Inverse problem in the LaMET framework

This paper demonstrates that the inverse problem in the LaMET framework, arising from noisy lattice data at intermediate harmonics ($\lambda \sim 5\text{--}15$), introduces significant uncertainties that cannot be resolved by simple asymptotic assumptions, thereby necessitating more sophisticated techniques and correcting the misconception that LaMET offers a direct $x$-dependence computation compared to short-distance factorization.

The Gist

The Big Picture: Trying to Reconstruct a Symphony from a Broken Recording

Imagine you are a music producer trying to reconstruct a full, beautiful symphony (the Parton Distribution Function, or PDF) that describes how the tiny particles inside a proton move.

You don't have the original recording. Instead, you have a very noisy, incomplete recording of just the first few seconds of the music (the Lattice QCD data).

The scientists in this paper are trying to figure out how to take that broken, noisy snippet and mathematically "fill in the blanks" to hear the whole song. They are testing a specific method called LaMET (Large Momentum Effective Theory).

The Problem: The "Missing Notes"

In the world of particle physics, to see the inside of a proton clearly, you need to look at it from a very high speed (high momentum). However, in computer simulations (Lattice QCD), looking at these high speeds is like trying to hear a whisper in a hurricane. The signal gets lost in the noise very quickly.

• The Data Gap: The computer simulations can only give us reliable data for the first few "notes" (Fourier harmonics). After that, the data becomes so noisy it's useless.
• The Inverse Problem: We have the beginning of the song, but we need to guess the rest of the song to understand the whole melody. This is called an "inverse problem." It's like trying to guess the ending of a movie just by watching the first 10 minutes, but the first 10 minutes are also slightly blurry.

The Common Mistake: The "Rigid Tail" Assumption

Many researchers have tried to solve this by saying, "We know how the song ends! It fades out smoothly like a exponential decay (a gentle slide into silence)."

They assume the missing notes follow a strict, predictable pattern (like a slide whistle going down). They fit their data to this pattern and claim, "Look, we have the whole song!"

The authors of this paper say: "Wait a minute."

They argue that:

1. We don't actually know how it fades out. The data gets too noisy before we can see the "fade out" clearly.
2. The assumption is dangerous. If you force the data to fit a specific "fade out" shape, you might be inventing a fake ending that doesn't match reality.
3. It doesn't matter as much as you think. Surprisingly, they found that how the song fades out at the very end (the asymptotic behavior) doesn't actually change the middle of the song (the part of the proton's structure we care about most) very much.

The Experiment: Trying Different "Guessing" Strategies

The authors ran a massive simulation using real data from a supercomputer. They tried to reconstruct the proton's structure using three different "guessing" strategies:

1. The "Rigid" Guess: Assuming the missing notes follow a strict mathematical rule (exponential decay).
2. The "Flexible" Guess (Gaussian Processes): Using a smart, flexible AI-like method that says, "I don't know the exact rule, but I know the notes shouldn't jump around wildly. Let me find the smoothest path that fits the data we have."
3. The "Backus-Gilbert" Method: A traditional mathematical trick that often gets confused and produces weird, jagged results when the data is noisy.

The Results:

• The "Rigid" guess and the "Flexible" guess produced very different results for the middle of the song (the part of the proton's structure at moderate values).
• The "Rigid" guess made the scientists feel very confident, but that confidence was a false sense of security. Because they forced the data to fit a rule, they underestimated how much they were actually guessing.
• The "Flexible" guess showed that the uncertainty (the "fuzziness" of the answer) is actually much bigger than people thought.

The Big Misconception: "Direct" vs. "Indirect"

There is a popular belief in the physics community that:

• LaMET allows you to calculate the proton's structure directly (like taking a high-res photo).
• Other methods only give you indirect clues (like guessing the photo from a blurry shadow).

The authors say this is a lie.

