AIM-PROBABILITY-0016 · Complete scoped heat-flow theorem

Heat Flow Leaves Two-Hidden-Unit Restricted Boltzmann Models

Manuscript 20 October 2026 · Online 20 October 2026

math.PRmath.AGstat.MLUnrefereed preprint

Abstract

Independent bit-flip heat flow preserves binary restricted Boltzmann machines with one hidden unit, by their classical representation as mixtures of two product distributions. We give explicit evolving parameters. With two hidden units and at least four visible units, the behavior is different: for every fixed positive vector of flip rates, a generic positive parameter leaves even the algebraic closure of the model immediately. A sixteen-by-sixteen integer Jacobian determinant proves transversality in the four-visible case. A coordinate-permutation argument covers arbitrary positive rates, and conditioning extends the obstruction to every larger visible dimension. The proof distinguishes a rank witness from a pointwise smoothness certificate. For the logarithmic-mean graph transport metric, this yields a generically nonzero residual in projected entropy flow. This is a complete scoped finite-state theorem, not a solution of the broad AIM question about PDE approximation. The preprint is unrefereed and makes no absolute priority claim.

Record

Affiliation
Mercury Software GmbH
Result
Complete scoped heat-flow theorem
Categories
math.PR · math.AG · stat.ML
Manuscript
20 October 2026
Online release
20 October 2026
Version
1.0
License
Creative Commons Attribution 4.0 International

Files and verification

The PDF is the canonical reading copy. The source archive contains the LaTeX manuscript, bibliography, reproducibility material, and audit documents without build artefacts.

Citation

Alper Ferudun, “Heat Flow Leaves Two-Hidden-Unit Restricted Boltzmann Models,” EulerSolve Research Papers, AIM-PROBABILITY-0016, 2026. https://doi.org/10.5281/zenodo.23489067.

BibTeX
@misc{ferudun2026rbmheat,
  author = {Ferudun, Alper},
  title = {Heat Flow Leaves Two-Hidden-Unit Restricted Boltzmann Models},
  year = {2026},
  howpublished = {EulerSolve Research Papers},
  url = {https://eulersolve.org/papers/aim-probability-0016/},
  doi = {10.5281/zenodo.23489067},
  note = {AIM-PROBABILITY-0016; unrefereed preprint}
}

More research papers

Show all 378 other papers

All 379 research papers →