{"id":"83918cdc-89f8-4e94-9c57-d6d357327ea0","arxiv_id":"2608.03376","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"An averaged, common-noise-free mean-field game equilibrium induces an approximate Nash equilibrium for a two-scale common-noise mean-field game, with error O(delta^(1/6)) or O(delta^(1/3)).","lead":"This paper shows how to build good approximate equilibria in mean-field games with common noise when the system has a fast and a slow time scale. The idea is to replace the fast, noisy component by its average, solve a simpler game, and then use that solution in the original game; the error shrinks as the time scales separate.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 3's Poisson-equation estimate needs ∂_t u and ∂_xx u for sources only assumed C^{1/2,1}_b in (t,x); Assumptions 1/3/6 do not provide this, so the strong-averaging estimates behind Theorems 1, 2, and 9 are not justified as stated.","rationale":"The paper's construction is coherent, and the independence assumptions highlighted by the reader are explicitly stated and indeed known to be necessary for strong convergence; I do not treat them as an internal flaw. The most load-bearing issue is instead the regularity gap in Lemma 3: the proof needs ∂_t u and ∂_xx u for a Poisson solution whose source is only C^{1/2,1}_b in (t,x), which the stated assumptions do not guarantee. Because Proposition 2, Lemma 2, and the central error estimates in Theorems 1, 2, and 9 all rely on this lemma, the quantitative Nash-error claim is not fully justified as written. This is a concrete, repairable gap—likely fixable by strengthening smoothness assumptions—so it does not warrant rejection, but it does confirm the need for a conditional verdict. The reader's separate concern about Assumption 3.1 from [29] is plausible but secondary; my concern is more precisely localized at inequality (108) and the Itô step in Lemma 3.","tokens_in":43934,"tokens_out":37890,"duration_ms":336091,"concrete_test":"Verify whether b and F1 under Assumptions 1/3/6 satisfy the hypotheses of [59, Lemma 4.1] used for inequality (108). As a direct analytical check, take the fast process to be an OU diffusion and choose Φ(t,x,y)=arctan(x)+y, a bounded function that is C^1 in x, 1/2-Hölder in t, and Lipschitz in y; compute u(t,x,y)=∫_0^∞ E[Φ(t,x,Y_s)]ds and test whether ∂_xx u is bounded. If u_xx is undefined or unbounded, Lemma 3's estimate fails under the stated assumptions. If it fails, re-run the proof of Theorem 1's Step 2 with b, F1, and \\hat f2 assumed C^{1,2}_b in (t,x) to confirm the δ^{1/6} rate is recovered.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The quantitative Nash-error estimates rest on Lemma 3 (Appendix 5.7), which asserts E[sup_t |H^δ_t|^2] ≤ Cδ for H^δ_t = ∫_0^t (F1(s,X^δ_s,Y^δ_s)−\\bar F1(s,X^δ_s))ds. The proof applies Itô's formula to a Poisson-equation solution u with L^Y u = Φ and claims, via [59, Lemma 4.1], the bounds |∂_t u|+|∂_x u|+|∂_xx u|+|∂_y u| ≤ C(1+|y|^ℓ), inequality (108). But the standing assumptions only place b, F1, and \\hat f2 in C^{1/2,1}_b([0,T]×R) in (t,x) (Assumptions 1(ii), 3, 6(ii), 7). Since L^Y is an operator in y alone, u inherits its t,x smoothness from Φ; with only 1/2-Hölder time regularity and C^1 space regularity, ∂_t u and ∂_xx u need not exist, so the Itô computation in Lemma 3 is not available. Proposition 2's bound on A^1_t, Lemma 2, and the A3/Step 2 bounds in Theorems 1 and 9 all invoke this lemma. If [59, Lemma 4.1] in fact requires more regularity than these assumptions provide, the proofs have a gap that would need additional C^{1,2}_x/C^1_t assumptions on b, F1, and \\hat f2, or a different averaging argument.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for constructing approximate Nash equilibria in mean-field games (MFGs) with common noise when the dynamics have two time scales. The common noise is carried by a fast, ergodic diffusion, while the slow variable is either optimally stopped or optimally controlled. The authors average the fast scale with respect to its invariant measure to obtain an effective MFG without common noise, and then show that an equilibrium of the effective problem induces a strong randomized epsilon-MFG equilibrium of the original two-scale problem. The Nash error is quantified as C*delta^(1/6) in a Poisson-equation regime and C*delta^(1/3) in an explicit Ornstein-Uhlenbeck regime. The paper also develops existence results for randomized equilibria of the effective problem via a linear-programming relaxation, and extends the construction