⚛️ high-energy theory

Spinning Conformal Correlators from Neural Networks

This paper constructs spinning conformal fields by combining neural networks with the embedding formalism to compute their correlation functions, demonstrating that a specific ensemble of independent and identically distributed neurons recovers the 4d Maxwell conformal field theory in the infinite-width limit.

Original authors: Manas Dogra, James Halverson, Joydeep Naskar

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Manas Dogra, James Halverson, Joydeep Naskar

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

In the vast landscape of modern physics, two distinct worlds often seem to speak different languages. On one side stands quantum field theory, the mathematical framework that describes the fundamental particles and forces of our universe, from the light that allows us to see to the glue that holds atomic nuclei together. On the other side sits the world of neural networks, the computational engines behind artificial intelligence that learn to recognize patterns, translate languages, and play games. For decades, these fields operated in separate silos, one focused on the fabric of reality and the other on the architecture of software. However, a surprising connection has emerged: the mathematical structures that govern how a neural network processes information are deeply similar to the equations that describe how particles interact in a vacuum. This link suggests that the way a computer learns might offer a new, unexpected way to understand the fundamental laws of nature.

At the heart of this connection is a concept known as conformal symmetry. In physics, this describes a universe where the laws of nature remain unchanged even if you stretch or shrink the distances between objects. It is a powerful idea that helps physicists solve complex problems about how particles behave at different scales. For years, researchers have used neural networks to model simple, scalar particles—those without any internal direction or spin. But the universe is filled with more complex entities, such as light and magnetic fields, which possess a property called spin. These spinning fields behave differently; they have directionality and orientation, making them much harder to describe mathematically. The challenge has been to find a way to use the flexible, pattern-matching power of neural networks to recreate these spinning fields and the intricate dance of their interactions.

A team of researchers has now taken a significant step forward by successfully constructing these spinning fields directly from neural networks. They did not simply use a computer to simulate a known physical system; instead, they built the physical laws themselves out of the network's architecture. By arranging the network's internal components in a specific, symmetrical way, they created a system that naturally produces the correct mathematical behavior for spinning particles. The researchers demonstrated that when they calculated how these artificial fields interacted with one another, the results perfectly matched the known rules of conformal field theory. They showed that the network could reproduce the precise patterns of interaction for two, three, and even four spinning particles coming together, a feat that requires a high degree of mathematical precision and structural integrity.

The most striking achievement of this work is the recreation of the free Maxwell theory, which describes the behavior of light and electromagnetism in four-dimensional space. In the real world, light is a free field, meaning its photons do not interact with each other directly; they simply pass through one another. The researchers found that by using a single, simple neural network, they could capture the basic two-point interaction of this field, but the network still contained hidden complexities that prevented it from being a perfect match for the real world. To solve this, they employed a technique involving a vast ensemble of many independent networks working in parallel. As they increased the number of these networks to infinity, the unwanted complexities vanished, and the system settled into a state that was indistinguishable from the true, free theory of light. The resulting model reproduced the exact correlation functions of the electromagnetic field, including the specific way the field strength relates to itself across space and time.

This work does more than just replicate known physics; it offers a new perspective on how physical laws might emerge from statistical systems. The researchers showed that the specific way the parameters of a neural network are shared and correlated determines which physical structures appear. By carefully tuning these connections, they could control which types of particle interactions were allowed and which were forbidden. This suggests that the architecture of a neural network is not just a tool for calculation, but a fundamental definition of a physical theory. The study confirms that the deep structural correspondence between machine learning and quantum field theory is robust enough to handle the complexity of spinning fields, opening the door to using these computational tools to explore more exotic and difficult-to-solve theories in the future.

The researchers were careful to note the boundaries of their success. Their construction successfully captures the local, gauge-invariant sector of the theory, which describes the observable effects of the field, but it does not yet address non-local features like magnetic monopoles or the specific transformation rules of the underlying potential. Furthermore, while the method works beautifully for four-dimensional space, it relies on specific properties of that dimension that do not hold in three or five dimensions, meaning the approach is not yet universal. Nevertheless, the ability to derive the exact behavior of a fundamental force like electromagnetism from the statistical properties of a neural network provides a concrete proof of concept. It demonstrates that the language of artificial intelligence can be used to write the laws of physics, offering a fresh and powerful lens through which to view the fundamental building blocks of our universe.

Technical Summary: Spinning Conformal Correlators from Neural Networks

Problem Statement
Recent developments in the Neural Network-Field Theory (NN/FT) correspondence have established that infinite-width neural networks can reproduce the correlators of free scalar conformal field theories (CFTs). However, a significant gap remained in extending this framework to operators with non-zero spin. While scalar fields were successfully constructed using the embedding formalism, the construction of spinning conformal fields—specifically those satisfying transversality and homogeneity conditions required by conformal symmetry—had not been systematically addressed within the NN/FT paradigm. Furthermore, it was unclear whether the specific architecture and parameter statistics of neural networks could reproduce the precise tensor structures and correlation functions of interacting or gauge-invariant spinning theories, such as the free Maxwell CFT in $d=4$ .

