Answers are grounded in this paper and cite their sources.
Theoretical and computational analysis demonstrates scaling and renormalization constraints in deep neural architectures, highlighting strict limits on quantum field reconstruction.
We develop a mathematical framework for neural field laws that separates width, head count, sequence length, depth, training, coarse-graining, parameter quotients, and quantum reconstruction. Under a declared conditional Gaussian representation, connected field correlators are determined by cumulants of the random covariance. Fixed-head attention can consequently remain non-Gaussian at infinite channel width, whereas independent head aggregation suppresses higher cumulants. A grading fan organizes distinct asymptotic paths, and a stationary covariance cocycle identifies the rate of initialization-memory decay. Training is treated separately through controlled Fokker--Planck evolution and Hamilton--Jacobi duality. Explicit coarse maps make RG claims testable. We give transport and composition bounds, exact Gaussian decimation, an exactly closed non-Gaussian Gamma family with computable scaling exponents, and a conditioned spectral estimate for approximate conjugacy. A constructive obstruction shows that zero one-step training risk need not control recursive coarse-graining; discarded-state memory and intermediate-scale support explain two sources of failed closure. Reflection positivity remains a separate requirement. We establish preservation results and negative witnesses, and exhibit a reflection-positive covariance mixture that fails clustering. Neural random fields therefore require additional hypotheses for positive-metric quantum reconstruction. Reproducible calculations include 32 fixed attention kernels and a supervised study of 24 trained models (12 LSTMs and 12 transformers) across three input tasks and three seeds. Intermediate-scale training reduces median eight-token Rademacher averaging error from 0.0665 to 0.0256 for the LSTM and from 0.2664 to 0.00493 for the transformer. All 180 sampled scale records satisfy the applicable trajectory-error bound. These results test recovery of a prescribed map and a compositional failure mechanism; they do not establish spontaneous universality or an isomorphism between generic neural architectures and RG flows.
No takes yet. Share an insight, caveat, or question.
Miquel Noguer Alonso (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: