{"id":"37609cbe-9017-46dc-9c0e-9236066a4723","arxiv_id":"2507.03206","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A new algorithm picks the support radius of weak-form test functions by locating the changepoint of an estimated integration error curve, and this radius lands near the parameter-error minimum in most tested cases.","lead":"This paper introduces a data-driven rule for choosing how wide the smoothing functions should be when identifying differential equations from noisy data. The rule estimates numerical integration error from the data alone and picks the width where that error stops shrinking, which often matches the width giving the most accurate parameter estimates.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The operational premise of the method — that the estimated integration-error changepoint brc(U) tracks the parameter-error-minimizing radius — is not just unproved but appears to fail in the paper's own high-noise logistic-OLS and FitzHugh-Nagumo examples, so the central claim of automatic radius…","rationale":"The reader's weakest assumption identifies the same two-link chain: Conjecture 2.2 plus the noisy-data substitution in Remark 2.9. My reading agrees, and the paper's own logistic-OLS and FitzHugh-Nagumo panels are preliminary evidence that the chain breaks at high noise. I would not move the verdict away from CONDITIONAL: the integration-error expansion in Lemma 2.3 and Proposition 2.4 is a genuine technical contribution, and the method does select useful radii in several settings (Duffing, Lorenz, and IRLS-corrected logistic growth). But the central claim as stated in the abstract is stronger than the evidence, and the proposed check would settle whether the advertised regime includes cases where the method fails. This is a correctness-risk concern, not a stylistic one, and it is addressable by adding an explicit validity condition or by softening the claim.","tokens_in":24286,"tokens_out":6186,"duration_ms":82200,"concrete_test":"For the logistic growth model under WENDy-OLS with 20% noise, M=500, p=16, run 1000 independent noise realizations. For each realization, compute brc via Algorithm 2.1 and compute the oracle radius r* = argmin_r E[||w0(r)-w*||^2] using the ground-truth parameter and averaging over realizations. Report the distribution of brc-r* and the fraction of realizations where brc lies outside the set of radii whose average relative error is within 20% of the minimum. If that fraction is substantial (e.g., >5%), then Conjecture 2.2 and Remark 2.9 cannot both hold in an advertised regime, and the paper should either restrict the claim or provide a validity condition before presenting the method as a general data-driven radius-selection rule.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim requires two unproved links: Conjecture 2.2, that the true integration-error changepoint rc equals argmin_r E[||w_hat(r)-w*||^2], and Remark 2.9, that replacing u* by noisy U in the estimator beint does not shift the changepoint. Conjecture 2.2 is stated under E[r(u,w)] = 0, but the residual decomposition (2.1) shows that parameter error is driven by eTheta, r0, eint, and noise; for WENDy-OLS the regression matrix is a nonlinear function of noisy data, so the residual is generally not zero-mean. The paper's own Figures 4 and 7 show that for logistic growth and FitzHugh-Nagumo under OLS at noise ratios above roughly 5%, brc falls outside the low-parameter-error region, while the Duffing and Lorenz cases work better. Remark 2.9 also has a concrete failure mechanism: I(U,S,mu) requires endpoint derivatives up to order 2S estimated by finite differences of noisy data; for S=1 and dt=0.02, the second-derivative finite-difference stencil amplifies noise by a factor of order 1/dt^2, so the estimated changepoint can shift relative to the true eint changepoint. Thus the selection rule is reliable only in a regime the paper does not characterize, and the abstract's claim that selected supports 'consistently align' with minimal parameter error overstates the evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-driven procedure for selecting the support radius of compactly supported test functions in weak-form parameter estimation (WENDy/WSINDy). The central idea is to approximate the numerical integration error induced by quadrature, using an Euler–Maclaurin expansion that is independent of the true parameters, and then to choose the radius where the estimated integration-error curve changes from exponential decay to stagnation. This radius is conjectured to coincide with the radius minimizing expected parameter error (Conjecture 2.2). The authors derive an asymptotic expansion of the integration error (Lemma 2.3), construct an efficient FFT-based estimator (Proposition 2.4), propose Algorithm 2.1, give practical guidance for piecewise polynomial test functions, and validate the approach on logistic growth, Duffing, Lorenz, and FitzHugh–Nagumo systems under WENDy-OLS and WENDy-IRLS, also comparing with their earlier Multi-scale-Global method.","tokens_in":24649,"tokens_out":3475,"duration_ms":46613,"significance":"If the central conjecture were established or its validity regime were clearly characterized, the method would give a practical answer to a long-standing hyperparameter question in weak-form scientific machine learning: how to select the test-function support from noisy data alone. The derivation of the integration-error estimator (Lemma 2.3 and Proposition 2.4) is a genuine contribution: the estimator is explicitly independent of the unknown parameters, computable via FFTs, and supported by extensive numerical experiments across several ODE systems, noise levels, and resolutions. The paper is also honest in stating that Conjecture 2.2 is unproved and left for future work. However, the paper's own numerical results contain configurations in which the selected radius does not align with the low-parameter-error region, so the advertised claim of consistent alignment is too strong. The main value of the paper is thus as a promising, well-engineered heuristic with a rigorous error-expansion foundation, rather than as a fully validated automatic selection rule.","major_comments":[{"comment":"Conjecture 2.2 is the load-bearing assertion of the paper: it identifies the changepoint of the monotone decrease of log(||eint(r)||) with the minimizer of expected parameter error. The assumptions stated before the conjecture are not satisfied by the WENDy-OLS problem in general. In the residual decomposition (2.1), the parameter error is driven by eTheta, r0, eint, and noise; when the regression matrix G is a nonlinear function of noisy data, the residual is not zero-mean and minimizing eint alone need not minimize E[||w_hat - w*||^2]. The paper does not provide a proof or a precise characterization of when the conjecture holds, nor does it bound the discrepancy. Since Algorithm 2.1 and the abstract's central claim rely on this conjecture, the manuscript needs either a proof under stated conditions, a quantitative bound on the deviation, or a clearly limited validity claim with supporting analysis.","section":"Section 2.2, Conjecture 2.2"},{"comment":"The transition from true trajectory u* to noisy measurements U inside the error estimator is asserted without proof. The estimator be(U,S,mu) requires finite-difference approximations of derivatives up to order 2S at the endpoints. For S=1 and dt=0.02, a second-derivative finite-difference stencil amplifies noise by a factor of order 1/dt^2, which can shift the estimated changepoint relative to the true integration-error changepoint. The remark claims this is 'substantially mitigated' because only boundary points are used, but no variance or bias analysis is provided. The paper needs either a quantitative analysis of this noise-induced shift or a numerical study that explicitly tests the sensitivity of brc to the finite-difference stencil order and dt across the noise regimes used in the experiments.","section":"Remark 2.9"},{"comment":"The paper's own results contradict the abstract's statement that selected supports 'consistently align with regions of minimal parameter estimation error'. For logistic growth under WENDy-OLS (Figure 4, left) and for FitzHugh–Nagumo under both OLS and IRLS (Figure 7), the dashed brc line falls outside the low-parameter-error region at noise ratios above roughly 5%, with the discrepancy growing with noise. The text in Section 4.1 acknowledges these cases but the abstract and Section 5 do not carry the caveat. The central claim must be reworded to state the observed validity regime, and the method should be accompanied by guidance (or a diagnostic) for detecting when the selected radius is unreliable.","section":"Section 4.1, Figures 4 and 7"},{"comment":"The changepoint of the estimated error curve is computed by Algorithm B.1 on log(be(r)), but the error metric E(k) in Algorithm B.1 divides by the function values ym. For the integration-error curve, which decays to near machine precision and can cross zero or become negative in finite-precision arithmetic, the relative-error formulation can be unstable and the