{"id":"7ab49199-fac4-4e86-bc83-54a6027d8e5e","arxiv_id":"2411.16952","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A Gaussian proposal built from the linearized KdV model makes rejection sampling of the truncated KdV Gibbs measure 1 to 6 orders of magnitude more efficient than uniform sampling.","lead":"This paper offers a faster way to generate random wave fields from the truncated KdV model, a statistical model used to study rogue waves. The new sampling scheme accepts far more candidate waves than the standard approach and runs in parallel, making extreme-wave searches cheaper.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Rejection constant M (Eq. 45) is found by Nelder-Mead from a Dirichlet-kernel start without a global-optimality check; if M is underestimated the sampler is biased and the reported 1-6 orders-of-magnitude improvement is inflated.","rationale":"The reader's weakest assumption is precisely the unverified global maximum of f/g (Eq. 45). This is the most load-bearing point because the sampler's exactness and the quantitative improvement claim both fail if M is too small. Our K=2 calculation shows the Dirichlet-kernel heuristic is not an H3 maximizer, casting doubt on the claim that the Nelder-Mead start is near the global max. A multi-start or large-sample check of f/(M g) would settle whether M is certified. Since the paper otherwise provides a coherent derivation and plausible numerical support, the appropriate verdict remains CONDITIONAL as the reader stated; no change is needed.","tokens_in":53,"tokens_out":14391,"duration_ms":253644,"concrete_test":"For a representative physical regime (e.g., K=16, β'=40, C3/C2=120), run a multi-start optimization of Eq. (40): use at least 100 random initial points on S^{2K-1} in addition to the Dirichlet kernel, and record the largest value found; also track the maximum observed f/(M g) over 10^8 rejection proposals. If any observed ratio exceeds 1, or the multi-start maximum exceeds the published M, the rejection constant is underestimated and the acceptance-rate improvements in Tables I and II must be recomputed with a certified M.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—exact rejection sampling with 1–6 orders of magnitude acceptance-rate improvement—depends on the rejection constant M = max_{x̂} f(x̂)/g(x̂) in Eq. (45). The paper computes M by Nelder-Mead initialized at the Dirichlet kernel (Eq. 47) and reports convergence in a few iterations, but no proof or numerical certificate shows this is the global maximum. If M is underestimated, the acceptance probability f/(M g) can exceed 1 for some states; an implementation that accepts with probability min(1, f/(M g)) then accepts those states with probability 1, biasing the sample away from the target and inflating the measured acceptance rates and hence the improvement factors in Tables I and II. The concern is not merely formal: for K=2, Appendix A gives H3 = 2π R2 R1^2 cos(2θ1−θ2), and with the energy constraint R1^2+R2^2 = E0/(2π), the H3 maximizer satisfies R2/R1 = 1/√2, not the equal-amplitude Dirichlet state (R2/R1 = 1). Since f/g ≈ exp(β C3 H3) for strong nonlinearity (Eq. 46), the Dirichlet kernel can lie away from the true maximizer, and a single Nelder-Mead run is not a trustworthy global optimization in dimension 2K−1 (up to 63 for K=32). The absence of any validation that max f/(M g) ≤ 1 on a large sample of proposals leaves the central claim unprotected.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a rejection sampling algorithm for the mixed canonical-microcanonical Gibbs measure of the truncated KdV equation. The proposal is an anisotropic Gaussian distribution projected onto the energy sphere, originally constructed by Sun and Moore (2023) to approximate the linear TKdV measure. The authors derive a closed-form expression for the proposal density g (Eq. 38), define a rejection sampler with constant M (Eq. 45), and report acceptance rates and improvements over uniform proposals for K=16 and K=32, parallel speedup, and examples of extreme wave fields. The central claim is that within experimentally relevant parameter regimes the proposal yields 1--6 orders of magnitude more accepted samples than the uniform proposal.","tokens_in":17685,"tokens_out":6517,"duration_ms":61609,"significance":"If the algorithm is correct as