{"id":"63b16463-a1ba-4ab4-9af7-64d114b2f0e6","arxiv_id":"2501.18715","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A data-driven library, chebgreen, learns continuous empirical Green's functions for unknown 1D linear PDEs and interpolates them across control parameters using manifold-based interpolation of singular functions.","lead":"This paper introduces chebgreen, a library that learns the Green's function of an unknown one-dimensional differential equation from input-output data and predicts it at new parameter values without solving the equation. The approach combines neural networks with Chebyshev polynomial representations to make the learned model continuous and mesh-independent, which could reduce the cost of exploring parameterized physical systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Interpolation guarantee is only asymptotic/local; the large Airy gaps (θ=1,5,10) are never checked against Lemma 1's π-radius, and the fractional-Laplacian text/figure disagree on whether θ* is unseen, so the central interpolation claim is not yet supported.","rationale":"The paper's main novelty is the manifold interpolation of continuous singular functions to unseen parameters. For that claim to hold, the lifted singular functions must be smooth enough in θ over the interpolation interval, and the finite gaps must be within the regime where the projection and retraction error rates are meaningful. Section 3 proves only local asymptotic rates: Proposition 1 gives cubic error as y→φ, Proposition 2 gives quadratic retraction error for small tangent vectors, and §3.3 merely states the Lagrange rate O(h^{n+1}) without constants or a combined bound. The numerical experiments use large gaps (Airy θ=1,5,10; advection-diffusion θ=1,2,3; fractional Laplacian θ=0.8,0.9,0.95). No computation of the geodesic distance between the reference basis and the other interpolants is reported, so Lemma 1's injectivity-radius condition is unverified. This is load-bearing: if the singular functions are not sufficiently smooth or the points lie outside the local regime, the interpolation could fail outside these examples. The fractional-Laplacian inconsistency (θ*=0.9 in the text, 0.85 in the figure caption) further weakens the empirical support because interpolating at a training node is not a test of interpolation. I do not claim the method is wrong; the code is available and the reported errors are positive. The proposed test—comparing interpolation error at small versus large parameter gaps—would directly show whether the local smoothness assumption controls the observed accuracy. If the small-gap error is much smaller, the concern is answered; if not, the central claim needs either a non-asymptotic analysis or more comprehensive experiments before acceptance. The reader's conditional verdict is appropriate, and my concern does not change it.","tokens_in":27051,"tokens_out":12405,"duration_ms":119868,"concrete_test":"For the Airy operator, run the full chebgreen pipeline with three training nodes spaced closely (θ=1,2,3, target θ=2.5) and with the paper's wide spacing (θ=1,5,10, target θ=7), using identical hyperparameters except θ; if the small-spacing test error is not substantially below the 2.76% reported for the wide spacing, the assumed smoothness of the lifted singular functions in θ is not supported and the Section 3 asymptotic analysis does not explain the reported accuracy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central interpolation claim is supported only by an asymptotic, local error analysis. Section 3.3 composes three rates—cubic projection error, O(h^{n+1}) Lagrange interpolation error, and quadratic QR-retraction error—but never bounds the constants or the radius within which these rates hold at the parameter separations used. Lemma 1 gives an injectivity radius π for the exponential map, yet the paper never checks that the learned singular-function quasimatrices at θ1=1 and θ3=10 in the Airy example lie inside the geodesic ball of radius π around the reference basis θ0=5. Without such a check, the tangent-space projection is not guaranteed to be a valid logarithmic-map approximation, and the Lagrange interpolation in the tangent space has no theoretical support at finite gaps. The empirical evidence is also thin: one of the three demonstrations, the fractional Laplacian, is internally inconsistent (Section 4.4 states interpolation to θ*=0.9, which is itself a training node, while Fig. 16's caption reports θ*=0.85), so it may not actually test interpolation to an unseen parameter. The central claim of high-fidelity interpolation to unseen parameters therefore rests on two examples and an unverified local-smoothness assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents chebgreen, a Python library that learns continuous approximations of Green's functions of one-dimensional linear PDEs from input-output data and interpolates them to new values of a control parameter. The pipeline trains rational neural networks to represent the Green's function and homogeneous solution, converts the learned bivariate function into a Chebyshev-based singular value expansion using a new Python implementation of chebfun, and interpolates the resulting quasimatrices on the tangent space of an infinite-dimensional Stiefel manifold. Numerical experiments cover Poisson, advection-diffusion, Airy, and fractional Laplacian problems, with and without noise; the reported test errors for interpolated models are below about 3%.","tokens_in":27341,"tokens_out":6173,"duration_ms":51292,"significance":"If the results hold, the paper offers a meaningful step beyond the discrete empirical Green's function interpolation of Praveen et al. [32]: the representation is continuous and mesh-independent, it applies to non-self-adjoint operators, and the code is publicly available. The paper also provides a rare external check: for the advection-diffusion case the interpolated Green's function is compared to an analytical expression with relative error 0.26%. The theoretical error analysis (cubic projection error, quadratic retraction error) is elegant but only local; the main support for the central claim is the empirical test-error evidence, which is reasonably strong for the Airy and advection-diffusion experiments.","major_comments":[{"comment":"The combined error analysis is asymptotic and local, and the paper does not verify the conditions under which the bounds hold at the parameter spacings used in the experiments. In particular, Lemma 1 guarantees the exponential map is a local diffeomorphism only inside a geodesic ball of radius π, but the manuscript never checks that the learned singular quasimatrices at θ1=1 and θ3=10 in the Airy example lie within that ball around the reference θ0=5. Please either add a numerical check (e.g., compute the norm of the lifted tangent vectors ΓU_j and ΓV_j, or approximate the geodesic distances, and compare with π) or explicitly state that the theoretical bounds are asymptotic and do not cover the finite parameter gaps used in the experiments.","section":"Section 3.3, Lemma 1"},{"comment":"The fractional Laplacian experiment is internally inconsistent: the text states 'interpolating to θ* = 0.9', but θ2 = 0.9 is one of the three training nodes, while the Fig. 16 caption reports 'Interpolated Green's function at θ* = 0.85'. Since the central claim concerns interpolation to unseen parameters, this example must be corrected and re-run; if the intended target is 0.85, the text and Table 1 should be updated consistently.","section":"Section 4.4"},{"comment":"The advection-diffusion experiment reports Ntrain = 237 input-output pairs, which conflicts with the default Nsamples = 100 (95% training) described in Section 4 and with the sample-size study in Section 4.1.1 that motivates using 100 samples. Please clarify the data budget for this experiment and reconcile the 'order of magnitude lower data' claim with the actual numbers.","section":"Section 4.2"},{"comment":"The theoretical error analysis covers only the interpolation steps for a given singular value expansion; it does not include the neural network approximation error or the Chebyshev truncation error. The statement that the quadratic retraction error 'dominates the error of the other two steps' is also not a complete error bound for the full pipeline. Please state clearly which sources of error are covered by the theory and which are only assessed empirically.","section":"Section 3.3"}],"minor_comments":[{"comment":"The periodic boundary condition is written as u(−π/2) = u(−π/2) on both sides; it should presumably be u(−π/2) = u(π/2).","section":"Equation (19)"},{"comment":"The notation 'using {[ΓU_j, ΓU_j]}' appears to have a typo; the second entry should likely be ΓV_j.","section":"Algorithm 2, Step 3.2"},{"comment":"There is a typo: 'consiedered' should be 'considered'.","section":"Section 2.7"},{"comment":"The x-axis of Fig. 10A is labeled 'Noise level' but the study in Section 4.1.1 is about sample count; please ensure the figure labels match the discussion.","section":"Section 4.1.1"}],"recommendation":"major_revision","confidential_remarks":"The fractional Laplacian inconsistency and the unverified injectivity-radius condition are the main blockers. The paper is otherwise suitable for Computer Physics Communications, and the open-source code and analytical check are strong points. Please ask the authors to either provide the numerical check for the Stiefel-manifold radius or temper the theoretical claims, and to correct the fractional Laplacian target value."