{"id":"55325fa1-189f-4524-8002-e022fc4e8d91","arxiv_id":"2412.04206","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A pipeline of biased sampling, multi-histogram reweighting, and power-law extrapolation of the SCGF estimates infinite-size rate functions, with accuracy that breaks down near biased-ensemble phase transitions.","lead":"Rare-event simulations with biased sampling, histogram-free reweighting, and power-law finite-size extrapolation are combined to estimate limiting large-deviation rate functions, demonstrated on binomial data and Erdős-Rényi random graphs. The method works when the scaling assumption holds, and fails predictably near phase transitions in the biased ensemble, an important caution for users.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The parametric Legendre-Fenchel transform (Eq. 14) assumes μ(q)=Ψ′(q), but μ and Ψ are extrapolated independently via Eq. (13); with q-dependent B(q) the derivative of a power-law SCGF contains a logarithmic correction, so Eq. (14) can yield a curve that is not the true LF transform.","rationale":"The reader's weakest assumption is the validity of the power-law scaling in Eq. (13). That is a legitimate concern, and the paper itself demonstrates a breakdown at q≈0.317. However, the concern raised here is more fundamental: even when Eq. (13) holds exactly for Ψ, the independent power-law fit for μ is mathematically inconsistent with the derivative of Ψ unless B is q-independent. This affects the final Legendre-Fenchel step in every application, not just near phase transitions. The paper does not discuss this consistency requirement, and its error propagation treats μ and Ψ as uncorrelated, which masks the problem. The proposed synthetic test would settle whether the parametric transform is actually valid under the paper's own assumptions. Since the method could be repaired by performing the Legendre-Fenchel transform numerically on the extrapolated SCGF rather than using the parametric shortcut, the appropriate verdict remains conditional: the paper's demonstration is plausible, but the presented final step is not justified as written, and the provided error bars may be incomplete. This reinforces the reader's CONDITIONAL verdict without changing it.","tokens_in":13385,"tokens_out":6958,"duration_ms":73625,"concrete_test":"Construct a synthetic model with known finite-size SCGF Ψ(q;N)=C(q)+A(q)N^{-B(q)} where B(q) varies with q, e.g., B(q)=0.5+0.1q. Generate exact biased samples from the implied distribution for several N, then apply the full pipeline including separate power-law fits for μ and Ψ. Compare the output of Eq. (14) with the true Legendre-Fenchel transform of the extrapolated SCGF C_Ψ(q); if the two curves disagree beyond statistical error, the parametric shortcut is invalid and the method should instead use a numerical LF transform of the extrapolated SCGF.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (14) evaluates the rate function at s = μ(q) as qμ(q) − Ψ(q). This is only the correct Legendre-Fenchel transform if the extrapolated μ(q) equals the derivative of the extrapolated Ψ(q). In the paper, μ(q;N) and Ψ(q;N) are fitted separately to the same power-law form g(N;q)=C(q)+A(q)N^{-B(q)} (Eq. 13). If the true finite-size SCGF were exactly of this form with a q-dependent exponent B(q), then its q-derivative would contain the term −A(q)B′(q) ln N · N^{-B(q)}, which is not a pure power law. Thus the separate power-law fit for μ is misspecified unless B(q) is constant in q. Consequently, the extrapolated C(q) from the μ-fit will generally not equal d/dq of the extrapolated SCGF. In that case, Eq. (14) is not the supremum over q, and the resulting I(s) is not the convex envelope of the extrapolated SCGF. The paper explicitly ignores the correlation between μ and Ψ in its error propagation (Section II C), so this inconsistency is not reflected in the quoted error bars. This concern is independent of whether Eq. (13) is the correct scaling; it is an internal consistency issue in the method as presented. The binomial validation may not expose the problem because there B(q)≈γ is effectively constant, but for generic systems, especially the ER graphs, B(q) is expected to vary with q.