stated, it is a valuable contribution: it produces independent samples from the target measure, is trivially parallelizable, and avoids MCMC tuning and burn-in. The use of a self-consistently chosen alpha* (Eq. 29) and the closed-form proposal density are strong points, and the acceptance-rate comparison to a uniform proposal is a fresh numerical contribution. The paper is honest about the heuristic nature of the optimization step, but that heuristic is load-bearing because the exactness of rejection sampling depends on a valid upper bound M.","major_comments":[{"comment":"The rejection constant M is found by Nelder-Mead from a Dirichlet-kernel initial guess, with no global optimality certificate. This is load-bearing: if M is underestimated, the acceptance probability f/(Mg) exceeds 1 for some states; accepting those with probability 1 biases the sampler and inflates the reported acceptance rates. The concern is not purely formal: for K=2, Appendix A gives H3 = 2π R2 R1^2 cos(2θ1−θ2), whose maximizer under the energy constraint has R2/R1 = 1/√2, while Eq. (47) has equal amplitudes R2/R1 = 1. Thus the initial guess is not the maximizer of the dominant factor in Eq. (46), even in the simplest case, and a single Nelder-Mead run in dimensions up to 63 does not establish the required upper bound. Please add a numerical certificate (e.g., multi-start or global optimization, or verification that max f/(Mg) ≤ 1 over a very large proposal sample) or a rigorous upper bound for Eq. (40).","section":"IV.A, Eqs. (44)-(45) and (47); Appendix A"},{"comment":"The acceptance-rate measurements lack error bars, and the extreme entries are based on very few accepted samples. For example, Table II reports acceptance rates as low as 2.1×10^{-9}, and the text states that for the asterisked cases the naive version could not produce 100 samples over weeks; the corresponding improvement factors are thus point estimates with unknown relative error. Since the central quantitative claim is the improvement factor (1–6 orders of magnitude), please report confidence intervals or bootstrap estimates, or at least the number of accepted samples underlying each entry. This is necessary for the reader to judge whether the 'orders of magnitude' statement holds in the extreme nonlinear regime.","section":"V.B, Tables I and II"},{"comment":"The paper asserts that the algorithm generates independent samples from the target measure, but the only evidence is the internal acceptance-rate computation; there is no independent check of the output distribution. A direct comparison of sampled moments (e.g., skewness and power spectra) against a long MCMC or SIR run in a few moderate cases would provide a strong safeguard against undetected bias caused by an invalid M. Without such a check, the exactness claim rests entirely on the unverified upper bound in Eq. (45).","section":"V.B, Algorithm 1 and Section IV.B"}],"minor_comments":[{"comment":"There is a grammatical error: 'include better parallelization properties' should be 'including better parallelization properties.'","section":"Abstract and Introduction"},{"comment":"In the Laplace transform formula L[t^{K-1}] = ∫ exp(-st) t^{K-1} dt, the integration variable is t (or u after the change of variables), not du; the displayed notation is inconsistent and should be corrected.","section":"III.B, Eq. (36)"},{"comment":"The notation Const is used repeatedly for different constants in Eqs. (27), (34), (37), (38), and (46). This is acceptable informally, but the authors should add a sentence reminding the reader that these constants are not equal, to avoid confusion in the ratio f/g derivation.","section":"II.C and III.A"},{"comment":"The parallel speedup is demonstrated on a single hardware platform (Apple M2 Ultra) with no details on how the timing was measured or whether the speedup curve is averaged over repeated runs. A brief description of the measurement procedure and variance would strengthen this section.","section":"V.C"},{"comment":"No code or data are provided. Since the rejection constant M is computed by a numerical optimization and the reported acceptance rates are stochastic estimates, supplying code or pseudorandom seeds would greatly improve reproducibility.","section":"Reproducibility"}],"recommendation":"major_revision","confidential_remarks":"The paper is a promising numerical methods contribution, and the central claim is likely robust in broad terms. However, the heuristic computation of M is the mathematical hinge of the exact sampler, and the current manuscript does not protect this point. I would be willing to accept after the authors add a verification step for the upper bound M, or an independent validation of the sampler's output distribution, together with uncertainty quantification for the reported acceptance rates. Encouraging code/data deposition would also be helpful."