{"id":"a92b41eb-f791-4bd8-af38-a46752fc0285","arxiv_id":"2501.00780","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A deep learning particle method is proposed that represents McKean-Vlasov SDE solutions as neural functions of Brownian motion and minimizes coefficient residuals, but representation is not proved to be expressive enough.","lead":"The authors train neural networks to represent the random paths of interacting-particle systems, using calculus rules to match the drift and noise terms without a fine time grid. The value would be faster, parallel simulations, but the paper does not prove such representations exist for the general equations it targets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ansatz (2.5) restricts solutions to deterministic functions of the current Brownian value; for non-degenerate diffusions no such F exists, so the residual hypotheses (3.2)-(3.3) are unattainable and the h-independent convergence claim rests on a false premise.","rationale":"The reader's weakest assumption is exactly the one that breaks the paper. Theorems 3.6, 3.7, and 3.9 are conditional stability estimates and, taken as conditional statements, are not the problem; the problem is that their hypotheses are not satisfiable for the proposed ansatz on generic non-degenerate MV-SDEs. The example of a one-dimensional OU process is not an MV-SDE, but it isolates the core mathematical point: if the method cannot represent the solution map of a linear SDE with constant diffusion, it cannot represent the solution map of a nonlinear mean-field SDE. The conditional estimates would become useful if Y were path-dependent (e.g., a functional of W_{[0,t]} or a neural SDE), and the shared-F i.i.d. construction is a reasonable way to avoid propagation of chaos once representability is restored. But as written, the paper's central claim—that the loss can be minimized independently of h and the error goes to zero—is unsupported by any convergence theorem, and the experiments do not report the residuals that the theorems require. I agree with the reader's REJECT verdict; no adjustment is needed.","tokens_in":22212,"tokens_out":16704,"duration_ms":176240,"concrete_test":"Analytically evaluate inf_F L(F) for dX_t=-X_t dt+dW_t, X_0=0, with L(F)=∫_0^T E[|F_t+1/2F_ww+F|^2+|F_w-1|^2]dt over F∈C^{1,2}, F(0,·)=0. For any F, the Malliavin derivative of Y_t=F(t,W_t) is D_sY_t=F_w(t,W_t) for s<t, whereas D_sX_t=e^{-(t-s)}; hence if L(F) is small, F_w(t,W_t) must approximate both 1 and the s-dependent kernel e^{-(t-s)}, which is impossible. Computing this infimum (or a positive lower bound) settles whether the residual hypotheses (3.5)-(3.6) are attainable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"All three main theorems (3.6, 3.7, 3.9) are conditional: if the residuals (3.2)-(3.3) (or (3.5)-(3.6), (3.8)-(3.9)) are small, then path and Wasserstein errors are small. The paper never shows the residuals can be made small, and in general they cannot, because of the ansatz Y_t=F(t,W_t) in (2.5). For each fixed t, F(t,W_t) is measurable with respect to the single random variable W_t, whereas the solution X_t of a non-degenerate SDE is a functional of the whole Brownian path. Concrete obstruction: take dX_t=-X_t dt+dW_t, X_0=0. Its solution X_t=∫_0^t e^{-(t-s)}dW_s has Var(X_t|W_t)=(1-e^{-2t})/2-(1-e^{-t})^2/t>0 for t>0, so no Borel F can satisfy X_t=F(t,W_t). Equivalently, with σ=1 the diffusion match forces F_w=1, hence F(t,w)=w+C(t), and the pseudo-drift C'(t) cannot match the state-dependent drift -w-C(t). The same obstruction holds for MV-SDEs with mean-field drift. The universal approximation theorem cited ([28]) only approximates functions of (t,w); it cannot approximate path functionals. The experiments report distributional MSE, not the residuals ε_2, ε_3, so they do not demonstrate that the hypotheses hold.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a deep-learning, PINN-type solver for McKean-Vlasov SDEs. The solution is represented as Y_t^n = F^n(t, W_t^n) (or by a shared F), and the Itô expansion of this ansatz is matched to the drift and diffusion coefficients of the MV-SDE through an L2 loss. The authors provide conditional error estimates in Lp and Wasserstein distance for three settings: interacting particle systems with separate networks, i.i.d. copies with a shared network, and self-interacting diffusions for stationary distributions. Numerical experiments on a Burgers-type MV-SDE, an fBm-driven model, and a stationary-distribution example compare the method with Euler-Maruyama particle simulation.","tokens_in":22544,"tokens_out":6577,"duration_ms":64516,"significance":"If