The Signal
Correlated Decoding Can Cost Surface Codes Effective Distance
A study posted to arXiv on October 19 finds that standard correlated decoding can worsen surface-code logical error rates far below threshold. The authors’ simulations show that an error-rate-dependent adjustment restores the decoder’s adva
Quantum.world Team · Written with AI · fact-checked ·
“Under circuit-level noise, a single first-pass edge can discount several correlated edges, even though their joint occurrence requires additional independent faults, so the decoder favors an over-discounted correction.”
César Hernando-Fuente, Joonas Majaniemi, Evert van Nieuwenburg and Felix Frohnert, arXiv paper
A paper submitted to arXiv on October 19 finds that a standard way to use correlations in surface-code decoding can make logical errors more likely when physical error rates are very low. In simulations, correlated matching and a correlated clustering decoder both lose accuracy against their uncorrelated versions in that regime. The authors say the effect amounts to a one-unit loss in effective code distance: the logical error rate falls with physical error probability more slowly than the code’s nominal distance predicts.
When a second pass discounts too much
Surface-code decoders often separate measurement outcomes into two graphs, one for each type of stabilizer check, and decode them independently. Correlated decoding tries to recover information lost in that split: after an initial pass, it adjusts the weights of edges in one graph using the correction selected in the other. The paper tests correlated matching, based on minimum-weight perfect matching, and the Local Clustering Decoder, a weighted union-find method.
The failure comes from how the standard update handles circuit-level noise. One edge selected on the first pass can be associated with several different error mechanisms, each linked to a different edge in the other graph. The decoder discounts all those partner edges, then can treat multiple discounted edges as though the same first-pass event made them jointly likely. In reality, explaining the whole correction may require additional independent faults, which should make it less probable.
The authors trace the problem to a concrete case that already appears at distance three: an XY error on a CZ gate creates correlated edges in both graphs. A second-pass correction can then choose a longer, logically wrong chain because its discounted edges look too cheap. Under the standard rule, the decoder’s assigned cost misses the probability penalty for an extra fault.
The low-error penalty shows up in the simulations
The simulations model d-round memory experiments on rotated surface codes under SI1000 circuit-level noise, with additional tests under uniform depolarizing noise. At a physical CZ-gate error rate of 10−5, distance-three correlated matching fails about 4.4 times as often as ordinary matching. More than 80% of those correlated-decoder failures come from syndromes that the uncorrelated decoder corrects.
As the physical error rate falls, the smallest fault configurations increasingly determine the logical error rate. For a distance-d code, the expected low-error scaling is proportional to p(d+1)/2; the standard correlated decoders instead trend toward p(d−1)/2. That shift is the reported unit loss of effective distance. The reversal is not apparent in the same way near threshold, where more numerous, larger fault patterns can dominate and correlated decoding may retain its advantage.
The study also finds a crossover for distance five: standard correlated matching becomes worse than uncorrelated matching around p = 4 × 10−5, and its error rate is roughly 1.5 times higher around 2 × 10−5. The authors report similar low-error degradation for the clustering decoder, suggesting that the issue lies in the reweighting rule rather than one particular decoding algorithm.
Regularization keeps the correction’s cost honest
The proposed fix scales each conditional probability used in reweighting by an adjustable factor, α. Lowering that factor at low physical error rates makes the discount less aggressive, adding a penalty when a correction uses several edges tied to the same first-pass edge. The team fits α as a function of error rate and tests the resulting rule against both decoder baselines.
At p = 10−4, regularized correlated matching reduces the logical error rate to about 0.55 times that of ordinary matching at distance five, and 0.35 times at distance seven. The standard correlated version reaches ratios of about 0.73 and 0.45 at those same distances. A ratio below one means fewer logical failures than the uncorrelated reference. The authors say the regularized update also restores the expected distance scaling in the cases studied.
The adjustment is computed during initialization, leaving the underlying matching and clustering routines unchanged, according to the paper. But the results are simulations, not a demonstration on quantum hardware, and the best regularization strength is an empirical fit. The authors leave open how to incorporate correlations broadly without compromising low-error performance; their specific finding is that every discounted edge still has to carry the cost of any additional independent fault needed to explain it.