{"id":"c198c7f1-2a33-4d6a-b9fd-6428387e1ba0","arxiv_id":"2502.03670","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A physics-informed neural network trained on Monte Carlo-averaged residuals approximates the expected value of linear SPDE solutions, tested on the stochastic heat equation in 144 settings.","lead":"Researchers train a neural network on random noise samples to approximate the average solution of linear stochastic heat equations without deriving the mean equation by hand. The method, called LEC, gives low errors in low dimensions, degrades as dimension grows, and comes with public code.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim depends on an unstated loss-decomposition identity for MC-averaged residuals; the paper neither proves it nor benchmarks against a mean-PDE baseline, leaving the convergence to E[u] empirically under-supported.","rationale":"The reader's weakest_assumption identifies the same load-bearing premise: the minimizer of the MC-averaged squared residual equals the mean-PDE solution only through the decoupling identity for linear additive noise. I agree with that diagnosis. The math itself is sound for the setting considered: because the network does not take W as input, E_W[(Au - f - gW)^2] = (Au - f - gE[W])^2 + g^2 Var(W), and the variance term does not affect the minimizer. But the paper never states this identity, and Section 4.5 acknowledges the absence of a formal convergence proof. This is a genuine support gap, not a fatal flaw. The empirical errors against analytical benchmarks are encouraging, but the lack of repeated-seed statistics and the absence of a baseline PINN on the mean PDE mean the evidence is weaker than the abstract implies. The 'simulator-free' claim is also an overclaim since training requires sampling noise realizations, but that is secondary to the central convergence claim. The proposed test--comparing LEC to a mean-PDE baseline--would directly settle whether stochastic training actually learns E[u] and would quantify any degradation due to finite-m averaging. Until that comparison and the derivation are provided, CONDITIONAL acceptance is the appropriate verdict, matching the reader's assessment.","tokens_in":11988,"tokens_out":9307,"duration_ms":87880,"concrete_test":"Train the same architecture and hyperparameters on the deterministic mean PDE dv/dt = nu*Laplacian(v) + theta*G(x)*E[W(xi(t))] with the same initial and boundary conditions, using the same sampling distributions for (t,x,nu,params). Compare relative L2 errors against the LEC results in Tables 1-4 for all 144 configurations, using at least 5 seeds per configuration. If the mean-PDE baseline matches LEC within, say, 10% relative error, the central claim is confirmed; if LEC is systematically worse, the finite-m averaging introduces bias or variance. Additionally, add a one-line derivation of the loss decomposition to the paper to close the theoretical gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that minimizing the Monte Carlo-averaged residual (Eq. 3.2) makes a single network converge to E[u]. This holds only if the squared residual decouples as E_W[(A u_theta - f - gW)^2] = (A u_theta - f - gE[W])^2 + g^2 Var(W), which requires u_theta to be independent of the noise realization W. The architecture in Section 3.2 satisfies this (inputs are (t,x,nu,params), not W), and for the linear additive-noise SPDEs considered the identity is elementary. However, the paper never states or proves this identity; Section 4.5 explicitly concedes 'lacks a formal convergence proof.' Without it, the claim is only a hypothesis supported by selected aggregate errors. The reported relative L2 errors (e.g., 0.0318 at d=2) are averaged over 36 configurations per dimension, not repeated seeds, so the variance across training runs is unknown. Moreover, no baseline PINN trained directly on the mean PDE (source theta*G*E[W]) is reported, so it is untested whether the stochastic training actually converges to E[u] rather than to some other regression target. The load-bearing gap is therefore the missing derivation and the missing baseline, not an observed failure.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Learned Expectation Collapser (LEC), a physics-informed neural network trained on a Monte Carlo-averaged residual loss for linear SPDEs with additive noise. The central claim, stated in Sections 1 and 5, is that a single feedforward network trained by minimizing the average squared residual over stochastic realizations converges to the expected solution E[u(t,x)] of the SPDE. The method is tested on the stochastic heat equation with Dirichlet boundary conditions across spatial dimensions d = 2, 4, 6, 8, three noise models (Brownian motion with drift, Ornstein-Uhlenbeck, compound Poisson), three forcing maps, and four Monte Carlo sample sizes, totaling 144 configurations. Relative L2 errors are reported in Tables 1-4, showing errors that grow with dimension and with noise/forcing complexity. The paper includes a public code repository and derives analytical benchmark expectations in Appendix A.","tokens_in":12171,"tokens_out":4914,"duration_ms":50250,"significance":"If the central claim is correct, the observation is useful and somewhat surprising: a standard PINN can estimate the mean solution of a linear SPDE without deriving E[W] or solving pathwise, using only samples