{"id":"c6b4bec8-56d0-4cf6-abc6-0058245e44ab","arxiv_id":"1908.04205","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Probabilistic amplitude shaping with product codes and hard-decision decoding achieves up to 2.7 dB gain and 1 bpcu spectral efficiency improvement over uniform signaling, with component-code parameters constrained by integrality conditions.","lead":"This paper applies probabilistic amplitude shaping to product codes with hard-decision decoding for fiber-optic systems, deriving design constraints for the component codes. Simulations on 256-QAM show up to 2.7 dB gain and 1 bit/channel use spectral efficiency improvement over uniform signaling.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 2.7 dB and 1 bpcu headline may be inflated by an unmatched comparison: Fig. 4 compares under-sampled PAS and uniform-PC envelopes without rate-matching the uniform baseline to each PAS SE.","rationale":"The reader identified the AWGN/GN-model assumption and lack of simulation artifacts as the weakest points. The AWGN assumption is explicitly stated and standard in this literature, so I do not treat it as the most load-bearing issue. The more specific risk to the central quantitative claim is that the 2.7 dB and 1 bpcu gains may be artifacts of the comparison protocol in Fig. 4: the uniform baseline is not rate-matched to the PAS operating points, only a small subset of possible uniform PC rates is simulated, and the maximum horizontal and vertical gaps of two under-sampled Pareto fronts need not occur at matched SE or SNR. This is a testable, concrete concern: a matched-SE baseline could be constructed by searching over all shortenings for the uniform PC, and the claimed gains re-measured. If the gains survive that check, the central claim is credible; if they shrink, the paper's headline overstates the benefit of PAS. Independently, the absence of code and an unspecified random interleaver make it impossible to verify the exact reported points, but that is a reproducibility problem rather than a demonstrated flaw. The reader's conditional verdict remains appropriate: the paper should be accepted only after the matched-SE baseline comparison is provided and the simulation details are released.","tokens_in":32,"tokens_out":30389,"duration_ms":453746,"concrete_test":"Recompute Fig. 4 with a rate-matched uniform baseline: for each PAS operating point (s, SNR, SE), select the uniform-signaling PC with the value of s (allowing any admissible shortening, not just the eight listed) whose uniform SE = 8R is closest to the PAS SE at that point, and record the SNR gap; conversely, at each SNR, record the SE gap to the best uniform code. If the maximum SNR gain drops materially below the reported 2.7 dB (e.g., by more than 0.5 dB) or the maximum SE gain drops below 1 bpcu, the headline claim is not supported as stated. The authors should also release the simulation code and the random-interleaver seed used for the reported points so the matched-SE comparison can be reproduced.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim rests on Fig. 4's comparison of PAS+PC+iBDD with uniform signaling. For a fixed component code (fixed s), PAS has strictly lower SE than uniform (SE_PAS = 2H(A)+2γ < 2(m-1)+2γ = SE_uniform, since H(A)<m-1), so the 2.7 dB SNR gain and the 1 bpcu SE gain must come from comparing PAS at one code rate to uniform at a different code rate. The paper does not specify how the uniform baseline is rate-matched to each PAS operating point, and it uses only eight values of s for the uniform front even though v=10, t=3 admits many more possible shortenings for uniform PCs. With two under-sampled Pareto fronts, the maximum horizontal gap can occur at a SE where no uniform code was simulated (or even exists), and the maximum vertical gap can occur at a different SNR. The reported maxima are therefore not established as gains at matched SE/SNR; they may be inflated by code-rate granularity and by comparing against a non-optimized baseline. This is not an accusation of incorrect simulation, but the comparison protocol must be explicit and fair before the headline '2.7 dB / 1 bpcu' can be taken at face value.