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe genuinely new piece here is not the proposal distribution—that's from Sun & Moore (2023)—but the decision to use it in rejection sampling and the measured acceptance-rate gains. That's a legitimate extension, and the paper is honest about the provenance. The numerical evidence, while not fully reproducible, is suggestive: acceptance rates above 90% in the linear case and improvement factors of 1–6 orders of magnitude in the physically interesting nonlinear regime. The exact K=2 formulas in Appendix A are a nice check on the H2/H3 computation, and the parallel speedup plot is a sensible demonstration.\n\nThe soft spot is the rejection constant M. The sampler is only exact if M >= max f/g. The paper computes M with Nelder-Mead from a Dirichlet-kernel start and reports convergence in a few iterations. That is not a global optimization certificate, and the K=2 example from the stress test shows the Dirichlet state is not the H3 maximizer once nonlinearity is strong. If M is underestimated, the acceptance probability can exceed one, the sampler becomes biased, and the measured improvement factors are inflated. The authors never check max f/(Mg) <= 1 on a large sample of proposals. This is a real gap, but I wouldn't call it fatal to the qualitative claim—the gains are so large that even a somewhat off M would leave the main conclusion intact. It does, however, undo the 'exact' label until fixed.\n\nThe other weaknesses are more minor. No code or data is shipped, so the numbers in Tables I and II can't be independently reproduced. The asterisked rows in Table II estimate naive acceptance rates from fewer than 100 samples, and there are no error bars anywhere. The paper claims advantages over MCMC but never actually benchmarks against an MCMC sampler, so that comparative claim is not demonstrated.\n\nOverall, this is solid, honest computational work with one load-bearing gap. It deserves peer review—a referee should push for a proper check of M (e.g., a dense random search or a deterministic bound) and for released code. With those, this would be a genuinely useful paper for anyone sampling TKdV ensembles for rogue-wave statistics.\n\nRecommendation: send to review, conditional on the authors addressing the M certification.","headline":"A useful rejection-sampling trick for the TKdV Gibbs measure with credible large speedups, but the 'exact' claim rests on an unverified global maximum.","tokens_in":18132,"tokens_out":2502,"would_cite":true,"duration_ms":25417,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65C05","65Y05","76B15","82B31"],"pacs":[],"model":"deepseek-v4-flash","headline":"A paper introduces a proposal distribution that makes rejection sampling of the truncated KdV Gibbs measure up to six orders of magnitude more efficient and nearly perfectly parallel.","keywords":["truncated KdV","Gibbs measure","rejection sampling","proposal distribution","rogue waves","anisotropic Gaussian","parallel sampling","extreme wave statistics"],"falsifier":"For the parameters in Tables I and II, scan the sphere densely (millions of independent draws from the proposal plus a multi-start global optimizer) and check whether any state yields $f(\\hat x)/g(\\hat x)$ greater than the reported $M$; if such a state exists, the acceptance probability would exceed 1 there and the sampler is biased. A cheaper check is to compare ensemble statistics (mean power spectrum, skewness) against a long independent MCMC run: disagreement beyond Monte Carlo error would indicate a misspecified rejection constant.","tokens_in":17136,"feed_emoji":"🌊","tokens_out":7640,"duration_ms":63938,"temperature":0.7,"pith_summary":"The paper aims to make rejection sampling of the truncated KdV Gibbs measure practical in the parameter regime