changepoint may depend sensitively on the lower tail of the curve. The paper does not discuss this numerical issue or justify the choice of relative error over an absolute error or a fit in log-space. A brief analysis or a robustness check (e.g., varying the changepoint algorithm's tolerance) would strengthen the reproducibility of brc.","section":"Section 2.3 and Definition 2.6"}],"minor_comments":[{"comment":"There is a typo 'Θ(U) ∈∈ R' in the definition of the discrete feature matrix; the double '∈' should be a single symbol.","section":"Page 4, equation (1.3) context"},{"comment":"The sentence 'We choose brc (defined as the change point of be(r)) as the reference test function radius. and construct the test function set K' contains a period before 'and'; this should be corrected.","section":"Page 6, Conjecture 2.2 and surrounding text"},{"comment":"In the paragraph beginning 'However, the MG approach can sometimes offer runtime advantages', the reference 'Lorenz system under WENDy-IRLS(Figure 11)' should be 'Figure 13'; the Lorenz results are shown in Figure 13, not Figure 11.","section":"Section 4.2, Figure 11 reference"},{"comment":"The text says 'As ∆t increases, higher values of p are required' but the discussion of the scaling is for the integration error only; a sentence clarifying that this is independent of noise effects would help avoid misinterpretation.","section":"Section 3.1, paragraph after Figure 3"},{"comment":"The figure caption says 'for the Lorenz system (left) and logistic growth model (right)' but the main text does not describe the plotted curves beyond the legend; adding a sentence explaining the colored lines and the takeaway about fM would improve readability.","section":"Appendix I, Figure 15"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a numerical analysis / scientific machine learning journal, and the derivation of the error estimator is a solid contribution. However, the central conjecture is unproved and the numerical validation shows clear failure cases; the authors should be pushed to either prove the conjecture under conditions, provide a failure diagnostic, or substantially soften the claims. The heavy citation of the authors' own prior work is not by itself a problem, but the novelty relative to WENDy and WSINDy should be made crisper in the revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this if you work on weak-form system identification. The method is real: the Euler-Maclaurin integration-error expansion in Lemma 2.3 and the FFT-based estimator in Proposition 2.4 are a genuine step forward, and Algorithm 2.1 gives a concrete heuristic for choosing the test-function radius without knowing the parameters. The SL-vs-MG runtime comparison in Section 4.2 shows a real speedup, especially under IRLS. The authors are also honest that Conjecture 2.2 is unproved and that a formal justification is left for future work.\n\nThe soft spots are in proportion. The load-bearing claim, that the changepoint of the estimated integration error equals the radius minimizing parameter error, is a conjecture, not a theorem. The residual decomposition in (2.1) shows why it should not hold exactly: parameter error is driven by noise and nonlinearity bias as well as integration error, so the zero-mean residual assumption in Conjecture 2.2 is violated in the OLS case. The paper's own Figures 4 and 7 show the alignment fails for logistic growth and FitzHugh-Nagumo under OLS at noise ratios above about five percent. That is not a minor caveat; it is the main claim failing in a nontrivial subset of the experiments. Remark 2.9, replacing the true trajectory with noisy data in the estimator, is plausible for modest noise but is not analyzed; endpoint finite differences for S=1 already require second derivatives, and noise amplification is a real concern. No code or data are provided, which makes it hard to check whether the changepoint detection is sensitive to the fitting tolerance in Algorithm B.1.