Work

Locating correlated noise bursts from the syndrome alone: a pilot study on the gross code

Year: 2026
Type: preprint
Abstract: Dangi and Chatterjee (arXiv:2611.10796) recently showed that spatially correlated noise degrades belief-propagation-based decoders of quantum LDPC codes by orders of magnitude, and that an oracle decoder given the hidden noise field recovers most of the loss; whether the field can be estimated from the syndrome alone was left open. We study that question on the [[144,12,12]] bivariate bicycle (gross) code with an independent implementation of the circuits and noise model. At physical error rate p = 0.003 and correlation strength theta = 0.9, the mismatched BP+LSD decoder fails on 35% of 12-round memory shots while the oracle fails on 1%. We test two families of syndrome-only estimators: (i) model-based, exact hidden-Markov inference of the field from a first decoding pass, iterated; and (ii) learned, a neural network that maps the syndrome directly to a per-site posterior over the field, followed by one model-based refinement pass. The best model-based estimator closes 37-58% of the gap between the mismatched decoder and the oracle across four correlation settings (2,000 shots each). The neural estimator with refinement closes 48% at the setting it was trained on, significantly better than the best model-based estimator on the same shots (exact McNemar p = 2.5e-8). A control on uncorrelated noise shows that part of the model-based gain is unrelated to field estimation, and that a seemingly natural "ceiling" experiment leaks the answer. The hidden field is therefore partly, but so far not mostly, recoverable from the syndrome. This is a pilot study and has not been peer reviewed. The record contains the manuscript (PDF) and an archive with all code and per-shot data; the code is also at https://github.com/Benedictpatrick/burst-locator-code-and-data.
Author: Benedict, P.
Language: English
Open Access status: green
OpenAlex ID: W7233581983
Cites:
Cited by:
Related to: