{"id":"072aebed-5457-40c8-98b2-843425ea3fb7","arxiv_id":"2504.17511","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new type of polar-code pre-transformation enables subcode ensemble decoding, which matches the error performance of larger list-size decoders at the same hardware list budget.","lead":"The paper brings a decoding method called subcode ensemble decoding, previously used for other code families, to polar codes. It reports frame error rate gains of 0.1 to 0.25 dB over standard list decoding for short block lengths.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Covering condition (1) is asserted but never checked; for M=2 and dp=2 it is impossible by cardinality, so reported FERs are unexplained unless the transmitted-codeword distribution is restricted.","rationale":"The reader's verdict correctly identifies the covering condition as the load-bearing assumption. My stress-test sharpens this concern with a concrete cardinality argument: for M=2 and dp=2, the union of the subcodes cannot equal the full code, so the reported ScED-2 results can only be explained if the transmitted codewords are drawn from a restricted subset or the depth is smaller than suggested. Since the paper does not state the depth used in the experiments, the transmitted-codeword distribution, or any coverage check, the central claim is not yet substantiated. The proposed PT classification and Theorem 1 are useful and plausible, and the reported gains are consistent with the method if covering holds, so a conditional verdict is appropriate. I would not change the reader's conditional assessment; the concern is addressable by the concrete membership/coverage test described above.","tokens_in":8376,"tokens_out":10369,"duration_ms":113603,"concrete_test":"Re-run the Fig. 6/7 simulations while logging each transmitted info word u. For each selected PT-C i, test whether the transmitted codeword x(u) lies in CT_i by evaluating the dp affine constraints (u_t = sum_{j<t} A_{t,j} u_j + b_t for t in T_C^i). Count violations over all Monte-Carlo trials. Additionally, sample 10^6 uniformly random info words u and compute the fraction outside all M subcodes; if any uncovered codeword is found, condition (1) fails and the reported FERs must be recomputed under uniform codeword transmission. The same membership test will reveal whether the original simulations used a restricted codeword distribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Every reported FER gain rests on the covering condition (1): each transmitted codeword must lie in at least one PT-C subcode. The selection algorithm in Sec. IV-B maximizes the number of decoded URPs and does not check or enforce (1). This is not a cosmetic gap. A PT-C of depth dp fixes dp information bits as affine functions of earlier bits, so each subcode has size 2^{kappa-dp}. For M=2, if dp=2 (the depth suggested by Sec. IV-A), the union of the two subcodes has size at most 2^{kappa-1} and cannot cover C. Even for dp=1, two random affine hyperplanes cover only about 3/4 of C, so (1) fails. For M=8, dp=2 is cardinality-feasible, but random selection is not shown to cover. The simulation section does not state dp, the transmitted-codeword distribution, or any coverage verification. If codewords were drawn uniformly, uncovered codewords would produce a FER floor far above 10^{-3}; if only a fixed codeword (e.g., the all-zero codeword) was transmitted, the results do not support the general claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper extends subcode ensemble decoding (ScED) to polar codes by introducing a new class of pre-transformations, called PT-C, which are used only at the decoder side. The authors propose a unified framework for pre-transformations (PT-A, PT-B, PT-C), prove a correctness theorem for SC decoding on a single subcode, describe a data-driven selection procedure for choosing an ensemble of PT-C subcodes, and report Monte-Carlo FER simulations showing that ScED with M SCL-L decoders outperforms stand-alone SCL-L decoding by 0.1 to 0.25 dB at FER 1e-3, with two SCL-L paths matching one SCL-2L decoder for a short 5G polar code.","tokens_in":8680,"tokens_out":5100,"duration_ms":52338,"significance":"If the reported gains are fully supported, the paper offers a practically relevant alternative to increasing the list size in SCL decoding, with potential hardware re-utilization or latency reduction. The paper has clear strengths: the classification of pre-transformations is useful, the proof of Theorem 1 is clean for the SC setting, the simulation setup includes an ML benchmark for the short code, and the empirical comparison is concrete. However, the central covering condition is neither enforced nor verified, the theoretical result does not cover the SCL decoders used in all reported simulations, and several simulation parameters (PT-C depth, transmitted-codeword distribution, URP sampling SNR) are unspecified. These gaps currently prevent the reader from assessing whether the reported FER gains hold for the general polar code or only for a restricted set of transmitted codewords.","major_comments":[{"comment":"The covering condition in Eq. (1) is load-bearing but is never verified. The design algorithm in §IV-B maximizes the number of decoded URPs and does not check that the union of the selected PT-C subcodes equals the original