to relaxed controls.","tokens_in":44282,"tokens_out":16611,"duration_ms":160773,"significance":"If the technical gaps are repaired, the paper makes a valuable contribution: it gives a constructive route from a tractable averaged MFG without common noise to an approximate equilibrium of a common-noise MFG, bypassing the master equation and providing explicit convergence rates. The strengths are substantial: the structural assumptions are stated clearly, the independence of the slow volatility from the fast variable is explicitly flagged and its necessity is acknowledged, the construction is non-circular because the effective equilibrium is an input rather than a disguised version of the original problem, and the paper contains new existence results for randomized stopping equilibria. The two regimes, with rates 1/6 and 1/3, are concrete and would be directly useful in applications such as the electricity-market entry-exit game mentioned in the introduction.","major_comments":[{"comment":"The proof of Lemma 3 applies Itô's formula to a solution u of L^Y u = Phi and asserts the bound |partial_t u| + |partial_x u| + |partial_xx u| + |partial_y u| <= C(1+|y|^ell). Under Assumptions 1, 3, 6, and 7, the source Phi = F1 - bar F1 (and similarly b - bar b, hat f2 - bar hat f2) is only C^{1/2,1}_b in (t,x), and since L^Y acts only in the y variable, u does not automatically acquire partial_t u or partial_xx u derivatives from Phi. The Itô computation, and hence the bound E[sup_t |H^delta_t|^2] <= C delta, is therefore not justified as stated. This estimate is load-bearing: it is used in Proposition 2 (bound on A^1_t), Lemma 2, and in the Step 2 bounds of Theorems 1 and 9, so the quantitative Nash-error estimates rest on it. The authors should either add regularity hypotheses (for example, C^{1,2}_b in (t,x) for b, F1, and hat f2) or replace the pointwise Itô argument with a regularization or weak-derivative argument that yields the same estimate.","section":"Appendix 5.7, Lemma 3 and inequality (108)"},{"comment":"The proof of Theorem 7 asserts that conv(R0) is a closed subset of the compact set R1. This is not established and is generally false: the set P_S of laws of strict stopping times need not be weakly closed, since a weak limit of strict stopping-time laws can be a genuinely randomized stopping law, and the finite convex hull of a non-closed set need not be closed. The separation argument should instead be run with the closure overline(conv)(R0); it then yields R subset of overline(conv)(R0), and together with Theorem 6 gives R = R1, which is sufficient for the representation in Theorem 8. The equality conv(R0) = R1 stated in Theorem 8 is stronger than what the argument establishes. Please revise Theorems 7 and 8 accordingly; as written, the key representation of LP occupation measures by randomized stopping kernels is not fully justified.","section":"Section 3, Theorems 7 and 8"}],"minor_comments":[{"comment":"In the chain of estimates for A^3_t, Lemma 3 is invoked for an integrand containing the factor 1_{(t,T]}(tau*). This application should be spelled out, for example by writing the stopped integral as H_{tau* wedge (k+1)Delta} - H_{tau* wedge kDelta} and using the sup-norm bound on H^delta; the current display is terse and could be misread as applying Lemma 3 to a discontinuous-in-time multiplier.","section":"Theorem 1, Step 2"},{"comment":"The two different spaces denoted V(C) in the Notation section appear with the same symbol; please use distinct symbols, for example mathcal V(C) and mathbb V(C), to avoid confusion between deterministic flows and F^c-progressively measurable flows.","section":"Section 1, Notation"},{"comment":"The sentence stating that Assumptions 1/3 or 2/4/5 'ensure that Assumption 3.1 in [29] holds' is an omitted verification. Please either provide a direct check of the conditions of Assumption 3.1 or give precise pointers to the corresponding inequalities, since this assertion is the bridge to the LP equilibrium existence theorem.","section":"Section 3, before Theorem 4"},{"comment":"There are several typographical slips, including 'equatio', 'regiume', inconsistent use of epsilon/epsilon in Definition 3, and broken hyphenation in 'H older'; these should be cleaned up in the revision.