the central representation were valid, the meshless formulation and the decoupling of the time-discretization error from the optimization error would be a useful contribution to MV-SDE simulation. The paper's conditional stability estimates are standard, and the use of the Fournier-Guillin i.i.d. Wasserstein bound in the shared-F case is appropriate. However, the method's main premise—that a solution can be written as a deterministic function of the current Brownian value—is false for generic non-degenerate diffusions, and the experiments do not report the residual quantities on which the theorems condition. The h-independence claim is therefore not supported. These issues affect the core theoretical claims of the paper.","major_comments":[{"comment":"The ansatz Y^{n,N}_t = F^n(t, W^n_t) in Eq. (2.5) is the load-bearing premise of the method, but it is not valid for non-degenerate SDEs. For fixed t, F^n(t, W^n_t) is measurable with respect to the single random variable W^n_t, whereas the solution X_t of (2.2) is in general a functional of the entire Brownian path. A concrete obstruction is the scalar SDE dX_t = -X_t dt + dW_t, X_0 = 0, whose solution X_t = ∫_0^t e^{-(t-s)} dW_s has Var(X_t | W_t) = (1-e^{-2t})/2 - (1-e^{-t})^2/t > 0 for t > 0, so no Borel F exists with X_t = F(t, W_t). Equivalently, with σ = 1 the diffusion matching condition forces F_w = 1, hence F(t,w) = w + C(t), and the resulting pseudo-drift C'(t) cannot match the state-dependent drift -w - C(t). The same obstruction carries over to the MV-SDEs considered. Consequently the residual hypotheses (3.2)-(3.3), (3.5)-(3.6), and (3.8)-(3.9) are generally not achievable, and the universal approximation result cited ([28]) only approximates functions of (t,w), not path functionals.","section":"§2.1, Eq. (2.5)"},{"comment":"The three main theorems are conditional stability statements: they assert that if the training residuals are bounded by ε_i, then the Lp path error and Wasserstein error are bounded by constants that vanish as ε_i → 0 (and N → ∞ where claimed). Since these residuals are exactly the loss terms L2 and L3 minimized in Algorithm 4.1, the theorems do not by themselves provide an a priori error estimate for the proposed method. An a priori estimate requires either a proof that the residuals can be driven to zero, or a representability result for the ansatz class, or an optimization guarantee. Neither is supplied. In view of the obstruction in Eq. (2.5), the h-independence claim in Sections 2.1 and 2.4 rests on an unverified and generally false premise.","section":"§3.3, Theorems 3.6, 3.7, 3.9"},{"comment":"The residual conditions such as (3.2) and (3.3) are stated as continuous-time integrals, but the loss in Algorithm 4.1 is a finite sum over training time points. The paper provides no estimate for the difference between the empirical loss and the continuous residuals, so the theorems do not imply that small training loss yields small ε_2 and ε_3. Hence the advertised independence from the time step h is not established; h still enters through the quadrature of the loss integrals. The meshless numerical experiments with deleted time points illustrate robustness but do not fill this gap.","section":"§4.1, Algorithm 4.1 and §3.3"},{"comment":"Theorem 3.9 states that C1 and C2 converge to 0 when T → ∞ and ε_1, ε_2, ε_3 → 0, but the proof yields constants that grow like exp((2L + C^2)t), so they blow up as t increases. No long-time argument is given to offset this exponential growth by the contraction in Hypothesis 3.8. The claimed convergence to the stationary distribution therefore does not follow from the displayed estimates; at best the bounds hold on compact time intervals with constants depending exponentially on T.","section":"§3.3, Theorem 3.9"}],"minor_comments":[{"comment":"The sentence beginning 'Under some mild conditions [9, 32],' is incomplete and should be finished or merged with the following sentence.","section":"§2.3"},{"comment":"The notation 'b^{P seudo}' and 'σ^{P seudo}' contains an obvious typo; it should be 'pseudo' throughout.","section":"§2.1, Eqs. (2.6)-(2.7)"},{"comment":"The loss weights λ_i in Algorithm 4.1 are not defined, and no normalization by the number of time points D is specified, making the relationship between the empirical loss and the integral residuals in (3.2)-(3.3) unclear.","section":"§4.1, Algorithm 4.1"},{"comment":"The training time for PEM is listed as '×' without explanation; a short note stating why this entry is not applicable would improve the table.","section":"§4.2, Table 