{"id":"be90137f-b276-47a4-a92b-2de7eb00bd8a","arxiv_id":"2607.17960","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"For a low-rate MET-LDPC code at very low SNR, SPA decoding is most robust to fixed-point precision, and Q8.4 is the practical hardware-efficient choice for SPA.","lead":"This paper compares three LDPC error-correction algorithms under fixed-point arithmetic for a quantum-key-distribution reconciliation code. It reports that SPA decoding is the most robust to low precision, and that Q8.4 is the most hardware-efficient format that still works well for SPA.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Q8.4 SPA recommendation depends on an unspecified fixed-point tanh/atanh realization; without an FPGA-realistic model, the hardware claim is unverified.","rationale":"The most load-bearing condition for the central claim is that the 'fixed-point SPA' is actually implementable at Q8.4 without hidden floating-point assistance. The reader's weakest assumption correctly identified the unspecified tanh/atanh realization; I agree. A machine-checked proof is absent, and the simulation code is not provided, so the only evidence is the FER plots. The paper's own Section I-C only guarantees quantization of inputs/storage, and Section I-D explicitly says SPA keeps message storage in fixed-point form, which is weaker than saying the nonlinear check-node computation is fixed-point. If Eq. (5) is evaluated in floating point, the Q8.4 result is not a hardware-relevant SPA. The concrete LUT test would settle this. The secondary internal inconsistency about NMS's best format (Section III 'Q16.8 and Q12.6 respectively' vs. conclusion 'Q16.8 optimal threshold for reduced-complexity decoders') also needs editorial fixing, but the SPA fixed-point realization is more fundamental because it undercuts the headline recommendation. Verdict remains CONDITIONAL; no change from reader.","tokens_in":7483,"tokens_out":4007,"duration_ms":33747,"concrete_test":"Re-implement the SPA check-node update with an explicit fixed-point tanh/atanh LUT representative of FPGA practice: e.g., 2^8-entry tables over [-8,8] for tanh(|m|/2) and atanh, with inputs/outputs quantized to Q8.4 and intermediate products accumulated/saturated in Q8.4. Run the same N=10000, R=0.02 MET-LDPC ensemble at SNR=-20..-15 dB, 2000 frames, and compare FER/mean-iterations with the reported Q8.4 and Q12.6 curves. If the LUT-based Q8.4 FER remains within, say, 0.1-0.2 dB of the reported SPA waterfall and still beats NMS Q16.8, the recommendation survives; if the waterfall shifts or the NMS gap closes, the hardware claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central hardware recommendation — SPA with Q8.4 as the best performance/scalability point — rests on the SPA decoder being genuinely fixed-point throughout. Sections I-C and I-D describe quantization of channel LLRs, saturation to [-Lmax, Lmax], and fixed-point message storage, but they never specify how Eq. (5)'s 2 atanh(prod tanh(...)) check-node update is evaluated in fixed point. In particular: is tanh(·/2) computed by a LUT, CORDIC, polynomial, or in floating point and then quantized? What are the input/output word lengths of that block, and how are the products inside the atanh accumulated and saturated? If the simulations compute Eq. (5) in floating point and only quantize stored messages, the 'Q8.4 SPA' results describe a hybrid float/fixed decoder, not an 8-bit datapath. Real FPGA SPA decoders must approximate tanh/atanh, and the approximation error at Q8.4 could lift the FER curve, shrink the waterfall, or alter the ranking vs. NMS. Since the conclusion explicitly invites hardware designers to adopt SPA Q8.4, the missing implementation detail is load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript compares three LDPC decoding algorithms—SPA, MSA, and NMS—under a common fixed-point quantization framework, targeting low-SNR reconciliation for CV-QKD. The authors use a single MET-LDPC code with rate 0.02, block length 10000, SNR range -20 to -15 dB, and four fixed-point formats (Q8.4, Q12.6, Q16.8, Q24.12) with a common clipping limit Lmax=8.0. They report FER and average iterations, concluding that SPA is the most reliable decoder, that Q16.8 is the lowest reliable precision for reduced-complexity decoders, that NMS outperforms MSA, and that SPA in Q8.4 offers the best performance/efficiency trade-off for hardware implementation.","tokens_in":7835,"tokens_out":6289,"duration_ms":56841,"significance":"If the central claim holds, the paper gives useful practical guidance to hardware designers: an 8-bit SPA with 4 fractional bits may be sufficient for low-SNR CV-QKD reconciliation without a large FER penalty, and Q16.8 NMS is a promising reduced-complexity fallback. The study's unified