{"id":"c29f18fa-0ebc-4484-bb65-d485413d401e","arxiv_id":"2411.13876","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A hybrid of optimized min-sum iteration, neural-network reliability boosting, and ordered-statistics decoding approaches the maximum-likelihood bound for short BCH codes.","lead":"This paper proposes a hybrid decoder for short BCH error-correcting codes that combines a fast iterative decoder with a slower, more accurate post-processor. It reports near-optimal error rates at lower complexity than existing schemes for three length-63 BCH codes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'interleaving' permutation used in the revised NMS is not an automorphism of the BCH codes, so Algorithm 2 as described may be decoding under a different code than intended.","rationale":"The reader's weakest assumption focuses on the DIA model's missing training procedure, which is a reproducibility concern but does not invalidate the algorithm if the model is properly trained. The load-bearing issue I identify is more severe: the paper explicitly states that the three permutation types belong to the automorphism group and uses them to define the revised NMS. The even/odd interleaving permutation is not in the automorphism group of primitive BCH codes of length 63, so Algorithm 2 as written applies a non-automorphism permutation while keeping H_s fixed. That means each permuted branch of the decoder is seeking codewords of a different code, and the inverse permutation of the resulting LLRs has no valid decoding interpretation. If this concern lands, the central performance claim of N-D-O is not trustworthy, regardless of DIA training. The concrete test I propose is simple and decisive: check whether π(c) satisfies H_s for arbitrary codewords. If it fails, the authors must either correct the permutation set or explain why the algorithm still works; without such a correction, the paper cannot be accepted as a valid decoding scheme. Thus I recommend keeping the verdict CONDITIONAL but with a more fundamental condition than the reader's, and I record a disagreement on the weakest assumption.","tokens_in":10508,"tokens_out":17815,"duration_ms":172622,"concrete_test":"Generate a random codeword c of the BCH(63,45) code using its generator matrix, apply the even/odd interleaving permutation π defined in Section III-B (new sequence = even-indexed bits 0,2,...,62 followed by odd-indexed bits 1,3,...,61), and compute H_s π(c)^T. If this product is nonzero for any codeword, π is not an automorphism. Run this for all three codes and, if it fails, inspect the source code linked in Section IV to determine whether H_s is also permuted during the dilated-input construction. If H_s is not permuted, the algorithm as described cannot decode correctly and the simulation curves need to be re-examined.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Section III-B claims that the three permutation types are drawn from the automorphism group of cyclic codes, but the interleaving permutation—concatenating even-indexed bits with odd-indexed bits—is not an automorphism for the BCH(63,k) codes used here. For a permutation to be applicable with an unpermuted parity-check matrix, it must map every codeword to another codeword; the even/odd interleaving does not satisfy this property for primitive BCH codes, whose automorphism group is the affine group (multiplications and translations), not arbitrary even/odd rearrangements. In Algorithm 2, each permuted input is decoded using the same H_s, which is only correct if the applied permutation is an automorphism. If it is not, the per-permutation NMS updates are solving a different decoding problem, and the later averaging and parity check (5) have no theoretical guarantee. The paper's reported gains could stem from an implementation that also permutes H_s (which is not described) or from a different interleaving definition, making the simulation results unverifiable from the text. This is more fundamental than the missing DIA training details because it challenges the validity of the core decoder itself.