Automorphism-Ensemble Decoding (AED)

• Automorphism-Ensemble Decoding (AED) is an ensemble decoding approach that leverages a code's automorphism group to break symmetry in belief propagation decoders.
• It employs techniques like row-addition, overcompleteness, or undercompleteness to disrupt decoder equivariance, yielding up to 0.3 dB gain in block-error rate and lower decoding latency.
• AED extends naturally to general linear codes, offering hardware-friendly parallelism and minimal complexity while ensuring valid codeword recovery without transmitter changes.

Automorphism-Ensemble Decoding (AED) is an ensemble-decoding architecture for modern error-correcting codes that leverages the code’s automorphism group to achieve substantial performance gains—specifically by mitigating symmetry-induced limitations of conventional iterative decoders. AED is most prominent for quasi-cyclic low-density parity-check (QC-LDPC) codes, but its theoretical framework extends naturally to general linear codes. By introducing symmetry-breaking in the receiver’s graphical decoding, AED achieves up to 0.3 dB block-error-rate gains and dramatic reductions in average decoding latency relative to other BP-based list decoders, without requiring transmitter modifications or latency overhead (Geiselhart et al., 2022).

1. Automorphism Group Structure and BP Equivariance

Let $C$ be a binary linear code of length $N$. Its automorphism group is

$$\mathrm{Aut}(C) = \{ \pi \in S_N : \pi(c) \in C \ \forall \ c \in C \},$$

with $S_N$ the symmetric group. For QC-LDPC codes, $N = nZ$ ($n$: block size, $Z$: circulant size), and the automorphism group includes the subgroup $Q_Z$ of length-$Z$ cyclic shifts

$$\pi_{d,Z}(i) = \begin{cases} i+d-Z, & \text{if } (i \bmod Z)+d \ge Z \ i+d, & \text{otherwise} \end{cases}, \quad d=0, \dots, Z-1.$$

Standard flooding or layered BP decoding on the canonical QC-Tanner graph is equivariant w.r.t. these shifts: for all $N$0,

$N$1

where $N$2 is the channel LLR vector. This equivariance implies that running multiple BP decoders in parallel on cyclically shifted input yields identical outputs modulo the shift, rendering naive ensemble approaches ineffective (Geiselhart et al., 2022).

2. Breaking Decoder Symmetry: Rationale and Techniques

Classical automorphism-based ensemble decoding operates by permuting the LLR input using a collection of distinct automorphisms $N$3, decoding via BP, inverting the permutation on the outputs, and selecting the best candidate (typically via an ML-metric). However, if the decoding algorithm is equivariant to $N$4—as with BP on unmodified QC-LDPC graphs—every permutation leads to the same decoded codeword and no ensemble gain is possible:

$N$5

Therefore, to realize ensemble gains, one must break this symmetry at the decoder, while preserving codeword validity.

Three decoder-side modifications for breaking QC equivariance are proposed (all on the lifted parity-check matrix $N$6):

• Row-addition: Replace row $N$7 by $N$8 for some pair $N$9.
• Overcomplete: Append linear combinations of existing rows as auxiliary checks.
• Undercomplete: Remove one or more rows (most simply, delete an entire $$\mathrm{Aut}(C) = \{ \pi \in S_N : \pi(c) \in C \ \forall \ c \in C \},$$0-row block in $$\mathrm{Aut}(C) = \{ \pi \in S_N : \pi(c) \in C \ \forall \ c \in C \},$$1).

The undercomplete variant merely disables one check node in the QC-Tanner graph. All three variants modify only the decoding factor graph, with the code definition (as a row-span of $$\mathrm{Aut}(C) = \{ \pi \in S_N : \pi(c) \in C \ \forall \ c \in C \},$$2) unchanged—guaranteeing codeword integrity at the receiver. This intervention is sufficient to destroy BP equivariance under QC shifts (Geiselhart et al., 2022).

3. AED Workflow and Ensemble Decoding Procedure

For QC-LDPC codes with circulant size $$\mathrm{Aut}(C) = \{ \pi \in S_N : \pi(c) \in C \ \forall \ c \in C \},$$3, AED operates as follows:

