Observable Optimization for Precision Theory: Machine Learning Energy Correlators

This paper demonstrates how machine learning-based simulation techniques can systematically optimize precision-theory-compatible observables, specifically identifying that isosceles right-triangle energy 3-point correlators maximize sensitivity to the top quark mass while remaining computable to high precision.

The Gist

Imagine you are a detective trying to solve a crime, but the crime scene is a particle collider where protons smash together at nearly the speed of light. The "clues" are the debris flying out of the crash. Your goal is to figure out the weight of a specific criminal: the top quark (the heaviest known elementary particle).

The problem? The debris is a chaotic, multi-dimensional mess. There are thousands of particles, moving in all directions with different energies. If you try to look at everything at once, it's too much data. If you try to look at just one simple thing (like the total energy), you lose too much detail.

This paper is about building a smart, automated detective that helps you find the perfect way to look at the debris to catch the top quark's weight.

Here is the breakdown of their method, using simple analogies:

1. The Problem: The "Black Box" vs. The "Blueprint"

In physics, there are two types of tools:

• The Black Box (Neural Networks): You can train a computer to look at the debris and say, "That looks like a top quark!" with incredible accuracy. But, the computer can't explain why. It's like a magic 8-ball. You can't use a magic 8-ball to write a textbook or prove a theory because no one understands the math behind the answer.
• The Blueprint (Precision Theory): Physicists can calculate very specific, simple shapes of debris using pure math. But these simple shapes often aren't sensitive enough to catch the top quark's weight accurately.

The Goal: The authors wanted to find a shape that is both simple enough to be calculated with pure math (a blueprint) and sensitive enough to catch the top quark (a good detective clue).

2. The Strategy: The "Map Maker" and the "Searcher"

The team used a two-step machine learning process to find this perfect shape.

Step 1: The Map Maker (Learning the Terrain)

Imagine the debris patterns exist in a giant, 3D landscape. Some areas are crowded with soft, low-energy particles (like a foggy valley), and some areas have high-energy particles (like a mountain peak).

• The top quark's weight is hidden in the mountain peaks.
• Standard computer programs (like a simple neural network) tend to get distracted by the foggy valleys because there are so many particles there. They ignore the peaks.
• The Fix: The authors taught their computer to ignore the fog and focus only on the energy peaks. They created a "weighted map" that highlights the high-energy parts of the collision.
• They used two different types of "Map Makers" (a Dense Neural Network and a Normalizing Flow) to learn the shape of this landscape perfectly.

Step 2: The Searcher (Finding the Best Angle)

Now that they have a perfect map of the debris, they need to find the best "viewing angle" to measure the top quark.

• Think of the debris as a triangle formed by three particles. You can measure this triangle in infinite ways: Is it equilateral (all sides equal)? Is it skinny? Is it wide?
• The authors used a technique called Neural Ratio Estimation (NRE). Imagine this as a "Shape Optimizer."
• The Optimizer tries millions of different triangle shapes. For each shape, it asks: "If I measure the universe using this specific triangle, how clearly can I see the top quark's weight?"
• It keeps the shape that gives the clearest, most precise answer.

3. The Discovery: The "Right-Angled Isosceles" Triangle

After searching through millions of possibilities, the computer found a winner.

The best shape to measure the top quark isn't a perfect equilateral triangle (which was the standard guess before). Instead, the optimal shape is a Right-Angled Isosceles Triangle.

• The Metaphor: Imagine a triangle where two sides are equal length, and the third side is longer, forming a perfect "L" shape (like the corner of a square).
• Specifically, the ratio of the sides is roughly 1 : 1 : 1.41 (which is the square root of 2).

4. Why This Matters

The most important part of this paper isn't just the triangle shape; it's the process.

