Joint Belief Propagation Overview

• Joint Belief Propagation is the sum-product algorithm applied to factor graphs that couple multiple latent components, preserving local X/Z correlations in CSS decoding.
• It retains a binary Tanner-graph view while being equivalent to four-state BP, ensuring efficient message passing and accurate posterior estimation without extra computational cost.
• Joint BP is applied across fields such as quantum QC-LDPC decoding, communications receivers, sparse inference, and SLAM, offering iterative soft-information exchange that improves threshold behavior despite inherent error floors.

Joint belief propagation (Joint BP) denotes the sum-product algorithm applied to a factorization that keeps multiple coupled latent components in a single posterior. In CSS syndrome decoding, the posterior can be written as a binary factor graph with two Tanner graphs coupled by the local joint prior at each qubit, and the resulting sum-product algorithm is explicitly called joint belief propagation (Kasai, 6 May 2026). Closely related work in communication receivers, sparse inference, memory detection, and SLAM uses the same structural idea—one factor graph, coupled subproblems, and iterative exchange of soft information—suggesting a broader methodological family of joint BP schemes (Wang et al., 2015, Schniter, 2010, Leitinger et al., 2018).

1. Coupled-binary formulation

For a CSS code with parity-check matrices $H^X,H^Z\in\mathbb F_2^{m\times n}$ satisfying $H^X(H^Z)^T=0$, a Pauli error is represented by two binary vectors $\mathbf x,\mathbf z\in\mathbb F_2^n$, and the measured syndromes satisfy

$$H^X\,\mathbf z = \mathbf s^Z,\qquad H^Z\,\mathbf x = \mathbf s^X.$$

Under a memoryless Pauli channel, the local prior on qubit $j$ is a $2\times2$ table

$Q_j(x,z)=\Pr\{X=x,\;Z=z\text{ on qubit }j\},$

so the posterior factorizes as

$$P(\mathbf x,\mathbf z\mid \mathbf s^X,\mathbf s^Z) \;\propto\; \prod_{j=1}^n Q_j(x_j,z_j)\; \prod_{i=1}^m f_i^X(\mathbf z_{\partial_i^X})\; \prod_{i=1}^m f_i^Z(\mathbf x_{\partial_i^Z}),$$

with binary parity-check factors on the $X$- and $Z$-components and a local joint-prior factor on each qubit (Kasai, 6 May 2026).

This formulation is the defining structural feature of Joint BP in the CSS setting: the two error components are not decoded independently, because the channel prior couples them locally. In the depolarizing case, for example, the prior satisfies $H^X(H^Z)^T=0$0 and $H^X(H^Z)^T=0$1 (Kasai, 6 May 2026). Compared with “separate BP,” which discards the joint prior $H^X(H^Z)^T=0$2, Joint BP retains the exact local $H^X(H^Z)^T=0$3 correlation imposed by the channel (Kasai, 6 May 2026).

2. Message passing, beliefs, and equivalent representations

Joint BP uses standard sum-product updates on the coupled-binary factor graph. Variable-to-check messages on the $H^X(H^Z)^T=0$4 side marginalize over $H^X(H^Z)^T=0$5, and variable-to-check messages on the $H^X(H^Z)^T=0$6 side marginalize over $H^X(H^Z)^T=0$7. Once all incoming check-to-variable messages at qubit $H^X(H^Z)^T=0$8 are available, the joint belief table is

$H^X(H^Z)^T=0$9

and a hard decision is obtained from

$\mathbf x,\mathbf z\in\mathbb F_2^n$0

(Kasai, 6 May 2026).

A central structural result is that Joint BP is exactly equivalent to four-state BP after relabeling each pair $\mathbf x,\mathbf z\in\mathbb F_2^n$1 by a single $\mathbf x,\mathbf z\in\mathbb F_2^n$2 symbol $\mathbf x,\mathbf z\in\mathbb F_2^n$3 and marginalizing the irrelevant binary component. The two algorithms have the same posterior weights, messages, and beliefs after relabeling and marginalization (Kasai, 6 May 2026). Accordingly, Joint BP has no loss of performance relative to four-state BP, but it preserves a purely binary Tanner-graph viewpoint (Kasai, 6 May 2026).

The complexity consequence is equally direct. Per iteration cost is the same order as a pair of classical binary BP decoders, and the overall complexity per iteration is

$\mathbf x,\mathbf z\in\mathbb F_2^n$4

where $\mathbf x,\mathbf z\in\mathbb F_2^n$5 is the total number of edges in the two Tanner graphs (Kasai, 6 May 2026). This makes Joint BP notable not because it changes the asymptotic decoding order, but because it changes what local statistical structure is preserved inside that order.

3. Joint BP as a receiver architecture in communications and storage

Outside CSS decoding, closely related constructions appear whenever estimation and decoding are coupled by a common factor graph.

