{"id":"152d6b08-cc2c-4faf-90f6-791010ece707","arxiv_id":"2510.09213","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"The two-phase Morphology-Adaptive Random Feature Method solves the multi-frequency Helmholtz inverse source problem by adaptive quadrature plus morphology-matched basis functions, reaching 1.3–16% relative l2 errors on 2D/3D benchmarks under 0–20% noise.","lead":"MA-RFM reconstructs hidden wave-producing sources by rewriting the Helmholtz equation as an integral, adapting its sampling grid, and adding basis functions shaped like the detected source geometry. On 2D/3D benchmarks with up to 20% noise it reports 1.3–16% relative errors while cutting integration cost versus Fourier-based inversion.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MA-RFM's 'morphology-adaptive' accuracy is contingent on the user supplying the correct template family; no benchmark uses an out-of-family shape.","rationale":"The reader's weakest assumption and the load-bearing concern here coincide: the second-stage enrichment is only as good as the user-supplied morphology family and the detection pipeline's ability to pick the right member. The paper never tests a shape outside the chosen family, so the 'adaptive morphology' claim is not yet supported for genuinely unseen geometries. I also weighed two other issues — the invalid analytic-continuation step in the proof of Theorem 2.1 and the inverse crime in data generation — but neither is as directly tied to the central claim as the template-family dependence. The proof issue affects the uniqueness theorem, not the numerical comparisons; the inverse crime affects absolute error levels but likely affects IRFM and MA-RFM similarly, so the efficiency and relative-accuracy story survives. The template-family dependence, by contrast, determines whether the method is a general adaptive technique or a shape-prior method. The recommended verdict stays CONDITIONAL: the reported benchmarks are plausible, but the advertised adaptivity needs an out-of-family test and code release before acceptance.","tokens_in":28036,"tokens_out":3691,"duration_ms":33918,"concrete_test":"Add a benchmark whose support is an ellipse (e.g. ((x-0.5)/0.3)^2 + ((y-0.5)/0.1)^2 ≤ 1) — not in the template bank F used by Algorithm 2. Generate boundary data with an independent forward solver different from Eq. (2) (e.g. a boundary-element or finite-difference discretization), add 10% noise as in Eq. (43), and run MA-RFM with F containing only circle/rectangle/Gaussian templates. Compare final E_l2 against IA-RFM and against a run where the correct ellipse basis is manually added. If out-of-family error ≈ IA-RFM error (≈20%+) while in-family error ≤13%, the 'morphology-adaptive' claim fails for unseen shapes; if out-of-family error stays ≤15%, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central numerical claim — MA-RFM accurately reconstructs complex/discontinuous sources — rests on the second stage's ability to enrich the basis using detected morphology. Algorithm 2 (lines 6–10) selects the basis type T_k from a user-provided family F, and Eqs. (33)–(36) generate basis functions only from that family. In every benchmark the template family is chosen to match the unknown source: circle-sigmoid (Ex. 4.3), truncated Gaussians (Ex. 4.4), circle+rectangle (Ex. 4.5), kidney SDF (Ex. 4.6), cone-ReLU (Ex. 4.7), torus SDF (Ex. 4.8). Detection itself is biased (Table 3: r̂=0.219 vs. true 0.2), yet the method still works because the correct family is included. If the true support is an ellipse, star polygon, or any shape not in F, the enrichment space cannot represent it and MA-RFM reduces to IA-RFM plus misspecified bump functions; the reported 10–16% errors are then not evidence for the general claim. This is a correctness concern about the method's advertised adaptivity, not about the benchmark computations as reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a two-phase random feature method for the multi-frequency Helmholtz inverse source problem. The first phase (IA-RFM) uses an iterative adaptive integration mesh, driven by the current source estimate's magnitude and gradient, to concentrate quadrature points near the support. The second phase (MA-RFM) adds hand-designed 'morphology' basis functions -- circle sigmoids, Gaussian bumps, cone-ReLUs, torus SDFs, etc. -- chosen after inspecting the IA-RFM solution, and then retrains a convex Tikhonov-regularized least-squares problem. Theoretical sections prove conditional uniqueness for multi-frequency Cauchy data (Theorem 2.1) and an optimal convergence-rate bound for the Tikhonov solution (Theorem 3.1). Numerical experiments cover smooth, discontinuous, disjoint, and 3D sources, reporting relative L2 errors of roughly 1.35%--16% at 5--10% noise, and compare against IRFM, a Fourier method, and PINN.","tokens_in":28303,"tokens_out":7877,"duration_ms":70181,"significance":"If the numerical results are reproducible, the integral-equation formulation with a hard radiation condition and a convex random-feature solve is a useful and efficient alternative to PINN-type approaches for this inverse-source class. The paper's honest reporting of noise sensitivity, its inclusion of 3D benchmarks, and its theoretical stability bound are strengths. However, the central \"morphology-adaptive\" claim is not yet substantiated: every experiment supplies the template family that matches the true source, no out-of-family shape is tested, and the reported gains are not isolated from the adaptive-integration component. The uniqueness proof in Appendix A also has a serious gap. The manuscript is a promising contribution, but it needs substantial revision before the advertised claims can be accepted.","major_comments":[{"comment":"The final step of the proof is not valid as written. After showing that the Fourier transform \\hat S vanishes on the spheres |\\xi|=k_j, the proof restricts to one complex line f(z)=\\hat S(z,0,\\dots,0) and then concludes that '\\hat S is zero on the whole complex domain C^d.' Vanishing of f on a set with an accumulation point only implies that the marginal \\int S(x_1,x') dx' is zero, not that S is identically zero. To obtain S=0 one would need to repeat the argument in every direction (or use a genuinely multi-dimensional uniqueness theorem). Since Theorem 2.1 is stated as the theoretical basis for the multi-frequency setup, this gap must be repaired before the theorem can be used.","section":"Appendix A, proof of Theorem 2.1"},{"comment":"The accuracy improvement is not isolated to the morphology stage. MA-RFM differs from the IRFM baseline in two ways: adaptive integration (IA-RFM) and the added morphology basis functions. For discontinuous sources, Tables 2 and 5 compare IRFM directly with MA-RFM, but no IA-RFM-only column is reported. Thus the gains (e.g., 22.78% to 10.50% in Table 2) conflate the effect of adaptive quadrature with the effect of the morphology bases. An ablation with IA-RFM alone on the same examples is necessary to support the paper's claim that the second stage is responsible for the improved reconstruction of discontinuities.","section":"§4, Tables 2 and 5"},{"comment":"The 'morphology-adaptive' property is currently conditional on the user supplying a template family F that contains the true shape. In every benchmark, the basis family is chosen to match the unknown source: circle-sigmoid (Ex. 4.3), truncated Gaussian (Ex. 4.4), circle+rectangle (Ex. 4.5), kidney SDF (Ex. 4.6), cone-ReLU (Ex. 4.7), and torus SDF (Ex. 4.8). In Example 4.8 the text explicitly says the choice was prompted by the detected point cloud forming a 'toroidal donut-like structure.' No test uses a source whose shape is not in F, and Algorithm 2 line 8 does not describe a mechanism for selecting among candidate families. The claimed adaptivity therefore reduces to a user-supplied shape prior. Please provide either an automatic family-selection rule or an out-of-family benchmark that demonstrates robustness when the template is misspecified.","section":"Algorithm 2, line 8; Examples 4.3–4.8"},{"comment":"The claim that the method 'reduces the computational cost by two to three orders of magnitude compared to traditional methods' is not supported by the data. Table 1 reports no wall-clock time; the reported metric n_integral is ambiguous (entries like '800 2' suggest a formatting error for powers). The number of integration points alone does not determine complexity, since the linear solve cost depends on the number of basis functions and observation points as well. Please report actual timings or a clearly specified operation count for each method, and make the units/format of Table 1 unambiguous.","section":"§5 Conclusion"},{"comment":"Random Feature Methods are stochastic, but every reported error is a single run with no seed specified and no variance estimate. The claimed accuracy differences between methods could be within run-to-run variability, especially at low noise levels. At minimum, provide mean and standard deviation over several random seeds (e.g., 5--10 seeds) for the main comparisons in Tables 2, 5, and 6. Without this, the results are not statistically reproducible.","section":"§4 (all numerical experiments)"}],"minor_comments":[{"comment":"The radiation condition is repeatedly spelled 'Sommerfield'; it should be 'Sommerfeld'.","section":"Throughout"},{"comment":"The domain definitions are 2D: V0 = [-0.5,0.5]×[-0.5,0.5] and Ω = [-0.75,0.75]×[-0.75,0.75], but the example is 3D. These should be Cartesian products of three intervals.","section":"Example 4.8"},{"comment":"In the displayed calculation for the second term, the source condition is written as s* = Ψ_M^T Ψ_M w (without the exponent ν). Also 'singular value deposition' should be 'singular value decomposition'.","section":"Proof of Theorem 3.1"},{"comment":"The relative error at θ_max=π/2 is 22.79%, which is large. The phrase 'still roughly determine the locations' is fair, but the abstract's claim of superior performance under limited aperture should be tempered.","section":"Example 4.9, Table 7"},{"comment":"The notation 'n_d Gauss points in the reference cell' is later used as n^d; please clarify in the text or notation.","section":"Algorithm 1, line 1"}],"recommendation":"major_revision","confidential_remarks":"The paper has a plausible core idea and a useful convex formulation, but the current evidence for its two main advertised properties -- morphology adaptivity and large computational speedups -- is incomplete. The uniqueness proof in Appendix A also needs a mathematical fix. I do not see grounds for rejection: the issues are addressable with additional experiments (ablation, repeated seeds, out-of-family test, timing) and a corrected proof. Scope-wise the paper fits math-ph/NA journals; I would encourage a major revision rather than acceptance at this stage."