{"id":"df6d6bec-485c-467a-840a-40fd8474e3a3","arxiv_id":"2507.21841","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A pipeline that fits noisy data with a smooth general-solution function, differentiates it with B-splines, and uses SVD to recover a constant-coefficient linear ODE.","lead":"This paper proposes a three-step technique that turns noisy measurements into an ordinary differential equation by fitting a smooth curve, computing its derivatives with splines, and reading off the equation from a derivative matrix. The method recovers known equations for a spring-mass system and for estrogen breakdown, and it targets scientists with messy experimental data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Section 2.3's SVD nullspace step is not identifiable when P exceeds the true order: the nullspace is multidimensional, so the last right-singular vector need not recover the true ODE coefficients.","rationale":"The reader's weakest_assumption (the method is limited to homogeneous linear constant-coefficient ODEs) is a scope restriction that the conclusion already acknowledges, not an internal flaw. The more load-bearing problem is the SVD nullspace identification in Section 2.3, which the reader's rationale flags but does not make the central concern. For the intended use of setting P above the true order, the nullspace is generically multidimensional, so the claimed uniqueness of C is not established and the abstract's accuracy/sparsity claims are not supported by the derivation. This does not prove the empirical spring-mass and EDC results are wrong; it means they cannot be attributed to the described algorithm without code or an identifiability analysis. A conditional verdict remains appropriate because the flaw is fixable (e.g., enforce a known order r, or select a sparse vector in the nullspace with a proper regularization argument), but the current manuscript needs that fix.","tokens_in":14741,"tokens_out":14547,"duration_ms":202134,"concrete_test":"Implement the complete pipeline on noiseless data from y' - 2y = 0 (or x'' + 2x' + x = 0) with P set to 7, using the authors' GA, B-spline, and SVD code. Compute the nullity of the gradient matrix and the recovered C. If the nullity is not 1, or if C is not proportional to the true coefficient vector (e.g., (2,-1,0)), or if C changes when the L columns selected for G~ are changed, then Eq. (26) does not identify the governing ODE.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central recovery claim rests on Eqs. (23)-(26), where C is taken as V(:,P+1), the last right-singular vector of the gradient matrix. This requires Γ=(P+1)-rank(G~) = 1. But when the user chooses P above the true order r, as the paper explicitly does ('a higher order ODE formulation with high degree of freedom'), the exact gradient matrix has rank r, so the nullspace has dimension P+1-r > 1. The last right-singular vector is then an arbitrary vector in that multidimensional nullspace, not the true coefficient vector. For example, for y=e^{2x} (true ODE y'-2y=0) and P=2, every column of G is proportional to (1,2,4)^T; the nullspace is two-dimensional and contains, e.g., (2,-1,0) and (4,0,-1), so SVD does not select a unique coefficient vector. The spline step does not fix this: Section 2.2 argues polynomial derivatives are linearly independent, which would make G full rank and G^T C=0 trivially solved only by C=0, while Section 2.3 needs rank deficiency. If instead L<P+1 columns are used, the null vector depends on the arbitrary column choice. The paper's own admission that the optimized GA coefficient set is not the ODE coefficient set compounds the problem: the approximate general solution satisfies the high-order trial ODE, and the subsequent SVD step has no mechanism to discard the spurious null directions. The reported sparse coefficients therefore appear to rely on the manual thresholding described in Figures 7 and 11 and Table 6 rather than on the claimed automatic sparsity.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a three-step method for discovering homogeneous linear constant-coefficient ordinary differential equations from noisy data. First, a genetic algorithm fits an approximate general solution of the assumed ODE family (exponential/sinusoidal eigenfunctions) to the data. Second, an adaptive B-spline approximates this fitted function to provide a polynomial form whose derivatives are used to build a gradient matrix. Third, singular value decomposition of a reduced gradient matrix yields a null vector that is taken as the ODE coefficient vector. The method is demonstrated on synthetic spring-mass data (overdamped, critically damped, underdamped) and on experimental photolytic degradation data for estrogenic chemicals. The paper claims high accuracy, noise robustness, and sparsity without regularization.","tokens_in":15071,"tokens_out":8843,"duration_ms":91139,"significance":"If