{"id":"ac3b2d70-bbd1-4563-9822-bd06be668c2c","arxiv_id":"2506.14756","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Testing multiple channel-estimate candidates and choosing the GRAND-based result with the highest posterior likelihood yields several dB BLER gains under channel estimation errors.","lead":"The authors propose decoding with several candidate channel estimates placed around a pilot-based estimate, then selecting the codeword with the highest posterior likelihood from a guessing-based decoder. This reduces block error rates under channel estimation errors by several decibels for 5G polar and CRC codes, potentially enabling shorter pilot sequences.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Selection by Eq. (8) is an unproven heuristic when comparing ORBGRAND outputs from different, mismatched channel candidates; all reported gains rest on it, so the central claim is not yet fully supported.","rationale":"The paper is an empirical proposal: the central claim is that testing multiple channel candidates and selecting by Eq. (8) gives BLER gains. The selection rule is exactly the novel mechanism, and it is the least supported link in the argument. The reader's weakest assumption identifies the same issue, and my reading agrees. Secondary gaps (no released code or data, Method 2 integrals given without derivation, no error bars) are real but do not by themselves undermine the central claim as strongly as the unproven model-selection criterion. The proposed ablation distinguishes whether the gain comes from the posterior rule or merely from decoding multiple candidates and choosing a valid codeword. Since the concern is addressable and does not contradict the reported results, the conditional verdict is appropriate rather than accept or reject.","tokens_in":8343,"tokens_out":6141,"duration_ms":63296,"concrete_test":"Reimplement Algorithm 1 on a fixed Monte Carlo set for the Fig. 2 CRC setting (σ_E^2=0.01, Rice K=10, MMSE, 16-QAM, M=5) at SNR around 16-20 dB. For each block, record (i) the candidate m* chosen by Eq. (8), (ii) the genie-optimal candidate that minimizes bit errors between q^(m) and the true transmitted codeword, and (iii) the BLER when the selection rule is instead uniform random choice among the M decoder outputs. Report the m*-vs-genie hit rate and the three BLER curves. High hit rate plus degraded random-selection BLER validates the posterior criterion; similar BLER under random selection would show multi-candidate decoding alone drives the gain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1's model-selection rule is the load-bearing step. For each candidate h^(m), ORBGRAND is run on LLRs λ^(m) built under h = h^(m), and Eq. (8) converts the output noise sequence z^(m) into P(Z = z^(m)) via B_l = e^{-|λ_l|}/(1+e^{-|λ_l|}). Those B_l are correct only if the candidate is the true channel. When the true h differs, the LLRs are miscalibrated by the residual term (h - h^(m)) x in the received signal, and no derivation shows that argmax_m P(m) is the maximum-likelihood or MAP model choice. In particular, Eq. (8) contains neither the model likelihood P(y | h^(m), c^(m)) nor the model prior P(h^(m)); it is a product of per-bit flip probabilities under one assumed channel. Remark 1 multiplies by the CEE density f_hE(Δ_C^(m)), which adds a prior on m but does not fix the missing likelihood normalization. Because every BLER curve in Figs. 2-4 is generated by this heuristic, the paper's central assertion that the 'most probable codeword' is selected is exactly the unproven point. The simulations could still be correct and robust; this is not a contradiction, but it is the fragile assumption that needs support.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two GRAND-based decoding methods to mitigate channel estimation errors (CEE) in flat-fading channels. Method 1 (Algorithm 1) tests multiple candidate channel estimates in the complex plane around the pilot-based estimate, computes LLRs for each candidate using ML, ZF, or MMSE detectors, runs ORBGRAND on each set of LLRs, and selects the candidate/codeword with the largest ORBGRAND posterior likelihood from Eq. (8). Method 2 refines the LLR computation by integrating over the residual CEE within each candidate's Voronoi cell, using the approach of [21]. The paper reports BLER simulations for [128,112] CRC and CA-Polar codes, showing gains of several dB over a baseline that ignores CEE and performance close to a genie-aided upper bound. The authors conclude that receivers could maintain error rates with shorter pilot sequences