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read. The paper does something genuinely new: it shows that for a two-scale MFG with common noise, an equilibrium of the averaged (no-common-noise) MFG can be lifted to an explicit ε-MFG equilibrium of the original game, with ε = C δ^{1/6} or C δ^{1/3} depending on regime. This is a practical route that avoids the master equation, and the authors are honest about the structural assumptions (volatility of the slow component independent of the fast variable, fast dynamics independent of the slow). They also prove existence of randomized stopping equilibria via linear programming, connecting their construction to the LP framework of their earlier papers. That part is solid and new in itself, and the self-citations to their LP framework are earned.\n\nThe proof strategy is clear: strong stochastic averaging bounds control the payoff difference. Theorems 1, 2, and 9 are the core, and the payoff decomposition is clean. I believe the main argument is correct in spirit.\n\nBut the soft spot is real. The stress-test note holds up: Lemma 3's proof applies Itô's formula to a function u solving L^Y u = Φ, and claims bounds on ∂_t u and ∂_xx u. Under the stated assumptions, b, F1, and \\hat f2 are only C^{1/2,1}_b in (t,x), so u inherits at best 1/2-Hölder time and C^1 space regularity. ∂_t u and ∂_xx u need not exist. The pinch is that Lemma 3 sits under Proposition 2, Lemma 2, and the A3/Step 2 estimates in Theorems 1 and 9. It is load-bearing. The fix is probably to strengthen the assumptions to C^{1,2} in (t,x) for those coefficients, or to replace the Itô computation with a smoothing argument. Until that is patched, the quantitative rates are not justified as written.\n\nThe paper is otherwise careful, and the nature of the flaw is a regularity mismatch, not a false conclusion. This is for researchers in MFG and stochastic averaging who want a constructive alternative to master equations; the gap is fixable, but it needs to be fixed. I would send it to a serious referee with a clear request to address Lemma 3, and I'd cite it once the fix is in place.","headline":"A genuinely new two-scale averaging route to approximate common-noise MFG equilibria, with a real but patchable regularity gap in the key averaging lemma.","tokens_in":44780,"tokens_out":3413,"would_cite":true,"duration_ms":28472,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49N80","60H10","91A16"],"pacs":[],"model":"deepseek-v4-flash","headline":"An equilibrium of the averaged game, played under the original two-scale noise, is an approximate Nash equilibrium of the full common-noise game, with error of order δ^{1/6} or δ^{1/3}.","keywords":["mean-field games","common noise","averaging principle","two-time-scale diffusions","approximate Nash equilibria","randomized optimal stopping","relaxed controls","stochastic averaging"],"falsifier":"Compute the maximal unilateral deviation in a small two-scale MFG with OU fast noise, using the paper's construction, and compare it with $δ^{{1/3}}$; a deviation that fails to vanish as δ→0, or that decays at a different rate, would falsify the quantitative claim. Alternatively, take a two-scale common-noise MFG in which the slow volatility depends on the fast variable—for example σ(t,x,Y_t^δ)—and check the strong convergence estimate sup_{0≤t≤T} E|X_t^δ−$X_t^{0}$|, which is known to fail in that case (Example 4.1 of [48]).","tokens_in":43741,"feed_emoji":"⚖️","tokens_out":7219,"duration_ms":71620,"temperature":0.7,"pith_summary":"This paper tackles mean-field games in which a large population of agents is exposed both to common noise and to dynamics on two time scales: a slow, controlled or stopped variable and a fast ergodic variable carrying the common noise. The central claim is that one does not need to solve the full common-noise game: take the effective game obtained by averaging all coefficients and payoffs against the stationary law of the fast variable, find any equilibrium there, then play the same randomized stopping or control strategy in the original two-scale system. For δ small enough this strategy is an ε-equilibrium of the original game, with ε = $Cδ^{{1/6}}$ in the Poisson-equation setting and ε = $Cδ^{{1/3}}$ when the fast variable is explicitly Ornstein–Uhlenbeck. If true, this gives a constructive, quantitative route to approximate equilibria in a class of common-noise MFGs that is otherwise analytically very hard.","feed_headline":"Averaging fast noise builds approximate Nash equilibria","feed_subtitle":"Solving the averaged game without common noise gives explicit epsilon-equilibria of the two-scale game.","key_machinery":"The load-bearing object is the effective mean-field game: the drift and payoff are averaged against the invariant measure of the fast ergodic component, and the original common