{"id":"f12b84bf-3fe6-4d17-a5ed-c630bfcf5515","arxiv_id":"2505.00508","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Weak RFM extends the random feature method to weak solutions by enforcing the governing equations through integrals against sinusoidal test functions and solving a linear least-squares system.","lead":"This paper introduces Weak RFM, a random-feature method that solves partial differential equations in weak form, using sinusoidal test functions and a linear least-squares system instead of neural network training. It reports faster and often more accurate results than physics-informed neural networks on several 2D and 3D benchmarks with non-smooth solutions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that a finite set of sinusoidal weak-form tests suffices is unsupported: Weierstrass uniform approximation does not control the derivatives appearing in Lφ, and the undefined window function makes density in the needed test space uncheckable.","rationale":"The reader's weakest assumption is the same one I would isolate: a finite set of sinusoidal test functions is asserted, via a uniform approximation theorem, to be sufficient to enforce the weak form. I agree. The concern is load-bearing because the abstract and Section 1 promise that the resulting solution 'satisfies the weak form by default'; if the finite test set is not dense in the norm that controls the adjoint operator L, that promise is false. Weierstrass's theorem in §3.1 is not the right tool: it gives uniform approximation of φ, not approximation of Lφ, and the weak form (5) uses Lφ. A related, independent error in the paper is the 'exact solution' of the static heat equation in §4.3: it is discontinuous across y=x and y=1-x, so the distributional Laplacian contains surface measures that the bounded RHS 9·sgn(...) cannot match. This strengthens the impression that the notion of weak solution is being applied too loosely, but it is not the central mechanism of the method; the finite-test density gap is. The reader's CONDITIONAL verdict already reflects the need to fix the theoretical support and the benchmark, and my analysis does not identify a different verdict. I would ask for (i) a precise definition of the window function, (ii) a convergence/density statement for the test family in the graph norm of L, and (iii) correction or removal of the invalid static-heat exact solution.","tokens_in":12147,"tokens_out":7676,"duration_ms":89970,"concrete_test":"Run WRFM on the 2D Helmholtz benchmark (9) with an explicit C∞ compactly supported window w (e.g. a tensor-product bump) and record the trained coefficients. For each K (the number of retained wave numbers), evaluate the out-of-sample weak residual r_l = |∫_Ω (u_M Lφ_l - f φ_l)| for wave numbers l = K+1,...,10K not used in (8), using the same w. If max_l |r_l| does not decrease as K increases, or is not comparable to the training residual, the finite test set is not enforcing the weak form. As a second check, repeat the whole experiment with a different window (e.g. a C^1 polynomial bump); if the computed solution and errors change materially, the unspecified 'window function' in §3.1 is a load-bearing parameter that must be prescribed. Both tests directly probe whether Eq. (8) encodes the weak problem.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the assertion in §3.1 that checking (5) for the finite test family φ_k = w_[x0,x1] sin(πk(x-x0)/(x1-x0)) is enough to make the computed u 'satisfy the weak form by default'. The only justification offered is the Weierstrass approximation theorem, which yields uniform approximation of a continuous function by trigonometric polynomials. Uniform convergence does not imply convergence of Lφ_k to Lφ, and the weak form (5) contains derivatives of φ through the adjoint operator L. Therefore the finite residual entering the loss of §3.2, or the rows of A1 in (8), does not control the true weak-form residual sup_{φ∈D} |∫_Ω (u Lφ - f φ)| unless the test span is dense in the graph norm of L. The window function w_[x0,x1] is never defined, so no such density statement can even be checked: with w a fixed bump, the products w·sin(k·) may fail to be dense in H^1_0; with w chosen differently for different k, the family may not form a usable test space. Since A2/A3 also impose boundary and interface conditions pointwise without a trace argument, the entire mechanism by which (8) is supposed to encode the weak problem is unverified. This is the central gap: if the finite test set is not dense in the