the method worked as described, it would offer a useful alternative to SINDy-style approaches for a restricted but common class of ODEs (homogeneous, linear, constant-coefficient), with the advantage of not requiring a pre-specified basis library and of smoothing noisy data before differentiation. The numerical results on the spring-mass system and the EDC case show plausible coefficient recovery in some regimes. However, the significance is limited by the narrowness of the assumed ODE class, the failure to provide code or data, and several theoretical gaps in the derivation, which I detail below. The paper is honest about some limitations (e.g., variable-coefficient and higher-order systems), but the central mechanism for coefficient extraction is not rigorously justified as written.","major_comments":[{"comment":"The rank-deficiency justification for the gradient matrix is internally inconsistent. The text states that the spline S is a polynomial of order P with P+1 linearly independent basis functions (x^0,...,x^P). If that is the case, the gradient matrix G (with rows representing derivatives of orders 0 through P) has full row rank P+1, so the equation G^T C = 0 admits only the trivial solution C=0, and the paper's asserted nullspace does not exist. If instead the B-spline order ω=P gives piecewise polynomials of degree P-1, then the P-th derivative row is identically zero, making C_P unidentifiable and the nullspace multidimensional. The assumption Γ=1 is therefore unsupported; when the user chooses P above the true order, as the paper explicitly does, the last right-singular vector is an arbitrary vector in a multidimensional nullspace and need not recover the true ODE coefficients. The reduced matrix G̃ is never defined concretely, and its column selection is arbitrary, which further undermines the reproducibility of the claimed extraction.","section":"2.3, Eqs. (22)-(26)"},{"comment":"The eigenfunction basis in Eq. (6) is not the most general solution of a homogeneous linear constant-coefficient ODE when eigenvalues have multiplicity or come in complex-conjugate pairs. For a repeated complex root, one requires independent coefficients for x^j e^{βx} cos(γx) and x^j e^{βx} sin(γx) for each power j; Eq. (6) couples the cosine and sine terms with a single coefficient D_q. This restriction narrows the function class used in the approximate general solution and may bias the subsequent ODE discovery, especially for underdamped oscillatory systems with repeated roots.","section":"2.1, Eq. (6)"},{"comment":"The EDC case study is not a clean demonstration on real noisy data. The authors state that the experimental data contain too few points for the genetic algorithm, so they logarithmically transform the data, generate 1000 new samples by linear approximation, add noise of scale 0.001 to those synthetic samples, and then run the full pipeline. The recovered rate constants in Table 7 are computed from data that were themselves derived from the same experimental observations used to define the 'original' first-order rate constants; the comparison is therefore partly tautological. The claim of noise robustness in real experimental settings is accordingly overstated for this case.","section":"3.2, Tables 6 and 7"},{"comment":"The claim that the method 'promotes sparsity in the solution without using any regularization techniques' is contradicted by the manual thresholding applied throughout the results. In Figure 7, coefficients below 10^-4 are set to zero; in Table 6, terms below 10^-5 are neglected; in Figure 11, the threshold epsilon=10^-4 is applied. This is a hard-thresholding post-processing step, which is a form of sparsity enforcement. The SVD step alone does not produce sparse coefficient vectors, as the null space vector is generally dense.","section":"Abstract and Section 3.1, Figures 7 and 11"},{"comment":"The reported coefficient recovery for the noiseless overdamped case is not accurate: the predicted ODE (2.03 d²x/dt² + 1.03 dx/dt + 1.00 x = 0) differs from the original (2 d²x/dt² + 4 dx/dt + 1.00 x = 0) by a factor of about four in the damping coefficient, yet Section 3.1 reports R²=1.0 for the regression of mass and damping factor (Figure 5). The discrepancy suggests a possible identifiability issue in the overdamped regime that the paper does not explain, and it weakens the general claim of high-accuracy ODE recovery.","section":"Table 3"}],"minor_comments":[{"comment":"The index range 'q = 0, 2, · · ·, Q' appears to be a typo; it should likely read 'q = 0, 1, · · ·, Q' or 'q = 1, 2, · · ·, Q'.","section":"2.1"},{"comment":"The word 'suuitable' should be corrected to 'suitable'.","section":"2.2"},{"comment":"The SVD dimensions are incorrect: for G̃ ∈ R^{(P+1)×L}, the correct factorization has U ∈ R^{(P+1)×(P+1)}, Σ ∈ R^{(P+1)×L}, and V ∈ R^{L×L}. As written, the dimensions of U and Σ are