by exploiting ORBGRAND's parallelizability.","tokens_in":8553,"tokens_out":4581,"duration_ms":44569,"significance":"If the reported gains are robust, the paper offers a novel and practically interesting trade-off: it trades extra computation, which GRAND can parallelize well, for a reduction in required pilot length or estimation accuracy. The algorithm is clearly stated in pseudocode, and the simulation settings are partly specified, which forms a reasonable starting point for reproduction. However, the central model-selection heuristic is not theoretically justified, and the simulation evidence is limited to a small number of configurations without statistical support. The contribution is therefore promising but not yet fully supported in its present form.","major_comments":[{"comment":"The selection rule m* = argmax_m P(m) is the load-bearing step of the paper, but P(m) is computed from LLRs lambda^(m) that are only correctly calibrated when the true channel equals the candidate h^(m). When the true channel differs, the LLRs are miscalibrated, and Eq. (8) does not contain the model likelihood P(y | h^(m), c) or a prior over h^(m); it is merely the probability of the guessed noise sequence under one assumed channel. The paper provides no derivation showing that comparing these values across different mismatched candidates selects the most probable codeword. The heuristic reasoning in Section III-A ('candidates farther from the true fading value result in larger CEEs') does not supply the missing normalization. Because every BLER curve in Figs. 2-4 is produced by this rule, the central claim rests on an unproven assumption. Please either derive (8)-based selection as a proper MAP/ML model comparison, or provide supporting analysis and empirical checks (e.g., comparing against a criterion that includes a model prior and per-candidate likelihood, or against CRC-verified candidates) to demonstrate that the heuristic does not distort the results.","section":"Section III-A, Algorithm 1 and Eq. (8)"},{"comment":"The integrals in Eq. (10) and the claimed closed-form expressions in terms of the error function are asserted without derivation. The Voronoi cell boundaries for the nine-candidate grid, the domain of integration in the complex plane, and the exact truncated Gaussian density f(Delta_h_E^(m)) are not specified. Without these details, the Fig. 4 results cannot be reproduced or checked. Please provide the integration limits for the candidate grid, the derivation of the closed forms, and the explicit form of the truncated Gaussian distribution used.","section":"Section IV, Method 2 / Eq. (10)"},{"comment":"The only baseline in Figs. 2-4 is a receiver that ignores CEE and uses the pilot-based estimate in the LLR computation. CEE-aware LLR computation already exists, e.g., the method of [21] referenced in Section II. Without a comparison to at least one such baseline, the claimed gains of 'several dBs' demonstrate an advantage over a deliberately naive receiver but not over existing state-of-the-art practice. Please add a CEE-aware LLR decoding baseline for the same codes and parameters to substantiate the central claim.","section":"Section IV, baseline comparison"},{"comment":"The simulation evidence is narrow and incomplete: only two parameter configurations are reported (sigma_E^2 = 0.01 with MMSE and 16-QAM for Method 1; sigma_E^2 = 0.1 with ML and QPSK for Method 2), no code is provided, no confidence intervals are given, and the number of Monte Carlo trials is not stated. The mechanisms for generating Rice fading, the CEE h_E, and the pilot-based estimate are not described in enough detail to allow replication. Please provide these details and, ideally, additional configurations to separate the effects of the code, detector, constellation, and sigma_E^2.","section":"Section IV, simulation details"}],"minor_comments":[{"comment":"The notation P(m) is overloaded: in Algorithm 1 it denotes the value computed from Eq. (8), but the text alternates between calling it 'posterior probability', 'posterior likelihood', and 'weight'. It should be made precise that Eq. (8) gives the probability of the guessed noise sequence given the candidate-specific LLRs, not the posterior probability of the codeword.","section":"Section III-A, Algorithm 1"},{"comment":"The independence assumption P(z^(m), Delta_C^(m)) = P(z^(m)) P(Delta_C^(m)) used for the weighted variant is stated without justification. Since this weighting is presented as an optional improvement, either a justification or a sensitivity analysis of the final BLER with