{"id":"ed06bf46-1d01-4420-86f1-698e8e8370bd","arxiv_id":"1908.01613","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"Finite-horizon mean field control and game problems can be approximated by training neural networks, with explicit error rates in population size, network width, and time step.","lead":"The paper proves error bounds for two neural-network numerical methods that solve finite-horizon mean field control and mean field game problems. A generalist should read it because it is among the first to give rigorous convergence rates for deep-learning approximations in this setting.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposition 8 drops the square root from Proposition 13, so the network-width rate in Theorem 3 is twice as fast as the proof supports.","rationale":"The paper's goal is to give an explicit, quantitative convergence rate for neural-network approximations of finite-horizon mean field control problems. The load-bearing part of that claim is Theorem 3's three-term error decomposition, especially the network-width rate ε2. The proof of Proposition 8 combines the approximation estimate of Proposition 10 with the stability estimate of Proposition 13. The stability estimate gives a square-root dependence: |J_N(v)-J_N(w)| ≤ C(Γ²+1/R)^{1/2}. The authors bound the inside quantity Γ²+1/R by C n_in^{-1/(3(d+1))}, but then omit the outer square root, concluding directly with the same exponent. This is a concrete, local algebraic error in a key step, not a philosophical objection or a disagreement with the overall framework. The consequence is that the stated width exponent is too optimistic by a factor of two; the proof actually supports n_in^{-1/(6(d+1))}. The reader's rationale explicitly identifies this same issue, although the reader's 'weakest assumption' field instead emphasizes the smoothness/linearity assumptions. Both concerns are real, but the dropped square root is the more direct threat to the theorem as stated, because it affects the central quantitative claim even within the paper's own assumptions. The concern is repairable: replacing the rate in Theorem 3 and Proposition 8 with the slower exponent restores correctness, and the numerical experiments remain suggestive. A less severe critique would not change the verdict, and a harsher one would ignore that the framework, the decomposition, and the qualitative convergence are sound. Therefore the reader's CONDITIONAL verdict remains appropriate, and this stress-test does not move it.","tokens_in":39985,"tokens_out":5918,"duration_ms":51918,"concrete_test":"Recompute the final step of Proposition 8 by substituting R=n_in^a and Γ=C0(1+n_in^a)n_in^{-3a/2} with a=1/(3(d+1)) into the bound (Γ²+1/R)^{1/2} from Proposition 13. The result is O(n_in^{-a/2})=O(n_in^{-1/(6(d+1))}), not O(n_in^{-a})=O(n_in^{-1/(3(d+1))}). If this slower rate is confirmed, Theorem 3's ε2(n_in) must be weakened accordingly, or an additional argument must be supplied to recover the claimed faster rate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim is Theorem 3's width rate ε2(n_in)=O(n_in^{-1/(3(d+1))}). In the proof of Proposition 8, the authors set R=n_in^{1/(3(d+1))} and Γ=C0(1+R)n_in^{-1/(2(d+1))}. Proposition 13 states |J_N(v)-J_N(w)| ≤ C(Γ²+1/R)^{1/2}. The displayed algebra in the proof of Proposition 8 correctly shows Γ²+1/R ≤ C1 n_in^{-1/(3(d+1))}, but then the proof stops without applying the square root. Since C(Γ²+1/R)^{1/2} ≤ C n_in^{-1/(6(d+1))}, the stated ε2 rate is not supported; the proof as written yields only n_in^{-1/(6(d+1))}. This does not destroy qualitative convergence, because the error still tends to zero as n_in grows, but it overstates the width convergence rate by a factor of two in the exponent. Because Theorem 3 is the paper's main quantitative guarantee, and the entire complexity comparison rests on this exponent, the rate statement in the theorem and in the abstract should be corrected or the proof revised.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops and tests two neural-network based numerical schemes for finite-horizon McKean-Vlasov control problems. Method 1 directly minimizes a sampled, time-discretized social cost over neural-network feedback controls. Method 2 rewrites a general McKean-Vlasov FBSDE as a forward control problem and minimizes a terminal-matching loss, so it also applies to mean field games. The central theoretical contribution is Theorem 3, which decomposes the error between the true mean field optimum and the proxy optimum into three explicit rates: O(N^{-1/max(d,4)} times a logarithmic factor in d=4) for the finite-population approximation, O(n_in^{-1/(3(d+1))}) for the