{"id":"d43a28c2-1386-49c0-8bf3-58fa36ecbbe9","arxiv_id":"2501.07977","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A Bayesian exchange-Monte-Carlo framework using the Ott-Antonsen analytical solution recovers Kuramoto model parameters from order-parameter time series, with accuracy limited by noise, system size, and a K-gamma identifiability degeneracy.","lead":"Researchers built a Bayesian method that estimates the coupling strength and frequency spread of a Kuramoto oscillator model using only the time series of the macroscopic order parameter, not individual oscillators. A smart generalist would read this because it targets practical synchronization systems, such as circadian cells or neural populations, where only aggregate signals are available.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The white-noise observation model (Eq. 8) is contradicted by the paper's own Fig. 5, so posterior widths and the empirical-Bayes selection of b are miscalibrated for real Kuramoto data; the paper discloses this and hedges appropriately, so the verdict stays CONDITIONAL.","rationale":"I agree with the reader that the white-noise assumption (Eq. 8) is the single most load-bearing assumption, because it determines all posterior widths and the empirical-Bayes selection of b. The paper itself provides direct evidence against it in Fig. 5 and shows its consequence in Figs. 4(g,h), where true parameters fall outside ±1 SD for every N including N=10^5. At the same time, the paper is unusually transparent: it explicitly states the discrepancy, analyzes the autocorrelation of the residual, and proposes a concrete OU-based alternative (Eqs. 38–39). The MAP point estimates do converge toward true values as N grows, and the benchmark experiments (Fig. 2) validate the internal consistency of the EMC implementation. The additional degeneracy along Eq. 33 is real but disclosed and analyzed in Appendix C. These supporting elements justify a CONDITIONAL verdict rather than rejection: the method is promising and honestly presented, but the reported posterior uncertainties are not yet trustworthy for Kuramoto-model data, and the separate identifiability of K and γ remains weak. My proposed test directly targets whether the authors' own proposed fix resolves the miscalibration, which is the key open question for the central claim.","tokens_in":20211,"tokens_out":2034,"duration_ms":19804,"concrete_test":"Re-run the N=10^5 Kuramoto-data experiments with the OU-based observation model (Eqs. 38–39) in place of Eq. 8, using the same EMC machinery. If the resulting |error|/SD ratios in the analogues of Figs. 4(g,h) move into the [0,1] band while MAP accuracy is retained, the white-noise assumption is confirmed as the source of posterior miscalibration; if the ratios stay above 1, the misspecification lies deeper in the OA forward model itself.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that OA-based Bayesian inference yields reliable estimates of K and γ, including posterior uncertainties, rests on the observation model yi = Rsol(ti;θ) + ξi with ξi i.i.d. Gaussian white noise (Eq. 8). The paper's own Fig. 5 shows that for data from the Kuramoto model (Eq. 1), the residual Rdif(t) = R(t) − Rsol(t;θ) has a positive autocorrelation decaying exponentially over τ up to ~20, contradicting the white-noise assumption. Since this assumption enters the likelihood (Eq. 10), it directly controls all posterior widths and the empirical-Bayes selection of b via the free energy (Eqs. 24–25). The consequence is visible in Figs. 4(g,h): the ratio |error|/SD exceeds 1 on average for all N, including N=10^5, so the reported ±1 SD intervals are systematically overconfident. The paper is explicit about this mismatch and proposes an OU-based replacement (Eqs. 38–39) as future work. A second, separate limitation is the strong posterior correlation along γ ≈ 0.33K + 0.0635 (Eq. 33), which the authors analyze in Appendix C as a flat direction of the forward model; as a result, the data alone do not sharply separately identify K and γ, only a combination. Both issues are disclosed and analyzed, so they weaken the practical claim but do not invalidate the careful point estimates or the conditional contribution of the method.