They explain that because the data is so noisy and incomplete, LaMET is also indirect. You cannot take a "direct photo" of the proton's structure at a specific point. You are still forced to make assumptions and smooth out the data. Whether you use LaMET or other methods, you are essentially trying to reconstruct a blurry image from a few pixels.

The Takeaway: Be Humble About the Unknown

The paper concludes with a call for honesty in science:

• Stop pretending we know the end of the story. The data we have right now ends before the "fade out" begins. We are guessing the ending.
• Don't trust the "Rigid" models. Assuming the missing data follows a perfect exponential curve hides the true uncertainty.
• We need better tools. We need more sophisticated math (like the flexible "Gaussian Process" methods they tested) to honestly admit how much we don't know.

In short: The scientists are saying, "We are trying to solve a puzzle with half the pieces missing. We can't just glue the missing pieces together based on what we think they should look like. We need to admit that the picture is still fuzzy, and we need better tools to measure just how fuzzy it is."

Technical Summary

1. Problem Statement

The paper addresses a critical challenge in calculating Parton Distribution Functions (PDFs) from first principles using Large Momentum Effective Theory (LaMET) and Short-Distance Factorization (SDF).

• The Core Issue: LaMET reconstructs light-cone PDFs ($q(x)$) by taking the Fourier transform of quasi-PDF matrix elements ($\langle P|O(z)|P\rangle$) computed on the lattice. However, lattice QCD data is only available over a finite range of spatial separations ($z$) due to exponential signal loss and noise at large $z$.
• The Inverse Problem: Reconstructing the PDF from a truncated, noisy Fourier dataset is an ill-posed inverse problem. The missing high-frequency harmonics (large $z$) must be extrapolated.
• Current Limitations:
  • Current lattice simulations typically lose signal at Fourier harmonics $\lambda = zP_z \sim 5-15$.
  • The asymptotic exponential decay regime (where the signal is dominated by the lightest meson) is often not reached before the noise overwhelms the signal.
  • Existing methods often rely on rigid, few-parameter exponential models to extrapolate the missing data. The authors argue this leads to underestimated uncertainties and model bias, particularly because the "transition region" ($\lambda \sim 5-15$) is where the signal is weakest but the extrapolation begins.
• Misconception: The paper challenges the prevailing view that LaMET allows for a "direct" calculation of the $x$-dependence of PDFs, whereas SDF only provides moments. The authors argue that both frameworks face mathematically similar inverse problems regarding the reconstruction of $x$-dependence from limited Fourier data.

2. Methodology

The authors utilize a realistic lattice dataset (proton unpolarized isovector PDF) with a pion mass of 358 MeV and momentum $P_z = 2$ GeV. They employ Gaussian Process Regression (GPR) as a flexible, non-parametric framework to explore the uncertainty of the inverse problem without relying on rigid parametric assumptions.

Key methodological steps include:

• Non-Parametric Reconstruction: They compare the standard Backus-Gilbert (BG) method with Gaussian Process Regression (GPR).
  • BG Method: They highlight that traditional BG reconstruction suffers from bias and irregular uncertainty behavior (e.g., artificial tightening of errors) unless heavily preconditioned with parametric models, which reintroduces model dependence.
  • GPR: They use kernels in $x$-space (Radial Basis Function and Logarithmic RBF) to enforce smoothness and correlation length, allowing for a more principled uncertainty quantification.
• Testing Asymptotic Constraints: They test the sensitivity of the $x$-reconstruction to different assumptions about the tail behavior (large $\lambda$):
  1. No explicit constraint: Pure GPR extrapolation.
  2. Exponential decay: Enforcing a decay rate based on meson masses ($r \approx 0.6 - 1.0$ fm).
  3. Gaussian decay: Enforcing a decay based on the proton radius.
  4. Rigid Parametric Model: Using a simple $A e^{-b\lambda}$ fit.
• Comparison: They systematically vary the kernel type, decay rates, and regularization parameters to map the "envelope" of possible reconstructions, rather than relying on a single best-fit curve.

