{"id":"79ecc0d5-a829-437f-8711-df8850979b2f","arxiv_id":"2507.16380","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A two-layer PINN can be trained by SGD to O(epsilon) loss with width independent of the number of samples, provided the target lies in a custom function class and the SGD trajectory does not explode.","lead":"This paper proves that SGD can train a two-layer physics-informed neural network to low loss with a width that depends on accuracy but not on sample count, under a custom function class assumption. The result narrows the gap between over-parameterization theory and narrow networks in practice, but the main theorem hinges on an unproven assumption that the SGD trajectory stays bounded.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The proof of Theorem 3.4 relies on unproved Assumption (3.1) (bounded SGD trajectory), which is absent from the theorem statement and abstract, so the central convergence claim is not established.","rationale":"The reader's weakest assumption is the right one. Equations (3.9)-(3.14) show that the bound on the SGD displacement uses Assumption (3.1) at the start, and the rest of the proof is a chain of estimates that cannot be started without it. This is not a stylistic complaint: the theorem statement and abstract omit (3.1), so the advertised claim is stronger than the proof. The numerical experiment in Section 5 cannot fill this gap because it tests one target and one initialization scale, while the theory claims a proof for all f in F. I also note the secondary issue that F is defined through Λ, which depends on m, so the interpretation of ||f||_F needs care; however, the decisive gap remains (3.1). If a proof of (3.1) from the stated hyperparameters can be supplied, the result may be recoverable; as submitted, it is not established, which supports the reader's REJECT verdict.","tokens_in":27086,"tokens_out":15431,"duration_ms":158136,"concrete_test":"Re-derive Theorem 3.3(a)-(c) with R(t)=max_i ||w_i^(t)|| left as a variable instead of invoking (3.1), and attempt to close an induction using only the stated hyperparameter choices η=Θ(ε/m), T≤T0, and the initial scale m^{-β}. Check whether the resulting recursive bound implies R(t)≤O(1) for all t≤T whenever m≥M; if the induction requires an additional condition not stated in Theorem 3.4, then the theorem must be made conditional on (3.1).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Assumption (3.1) is load-bearing at the very first estimate of the optimization proof. In Theorem 3.3(a), inequality (3.9) bounds |∂ψ/∂w_ij| by O(1)|a_i^(0)| \"using (2.7) and assumption (3.1)\". This is where boundedness of w_i^(t) and ψ controls powers of (w_i^(t)^T x + b_i^(0)) appearing in the derivative. If (3.1) fails, the right side can grow with t, so the displacement bound (3.14), and everything downstream ((3.15)-(3.16), (3.23)-(3.28), and the final O(ε) estimate (3.45)), lacks justification. Theorem 3.4's statement does not include (3.1), and the abstract advertises an unconditional width-dependent guarantee; the authors explicitly state after (3.1) that they cannot prove it. Thus the main theorem, as written, is not established. The result could be restored by a proof of an a priori trajectory bound for η=Θ(ε/m), T≤T0, or by explicitly restating the theorem as conditional on (3.1).","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies SGD training of two-layer physics-informed neural networks for Poisson's equation on the unit ball, with a cubic-ReLU activation and randomized initialization with scales m^{-α} and m^{-β}. It introduces a problem-dependent function class F and its finite-width discretization F_m, proves that a pseudo-network close to the target exists, bounds the distance between the true PINN and the pseudo-network along the SGD trajectory, and concludes in Theorem 3.4 that width m ≥ M(ε, f) yields average training loss O(ε) with η = Θ(ε/m), independent of the sample count N. A Rademacher-complexity generalization bound is derived in Theorem 4.2, and a d = 3 Poisson experiment is reported. The main advertised contribution is the removal of over-parameterization, i.e., a width requirement that does not grow with the number of training samples.","tokens_in":27281,"tokens_out":17515,"duration_ms":183395,"significance":"The target is important and the approach is creative: the pseudo-network is a linearization of the PINN, the approximation lemmas are explicit, and no constants are fitted to experiments. If the main theorem were fully proved, it would be a substantial step toward explaining narrow-PINN training. However, the proof of the central theorem depends on an unproved boundedness assumption and on an m-dependent function class whose norm is not controlled uniformly; both issues are structural for the claimed width-independent guarantee.","major_comments":[{"comment":"Assumption (3.1), namely ||w_i^(t)|| = O(1) and |ψ(x;W(t))| = O(1) for t = 1,...,T, is used at the