respect to this assumption should be included.","section":"Remark 1"},{"comment":"The parameter values for Fig. 4 (sigma_E^2 = 0.1, ML detector, QPSK) appear only in the text, not in the figure caption. Include them in the caption for readability.","section":"Section IV, Fig. 4"},{"comment":"The zero-forcing expression is correct but could be cleaner: the equalized output is y_ZF = x + (h_E / h_hat) x_i + n_i / h_hat; the current notation is understandable but slightly confusing because of the repeated index i in the text description.","section":"Section II-B, Eq. (4)"},{"comment":"The text contains some typographical issues, including 'V orono¨ı' with a misplaced space, 'algrithm' in Remark 1, and 'the the' style slips. A careful proofreading pass is recommended.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The self-citation pattern is not problematic: Duffy and Medard are principal authors of the GRAND/ORBGRAND literature, and the candidate-selection contribution in this manuscript is separable from the prior decoding machinery. The paper's fit to cs.IT is reasonable but leans toward a systems/communication-technique contribution; the theoretical content is thin. The main risk is that the selection heuristic in Eq. (8) is not justified across mismatched channel candidates; if the authors can provide a derivation or strong empirical validation (including a comparison to CEE-aware baselines), the contribution would be significantly strengthened. I would not reject on novelty grounds, but the current evidence does not yet support the strength of the abstract's claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you care about GRAND or about trading pilot overhead for decoder complexity. The idea is simple: instead of refining the channel estimate, run ORBGRAND under several candidate estimates in the complex plane and pick the candidate whose decoded output has the highest ORBGRAND posterior likelihood. Method 2 folds residual CEE distributions into LLRs via Voronoi cells. That is genuinely new relative to the cited work, and the reported BLER gains—several dB, closely approaching genie-aided—are striking if they reproduce.\n\nThe paper is clearly written and the simulations are coherent, but they are narrow: two code families at [128,112], one constellation and one CEE variance per method, no code, no data, no confidence intervals. The closed-form integrals in Method 2 are asserted without derivation; a referee should ask for them. The baseline ignores CEE entirely rather than using CEE-aware LLRs from [21], so part of the headline gain is against a strawman. Still, the gain is the point, and the genie bound is informative.\n\nThe load-bearing soft spot is the selection rule. Equation (8) is a product of per-bit flip probabilities under the assumed candidate; it is not a proper model posterior because it lacks the model likelihood and a prior over candidates. The stress-test note is right that no argument shows argmax_m P(m) is maximum-likelihood or MAP over candidates. That said, this is a heuristic, not a contradiction—GRAND itself is a heuristic decoder that works in practice, and the empirical gain is consistent across the tested settings. But the paper overstates it as selecting the 'most probable codeword.' A referee should ask for either an analytical argument or a broader simulation sweep (different constellations, CEE variances, code rates) to back the heuristic.\n\nMy verdict: conditional accept. The core idea is new and useful; the evidence is suggestive but not fully reproducible. I would send it to peer review. A cleared-up version would be a solid journal paper. I'd bring it to reading group as a nice example of a pragmatic receiver design.","headline":"A clever, plausible receiver-side trick for coping with channel estimation errors, with real but narrow simulation evidence; the selection heuristic is not proven but the paper is worth refereeing.","tokens_in":9170,"tokens_out":2393,"would_cite":true,"duration_ms":22607,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A40","94B35"],"pacs":[],"model":"deepseek-v4-flash","headline":"ORBGRAND's posterior likelihood can choose among neighboring fading estimates, reclaiming several dB of block-error-rate loss from imperfect channel estimation.","keywords":["channel estimation error","ORBGRAND","guessing random additive noise decoding","fading channel","pilot-based channel estimation","block error rate","CA-Polar codes","CRC codes"],"falsifier":"Take a simulated block with known transmitted codeword and known true fading coefficient, run