REVIEW 2 major objections 5 minor 4 references
Multilevel Picard approximations for McKean-Vlasov stochastic differential equations with nonconstant diffusion
T0 review · 2 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash
Pith's one-line read This paper proves that a multilevel Picard algorithm approximates McKean–Vlasov SDEs with nonconstant diffusion in $L^2$ with cost polynomial in the dimension and in $1/\epsilon$.
desk verdict Solid MLP extension to nonconstant-diffusion McKean–Vlasov SDEs; the theorem is likely true, but equation (63) states a false equality that must be downgraded to an inequality. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The scheme is built on a Picard iteration $X_n$ whose defining identity is re-expressed as a telescoping sum over levels $\ell$ of differences of expected drift and diffusion terms, with $\mathbb{E}[\mu^d(x,X_\ell)]$ and $\mathbb{E}[\sigma^d(x,X_\ell)]$ replaced by sample averages using independent Brownian paths. A random time $tu$ with $u$ uniform on $[0,1]$ approximates the Riemann integral, and the left-step function $\lfloor s\rfloor_K$ places everything on the grid $\{0,T/K,\dots,T\}$. The central mechanism is storing the whole process on this grid, so the stochastic integral $\int_0^t [\sigma^d(X_\ell,\tilde X_\ell)-\sigma^d(X_{\ell-1},\tilde X_{\ell-1})]\,dW$ can be evaluated; this is what distinguishes the nonconstant-diffusion case.
What would settle it
Fix $d$ and any coefficients satisfying (2), set $m=n$ and $K=n^n$, and measure $\sup_{t\in[0,T]}\|X^{d,0}_{n,n,n}(t)-X^{d,0}_{\infty}(t)\|_{L^2}$ for increasing $n$; Theorem 1 predicts this error decays super-polynomially while the cost stays consistent with a bound like $C^d_{n,n,n^n}\le(8n)^{2n}d^{c+1}$. If the measured error fails to converge, or if the smallest $n$ achieving a given $\epsilon$ grows faster than any polynomial in $1/\epsilon$, the theorem's conclusion is contradicted.
Extended reading notes
Core claim
Under the Lipschitz and growth conditions (2)–(3) with a single constant $c$, Theorem 1 proves that for every dimension $d$ and tolerance $\epsilon$ there is a level $n_{d,\epsilon}$ such that the MLP approximation (4) with $n=m=K=n_{d,\epsilon}$ satisfies $\sup_{t\in[0,T]}\|X^{d,0}_{n_{d,\epsilon},n_{d,\epsilon},n_{d,\epsilon}}(t)-X^{d,0}_{\infty}(t)\|_{L^2}<\epsilon$, and its computational cost is bounded by $(cd^c+1)^6 C_\delta \epsilon^{-(4+\delta)}$ for any $\delta\in(0,1)$. The new ingredient is that the scheme stores the whole discrete trajectory $(X^{d,\theta}_{n,m,K}(kT/K))_{k=1,\dots,K}$ rather than a single time-point value, which lets it approximate the stochastic integral against the nonconstant $\sigma^d$ via Monte Carlo; the constant-diffusion MLP of [38] needed no such storage.
Load-bearing premise
The whole argument depends on a single Lipschitz bound $c$ that stays fixed as the dimension grows; if the drift and diffusion coefficients become more sensitive with $d$, the promised polynomial cost bound no longer follows.
Editorial extensions
If this is right
- For any family of McKean–Vlasov SDEs satisfying the uniform Lipschitz condition (2), the MLP scheme delivers the whole time trajectory on a $K$-point grid with $L^2$ error below $\epsilon$ and cost $O(d^{6c}\epsilon^{-(4+\delta)})$, so exponential cost growth in the dimension is avoided.
- In the two test problems, a mean-field Ornstein–Uhlenbeck process and a multidimensional geometric Kuramoto model, the dimension-adjusted $L^2$ error stays roughly flat as $d$ goes from 10 to 1000, consistent with the theoretical bound.
- Lemma 3 gives a stand-alone existence and uniqueness result and a rate $O(\sqrt{T/K})$ for time discretization of the McKean–Vlasov equation, so the discretization error is quantitatively controlled.
- Because the method outputs the full path, path-dependent quantities such as occupation times or hitting statistics can be computed from the same run without restarting the algorithm.
Reading between the lines
- This reader notes that the theorem's polynomial-in-$d$ conclusion rests on the Lipschitz constant $c$ being independent of $d$; for a family of equations in which $c=c_d$ grows with $d$, the cost bound $(cd^c+1)^6$ would become super-polynomial, so stress-testing the algorithm on such families would clarify how far the no-curse statement extends.
- The trajectory storage needed for the stochastic integral suggests the scheme's memory footprint grows like $Kd$ times the number of samples; variance-reduction techniques for the $\sigma$-differences, for instance antithetic or control-variate sampling, might reduce the constants, an extension the paper does not explore.
- The same telescoping construction should carry over to McKean–Vlasov equations with common noise or with jump terms, since both reintroduce stochastic integrals whose evaluation requires the stored path; testing this would be a natural sequel.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper introduces a multilevel Picard (MLP) approximation for McKean–Vlasov SDEs with nonconstant diffusion coefficient. The main result, Theorem 1, states that the L2 approximation error of the scheme (4) can be made smaller than any prescribed epsilon with a computational cost bounded by (c d^c + 1)^6 C_delta epsilon^{-(4+delta)}, hence polynomially in dimension d and in 1/epsilon. The proof is organized into separate steps: existence and discretization error for the time-discretized Picard iteration (Lemma 3), measurability and distributional properties, bias, statistical error, global error, and cost analysis. Two numerical experiments, a mean-field Ornstein-Uhlenbeck model and a multidimensional geometric Kuramoto model, are reported for dimensions up to 1000, with code made available.
Significance. If Theorem 1 is fully established, the paper provides the first explicit MLP scheme for McKean–Vlasov SDEs with nonconstant diffusion that provably overcomes the curse of dimensionality in the L2 sense, extending the constant-diffusion result of [38]. The proof is detailed and modular, and the numerical section demonstrates practical performance in dimensions up to 1000, which is substantially higher than most prior work. The main caveat is that the polynomial-in-d conclusion is tied to the assumption that the Lipschitz constant c in (2)-(3) and the cost bounds in (5) do not grow with d; this is a genuine scope limitation rather than an internal inconsistency. The code release and the explicit error/cost bounds are strengths that should be preserved in revision.
major comments (2)
- [Section 5, Step 5, Eq. (63)] The displayed identity in Eq. (63) is false as stated. For conditionally independent summands Y_ell, the conditional variance of the sum is the sum of the conditional variances, so sqrt(Var(sum | G)) = sqrt(sum Var(Y_ell | G)), which is generally strictly less than sum sqrt(Var(Y_ell | G)). The subsequent use of (63) in the statistical-error bound (67) can be repaired by replacing this equality with the inequality sqrt(Var(sum | G)) <= sum sqrt(Var(Y_ell | G)) and applying the triangle inequality for the L2 norm. Because (63) is the technical basis of the statistical error recursion, this correction is necessary for the proof as written to be rigorous.
- [Section 5, Step 5, Eq. (67)] Equation (67) as typeset is dimensionally inconsistent: the left-hand side is the squared L2 distance, while the right-hand side is an unsquared sum of square roots of time integrals. The derivation that follows only requires the unsquared estimate ||X - E^G X||_{L2} <= sum ...; please correct the display accordingly, either by removing the square on the left-hand side or by adding an outer square on the right-hand side and adjusting the subsequent inequalities.
minor comments (5)
- [Section 5, heading] The proof section is headed "Proof of Lemma 1" but it proves Theorem 1; please rename it to "Proof of Theorem 1".
- [Section 5, Step 1, item (C) and Eq. (52)] The discretization-error bound stated in Step 1 (C) and used in (52) has the exponent e^{4(√T+1)(T+1)c^2}, whereas the corresponding proved statement in Lemma 3(iv) has e^{5(√T+1)^2(T+1)c^2}. Please reconcile the constants or provide a derivation for the sharper exponent.
- [Section 2.2, Table 2] For the Kuramoto example, the 'true' solution is approximated by the Euler-Maruyama scheme combined with the Taylor-type approximations (16)-(17), not by an independent exact solution. This should be stated explicitly in the text and taken into account when interpreting the reported errors.
- [Algorithm 1, line 24] The indexing in line 24 is suspicious: for j=0, the expression floor((j-1)u)+1 can evaluate to 0 when u>0, which may be outside the intended array range. Please check the off-by-one behavior for j=0.
- [Throughout the proof] Several displays contain unbalanced parentheses, for example Eq. (56) and the expression X_{j,m,K}^d(⌞s⌟_K)) in Eq. (57). A careful proofreading pass is needed.
