Extragradient methods for mean field games of controls and mean field type FBSDEs

Charles Meynard (LJAD)

REVIEW 3 major objections 5 minor 1 cited by

Extragradient methods for mean field games of controls and mean field type FBSDEs

T0 review · 3 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read This paper shows that monotone mean-field FBSDEs—including mean-field games of controls—can be solved by an extragradient iteration on a Hilbert space of controls, with exponential convergence under sufficiently strong monotonicity.

desk verdict A clean Hilbertian-inverse reformulation and a solid averaged convergence theorem, but the exponential last-iterate result has an algebra gap and the wellposedness of the target system is borrowed from an unpublished preprint. read the letter →

arxiv 2602.14621 v3 pith:QN6NKZK6 submitted 2026-02-16 math.OC

Charles Meynard (LJAD) This is my paper · ORCID

classification math.OC MSC 49N8060H1065K1535Q89

keywords mean-fieldgamesofcontrolsFBSDEsextragradientmethodmonotonevariationalinequalitydisplacementmonotonicitycommonnoisefictitiousplaynumericalscheme

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that solutions of monotone mean-field forward-backward stochastic differential equations, including mean-field games of controls, can be represented as the unique zero of a monotone operator on a Hilbert space of square-integrable adapted control processes. It then proves that an explicit extragradient iteration on that space converges: the averaged iterates achieve a 1/n rate, and under stronger monotonicity the last iterate converges exponentially fast. The method is purely probabilistic and extends without rate degradation to mean-field FBSDEs with common noise. A sympathetic reader would care because this gives a simulation-based numerical scheme with provable convergence rates for a class of coupled mean-field problems whose wellposedness rests on monotonicity rather than short time horizons.

What carries the argument

The carrying object is the monotone variational inequality on the Hilbert space H_T = (L²-adapted controls, E∫_0^T |·|² dt). The paper constructs a monotone operator v whose unique zero is exactly the mean-field game equilibrium or FBSDE solution, proves v is Lipschitz and c_L-strongly monotone under displacement or L2 monotonicity, and then applies the classical extragradient iteration with two evaluations per step: a look-ahead half-step α_{n+1/2} = α_n − γ v(α_n) and a full step α_{n+1} = α_n − γ v(α_{n+1/2}). The engine of the proof is a Hilbert-space version of a standard extragradient inequality that converts strong monotonicity of v into a contraction estimate, yielding either average

What would settle it

Exhibit a Lipschitz, L2-monotone mean-field FBSDE satisfying Hypotheses 2.13–2.14 whose decoupling field blows up before the horizon T; then the existence premise behind α* fails and the convergence theorems in Section 2 have no target. A simpler check: run the finite-dimensional truncation of the algorithm with the exact oracle v on a strongly monotone instance with γ below the stated threshold and test whether the last-iterate error decreases geometrically; a violation would indicate the contraction constant in Theorem 2.11 is miscomputed.

Watch

Extended reading notes

Core claim

The central discovery is that the equilibrium control of a displacement-monotone mean-field game of controls, and more generally the solution of an L2-monotone mean-field FBSDE, can be characterized as the unique zero of an explicitly constructed monotone Lipschitz operator v on the space H_T of adapted square-integrable controls. For mean-field games, v(α) = ∇αL(X^α, α, L(X^α, α)) − U^α, where U^α is the backward value process generated by α; for general monotone FBSDEs, v(α) = F_u^{-1}(X^α, α) − U^α. The extragradient update α_{n+1} = α_n − γ v(α_n − γ v(α_n)) is explicit at each step, avoiding the implicit fixed-point solve of a naive proximal method. Theorems 2.9 and 2.19 show that avera

Load-bearing premise

The convergence results assume the target FBSDE already has a unique strong solution with a Lipschitz decoupling field over the whole interval [0,T]; the manuscript supplies an a priori estimate and cites prior work for that existence, so if that wellposedness fails, the limit α* that the iterates approach is undefined.

Editorial extensions

If this is right

For displacement-monotone mean-field games of controls, the equilibrium can be computed by simulating forward paths for a candidate control, solving a decoupled backward SDE for the value process, and updating the control with the extragradient rule—no PDE discretization is required.
In the strongly monotone regime, reaching precision ε needs O(log(1/ε)) iterations in the deterministic-oracle setting, so the per-iteration Monte Carlo cost dominates and the iteration count does not grow with state dimension.
Averaged iterates converge at rate O(1/n) under only Lipschitz continuity and monotonicity, providing a fallback when the extra strong-monotonicity condition for exponential convergence is absent.
Adding a common noise leaves the convergence rates unchanged; it only enlarges simulation cost because conditional laws must be propagated along each common-noise trajectory.
In semi-monotone short-horizon settings, a decreasing-step variant converges at a slower algebraic rate and removes the need to know the Lipschitz norm of v in advance.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The control-space formulation makes this algorithm a probabilistic counterpart of fictitious play: both are monotone variational inequalities, so convergence insights may transfer between the learning and numerical communities; the paper only sketches this link.
For general FBSDEs the practical bottleneck is the Hilbertian inverse F_u^{-1}; the method is directly implementable mainly when F does not depend on the law of the backward process or when the inverse is explicit—a limitation the author acknowledges in Remark 2.20.
The author leaves open in Remark 2.12 whether exponential last-iterate convergence holds under monotonicity in the forward variable X; a natural testable extension is to run the same algorithm on strongly X-monotone examples and look for a contraction constant numerically.
If the wellposedness of the target FBSDE with a Lipschitz decoupling field is established independently and self-containedly, the conditional convergence theorems become unconditional for the full L2-monotone class.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Here's my read of Meynard's paper. The genuinely new piece is the explicit Hilbertian-inverse reformulation of mean-field FBSDEs: writing the equilibrium as a zero of a monotone operator on the control space, with the backward process defined through the inverse of the forward driver. That is a clean idea and it does make the solver explicit in a way standard formulations aren't. The averaged O(1/n) convergence theorem (Thm 2.9 and its FBSDE analog) follows the standard variational inequality argument and holds up. The common-noise extension is natural, and the numerical section uses an explicit solution, which is a good sanity check.

The soft spots are real. First, the exponential last-iterate theorem (Thm 2.11) does not close. The displayed contraction factor is 1−2γc_L+4γ²c_L‖v‖²_Lip, and under the stated bound γ < min(1/(2‖v‖_Lip), c_L/‖v‖²_Lip) there are parameter regimes where that factor exceeds 1. The proof also goes through a factor with ‖v‖_Lip without the square in one step, so there is an algebra slip. The rate statement is not supported by the displayed inequalities.

Second, the wellposedness of the target FBSDE (2.8) under L2-monotonicity is not actually proved in the paper. Theorem 2.15 is imported from the author's own preprint [39], and Appendix A.1 only establishes an a priori estimate and then asserts, "it follows naturally from ideas introduced in [14,39]" that a Lipschitz decoupling field exists. No continuation argument is given. So the α* the convergence theorems talk about is only defined conditional on an external result. If that result is correct, the rest stands; but the paper should either prove it or state the main theorems as conditional on [39].

The numerical section is illustrative rather than empirical: no error bars, no baseline comparison, no code. That is minor relative to the two issues above.

Overall, the paper deserves a serious referee, but it needs major revision. The author is honest about limits; Remarks 2.20 and 2.21 acknowledge the inverse is nontrivial and the exponential result may not extend to other monotonicity regimes. I'd send it to review, with instructions to fix the contraction-factor step and make the wellposedness dependence transparent. I wouldn't cite the exponential claim until it's repaired.

Referee Report

3 major / 5 minor

Summary. The paper proposes an extragradient-type algorithm for monotone mean-field FBSDEs and mean-field games of controls. It recasts the equilibrium as a zero of a monotone operator on the Hilbert space of controls, introduces a two-step iteration, and proves O(1/n) convergence of averaged iterates under Lipschitz/strong-monotonicity assumptions (Theorems 2.9 and 2.19). It further claims exponential last-iterate convergence in the strongly monotone case (Theorem 2.11), extends the scheme to FBSDEs with common noise, and reports numerical tests on an explicitly solvable example. The abstract's headline claim is exponential convergence, but as written that theorem is not proved; the wellposedness of the target FBSDE is also imported from a same-author preprint rather than established in this manuscript.

Significance. If fully established, the paper would be a useful bridge between variational-inequality optimization and probabilistic numerics for mean-field FBSDEs. The averaged-convergence argument follows a standard VI template and is credible; the explicit dependence on the monotonicity constant, the treatment of common noise, and the numerical validation are assets. However, the main novelty claimed in the abstract—exponential convergence of the last iterate—is not supported by the displayed proof, and the target solution is not shown to exist inside the manuscript. These are load-bearing gaps, so the contribution as it stands is conditional. The numerical section is reassuring but secondary and also contains sign/index inconsistencies.

major comments (3)

[§2.2, Theorem 2.15, Remark 2.21, Lemma A.1] The convergence theorems are statements about the unique strong solution of (2.8), but that object is not constructed in the manuscript under the assumptions used. Theorem 2.15's proof cites [39, Lem. 3.21 & 3.9], and Lemma A.1 derives only a priori estimates for differences of hypothetical solutions; its final sentence asserts the existence of a Lipschitz decoupling field 'from ideas introduced in [14,39]' without giving the continuation argument. A uniqueness/a priori bound does not by itself produce a global strong solution for non-gradient L2-monotone FBSDEs on arbitrary T, and the cited [39] is a preprint by the same author. Since α* in Theorems 2.9, 2.11, and 2.19 is defined via (2.8), the main convergence results are conditional on an unproved wellposedness statement. The same gap appears in Lemma 2.5 for the MFG FBSDE (2.4).
[§2.1.2, proof of Theorem 2.11] The claimed exponential rate is not established. The transition after the inequality involving -γ²||v(α_{i+1/2})||² - 2γ³c_L||v(α_i)||² + 2γ³||v||_Lip||v(α_i)||||v(α_{i+1/2})|| is not justified: the three terms are not bounded by γ²(γ||v||²_Lip/c_L - 1)||v(α_{i+1/2})||² under the stated hypotheses. The cross term is positive and the negative quadratic in ||v(α_i)|| does not control it with the displayed constant. In addition, the final rate uses (1 - 2γc_L + 4γ²c_L||v||²_Lip)^n, but the condition γ < min(1/(2||v||_Lip), c_L/||v||²_Lip) does not make this factor lie in (0,1): for example, L=2, c_L=1.9, and γ=0.2 satisfy the condition but give the displayed factor 1.456. Thus the theorem is unproved and the headline exponential convergence is not supported. The proof also switches between ||v||_Lip and ||v||²_Lip in the displayed rate, which needs clarification.
[§3, Lemma 3.3 and Corollary 3.4] The common-noise extension inherits the same wellposedness gap. Lemma 3.3 refers to [39, Thm. 3.33 and §3.4.1] for existence of a unique strong solution with Lipschitz decoupling field. Corollary 3.4 then proves convergence of the algorithm to that solution. The convergence argument is a reasonable adaptation of the no-common-noise case, but the target object is again not proved to exist within the manuscript. If the preprint [39] is not accepted as an independent reference, the common-noise convergence statement is conditional in the same way as the base-case theorems.

minor comments (5)

[General] There are numerous typos and infelicities: 'genrality', 'litterarure', 'remind' for 'recall', and repeated or misformatted references. Reference [22] and [23] appear to be the same paper. A careful copyedit is needed.
[Theorem 2.11] The statement uses α^0 while the algorithm in Theorem 2.9 initializes at α^1; the proof also shifts indices. The indexing should be unified and the statement should state whether the rate holds for n≥0 or n≥1.
[Final rate in proof of Theorem 2.11] The contraction factor is written inconsistently: earlier in the proof it is 1 - 2γc_L + 4γ²c_L||v||_Lip, while the final displayed rate uses 1 - 2γc_L + 4γ²c_L||v||²_Lip. This is more than a typo because the admissibility conditions are different in the two cases.
[Section 4, Algorithm 1] The discrete forward dynamics in Algorithm 1 read X^{i,j} = X^{i,j-1} + Δt α^{i,j}, whereas the continuous parametrization in (2.10) is X_t = X_0 - ∫_0^t α_s ds. The sign convention should be clarified or corrected; as written the discrete scheme does not match the continuous equation used in the convergence theorems.
[Section 4.2.1] The numerical example takes c=0 and f(x)=atan(x-1), and states that this lies in the monotone regime. This is not immediate from Hypothesis 2.14, especially because the forward driver is -aU and the monotonicity condition couples F and G. The authors should verify and state precisely why the assumptions of the theorems are satisfied for this example.

