{"id":"e8fa6784-8efd-4e8f-9b27-38c3b88fe35a","arxiv_id":"2505.09978","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A concatenated-coding construction with an improved MRIP ordering from inner-code soft outputs reaches near-ML block error rates for short low-rate codes at lower A* decoding complexity.","lead":"By using a short inner code's soft outputs to build a better reliability ordering, this paper decodes short low-rate error-correcting codes near the optimal limit with less search. Practitioners building ultra-reliable low-latency links may get faster, lower-power decoders for 128-bit codes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central efficiency claim rests on an unverified minimum distance for the convolutional-inner concatenated codes; d_min is only inferred from 0.5 dB slopes in Sec. V.D.","rationale":"The paper's own limitation statement in Section V.D ('The minimum distances of other concatenated codes are unknown') marks exactly where the central claim is least secure. The improved-MRIP decoding can only be called 'much more efficient' than eBCH if the concatenated code is at least as strong as the eBCH code; otherwise the comparison is between a weaker code decoded near its ML bound and a stronger code decoded with a restricted search, which does not establish a general efficiency advantage. The reader identified this same weakest assumption, and I agree. I found no independent internal inconsistency in the modified stack design or in the LLR-ordering argument, and the reported simulated points do support the narrow claim at the tested SNRs. The missing element is a concrete d_min computation, which is fixable rather than fatal. The verdict therefore remains CONDITIONAL, unchanged from the reader's assessment.","tokens_in":15406,"tokens_out":8911,"duration_ms":96622,"concrete_test":"Compute the true minimum distance of each proposed concatenated code, especially the (128,36) code with the (2,1,6) inner code, by a constrained trellis or Viterbi search over the outer (16,9) Reed-Solomon codewords, or by running Magma's MinimumDistance on the explicit generator matrix. If d_min is at least 32, the slope-based inference in Section V.D is corroborated; if d_min is less than 32, the 'close to eBCH ML bound' claim will fail at higher SNR even though the decoder is near its own ML bound, and the efficiency comparison must be restricted to the reported SNR range.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's strongest claim is that the (128,36) concatenated code with the (2,1,6) inner code achieves BLER close to the eBCH ML bound at 3.0 and 3.5 dB with lower decoding complexity. That claim is only established for the two simulated SNR points unless the code's true minimum distance is comparable to the eBCH code's d_min=32. The paper explicitly does not compute d_min for the convolutional-inner constructions: Section V.D states, 'The minimum distances of other concatenated codes are unknown,' and then infers from the slopes of ML-bound curves over only 0.5 dB that 'their minimum distances are close to the comparable eBCH codes.' This inference is weak: in the low-SNR regime, the slope of a BLER curve depends on the error coefficient as well as on d_min, and the cited d_min*R values for the eBCH codes (11, 9, and 8.25) are close enough that a 0.5 dB slope comparison cannot reliably distinguish, say, d_min=24 from d_min=32. If the true d_min is smaller, the decoder can indeed be near its own ML bound while the code itself becomes strictly weaker than eBCH at higher SNR, so the headline claim of being 'much more efficient' would not generalize beyond the reported SNR grid. The product bound used for the (8,4,4) inner code (8x4=32) does not apply to the convolutional inner codes, since the inner map is not symbol-wise; the free distance of the convolutional code gives only a weak lower bound and does not by itself establish comparability with 32.