Cross-Correlation Decoder in CSS Codes

• Cross-Correlation Decoder is a strategy for CSS codes that exploits statistical correlations between X and Z errors, leading to improved error thresholds.
• It incorporates erasure masks in Z-error decoding by assigning zero cost to errors paired with detected X faults, using modified minimum-weight perfect matching.
• Empirical results demonstrate significant threshold gains in depolarizing noise, notably benefiting triangular toric and asymmetric surface code constructions with modest overhead.

A cross-correlation decoder is a decoding strategy for Calderbank-Shor-Steane (CSS) quantum error-correcting codes that directly exploits the statistical correlation between the bit-flip ($X$) and phase-flip ($Z$) components of physical errors under the single-qubit depolarizing noise model. In contrast to standard CSS decoders, which treat the $X$ and $Z$ parts of the error independently, the cross-correlation decoder uses information obtained during $X$-error decoding to inform and constrain the $Z$-error decoding, effectively increasing logical error thresholds, especially in asymmetric surface code constructions. This decoder achieves threshold improvements with only modest computational overhead and can be efficiently implemented using variants of the minimum-weight perfect matching algorithm (Delfosse et al., 2014).

1. Noise Model and Error Decomposition

Consider an $n$-qubit CSS code subjected to independent single-qubit depolarizing noise with probability $p$ per qubit. At each site $i$, errors $E_i$ are drawn from $Z$0 with

$Z$1

Each Pauli error $Z$2 decomposes as $Z$3 with $Z$4 and $Z$5. Marginals are therefore

$Z$6

The conditional distribution $Z$7 is characterized as follows: $Z$8 This highlights the key correlation: where an $Z$9-error occurs, the $X$0-error is no longer distributed as a binary symmetric channel (BSC), but is instead an erasure, i.e., completely random.

2. Standard Versus Correlated Decoding Approaches

In conventional CSS decoders, $X$1 and $X$2 are estimated independently using syndrome information associated with the dual and primal graphs, respectively. Each syndrome component—$X$3 (for $X$4-errors) and $X$5 (for $X$6-errors)—labels the locations of error endpoints. Decoding is typically performed by matching these endpoints via minimum-weight perfect matching (PMA), where the paths correspond to likely error chains.

The cross-correlation decoder augments this process by, after estimating $X$7, propagating that information to the $X$8-error step. The $X$9-error estimate defines an erasure set for the $Z$0-decoding: qubits with detected $Z$1-errors provide no further $Z$2-information except that $Z$3 is uniformly random at those locations.

3. Cross-Correlation Cost Function and Erasures

Let $Z$4 denote the set of edges in the primal graph $Z$5 where the estimated $Z$6-error acts nontrivially. For a candidate $Z$7-error support $Z$8, the cross-correlation cost is

$Z$9

This cost reflects the log-likelihood ratio for $X$0: placing a $X$1-error on a site that already has an $X$2-error (i.e., an erasure) incurs zero cost, while placing a $X$3-error elsewhere incurs unit cost. Thus, the decoder seeks a $X$4-error configuration that minimizes $X$5, i.e., the number of $X$6-errors on non-erased sites.

4. Algorithmic Integration with Minimum-Weight Perfect Matching

The full decoding scheme, termed "Correlated-PMA" in the source, consists of the following steps:

1. $X$7-Error Decoding: Construct the weighted distance graph over $X$8 (dual-graph syndrome) using dual-graph distances. Solve the minimum-weight matching and extract shortest paths, returning the estimated $X$9 as the symmetric difference of these paths.
2. $Z$0-Error Decoding with Erasures: Treat the previously determined $Z$1 as the erasure set $Z$2. For each pair in $Z$3 (primal-graph syndrome), define the $Z$4-distance as $Z$5, minimizing path cost with erased edges at cost zero. Solve for the minimum-weight matching over $Z$6 with these adjusted distances, then recover $Z$7 as the symmetric difference of the corresponding $Z$8-geodesic paths.
3. Return Final Estimate: The decoder outputs $Z$9 as the Pauli error estimates.

All graph constructions and path extractions utilize $n$0 BFS or Dijkstra calls; matching is efficiently handled using optimized algorithms such as Blossom V, with overall runtime of $n$1 to $n$2 for families of dimension $n$3.

5. Performance and Threshold Improvements

Empirical testing on triangular toric codes of length $n$4 under depolarizing noise reveals a significant threshold gain:

• Standard PMA (independent $n$5/$n$6 decoding): depolarizing threshold $n$7
• Correlated-PMA: depolarizing threshold $n$8

On square toric codes, correlated decoding provides a slight improvement over the standard PMA threshold $n$9, though the optimal threshold with full degeneracy exploitation approaches $p$0. The improvement is particularly notable in asymmetric surface code constructions where primal and dual connectivities differ substantially.

Code Type	Standard PMA Threshold	Correlated-PMA Threshold
Triangular toric	$p$1	$p$2
Square toric	$p$3	Slight gain

6. Computational Complexity and Resource Requirements

The cross-correlation decoding process essentially doubles the number of perfect matching calls compared to independent decoding. However, additional overhead—marking erasures and extracting $p$4-geodesics—is negligible ($p$5 in both cases). The principal runtime is dominated by matching and path-finding, remaining within feasible bounds for practical code sizes. Memory requirements increase only by a single erasure bit-mask.

Trade-offs include:

• Slightly higher latency (two PMAs per component instead of one)
• Reliance on fidelity of the $p$6 estimate; an incorrect $p$7-error decoding can increase erasure rates during $p$8-decoding and degrade local performance

A plausible implication is that the decoder's effectiveness in approaching hashing-bound rates is pronounced for highly asymmetric CSS constructions.

7. Significance and Context Within Quantum Error Correction

The cross-correlation decoder for CSS codes demonstrates a concrete technique by which physical error model correlations can be leveraged algorithmically to enhance error thresholds. In contrast to decoders that ignore $p$9 correlations, this approach directly encodes knowledge of the underlying noise structure into the decoding metric, yielding measurable improvements especially in tailored code architectures. The methods are computationally accessible, make minimal additional demands on resources, and are compatible with widely implemented minimum-weight matching solvers. For depolarizing noise on surface codes, the correlated decoder narrows the gap to fundamental code capacity thresholds (Delfosse et al., 2014).