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an iterative decoding framework for short BCH codes based on three components: a heuristic optimization of the parity-check matrix (H_s) with reduced density and more regular column weights, a revised normalized min-sum (NMS) decoder that concurrently applies three types of permutations per iteration, and a neural-network-based reliability enhancement (DIA) followed by order-1 ordered statistics decoding (OSD). Simulation results for the BCH (63,36), (63,39), and (63,45) codes are presented, claiming that the hybrid N-D-O(4,1) decoder approaches the maximum-likelihood lower bound with lower complexity than competing iterative decoders. The paper also reports the undetected error rate of NMS for these codes and argues that it forms a performance floor for any decoder that uses parity-check-based stopping.","tokens_in":10791,"tokens_out":7095,"duration_ms":68639,"significance":"If the claims hold, the paper would make a useful contribution to iterative decoding of short BCH codes: the optimized parity-check matrix construction, the convergence acceleration by concurrent permutations, and the DIA-based reliability boosting for OSD are all potentially valuable. The explicit reporting of undetected FER as a floor for iterative decoders is a useful observation. Strengths include the use of external ML benchmarks from [19], comparison with several published decoders, an explicit complexity table, and a promise of source code on GitHub. However, the central decoder as described has a theoretical gap (the interleaving permutation is not an automorphism of the codes used), and the DIA model is not reproducible from the manuscript because no training procedure is given. These issues affect the validity of the main performance and complexity claims and need to be addressed before the paper can be accepted.","major_comments":[{"comment":"The paper states that the three types of permutations are drawn from the automorphism group of cyclic codes, but the interleaving permutation (concatenating bits at even indices with those at odd indices) is not an automorphism of the primitive BCH(63,k) codes used in the simulations. For these codes, the permutation automorphism group is the affine group on GF(64), consisting of permutations of the form x -> a x + b, while the even/odd interleaving does not have this form and is not shown to preserve the code. In Algorithm 2, each permuted input is decoded using the same unpermuted H_s, and the parity check in (5) is applied after inverse permutation and averaging; this is only justified if every applied permutation maps codewords to codewords. As written, the per-permutation NMS updates are decoding under a different code, so the theoretical basis for the convergence and parity-check termination of Algorithm 2 is invalid. Please either prove that the interleaving permutation preserves each of the BCH codes considered, replace it with a third automorphism, or explicitly permute H_s accordingly and rerun the simulations. This is load-bearing for the reported performance of the revised NMS.","section":"Section III-B, Algorithm 2"},{"comment":"The DIA model is a central component of the hybrid decoder, yet the manuscript only specifies its input and output shapes (6300×5×1 to 100×63) and states 'Assuming a well-trained DIA model'. No training procedure is given: the training data set (which SNR values, which code, how failures are collected), the labels, the loss function, the optimizer, the hyperparameters, and the validation strategy are all absent. Since [17] is cited for the rationale, the reader cannot determine whether the DIA model used here is trained on BCH failures, on LDPC failures, or on a mixture, and whether it generalizes over the entire SNR range. The performance gain attributed to DIA in Figs. 1-3 and the ablation claim in Section IV-A therefore are not reproducible from the manuscript. Please provide complete training details and, ideally, release the trained model weights.","section":"Section IV-A and Appendix A"},{"comment":"The central claim that N-D-O(4,1) 'approaches the lower bound determined by the undetected FER and ML curves' rests entirely on simulation curves, but the paper reports no number of simulated frames, no seeds, and no confidence intervals. For FER values around 1e-5 and below, which are shown in several figures, the absence of statistical uncertainty makes it impossible to judge whether the curves truly approach the claimed floor. Please report the number of trials, the stopping rule for the simulations, and error bars or a bootstrap confidence band, at least for the proposed decoder and the ML benchmark.","section":"Section IV-A, Eq. (6) and Figs. 1-3"},{"comment":"For the BCH (63,36) code, the optimized H_s increases the number of length-4 cycles from 5,909 to 114,196, and the paper asserts that 'the adverse effects of the increased number of length-4 cycles are mitigated by the row redundancy'. No evidence is provided for this assertion; a direct comparison with a redundant parity-check matrix that has a comparable number of rows but fewer 4-cycles would be needed. Without such a comparison, the advantage of the proposed optimization for this code is not established.","section":"Section III-A, Table I"},{"comment":"The complexity comparison ignores the DIA model and the OSD component of the hybrid, stating only that they are 'safely ignored due to their relatively minor workload'. This is not justified quantitatively: order-1 OSD requires sorting