simulation setup—same parity-check matrix, same quantization pipeline, same SNR grid—is a genuine strength, and the disclosure of the NMS α sweep is transparent. The equations for SPA, MSA, and NMS are standard. However, the quantitative conclusions currently rest on an incompletely specified fixed-point SPA model, a parameter-tuned NMS comparison, and no floating-point baseline or statistical uncertainty estimates, so the practical hardware recommendation is not yet fully supported.","major_comments":[{"comment":"The central recommendation to implement SPA with Q8.4 assumes that the SPA check-node update itself is computed in fixed-point arithmetic, but the manuscript never specifies how Eq. (5), 2 atanh(prod tanh(...)), is evaluated. The text in Sec. I-C says SPA 'keeps message storage in fixed-point form,' which suggests the nonlinear functions may be computed in floating point and quantized only at message boundaries. If so, the reported Q8.4 SPA results describe a hybrid float/fixed decoder, not an 8-bit datapath. Real FPGA implementations of tanh/atanh use LUTs, CORDICs, or polynomial approximations with finite intermediate word lengths, and those approximations can shift the FER curve and potentially change the ranking versus NMS. Please specify the fixed-point realization of every intermediate in Eq. (5), including input/output word lengths, product accumulation/saturation, and the tanh/at","section":"Secs. I-C/I-D, Eq. (5)"},{"comment":"The NMS normalization factor α=0.65 was selected by sweeping α from 0.40 to 0.85 on the same code, the same decoder, and the same SNR range used in the main experiments. MSA corresponds to α=1 and SPA has no such tunable parameter. The comparison 'NMS clearly outperformed MSA' and the Q16.8 threshold for reduced-complexity decoders are therefore not parameter-free. Please report the sensitivity of the NMS results to α, or use a protocol that selects α on held-out SNR points, before drawing general conclusions about the NMS/MSA ranking.","section":"Sec. II, Eq. (7)"},{"comment":"There is no floating-point baseline in the reported curves, so the claim that Q8.4 SPA 'preserves excellent decoding performance' cannot be quantitatively supported—one cannot tell how much degradation is due to quantization. In addition, only 2000 frames are simulated per SNR point with no error bars; near a waterfall, this makes differences in FER and the claimed optimum precisions statistically fragile. Finally, 'hardware efficiency' is inferred only from bit width; no area, throughput, memory, or power estimates are given. Please add a floating-point reference, confidence intervals or enough frames for stable FER estimates, and at least first-order complexity estimates to substantiate the 'best balance' conclusion.","section":"Secs. III--IV"},{"comment":"There is an internal inconsistency in the reported optimum precisions. Sec. III says 'both MSA and NMS achieve their most suitable operating condition under the Q16.8 and Q12.6 formats respectively,' but the conclusion states that 'Q16.8 emerged as the optimal precision threshold' for reduced-complexity decoders. If NMS is best at Q12.6, Q16.8 is not the lowest reliable format. Also, Sec. III describes SPA as 'reaching its quantization-limited behavior at Q8.4' while Sec. IV recommends Q8.4 as preserving excellent performance. Please clarify the criterion used for 'optimal' and align the statements in Secs. III and IV.","section":"Sec. III"}],"minor_comments":[{"comment":"The manuscript says decoder performance was evaluated in terms of FER and average number of iterations, but no figure or table presents the average iteration results. Please include them or remove the claim.","section":"Sec. II"},{"comment":"The caption reads 'BIT ERROR RATE (FER)'; the metric is frame error rate. Use the correct term consistently.","section":"Fig. 1 caption"},{"comment":"The statement that 'increased fractional precision reduces the available dynamic range in fixed-width accumulators' is not applicable as written: Q24.12 has a larger dynamic range than Q16.8 because the total word length also increases. Clarify the intended comparison.","section":"Sec. III"},{"comment":"The title mentions QC-MET-LDPC, but the quasi-cyclic lifting/construction is never described. Either define the quasi-cyclic structure or adjust the title to 'MET-LDPC'.","section":"Title and Sec. I-B"},{"comment":"Reference [16] is incomplete: it lacks the full title and publisher details. Please complete it.","section":"Ref. [16]"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the central empirical question is relevant, but the load-bearing details omitted from the fixed-point SPA model and the lack of a floating-point baseline make the current version unsuitable for acceptance. The issues