{"id":"6105683c-ea93-477a-bb2a-1abef81581a2","arxiv_id":"2506.15481","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A hybrid deep learning method solves high-dimensional parabolic PDEs by combining coarse stochastic trajectory sampling with local antithetic residual estimation, demonstrated up to dimension 10,000.","lead":"This paper introduces a 'shotgun method' for solving high-dimensional parabolic partial differential equations, which estimates the PDE residual at points sampled from coarse stochastic trajectories using a random antithetic finite-difference scheme. It reports accurate results on problems up to 10,000 dimensions and long time horizons, with less computational cost than full-trajectory deep FBSDE methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Appendix A's proof of the key estimator error bound (20) is dimensionally inconsistent: it computes the variance of Δt·ε rather than ε, and §5.2's convergence argument is explicitly heuristic and assumes noise independent of θ; the mechanism enabling coarse-N shotgun training is therefore unproven.","rationale":"The reader's weakest assumption already identified the residual estimator's noise behavior and the Appendix A contradiction as the critical point. My stress-test confirms and sharpens this: the appendix's scaling error is not merely a typo, because the method's reliance on large M (and hence on O(1) variance) is central to the design. The paper itself says the convergence explanation is only heuristic, and the heuristic's assumptions (residuals linear in θ, perturbation independent of θ) do not hold for a neural network whose second derivatives enter the residual noise. This is a genuine proof gap, not an outside-consensus disagreement. However, the numerical evidence in Section 4, especially the systematic M-dependence in Table 1 and Figure 5, is consistent with the intended O(1) variance behavior, and the experiments suggest the method may work even without a rigorous proof. I therefore do not recommend changing the conditional verdict: the paper should either fix the Appendix A derivation or clearly mark eq. (20) as conjectural, and it should provide the empirical variance check or a more realistic noise model for SGD. I agree with the reader that the issues are addressable and do not warrant rejection.","tokens_in":15749,"tokens_out":9940,"duration_ms":97990,"concrete_test":"Freeze a trained uθ from the BSB d=100, T=1 experiment, sample 10^4 points (t,x) from the coarse trajectory distribution of Algorithm 2, and estimate Var(r) of the Algorithm 1 residual (18) for Δt = 4^{-3}, 4^{-4}, 4^{-5}, both with M=1 and with M=16,64. If Var(r) scales as O(Δt^2) rather than O(1), or if M-averaging does not reduce the effective SGD noise variance by ≈M^{-1} at the converged parameters, then eqs. (20)-(22) and the §5.2 explanation fail. The corresponding analytical check is to re-derive Appendix A while keeping the explicit 1/(2Δt) prefactor of eq. (18).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The shotgun method's central mechanism is eqs. (18)-(22): the single-step residual estimator has O(Δt) bias and O(1) variance, so averaging M local samples reduces the variance to O(M^{-1}), and this noise is argued to be harmless during SGD. The only proof offered, Appendix A, is inconsistent with eq. (20). In eq. (60), the leading stochastic term is e2 = (1/2) η^T H η − (1/2) Tr[H] Δt with η ~ N(0, Δt I). With δ = Δt^{-1/2}η this equals (Δt/2)(δ^T H δ − Tr[H]) and has variance O(Δt^2), as the appendix itself concludes at eq. (68). But the residual in eq. (18) contains (1/(2Δt))η^T H η − (1/2)Tr[H] = (1/2)(δ^T H δ − Tr[H]), which has O(1) variance. The appendix therefore proves the variance of Δt·ε, not of ε; claim (20) is left unproved and in fact the appendix's own bound contradicts (20). The only remaining support, §5.2, is explicitly 'non-rigorous' and additionally assumes residuals are linear in θ and that the perturbation ε_k^{(n)} is independent of θ, while Algorithm 1's ε depends on ∇²u_θ. Thus the theoretical basis for tolerating M≪d and coarse N is missing. The numerical results may still stand, but the central claim is currently supported only by heuristic and experiments, not by the presented analysis.