the LLRs and performing Gaussian elimination, and the DIA neural network adds inference cost per failed NMS frame. Since the paper's headline claim is the best trade-off between performance and complexity, the complexity ratio for N-D-O(4,1) should include these contributions or be rephrased as the complexity of the NMS component only.","section":"Section IV-B, Table II"}],"minor_comments":[{"comment":"In the discussion of the BCH (63,45) code, the text refers to 'Fig. 6' for the FER and BER curves, but the corresponding figures are Fig. 3(a) and Fig. 3(b).","section":"Section IV-A"},{"comment":"The decoder name is written inconsistently as 'MBBD' in Fig. 3 and 'MBBP' in Table II and the text; please unify the notation.","section":"Section IV-A and Fig. 3"},{"comment":"The text says 'except for BP-RDD' when comparing complexity ratios; this appears to be a typo for 'BP-RNN'.","section":"Section IV-B"},{"comment":"The phrase 'the maxima of the ML and undetected FER curves serve as a lower bound' is imprecise; a lower bound is the pointwise maximum (upper envelope) of the two curves, not 'maxima'.","section":"Section IV-A"},{"comment":"The statement that alpha = 0.78 was 'optimized based on the shape of the dilated input and I = 4' does not describe how the optimization was performed; please state the search procedure and the criterion used.","section":"Section IV"},{"comment":"The URL 'https://github.com/lgw-frank/Short BCH Decoding OSD' contains spaces; use a URL-encoded or clickable link so that the promised source code is accessible.","section":"GitHub link in Section IV"},{"comment":"The concluding sentence of the abstract and Section IV-B claim the 'best trade-off among existing decoders for short BCH codes'; given that only three BCH codes and specific parameter settings are tested, this statement should be qualified to avoid overgeneralization.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"The paper extends the authors' prior work [17], and the DIA component is described only by reference to that paper. If this manuscript is intended as a self-contained journal submission, the lack of training details and the possible dependence on an unpublished or not-yet-available model are a risk. The GitHub link appears malformed, which may also be an issue for reproducibility. Please verify that the source code and trained model are actually available to reviewers."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper is a genuine incremental advance, not a breakthrough. The one clearly new result is the undetected-error characterization for short BCH codes under iterative decoding. The hybrid decoder itself is a sensible assembly of known parts (redundant parity-check matrix, automorphism permutations, neural reliability boosting, OSD post-processing), and the simulations suggest it works. But the reproducibility gaps — especially the DIA model — are serious enough that I wouldn't trust the quantitative claims yet.\n\nWhat's good: The undetected FER/BER curves are useful and, as far as I know, a first for these codes. They show that the undetected errors set a floor that sits above the ML bound at high SNR, which matters for anyone trying to push short BCH decoders closer to ML. The revised NMS with three permutations per iteration is a reasonable twist on mRRD/PBP, and the reported 4-iteration convergence is impressive if it holds. The H_s construction for (63,39) and (63,45) genuinely reduces 4-cycles and regularizes column weights; the (63,36) case increases 4-cycles but the authors argue redundancy compensates — plausible but not proven.\n\nSoft spots, in rough order of severity. First, the DIA model is a black box. 'Assuming a well-trained DIA model' is not acceptable in a paper where the headline hybrid depends on it. No training data, loss function, hyperparameters, or SNR range; the appendix gives only the conv/dense shape and some validation plots for (63,45). Without that, no one can reproduce the N-D-O curves or know whether the gain over OSD(1) is real. Second, no error bars or seed information anywhere; many comparisons are within 0.2–0.4 dB, which can be noise. Third, the complexity ratio in Table II conveniently ignores the DIA and OSD costs for N-D-O while counting full iteration loads for competitors; the 'best trade-off' claim overreaches. Fourth, the GitHub link contains spaces and is broken. Minor: the Frobenius permutation definition looks garbled ('2^i mod N' should be '2i mod N').\n\nOn the stress-test note: I checked the interleaving permutation. It's not an arbitrary even/odd split; it's multiplication by 32 modulo 63, which is a valid multiplier automorphism for these BCH codes. So that concern doesn't hold.