1. Inputs: $$\mathrm{Aut}(C) = \{ \pi \in S_N : \pi(c) \in C \ \forall \ c \in C \},$$4 (channel LLRs), set of QC shifts $$\mathrm{Aut}(C) = \{ \pi \in S_N : \pi(c) \in C \ \forall \ c \in C \},$$5, modified check matrix $$\mathrm{Aut}(C) = \{ \pi \in S_N : \pi(c) \in C \ \forall \ c \in C \},$$6 ($$\mathrm{Aut}(C) = \{ \pi \in S_N : \pi(c) \in C \ \forall \ c \in C \},$$7 with a row dropped), maximum BP iterations $$\mathrm{Aut}(C) = \{ \pi \in S_N : \pi(c) \in C \ \forall \ c \in C \},$$8.
2. Parallel Decoding: For each $$\mathrm{Aut}(C) = \{ \pi \in S_N : \pi(c) \in C \ \forall \ c \in C \},$$9,
   • Permute: $S_N$0,
   • Decode: $S_N$1,
   • Un-permute: $S_N$2,
   • Metric: $S_N$3.
3. Selection: Output candidate $S_N$4 for $S_N$5.

Since the code is unchanged at the transmitter, and all decoders run in parallel, the worst-case latency is that of a single BP decoder with $S_N$6 iterations plus negligible overhead for permutations and metric calculation (Geiselhart et al., 2022).

4. Performance Benchmarks and Complexity

AED, evaluated on CCSDS, 802.11n, and 5G QC-LDPC codes under BI-AWGN channels and floating-point SPA (flooding), achieves:

• Block-error-rate (BLER) gain: $S_N$7–$S_N$8 dB at BLER ≈ $S_N$9 over standard BP decoding,
• Comparison to Saturated BP (SBP) List Decoding: Matches SBP's gain, but reduces average decoding latency by more than $N = nZ$0 (for CCSDS(128,64): AED-16 average 3.4 iterations vs. SBP average 28.5 at $N = nZ$1 dB).
• Throughput: No increase in worst-case latency, as all $N = nZ$2 BP decoders are run in parallel.

No code modifications, no added edges except for the dropped (or added) row in $N = nZ$3, and minimal hardware changes—one check node is disabled in hardware—are required. Storage is minimal: a single $N = nZ$4 and a per-decoder flag for the disabled check node suffice. Each candidate outputs a valid codeword (checked against the original $N = nZ$5 for the undercomplete variant) (Geiselhart et al., 2022).

Code Family	Blocklength $N = nZ$6	Rate $N = nZ$7	Circulant $N = nZ$8	AED gain at BLER $N = nZ$9
CCSDS	128	1/2	16	≈ 0.20 dB
CCSDS	256	1/2	32	≈ 0.20 dB
802.11n	648	5/6	27	≈ 0.20 dB
5G BG2	132	1/2	11	≈ 0.30 dB
5G BG2	264	1/2	22	≈ 0.20 dB

A similar gain to SBP is achieved with dramatically reduced average iteration count and hardware complexity (Geiselhart et al., 2022).

5. Practical Implementation and Design Remarks

AED’s core advantage is its simplicity: all BP decoders run identical code structure (except the disabled row), and no on-the-fly Tanner graph permutations or edge rewirings are needed. The ensemble size $n$0 (typically $n$1) is hardware-friendly, and the permutation logic constitutes trivial reindexing. Average latency is reduced due to consistent early stopping across ensemble branches, as soon as any constituent BP instance finds a valid codeword.

In contrast to list-SBP or BP-based permutation decoding that require complex reliability metrics, sorting, and potentially many extra decoding attempts, AED's selection is purely a Euclidean-metric-based winner-takes-all among a modest list. This regularity is advantageous for hardware pipelining and area efficiency.

6. Extension to Generalized and Non-QC Codes

While the mechanism of AED for QC-LDPC hinges on breaking the decoder's natural equivariance under code automorphisms, the general principle extends to other linear codes and automorphism-rich structures (including cyclic and Reed–Muller codes), provided one ensures candidate-generating permutations are “non-absorbed” by the base decoder's symmetry group. In all cases, the theoretical foundation is that ensemble decoding gains are only realized if the ensemble contains genuinely diverse decoding “views,” not factorizable by decoder equivariance (Geiselhart et al., 2022).

7. Summary and Impact

Automorphism-Ensemble Decoding, by purposely breaking unintended symmetry in the BP decoder’s check-matrix, converts absorbed automorphisms into meaningful decoding diversity. On prominent industrial QC-LDPC codes (CCSDS, 802.11n, 5G), this technique reliably yields 0.2–0.3 dB gains at BLER $n$2, with no transmitter change and no increase in worst-case latency. AED dramatically reduces average latency compared to SBP list decoding and is hardware efficient owing to its regularity and low memory requirement.

AED sits at an attractive trade-off point in the performance–complexity–latency landscape for short-block LDPC codes: simple, parallelizable, and predictable, while substantially closing the BP–ML gap (Geiselhart et al., 2022).