of the noise process. The paper's strengths are its reproducible public code, the breadth of the 144-configuration benchmark, and the fact that the underlying loss-decomposition identity for linear additive noise is correct. However, the manuscript is currently an empirical demonstration rather than a supported claim: the key identity that makes the stochastic residual minimizer equal to the mean-PDE minimizer is never stated, and no baseline PINN trained directly on the mean PDE is reported. These omissions leave the central claim load-bearing but unproven and empirically under-supported relative to the strength of the conclusion.","major_comments":[{"comment":"The central claim that minimizing the Monte Carlo-averaged residual yields E[u] rests on the pointwise identity E_W[(A u_theta - f - gW)^2] = (A u_theta - f - g E[W])^2 + g^2 Var(W), which holds when u_theta is independent of the noise realization W. This identity is never stated or proved in the paper. Please add an explicit derivation showing that the variance term is independent of u_theta, so the minimizers of the stochastic residual coincide with the minimizers of the deterministic mean-PDE residual, and state the regularity assumptions (finite second moments, integrability) needed for the argument. Without this derivation, the statement in Section 5 that vanilla PINNs 'converge to the expected solution' is a hypothesis, especially since Section 4.5 concedes the method 'lacks a formal convergence proof.'","section":"Section 3.2, Eq. (3.2)"},{"comment":"No baseline PINN trained directly on the mean PDE with source term vartheta G(x) E[W(xi(t))] is reported. Such a baseline is the natural control: if the stochastic residual training truly recovers E[u], it should match the accuracy of a PINN trained on the deterministic mean equation with the same architecture, collocation points, and training budget. Without this comparison, the reader cannot distinguish between the proposed stochastic-averaging mechanism and the trivial possibility that either training objective yields a similar approximate solution to the mean PDE. Please add this baseline and report the comparison, at least for a representative subset of configurations.","section":"Section 4, Tables 1-4"},{"comment":"The reported error statistics are averages over 36 configurations per dimension, not over independent training runs with different random seeds. The standard deviations in Tables 1-4 therefore reflect variation across noise models, forcing functions, and Monte Carlo settings, not the stochasticity of neural-network optimization. To support the claim that the model 'consistently' learns the expected solution, please either report the number of seeds per configuration and give mean +/- std over seeds for at least a representative subset, or explicitly state that each configuration was trained once and present the variability across seeds as a separate quantity.","section":"Section 4.1, Tables 1-2"}],"minor_comments":[{"comment":"The equation after taking expectations appears typographically corrupted: 'partial_nu E[u(t,x)] / partial t' should read 'partial E[u(t,x)] / partial t = nu Delta E[u] + vartheta G(x) E[W(xi(t))]'. Please correct the notation.","section":"Section 3.1"},{"comment":"The loss-weight assignment in the algorithm is inconsistent with Section 3.2: line 25 uses lambda_B for the initial-condition loss and lambda_I for the boundary loss, while the text defines lambda_I for initial and lambda_B for boundary. Please correct the algorithm and confirm that the code matches the intended weighting.","section":"Algorithm 1, line 25"},{"comment":"The sentence 'Jin et al. [33] present fractional PINNs...' cites reference [33], which is a paper on masked autoencoders by Kim, Choi, Widemann, and Zohdi, not a fractional-PINN paper by Jin et al. Please check the citation and either correct the reference or the attribution.","section":"Section 2.1, reference [33]"},{"comment":"The phrase 'abalition studies' appears to be a typo for 'ablation studies'. Please proofread the manuscript for similar typographical errors.","section":"Section 3.2"},{"comment":"The text says the exponential-linear forcing 'slightly outperforms linear in most cases,' but in d=2 the linear forcing has lower mean error (0.0294 vs 0.0308). Consider qualifying this statement to avoid over-generalization.","section":"Section 4.3, Table 3"}],"recommendation":"major_revision","confidential_remarks":"This is a borderline case: the core observation is simple and likely correct for the linear additive-noise setting, but the manuscript's framing as a new framework and its strong convergence claim currently outrun the evidence. The missing identity and missing mean-PDE baseline are fixable within the scope of the paper, so I recommend major revision rather than rejection. I would also encourage the editor to ensure the final version addresses the seed-variability question, as the current aggregate statistics over configurations do not establish reproducibility across training runs."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is an honest, workmanlike empirical paper whose core trick is a one-line identity the authors never write down. For linear additive-noise SPDEs, training a single PINN on the average squared residual over noise draws lands on E[u] because E_W[(A u_theta - f - gW)^2] = (A u_theta - f - gE[W])^2 + g^2 Var(W), and the variance term is constant in u_theta. That is the whole mechanism. It is correct for the equations they consider, but the paper neither states the identity nor benchmarks against a PINN trained directly on the mean PDE.