{"id":"63f1a8eb-8f42-4ddf-80ed-8095e8fabeaa","arxiv_id":"2508.21220","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Even in almost-Gaussian discrete signaling systems, the Gaussian approximation underestimates the true information rate, and for nonlinear continuous systems its accuracy depends on how it is applied.","lead":"This paper tests how accurate a common shortcut, the Gaussian approximation, is at measuring information flow in living cells and other noisy systems. It finds the shortcut can be badly wrong for discrete systems even when they look nearly random, and gives guidance on when exact calculations are needed.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PWS ground-truth convergence (sample size, trajectory length, Euler–Maruyama time step) is not demonstrated, so the reported deviations—especially the low-copy-number optimum—could be artifacts.","rationale":"The reader's weakest assumption identifies the PWS ground truth as the load-bearing point, and I agree. The manuscript itself states that PWS is exact only for infinite sample size and, for the Langevin extension, infinitesimal time step (Eqs. 9 and 11), yet none of the required convergence limits are quantified. This matters especially for the low-copy-number optimum in Fig. 1b, which is a surprising, qualitative claim and is explicitly left for future work. Because the paper's contribution is to quantify deviations from an 'exact' benchmark, the absence of convergence evidence leaves the central numerical claims conditional. I would keep the reader's CONDITIONAL verdict; the concern does not overturn the paper's theoretical framing or its qualitative directions, but it does mean the specific quantitative findings should be accepted only after the proposed convergence/independent-check is supplied.","tokens_in":15145,"tokens_out":13819,"duration_ms":150618,"concrete_test":"Recompute the mutual information rate for the discrete linear network at the parameters of Fig. 1b (e.g., κ=1, λ=1, ρ=10, μ=1) using an independent high-accuracy method—numerical integration of the stochastic filtering equation as in Moor & Zechner [4]—and compare with PWS and the discrete approximation. Also run PWS with N=10^4, 10^5, 10^6 and T=10, 20, 50, 100 to verify stability and confirm that the peak at κ≈1 persists. If the independent exact value does not lie on the reported PWS curve, the ground-truth claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claims rest on PWS as exact ground truth, but the paper does not demonstrate convergence of the three limits that define PWS: (i) the number of Monte Carlo samples N in Eq. (9), (ii) the trajectory length T in the rate limit of Eq. (4), and (iii) for the Langevin extension, the discretization time step Δt in the Onsager–Machlup path weight Eq. (11). Reported rates have no error bars, no sample sizes, no trajectory lengths, and no convergence checks. This is not merely a reproducibility gap: the distinctive results—the nonmonotonic maximum at low input copy number in Fig. 1b, the claimed accuracy of the discrete approximation for output copy numbers much smaller than one, and the signed deviations in Figs. 3–4—are exactly the quantities that finite-T and finite-Δt biases can corrupt. The paper itself flags the low-copy-number optimum as 'surprising' and leaves it for future work, but without convergence evidence it cannot be distinguished from a PWS artifact.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript evaluates the Gaussian approximation (GA) to the mutual information rate in two model problems, using Path Weight Sampling (PWS) as an exact benchmark. For a linear discrete reaction network, the LNA-based GA is shown to be a lower bound that is not tight even at large copy numbers, whereas a recently proposed 'discrete approximation' agrees with PWS except at low input copy numbers, where PWS indicates a nonmonotonic maximum. For a nonlinear Langevin system with Hill-type activation, the LNA-based GA overestimates and an empirical Gaussian estimate underestimates the exact rate, with errors increasing with gain and decreasing with response time. The paper derives closed-form expressions, discusses the mechanism behind the discreteness effect, and gives practical guidance on when approximate methods suffice.","tokens_in":15430,"tokens_out":9008,"duration_ms":90757,"significance":"If the numerical results are reliable, the paper provides a useful quantitative caution against routine use of the Gaussian approximation