1"},{"comment":"The claim that omitting propagation of chaos is 'a theoretical breakthrough' is overstated; the shared-F setting still requires the empirical measure \\v{\\nu}^{(N)}_t to approximate \\nu_t, and convergence is only supplied through the i.i.d. Wasserstein bound of Lemma 3.4.","section":"§2.2"},{"comment":"The experiments report the distributional MSE, not the residual quantities ε_2 and ε_3 from (3.2)-(3.3). Reporting the training residuals would directly test the hypotheses on which the theorems condition.","section":"§4.2.1"}],"recommendation":"reject","confidential_remarks":"The central theoretical premise is not merely unproved but false for generic non-degenerate SDEs, and this affects all three main theorems. The paper could become viable only with a substantially different ansatz, such as a path-dependent representation or a rigorous reduction to a Fokker-Planck formulation, together with a proper treatment of the time-discretized loss. I do not see a fix within the scope of the current manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things up front about arXiv:2501.00780. The shared-network construction — one F_θ applied to i.i.d. Brownian paths, with empirical-measure error handled by Fournier-Guillin instead of propagation of chaos — is genuinely new and structurally cleaner than the standard particle-plus-Euler approach. The problem is the central claim: the advertised h-independent convergence rests on an ansatz that cannot represent non-degenerate diffusions, so the residual hypotheses in Theorems 3.6, 3.7, and 3.9 are generally unattainable.\n\nCredit where it is earned. The conditional estimates are correctly argued with standard tools (BDG, Minkowski, Gronwall), the constants are honestly exponential in T, and as a posteriori stability bounds they are legitimate. The shared-F idea turns the interacting particle problem into Monte Carlo approximation of the law of F_θ(t, W_t) and cleanly drops the propagation-of-chaos bookkeeping; that is worth keeping. The deletion experiments show real robustness to non-uniform grids, the most convincing empirical point in the paper.\n\nThe soft spot is load-bearing. Equation (2.5) restricts each particle to Y_t = F(t, W_t), a function of the current Brownian value. A non-degenerate diffusion is a functional of the whole Brownian path. Take dX_t = -X_t dt + dW_t, X_0 = 0; the solution is ∫_0^t e^{-(t-s)} dW_s, and Var(X_t | W_t) = (1-e^{-2t})/2 - (1-e^{-t})²/t > 0 for every t > 0, so no F satisfies X_t = F(t, W_t). With σ = 1 the diffusion match forces F(t,w) = w + C(t), and then the drift residual C'(t) + w + C(t) has L² norm at least √t; it cannot be trained away. The same obstruction holds for mean-field drifts. The paper's Remark after Theorem 3.6 asserts the residuals can be controlled by thresholding the loss — that is exactly the claim that fails — and the Note in Section 3.3, which justifies approximation by tanh networks, covers only functions of (t, w), not path functionals. The experiments report distributional MSE, never the residuals ε₂, ε₃ the theory controls, and the baselines (h = 1, M = 2) are weak, with no error bars or code.\n\nThe fix is clear: give F_θ access to the whole path, or restrict the scope to SDEs admitting a Lamperti-type reduction. As written, the method is a conditional residual scheme without a convergence theorem. Anyone working on PINN solvers for SDEs should read this as a cautionary example, because the failure mode — an inexpressive ansatz presented as a solver — is common. It deserves a serious referee: the obstruction is subtle and the shared-network idea has merit, but the expectation should be heavy revision or rejection. I would cite it as related work and bring it to reading group.","headline":"A clean shared-network i.i.d. construction and honest conditional stability bounds are undercut by an ansatz, Y=F(t,W_t), that cannot represent non-degenerate diffusions, so the claimed h-independent convergence does not go through as stated.","tokens_in":23060,"tokens_out":16802,"would_cite":true,"duration_ms":141442,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60H10","65C30","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Training neural functions of Brownian paths to match Itô drift and diffusion gives a time-step-independent solver for McKean-Vlasov SDEs.","keywords":["McKean-Vlasov equation","interacting particle system","physics-informed neural networks","meshless solver","propagation of chaos","Wasserstein distance","fractional Brownian motion"],"falsifier":"Consider the scalar linear SDE $dX_t=-X_t\\,dt+e^{-t}\\,dW_t$ with $X_0=0$, whose solution depends on the whole Brownian path and not only on $W_t$. Train the proposed network $F_\\theta(t,W_t)$ on this equation; if the minimal drift and diffusion residuals in (3.2)-(3.3) cannot be driven below a strictly positive floor while an Euler-Maruyama reference converges, the claim that error is controlled solely by the training loss fails.","tokens_in":21943,"feed_emoji":"🧠","tokens_out":7710,"duration_ms":73624,"temperature":0.7,"pith_summary":"The paper introduces a way to solve McKean-Vlasov stochastic differential equations without a fixed time mesh. Instead of Euler steps on interacting particles, it writes each particle path as a neural network applied to that particle's Brownian motion, uses Itô's formula to read off what drift and diffusion the network path is actually following, and trains the network until those pseudo coefficients match the true mean-field drift and diffusion. The authors claim the resulting error is controlled by the training loss, not by the time step size, and that in the one-network version the target is the MV-SDE itself, so propagation of chaos is not needed. This would matter because traditional particle solvers slow down sharply on long time horizons or strongly nonlinear coefficients, whereas this method is meshless, parallelizable, and extends to noises like fractional Brownian motion. The self-interacting variant also approximates stationary distributions from a single trajectory without storing long sample histories.","feed_headline":"Neural paths approximate MV-SDEs without a time mesh","feed_subtitle":"Training F(t,W_t) to match drift and diffusion makes error depend on loss, not step size, and skips propagation of chaos.","key_machinery":"The central object is the pseudo MV-SDE obtained by applying Itô's formula to the ansatz $Y^{n,N}_t=F^n(t,W^n_t)$. Its drift is $b^{\\mathrm{pseudo}}_{n,t}=F^n_1(t,W^n_t)+\\frac12 F^n_{22}(t,W^n_t)$ and its diffusion is $\\sigma^{\\mathrm{pseudo}}_{n,t}=F^n_2(t,W^n_t)$. The loss is the sum of the initial mismatch and the $L^2$ discrepancies between these pseudo coefficients and the true $b$ and $\\sigma$ evaluated at $Y^{n,N}_t$ and the empirical measure $\\hat\\nu^{(N)}_t$. The proofs feed those residuals through stochastic-integral estimates and a nonlinear Gronwall lemma to obtain sup-in-time $L^p$ and Wasserstein bounds; in the shared-$F$ case, an empirical-measure Wasserstein bound controls the sampling error.","core_discovery":"On the paper's own terms, the central claim is that an MV-SDE solution can be approximated by neural paths $Y^{n,N}_t=F^n(t,W^n_t)$, where the network is trained so that the Itô-formula pseudo coefficients match the true coefficients: the pseudo drift $F^n_1+\\frac12F^n_{22}$ is matched to $b(t,Y^{n,N}_t,\\hat\\nu^{(N)}_t)$ and the pseudo diffusion $F^n_2$ is matched to $\\sigma(t,Y^{n,N}_t,\\hat\\nu^{(N)}_t)$. Theorems 3.6, 3.7, and 3.9 then bound the sup-in-time $L^p$ error and the Wasserstein distance by constants that vanish as the initial, drift, and diffusion residuals go to zero, and those constants do not involve the discretization step $h$. In the shared-$F$ version, the outputs for $N$ independent Brownian motions are i.i.d., so the empirical measure approximates the true law by the standard i.i.d. Wasserstein bound and propagation of chaos is not invoked.","pith_inferences":["The representability assumption $Y_t=F(t,W_t)$ is the point most worth testing: for non-degenerate diffusions the true path generally depends on the whole Brownian past, so a path-dependent equation would expose an irreducible lower bound on the residuals in (3.2)-(3.3).","If the meshless error control survives uneven sampling, the same loss could be used for parameter estimation or data assimilation from irregularly observed SDE trajectories, a natural extension the paper does not develop.","The $h$-independence statement concerns the training loss; the loss itself is evaluated at sampled Brownian points, so a careful quadrature-error analysis would be needed to confirm the claim for arbitrarily sparse or randomly deleted observation times.","The shared-$F$ formulation's slower Monte Carlo convergence suggests a hybrid: use the interacting-particle loss for a few networks to get fast distributional accuracy and then fine-tune a single shared network for pathwise predictions."],"forward_implications":["The training loss replaces the Euler step-size error, so Brownian trajectories can be sampled at coarse, uneven, or partially missing time points while keeping the same error bounds.","In the single-network version, the approximation targets