REVIEW 2 major objections 5 minor 27 references
On the rate of the vanishing viscosity approximation for Mean Field Games with nonlocal coupling
T0 review · 2 major / 5 minor · reviewed 2026-07-09 · glm-5.2
Pith's one-line read Viscosity vanishing in mean field games gets explicit rates
desk verdict Solid contribution improving HJ rates and adding density rates for vanishing viscosity in MFGs; one stress-test concern does not land, one real limitation remains 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 argument proceeds in three stages. First, a duality argument testing the difference of two viscous solutions against the density difference yields an O(epsilon) bound on the monotonicity integral and on the weighted gradient squared integral (Theorem 4.2), using semiconcavity (Lemma 4.1) and the maximum principle for the Fokker-Planck equation. Second, the adjoint method is applied to the viscosity derivative of the value function, using the Bernstein method to control the Hessian term and assumptions (F1)/(F2) to control the coupling derivative, yielding the O(epsilon^{1/2}) rate for the value function (Theorem 4.6, Steps 1-2). Third, the value function rate is transferred to the drift,
What would settle it
If one could construct a Mean Field Game satisfying all the stated structural assumptions (H1)-(H4), (F), and monotonicity (1.5)-(1.6), but where the W^{2,infinity} bound (4.4) fails, the density convergence rate O(epsilon^{1/8}) would not apply. More critically, if the coupling regularity assumptions (F1)/(F2) were shown to be incompatible with the monotonicity conditions for any non-trivial coupling, the value function rate would also fail.
Extended reading notes
Core claim
The paper establishes that the vanishing viscosity approximation for first-order Mean Field Games with nonlocal coupling converges at rate O(epsilon^{1/2}) for the value function (in L^1 and L^infinity) and at rate O(epsilon^{1/8}) for the density in the 2-Wasserstein distance. The key technical innovation is an O(epsilon) bound on the Lasry-Lions monotonicity functional integral involving the coupling difference weighted by the density difference, combined with a dual/adjoint method exploiting semiconcavity of the value function and uniform convexity of the Hamiltonian. This bound, together with assumptions (F1) or (F2) translating the monotonicity estimate into L^infinity or L^1 control of
Load-bearing premise
The density convergence rate O(epsilon^{1/8}) requires the viscous value function to have a second spatial derivative bounded uniformly in epsilon (condition 4.4). This is not guaranteed by the general framework and holds only under additional specific conditions such as convex data, small terminal data, or short time horizons.
Editorial extensions
If this is right
- Numerical schemes for first-order MFGs that introduce artificial viscosity can now be equipped with explicit error bounds of order epsilon^{1/2} for the value function and epsilon^{1/8} for the density, guiding mesh-size and viscosity parameter choices.
- The partial decoupling strategy—using the regularizing coupling to first establish HJ convergence independently of the density—may extend to other forward-backward PDE systems where the forward component's drift depends on the backward component's gradient.
- The gap between the value function rate (epsilon^{1/2}) and the density rate (epsilon^{1/8}) suggests that the density convergence is bottlenecked by the gradient estimate (epsilon^{1/4} in L^2); improving the gradient rate would directly improve the density rate.
- The selection principle interpretation—that viscous approximations select a privileged inviscid solution—is now quantitative, allowing one to assess how small epsilon must be for the selected solution to approximate the inviscid one to a given tolerance.
Reading between the lines
- The O(epsilon^{1/8}) density rate is likely not sharp. The chain of inequalities loses rate at each step: O(epsilon^{1/2}) for u, O(epsilon^{1/4}) for the drift in L^2, then square-root via Wasserstein to get epsilon^{1/8}. A more direct argument bypassing the Wasserstein square-root step might recover a better rate.
- The assumption (F1)/(F2) on the coupling—that the monotonicity integral bound implies an L^infinity or L^1 bound on the coupling difference—is essentially a regularity assumption on the solution operator of an elliptic equation. Characterizing the minimal regularity on the convolution kernel for which these hold would clarify the boundary of applicability.
- If one could establish the W^{2,infinity} bound (4.4) under weaker conditions—for instance, for longer time horizons or non-convex data—the density convergence result would apply to a substantially broader class of MFG systems.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This manuscript studies quantitative convergence rates for the vanishing viscosity approximation of first-order, time-dependent Mean Field Games (MFGs) with a nonlocal regularizing coupling. Under standard structural assumptions on the Hamiltonian (H1)-(H4) and regularizing properties of the coupling F, the authors improve the previously known convergence rate for the value function u^ε from O(ε^{1/4}) to O(ε^{1/2}) in both L^1 and L^∞. Furthermore, by exploiting duality methods and the semiconcavity of the solutions, they establish a quantitative convergence rate for the density m^ε in the 2-Wasserstein distance, W_2(m^ε, m) ≤ C ε^{1/8}, provided an additional uniform W^{2,∞} bound on u^ε holds. The proofs rely on integral methods, Bernstein-type estimates, and stability estimates for linear Fokker-Planck equations.
Significance. The paper makes a solid contribution to the analysis of MFGs by providing the first quantitative convergence rate for the Fokker-Planck component in the vanishing viscosity limit under these structural assumptions. The improvement of the rate for the Hamilton-Jacobi equation from O(ε^{1/4}) to O(ε^{1/2}) is a notable sharpening of the results in [TZ25]. The strategy of partially decoupling the system—first transferring the HJ convergence rate to the drift field, and then propagating it to the Fokker-Planck equation via stochastic characteristics—is elegant and exploits the regularizing structure of the coupling effectively. The authors clearly delineate the conditions under which the additional W^{2,∞} bound (Assumption 4.4) is satisfied, such as convex data or short time horizons, which provides concrete, falsifiable applicability criteria for the main results.
major comments (2)
- §4, Proof of Theorem 4.6, Step 2 (Eq. 4.2): The skeptic's concern regarding the uniform L^∞ bound on the adjoint state ρ^ε is valid and represents a load-bearing gap in the current proof. In the derivation of the L^1 estimate (4.2), the authors bound the term ∫|∂_ε u^ε| dx by duality against the adjoint ρ^ε solving (4.6). The text states: 'Since ∥∆u^ε∥_{L^1(Q_T)}, ∥ρ^ε∥_{L^∞(Q_T)} are bounded independent of ε, we get...'. While the uniform bound on ∥∆u^ε∥_{L^1} is justified by Lemma 4.1 (semiconcavity), the uniform L^∞ bound on ρ^ε is not established anywhere in the manuscript. The adjoint ρ^ε solves a Fokker-Planck equation with drift D_pH(x, Du^ε) and vanishing diffusion ε. A uniform L^∞ bound on ρ^ε requires controlling the divergence of the drift, which involves D^2_{pp}H · D^2u^ε. The semiconcavity bound from Lemma 4.1 only provides an upper bound on ∆u^ε, not |∆u^ε|. The authors do
- not provide a parabolic Harnack-type argument or any other justification for why ∥ρ^ε∥_{L^∞} remains bounded independently of ε as the diffusion vanishes. If ∥ρ^ε∥_{L^∞} grows as ε → 0, the L^1 estimate (4.2) does not close. This issue must be addressed, either by providing the missing estimate for ρ^ε or by modifying the duality argument. Note that this concern does not affect the L^∞ estimate (4.1) in Step 1, where the initial datum for the adjoint is a probability measure and the estimate relies on the L^1 conservation of ρ^ε rather than its L^∞ norm.
minor comments (5)
- §1.3, Main contributions: The text mentions 'new quantitative convergence of gradients in L^∞_t(L^2_x)'. It would be helpful to explicitly state the rate O(ε^{1/4}) here, matching the statement in Theorem 4.6 (4.3), for consistency and immediate impact.
- §4, Theorem 4.2: The proof involves a double parameter η ≤ ε and then lets η → 0. The notation and the logic flow here are slightly terse; expanding the explanation of how the limit η → 0 recovers the estimates with m (rather than m^η) would improve readability.
- §4, Proof of Theorem 4.6, Step 1: The claim that '2ε ∫∫ |D^2u^ε|^2 ρ^ε dxdt ≤ C(...)' is justified via the Bernstein method on ω^ε = |Du^ε|^2. The duality argument transferring the bound from the ω^ε equation to the integral against ρ^ε is sketched very briefly. A few additional lines detailing this transfer would strengthen the rigor of this key step.
- Remark 4.5: The reference to [TZ25, Remark 7.1] for the verification of (F1)-(F2) is noted. It might be beneficial to briefly sketch the core idea of why the convolution structure F[m] = f(x, k ⋆ m) ⋆ k satisfies these assumptions, to make the manuscript more self-contained.
- Typo in §4, Proof of Theorem 4.6, Step 2: 'Since ∥∆u^ε∥_{L^1(Q_T)}, ∥ρ^ε∥_{L^∞(Q_T)} are bounded independent of ε, we get, integrating in ε, the following bound ∫ |(u^ε - u)(τ, x)| dx ≤ C_1 ε + 2C_2 √ε.' The term C_1 ε appears to be a typo, as the preceding estimate yields a bound of C_1 + C_2/√ε for ∂_ε u^ε, which upon integration in ε should give C_1 ε + 2C_2 √ε. This is consistent, but the intermediate display equation showing the bound on ∂_ε u^ε (currently '≤ C_1 + C_2/√ε') should be clearly stated before the integration step.