and gives concrete guidance. Its analytic derivations are self-consistent, and benchmarking against an independent exact Monte Carlo method is appropriate. The extension of PWS to Langevin dynamics is a useful methodological contribution. However, the central quantitative claims—especially the low-copy-number optimum and the signed deviations in the nonlinear system—rest entirely on PWS estimates for which no convergence evidence or uncertainty is reported. As it stands, the paper's empirical conclusions are plausible but not fully substantiated.","major_comments":[{"comment":"The PWS benchmark is defined by three limits: N→∞ in Eq. (9), T→∞ in Eq. (4), and Δt→0 in Eq. (11). None of these limits is documented for any figure: no sample sizes, trajectory lengths, time steps, convergence checks, or error bars are reported. This is load-bearing because the central quantitative claims—the non-tight lower bound in Fig. 1a, the nonmonotonic optimum in Fig. 1b, and the signed deviations in Figs. 3–4—are deviations from PWS. Without evidence that the PWS estimates have converged, the reported deviations could be numerical artifacts. Please add convergence diagnostics and error bars/confidence intervals for all PWS points.","section":"§II.C, Eqs. (9), (11); §III.A–B"},{"comment":"The low-copy-number maximum of the information rate is a central new observation, yet the paper itself states that 'a precise characterization of this finding' is left for future work. Because both analytic approximations are independent of κ, the entire phenomenon rests on the PWS points for small κ. The figure shows no error bars and no dependence on trajectory length or sample size. The authors should show that the nonmonotonicity is stable under increasing T and N, and provide uncertainty estimates, before concluding that low input copy numbers maximize the rate.","section":"§III.A, Fig. 1b"},{"comment":"The empirical Gaussian rate is computed by Welch's method from simulated trajectories. Finite-sample spectral and coherence estimates are biased; the manuscript does not report window length, overlap, number of independent realizations, or statistical uncertainty. The claimed lower-bound behavior relies on the asymptotic Mitra–Stark argument, so the numerical verification in Fig. 4 must be shown not to be an artifact of spectral estimation bias. Please provide the estimation parameters and error bars, and ideally a check that the empirical Gaussian estimate converges to the true spectral coherence from below as data length increases.","section":"§III.B, Eq. (24), Fig. 4"},{"comment":"For diffusive systems, the path weight Eq. (11) is valid only in the limit Δt→0 and, for multiplicative noise, up to a normalization constant that depends on x_i and therefore does not cancel in Eq. (9). The paper does not report Δt or a convergence study for the Euler–Maruyama discretization. While the case studies appear to use additive noise (constant σ), the general claim that PWS has been extended to diffusive dynamics needs a statement of the discretization error and a convergence check for the specific systems studied.","section":"§II.C, Eq. (11)"}],"minor_comments":[{"comment":"Typo: 'Euler-Mayurama' should be 'Euler–Maruyama'.","section":"§II.C"},{"comment":"The displayed estimator appears to be missing an explicit 1/N prefactor; as written it reads like a sum rather than an average.","section":"§II.C, Eq. (9)"},{"comment":"The colorbar labels are hard to parse ('10 1 100 101'); they should be formatted as 10^{-1}, 10^0, 10^1, and the units of τx should be stated.","section":"Figs. 3–4"},{"comment":"No code or data availability statement is provided. For reproducibility, the PWS implementation, spectral estimation parameters, and simulation details should be included in an appendix or public repository.","section":"General"},{"comment":"The phrase 'even when the system's statistics are nearly Gaussian' is not quantified. A brief measure of non-Gaussianity (e.g., skewness of the stationary distribution or a path-space diagnostic) would strengthen the claim.","section":"Abstract/Introduction"}],"recommendation":"major_revision","confidential_remarks":"The manuscript makes strong quantitative claims based on Monte Carlo estimates, yet no error bars or convergence diagnostics are reported. This is surprising for a study whose main message is about the size of deviations from an approximation. I would encourage the editor to require a reproducibility appendix with the PWS and spectral estimation parameters before publication, as the current form does not allow the reader to verify the central numerical findings."