very first step of the optimization proof: Eq. (3.9) bounds ∂ψ/∂w_ij by O(1)|a_i^(0)| \"using (2.7) and assumption (3.1)\", and this bound feeds into (3.13)-(3.14), then into (3.28), (3.41), and the final O(ε) estimate (3.45). The authors explicitly state after (3.1) that they cannot prove this condition. Theorem 3.4 does not include (3.1) as a hypothesis, and the abstract states the width guarantee unconditionally. Therefore the central convergence claim is not established as written. The result could be repaired by proving an a priori trajectory bound for η = Θ(ε/m) and T ≤ T0, or by restating the theorem and abstract conditionally on (3.1).","section":"§3, (3.1), (3.9), (3.14), Theorem 3.4"},{"comment":"The function class F and its norm depend on m through the initialization region Λ in (2.10) and the density p(θ). Consequently ||f||_F appearing in the width threshold M of Theorem 3.4 is an m-dependent quantity for a fixed target f. The claim in Remark 2.1 that all polynomials vanishing at zero belong to F is only demonstrated with representation coefficients that themselves depend on m; for a fixed nonzero f, a rescaling argument gives ||f||_F of order m^{α+2β} unless additional m-dependence is introduced. Since M contains powers of ||f||_F, the statement that the required width depends only on ε and the problem is not justified. The authors need either a uniform-in-m representation with bounded F-norm, or an explicit solution of the implicit inequality m ≥ M(ε, ||f||_F(m)).","section":"§2.4, (2.10)-(2.13), Remark 2.1; Theorem 3.4"},{"comment":"The generalization theorem inherits the unproved assumption (3.1) through Theorem 3.4, and its sample-size threshold N0 depends on the term m^{-α-2β}||f||_F + 1. If the F-norm growth described in the previous comment occurs, then N0 is not manifestly bounded independently of m. Thus the claim that N0 has an upper bound independent of m requires the same missing norm control.","section":"§4, Theorem 4.2"}],"minor_comments":[{"comment":"The functions γ1, γ2, γ3 are not globally 1-Lipschitz, although the proof invokes Lemma 2.5(c) with a global Lipschitz constant. On the bounded range of the affine arguments the Lipschitz constants are O(1), so the argument can likely be repaired by stating local Lipschitz bounds, but the current wording is inaccurate.","section":"§4, Theorem 4.1"},{"comment":"The summation in Eq. (3.46) is written with the index \"T = 1\" rather than t = 0 to T-1; this appears to be a typographical error.","section":"§3, Eq. (3.46) and surrounding text"},{"comment":"The numerical study reports single runs without error bars or repeated-initialization statistics, so it does not directly validate the probabilistic statements in Theorems 3.4 and 4.2.","section":"§5, Table 1 and Figure 1"}],"recommendation":"reject","confidential_remarks":"The main gap is explicitly acknowledged by the authors in the text, so the issue is not a hidden technicality. The function-class dependence on m also affects the novelty claim of a sample-independent width threshold. I would be willing to reconsider a substantially revised version that proves the trajectory bound or states the theorem conditionally with a correspondingly weakened abstract."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a real new result, but the main theorem as stated is not established. The proof chain starts with Assumption (3.1), a bounded-trajectory condition. The authors explicitly say they cannot prove it, yet Theorem 3.4's statement and abstract present the convergence guarantee as unconditional. That is the load-bearing gap, and the stress-test note is right about where it lands.\n\nWhat is genuinely new: for two-layer cubic-ReLU PINNs on Poisson's equation, the required width is independent of the number of training samples, provided the target lies in their class F. This is not present in the cited PINN literature, which needs width polynomial in N. The technique follows known FNN lazy-training and Barron-space analyses, but adapting it to the PINN Laplacian structure is a real contribution. The approximation part is clean: standard existence estimates parameterized by epsilon, no fitted constants, no back-fitting. The pseudo-network is a linearization, not a smuggled conclusion. The Rademacher generalization section is standard but coherent, and the single numerical experiment is modest yet consistent with the claimed N-independence.\n\nWhere it goes soft: Assumption (3.1) is used in Theorem 3.3(a) to bound derivatives, and everything downstream depends on those bounds. If trajectories can grow, the gradient bounds fail and the convergence proof collapses. The authors admit this directly in the text. The theorem statement should either include the assumption or be labeled conditional. The result could be repaired by proving an a priori trajectory bound for eta = Theta(epsilon/m) and T up to T0, or by restating the theorem as conditional. There is also a subtler issue worth raising: the function class F is defined using Lambda, which depends on m through m^{-alpha} and m^{-beta}. Since the width threshold M depends on the problem and F changes with m, the claim that the width threshold depends only on epsilon and the problem needs reconciliation. That is not fatal, but it should be clarified.