code C. For M=2 and depth dp=2 this is impossible by cardinality: each subcode has size 2^{κ-2}, so the union has size at most 2^{κ-1} < |C|; even for dp=1, two random affine hyperplanes do not in general cover C. The manuscript must state the PT-C depth dp used in the final ensembles of §V, verify condition (1) for the selected ensembles (e.g., an exact coverage check for the short code or a constructive covering argument), and report which codewords are transmitted in the simulations. If codewords were drawn uniformly from C, uncovered codewords would produce a FER floor far above 1e-3; if only codewords lying in the union were transmitted, the reported results do not support the general claim.","section":"§III-B, Eq. (1); §IV-B; §V"},{"comment":"Theorem 1 concerns only SC decoding, and its proof relies on the two decoders making the same decisions up to the target bit; this argument does not directly extend to SCL decoding, where list branching, path-metric comparisons, and the final CRC-based selection change the decision process. The paper explicitly states that SCL analysis is future work, yet all reported gains in §V are for ScED-M-SCL-L. The theorem therefore cannot be cited as theoretical support for the headline SCL-based claim, which currently rests solely on simulations. Please either provide an SCL analysis or clearly label the theoretical contribution as SC-only and present the SCL results as purely empirical.","section":"§III-B, Theorem 1"},{"comment":"The ensemble selection procedure is based on decoding a set of 1000 URPs-SCL-L, but the operating SNR for sampling these URPs is not reported and no independent validation set is used to check for overfitting of the union-maximization heuristic. Since the selected ensemble is then used directly in the reported FER curves, please specify the URP sampling SNR, the number of training URPs, and provide a validation check (e.g., performance on a fresh URP set or a coverage check on the codeword space) to show that the observed gains are not an artifact of the particular training set.","section":"§IV-B"}],"minor_comments":[{"comment":"The captions are inconsistent with the text: §IV-A says 192 random PT-Cs are selected per 100 URPs, but the Fig. 5 caption says 190 new decoders; please make the numbers consistent.","section":"Fig. 4 and Fig. 5 captions"},{"comment":"The text states that URPs-SC are sampled at Eb/N0 = 2.5 dB, but the Fig. 4 caption omits this SNR information; please add it to the caption.","section":"§IV-A"},{"comment":"The statement \"If x̂_i ∈ C_i, the ith subcode decoding converged successfully\" is most natural for iterative decoders such as BP; for SCL decoding the analog of successful convergence should be defined more precisely, since SCL always outputs a codeword in the subcode by construction.","section":"§II-C"},{"comment":"The path-metric condition is only referenced via [18, Eq. (10)]; please reproduce the condition or give a self-contained definition so that the selection criterion is understandable without consulting the reference.","section":"§IV-B"},{"comment":"The FER plots do not show confidence intervals or error bars; given the statement that at least 1000 frame errors are simulated per point, please add standard errors or state the uncertainty explicitly.","section":"§V"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a coding theory journal and the ScED idea for polar codes is interesting. The main concern is that the reported FER gains are not reproducible or interpretable without specifying the PT-C depth, the transmitted-codeword distribution, and, most importantly, a verification of the covering condition in Eq. (1). The M=2/dp=2 case seems cardinality-impossible, so the authors need to either change the ensemble parameters, verify coverage exactly, or restrict the claims accordingly. I would encourage a revision that addresses these issues; the underlying approach is not unreasonable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper extends subcode ensemble decoding (ScED) to polar codes by introducing PT-C, a pre-transformation used only in decoding. That's genuinely new, and the classification of PTs (A/B/C) is a useful organizing frame. The simulation work is substantial: tests across rates and lengths, Monte Carlo with at least 1000 frame errors per point, and comparison against SCL with multiple list sizes and ML where possible. If the results held as stated, ScED would be a credible, hardware-friendly way to get better FER at fixed list size.\n\nThe soft spots are proportional to how much they matter, and the biggest one matters a lot. Condition (1) -- the union of the selected subcodes must equal the transmitted code -- is the entire basis of ScED. The design algorithm in Sec. IV-B picks subcodes to maximize decoded URPs, but never checks coverage. For the headline ScED-2 results, it cannot be satisfied. Each PT-C of depth dp removes dp degrees of freedom, so two subcodes of depth 2 contain at most 2/4 of the codewords, and even depth 1 with two subcodes covers at most 3/4 of the space. The simulations therefore only make sense if a single fixed codeword (almost certainly the all-zero word) is transmitted. If that is what was done, the FER curves reflect the decoder's behavior for that one codeword and do not support general claims. The paper never states which codewords were transmitted, so this is not a cosmetic omission; it is load-bearing.