swapped.","section":"2.3, Eq. (24)"},{"comment":"The statement that the last column of V corresponds to the 'least squares solution with minimum Euclidean norm' is not accurate. The minimum-norm solution to G̃^T C = 0 is the zero vector, while the last right-singular vector is the direction of the smallest singular value; the citation to Trefethen and Bau does not support the usage as stated.","section":"2.3"},{"comment":"The error metric used in the SINDy comparison is not defined; it is unclear whether the values correspond to mean squared error of the coefficients, of the trajectory, or of some other quantity. This makes the comparison difficult to interpret.","section":"3.1, Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper's scope is considerably narrower than the abstract suggests: the method assumes a homogeneous linear constant-coefficient ODE, and the approximate general solution is constructed from exactly that solution family, so the 'discovery' is essentially parameter estimation within a pre-chosen function class. The central technical derivation in Section 2.3 has a rank-deficiency flaw that invalidates the coefficient extraction as described; the numerical results might still be reproducible with a corrected implementation, but the manuscript as written does not provide enough detail. The EDC case study is weakened by the synthetic interpolation and noise addition. If the authors can fix the derivation, provide code, and re-frame the claims to match the actual scope, the paper could become acceptable, but in its current form I would not recommend acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes a new pipeline for discovering homogeneous linear constant-coefficient ODEs from noisy data: fit a general-solution form with a genetic algorithm, spline it, compute derivative rows, and read off coefficients as the last right-singular vector of the gradient matrix. That two-step idea is genuinely new relative to SINDy, weak-SINDy, and Bayesian spline learning, and the spring-mass results at the correct ODE order are accurate. The authors are also honest that the GA's coefficient set is not the governing equation. The writing is mostly clear.\n\nThe problem is the load-bearing math in Section 2.3. The paper says the spline's derivative rows are linearly independent, which would make the full gradient matrix full rank and G^T C = 0 solvable only by C = 0. To get a non-trivial nullspace, they subsample columns to get a rank-deficient matrix, but the nullspace of that reduced matrix is not unique and depends on which columns you pick. When the chosen order P exceeds the true order, which the paper explicitly does (P=7 for first-order EDC systems), the nullspace has dimension P+1-r > 1, and the last right-singular vector is an arbitrary vector in that subspace, not the true coefficients. The paper acknowledges Gamma>1 and says it takes the last column, but that does not select the true vector. So the central recovery claim is unsupported.\n\nThe EDC case is also weaker than it looks. The experimental data are sparse, so the authors log-transform, add 1000 new points by linear interpolation in log space, then invert the transform and add noise. That synthetic augmentation essentially enforces exponential decay, so the recovered first-order rate constants are fitted to the augmented data, not discovered from independent evidence. The reported sparsity comes partly from manual thresholding: Table 6 keeps second-order terms like 0.0025, which are not negligible relative to the 10^-4 threshold used for visualization. And the SINDy comparison is incomplete; for noise 0.1, critical damped, SINDy's error is 1.49e-06 versus their 0.000152, an order-of-magnitude worse performance that is not discussed. No code or data are provided.\n\nWho gains from this? Readers interested in equation discovery might find it a useful worked example of how smooth-then-differentiate approaches can identify correct coefficients when the model order is known, and how they break down otherwise. It deserves a serious referee because the basic pipeline is novel and the spring-mass results suggest the approach could be salvaged with a proper order-selection or rank-revealing step. But as written, the central claim about recovering governing ODEs from noisy data without a basis library is not established.