neural-network width, and O(sqrt(Delta t)) for the time discretization. The proof proceeds through three steps: finite-agent approximation via a result from the authors' book, neural-network approximation of the decoupling feedback using a derivative-aware periodic approximation theorem, and an Euler scheme error bound. Several numerical examples, including common-noise and mean-field-game cases, are compared with PDE or ODE benchmarks.","tokens_in":40167,"tokens_out":10533,"duration_ms":111436,"significance":"If the stated rates were established, this would be a valuable contribution: it gives explicit, non-asymptotic convergence rates for neural-network methods in finite-horizon mean field control, extends the methodology to common noise, and provides a genuine algorithmic alternative to PDE-based solvers. The paper has real strengths: the error decomposition is clean; the neural approximation step is novel in that it approximates a function and its derivatives simultaneously with controlled Lipschitz constants; the time-discretization estimate is proved with constants independent of the number of agents; and the numerical study is broad and honestly presented. However, the main width rate in Theorem 3 is not supported by the proof as written: Proposition 13 gives only a square-root bound, so the proof yields the slower rate n_in^{-1/(6(d+1))} unless Proposition 13 is strengthened. This is a load-bearing quantitative claim, not a cosmetic issue. The paper also relies on strong classical regularity of the decoupling field, and several test cases fall outside that scope; the authors acknowledge this for some but not all of those cases.","major_comments":[{"comment":"In the proof of Proposition 8, the authors set R = n_in^{1/(3(d+1))} and correctly show that Gamma^2 + 1/R <= C_1 n_in^{-1/(3(d+1))}. However, Proposition 13 bounds |J_N(v) - J_N(w)| by C (Gamma^2 + 1/R)^{1/2}, not by C (Gamma^2 + 1/R). Therefore the displayed conclusion should be J_N(hat v) >= J_N(hat phi) - K_2 n_in^{-1/(6(d+1))}, not the stated exponent -1/(3(d+1)). This discrepancy propagates directly to the rate epsilon_2(n_in) in Theorem 3 and to the associated claims in the introduction and abstract. The theorem can be repaired by changing the width rate to O(n_in^{-1/(6(d+1))}), or by proving a stronger, square-root-free version of Proposition 13; as written, the stated rate is unsupported.","section":"Section 3.2, proof of Proposition 8"},{"comment":"Proposition 15 is stated under condition (29), which requires |partial_t phi(t,x)| <= C_1 |(t,x)| in addition to boundedness of partial_x phi and partial^2_xx phi. For the network hat phi produced by Proposition 8, the proof of Proposition 10 supplies bounds on the Lipschitz constants of phi, partial_x phi, and partial^2_xx phi, but it does not establish the stated growth condition on partial_t phi; in fact the scaling construction gives partial_t phi bounded uniformly, not necessarily vanishing at (0,0). The Ito estimates in Appendix D appear to need only boundedness of partial_t phi, so the mismatch is likely repairable, but it must be addressed: either verify (29) for the constructed network or weaken the assumption in Proposition 15 to the boundedness actually used.","section":"Section 3.3, Proposition 15 and its application to Theorem 3"}],"minor_comments":[{"comment":"The text states that test cases 5 and 6 fall outside the scope of the theory, but test case 2 also falls outside, since its terminal cost g(x) = min(|x-xi_1|, |x-xi_2|) is non-smooth. Please make the limitation statement cover all three cases explicitly.","section":"Section 5, numerical results"},{"comment":"After correcting the rate in Proposition 8 and Theorem 3, the discussion of the curse of dimensionality and of possible improvements should be updated to reflect the actual width exponent n_in^{-1/(6(d+1))} rather than the stated n_in^{-1/(3(d+1))}.","section":"Remark 4 and Remark 11"},{"comment":"The proof of Proposition 7 quotes inequality (36) from [19, Theorem 6.17] and gives only a short derivation of the reduction to closed-loop controls. Since this step supplies the finite-population rate epsilon_1, please state the precise modification of the referenced theorem or provide the full argument, so that the main theorem is self-contained.","section":"Appendix B"},{"comment":"There are minor typographical issues, including 'Apppendix A' in Remark 12 and the inconsistent spelling 'Bachmann-Landau'; these should be cleaned up in a revision.