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this paper extends the group's earlier discrete empirical Green's function interpolation to a continuous, mesh-independent setting, and to non-self-adjoint 1D operators. That is a real step forward, and the implementation is public. But the headline claim—high-fidelity interpolation to unseen parameters—rests on two clean examples plus one demonstration that is internally inconsistent about whether the target parameter was actually unseen. Still, it deserves a serious referee.\n\nWhat is genuinely new: the SVE of a rational-NN-learned Green's function is computed via a Python chebfun2 implementation, yielding quasimatrices; interpolation then happens in the tangent space of an infinite-dimensional Stiefel manifold, with a projection they prove is orthogonal and an error analysis giving cubic projection error, O(h^{n+1}) Lagrange error, and quadratic QR-retraction error. That is more theory than most operator-learning papers carry. The advection-diffusion example includes an analytical comparison (relative error 0.26%), which is the right kind of check. The noise-robustness experiments are also a plus.\n\nSoft spots, in proportion: the theoretical error bounds are local and asymptotic; the paper never checks whether the parameter separations used (Airy theta=1,5,10, reference at 5) actually lie within the injectivity radius pi of Lemma 1. Without that check, the tangent-space projection is a plausible heuristic rather than a guaranteed approximation at the gaps tested. The fractional Laplacian section states interpolation to theta*=0.9, which is itself a training node, while the figure caption reports theta*=0.85; if the text is right, that demonstration does not test interpolation to an unseen parameter at all. The data-efficiency claim (order-of-magnitude fewer samples vs. [32]) is asserted but not benchmarked against the prior method. And the advection-diffusion experiment uses Ntrain=237 while the default says Nsamples=100; that needs an explanation.\n\nNone of this sinks the method; all of it is addressable. The central idea—continuous SVE plus manifold interpolation—is sound, and the paper is honest about the scope (1D, linear, one parameter).\n\nWho this is for: people working on operator learning, surrogate modeling from experimental data, or Green's function discovery. I would send it to a referee who knows Stiefel manifold interpolation and ask for the fixes above. I would not desk-reject it.","headline":"Genuine continuous extension of EGF interpolation with public code and a real error analysis, but the unseen-parameter claim needs an inconsistency fixed and a check on the injectivity radius before it fully lands.","tokens_in":27832,"tokens_out":2320,"would_cite":true,"duration_ms":22195,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65D05","65F25","65R20","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes that learned continuous Green's functions for unknown one-dimensional linear operators can be interpolated to unseen control-parameter values with test errors under about 3 percent, including for non-self-adjoint…","keywords":["Empirical Green's function","operator learning","singular value expansion","Chebyshev polynomials","quasimatrices","Stiefel manifold interpolation","rational neural networks","parametric linear PDEs"],"falsifier":"Learn Green's functions for $d^2u/dx^2 + \\theta^2 u = f$ at $\\theta=4.9$ and $\\theta=5.1$, where the paper shows the first two singular functions swap order, interpolate to $\\theta^*=5.0$ with the paper's Algorithm 2, and compare against a directly learned target at 5.0; a test error far above the roughly 3 percent seen in the paper's Airy experiments would show that the sign- and order-corrected tangent-space interpolation does not track the crossing.","tokens_in":26869,"feed_emoji":"🧮","tokens_out":11027,"duration_ms":95259,"temperature":0.7,"pith_summary":"The paper seeks to show that the solution operator of an unknown one-dimensional linear system, parameterized by a scalar control variable, can be learned from force-response data and then predicted at parameter values never seen in training. Its route is to learn a continuous empirical Green's function, the impulse response of the hidden operator, as a rational neural network, rewrite it as a singular value expansion in a Chebyshev basis, and interpolate the left and right singular functions through the tangent space of an infinite-dimensional Stiefel manifold. On the paper's benchmarks the interpolated Green's functions reach test errors of roughly 0.3 to 3 percent, the method handles non-self-adjoint