⚛️ lattice

Tackling the noisy truncated moment problem with Gaussian Processes: An application to parton distribution functions

This paper introduces a novel Gaussian Process-based Bayesian framework that robustly reconstructs parton distribution functions from noisy truncated moments, utilizing new tools to correct posterior bias and learn correlation lengths, thereby enabling the application of theoretical first-principles data from lattice QCD to PDF recovery.

Original authors: Rohith Karur

Published 2026-09-22

📖 4 min read🧠 Deep dive

CC BY 4.0

Original authors: Rohith Karur

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

Inside every atom's core, protons and neutrons are not solid, indivisible spheres but bustling cities of smaller particles called quarks and gluons. These constituents, known collectively as partons, zip around at nearly the speed of light, carrying the momentum that holds the nucleus together. To understand how these particles behave, physicists rely on a map called a parton distribution function. This map does not show a single location but rather describes the probability of finding a specific parton carrying a certain fraction of the total momentum. The challenge is that these maps are incredibly difficult to draw directly. While experiments can probe these particles by smashing them together, the resulting data is often sparse and noisy, offering only a few scattered clues about the underlying structure.

For decades, scientists have tried to reconstruct these full maps from limited data points, a mathematical challenge known as the truncated moment problem. Imagine trying to guess the shape of a hidden object by only feeling a few of its edges; without enough information, the guess could be wildly wrong. In the world of particle physics, this problem is compounded by the fact that the data itself contains uncertainty. Recent advances in theoretical physics have opened a new door: using supercomputers to calculate these momentum fractions from first principles, a method called lattice quantum chromodynamics. However, these calculations also come with noise and limited data points. The question remains: how can researchers take a handful of noisy numbers and reliably reconstruct the complex, smooth curve that describes the particle's inner life?

In a new study, researchers at the University of California, Berkeley, and Lawrence Berkeley National Laboratory have developed a sophisticated statistical tool to solve this puzzle. They turned to a method called Gaussian Process Regression, a powerful technique from the field of machine learning that is particularly good at making predictions when data is scarce and uncertain. Instead of forcing the data to fit a rigid, pre-determined shape, their approach treats the unknown map as a flexible curve that can bend and twist to match the evidence, while respecting the laws of physics. The team tested their framework on five different types of particle data, ranging from the well-understood "valence" quarks that define a proton's identity to the more chaotic "sea" quarks and gluons that constantly pop in and out of existence.

The researchers found that their new method could robustly reconstruct these particle maps using a moderate number of data points, even when those points were noisy. A key part of their success was the development of new tools to correct for a subtle bias that often creeps into such calculations. When data is imperfect, standard methods can sometimes skew the average result, making the final map look slightly different from reality. The team introduced a technique to systematically remove this skew, ensuring that the reconstructed map centers on the true value. They also created a flexible way to adjust how the map behaves at different scales, allowing the model to learn the right amount of detail needed for each specific type of particle.

The study revealed that not all particles are equally easy to map. The valence quarks, which are more tightly constrained by the rules of physics, were reconstructed with high precision using standard tools. However, the gluons and sea quarks, which are far more chaotic and less constrained, required the researchers to use their more advanced, flexible tools. For these difficult cases, the team had to introduce specific adjustments to handle the wild fluctuations that occur when the particles carry very little momentum. By combining these adjustments with their bias-correction method, they were able to produce reliable maps even for the most unruly datasets.

This work offers a promising path forward for the future of particle physics. As supercomputers generate more and more theoretical data about the inner workings of protons and neutrons, the ability to turn those noisy numbers into clear, accurate maps will become increasingly vital. The framework developed in this study provides a way to extract the most reliable information possible from limited data, bridging the gap between raw theoretical calculations and the physical reality of the subatomic world. By refining how scientists interpret these fundamental building blocks, the research paves the way for a deeper understanding of the strong nuclear force that binds our universe together.

Technical Summary: Tackling the Noisy Truncated Moment Problem with Gaussian Processes

Problem Statement
This work addresses the noisy truncated Hausdorff moment problem, specifically in the context of recovering Parton Distribution Functions (PDFs) from their moments. In Quantum Chromodynamics (QCD), PDFs describe the momentum fraction $x$ carried by partons (quarks and gluons) within a hadron. While Lattice QCD (LQCD) and experimental data increasingly provide the first few moments of these distributions ( $\langle x^n \rangle$ ), reconstructing the continuous function $f(x)$ from a finite, noisy set of moments is an ill-posed inverse problem.

The challenge is twofold:

Truncation: A finite number of moments does not uniquely determine a function with compact support on $[0,1]$ .
Noise: Experimental and LQCD data contain uncertainties. Standard reconstruction methods often fail or become unstable when noise is introduced, as the solution space explodes combinatorially. Furthermore, PDFs exhibit power-law divergences ( $x^\alpha$ ) at low $x$ , which complicates convergence and requires specific handling of singularities.