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this if you work on information rates in biochemical networks. The genuinely new piece is the PWS extension to Langevin/diffusive systems, and the finding that in a nonlinear continuous model the LNA-based Gaussian rate overestimates while the empirical Gaussian estimate underestimates the exact rate. That bracketing result is clean and practically useful. The discrete-system part largely repeats what this group already established in Refs. [4,5,24]—the Gaussian lower bound is not new—but it is a clear write-up and the practical recommendations at the end are sensible.\n\nThe soft spot is exactly what the stress-test note flags. PWS is exact only in the limits N→∞ in Eq. (9), T→∞ in Eq. (4), and Δt→0 in Eq. (11). I could not find any reporting of N, T, Δt, number of samples, or convergence checks anywhere in the paper or figures. No error bars. That is not merely a reproducibility gap: it directly affects the two most interesting quantitative outcomes. The nonmonotonic maximum at low input copy number in Fig. 1b is 'surprising' and left for future work—but without convergence evidence it could be a finite-sample or finite-time artifact. Likewise the claimed accuracy of the discrete approximation for output copy numbers much smaller than one needs a check. I am not saying the results are wrong; I am saying the paper does not currently let you tell.\n\nThe derivations in the appendices look self-consistent: Eq. (A19) reduces to the known linear expression Eq. (A22), and the Mitra–Stark lower-bound argument is applied correctly, with the caveat noted that the LNA-based estimate does not inherit the bound. Self-citation is heavy but mostly legitimate since the benchmark PWS is their own method, published separately.\n\nThe paper is honest and clearly written; the authors flag the open questions themselves. I would send it to a serious referee, but conditionally: the referee should demand convergence diagnostics and error bars, and a robustness check on the low-copy-number maximum. As is, I would not build on Fig. 1b without seeing those details.","headline":"Useful, clearly written, and honest, but the missing convergence diagnostics for the PWS ground truth make the novel quantitative claims—especially the low-copy-number optimum—provisional.","tokens_in":15845,"tokens_out":2034,"would_cite":true,"duration_ms":20190,"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":"For discrete biochemical networks the Gaussian approximation never reaches the exact information rate, and in nonlinear continuous systems its bias direction depends on the linearization route.","keywords":["mutual information rate","Gaussian approximation","linear noise approximation","Path Weight Sampling","stochastic reaction networks","Langevin dynamics","information theory","biochemical signaling"],"falsifier":"Recompute the four-reaction network's information rate with an independent exact method—e.g., numerical integration of the stochastic filtering equation—at large copy numbers (s̄,x̄ ≥ 100) and compare with the LNA closed form; agreement would refute the claim that the Gaussian bound never tightens. For the nonlinear Langevin system, scan parameters with fast response and high gain and check whether the empirical Gaussian estimate ever exceeds the PWS rate; the paper predicts it never does.","tokens_in":15118,"feed_emoji":"📡","tokens_out":10378,"duration_ms":94440,"temperature":0.7,"pith_summary":"This paper tests when the standard shortcut for computing the mutual information rate—the amount of information shared per unit time between input and output trajectories—can be trusted for systems that are not actually Gaussian. Using Path Weight Sampling, an exact Monte Carlo method, the authors show that for a simple discrete biochemical reaction network the Gaussian approximation is only a lower bound and stays a lower bound even when copy numbers are large and statistics look nearly Gaussian. For a continuous nonlinear Langevin system, the Gaussian approximation is biased in opposite directions depending on how it is applied: linearizing the dynamics analytically overestimates the true rate, while estimating spectra from simulations underestimates it. The practical payoff is a set of concrete usage conditions: rely on the Gaussian approximation for weak nonlinearity or slow output