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Bayesian framework for estimating the coupling strength K and heterogeneity γ of a Kuramoto model from time series of the macroscopic order parameter R(t) alone. The forward model is the Ott-Antonsen analytical solution R_sol(t;θ) (Eq. 29) under a Lorentzian frequency distribution, and the observation model is R(t) = R_sol(t;θ) plus Gaussian white noise (Eq. 8). Parameters are estimated by exchange Monte Carlo (EMC); the noise scale b is selected by empirical Bayes via the marginal likelihood (Eqs. 24–25), and K and γ are obtained as MAP estimates. The paper validates the approach in a benchmark where data are generated from Eq. (8), showing convergence to truth as σ→0, and then tests it on Kuramoto-model data with N = 100–10^5, reporting point estimates that approach the true values as N grows but posterior intervals that miss the true parameters at all N. The authors diagnose the latter as a violation of the white-noise assumption and propose an OU-process observation model as future work.","tokens_in":20500,"tokens_out":6430,"duration_ms":60688,"significance":"The paper contributes an original combination of the OA analytical solution with EMC-based Bayesian inference for macroscopic observables, and the implementation is supported by open code and a careful benchmark. The distinction between benchmark consistency and model-mismatch behavior is clearly drawn, and the autocorrelation analysis in Fig. 5 is a valuable diagnostic. However, the central advantage of Bayesian inference — calibrated posterior uncertainties — is not achieved for the actual Kuramoto model: Figs. 4(g,h) show systematic overconfidence at all system sizes, a direct consequence of the white-noise assumption in Eq. (8). The identifiability ridge between K and γ (Eq. 33) further limits the practical significance of separate point estimates. If the observation model is repaired or the claims are restricted accordingly, the method could have real utility for biological and physical synchrony data.","major_comments":[{"comment":"The observation model y_i = R_sol(t_i;θ) + ξ_i with i.i.d. Gaussian ξ_i (Eq. 8) is contradicted by the paper's own autocorrelation results: Fig. 5(b) shows C(τ) decaying exponentially over τ ≲ 20, not a delta function. Because Eq. (8) enters the likelihood (Eq. 10), the posterior widths and the empirical-Bayes selection of b via F(b) (Eqs. 24–25) are miscalibrated for data generated from Eq. (1). The consequence is visible in Figs. 4(g,h): the ratio |error|/SD exceeds 1 on average for every N, including N=10^5, so the reported ±1 SD intervals are systematically overconfident. Since the central advantage of the Bayesian framework is reliable uncertainty quantification, this model mismatch is load-bearing. The proposed OU replacement (Eqs. 38–39) is presented only as future work and is not evaluated; the main text should either implement and test such a model (or an equivalent correction) or explicitly restrict the claims to point estimation with the caveat that posterior uncertainties are not trustworthy for finite-N Kuramoto data.","section":"§III C, Figs. 4(g,h) and Fig. 5(b)"},{"comment":"The benchmark in Sec. III B validates internal consistency, but it is self-referential: the same model (Eq. 8) is used for both data generation and inference, so it does not validate the forward model against the Kuramoto dynamics. The real-data experiments in Sec. III C show that point estimates of K and γ converge toward the true values as N increases, but the systematic overconfidence identified above is not captured by the benchmark. The paper should add a test of posterior calibration under model mismatch — e.g., a probability-integral-transform histogram or coverage curve across many simulated datasets — or clearly state that the reported credible intervals apply only to the white-noise observation model.","section":"§III B and §III C"},{"comment":"The strong posterior correlation γ ≈ 0.33K + 0.0635 (Eq. 33) and the local flat-direction analysis in Appendix C indicate that K and γ are not separately identifiable from a single R(t) trajectory: the data constrain a particular combination much more strongly than the individual parameters. The paper analyzes this as a property of the forward model, but it does not discuss what it implies for the reported MAP values of K and γ, which are determined along the ridge largely by the prior and by the exact location of the ridge. Please add an explicit identifiability statement and report credible intervals for the identified combination (e.g., γ − 0.33K or the