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a 'shotgun method' for solving high-dimensional parabolic PDEs of the form (1), combining ideas from deep FBSDE formulations and PINNs. Instead of simulating full trajectories at a fine time step, the method draws collocation points from coarsely discretized SDE trajectories and estimates the PDE residual at each point with a local antithetic single-step estimator, Algorithm 1. The residual estimator is averaged over M local samples with local step Δt, while the trajectory step ℓ = T/N can be much larger than Δt. The DNN u_θ(t,x) is trained by minimizing the squared residuals plus a terminal loss, as described in Algorithm 2. Numerical experiments include a 1D Laplace problem, a 100D HJB equation with T = 1 and T = 10, a 100D Allen–Cahn equation, and the Black–Scholes–Barenblatt equation in dimensions up to 10000 and with T = 10. The results are compared with the SDE matching method of [20] and with reference solutions, and the authors claim competitive accuracy at lower backward-propagation cost. The theoretical discussion in Appendix A aims to justify the bias and variance of the residual estimator, and Section 5.2 gives a heuristic explanation for why large residual variance is tolerable during training.","tokens_in":16096,"tokens_out":15811,"duration_ms":150499,"significance":"If the numerical results are taken at face value, the method is a useful practical contribution: it decouples the trajectory discretization from the residual accuracy, which could reduce the cost of long-time high-dimensional problems, and it is tested against external references (closed-form BSB solution, Monte Carlo HJB reference, known Allen–Cahn value). The paper reports reproducible experimental setups and demonstrations up to d = 10000 and T = 10, which is beyond most existing deep FBSDE experiments. The main weakness is theoretical: the proof of the key estimate (20) is stated incorrectly in Eq. (60), its validity is restricted to Δt ≪ 1/d while the experiments use dΔt ≈ 10, and the convergence explanation in Section 5.2 is explicitly heuristic with assumptions not satisfied by the algorithm. These issues do not invalidate the numerical evidence, but they limit the strength of the central claim that the method's coarse-N behavior is understood.","major_comments":[{"comment":"The displayed expansion is for Δt·ε rather than for ε. With η = √Δt δ, the second term in Eq. (60) is (Δt/2)(δ^T H δ − Tr[H]), whose variance is O(Δt^2); this is Var(Δt·ε), not Var(ε). The corresponding term for ε in Eq. (18) would be (1/(2Δt))η^T H η − (1/2)Tr[H], which has O(1) variance. Thus as written Eq. (60) is inconsistent with Eq. (20), but the inconsistency disappears if the left-hand side is corrected to Δt·ε (or if the right-hand side is divided by Δt). Since Eqs. (67)--(68) then imply Eq. (20), this is a notation error rather than a mathematical contradiction; please fix Eq. (60) and define 'εΔt' explicitly.","section":"Appendix A, Eq. (60)"},{"comment":"The proof of (20) assumes ∥η∥ ≪ 1, which for σ of order I requires Δt ≪ 1/d. In the high-dimensional BSB tests (§4.4.1) the paper uses d = 10000 and Δt = 4^{-5}, so dΔt ≈ 9.8 and E∥η∥^4 = d(d+2)Δt^2 is not small; the tail term bound in Eq. (65) is therefore not O(Δt^2) in this regime. Consequently the claims E[ε] = O(Δt) and Var(ε) = O(1) are not established for the parameter values used in the experiments. The authors should either restrict the theoretical statement to Δt ≪ 1/d, replace the Taylor argument with a dimension-aware bound, or provide a numerical measurement of E[ε] and Var(ε) as functions of d at fixed Δt. Without one of these, Eq. (20) does not justify taking M ≪ d in the tested regime.","section":"Appendix A (Eqs. (60)-(68)) and §4.4"},{"comment":"The explanation of why large-variance residuals are tolerable relies on assumptions that are not satisfied by Algorithm 2: the residuals f_k(θ) are taken to be linear, and the perturbations ε_k^{(n)} are assumed independent of θ. In the shotgun method, the perturbation in Eq. (18) depends on ∇²u_θ through the local samples η, so it is parameter-dependent. The section is explicitly non-rigorous, but the paper should state more clearly that no convergence guarantee is proved for the actual algorithm and that Eq. (58) is only an idealized model. This distinction matters because the abstract and Section 5.1 present the tolerance of coarse N as a feature of the method.","section":"§5.2"}],"minor_comments":[{"comment":"The line 'Sample η ∼ N(0, √Δt I_d)' is not a valid variance specification; it should read η ∼ N(0, Δt I_d) (or state that √Δt is the standard deviation).","section":"Algorithm 1, line 5"},{"comment":"The sentence 'the total number of collocation points is (N + 2)M' omits the number of trajectories M1 and the fact that each residual evaluation uses two points x±; please clarify whether this count is per trajectory and how the terminal condition is counted.","section":"§3.2"},{"comment":"The captions for panels (b) and (c) are swapped relative to the definitions f_2 = sin and f_3 = e^{0.5x} in Section 4.1.","section":"Figure 1"},{"comment":"The citation [4] for the randomized Hessian approximation points to the 1000 Genomes consortium paper; this appears to be an incorrect reference.","section":"Remark 1"},{"comment":"The reference solution is computed with 10^5 Monte Carlo samples per (t,x), but no standard deviation of this reference is reported; please include a maximum standard error so the reader can gauge the accuracy of the reference.","section":"§4.2"},{"comment":"The Allen–Cahn test reports only the scalar u_θ(0,x0) and no error curve; please add trajectory errors or a comparison with a reference to make the result quantitative.","section":"§4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for the Journal of Scientific Computing, and the numerical results are interesting. The main issue is that the theoretical section needs careful revision: the missing Δt in Eq. (60) is easy to fix, but the dimension restriction Δt ≪ 1/d is a substantive gap relative to the experiments. The benchmarks appear external and no circularity is apparent."