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a numerical method to estimate infinite-size large-deviation rate functions from finite-size simulation data. The method combines biased importance sampling with multi-histogram (Ferrenberg–Swendsen) reweighting to compute the scaled cumulant generating function (SCGF) and its derivative, the tilted mean, as functions of the bias parameter q. These finite-size quantities are then extrapolated to the thermodynamic limit using a power-law fit (Eq. 13), and the rate function is obtained via a parametric Legendre–Fenchel transform (Eq. 14). The approach is demonstrated on a scale-dependent binomial model and on the largest connected component of Erdős–Rényi random graphs at connectivities c = 0.5 and c = 2.0, with comparisons to exact analytical rate functions. The paper also derives a detailed error-propagation scheme for the reweighted estimators in Appendix B.","tokens_in":13754,"tokens_out":5574,"duration_ms":53883,"significance":"If valid, the method gives a generally applicable, histogram-free route to limiting rate functions with propagated statistical errors, provided finite-size scaling of the form of Eq. (13) holds. The paper's strengths include the use of two nontrivial benchmark systems with known analytical solutions, the derivation of error-propagation formulas for the reweighted estimators, and the honest report of failures at biased-ensemble phase transitions and in non-convex regimes. The central methodological assumption—the power-law extrapolation—is clearly stated and its breakdown is demonstrated in one of the test cases, which is informative rather than hidden. However, the validity of the parametric Legendre–Fenchel step rests on an internal consistency condition between the separately extrapolated SCGF and tilted mean that is not checked or enforced in the paper.","major_comments":[{"comment":"The parametric Legendre–Fenchel transform in Eq. (14) is only equivalent to the true Legendre–Fenchel transform if the extrapolated μ(q) equals the derivative of the extrapolated Ψ(q). The paper fits μ(q;N) and Ψ(q;N) independently via Eq. (13). If the exponent B(q) is q-dependent, the derivative of a power-law SCGF contains a logarithmic correction term of the form −A(q)B′(q) ln N · N^{−B(q)}, which is not a pure power law. Hence the separate power-law fit for μ is misspecified unless B(q) is constant, and the extrapolated values will not generally satisfy the relation μ(q)=Ψ′(q). In that case, Eq. (14) does not compute the supremum in Eq. (6), and the resulting I(s) is not the convex envelope of the extrapolated SCGF. The paper explicitly ignores the correlation between μ and Ψ in the error propagation (Section II C), so the quoted error bars do not include this systematic inconsistency. The binomial example may not expose the problem because there B(q) is effectively constant, but for generic systems, especially the ER graphs, B(q) is expected to vary. The authors should either enforce consistency (e.g., derive μ from a single fit of Ψ or fit μ and Ψ jointly) or provide a quantitative argument that the inconsistency is negligible in the cases studied.","section":"Section II C, Eq. (14)"},{"comment":"The power-law scaling ansatz in Eq. (13) is adopted without derivation, and the paper itself shows a clear breakdown in the ER c=0.5 case around q≈0.317, where a biased-ensemble phase transition produces a double-peaked distribution (Fig. 5 inset) and makes the power-law fit unreliable (Fig. 6), leading to a gap in the estimated rate function (Fig. 7). No diagnostic is provided to detect such a breakdown when no analytical solution is available. This weakens the claim of 'rather general applicability' of the method: a practitioner applying it to a new system has no way to know whether the extrapolation is trustworthy. The authors should discuss concrete checks, such as comparing fits obtained from different subsets of system sizes, examining the residuals of the fit, or using a model-selection criterion to test the adequacy of the power-law form.","section":"Section III B, Eq. (13)"}],"minor_comments":[{"comment":"The phrase 'Ψ (s; N is readily obtained' should read 'Ψ(q;N) is readily obtained'.","section":"Section II C, after Eq. (12)"},{"comment":"The caption contains a typo: 'the larges connected component' should be 'the largest connected component'.","section":"Figure 5 caption"},{"comment":"The phrase 'a a rather large statistical error estimate' contains a doubled