1. No Magic Left Behind: The computer used machine learning to find the triangle, but the final result is just a simple geometric definition: "Measure triangles with sides in a 1:1:√2 ratio."
2. Pure Math: Because the result is just a simple shape, physicists can now go back to their chalkboards and calculate the top quark's mass using pure theory, without needing the computer to explain itself.
3. Future Proof: This method can be used for any particle, not just the top quark. It's a new way to design experiments: let AI explore the infinite possibilities of how to look at data, find the best "lens," and then hand that lens to the human theorists.

Summary Analogy

Imagine you are trying to hear a specific whisper in a noisy stadium.

• Old Way: You put on a generic noise-canceling headset (standard observables). It helps, but you still miss the whisper.
• Black Box Way: You hire a super-smart AI that listens and says, "I hear the whisper!" but it won't tell you how it filtered the noise. You can't trust it for official records.
• This Paper's Way: You use the AI to analyze the stadium's acoustics. The AI realizes that if you stand in a specific spot and tilt your head at a 45-degree angle, the whisper becomes crystal clear.
• The Result: You tell everyone, "Stand here and tilt your head 45 degrees." Now, anyone can do it, and you can prove mathematically why that angle works.

The paper found that 45-degree angle for the top quark: a right-angled isosceles triangle.

Technical Summary

1. Problem Statement

In collider physics, there is a fundamental tension between observables that are useful for specific tasks (e.g., distinguishing quark/gluon jets or detecting anomalies) and observables that are theoretically calculable to high precision.

• The Gap: Many powerful observables, such as neural network outputs, are impossible to calculate from first principles (requiring simulation) and thus cannot be directly compared to precision theory. Conversely, many theoretically calculable observables (e.g., high-point scattering amplitudes) are difficult to measure or lack sensitivity to specific physical parameters.
• The Goal: The authors aim to bridge this gap by systematically exploring the space of calculable observables to find the specific "marginalization" (projection) of a multi-dimensional distribution that is optimally sensitive to a physical parameter (in this case, the top-quark mass, $m_t$).
• The Challenge: Directly searching the space of observables using Monte Carlo (MC) simulation data is computationally prohibitive. Furthermore, standard density estimation techniques often fail to capture the high-energy regions of a distribution that are most sensitive to mass parameters, as they prioritize high-probability (soft) regions.

2. Methodology

The authors propose a two-step Machine Learning (ML) framework to optimize precision observables without the final observable depending on the ML model itself.

Step 1: Learning the Multi-Dimensional Distribution (Density Estimation)

The authors focus on the 3-point Energy-Energy Correlator (EEEC), a theoretically calculable observable in QCD that describes the angular correlation of energy flow. They aim to learn the differential distribution $EEEC(\zeta_1, \zeta_2, \zeta_3, m_t)$ from $e^+e^-$ collision simulations (Pythia 8).

• Data: Simulated top-quark jets at $\sqrt{s} = 2$ TeV. The input space is 5-dimensional: three angular variables ($\zeta_1, \zeta_2, \zeta_3$), the normalized energy product ($\tilde{E}$), and the MC top mass ($m_t$).
• Approaches Tested:
  1. Dense Neural Network (DNN): Trained to minimize the Negative Log-Likelihood (NLL).
     • Innovation: The authors introduce a physics-motivated loss function. Instead of learning the probability density $p(\vec{x})$, they learn the energy-weighted distribution directly. Since the EEEC is an energy-weighted integral of the cross-section, the loss function re-weights samples by their energy product ($\tilde{E}$). This forces the network to focus on the high-energy regions critical for mass discrimination, which standard probability density learning often ignores.
  2. Normalizing Flows (NF): A more complex architecture using invertible maps to learn the probability density.
     • Result: While NFs can learn the distribution without explicit energy weighting, they were found to be less accurate than the DNN with the weighted loss in capturing the specific features relevant to $m_t$ discrimination and were computationally slower to evaluate.
• Outcome: The DNN with the energy-weighted loss successfully learns an analytic surrogate, $EEEC_\phi(\vec{\zeta}, m_t)$, which can be marginalized to any lower-dimensional shape.