projection along the flat direction), not only marginal intervals for K and γ.","section":"§III C and Appendix C"}],"minor_comments":[{"comment":"The phrase 'flushing of fireflies' should be 'flashing of fireflies' (both occurrences).","section":"Abstract and Introduction"},{"comment":"The text refers to 'Figs. 3 (C-3), (D-3)' but Fig. 3 has no panel D; this appears to be a typo for panels (A-3)–(C-3).","section":"§III C 3"},{"comment":"The definition 'ˆσ := 1/b2\\,_l' is dimensionally incorrect; since b = 1/σ², it should be 1/√b_l.","section":"Fig. 1 caption"},{"comment":"The condition 'i/nθ is an integer' should be stated as 'i is a multiple of nθ' to avoid ambiguity about integer division.","section":"Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The work is likely suitable for a physics/data-analysis journal if the model-mismatch issue is addressed. The authors are transparent about limitations, and the code availability is a plus. I would not recommend rejection because the point-estimation machinery appears sound; the needed work is to calibrate uncertainties or reframe the contribution as point estimation plus an honest diagnostic of the observation-model mismatch."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the genuinely new bit is using the Ott–Antonsen analytical solution as the forward model for Bayesian parameter estimation from the Kuramoto order parameter alone; that combination is not in the cited literature. Second, the paper is unusually honest about where it fails: its own Fig. 5 shows the residual between the Kuramoto order parameter and the OA solution is autocorrelated, not white noise, so the posterior intervals in the finite-N experiments are systematically too tight (Fig. 4g,h). The authors say so and propose an OU-process replacement as future work.\n\nThe execution is solid. The exchange Monte Carlo implementation is standard for Bayesian spectroscopy, and the benchmark where data are generated from the same model is a legitimate internal consistency check. The MAP estimates converge toward the true values as N grows, and the paper does not pretend otherwise. Appendix C's analysis of the K–gamma degeneracy is a real asset: it identifies a flat direction in the forward model and shows the empirical regression slope matches the numerically computed flat direction. That is careful, reproducible work.\n\nThe soft spots are real but disclosed. The white-noise model enters the likelihood directly, so the posterior widths and the empirical-Bayes selection of b are miscalibrated whenever the data come from an actual finite-N Kuramoto system. This is not a minor technicality; it means the uncertainty quantification is not trustworthy for the stated application. The degeneracy is more fundamental: the data constrain one combination of K and gamma, not the two parameters separately, which weakens the headline claim of estimating coupling strength and heterogeneity independently. Both issues are analyzed in the paper, so I would not call it misleading—just incomplete. Reproducibility is decent in principle; there is a GitHub repository, but without a pinned commit or environment, so I would not call it fully reproducible as shipped.\n\nWho gets value: anyone working on macro-scale inference for oscillator populations, especially circadian or neuroscience applications where individual oscillators aren't trackable. The paper is worth refereeing; the fix is either to move the OU model into the main results or to reframe the contribution as point estimation along the degenerate direction with uncertainties explicitly uncalibrated. I would send it to review with that expectation.","headline":"A competent, honest method paper: novel use of the Ott-Antonsen solution as a Bayesian forward model, but the white-noise observation model makes the posterior intervals overconfident and the K-gamma degeneracy limits separate identifiability.","tokens_in":21096,"tokens_out":2908,"would_cite":true,"duration_ms":30866,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["05.45.Xt"],"model":"deepseek-v4-flash","headline":"A Bayesian framework using the Ott-Antonsen solution and exchange Monte Carlo estimates the coupling strength and heterogeneity of a Kuramoto population from time series of a single macroscopic quantity, the order parameter.","keywords":["Kuramoto model","Ott-Antonsen ansatz","Bayesian inference","exchange Monte Carlo","order