operators such as advection-diffusion, and the data budget of about 100 input-output pairs is an order of magnitude smaller than in the earlier discrete approach. If the claim holds, unknown physical systems described by parametric linear PDEs can be modeled mesh-independently from modest, noisy measurements.","feed_headline":"Learned Green's functions interpolate to unseen parameter values","feed_subtitle":"Advection-diffusion, Airy, and fractional-Laplacian benchmarks stay below 3 percent test error.","key_machinery":"The load-bearing object is the singular value expansion of the learned Green's function, $G_\\theta(x,s) = U(s;\\theta)\\Sigma(\\theta)V^*(x;\\theta)$, where $U$ and $V$ are quasimatrices, meaning matrices whose columns are $L^2$ functions represented as Chebyshev interpolants, built by a two-dimensional Chebyshev factorization. Around this object the algorithm constructs an infinite-dimensional analogue of the Stiefel manifold, the space of quasimatrices with orthonormal columns. The essential mechanism is tangent-space interpolation: project each model's singular-function quasimatrix onto the tangent space at a reference parameter, interpolate those tangent points and the singular values with Lagrange polynomials, and retract via quasimatrix QR factorization. The paper supplies the projection formula and proves that it is orthogonal, and its error analysis identifies the QR retraction as the dominant error source.","core_discovery":"The central discovery claimed is that a continuous empirical Green's function comes with a manifold-interpolation rule that lets it travel in parameter space: the singular functions of the learned Green's function for each sampled parameter $\\theta_j$ are treated as points on the Stiefel manifold of quasimatrices, lifted to the tangent space at a reference parameter $\\theta_0$ by the projection $\\Psi \\mapsto \\Psi - \\Phi\\,\\mathrm{sym}(\\Phi^*\\Psi)$, interpolated with Lagrange polynomials, and mapped back by quasimatrix QR factorization, while the singular values are interpolated separately. The paper proves that this projection is the orthogonal projection onto the tangent space, gives error bounds of cubic order for the projection and quadratic order for the QR retraction, and handles mode swaps and sign flips by matching inner products against the reference basis. In the paper's account the result is high-fidelity interpolation for the Poisson, advection-diffusion, Airy, and fractional-Laplacian operators, including extrapolation to $\\theta=9$ from models at 6, 7, and 8, with the continuous Chebyshev representation making the pipeline independent of any particular spatial mesh.","pith_inferences":["Stress-test extension: interpolate across the singular-value crossing that the paper identifies for the Helmholtz operator at $\\theta_{\\mathrm{crit}}=\\sqrt{5/2}\\,\\pi$; if the test error spikes at the crossing, the mode-matching assumption, rather than the manifold geometry, is the bottleneck.","Extension to real sensor data: because the singular value expansion is continuous, the pipeline should in principle accept force-response measurements taken at different sensor locations for different parameter values without first resampling to a common grid; the paper allows arbitrary sensor locations but does not demonstrate this cross-parameter mismatch.","Higher-dimensional extension: the same tangent-space interpolation would have to be retried on tensor decompositions such as chebfun3/Tucker or tensor-train forms, and the Eckart-Young-Mirsky obstruction the paper flags suggests a different low-rank manifold will be needed there."],"forward_implications":["For one-dimensional linear operators with a scalar parameter, the pipeline yields a continuously queryable Green's function from about 100 force-response pairs without writing down the governing equation.","Interpolated test errors are 0.54 percent for advection-diffusion at $\\theta=2.5$, 2.76 percent for Airy at $\\theta=7$, and 0.99 percent for the fractional Laplacian at $\\theta=0.85$; extrapolation to Airy $\\theta=9$ gives 2.28 percent.","The method extends to non-self-adjoint and nonlocal operators, going beyond the self-adjoint case treated in the prior discrete empirical-Green's-function work.","Training with 50 percent output noise leaves interpolated test errors below about 3.3 percent on the three benchmarks, and the advection-diffusion test error stays below 5 percent even at 100 percent noise.","Because the representation lives in a Chebyshev basis rather than on a fixed sensor grid, the stored Green's function can be evaluated at arbitrary locations and kept to high precision."],"supporting_citations":[{"why":"supplies the discrete empirical-Green's-function