Simulated Author's Rebuttal
We thank the referee for the careful reading and the constructive assessment. The referee raises one major concern regarding the uniform L^∞ bound on the adjoint state ρ^ε in Step 2 of the proof of Theorem 4.6, which we address below.
read point-by-point responses
-
Referee: The referee identifies a load-bearing gap in Step 2 (Eq. 4.2) of the proof of Theorem 4.6: the uniform L^∞ bound on the adjoint state ρ^ε is not established. The adjoint solves a Fokker-Planck equation with vanishing diffusion ε, and semiconcavity only controls ∆u^ε from above, not |∆u^ε|, so a uniform L^∞ bound on ρ^ε is not justified. If ∥ρ^ε∥_{L^∞} grows as ε→0, the L^1 estimate (4.2) does not close. The referee notes this does not affect Step 1 (4.1), where L^1 conservation suffices.
Authors: The referee is correct that the uniform L^∞ bound on ρ^ε is not currently justified in the manuscript, and we acknowledge this gap in Step 2. We have carefully re-examined the argument and agree that the semiconcavity estimate from Lemma 4.1 provides only an upper bound on ∆u^ε, which is insufficient to control div(D_pH(x,Du^ε)) and hence insufficient to establish a uniform L^∞ bound on ρ^ε via standard maximum principle or Harnack-type arguments for the adjoint Fokker-Planck equation with vanishing diffusion. We will revise the proof of (4.2) to eliminate the dependence on ∥ρ^ε∥_{L^∞}. The key observation is that the duality argument in Step 2 can be restructured to use an L^1-adjoint rather than an L^∞-adjoint. Specifically, instead of taking ρ(τ) = sgn(∂_ε u^ε(τ)) ∈ L^∞ and bounding ∫|∂_ε u^ε| dx by duality against ρ^ε with the L^∞ norm of ρ^ε appearing, one can work directly with the L^1 norm of the adjoint state (which is conserved and equals 1) combined with the L^1 bound on ∆u^ε from Lemma 4.1 and the L^1 bound on ∂_ε F[m^ε] from Corollary 4.4. Concretely, the term ∫∫ ∆u^ε ρ^ε dxdt is bounded by ∥∆u^ε∥_{L^1} · ∥ρ^ε∥_{L^∞} in the current draft, but can alternatively be controlled by ∥∆u^ε∥_{L^1} · ∥ρ^ε∥_{L^1} = ∥∆u^ε∥_{L^1} when the adjoint is initialized as a probability measure (as in Step 1), using the mass conservation property of the Fokker-Planck equation. This mirrors the structure already used successfully in Step 1, where the referee correctly notes the argument closes because it relies on L^1 conservation rather than L^∞ bounds. We will rewrite Step 2 to make this explicit, removing the unjustified appeal to ∥ρ^ε∥_{L^∞}. We note that this modification does not affect the final rate O(√ε) for (4.2), nor does it impact the subsequent estimates (4.3) and (4 revision: yes
Circularity Check
No significant circularity found; derivation chain is self-contained with standard PDE estimates
full rationale
The paper's derivation chain proceeds as follows: (1) Lemma 4.1 establishes semiconcavity, L^1 bounds on Δu^ε, and Lipschitz bounds via known results cited to [CG19, CGM23, CP20, Lio82] — these are standard PDE estimates with alternative external references provided, not self-referential definitions. (2) Theorem 4.2 derives the monotonicity integral bound ∫(F[m^ε]-F[m])(m^ε-m) ≤ Cε by testing the difference equations against -μ and w, using uniform convexity (H4), the L^∞ bound on m^ε (maximum principle, citing [Eva10b, LBL19]), and Lemma 4.1. This is a genuine derivation from structural assumptions, not a renaming. (3) Theorem 4.6 Step 1 (eq. 4.1) uses the adjoint method of Evans [Eva10a]: the equation for ∂_εu^ε is obtained by differentiating the HJB equation (citing Lemma 2.1 in [CG25], which is a straightforward differentiation), and the Bernstein method controls the D²u^ε term via duality with ρ^ε. The rate O(√ε) follows from |∂_εu^ε| ≤ C√ε integrated over ε ∈ [0,ε], yielding O(ε^{3/2}) ⊂ O(√ε) — the claim is weaker than what the estimate gives, not forced. (4) Equation (4.3) uses the mean value theorem for D_pH, the interpolation inequality ∥Du^ε-Du∥_{L^2}² ≤ ∥Δ(u^ε-u)∥_{L^1}∥u^ε-u∥_{L^∞}, and (4.1)/(4.2) — standard functional analysis, not circular. (5) Equation (4.5) applies Corollary 3.2 (a self-contained linear FP estimate proved via Itô's formula and Gronwall) with ∥b^ε-b∥_{L^1} ≤ Cε^{1/4} from (4.3), yielding W_2 ≤ C(√ε + ε^{1/8}) = O(ε^{1/8}). Each step uses outputs of prior steps as inputs, which is normal mathematical reasoning. The self-citations ([CG25], [CG19], [CGM23]) are for standard semiconcavity/differentiation results with alternative references ([CP20], [Lio82]) provided. The skeptic's concern about the unestablished ∥ρ^ε∥_{L^∞} bound in Step 2 of (4.2) is a correctness gap, not a circularity issue — ρ^ε is defined by the adjoint equation (4.6), not in terms of the quantity being estimated. No step reduces to its inputs by construction.
