Compton Form Factor Extraction using Quantum Deep Neural Networks

This paper demonstrates that quantum-inspired deep neural networks (QDNNs) offer superior predictive accuracy and tighter uncertainties compared to classical methods for extracting Compton form factors from JLab experimental data, establishing them as an efficient tool for future multidimensional studies of hadronic structure.

The Gist

Imagine the inside of a proton (a tiny particle inside an atom) as a bustling, three-dimensional city. Physicists want to map this city: they want to know where the "citizens" (quarks and gluons) are, how fast they are moving, and how they are arranged in space. This map is called a Generalized Parton Distribution (GPD).

However, you can't take a photograph of this city directly. Instead, scientists shoot high-energy electrons at protons (like throwing a ball at a moving target) and watch how the light scatters. This is called Deeply Virtual Compton Scattering (DVCS). The data they get is like a blurry, noisy shadow of the city. To turn that shadow into a clear map, they have to solve a very difficult math puzzle called "deconvolution."

The "ingredients" needed to solve this puzzle are called Compton Form Factors (CFFs). Think of CFFs as the secret recipe numbers that, when plugged into the physics equations, recreate the shadow the scientists see.

The Problem: The Shadow is Fuzzy

For years, scientists have used standard computer programs (Classical Deep Neural Networks, or CDNNs) to guess these recipe numbers. It's like trying to tune a radio to find a clear station. Sometimes the signal is clear, but often it's full of static (noise) and the station is hard to find, especially in areas where the data is sparse or the signal is weak.

The New Idea: A Quantum-Inspired Radio

The authors of this paper asked: What if we used a different kind of tuner? They tried using Quantum Deep Neural Networks (QDNNs).

Don't worry, they didn't use a real quantum computer (which is currently very fragile and noisy). Instead, they built a simulator on a regular supercomputer that acts like a quantum computer.

• The Analogy: Imagine a classical computer is like a standard flashlight. It shines a beam of light in a straight line. A quantum-inspired computer is like a flashlight that can also split its beam into many different colors and angles simultaneously, allowing it to "see" patterns in the dark that a straight beam misses.
• The Mechanism: The QDNN uses "entanglement" (a quantum concept where parts of a system are linked in a way classical parts aren't) to find hidden connections in the noisy data that the classical computer might miss.

What They Did

1. The Test Drive (Pseudodata): Before trying this on real data, they created a "fake" universe. They invented the true recipe numbers (CFFs) and then generated fake experimental data with known errors. This is like a flight simulator: they knew exactly where the plane should be, so they could test if their new navigation system (QDNN) was better than the old one (CDNN).
2. The Race: They ran both the Classical and Quantum models against this fake data.
   • Result: The Quantum model (QDNN) was often more accurate and gave much tighter, more precise results. It was better at ignoring the "static" and finding the true signal.
3. The "Traffic Light" (The Qualifier): They realized that the Quantum model isn't always the winner. Sometimes the Classical model is better. So, they created a simple "traffic light" metric (called the DVCS Quantum Qualifier).
   • This tool looks at the data and asks: "Is this data noisy and complex?"
   • If Yes: It turns the light green for the Quantum model.
   • If No: It turns the light green for the Classical model.
   • This ensures they always use the best tool for the specific job.

The Real-World Test

They took this "smart traffic light" system and applied it to real data from the Jefferson Lab (a major physics lab in Virginia).

• They analyzed thousands of data points.
• For about 60% of the data, the Quantum model was the clear winner, providing a much clearer map of the proton's interior.
• For the rest, they used the Classical model.
• They combined all these best guesses into a single, global map.

The Conclusion

The paper claims that by using these "Quantum-inspired" tools, they were able to extract the recipe numbers (CFFs) with less uncertainty (a clearer picture) than previous methods.

• Key Takeaway: The Quantum approach didn't just give a slightly better answer; it acted as a "self-correcting" mechanism that stabilized the results, especially in the messy, noisy parts of the data where classical methods usually struggle.
• Future: They say this method is ready to be used on real quantum computers once those machines mature, but for now, the simulation proves the concept works.

In short: They built a smarter, more flexible way to decode the blurry shadows of subatomic particles, resulting in a sharper, more detailed map of the proton's internal structure.

Technical Summary

1. Problem Statement

The extraction of Compton Form Factors (CFFs) from Deeply Virtual Compton Scattering (DVCS) data is a critical step in mapping the 3D structure of hadrons (Generalized Parton Distributions or GPDs). However, this process faces significant challenges:

• Ill-posed Deconvolution: CFFs are convolutions of GPDs with calculable coefficients. Recovering CFFs from experimental cross-section data is an ill-posed inverse problem.
• Model Bias: Traditional "global fits" rely on specific functional ansätze (e.g., VGG, GK models) which introduce theoretical bias. "Local fits" avoid this but suffer from limited predictive power and large systematic uncertainties due to sparse data.
• Data Sparsity and Noise: DVCS measurements, particularly at Jefferson Lab (JLab), often involve sparse kinematic bins and significant experimental uncertainties (statistical and systematic).
• Limitations of Classical ML: While classical Deep Neural Networks (CDNNs) have been used to reduce bias, they can still struggle with overfitting in high-noise, sparse regimes and may lack the capacity to capture complex interference patterns inherent in quantum scattering amplitudes.

The authors investigate whether Quantum Deep Neural Networks (QDNNs)—specifically hybrid quantum-classical architectures—can offer superior performance in extracting CFFs compared to classical counterparts.

2. Methodology

A. Theoretical Framework

The study utilizes the twist-2 Belitsky–Kirchner–Müller (BKM10) formalism.

