Automorphism Ensemble Decoding (AED)

• Automorphism Ensemble Decoding (AED) is an ensemble method that leverages code automorphisms to create multiple symmetry-altered decoding views from a single received word.
• It improves performance in short blocklength regimes by running parallel decoders on permuted inputs, yielding near-ML candidate lists without full list-decoding overhead.
• AED balances trade-offs among decoding accuracy, hardware latency, and complexity by exploiting non-invariant decoder properties to select the most likely valid codeword.

Searching arXiv for recent AED papers to ground the article and citations. {"query":"Automorphism Ensemble Decoding polar codes LDPC arXiv", "max_results": 10} Automorphism Ensemble Decoding (AED) is an ensemble-decoding method that exploits code automorphisms to generate multiple symmetry-related decoding views of the same received word and then selects a final estimate from the resulting candidate list. In the short-blocklength regime, it has been developed as a way to improve suboptimal constituent decoders without incurring the full path-management overhead of list decoding, and for pre-transformed polar codes it can improve performance for a fixed list size by running multiple parallel SCL decodings on permuted received words, yielding a list of estimates from which the final estimate is selected; the same abstract also states that AED is limited to appropriately designed polar codes (Krieg et al., 2024, Lulei et al., 24 Apr 2025).

1. Core decoding mechanism

AED is a particular realization of the broader ensemble-decoding framework in which $M$ independent constituent decoders operate in parallel on transformed versions of the same observation, each proposing a codeword candidate, after which a final ML-in-the-list decision is taken over the valid outputs (Krieg et al., 2024). In the standard classical formulation, branch $j$ uses an automorphism $\pi_j \in \operatorname{Aut}(\mathcal C)$ and computes

$$\hat{c}_j = \pi^{-1}_j\left(\operatorname{Dec}(\pi_j(\mathbf{L}_{\mathrm{ch}}))\right).$$

For parity-check-defined codes, only valid candidates are retained; one explicit validity condition used in the literature is the zero-syndrome test

$s_j=\hat{c}_jH^\mathrm{T}=\mathbf{0}.$

The final output is then chosen by a channel-metric rule over the surviving list (Krieg et al., 2024).

The defining feature of AED is that it does not average soft outputs across branches. Instead, it uses automorphisms to create multiple transformed decoding problems, collects hard codeword candidates, and performs a final selection. This is why AED is often described both as a permutation-based decoder and as a list-of-candidates decoder. For LDPC decoding, the same operation can also be viewed as a decoder transformation: applying an automorphism to the observation is equivalent to decoding with a column-permuted parity-check representation, so AED becomes a special case of multiple-bases belief propagation (MBBP) in which the branch parity-check matrices are related by automorphism-induced column permutations (Geiselhart et al., 2022).

2. Automorphisms, equivariance, and diversity conditions

The algebraic object underlying AED is the automorphism group of the code,

$$\operatorname{Aut}(\mathcal{C})= \left\{\pi : \pi(\mathbf{c}) \in \mathcal{C} \; \forall \mathbf{c} \in \mathcal{C}\right\},$$

that is, the set of coordinate permutations mapping codewords to codewords (Krieg et al., 2024). AED relies on the fact that these symmetries preserve code membership while altering the coordinate view presented to the decoder.

However, code symmetry alone is not sufficient. AED is useful only when the decoder is not equivariant to the same transformations. One formulation given in the literature is that AED works “as long as the decoder is not invariant to this transformation,” and therefore “the code needs to have more symmetries than the decoder for AED to work” (Krieg et al., 2024). For QC-LDPC codes this issue is especially sharp: standard BP on the conventional factor graph is equivariant to the quasi-cyclic automorphisms, so naive QC permutations generate no decoding diversity. This is why receiver-side symmetry breaking by row deletion, row addition, or row summation was proposed for QC-LDPC AED (Geiselhart et al., 2022).