Methodology
The authors extend the embedding formalism of CFTs to spinning operators by constructing specific neural network architectures. The methodology relies on the following principles:

Embedding Formalism: The authors utilize the isomorphism between the conformal group on $\mathbb{R}^d$ and the Lorentz group $SO(d+1, 1)$ acting on an embedding space $\mathbb{R}^{d+1,1}$ . Spinning primaries are lifted to homogeneous tensor operators on the projective null cone (PNC).
Neural Network Ansatz: To represent a spinning primary of spin $J$ and scaling dimension $\Delta$ , the authors propose a homogeneous neural network architecture. For a vector field ( $J=1$ ), the ansatz is:
$A(X, Z) = (\tilde{\Theta} \cdot X)^{-\Delta-1} \left[ (\Theta \cdot X)(\eta \cdot Z) - (\eta \cdot X)(\Theta \cdot Z) \right]$
where $\Theta, \tilde{\Theta}, \eta$ are independent Gaussian random vectors. The term in the brackets ensures the field is transverse ( $X \cdot A = 0$ ) and invariant under the shift $Z \to Z + \alpha X$ , satisfying the requirements for a section on the PNC.
Correlator Computation: Correlators are computed by averaging over the probabilistic parameters of the network. The authors explicitly calculate two-, three-, and four-point functions for various spins and dimensions, checking for the correct conformal tensor structures (e.g., $H_{ij}$ and $V_{i,jk}$ building blocks) and scaling behaviors.
Maxwell CFT Construction: To recover the free Maxwell theory in $d=4$ , the authors address the issue that the vector potential $A_\mu$ is not a conformal primary (it has anomalous scaling and gauge dependence). They construct a "Maxwell-like" theory using the vector ansatz with $\Delta=1$ . Recognizing that a single network channel yields non-Gaussian higher-point functions, they employ a large- $N$ ensemble of $N$ independent channels. By taking the limit $N \to \infty$ followed by $\Delta \to 1$ , the connected non-Gaussian contributions are suppressed, and the theory becomes Gaussian, reproducing the exact Wick contractions of free Maxwell theory.

Key Contributions and Results

Construction of Spinning NN-CFTs: The paper successfully constructs neural network architectures that generate spinning conformal fields. It demonstrates that the two-point functions for spin-1 and spin-2 fields match the required CFT tensor structures (specifically the projector structure for spin-1).
Verification of Correlators: The authors explicitly compute three- and four-point functions.
- For three-point functions of vectors, the results match the expected CFT forms, decomposing into linear combinations of the standard tensor structures $H_{ij}V_{k,ij}$ .
- For four-point functions, the architecture produces correlators that are linear combinations of tensor structures with coefficients depending on conformal cross-ratios $u$ and $v$ . The paper shows that different permutations of network parameters yield different valid CFT tensor structures, effectively controlling the Operator Product Expansion (OPE) data.
Recovery of 4d Maxwell CFT:
- The authors construct a "Maxwell-like" theory where the field strength two-point function matches that of $d=4$ Maxwell theory upon restriction to the Poincaré section.
- Crucially, they show that a single-channel NN-CFT is non-Gaussian and fails to reproduce the higher-point functions of the free Maxwell theory.
- By introducing an ensemble of $N$ i.i.d. channels and taking the large- $N$ limit, the theory becomes Gaussian. In this limit, the connected parts of higher-point correlators vanish, and the full set of correlators is determined by Wick's theorem, exactly reproducing the local gauge-invariant sector of the free Maxwell CFT (including the correct normalization of the field strength correlator).
Handling Non-Primary Fields: The paper addresses the subtlety that the Maxwell field strength is a primary only at $\Delta=1$ , while the vector potential is not. The construction involves a rescaling limit ( $\Delta \to 1$ ) that decouples the non-primary vector potential while retaining the correct field strength correlators.

Significance and Claims
The paper claims to provide the first systematic construction of spinning conformal fields directly from neural networks, extending the NN/FT correspondence beyond scalar theories. The primary significance lies in demonstrating that:

Neural network architectures can encode the specific tensor structures required by conformal symmetry for spinning operators.
The sharing and statistics of network parameters act as a mechanism to control which tensor structures and OPE data appear in the resulting CFT.
The large- $N$ limit of neural network ensembles provides a robust method to isolate Gaussian sectors (Generalized Free Fields) from non-Gaussian architectures, successfully recovering the free Maxwell CFT in $d=4$ .

The authors note that while they have successfully constructed the local gauge-invariant sector, the approach currently does not address non-local observables (like Wilson lines) or the transformation laws of the gauge potential itself, as the potential is effectively decoupled in the final construction. They suggest that a systematic understanding of the map between network data and CFT OPE data could provide a route to reverse-engineer more general CFTs, but this remains a direction for future work.

More like this