{"id":"dc951d4c-44ca-41c2-83c6-68043e4fabb4","arxiv_id":"2507.23426","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ODR-BINDy recovers governing ODEs from noisy data by relaxing the model equation as a soft constraint and selecting terms through Bayesian evidence.","lead":"This paper introduces ODR-BINDy, a method that learns differential equations from noisy time-series data by treating the model equation as a soft constraint and using orthogonal distance regression with Bayesian model selection. If it works as claimed, it would make automatic equation discovery substantially more robust to noise and sparse sampling in chaotic systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stated limitation in Sec. 4 contradicts the headline claim: the soft-constraint loss (Eq. 6) assumes η is i.i.d. Gaussian with constant variance, but decaying signals violate this and ODR-BINDy is conceded to fail on them, so 'consistently outperforms all existing methods' is unsupported.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing point: the i.i.d. Gaussian model-error hypothesis in Eq. (6) is what makes the soft constraint tractable, and Section 4 already demonstrates a setting where it fails. That failure is not a peripheral caveat; it sits in a standard dynamical regime (decaying oscillations) and directly undercuts the abstract's claim that ODR-BINDy 'consistently outperforms all existing methods.' The paper deserves credit for reporting the limitation and for making code and data available, and the mathematical core appears coherent. The conditional verdict remains appropriate: the method may well be strong on sustained periodic and chaotic signals, but the headline claim must be scoped to those regimes and backed by a fair comparison on decaying signals before it can be accepted as stated. This stress-test therefore reinforces the reader's conditional verdict rather than moving it to rejection, because the limitation is explicit and the failure mode is understood.","tokens_in":20898,"tokens_out":6642,"duration_ms":70057,"concrete_test":"Reproduce the decaying-signal experiment of Fig. 6 on a damped cubic oscillator and on a linear damped oscillator ẍ + 2γω₀ẋ + ω₀²x = 0, using the same Δt = 0.05, σx = 0.084, a cubic-polynomial library, and per-package tuned hyperparameters as in Table A1 for ODR-BINDy, modified-SINDy, DySMHO, and Ensemble-SINDy across at least 20 noise seeds. If ODR-BINDy fails to recover the correct sparsity while any baseline succeeds, the 'consistently outperforms all existing methods' claim is refuted; if all methods fail, the claim should be explicitly restricted to non-decaying signals.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing assumption for the central claim is the stochastic-relaxation model in Eq. (6): the discretisation/model residual η = L∂t X − LI Θ(X)Ξ is treated as i.i.d. Gaussian noise with a single variance σ∂t. On that assumption, the soft constraint lets the denoised trajectory follow data beyond the Lyapunov time, which is the paper's claimed advantage over hard-constraint methods. Section 4 and Fig. 6 concede that this assumption is already violated for a damped cubic oscillator: the truncation error decays with the signal and is correlated with it, and ODR-BINDy then selects an oversimplified model. Because the success-rate experiments in §3 cover only sustained oscillatory/chaotic signals (Van der Pol, Lorenz63, Rössler), the paper's abstract-level claim of consistent superiority over all existing methods is not supported in a common dynamical regime, and the conclusion overstates what the evidence shows. The appendix comparison to SIDDs is also a reading of the original figure rather than a direct run, and DAHSI is not in the main comparison; these reinforce that 'all existing methods' is not established. The concern is not that the method is internally inconsistent—the derivation is coherent—but that the scope of the empirical claim is broader than the tested and conceded regimes.