important clues." Surprisingly, the paper shows that you don't.

• Analogy: Imagine trying to identify a song. You could listen to every single sound wave (the original 2,600 cells), or you could just listen to the melody, the rhythm, and the bass line (the 46 numbers). The melody tells you everything you need to know about the song, without the noise.
• The Result: When they tested this on a computer, the 46-number list worked just as well (and sometimes even better) at spotting "new physics" (the crime) than the massive 2,600-cell list.

2. It's "Quantum-Ready"
New quantum computers are very powerful but very small (they have very few "qubits," or memory slots). They can't handle the 2,600-cell map at all.

• The Fit: The RMM-C46 list has exactly 46 numbers. This fits perfectly into the limited memory of current quantum computers. It's like shrinking a giant suitcase down to the size of a carry-on bag so you can finally get it on the plane.

3. It's Easy to Understand
Because the 46 numbers are based on real physics (like "total energy" or "distance between particles"), scientists can look at the list and say, "Ah, the 'Jet Energy' number is high, which means we found a heavy particle!"

• Contrast: Other methods use "black box" AI that gives you a result but doesn't explain why. RMM-C46 is like a transparent window; you can see exactly what the computer is looking at.

The Bottom Line

The authors took a messy, huge, and slow way of looking at particle collisions and turned it into a compact, efficient, and crystal-clear summary.

• Before: A giant, confusing library with thousands of empty books.
• After: A single, well-organized index card with 46 key facts that tells the whole story.

This new method helps scientists find "New Physics" (like the mysterious Higgs boson or even stranger particles) faster, cheaper, and with the help of the next generation of quantum computers. It's a win for physics, a win for computers, and a win for understanding the universe.

Technical Summary

1. Problem Statement

High-energy physics (HEP) analyses at the Large Hadron Collider (LHC) increasingly rely on Machine Learning (ML) to detect subtle deviations from the Standard Model (SM). However, existing event representations face significant bottlenecks:

• High Dimensionality & Redundancy: The Rapidity-Mass Matrix (RMM) is a structured, physics-grounded representation encoding pairwise correlations (invariant mass, rapidity difference, transverse energy) among jets, $b$-jets, leptons, photons, and missing transverse energy (MET). While effective, a full RMM (e.g., $51 \times 51$) contains $\sim 2,600$ entries. Even optimized versions have $\sim 1,287$ entries.
• Sparsity Issues: To accommodate events with varying object multiplicities, the RMM uses fixed-size slots, resulting in many zero-padded entries. In unsupervised anomaly detection (e.g., using Autoencoders), these zeros can dominate the loss function, forcing models to learn padding artifacts rather than physical signals.
• Quantum Incompatibility: Current Near-term Noisy Intermediate-Scale Quantum (NISQ) devices have severely limited qubit counts ($O(10)$). Encoding high-dimensional RMM data directly is impossible without aggressive, often physics-agnostic, preprocessing.
• Interpretability Loss: Many dimensionality reduction techniques (e.g., generic autoencoders or learned embeddings) strip away the explicit mapping to physical observables, making the resulting features difficult to interpret physically.

2. Methodology: The RMM-C46 Representation

The authors propose RMM-C46, a compact, physics-driven representation that reduces the input dimensionality from $\sim 1,000+$ to 46 features while preserving the block structure and physical semantics of the original matrix.

Construction Logic

The method partitions the full RMM into 46 disjoint zones, each corresponding to a specific physical quantity or interaction type. These zones are aggregated into single scalar features using two schemes:

1. Additive Aggregation: Summing all entries in a zone.
2. Frobenius Aggregation (Default): Calculating the Frobenius norm ($\sqrt{\sum R_{ij}^2}$) of the zone. This emphasizes large entries (hard/boosted objects) and provides a strictly non-negative "energy-like" measure.