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper applies probabilistic amplitude shaping (PAS) to product codes (PCs) with hard-decision decoding. In Section III, the authors derive integer-feasibility conditions (Eqs. (5) to (8)) linking the PC component-code shortening parameter s to the PAS tuning parameter gamma, and tabulate feasible parameters for v=10, t=3 in Table I. They propose a random interleaver to spread bit-levels across PC component codewords and apply both iterative bounded-distance decoding (iBDD) and iBDD with combined reliability (iBDD-CR). Section IV reports AWGN simulations at block-error probability 10^-3 for 256-QAM, comparing PAS with PCs to uniform signaling and to staircase codes (SCCs). The headline claims are gains up to 2.7 dB and spectral efficiency improvement up to about 1 bit/channel use over uniform signaling, plus up to 0.3 dB from iBDD-CR and a statement that PAS with PCs and iBDD-CR closes the gap to PAS with SCCs and iBDD.","tokens_in":10541,"tokens_out":14607,"duration_ms":139909,"significance":"The necessary-condition derivation is self-contained and does not rely on curve fitting; it gives a simple design rule that should transfer to other component-code parameters. The observation that PCs offer 205 feasible rate points versus 40 for SCCs is a concrete practical advantage. Extending iBDD-CR to PAS is a useful step, and the authors are transparent that their operating points are optimized for the Hamming-metric AIR, so the iBDD-CR gains are conservative. If the headline gains survive a rate-matched comparison and the simulation details are made reproducible, this would be a solid contribution to high-throughput coded modulation with hard-decision decoding. In its current form, the numerical claims are not yet independently verifiable from the manuscript.","major_comments":[{"comment":"The headline claim of up to 2.7 dB gain and 1 bpcu SE improvement is not supported by a clearly defined comparison. The text does not state how the uniform-signaling baseline is matched to each PAS operating point in terms of spectral efficiency or code rate. For a fixed component code (fixed s), uniform signaling has SE 2(m-1+gamma), while PAS has SE 2H(A)+2gamma with H(A)<m-1, so the uniform baseline has higher SE at the same s; any gain at fixed SE must come from comparing PAS at one code rate with uniform at a different code rate. The uniform envelope in Fig. 4 is built from only eight shortening values (s=77, 289, 447, 605, 661, 727, 759, 797), although uniform PCs allow many more shortenings (they are not restricted by Eq. (8)). The paper also does not specify how the two envelopes are interpolated between the simulated points. As a result, the reported maxima may occur at a SE or SNR where no uniform code was simulated, and the 2.7 dB / 1 bpcu figures may be inflated by code-rate granularity rather than by shaping. Please define the comparison protocol precisely, e.g., rate-match the uniform code to each PAS operating point (nearest feasible rates above and below), and report gains for matched pairs; or explicitly state the envelope interpolation and the sensitivity of the maxima to the baseline code set.","section":"Section IV, Fig. 4"},{"comment":"The numerical results cannot be independently verified from the manuscript. The random interleaver that distributes the bit-levels among PC component codewords is described only as 'a random interleaver' with no construction, seed, or permutation length; since the authors argue this interleaver is necessary for symmetric PCs, the simulation outcomes depend on an unspecified implementation. In addition, no code, data, or error bars are provided, and the number of simulated blocks needed to estimate the 10^-3 block-error probability is not stated. Please specify the interleaver (e.g., a fixed permutation generated by a described pseudo-random process), the number of trials, and the uncertainty in the reported operating SNRs, or make the simulation code/data available.","section":"Sections III and IV"}],"minor_comments":[{"comment":"The caption lists PC shortening values '77, 266, 447, 535, 605'; elsewhere in the text and Fig. 3 the second value is 261. Please correct the typo.","section":"Section IV, Fig. 5 caption"},{"comment":"The SCC shortening parameters are inconsistent: the text gives '63, 274, 431, 519, 591' in one place, '63, 271, 431, 591, 647, 711, 743, 783' in another, and the Fig. 5 caption gives '63, 247, 431, 519, 591'. Please unify these lists.","section":"Section IV"},{"comment":"In the description of the PAS encoder, 'The vector u is parsed to u^{gamma n} and u^k' should read 'parsed