Gaussian Processes for Inferring Parton Distributions

This paper demonstrates that Gaussian Process Regression provides a robust, non-parametric Bayesian framework for reconstructing parton distribution functions from lattice QCD data, offering controlled uncertainty estimates and reduced model bias compared to traditional functional forms.

The Gist

The Cosmic Jigsaw Puzzle: How We Map the Inside of a Proton

Imagine you are trying to draw a detailed map of a bustling, crowded city—let’s call it "Proton City"—but there’s a catch: you aren't allowed to actually go there.

Instead, you are standing on a distant hilltop, and all you can see are the blurry, flickering lights of the cars moving through the streets. You can see how many lights pass by every minute, and you can see how bright the clusters are, but you can’t see the streets, the buildings, or the individual drivers.

Your goal is to reconstruct the exact layout of the city (the Parton Distribution Functions, or PDFs) using only those blurry, flickering light patterns (the Lattice QCD data).

This paper describes a new, high-tech way to solve this "blurry light" problem using a mathematical tool called Gaussian Process Regression (GPR).

---

The Problem: The "Ill-Posed" Mystery

In science, we call this an "ill-posed problem." This is a fancy way of saying there are too many ways to be wrong.

If you see a bright flash of light, was it one giant truck, or was it ten small motorcycles passing at once? Without more information, both answers are technically "correct" based on what you saw. If you just guess, you might draw a city with streets that don't exist, or miss entire neighborhoods. In physics, if we guess wrong, our understanding of how the universe is held together falls apart.

The Solution: The "Smart Sketch Artist" (Gaussian Processes)

To solve this, the researchers used Gaussian Processes. Think of GPR not as a rigid stencil, but as a highly intuitive sketch artist.

If you show a traditional computer program a few dots, it tries to connect them with a stiff, straight ruler. If the dots don't line up perfectly, the ruler breaks or creates a jagged, ugly shape.

The GPR Sketch Artist, however, has "intuition" (which scientists call a Prior). Before the artist even sees the data, they have a sense of what a city should look like:

• "Streets are usually smooth, not zig-zagging like lightning bolts."
• "Traffic usually flows in patterns, not random explosions."

This "intuition" allows the artist to fill in the gaps. When the data is blurry or missing (like the dark outskirts of the city), the artist doesn't just make up a fake street; they draw a faded, shaky line to say, "I think a street is here, but I'm not very sure." This is what scientists call "controlled uncertainty."

The Experiment: The "Stress Test"

The researchers didn't just jump into real, messy data. They performed a "Closure Test."

They created a "Fake City" where they already knew exactly where every street and car was. Then, they intentionally blurred the view and gave it to their GPR Sketch Artist.

• Did the artist find the real streets? Yes.
• Did the artist admit when they were guessing? Yes.

They also played "What If?" They gave the artist different types of intuition (different Kernels). One artist might think cities are very smooth; another might think they are more complex. They used a technique called "Model Averaging"—essentially asking a whole room of different artists to draw the city and then combining their drawings into one "Master Map" that accounts for everyone's different opinions.

Why Does This Matter?

Protons are the building blocks of everything. Understanding the "traffic patterns" of the quarks and gluons inside them is fundamental to understanding the very fabric of reality.

By using this GPR method, scientists can now look at the "blurry lights" from supercomputers (Lattice QCD) and create much more reliable, honest, and accurate maps of the subatomic world. They aren't just guessing; they are providing a map that tells you exactly where they are certain and exactly where they are still in the dark.

Technical Summary

Technical Summary: Gaussian Processes for Inferring Parton Distributions

Title: Gaussian Processes for Inferring Parton Distributions
Authors: Yamil Cahuana Medrano, Hervé Dutrieux, Joseph Karpie, Kostas Orginos, Savvas Zafeiropoulos (HadStruc Collaboration)
Target Journal: JHEP

---

1. The Problem: The Ill-Posed Inverse Problem in Lattice QCD

The extraction of Parton Distribution Functions (PDFs)—which describe the longitudinal momentum fraction ($x$) of quarks and gluons within a hadron—from Lattice Quantum Chromodynamics (LQCD) data is a fundamentally ill-posed inverse problem.