Circularity Check

1 steps flagged · score 4.0 of 10

Convergence target α* depends on existence of a unique solution of (2.8), which is imported from the author's own preprint [39] rather than proved in this paper.

self citation load bearing [Theorem 2.15 (p. 12) and Lemma A.1 (Appendix, p. 29)]
"The existence of a strong solution is a direct consequence of [39] Lemma 3.21 combined with Lemma 3.9. ... Since this a priori estimate is valid for any initial condition and on any time interval, it follows naturally from ideas introduced in [14, 39] that the FBSDE (A.1) admits a Lipschitz decoupling field."

The convergence theorems (2.9, 2.11, 2.19) all quantify convergence to α*, which is defined as the unique strong solution of the FBSDE (2.8). Existence of that object under Hypothesis 2.14 is exactly the content of Theorem 2.15, whose proof is only a pointer to the author's own preprint [39]. The self-contained argument in Lemma A.1 establishes merely an a priori estimate for differences of two hypothetical solutions and then asserts, rather than derives, that a Lipschitz decoupling field exists. Thus the limiting object of the central claim is supplied by a self-citation; if [39] is not accepted as independent, α* is undefined and the exponential-convergence statements are vacuous. This is load-bearing self-citation, though not a fit of the predicted quantity.

full rationale

The extragradient convergence analysis itself is not circular: Theorems 2.9 and 2.11 prove contraction / O(1/n) rates for a Lipschitz strongly monotone operator v, using only the generic extragradient inequality (Lemma 1.4) and the strong-monotonicity identity for v(α*)=0. The numerical section tests the scheme against an explicitly constructed solution of a linear-quadratic FBSDE, not against data used to fit parameters. The only serious circularity concern is the wellposedness prerequisite: the target α* is the unique solution of (2.8), and existence is imported from the author's own preprint [39], with Appendix A.1 giving only an a priori estimate and a one-sentence appeal to 'ideas introduced in [14,39]'. This makes the central result conditional on a self-citation chain, but the rate derivation has independent mathematical content. Hence score 4 rather than a higher score: no prediction reduces by construction, and the convergence argument would be valid if the imported wellposedness theorem were independently established.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No parameters are fitted to data in the core proofs; the only tunable constants are user-chosen step sizes and discretization settings. The central claim depends on strong-monotonicity hypotheses, on an external/self-cited wellposedness result, and on the availability of the Hilbertian inverse F_u^{-1}. No new physical or ontological entities are proposed.

free parameters (4)

step size γ = γ = 0.08 in Section 4.2.2; otherwise constrained by γ ≤ 1/||v||_Lip in Theorem 2.9 or γ < min(1/(2||v||_Lip), c_L/||v||²
Convergence rates and numerical behavior depend quantitatively on the step-size choice; no adaptive step-size rule is supplied.
Monte Carlo sample size Np = Np = 10000
Used in the numerical illustration; regression and error-estimator accuracy depend on it.
time steps Nt = Nt = 100
Used in the numerical illustration; temporal discretization error is separate from the algorithmic convergence rate.
Hermite polynomial truncation = first 10 Hermite polynomials
Chosen regression basis for conditional expectations; no selection criterion is given.

assumptions (5)

ad hoc to paper Existence of a unique strong solution and Lipschitz decoupling field for L2-monotone FBSDE (2.8) on arbitrary time interval
Appendix A.1 proves an a priori estimate then asserts existence from 'ideas introduced in [14,39]'; this is load-bearing for all convergence theorems.
domain assumption Hypothesis 2.14: F, G, g are Lipschitz; g is displacement monotone; (F,G) is c_F-strongly monotone in U
Defines the regime of the method and is used to prove monotonicity of the variational-inequality operator v.
domain assumption Hypothesis 2.2: displacement monotonicity of ∇g and strong monotonicity of (∇xL, ∇αL) in the control with constant c_L
Used for the MFG-of-controls reformulation and for exponential convergence of the last iterate.
standard math Known wellposedness of decoupled Lipschitz FBSDEs (Zhang [46], Theorem 4.3.1)
Invoked to define U^α and Z^α for a given control α in (2.7), (2.10), and (3.3).
standard math Proposition 1.2 / Corollary 1.3 (adapted from [43]): Hilbertian inequality implies pointwise inequality
Used to derive convexity and strong convexity statements from Hilbertian monotonicity assumptions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Extragradient methods for mean field games of controls and mean field type FBSDEs." pith.science (2026). https://pith.science/paper/QN6NKZK6