For polar codes, the same theme appears in a more refined form. SC-based AED is not improved by automorphisms that are absorbed by SC decoding, and the complete SC-invariant affine automorphism subgroup for decreasing polar codes was characterized as a BLTA subgroup computable from the code’s information set (Ye et al., 2022). This immediately implies that AED performance depends not simply on the size of the automorphism group, but on the quotient between code symmetries and decoder-invariant symmetries. A related line of work on full automorphism groups of polar codes emphasizes that useful AED design may require going beyond previously studied affine subclasses, including non-affine automorphisms for some families (Ma et al., 2024).

3. Representative realizations across code families

AED has been instantiated with different constituent decoders and different transformed objects, but the ensemble structure remains stable across settings.

Setting	Branch transformation	Final selection
Short LDPC codes	Permute $\mathbf{L}_{\mathrm{ch}}$, decode by BP, inverse-permute output	ML-in-the-list over zero-syndrome candidates
Polar and RM codes	Permute received word, decode by SC, SCL, BP, or Fast-SSC variant	Most likely candidate or branch metric
Quantum LDPC / CSS codes	Transform syndrome by $s_A = U_A \cdot s$, decode in transformed representation	Minimum-weight or prior-likelihood candidate

Within short-length LDPC decoding, AED is one of the structure-exploiting ensemble methods that “typically show greater performance improvements” than perturbation-based alternatives such as noise-aided or saturation-based ensembles (Krieg et al., 2024). On the $(63,6)$ simplex code, AED-8 is reported as the best ensemble and as achieving near-ML performance; on the $(273,191)$ projective-geometry LDPC code, AED and MBBP are essentially tied and are the best performers; on the $j$0 5G QC-LDPC code, AED still improves over BP but only modestly because the useful symmetry mismatch is weak (Krieg et al., 2024).

For QC-LDPC codes, the explicit symmetry-breaking construction makes AED operational on standardized families without changing the transmitted code. Using an undercomplete parity-check matrix formed, for example, by deleting one check, the method yields gains of $j$1 to $j$2 dB over conventional BP decoding on CCSDS, 802.11n, and 5G codes, while the paper also reports more than an eightfold reduction in average decoding latency compared with similarly performing saturated BP (Geiselhart et al., 2022).

For short polar-like codes, AED-compatible constructions are often tied to Reed–Muller structure. Partially-Symmetric Reed-Muller (PS-RM) codes were introduced as RM-polar and $j$3-symmetric codes with $j$4, and for $j$5 they are reported to exist for almost all code dimensions. Their SC absorption-group analysis shows that valuable permutations for AE-SC always exist in the range $j$6, and a $j$7 PS-RM code under AE-8-SC is reported to outperform a 5G polar code under CA-SCL-8 by about $j$8 dB (Pillet et al., 2022). Shortened polar codes provide a more constrained variant: their automorphism group is “limited but non-empty,” which still makes AE decoding possible, and the reported $j$9 shortened polar examples show AE-4-BP outperforming SCL-4 by $\pi_j \in \operatorname{Aut}(\mathcal C)$0 dB at $\pi_j \in \operatorname{Aut}(\mathcal C)$1 while lowering average decoding execution time in the low-BLER regime (Pillet et al., 2024).

4. Polar-code-specific design program

AED for polar codes has developed into a code-design program as much as a decoding program. A central negative result is that SC decoding is invariant with respect to the classical lower-triangular affine automorphisms, so conventional polar-code symmetries are often useless for SC-based AED. This motivated the design of polar codes with non-LTA automorphisms, especially upper-triangular linear automorphisms, and those constructions were reported to make SC-based AE effective, with examples outperforming conventional polar codes under SCL decoding while keeping latency comparable to SC decoding (Pillet et al., 2021).

