{"id":"8dddbe74-98b3-4aca-adb0-e2edc2242f30","arxiv_id":"2509.08079","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"For a new discrete soft-decision channel with linearly ordered reliabilities, the paper derives closed-form error and success exponents for random-code ML decoding and proves soft-decision decoding strictly outperforms hard-decision decoding.","lead":"This paper introduces a discrete soft-decision channel called the linear reliability channel, where the soft information is the rank ordering of received symbol reliabilities, and derives exact error exponents for maximum likelihood decoding. It gives a quantitative comparison of soft-versus hard-decision decoding and positions ORBGRAND as the exact ML decoder for this channel.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 5's flatness scale is too small to support the claimed order-statistic linearity; the LRC-to-continuous-channel approximation is not proven.","rationale":"The reader's weakest assumption identifies exactly the load-bearing concern: the approximation theorem connecting the LRC to continuous-noise channels is under-proven. The internal LRC analysis—guesswork LDPs, error exponents, and the ordering of hard/soft exponents—appears mathematically sound and detailed; the concern is confined to the motivating approximation. I agree with the reader's conditional verdict: the paper should be accepted conditionally, with the approximation step requiring rigorous treatment before the broad motivating claim is taken as established. The proposed concrete test would settle the matter by demonstrating that the flatness on the sigma^{-3/2} scale cannot extend to the Theta(1/sigma) quantile range, or, if it somehow does, by showing the contrary. My verdict is UNCHANGED because the reader's CONDITIONAL assessment already captures this risk and no new evidence shifts it.","tokens_in":26794,"tokens_out":16219,"duration_ms":187562,"concrete_test":"Take f0 to be the standard normal density. Compute f_L(l) exactly and define Q(p)=F_{|L|}^{-1}(p) for fixed p in (0,1) (e.g., p=0.5, p=0.9). Show that lim_{sigma->infty} f_L(Q(p))/f_L(0) = exp(-c_p^2/8) < 1, where c_p = sigma Q(p) is independent of sigma. Then compare Q(p) with the linearization Q_lin(p)=p/(2 f_L(0)) and verify that |Q(p)-Q_lin(p)|/Q(p) does not tend to 0 as sigma->infty. If true, Theorem 5's local flatness cannot imply the order-statistic linearity claimed in Section III-A.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's core claim that the LRC approximates continuous-noise channels at high noise variance rests on Section III-A. Theorem 5 proves sup_{|l|<=eps}|f_L(l)-f_L(0)|=O(1) only for eps=O(sigma^{-3/2}). But the order-statistic quantiles of |L| that matter for a block of length n lie at r=Theta(1/sigma): for any fixed fraction p of bits (e.g., p=0.5), the p-th quantile of |L| is c_p/sigma for a constant c_p determined by the density shape. Since eps*r^{-1}=O(sigma^{-1/2})->0, the flatness interval is asymptotically negligible relative to the quantile range. The bridge in Section III-A ('It follows from a standard result on order statistics...') requires the quantile function F^{-1}(p) to be approximately linear over a non-vanishing range of p, but local flatness at 0 with f_L''(0)=O(sigma^3) gives a curvature correction at r=Theta(1/sigma) that is O(1) relative to the linear term, not o(1). Thus the sorted reliabilities of the continuous channel are not shown to approach the linear reliability profile of the LRC; in fact, for Gaussian-like f_L the relative variation of the density over the quantile range is a constant (depending on p) as sigma grows, so the approximation does not improve with sigma for any fixed fraction of bits. This gap does not affect the internal LRC analysis (Theorems 9-20), but it invalidates the approximation claim as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the linear reliability channel (LRC), a discrete soft-decision channel in which the bit reliabilities are exactly linear under a uniformly random permutation. It claims that the LRC approximates a general class of continuous-noise channels at high noise variance, and it analyzes maximum-likelihood decoding in the LRC. The main technical results are explicit GRAND-based ML decoders (Theorems 9 and 10), large deviation principles for the guesswork of the soft- and hard-decision noise processes (Theorems 11, 12, and 15), and consequent error and success exponents for random-code ensembles (Theorem 18 applied to the LRC, with ordering results in Propositions 19-20). The proofs are detailed, including a long appendix for the hard-decision scaled cumulant generating function and for the ordering of critical rates.","tokens_in":27057,"tokens_out":23338,"duration_ms":277418,"significance":"If the results stand, the paper makes a valuable contribution: it provides one of the first discrete soft-decision channel models for which the soft- and hard-decision ML decoders and their error/success exponents can be analyzed explicitly, and it gives a quantitative comparison of soft versus hard decision. The use of published guesswork/GRAND tools is appropriate and not circular: the LRC-specific calculations are new derivations against that framework. The paper also contains substantial technical work in the form of the hard-decision sCGF and the critical-rate ordering proof. However, the advertised