\n\nBottom line: the paper deserves a serious referee. The undetected-error result alone is worth publishing, and the decoder is plausible enough to warrant effort. The revision needs DIA training details, a working code link, error bars, and a less inflated complexity claim. I'd send it to review and ask for those fixes.","headline":"A plausible hybrid decoder for short BCH codes with a genuinely new undetected-error characterization, undone by missing DIA training details and a broken code link.","tokens_in":11309,"tokens_out":6009,"would_cite":true,"duration_ms":52250,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94B35","94B15"],"pacs":[],"model":"deepseek-v4-flash","headline":"A four-iteration hybrid decoder approaches the maximum-likelihood lower bound for short BCH codes.","keywords":["BCH codes","iterative decoding","normalized min-sum","parity-check matrix optimization","random permutations","ordered statistics decoding","neural network decoding","undetected errors"],"falsifier":"Reproduce N-D-O(4,1) on the BCH(63,36) code by training the DIA model from scratch with a disclosed procedure and compare the FER at Eb/N0 = 3 dB to the published curve; if the curve cannot be matched, or if the undetected-FER lower bound is crossed, the central claim is false.","tokens_in":10292,"feed_emoji":"📡","tokens_out":7213,"duration_ms":49546,"temperature":0.7,"pith_summary":"This paper proposes a decoding chain for short BCH codes that combines a revised normalized min-sum (NMS) decoder with a neural-network reliability booster and a low-order ordered-statistics decoder (OSD). The authors claim that this hybrid, called N-D-O(4,1), reaches the lower bound set by the maximum-likelihood (ML) curve and the undetected-error curve across the SNR range of interest, while using only four NMS iterations and an order-1 OSD post-processor. They further claim that this gives the best performance-versus-complexity trade-off among existing iterative decoders for these codes. The practical stake is that short BCH codes could be decoded near optimally without the latency of full OSD or the many iterations of earlier permuted belief-propagation schemes.","feed_headline":"Hybrid decoder nears maximum-likelihood limit for short BCH codes","feed_subtitle":"Four NMS iterations, a neural reliability boost, and order-1 OSD give near-ML decoding at lower cost.","key_machinery":"The central mechanism is the N-D-O(4,1) chain: an optimized parity-check matrix H_s, a revised NMS that applies three types of permutations (even-odd interleaving, Frobenius mapping 2i mod N, and cyclic shifts with step 21) to each received block per iteration, a DIA neural network that turns failed NMS trajectories into improved bit reliabilities, and an order-1 OSD that performs the final correction. The optimized H_s reduces density and column-weight spread while raising redundancy; the concurrent permutations let four NMS iterations mimic the diversity that earlier schemes obtained with hundreds of iterations; DIA supplies OSD with better soft inputs than the raw LLRs.","core_discovery":"The central claim is that iterative decoding of short BCH codes does not need many iterations or high-order OSD if the parity-check matrix is optimized, the message-passing is diversified with automorphism permutations each iteration, and the failed-decoding trajectories are re-weighted by a trained neural network. Concretely, the authors construct a redundant, low-density parity-check matrix H_s with balanced column weights by binary-summing and cyclically shifting rows of the standard H, then run four iterations of NMS in which each iteration processes nine permuted versions of the input (interleaved, Frobenius, and cyclic-shifted). On NMS failures, a decoding-information-aggregation (DIA) neural network converts the per-iteration LLR trajectories into enhanced reliability measurements, and an order-1 OSD then re-decodes using those measurements. The resulting N-D-O(4,1) FER curves track the lower bound formed by the undetected FER and ML curves for BCH(63,36), BCH(63,39), and BCH(63,45), and the paper reports that it outperforms existing permuted BP decoders at a fraction of their complexity.","pith_inferences":["Beyond the paper's claims: if the undetected-error floor is fundamental rather than an artifact of the specific NMS decoder, then any future iterative decoder for short BCH codes must either supply soft information that OSD can refine or accept a hard floor above ML; the paper's own complexity comparison suggests the floor is worth measuring as a standard benchmark.","Beyond the paper's claims: the same H_s construction could be tested on q-ary alphabets or other cyclic structures, since the row-shift and binary-sum operations only use the cyclic automorphism, and a natural experiment is to check whether the complexity gain persists for Reed-Solomon codes.","Beyond the paper's claims: the DIA model's