appear addressable in revision: adding implementation-level detail for Eq. (5), a sensitivity analysis for α, confidence intervals, and a floating-point reference would substantially strengthen the claims. I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take: this is a useful, narrowly-scoped simulation study comparing SPA, MSA, and NMS under four fixed-point formats for a low-rate MET-LDPC code in the CV-QKD low-SNR regime. The ranking (SPA most robust, NMS better than MSA among simplified decoders) is plausible and consistent with standard LDPC theory. The paper does not introduce new mechanisms, but the specific combination — one code, one SNR range, four formats, unified graph/matrix/quantization pipeline — is new in the cited literature and directly relevant to hardware designers.\n\nWhat it does well: the experimental design is uniform across decoders. Same parity-check matrix, same stopping criterion, same clipping and quantization for a given format. That makes the comparison cleaner than most. The choice of a rate-0.02 MET-LDPC code at SNR around -20 to -15 dB matches the CV-QKD reconciliation regime, and the paper correctly emphasizes that numerical precision interacts with the decoder rule rather than being a separable parameter.\n\nSoft spots, in order of severity. First, the internal contradiction about precision thresholds: the results section says MSA is most suitable at Q16.8 and NMS at Q12.6, while the abstract and conclusion state Q16.8 was the lowest consistent precision for reduced-complexity decoders. That is not a small typo; the central \"Q16.8 threshold\" claim depends on it. Second, the stress-test concern is real: nowhere does the paper specify how the SPA check-node update in Eq. (5) is computed in fixed point. If tanh/atanh are evaluated in floating point and only stored messages are quantized, the \"Q8.4 SPA\" results describe a hybrid float/fixed decoder, not an 8-bit datapath, and the hardware recommendation loses its basis. This needs to be stated explicitly. Third, the NMS-vs-MSA comparison is partly constructed: α=0.65 was chosen by sweeping on the same code and SNR range, so it is an optimized result for NMS, and MSA gets no equivalent tuning. Fourth, no error bars on FER from 2000 frames and no floating-point baseline make the quantitative thresholds harder to trust; a floating-point curve would also anchor the fixed-point degradation.\n\nThe SPA-is-best conclusion likely survives these fixes. The specific precision recommendations might not. The paper is for hardware engineers implementing CV-QKD reconciliation, not for a theory audience. It deserves a serious referee, but the referee should be asked to check the contradiction, request a floating-point baseline, and require the SPA arithmetic to be specified.\n\nMy recommendation: engage with it, send to peer review, but expect revision.","headline":"Plausible fixed-point comparison for CV-QKD reconciliation hardware, but the precision-threshold claims are weakened by an internal contradiction and an unspecified SPA datapath.","tokens_in":8312,"tokens_out":2760,"would_cite":false,"duration_ms":22282,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a fixed-point sum-product decoder at Q8.4 precision offers the best reliability-to-hardware-cost balance for low-SNR LDPC decoding in CV-QKD reconciliation.","keywords":["fixed-point LDPC decoding","CV-QKD reconciliation","sum-product algorithm","normalized min-sum","MET-LDPC codes","low-SNR decoding","hardware implementation","frame error rate"],"falsifier":"Implement an SPA decoder on an FPGA using a concrete tanh/atanh lookup table with Q8.4 message storage and the same MET-LDPC code at SNR around −17 dB; if the measured FER degrades noticeably relative to the paper's simulated Q8.4 curve, or if a different tanh approximation changes the decoder ranking, the central recommendation would not carry over to hardware.","tokens_in":7424,"feed_emoji":"🔐","tokens_out":1632,"duration_ms":15557,"temperature":0.7,"pith_summary":"The paper sets out to determine which LDPC decoding algorithm and which fixed-point number format should be used when moving continuous-variable quantum key distribution reconciliation from floating-point simulation to hardware. It compares SPA, MSA, and NMS under the same parity-check matrix, graph representation, clipping policy, and quantization pipeline, sweeping four fixed-point formats. The central claim is that SPA preserves excellent decoding performance even at Q8.4, while reduced-complexity decoders need Q16.8 to operate reliably, where NMS clearly beats MSA. A sympathetic reader would care because these choices directly affect memory footprint, arithmetic cost, and achievable secret-key rate in practical CV-QKD systems.","feed_headline":"SPA at 8-bit