• Forward Model: The cross-section is modeled as a coherent sum of the Bethe–Heitler (BH) amplitude (known exactly) and the DVCS amplitude.
• Parameters: The fit extracts four effective, kinematic-dependent quantities: the real parts of three CFFs ($\Re e H, \Re e E, \Re e \tilde{H}$) and a $\phi$-independent DVCS-squared term.
• Loss Function: A physics-informed loss function minimizes the $\chi^2$ between the predicted cross-section (derived from the BKM10 formalism using the network's output) and the experimental/pseudodata cross-section.

B. Neural Network Architectures

The authors compare two models trained on identical physics-informed losses:

1. Classical DNN (CDNN): A standard feedforward network with 8 hidden layers (64 neurons each) and ReLU activations.
2. Quantum DNN (QDNN): A hybrid architecture implemented using PennyLane on a state-vector simulator.
   • Encoding: Classical inputs ($x_B, Q^2, t$) are embedded into a 6-qubit Hilbert space via angle embedding.
   • Circuit: The core consists of Strongly Entangling Layers (SELs).
   • Readout: Pauli-Z expectation values are measured and passed to a classical post-processing head to output the four CFFs.
   • Optimization: The QDNN uses the parameter-shift rule for gradient computation.

C. Training and Validation Strategy

• Pseudodata Generation: To perform "closure tests" (where the ground truth is known), the authors generated 1,000 noisy replicas of pseudodata based on JLab experimental kinematics and a known generating function for CFFs.
• Error Metrics: Performance was evaluated using:
  • Accuracy: Deviation from the true CFF values.
  • Precision: Variance across replicas.
  • Algorithmic Error: Uncertainty inherent to the optimization algorithm.
  • Methodological Error: Uncertainty arising from model assumptions.
• Quantum Qualifier ($\hat{\Xi}_{DVCS}$): A novel, data-driven metric was developed to predict a priori whether a QDNN or CDNN would perform better for a specific kinematic bin. It relies on two features:
  1. Nonlinearity ($N$): The deviation of the cross-section dependence on the azimuthal angle $\phi$ from a linear trend.
  2. Scaled Experimental Error ($\epsilon_s$): The magnitude of experimental uncertainty relative to the signal.

D. Experimental Application

The methodology was applied to real JLab data (Hall A and Hall B). The Quantum Qualifier was used to select the optimal model (QDNN or CDNN) for each local kinematic bin. These local extractions were then aggregated to train a global DNN parameterization.

3. Key Contributions

1. First QDNN Extraction of CFFs: This is the first application of quantum-inspired deep learning to extract Compton Form Factors from DVCS data.
2. Quantum Qualifier Metric: The introduction of $\hat{\Xi}_{DVCS}$ provides a practical, heuristic tool to determine when quantum architectures are advantageous. The study found that QDNNs excel in regions with high experimental noise and high nonlinearity in the $\phi$-dependence.
3. Optimization of Quantum Circuits: The paper identifies and mitigates specific quantum training challenges, such as barren plateaus (vanishing gradients). They implemented:
   • Layer-wise depth growth.
   • Depth-scaled initialization.
   • Tunable entanglement ranges (nearest-neighbor vs. full).
4. Like-for-Like Benchmarking: The study ensures a fair comparison by keeping the forward model, loss function, and input/output dimensions identical between CDNN and QDNN, isolating the performance difference to the architecture itself.

4. Key Results

• Performance on Pseudodata:
  • The Full QDNN (optimized) significantly outperformed the CDNN in precision (narrower uncertainty bands) and accuracy for $\Re e E$ and $\Re e \tilde{H}$.
  • The QDNN reduced the global $\chi^2$ metric ($M_{\chi^2}$) by approximately 64% compared to the baseline CDNN on pseudodata.
  • While the initial "Basic QDNN" suffered from higher algorithmic error due to optimization difficulties, the "Full QDNN" (with mitigations) achieved algorithmic error levels comparable to the CDNN while retaining superior precision.
• Experimental Data Application:
  • The Quantum Qualifier predicted that the QDNN would outperform the CDNN in ~60% of the experimental kinematic bins.
  • The resulting global fit showed substantially reduced uncertainties (tighter error bands) compared to previous extractions (e.g., $\chi$DNN, KM15).
  • The extracted CFFs were consistent with the KM15 global analysis for $\Re e E$ and $\Re e \tilde{H}$, but showed a systematic difference in the $t$-dependence of $\Re e H$, suggesting the QDNN captures latent correlations differently.
• Robustness: The QDNN demonstrated superior stability in sparse, high-noise regions where classical networks tended to overfit statistical fluctuations.

5. Significance and Conclusion

• New Paradigm for Hadronic Physics: The study establishes QDNNs as a viable, complementary tool to classical ML for extracting hadronic structure parameters. It suggests that the implicit regularization provided by the unitary structure of quantum circuits helps prevent overfitting in data-scarce regimes.
• Efficiency: The QDNN achieves higher precision with fewer trainable parameters (876 vs. 33,796 for the CDNN) by leveraging the exponential feature space of the Hilbert space.
• Future Outlook: The framework is scalable to full 8-CFF extractions and can incorporate polarization observables. While current results use simulators, the architecture is hardware-compatible, paving the way for execution on future Noisy Intermediate-Scale Quantum (NISQ) devices.
• Practical Utility: The Quantum Qualifier offers a concrete strategy for experimentalists to optimize their analysis pipelines, dynamically selecting the best algorithm based on data quality and kinematic complexity.

In summary, the paper demonstrates that quantum-inspired neural networks, when properly optimized and guided by data-driven selection metrics, can extract Compton Form Factors with higher precision and reduced model bias than classical deep learning methods, particularly in the challenging, noisy environment of DVCS experiments.