where anomalous, extreme waves occur. It replaces the spectrally uniform proposal with an anisotropic Gaussian distribution derived from the linearized system, which is nearly as easy to sample but much closer to the target measure. The paper reports that this proposal yields 1 to 6 orders of magnitude more accepted samples than the naive uniform proposal, with acceptance rates above 80% in the linear case and improvements as large as 2.5 million-fold in the strongest tested case. If correct, the sampler gives independent, uncorrelated wave-field samples at high throughput and with near-optimal parallel speedup, making it a practical tool for studying rogue-wave statistics.","feed_headline":"Rogue-wave sampling gains up to six orders of magnitude","feed_subtitle":"A new proposal distribution makes rejection sampling of the TKdV Gibbs measure practical and fully parallel.","key_machinery":"The machinery is the proposal density $g$ on the unit sphere $S^{2K-1}$, obtained by projecting an anisotropic Gaussian with variances $\\sigma_k^2 = 1/(1+\\alpha^*\\beta' k^2/K^3)$, where the self-consistent parameter $\\alpha^*$ is the unique root of $F(\\alpha) = 1 - (\\alpha/K) \\sum_{k=1}^K 1/(1+\\alpha\\beta' k^2/K^3)$. Drawing from $g$ is one Gaussian draw plus a radial projection, and the acceptance step uses the closed-form ratio $f(\\hat x)/g(\\hat x) = \\exp(-\\beta H_K[\\hat U(\\hat x)]) \\left(1 + (\\alpha^* \\beta'/K^3) \\sum_k k^2(\\hat x_k^2 + \\hat x_{K+k}^2)\\right)^K$. $H_2$ and $H_3$ are evaluated in $O(K^2)$ time, with $H_3$ computed by a double-summation rearrangement of the triple-sum formula, and the rejection constant $M$ is chosen by Nelder-Mead maximization of the ratio, initialized at the Dirichlet kernel.","core_discovery":"The central claim is that the Gibbs density $f$ of the TKdV system, supported on the energy sphere, can be sampled exactly by rejection when the proposal is the density $g$ induced by an anisotropic Gaussian with per-mode variances $\\sigma_k^2 = 1/(1+\\alpha^* \\beta' k^2/K^3)$. Because $g$ was constructed to match the linearized ($C_3=0$) Gibbs measure in the large-$K$ limit, the ratio $f/g$ stays close to constant, and the acceptance probability $f/(M g)$ is high; numerical tests show acceptance rates above 80% for linear TKdV and improvements over the naive uniform proposal of 1--6 orders of magnitude in the skewness-producing regimes where extreme waves appear. The sampler also reproduces the expected physics: linear cases give symmetric, nearly Gaussian surface displacements, while increasing the nonlinearity ratio $C_3/C_2$ produces positive skewness, and 1-in-500 sampled wave fields exceed the $4\\sigma$ rogue-wave threshold by up to 24%.","pith_inferences":["A natural extension is to add a cubic tilt to the proposal to match the skewness contribution of $H_3$, which would raise acceptance rates in the high-nonlinearity regime where the current method degrades.","The non-monotone improvement factors seen at $K=32$ suggest the ratio $f/g$ has exploitable structure, so a tighter analytic bound on $M$ may be derivable.","The sampler's independence makes it a clean generator of training data for machine-learning models that flag pre-rogue states, a direction the paper lists as future work.","Because $g$ is derived from the linear system but works beyond it, the approach may transfer to other Hamiltonian truncations where a quadratic energy dominates the Gibbs weight."],"forward_implications":["For linear TKdV the acceptance rate exceeds 80%, so the sampler is nearly free in that regime.","In the nonlinear regimes where 1-in-500 events exceed the $4\\sigma$ threshold, the improved proposal still beats the uniform one by 1--3 orders of magnitude, making extreme-event searches feasible.","Because accepted samples are independent, the algorithm parallelizes with essentially no communication, achieving 80% of optimal speedup on 16 cores.","The method yields exact samples from the Gibbs measure, so histograms, spectra, and extremal statistics computed from them are unbiased.","The same proposal can be used in sampling-importance resampling, with no need for the global rejection constant."],"supporting_citations":[{"why":"Derives the anisotropic Gaussian proposal and its convergence to the linear TKdV Gibbs measure, which the sampler is built