@misc{pith2026260214621,
  author       = {Pith},
  title        = {Pith review of: Extragradient methods for mean field games of controls and mean field type FBSDEs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QN6NKZK6}},
  note         = {Machine review of arXiv:2602.14621}
}

read the original abstract

In this paper we present a numerical scheme to solve coupled mean field forward-backward stochastic differential equations driven by monotone vector fields. This is based on an adaptation of so called extragradient methods by characterizing solutions as zeros of monotone variational inequalities in a Hilbert space. We first introduce the procedure in the context of mean field games of controls and highlight its connection to the fictitious play. Under sufficiently strong monotonicity assumptions, we demonstrate that the sequence of approximate solutions converges exponentially fast. Then we extend the method and main results to general forward backward systems of stochastic differential equations that do not necessarily stem from optimal control.

Figures

Figures reproduced from arXiv: 2602.14621 by the authors.

**Figure 1.** log-linear plot of the last iterate error in function of the number of iterations After a small number of iterations, the linear fit is almost perfect. This emphasizes that the last iterates converge exponentially fast when the conditions for applying Theorem 2.11 are met. Let us emphasize that the number of iterations may appear large only because we considered a time interval of [0, 10] and even then, computation… view at source ↗

**Figure 3.** θ(t) 0 2 4 6 8 10 -2.5 -2 -1.5 -1 -0.5 0 Time Plot of the functions t 7→ θ(t), t 7→ η(t) in blue and of the computed values using our algorithm with the red crosses. Since εerror is very small (∼ 10−13) the difference between the true curve and the computed values does not come from the number of iterations but is rather inherent to the discretization (both in time and on the space of random paths) of the problem. 5… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Kullback-Leibler Mirror-Prox for Measure-Valued Variational Inequalities and Mean-Field Equilibria
math.OC 2026-08 conditional novelty 6.0 of 10

A relative-entropy extragradient algorithm provably computes static mean field equilibria, with explicit trade-offs between grid fineness and iteration count.

Reference graph

Works this paper leans on

45 extracted references · cited by 1 Pith paper

[39]

C. Meynard. Monotone solutions to mean ﬁeld games maste r equation in the l2-monotone setting, 2025

2025
[1]

Mean ﬁeld games: Convergence of a ﬁnite diﬀerence method

Yves Achdou, Fabio Camilli, and Italo Capuzzo-Dolcetta . Mean ﬁeld games: Convergence of a ﬁnite diﬀerence method. SIAM Journal on Numerical Analysis , 51(5):2585–2612, 2013

2013
[2]

Mean ﬁeld games of control s: Finite diﬀerence approximations

Yves Achdou and Ziad Kobeissi. Mean ﬁeld games of control s: Finite diﬀerence approximations. Mathematics in Engineering, 3(3):1–35, 2021

2021
[3]

Convergence of a ﬁnite diﬀerence scheme to weak solutions of the system of partial diﬀerential equations arising in mean ﬁeld games

Yves Achdou and Alessio Porretta. Convergence of a ﬁnite diﬀerence scheme to weak solutions of the system of partial diﬀerential equations arising in mean ﬁeld games. SIAM Journal on Numerical Analysis , 54(1):161–186, 2016

2016
[4]

Graves, Houzhi Li, Jean-Fran çois Chassagneux, Franccois Delarue, and René A

Andrea Angiuli, Christy V. Graves, Houzhi Li, Jean-Fran çois Chassagneux, Franccois Delarue, and René A. Car- mona. Cemracs 2017: numerical probabilistic approach to mf g. ESAIM: Proceedings and Surveys , 2018

2017
[5]

Adaptive extra-gradient methods for min-max optimization and games

Kimon Antonakopoulos, Elena Veronica Belmega, and Pana yotis Mertikopoulos. Adaptive extra-gradient methods for min-max optimization and games. In International Conference on Learning Representations , 2021

2021
[6]

Sifting through the noise: Universal ﬁrst-order methods for stochastic var iational inequalities

Kimon Antonakopoulos, Thomas Pethick, Ali Kavis, Panay otis Mertikopoulos, and Volkan Cevher. Sifting through the noise: Universal ﬁrst-order methods for stochastic var iational inequalities. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. W ortman Vaughan, editors, Advances in Neural Information Processing Systems , volume 34, pages 13099–1...
[7]

A tight and uniﬁed analysis of gradient-based methods for a whole spectrum of diﬀerentiab le games

W aïss Azizian, Ioannis Mitliagkas, Simon Lacoste-Juli en, and Gauthier Gidel. A tight and uniﬁed analysis of gradient-based methods for a whole spectrum of diﬀerentiab le games. In Silvia Chiappa and Roberto Calandra, editors, Proceedings of the Twenty Third International Conference o n Artiﬁcial Intelligence and Statistics , volume 108 of Proceedings of...