Algorithm 1, and compare the candidate it selects with the one whose decoded word has the fewest bit errors. If, at $\\sigma_E^2 = 0.1$, the argmax-posterior rule selects the best candidate no more often than the pilot-only estimate would, or if the reported BLER gain disappears when $\\sigma_E$ is set larger than the candidate grid spacing, the central claim is refuted.","tokens_in":8016,"feed_emoji":"📡","tokens_out":10309,"duration_ms":84126,"temperature":0.7,"pith_summary":"This paper asks whether a receiver that already has a noisy pilot-based estimate of a fading channel can decode more reliably without getting a better estimate. Its answer is yes: instead of decoding only with the pilot estimate, the receiver runs ORBGRAND, a soft-input decoder that generates candidate noise sequences in reliability order, on several nearby channel candidates in parallel and keeps the codeword whose ORBGRAND posterior likelihood is highest. In simulations with [128,112] CRC and CA-Polar codes under Rice fading, this candidate-testing approach gives several dB of block-error-rate gain over decoding that ignores channel estimation error, approaching a genie-aided benchmark that knows the transmitted codeword. A refined version, which accounts for residual estimation error inside each candidate's Voronoi cell when forming LLRs, adds roughly another 2 dB in the tested CRC case. The practical stake is that pilots could be shortened, raising the data rate, without sacrificing error rate.","feed_headline":"Trying nearby channel guesses wins several dB in decoding","feed_subtitle":"That could shrink pilot overhead and raise data rates without sacrificing error rate.","key_machinery":"The load-bearing object is the ORBGRAND posterior likelihood, $P(\\vec Z = \\vec z) = \\prod_{l:z_l=0}(1-B_l) \\prod_{l:z_l=1} B_l$ with $B_l = e^{-|\\lambda_l|}/(1+e^{-|\\lambda_l|})$, computed from the LLR vector produced under each channel candidate. It acts as a model comparator: among the parallel decode attempts, the candidate whose associated guessed noise sequence has the largest posterior likelihood is trusted. Around it, the ML, ZF, and MMSE LLR approximations translate each candidate into soft inputs, and Method 2's additional machinery is the Voronoi-cell-truncated Gaussian integration of residual CEE inside the LLR computation, following the imperfect-CSI LLR approach.","core_discovery":"The central claim is that ORBGRAND's posterior likelihood formula can be repurposed as a selection rule over channel estimates. Given a pilot-based estimate $\\hat h$, the receiver forms $M$ candidates $\\hat h^{(m)} = \\hat h + \\Delta_C^{(m)}$ placed on a grid scaled by the known CEE standard deviation $\\sigma_E$, computes LLRs for each candidate with an ML, ZF, or MMSE detector, decodes each candidate's LLR vector with ORBGRAND, and selects the codeword/candidate pair with the largest posterior probability $P(m)$ from Eq. (8), optionally weighted by the Gaussian density of the CEE. The paper reports that this rule recovers most of the loss caused by channel estimation error: BLER gains over 5 dB relative to ignoring CEE in the 15-20 dB SNR range for [128,112] CRC and CA-Polar codes at $\\sigma_E^2 = 0.01$, approaching genie-aided candidate selection. Method 2 replaces each candidate's LLRs by likelihoods integrated over the residual CEE inside the candidate's Voronoi cell, following imperfect-CSI LLR treatment, and is reported to add about 2 dB over Method 1 at $\\sigma_E^2 = 0.1$ for a [128,112] CRC code.","pith_inferences":["If the posterior-likelihood selection is well calibrated, the same rule could be iterated: center a finer candidate grid on the winning estimate and re-decode, a refinement the paper does not test.","The method's data-rate benefit should be compared against simply increasing pilot length by the equivalent number of pilot symbols; the paper demonstrates lower BLER at a fixed CEE but does not run that end-to-end rate comparison.","For time-varying channels, the winning candidate from one block could seed the candidate set of the next block, a natural online-tracking extension of the paper's stated future direction.","The candidate-placement rule assumes the receiver knows $\\sigma_E^2$ and that the CEE is Gaussian; a stress test with mismatched $\\sigma_E^2$ or non-Gaussian CEE would show how fragile the selection score is."],"forward_implications":["In the settings simulated, Method 1 gives block-error-rate gains over 5 dB relative to ignoring channel estimation error for [128,112] CRC and CA-Polar codes, approaching the genie-aided