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces ODR-BINDy, a Bayesian sparse model-discovery method that treats the discretized governing equation as a soft constraint through orthogonal distance regression. The method simultaneously denoises the trajectory and estimates parameters, with Bayesian evidence used for sparse model selection. The authors compare ODR-BINDy with Ensemble-SINDy, modified-SINDy, DySMHO, DAHSI, and SIDDs on the Van der Pol, Lorenz63, and Rössler systems, reporting higher success rates in recovering the correct model at high noise levels and short data lengths, and they attribute the improvement to stochastic relaxation of the model equation. Section 4 acknowledges a significant limitation for decaying signals, where the i.i.d. Gaussian model-error assumption is violated.","tokens_in":21227,"tokens_out":6038,"duration_ms":58750,"significance":"If the central claims are supported, ODR-BINDy would be a valuable contribution: it addresses error-in-variable issues in SINDy, provides a principled Bayesian approach to sparsity, and enables assimilation of trajectories longer than the Lyapunov time by relaxing the model constraint. The paper is transparent about its limitations and provides code and data links. However, the headline claim of consistent superiority over all existing methods is not supported by the evidence as presented: the benchmark is explicitly not fair, DAHSI and SIDDs are not directly compared in the main experiments, and the method is conceded to fail on decaying signals. The core derivation is internally coherent, so the result is a defensible method for sustained oscillatory/chaotic signals, but the scope of the empirical claims must be narrowed or the missing evidence supplied.","major_comments":[{"comment":"The manuscript itself concedes in Section 4 and Fig. 6 that for a damped cubic oscillator ODR-BINDy fails to recover the correct model because the time-invariant σ∂t model of the discretization error is violated. This directly contradicts the abstract's claim that ODR-BINDy 'consistently outperforms all existing methods in recovering the correct model from sparse and noisy datasets.' The success-rate experiments in Section 3 cover only sustained oscillatory/chaotic signals, so the general claim is not supported by the evidence. The authors should either restrict the claim to the tested regime or provide an extension that handles time-varying, signal-correlated model error.","section":"§4, Fig. 6"},{"comment":"The performance comparison is not a fair benchmark, as the authors acknowledge in Section 3. DySMHO is given a ten-times higher sampling rate (Δt=0.001 vs 0.01); Ensemble-SINDy and modified-SINDy are run without the constant term in the Lorenz63 library; DAHSI is excluded from the main results; SIDDs is not run but represented by a rough estimate from Fig. 5.4 of [18]; and the number of runs differs across methods (64, 20, 64, 64). Given these asymmetries, the conclusion that ODR-BINDy is 'superior... compared to all current SINDy variants' cannot be drawn from the presented evidence. A fair comparison with matched data, libraries, and run counts, or a clearly scoped claim, is required.","section":"§3, Table A1"},{"comment":"The central hyperparameter σ∂t, and to a lesser extent σp, is chosen per system in Table A1 (σ∂t = 1e-2, 1e-3, 5e-3; σp = 10, 100, 50) rather than independently measured or selected by the evidence. Section 2.4 itself advises optimizing σ∂t to maximize evidence. Because σ∂t controls the strength of the stochastic relaxation that underlies the method's advantage, the reported success rates may partly reflect favorable hyperparameter choices. The paper should report sensitivity to σ∂t or demonstrate that evidence-based tuning recovers the same performance.","section":"§2.4, Table A1"},{"comment":"The i.i.d. Gaussian assumption for η is load-bearing for the stochastic relaxation argument. Section 4 shows it is violated for decaying signals, and for chaotic systems the discretization error is deterministic and correlated in time. The paper should more explicitly frame Eq. (6) as an approximate noise model, state the conditions under which it is adequate, and avoid presenting the method as robust in regimes where the assumption is known to fail.","section":"§2.1, Eq. (6)"}],"minor_comments":[{"comment":"There are numerous typos, including 'DySHMO' in §2.3.1, 'Orthongal' in §2.2, 'indepedently' in the Introduction, 'existance' in §3, 'Hyperparmeter' in §A.1, 'approxmation' in §A.5, and 'wtih' in §A.6.","section":"General"},{"comment":"The SIDDs curve is described as 'a rough estimate from fig 5.4 of [18]'; this should be stated directly in the figure caption, and the curve should not be used for quantitative comparisons.","section":"Fig. A4"},{"comment":"Table A1 would