{"id":"260cc9f8-4678-4e3a-995b-68dc466b2faa","arxiv_id":"2607.09602","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Soft CPA decoding of RM codes is exact-marginal and symmetric; density evolution (with hard-decision approximations) captures its rapid soft-information collapse and yields vanishing error only at vanishing rate.","lead":"This paper builds a density-evolution model for soft-decision collapsed projection-aggregation decoding of Reed-Muller codes on the AWGN channel and proves the decoder returns exact marginals and is symmetric. The analysis explains the decoder's fast mean/variance collapse and shows vanishing error only for vanishing-rate RM codes.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Independence of subspace messages is load-bearing for the asymptotic vanishing claim, and the paper already flags that finite-length dependence invalidates the quantitative bound.","rationale":"The Reader correctly isolates the independence assumption (Eqs. 47–51) as the weakest link supporting the strongest claim. My reading confirms that this is the single load-bearing gap: every other approximation (hard-decision FHT, Gaussian CLT under independence, vanishing-rate restriction) is either acknowledged or secondary. Because the paper already flags the finite-length dependence problem and never supplies a limiting-independence argument, the asymptotic claim remains conditional on an unproven hypothesis. No stronger internal contradiction appears, so the Reader’s CONDITIONAL verdict is unchanged; the concrete covariance test above would settle whether the gap closes or remains open.","tokens_in":23038,"tokens_out":555,"duration_ms":7260,"concrete_test":"Fix a vanishing-rate family (e.g., RM(m,2) or RM(m,⌊log m⌋)) and, for successive m, compute the empirical covariance matrix of the n_B aggregated messages u_i that enter the average (5). If the average pairwise correlation remains bounded away from zero (or decays slower than 1/n_B) as m\to∞, the variance reduction claimed in (49)–(51) fails and the asymptotic vanishing argument does not hold.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim (Section V final paragraph) is that soft-decision CPA achieves vanishing error probability for vanishing-rate RM codes. That claim is obtained by feeding the one-iteration DE model of Section IV into the asymptotic arguments of Section V. The DE model produces a Gaussian LLR whose variance is Var[u]/n_B (Eqs. 47–51) only under the explicit independence assumption among the n_B subspace messages. The paper itself records (p. 13) that the messages are dependent at finite length, that the true variance is therefore larger, and that the resulting error-probability bound (52) is unusable. The same independence is invoked for the asymptotic step: once n_B\to∞ (Proposition 3) the variance vanishes and the error probability vanishes. Because the paper never proves that the dependence becomes negligible as m\to∞ (or that the CLT still holds under the residual dependence), the vanishing-error conclusion rests on an unproven limiting independence that the finite-length analysis already shows fails. The hard-decision approximation of FHT (Eqs. 31–33) is secondary; the independence gap is the single point that, if it does not close, collapses the asymptotic claim.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper analyzes soft-decision collapsed projection-aggregation (CPA) decoding of Reed-Muller codes over the BIAWGN channel. It proves that CPA recovers the exact marginal probability of each bit (Proposition 1 and the derivation leading to Eq. (13)) and is symmetric with respect to the transmitted codeword (Lemma 1 and Proposition 2). From these properties a density-evolution model is constructed; the projection and FHT steps are replaced by hard-decision BSC approximations (Eqs. (31)–(33)), the aggregation messages are bounded by mean μ′ and variance (σ′)² (Eqs. (45)–(46)), and the average over n_B subspaces is treated as Gaussian by the central-limit theorem under an independence assumption (Eqs. (47)–(51)). Numerical histograms for RM(7,3) and RM(8,3) show that the model qualitatively tracks the rapid collapse of mean and variance. An asymptotic argument then claims that, for vanishing-rate RM codes (r finite), one-iteration CPA yields vanishing bit-error probability once n_B→∞ (Proposition 3 and the final paragraph of Section V).","tokens_in":23337,"tokens_out":907,"duration_ms":9247,"significance":"Exact-marginal and symmetry proofs for soft CPA fill a documented gap left by earlier hard-decision analyses of RPA/CPA. The density-evolution construction is the first attempt to track soft-message densities for this decoder family and correctly identifies the mean/variance collapse that underlies the observed fast convergence. The asymptotic vanishing-error claim for vanishing-rate codes is of genuine interest if it can be made rigorous, because it would place soft CPA on the same theoretical footing as the hard-decision results of [16], [17]. The paper is therefore a useful first step, even though the quantitative DE bounds remain loose and the asymptotic step rests on an unproven independence limit.","major_comments":[{"comment":"The asymptotic vanishing-error claim (Section