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the shotgun method is a legitimate incremental contribution. It replaces full-trajectory training in deep FBSDE solvers with coarse trajectory sampling plus a local antithetic residual estimator, and the numerics show real gains: on a 100-d BSB equation, N=10 with M=16 matches SDE matching at N=192, with lower backprop cost; they push to d=10000, T=10. That is worth a referee's time.\n\nWhat is genuinely new is eqs. (18)-(22): the antithetic average of two one-sided Euler steps gives a single-step residual estimate whose bias is O(Δt) and whose variance is O(1), with the variance reducible by averaging M local samples. That is a clean hybrid of PINN-style residual evaluation and FBSDE data distributions, and it is distinct from the random finite-difference tricks in [11,17]. The paper also deserves credit for testing against exact or high-accuracy references and for openly flagging the L2-loss limitation in Section 5.4.\n\nNow the soft spots. First, the stress-test's central complaint about Appendix A does not land. The appendix writes εΔt, derives Var(εΔt)=O(Δt^2), and that immediately implies Var(ε)=O(1) for the estimator error ε. The notation is sloppy and easy to misread—the paper should have said 'we analyze Δt·ε'—but the claimed bound (20) is not contradicted. The real theoretical gap is Section 5.2. The argument there assumes residuals are linear in θ and that the perturbation is independent of θ, both false in the actual algorithm since the noise is built from ∇²u_θ. The paper calls it non-rigorous, which is honest, but the mechanism that lets M=64 work in d=10000 is still unexplained beyond a toy model and experiments. That is a genuine gap, though not fatal for an engineering contribution.\n\nSecond, reproducibility is incomplete. No code is released, only 'available upon request,' and the comparisons are thin: only SDE matching is benchmarked in detail, there are no runtime tables, and the Allen-Cahn case reports a single number without error bars. The d=10000 test is on a different GPU.\n\nOverall, the numerical evidence is credible and the method is practical. The theory is heuristic at the convergence level, and the appendix needs a rewrite for clarity. I'd send it to review and ask for code, a cleaner statement of the variance scaling, and at least one comparison to diffusion loss or DeepMartNet. The stress-test note should not be the basis for a desk reject.","headline":"Incremental but real method for high-dimensional parabolic PDEs; numerics are strong, theory is partial, and the Appendix A inconsistency the stress-test claims is a misreading.","tokens_in":16595,"tokens_out":10809,"would_cite":true,"duration_ms":94176,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["65C30","65M99"],"pacs":[],"model":"deepseek-v4-flash","headline":"A deep 'shotgun method' solves high-dimensional parabolic PDEs, including nonlinear ones, by estimating residuals from coarse trajectory samples with local antithetic noise, matching fine-step deep FBSDE methods in accuracy.","keywords":["high-dimensional parabolic PDEs","forward-backward stochastic differential equations","deep learning","shotgun method","antithetic variates","randomized finite differences","physics-informed neural networks","curse of dimensionality"],"falsifier":"Rerun the 100-dimensional Hamilton–Jacobi–Bellman or Black–Scholes–Barenblatt test with M = 1 and with M = 1024 at the same $\\Delta t$ and compare relative errors; if larger M does not reduce the error or if the M = 1 run diverges, the claim that averaging local noises controls the estimator's variance is falsified. Alternatively, measure the estimator bias and variance directly on a known quadratic solution v: if the bias does not decrease with $\\Delta t$ or the variance does not decrease with M, the error estimate (20) fails.","tokens_in":15559,"feed_emoji":"🎯","tokens_out":8877,"duration_ms":88399,"temperature":0.7,"pith_summary":"The paper proposes a deep learning method that solves high-dimensional parabolic PDEs without simulating full fine-discretized stochastic trajectories. Instead, it samples collocation points from coarsely