2020

Show all 45 references

[8]

A universal algorithm for var iational inequalities adaptive to smoothness and noise

Francis Bach and Kﬁr Y Levy. A universal algorithm for var iational inequalities adaptive to smoothness and noise. In Alina Beygelzimer and Daniel Hsu, editors, Proceedings of the Thirty-Second Conference on Learning Th eory, volume 99 of Proceedings of Machine Learning Resear...

2019
[9]

Bensoussan, S.C.P

A. Bensoussan, S.C.P. Yam, and Z. Zhang. W ell-posedness of mean-ﬁeld type forward–backward stochastic diﬀer- ential equations. Stochastic Processes and their Applications , 125(9):3327–3354, 2015

2015
[10]

Mean Field Games and Mean Field Type Control Theory

Alain Bensoussan, Jens Frehse, and Philip Yam. Mean Field Games and Mean Field Type Control Theory . Springer, New York, 2013

2013
[11]

Bertucci and C

C. Bertucci and C. Meynard. Noise through on additionna l variable for mean ﬁeld games master equation on ﬁnite state space, 2024

2024
[12]

Bertucci and C

C. Bertucci and C. Meynard. A study of common noise in mea n ﬁeld games, 2024

2024
[13]

Monotone solutions for mean ﬁeld gam es master equations: continuous state space and common noise

Charles Bertucci. Monotone solutions for mean ﬁeld gam es master equations: continuous state space and common noise. Communications in Partial Diﬀerential Equations , 48(10-12):1245–1285, 2023

2023
[14]

On lipschitz solutions of mean ﬁeld games master equations

Charles Bertucci, Jean-Michel Lasry, and Pierre-Loui s Lions. On lipschitz solutions of mean ﬁeld games master equations. Journal of Functional Analysis , 287(5):110486, 2024

2024
[15]

Chapter 15 - an algorithmic guide fo r ﬁnite-dimensional optimal control problems

Jean-Baptiste Caillau, Roberto Ferretti, Emmanuel Tr élat, and Hasnaa Zidani. Chapter 15 - an algorithmic guide fo r ﬁnite-dimensional optimal control problems. In Emmanuel T rélat and Enrique Zuazua, editors, Numerical Control: Part B , volume 24 of Handbook of Numerical Anal...

2023
[16]

Cardaliaguet, F

P. Cardaliaguet, F. Delarue, J.-M. Lasry, and P.-L. Lio ns. The master equation and the convergence problem in mean ﬁeld games. Princeton University Press , 2019

2019
[17]

Monot one solutions of the master equation for mean ﬁeld games with idiosyncratic noise

Pierre Cardaliaguet and Panagiotis Souganidis. Monot one solutions of the master equation for mean ﬁeld games with idiosyncratic noise. SIAM Journal on Mathematical Analysis , 54(4):4198–4237, 2022

2022
[18]

Learnin g in mean ﬁeld games: The ﬁctitious play

Cardaliaguet, Pierre and Hadikhanloo, Saeed. Learnin g in mean ﬁeld games: The ﬁctitious play. ESAIM: COCV , 23(2):569–591, 2017. 28 MEYNARD CHARLES

2017
[19]

Mean ﬁeld forward-b ackward stochastic diﬀerential equations

René Carmona and François Delarue. Mean ﬁeld forward-b ackward stochastic diﬀerential equations. Electronic Communications in Probability , 18:1 – 15, 2013

2013
[20]

Probabilistic Theory of Mean Field Games with Applications II

René Carmona and François Delarue. Probabilistic Theory of Mean Field Games with Applications II. Springer, 2018

2018
[21]

Mea n ﬁeld games with common noise

René Carmona, François Delarue, and Daniel Lacker. Mea n ﬁeld games with common noise. The Annals of Proba- bility, 44(6):3740 – 3803, 2016

2016
[23]

Convergence analys is of machine learning algorithms for the numerical solu- tion of mean ﬁeld control and games: Ii–the ﬁnite horizon cas e