This design pressure appears again in rate compatibility. For AED-compatible polar families with a prescribed useful automorphism group, a one-bit granular rate-compatible sequence cannot exist; the reason is that information-set enlargements must occur in symmetry-preserving packets rather than one arbitrary synthetic channel at a time. By allowing larger dimension steps, symmetric partial-order constructions and explicit $\pi_j \in \operatorname{Aut}(\mathcal C)$2-expansion rankings can produce rate-compatible sequences in which every code retains the prescribed automorphism structure (Geiselhart et al., 2023).

A more flexible data-driven version of this idea was later developed for nested symmetric polar codes. There, theoretic results on nested subcodes and supercodes show that symmetry can survive length nesting, and a shortest-path design procedure, augmented with zero padding, was used to build a fully nested, rate-compatible sequence for AE-SC-8. The reported construction outperforms existing code designs for AED and also the 5G polar code under CRC-aided SCL decoding in the average metric used in that study (Geiselhart et al., 2024).

At the decoder level, complexity reduction has become a parallel theme. For SC-based AED, sequential activation with early stopping was proposed under the name dynamic automorphism ensemble decoding. The stopping rule uses the best-so-far SC path metric

$\pi_j \in \operatorname{Aut}(\mathcal C)$3

and stage-dependent thresholds $\pi_j \in \operatorname{Aut}(\mathcal C)$4, stopping when $\pi_j \in \operatorname{Aut}(\mathcal C)$5. For various RM and polar codes of length $\pi_j \in \operatorname{Aut}(\mathcal C)$6, and for BLER below $\pi_j \in \operatorname{Aut}(\mathcal C)$7, the paper reports average-complexity reductions of at least $\pi_j \in \operatorname{Aut}(\mathcal C)$8 and up to $\pi_j \in \operatorname{Aut}(\mathcal C)$9 relative to the original AED complexity, with negligible degradation in BLER (Charles et al., 30 Apr 2026).

5. Complexity, latency, and hardware trade-offs

AED is structurally attractive because its constituent decoders are independent. Comparative work on short LDPC codes therefore emphasizes “low latency through parallelism,” since the branches can run concurrently with limited control overhead (Krieg et al., 2024). In QC-LDPC AED, the same property allows graph-symmetry breaking without increasing worst-case decoding latency, and the receiver-side changes can be as small as deleting one parity check from the decoding graph while retaining the original code for final validity filtering (Geiselhart et al., 2022).

For short polar codes, the strongest implementation case is the 12 nm FinFET Fast-SSC-based AED architecture developed for 6G URLLC. That work replaced conventional ML-in-the-list candidate selection by a branch-local path metric, so the final candidate is chosen by the smallest PM rather than by correlating each codeword with the stored channel LLR vector. This removes a large LLR buffer that the paper states can account for up to $$\hat{c}_j = \pi^{-1}_j\left(\operatorname{Dec}(\pi_j(\mathbf{L}_{\mathrm{ch}}))\right).$$0 of total decoder area. In the reported silicon comparison, the AED implementation outperforms state-of-the-art SCL decoders by up to $$\hat{c}_j = \pi^{-1}_j\left(\operatorname{Dec}(\pi_j(\mathbf{L}_{\mathrm{ch}}))\right).$$1 in latency, $$\hat{c}_j = \pi^{-1}_j\left(\operatorname{Dec}(\pi_j(\mathbf{L}_{\mathrm{ch}}))\right).$$2 in area efficiency, and $$\hat{c}_j = \pi^{-1}_j\left(\operatorname{Dec}(\pi_j(\mathbf{L}_{\mathrm{ch}}))\right).$$3 in energy efficiency, while providing the same or even better error-correction performance (Kestel et al., 2023).