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a low-complexity decoding scheme for short, low-rate binary linear block codes, with a focus on (128,36) codes. The key idea is to replace the conventional MRIP frame, which is based on channel observations, with an improved MRIP frame based on log-likelihood ratios obtained from a soft-in soft-out decoder for an inner code in a concatenated construction. The authors construct several (128,36) codes by concatenating a (16,9) Reed-Solomon code over GF(16) with rate-1/2 inner codes, including a (8,4,4) extended Hamming code, a (16,8,5) block code, and (2,1,4) and (2,1,6) convolutional codes. They also introduce a modified stack for A* decoding that does not require ordering, prove a search-order property for it in Theorem 1, and propose a codeword-independent stopping threshold. Simulations report BLER values close to eBCH ML bounds at Eb/N0 = 3 dB and 3.5 dB using path constraint lambda = 4, and the paper claims that the concatenated codes are much more efficient than the (128,36) eBCH code, which requires lambda = 5 to approach the same bounds. Additional results are given for codes of length 128 and 130 at other rates.","tokens_in":15801,"tokens_out":13042,"duration_ms":125096,"significance":"The improved MRIP idea is a genuinely useful contribution: it shows that soft information from an inner code can produce a more reliable basis for tree-search decoding of concatenated codes, potentially reducing the path constraint needed for near-ML decoding. The modified stack without ordering is simple, and the proof of Theorem 1 is explicit. The BLER results at the tested Eb/N0 points are concrete, falsifiable, and clearly presented. The paper is also honest in disclosing that the minimum distances of the convolutional-inner codes are unknown and that the threshold alpha is chosen by simulation. If the distance properties are confirmed and a matched-complexity comparison is supplied, the near-ML results for lambda = 4 would be a meaningful advance for low-rate short-block decoding. In its present form, the contribution is promising, but the headline efficiency claim is not yet established.","major_comments":[{"comment":"The efficiency claim in V.A.2 depends on the concatenated codes having minimum distances comparable to the benchmark eBCH code. Section V.D explicitly states that the minimum distances of the convolutional-inner concatenated codes are unknown and then infers from the slopes of the ML-bound curves over the 0.5 dB interval from 3 dB to 3.5 dB that 'their minimum distances are close to the comparable eBCH codes.' This inference is weak: at these SNR values the BLER slope is affected by the error coefficient and by the operating point, and the d_min R products of the eBCH codes (11, 9, and 8.25) are close enough that a 0.5 dB slope comparison cannot reliably distinguish, for example, d_min = 24 from d_min = 32. The product bound used for the (8,4,4) inner code (8 x 4 = 32) does not apply to the convolutional inner codes because the inner map is not symbol-wise. The authors should compute, or provide an analytic lower bound on, d_min for the convolutional-inner constructions (for example, from the active distances of the convolutional code and the RS outer code), or explicitly restrict the headline claim to the simulated Eb/N0 points.","section":"V.D"},{"comment":"The headline claim that the concatenated codes are 'much more efficient' than the (128,36) eBCH code is not supported by a matched-complexity comparison. The BLER improvement for eBCH with lambda = 5 is quoted in V.A.2 (about 4.6e-5 at 3 dB and 7e-6 at 3.5 dB), but the only complexity results in the paper, Section III.C and Fig. 2, are for eBCH with lambda = 4. No complexity curves are reported for the proposed concatenated codes. To conclude lower complexity at equal or better BLER, the authors should report the number of visited edges and comparisons (or real-number operations) for eBCH with PC-out-5 and for each concatenated decoder with PC-out-4 at the same target BLER points.","section":"III.C and V.A.2"},{"comment":"The 'ML bound' used as the benchmark appears to be the empirical lower bound described in V.A.3, obtained by counting decoded codewords c_best with M(r,c_best) greater than the metric of the correct codeword. That procedure gives a lower bound on the ML error probability, not the ML bound itself, and its tightness is not demonstrated in the paper. Since the central observation in V.A.2 is that the (2,1,6)-inner concatenated decoder operates close to the eBCH ML bound, the authors should state explicitly how the eBCH ML-bound curve in Fig. 7 was obtained and provide evidence of its tightness (for example, by checking agreement with known ML bounds for small codes or with exhaustive search). A