on.","marker":"[28]"},{"why":"Introduced the canonical-microcanonical TKdV Gibbs measure and the laboratory validation of anomalous wave statistics that motivates the sampling problem.","marker":"[16]"},{"why":"Established the statistical matching condition that sets the downstream inverse temperature, the context the sampler assumes.","marker":"[11]"},{"why":"Provides the triple-summation formula for $H_3$ that the paper rearranges into the $O(K^2)$ evaluation used in the acceptance ratio.","marker":"[35]"},{"why":"Supplies the standard rejection sampling framework the algorithm implements.","marker":"[36]"},{"why":"The Nelder-Mead optimizer used to compute the rejection constant $M$.","marker":"[39]"},{"why":"Defines the $4\\sigma$ threshold for rogue waves used to interpret the extreme sampled wave fields.","marker":"[7]"}],"fun_headline_variants":["Rogue-wave sampling acceptance up 6 orders of magnitude","New proposal distribution yields 1e6x more accepted samples","Parallel rejection sampler for rogue waves gets major boost","Extreme-wave simulation sees 6-order improvement via novel prior","Better sampling for KdV extreme events: 1e6-fold acceptance"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Nelder-Mead search starting from the Dirichlet kernel finds the true global maximum of the ratio $f/g$ on the sphere, since no proof or verification of global optimality is provided for the rejection constant $M$.","fun_headline_variants_meta":{"raw":{"variants":["Rogue-wave sampling acceptance up 6 orders of magnitude","New proposal distribution yields 1e6x more accepted samples","Parallel rejection sampler for rogue waves gets major boost","Extreme-wave simulation sees 6-order improvement via novel prior","Better sampling for KdV extreme events: 1e6-fold acceptance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000818,"raw_usage":{"total_tokens":3544,"prompt_tokens":866,"completion_tokens":2678,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":482,"completion_tokens_details":{"reasoning_tokens":2594}},"tokens_in":482,"tokens_out":2678,"duration_ms":19864,"temperature":1.0,"reasoning_tokens":2594,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:42:01.274182+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For the parameters in Tables I and II, scan the sphere densely (millions of independent draws from the proposal plus a multi-start global optimizer) and check whether any state yields $f(\\hat x)/g(\\hat x)$ greater than the reported $M$; if such a state exists, the acceptance probability would exceed 1 there and the sampler is biased. A cheaper check is to compare ensemble statistics (mean power spectrum, skewness) against a long independent MCMC run: disagreement beyond Monte Carlo error would indicate a misspecified rejection constant.","supporting_citations":[{"cited_title":"Lawrence, K","cited_arxiv_id":null,"evidence_quote":"Derives the anisotropic Gaussian proposal and its convergence to the linear TKdV Gibbs measure, which the sampler is built on."},{"cited_title":"Viotti and F","cited_arxiv_id":null,"evidence_quote":"Introduced the canonical-microcanonical TKdV Gibbs measure and the laboratory validation of anomalous wave statistics that motivates the sampling problem."},{"cited_title":"Dematteis, T","cited_arxiv_id":null,"evidence_quote":"Established the statistical matching condition that sets the downstream inverse temperature, the context the sampler assumes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the triple-summation formula for $H_3$ that the paper rearranges into the $O(K^2)$ evaluation used in the acceptance ratio."},{"cited_title":"Ross, A First Course in Probability: 8th edition (Pearson Education, 2010)","cited_arxiv_id":null,"evidence_quote":"Supplies the standard rejection sampling framework the algorithm implements."},{"cited_title":"We find that the numerical optimization typically converges in only a few iterations, confirming that Eq","cited_arxiv_id":null,"evidence_quote":"The Nelder-Mead optimizer used to compute the rejection constant $M$."},{"cited_title":"Haver, A possible freak wave event measured at the Draupner Jacket January 1 1995, in Rogue waves, Vol","cited_arxiv_id":null,"evidence_quote":"Defines the $4\\sigma$ threshold for rogue waves used to interpret the extreme sampled wave fields."}],"review_version":1}