benefit from reporting the number of runs per method and error bars; currently modified-SINDy has only 20 runs, which limits the reliability of the success-rate differences.","section":"Table A1"},{"comment":"In Eq. (2), σx^2 is introduced as an Ñ×D matrix, but in Eq. (6) it is used as a scalar; please clarify the notational convention.","section":"Eq. (2) and Eq. (6)"},{"comment":"The reference list is incomplete: several entries lack article titles or page numbers (e.g., [1], [3], [17], [18]), which hinders reproducibility of the literature review.","section":"References"},{"comment":"The phrase 'consistently outperforms all existing methods' appears in both the abstract and the conclusion, but it is not supported by the scope of the experiments; it should be replaced by a scope-limited statement consistent with Sections 3 and 4.","section":"Abstract and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a genuinely interesting method and the core derivation is coherent, so I do not recommend rejection. However, the promotional claims are considerably broader than the evidence, and the benchmark asymmetries are serious enough that the central comparative claim needs major reworking. The authors have been unusually candid about limitations, which makes the revision tractable: they can either run a fairer comparison and address the decaying-signal case, or clearly delimit the scope of the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something real. ODR-BINDy's specific combination—orthogonal distance regression, stochastic relaxation of the discretized equation, and Bayesian evidence for model selection—is new, and the empirical results on sustained oscillatory and chaotic systems are plausible. The Lorenz63 recovery at 30% noise is a genuine result, and the derivations of the loss and Laplace-approximated evidence are coherent. Credit where due: the author ships code and data, acknowledges that the loss is equivalent to DAHSI's, and, unusually for this literature, states a limitation up front in Section 4. Decaying signals break the constant-variance iid Gaussian model-error assumption, and the paper concedes that ODR-BINDy then selects an oversimplified model. That is honest and important.\n\nThe soft spots are proportionate but real. The abstract's \"consistently outperforms all existing methods\" is not supported by the evidence in the paper. The benchmark is explicitly not fair: DySMHO gets ten times the sampling rate, modified-SINDy and Ensemble-SINDy lose the constant term in the Lorenz library, DAHSI only appears in an appendix under a generously defined success metric, and the SIDDs comparison is digitized from a figure in another paper. The hyperparameters σ∂t and σp are set per system in Table A1, not independently measured, and the text says hyperparameter optimization is advisable; some of the reported advantage may therefore come from favorable tuning. These issues weaken the comparative claim, but they do not invalidate the method. The stochastic-relaxation idea is the core contribution, and the failure on decaying signals is a scope limitation, not a fatal flaw.\n\nWho should read this: anyone working on sparse model discovery, data assimilation, or SINDy variants. The paper gives a useful method with a clearly stated regime of validity—sustained oscillatory or chaotic signals with noise up to roughly 30–40%. It deserves a serious referee, but the referee should require the claims to be scaled back and the benchmarks to be made fairer—ideally including DAHSI in the main runs and reporting results on decaying signals as a known failure mode. I would send it to peer review, with the expectation of major revision rather than rejection.","headline":"A solid within-subfield advance that pairs ODR with Bayesian evidence for SINDy, but the headline claim of consistent superiority over all existing methods is undercut by the paper's own limitations section and by a benchmark that is explicitly stacked in its favor.","tokens_in":21709,"tokens_out":1882,"would_cite":true,"duration_ms":22262,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F15","62J02"],"pacs":[],"model":"deepseek-v4-flash","headline":"ODR-BINDy treats the model equation as a soft constraint and recovers governing equations from sparse, noisy data, including Lorenz63 with up to 30% noise.","keywords":["SINDy","model discovery","error-in-variable","data