LQCD provides matrix elements in coordinate space (or Ioffe time $\nu$ space), which are related to PDFs via integral transforms (e.g., Fourier transforms for pseudo-PDFs). Because lattice data is discrete, finite in range, and contains statistical noise, there are infinitely many continuous functions that can satisfy the data. Traditional methods often rely on rigid parametric functional forms (e.g., $Nx^\alpha(1-x)^\beta$), which introduce model bias, or neural networks, which can be difficult to interpret regarding uncertainty. The challenge is to find a regularization method that is flexible enough to avoid bias while providing controlled, reliable uncertainty estimates.

2. Methodology: Gaussian Process Regression (GPR)

The authors propose a non-parametric Bayesian framework using Gaussian Process Regression (GPR) to reconstruct PDFs.

• Bayesian Framework: Instead of assuming a specific functional form, the authors use a Gaussian Process as a prior distribution $P[q|I]$. This prior encodes smoothness, correlations, and physical constraints without restricting the solution to a low-dimensional parameter space.
• Kernel Engineering: The core of GPR is the covariance kernel $K(x, x')$, which defines how the function correlates at different points. The paper explores several sophisticated kernel types:
  • Radial Basis Functions (RBF): Standard smooth kernels.
  • Gibbs/CDGP Kernels: Non-stationary kernels where the correlation length $l(x)$ varies, allowing for decorrelation in the low-$x$ regime where data is sparse.
  • Log-RBF/Polylogarithmic Kernels: Designed via the "kernel trick" to allow for divergent behavior at $x \to 0$, mimicking the expected physical behavior of PDFs.
  • Changepoints: Smoothly transitioning between different kernels to handle different physical regimes (e.g., DGLAP vs. ERBL).
• Constraints: Physical requirements, such as the PDF vanishing at $x=1$ or normalization sum rules, are incorporated into the prior using the Woodbury identity to modify the kernel.
• Hyperparameter Optimization & Model Averaging: The authors move beyond fixing hyperparameters ($\theta$) by using Importance Sampling Monte Carlo to integrate over the hyperparameter space. To handle the uncertainty arising from the choice of different kernels/models, they implement Bayesian Model Averaging (BMA), weighting different models using Information Criteria (PAIC, BTIC, and BAIC).

3. Key Contributions

• Systematic Non-Parametric Framework: Establishes GPR as a robust alternative to parametric fits, reducing model dependence.
• Information Quantification: Introduces the use of Kullback–Leibler (KL) divergence to provide a point-by-point metric of how much information the data actually provides versus the prior.
• Comprehensive Kernel Study: Provides a rigorous mathematical derivation and testing of various kernels, specifically identifying "pathological" kernels (like certain Jacobi or Rank-1 polynomial kernels) that fail to provide reliable error estimates.
• Closure Tests: Uses synthetic data (generated from NNPDF4.0) to prove that the method can accurately reconstruct the "true" function and its variance.

4. Results

• Synthetic Data Tests: The GPR method successfully reproduces the central values of the NNPDF4.0 distributions. Crucially, the error bands are consistent with (or slightly more conservative than) the true variance, ensuring that the method does not underestimate uncertainty.
• Information Gain: KL divergence analysis shows that information gain is highest in the high-$x$ and low-$\nu$ regions. It also demonstrates that even with limited Ioffe time ($\nu_{max}$), GPR can capture significant information, though uncertainty naturally grows in the unconstrained low-$x$ regime.
• Model Averaging Robustness: The Bayesian Model Averaging results are remarkably consistent across different information criteria (PAIC, BTIC, BAIC). This demonstrates that the "scientist-made choices" (kernel selection) do not lead to wildly different physical conclusions when handled via a proper Bayesian weighting.
• Lattice Data Application: When applied to real lattice QCD data (isovector proton PDF), the method provides a reliable reconstruction that accounts for the specific noise and correlation structures of the lattice ensemble.

5. Significance

This work provides a high-level statistical toolkit for the hadron structure community. By moving from "fitting a function" to "inferring a distribution," the authors offer a path toward quantifiable and improvable uncertainty. The ability to use KL divergence to map "information density" allows lattice theorists to understand exactly how much more momentum/Ioffe time data is required to resolve specific kinematic regions, making it a vital tool for the design of future lattice computations.