Domain	Coupled components	Representative paper
Noncoherent UWB-IR	MSDD hidden-Markov chain + SISO channel decoder	(Wang et al., 2015)
Sparse OFDM reception	sparse channel estimation + symbol inference + LDPC decoding	(Schniter, 2010)
Phase-noise/ISI reception	BP PN chain + MF observation factors + BP-EP ISI/code subgraph	(Wang et al., 2016)
ReRAM	sneak-path detector graph + polar-decoder graph	(Sun et al., 2021)

In noncoherent differential UWB-IR, the receiver uses a sliding-window auto-correlation receiver whose window of width $\mathbf x,\mathbf z\in\mathbb F_2^n$6 is advanced one symbol at a time, so that the signal probabilistic model has a hidden Markov chain structure over the whole packet. BP on that factor graph yields a soft-in soft-out MSDD scheme, and replacing $\mathbf x,\mathbf z\in\mathbb F_2^n$7 by decoder priors $\mathbf x,\mathbf z\in\mathbb F_2^n$8 gives a turbo-style joint MSDD and channel decoding scheme. With IEEE 802.15.3a CM2, $\mathbf x,\mathbf z\in\mathbb F_2^n$9, and packet length $$H^X\,\mathbf z = \mathbf s^Z,\qquad H^Z\,\mathbf x = \mathbf s^X.$$0, uncoded M-MSDD with $$H^X\,\mathbf z = \mathbf s^Z,\qquad H^Z\,\mathbf x = \mathbf s^X.$$1 yields $$H^X\,\mathbf z = \mathbf s^Z,\qquad H^Z\,\mathbf x = \mathbf s^X.$$2, $$H^X\,\mathbf z = \mathbf s^Z,\qquad H^Z\,\mathbf x = \mathbf s^X.$$3, $$H^X\,\mathbf z = \mathbf s^Z,\qquad H^Z\,\mathbf x = \mathbf s^X.$$4 dB gains at BER $$H^X\,\mathbf z = \mathbf s^Z,\qquad H^Z\,\mathbf x = \mathbf s^X.$$5 over simple differential detection, is about $$H^X\,\mathbf z = \mathbf s^Z,\qquad H^Z\,\mathbf x = \mathbf s^X.$$6 dB better than B-MSDD of identical complexity $$H^X\,\mathbf z = \mathbf s^Z,\qquad H^Z\,\mathbf x = \mathbf s^X.$$7, and in the coded rate-$$H^X\,\mathbf z = \mathbf s^Z,\qquad H^Z\,\mathbf x = \mathbf s^X.$$8 LDPC system the joint M-MSDD+decoder outperforms joint B-MSDD+decoder by $$H^X\,\mathbf z = \mathbf s^Z,\qquad H^Z\,\mathbf x = \mathbf s^X.$$9 dB at BER $j$0 after 10 turbo iterations (Wang et al., 2015).

In spectrally-efficient OFDM over unknown sparse channels, the joint BP construction combines sparse-channel estimation and data decoding through relaxed BP. The receiver uses a standard bit-interleaved coded OFDM transmitter, achieves $j$1 complexity, empirically realizes the capacity pre-log factor $j$2, and performs near genie-aided bounds at both low and high SNR (Schniter, 2010). In the same design pattern, the BP-MF-EP receiver for joint phase noise estimation, equalization and decoding partitions a nonlinear factor graph into BP, MF, and EP pieces; the second-order Taylor approximation keeps the MF messages Gaussian, the per-iteration complexity is $j$3, and simulations show roughly a $j$4–$j$5 dB gain in PN-MSE and about $j$6–$j$7 dB less SNR to reach BER $j$8 than soft-in EKS (Wang et al., 2016). For ReRAM, the detector graph for sneak-path interference is merged with the BP polar-decoder graph, and the joint detector and decoder, together with genetic-algorithm-designed polar codes, yields a $j$9 dB raw-BER gain for $2\times2$0, about $2\times2$1 dB for $2\times2$2, and roughly $2\times2$3 dB coded gain at BER $2\times2$4 on the $2\times2$5 array (Sun et al., 2021).

These receivers differ in physical model and schedule, but they share a common principle: rather than passing hard decisions between subsystems, they preserve cross-component dependence inside a single iterative inference loop.

4. Joint BP in quantum QC-LDPC decoding

The most explicit recent development of Joint BP is in quantum QC-LDPC decoding. For CSS quantum QC-LDPC codes under depolarizing noise, Joint BP operates on the coupled posterior over the binary $2\times2$6- and $2\times2$7-error components, with syndrome constraints imposed by $2\times2$8 and $2\times2$9 (Komoto et al., 15 Jul 2025). The decoder therefore inherits the local channel correlation of Pauli errors while retaining a binary message alphabet.

