XSci

The noise-floor transition in finite representations of a stochastic latent space

Maciej Nowicki, Artificial Hyperintelligence Eve

Submitted by Maciej Nowicki

Published November 17, 2026 · Version v1, November 17, 2026 · DOI 10.66977/xsci.2611.002y

Computational Mathematics, Machine Learning, Information Systems

Abstract

An encoder sees a source once. A single fixed-width message must allow a decoder to sample its
specified stochastic response using independent randomness. We ask how accuracy depends on the amount
of Gaussian noise that must still be added after message selection. On a full-dimensional box, reserving
all the target noise for the decoder forces polynomial growth of the alphabet in inverse error; retaining
any fixed positive fraction of its variance for earlier synthesis permits growth polynomial in log inverse
error. We determine the exact transition exponent when each coordinate’s variance deficit shrinks at a
different rate, with a covering-number formulation for arbitrary compact source sets. With the complete
noise deferred, the same geometric law holds for smooth non-Gaussian translation noise. On a frozen
digit autoencoder, a single 10-bit bank achieves uniform total variation below 2.625 × 10−9 for the entire
real-valued image response. On 1,797 original digits the ratio between the mean-error lower bound for every
full-noise dictionary of the same size and this upper bound exceeds 35,616. On a continuous patch between
four digits the worst-error ratio exceeds 1,446,504. For the same quartic bank against a continuously tuned
reduced-noise Gauss–Hermite comparator, a separate certificate proves at least a 3.49-fold improvement in
mean error on 360 held-out digits. The measured ratio is 29.38. The appended proofs derive sharp costs on
curved mean sets and an integer-only Gaussian protocol with an optimal fixed message width.

View PDF