Self-Corrected Min-Sum Algorithm

• The Self-Corrected Min-Sum algorithm is a low-complexity iterative decoding scheme for LDPC codes that modifies the traditional Min-Sum approach by erasing messages with sign inconsistencies.
• It achieves near-optimal error performance close to the Sum–Product decoder while retaining computational simplicity and robustness against noise variance mismatches.
• The algorithm leverages localized sign-flip detection to restore symmetric Gaussian message properties, facilitating effective density evolution analysis and practical hardware integration.

The Self-Corrected Min-Sum (SCMS) algorithm is a low-complexity iterative decoding scheme for low-density parity-check (LDPC) codes. It is derived as a simple yet effective modification of the traditional Min-Sum algorithm. SCMS achieves near-optimal performance, approaching that of the Sum–Product (SP) decoding algorithm, by pruning unreliable messages via sign-flip detection at the variable node updates. This design restores the symmetric Gaussian properties required for classical message density analyses, while retaining the implementation advantages of Min-Sum, namely computational simplicity and insensitivity to noise variance estimation errors (0803.1090).

1. Algorithmic Foundation and Rationale

In standard message-passing LDPC decoding, the Sum–Product (SP) algorithm realizes optimal performance through nonlinear computations: each check-node message is formed as an exact "tanh-rule" combination of all incoming variable-node messages, requiring transcendental function evaluations per edge. However, SP is computationally expensive and sensitive to mismatch in assumed channel noise variance.

The Min-Sum algorithm simplifies check-node processing by replacing the SP tanh-combination with a sign product and selection of the minimum-magnitude among inputs. This eliminates transcendental operations and noise-variance dependence but leads to overestimated check-node messages, causing their distributions to deviate from the needed symmetric Gaussianity.

The SCMS algorithm addresses this by introducing a local message-erasure rule at the variable node level. Specifically, any variable-to-check message whose sign fluctuates between two consecutive iterations is erased (set to zero). This process "cleans" unreliable or inconsistent branches in the computation tree. Unlike previous Min-Sum corrections, SCMS does not modify the check-node rule but instead prunes unreliable extrinsic information from the variable nodes, resulting in check nodes receiving messages that are again well-approximated as symmetric Gaussian, and yielding quasi-optimal decoding performance (0803.1090).

2. Detailed Update Rules

Let $\gamma_n$ denote the channel-derived a priori log-likelihood ratio (LLR) at variable node $n$, while the messages $\alpha_{m,n}^l$ (variable-to-check) and $\beta_{m,n}^l$ (check-to-variable) propagate along edge $(n,m)$ at iteration $l$. The SCMS protocol is as follows:

2.1 Initialization

$$\alpha_{m,n}^0 = \gamma_n,\quad \tilde\gamma_n^0 = \gamma_n$$

2.2 Check-Node Update (Min-Sum Rule)

$$\beta_{m,n}^l = \left(\prod_{n' \in \mathcal{H}(m)\setminus \{n\}} \operatorname{sign}(\alpha_{m,n'}^{l-1})\right) \min_{n' \in \mathcal{H}(m)\setminus \{n\}} \left| \alpha_{m,n'}^{l-1} \right|$$

where $\mathcal{H}(m)$ is the set of variable nodes adjacent to check node $m$.

2.3 A Posteriori Update

$$\tilde\gamma_n^l = \gamma_n + \sum_{m\in\mathcal{H}(n)} \beta_{m,n}^l$$

2.4 Variable-Node Update with Self-Correction

For edge $(n, m)$:

• Compute temporary estimate: $$\alpha_{m,n}^{\text{tmp},l} = \tilde\gamma_n^l - \beta_{m,n}^l$$.
• If $$\operatorname{sgn}(\alpha_{m,n}^{\text{tmp},l}) = \operatorname{sgn}(\alpha_{m,n}^{l-1})$$, set $\alpha_{m,n}^l := \alpha_{m,n}^{\text{tmp},l}$.
• Else, set $\alpha_{m,n}^l := 0$ (erase).

Formally,

$$\alpha_{m,n}^l = \begin{cases} \tilde\gamma_n^l - \beta_{m,n}^l & \text{if } (\tilde\gamma_n^l - \beta_{m,n}^l)\,\alpha_{m,n}^{l-1} > 0\ 0 & \text{otherwise} \end{cases}$$

Sign detection is defined such that $\operatorname{sgn}(0)$ may be interpreted as both $+1$ and $-1$, permitting unconditional updating of a zero message.