Assumptions & free parameters
assumptions (4)
- domain assumption Uniform convexity of H in the gradient variable: θ Id ≤ D²_pp H(x,p) ≤ Θ Id (H4)
- domain assumption Regularizing nonlocal coupling F maps C([0,T];P(T^d)) to C(Q_T) with range in bounded L^∞_t(W^{2,∞}_x) (F)
- domain assumption Lasry-Lions monotonicity conditions (1.5)-(1.6)
- ad hoc to paper Additional W^{2,∞} bound on u^ε independent of ε (4.4)
Cite this review
Pith. "Pith review of On the rate of the vanishing viscosity approximation for Mean Field Games with nonlocal coupling." pith.science (2026). https://pith.science/paper/LJ6UAZWY
@misc{pith2026260707526,
author = {Pith},
title = {Pith review of: On the rate of the vanishing viscosity approximation for Mean Field Games with nonlocal coupling},
year = {2026},
howpublished = {\url{https://pith.science/paper/LJ6UAZWY}},
note = {Machine review of arXiv:2607.07526}
}
read the original abstract
We study quantitative convergence rates of the vanishing viscosity approximation of first-order time-dependent Mean Field Games with regularizing coupling acting in the Hamilton-Jacobi equation. Under standard structural assumptions on the Hamiltonian ensuring convergence of the vanishing viscosity approximation, previous results provide either qualitative convergence for the two unknown of the system or quantitative estimates merely for solutions of the Hamilton-Jacobi equation. In this work, we improve these convergence rates under the same assumptions and establish, in addition, quantitative estimates for the convergence of the associated forward Fokker-Planck equation. As a consequence, we obtain quantitative convergence rates for the full Mean Field Game system, thus extending and strengthening the existing theory for the first-order limit.
Reference graph
Works this paper leans on
-
[1]
S. Bianchini and A. Bressan. Vanishing viscosity solutions of nonlinear hyperbolic systems. Ann. Math. (2) , 161(1):223--342, 2005
work page 2005
-
[2]
P. Bonicatto, G. Ciampa, and G. Crippa. On the advection-diffusion equation with rough coefficients: weak solutions and vanishing viscosity. J. Math. Pures Appl. (9) , 167:204--224, 2022
work page 2022
-
[3]
E. Bru \'e and Q.-H. Nguyen. Advection diffusion equations with Sobolev velocity field. Commun. Math. Phys. , 383(1):465--487, 2021
work page 2021
-
[4]
P. Cannarsa and R. Capuani. Existence and uniqueness for mean field games with state constraints. In PDE models for multi-agent phenomena. Selected papers based on the presentations at the workshop, Rome, Italy, November 28 -- December 2, 2016 , pages 49--71. Cham: Springer, 2018
work page 2016
-
[5]
M. Colombo, G. Crippa, and M. Sorella. Anomalous dissipation and lack of selection in the Obukhov - Corrsin theory of scalar turbulence. Ann. PDE , 9(2):48, 2023. Id/No 21
work page 2023
-
[6]
L.-P. Chaintron and S. Daudin. Optimal rate of convergence in the vanishing viscosity for uniformly convex Hamilton - Jacobi equations. Preprint, arXiv :2506.13255, 2025
work page Pith review arXiv 2025
-
[7]
M. Cirant and A. Goffi. On the existence and uniqueness of solutions to time-dependent fractional MFG . SIAM J. Math. Anal. , 51(2):913--954, 2019
work page 2019
-
[8]
M. Cirant and A. Goffi. Maximal L^q -regularity for parabolic H amilton- J acobi equations and applications to M ean F ield G ames. Ann. PDE , 7(2):Paper No. 19, 40, 2021
work page 2021
Show all 27 references
-
[9]
Cirant and A
M. Cirant and A. Goffi. Convergence rates for the vanishing viscosity approximation of Hamilton - Jacobi equations: the convex case. Preprint, arXiv :2502.15495, to appear in Indiana Univ. Math. J., 2025