interpolation scheme that this work generalizes to continuous quasimatrices","marker":"[32]"},{"why":"provides the greenlearning rational-neural-network method used to learn a continuous Green's function from force-response data","marker":"[31]"},{"why":"supplies the two-dimensional Chebyshev approximation whose factorization becomes the continuous singular value expansion","marker":"[34]"},{"why":"gives the Householder triangularization of a quasimatrix used to produce orthonormal singular-function factors","marker":"[44]"},{"why":"establishes the infinite-dimensional Stiefel manifold, its tangent space, and quasimatrix QR factorization as a retraction","marker":"[43]"},{"why":"contributes the tangent-space logarithmic and exponential map interpolation idea adapted by the algorithm","marker":"[46]"},{"why":"originates the manifold interpolation strategy for adapting reduced-order models to new parameter values","marker":"[33]"},{"why":"motivates the Gaussian-process forcing ensemble and the modest sample count through theoretical learning guarantees","marker":"[26]"},{"why":"provides the injectivity-radius result on finite Stiefel manifolds that the paper extends to the Hilbert manifold case","marker":"[48]"}],"fun_headline_variants":["chebgreen: learn Green's functions, interpolate across parameters","Empirical Green's functions that interpolate to unseen parameter values","Mesh-free Green's function library with parameter-space interpolation","Interpolating Green's functions in control-parameter space via Chebyshev"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that the left and right singular functions of the Green's function vary smoothly enough across the sampled parameter values for Lagrange interpolation inside a single tangent space to represent the unseen parameter, and the paper's error analysis establishes this only asymptotically as the parameter gap goes to zero rather than at the tested gaps such as Airy $\\theta$ from 1 to 10.","fun_headline_variants_meta":{"raw":{"variants":["chebgreen: learn Green's functions, interpolate across parameters","Empirical Green's functions that interpolate to unseen parameter values","Mesh-free Green's function library with parameter-space interpolation","Interpolating Green's functions in control-parameter space via Chebyshev"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000783,"raw_usage":{"total_tokens":3431,"prompt_tokens":891,"completion_tokens":2540,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":2467}},"tokens_in":507,"tokens_out":2540,"duration_ms":18049,"temperature":1.0,"reasoning_tokens":2467,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T22:44:39.984750+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Learn Green's functions for $d^2u/dx^2 + \\theta^2 u = f$ at $\\theta=4.9$ and $\\theta=5.1$, where the paper shows the first two singular functions swap order, interpolate to $\\theta^*=5.0$ with the paper's Algorithm 2, and compare against a directly learned target at 5.0; a test error far above the roughly 3 percent seen in the paper's Airy experiments would show that the sign- and order-corrected tangent-space interpolation does not track the crossing.","supporting_citations":[{"cited_title":"Praveen, N","cited_arxiv_id":null,"evidence_quote":"supplies the discrete empirical-Green's-function interpolation scheme that this work generalizes to continuous quasimatrices"},{"cited_title":"Boull´ e, C","cited_arxiv_id":null,"evidence_quote":"provides the greenlearning rational-neural-network method used to learn a continuous Green's function from force-response data"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"gives the Householder triangularization of a quasimatrix used to produce orthonormal singular-function factors"},{"cited_title":"Energy-adaptive Riemannian optimization on the Stiefel manifold","cited_arxiv_id":"2108.09831","evidence_quote":"establishes the infinite-dimensional Stiefel manifold, its tangent space, and quasimatrix QR factorization as a retraction"},{"cited_title":"Sternfels, C","cited_arxiv_id":null,"evidence_quote":"contributes the tangent-space logarithmic and exponential map interpolation idea adapted by the algorithm"},{"cited_title":"Boull´ e, A","cited_arxiv_id":null,"evidence_quote":"motivates the Gaussian-process forcing ensemble and the modest sample count through theoretical learning guarantees"},{"cited_title":"On the Injectivity Radius of the Stiefel Manifold: Numerical investigations and an explicit construction of a cut point at short distance","cited_arxiv_id":"2403.03782","evidence_quote":"provides the injectivity-radius result on finite Stiefel manifolds that the paper extends to the Hilbert manifold case"}],"review_version":1}