Virasoro Symmetry in Neural Network Field Theories

Abstract: Neural Network Field Theories (NN-FTs) can realize global conformal symmetries via embedding space architectures. These models describe Generalized Free Fields (GFFs) in the infinite width limit. However, they typically lack a local stress-energy tensor satisfying conformal Ward identities. This presents an obstruction to realizing infinite-dimensional, local conformal symmetry typifying 2d Conformal Field Theories (CFTs). We present the first construction of an NN-FT that encodes the full Virasoro symmetry of a 2d CFT. We formulate a neural free boson theory with a local stress tensor $T(z)$ by properly choosing the architecture and prior distribution of network parameters. We verify the analytical results through numerical simulation; computing the central charge and the scaling dimensions of vertex operators. We then construct an NN realization of a Majorana Fermion and an $\mathcal{N}=(1,1)$ scalar multiplet, which then enables an extension of the formalism to include super-Virasoro symmetry. Finally, we extend the framework by constructing boundary NN-FTs that preserve (super-)conformal symmetry via the method of images.

• The paper introduces a novel Log-Kernel architecture that constructs NN field theories with a local stress tensor and complete Virasoro symmetry.
• It demonstrates accurate recovery of conformal data, including correlation functions, scaling dimensions, and central charge matching theoretical predictions.
• The study extends its framework to fermionic and supersymmetric models, paving the way for simulations of non-perturbative quantum field phenomena.

Virasoro Symmetry in Neural Network Field Theories: A Technical Overview

Introduction and Theoretical Context

The paper "Virasoro Symmetry in Neural Network Field Theories" (2512.24420) addresses the explicit construction of Neural Network Field Theories (NN-FTs) encoding the full (infinite-dimensional) Virasoro symmetry algebra of two-dimensional Conformal Field Theory (2D CFT). Standard approaches have established the emergence of global conformal symmetries in NN-FTs by leveraging infinite-width limits and embedding space constructions, which realize Generalized Free Fields (GFFs). However, such models do not generally admit a local, conserved stress-energy tensor, hence lacking the generators required for Virasoro symmetry—a key ingredient for 2D critical phenomena and string theory applications.

This work overcomes the aforementioned limitations through the construction of a Log-Kernel (LK) network architecture, delivering a field-theoretic model that supports both a local stress tensor and Virasoro symmetry. Extensions to fermionic models and boundary conditions are also developed, including realizations of the $\mathcal{N}=(1,1)$ super-Virasoro algebra.

The Log-Kernel Architecture and Emergent Local CFT Structure

The LK network is engineered as a superposition of random Fourier features on $\mathbb{C}$, tuned such that its covariance kernel—after ensemble averaging—corresponds to that of the 2D free boson:

$\langle \phi(z)\phi(0) \rangle \propto -\ln|z|^2$

To ensure scale invariance and global conformal symmetry, the spectral density of the Fourier weights is set to $p(k)\propto |k|^{-2}$, uniquely fixing the architecture to match the CFT’s scaling requirements.

Finite spectral cutoffs (both IR and UV) are introduced as regularizers; within their admissible window, the kernel exhibits the required scale and translation covariance. The variance normalization of the random weights is a critical hyperparameter, calibrated to reproduce the canonical normalization for the boson.

Construction and Statistics of Virasoro Generators

Central to the paper’s contribution is the realization of Virasoro generators $L_n$ as stochastic functionals of the neural ensemble. By expanding the field in angular (Laurent) modes, the current $J(z) = i\partial_z\phi$ is decomposed into Gaussian random variables $\alpha_n$. The Virasoro operators are then constructed as normal-ordered bilinears of these modes.

Through ensemble statistics and explicit functional integration, the commutation relations of the Virasoro algebra are recovered:

$$[L_m, L_n] = (m-n)L_{m+n} + \frac{c}{12}m(m^2-1)\delta_{m+n,0}$$

The variance of the mode coefficients sets the central charge normalization. Simulation of the LK network yields a central charge $c_{\text{exp}}=0.9787\pm0.0421$, matching the expected $c=1$ within $2.2\%$ precision.

