{"id":"233f6111-8cdf-417f-b301-05b74bd86b37","arxiv_id":"2512.17884","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A regularized random Fourier feature model with finite-element recovery learns PDE solution operators from noisy data, with a conditioning guarantee when the number of features scales like m log m.","lead":"Operator learning methods that map PDE inputs to outputs often become unstable under noisy data and expensive to train. This paper adds a frequency-weighted penalty and a finite-element reconstruction step to random Fourier feature models, reporting lower test errors on six benchmark PDE problems and faster training than full kernel methods.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.1's condition (6) is not verified in the experimental regime; the conditioning guarantee therefore does not apply to the reported RRFF results.","rationale":"Good-faith reading: the paper's empirical contribution is substantial — released code, 20-trial averages on six PDE benchmarks, and a consistent improvement of RRFF over RFF. Theorem 3.1's proof is internally coherent given (5)–(6). However, the central theoretical claim is that a well-conditioned random feature matrix is guaranteed with N ~ m log m, and the paper explicitly connects this to estimation/generalization and to the method's numerical behavior. Condition (6) is the linchpin: it controls both the bias term ∥B∥_2 ≤ η and the row-sum bounds in the Bernstein argument. Since the published experimental parameters are exactly where (6) appears to be violated (mφ(κ) ≫ 1), the theory does not cover the experiments. This does not refute the empirical RRFF-vs-RFF comparison, but it does mean the paper's explanation for that comparison is unsupported. The reader's CONDITIONAL verdict is the right level: the gap is verifiable and fixable, not a demonstrated counterexample. I would not adjust the verdict. My only refinement is that κ should be the training-input separation in R^n, which strengthens rather than weakens the concern.","tokens_in":22460,"tokens_out":13289,"duration_ms":138025,"concrete_test":"Using the released code and the exact training inputs (the rows of A, with the same noise protocol as Appendix A), compute κ = min_{j≠j′} ∥u_j − u_{j′}∥_2 for a representative experiment per table (e.g., Darcy Flow, Table 3, and Advection I, Table 1). For each (ν, σ), compute φ(κ) and the smallest η_min = m φ(κ) required by (6). Then check whether there exists η ∈ (η_min, 1) satisfying N ≥ 6 η^{−2} m log(2m/δ) with δ=0.01 (the constant C=6 from the proof). If no such η exists, Theorem 3.1 gives no guarantee for that configuration. Report κ, φ(κ), η_min, and the required N for η=0.5 and η=0.9.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The core theoretical result, Theorem 3.1, is conditional on the scale condition (6): φ(κ) := (σ√νκ)^{ν/2} K_{ν/2}(σ√νκ) / (2^{ν/2−1}Γ(ν/2)) ≤ η/m with η<1. Because φ(t) is the characteristic function of the Student-t feature distribution, φ(0)=1 and φ decreases with ∥t∥; to satisfy (6) one needs the minimum separation κ among the rows of A to be large enough that φ(κ) ≤ η/m. In the experiments, the rows of A are the training input functions (possibly noisy), and m is the training-set size (≈10^3–10^4). With the reported scale parameters (e.g., σ=√2×10^−5 for Darcy/Helmholtz, σ=0.02–0.2 elsewhere), the argument σ√νκ is small for any reasonable separation of normalized PDE input functions, so φ(κ) is close to 1. Then mφ(κ) ≫ 1, so no η∈(0,1) can meet (6); even if mφ(κ)<1, the η forced by (6) would be so large that the sample-size condition (5) N ≥ Cη^{−2} m log(2m/δ) fails for the reported N. The paper never computes κ or checks (6). Thus the proof's central premise is not established in the regime where RRFF is tested, and the statement that the conditioning guarantee 'yields estimation and generalization bounds' is not connected to the experiments. The reader's concern is valid; if anything the failure is more direct because κ refers to training-input separation in R^n, not spatial grid spacing.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a regularized random Fourier feature method (RRFF) and a finite-element variant (RRFF-FEM) for learning PDE solution operators from noisy data. Random feature weights are drawn from multivariate Student's t distributions, and a frequency-weighted Tikhonov term is added to the least-squares objective. The main theoretical result, Theorem 3.1, gives a high-probability bound on the deviation of the normalised random feature Gram matrix from the identity under conditions (5)–(6), and the paper states that this yields estimation and generalization bounds. Extensive experiments on advection, Burgers, Darcy, Helmholtz, Navier–Stokes, and structural mechanics benchmark problems compare RRFF/RRFF-FEM with unregularized RFF, kernel methods, and neural operators.","tokens_in":22836,"tokens_out":10416,"duration_ms":113545,"significance":"If the theoretical guarantee held in the parameter regime used in the experiments, the paper would be a useful contribution: it unifies and extends prior random-feature conditioning results to Student-t features, gives the advertised N = O(m log m) scaling, and the benchmark suite is unusually broad and accompanied by code. The proof of Theorem 3.1 appears internally consistent under its stated hypotheses. However, the main load-bearing condition is not verified and is implausible for the reported experimental parameters, and the estimation/generalization consequences are only asserted. The numerical study is substantial and the empirical improvements are interesting, but the advertised provable well-conditioning is not actually connected to the experiments in the present form.","major_comments":[{"comment":"Condition (6) is load-bearing but never verified, and it is not satisfied by the reported experimental parameter values. Since φ(t) is the characteristic function of the feature distribution, φ(0)=1 and φ decreases with ∥t∥; therefore (6) requires the minimum separation κ among the rows of A — the training input functions, not the spatial grid — to be large enough that φ(κ) ≤ η/m. For the Gaussian case and the reported Darcy/Helmholtz scale σ=√2×10^{-5}, (6) reads exp(-σ²κ²/2) ≤ η/m, so for m=10^3–10^4 one needs κ of order 10^5; the discretized PDE input vectors have Euclidean norms of order 10–10^3, so this is not attainable. The paper never computes κ or checks (6). Even if κ could be chosen large, the very small η forced by (6) would make condition (5) require far more than the reported N. Consequently Theorem 3.1 does not apply in the regime where RRFF is tested, and the statement th","section":"§3.2, Theorem 3.1, Eq. (6)"},{"comment":"The paper asserts that the concentration property can be used to derive estimation and generalization bounds 'similar to [18]', but no statement or proof is given. This is not a routine corollary: the estimator in (3) solves a regularized, frequency-weighted least-squares problem with Student-t features, whereas [18] analyzes unregularized minimum-norm interpolation with Cauchy features. The abstract and introduction advertise these guarantees, so the missing derivation is load-bearing. Either state the precise bounds and prove them, or explicitly limit the theoretical claim to the conditioning result.","section":"§3.2, final paragraph"},{"comment":"The regularization parameter α is selected by minimizing the reported average relative test error ('we tested a range of values for α to find the optimal α such that the RRFF method minimizes the test error', Figures 8 and 12), and Tables 7 and 10 report the corresponding errors. This is a test-set leak: the reported 'test' errors no longer measure genuine generalization. The observed improvement of RRFF over RFF may therefore be inflated. Use a held-out validation split or nested cross-validation to choose α, and report test errors only for the final model.","section":"§4.2–§4.3, Tables 7 and 10"}],"minor_comments":[{"comment":"The normalization factor appears to be σ π^{d/2} ν^{d/2} Γ(ν/2), but a d-dimensional scale parameter σ should give a factor σ^d π^{d/2} ν^{d/2} Γ(ν/2). Please correct the density or clarify the notation for σ.","section":"§3, definition of ρ(ω)"},{"comment":"The symbol m is overloaded: in Section 2 it denotes the output dimension, while in Assumption 1 and Theorem 3.1 it denotes the number of training samples. This makes the statement of condition (6) confusing and should be fixed by renaming one of the two quantities.","section":"§2 vs §3.2"},{"comment":"The displayed Lipschitz condition contains a typo: ∥u2−u2∥ should presumably be ∥u1−u2∥.","section":"§2, Lipschitz assumption"},{"comment":"The phrase 'consistently outperform' is stronger than the data support for Navier–Stokes with ν=2,3, where the reported differences are only 0.04–0.06 percentage points. Reporting standard deviations or confidence intervals over the 20 trials would help assess these differences.","section":"§4.7, Table 7"}],"recommendation":"major_revision","confidential_remarks":"The paper has a promising method and an unusually broad numerical study, but the central theoretical claim is currently disconnected from the experiments. The gap is fixable: either verify the scale condition for the actual data, provide a theorem that covers the small-σ regime, or sharply circumscribe the claim that Theorem 3.1 explains the empirical results. The missing estimation/generalization proof and the test-set-leaked α selection should also be addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a genuinely useful incremental contribution to the random-feature operator-learning program, and the experiments consistently show the regularized version helping on noisy benchmarks. The main caveat is that the headline theorem's condition (6) is never checked in the parameter regime they test, and it looks likely to be violated there.\n\nWhat is new: a concentration bound for random feature matrices with Student-t weights (which recovers Gaussian and Cauchy as edge cases), a frequency-weighted Tikhonov regularizer, and pairing with finite element reconstruction. The proof of Theorem 3.1 is coherent under its stated assumptions; the matrix Bernstein argument is clean. The numerical study is extensive—six PDE families, multiple ν, both raw RFF and FEM-recovered predictions, and released code. The qualitative claim that regularization suppresses noise is supported by the error tables and figures.\n\nThe soft spots are real but not fatal. Most important: condition (6) requires the Matérn characteristic function at the minimum separation κ to be ≤ η/m. With the reported scales (σ around 10^-2 to 10^-5), σκ is tiny for normalized input grids, φ(κ) is close to 1, and mφ(κ) ≫ 1. The paper never computes κ or reports whether (6) holds. So the conditioning theorem, as stated, does not connect to the experiments—the concentration guarantee may still hold, but it is not the reason we can trust the numerics. That needs fixing, either by verifying (6) numerically or by proving a different condition that covers small scales. I also note that α is selected by minimizing test error in Sections 4.2 and 4.3, which makes the reported test errors optimistic; Tables 7 and 10 lack standard deviations; and the neural operator comparisons are noise-free while kernel baselines' noise protocol is not stated. These are addressable with a moderate revision.\n\nWho gets value: anyone working on random features, kernel approximations, or surrogate models for PDEs with noisy data. The RRFF-FEM recipe is easy to implement and likely to become a baseline. It deserves a serious referee—I would send it out.\n\nRecommendation: send to peer review, asking the authors to check condition (6), move α selection to a validation split, add error bars, and match the noise protocols across baselines.","headline":"A solid, incremental random-feature method for noisy operator learning; the main theorem's condition (6) appears violated in the tested regime, but the empirical claims and Student-t generalization are worth referee time.","tokens_in":23343,"tokens_out":2112,"would_cite":true,"duration_ms":22338,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","68T05"],"pacs":[],"model":"deepseek-v4-flash","headline":"Frequency-weighted Tikhonov regularization makes random Fourier feature matrices well-conditioned at N ~ m log m features, and RRFF-FEM then learns noisy PDE operators more accurately and faster than unregularized random features.","keywords":["operator learning","random Fourier features","Tikhonov regularization","Student's t distribution","Matérn kernel","condition number","finite element interpolation","noisy PDE data"],"falsifier":"Compute the left-hand side of condition (6) with the experiment's parameters — take σ and ν from the parameter tables, κ equal to the smallest grid spacing, and m the number of output collocation points. In several reported settings σκ is small and the Matérn characteristic function is close to 1, so the product with m is much larger than η; checking whether the empirical spectral norm ‖(1/N)AA* − I_m‖₂ nevertheless stays below 2η with probability at least 1−δ would settle whether the concentration holds despite the unverified hypothesis.","tokens_in":22286,"feed_emoji":"🧮","tokens_out":9197,"duration_ms":89789,"temperature":0.7,"pith_summary":"The paper is trying to establish that a random Fourier feature model for operator learning can be made stable and accurate in the presence of noisy data by adding a frequency-weighted Tikhonov penalty, without sacrificing the speed that made random features attractive. It draws the random frequencies from multivariate Student's t distributions, a family that includes both Cauchy and Gaussian limits, and pairs the learned map with a finite element interpolant so predictions are functions on the output domain. The main theoretical result is a high-probability bound on the extreme singular values of the random feature matrix: if the number of features N grows like m log m in the number of training samples m, and a scale condition on the Matérn kernel is met, the matrix is well conditioned. Across advection, Burgers', Darcy, Helmholtz, Navier-Stokes, and structural mechanics benchmarks, the regularized method reports lower test error and faster training than the unregularized random feature baseline, and accuracy comparable to kernel and neural operator methods on the tested cases.","feed_headline":"Regularized random features beat plain random features on noisy PDEs","feed_subtitle":"Frequency-weighted penalty cuts high-frequency noise, improving test error and training time on six benchmark PDEs.","key_machinery":"The load-bearing object is the random feature matrix A∈C^{m×N} with entries exp(i⟨ω_k, x_j⟩). Its normalized Gram matrix (1/N)AA* concentrates around the kernel matrix whose entries are the Matérn characteristic function ϕ_ω(x_j−x_k) = E[exp(i⟨ω, x_j−x_k⟩)], the Fourier transform of the Student-t sampling density. Because this characteristic function is positive, decreasing in distance, and bounded by its value at the minimum spacing κ, standard row-sum bounds for Hermitian matrices control the deviation of the expected Gram matrix from the identity; a matrix concentration inequality then converts the per-entry bounds into the high-probability spectral concentration of Theorem 3.1. The secon","core_discovery":"On the paper's own terms, the central claim is Theorem 3.1: with feature weights drawn from the multivariate Student's t density (scale σ, degrees of freedom ν), if N ≥ C η^{-2} m log(2m/δ) and the Matérn characteristic function evaluated at the minimum separation κ of the training points is at most η/m, then with probability at least 1−δ the normalized Gram matrix satisfies ‖(1/N)AA* − I_m‖₂ ≤ 2η. This concentration makes the least-squares system well conditioned and the pseudoinverse well defined, which the paper says yields estimation and generalization bounds. The method itself solves, for each output component, a regularized least squares problem whose penalty ∑‖ω_k‖_2^p |x_k|² suppress","pith_inferences":["A natural parameter-selection rule follows from condition (6): pick σ and ν so that the Matérn characteristic function at the grid spacing falls below η/m; tuning σ this way could bring the experiments into the regime where the theorem applies.","Because ν=1 gives Cauchy and ν→∞ gives Gaussian weights, the Student-t family offers a one-parameter interpolation between heavy-tailed and light-tailed sampling, which could be matched to the regularity of a given PDE's solution operator.","The frequency penalty ∑‖ω_k‖^p |x_k|² can be read as an RKHS Sobolev-norm penalty; this connection suggests a testable comparison against direct Sobolev-space kernel methods on the same noisy benchmarks.","The FEM recovery map decouples the number of training samples from the output grid resolution; a competing implementation could replace it with learned recovery maps and measure how much of the reported gain comes from the interpolation step."],"forward_implications":["With N ~ m log m features and the scale condition met, the feature matrix is spectrally concentrated around the identity, so the regularized least-squares problem is stable and the pseudoinverse is well defined.","The frequency-weighted penalty suppresses high-frequency noise, which the experiments indicate lowers test error on all benchmark PDEs compared to the unregularized random feature model.","Because RRFF-FEM's output is a finite element interpolant, the learned operator can be evaluated at arbitrary points of the output domain, not just the training grid.","The experiments indicate that RRFF retains the speed advantage of random features, with training times below those reported for Matérn and RBF kernel methods on the same problems."],"fun_headline_variants":["Regularized random features tame noisy PDE operators","Frequency-weighted penalty boosts noisy PDE learning","RRFF-FEM: robust operator learning with guarantees","Noise-robust random Fourier features for operator learning"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The main guarantee requires the kernel's value at the shortest distance between any two training points to be at most η/m; in the paper's experiments the frequency scale is so small that this value is near 1, so the theorem's hypotheses appear to be unmet in the very regime where the numerical gains are shown, and the stated estimation/generalization bounds are inherited from earlier work rather than proved.","fun_headline_variants_meta":{"raw":{"variants":["Regularized random features tame noisy PDE operators","Frequency-weighted penalty boosts noisy PDE learning","RRFF-FEM: robust operator learning with guarantees","Noise-robust random Fourier features for operator learning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000212,"raw_usage":{"total_tokens":1275,"prompt_tokens":787,"completion_tokens":488,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":531,"completion_tokens_details":{"reasoning_tokens":429}},"tokens_in":531,"tokens_out":488,"duration_ms":6140,"temperature":1.0,"reasoning_tokens":429,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T15:07:12.900346+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the left-hand side of condition (6) with the experiment's parameters — take σ and ν from the parameter tables, κ equal to the smallest grid spacing, and m the number of output collocation points. In several reported settings σκ is small and the Matérn characteristic function is close to 1, so the product with m is much larger than η; checking whether the empirical spectral norm ‖(1/N)AA* − I_m‖₂ nevertheless stays below 2η with probability at least 1−δ would settle whether the concentration holds despite the unverified hypothesis.","supporting_citations":[],"review_version":1}