noise disappears from the effective problem. The argument is carried by strong stochastic averaging estimates—an order-1/2 sup-in-time L2 bound in the Poisson-equation regime and an order-1/3 bound from a discretization argument in the explicit OU regime—which let payoff differences under unilateral deviations be controlled term by term. To make the construction feed on existing existence theory, randomization is encoded through regular probability kernels for stopping and relaxed controls, and the continuous-time linear-programming formulation of MFGs is used to prove that equilibria of the effective game exist.","core_discovery":"The paper establishes that the common-noise MFG with two time scales can be approximated by an averaged MFG without common noise: every equilibrium of the effective game, interpreted through the same randomized strategy, induces a strong randomized ε-equilibrium of the original game. The Nash error is controlled by strong stochastic averaging estimates—L1/L2 convergence of the slow process to its averaged limit—and the time-scale parameter δ directly bounds the loss under unilateral deviations. For optimal stopping the rates are $δ^{{1/6}}$ (Poisson-equation regime) and $δ^{{1/3}}$ (explicit OU regime); for controlled diffusions the relaxed-control version gives $δ^{{1/6}}$. The paper also proves existence of effective equilibria with randomized stopping via the linear-programming formulation, and shows that LP occupation measures are exactly represented by randomized stopping kernels.","pith_inferences":["A natural stress test is whether the rates 1/6 and 1/3 are sharp; a simple linear-quadratic two-scale example could be solved exactly and compared with the predicted error, which is not done in the paper.","The structural independence assumptions (volatility independent of the fast variable, fast dynamics independent of the slow variable) look necessary for the L1/L2 control of paths, but the effective equilibrium might still approximate payoffs in weaker metrics, at the cost of losing the Nash-error guarantee.","Because the construction needs only an equilibrium of the averaged game, the same scheme could be coupled with numerical LP solvers for entry-exit or electricity-market games where common weather or demand shocks act on the fast scale.","Extending the fast dynamics to non-Gaussian ergodic noise, such as Lévy processes or piecewise-deterministic Markov processes, would likely preserve the scheme as long as a strong averaging estimate with a quantified rate is available."],"forward_implications":["Approximate equilibria for two-scale common-noise MFGs can be constructed by solving a simpler averaged problem, bypassing the master equation or a direct fixed-point argument for the full game.","The error estimate quantifies the cost of ignoring fast fluctuations: the same randomized strategy achieves ε = O(δ^{1/6}) or O(δ^{1/3}), so the approximation improves as the time-scale separation grows.","In optimal stopping MFGs, where pure-strategy equilibria may fail to exist, the randomized-stopping framework guarantees existence of effective equilibria and therefore of approximate two-scale equilibria.","In controlled MFGs, an effective equilibrium with relaxed control yields a two-scale ε-equilibrium; under a convexity condition the strict-control version also works.","The LP representation gives a computational handle: one can solve the averaged game by linear programming and then simulate the original multiscale dynamics under the recovered randomized strategy."],"supporting_citations":[{"why":"Supplies the Poisson-equation strong averaging estimates of order 1/2 that drive the δ^{1/6} Nash-error bounds.","marker":"[59]"},{"why":"Supplies the order-1/3 strong convergence result for slow–fast McKean–Vlasov diffusions used in the explicit OU regime.","marker":"[58]"},{"why":"Introduces the relaxed solution approach for mean-field games of optimal stopping on which the effective problem and its LP formulation rest.","marker":"[13]"},{"why":"Provides the continuous-time LP formulation and existence results for control and optimal-stopping MFGs, including the equilibrium existence theorem used for the effective game.","marker":"[29]"},{"why":"Gives sufficient conditions for existence of effective MFG equilibria with strict stopping, cited in Remark 2.","marker":"[30]"},{"why":"Proves strong convergence of the averaging principle and supplies Example 4.1 showing that dependence of the slow volatility on the fast variable blocks strong convergence, justifying a