Methodology
The author proposes a Bayesian framework centered on Gaussian Process Regression (GPR) to solve this problem. The core hypothesis is that the unknown PDF $f(x)$ can be modeled as a Gaussian Process characterized by a mean function $m(x)$ and a covariance kernel $K(x, x')$ .

Key methodological components include:

Analytical Posterior Formulation: The author derives the posterior distribution $p(f|y, \Theta)$ , where $y$ represents the observed moments with covariance $C_y$ , and $\Theta$ represents hyperparameters. The posterior mean and covariance are derived analytically using the Woodbury matrix identity, treating the moment integral operator as a linear constraint (or linearized for non-linear cases).
Hyperparameter Sampling: Instead of finding a single Maximum A Posteriori (MAP) estimator, the framework samples the hyperparameter space (including kernel variance $\sigma$ , correlation length $\ell$ , and power divergence exponent $\alpha$ ) using Hamiltonian Monte Carlo (HMC), specifically the No-U-Turn Sampler (NUTS). This allows for a full characterization of the posterior uncertainty.
Novel Tools for Bias and Flexibility:
- Bias Correction: The author identifies that data noise induces a mean bias in the posterior. They introduce a technique to correct this by progressively deflating the data covariance matrix ( $C_y \to C_y/\lambda$ ) and extrapolating to the zero-noise limit to obtain an unbiased mean posterior.
- Adaptive Kernels: Standard stationary kernels (e.g., Squared Exponential, Matérn) are found insufficient for the multiscale nature of PDFs. The author introduces Gibbs-style kernels where the correlation length $\ell(x)$ is a learnable function of $x$ (parameterized by polynomials or specific forms like a "cosh dip"). This allows the model to automatically adapt to different correlation scales in low- $x$ versus high- $x$ regions.
- Power Divergence Priors: To handle the $x^\alpha$ behavior at low $x$ , the kernel is modified as $K(x, x') \to x^\alpha K(x, x') (x')^\alpha$ , with $\alpha$ treated as a sampled hyperparameter.

Key Results
The framework was tested on five datasets derived from phenomenological models (NNPDF30 and FANTO10) covering valence quarks, gluons, and sea quarks for both nucleons and pions. These datasets varied in constraint levels (from well-constrained valence PDFs to severely unconstrained gluon/sea PDFs) and noise levels.

Valence Quark PDFs: For constrained valence datasets, the GPR framework robustly reconstructs PDFs using a moderate number of moments.
- Kernel Performance: Short correlation scale kernels (e.g., Matérn $\nu=1/2$ ) failed to converge to the true variance, while long correlation scale kernels (e.g., Log-SE) missed fine details. Medium correlation scale kernels (e.g., Matérn $\nu=3/2$ ) and Gibbs-style kernels provided the best balance, capturing both global trends and local features.
- Bias Correction: The bias correction technique was shown to be essential for datasets with high noise (e.g., FANTO10), effectively eliminating systematic shifts in the reconstructed mean.
Gluon and Sea Quark PDFs: These datasets are "severely unconstrained" with large uncertainties and strong power divergences at low $x$ .
- Standard Kernels: Out-of-the-box kernels failed, often producing unphysical oscillatory solutions or failing to capture the low- $x$ divergence.
- Specialized Strategies: Successful reconstruction required a combination of:
  - Logarithmic Grids: To resolve the $x \ll 1$ region.
  - Power Divergence Priors: Enforcing the $x^\alpha$ behavior.
  - Gibbs "Cosh Dip" Kernels: A specific kernel form that enforces a "dip" in correlation length at low $x$ , reflecting the distinct physics of the low- $x$ regime.
  - Hyperparameter Restrictions: Constraining the sampling ranges of $\sigma$ and $\alpha$ was necessary to prevent the sampler from converging to unphysical solutions.

Significance and Claims
The paper claims that the proposed GPR framework offers a robust, model-agnostic approach to the truncated moment problem that is superior to conventional methods when dealing with noisy, sparse data.

Robustness: The framework successfully reconstructs PDFs across a spectrum of divergence degrees ( $x^\alpha$ ) and noise levels, where standard parametric fits or non-Bayesian methods struggle.
Resolution of Persistent Issues: The novel tools introduced—specifically the bias correction method and the learnable Gibbs kernels—resolve systematic issues (such as mean bias and inability to capture multiscale correlations) that are difficult to address via conventional means.
Application to LQCD: The author states that this framework is directly applicable to moments obtained from Lattice QCD. As LQCD calculations improve in volume and quality (e.g., via gradient flow techniques), this method provides a pathway to reconstruct PDFs from first principles, complementing existing experimental and theoretical determinations.

The work concludes that while trade-offs exist in parameter selection, the GPR framework is a powerful tool for extracting physical distributions from limited moment data, with potential applications extending beyond particle physics to any field facing the finite moment problem.

More like this