A different hardware point on the design spectrum is serial AED for Polar-code-based physical unclonable functions. There, one SC decoder is reused across multiple automorphism attempts, and the number of AED candidates is scaled by cascaded and recursive interleavers rather than by instantiating many decoders. The same work combines this with aggressive 3-bit quantization and reports that, at BLER $$\hat{c}_j = \pi^{-1}_j\left(\operatorname{Dec}(\pi_j(\mathbf{L}_{\mathrm{ch}}))\right).$$4, the resulting coding scheme achieves the same failure rate as a BCH-based baseline while requiring $$\hat{c}_j = \pi^{-1}_j\left(\operatorname{Dec}(\pi_j(\mathbf{L}_{\mathrm{ch}}))\right).$$5 fewer codeword bits for the same $$\hat{c}_j = \pi^{-1}_j\left(\operatorname{Dec}(\pi_j(\mathbf{L}_{\mathrm{ch}}))\right).$$6 payload bits (Rübenacke et al., 10 Oct 2025).

These hardware results do not eliminate AED’s main trade-off. Parallel AED suppresses latency growth but scales area with the ensemble size, while serial AED suppresses area but increases execution time. The literature therefore treats AED not as a single complexity point but as an architecture family whose operating point depends on how automorphisms are generated, how candidates are ranked, and whether the branches are parallel, sequential, or early-stopped.

6. Generalizations, neighboring methods, and present limits

AED has already been generalized in several directions. Generalized AED (GAED) replaces permutation automorphisms by arbitrary linear bijective self-maps of the code space,

$$\hat{c}_j = \pi^{-1}_j\left(\operatorname{Dec}(\pi_j(\mathbf{L}_{\mathrm{ch}}))\right).$$7

and preprocesses the observation by box-plus combinations rather than simple permutations. In the binary case, if $$\hat{c}_j = \pi^{-1}_j\left(\operatorname{Dec}(\pi_j(\mathbf{L}_{\mathrm{ch}}))\right).$$8, the transformed LLRs are computed as

$$\hat{c}_j = \pi^{-1}_j\left(\operatorname{Dec}(\pi_j(\mathbf{L}_{\mathrm{ch}}))\right).$$9

This strictly enlarges the automorphism search space beyond classical scaled permutations, but the same work also stresses that sparse transforms are preferable because dense box-plus preprocessing reduces LLR magnitudes (Mandelbaum et al., 2023).

A second generalization is structural rather than algebraic. Constituent automorphism decoding for Reed–Muller codes applies automorphism-ensemble decoding not only at the root of the Plotkin tree but also to selected internal Plotkin constituents. The resulting decoder targets the constituents with the highest first-error probability and is reported to achieve state-of-the-art performance-versus-complexity trade-offs against GMC, conventional AE, and tested SCL decoders (Qu et al., 2024).

A third extension is quantum. AutDEC for quantum LDPC codes applies automorphisms of CSS codes or of the detector error model, using the relation

$s_j=\hat{c}_jH^\mathrm{T}=\mathbf{0}.$0

to transform syndromes before running parallel BP decoders. The reported performance is comparable to BP-OSD-0 on Quantum Reed–Muller and Bivariate Bicycle codes, with lower time overhead than OSD-style post-processing (Koutsioumpas et al., 3 Mar 2025).

The principal limitations of AED are consistent across these variants. First, it depends on a sufficiently rich automorphism group, which is why the comparative LDPC literature identifies AED as requiring “Code Automorphisms” and “Non-Equivariance” (Krieg et al., 2024). Second, the useful symmetry must exceed decoder symmetry; otherwise branch diversity collapses, as in unmodified QC-LDPC BP. Third, for polar codes the method is constrained by code design: AED can improve fixed-list decoding by running parallel SCL decodings on permuted received words, but it is limited to appropriately designed polar codes, which is precisely the gap that subcode ensemble decoding was proposed to relax (Lulei et al., 24 Apr 2025).

In that sense, AED now occupies a well-defined position in the decoding landscape. It is a symmetry-exploiting ensemble method whose effectiveness is governed by the interaction between code automorphisms, decoder equivariance, and branch-selection machinery. Where those three elements are aligned, AED approaches near-ML performance with architectures that are often more parallel, more local, and more implementation-efficient than list-based alternatives.