Circularity Check
No significant circularity: the MLP convergence proof is self-contained, with only minor, non-load-bearing self-citations to published technical lemmas.
full rationale
The central derivation in Theorem 1 is not circular in the sense defined here. The MLP scheme (4) is explicitly constructed as a Monte Carlo/telescoping approximation of the Picard iteration (9)-(10), and the proof establishes the L2 error by combining a bias bound (62), a statistical-error bound (67)-(70), and a time-discretization bound (52) obtained from the paper's own Lemma 3. The complexity bound is not a renamed input: n_{d,epsilon} is defined by the error criterion (78), and C_delta in (81) is a supremum over n of cost times error^{4+delta}, whose finiteness follows from the previously proved error decay (79) and cost bound (77). Thus the conclusion is not equivalent to its assumptions by construction. The cited results [36, Lemma 3.9] and [38, Corollary 2.3] share an author with the present paper, but they are published proof-based technical lemmas used as black boxes: a discrete Gronwall-type recursion and a cost recursion. They do not contain the nonconstant-diffusion convergence claim, and the central result does not reduce to them. This is a minor self-citation, not load-bearing circularity. The fixed-Lipschitz-constant assumption (2)-(3) is a scope limitation: the polynomial-in-d complexity conclusion requires c not growing with d, but this is a limitation of scope, not a circularity. Separately, the suspicious equality in Eq. (63) appears to be a proof-correctness concern rather than a circularity pattern; it does not change this verdict.
Assumptions & free parameters
assumptions (4)
- standard math Standard filtered probability space with independent Brownian motions and uniform random variables, satisfying the usual conditions.
- domain assumption Lipschitz and polynomial growth conditions (2)-(3) on mu^d, sigma^d, and xi^d, with constant c independent of d.
- domain assumption Computational cost oracle (5): evaluating mu, sigma, and generating a scalar random variable costs at most c d^c.
- standard math Background lemmas: [36, Lemma 3.9] and [38, Corollary 2.3] used in the proof of Theorem 1.
Cite this review
Pith. "Pith review of Multilevel Picard approximations for McKean-Vlasov stochastic differential equations with nonconstant diffusion." pith.science (2026). https://pith.science/paper/YMUB23CK
@misc{pith2026250203205,
author = {Pith},
title = {Pith review of: Multilevel Picard approximations for McKean-Vlasov stochastic differential equations with nonconstant diffusion},
year = {2026},
howpublished = {\url{https://pith.science/paper/YMUB23CK}},
note = {Machine review of arXiv:2502.03205}
}
abstract
We introduce multilevel Picard (MLP) approximations for McKean--Vlasov stochastic differential equations (SDEs) with nonconstant diffusion coefficient. Under standard Lipschitz assumptions on the coefficients, we show that the MLP algorithm approximates the solution of the SDE in the $L^2$-sense without the curse of dimensionality. The latter means that its computational cost grows at most polynomially in both the dimension and the reciprocal of the prescribed error tolerance. In two numerical experiments, we demonstrate its applicability by approximating McKean--Vlasov SDEs in dimensions up to 1000.
Reference graph
Works this paper leans on
-
[1]
[1]Acebr ´on, J. A., Bonilla, L. L., P ´erez Vicente, C. J., Ritort, F., and Spigler, R.The Kuramoto model: A simple paradigm for synchronization phenomena.Rev. Mod. Phys. 77(Apr 2005), 137–185. [2]Agarwal, A., Amato, A., dos Reis, G., and Pagliarani, S.Numerical approximation of McKean- Vlasov SDEs via stochastic gradient descent.arXiv preprint arXiv:231...
-
[1057]
[34]Han, J., Hu, R., and Long, J.Learning high-dimensional McKean–Vlasov forward-backward stochastic differential equations with general distribution dependence.SIAM Journal on Numerical Analysis 62, 1 (2024), 1–24. [35]Hu, K., Ren, Z., ˇSiˇska, D., and Szpruch, L.Mean-field Langevin dynamics and energy landscape of neural networks.Annales de l’Institut H...
arXiv 2024
-
[2018]
[19]Carrillo, J. A., Gvalani, R. S., Pavliotis, G. A., and Schlichting, A.Long-time behaviour and phase transitions for the McKean–Vlasov equation on the torus.Archive for Rational Mechanics and Analysis 235(2020), 635–690. [20]Chazelle, B., Jiu, Q., Li, Q., and W ang, C.Well-posedness of the limiting equation of a noisy consensus model in opinion dynamic...
arXiv 2020
-
[2019]
Mathematics and Financial Economics 12(2018), 335–363
[17]Cardaliaguet, P., and Lehalle, C.-A.Mean field game of controls and an application to trade crowding. Mathematics and Financial Economics 12(2018), 335–363. [18]Carmona, R., and Delarue, F.Probabilistic theory of mean field games with applications I-II. Probability theory and stochastic modelling volume 83+84. Springer, Cham,
work page 2018
Reviewed August 9, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.