approximation of continuous-noise channels by the LRC is not established as stated; this weakens the motivational claim, although the internal LRC analysis is largely independent of that approximation.","major_comments":[{"comment":"The proof that Lambda'_Z(alpha) in (0, ln 2) is incomplete. The implication '0<Lambda_Z(alpha)<alpha ln2 for alpha>0, and hence Lambda'_Z(alpha)<ln2' is not valid for an arbitrary convex function with Lambda_Z(0)=0; a strictly convex function such as f(alpha)=a alpha^2+b alpha with a+b<ln2 can have f'(1)>ln2. This inference is used in Proposition 19 to locate the critical rates. The claim may be true for the specific sCGFs, but a direct proof is needed, or the argument should be restricted to alpha=1 where Appendix B already gives an explicit formula.","section":"Section V-B, Proposition 17"}],"minor_comments":[{"comment":"The Laplace and uniform distributions are listed as examples satisfying Assumption 4, but neither is strictly log-concave and C^4; the Laplace density is not C^1 at 0. The assumption covers normal and logistic but not the stated examples. Please correct the statement or relax the assumption.","section":"Section III-A, Assumption 4"},{"comment":"There are many mislabeled cross-references: 'Theorem 4' for Assumption 4, 'Theorem 6' for Lemma 6, and several 'Theorem 25'-'Theorem 33' for the corresponding lemmas in the appendices. Please renumber consistently.","section":"Throughout"},{"comment":"The notation w_tau(z)=sum_{i: tau(z)_i=1} i is ambiguous because tau is a permutation of [n], not a map on vectors. Please define the action explicitly, e.g., w_tau(z)=sum_{i:z_{tau(i)}=1} i or the equivalent.","section":"Definition 1 and proof of Lemma 2"},{"comment":"The text says there are n(n+1)/2 logistic-weight types; there are n(n+1)/2+1 possible weights (from 0 to n(n+1)/2). Also, in Section V-A 'wieldy' should be 'unwieldy'.","section":"Section II"}],"recommendation":"major_revision","confidential_remarks":"The paper's internal LRC analysis appears sound and is a solid contribution to the guesswork/GRAND literature. The binding issue is the approximation claim in Section III-A: as written, Theorem 5 does not support the assertion that the sorted reliabilities of continuous-noise channels become linear over any fixed fraction of the block. If the authors can either prove a genuinely uniform approximation over a non-vanishing quantile range or re-scope the paper to present the LRC as a stylized model motivated by the linear-reliability heuristic, the manuscript would be much stronger. The other proof gap (Proposition 17) is local and likely fixable directly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The LRC is a genuinely new object and the internal analysis is the real contribution. The decoder characterizations in Theorems 9-10 are clean, the guesswork LDPs are serious mathematics, and the exponent ordering results follow in a well-structured way. Appendix A, where the hard-decision sCGF is derived via saddle-point bounds on elementary symmetric polynomials, is particularly careful. I'd trust these results conditional on checking the details; the Rényi entropy ordering and the critical-rate argument are coherent. The paper also doesn't lean improperly on the authors' own GRAND/guesswork framework: those prior results are published, parameter-free, and applied rather than assumed.\n\nThe soft spot is Section III-A, and it is not minor. Theorem 5 proves local flatness of the LLR density only on a window of size O(sigma^{-3/2}), while the order-statistic quantiles that matter for a length-n block sit at scale Theta(1/sigma). That window is asymptotically negligible relative to the quantile range, so the leap to \"the sorted reliabilities are asymptotically linear\" via a generic order-statistics citation is unsupported. Worse, the curvature correction at the quantile scale is O(1) relative to the linear term, not o(1); for any fixed fraction of bits, the approximation does not improve as sigma grows. So the motivating claim in the abstract is not just unproven; the scaling evidence points the other way. The internal LRC analysis survives this, because it never needs the approximation, but the paper as advertised needs either a real theorem with the right scaling or a much more modest motivational statement.\n\nMinor remarks: the beta comparison values in Fig. 7 are free parameters and should be stated; no code or formal verification is provided, which is normal for this kind of theory paper.\n\nBottom line: this deserves serious referee time. The channel itself and the exact exponents are worth having in the literature, but the approximation claim should be heavily revised or removed before publication. I'd raise the Section III-A issue explicitly in review and ask the authors to either prove a quantile-scale result or narrow the claim. The rest of the paper can go through with standard checking.","headline":"Genuinely new discrete channel with exact ML exponents, but the abstract's claim that it approximates continuous-noise channels at high variance is not proven and looks wrong as stated.","tokens_in":668,"tokens_out":644,"would_cite":true,"duration_ms":30405,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A15","94A17","60F10","94B35"],"pacs":[],"model":"deepseek-v4-flash","headline":"The linear reliability channel makes soft-decision ML decoding exactly analyzable, and