lack of training details means a direct reproducibility test would be to train it on the same trajectory format with different random seeds and see whether the pre-CDF advantage over the I-th iteration remains, separating the method from the particular trained instance.","Beyond the paper's claims: because the paper assumes a well-trained DIA model, a reader should treat the near-ML FER curves as conditional on that assumption; reporting the model's training data and loss would let others verify whether the curves are reproducible."],"forward_implications":["For the three BCH codes tested, N-D-O(4,1) approaches the lower bound set by undetected FER and ML curves over the whole SNR range; for BCH(63,36) the hybrid adds about 0.8 dB over NMS alone.","The hybrid matches or beats the BER/FER of mRRD with multiple subdecorators and of RRD, while running at 0.088 of the mRRD(1) complexity ratio shown in Table II.","Undetected errors of the NMS decoder set a real floor: for these short codes the undetected FER intersects the ML curve, so any decoder that relies on parity checks alone cannot go below that floor.","The DIA reliability boosting, originally designed for LDPC codes, carries over to HDPC codes such as BCH, improving the number of correct bits in OSD's most reliable basis.","The paper anticipates that longer BCH codes and binary images of Reed-Solomon codes are natural extensions, with more NMS iterations and higher OSD order."],"supporting_citations":[{"why":"Supplies the BCH code parameters, hard-decision decoding baselines, and maximum-likelihood simulation curves used as the lower-bound reference.","marker":"[19]"},{"why":"Provides the decoding-information-aggregation neural-network reliability-boosting method that the hybrid applies to NMS failures.","marker":"[17]"},{"why":"Defines ordered statistics decoding, the post-processor whose order-1 version completes the hybrid.","marker":"[15]"},{"why":"Introduces the modified random redundant decoding (mRRD) decoder and its parallel-subdecoder complexity, the main performance and complexity baseline.","marker":"[6]"},{"why":"Introduces random redundant decoding (RRD), whose permutation approach and BER curves are compared against NMS and N-D-O.","marker":"[5]"},{"why":"Proposes the permuted belief-propagation (PBP) decoder used as a serial-mode baseline in the BER comparisons.","marker":"[8]"},{"why":"Provides the neural belief-propagation decoder baselines (BP-RNN, mRRD-RNN) that the hybrid outperforms in FER and BER.","marker":"[13]"},{"why":"Supplies the multiple-bases belief-propagation (MBBP) decoder, a complexity benchmark with square parity-check matrices.","marker":"[7]"},{"why":"Gives the EPCM binary-circulant-matrix baseline and the cyclic-permutation viewpoint used to justify the permutation set.","marker":"[20]"}],"fun_headline_variants":["Four-iteration BCH decoder hits near-ML with neural boost","Permuted NMS plus neural reweighting nears ML for short BCH","Short BCH codes: four iterations to near-ML decoding","Hybrid BCH decoder: permutations, neural net, order-1 OSD","BCH decoding near the ML limit at a fraction of the cost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the DIA neural network is well-trained and generalizes across SNRs, since the paper only specifies its input and output shapes and gives no training procedure; if the trained model cannot be reproduced or does not transfer, the hybrid's gain over plain order-1 OSD collapses.","fun_headline_variants_meta":{"raw":{"variants":["Four-iteration BCH decoder hits near-ML with neural boost","Permuted NMS plus neural reweighting nears ML for short BCH","Short BCH codes: four iterations to near-ML decoding","Hybrid BCH decoder: permutations, neural net, order-1 OSD","BCH decoding near the ML limit at a fraction of the cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000652,"raw_usage":{"total_tokens":3010,"prompt_tokens":983,"completion_tokens":2027,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":1931}},"tokens_in":599,"tokens_out":2027,"duration_ms":40569,"temperature":1.0,"reasoning_tokens":1931,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:47:23.056742+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce N-D-O(4,1) on the BCH(63,36) code by training the DIA model from scratch with a disclosed procedure and compare the FER at Eb/N0 = 3 dB to the published curve; if the curve cannot be matched, or if the undetected-FER lower bound is crossed, the central claim is false.","supporting_citations":[{"cited_title":"Boosting ordered statistics decoding of short LDPC codes with simple neural network models,","cited_arxiv_id":null,"evidence_quote":"Provides the decoding-information-aggregation neural-network reliability-boosting method that the hybrid applies to NMS failures."}],"review_version":1}