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful numerical paper that does what it claims on the tested benchmarks, but the 'morphology-adaptive' part is closer to 'shape-prior-assisted' — in every experiment the template family is chosen to match the true support, and there is no out-of-family test. The proof of Theorem 2.1 has a gap, though it is easily repairable.\n\nWhat is genuinely new: combining the integral-equation formulation of the Helmholtz inverse source problem with RFM and an adaptive quadrature (IA-RFM), then enriching the basis with morphology functions whose parameters are estimated from the rough first-stage solution. That two-phase synthesis is not in the cited RFM literature. The benchmark set is broad — disc, two circles, rectangle+circle, kidney, 3D cones, torus, limited aperture — and the error tables are honestly reported. The Tikhonov analysis in Theorem 3.1 is standard but clean, and the controlled verification in Example 4.1 is a nice addition.\n\nThe soft spots, in order of importance. First, the template family F in Algorithm 2 is user-chosen per test case and always matches the true shape. The first stage does detect location and scale — even though that detection is biased, e.g. r̂=0.219 vs 0.2 in Example 4.3 — but it does not choose the topology of the support. That makes 'morphology-adaptive' an overstatement; the real claim is 'adaptive within a shape family the user provides.' An out-of-family test (say an ellipse or star-shaped support) is needed before the general claim can be trusted. Second, no seeds or variance are reported for a method that is random by construction. Third, data are generated with the same integral operator used in the inversion (mild inverse crime); independent forward data would strengthen the numbers. Fourth, the proof of Theorem 2.1 as written is wrong: knowing the restriction of \\hat{S} to one complex line vanishes at z=k_j does not imply \\hat{S} vanishes as a function of several variables. The fix is to run the same argument along every direction, so the theorem is likely true, but the current text is not a proof. Fifth, the conclusion's 'two to three orders of magnitude' cost reduction is supported mainly by one comparison against the Fourier method, and 10–16% relative errors at 10% noise are decent, not 'high accuracy.'\n\nWho should read it: anyone working on random feature methods or mesh-free inverse problems. It deserves a serious referee. I would send it to review with a request for multi-seed statistics, code/data release, an out-of-family morphology test, and a corrected proof.","headline":"Solid two-phase RFM paper with a real 'shape-prior-assisted' caveat and a repairable proof gap — worth a serious referee.","tokens_in":28891,"tokens_out":6515,"would_cite":true,"duration_ms":59758,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35R30","65N35"],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-phase random feature method reconstructs sharp and disjoint Helmholtz sources at a fraction of traditional cost.","keywords":["Helmholtz equation","inverse source problem","random feature method","Tikhonov regularization","adaptive quadrature","morphology basis","multi-frequency","integral equation"],"falsifier":"Run the method on a source whose support is a shape not in the paper's families, such as a thin spiral arm or a disconnected set of random polygons, with 10% noise; if the relative L2 error does not improve over the plain IA-RFM (or stays above roughly 20%), the adaptive-morphology claim loses its support.","tokens_in":27759,"feed_emoji":"🎯","tokens_out":3005,"duration_ms":28103,"temperature":0.7,"pith_summary":"This paper proposes the Morphology-Adaptive Random Feature Method (MA-RFM) for the inverse source problem of the Helmholtz equation, targeting sources with discontinuous or geometrically complex supports. The method first reformulates the problem as an integral equation using the fundamental solution, then applies an adaptive iteration that concentrates quadrature points where the source magnitude or gradient is large. In a second phase, it adds morphology-matched smooth basis functions, such as sigmoid or Gaussian bumps, to represent sharp interfaces. The authors report relative L2 errors around 10–16% even with 10% measurement noise, and computational speedups of two to three orders of magnitude over uniform-grid Fourier or fixed random-feature approaches. If correct, this offers a mesh-free, convex alternative to neural-network solvers for a class of ill-posed inverse problems.","feed_headline":"Random feature method cuts inverse-source cost by orders of magnitude","feed_subtitle":"Recovers sharp, disjoint Helmholtz sources from noisy boundary data at a fraction of the quadrature points.","key_machinery":"The central machinery is a two-phase basis- and mesh-adaptive random feature method built on the integral equation formulation of the Helmholtz equation. The fundamental solution (Hankel function in 2D, exponential over distance in 3D) serves as a kernel that makes the radiation condition a hard constraint and reduces the unknown to the source term alone. The first phase uses a dual-criterion indicator combining the absolute value and gradient of the current source reconstruction to iteratively refine a tensor-product Gauss quadrature mesh, concentrating samples where the source is large or varies rapidly. The second phase enriches the random-feature span with smooth approximations to level-","core_discovery":"The authors claim that the multi-frequency Helmholtz inverse source problem, even when the source has discontinuous or complicated geometry, can be solved accurately and cheaply by a two-stage random feature algorithm. In the first stage, IA-RFM iteratively refines a Gauss quadrature mesh using the absolute value and gradient of the current source approximation, concentrating integration points near the true support. In the second stage, based on the detected support and its boundary, additional basis functions are constructed from signed-distance level sets of simple shapes (circle, rectangle, kidney, torus) with sigmoid, tanh, Gaussian, or ReLU profiles. The resulting linear system is Tikh","pith_inferences":["The method's success depends on the user choosing a template family of shapes that matches the true source geometry; the paper tests only shapes from the chosen families, so a source outside these families might not benefit from the second phase at all.","The integral-equation reformulation is not limited to the Helmholtz equation: any linear PDE with a known fundamental solution could be treated similarly, though the adaptive morphology step would need a shape dictionary for that problem's singularities.","A natural testable extension is to let the second phase search over a broader dictionary of level-set functions (for example, arbitrary implicit curves) rather than pre-selected simple shapes, and to measure how the reconstruction improves when the dictionary contains the true shape versus when it does not.","The stability theorem suggests that the reconstruction error scales almost linearly with noise when the source-condition parameter is near one; this could guide practical error bars for reconstructions in medical or acoustic imaging."],"forward_implications":["If correct, discontinuous and complex-shaped sources can be recovered from multi-frequency Cauchy data with only boundary measurements, without dense volumetric meshes.","The two to three orders of magnitude reduction in integration points would make 3D inverse source problems practical on ordinary hardware.","Because the discretized problem is strictly convex, the method avoids local minima that plague neural-network solvers and provides a unique solution for any regularization parameter.","The stability bound gives a principled way to choose the regularization parameter from the estimated noise level and model error, which could be extended to other linear inverse problems.","The morphology-adaptation idea suggests that a small dictionary of level-set-based basis functions can vastly improve random feature methods for PDEs with singular or piecewise-smooth coefficients."],"fun_headline_variants":["MA-RFM: random features that adapt to irregular source shapes","Two-phase random feature method homes in on complex source geometry","Morphology-adaptive random features resolve sharp Helmholtz sources","MA-RFM: two-stage random features for discontinuous sources","Adaptive random feature method recovers disjoint sources from noise"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The second phase assumes that the user-chosen template family of shapes, together with the detection thresholds, identifies the true support well enough that the added basis functions can represent it; if the true shape is not in the family, the enrichment cannot improve the reconstruction.","fun_headline_variants_meta":{"raw":{"variants":["MA-RFM: random features that adapt to irregular source shapes","Two-phase random feature method homes in on complex source geometry","Morphology-adaptive random features resolve sharp Helmholtz sources","MA-RFM: two-stage random features for discontinuous sources","Adaptive random feature method recovers disjoint sources from noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000709,"raw_usage":{"total_tokens":3059,"prompt_tokens":802,"completion_tokens":2257,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":2174}},"tokens_in":546,"tokens_out":2257,"duration_ms":22834,"temperature":1.0,"reasoning_tokens":2174,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T10:40:15.339403+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the method on a source whose support is a shape not in the paper's families, such as a thin spiral arm or a disconnected set of random polygons, with 10% noise; if the relative L2 error does not improve over the plain IA-RFM (or stays above roughly 20%), the adaptive-morphology claim loses its support.","supporting_citations":[],"review_version":1}