\n\nThe citation pattern looks fine. Prior PINN work is cited for the polynomial-in-N width requirement, and the FNN analogy is cited. There is no circularity burden.\n\nThis paper is for researchers in ML theory for PINNs and possibly numerical analysis. It is a serious paper, not a desk reject, but it needs a referee to push on the missing trajectory bound and on the F-Lambda dependence. I would send it to peer review with those instructions.","headline":"Worth a serious referee, but the central optimization guarantee is conditional on a non-explosion assumption the authors admit they cannot prove, so the paper currently overclaims.","tokens_in":27756,"tokens_out":1510,"would_cite":false,"duration_ms":19124,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68Q32","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"For two-layer PINNs, SGD reaches O(epsilon) average loss with width that need not grow with the sample count.","keywords":["physics-informed neural networks","stochastic gradient descent","two-layer networks","optimization","generalization","Rademacher complexity","over-parameterization","Poisson equation"],"falsifier":"Run SGD with the widths, learning rate, and iteration counts specified in Theorem 3.4 on the paper's own test problem ($d=3$, $f=x_1^2+x_2^2+x_3^2$) and record $\\sup_t \\max_i \\|w_i^{(t)}\\|_2$, $\\sup_t |\\psi(x;W^{(t)})|$, and the average training loss. A run in which the average training loss stays above $C\\epsilon$ after $T$ iterations, or in which either supremum exceeds its $O(1)$ bound, would show the stated theorem's assumptions or conclusion fail for that configuration.","tokens_in":26842,"feed_emoji":"📉","tokens_out":12901,"duration_ms":133544,"temperature":0.7,"pith_summary":"Physics-informed neural networks solve Poisson's equation by minimizing the squared residual of the PDE, and this paper proves that plain SGD can drive that residual small without the enormous widths required by over-parameterization theory. The main theorem (Theorem 3.4) states that if the target right-hand side $f$ belongs to a structured function class $F$ (which includes every polynomial vanishing at the origin), then a width $m$ larger than a threshold $M$ depending only on $\\epsilon$ and the problem ensures that after $T$ SGD iterations the average squared training loss is at most $O(\\epsilon)$. Theorem 4.2 extends the same $O(\\epsilon)$ bound to the expected loss once the number of training samples is sufficiently large, using Rademacher complexity. On this view, the bottleneck for training PINNs is the regularity of the target function, not the size of the dataset.","feed_headline":"SGD trains two-layer PINNs to small error with width fixed by the PDE","feed_subtitle":"New proof says the number of training samples does not set the required network size.","key_machinery":"The load-bearing object is a pseudo network $g(x;W)$, defined as the linearization of the PINN $\\psi$ in the hidden-layer weights: the random output coefficients $a_i^{(0)}$, biases $b_i^{(0)}$, and ReLU activation patterns are frozen, and only the first-order dependence of $\\psi$ on $w_i$ is kept. Because $g$ is affine in $W$, its squared loss is convex, so the SGD analysis can use convex-optimization descent on this surrogate and then bound the discrepancy $\\psi-g$ along the trajectory. The function class $F$ and its finite-width discretization $F_m$, built from the random basis $\\zeta(x;\\theta)$ that comes from applying $\\Delta$ to the network ansatz, provide the approximation step: a small-norm weight matrix $W^*$ near initialization approximates $f$ to accuracy $\\epsilon$. Theorem 3.3 bounds $|\\psi-g|$ and the difference of their gradients, Theorem 4.1 bounds the Rademacher complexity of the network class, and the proof chain assumes the SGD trajectory stays bounded (Assumption (3.1)), which the authors state they cannot prove.","core_discovery":"The paper's central claim is that the network width needed for SGD convergence is independent of the sample count. For the two-layer ReLU PINN $\\psi = \\Delta\\varphi$ with $\\varphi = (\\|x\\|_2^2-1)\\tilde\\varphi$ on the unit ball, if the right-hand side $f$ lies in $F$ and Assumption (3.1) holds, then for every $\\epsilon>0$ there is a width threshold $M$ such that for $m\\ge M$, with learning rate $\\eta=\\Theta(\\epsilon/m)$ and iteration count $T$ in a specified interval, the average training loss satisfies $\\mathbb{E}_X\\mathbb{E}_{\\mathrm{sgd}}\\bigl[\\frac{1}{T}\\sum_{t=0}^{T-1} L(\\psi(\\cdot;W^{(t)}))\\bigr] \\le O(\\epsilon)$ with probability at least $1-\\delta$. Theorem 4.2 adds that if $N\\ge N_0$, the same