Conformal Fields from Neural Networks

The Big Picture

Imagine designing a building that looks identical from every angle, not just rotated but stretched, squeezed, and viewed from any distance. That’s the challenge facing physicists who study conformal field theories (CFTs): mathematical frameworks invariant under rotation, translation, and rescaling. These theories appear everywhere from the physics of boiling water to the deep structure of string theory, yet constructing new ones from scratch has always been brutally hard.

A team at Northeastern University and East China Normal University has found an unexpected shortcut: build conformal fields out of neural networks. Not train a network to approximate a CFT, but actually use the architecture and parameter statistics of a neural network as the literal definition of a new quantum field theory. The mathematics fit together cleanly.

Halverson, Naskar, and Tian show that by choosing a neural network built from homogeneous functions and restricting it to a carefully chosen surface in higher-dimensional space, you automatically get a conformally symmetric field theory. The result opens a new factory for exactly solvable CFTs, complete with computable predictions for field interactions and particle-like states.

Neural networks, viewed as parameter-dependent functions rather than learning machines, can serve as direct constructions of conformal field theories. Integration over parameters plays the role of the path integral.

How It Works

The construction rests on a geometric trick called the embedding formalism, originally due to Dirac. Conformal symmetry in D dimensions is secretly equivalent to rotational symmetry in D+2 dimensions. So instead of working in physical spacetime, you lift everything to a higher-dimensional space, then project back down onto the projective null cone: the set of lightlike directions in that larger space.

Neural networks enter here. The researchers define a network Φ(X) on the full (D+2)-dimensional space with one crucial constraint: the network must be homogeneous, meaning scaling the input by λ scales the output by λ^Δ for a fixed exponent Δ. Homogeneous activation functions like monomials x^n make this straightforward to build.

Restricting the network to the projective null cone produces a field that automatically obeys all transformation laws of a conformal field in D dimensions. Δ becomes the conformal dimension, the number encoding how the field responds to rescaling.

The workflow has four steps:

1. Choose an architecture — a neural network with homogeneous activations, specific depth and width
2. Choose a parameter distribution P(Θ) — the analog of choosing an action in conventional field theory
3. Restrict to the null cone — project the network from D+2 down to D dimensions
4. Compute correlators — integrate over network weights instead of field configurations

That last step is the computational payoff. For shallow networks with simple parameter distributions, the integrals are tractable, sometimes exactly. The paper computes exact four-point correlation functions in multiple examples: measurements of how four points simultaneously influence each other, where most of the rich physics (operator spectra, OPE coefficients, conformal block data) lives.

A particularly clean result comes from the infinite-width Gaussian process limit. When a network is made infinitely wide with Gaussian-distributed weights, its outputs become a Gaussian process, which is effectively a free field theory. The researchers use this to construct generalized free CFTs, including an explicit realization of the free boson reproduced from first principles.

Deep networks add structure layer by layer. Each layer produces its own conformal field, and the authors derive recursion relations linking conformal dimensions and four-point functions of successive layers. Going deeper corresponds to RG-flow-like transformations, a connection to the renormalization group that looks worth pursuing seriously.

The paper also borrows tools from particle physics. Techniques developed for Feynman integrals apply directly to the parameter-space integrals here, and the exchange goes both ways: collider physics tools help construct conformal theories, while the neural network picture may offer new intuition for amplitude calculations.

Why It Matters

CFTs occupy a foundational role in theoretical physics. The renormalization group, which explains why vastly different physical systems share the same critical behavior, can be understood as flows between CFT fixed points. The conformal bootstrap uses consistency conditions to constrain and solve CFTs without writing down an explicit Lagrangian. It is one of the most productive programs in modern theoretical physics. The neural network construction offers something different: explicit, computable realizations where spectra and correlators can be extracted directly.

The two approaches are complementary. Where bootstrap methods derive powerful bounds from symmetry and unitarity, the neural network construction builds specific examples from the ground up. A natural next step is combining them, using bootstrap constraints to filter which neural network architectures yield physically sensible theories.

The paper leaves reflection positivity (the Euclidean analog of unitarity, ensuring probabilities stay positive) as an open problem. Solving it could enable systematic classification of CFTs accessible via this construction. On the numerical side, Monte Carlo sampling over parameter space provides an approximation strategy for architectures where exact integrals are out of reach, effectively treating neural network inference as a simulation tool for quantum field theory.

By treating neural network architectures as homogeneous functions on a higher-dimensional null cone, this work turns machine learning machinery into a construction kit for conformal field theories. It yields exact four-point functions, free field limits, and recursive deep-network extensions that open a new path into quantum field theory.