article; it should read 'a rather large statistical error estimate'.","section":"Section IV"},{"comment":"The condition 'with b(s) ≥ 0' is written on the same line as the equation; consider placing it on a separate line for clarity.","section":"Section II A, Eq. (4)"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about Eq. (14) is well-founded and, in my reading, a genuine load-bearing issue. The paper's validation on the binomial model does not rule it out because the scaling exponent there is effectively q-independent. The issue is fixable by revising the extrapolation procedure to enforce consistency between μ and Ψ or by clearly stating the constant-B assumption and justifying it for the targeted applications. The paper otherwise makes a useful contribution in combining multi-histogram reweighting with finite-size extrapolation and in providing explicit error-propagation formulas. I recommend major revision rather than rejection, since the core idea is sound and the present flaw appears repairable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a genuinely useful methods paper for the large-deviation simulation crowd. The new thing is the combination: you run importance sampling at a few biases, use histogram-free multi-histogram reweighting to get μ(q;N) and Ψ(q;N) at arbitrary q, then extrapolate each to N→∞ with a power-law fit, and then apply the parametric Legendre-Fenchel transform. The error-propagation formulas in Appendix B are a real addition — I haven't seen them in print either. The two testbeds (binomial, ER largest component) are sensible, and the paper is honest about where things break: the phase transition in the biased ER ensemble at c=0.5, and the non-convexity at c=2.0.\n\nThe soft spots are real but not fatal. The power-law ansatz is ad hoc, and there's no diagnostic for when it fails beyond 'the fit looks bad'. No code or data are provided, so reproducing the tests is harder than it should be. And there's an internal consistency issue that the paper doesn't address. Eq. (14) uses s = μ(q), but μ and Ψ are extrapolated separately via Eq. (13). If the true finite-size Ψ(q;N) were exactly C(q)+A(q)N^{-B(q)} with B(q) q-dependent, then Ψ'(q;N) contains a term proportional to ln N, which is not a power law. So fitting μ(q;N) to a pure power law is misspecified, and the extrapolated C_μ(q) will not generally equal dC_Ψ/dq. In that case Eq. (14) is not the true LF transform, even if each extrapolation looks fine individually. The binomial example wouldn't reveal this because B is effectively constant there. This is a theoretical gap that a referee should push on.\n\nOverall, it's a solid demonstration of a plausible pipeline, with clear limitations stated. The internal consistency question is the one thing I'd want resolved before relying on the method for a system without a known answer. It deserves a serious referee, and I'd cite it for the error formulas and the reweighting trick.","headline":"Useful methods paper for large-deviation rate functions, with honest validation, but the parametric LF transform has an internal consistency issue that needs attention.","tokens_in":14242,"tokens_out":3385,"would_cite":true,"duration_ms":34009,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["05.10.-a","02.50.-r"],"model":"deepseek-v4-flash","headline":"The paper shows that limiting large-deviation rate functions can be estimated without histograms and without committing to one sampling algorithm, by combining multi-histogram reweighting with power-law extrapolation in system size.","keywords":["large deviations","rate function","rare events","importance sampling","multi-histogram reweighting","scaled cumulant generating function","finite-size scaling","Erdős–Rényi random graphs"],"falsifier":"Run the identical pipeline on the binomial model with $r(N)=r_\\infty+c/\\ln N$, where the exact finite-size and limiting rate functions are known analytically and the finite-size correction is logarithmic rather than power-law. If the power-law extrapolation returns a rate function that deviates from the exact one by more than the propagated errors, the general claim fails; if it still converges, the power-law requirement is weaker than the paper assumes.","tokens_in":13162,"feed_emoji":"📊","tokens_out":13862,"duration_ms":122411,"temperature":0.7,"pith_summary":"Rare events are exponentially