V, final paragraph) rests on the CLT representation (Eqs. (47)–(51)) that treats the n_B subspace messages as independent, so that Var[¯l]=Var[u]/n_B. The manuscript itself records (p. 13) that finite-length dependence inflates the true variance and renders the error-probability bound (52) unusable. No argument is supplied that residual dependence vanishes as m→∞ (or that a CLT still holds under the residual dependence). Without such an argument the step “n_B→∞ ⇒ variance→0 ⇒ error→0” is incomplete; the claim therefore needs either a rigorous independence proof or a weaker statement that does not rely on the CLT variance formula.","section":null},{"comment":"The hard-decision approximation of FHT (Eqs. (31)–(33)) produces only an upper bound on the true soft-FHT error rate. While this is acknowledged, the subsequent mean lower bound μ′ and variance upper bound (σ′)² (Eqs. (45)–(46)) inherit the same looseness. Because these bounds are fed directly into the asymptotic analysis of Section V, the paper should either quantify the gap between hard-decision and soft FHT for order-1 RM codes or replace the hard-decision step by a soft analysis that still yields a positive mean and bounded variance.","section":null}],"minor_comments":[{"comment":"Figures 1–4 would be clearer if the empirical histograms and the DE Gaussians were overlaid on the same axes rather than plotted separately.","section":null},{"comment":"Notation for the projected LLR (l/B_i versus l/1) is inconsistent between Sections II and III; a single convention would help the reader.","section":null},{"comment":"The early-stopping threshold θ is mentioned but never given a numerical value in the simulation captions; adding it would improve reproducibility.","section":null},{"comment":"A few typographical slips remain (e.g., “thistanh” on p. 4, missing spaces around operators).","section":null}],"recommendation":"major_revision","confidential_remarks":"The independence gap identified by the stress-test is real and load-bearing for the strongest claim; a major-revision decision is therefore appropriate. The exact-marginal and symmetry proofs are solid and should be retained. If the authors can close the independence argument (or weaken the asymptotic statement accordingly), the paper would be a solid contribution to the theoretical literature on PA decoding."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This paper gives the first density-evolution look at soft-decision CPA decoding of RM codes over BIAWGN. The two clean results are that CPA recovers exact marginals (Proposition 1 plus the rewrite of the aggregation step) and that the decoder is symmetric, so the all-zero assumption is legitimate. Those proofs are careful and fill a gap left by the earlier hard-decision analyses of RPA/CPA.\n\nThey then build a DE recursion by replacing the soft projection and FHT with hard-decision BSC approximations, derive mean/variance bounds on the aggregated messages, and invoke a CLT under subspace independence. The simulations (RM(7,3) and (8,3)) show that the model qualitatively tracks the rapid collapse of both mean and variance that people observe in practice; that is useful intuition for why CPA converges so fast. The asymptotic claim is correctly limited to vanishing rate: once the number of subspaces goes to infinity the variance vanishes and Pe vanishes.\n\nThe soft spot is exactly the one the stress-test flags. Independence of the n_B subspace messages is load-bearing for both the Gaussian variance formula and the asymptotic step. The authors themselves note (around the figures) that finite-length dependence inflates the true variance, so their error-probability bound is unusable. They never prove that the dependence becomes negligible as m\to∞, so the vanishing-error conclusion sits on an unclosed gap. The hard-decision FHT approximation is secondary and already presented as an upper bound. None of this is fatal; it just means the quantitative DE is only qualitative and the asymptotic is conditional.\n\nThe paper is for people already working on RM or polar decoding theory who want analytic tools for PA-style algorithms. It is not a new practical decoder and does not claim capacity at positive rates. Math and citations look solid; no circularity. I would send it to referees—the core proofs and the DE framework are worth the discussion, and the limitations are already visible.","headline":"First soft DE treatment of CPA for RM codes; clean exact-marginal/symmetry proofs, but the vanishing-rate asymptotic rests on an independence assumption the authors already flag as imperfect at finite length.","tokens_in":23909,"tokens_out":504,"would_cite":true,"duration_ms":14696,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94B35","94B70"],"pacs":[],"model":"grok-4.5","headline":"Soft-decision collapsed projection-aggregation decoding of Reed-Muller codes has a density-evolution model that shows vanishing error at vanishing rate.","keywords":["Reed-Muller codes","collapsed projection-aggregation","density evolution","soft-decision decoding","BIAWGN channel","asymptotic analysis","exact marginals"],"falsifier":"Run the density-evolution recursion and the actual soft CPA decoder on a sequence of RM codes whose rates tend to zero (fixed order, growing length) and check whether both the predicted and the observed bit-error rates tend to zero at the same noise levels.","tokens_in":23936,"feed_emoji":"📡","tokens_out":858,"duration_ms":8917,"temperature":0.7,"pith_summary":"Reed-Muller codes are known to achieve capacity on several channels, and collapsed projection-aggregation (CPA) decoding has already been observed to come close to maximum-likelihood performance in practice. This paper supplies the first soft-decision density-evolution analysis of CPA over the binary-input AWGN channel. The authors first prove that each iteration of CPA recovers an exact marginal probability and that the decoder is symmetric, so the usual all-zero-codeword assumption is valid. They then build an approximate density-evolution recursion by replacing soft projection and fast-Hadamard decoding with hard-decision counterparts, track the resulting Gaussian means and variances, and show that both quantities collapse rapidly. The same recursion yields an asymptotic statement: when the code rate itself vanishes, the bit-error probability under CPA also vanishes. The analysis therefore supplies both a qualitative explanation for the decoder’s fast convergence and a first rigorous regime in which soft CPA succeeds.","feed_headline":"CPA decoding of RM codes vanishes when rate vanishes","feed_subtitle":"Density evolution proves soft projection-aggregation succeeds on vanishing-rate Reed-Muller codes","key_machinery":"Density evolution for CPA: after proving exact-marginal and symmetry properties, the authors replace soft projection/FHT by hard-decision counterparts, obtain a Gaussian density for the aggregated LLR via the central-limit theorem, and iterate the resulting mean-and-variance recursion.","core_discovery":"Soft-decision CPA decoding of Reed-Muller codes returns exact marginals, is symmetric, and admits a density-evolution model (under hard-decision approximations of projection and FHT) whose asymptotic analysis proves that the decoder achieves vanishing error probability whenever the code rate vanishes.","pith_inferences":["If a refined analysis that accounts for subspace dependence can restore a usable finite-length error bound, the same density-evolution engine would immediately yield concrete design rules for practical code lengths.","The exact-marginal property suggests that CPA can be viewed as a special case of belief propagation on a highly structured factor graph; tools from the BP literature may therefore transfer directly.","Extending the hard-decision approximation of the FHT step to a soft Gaussian approximation would tighten the mean lower bound and possibly enlarge the asymptotic success region beyond vanishing rate."],"forward_implications":["The same density-evolution framework can be used to design reduced-complexity soft PA variants by pruning subspaces while still guaranteeing the mean/variance collapse.","Because CPA recovers exact marginals, broadcast (rather than extrinsic) updates are information-theoretically justified and can replace more expensive extrinsic schedules.","The vanishing-rate success region supplies a concrete asymptotic benchmark against which future soft RM decoders can be compared.","The rapid collapse of both mean and variance explains the empirically observed early-stopping behaviour of CPA and can guide threshold design."],"fun_headline_variants":["Density evolution shows CPA vanishes error for vanishing-rate RM codes","Soft CPA on RM codes yields vanishing error as rate vanishes","CPA density model proves vanishing error when RM code rate vanishes","Soft-decision CPA succeeds with vanishing error for vanishing RM rates","Exact-marginal CPA decoding vanishes error probability as RM rate vanishes"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The analysis treats the messages returned by the different subspaces as independent so that the central-limit theorem can be applied; the paper itself notes that finite-length dependence inflates the true variance.","fun_headline_variants_meta":{"raw":{"variants":["Density evolution shows CPA vanishes error for vanishing-rate RM codes","Soft CPA on RM codes yields vanishing error as rate vanishes","CPA density model proves vanishing error when RM code rate vanishes","Soft-decision CPA succeeds with vanishing error for vanishing RM rates","Exact-marginal CPA decoding vanishes error probability as RM rate vanishes"]},"model":"grok-4.5","effort":"low","cost_usd":0.00353,"raw_usage":{"total_tokens":1149,"prompt_tokens":745,"num_sources_used":0,"completion_tokens":67,"cost_in_usd_ticks":35300000,"prompt_tokens_details":{"text_tokens":745,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":337,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":745,"tokens_out":67,"duration_ms":5340,"temperature":1.0,"reasoning_tokens":337,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T01:48:01.515241+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the density-evolution recursion and the actual soft CPA decoder on a sequence of RM codes whose rates tend to zero (fixed order, growing length) and check whether both the predicted and the observed bit-error rates tend to zero at the same noise levels.","supporting_citations":[],"review_version":1}