⚛️ phenomenology

Simplex Demixing: Disentangling Multiple Light-Flavor Jets at Colliders

This paper introduces "simplex demixing," a machine-learning framework that enables the data-driven extraction and operational definition of multiple light-flavor jet categories (such as up-quark, down-quark, and gluon jets) from collider data by identifying maximally separable features within a bounded geometric structure.

Original authors: Gregorio de la Fuente, Jesse Thaler

Published 2026-07-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Gregorio de la Fuente, Jesse Thaler

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are at a massive, chaotic party where thousands of people are shouting over each other. You can't hear individual conversations, but you can hear the general hum of the room. In the world of particle physics, specifically at the Large Hadron Collider (LHC), scientists face a similar problem. They smash protons together at nearly the speed of light, creating a storm of tiny particles that fly out in tight bundles called "jets." These jets are like the party noise: a messy mixture of different types of particles (quarks and gluons) that have been scrambled together.

For a long time, physicists have tried to figure out which specific "flavor" of particle started each jet, much like trying to guess who started a specific conversation in a noisy room. The problem is that the standard way of labeling these particles relies on computer simulations that aren't perfect. It's like trying to identify a person by a blurry photo; you might get the general idea, but you can't be sure. Scientists want a way to look at the actual data from the experiments and say, "Ah, this group of particles came from a down-quark, and that one came from a gluon," without needing a perfect simulation to tell them what to look for. They need a way to "demix" the noise to hear the individual voices.

This is where a new paper comes in, introducing a clever mathematical trick called "simplex demixing." Think of the different types of jets as different colors of paint. If you mix red, blue, and yellow paint together, you get a muddy brown. If you have a bucket of muddy brown, it's usually impossible to figure out exactly how much red, blue, and yellow went into it. However, if you have several different buckets of mud, each with a slightly different shade because they were mixed in different ratios, you can work backward. By looking at the edges of the color spectrum in all those buckets, you can mathematically reconstruct the pure red, pure blue, and pure yellow paints that started it all.

The authors of this paper, Gregorio de la Fuente and Jesse Thaler, have built a machine-learning framework that does exactly this for particle jets. Instead of paint colors, they are dealing with complex data about how particles move and interact. They trained a computer to look at many different "mixtures" of jets (created from simulated data that mimics real collider conditions) and asked it to find the most distinct, pure categories hidden inside the mess.

Here is what they found:

The Geometry of Mixing: They discovered that if you have a certain number of pure jet types (let's say 7 different flavors like up-quarks, down-quarks, strange-quarks, and gluons), the computer's answers, when plotted on a graph, naturally form a specific shape called a "simplex." If there are 7 types, the shape is a 6-dimensional version of a pyramid. The corners (or vertices) of this shape represent the pure, unmixed jet flavors.
The "Tag-and-Probe" Trick: To test this, they didn't just use random data. They used a strategy called "tag-and-probe." Imagine you have a pair of twins (two jets from the same event). You use a standard, supervised computer program to guess the flavor of one twin (the "tag"). If the program is confident, you assume the other twin (the "probe") is likely the same flavor. By doing this millions of times, they created 14 different buckets of jet mixtures, each with a slightly different recipe of flavors.
The Results: When they ran their "simplex demixing" algorithm on these 14 buckets, the computer successfully identified 7 distinct corners in the data. These corners corresponded very well to the seven light-flavor jets they were looking for: down, up, strange, anti-strange, and gluon jets.
- Strong Identification: The algorithm was very good at finding the down, up, strange, anti-strange, and gluon flavors. The data for these was clear enough that the computer could isolate them almost perfectly.
- Weak Identification: However, the anti-down and anti-up flavors were much harder to spot. The computer found them, but they were "contaminated" by gluons. This happened because there were simply fewer anti-down and anti-up jets in the data, and they didn't have enough unique features to stand out clearly from the crowd. The paper suggests that with more data or better detectors, these might become clearer, but right now, they are only weakly identifiable.

The paper also showed that once they found these 7 "pure" corners, they could reconstruct the properties of each flavor. For example, they could look at how many particles are in a jet or how much electric charge it carries, and see that the "up-quark" jets had a positive charge while "down-quark" jets had a negative charge, just as physics predicts.

It is important to note that this study was a "proof-of-concept" using simulated data from a computer program called Pythia, which mimics how particles behave. While the results are promising and show that the method works in theory, the authors acknowledge that real-world detectors at the LHC aren't perfect. Real detectors might struggle to tell the difference between certain particles (like pions and kaons), which could make it harder to separate the flavors in actual experiments.

In short, this paper doesn't claim to have solved the problem of identifying every single jet in a real collider today. Instead, it offers a powerful new tool—a mathematical way to untangle the mess. It suggests that even if we can't label every single jet perfectly, we can still figure out the "pure" recipes of the different particle flavors hidden inside the data, opening the door to a more data-driven understanding of the subatomic world.

Technical Summary: Simplex Demixing: Disentangling Multiple Light-Flavor Jets at Colliders

Problem Statement
Assigning flavor labels (e.g., "quark" or "gluon") to observed jets at the Large Hadron Collider (LHC) is a fundamental challenge in collider physics. Traditional methods rely on hand-crafted observables or supervised deep learning classifiers trained on Monte Carlo (MC) simulations, which assume a perfect correlation between the parton-level flavor from the hard process and the hadron-level jet measurement. This assumption is known to fail due to non-perturbative effects and color correlations. While a "hadron-level" operational definition for quark and gluon jets was previously established using two mixed samples, it is limited to extracting only two jet categories. There is currently no robust, data-driven generalization for disentangling multiple jet flavors (e.g., distinguishing up, down, strange, and gluon jets) from mixed collider data without relying on generator-level truth labels.