loose lower bound would make the near-ML claim vacuous.","section":"V.A.3 / Fig. 7"},{"comment":"The stopping threshold M_TH,alpha in Eq. (5) is a free parameter chosen by simulation on the same codes: the paper states, 'In practice, we usually choose alpha to be 0.045 or 0.05 which is obtained through simulation.' The paper does not quantify P(alpha), the probability that the early-stopping test declares a non-ML codeword to be ML, nor does it report the sensitivity of the complexity and BLER results to alpha. If any of the efficiency comparisons use alpha = 0.05, the claim depends on this fitted parameter. Please provide a sensitivity analysis over alpha (for example, alpha = 0, 0.03, 0.05, 0.07) and, if possible, justify the choice on a code or SNR grid separate from the test points used in the headline comparison.","section":"II.E / V.A.4"}],"minor_comments":[{"comment":"There are typos in the abstract and introduction: 'ordered statics decoding' should be 'ordered statistics decoding', and 'URLCC' should be 'URLLC'.","section":"I / Abstract"},{"comment":"The sentence comparing slopes contains a duplicated phrase: 'for the three codes from Eb/N0 of 3 dB to 3.5 dB' appears twice and should be rewritten.","section":"V.D"},{"comment":"The BLER and CCDF plots do not report the number of simulation trials or confidence intervals; given the differences at the 1e-5 level, this information would help assess the statistical significance of the comparisons.","section":"Figures 1, 5, 7"},{"comment":"The ordered-statistics Gaussian approximation in Eqs. (7)-(11) assumes independent LLRs, an assumption the authors later note is violated; the text should clarify that this analysis is illustrative and that the main P(j|MRIP) results come from direct simulation.","section":"IV.C.2"},{"comment":"The exception clause in Theorem 1 ('except for the path c'' of which d_H(c''^{k-1}_0, z^{k-1}_0) = i+1 where c''_{k-1} != z_{k-1}') is difficult to parse; stating the exception in words or as a corollary after the proof would improve readability.","section":"III.B"},{"comment":"The complexity measure counts 'real-number operations' as visited edges plus comparisons, but the memory and push/pop costs of a 60000-node stack are not included; clarifying what operations are counted would make the complexity comparison easier to interpret.","section":"III.C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for a communications/information theory journal. The main risk is overclaiming efficiency based on an unmatched baseline and unverified distance properties. I recommend asking for a computation or analytic lower bound on d_min for the convolutional-inner concatenated codes, or a clear restriction of the claims to the simulated SNR range, plus a matched-complexity comparison for lambda = 5. The paper would also benefit from a brief reproducibility statement, as no simulation code or data are provided."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real new idea here is to build the MRIP frame from SISO decoder LLRs of the inner code rather than raw channel LLRs, and to use that improved frame with A* / PC-out decoding. That is a sensible and useful engineering contribution, and the paper shows it convincingly: the P(j|MRIP) and CCDF curves are the right way to demonstrate why the frame helps, and the BLER results at the tested Eb/N0 points look credible. The modified stack without ordering is also a modest but solid addition, and the theorem about the search order is clearly stated and proved.\n\nThe paper is honest where many are not: it explicitly says the minimum distances of the convolutional-inner concatenated codes are unknown. But then it leans on a weak inference to make the headline claim. The slopes of ML-bound curves over only 0.5 dB (3.0 to 3.5 dB) are not a reliable way to distinguish d_min=24 from d_min=32, especially at these low SNR values where error coefficients matter. The (8,4,4) inner code gives a clean product bound of 32, but the convolutional inner codes do not. If a (128,36) convolutional-inner code has d_min closer to 24, its BLER will flatten above the eBCH curve outside the simulated range, which undercuts the generality of the efficiency claim.