parameter","parameter estimation","marginal likelihood","synchronization"],"falsifier":"Re-run the exchange Monte Carlo inference on a long finite-Kuramoto trajectory using the Ornstein-Uhlenbeck observation model of Eq. (38) and measure the empirical coverage of the 68% posterior intervals; if the coverage remains below the nominal level for $N \\geq 10^4$, the macroscopic-only identification strategy has not yet achieved calibrated uncertainty.","tokens_in":19948,"feed_emoji":"🔄","tokens_out":5354,"duration_ms":48308,"temperature":0.7,"pith_summary":"This paper claims that Bayesian inference can recover two key parameters of a synchronizing population—the coupling strength K and the frequency heterogeneity γ—using only the time series of the Kuramoto order parameter R(t), without data on individual oscillators. The forward model is the exact Ott-Antonsen solution for the order parameter in the thermodynamic limit, and the posterior and marginal likelihood are computed by exchange Monte Carlo. In benchmark data generated from the same model, estimates converge to the truth as observational noise shrinks, with calibrated posterior intervals. In finite-size Kuramoto simulations, the maximum a posteriori estimates converge to the true values as the number of oscillators grows, but the posterior intervals are too narrow because the residual fluctuations are not white noise. The paper diagnoses this mismatch with an autocorrelation analysis and proposes an Ornstein-Uhlenbeck observation model as the next step.","feed_headline":"Bayesian fit reads coupling strength from sync curves","feed_subtitle":"Ott-Antonsen solution plus exchange Monte Carlo turns order-parameter time series into model parameters and error bars.","key_machinery":"The carrying mechanism is the Ott-Antonsen ansatz: in the thermodynamic limit with a Lorentzian frequency distribution, the order parameter obeys the closed ODE $\\dot{R} = -\\lambda R + (\\lambda - \\gamma)R^3$ with $\\lambda = \\gamma - K/2$, whose non-constant solution $R_{\\mathrm{sol}}(t; \\theta)$ becomes the deterministic backbone of the likelihood. Around it sits a Gaussian likelihood with precision $b$, exchange Monte Carlo over replicas at inverse temperatures $b_l$ to sample the posterior and compute the marginal likelihood by chaining of importance sampling, and empirical Bayes to pick $b$. The same machinery reveals a near-ridge direction in $(K, \\gamma)$ along which the solution shape barely changes, explaining the strong positive correlation in the posterior samples.","core_discovery":"The central discovery is that parameter estimation of the Kuramoto model can be performed from macroscopic quantities alone by attaching a Gaussian likelihood to the Ott-Antonsen analytical solution of the order parameter. Under the assumption $y_i = R_{\\mathrm{sol}}(t_i; \\theta) + \\xi_i$ with $\\xi_i$ independent Gaussian noise, the authors set the noise precision $b = 1/\\sigma^2$ as an unknown hyperparameter, select it by empirical Bayes (maximizing the marginal likelihood over a ladder of inverse temperatures), and take the MAP estimate of $\\theta = (\\gamma, \\lambda)$ from the selected replica. Exchange Monte Carlo supplies both the posterior samples and, via chaining of importance sampling, the marginal likelihood. In the benchmark case the method is internally consistent, and in the Kuramoto-model case the MAP estimates approach the true $(K, \\gamma)$ as $N$ increases. The paper also demonstrates that the residual $R(t) - R_{\\mathrm{sol}}(t)$ is not white noise but has an exponentially decaying autocorrelation, so the white-noise likelihood understates uncertainty; this is not treated as a failure of the estimation approach but as evidence that the observation model should be upgraded to an Ornstein-Uhlenbeck process.","pith_inferences":["The near-linear ridge in the posterior means that $K$ and $\\gamma$ are individually identifiable only up to the combination preserving $R_{\\mathrm{sol}}(t)$; a single desynchronization curve is unlikely to separate them unless the prior breaks the ridge or the observable is changed.","The exponentially decaying autocorrelation of residuals suggests that a two-dimensional Ornstein-Uhlenbeck reduction of finite-size Kuramoto dynamics could act as an explicit stochastic forward model, turning the proposed Eq. (38) into a principled likelihood rather than a heuristic fix.","The same estimation design