Same WZW current algebra, two neural realizations, different errors: level–rank duality carries charged correlators across paired theories
Synopsis
Neural network field theory writes field-theoretic correlators as expectations over network parameters, yet one continuum current algebra can be realized by different networks. This work builds positive-level WZW current algebras from bosonic Wakimoto fields and fermionic Fourier–Berezin fields, and uses level–rank duality to connect the charged sectors of paired theories. At fixed regulator and finite width, direct and coset constructions differ in error, spectral distribution and regulator choice shift accuracy, and source-sector measurements predict paired-sector responses without measuring the target preparation.
Interpretation
One continuum current algebra can be realized by two kinds of neural network: a bosonic Wakimoto construction assembles affine currents from scalar fields and commuting ghost fields, while a fermionic Fourier–Berezin construction yields row, column and number currents from bilinears of Grassmann-valued fields. Earlier neural network field theory reproduced continuum kernels mainly at the level of free-field and conformal two-point structure; higher connected moments, shared-parameter correlations and regulator dependence of the finite ensemble were not examined alongside a duality. Taking the infinite-width limit at fixed regulator and then removing the annular cutoff, both constructions recover the target affine current algebra; the fermionic double-pole coefficient gives positive levels, with row and column levels counted by flavor multiplicities.
Level–rank duality pairs the charged intermediate channels of two WZW theories, so observations in one sector can predict correlators in the other. The duality was previously an exact representation-theoretic relation; paired conformal blocks need not be equal as functions, and it does not by itself identify the parameters or finite approximation schemes of two neural readouts. In the six-channel fermionic construction, the charge-two tensor product splits into two intermediate channels, and affine sewing agrees with the Knizhnik–Zamolodchikov recursion through grade eight for the row theory and grade four for the column theory; source-trained predictors reach mean errors below one percent against paired preparation responses without target-preparation measurements.
Finite-width error depends on the realization: direct Wakimoto and coset constructions differ at equal feature width, and a spectral-tilt parameter changes accuracy without increasing width or sampled parameters per neuron. Both constructions give the same target current sector in the ordered continuum limit, so physical equivalence alone does not determine approximation quality at finite width. Across three regulator profiles, the direct construction has the smaller regulated Gaussian error on the fixed insertion set; at each fixed width a spectral exponent minimizes the stress-tensor correlation error, and that optimum varies with regulator, correlator and width.
Perspective
The scope is neural realizations of positive-level WZW current algebras; numerical comparisons cover current and stress-tensor correlators of the direct Wakimoto and coset constructions at fixed regulator parameters, a fixed insertion set and specified feature widths. The charged-sector transfer experiments use charge-one and selected charge-two insertions in the six-channel fermionic construction and compare against independent KZ transport.
The optimal spectral exponent at finite width changes with correlator, regulator and width, so a value selected at one width need not remain optimal at another. The direct construction has the smaller error on the stress-tensor correlators examined here, and that observation does not extend to all observables. Amplitude calibration lowers the fitted-data error but worsens held-out and validation errors at some widths. In finite-mode compression, optimized subspaces beat conventional mode truncation in only two of five settings, and the finite reference itself carries truncation error.