unlikely, so estimating their probabilities requires specialized sampling, and numerical estimates are always made at finite system size. This paper demonstrates a recipe that estimates the infinite-size rate function—the exponential decay rate of those probabilities—without ever building a histogram and without tying the user to one rare-event algorithm. The recipe combines multi-histogram reweighting of biased samples with a power-law extrapolation in system size, then applies a Legendre-Fenchel transform to the extrapolated scaled cumulant generating function. Tests on a binomial variable and on the largest connected component of Erdős–Rényi random graphs reproduce the known exact rate functions, except where a phase transition in the biased ensemble breaks the assumed scaling and leaves a gap in the estimate.","feed_headline":"Rare-event rate functions estimated directly at infinite system size","feed_subtitle":"A histogram-free pipeline plus power-law scaling turns any biased sampler into limiting rate function estimates.","key_machinery":"The machine is a four-step chain. Step one is exponentially biased sampling: for each bias value $q$, data are drawn from the tilted distribution $\\tilde p(S;N,q)\\propto p(S;N)e^{qS}$. Step two is histogram-free multi-histogram reweighting, which recombines all biased data to estimate $\\langle e^{qS}\\rangle_{p(S;N)}$ and the tilted mean $\\mu(q;N)=\\langle S\\rangle_{\\tilde p}/N$ at any $q$. Step three is a per-$q$ power-law fit $g(N;q)=C(q)+A(q)N^{-B(q)}$ that extrapolates both $\\Psi(q;N)$ and $\\mu(q;N)$ to $N=\\infty$. Step four is the parametric Legendre-Fenchel transform $I(\\mu(q))=q\\mu(q)-\\Psi(q)$, which converts each extrapolated pair into one point of the rate function with arbitrary density in $s$. The load-bearing identity is $\\Psi'(q)=\\mu(q)$, which makes the transform parametric and avoids solving a supremum for every $s$.","core_discovery":"The central discovery is that the limiting rate function $I(s)$ can be obtained from finite-size biased samples through the chain: estimate the scaled cumulant generating function $\\Psi(q;N)$ and its derivative $\\mu(q;N)$ at many bias values $q$ via histogram-free reweighting; fit each to a power law $C(q)+A(q)N^{-B(q)}$; take $C(q)$ as the $N\\to\\infty$ value; and evaluate the parametric Legendre-Fenchel transform $I(\\mu(q))=q\\mu(q)-\\Psi(q)$. Because the reweighting is done on raw data points rather than binned data, the resolution in $s$ is limited only by the number of $q$ values and by how well the biased samples cover the support. The binomial test and the Erdős–Rényi random graph test show agreement with exact analytical rate functions when the power-law scaling holds. At connectivity $c=0.5$, a biased-ensemble phase transition near $q\\approx0.317$ produces a double-peaked distribution, the power-law fit breaks down, and the estimated rate function acquires a gap; at $c=2.0$, the transform correctly returns the convex envelope but not concave parts of the rate function.","pith_inferences":["If the power-law assumption holds for a new model, the same pipeline should apply to any intensive observable satisfying a large-deviation principle, such as time-averaged currents in driven lattice gases, because the reweighting step never uses the model's dynamics.","The failure at biased-ensemble phase transitions could be turned into a diagnostic: a double-peaked reweighted distribution at some $q$ warns that the power-law fit at that $q$ will not extrapolate, so the gap could be forecast before fitting.","Since the Legendre-Fenchel transform always returns the convex envelope, the method cannot see concave regions of a non-convex rate function; combining it with a direct extrapolation of empirical rate functions would separate true concavity from numerical failure.","The error propagation currently ignores the correlation between $\\mu(q)$ and $\\Psi(q)$; accounting for that correlation is a straightforward extension that should tighten the quoted error bars, especially near the transition."],"forward_implications":["No binning is needed: the resolution of the estimated rate function is set by the number of bias values $q$, so any region of $s$ can be refined with extra $q$ values rather than new histograms.","The power-law extrapolation separates finite-size effects from the limiting result, so