\n\nSmaller issues: Theorem 1 is only for SC, while all results are SCL; that is fine as motivation but should be labeled. The depth used in the final ensembles is never stated, which hurts reproducibility. The selection heuristic is tuned on a finite set of URPs and could overfit, but that is a tuning procedure, not circularity.\n\nThe PT-C framework and the analysis are genuinely useful, and the authors are clearly knowledgeable. The missing coverage verification and unspecified simulation setup are fixable: they could either prove that the ensemble covers the code, or restrict the claims to a known transmitted word and say so. As submitted, the evidence does not support the general performance claim.\n\nI would send this to peer review: the core idea has enough value, and the gaps are addressable. A careful referee should push for coverage verification and a precise statement of the transmitted-codeword distribution. I would also bring it to a reading group, because the covering issue is an instructive example of a load-bearing assumption slipping through.","headline":"A useful new framework for ScED in polar codes, but the reported gains rest on an unverified -- and for M=2 impossible -- covering condition, so the performance claims are not yet supported.","tokens_in":9168,"tokens_out":11095,"would_cite":true,"duration_ms":101335,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Subcode ensemble decoding lifts polar codes 0.1–0.25 dB at fixed list size, the paper reports.","keywords":["polar codes","subcode ensemble decoding","pre-transformed polar codes","successive cancellation list decoding","dynamic frozen bits","short block length regime","frame error rate","ML-in-the-list decision"],"falsifier":"Take one optimized ScED ensemble (e.g., ScED-8-SCL-8 for $C_T(256,128)$) and test whether every codeword of $C$ lies in at least one subcode $C_i$; if any codeword is missing, transmit it and observe that no path can return it, giving a lower FER bound that would show up as a floor in longer simulations.","tokens_in":8201,"feed_emoji":"📡","tokens_out":6365,"duration_ms":56733,"temperature":0.7,"pith_summary":"Subcode ensemble decoding (ScED) was originally a way to run several belief-propagation decoders on different subcodes of a linear code; this paper brings the idea to polar codes. The key move is a new class of pre-transformations, called PT-C, that are used only at decoding: they turn selected information bits of the polar graph into dynamic frozen bits, so each PT-C defines a subcode the decoder can decode on, while encoding stays untouched. The paper claims that with a small optimized set of such subcodes, SCL-based ScED outperforms stand-alone SCL decoding at the same list size, by 0.1 dB to 0.25 dB at a frame error rate of $10^{-3}$, and that for the short 5G code $C_T(64,32)$, two list-$L$ subcode decoders match one list-$2L$ decoder. The reason to care is hardware: doubling a list size more than quintuples decoder area, so spreading the same total list budget over parallel paths is a practical way to buy performance.","feed_headline":"Subcode ensembles lift polar codes 0.1–0.25 dB at fixed list size","feed_subtitle":"Two list-L subcode decoders match one list-2L decoder for the short 5G code, without the area cost.","key_machinery":"The load-bearing object is the PT-C pre-transformation: an affine map $T(u_p) = u_p A + b$, with upper triangular $A$ and target bits chosen inside the information set, applied only on the decoding side. Applied to the polar graph, it turns target information bits into dynamic frozen bits, defining a subcode $C_T$ of the transmitted polar code; because $A$ is upper triangular, the decoder can evaluate these constraints sequentially. The ensemble design then relies on two conditions: the subcodes must jointly cover the original code (union over $i$ of $C_i = C$), and the $M$ subcodes should be chosen to maximize how many recorded undecodable received patterns they jointly correct, subject to the subcode estimate having a lower path metric so it survives the ML-in-the-list decision.","core_discovery":"The paper's central claim is that polar subcodes created by PT-C pre-transformations form an effective ensemble decoder: run $M$ parallel SCL decoders, each on a different joint graph, collect the candidate codewords, and pick the one maximizing the channel log-likelihood among valid codewords of the original polar code. A theorem shows that in the SC regime, whenever stand-alone SC decoding succeeds on a transmitted codeword that lies in the subcode, the subcode decoder also succeeds; hence an ensemble whose subcodes cover the whole code is never worse than SC decoding, and the empirical results extend this advantage to SCL decoding. Concretely, $\\mathrm{ScED}\\text{-}M\\text{-}\\mathrm{SCL}\\text{-}L$ beats $\\mathrm{SCL}\\text{-}L$ by 0.1–0.25 dB at FER $10^{-3}$ across the tested 5G polar codes, and two SCL-$L$ subcode decoders match one SCL-$2L$ decoder for $C_T(64,32)$, while eight are needed for the longer rate-$1/4$ and rate-$1/2$ codes.","pith_inferences":["A direct verification of the covering condition (1) for the optimized ensembles would determine whether the reported FER is upper-bounded by uncorrectable codewords; the selection algorithm in