René Carmona and Mathieu Laurière. Convergence analys is of machine learning algorithms for the numerical solu- tion of mean ﬁeld control and games: Ii–the ﬁnite horizon cas e. The Annals of Applied Probability , 32(6):4065–4105, 2022

2022
[24]

Numerical method for FBSDEs of McKean–Vlasov type

Jean-François Chassagneux, Dan Crisan, and François D elarue. Numerical method for FBSDEs of McKean–Vlasov type. The Annals of Applied Probability , 29(3):1640–1684, 2019

2019
[25]

Crandall, Hitoshi Ishii, and Pierre-Louis L ions

Michael G. Crandall, Hitoshi Ishii, and Pierre-Louis L ions. User’s guide to viscosity solutions of second order pa rtial diﬀerential equations. Bulletin of the American Mathematical Society , 27(1):1–67, 1992

1992
[26]

Explora tion noise for learning linear-quadratic mean ﬁeld games

François Delarue and Athanasios Vasileiadis. Explora tion noise for learning linear-quadratic mean ﬁeld games. Mathematics of Operations Research , 50(3):1762–1831, 2025

2025
[27]

Gangbo, A

W. Gangbo, A. Mészáros, C. Mou, and J. Zhang. Mean ﬁeld ga mes master equations with non-separable hamilto- nians and displacement monotonicity. Annals of Probability , 50, 2022

2022
[28]

A regression-based monte carlo method to solve back- ward stochastic diﬀerential equations

Emmanuel Gobet, Jean-Philippe Lemor, and Xavier W arin . A regression-based monte carlo method to solve back- ward stochastic diﬀerential equations. The Annals of Applied Probability , 15(3):2172–2202, 2005

2005
[29]

Stochastic extragradient: General analysis and improved rates

Eduard Gorbunov, Hugo Berard, Gauthier Gidel, and Nico las Loizou. Stochastic extragradient: General analysis and improved rates. In Gustau Camps-Valls, Francisco J. R. R uiz, and Isabel Valera, editors, Proceedings of The 25th International Conference on Artiﬁcial Intelligenc...

2022
[30]

Learning in nonatomic games, part i: Finite action spaces and population games

Saeed Hadikhanloo, Rida Laraki, Panayotis Mertikopou los, and Sylvain Sorin. Learning in nonatomic games, part i: Finite action spaces and population games. Journal of Dynamics and Games , 9(4):433–460, 2022

2022
[31]

Mészáros

Joe Jackson and Alpár R. Mészáros. Quantitative conver gence for displacement monotone mean ﬁeld games of control, 2025

2025
[32]

G. M. Korpelevich. The extragradient method for ﬁnding saddle points and other problems. Ekonomika i Matem- aticheskie Metody , 12:747–756, 1976

1976
[33]

N. V. Krylov. Lectures on Elliptic and Parabolic Equations in Hölder Spac es, volume 12 of Graduate Studies in Mathematics. American Mathematical Society, 1996

1996
[34]

Mean ﬁeld ga mes

Jean-Michel Lasry and Pierre-Louis Lions. Mean ﬁeld ga mes. Japanese Journal of Mathematics , 2(3):229–260, 2007

2007
[35]

Generalized condi tional gradient and learning in potential mean ﬁeld games

Pierre Lavigne and Laurent Pfeiﬀer. Generalized condi tional gradient and learning in potential mean ﬁeld games. Applied Mathematics & Optimization , 88(3), 2023

2023
[36]

P.-L. Lions. Lectures at collège de france, 2007-2011

2007
[37]

Lions and P

P.-L. Lions and P. Souganidis. Extended mean-ﬁeld game s. Rendiconti Lincei, 3:611–625, 2020

2020
[38]

Mészáros M

Alpár R. Mészáros M. Bansil and C. Mou. Global well-pose dness of displacement monotone degenerate mean ﬁeld games master equations, 2023

2023
[40]

W ellposedness of seco nd order master equations for mean ﬁeld games with nonsmooth data

Chenchen Mou and Jianfeng Zhang. W ellposedness of seco nd order master equations for mean ﬁeld games with nonsmooth data. Memoirs of the American Mathematical Society , 2019

2019
[41]

Mean ﬁeld games of cont rols: Propagation of monotonicities, 2022

Chenchen Mou and Jianfeng Zhang. Mean ﬁeld games of cont rols: Propagation of monotonicities, 2022

2022
[42]