Methodology: Simplex Demixing
The authors introduce "simplex demixing," a machine-learning framework designed to extract $T$ operational topics (jet flavors) from $M$ data samples (mixtures), where $T \leq M$ . The method generalizes the geometric interpretation of the two-mixture operational definition to higher dimensions.

Geometric Framework:
- The authors train a classifier on $M$ mixed jet samples using categorical cross-entropy loss.
- In the asymptotic limit, the classifier's output distribution, $\vec{c}(x)$ , is proven to lie within a $(T-1)$ -simplex embedded in the probability $(M-1)$ -simplex.
- The vertices of this $(T-1)$ -simplex correspond to "anchor regions" in phase space where a single operational topic is pure (nonzero) while all others vanish. These vertices represent the hadron-level definition of the jet flavors.
- The mixing fractions of the topics in the $M$ samples are encoded in the positions of these vertices.
Machine Learning Architecture:
- To robustly learn this geometry and avoid distortions from mismodeling, the authors propose a three-stage training procedure compatible with any base classifier architecture (e.g., Particle Flow Networks):
  - Stage 1 (Learn): The classifier learns a representation of the data with vertices fixed at the corners of the probability simplex.
  - Stage 2 (Shape): Vertices are unfrozen and pulled toward the data distribution via an auxiliary "edge loss" (sum of Euclidean distances between vertices) to enforce the simplex geometry.
  - Stage 3 (Prune): An $L_1$ (Lasso) loss is applied to vertex weights to sparsify the representation, removing redundant vertices and retaining only the $T$ active vertices corresponding to the true operational topics.
Validation and Interpretation:
- Since the operational topics are defined purely by data geometry, they may not perfectly align with specific "truth" labels (e.g., Pythia flavors) if those labels are not mutually irreducible.
- The authors introduce a probabilistic framework to quantify the alignment between inferred topics and known categories using conditional probabilities $p(t|c)$ and quasi-probabilities $p(c|t)$ , calculated via Moore-Penrose pseudoinverses of the mixing fraction matrices.

Key Results

Toy Study (Three Mixtures):
- The method was validated on synthetic mixtures of down-quark, up-quark, and gluon jets generated by Pythia.
- The classifier successfully recovered the mixing fractions and identified the three operational topics.
- When one flavor (gluons) was removed from the mixtures, the classifier geometry correctly collapsed from a 2-simplex (triangle) to a 1-simplex (line segment), demonstrating the method's ability to infer the number of latent topics.
- The inferred topics showed a near-diagonal correspondence with the Pythia truth labels, with deviations attributed to finite statistics and the non-zero reducibility factor between Pythia quark and gluon definitions.
Proof-of-Concept Physics Application (Dijet Events):
- The authors applied the method to a quasi-realistic scenario involving 14 jet mixtures derived from dijet events at HL-LHC statistics ( $\sqrt{s}=14$ TeV).
- Using a "tag-and-probe" strategy, they constructed mixtures where the "tag" jet was classified by a supervised tagger, and the "probe" jet was assigned to a mixture based on the tag's prediction.
- Training on 14 mixtures, the classifier distribution was found to be bounded by a 6-simplex, indicating $T=7$ operational topics.
- Strongly Identifiable Topics: Five topics ( $T_1, T_3, T_5, T_6, T_7$ ) were strongly associated with Pythia down, up, strange, anti-strange, and gluon flavors, respectively.
- Weakly Identifiable Topics: Two topics ( $T_2, T_4$ ) were associated with anti-down and anti-up flavors but were heavily contaminated by gluons due to the low abundance of these flavors in the dataset and weak correlations.
- Substructure Extraction: The authors successfully inverted the mixing fractions to extract substructure distributions (constituent multiplicity, 2-subjettiness, and jet charge) for each operational topic. The extracted distributions for the strongly identifiable topics matched their corresponding Pythia truth distributions, including the separation of jet charges for up vs. strange quarks.

Significance and Claims
The paper claims that simplex demixing provides a practical, data-driven pathway to define and extract multiple jet flavor categories directly from collider data, bypassing the ambiguities of generator-level definitions. Key contributions include:

Generalization: Extending the operational definition of jet flavors from two categories to an arbitrary number $T$ .
Geometric Insight: Establishing that the output of a classifier on mixed samples is bounded by a simplex whose vertices define the pure topics.
Data-Driven Extraction: Demonstrating that light-flavor jet categories (up, down, strange, anti-strange, gluon) emerge clearly at the ensemble level, even if tagging individual jets is difficult.
Substructure Recovery: Showing that the method can recover the substructure distributions of these latent topics, enabling the study of flavor-dependent jet properties without labeled training data.

The authors note that while the proof-of-concept assumes perfect detector information (specifically $\pi/K/p$ discrimination), the framework is designed to be applied to experimental data. They acknowledge that the identifiability of specific flavors depends on their relative abundance and the available hadron-level information, and that future work is needed to address realistic detector limitations and the selection of the number of topics $T$ in a domain-agnostic manner.

More like this