\n\nA second soft spot is the complexity comparison. The abstract promises codes that are 'much more efficient' than eBCH, but the BLER for eBCH is quoted with lambda=5 while complexity is only shown for lambda=4. That is not a matched comparison. Fixing it is straightforward: report complexity for eBCH PC-out-5 at the same Eb/N0 points, or at least give the search-edge counts.\n\nMinor issues: the stopping threshold alpha in Eq. (5) is tuned by simulation on the same codes, so it is a fitted parameter rather than a predicted one; no code or data is shipped; there are no error bars, which matters for BLER around 1e-5. These are fixable and do not sink the narrow claim.\n\nThe central mechanism holds up for the specific codes and SNRs tested. The paper deserves serious peer review. I would ask the authors to compute or bound the actual d_min for the convolutional-inner constructions, add matched-complexity eBCH baselines, and be more careful about extrapolating from slopes.","headline":"Improved MRIP from inner-code SISO LLRs is a genuinely new twist with solid simulation support, but the headline 'much more efficient than eBCH' claim leans on an unverified minimum distance for the convolutional-inner codes.","tokens_in":16276,"tokens_out":1074,"would_cite":true,"duration_ms":12130,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94B35","94B65","94B12"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes a way to decode short, low-rate block codes with near-maximum-likelihood accuracy at reduced complexity, by using a concatenated coding structure to build a better reliability ordering.","keywords":["short block codes","low-rate codes","ordered statistics decoding","A* decoding","MRIP frame","concatenated codes","soft-in soft-out decoding","block error rate"],"falsifier":"Compute the exact minimum distance of the (128,36) concatenated code using the (16,9) Reed-Solomon outer code over GF(16) and a (2,1,6) convolutional inner code. If $d_{\\min}$ turns out to be significantly smaller than the eBCH code's $d_{\\min}=32$, then the BLER curves will flatten at a higher error floor, contradicting the claim that the concatenated code matches the eBCH code's performance at low error rates. A direct computation would settle this without relying on slope inference.","tokens_in":15214,"feed_emoji":"","tokens_out":4256,"duration_ms":30020,"temperature":0.7,"pith_summary":"The paper addresses the problem of decoding short, low-rate linear block codes, where standard near-optimum decoders like ordered statistics decoding (OSD) or A* decoding become impractical because the required search space is too large. Its central proposal is to use a concatenated coding structure—a Reed-Solomon outer code with a short inner code—and to derive the decoder's reliability ordering (the \"improved MRIP frame\") from soft-in soft-out (SISO) decoding of the inner code instead of from raw channel measurements. This improved ordering places fewer errors in the most reliable positions, so the same near-maximum-likelihood performance can be achieved with a smaller search order (lower $\\lambda$). For a benchmark (128,36) extended BCH code, the paper constructs several (128,36) concatenated codes and reports that the one using a (2,1,6) convolutional inner code achieves block error rates close to the maximum-likelihood bound at $E_b/N_0 = 3$ dB and 3.5 dB, while the eBCH code needs a higher search order to come close. If correct, this offers a practical way to decode low-rate short codes with near-optimal performance at significantly lower complexity.","feed_headline":"Concatenated codes shrink the search space for near-ML decoding","feed_subtitle":"For short, low-rate block codes, an improved reliability ordering from inner-code SISO decoding cuts decoding complexity while staying…","key_machinery":"The central object is the improved MRIP frame: instead of ordering received symbols by the magnitude of their channel LLRs, the decoder permutes the received vector (and the generator matrix) according to the magnitudes of LLRs output by a soft-in soft-out (SISO) decoder for the inner code of the concatenated code. This ordering is used to establish the systematic generator matrix and the hard-decision vector for A* decoding with a path constraint (PC-out-$\\lambda$). The paper also introduces a modified stack for A* decoding that requires no sorting—nodes are appended at the bottom—and proves (Theorem 1) that this stack searches goal nodes in order of their Hamming distance from the