appropriate norm, WRFM can produce a function with small finite-test residual that is not a weak solution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Weak Random Feature Method (WRFM), a mesh-free least-squares solver for linear PDEs in weak form. The solution is represented as a partition-of-unity combination of random feature functions, and the discrete equations are assembled from finitely many sinusoidal test functions multiplied by an unspecified window function, together with pointwise boundary and interface conditions. The method is tested on 2D Helmholtz, 2D static heat, 3D Poisson, and 3D heat problems and compared with PINN and weak adversarial networks (WAN). The central claims are that WRFM enforces the weak form by construction and achieves comparable or better accuracy with significantly fewer parameters and less computation time, all on CPU.","tokens_in":12336,"tokens_out":11727,"duration_ms":126308,"significance":"If the theoretical gap were closed, WRFM would be an interesting contribution: it is mesh-free, requires no iterative neural-network training, and solves a linear least-squares system, potentially offering an efficient alternative to PINN and WAN for problems with low-regularity data. The paper also makes a useful comparison with existing machine-learning PDE solvers and includes a sensitivity study for the main hyperparameters. However, the manuscript is not yet reliable: the finite-test-function justification is incomplete, and several benchmark reference solutions are not weak solutions of the displayed equations because distributional delta terms are omitted. No code or reproducibility artifacts are provided, which further weakens the empirical claims.","major_comments":[{"comment":"The window function w_[x0,x1] is never defined, so the claimed representation of D(R) test functions is not checkable. More importantly, the Weierstrass approximation argument is insufficient: uniform approximation of a continuous test function by sine series does not imply convergence of Lφ_k to Lφ, since L contains derivatives of φ through the adjoint operator. No error estimate is given relating the finite-test-set residual entering Eq. (8) to the true weak-form residual sup_{φ∈D} |∫_Ω (u Lφ - f φ)|. Without density in the graph norm of L, or an explicit residual bound, a function with small finite-test residual need not be a weak solution. The claim that the computed solution 'satisfies the weak form by default' is therefore unsupported.","section":"Section 3.1, Eq. (6)"},{"comment":"The reference functions used in several benchmarks are not weak solutions of the stated equations in the sense of Eq. (5). For Eq. (9), u_e = sinh(|x|) cos(y²) has a jump of 2 cos(y²) in ∂_x u at x = 0, so Δu_e + u_e equals the displayed RHS plus 2 cos(y²)δ(x). For Eq. (11), the boundary-data function (x - I[x-1/2])² sin(y)e^{-z} has a derivative jump across the internal plane x = 1/2, producing a δ(x-1/2) term in its Laplacian. For Eq. (12), delta terms appear along x = 0 and y = 0. In each case the stated smooth RHS lacks these distributions, so the experiments do not actually test approximation of a weak solution of the displayed problem.","section":"Sections 4.2, 4.4, 4.5, Eqs. (9), (11), (12)"},{"comment":"The displayed 'weak solution' for the static heat equation is discontinuous along y = x and x + y = 1; for example, at (0.8, 0.8) the one-sided limits are approximately +0.31 and -0.31. Its distribution Laplacian therefore contains line Dirac terms that are not present in the RHS 9 sgn((x-y)(y+x-1)), so it does not satisfy the weak form of Eq. (10). The reported L∞ errors around 0.8-1.2 are dominated by smearing these discontinuities, and the claim that PINN and WRFM 'successfully solve' the problem is not supported by comparison with an invalid reference solution.","section":"Section 4.3, Eq. (10)"},{"comment":"Boundary and interface conditions are imposed as pointwise equalities at subdomain endpoints or along axis partitions, but no trace argument or quadrature error analysis connects these discrete conditions to the weak boundary integral in Eq. (5). No convergence study in the number of test functions K or in quadrature order is reported; Section 4.6 only varies the number of subdomains S and feature functions J. The paper therefore does not establish that increasing the discrete test set drives the weak-form residual to zero.","section":"Sections 3.2 and 4.1"}],"minor_comments":[{"comment":"The symbol I is called the 'signal function'; it should be called the indicator function, and the term should be defined consistently.","section":"Section 2.2"},{"comment":"The symbol φ^c_[x0,x1] is used but never defined; the compactly supported window used in numerical examples needs an explicit formula or reference.","section":"Section 3.1, Eq. (6)"},{"comment":"The paper states that each experiment is repeated three times and averages are reported, but no standard deviations are given in Tables 1-5; the sensitivity study in Section 4.6 reports standard deviations only for the Helmholtz example.","section":"Tables 1-5"},{"comment":"Time comparisons are not fully apples-to-apples: WRFM runs on CPU while PINN and WAN use a GPU, so the reported speed advantages should be interpreted with caution.","section":"Section 4.1"},{"comment":"Decimal commas (e.g., 0,169) are inconsistent with standard English mathematical notation and should be replaced by decimal points; there are also typographical errors such as 'Poisson‘s equation'.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The core numerical idea is worth pursuing, but the manuscript requires substantial revision: the theoretical justification of the finite test set must be made rigorous, and the benchmark equations and reference solutions need to be corrected so that the displayed weak forms are actually satisfied by the reference functions. I would encourage a revision rather than rejection, since the identified issues are fixable within the scope of the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nWhat you should know: this paper extends the random feature method (RFM) to weak solutions by building a least-squares system from integrated-by-parts adjoint tests and Fourier test functions. When the pieces work, it is genuinely faster and uses far fewer parameters than PINN/WAN on the 2D and 3D examples shown. The Helmholtz and Poisson results are suggestive. Credit where due: the authors do not tune coefficients to match reference solutions; the coefficients come from solving a linear system, so the comparison is not circular. The paper is a natural extension rather than a leap, but it is not a copy of prior work.\n\nThe soft spots are real. The theoretical justification has a load-bearing gap. They assert that a finite set of sinusoidal test functions multiplied by an undefined window function w[x0,x1] suffices to certify a weak solution. Uniform convergence by Weierstrass does not control convergence of Lφ_k to Lφ, because L contains derivatives. Without density in the graph norm of L, the finite row-block A1 does not control the true weak residual. The window function is never defined, so the claim is not checkable. Also missing: a trace argument for pointwise boundary conditions in a weak formulation, and a statement of well-posedness.\n\nWorse: the 'exact solution' for the 2D static heat problem (10) is not a weak solution of the stated equation. It is discontinuous across y=x and y=1-x, so its distributional Laplacian carries surface delta terms that cannot equal the bounded function 9·sgn(...), and it does not satisfy the stated zero boundary data on the inner square. That benchmark should be removed or corrected.\n\nThe paper also never runs standard RFM as a baseline, despite the method being an extension of it, and ships no code or data. Both are omissions that weaken the empirical claims.\n\nWho is this for: researchers using RFM or kernel/random-feature methods for PDEs, and anyone wanting a cheap linear-algebra alternative to PINN/WAN on low-regularity problems. The core idea may be salvageable, but as submitted the central claim is not supported. I would send it to peer review anyway—the method is novel enough and the experiments are detailed enough to warrant a serious referee, with a request for major revision. If the authors fix the density argument, define the window, correct the benchmark, and add an RFM baseline, the paper could become a solid contribution.","headline":"Weak RFM is a plausible, fast linear-solve extension of random feature methods to weak solutions, but the test-function density argument is unproven and one benchmark's exact solution is not actually a weak solution.","tokens_in":12995,"tokens_out":3210,"would_cite":false,"duration_ms":31286,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N35","65M70","35D30","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that reformulating the random feature method around the weak form of a PDE yields a mesh-free solver that computes approximate weak solutions of low-regularity problems from a single least-squares system, with accuracy…","keywords":["partial differential equations","weak solutions","random feature method","least squares","test functions","partition of unity","physics-informed neural networks","weak adversarial networks"],"falsifier":"Take the 2D Helmholtz problem of Section 4.2 with the optimal hyperparameters, then double $K_x,K_y$ while holding $J,S$ fixed and measure the $L^2$ error against the exact weak solution. If the error plateaus and stops shrinking while the finite-set residual is small, the finite test-function set is not capturing the full weak problem; alternatively, evaluating the residual against a test function with frequency outside the chosen set and large amplitude should reveal whether the weak form holds beyond the finite set.","tokens_in":11772,"feed_emoji":"📐","tokens_out":6343,"duration_ms":61861,"temperature":0.7,"pith_summary":"The paper proposes Weak Random Feature Method (WRFM), a mesh-free numerical solver aimed at linear PDEs whose data or boundary conditions are too rough to admit classical solutions. It claims that replacing the strong PDE residual with integrals against a finite set of sine test functions, supported by window functions, turns the weak problem into one least-squares linear system. On benchmark problems in 2D and 3D with nonsmooth or discontinuous data, WRFM reportedly matches or beats weak-adversarial-network and physics-informed baselines in accuracy while using fewer parameters and much less compute. If the claim holds, WRFM gives a simple non-iterative route to approximate weak solutions, the class needed for many fluid and material models.","feed_headline":"Weak random features solve PDEs with no classical solution","feed_subtitle":"By testing equations against sine waves, the method skips neural training and beats PINN and WAN on nonsmooth benchmarks.","key_machinery":"The load-bearing object is the weak-form linear system of Eq. (8). Its first block has entries $\\int_{\\Omega_n} \\phi_j(x)\\, \\mathcal{L}^*\\varphi_i^n(x)\\,dx$ with $\\mathcal{L}^*$ the formal adjoint of the PDE operator; the second block pins boundary values at block endpoints; the third block enforces $C^0$ continuity between neighboring blocks. The test functions are windowed sine products, $\\varphi_i^n = w_{[x_0^n,x_1^n]}\\sin\\!\\big(\\pi i(x-x_0^n)/(x_1^n-x_0^n)\\big)$ in 1D with tensor-product analogues in higher dimensions. The window truncates support to the block; the sinusoids supply the density argument that a finite set of test functions approximates the full space $D(\\Omega)$. This construction is what removes derivative requirements on $u$ and eliminates the inter-block differentiability conditions present in standard RFM.","core_discovery":"The central discovery is that the random feature method can be extended from strong solutions to weak solutions by swapping the pointwise equation $Lu=f$ for the adjoint-form identities $\\int_\\Omega u\\, L^*\\varphi\\,dx = \\int_\\Omega f\\,\\varphi\\,dx$ and the corresponding boundary identities. The solution is still a partition-of-unity sum of random features $\\sigma(k_m\\cdot x+b_m)$; the new ingredient is the test-function set $\\varphi_k = w_{[x_0,x_1]}\\sin\\!\\big(\\pi k(x-x_0)/(x_1-x_0)\\big)$ (and tensor products in 2D/3D), which by Weierstrass approximation the authors take to span all compactly supported test functions. The coefficients solve the linear system (8) whose interior block is the Gram-like matrix of $\\int_{\\Omega_n}\\phi_j\\, L^*\\varphi_i\\,dx$. The method enforces boundary values and interface continuity but, unlike strong RFM, imposes no differentiability across subdomains, which is exactly why it tolerates nonsmooth solutions.","pith_inferences":["If a rigorous error bound were added linking the finite-set test residual to the true weak-form residual, WRFM would become a provably convergent mesh-free weak solver; the paper supplies only a Weierstrass density argument, so convergence rates are an open question.","The same windowed-sine test-function ansatz could be applied to nonlinear PDEs by minimizing the weak residual, though the coefficients would then come from a nonlinear least-squares problem rather than a single linear solve.","The adversarial idea used in WAN could be imported here as the authors suggest: generate test functions adaptively rather than by tensor product, which might remove the exponential growth in dimension.","Replacing the window with smoother or adaptive partitions could improve accuracy for discontinuous solutions; the current experiments show errors concentrated along discontinuity lines, so window placement is likely a leverage point."],"forward_implications":["Approximate weak solutions to linear PDEs with nonsmooth boundary data or right-hand sides can be computed by one linear solve, with no iterative network training.","For the tested 2D Helmholtz, 2D static heat, 3D Poisson, and 3D heat problems, WRFM reaches accuracy comparable to or better than PINN and WAN with far fewer parameters and, in 2D, roughly an order-of-magnitude speedup on CPU only.","Because the system imposes no cross-block differentiability, the method localizes nonsmooth features better than strong-form solvers and does not smear kinks as much.","The number of test functions is a tensor product in dimension, so current WRFM scalability in 3D is limited by the size of the linear system; the authors identify this as the main current limitation.","Accuracy depends mainly on the product $M=JS$ (features per block times number of blocks), with a broad plateau of good values, so hyperparameters are reportedly easy to choose."],"supporting_citations":[{"why":"Supplies the base random feature method (random features, partition of unity, least-squares linear system) that WRFM modifies.","marker":"[15]"},{"why":"Justifies replacing the infinite set of compactly supported test functions by finite trigonometric combinations via the approximation theorem.","marker":"[24]"},{"why":"Introduces the weak adversarial network baseline to which WRFM compares in accuracy, time, and parameter count.","marker":"[25]"},{"why":"Defines the physics-informed neural network baseline used for the nonsmooth benchmark problems.","marker":"[11]"}],"fun_headline_variants":["Weak RFM solves nonsmooth PDEs without neural training","Random features meet weak forms for faster PDE solving","Skipping strong solutions: Weak RFM for nonsmooth problems","Adjoint-form random features solve weak PDEs efficiently","Weak RFM matches PINN accuracy with far less compute"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that satisfying the weak equation on a finite list of chosen sine-shaped test functions is enough to force the computed function close to the real weak solution; the paper justifies this by approximation but does not quantify how many test functions are needed or bound the error.","fun_headline_variants_meta":{"raw":{"variants":["Weak RFM solves nonsmooth PDEs without neural training","Random features meet weak forms for faster PDE solving","Skipping strong solutions: Weak RFM for nonsmooth problems","Adjoint-form random features solve weak PDEs efficiently","Weak RFM matches PINN accuracy with far less compute"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000295,"raw_usage":{"total_tokens":1739,"prompt_tokens":997,"completion_tokens":742,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":662}},"tokens_in":613,"tokens_out":742,"duration_ms":7429,"temperature":1.0,"reasoning_tokens":662,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:41:02.607424+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the 2D Helmholtz problem of Section 4.2 with the optimal hyperparameters, then double $K_x,K_y$ while holding $J,S$ fixed and measure the $L^2$ error against the exact weak solution. If the error plateaus and stops shrinking while the finite-set residual is small, the finite test-function set is not capturing the full weak problem; alternatively, evaluating the residual against a test function with frequency outside the chosen set and large amplitude should reveal whether the weak form holds beyond the finite set.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the base random feature method (random features, partition of unity, least-squares linear system) that WRFM modifies."},{"cited_title":"Weierstrass, ¨Uber die analytische Darstellbarkeit sogenannter willk¨ urlicher Functionen einer reellen Ver¨ anderlichen, Sitzungsberichte der Akademie zu Berlin 1 (1885) 633–639","cited_arxiv_id":null,"evidence_quote":"Justifies replacing the infinite set of compactly supported test functions by finite trigonometric combinations via the approximation theorem."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the weak adversarial network baseline to which WRFM compares in accuracy, time, and parameter count."},{"cited_title":"Raissi, P","cited_arxiv_id":null,"evidence_quote":"Defines the physics-informed neural network baseline used for the nonsmooth benchmark problems."}],"review_version":1}