assimilation","chaos","Bayesian evidence","orthogonal distance regression","stochastic relaxation"],"falsifier":"Run ODR-BINDy on a known damped cubic oscillator with 10% measurement noise and compute the residual $\\eta = L_{\\partial t} X - L_I \\Theta(X)\\Xi$ in sliding time windows; if its variance is not flat but decays and correlates with the state, the constant-$\\sigma_{\\partial t}^2$ assumption is violated, and the paper's own example predicts the method will return an oversimplified equation.","tokens_in":2005,"feed_emoji":"📈","tokens_out":5445,"duration_ms":135267,"temperature":0.7,"pith_summary":"This paper argues that standard SINDy-style model discovery is limited not by the noise itself but by how the model equation is imposed during fitting. Existing nonlinear optimisation methods such as modified-SINDy and DySMHO enforce the discretised equation as a hard constraint, which prevents assimilation of chaotic data for longer than the Lyapunov time. ODR-BINDy treats the equation as a soft constraint by modelling the discretisation residual as Gaussian noise, so the denoised trajectory can stay close to the data over arbitrarily long records. It then uses Bayesian evidence to select the sparse model. In Lorenz63, Rössler, and Van der Pol benchmarks, the paper reports that ODR-BINDy consistently recovers the correct equation where comparable methods fail, reliably recovering Lorenz63 from data with up to 30% measurement noise.","feed_headline":"Equation as soft target recovers chaos from 30% noise","feed_subtitle":"ODR-BINDy treats the model equation as a soft constraint, unlocking long noisy chaotic time series.","key_machinery":"The central object is the soft-constrained loss of Eq. (6): $L(X,\\Xi) = \\frac{1}{2\\sigma_x^2}\\|\\hat{X} - X\\|_F^2 + \\frac{1}{2\\sigma_{\\partial t}^2}\\|L_{\\partial t} X - L_I \\Theta(X)\\Xi\\|_F^2 + \\frac{1}{2\\sigma_p^2}\\|\\Xi\\|_F^2$. Each prefactor is the inverse of a noise or prior variance: $\\sigma_x^2$ for measurement noise, $\\sigma_{\\partial t}^2$ for the equation residual, and $\\sigma_p^2$ for the parameter prior. Optimising this loss is an orthogonal distance regression because both the state and the derivative residual are penalised in their own noise metrics. The same loss is equivalent to weak-constraint 4D-Var data assimilation and to a physics-informed neural network loss. Model selection is carried out by the Bayesian evidence computed with a Laplace approximation and a Gauss-Newton Hessian, which drives the greedy deletion of dictionary terms.","core_discovery":"The paper's central claim is that stochastic relaxation of the model equation is the key to overcoming the error-in-variable problem in data-driven model discovery. Where earlier methods enforce $L_{\\partial t} X - L_I \\Theta(X)\\Xi = 0$ exactly, ODR-BINDy writes the residual as an independent Gaussian noise term $\\eta \\sim \\mathcal{N}(0, \\sigma_{\\partial t}^2)$, so the equation becomes a soft target in the loss. This change allows the fitted trajectory to follow the observed data beyond the Lyapunov time, even for chaotic systems, and thereby lets the algorithm exploit the full information in long noisy time series. The same optimisation solves for the denoised states and the parameters, and a greedy Bayesian-evidence search provides the sparse model selection. The paper demonstrates on three benchmark systems that this combination recovers the true sparsity pattern more reliably and uses fewer data points than Ensemble-SINDy, modified-SINDy, DySMHO, and DAHSI.","pith_inferences":["A concrete extension the paper does not explore is replacing the constant $\\sigma_{\\partial t}^2$ with a state-dependent variance; its own damped-oscillator example suggests that this should rescue the method for decaying signals.","Because the loss is built from an explicit noise model, the same framework should accommodate non-Gaussian measurement noise by substituting the corresponding energy function, a change the paper notes but does not implement.","The demonstrated soft-constraint assimilation is the same mechanism used in weak-constraint 4D-Var, so the method could plausibly serve as a parameter-estimation tool for operational data assimilation once the sparse model is identified."],"forward_implications":["A single optimisation can assimilate chaotic time series much longer than the Lyapunov time, so the previous practice