into'.","section":"Section II"},{"comment":"The sets S_l^0 and S_l^1 are described as 'sets of size 2^m ASK symbols'; since the two sets partition the 2^m-ary ASK alphabet by one bit-level, each should have size 2^{m-1}. Please check the typesetting.","section":"Eq. (9)"},{"comment":"The phrases 'up to 2.7 dB' and 'up to 1 bpcu' are not qualified in the abstract/conclusion; Section IV notes that only 5 of 205 feasible operating points were simulated in Fig. 3 and 8 in Fig. 4. Please state explicitly that these maxima are over the tested subset of component codes.","section":"Abstract and Conclusion"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Firas,\n\nQuick read of the Sheikh et al. paper. The core contribution is straightforward: take the PAS scheme the group applied to staircase codes in [18], carry it over to product codes, and derive the necessary integrality conditions (Eqs. 5–8) for the BCH component code parameters. That derivation is clean and correct, and the observation that PCs give much finer SE granularity than SCCs (205 feasible s values versus 40) is a genuine practical point. The random interleaver for spreading bit levels across row/column codes is a sensible fix, and the iBDD-CR comparison is a nice extra. The paper is honestly written: the AWGN modeling assumption is flagged, and the conclusion admits that the AIR optimization for hybrid decoding is left open.\n\nWhere I part ways with the abstract is the 'up to 2.7 dB and 1 bpcu' headline. Those numbers come from Fig. 4, and the comparison protocol there is not explicit. The text says the PCs are chosen to have roughly the same code rate as the SCCs in [18], but it never says how the uniform baseline is rate-matched to each PAS operating point. Since PAS reduces SE relative to uniform signaling at the same code rate (H(A) < m-1), comparing PAS and uniform at the same s means comparing different SEs. To claim a 2.7 dB gain, you need to show the uniform PC at the same SE as the PAS point. The figure uses only eight s values for the uniform front, though the code parameters admit many more, so the envelope is coarse. The maximum horizontal gap could easily be an artifact of an under-sampled uniform curve. This is not an accusation of cooking the numbers, but the paper needs to pin down the comparison before those headline numbers are cited.\n\nThe other weakness is reproducibility: no code, no data, no error bars, and the random interleaver is only described as 'random', which doesn't allow someone to re-run the simulation. For a paper whose main evidence is simulation, that's a real gap.\n\nWho should read this: people working on shaping with hard-decision BCH codes for optical transceivers. They will get the feasibility conditions and the PC vs SCC trade-off. It deserves peer review, but I would ask for a revision that spells out the baseline selection and ideally provides artifacts.","headline":"Solid incremental extension of PAS to product codes with a correct feasibility derivation, but the headline 2.7 dB/1 bpcu gains are not rigorously established because the uniform baseline comparison is under-specified and likely not rate-matched.","tokens_in":11054,"tokens_out":6982,"would_cite":true,"duration_ms":67656,"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 shows that probabilistic amplitude shaping can be combined with hard-decision-decoded product codes to gain up to 2.7 dB and one bit per channel use over uniform signaling in high-throughput fiber-optic systems.","keywords":["probabilistic amplitude shaping","product codes","hard-decision decoding","iterative bounded distance decoding","fiber-optic communications","spectral efficiency","coded modulation","staircase codes"],"falsifier":"Run the PAS-PC scheme with the tabulated parameters on a fiber testbed or split-step simulation with realistic launch powers; if the measured SNR gain over uniform signaling is substantially below 2.7 dB at $P_e=10^{-3}$, the AWGN modeling assumption is the point of failure. A second check: pick a shortened BCH component code with a shortening value $s$ that violates the integer-feasibility condition (6)-(8); if it still yields the designed spectral efficiency and target error rate, the paper's necessary conditions are not necessary.","tokens_in":10092,"feed_emoji":"📶","tokens_out":6984,"duration_ms":60913,"temperature":0.7,"pith_summary":"Probabilistic amplitude shaping (PAS), which