Mean ﬁeld games of cont rols: Propagation of monotonicities

Chenchen Mou and Jianfeng Zhang. Mean ﬁeld games of cont rols: Propagation of monotonicities. Probability, Uncertainty and Quantitative Risk , 7(3):247–274, 2022

2022
[43]

Mészáros and C

A. Mészáros and C. Mou. Mean ﬁeld games systems under dis placement monotonicity. SIAM Journal on Mathe- matical Analysis, 56(1):529–553, 2024

2024
[44]

Dual extrapolation and its applicatio ns to solving variational inequalities and related problem s

Yurii Nesterov. Dual extrapolation and its applicatio ns to solving variational inequalities and related problem s. Mathematical Programming, 109(2):319–344, 2007

2007
[45]

Tyrrell Rockafellar and Peter R

R. Tyrrell Rockafellar and Peter R. W olenski. Convexit y in hamilton–jacobi theory i: Dynamics and duality. SIAM Journal on Control and Optimization , 39(5):1323–1350, 2000

2000
[46]

Backward Stochastic Diﬀerential Equations: From Linear to Fully Nonlinear Theory , volume 86 of Probability Theory and Stochastic Modelling

Jianfeng Zhang. Backward Stochastic Diﬀerential Equations: From Linear to Fully Nonlinear Theory , volume 86 of Probability Theory and Stochastic Modelling . Springer, 2017. A. Wellposedness of a monotone FBSDE Lemma A.1. Let (F, G, g ) : R2d×P 2(R2d)→ Rd be Lipschitz function...

2017

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [39]

C. Meynard. Monotone solutions to mean ﬁeld games maste r equation in the l2-monotone setting, 2025

2025

[2] [1]

Mean ﬁeld games: Convergence of a ﬁnite diﬀerence method

Yves Achdou, Fabio Camilli, and Italo Capuzzo-Dolcetta . Mean ﬁeld games: Convergence of a ﬁnite diﬀerence method. SIAM Journal on Numerical Analysis , 51(5):2585–2612, 2013

2013

[3] [2]

Mean ﬁeld games of control s: Finite diﬀerence approximations

Yves Achdou and Ziad Kobeissi. Mean ﬁeld games of control s: Finite diﬀerence approximations. Mathematics in Engineering, 3(3):1–35, 2021

2021

[4] [3]

Convergence of a ﬁnite diﬀerence scheme to weak solutions of the system of partial diﬀerential equations arising in mean ﬁeld games

Yves Achdou and Alessio Porretta. Convergence of a ﬁnite diﬀerence scheme to weak solutions of the system of partial diﬀerential equations arising in mean ﬁeld games. SIAM Journal on Numerical Analysis , 54(1):161–186, 2016

2016

[5] [4]

Graves, Houzhi Li, Jean-Fran çois Chassagneux, Franccois Delarue, and René A

Andrea Angiuli, Christy V. Graves, Houzhi Li, Jean-Fran çois Chassagneux, Franccois Delarue, and René A. Car- mona. Cemracs 2017: numerical probabilistic approach to mf g. ESAIM: Proceedings and Surveys , 2018

2017

[6] [5]

Adaptive extra-gradient methods for min-max optimization and games

Kimon Antonakopoulos, Elena Veronica Belmega, and Pana yotis Mertikopoulos. Adaptive extra-gradient methods for min-max optimization and games. In International Conference on Learning Representations , 2021

2021

[7] [6]

Sifting through the noise: Universal ﬁrst-order methods for stochastic var iational inequalities

Kimon Antonakopoulos, Thomas Pethick, Ali Kavis, Panay otis Mertikopoulos, and Volkan Cevher. Sifting through the noise: Universal ﬁrst-order methods for stochastic var iational inequalities. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. W ortman Vaughan, editors, Advances in Neural Information Processing Systems , volume 34, pages 13099–1...

[8] [7]

A tight and uniﬁed analysis of gradient-based methods for a whole spectrum of diﬀerentiab le games

W aïss Azizian, Ioannis Mitliagkas, Simon Lacoste-Juli en, and Gauthier Gidel. A tight and uniﬁed analysis of gradient-based methods for a whole spectrum of diﬀerentiab le games. In Silvia Chiappa and Roberto Calandra, editors, Proceedings of the Twenty Third International Conference o n Artiﬁcial Intelligence and Statistics , volume 108 of Proceedings of...

2020