\n\nWhat is new: applying vanilla PINNs with MC-averaged residuals directly to estimate the expectation is something I have not seen in the cited literature, and the paper tests it over 144 configurations (d = 2,4,6,8; three noise models; three forcing maps). The analytical benchmarks in Appendix A check out, the code is public, and the limitations section explicitly concedes the lack of a convergence proof. That honesty counts.\n\nWhere it is soft. First, the missing derivation is load-bearing; without it the central claim is a hypothesis, not a result. Second, no baseline against a deterministic mean-PDE PINN, so the empirical support is indirect - we only know the network gets low L2 error to analytical E[u], not that the stochastic version beats or matches the obvious alternative. Third, the reported standard deviations are over configurations, not training seeds, so run-to-run variance is unknown. Fourth, 'simulator-free' overstates it: they sample noise realizations during training, they just avoid PDE discretization.\n\nThe paper is clearly written, the scope is honestly limited to linear additive noise, and the authors do not oversell the theory. The conclusion's 'converge' is stronger than what finite-capacity training for 10k epochs demonstrates, but the aggregate errors (3% in 2D, 47% in 8D) are plausibly explained by the identity above plus capacity limits and the shrinking signal.\n\nVerdict: send it to review, but a serious referee should require the one-line derivation, a mean-PDE baseline, and seed-wise error bars before the claim is taken as established. I would not cite it for the method; I might cite it as empirical evidence that MC-averaged residual PINNs implicitly solve the mean equation.","headline":"An honest, workmanlike empirical paper whose core mechanism is a one-line identity the authors never write down; the claim is plausible and the gaps are fixable.","tokens_in":12798,"tokens_out":3200,"would_cite":false,"duration_ms":30201,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H15","65M75","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"One feedforward network trained on averaged noise residuals can approximate the expected solution of a linear stochastic PDE.","keywords":["Physics-Informed Neural Networks","Stochastic Partial Differential Equations","Expected value estimation","Monte Carlo sampling","Stochastic heat equation","Mean-field approximation","High-dimensional PDEs","Additive noise"],"falsifier":"Train LEC on a zero-mean noise process with large variance, such as $\\xi(t)=\\sigma W_t$ with no drift, so that $\\mathbb{E}[W(\\xi(t))]=0$; the averaged residual loss then reduces to the deterministic heat residual plus a constant, so the network should reproduce the pure initial-boundary-value solution. If the learned function deviates from that solution, the minimizer is not the expectation.","tokens_in":11690,"feed_emoji":"🎲","tokens_out":7617,"duration_ms":69021,"temperature":0.7,"pith_summary":"This paper claims that an ordinary physics-informed neural network, trained by averaging its PDE residual over many random noise samples, learns the expected value of the solution of a linear stochastic partial differential equation. The authors introduce the Learned Expectation Collapser (LEC), which samples space-time points and noise parameters together during training and minimizes a Monte Carlo-averaged residual loss. They test this on the stochastic heat equation with additive noise across 144 configurations, varying spatial dimension, noise process, forcing function, and number of samples. The empirical pattern is that the network tracks the analytical expectation closely in low dimensions, with relative L2 error roughly doubling every two dimensions, while larger Monte Carlo batches mainly stabilize training. The paper is explicit that this is an empirically supported hypothesis rather than a proven theorem.","feed_headline":"Noise-sampled training makes neural nets learn the mean SPDE solution","feed_subtitle":"Averaging the PDE residual over random noise samples recovers the mean solution without discretizing the domain.","key_machinery":"The load-bearing object is the Monte Carlo-averaged residual loss with the network input restricted to distribution parameters rather than noise realizations. At each training step the same network output $u_\\theta(t,x)$ is scored against $m$ independent samples $W_i$, so the squared residual expands into the residual at the mean forcing plus a noise-variance term that does not depend on $u_\\theta$. That decoupling is what lets the minimizer align with the solution of the mean PDE. The network architecture is a fully connected tanh network whose width and depth grow slowly with dimension, and the input includes the noise model's parameters ($\\mu,\\sigma$ etc.) sampled from fixed ranges, which lets one network represent a family of expectation solutions.","core_discovery":"On the paper's own terms, the discovery is that a single feedforward network trained to minimize $$L = \\frac{1}{m}\\sum_{i=1}^m \\|\\partial_t u_\\$\\theta$ - \\nu\\$\\Delta$ u_\\$\\theta$ - \\vartheta