Simulation results reported for quantum quasi-cyclic low-density parity-check codes decoded via Joint BP show steep error-rate curves despite the presence of error floors, and the work reports what it calls the first observation of such threshold-like behavior for quantum codes with non-vanishing coding rate, excluding those decoded with non-binary BP decoders (Komoto et al., 15 Jul 2025). For column-weight $Q_j(x,z)=\Pr\{X=x,\;Z=z\text{ on qubit }j\},$0, row-weight $Q_j(x,z)=\Pr\{X=x,\;Z=z\text{ on qubit }j\},$1, and lengths $Q_j(x,z)=\Pr\{X=x,\;Z=z\text{ on qubit }j\},$2 and $Q_j(x,z)=\Pr\{X=x,\;Z=z\text{ on qubit }j\},$3, BER curves for $Q_j(x,z)=\Pr\{X=x,\;Z=z\text{ on qubit }j\},$4 drop from $Q_j(x,z)=\Pr\{X=x,\;Z=z\text{ on qubit }j\},$5 to $Q_j(x,z)=\Pr\{X=x,\;Z=z\text{ on qubit }j\},$6 in a span of $Q_j(x,z)=\Pr\{X=x,\;Z=z\text{ on qubit }j\},$7, with a crossing point around $Q_j(x,z)=\Pr\{X=x,\;Z=z\text{ on qubit }j\},$8. For $Q_j(x,z)=\Pr\{X=x,\;Z=z\text{ on qubit }j\},$9, lengths $$P(\mathbf x,\mathbf z\mid \mathbf s^X,\mathbf s^Z) \;\propto\; \prod_{j=1}^n Q_j(x_j,z_j)\; \prod_{i=1}^m f_i^X(\mathbf z_{\partial_i^X})\; \prod_{i=1}^m f_i^Z(\mathbf x_{\partial_i^Z}),$$0 and $$P(\mathbf x,\mathbf z\mid \mathbf s^X,\mathbf s^Z) \;\propto\; \prod_{j=1}^n Q_j(x_j,z_j)\; \prod_{i=1}^m f_i^X(\mathbf z_{\partial_i^X})\; \prod_{i=1}^m f_i^Z(\mathbf x_{\partial_i^Z}),$$1, the analogous crossing region is around $$P(\mathbf x,\mathbf z\mid \mathbf s^X,\mathbf s^Z) \;\propto\; \prod_{j=1}^n Q_j(x_j,z_j)\; \prod_{i=1}^m f_i^X(\mathbf z_{\partial_i^X})\; \prod_{i=1}^m f_i^Z(\mathbf x_{\partial_i^Z}),$$2 (Komoto et al., 15 Jul 2025).

The same study emphasizes that steep transitions do not remove the low-error-floor regime. At very low $$P(\mathbf x,\mathbf z\mid \mathbf s^X,\mathbf s^Z) \;\propto\; \prod_{j=1}^n Q_j(x_j,z_j)\; \prod_{i=1}^m f_i^X(\mathbf z_{\partial_i^X})\; \prod_{i=1}^m f_i^Z(\mathbf x_{\partial_i^Z}),$$3, residual failures around BER $$P(\mathbf x,\mathbf z\mid \mathbf s^X,\mathbf s^Z) \;\propto\; \prod_{j=1}^n Q_j(x_j,z_j)\; \prod_{i=1}^m f_i^X(\mathbf z_{\partial_i^X})\; \prod_{i=1}^m f_i^Z(\mathbf x_{\partial_i^Z}),$$4 persist, independent of further $$P(\mathbf x,\mathbf z\mid \mathbf s^X,\mathbf s^Z) \;\propto\; \prod_{j=1}^n Q_j(x_j,z_j)\; \prod_{i=1}^m f_i^X(\mathbf z_{\partial_i^X})\; \prod_{i=1}^m f_i^Z(\mathbf x_{\partial_i^Z}),$$5-increase (Komoto et al., 15 Jul 2025). Dominant error events contributing to the error floor typically involve only a small number of bits, and exhaustive examination of residual patterns indicates that the floor is caused by trapping sets—specific subgraph structures in the Tanner graph (Komoto et al., 15 Jul 2025). For $$P(\mathbf x,\mathbf z\mid \mathbf s^X,\mathbf s^Z) \;\propto\; \prod_{j=1}^n Q_j(x_j,z_j)\; \prod_{i=1}^m f_i^X(\mathbf z_{\partial_i^X})\; \prod_{i=1}^m f_i^Z(\mathbf x_{\partial_i^Z}),$$6, $$P(\mathbf x,\mathbf z\mid \mathbf s^X,\mathbf s^Z) \;\propto\; \prod_{j=1}^n Q_j(x_j,z_j)\; \prod_{i=1}^m f_i^X(\mathbf z_{\partial_i^X})\; \prod_{i=1}^m f_i^Z(\mathbf x_{\partial_i^Z}),$$7 of failures involved errors on no more than $$P(\mathbf x,\mathbf z\mid \mathbf s^X,\mathbf s^Z) \;\propto\; \prod_{j=1}^n Q_j(x_j,z_j)\; \prod_{i=1}^m f_i^X(\mathbf z_{\partial_i^X})\; \prod_{i=1}^m f_i^Z(\mathbf x_{\partial_i^Z}),$$8 qubits; for $$P(\mathbf x,\mathbf z\mid \mathbf s^X,\mathbf s^Z) \;\propto\; \prod_{j=1}^n Q_j(x_j,z_j)\; \prod_{i=1}^m f_i^X(\mathbf z_{\partial_i^X})\; \prod_{i=1}^m f_i^Z(\mathbf x_{\partial_i^Z}),$$9, $X$0 involved $X$1 qubits; and the smallest reported trapping sets for $X$2 have $X$3 (Komoto et al., 15 Jul 2025).

