A Factor-Graph Formulation of CSS Syndrome Decoding: Joint BP and Four-State BP

Abstract: For CSS syndrome decoding, the two check matrices impose binary parity-check constraints on the two Pauli error components. The posterior can therefore be written as a binary factor graph with two Tanner graphs coupled by the local joint prior at each qubit. We call the sum-product algorithm on this factorization joint belief propagation (joint BP). Joint BP retains the local channel correlation between the two Pauli components. This note compares joint BP with the four-state Pauli-label factor graph used for four-state BP. The two algorithms are shown to have the same posterior weights, messages, and beliefs after relabeling the four local Pauli states and marginalizing the irrelevant binary component.

• The paper shows that joint BP and four-state BP yield identical CSS decoding performance when local X/Z error correlations are preserved.
• It establishes an invertible mapping between binary and F4 factor-graph representations, ensuring equivalent posterior inference and message updates.
• The work highlights that retaining joint prior information improves decoding over separate BP, enabling efficient adaptations of classical LDPC techniques.

A Factor-Graph Formulation for CSS Syndrome Decoding: Joint BP and Four-State BP

Overview

This paper analyzes two distinct but equivalent factor-graph representations for the Sum-Product Algorithm (SPA, or Belief Propagation, BP) in the context of Calderbank-Shor-Steane (CSS) codes for quantum error correction. The focus is on syndrome decoding, which seeks to infer Pauli error configurations consistent with measurement outcomes. The core result is an equivalence theorem: joint BP on a coupled binary factor graph and four-state BP on an $\mathbb{F}_4$ factor graph produce identical posterior beliefs, message updates, and decoding decisions up to relabeling and marginalization. This result clarifies the role of local $X/Z$ correlations and demonstrates that, for CSS codes, the binary factor-graph representation is both natural and sufficient for representing all channel correlations, challenging the widespread association of four-state (quaternary) BP with improved correlation handling.

Factor-Graph Structures for CSS Codes

CSS codes are defined by two binary check matrices, $H^X$ and $H^Z$, operating on $n$ physical qubits. The standard approach to syndrome decoding involves calculating the posterior probability of Pauli errors given the syndrome. In the factor-graph formalism, this factorization can be represented in two ways:

1. Coupled Binary Factor Graph (Joint BP):
   • The variables are pairs $(x_j, z_j)$, for $x_j, z_j \in \mathbb{F}_2$, representing $X$ and $Z$-components of Pauli errors.
   • The local factor $Q_j(x_j, z_j)$ captures the joint prior for physical qubit $X/Z$0, including all local channel correlations.
   • Tanner graphs for $X/Z$1 and $X/Z$2 are connected locally at each qubit by $X/Z$3.
   • SPA proceeds on this factor graph, updating variable-to-factor and factor-to-variable messages using the full joint local information.
2. Four-State Factor Graph ($X/Z$4 or Pauli Label, Four-State BP):
   • The pair $X/Z$5 is relabeled as $X/Z$6, where $X/Z$7.
   • Nodes represent single four-state variables per qubit; the prior is $X/Z$8, corresponding to the same joint prior in the new label.
   • Check nodes depend on appropriate binary projections of $X/Z$9, i.e., $H^X$0 or $H^X$1 as needed for $H^X$2 or $H^X$3.
   • SPA proceeds analogously, with all local structure preserved.

An alternative, less expressive baseline is separate BP, where the local prior $H^X$4 is replaced by marginals $H^X$5 and $H^X$6. This approach discards $H^X$7 correlations, and its two decoders are completely disconnected.

Theoretical Results: Equivalence of Joint BP and Four-State BP

A formal equivalence theorem demonstrates that joint BP (binary factor graph with joint prior) and four-state BP ($H^X$8 factor graph) perform the same computation for CSS posterior inference, up to an invertible relabeling and marginalization of messages. Specifically:

• The posterior weights assigned to any error configuration are identical under the mapping $H^X$9.
• Each sum-product message in the four-state representation, when marginalized over the unused binary component, corresponds exactly to its binary counterpart.
• The variable node beliefs in both representations coincide after relabeling, ensuring identical hard decisions when applying MAP selection.
• The entire SPA message-passing process is preserved; not just the final estimates, but all intermediate messages and recursion steps.

The proof leverages the observation that, for CSS codes, check constraints in the four-state graph depend on only one binary component per node and can be separated. Therefore, summing out the irrelevant component in variable-to-check messages under SPA exactly produces the coupled-binary update.

Numerical and Algorithmic Implications

• Local $H^Z$0 Correlations: Both representations fully capture and utilize local channel correlations between $H^Z$1 and $H^Z$2 errors at each qubit, provided the joint prior $H^Z$3 is retained.
• Complexity: Both joint BP and four-state BP compute with the same information content. The choice of representation (binary or quaternary) affects only the state labeling—not decoding strength or accuracy.
• Message Updates: The paper presents log-likelihood-ratio (LLR) domain formulas for joint BP, facilitating efficient implementation and confirming that all standard practical SPA/PB optimizations (variable/factor scheduling, log-domain updates) apply straightforwardly to the coupled binary representation.

Practical and Theoretical Implications

The result refines the common understanding of decoding quantum LDPC codes:

• CSS Codes: For CSS codes, the coupled binary factor graph with the joint local prior is the most natural factor-graph representation, preserving both the Tanner graph structure and all local correlations. There is no advantage to representing the problem in $H^Z$4; the "quaternary" BP is just a relabeling.
• General Stabilizer Codes: For non-CSS stabilizer codes, checks generally couple both $H^Z$5 and $H^Z$6 components nonlocally, and a quaternary or higher-field factor-graph representation is indeed natural, as joint/coupled binary structure may not capture correlations arising from non-CSS constraints.
• Algorithm Design: Optimizations and analysis techniques from the classical LDPC literature (density evolution, irregular ensemble design, etc.) carry seamlessly to joint BP for quantum LDPC codes, provided the joint prior is retained.
• Comparison to Separate BP: The wide practice in quantum LDPC decoding literature of using "binary BP" as separate BP—with independent decoders for $H^Z$7 and $H^Z$8—may discard correlation information and lead to suboptimal decoding in correlated Pauli channels. This work clarifies that binary representation alone is not the source of suboptimality, but rather the loss of joint prior information.

Future Directions

• Implementation: The joint BP representation facilitates efficient message scheduling and leverages existing classical BP infrastructure, suggesting practical speedups or memory reductions for syndrome decoders in CSS-based fault-tolerant quantum codes.
• Code Design: Understanding the equivalence allows optimization and ensemble design techniques developed for binary LDPC codes to be adapted for quantum CSS codes with full fidelity.
• Beyond CSS: The sharp distinction elucidated in this work motivates the development of similarly natural and efficient graphical decoding formulations for non-CSS quantum codes and for general Pauli channels.

Conclusion

This paper rigorously delineates the relationship between binary and four-state BP for CSS syndrome decoding. It establishes that, as long as the joint channel statistics are respected in the factor graph, binary joint BP and four-state BP are algorithmically identical for CSS codes. The distinguishing feature for accurate quantum LDPC decoding is the retention of local $H^Z$9 error correlations, not the superficial use of binary or quaternary messages. These results support adopting the coupled binary factor-graph perspective as the standard for CSS codes, with implications for both practical decoder implementation and the interpretation of simulation results in the quantum LDPC literature.