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The square-root discrepancy in Proposition 8 is real and central, but it is fixable within the manuscript's scope by correcting the width exponent in Theorem 3 to -1/(6(d+1)) or by strengthening Proposition 13. I therefore recommend major revision rather than rejection. The finite-population step is imported from the authors' own book, which limits novelty but is not circular. The numerical experiments are useful even though several test cases fall outside the theoretical assumptions; the authors should make the scope limitations uniform and explicit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere is the quick version. Carmona and Lauri\\`ere propose two neural-network methods for finite-horizon mean field control and games: Method 1 minimizes the MFC cost directly; Method 2 recasts a McKean-Vlasov FBSDE as a control problem. That second step is a genuine extension of the deep BSDE method to mean-field settings, and it handles common noise in a way the earlier literature did not. The paper's real contribution is a three-step error decomposition \\u2014 finite population, network width, time discretization \\u2014 with explicit rates. Nobody had done this for this class. The numerical experiments are suggestive: they benchmark against PDE or analytic solutions, include common-noise examples, and are honest that test cases 5 and 6 lie outside the theory.\n\nThe main quantitative claim is the network-width rate in Theorem 3, \\epsilon_2(n_in) = O(n_in^{-1/(3(d+1))}). The proof has a genuine gap. Proposition 13 yields |J_N(v)-J_N(w)| \\le C(\\Gamma^2+1/R)^{1/2}. In the proof of Proposition 8, the authors set R = n_in^{1/(3(d+1))} and \\Gamma = C_0(1+R)n_in^{-1/(2(d+1))}, and correctly show \\Gamma^2+1/R \\le C_1 n_in^{-1/(3(d+1))}. But then they stop without taking the square root. The correct conclusion is \\epsilon_2 = O(n_in^{-1/(6(d+1))}). The stated rate is too fast by a factor of two in the exponent. This is not a philosophical quibble: it is the theorem's headline quantitative guarantee. The good news is qualitative convergence survives \\u2014 the error still goes to zero \\u2014 and the gap is likely repairable by revising the rate or sharpening the stability estimate.\n\nOther soft spots, in order of importance. The regularity assumptions (B2, C2, linear dynamics in A1) are strong, and the authors' own test cases 5 and 6 fall outside them. The finite-population step is imported almost unchanged from the authors' book, which is fine, but it means the genuinely new part is narrower. No code or data are provided, and the SGD/optimization error is acknowledged but not analyzed.\n\nOn balance I would send this to referees. It is a serious method paper with a real but contained error in one rate statement. My recommendation: accept after a revision that fixes the exponent and clarifies which test cases the theory actually covers.\n\nLet me know if you want to discuss.","headline":"A serious paper and a real framework, but the main network-width rate in Theorem 3 is overstated: the proof drops the square root from Proposition 13, so the stated n^{-1/(3(d+1))} should be n^{-1/(6(d+1))}.","tokens_in":40767,"tokens_out":2451,"would_cite":true,"duration_ms":21207,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49N80","93E20","65K10","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A neural-network proxy provably reaches the mean field control optimum within an explicit error.","keywords":["mean field control","mean field games","McKean-Vlasov dynamics","neural network approximation","forward-backward SDE","convergence rate","common noise","stochastic gradient descent"],"falsifier":"Solve a smooth linear-quadratic mean field control problem with known optimum, at dimension $d=4$, and measure the deficit $\\inf_\\alpha J(\\alpha) - \\inf_\\phi \\tilde J_N(\\phi)$ as $N$ doubles with $n_{\\mathrm{in}}$ and $\\Delta t$ fixed and large; if the deficit does not decay like $N^{-1/4}\\sqrt{\\ln N}$, the bound is wrong. In the other direction, running the same test with a nonsmooth terminal cost, such as the minimum of distances to two targets, tests whether the smoothness assumption is truly needed for the claimed rate.","tokens_in":39678,"feed_emoji":"🎯","tokens_out":6962,"duration_ms":65352,"temperature":0.7,"pith_summary":"This paper claims that a practical machine-learning scheme for finite-horizon mean field control comes with a provable accuracy certificate. The scheme replaces the true control problem, where a continuum of agents choose open-loop controls, by a proxy in which N agents follow one neural-network feedback rule on a discrete time grid. The main theorem bounds how far the proxy optimum can sit below the true optimum by an explicit sum of three errors: the finite-population error, the neural-network