candidate-selection benchmark.","The procedure is defined for ML, ZF, and MMSE LLR computations, so the candidate-selection idea is not tied to a single detector type.","Because ORBGRAND is highly parallelizable, the M parallel decode attempts add manageable computational overhead in hardware-oriented implementations.","Method 2, which integrates residual estimation error over each candidate's Voronoi cell when forming LLRs, adds roughly 2 dB beyond Method 1 in the tested CRC configuration."],"supporting_citations":[{"why":"Supplies ORBGRAND, including the ordered-reliability query order and the posterior likelihood expression Eq. (8) on which the candidate-selection rule is built.","marker":"[25]"},{"why":"Supplies the LLR formulation for channels with imperfect channel state information, which Method 2 adapts by integrating over residual CEE within each Voronoi cell.","marker":"[21]"},{"why":"Establishes CRC codes as usable error-correcting codes with GRAND, providing one of the two code families used in the numerical evaluation.","marker":"[31]"},{"why":"Establishes 5G NR CA-Polar maximum-likelihood decoding with GRAND, providing the other code family used in the numerical evaluation.","marker":"[32]"},{"why":"Supplies the fading-channel detector setup (ML, ZF, MMSE LLRs) used to generate soft inputs from noisy channel estimates.","marker":"[28]"}],"fun_headline_variants":["Guessing channel estimates boosts decoding by 5 dB","GRAND channel guesses deliver 5 dB gain","Testing channel estimate neighbors improves error correction","ORBGRAND selects channel estimate via posterior likelihood"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's selection step assumes that the confidence score ORBGRAND assigns to a decoded word is still a fair way to compare decodes made under different, partly wrong channel guesses. No proof is given that picking the highest score finds the best codeword; if the score is miscalibrated for guesses far from the true channel, the gains rest on an unproven heuristic.","fun_headline_variants_meta":{"raw":{"variants":["Guessing channel estimates boosts decoding by 5 dB","GRAND channel guesses deliver 5 dB gain","Testing channel estimate neighbors improves error correction","ORBGRAND selects channel estimate via posterior likelihood"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0007,"raw_usage":{"total_tokens":3205,"prompt_tokens":1033,"completion_tokens":2172,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":649,"completion_tokens_details":{"reasoning_tokens":2113}},"tokens_in":649,"tokens_out":2172,"duration_ms":16376,"temperature":1.0,"reasoning_tokens":2113,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:09:31.883138+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a simulated block with known transmitted codeword and known true fading coefficient, run Algorithm 1, and compare the candidate it selects with the one whose decoded word has the fewest bit errors. If, at $\\sigma_E^2 = 0.1$, the argmax-posterior rule selects the best candidate no more often than the pilot-only estimate would, or if the reported BLER gain disappears when $\\sigma_E$ is set larger than the candidate grid spacing, the central claim is refuted.","supporting_citations":[{"cited_title":"Ordered reliability bits guessing random additive noise decoding,","cited_arxiv_id":null,"evidence_quote":"Supplies ORBGRAND, including the ordered-reliability query order and the posterior likelihood expression Eq. (8) on which the candidate-selection rule is built."},{"cited_title":"The LLR metric for q-ary LDPC codes with MPSK modulation over Rayleigh channels with im- perfect CSI,","cited_arxiv_id":null,"evidence_quote":"Supplies the LLR formulation for channels with imperfect channel state information, which Method 2 adapts by integrating over residual CEE within each Voronoi cell."},{"cited_title":"CRC codes as error correction codes,","cited_arxiv_id":null,"evidence_quote":"Establishes CRC codes as usable error-correcting codes with GRAND, providing one of the two code families used in the numerical evaluation."},{"cited_title":"5G NR CA-Polar maximum likelihood decoding by GRAND,","cited_arxiv_id":null,"evidence_quote":"Establishes 5G NR CA-Polar maximum-likelihood decoding with GRAND, providing the other code family used in the numerical evaluation."},{"cited_title":"GRAND for fading channels using pseudo-soft information,","cited_arxiv_id":null,"evidence_quote":"Supplies the fading-channel detector setup (ML, ZF, MMSE LLRs) used to generate soft inputs from noisy channel estimates."}],"review_version":1}