Step 2: Observable Optimization via Neural Ratio Estimation (NRE)

Once the surrogate distribution is learned, the authors search for the optimal "shape" (marginalization) of the EEEC that minimizes the uncertainty in inferring $m_t$.

• Parameterization: They define a family of triangular shapes on the sphere parameterized by side ratios $(n_1, n_2)$ and a smearing window $\delta_a$. The shape is defined as $EEEC(\zeta_1, n_1, n_2, m_t)$.
• Neural Ratio Estimation (NRE):
  • NRE is used to solve the inverse problem: inferring the posterior $p(m_t | \text{shape})$.
  • A classifier network is trained to distinguish between samples drawn from the joint distribution $p(\text{shape}, m_t)$ and the product of marginals $p(\text{shape})p(m_t)$.
  • The classifier output approximates the likelihood ratio, allowing the extraction of the posterior distribution for $m_t$ for any given shape.
• Optimization Metric: The authors define an error metric combining bias ($\Delta m^{Dev}_t$) and variance ($\Delta m^{Shape}_t$) derived from the posterior (Maximum A Posteriori and 95% Highest Posterior Density intervals). They perform a grid search over $(n_1, n_2)$ to find the shape that minimizes the combined Root Mean Square (RMS) error.

3. Key Contributions

1. Systematic Observable Optimization: Demonstrated a workflow where ML is used only to explore the space of observables and identify an optimal definition. The final output is a purely analytical observable definition that can be computed to high precision and compared to data without any reference to the ML model.
2. Physics-Informed Loss Functions: Introduced a novel loss function for density estimation that explicitly re-weights samples by their energy. This solved the issue of standard DNNs failing to learn the high-energy tails of distributions relevant for precision measurements.
3. Application to Top Mass: Successfully applied this framework to the 3-point EEEC, a theoretically challenging but calculable observable, to optimize sensitivity to the top-quark mass.
4. Validation: Validated the ML-derived uncertainty estimates against classical polynomial fitting methods, showing that the NRE posterior uncertainties accurately reflect the statistical limitations of the data.

4. Results

• Optimal Shape: The grid search identified an optimal marginalization corresponding to isosceles triangles with a side ratio of approximately $1 : 1 : \sqrt{2}$ (right-angled isosceles triangles).
  • Specifically, the optimal parameters found were $(n_1, n_2) \approx (1.02, 1.99)$.
• Performance:
  • For a true top mass of $m_t = 172.5$ GeV, the optimal shape yielded an inferred mass of $172.47 \pm 0.31$ GeV (using the NRE posterior width).
  • Classical fitting methods on the same data yielded $172.62 \pm 0.28$ GeV, confirming the reliability of the ML-derived uncertainty.
  • The optimal shape showed significantly better sensitivity (lower bias and variance) compared to equilateral triangles (the standard choice in previous literature) or other tested shapes.
• Model Comparison: The DNN with the energy-weighted loss outperformed the Normalizing Flow in reproducing the specific features of the EEEC distribution needed for mass extraction, while being faster to evaluate.

5. Significance

• Bridging Theory and Experiment: This work provides a concrete pathway to discover new precision observables that are both theoretically calculable (allowing for high-order QCD predictions) and experimentally optimal.
• Beyond "Black Box" ML: Unlike typical ML applications in physics where the neural network is the observable (and thus uncalculable), this approach uses ML as a discovery tool to define a new, calculable observable. The final result is transparent and independent of the training data.
• Generalizability: While demonstrated for the top mass and EEECs, the framework is general. It can be applied to any multi-dimensional calculable observable space to optimize sensitivity to couplings, masses, or for anomaly detection, potentially revolutionizing how precision measurements are designed at future colliders.
• Statistical Rigor: The paper establishes a method to quantify the "goodness" of an observable shape using posterior widths, offering a rigorous alternative to ad-hoc shape selection.