3. Gaussian-Approximation Analysis

Under the classical Gaussian-approximation framework for message-passing decoders, three key properties are crucial:

• (DG) Density Gaussianity: message densities are i.i.d. symmetric Gaussian,
• (CNP) Check-Node Particle: each check-node output sign is the product of input signs,
• (VNP) Variable-Node Particle: variable updates are the sum of a priori LLR and incoming check-node messages.

For SP and Min-Sum, (CNP) and (VNP) hold; (DG) is violated by Min-Sum due to check-node overestimation and non-Gaussian outputs. The SCMS erasure mechanism eliminates messages prone to sign fluctuation, empirically restoring symmetric Gaussianity to the check-node outputs. This enables density evolution tracking via a one-dimensional recursion analogous to SP: $P_e^{l+1} = \varphi(P_e^{l})$ where

$$\varphi(x) = \sum_{i=2}^{d_v} \lambda_i\, Q\left( \sqrt{ \frac{1}{\sigma^2/2} + (i-1)\, Q^{-1}\left( \frac{1-\rho(1-2x)}{2} \right) } \right)$$

with $\lambda(x) = \sum\lambda_i x^{i-1}$, $\rho(x) = \sum\rho_j x^{j-1}$, and $$Q(x)=\frac{1}{\sqrt{2\pi}} \int_x^\infty e^{-t^2/2}\, dt$$. This alignment explains observed quasi-optimal performance of SCMS (0803.1090).

4. Complexity Considerations

Decoder	Per-Edge Operations	Additional Resources
Sum–Product	2 $\tanh(\cdot)$, 1 mult, 2 summations	Noise-variance dependent
Min-Sum	1 comparison over $d_c$, sign-product	None
SCMS	Min-Sum + sign-comparison, zeroing	1-bit sign storage per edge

SCMS requires only the operations of Min-Sum with the addition of one sign comparison and conditional zeroing per edge per iteration—implemented as a single bitwise XOR and multiplexing operation. The required memory increase is negligible, limited to storing the previous sign for each edge (0803.1090).

5. Performance Evaluation

Monte-Carlo simulations evaluated SCMS on both irregular LDPC ensembles and practical quasi-cyclic (QC-LDPC) codes such as IEEE-802.16e (WiMAX), across the AWGN channel using BPSK/QPSK modulation at rate $1/2$.

• For irregular LDPCs (length 2304, 8064), SCMS maintains bit error rate (BER) and frame error rate (FER) within 0.05 dB (2304 bits) and 0.10 dB (8064 bits) of SP. At high SNR, SCMS can surpass SP in the error floor regime where SP exhibits $10^{-6}$ FER.
• Normalized Min-Sum (Min-Sum with optimized normalization) approximates SCMS at low SNR but exhibits a sharper FER floor.
• For WiMAX QC-LDPC (length 2304), fixed-point SCMS (4-bit dynamic range, 0.25 step) is within 0.1 dB of floating-point SP up to FER $10^{-3}$, and can outperform SP in error-floor regions (0803.1090).

6. Implementation Pseudocode

A compact description of a single SCMS iteration is as follows:

for each check node m: for each neighbor n in H(m): sign_mn = ∏_{n'≠n} sign(α_{m,n'}) min_mn = min_{n'≠n} |α_{m,n'}| β_{m,n} = sign_mn · min_mn for each variable node n: compute \tilde γ_n = γ_n + ∑_{m∈H(n)} β_{m,n} for each neighbor m in H(n): α_tmp = \tilde γ_n − β_{m,n} if sign(α_tmp) == sign(α_{m,n}) or α_{m,n} == 0 α_{m,n} = α_tmp else α_{m,n} = 0

This demonstrates that SCMS can be adapted to existing Min-Sum hardware with minimal modification. Only one extra bit per edge (to store the previous sign) and a per-edge sign comparison are required (0803.1090).

7. Interpretation and Broader Implications

SCMS achieves nearly all of the SP decoder's coding gain at considerably reduced computational complexity and with complete insensitivity to noise-variance mismatch. The local sign-based message erasure is both effective and resource-light, requiring only marginal hardware and memory augmentation. The inherent generality of the erasure principle—removing sign-flip messages—may be relevant in other iterative decoders on cyclic graphs. Empirical evidence also suggests that in fixed-point implementations, SCMS not only preserves decoding performance but can outperform both Min-Sum variants and SP in the error-floor regime for practical codes such as those used in WiMAX (0803.1090).