explicit error exponents show soft decisions strictly outperform hard decisions at every code rate.","keywords":["linear reliability channel","soft-decision decoding","error exponents","guesswork","large deviations","logistic weight","maximum likelihood decoding","random codes"],"falsifier":"Simulate a binary-input channel with normal or logistic noise at large variance, sort the absolute LLRs of length-n blocks, and compare the empirical spacings of the initial order statistics with the LRC's exactly linear spacings β/n for a β fitted from the first spacing. If the spacings diverge systematically from constant as σ grows, over the index range where the sorted values concentrate, the approximation claim would fail; the gap would be expected because Theorem 5 controls only the interval |l| ≤ O(σ^{-3/2}), while the relevant quantiles sit at l = Θ(1/σ).","tokens_in":26601,"feed_emoji":"📡","tokens_out":8688,"duration_ms":100203,"temperature":0.7,"pith_summary":"The paper introduces a discrete channel model, the linear reliability channel (LRC), in which the soft information available to a decoder is exactly the rank ordering of per-bit reliabilities. The motivation is that for common continuous-noise channels at high noise variance, sorted reliability magnitudes look approximately linear, so the LRC is offered as a faithful discrete proxy for that regime. On the LRC, the paper shows that hard- and soft-decision maximum-likelihood decoding reduce to simple guessing algorithms, derives large-deviation rate functions for the number of guesses, and turns these into explicit error and success exponents for random codes. The headline result is quantitative: soft-decision decoding strictly outperforms hard-decision decoding at every code rate, with the largest gap at intermediate noise levels. If the approximation is right, the LRC provides a clean benchmark for real soft-decision algorithms and a new statistical object, the logistic weight, for code design.","feed_headline":"Soft-decision decoding beats hard-decision at every code rate","feed_subtitle":"A discrete model with exactly linear reliabilities turns the soft-information advantage into computable error exponents.","key_machinery":"The LRC itself is the central object: a uniformly random permutation τ orders the bit reliabilities exactly as βτ(i)/n, so the soft information is combinatorial rather than metric. Two statistics carry the decoding analysis: the logistic weight w_τ(x)=Σ_{i:τ(x)_i=1} i, the soft-decision type statistic, and the Hamming weight, the hard-decision type statistic. The analytic workhorse is the large-deviations guesswork framework: the scaled cumulant generating functions of the optimal guessing processes are expressed through Rényi entropy rates, and their Legendre transforms give rate functions. A random-code large-deviations channel coding theorem then converts those rate functions into the err","core_discovery":"The paper introduces the linear reliability channel (LRC), where each channel use draws a uniformly random permutation τ and bit i flips with probability e^{-βτ(i)/n}/(1+e^{-βτ(i)/n}); the LLR magnitude of bit i is exactly βτ(i)/n, so the soft information is just the permutation. In this model, guessing noise patterns in increasing logistic weight is the soft-decision ML decoder and guessing by Hamming weight is the hard-decision ML decoder. The paper computes scaled cumulant generating functions for the optimal guessing processes, proves large-deviation rate functions for both decoders, and applies a random-code coding theorem to obtain explicit error and success exponents. Its central resu","pith_inferences":["If the LRC approximation holds for continuous-noise channels, then the practical performance of ORBGRAND on such channels becomes a corollary of exact ML optimality on the LRC, and the LRC exponents provide a benchmark for how much performance is lost by approximate soft-decision algorithms.","The paper's Rényi-entropy ordering argument is not obviously LRC-specific: any channel whose posterior noise distribution is majorized by its prior should show the same strict ordering of soft- versus hard-decision exponents. Constructing another discrete channel with that majorization property and checking the ordering would be a direct test.","The logistic-weight viewpoint suggests that code design for soft decisions should maximize minimum logistic weight rather than Hamming distance; one could test this by building small LRC codes and comparing ML block-error rates.","The BSC-vs-LRC parameter mapping is a heuristic tied to one Rényi order; because the rate functions have different curvature, no single-parameter map from β to p can match exponents at all rates, so a multi-order comparison would give a sharper equivalence."],"forward_implications":["Exact ML decoding on the LRC is fully characterized: soft-decision decoding guesses noise by increasing logistic weight, hard-decision decoding by Hamming weight, and neither guessing order depends on the noise parameter β.","Below capacity, the error probability decays exponentially with closed-form, computable exponents for both decoders; above capacity, the success probability decays exponentially as well.","Soft-decision decoding strictly outperforms hard-decision decoding across the rate range: better error exponents, better success exponents, larger capacity, and a later critical-rate