bound holds for the expected loss. The width threshold $M$ and the sample threshold $N_0$ depend on the problem and the accuracy, not on one another in the over-parameterized way.","pith_inferences":["Extension a reader could test: record $\\sup_t \\max_i \\|w_i^{(t)}\\|_2$ and $\\sup_t |\\psi(x;W^{(t)})|$ during the experiments; if these stay $O(1)$, Assumption (3.1) is empirically satisfied and the remaining gap is a formal proof of that bound.","Implicit in the construction of $F$ through the random basis $\\zeta$ is that the relevant quantity is a Barron-type norm of the transformed right-hand side under the PDE operator; analogous sample-independent width bounds for other second-order PDEs would require a similar operator-commutator argument, which the paper does not supply.","A practical consequence the authors do not draw is that architecture cost and data cost separate cleanly: fix the width from the target's structure, then add samples to shrink the generalization gap.","Reading caveat: the abstract's unconditional phrasing is stronger than the proved statement, since the proof relies on trajectory boundedness that the authors say they cannot yet establish."],"forward_implications":["Training two-layer PINNs by SGD no longer requires the width to grow polynomially with the number of training samples; the earlier $\\Omega(N^2)$-neuron regime is replaced by a threshold $m = \\Omega(c_f/\\epsilon^p)$.","Both the average training loss and the average expected loss can be pushed below $O(\\epsilon)$, with explicit choices of learning rate, iteration count, and minimum sample size.","The standard initialization scale $(\\alpha,\\beta)=(0,1/2)$ falls inside the condition $\\alpha+3\\beta>1$, so the theoretical guarantees cover the parameter setting used in the paper's numerical experiments.","The authors state the analysis is carried out for Poisson's equation on the unit ball but can be adapted to other types of PDEs and domains where a boundary-annihilating ansatz is available."],"supporting_citations":[{"why":"Over-parameterized two-layer PINN result requiring $\\Omega(N^2)$ neurons; the baseline that this paper's width threshold is designed to beat.","marker":"Gao et al., 2023"},{"why":"Earlier optimization and generalization theory for two-layer PINNs under over-parameterization; the setting this paper revisits without width growth.","marker":"Luo and Yang, 2020"},{"why":"Numerical evidence that narrow two-layer PINNs can reach low loss, the gap that motivates the sample-independent theory.","marker":"Grossmann et al., 2024"},{"why":"Source of the special-function learning approach and of the Rademacher contraction corollary used in the generalization theorem.","marker":"Allen-Zhu et al., 2019a"},{"why":"Prototype result that learning polynomials by two-layer ReLU networks needs width $\\Omega(1/\\epsilon^3)$, independent of sample size.","marker":"Andoni et al., 2014"},{"why":"Standard Rademacher complexity bounds and generalization inequalities on which Theorem 4.2 relies.","marker":"Shalev-Shwartz and Ben-David, 2014"}],"fun_headline_variants":["Two-layer PINN width need not scale with training data","SGD trains PINNs without over-parametrization, width fixed by problem","PINN convergence: width depends on problem, not sample size","No over-parametrization needed for SGD-trained two-layer PINNs","Width threshold for PINN SGD set by accuracy, not data size"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof assumes the SGD trajectory stays bounded: at every step the hidden-layer weights and the network output remain within an absolute constant, and the authors say they cannot prove that this holds.","fun_headline_variants_meta":{"raw":{"variants":["Two-layer PINN width need not scale with training data","SGD trains PINNs without over-parametrization, width fixed by problem","PINN convergence: width depends on problem, not sample size","No over-parametrization needed for SGD-trained two-layer PINNs","Width threshold for PINN SGD set by accuracy, not data size"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000823,"raw_usage":{"total_tokens":3592,"prompt_tokens":927,"completion_tokens":2665,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":2581}},"tokens_in":543,"tokens_out":2665,"duration_ms":18377,"temperature":1.0,"reasoning_tokens":2581,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:11:43.427752+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SGD with the widths, learning rate, and iteration counts specified in Theorem 3.4 on the paper's own test problem ($d=3$, $f=x_1^2+x_2^2+x_3^2$) and record $\\sup_t \\max_i \\|w_i^{(t)}\\|_2$, $\\sup_t |\\psi(x;W^{(t)})|$, and the average training loss. A run in which the average training loss stays above $C\\epsilon$ after $T$ iterations, or in which either supremum exceeds its $O(1)$ bound, would show the stated theorem's assumptions or conclusion fail for that configuration.","supporting_citations":[],"review_version":1}