\n\nNone of this kills the paper. The estimator is derived independently of the true parameters, and the method works well for Duffing and Lorenz across noise levels. The paper would be stronger if it weakened the abstract's \"consistently align\" claim, reported the critical radius with error bars across trials, and shipped code. As it stands, it deserves serious refereeing, and the limitations are mostly acknowledged even though the abstract overreaches.","headline":"Useful heuristic with real technical content, but the central conjecture is unproved and the paper's own high-noise examples show the radius-selection claim failing.","tokens_in":25163,"tokens_out":3159,"would_cite":true,"duration_ms":33865,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["46F05","62FXX","62JXX","65L09","68Q32"],"pacs":[],"model":"deepseek-v4-flash","headline":"In weak-form system identification, the radius of the test functions that minimizes numerical integration error also minimizes parameter-estimation error, and it can be identified from noisy data alone.","keywords":["weak-form parameter estimation","test function construction","integration error","changepoint detection","Euler-Maclaurin expansion","compactly supported test functions","WENDy","noise robustness"],"falsifier":"Run WENDy-OLS on logistic-growth data with high noise (e.g., 40 percent) and compare the radius selected by Algorithm 2.1 to the radius giving the smallest median parameter error over many realizations; the paper's own Figure 4 shows these already diverge there, so a single clean reproduction of that divergence under the paper's stated conditions would settle Conjecture 2.2.","tokens_in":24055,"feed_emoji":"📈","tokens_out":7657,"duration_ms":85567,"temperature":0.7,"pith_summary":"This paper tries to settle how to choose the width of the compactly supported test functions in weak-form scientific machine learning without tuning by hand. It claims that the radius at which the quadrature-error curve $\\log\\|e_{\\mathrm{int}}(r)\\|$ stops decreasing is the same radius that minimizes expected parameter-estimation error, and that this changepoint can be estimated from noisy data alone. Concretely, it builds an estimator of the integration error from an Euler–Maclaurin expansion whose endpoint terms are computed by finite differences, then reads off the changepoint with a two-line fit. If the claim is right, WENDy-style parameter estimation no longer needs a manual support-size choice, and the resulting test functions can be built faster than with the earlier multiscale-global basis.","feed_headline":"A bend in the error curve picks the best test-function radius","feed_subtitle":"A data-driven changepoint rule removes the manual support-size choice in weak-form parameter estimation.","key_machinery":"The load-bearing object is the surrogate integration-error curve $\\hat e(U,r)$ from Proposition 2.4, together with Conjecture 2.2 that its changepoint equals the parameter-error-minimizing radius. The estimator is built from Lemma 2.3's asymptotic expansion of the trapezoidal error, $\\hat e_{\\mathrm{int}}=(1/\\sqrt{T})PF\\hat\\Psi I$, where $P$ selects rows at test-function centers, $F$ is the DFT matrix, $\\hat\\Psi$ is the diagonal matrix of Fourier coefficients of the reference function, and $I$ contains endpoint values and finite-difference endpoint derivatives from the Euler–Maclaurin series truncated at order $S$. For the piecewise polynomial family, Proposition 3.1 gives those Fourier coefficients in closed form through Bessel functions, so the sweep over radii is cheap. The changepoint is located by Algorithm B.1, which fits two line segments to the log curve and picks the break minimizing relative fitting error; this mechanism converts an unobservable quantity (expected parameter error) into an observable curve feature.","core_discovery":"The paper's central claim is Conjecture 2.2: for the WENDy estimator with test-function radius $r$, the expected parameter error $\\mathbb{E}[\\|\\hat w(r)-w^\\star\\|^2]$ is minimized at the radius $r_c$ where the monotone decrease of $\\log\\|e_{\\mathrm{int}}(r)\\|$ stagnates. The supporting discovery is that this stagnation radius can be found without knowing $w^\\star$ or the clean trajectory $u^\\star$: Proposition 2.4 gives an estimator $\\hat e_{\\mathrm{int}}$ based on a truncated Euler–Maclaurin expansion of the trapezoidal-rule error, the closed-form Fourier coefficients of the piecewise-polynomial reference function $\\psi(t;r,p)=C(r-t)^p(r+t)^p$, and boundary derivatives computed by finite differences. Algorithm 2.1 sweeps $r$, applies a two-segment changepoint detector to $\\log \\hat e(U,r)$, and returns that radius as the construction radius. The numerical sections present this as consistently landing in low parameter-error regions, with the qualification that under WENDy-OLS the alignment degrades for strong nonlinearities at high noise.","pith_inferences":["An unstated consequence is that the same changepoint rule could select support sizes for weak-form PDE learning and model discovery, where manual scale choice is a known practical bottleneck.","A testable extension is to iterate the radius