discretized FBSDE trajectories and estimates the PDE residual at each point by a one-step Euler–Maruyama scheme with antithetic variates, averaging multiple independent noise draws. The central claim is that this 'shotgun' residual estimator makes a small number of coarse time steps (N = 10 for T = 1) sufficient for accuracy comparable to existing deep FBSDE solvers using hundreds of steps, and that the method remains accurate up to dimension 10000 and time horizon 10. The paper demonstrates this on the Hamilton–Jacobi–Bellman, Allen–Cahn, and Black–Scholes–Barenblatt equations, including nonlinear versions. A sympathetic reader would care because the method decouples local accuracy from trajectory length, addressing the long-time and high-dimensional limitations of existing deep FBSDE methods.","feed_headline":"One-step residuals solve 10,000-dimensional parabolic PDEs","feed_subtitle":"Local antithetic noise estimates match a 192-step solver with only 10 coarse time steps.","key_machinery":"The load-bearing object is the single-step residual estimator in equation (18), built from an antithetic pair of one-step Euler–Maruyama updates: from a point $(t,x)$ and a noise draw $\\eta\\sim N(0,\\Delta t I_d)$, it forms $x^+=x+\\mu\\Delta t+\\sigma\\eta$ and $x^-=x+\\mu\\Delta t-\\sigma\\eta$, then computes the central difference quotient $[v(t+\\Delta t,x^+)+v(t+\\Delta t,x^-)-2v(t,x)]/(2\\Delta t)-\\phi$. This estimator cancels the gradient terms $\\nabla v(t,x)$ entirely, avoiding Hessian evaluation, and averaging $M$ independent noise draws reduces its variance. It is combined with a coarse trajectory sampler in which each trajectory uses a random shift $\\delta\\sim\\mathrm{Uniform}(0,\\ell)$ so that the sampled time points have uniform marginal distribution, and each of a modest number of trajectories supplies collocation points for the residual loss.","core_discovery":"The central claim is that a single step of the Euler–Maruyama discretization of the FBSDE system, used with an antithetic pair of Brownian increments $\\eta$ and $-\\eta$, yields a stochastic estimator of the PDE residual at a point without evaluating second derivatives. Concretely, for a test function $v$, equation (18) gives $$\\frac{v(t+\\$\\Delta$ t,x^+)+v(t+\\$\\Delta$ t,x^-)-2v(t,x)}{2\\$\\Delta$ t}-\\phi=(\\mathcal{L}[v](t,x)-\\phi)+\\varepsilon,$$ where the gradient terms $\\nabla v$ cancel, so the estimator is a randomized central finite difference. The paper claims that as $\\Delta t\\to 0$, the estimator's bias is $O(\\Delta t)$ and its variance is $O(1)$, and that averaging $M$ independent noise draws reduces variance at rate $O(1/M)$; this is what lets the method use a coarse trajectory step $\\ell=T/N$ together with a smaller local step $\\Delta t$. The training loss is a PINN-style squared residual over points taken from trajectories with a random time shift, plus terminal matching. The paper reports that N = 10 and M = 16 matches the SDE matching method with N = 192 in accuracy at lower backward-propagation cost, and that accuracy persists for d up to 10000 and T = 10.","pith_inferences":["A testable extension: applying the same antithetic residual estimator to elliptic or time-independent PDEs, beyond the 1-D Laplace illustration, would show whether the machinery generalizes outside the parabolic setting.","The Section 5.2 heuristic implies that the learning-rate schedule, not just batch size, controls how estimator noise enters the final error; a comparison of constant versus decaying learning rate at fixed M would test this directly.","Because local noise averaging is independent per collocation point, the estimator could be combined with quasi-Monte Carlo or stratified sampling of $\\eta$ to reduce the needed batch M; the paper does not explore this."],"forward_implications":["Coarse trajectory discretization suffices: on the 100-dimensional Black–Scholes–Barenblatt equation, N = 10 with local batch 16 matches the accuracy of the SDE matching method with N = 192, so solver cost no longer scales with trajectory fineness.","Long horizons become tractable: the method solves the 100- and 1000-dimensional Black–Scholes–Barenblatt equation with T = 10 at relative error within about 2%, a regime where trajectory-matching methods need very many steps.","Extreme dimensions are reachable: tests run up to d = 10000 with local batch M = 64, showing that M can stay far smaller than d.","The residual estimator is gradient-free whenever the PDE coefficients do not depend on $\\nabla u$, which removes the $O(d^2)$ Hessian cost that limits PINN for high-dimensional problems.","Because residuals at different collocation points are independent, backward propagation avoids accumulating DNN outputs along trajectories, lowering training cost relative to FBSNN and SDE matching methods."],"supporting_citations":[{"why":"This reference