The 46 Feature Components

The 46 variables are derived as follows:

• 1 Global MET: The magnitude of missing transverse energy.
• 5 Transverse Energy ($E_T$) Terms: Aggregated diagonal entries for each object class (Jets, $b$-jets, Muons, Electrons, Photons).
• 5 Transverse Mass ($T$) Terms: Aggregated entries from the first row (MET vs. object class).
• 5 Longitudinal/Lorentz ($L$) Terms: Aggregated entries from the first column (object class vs. longitudinal kinematics).
• 15 Rapidity Difference ($h$) Zones: Pairwise rapidity differences for all combinations of object classes (e.g., jet-jet, jet-$b$-jet, muon-electron).
• 15 Invariant Mass ($m$) Zones: Pairwise invariant masses for all combinations of object classes.

This structure ensures that the representation retains the block-diagonal factorization of the original physics matrix, maintaining interpretability.

3. Key Contributions

• Physics-Informed Compression: Unlike generic dimensionality reduction, RMM-C46 explicitly aggregates physically correlated regions (e.g., "total jet-jet invariant mass activity") into well-defined scalar quantities.
• Quantum Readiness: The dimensionality of 46 aligns naturally with the capacity of near-term quantum devices, allowing for direct angle or amplitude encoding without complex latent space mapping.
• Handling Sparsity: By aggregating zones, the method effectively handles the zero-padding issue inherent in fixed-size matrices, preventing models from learning artifacts of missing data.
• Open Source: The implementation is made publicly available via a GIT repository.

4. Experimental Results

The authors evaluated RMM-C46 using Monte Carlo simulations of proton-proton collisions at $\sqrt{s} = 13.6$ TeV.

• Signal Processes: Exotic cascades ($X \to SH \to HHH$) and direct di-Higgs ($X \to HH$) with resonance masses $m_X = 0.5\text{--}2.0$ TeV.
• Backgrounds: Inclusive $t\bar{t}$ (dominant) and $WZ$+jets.

Supervised Learning (Binary Classification)

• Task: Distinguish $X \to SH$ signal from $t\bar{t}$ background using a Multi-Layer Perceptron (MLP).
• Result: RMM-C46 achieved an AUC of 0.999, slightly outperforming the full RMM (AUC 0.998).
• Implication: The compact representation preserves nearly all discriminative power while reducing input size by two orders of magnitude.

Unsupervised Learning (Anomaly Detection)

• Task: Train Autoencoders (AE) and Variational Autoencoders (VAE) on background ($t\bar{t}$) only; detect signal via reconstruction error.
• Result:
  • Autoencoder: RMM-C46 achieved AUC 0.9995, significantly outperforming the full RMM (AUC 0.9865).
  • VAE: RMM-C46 consistently matched or exceeded the full RMM across all resonance masses, showing superior scaling at higher masses.
• Interpretation: The removal of sparse, noisy, and redundant zero-padded regions in the full RMM allows the compact representation to learn a cleaner, more faithful manifold of the background physics.

Correlation Analysis

Correlation matrices of the 46 features showed strong block structures corresponding to object classes (e.g., jet-jet correlations), confirming that the physical factorization of the event topology is preserved.

5. Significance and Future Outlook

• Computational Efficiency: RMM-C46 drastically reduces training time and memory footprint, enabling rapid prototyping of ML architectures.
• Quantum-Ready: It provides a realistic pathway for Quantum Machine Learning (QML) in HEP. The 46 features can be mapped to $\sim 10$ qubits using efficient encoding protocols, bridging the gap between high-dimensional collider data and NISQ hardware constraints.
• Scalability for HL-LHC: As the High-Luminosity LHC generates unprecedented data volumes, RMM-C46 offers a standardized, interpretable, and computationally efficient interface for both classical and hybrid quantum-classical analyses.
• Interpretability: By maintaining a direct link to physical observables (mass, rapidity, energy), the method avoids the "black box" nature of learned embeddings, facilitating robust physics validation.

In conclusion, RMM-C46 represents a paradigm shift from raw, high-dimensional data ingestion to physics-driven feature engineering, offering a scalable, interpretable, and quantum-compatible solution for next-generation collider physics.