responses, and switch to exact methods for low input copy numbers or fast, strongly nonlinear systems.","feed_headline":"Gaussian approximation misses bits even when systems look Gaussian","feed_subtitle":"A loose lower bound in discrete networks, a two-way bias in nonlinear ones—here's when to trust the shortcut.","key_machinery":"The load-bearing object is the mutual information rate R(S,X), the long-time limit of trajectory mutual information. Path Weight Sampling (PWS) estimates it exactly by Monte Carlo: sample input trajectories s_T from P(s_T), output trajectories x_T from P(x_T|s_T), and average log P(x_T|s_T) − log P(x_T); for Langevin dynamics the path weight is the Onsager–Machlup action. The Gaussian approximation replaces trajectory statistics by a jointly Gaussian model, reducing the rate to a frequency integral over the coherence φ_sx(ω)=|S_sx(ω)|²/(S_ss(ω)S_xx(ω)). The two variants differ in how the spectra are obtained: analytically from the linear noise approximation (LNA), or numerically from simulat","core_discovery":"The paper's central claim is that the Gaussian approximation of the mutual information rate—exact only for linear systems with additive Gaussian noise—systematically misestimates the true rate in non-Gaussian systems. In a four-reaction linear network, the LNA-based Gaussian rate R=(λ/2)(√(1+ρ/λ)−1) is a strict lower bound that does not tighten as copy numbers grow; a reaction-based discrete approximation R=(λ/2)(√(1+2ρ/λ)−1) matches exact PWS results over nearly all parameters, except low input copy numbers where the true rate peaks above both. In a continuous Langevin system with a Hill-type nonlinearity, linearizing dynamics via the LNA overestimates the exact rate, while estimating spect","pith_inferences":["The distinction between reaction-based and state-based readouts implies the Gaussian bound is the information accessible to a decoder that cannot resolve individual reaction events; a testable prediction is that a downstream decoder with finer time resolution approaches the higher PWS rate.","The opposite bias directions likely extend beyond Hill kinetics to any saturating input–output relation, so repeating the comparison with Michaelis–Menten or exponential saturation would test whether the asymmetry is generic.","The low-input-copy-number peak suggests intrinsic noise can carry information rather than only degrade it; if real signaling networks tune input copy numbers near this optimum, that would be evidence of noise exploitation."],"forward_implications":["For linear discrete reaction networks, the Gaussian approximation is a strict lower bound on the true information rate and never becomes tight at high copy numbers; estimates from it should be interpreted as a floor, not a prediction.","The discrete approximation of [4] is accurate over a wide parameter range (output copy numbers down to ≪1 and input copy numbers ≳10), making it a cheap replacement for exact computation in those regimes.","For continuous nonlinear systems, the empirical Gaussian estimate is a lower bound for Gaussian inputs, while the LNA-based estimate is not bounded; across the tested parameters the empirical route is closer to the exact rate.","Practical usage rule: Gaussian approximations are reliable for small gain or slow output response; exact methods like PWS are needed for low input copy numbers (≲10) in discrete networks and for fast, strongly nonlinear continuous systems.","The true information rate of the discrete network has an optimum at low input copy number (s̄ ≈ 1), above both analytic approximations, so reducing input copy number can increase information transmission in this regime."],"supporting_citations":[{"why":"Supplies the analytical Gaussian information-rate formula for the linear motif and the spectral formula being tested.","marker":"[1]"},{"why":"Provides the reaction-based discrete approximation whose accuracy is compared against PWS and the Gaussian result.","marker":"[4]"},{"why":"Introduces Path Weight Sampling, the exact Monte Carlo method used as ground truth and extended here to Langevin dynamics.","marker":"[5]"},{"why":"Gives the theoretical result that a Gaussian model with matching covariance is a lower bound on the mutual information for a Gaussian input with a non-Gaussian output, underpinning the empirical-Gaussian lower bound.","marker":"[23]"},{"why":"Identifies the root cause of the Gaussian approximation's failure: distinct reaction events are lumped into one continuous noise term