supplies the Pardoux–Peng FBSDE theory that connects the parabolic PDE to the FBSDE system, which is the foundation of all deep FBSDE methods reviewed and modified here.","marker":"[13]"},{"why":"This reference introduces the Deep BSDE approach and the high-dimensional test problems, including Hamilton–Jacobi–Bellman and Allen–Cahn, that the shotgun method is compared against.","marker":"[5]"},{"why":"This reference presents the FBSNNs method whose single-step discretization is the direct precursor of the shotgun residual estimator.","marker":"[14]"},{"why":"This reference introduces the SDE matching method that serves as the main baseline; the paper's central accuracy comparison (N = 10 versus N = 192) is against this work.","marker":"[20]"},{"why":"This reference defines the PINN loss that the shotgun method adapts, replacing Hessian-based residuals with the stochastic residual estimator.","marker":"[15]"},{"why":"This reference provides the hard terminal-constraint trick used in the Black–Scholes–Barenblatt tests to enforce the terminal condition $u_\\theta(T,x)=g(x)$.","marker":"[9]"}],"fun_headline_variants":["Shotgun method: one-step residual estimator tackles 10,000-D PDEs","Coarse steps, fine accuracy: shotgun solves parabolic PDEs","10,000-dimensional parabolic PDEs solved with one Euler step per residual","Antithetic Brownian pairs enable one-step residual estimation for 10,000-D PDEs","Deep shotgun method: one-step antithetic residual estimator for 10,000-D PDEs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that the noisy residual estimate, computed at one finite local step size and with a limited batch of noise draws, is an accurate enough training signal that minimizing it steers the network to the true PDE solution; the paper supports this only by a non-rigorous small-learning-rate heuristic and by numerical examples.","fun_headline_variants_meta":{"raw":{"variants":["Shotgun method: one-step residual estimator tackles 10,000-D PDEs","Coarse steps, fine accuracy: shotgun solves parabolic PDEs","10,000-dimensional parabolic PDEs solved with one Euler step per residual","Antithetic Brownian pairs enable one-step residual estimation for 10,000-D PDEs","Deep shotgun method: one-step antithetic residual estimator for 10,000-D PDEs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000911,"raw_usage":{"total_tokens":3904,"prompt_tokens":928,"completion_tokens":2976,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":2871}},"tokens_in":544,"tokens_out":2976,"duration_ms":20281,"temperature":1.0,"reasoning_tokens":2871,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:34:21.261892+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the 100-dimensional Hamilton–Jacobi–Bellman or Black–Scholes–Barenblatt test with M = 1 and with M = 1024 at the same $\\Delta t$ and compare relative errors; if larger M does not reduce the error or if the M = 1 run diverges, the claim that averaging local noises controls the estimator's variance is falsified. Alternatively, measure the estimator bias and variance directly on a known quadratic solution v: if the bias does not decrease with $\\Delta t$ or the variance does not decrease with M, the error estimate (20) fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"This reference supplies the Pardoux–Peng FBSDE theory that connects the parabolic PDE to the FBSDE system, which is the foundation of all deep FBSDE methods reviewed and modified here."},{"cited_title":"Proceedings of the National Academy of Sciences 115(34), 8505–8510 (2018)","cited_arxiv_id":null,"evidence_quote":"This reference introduces the Deep BSDE approach and the high-dimensional test problems, including Hamilton–Jacobi–Bellman and Allen–Cahn, that the shotgun method is compared against."},{"cited_title":"In: Peter Carr Gedenkschrift: Research Advances in Mathematical Finance, pp","cited_arxiv_id":null,"evidence_quote":"This reference presents the FBSNNs method whose single-step discretization is the direct precursor of the shotgun residual estimator."},{"cited_title":"Journal of Computational Physics 470, 111557 (2022)","cited_arxiv_id":null,"evidence_quote":"This reference introduces the SDE matching method that serves as the main baseline; the paper's central accuracy comparison (N = 10 versus N = 192) is against this work."},{"cited_title":"Journal of Computational physics 378, 686–707 (2019)","cited_arxiv_id":null,"evidence_quote":"This reference defines the PINN loss that the shotgun method adapts, replacing Hessian-based residuals with the stochastic residual estimator."},{"cited_title":"SIAM Journal on Scientific Computing43(6), B1105– B1132 (2021)","cited_arxiv_id":null,"evidence_quote":"This reference provides the hard terminal-constraint trick used in the Black–Scholes–Barenblatt tests to enforce the terminal condition $u_\\theta(T,x)=g(x)$."}],"review_version":2}