G(x) W_i(\\xi(t))\\|^2 + \\lambda_I \\|u_\\$\\theta$(0,x)-\\vartheta G(x)\\|^2 + \\lambda_B \\|u_\\$\\theta$\\|^2_{\\partial\\$\\Omega$}$$ converges to $\\mathbb{E}[u(t,x)]$, the expectation of the SPDE solution, even though the network never sees the noise path $W_i$ and the loss contains no explicit expectation. The paper verifies this for the stochastic heat equation with additive noise, using forcing maps $W(\\xi)=\\xi$, $W(\\xi)=e^{-t}\\xi$, and $W(\\xi)=\\xi^2$, and three noise processes whose analytical means are known. Across all 144 runs the network reduces loss and error, with best accuracy at dimension 2 (mean relative L2 error 0.032) and growing error at higher dimensions (0.47 at d=8). The authors state the result as a hypothesis confirmed empirically, not as a convergence theorem.","pith_inferences":["The decoupling argument suggests the method should transfer to any linear SPDE with additive noise whose deterministic part is well posed, e.g. advection-diffusion or wave equations; the noise model would enter only through its mean and variance.","Because the loss minimizer depends on the noise only through its mean plus a constant variance, differences in accuracy between smooth and jump noise probably reflect optimization difficulty, not a change in the target expectation; testing with identical variance across noise models would separate the two.","A natural testable extension is to compare LEC's expectation against a Monte Carlo average of pathwise PINN solutions; if the two disagree, the single-network minimizer is capturing a different regression quantity than the true expectation.","The observed dimension scaling suggests that variance-reduced sampling of collocation points, or a better parameterization of the high-dimensional input, might push the method beyond d=8; the paper leaves this open."],"forward_implications":["If the claim holds, the expected solution of a linear SPDE with additive noise can be approximated by sampling the noise during training, without deriving $\\mathbb{E}[W]$ or solving pathwise.","The same trained network can supply expectations across a range of noise-model parameters, because those parameters are inputs rather than fixed experiment constants.","Accuracy degrades predictably with spatial dimension, with relative L2 error roughly doubling every two dimensions, so the method's practical range is low-to-moderate dimension unless sampling or architecture changes are made.","Increasing the Monte Carlo batch size does not dramatically change final accuracy but visibly stabilizes the loss and error trajectories during training."],"supporting_citations":[{"why":"Supplies the residual-based pointwise sampling loss that LEC's Monte Carlo-averaged objective extends.","marker":"[23]"},{"why":"Establishes the physics-informed neural network approach of adding PDE residual, initial, and boundary penalties to the loss.","marker":"[22]"},{"why":"Provides the analytical solution and Feynman-Kac representation of the stochastic heat equation used as ground truth expectation.","marker":"[17]"},{"why":"Represents the per-noise-realization neural SPDE solver that LEC contrasts with by using a single network.","marker":"[27]"},{"why":"Represents the modal-space PINN approach for SPDEs that LEC avoids by not requiring basis or covariance decompositions.","marker":"[28]"}],"fun_headline_variants":["Noise-sampled training teaches neural nets SPDE means","Neural nets learn mean SPDE solutions from noise samples","Random noise sampling turns PDE residuals into expectations","Physics-informed net estimates SPDE means by noise averaging","Single net, many noise samples: SPDE mean emerges"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the minimizer of the averaged squared residual equals the solution of the mean PDE; this requires the network output to be independent of the noise realization, so that the noise variance adds only a constant to the loss.","fun_headline_variants_meta":{"raw":{"variants":["Noise-sampled training teaches neural nets SPDE means","Neural nets learn mean SPDE solutions from noise samples","Random noise sampling turns PDE residuals into expectations","Physics-informed net estimates SPDE means by noise averaging","Single net, many noise samples: SPDE mean emerges"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000186,"raw_usage":{"total_tokens":1361,"prompt_tokens":1018,"completion_tokens":343,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":634,"completion_tokens_details":{"reasoning_tokens":265}},"tokens_in":634,"tokens_out":343,"duration_ms":3796,"temperature":1.0,"reasoning_tokens":265,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T04:08:47.394254+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train LEC on a zero-mean noise process with large variance, such as $\\xi(t)=\\sigma W_t$ with no drift, so that $\\mathbb{E}[W(\\xi(t))]=0$; the averaged residual loss then reduces to the deterministic heat residual plus a constant, so the network should reproduce the pure initial-boundary-value solution. If the learned function deviates from that solution, the minimizer is not the expectation.","supporting_citations":[{"cited_title":"Bertini, N","cited_arxiv_id":null,"evidence_quote":"Provides the analytical solution and Feynman-Kac representation of the stochastic heat equation used as ground truth expectation."}],"review_version":1}