2025 arXiv
-
[10]
Cecchin and A
A. Cecchin and A. Goffi. Convergence rates for the vanishing viscosity approximation of fully nonlinear, non-convex, second-order Hamilton - Jacobi equations via nonlinear convolutions. Calc. Var. Partial Differ. Equ. , 65(5):11, 2026. Id/No 174
2026
-
[11]
Camilli, A
F. Camilli, A. Goffi, and C. Mendico. Quantitative and qualitative properties for Hamilton - J acobi PDE s via the nonlinear adjoint method, to appear in Ann. Sc. Norm. Super. Pisa, Cl. Sci., 2023
2023
-
[12]
Cirant and A
M. Cirant and A. R. M \'e sz \'a ros. Long Time Behavior and Stabilization for Displacement Monotone Mean Field Games . Preprint, arXiv :2412.14903, 2025
2025
-
[13]
Cardaliaguet and A
P. Cardaliaguet and A. Porretta. An introduction to mean field game theory. In Mean field games. Cetraro, Italy, June 10--14, 2019. Lecture notes given at the summer school , pages 1--158. Cham: Springer; Florence: Fondazione CIME, 2020
2019
-
[14]
L. C. Evans. Adjoint and compensated compactness methods for H amilton- J acobi PDE . Arch. Ration. Mech. Anal. , 197(3):1053--1088, 2010
2010
-
[15]
L. C. Evans. Partial differential equations , volume 19 of Graduate Studies in Mathematics . American Mathematical Society, Providence, RI, second edition, 2010
2010
-
[16]
Jameson Graber and A
P. Jameson Graber and A. R. M \'e sz \'a ros. On monotonicity conditions for mean field games. J. Funct. Anal. , 285(9):45, 2023. Id/No 110095
2023
-
[17]
Gangbo and A
W. Gangbo and A. Tudorascu. Weak KAM theory on the Wasserstein torus with multidimensional underlying space. Commun. Pure Appl. Math. , 67(3):408--463, 2014
2014
-
[18]
Huang, R
M. Huang, R. P. Malham\'e, and P. E. Caines. Large population stochastic dynamic games: closed-loop M c K ean- V lasov systems and the N ash certainty equivalence principle. Commun. Inf. Syst. , 6(3):221--251, 2006
2006
-
[19]
H. Kunita. Stochastic differential equations and stochastic flows of diffeomorphisms. Ecole d' \'e t \'e de probabilit \'e s de Saint - Flour XII - 1982, Lect . Notes Math . 1097, 143-303 (1984)., 1984
1982
-
[20]
Le Bris and P.-L
C. Le Bris and P.-L. Lions. Parabolic equations with irregular data and related issues---applications to stochastic differential equations , volume 4 of De Gruyter Series in Applied and Numerical Mathematics . De Gruyter, Berlin, 2019
2019
-
[21]
P.-L. Lions. Generalized solutions of H amilton- J acobi equations , volume 69 of Research Notes in Mathematics . Pitman (Advanced Publishing Program), Boston, Mass.-London, 1982
1982
-
[22]
Lasry and P.-L
J.-M. Lasry and P.-L. Lions. Mean field games. Jpn. J. Math. , 2(1):229--260, 2007
2007
-
[23]
Li and D
H. Li and D. Luo. Quantitative stability estimates for Fokker - Planck equations. J. Math. Pures Appl. (9) , 122:125--163, 2019
2019
-
[24]
Navarro-Fern\' a ndez, A
V. Navarro-Fern\' a ndez, A. Schlichting, and C. Seis. Optimal stability estimates and a new uniqueness result for advection-diffusion equations. Pure Appl. Anal. , 4(3):571--596, 2022
2022
-
[25]
Tang and Y
W. Tang and Y. P. Zhang. The convergence rate of vanishing viscosity approximations for mean field games. SIAM J. Math. Anal. , 57(3):3217--3254, 2025
2025
-
[26]
C. Villani. Optimal transport. Old and new , volume 338 of Grundlehren Math. Wiss. Berlin: Springer, 2009
2009
-
[27]
W. Yu, Q. Du, and W. Tang. Beyond separability: convergence rate of vanishing viscosity approximations to mean field games via FBSDE stability. Preprint, arXiv :2505.18529, 2025
2025 arXiv
Reviewed July 9, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.