in the LNA.","marker":"[24]"},{"why":"Supplies the linear noise approximation used to derive the LNA-based Gaussian estimates and the linearized input-output mapping.","marker":"[27]"},{"why":"Provides the Onsager–Machlup path weight used to evaluate PWS for diffusive Langevin systems.","marker":"[28]"},{"why":"Confirms that for linear networks the LNA power spectra match the empirical spectra, isolating nonlinearity as the source of the two routes' divergence.","marker":"[33]"}],"fun_headline_variants":["Gaussian info-rate shortcut fails even in near-Gaussian systems","Even near-Gaussian stats can't save the Gaussian info-rate shortcut","Loose lower bound, two-way bias: Gaussian info-rate approximations","Exact mutual info vs Gaussian shortcut: the gap you can't ignore","Exact vs Gaussian info rate: gap widens with nonlinearity"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The conclusions treat the Path Weight Sampling results as ground truth, but PWS is exact only with infinitely many Monte Carlo samples and, for the Langevin extension, an infinitesimal time step; the paper does not report convergence checks, so any systematic bias in the PWS estimates would change the measured deviations of the Gaussian approximation.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian info-rate shortcut fails even in near-Gaussian systems","Even near-Gaussian stats can't save the Gaussian info-rate shortcut","Loose lower bound, two-way bias: Gaussian info-rate approximations","Exact mutual info vs Gaussian shortcut: the gap you can't ignore","Exact vs Gaussian info rate: gap widens with nonlinearity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002014,"raw_usage":{"total_tokens":7700,"prompt_tokens":765,"completion_tokens":6935,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":6858}},"tokens_in":509,"tokens_out":6935,"duration_ms":51813,"temperature":1.0,"reasoning_tokens":6858,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:27:52.657819+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the four-reaction network's information rate with an independent exact method—e.g., numerical integration of the stochastic filtering equation—at large copy numbers (s̄,x̄ ≥ 100) and compare with the LNA closed form; agreement would refute the claim that the Gaussian bound never tightens. For the nonlinear Langevin system, scan parameters with fast response and high gain and check whether the empirical Gaussian estimate ever exceeds the PWS rate; the paper predicts it never does.","supporting_citations":[{"cited_title":"The independent Gaussian white noise process ηs(t) sum- marizes all reactions that contribute to fluctuations in S","cited_arxiv_id":null,"evidence_quote":"Supplies the analytical Gaussian information-rate formula for the linear motif and the spectral formula being tested."},{"cited_title":"Any difference between the exact information rate and the Gaussian information rate must then be a result of the Gaussian approximation","cited_arxiv_id":null,"evidence_quote":"Provides the reaction-based discrete approximation whose accuracy is compared against PWS and the Gaussian result."},{"cited_title":"Mutual information between in- and output trajectories of biochemical networks","cited_arxiv_id":"0901.0280","evidence_quote":"Introduces Path Weight Sampling, the exact Monte Carlo method used as ground truth and extended here to Langevin dynamics."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the theoretical result that a Gaussian model with matching covariance is a lower bound on the mutual information for a Gaussian input with a non-Gaussian output, underpinning the empirical-Gaussian lower bound."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Identifies the root cause of the Gaussian approximation's failure: distinct reaction events are lumped into one continuous noise term in the LNA."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the linear noise approximation used to derive the LNA-based Gaussian estimates and the linearized input-output mapping."},{"cited_title":"State- versus Reaction-Based Information Processing in Biochemical Networks","cited_arxiv_id":"2505.13373","evidence_quote":"Provides the Onsager–Machlup path weight used to evaluate PWS for diffusive Langevin systems."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Confirms that for linear networks the LNA power spectra match the empirical spectra, isolating nonlinearity as the source of the two routes' divergence."}],"review_version":1}