Decomposing neural decoders: an exactly-anchored benchmark for quantum error correction

Abstract Minimum-weight perfect matching (MWPM) is the dominant surface-code decoder but is suboptimal under depolarizing noise, because it discards the X – Z correlation that Y errors induce and is not degeneracy-aware. We present an exactly anchored diagnostic benchmark for the code-capacity depolarizing channel, not a case for neural decoding. Five decoders (MWPM, correlated matching, belief-matching, belief propagation with ordered-statistics decoding (BP+OSD) and a lightweight multilayer perceptron (MLP)) are compared with an exact analytic maximum-likelihood (ML) reference at distance d = 3 , with exact per-syndrome ML decisions on sampled syndromes at d = 5 , and with a matched-marginal independent-noise control. Measuring MWPM’s suboptimality by the relative logical-error-rate (LER) reduction that ML achieves over it, the difference between the depolarizing and control channels is 19–32 percentage points (pp) above an 11%–14% floor at d = 3 , and 33–60 pp above an 8%–10% floor at d = 5 . At d = 3 the MLP recovers 90%–95% of the MWPM-to-ML gap, more than belief-matching (82%–89%), BP+OSD (74%–80%) or correlated matching (46%–58%). At d = 5 belief-matching and BP+OSD agree closely yet stay measurably above ML, recovering 79%–91% of the gap, while the MLP’s advantage over MWPM depends on the training budget (crossover near 3.5 × 10 5 samples at p = 0.08 ) and, when adequately trained, recovers at most 17%. In simulated circuit-level memory experiments at d = 3 and 5, the classical decoders keep their advantage over MWPM, the matched-marginal split becomes decomposition-dependent, and the MLP recovers 25% of the gap at d = 3 , T = 2 and, over five training seeds and budgets up to 9 × 10 5

Authors

Institutions

Publication Details

Journal
Quantum Science and Technology
Published
2026-11-16
DOI
https://doi.org/10.1088/2058-9565/aebd16
Primary Topic
Quantum Computing Algorithms and Architecture
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Decomposing neural decoders: an exactly-anchored benchmark for quantum error correction

Mohammad Walid Charrwi
1 links
Quantum Science and Technology
Quantum Computing Algorithms and Architecture
article

Decomposing neural decoders: an exactly-anchored benchmark for quantum error correction

Mohammad Walid Charrwi
article en

Abstract

Abstract Minimum-weight perfect matching (MWPM) is the dominant surface-code decoder but is suboptimal under depolarizing noise, because it discards the X – Z correlation that Y errors induce and is not degeneracy-aware. We present an exactly anchored diagnostic benchmark for the code-capacity depolarizing channel, not a case for neural decoding. Five decoders (MWPM, correlated matching, belief-matching, belief propagation with ordered-statistics decoding (BP+OSD) and a lightweight multilayer perceptron (MLP)) are compared with an exact analytic maximum-likelihood (ML) reference at distance d = 3 , with exact per-syndrome ML decisions on sampled syndromes at d = 5 , and with a matched-marginal independent-noise control. Measuring MWPM’s suboptimality by the relative logical-error-rate (LER) reduction that ML achieves over it, the difference between the depolarizing and control channels is 19–32 percentage points (pp) above an 11%–14% floor at d = 3 , and 33–60 pp above an 8%–10% floor at d = 5 . At d = 3 the MLP recovers 90%–95% of the MWPM-to-ML gap, more than belief-matching (82%–89%), BP+OSD (74%–80%) or correlated matching (46%–58%). At d = 5 belief-matching and BP+OSD agree closely yet stay measurably above ML, recovering 79%–91% of the gap, while the MLP’s advantage over MWPM depends on the training budget (crossover near 3.5 × 10 5 samples at p = 0.08 ) and, when adequately trained, recovers at most 17%. In simulated circuit-level memory experiments at d = 3 and 5, the classical decoders keep their advantage over MWPM, the matched-marginal split becomes decomposition-dependent, and the MLP recovers 25% of the gap at d = 3 , T = 2 and, over five training seeds and budgets up to 9 × 10 5

Quantum Science and TechnologyVol. 11(4)
Hill College (US)
Quantum Computing Algorithms and Architecture
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon