{"id":"8ea0284a-276a-4b43-bf82-2f3caf415b07","arxiv_id":"2412.02586","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An adaptive neural network subspace method, using tensor neural networks and a posteriori error estimators, solves 2D elliptic PDEs with singularities and interface discontinuities to relative errors as low as 1e-9.","lead":"Neural networks provide the trial space, and the PDE solution is found by solving a Galerkin linear system while training the network's parameters with either the Ritz energy or a residual-based error estimator as the loss. The authors report relative errors as low as 1e-5 to 1e-9 on 2D problems with singularities and discontinuous coefficients, and argue that high-accuracy quadrature, not network size, is the key.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim depends on hand-constructed quadrature that exploits the exact location and exponent of the singularity; the paper only tests this bespoke setting, so the general high-accuracy claim is not established.","rationale":"The paper's main thesis—that integration error controls machine-learning accuracy—is convincingly supported by the negative control, and the theoretical error decomposition (2.15) is standard. My stress-test focused on the condition under which ε_int is actually small. The quadrature rules used in the experiments are bespoke: they exploit the exact singularity lines and exponent in Section 3 and the straight or circular interface geometry in Section 4. This makes the demonstrated high accuracy conditional on oracle knowledge of the PDE's singular structure. The paper does not provide an automatic way to detect the singular set or choose the weighted quadrature, so the abstract's general claim overstates the demonstrated scope. This reinforces the reader's CONDITIONAL verdict rather than overturning it. I did not find an internal inconsistency in the main derivations: Theorem 2.1 holds for any trial set, and the a posteriori bounds in Section 4 are correctly derived. The absence of code, error bars, and relevant baselines remains a separate reproducibility concern that also supports CONDITIONAL. The proposed test—perturbing the singularity location while keeping the quadrature fixed—directly establishes whether the method's accuracy is an artifact of the tailored quadrature.","tokens_in":27953,"tokens_out":16711,"duration_ms":161239,"concrete_test":"Re-run the singular problem of Section 3 with the exact singularity moved to x1=0.37 while keeping the quadrature split at x1=0.5, using the same TNN, rank p=100, and optimizers. If the final test-point relative error degrades from about 2e-5 to O(1e-2), the method's high accuracy is contingent on a priori knowledge of the singularity location. As a complement, verify quadrature convergence in the original setting by varying Gauss-Jacobi orders n=50, 100, 200, 400 and composite Gauss-Legendre subintervals 50, 100, 200, and check that the reported 2.19e-5 is not limited by unconverged quadrature.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The error decomposition (2.15) makes the high-accuracy claim conditional on ε_int being negligible, but the paper only demonstrates this for integrands whose singular set and singular exponent are known in advance. In the singular example (Section 3), the quadrature splits the domain exactly at x1=1/2 and x2=1/2 and uses (α,β)-Gauss-Jacobi rules with α,β=-2/3, i.e., the exact singularity exponent. The paper's own negative control—4000 uniform Gauss-Legendre points per dimension degrade the test-point error from 2.19e-5 to 4.60e-2—shows that accuracy is not a property of the NN subspace or the loss alone, but of the tailored quadrature. The interface examples use composite Gauss-Lobatto grids aligned with the straight interface x=2/3 (or polar coordinates for the circle in Section 4.3), again requiring problem-specific geometry. No experiment tests a singularity of unknown location or exponent, nor a curved interface not aligned with the quadrature grid, so the central claim as stated—'solving PDEs with singularities and discontinuous coefficients with high accuracy'—is only supported when such prior knowledge is available. Without a mechanism to construct the required quadrature automatically, ε_int in (2.15) is uncontrolled in the general case.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an adaptive neural network subspace method for second-order elliptic PDEs. A neural-network-generated basis spans a finite-dimensional subspace, a Galerkin/Ritz solve determines the coefficients, and training updates the network parameters. The authors give an error decomposition into approximation error, numerical integration error, and optimization error, and they argue that high-accuracy quadrature, rather than network expressivity alone, is the key to accurate PDE solutions. They use either a Ritz loss or an a posteriori error-estimator loss derived from a complementarity identity, and they demonstrate the method on a singular Laplace problem and on elliptic interface problems with discontinuous coefficients. Reported test-point relative errors are around 2e-5 for the singular problem and 1e-8 to 1e-9 for straight-interface problems, with a curved-interface example reaching about 5.6e-7. A negative control with uniform Gauss-Legendre quadrature degrades the singular-problem error to 4.6e-2, supporting the paper's emphasis on integration accuracy.","tokens_in":28074,"tokens_out":6846,"duration_ms":71267,"significance":"If the claims are accepted, the paper makes a useful and partially falsifiable point: for low-dimensional elliptic problems, the accuracy ceiling of neural-network methods can be set by the quadrature used to evaluate the loss, not by network capacity. The numerical evidence is strong within the tested problem classes, and the explicit negative control (4000 uniform Gauss-Legendre points per dimension versus tailored Gauss-Jacobi quadrature) is a valuable experiment. The complementarity-based a posteriori estimator (Lemma 3.2) gives an exact identity that clearly explains why the residual-based loss is a sensible objective. The main limitation is that the reported high accuracy relies on quadrature rules constructed from exact knowledge of the singularity location/exponent and of the interface geometry; the paper does not yet provide an automatic mechanism for the general case, so the broad high-accuracy claim is conditional on this prior knowledge.","major_comments":[{"comment":"The singular-problem experiment uses (α,β)-Gauss-Jacobi quadrature with α,β=-2/3 after splitting the domain at x1=1/2 and x2=1/2, i.e., the exact location and exponent of the singularity are encoded in the quadrature rule. The paper gives no procedure for constructing such quadrature when the singularity is unknown in advance, so ε_int in inequality (2.15) is uncontrolled in the general case. The negative control in Section 3 (4000 uniform Gauss-Legendre points per dimension giving 4.601047e-2 versus 2.190126e-5 with Gauss-Jacobi quadrature) confirms that the reported high accuracy is a property of the tailored quadrature rather than of the NN subspace or the loss alone. Please add an automatic quadrature-error control strategy or explicitly state that the high-accuracy claim is restricted to problems whose singularity location and exponent are known and used in the quadrature.","section":"Section 3, quadrature for loss (3.23)"},{"comment":"The central error decomposition (2.15) is introduced through inequality (2.13), but ε_int is never defined, and the relation between the quadrature error in the loss functional and the energy-norm difference ∥u^(1)_NN - u^(2)_NN∥_a is not proved. As written, (2.13) is a heuristic statement ('ε_int × some norm of u^(1)_NN'), so the conclusion that integration error always controls machine learning accuracy is not established by the analysis. A rigorous version should specify the quadrature error in a suitable dual norm and the stability constant of the Ritz problem; the numerical section could then be connected to this estimate quantitatively.","section":"Section 2, Eqs. (2.13)-(2.15)"},{"comment":"The interface experiments are also built on problem-specific quadrature: composite Gauss-Lobatto rules aligned with the straight interfaces x1=2/3 and x1=2/9, and polar-coordinate quadrature for the circular interface in (4.22). For a curved interface that is not aligned with the quadrature grid and is not given by a simple parametrization, the loss (4.26) cannot be evaluated accurately by the described schemes, so the interface high-accuracy claim is conditional on the interface geometry being known and quadrature-compatible. Please state this limitation and, if the general claim is retained, provide a strategy for automatic interface quadrature (e.g., interface-fitted quadrature or local refinement).","section":"Section 4.3, loss (4.26) and quadrature"}],"minor_comments":[{"comment":"Algorithm 2 heading and Step 2 contain typos ('methd', 'wiht'); Section 3 contains 'It ie easy'; Section 4.1 contains 'adpative' and uses 'un' instead of u_N in the proof of Theorem 4.1; please correct all typographical errors.","section":"Throughout"},{"comment":"The step from (2.11) to (2.12) is the parallelogram law (polarization identity), not the binomial theorem; please rephrase.","section":"Section 2, Eqs. (2.11)-(2.12)"},{"comment":"Theorem 4.1 involves the constants λmin and λ1,Γ, but the loss (4.14) drops the multiplicative factor (1+1/λmin+1/λ1,Γ). Since the factor is independent of u_N, the loss remains a valid objective, but the relation between L(u_N) and ∥u-u_N∥_a should be stated as an equivalence with a constant rather than as an equality.","section":"Section 4, Theorem 4.1 and loss (4.14)"},{"comment":"The abstract mentions the hypercircle technique, but Section 3 derives the estimator via Green's formula and the complementarity identity (3.14); the connection to the hypercircle/complementarity method of [28] should be made explicit in the text.","section":"Abstract and Section 3"},{"comment":"Table 6 reports etest = 8.96e-01 for standard PINN and 5.20e-04 for nDS-PINN without specifying the network sizes, training schedules, or quadrature used for those baselines; please indicate whether these numbers are taken verbatim from [37] and cite the corresponding settings.","section":"Table 6"},{"comment":"The notation in the definition of f and the exact solution (3.21) uses x for the point and later x1, x2 for components; please make the notation uniform.","section":"Section 3, Eq. (3.21)"}],"recommendation":"major_revision","confidential_remarks":"The numerical study is extensive and the error numbers are impressive, but no code or reproducibility artifact is provided, which limits verification of the exact training schedules and quadrature counts. The relationship to the authors' earlier TNN work [30,32] should be clarified: much of the error analysis in Section 2 is standard Hilbert-space material, while the new contribution is the low-dimensional singular/interface application with tailored quadrature. In its current form, the title and abstract overstate the generality of the high-accuracy claim; a major revision should align the claims with the demonstrated scope and address the missing automatic quadrature-error control."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper makes a clean, mostly convincing case that the limiting factor in neural PDE solvers is not the network but the quadrature, and it demonstrates real accuracy on singular and interface problems by replacing Monte Carlo with problem-adapted Gauss rules. It deserves a serious referee, with conditions.\n\nWhat is actually new: the authors take the TNN subspace approach from their earlier work, add an a posteriori error-estimator loss via the hypercircle technique, and show that the combination works in 2D for a line-singularity problem and for several interface problems with discontinuous coefficients. The most valuable experiment is the negative control. With the exact same network and optimizer, 4000 uniform Gauss-Legendre points per dimension give a test-point error of 4.6e-2, while Gauss-Jacobi quadrature tailored to the x^(-2/3) singularity gives 2.19e-5, and the interface cases reach 1e-8 to 1e-9. That isolates integration error as the dominant term in inequality (2.15) about as cleanly as one could ask. The ResNet variant is also a sensible robustness check.\n\nSoft spots, in proportion. First, the quadrature is hand-built for known singular locations and exponents. The line singularity sits exactly at x1 = 1/2 and x2 = 1/2 with exponent -2/3; the interface is straight or circular and aligned with the quadrature grid. There is no mechanism to discover those features automatically, so the general claim about solving PDEs with singularities is only supported in this bespoke setting. I think that is a real scope limitation, though the paper's stated point—integration error controls accuracy—is still proven by the negative control. Second, the singular example uses only the Ritz loss; the advertised a posteriori-estimator loss is not exercised there. That weakens the narrative that the estimator is what drives the adaptivity, even if the estimator performs well in the interface tests. Third, no code or data are provided, and every reported error is a single run with no error bars. At the stated accuracies, a few independent runs would materially increase confidence. Fourth, the comparison set is thin: for the multi-material problem they compare against two PINN variants, but not against their own prior TNN method or other subspace NN baselines. The text also has typos and refs [6] and [21] carry the same arXiv ID—likely a copy-paste slip, but worth fixing.\n\nSum: the central claim is supported inside the tested regime. I would accept this for review and ask for code/data, repeated runs, and a paragraph that honestly addresses how the required quadrature could be constructed when the singular set is unknown. For anyone working in scientific machine learning, this is a useful and citable data point.","headline":"The paper's strongest point is the controlled experiment showing quadrature accuracy, not network architecture, limits neural PDE solvers, and it earns a conditional peer-review pass despite the hand-crafted quadrature and missing baselines.","tokens_in":794,"tokens_out":1652,"would_cite":true,"duration_ms":34778,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65N30","65N25","65L15","65B99"],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural-network solver for PDEs reaches high accuracy only when the integrals in its loss are computed accurately, not merely when the network is big.","keywords":["neural network subspace","tensor neural network","a posteriori error estimator","hypercircle technique","Gauss-Jacobi quadrature","elliptic interface problems","singular solutions","Ritz loss"],"falsifier":"Use the same subspace training on the singular Laplace problem but with an off-center singularity, say at $x_1=0.37$, and only uniform Gauss-Legendre quadrature; if the relative test error does not approach the 1e-5 level as the number of quadrature points grows, the paper's claim that integration accuracy is the controlling factor would be falsified.","tokens_in":27622,"feed_emoji":"🧮","tokens_out":5606,"duration_ms":53490,"temperature":0.7,"pith_summary":"The paper argues that the accuracy of a neural-network solution to a PDE is governed less by the network's expressive power than by the accuracy of the numerical integrations inside the loss function. It builds a finite-dimensional subspace whose basis functions are outputs of a neural network, obtains the discrete solution by a Galerkin projection, and trains by adaptively updating the subspace with a Ritz loss or a hypercircle-based a posteriori error estimator. Tested on a Laplace problem with line singularities and on elliptic interface problems with discontinuous and highly contrasted coefficients, the method achieves relative errors on uniform test grids of roughly 2.19e-5 for the singular case and 1e-8 to 1e-9 for the interface cases. The negative control is explicit: using 4000 uniform Gauss-Legendre points per dimension instead of singularity-adapted Gauss-Jacobi quadrature degrades the singular-problem test error from 2.19e-5 to 4.6e-2. A sympathetic reading is that high-accuracy integration, not network size, is the load-bearing ingredient.","feed_headline":"Quadrature error, not network size, decides neural PDE accuracy","feed_subtitle":"A neural subspace method hits 1e-8 errors on interface problems once loss integrals use matched quadrature.","key_machinery":"The central mechanism is the neural-network subspace $V_p := \\mathrm{span}\\{\\varphi_j(x;\\theta),\\,j=1,\\dots,p\\}$ built from the outputs of a network, over which a Galerkin or Ritz problem is solved exactly by assembling stiffness matrix and right-hand side; training then updates the parameters $\\theta$ by minimizing either the Ritz energy or the a posteriori error estimator $\\eta(u_p,y)$ from the hypercircle and complementarity theory, which upper-bounds the energy error. The quadrature rules that carry the argument are Gauss-Jacobi quadrature for singular factors of the form $(x-\\tfrac12)^{-2/3}$, composite Gauss-Legendre for regular terms, and composite Gauss-Lobatto on subdomains and interfaces; these make the integration error $\\varepsilon_{int}$ negligible for the non-polynomial neural-network integrands. The a posteriori loss also measures the dual error through the identity $\\eta^2(\\psi,y)=\\|u-\\psi\\|_a^2+\\|y^*-y\\|_*^2$, so minimizing it drives both the primal and dual approximations.","core_discovery":"On its own terms, the paper's central discovery is that neural-network subspace approximation can solve second-order elliptic boundary value problems with singular solutions and discontinuous coefficients to high accuracy, provided the integrals in the loss are evaluated by quadrature rules matched to the integrand's singularities or the interface. The analysis decomposes the final error into three terms: the approximation error $\\inf_{v\\in V_{NN}}\\|u-v\\|_a$, the integration error $\\varepsilon_{int}$, and the optimization error $\\varepsilon_{opt}$, and concludes that integration error controls the accuracy when Monte-Carlo or sampling methods are used. In the singular example, Gauss-Jacobi quadrature for the $(x_i-\\tfrac12)^{-2/3}$ factors yields a test-point relative error of 2.19e-5, whereas uniform Gauss-Legendre quadrature with 4000 points per dimension yields 4.6e-2. For interface problems, the a posteriori error-estimator loss with composite Gauss-Lobatto integration gives test-point errors of order 1e-8 to 1e-9, including a circular-inclusion problem with non-homogeneous boundary data. The natural conclusion drawn in the paper is that improving the accuracy of the integrations in the loss is the essential step toward high-accuracy neural PDE solvers.","pith_inferences":["A natural next test is an interface problem whose geometry is not aligned with tensor-product quadrature, such as an ellipse or a moving front; the method would need an adaptive quadrature construction to keep the integration error negligible.","Because the a posteriori loss also measures the dual error, it could be repurposed as a stopping criterion or as an indicator for choosing the subspace dimension $p$ rather than fixing it.","The quadrature-first principle suggests that extending the method to time-dependent or nonlinear PDEs will require quadrature rules matched to the evolving solution's singularities rather than simply larger networks."],"forward_implications":["Replacing Monte-Carlo loss evaluation with matched quadrature should improve the accuracy of existing neural PDE methods even when their architecture is unchanged.","The a posteriori error-estimator loss can act as a built-in error indicator, so the subspace can be refined adaptively without a separate mesh or an external error estimator.","On the singular benchmark, the machine-learned subspace reaches a test-point relative error around 2e-5, whereas the adaptive linear and quadratic finite element results reported in the paper sit near 2.3e-3 in relative L2 error.","Interface problems with coefficient contrasts up to 4000 and high-frequency solution components are solved to test-point errors of order 1e-6 to 1e-9, indicating that the adaptively updated subspace can handle low global smoothness."],"supporting_citations":[{"why":"Supplies the deep Ritz baseline whose Monte-Carlo loss integration is the target of the paper's integration-error criticism.","marker":"[9]"},{"why":"Provides the hypercircle and complementarity a posteriori error estimates, including the identity used to build the posterior error-estimator loss.","marker":"[28]"},{"why":"Introduces tensor neural networks and their high-accuracy numerical integration, the foundation of the subspace and quadrature approach.","marker":"[30]"},{"why":"Establishes the TNN-based machine learning method with a posteriori error estimators that the present paper adapts to singular and interface problems.","marker":"[32]"},{"why":"Supplies the Gauss-Jacobi quadrature formula and its exactness properties used to compute singular integrals in the loss.","marker":"[25]"},{"why":"Provides the singular Laplace benchmark problem with known exact solution that the paper uses to demonstrate the effect of quadrature accuracy.","marker":"[22]"},{"why":"Supplies convergence theory for elliptic and parabolic interface problems and the background remark that the gradient is not in $H(\\mathrm{div},\\Omega)$ across the interface.","marker":"[5]"},{"why":"Gives a physics-specialized neural network baseline that the paper compares against on the multi-material diffusion problem.","marker":"[35]"},{"why":"Gives a physics-informed neural network baseline whose test error is compared with the proposed method in the multi-material example.","marker":"[37]"}],"fun_headline_variants":["Neural PDE accuracy hinges on integral quadrature, not network size","Adaptive neural subspace method: integration error sets the limit","For neural PDE solvers, quadrature precision outranks network depth","Matched quadrature yields 1e-8 errors in neural PDE interface problems"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that for each test problem a quadrature rule can be chosen so that the integration error in the loss is negligible for the neural-network integrands; if no such rule is available or the rule does not match the singularity or interface, the high accuracy claimed by the paper collapses.","fun_headline_variants_meta":{"raw":{"variants":["Neural PDE accuracy hinges on integral quadrature, not network size","Adaptive neural subspace method: integration error sets the limit","For neural PDE solvers, quadrature precision outranks network depth","Matched quadrature yields 1e-8 errors in neural PDE interface problems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000513,"raw_usage":{"total_tokens":2458,"prompt_tokens":878,"completion_tokens":1580,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":1503}},"tokens_in":494,"tokens_out":1580,"duration_ms":13011,"temperature":1.0,"reasoning_tokens":1503,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:18:41.877945+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Use the same subspace training on the singular Laplace problem but with an off-center singularity, say at $x_1=0.37$, and only uniform Gauss-Legendre quadrature; if the relative test error does not approach the 1e-5 level as the number of quadrature points grows, the paper's claim that integration accuracy is the controlling factor would be falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the deep Ritz baseline whose Monte-Carlo loss integration is the target of the paper's integration-error criticism."},{"cited_title":"Vejchodsk` y, Complementarity based a posteriori error estimates and their properties, Math","cited_arxiv_id":null,"evidence_quote":"Provides the hypercircle and complementarity a posteriori error estimates, including the identity used to build the posterior error-estimator loss."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Gauss-Jacobi quadrature formula and its exactness properties used to compute singular integrals in the loss."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the singular Laplace benchmark problem with known exact solution that the paper uses to demonstrate the effect of quadrature accuracy."},{"cited_title":"Chen and J","cited_arxiv_id":null,"evidence_quote":"Supplies convergence theory for elliptic and parabolic interface problems and the background remark that the gradient is not in $H(\\mathrm{div},\\Omega)$ across the interface."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives a physics-specialized neural network baseline that the paper compares against on the multi-material diffusion problem."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives a physics-informed neural network baseline whose test error is compared with the proposed method in the multi-material example."}],"review_version":1}