\n\nMy recommendation: send it to peer review, but with major revisions required. The authors need to fix or replace the null-space identification, add code and data, and redo the EDC study without synthetic point injection.","headline":"New smoothing-plus-SVD pipeline for linear ODE discovery that works when the order is known but lacks a valid null-space identification when the order is over-specified, and the empirical case studies lean on manual thresholding and synthetic data.","tokens_in":15636,"tokens_out":8218,"would_cite":false,"duration_ms":94981,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["34A30","65D07","65F20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a homogeneous linear constant-coefficient ODE can be recovered from noisy data by fitting its general solution, smoothing with a B-spline, and taking the null space of the resulting gradient matrix.","keywords":["ODE discovery","sparse identification","noisy data","B-spline approximation","singular value decomposition","genetic algorithm","model identification","inverse modeling"],"falsifier":"Generate clean data from $y''' - 6y'' + 11y' - 6y = 0$, run the pipeline with a higher order such as $P=7$, and check whether the null-space vector converges to $(-6, 11, -6, 1)$ up to scale as noise tends to zero; failure to converge within the reported tolerances would refute the central claim that the method recovers homogeneous linear constant-coefficient ODEs from noisy data.","tokens_in":1591,"feed_emoji":"📐","tokens_out":2452,"duration_ms":116903,"temperature":0.7,"pith_summary":"The paper proposes a three-step pipeline that recovers a governing ordinary differential equation from noisy measurements without pre-specifying a library of candidate basis functions. It assumes the true equation is a homogeneous linear constant-coefficient ODE, and works by fitting the analytical general solution of that ODE family to the data, smoothing the fitted curve with an adaptive B-spline, and reading the ODE coefficients out of the null space of a gradient matrix built from spline derivatives. On spring-mass benchmarks and on experimental photolysis data for estrogen-disrupting chemicals, the recovered equations match the expected second-order or first-order forms and yield physical parameters such as mass, damping, and rate constants. The paper reports that sparsity in the recovered coefficients emerges without any regularization, that the method tolerates Gaussian noise better than derivative-based sparse regression baselines, and that higher-order or variable-coefficient systems remain outside its current scope.","feed_headline":"Spline plus null space recovers ODEs from noisy data","feed_subtitle":"A fit-then-differentiate pipeline finds linear constant-coefficient equations and physical parameters without regularization.","key_machinery":"The load-bearing object is the gradient matrix $\\mathbf G$ whose entries are spline derivatives $G_{p,k}=d^p y(x_k)/dx^p$; because the spline is a polynomial of order $P$, its derivatives $x^0, x^1, \\dots, x^P$ are linearly independent, so when the data obey a unique governing equation the reduced matrix $\\tilde{\\mathbf G}$ is rank-deficient with a one-dimensional null space. Singular value decomposition $\\tilde{\\mathbf G}=\\mathbf U\\boldsymbol\\Sigma\\mathbf V^\\top$ yields the coefficient vector $\\mathbf C$ as the last column of $\\mathbf V$. The approximate general solution step and the adaptive-knot B-spline fit exist to make this null-space extraction possible: the former removes noise by fitting the general solution form of the assumed ODE family, and the latter supplies a polynomial whose derivatives are well-conditioned and linearly independent.","core_discovery":"The central claim is that for data generated by a homogeneous linear ODE of order $P$ with constant coefficients, the coefficients $C_p$ can be estimated accurately from noisy measurements by a three-stage procedure: first, optimize the coefficient set $\\hat C$ of the general solution form $$y=\\sum_q D_q\\left(\\sum_{j=0}^{\\alpha_q}x^j\\right)$e^{{\\beta_q x}}$(\\cos(\\gamma_q x)+\\sin(\\gamma_q x))$$ using a genetic algorithm and a pseudoinverse fit; second, approximate that smooth function by a B-spline with adaptive knot selection so that the derivatives entering the gradient matrix are linearly independent; third, assemble the gradient matrix $\\mathbf G$ from spline derivatives, compute its singular value decomposition, and take the ODE coefficients as the right-singular vector spanning the null space. The paper argues that this recovers the governing equation without a basis library or regularization, produces sparse coefficient vectors by construction, and remains accurate when Gaussian noise is added to the data.","pith_inferences":["The method's smoothing step effectively pre-filters the data with a model from the target ODE family, so for genuinely nonlinear systems a 'discovered' linear ODE should be read as a local linearization rather than the true law.","The genetic-algorithm fit of the general solution is only a means to a smooth curve; a cheaper deterministic smoothing approach could replace it and make the pipeline faster and more reproducible.","The EDC case study adds 1000 synthetic points by linear interpolation of a small number of experimental points before discovery; a natural test is to vary the original sample count and noise level to see when this interpolation distorts the recovered rate constants.","The