Spectrum and Scaling of Neural Vertex Operators

Neural Vertex Operators (NVOs) are constructed as nonlinear readouts of the field ensemble, $\mathcal{V}_a(z) = :e^{ia\phi(z)}:$, with normal ordering implemented via vacuum subtraction. Their two-point functions display precise power-law scaling:

$$\langle \mathcal{V}_a(z)\mathcal{V}_{-a}(0)\rangle \sim |z|^{-2a^2}$$

This behavior is confirmed numerically for a range of charges $a$, with extracted scaling dimensions $\Delta_a$ agreeing with the theoretical $a^2$ spectrum to within $<1\%$ for $a\leq1$.

Figure 1: Log-log plot of the Neural Vertex Operator two-point function, demonstrating correspondence between the measured and theoretical scaling dimensions of the free boson.

Fermionic Neural Architectures and Super-Virasoro Symmetry

A parallel construction is implemented for fermionic degrees of freedom via Grassmann-valued network parameters, producing a Neural Majorana Fermion (NMF) with propagator converging to the Cauchy kernel $\langle\psi(z)\psi(w)\rangle = 1/(z-w)$, characteristic of the $c=1/2$ Ising CFT.

By combining bosonic and fermionic network realizations, the $\mathcal{N}=(1,1)$ super-Virasoro algebra is reconstructed. Mode expansions for supercurrents and stress tensors, along with their algebraic anticommutators, are evaluated using ensemble averages, matching the super-Virasoro structure with central charge $c=3/2$.

Realization of Boundary CFTs and Method of Images

The LK architecture generalizes naturally to domains with boundaries (upper half-plane) by employing the method of images. For the bosonic field, boundary conditions (Dirichlet/Neumann) are enforced by pairing each random feature with its reflection; for the fermionic sector, spin structure is controlled via reflected Grassmann parameters.

Explicit expressions are given for the resulting ensemble-averaged propagators, and the corresponding boundary kernel is shown to preserve global and local (super-)conformal symmetry, including compatibility with the NS/R sectors.

Finite-Width Corrections and Emergent Interactions

Though the infinite-width limit yields a strictly Gaussian free theory, the paper analyzes finite-width corrections by measuring the connected four-point function (excess kurtosis) across varying network widths. A clear $1/N$ decay is observed for the interaction strength, consistent with the theoretical prediction that finite $N$ generates weakly non-Gaussian, interacting effective field theories.

Figure 2: Scaling of the connected four-point function $G_{4c}$ (measuring interaction strength) as a function of network width $N$, showing precise $1/N$ suppression and rapid convergence to a Gaussian fixed point.

Implications and Speculative Outlook

This work rigorously demonstrates that neural architectures can instantiate local, infinite-dimensional conformal symmetry (Virasoro, super-Virasoro) in field-theoretic kernels, including the precise spectrum and central charge. The technical realization of both bulk and boundary CFTs, and their fermionic and supersymmetric extensions, establishes NN-FTs as a laboratory for simulating and studying universal features of critical 2D systems.

At the practical level, this framework enables new forms of simulation for non-perturbative quantum field theoretic phenomena—potentially benefiting studies in statistical mechanics, condensed matter, and string theory. The inclusion of interactions via finite-width effects directly connects with RG dynamics and neural effective actions, suggesting a pathway toward simulating nontrivial CFTs and dualities (e.g., bosonization, mirror symmetry) with neural architectures. Extension to engineered minimal or interacting models remains a leading challenge.

Conclusion

The paper achieves the first explicit construction of NN-based field theories with exact Virasoro symmetry by rigorously tuning network priors and architecture. Through both analytical derivation and large-scale simulations, it verifies the reproduction of all key data of the free boson CFT, including correlation functions, spectrum, and boundary behavior, and extends these results to fermionic and supersymmetric models. This establishes a new, robust formalism for the intersection of neural networks and quantum critical field theories, with both theoretical and algorithmic implications for ongoing research in machine learning and mathematical physics.