of limiting the assimilation horizon to the Lyapunov time is no longer required.","Bayesian evidence can replace coefficient-magnitude thresholding in model discovery, reducing the risk of discarding small but physically significant terms.","The loss structure is independent of the time-discretisation scheme, so finite differences, weak-form collocation, or other linear derivative operators can be swapped in without changing the algorithm.","The algorithm's computational complexity is polynomial in the library size, and it converges in under ninety seconds on a laptop for the Lorenz63 example, making the approach practical for routine model discovery.","The documented failure on decaying signals shows that the next improvement is to model the equation residual as time-varying and signal-correlated rather than i.i.d. Gaussian."],"supporting_citations":[{"why":"Defines the sparse identification of nonlinear dynamics (SINDy) problem that ODR-BINDy extends.","marker":"[1]"},{"why":"Supplies the ensemble linear-regression baseline and the bootstrap strategy used for initial parameter guesses.","marker":"[5]"},{"why":"Provides the Bayesian evidence and sparsity-inducing prior framework that ODR-BINDy inherits.","marker":"[6]"},{"why":"Modified-SINDy is a primary comparison baseline that enforces the equation as a hard constraint via ODE integration.","marker":"[15]"},{"why":"DAHSI is the existing method that also uses a soft constraint and shares the loss structure with ODR-BINDy.","marker":"[16]"},{"why":"DySMHO is a moving-horizon hard-constraint baseline that limits assimilation to less than the Lyapunov time.","marker":"[17]"},{"why":"SIDDs is a discretise-then-optimise hard-constraint baseline whose reported Lorenz63 results ODR-BINDy is compared against.","marker":"[18]"},{"why":"Supplies the orthogonal distance regression machinery used for the nonlinear least-squares optimisation.","marker":"[19]"},{"why":"Provides the weak-constraint 4D-Var background that frames the soft-constraint loss and the Lyapunov-time issue.","marker":"[21]"}],"fun_headline_variants":["Soft constraint lets ODR-BINDy see chaos in 30% noise","Equation as soft target beats SINDy on noisy chaos","ODR-BINDy: soft model constraint unlocks long noisy time series","Beyond Lyapunov time: soft equation recovers chaotic laws","Treating the model as a soft target fixes error-in-variable"],"cache_read_input_tokens":23808,"weakest_assumption_plain":"The load-bearing assumption is that the residual between the discretised model equation and the data is small, independent, identically distributed Gaussian noise with constant variance; when that residual grows or correlates with the signal, as the paper shows for decaying signals, the method selects an oversimplified model.","fun_headline_variants_meta":{"raw":{"variants":["Soft constraint lets ODR-BINDy see chaos in 30% noise","Equation as soft target beats SINDy on noisy chaos","ODR-BINDy: soft model constraint unlocks long noisy time series","Beyond Lyapunov time: soft equation recovers chaotic laws","Treating the model as a soft target fixes error-in-variable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00018,"raw_usage":{"total_tokens":1344,"prompt_tokens":1029,"completion_tokens":315,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":224}},"tokens_in":645,"tokens_out":315,"duration_ms":3420,"temperature":1.0,"reasoning_tokens":224,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:44:07.447952+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ODR-BINDy on a known damped cubic oscillator with 10% measurement noise and compute the residual $\\eta = L_{\\partial t} X - L_I \\Theta(X)\\Xi$ in sliding time windows; if its variance is not flat but decays and correlates with the state, the constant-$\\sigma_{\\partial t}^2$ assumption is violated, and the paper's own example predicts the method will return an oversimplified equation.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DAHSI is the existing method that also uses a soft constraint and shares the loss structure with ODR-BINDy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SIDDs is a discretise-then-optimise hard-constraint baseline whose reported Lorenz63 results ODR-BINDy is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the orthogonal distance regression machinery used for the nonlinear least-squares optimisation."}],"review_version":1}