3. Key Contributions

1. Identification of the Critical Transition Region: The authors demonstrate that the uncertainty in the PDF reconstruction is dominated by the intermediate Fourier harmonics ($\lambda \approx 8 - 15$), not the asymptotic tail. This is the region where lattice data is noisy and the asymptotic exponential regime has not yet been established.
2. Debunking the "Direct" Calculation Myth: They prove that the claim "LaMET gives direct $x$-dependence while SDF does not" is inaccurate. Both frameworks are limited by the same Fourier truncation. The inability to resolve sharp features in the light-cone PDF is due to the smearing kernel width ($\sim \Lambda_{QCD}/P_z$) and missing large-$z$ information, a limitation shared by both approaches.
3. Insignificance of Exact Asymptotic Behavior: The study shows that for the phenomenologically relevant range ($x > 0.2$), the specific choice of the asymptotic decay rate (exponential vs. Gaussian, or the exact value of the decay radius) has a minimal impact on the central value and uncertainty of the reconstructed PDF.
4. Superiority of Flexible Priors: They show that rigid parametric extrapolations (common in literature) fail to capture the true uncertainty. GPR with flexible priors reveals that the uncertainty is significantly larger than previously estimated when the transition region is not well-constrained.

4. Key Results

• Uncertainty Dominance: The variation in the reconstructed PDF is driven by the choice of the $x$-space kernel (e.g., RBF vs. Log-RBF) and the treatment of the intermediate $\lambda$ region, rather than the specific form of the asymptotic tail.
• Bias in Rigid Models: Simple parametric fits (e.g., $A e^{-0.12\lambda}$) can produce results that are in $1\sigma$ tension with more flexible GPR reconstructions, leading to a false sense of precision.
• Imaginary Component: The paper analyzes the imaginary part of the matrix element (which is less convergent than the real part). Even with highly non-convergent data, GPR with physical priors can provide a stable reconstruction, demonstrating that discarding this data (as some suggest) is a waste of resources.
• Upper Bound Validation: The authors compare their results to a "rigorous upper bound" derived in previous literature (Ref. [23]). They find that while individual rigid extractions often underestimate uncertainty, the envelope of all reasonable reconstructions aligns well with this upper bound, validating the existence of a significant, unavoidable uncertainty floor.
• Smearing Limit: The results confirm that LaMET effectively filters the light-cone PDF through a smearing kernel of width $\sim \Lambda_{QCD}/P_z$. Therefore, reconstructing sharp features at a specific $x$ is impossible without infinite momentum or infinite $z$-range.

5. Significance and Conclusion

• Paradigm Shift in Uncertainty Quantification: The paper argues that the community must move away from "best-fit" parametric extrapolations toward methods that explicitly account for the ill-posed nature of the inverse problem. The uncertainty in the transition region ($\lambda \sim 5-15$) is the bottleneck, not the asymptotic tail.
• Call for Sophisticated Techniques: The authors stress the need for more sophisticated techniques (like GPR with varied priors) to properly quantify uncertainties in both LaMET and SDF.
• Conservative Conclusions: The study uses a heavier-than-physical pion mass (358 MeV), which actually improves the signal-to-noise ratio compared to physical masses. The authors conclude that if the inverse problem is significant under these "favorable" conditions, it will be even more severe at physical pion masses.
• Future Directions: The paper calls for a serious assessment of uncertainty propagation in the transition region and suggests that future lattice calculations must extend to higher $\lambda$ (beyond $\sim 15$) to truly resolve the inverse problem, or accept that the PDF reconstruction will always be limited by the smearing inherent in the finite momentum and finite $z$ data.

In summary, this work provides a rigorous mathematical and numerical demonstration that the inverse problem is the dominant source of uncertainty in current LaMET calculations, and that the precise asymptotic behavior of the tail is less critical than the treatment of the noisy, intermediate data region.