transition.","The critical-rate transition, where the error exponent changes from linear to strictly convex, receives an intuitive decoder-level interpretation: below it errors come from atypically early spurious guesses, above it from atypically unlikely noise effects.","Matching the LRC to a BSC by equal average guesswork gives a heuristic mapping from β to bit-flip probability, showing how much softer soft information makes a channel look."],"supporting_citations":[{"why":"Defines ORBGRAND and the logistic weight, the algorithm that the paper shows becomes an exact ML decoder on the LRC.","marker":"[2]"},{"why":"Supplies the large-deviations guesswork theorem connecting scaled cumulant generating functions to Rényi entropy rates, used for both sCGF derivations.","marker":"[16]"},{"why":"Provides the random-code GRAND channel coding theorem used to convert guesswork LDPs into explicit error and success exponents.","marker":"[17]"},{"why":"Supplies the asymptotic count of sequences by logistic weight, used as the LRC analogue of Stirling's approximation for the binomial coefficient.","marker":"[18]"},{"why":"Gives the guesswork-moment inequality in terms of Rényi entropy used in the soft-decision sCGF proof.","marker":"[20]"},{"why":"Identifies the critical-rate phenomenon that the paper locates for each decoder and then orders between hard- and soft-decision decoding.","marker":"[28]"},{"why":"Is the standard order-statistics result invoked to pass from local flatness of the LLR density to approximate linearity of sorted reliabilities.","marker":"[29]"},{"why":"Supplies the Maclaurin/Newton inequalities used to prove monotonicity of hard-decision noise probabilities and log-concavity in the appendix.","marker":"[30]"}],"fun_headline_variants":["Soft info as permutation yields exact error exponents","Rank-order soft info: exact ML exponents for LRC","Exact error exponents from rank-ordered soft information","Permutation channel: soft-decision gain quantified"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The claim that the LRC approximates real continuous-noise channels rests on assuming that the shape of the log-likelihood-ratio density near zero controls the sorted reliability values across the meaningful initial range; the paper proves flatness only in a small shrinking interval around zero and cites a standard order-statistics result to make that bridge without quantified scaling conditions.","fun_headline_variants_meta":{"raw":{"variants":["Soft info as permutation yields exact error exponents","Rank-order soft info: exact ML exponents for LRC","Exact error exponents from rank-ordered soft information","Permutation channel: soft-decision gain quantified"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001305,"raw_usage":{"total_tokens":5125,"prompt_tokens":680,"completion_tokens":4445,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":424,"completion_tokens_details":{"reasoning_tokens":4384}},"tokens_in":424,"tokens_out":4445,"duration_ms":36775,"temperature":1.0,"reasoning_tokens":4384,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T21:18:50.505161+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a binary-input channel with normal or logistic noise at large variance, sort the absolute LLRs of length-n blocks, and compare the empirical spacings of the initial order statistics with the LRC's exactly linear spacings β/n for a β fitted from the first spacing. If the spacings diverge systematically from constant as σ grows, over the index range where the sorted values concentrate, the approximation claim would fail; the gap would be expected because Theorem 5 controls only the interval |l| ≤ O(σ^{-3/2}), while the relevant quantiles sit at l = Θ(1/σ).","supporting_citations":[{"cited_title":"Ordered reliability bits guessing random additive noise decoding,","cited_arxiv_id":null,"evidence_quote":"Defines ORBGRAND and the logistic weight, the algorithm that the paper shows becomes an exact ML decoder on the LRC."},{"cited_title":"Guesswork, large deviations, and shannon entropy,","cited_arxiv_id":null,"evidence_quote":"Supplies the large-deviations guesswork theorem connecting scaled cumulant generating functions to Rényi entropy rates, used for both sCGF derivations."},{"cited_title":"Partitions into distinct parts with bounded largest part,","cited_arxiv_id":null,"evidence_quote":"Supplies the asymptotic count of sequences by logistic weight, used as the LRC analogue of Stirling's approximation for the binomial coefficient."},{"cited_title":"An inequality on guessing and its application to sequential decoding,","cited_arxiv_id":null,"evidence_quote":"Gives the guesswork-moment inequality in terms of Rényi entropy used in the soft-decision sCGF proof."},{"cited_title":"A simple derivation of the coding theorem and some applications,","cited_arxiv_id":null,"evidence_quote":"Identifies the critical-rate phenomenon that the paper locates for each decoder and then orders between hard- and soft-decision decoding."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the standard order-statistics result invoked to pass from local flatness of the LLR density to approximate linearity of sorted reliabilities."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Maclaurin/Newton inequalities used to prove monotonicity of hard-decision noise probabilities and log-concavity in the appendix."}],"review_version":1}