redistributes the probabilities of constellation points to approach the Shannon limit, has mostly been paired with soft-decision LDPC codes. This paper shows that PAS also works with product codes decoded by hard decisions, the low-complexity decoding choice for high-throughput fiber transponders, provided the component-code parameters obey a rate-matching feasibility condition and the amplitude and sign bits are spread evenly across the code array by a random interleaver. For 256-QAM over AWGN, the resulting scheme beats uniform signaling by up to 2.7 dB in signal-to-noise ratio and adds up to 1 bit per channel use of spectral efficiency at a block error rate of $10^{-3}$. The paper also demonstrates that a recently introduced decoder, iterative bounded-distance decoding with combined reliability (iBDD-CR), adds another 0.3 dB and lets product codes match or outperform staircase codes of the same rate.","feed_headline":"Probabilistic shaping adds 2.7 dB to fiber product codes","feed_subtitle":"A rate-matching rule plus a random interleaver make shaping compatible with hard-decision decoding.","key_machinery":"The central object is the PAS-PC encoding chain: a constant-composition distribution matcher (CCDM) generates shaped amplitudes; a random interleaver distributes the amplitude bits and the uniform sign/parity bits equally across all rows of the systematic product code array; and the systematically encoded parity bits are mapped to sign bits. The load-bearing identity is the rate balance $R = \\tilde{k}^2/\\tilde{n}^2 = (m-1+\\gamma)/m$, which together with $n = \\tilde{k}^2/(m-1+\\gamma)$ converts the component-code shortening parameter $s$ into a discrete set of feasible spectral efficiencies $\\gamma$; for $(v,t)=(10,3)$ and 16-ASK the paper finds 205 such values, versus 40 for staircase codes. The improvement mechanism in the second result is iBDD-CR decoding, which propagates channel-reliability information between row and column decoders.","core_discovery":"The paper establishes that PAS is compatible with bit-wise hard-decision decoding of product codes, provided the component code parameters satisfy an integer-feasibility constraint derived from the PAS rate equation: for a PC with shortened BCH component code of length $\\tilde{n}$ and information length $\\tilde{k}$, the symbol block length $n = \\tilde{k}^2/(m-1+\\gamma)$ must be a positive integer and $\\gamma n$ a non-negative integer. Over 256-QAM with a Maxwell-Boltzmann amplitude distribution, this yields operating points whose achievable rate $2H(A)+2\\gamma$ approaches the shaped hard-decision information rate $R_{\\mathrm{HDD}}$, and the simulated operating points show up to 2.7 dB SNR gain and about 1 bpcu spectral efficiency gain over uniform signaling at $P_e=10^{-3}$. With the iBDD-CR decoder, the same PAS-PC scheme gains up to 0.3 dB over plain iBDD and matches or beats staircase codes of equal rate at several shortening parameters.","pith_inferences":["The same rate-balance argument should transfer to other product-like code families (e.g., braided BCH or half-product codes), so the feasibility condition may serve as a design rule beyond PCs and SCCs.","Since the shaping parameter $\\lambda$ is optimized for the hard-decision achievable rate rather than for the hybrid decoder, the 0.3 dB iBDD-CR gain is likely conservative; re-optimizing for the hybrid metric could move the operating points further.","A hardware-oriented prediction: a deterministic interleaver that spreads bit levels uniformly, rather than a random permutation, should achieve the same shaping gain while saving memory, a testable implementation choice.","The comparison is made at $P_e=10^{-3}$; if deep-space or long-haul submarine links require $P_e$ near $10^{-9}$, the error floor of the shaped PC should be assessed before extrapolating the 2.7 dB gain."],"forward_implications":["PAS can be added to existing hard-decision product-code transceivers without changing the decoder, making shaping available to very high-throughput and low-latency fiber links.","The feasibility condition defines the achievable spectral efficiency grid for a given component-code family; product codes offer 205 rates versus 40 for staircase codes at $(v,t)=(10,3)$, enabling finer rate adaptation.","At a block error rate of $10^{-3}$, shaped PCs outperform uniform PCs by up to 2.7 dB and 1 bit/channel use, with the gain increasing at