hard-decision MRIP vector, which gives it similar properties to OSD ordering without the comparison overhead.","core_discovery":"The paper claims that for short, low-rate block codes, replacing the conventional MRIP frame—built from the magnitudes of received symbols—with an improved MRIP frame built from the log-likelihood ratios (LLRs) produced by SISO decoding of an inner code can substantially reduce decoding complexity without sacrificing error performance. Specifically, for (128,36) concatenated codes using a (16,9) Reed-Solomon outer code over GF(16) and rate-1/2 inner codes, the A* decoder with path constraint PC-out-$\\lambda$ for $\\lambda = 4$ achieves BLER close to the maximum-likelihood bound for the (128,36) eBCH code. The best result comes from the (2,1,6) convolutional inner code: at $E_b/N_0 = 3$ dB, the BLER is about $3.5 \\times 10^{-5}$, while the eBCH code requires $\\lambda = 5$ to reach about $4.6 \\times 10^{-5}$. The paper also reports that the improved MRIP frame reduces the probability of errors in the MRIP positions, and that the complexity of computing the LLRs from the inner code is negligible compared to the tree search.","pith_inferences":["A natural extension is to apply the improved MRIP idea to other outer/inner code pairs, such as outer codes over larger fields or inner codes with higher memory convolutional codes, to see if the same complexity reduction holds for longer block lengths or higher rates.","The paper's slope-based inference that the minimum distances of the concatenated codes are close to the eBCH codes is indirect; a direct computation of $d_{\\min}$ for the (2,1,6)-based constructions would settle whether the near-ML performance persists at lower error rates than those simulated.","If the improved MRIP works as described, it could be combined with other reduced-complexity decoding techniques such as OSD variants, not just A* with PC-out, potentially giving a general recipe for low-rate short codes.","The use of a threshold $M_{TH,\\alpha}$ that does not require knowing $d_{\\min}$ suggests a practical path for codes where the minimum distance is unknown, but the paper leaves open the question of how to choose $\\alpha$ without simulation for a new code family."],"forward_implications":["If the claim holds, a (128,36) concatenated code with a (2,1,6) convolutional inner code can be decoded with $\\lambda = 4$ to achieve near-ML performance, whereas the (128,36) eBCH code requires $\\lambda = 5$ or larger, so the concatenated structure cuts the search space from $\\sum_{j=0}^{5} \\binom{36}{j}$ candidates to $\\sum_{j=0}^{4} \\binom{36}{j}$ candidates while also reducing per-node operati","The improved MRIP frame lowers the probability of errors in the 36 MRIP positions, and the paper's simulations show that the (2,1,6) inner code yields the lowest such probability among the considered inner codes at $E_b/N_0 = 3$ dB.","The approach is most beneficial for low-rate codes (rate below 1/2); the paper reports that for a rate-1/2 (130,65) concatenated code, the advantage over the (128,64) eBCH code is not significant, suggesting the method's efficiency is tied to low-rate applications.","The modified stack design is a separate contribution: it removes the need for ordering comparisons in A* decoding, at the cost of requiring a larger stack, which the paper argues is acceptable in modern technology. It achieves BLER similar to a conventional stack of half the size, with lower complexity.","The decoding complexity of obtaining LLRs from the inner code is claimed to be negligible compared to the tree search when $\\lambda \\geq 3$."],"supporting_citations":[{"why":"Defines OSD and the MRIP concept; supplies the asymptotic result that OSD-$\\lambda$ is near-optimum for $\\lambda \\geq \\lceil d_{\\min}/4 \\rceil - 1$, which motivates the paper's complexity comparison.","marker":"[2]"},{"why":"Introduces the path constraint (PC-$\\lambda$) that combines OSD ideas with A* decoding; the paper's PC-out-$\\lambda$ is a further simplification.","marker":"[8]"},{"why":"Defines PC-out-$\\lambda$, the exact decoding variant used in the paper's simulations.","marker":"[9]"},{"why":"Source for the (2,1,4) and (2,1,6) convolutional codes used as inner codes in the concatenated constructions.","marker":"[11]"},{"why":"The BCJR algorithm is used for SISO decoding of the convolutional