key structural assumption.","marker":"[48]"},{"why":"Shows that pure-strategy equilibria may fail in optimal-stopping MFGs, motivating the passage to randomized stopping.","marker":"[10]"},{"why":"Provides the probabilistic theory of MFGs with common noise, including disintegration and immersion results used to represent randomized stopping kernels.","marker":"[19]"},{"why":"Supplies the notion of approximate equilibrium adopted in the definitions of ε-MFG equilibrium.","marker":"[1]"}],"fun_headline_variants":["Mean-field games solved by averaging fast noise","Two-scale MFG: approximate equilibria from averaged game","Fast noise averaged, Nash equilibria approximated","Effective game approximates common-noise MFG equilibria","Averaging fast dynamics yields epsilon-Nash in MFG"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The slow variable's volatility must not depend on the fast variable, and the fast dynamics must not depend on the slow variable; if either fails, the strong convergence estimates that carry the Nash-error bound are known to break down.","fun_headline_variants_meta":{"raw":{"variants":["Mean-field games solved by averaging fast noise","Two-scale MFG: approximate equilibria from averaged game","Fast noise averaged, Nash equilibria approximated","Effective game approximates common-noise MFG equilibria","Averaging fast dynamics yields epsilon-Nash in MFG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1263,"prompt_tokens":874,"completion_tokens":389,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":490,"completion_tokens_details":{"reasoning_tokens":314}},"tokens_in":490,"tokens_out":389,"duration_ms":3814,"temperature":1.0,"reasoning_tokens":314,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:50:38.884194+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the maximal unilateral deviation in a small two-scale MFG with OU fast noise, using the paper's construction, and compare it with $δ^{{1/3}}$; a deviation that fails to vanish as δ→0, or that decays at a different rate, would falsify the quantitative claim. Alternatively, take a two-scale common-noise MFG in which the slow volatility depends on the fast variable—for example σ(t,x,Y_t^δ)—and check the strong convergence estimate sup_{0≤t≤T} E|X_t^δ−$X_t^{0}$|, which is known to fail in that case (Example 4.1 of [48]).","supporting_citations":[{"cited_title":"R \\\"o ckner and L","cited_arxiv_id":null,"evidence_quote":"Supplies the Poisson-equation strong averaging estimates of order 1/2 that drive the δ^{1/6} Nash-error bounds."},{"cited_title":"R \\\"o ckner, X","cited_arxiv_id":null,"evidence_quote":"Supplies the order-1/3 strong convergence result for slow–fast McKean–Vlasov diffusions used in the explicit OU regime."},{"cited_title":"Bouveret, R","cited_arxiv_id":null,"evidence_quote":"Introduces the relaxed solution approach for mean-field games of optimal stopping on which the effective problem and its LP formulation rest."},{"cited_title":"Dumitrescu, M","cited_arxiv_id":null,"evidence_quote":"Provides the continuous-time LP formulation and existence results for control and optimal-stopping MFGs, including the equilibrium existence theorem used for the effective game."},{"cited_title":"Dumitrescu, M","cited_arxiv_id":null,"evidence_quote":"Gives sufficient conditions for existence of effective MFG equilibria with strict stopping, cited in Remark 2."},{"cited_title":"Liu , Strong convergence of principle of averaging for multiscale stochastic dynamical systems , Communications in Mathematical Sciences, 8 (2010), pp","cited_arxiv_id":null,"evidence_quote":"Proves strong convergence of the averaging principle and supplies Example 4.1 showing that dependence of the slow volatility on the fast variable blocks strong convergence, justifying a key structural assumption."},{"cited_title":"Bertucci , Optimal stopping in mean field games, an obstacle problem approach , Journal de Math \\'e matiques Pures et Appliqu \\'e es, 120 (2018), pp","cited_arxiv_id":null,"evidence_quote":"Shows that pure-strategy equilibria may fail in optimal-stopping MFGs, motivating the passage to randomized stopping."},{"cited_title":"Carmona and F","cited_arxiv_id":null,"evidence_quote":"Provides the probabilistic theory of MFGs with common noise, including disintegration and immersion results used to represent randomized stopping kernels."},{"cited_title":"Ahuja, W","cited_arxiv_id":null,"evidence_quote":"Supplies the notion of approximate equilibrium adopted in the definitions of ε-MFG equilibrium."}],"review_version":2}