the returned rate function describes $N\\to\\infty$ rather than the largest simulated system.","Because the reweighting step consumes only raw data points, any algorithm that generates exponentially biased samples can feed the same post-processing pipeline.","The paper's error-propagation formulas turn the final rate function into an estimate with variances, so output can be quoted with confidence intervals.","A biased-ensemble phase transition is a detectable failure mode: the power-law fit deteriorates exactly where the reweighted distribution becomes bimodal, so the affected range of $q$ can be identified and excluded."],"supporting_citations":[{"why":"Supplies the weighted multi-histogram estimator used in Eq. (10) to turn all biased samples into expectation values at arbitrary bias $q$ without binning.","marker":"[36]"},{"why":"Shows that SCGF estimators from cloning follow a power-law finite-size scaling, the precedent for the extrapolation ansatz of Eq. (13).","marker":"[9]"},{"why":"Reports power-law finite-size scaling of large-deviation estimators, the second supporting reference for the extrapolation ansatz.","marker":"[39]"},{"why":"Documents the breakdown of finite-size scaling at biased-ensemble phase transitions that the paper reproduces for Erdős–Rényi graphs.","marker":"[40]"},{"why":"Provides the exact rate function for the largest connected component of Erdős–Rényi random graphs, the benchmark for comparison.","marker":"[47]"},{"why":"Describes the biased sampling scheme used to generate the Erdős–Rényi data points.","marker":"[5]"}],"fun_headline_variants":["Extrapolate rare-event rate functions to infinite size","Infinite-size rate functions from biased rare-event samples","Power-law fit yields exact large-deviation rate limits","Legendre-Fenchel path to infinite-size large-deviation rates","Finite biased ensembles scaled to exact rate-function limits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole scheme depends on the assumption that the finite-size corrections of the two extrapolated quantities shrink as a power of the system size; the paper gives no general proof of this and shows a case where it breaks.","fun_headline_variants_meta":{"raw":{"variants":["Extrapolate rare-event rate functions to infinite size","Infinite-size rate functions from biased rare-event samples","Power-law fit yields exact large-deviation rate limits","Legendre-Fenchel path to infinite-size large-deviation rates","Finite biased ensembles scaled to exact rate-function limits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000275,"raw_usage":{"total_tokens":1668,"prompt_tokens":997,"completion_tokens":671,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":592}},"tokens_in":613,"tokens_out":671,"duration_ms":7700,"temperature":1.0,"reasoning_tokens":592,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T21:38:25.263074+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the identical pipeline on the binomial model with $r(N)=r_\\infty+c/\\ln N$, where the exact finite-size and limiting rate functions are known analytically and the finite-size correction is logarithmic rather than power-law. If the power-law extrapolation returns a rate function that deviates from the exact one by more than the propagated errors, the general claim fails; if it still converges, the power-law requirement is weaker than the paper assumes.","supporting_citations":[{"cited_title":"Kumar, J","cited_arxiv_id":null,"evidence_quote":"Supplies the weighted multi-histogram estimator used in Eq. (10) to turn all biased samples into expectation values at arbitrary bias $q$ without binning."},{"cited_title":"Guevara Hidalgo, T","cited_arxiv_id":null,"evidence_quote":"Shows that SCGF estimators from cloning follow a power-law finite-size scaling, the precedent for the extrapolation ansatz of Eq. (13)."},{"cited_title":"Nemoto, E","cited_arxiv_id":null,"evidence_quote":"Reports power-law finite-size scaling of large-deviation estimators, the second supporting reference for the extrapolation ansatz."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the breakdown of finite-size scaling at biased-ensemble phase transitions that the paper reproduces for Erdős–Rényi graphs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes the biased sampling scheme used to generate the Erdős–Rényi data points."}],"review_version":1}