{"id":"eb9e084c-0631-41d0-8ed4-f98a4f97a487","arxiv_id":"2602.14621","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An extragradient algorithm solves monotone mean-field FBSDEs by reformulating them as variational inequalities, with O(1/n) averaged and exponential last-iterate convergence under strong monotonicity.","lead":"This paper adapts classical extragradient iterations to solve coupled mean-field forward-backward stochastic equations by recasting the problem as a monotone variational inequality in a Hilbert space. It proves O(1/n) averaged convergence and, under stronger monotonicity, exponential last-iterate convergence, with numerical experiments on a scalar test case.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Main exponential-convergence result presupposes unique strong solution of (2.8); existence is imported from self-cited [39] with only an a priori estimate and a handwave in Lemma A.1.","rationale":"The reader's weakest_assumption correctly identifies the wellposedness of (2.8) as the linchpin. I independently examined the other flagged issue — the contraction factor in Theorem 2.11 — and found that it closes: with c_L ≤ ∥v∥_Lip (a consequence of Lipschitz strong monotonicity) and γ < min(1/(2∥v∥_Lip), c_L/∥v∥_Lip^2), the squared-contraction factor 1−2γc_L+4γ^2 c_L ∥v∥_Lip lies in (0,1); the final display in the paper appears to contain a typo (squaring ∥v∥_Lip). So the remaining load-bearing concern is the unproved existence of the solution α*. The manuscript itself flags this gap in Lemma 2.5 ('we couldn't find this exact result in the literature') and in Lemma A.1's reliance on 'ideas introduced in [14,39]'. Because the convergence analysis is conditional on that existence, and because the referenced result is in a self-cited unpublished preprint, an independent proof or a precise citation to a published theorem is required. This does not overturn the reader's CONDITIONAL verdict; it reinforces it as the appropriate assessment.","tokens_in":31756,"tokens_out":14825,"duration_ms":124192,"concrete_test":"Provide a self-contained continuation proof of wellposedness for (2.8)/(A.1) under Hypotheses 2.13/2.14. Specifically: (i) prove existence of a Lipschitz decoupling field on a short interval [T−δ,T] via the standard contraction argument for Lipschitz FBSDEs (e.g., as in [46] Theorem 4.3.1); (ii) show the a priori estimate in Lemma A.1 (the bound ∥U0−V0∥^2 ≤ C e^{CT} ∥X−Y∥^2) provides a uniform Lipschitz constant for the decoupling field that allows gluing intervals up to arbitrary T or, alternatively, exhibit a counterexample with c_F T large where the estimate blows up and no global strong solution exists. If the continuation step cannot be completed without additional smallness assumptions on c_F T, the stated convergence theorems must be restricted to short horizons.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim — exponential convergence of the extragradient iterates to α* (Theorems 2.9, 2.11, 2.19) — is a conditional statement: it assumes that the target FBSDE (2.8) admits a unique strong solution with Lipschitz decoupling field. That wellposedness is not proved in the manuscript. Theorem 2.15's proof cites 'Lemma 3.21 combined with Lemma 3.9' of the author's own preprint [39], and Lemma 2.5 states 'we couldn't find this exact result in the literature' before a one-line appeal to 'standard techniques' (Appendix A.1). Lemma A.1 establishes only an a priori estimate for the difference of two hypothetical solutions (inequality 2.14), then asserts that 'it follows naturally from ideas introduced in [14,39]' that a Lipschitz decoupling field exists. No continuation/contraction argument is given that would promote the a priori bound to global existence for arbitrary T with the stated L2-monotonicity (Hypothesis 2.14). If such wellposedness fails — e.g., if the non-gradient L2-monotone condition does not yield the promised decoupling-field regularity, or if [39] proves only a weaker notion of solution — then the α* appearing in the convergence rate is undefined, and the central exponential-convergence claim is vacuous. This is not merely a citation problem: it is the foundational assumption on which the entire algorithm's target rests.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":32151,"tokens_out":19055,"duration_ms":190783,"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":[{"comment":"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).","section":"§2.2, Theorem 2.15, Remark 2.21, Lemma A.1"},{"comment":"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.","section":"§2.1.2, proof of Theorem 2.11"},{"comment":"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.","section":"§3, Lemma 3.3 and Corollary 3.4"}],"minor_comments":[{"comment":"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.","section":"General"},{"comment":"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.","section":"Theorem 2.11"},{"comment":"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":"Final rate in proof of Theorem 2.11"},{"comment":"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":"Section 4, Algorithm 1"},{"comment":"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.","section":"Section 4.2.1"}],"recommendation":"major_revision","confidential_remarks":"The central existence results are delegated to the author's own preprint [39], which appears to be unpublished. The editor may wish to verify the status and independence of that preprint before further review. In addition, the proof of Theorem 2.11 should be checked by an optimization/monotone-operator specialist: the displayed algebra does not appear to yield the claimed contraction, and the final rate expression is internally inconsistent. These are fixable in principle, but they are central to the paper's advertised contribution, so a major revision is appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","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.\n\nThe 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.\n\nSecond, 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].\n\nThe numerical section is illustrative rather than empirical: no error bars, no baseline comparison, no code. That is minor relative to the two issues above.\n\nOverall, 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.","headline":"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.","tokens_in":32603,"tokens_out":4550,"would_cite":false,"duration_ms":40271,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49N80","60H10","65K15","35Q89"],"pacs":[],"model":"deepseek-v4-flash","headline":"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.","keywords":["mean-field games of controls","mean-field FBSDEs","extragradient method","monotone variational inequality","displacement monotonicity","common noise","fictitious play","numerical scheme"],"falsifier":"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.","tokens_in":31624,"feed_emoji":"🎲","tokens_out":7247,"duration_ms":68959,"temperature":0.7,"pith_summary":"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.","feed_headline":"Extragradient solver for mean-field games converges exponentially","feed_subtitle":"A probabilistic control-space scheme with provable rates, unchanged by common noise.","key_machinery":"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","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"fun_headline_variants":["Extragradient converges exponentially for mean-field control games","Explicit extragradient for MFGs and FBSDEs: exponential speed","Mean-field FBSDEs solved exponentially fast via extragradient","Extragradient method: exponential convergence for MFGs and FBSDEs"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Extragradient converges exponentially for mean-field control games","Explicit extragradient for MFGs and FBSDEs: exponential speed","Mean-field FBSDEs solved exponentially fast via extragradient","Extragradient method: exponential convergence for MFGs and FBSDEs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000466,"raw_usage":{"total_tokens":2130,"prompt_tokens":682,"completion_tokens":1448,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":426,"completion_tokens_details":{"reasoning_tokens":1371}},"tokens_in":426,"tokens_out":1448,"duration_ms":11515,"temperature":1.0,"reasoning_tokens":1371,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T23:08:17.595956+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}