could be tested on real collective oscillations where only population-level signals are available, such as leaf circadian luminescence, with the caveat that the observation noise must first be characterized to reach calibrated uncertainties."],"forward_implications":["If the method is right, coupling strength and heterogeneity of a Kuramoto-like population can be monitored from population-level observables such as the averaged macroscopic signal, without resolving individual oscillators.","The empirical-Bayes selection of the noise precision $b$ yields an automatic estimate of observational noise, so the framework can in principle be applied to noisy experimental desynchronization curves.","Because the marginal likelihood is computed, the same machinery can compare competing coupled-oscillator models (for instance Kuramoto versus Sakaguchi-Kuramoto) on a given dataset.","The failure of the white-noise model at finite $N$ is a concrete diagnostic: uncertainty from the current observation model is optimistic, and the proposed Ornstein-Uhlenbeck noise model is the natural follow-up."],"supporting_citations":[{"why":"Supplies the Ott-Antonsen closed-form solution for the order parameter that serves as the paper's deterministic forward model.","marker":"[22]"},{"why":"Introduces the exchange Monte Carlo method whose replica-exchange scheme is used to sample the posterior and compute marginal likelihood.","marker":"[26]"},{"why":"Establishes the Bayesian spectral-deconvolution application of exchange Monte Carlo that the paper adapts to parameter estimation.","marker":"[27]"},{"why":"Provides the chaining-of-importance-sampling technique used to calculate the marginal likelihood from replica samples.","marker":"[36]"},{"why":"Gives the finite-size stochastic reduction of the Kuramoto-Sakaguchi model that supports the proposed Ornstein-Uhlenbeck observation model.","marker":"[43]"},{"why":"Presents the alternative macroscopic-observation estimation approach based on linear and nonlinear response theory that this work contrasts with.","marker":"[24]"}],"fun_headline_variants":["Bayesian estimation of oscillator coupling from order parameter","Sync waves reveal coupling strength via Bayesian inference","Order parameter time series enough for oscillator parameter fit","Bayesian method infers oscillator coupling from macroscopic data","Order parameter alone suffices for oscillator parameter inference"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole uncertainty calculus rests on the assumption that the observed order parameter equals the thermodynamic-limit Ott-Antonsen solution plus independent Gaussian white noise; the paper's own finite-size simulations show the residuals are correlated, so the posterior intervals are overconfident.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian estimation of oscillator coupling from order parameter","Sync waves reveal coupling strength via Bayesian inference","Order parameter time series enough for oscillator parameter fit","Bayesian method infers oscillator coupling from macroscopic data","Order parameter alone suffices for oscillator parameter inference"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000757,"raw_usage":{"total_tokens":3361,"prompt_tokens":936,"completion_tokens":2425,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":552,"completion_tokens_details":{"reasoning_tokens":2368}},"tokens_in":552,"tokens_out":2425,"duration_ms":16002,"temperature":1.0,"reasoning_tokens":2368,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:30:03.403603+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the exchange Monte Carlo inference on a long finite-Kuramoto trajectory using the Ornstein-Uhlenbeck observation model of Eq. (38) and measure the empirical coverage of the 68% posterior intervals; if the coverage remains below the nominal level for $N \\geq 10^4$, the macroscopic-only identification strategy has not yet achieved calibrated uncertainty.","supporting_citations":[{"cited_title":"Pietras and A","cited_arxiv_id":null,"evidence_quote":"Introduces the exchange Monte Carlo method whose replica-exchange scheme is used to sample the posterior and compute marginal likelihood."},{"cited_title":"Nagata, S","cited_arxiv_id":null,"evidence_quote":"Gives the finite-size stochastic reduction of the Kuramoto-Sakaguchi model that supports the proposed Ornstein-Uhlenbeck observation model."}],"review_version":1}