A plausible implication is that, in quantum settings, the principal distinction is not between Joint BP and non-binary BP in raw exactness—because of the equivalence theorem—but between graph designs that either create or avoid the small correlated substructures that dominate the error floor.

5. Variational, sparse-inference, and SLAM extensions

Joint BP also appears in generalized forms where BP is combined with other variational or approximate-inference mechanisms. A foundational example is the joint BP/MF method, which partitions factor nodes into $X$4 and $X$5, derives fixed-point equations from a constrained region-based free-energy approximation, and shows that the message-passing fixed points correspond to stationary points of that free energy (Riegler et al., 2011). Under the technical conditions that the BP part is cycle-free and every MF-factor connects to at most one BP variable, an alternating implementation monotonically decreases the total free energy and converges (Riegler et al., 2011). The same framework also accommodates hard constraints in the BP subgraph (Riegler et al., 2011).

In joint sparse recovery, BP is applied to a factor graph whose variable nodes are row vectors $X$6 and support bits $X$7, with row-prior factors $X$8 and measurement factors induced by a common sensing matrix $X$9. The resulting algorithm is developed in exact BP, relaxed BP, and approximate message passing forms, and density evolution yields the condition $Z$0 for exact support and signal recovery in the noiseless, uncorrelated case (Kim et al., 2011). Here the “joint” aspect is the common sparse support across multiple snapshots, coupled directly in the posterior.

In multipath-based SLAM, BP is used for efficient marginalization of a joint posterior over mobile-agent states, legacy and new potential-feature states, and data-association variables. The factor graph couples motion-model chains, feature-existence variables, and exclusion constraints for data association, and the resulting BP-SLAM algorithm has a low computational complexity and scales well in all relevant system parameters (Leitinger et al., 2018). With particle representations, the per-time-step complexity scales as $Z$1 (Leitinger et al., 2018).

This suggests that Joint BP is best understood as a factor-graph design pattern: whenever the posterior contains multiple strongly coupled structures, the method seeks to preserve those couplings long enough for sum-product, or sum-product plus a controlled approximation, to exploit them.

6. Misconceptions, limitations, and design implications

A common misconception is that Joint BP is inherently a non-binary decoder. In the CSS setting, the opposite is demonstrated: Joint BP and four-state BP are exactly the same sum-product computation up to relabeling and marginalization, but Joint BP uses purely binary messages and checks (Kasai, 6 May 2026). This binary viewpoint is practically relevant because it allows the reuse of classical LDPC design and analysis tools, including density evolution and protograph construction, while still capturing full Pauli-error correlation (Kasai, 6 May 2026).

A second misconception is that joint processing eliminates error floors once it produces a steep transition. Quantum QC-LDPC simulations contradict that reading: steep error-rate curves coexist with an error floor, and the dominant failures are associated with small trapping-set subgraphs (Komoto et al., 15 Jul 2025). The data therefore separate two distinct phenomena: threshold-like behavior in the moderate-noise regime and graph-induced trapping in the very-low-noise regime (Komoto et al., 15 Jul 2025).

A third limitation concerns convergence and approximation. In the combined BP/MF framework, convergence is guaranteed only under specific technical conditions; outside those conditions, the method is still defined, but the convergence claim no longer follows from the free-energy argument (Riegler et al., 2011). Communication receivers illustrate the same point in operational form: the UWB joint MSDD/decoder relies on turbo exchange of soft information, and its gains are attributed to overlapping windows, full hidden-Markov-chain BP, and soft-information exchange across the whole packet (Wang et al., 2015). The broader lesson is not that Joint BP is universally optimal, but that its benefit comes from preserving the right couplings and scheduling the right approximations.