Sec. IV-B maximizes URP coverage but does not prove union equality.","Because PT-C changes only the decoding graph, the same ensemble construction should transfer to other polar-style decoders (BP, soft cancellation) and to PAC or row-merged variants, where the dynamic-frozen-bit structure already exists.","The SC theorem suggests the gain mechanism is that subcode decoders can succeed on error patterns that confuse the full-code decoder; combining PT-C ensembles with automorphism ensembles would mix two orthogonal sources of diversity, although the paper does not test this."],"forward_implications":["For every tested 5G polar code, $\\mathrm{ScED}\\text{-}M\\text{-}\\mathrm{SCL}\\text{-}L$ outperforms stand-alone $\\mathrm{SCL}\\text{-}L$ by 0.1 to 0.25 dB at FER $10^{-3}$ with the same list size per path.","For the short code $C_T(64,32)$, two SCL-$L$ subcode decoders match the FER of one SCL-$2L$ decoder, for $L \\in \\{8,16\\}$; longer rate-$1/4$ and rate-$1/2$ codes need eight paths to match SCL-$2L$.","Because SCL-$2L$ costs more than five times the chip area of SCL-$L$, the matching performance can be obtained by reusing $L$-sized decoder hardware across paths, or by computing paths in parallel to cut latency.","PT-C depth trades subcode size against error correction: depth $d_p = 2$ appears the sweet spot in the paper's analysis, since deeper subcodes cover only $2^{k-d_p}$ codewords and require more paths to maintain covering."],"supporting_citations":[{"why":"defines subcode ensemble decoding and the covering condition (1) that the ensemble must satisfy.","marker":"[7]"},{"why":"introduces SCL list decoding of polar codes, the baseline decoder the ensemble is compared against.","marker":"[2]"},{"why":"shows how polar subcodes can be represented by pre-transformations, the basis for the PT-C construction.","marker":"[17]"},{"why":"provides row-merged polar codes and dynamic frozen bits, the mechanism PT-C repurposes for decoding.","marker":"[16]"},{"why":"supplies the heuristic for choosing a small subset of decodings that jointly correct many recorded error patterns.","marker":"[8]"},{"why":"gives the path-metric formula used to require that a subcode estimate can win the ML-in-the-list decision.","marker":"[18]"},{"why":"documents the hardware area cost of doubling the SCL list size, motivating fixed-list-size ensemble decoding.","marker":"[5]"}],"fun_headline_variants":["Subcode ensembles lift polar codes 0.1–0.25 dB","Polar codes gain 0.1–0.25 dB with subcode ensembles","Match list-2L polar decoding with two list-L passes","Subcode ensemble decoding beats list doubling for polar codes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the selected PT-C subcodes jointly cover the whole polar code (condition (1)); the paper asserts this following [7], but the design algorithm only maximizes how many recorded error patterns are decoded and never verifies that every codeword belongs to at least one subcode, so a transmitted codeword outside all subcodes can never be decoded correctly.","fun_headline_variants_meta":{"raw":{"variants":["Subcode ensembles lift polar codes 0.1–0.25 dB","Polar codes gain 0.1–0.25 dB with subcode ensembles","Match list-2L polar decoding with two list-L passes","Subcode ensemble decoding beats list doubling for polar codes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000245,"raw_usage":{"total_tokens":1565,"prompt_tokens":1001,"completion_tokens":564,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":487}},"tokens_in":617,"tokens_out":564,"duration_ms":5775,"temperature":1.0,"reasoning_tokens":487,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:39:21.756959+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take one optimized ScED ensemble (e.g., ScED-8-SCL-8 for $C_T(256,128)$) and test whether every codeword of $C$ lies in at least one subcode $C_i$; if any codeword is missing, transmit it and observe that no path can return it, giving a lower FER bound that would show up as a floor in longer simulations.","supporting_citations":[{"cited_title":"List decoding of polar codes,","cited_arxiv_id":null,"evidence_quote":"introduces SCL list decoding of polar codes, the baseline decoder the ensemble is compared against."},{"cited_title":"Polar subcodes,","cited_arxiv_id":null,"evidence_quote":"shows how polar subcodes can be represented by pre-transformations, the basis for the PT-C construction."},{"cited_title":"Row-merged polar codes: Analysis, design, and decoder implementation,","cited_arxiv_id":null,"evidence_quote":"provides row-merged polar codes and dynamic frozen bits, the mechanism PT-C repurposes for decoding."},{"cited_title":"Ensemble belief propagation decoding for short linear block codes,","cited_arxiv_id":null,"evidence_quote":"supplies the heuristic for choosing a small subset of decodings that jointly correct many recorded error patterns."},{"cited_title":"LLR-based Successive Cancellation List Decoding of Polar Codes,","cited_arxiv_id":null,"evidence_quote":"gives the path-metric formula used to require that a subcode estimate can win the ML-in-the-list decision."},{"cited_title":"Automorphism ensemble polar code decoders for 6G URLLC,","cited_arxiv_id":null,"evidence_quote":"documents the hardware area cost of doubling the SCL list size, motivating fixed-list-size ensemble decoding."}],"review_version":1}