precision wins low-SNR decoder comparison","feed_subtitle":"Q8.4 sum-product decoding balances reliability and cost for CV-QKD; Q16.8 normalized min-sum is the simpler fallback.","key_machinery":"The central mechanism is the fixed-point quantization pipeline, where real-valued messages are quantized as xq = round(x*2^nf), clipped to [−Lmax, Lmax] with Lmax=8.0, and saturated to the signed integer range of the chosen Qnt.nf format. The three decoders share this pipeline and differ only in the check-node update rule: SPA uses the exact tanh/atanh product rule, MSA uses the min-sum approximation, and NMS scales the min-sum output by α=0.65. This design isolates the interaction between arithmetic precision and decoding rule.","core_discovery":"Under a unified low-SNR fixed-point decoding framework, the sum-product algorithm (SPA) consistently achieved the lowest frame error rate and the clearest waterfall behavior across all tested precisions, remaining stable even at Q8.4. For reduced-complexity decoders, Q16.8 was the lowest precision that maintained reliable operation, and normalized min-sum (NMS) with normalization factor 0.65 outperformed plain min-sum (MSA). The practical conclusion is that SPA combined with Q8.4 is the most favorable operating point when both decoding reliability and implementation scalability are considered.","pith_inferences":["The paper's SPA results assume a particular fixed-point realization of the tanh/atanh check-node update; in real hardware a lookup-table or polynomial approximation may alter the Q8.4 recommendation, so a companion hardware-emulation study would be a natural next step.","The precision thresholds found here for a rate-0.02, length-10000 MET-LDPC code may shift with code rate, block length, or different MET-LDPC degree profiles, meaning the Q8.4/Q16.8 guidance is best treated as a starting point rather than a universal rule.","A testable extension would be to measure decoding throughput and resource utilization on an FPGA for SPA Q8.4 versus NMS Q16.8, converting the FER-based ranking into an explicit area-latency-reliability trade-off.","Because SPA proved robust at low precision, combining it with learned or optimized quantization of the nonlinear check node could push the practical operating point below Q8.4 while preserving waterfall performance."],"forward_implications":["Hardware designers can target an SPA decoder with Q8.4 fixed-point arithmetic for low-SNR CV-QKD reconciliation, reducing memory and datapath width compared to wider formats.","For simplified implementations, Q16.8 NMS is a viable reduced-complexity alternative that outperforms MSA at the same precision.","Wider fixed-point formats beyond the identified thresholds do not improve FER and can introduce saturation issues in intermediate operations.","Decoder selection must consider numerical precision jointly with the decoding rule, since MSA is markedly more sensitive to quantization than SPA.","The ranking SPA > NMS > MSA under fixed-point constraints provides a concrete baseline for FPGA-oriented LDPC decoder design in the negative-SNR regime."],"fun_headline_variants":["SPA at 8-bit fixed point beats min-sum decoders for CV-QKD","Q8.4 SPA delivers best reliability-cost balance for CV-QKD","Low-precision SPA leads fixed-point LDPC comparison for CV-QKD","NMS at Q16.8 emerges as simpler decoder fallback for CV-QKD"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the paper's fixed-point model—round-to-nearest quantization, Lmax=8.0 clipping, and the implicit fixed-point implementation of SPA's tanh/atanh update—faithfully represents the arithmetic that a real FPGA decoder would perform.","fun_headline_variants_meta":{"raw":{"variants":["SPA at 8-bit fixed point beats min-sum decoders for CV-QKD","Q8.4 SPA delivers best reliability-cost balance for CV-QKD","Low-precision SPA leads fixed-point LDPC comparison for CV-QKD","NMS at Q16.8 emerges as simpler decoder fallback for CV-QKD"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1135,"prompt_tokens":631,"completion_tokens":504,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":375,"completion_tokens_details":{"reasoning_tokens":416}},"tokens_in":375,"tokens_out":504,"duration_ms":4805,"temperature":1.0,"reasoning_tokens":416,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T16:30:01.905569+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Implement an SPA decoder on an FPGA using a concrete tanh/atanh lookup table with Q8.4 message storage and the same MET-LDPC code at SNR around −17 dB; if the measured FER degrades noticeably relative to the paper's simulated Q8.4 curve, or if a different tanh approximation changes the decoder ranking, the central recommendation would not carry over to hardware.","supporting_citations":[],"review_version":1}