selection as parameter estimates improve, since the paper fixes the estimator while sweeping radii but WENDy-IRLS updates the covariance during iterations.","The paper's own high-noise OLS results suggest the changepoint alignment depends on the regression bias being small; this predicts that the radius rule is most reliable when paired with a covariance-corrected or debiased estimator.","Because the estimator needs only endpoint derivatives, it may transfer to non-uniform or streaming time grids, though the paper only treats uniform grids."],"forward_implications":["Manual tuning of the test-function support size is no longer needed: the critical radius is computed from the data by Algorithm 2.1, requiring neither the true parameters nor noise-free data.","Across the tested systems the selected radius aligns with low parameter error under both WENDy-OLS and WENDy-IRLS in most settings, with the strongest misalignments appearing under OLS for logistic growth and FitzHugh–Nagumo at high noise.","The single-scale-local construction is consistently faster than the multiscale-global construction, with median IRLS runtime improvements of roughly 60–70 percent on several of the test systems.","For piecewise polynomial test functions, the paper gives a practical order guideline: choosing $p$ between 16 and 22 is sufficient for the tested resolutions, with integration error scaling at least as $O(\\Delta t^p)$ for even $p$."],"supporting_citations":[{"why":"Supplies the Euler–Maclaurin formula used to derive the asymptotic integration-error expansion in Lemma 2.3.","marker":"[1]"},{"why":"Introduces the WENDy-IRLS estimator and the Multi-scale-Global test-function construction that the single-scale-local method is compared against.","marker":"[3]"},{"why":"Provides the quadrature and Euler–Maclaurin error expansion used in the proof of Lemma 2.3.","marker":"[7]"},{"why":"Provides the weak-form foundation, the overlapping-test-function and FFT strategies, and the two-line changepoint detection algorithm (Algorithm B.1).","marker":"[23]"},{"why":"Establishes the analysis of how test-function regularity controls discretization error, motivating the order-selection guidance.","marker":"[24]"},{"why":"Introduces integration with compactly supported test functions to avoid differentiating data, the historical basis of the weak-form residual.","marker":"[44]"}],"fun_headline_variants":["Data-driven test radius: changepoint of integration error picks it","Weak-form learning: use error curve's bend to set test-function size","Find weak-form test radius from integration error stagnation without parameters","Automatic support size for test functions via error-change detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method works only if the radius where the estimated integration error stops decreasing is the same radius that minimizes parameter error, and only if noisy measurements can stand in for the true trajectory in finding that radius without moving it.","fun_headline_variants_meta":{"raw":{"variants":["Data-driven test radius: changepoint of integration error picks it","Weak-form learning: use error curve's bend to set test-function size","Find weak-form test radius from integration error stagnation without parameters","Automatic support size for test functions via error-change detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000507,"raw_usage":{"total_tokens":2480,"prompt_tokens":964,"completion_tokens":1516,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":1444}},"tokens_in":580,"tokens_out":1516,"duration_ms":13773,"temperature":1.0,"reasoning_tokens":1444,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:15:59.903637+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run WENDy-OLS on logistic-growth data with high noise (e.g., 40 percent) and compare the radius selected by Algorithm 2.1 to the radius giving the smallest median parameter error over many realizations; the paper's own Figure 4 shows these already diverge there, so a single clean reproduction of that divergence under the paper's stated conditions would settle Conjecture 2.2.","supporting_citations":[{"cited_title":"Atkinson, An Introduction to Numerical Analysis , Wiley, New York, NY, 2nd ed., 1989","cited_arxiv_id":null,"evidence_quote":"Supplies the Euler–Maclaurin formula used to derive the asymptotic integration-error expansion in Lemma 2.3."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes the analysis of how test-function regularity controls discretization error, motivating the order-selection guidance."}],"review_version":1}