Learning of Statistical Field Theories

Shreya Shukla, Abhijith Jayakumar, Andrey Y. LokhovarXiv 2025 · DOI · arXiv

Abstract

Recovering microscopic couplings directly from data provides a route to solving the inverse problem in statistical field theories, one that complements the traditional—often computationally intractable—forward approach of predicting observables from an action or Hamiltonian. Here, we propose an approach for the inverse problem that uniformly accommodates systems with discrete, continuous, and hybrid variables. We demonstrate accurate parameter recovery in several benchmark systems—including Wegner's Ising gauge theory, $\phi^4$ theory, Schwinger and Sine-Gordon models, and mixed spin-gauge systems, and show how iterating the procedure under coarse-graining reconstructs full non-perturbative renormalization-group flows. This gives direct access to phase boundaries, fixed points, and emergent interactions without relying on perturbation theory. We also address a realistic setting where full gauge configurations may be unavailable, and reformulate learning algorithms for multiple field theories so that they are recovered directly using observables such as correlations from scattering data or quantum simulators. We anticipate that our methodology will find widespread use in practical learning of field theories in strongly coupled regimes where analytical tools might fail.

Introduction

In theoretical physics, the standard paradigm is the "forward problem": starting with a mathematical description of energy (an action or Hamiltonian) and calculating observables like magnetism or particle scattering. However, in "strongly coupled" regimes—where interactions are intense—this calculation often becomes computationally intractable.

Conversely, the "inverse problem"—recovering model parameters from experimental data—is frequently even more challenging. Most existing methods require "full field configurations" (complete snapshots of every particle or spin), which real-world experiments like neutron scattering or quantum simulators cannot provide. Furthermore, the diversity of physical systems, ranging from discrete spins to continuous fields, has historically hindered the development of a unified learning algorithm.

Conceptual overview of the inverse problem and statistical efficiency of parameter recovery.

A Unified Framework for the Inverse Problem

To bridge this gap, we propose a unified framework that handles discrete, continuous, and hybrid variables by leveraging two key statistical techniques:

  • Pseudolikelihood (PL) for Discrete Systems: Instead of calculating the impossible-to-compute partition function, PL focuses on the conditional probability of a single variable given its neighbors.
  • Score Matching (SM) for Continuous Fields: This technique matches the "score" (the gradient of the log-probability density) of the model to the data, bypassing the need for normalization constants.

As shown in Figure 1, this approach achieves high statistical efficiency. The error in parameter recovery scales as $1/\sqrt{N}$ (where $N$ is the number of samples), which is the theoretical gold standard for statistical estimators. This efficiency holds across different models, including the 1D Ising model and $\phi^4$ theory.

Learning from Correlations

A pivotal contribution of this work is the reformulation of these learning objectives in terms of statistical moments (e.g., $\langle \phi^2 \rangle, \langle \phi^4 \rangle$). For many critical theories, such as $\phi^4$ or Ising models, the physics can be learned using only low-order correlation data. This is a significant advancement because it allows for the recovery of microscopic couplings without requiring full snapshots of the system, making the method applicable to realistic experimental data like scattering results.

Reconstructing Renormalization Group Flows

By iteratively coarse-graining the data (averaging groups of variables) and re-learning the parameters at each step, our framework can map out how a theory "flows" across different scales. This process, known as the Renormalization Group (RG), is usually reserved for complex analytical perturbation theory.

Non-perturbative RG flows and phase boundary detection across various field theories.

Our data-driven approach is non-perturbative, meaning it remains accurate even when interactions are extremely strong. As illustrated in Figure 5, we can perfectly reconstruct RG flows for theories like the Sine-Gordon model and the Massive Schwinger model. Furthermore, this method allows for precise phase boundary detection. For instance, in $\phi^4$ theory, we can locate phase transitions by identifying "turnaround points" in the RG flow of the mass parameter: if the mass increases under coarse-graining, the system is in a symmetric phase; if it becomes more negative, the symmetry is broken.

Key Results and Applications

The methodology was benchmarked against high-fidelity synthetic datasets generated via Monte Carlo simulations. Key findings include:

  • High Accuracy: Parameter recovery errors were consistently below 1% across various models.
  • Scalability: While the complexity of the forward problem grows exponentially in strong-coupling regimes, the inverse learning task only grows polynomially with the number of sites.
  • Hybrid Systems: For complex gauge-matter models, we demonstrated that parameters can be learned sequentially—solving for the matter part first, then the gauge part—without a loss in accuracy.

Conclusion

This framework provides a robust, data-driven route to solving the inverse problem in statistical field theories. By moving beyond the limitations of perturbation theory and the requirement for full field configurations, it opens new doors for understanding strongly coupled systems. Future work will explore extending this methodology to non-equilibrium dynamics and developing automated, graph-agnostic coarse-graining schemes.

Summary updated

Abstract

Recovering microscopic couplings directly from data provides a route to solving the inverse problem in statistical field theories, one that complements the traditional-often computationally intractable-forward approach of predicting observables from an action or Hamiltonian. Here, we propose an approach for the inverse problem that uniformly accommodates systems with discrete, continuous, and hybrid variables. We demonstrate accurate parameter recovery in several benchmark systems-including Wegner's Ising gauge theory, $ϕ^4$ theory, Schwinger and Sine-Gordon models, and mixed spin-gauge systems, and show how iterating the procedure under coarse-graining reconstructs full non-perturbative renormalization-group flows. This gives direct access to phase boundaries, fixed points, and emergent interactions without relying on perturbation theory. We also address a realistic setting where full gauge configurations may be unavailable, and reformulate learning algorithms for multiple field theories so that they are recovered directly using observables such as correlations from scattering data or quantum simulators. We anticipate that our methodology will find widespread use in practical learning of field theories in strongly coupled regimes where analytical tools might fail.
BibTeX
@misc{shukla2025learning,
  author = {Shreya Shukla and Abhijith Jayakumar and Andrey Y. Lokhov},
  title = {{Learning of Statistical Field Theories}},
  year = {2025},
  eprint = {2511.09859},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.09859},
  howpublished = {arXiv},
}
Figure