inner codes to obtain the LLRs for the improved MRIP frame.","marker":"[15]"},{"why":"Provides the benchmark context: it shows that (128,36) eBCH codes with OSD-5 beat turbo, LDPC, and polar codes of the same length, establishing the eBCH code as the comparison target.","marker":"[1]"},{"why":"Supplies the early stopping criterion based on $M_{TH,\\hat{c}}$ that the paper uses to reduce decoding complexity.","marker":"[14]"}],"fun_headline_variants":["LLR-based ordering slashes decoding complexity for short codes","Inner-code LLRs guide smarter search for near-ML decoding","Concatenated structure cuts search space in low-rate decoding","Improved MRIP frame from SISO decoding trims tree search"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that the concatenated codes can match or beat the eBCH benchmark at low complexity depends on their true minimum distances being comparable to the eBCH codes, but the paper does not compute $d_{\\min}$ for the convolutional-code-based constructions; it only infers from the slopes of the simulated ML-bound curves that the distances are close. If the true minimum distance is smaller, the error-rate curves will flatten below the eBCH code at lower error rates than the simulated range, undermining the general efficiency claim.","fun_headline_variants_meta":{"raw":{"variants":["LLR-based ordering slashes decoding complexity for short codes","Inner-code LLRs guide smarter search for near-ML decoding","Concatenated structure cuts search space in low-rate decoding","Improved MRIP frame from SISO decoding trims tree search"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000792,"raw_usage":{"total_tokens":3539,"prompt_tokens":1041,"completion_tokens":2498,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":657,"completion_tokens_details":{"reasoning_tokens":2440}},"tokens_in":657,"tokens_out":2498,"duration_ms":19027,"temperature":1.0,"reasoning_tokens":2440,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:19:40.573937+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the exact minimum distance of the (128,36) concatenated code using the (16,9) Reed-Solomon outer code over GF(16) and a (2,1,6) convolutional inner code. If $d_{\\min}$ turns out to be significantly smaller than the eBCH code's $d_{\\min}=32$, then the BLER curves will flatten at a higher error floor, contradicting the claim that the concatenated code matches the eBCH code's performance at low error rates. A direct computation would settle this without relying on slope inference.","supporting_citations":[{"cited_title":"Soft-decision decoding of linear block codes based on ordered statistics,","cited_arxiv_id":null,"evidence_quote":"Defines OSD and the MRIP concept; supplies the asymptotic result that OSD-$\\lambda$ is near-optimum for $\\lambda \\geq \\lceil d_{\\min}/4 \\rceil - 1$, which motivates the paper's complexity comparison."},{"cited_title":"Tree-search decoding with path constraints for linear block codes,","cited_arxiv_id":null,"evidence_quote":"Introduces the path constraint (PC-$\\lambda$) that combines OSD ideas with A* decoding; the paper's PC-out-$\\lambda$ is a further simplification."},{"cited_title":"Tree-search decoding using reduced-size stacks,","cited_arxiv_id":null,"evidence_quote":"Defines PC-out-$\\lambda$, the exact decoding variant used in the paper's simulations."},{"cited_title":"Lin and D","cited_arxiv_id":null,"evidence_quote":"Source for the (2,1,4) and (2,1,6) convolutional codes used as inner codes in the concatenated constructions."},{"cited_title":"Optimal decoding of linear codes for minimizing symbol error rate (corresp.),","cited_arxiv_id":null,"evidence_quote":"The BCJR algorithm is used for SISO decoding of the convolutional inner codes to obtain the LLRs for the improved MRIP frame."},{"cited_title":"Short block-length codes for ultra-reliable low latency communications,","cited_arxiv_id":null,"evidence_quote":"Provides the benchmark context: it shows that (128,36) eBCH codes with OSD-5 beat turbo, LDPC, and polar codes of the same length, establishing the eBCH code as the comparison target."},{"cited_title":"An improvement to generalized-minimum- distance decoding,","cited_arxiv_id":null,"evidence_quote":"Supplies the early stopping criterion based on $M_{TH,\\hat{c}}$ that the paper uses to reduce decoding complexity."}],"review_version":1}