lower code rates.","Switching the decoder to iBDD-CR recovers up to 0.3 dB for the same PAS-PC setup and closes the performance gap to staircase codes at several rates.","The optimal operating point for each component code sits at the crossing of the achievable-rate curve $2H(A)+2\\gamma$ with the shaped hard-decision information rate; at that point the simulated PC back-off is around 1.6 dB at $P_e=10^{-3}$."],"supporting_citations":[{"why":"Supplies the PAS architecture and the rate-balance relation that the feasibility condition is built on.","marker":"[16]"},{"why":"Introduces the PAS-with-HDD design for staircase codes that this paper adapts to product codes and compares against.","marker":"[18]"},{"why":"Defines the iBDD-CR decoding algorithm used to obtain the improved operating points.","marker":"[6]"},{"why":"Provides the constant-composition distribution matching used to generate shaped amplitudes.","marker":"[23]"},{"why":"Derives the hard-decision achievable rate $R_{\\mathrm{HDD}}$ used to optimize the shaping parameter.","marker":"[22]"},{"why":"Establishes the Gaussian noise model that justifies treating the fiber channel as AWGN.","marker":"[21]"}],"fun_headline_variants":["PAS gives product codes 2.7 dB gain in fiber systems","Shaping adds 2.7 dB to product codes with hard decoding","2.7 dB gain: probabilistic shaping meets product codes","Product codes + PAS: 2.7 dB and 1 bpcu gains over uniform","Hard-decision product codes gain 2.7 dB with PAS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The fiber-optic channel is modeled as AWGN through the Gaussian noise model, so the 2.7 dB gain is computed for an idealized channel and may shrink under nonlinear fiber effects.","fun_headline_variants_meta":{"raw":{"variants":["PAS gives product codes 2.7 dB gain in fiber systems","Shaping adds 2.7 dB to product codes with hard decoding","2.7 dB gain: probabilistic shaping meets product codes","Product codes + PAS: 2.7 dB and 1 bpcu gains over uniform","Hard-decision product codes gain 2.7 dB with PAS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000683,"raw_usage":{"total_tokens":3063,"prompt_tokens":869,"completion_tokens":2194,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":2098}},"tokens_in":485,"tokens_out":2194,"duration_ms":15514,"temperature":1.0,"reasoning_tokens":2098,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:12:04.296698+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the PAS-PC scheme with the tabulated parameters on a fiber testbed or split-step simulation with realistic launch powers; if the measured SNR gain over uniform signaling is substantially below 2.7 dB at $P_e=10^{-3}$, the AWGN modeling assumption is the point of failure. A second check: pick a shortened BCH component code with a shortening value $s$ that violates the integer-feasibility condition (6)-(8); if it still yields the designed spectral efficiency and target error rate, the paper's necessary conditions are not necessary.","supporting_citations":[{"cited_title":"Bandwidth efficient and rate-matched low-density parity-check coded modulation,","cited_arxiv_id":null,"evidence_quote":"Supplies the PAS architecture and the rate-balance relation that the feasibility condition is built on."},{"cited_title":"Probabilistic amplitude shaping with hard decision decoding and staircase codes,","cited_arxiv_id":null,"evidence_quote":"Introduces the PAS-with-HDD design for staircase codes that this paper adapts to product codes and compares against."},{"cited_title":"Refined reliability combining for binary message passing decoding of product codes,","cited_arxiv_id":null,"evidence_quote":"Defines the iBDD-CR decoding algorithm used to obtain the improved operating points."},{"cited_title":"Constant composition distribution match- ing,","cited_arxiv_id":null,"evidence_quote":"Provides the constant-composition distribution matching used to generate shaped amplitudes."},{"cited_title":"Achievable rates for probabilistic shaping,","cited_arxiv_id":null,"evidence_quote":"Derives the hard-decision achievable rate $R_{\\mathrm{HDD}}$ used to optimize the shaping parameter."},{"cited_title":"A simple and effective closed-form GN model correction formula accounting for signal non-gaussian distribution,","cited_arxiv_id":null,"evidence_quote":"Establishes the Gaussian noise model that justifies treating the fiber channel as AWGN."}],"review_version":1}