the MV-SDE directly from i.i.d. Brownian paths, so propagation of chaos is not needed as an intermediate step.","The same Itô-formula loss works for other noises; replacing the chain rule yields a solver for McKean-Vlasov SDEs driven by fractional Brownian motion.","The self-interacting one-path version approximates the stationary distribution from a single trajectory without storing long histories of samples.","Because all training points are evaluated inside one loss, the method can exploit GPU acceleration for large particle counts."],"supporting_citations":[{"why":"Supplies the baseline convergence rate $O(1/N+h)$ for Euler particle methods that the new method claims to improve by removing the $h$-dependence.","marker":"[1]"},{"why":"Gives the Wasserstein bound for empirical measures of i.i.d. samples used in the shared-$F$ error analysis.","marker":"[11]"},{"why":"Provides the tanh-network approximation result used to justify representing $F(t,W_t)$ in $C^{1,2}$.","marker":"[28]"},{"why":"Establishes convergence of self-interacting diffusions to stationary distributions, which the self-interacting variant relies on.","marker":"[9]"},{"why":"Provides the residual-minimization PINN paradigm that motivates the loss design.","marker":"[26]"},{"why":"Source of the Lipschitz hypothesis and particle/quantization schemes for McKean-Vlasov equations used as comparison.","marker":"[22]"},{"why":"Gives the Burgers-equation benchmark with explicit distribution used in the numerical experiments.","marker":"[6]"}],"fun_headline_variants":["Neural network method solves McKean-Vlasov without time mesh","Deep learning eliminates time step for McKean-Vlasov SDEs","No time mesh needed: deep learning for MV-SDEs","Meshless deep learning solver for McKean-Vlasov equations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that at every time $t$ the approximate particle path is fully determined by the current value of its driving Brownian motion; for a generic diffusion the true path depends on the whole Brownian past, and when that happens the trained residuals in the proof cannot be made arbitrarily small.","fun_headline_variants_meta":{"raw":{"variants":["Neural network method solves McKean-Vlasov without time mesh","Deep learning eliminates time step for McKean-Vlasov SDEs","No time mesh needed: deep learning for MV-SDEs","Meshless deep learning solver for McKean-Vlasov equations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000654,"raw_usage":{"total_tokens":3039,"prompt_tokens":1028,"completion_tokens":2011,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":1934}},"tokens_in":644,"tokens_out":2011,"duration_ms":13966,"temperature":1.0,"reasoning_tokens":1934,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:42:25.803251+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Consider the scalar linear SDE $dX_t=-X_t\\,dt+e^{-t}\\,dW_t$ with $X_0=0$, whose solution depends on the whole Brownian path and not only on $W_t$. Train the proposed network $F_\\theta(t,W_t)$ on this equation; if the minimal drift and diffusion residuals in (3.2)-(3.3) cannot be driven below a strictly positive floor while an Euler-Maruyama reference converges, the claim that error is controlled solely by the training loss fails.","supporting_citations":[{"cited_title":"Antonelli and A","cited_arxiv_id":null,"evidence_quote":"Supplies the baseline convergence rate $O(1/N+h)$ for Euler particle methods that the new method claims to improve by removing the $h$-dependence."},{"cited_title":"Fournier and A","cited_arxiv_id":null,"evidence_quote":"Gives the Wasserstein bound for empirical measures of i.i.d. samples used in the shared-$F$ error analysis."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the tanh-network approximation result used to justify representing $F(t,W_t)$ in $C^{1,2}$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes convergence of self-interacting diffusions to stationary distributions, which the self-interacting variant relies on."},{"cited_title":"Raissi, P","cited_arxiv_id":null,"evidence_quote":"Provides the residual-minimization PINN paradigm that motivates the loss design."},{"cited_title":"Particle method and quantization-based schemes for the simulation of the McKean-Vlasov equation","cited_arxiv_id":"2212.14853","evidence_quote":"Source of the Lipschitz hypothesis and particle/quantization schemes for McKean-Vlasov equations used as comparison."},{"cited_title":"Bossy and D","cited_arxiv_id":null,"evidence_quote":"Gives the Burgers-equation benchmark with explicit distribution used in the numerical experiments."}],"review_version":1}