same fit-smooth-nullspace design could be extended to nonhomogeneous or weakly nonlinear equations by fitting a particular solution along with the homogeneous general solution, but the paper does not do so."],"forward_implications":["Equations of the form $\\sum_{p=0}^P C_p y^{(p)}=0$ can be identified from noisy time-series without choosing a candidate basis beforehand.","Physical parameters can be read off the recovered coefficients: spring-mass tests recover mass and damping, with the best noisy-data agreement in underdamped responses, and photolysis tests recover first-order rate constants close to the literature values.","The recovered coefficients are sparse by construction: higher-order terms fall below threshold in both case studies with no regularization term added.","The same pipeline extends naturally to second-order harmonic-oscillator-type systems and to first-order degradation kinetics, while the authors state that higher-order and variable-coefficient systems remain limitations.","Because the intermediate genetic-algorithm coefficients are nonphysical, the method's output is the final null-space vector, not the intermediate smooth fit."],"supporting_citations":[{"why":"Introduces the sparse-regression baseline whose need for a candidate basis and direct derivative estimation motivates the paper's no-library, no-regularization contrast.","marker":"[9]"},{"why":"Software implementation used to produce the SINDy-STLSQ comparison numbers in Table 4.","marker":"[24]"},{"why":"Weak-form sparse regression variant used as an additional comparison method in Table 4.","marker":"[25]"},{"why":"Wronskian determinant result used to justify that polynomial spline derivatives are linearly independent.","marker":"[17]"},{"why":"SVD null-space computation used to extract the ODE coefficient vector from the reduced gradient matrix.","marker":"[21]"},{"why":"Experimental dataset of UV photolysis of estrogen-disrupting chemicals whose apparent first-order kinetics are recovered in Case Study 2.","marker":"[26]"},{"why":"Genetic algorithm optimizer used in the approximate general solution step.","marker":"[7]"}],"fun_headline_variants":["Spline and null space: noiseless ODE discovery","Fit smooth spline, then null space gives ODE","Noisy data to sparse ODEs via spline null space","Unsupervised ODE fits without regularization","Gradient matrix null space extracts ODE coefficients"],"cache_read_input_tokens":17664,"weakest_assumption_plain":"The load-bearing premise is Eq. (1), that the governing equation is a homogeneous linear ODE of order $P$ with constant coefficients; the authors' own conclusion notes that higher-order and variable-coefficient systems remain limitations.","fun_headline_variants_meta":{"raw":{"variants":["Spline and null space: noiseless ODE discovery","Fit smooth spline, then null space gives ODE","Noisy data to sparse ODEs via spline null space","Unsupervised ODE fits without regularization","Gradient matrix null space extracts ODE coefficients"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000359,"raw_usage":{"total_tokens":1957,"prompt_tokens":970,"completion_tokens":987,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":908}},"tokens_in":586,"tokens_out":987,"duration_ms":9996,"temperature":1.0,"reasoning_tokens":908,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:19:21.172825+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate clean data from $y''' - 6y'' + 11y' - 6y = 0$, run the pipeline with a higher order such as $P=7$, and check whether the null-space vector converges to $(-6, 11, -6, 1)$ up to scale as noise tends to zero; failure to converge within the reported tolerances would refute the central claim that the method recovers homogeneous linear constant-coefficient ODEs from noisy data.","supporting_citations":[{"cited_title":"Discovering governing equations from data by sparse identification of nonlinear dynamical systems","cited_arxiv_id":null,"evidence_quote":"Introduces the sparse-regression baseline whose need for a candidate basis and direct derivative estimation motivates the paper's no-library, no-regularization contrast."},{"cited_title":"Wronskian determinants and the zeros of certain functions","cited_arxiv_id":null,"evidence_quote":"Wronskian determinant result used to justify that polynomial spline derivatives are linearly independent."},{"cited_title":"Trefethen and D","cited_arxiv_id":null,"evidence_quote":"SVD null-space computation used to extract the ODE coefficient vector from the reduced gradient matrix."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Experimental dataset of UV photolysis of estrogen-disrupting chemicals whose apparent first-order kinetics are recovered in Case Study 2."},{"cited_title":"Genetic algorithm","cited_arxiv_id":null,"evidence_quote":"Genetic algorithm optimizer used in the approximate general solution step."}],"review_version":1}