approximation error, and the Euler time-discretization error. If the theorem is right, neural networks are not just a heuristic for high-dimensional mean field problems: the optimization one actually runs is guaranteed to stay within a computable distance of the exact optimum. The paper also offers a second method for general McKean-Vlasov forward-backward SDE systems, covering mean field games and common noise, backed by numerical experiments rather than a rate theorem.","feed_headline":"Neural-network control now comes with proven error rates","feed_subtitle":"A finite-population, finite-network, discrete-time proxy is guaranteed to stay within an explicit gap of the true mean field optimum.","key_machinery":"The load-bearing object is the decoupling field $V(t,x)=U(t,x,\\mu_t)$, the value of the master field $U$ along the optimal flow; its smoothness is what lets the optimal open-loop control be replaced by a Lipschitz feedback function $\\hat v(t,x)=\\hat\\alpha(t,x,\\mu_t,V(t,x))$. The proof then chains three approximations: (i) replacing the mean field by $N$ interacting particles, using a propagation-of-chaos estimate from the FBSDE theory to obtain $\\epsilon_1(N)$; (ii) approximating $\\hat v$ by a one-hidden-layer neural network with a periodic $C^3$ activation function, using simultaneous approximation of the function and its first and second derivatives, giving $\\epsilon_2(n_{\\mathrm{in}})$; (iii) discretizing time with an Euler scheme, with a strong-error lemma whose constants are independent of $N$, giving $\\epsilon_3(\\Delta t)$. The crucially named object is the class $\\mathcal{N}^\\psi$ of neural networks with periodic activation, chosen so that the approximating network inherits controlled Lipschitz constants for itself and its derivatives.","core_discovery":"Theorem 3 states that for the finite-horizon McKean-Vlasov control problem, the minimum over controls $\\alpha$ of $J(\\alpha)$ is at least the minimum over neural-network feedback controls $\\phi$ of the $N$-agent, discrete-time cost $\\tilde J_N(\\phi)$, minus $\\epsilon(N,n_{\\mathrm{in}},\\Delta t)$. Here $\\epsilon = \\epsilon_1(N)+\\epsilon_2(n_{\\mathrm{in}})+\\epsilon_3(\\Delta t)$, with $\\epsilon_1(N)=O(N^{-1/\\max(d,4)}(1+\\ln N)^{\\mathbf{1}_{d=4}})$, $\\epsilon_2(n_{\\mathrm{in}})=O(n_{\\mathrm{in}}^{-1/(3(d+1))})$, and $\\epsilon_3(\\Delta t)=O(\\sqrt{\\Delta t})$, the constants depending only on problem data and the activation function. In words, the proxy problem that one actually implements, a central planner minimizing the social cost of $N$ agents all using the same neural-network feedback law on a discrete time grid, yields a value that is no more than $\\epsilon$ below the true optimal cost. The proof decomposes the error into three manageable pieces, each tied to one approximation step; the same machinery yields an upper bound on the cost of the learned feedback control when it is evaluated in the original problem.","pith_inferences":["The theoretical bottleneck is the neural-network approximation of a function together with its derivatives; improving that step, for example with multilayer architectures that come with derivative-aware rates, would directly reduce $\\epsilon_2$ and soften the curse of dimensionality.","One could use the theorem as an empirical regularity test: run Algorithm 1 on problems with nonsmooth decoupling fields and compare the observed decay rates with the predicted ones, revealing where the smoothness assumptions are really needed.","The second method's reformulation of McKean-Vlasov FBSDEs as forward control problems suggests that convergence rates for the first method could be extended to full FBSDE solvers, giving mean field game algorithms the same kind of certificate once an analogue of Proposition 13 is proved for the shooting formulation.","The common-noise test cases indicate the techniques extend beyond the theorem's scope; a formal extension to conditional McKean-Vlasov dynamics would likely require measure-valued Lipschitz conditions of the same flavor as Assumptions B2-C2."],"forward_implications":["For any problem satisfying the regularity assumptions, running Algorithm 1 with sufficiently many agents, neurons, and time steps is guaranteed to produce a cost within an explicit distance of the exact mean field control optimum.","Because the method never discretizes the state space, the convergence proof is dimension-adaptive: the rate degrades only through the exponents $1/\\max(d,4)$ and $1/(3(d+1))$.","The finite-$N$ cost with a common feedback rule is a valid variational proxy: the gap between the $N$-agent value and the true mean field value is controlled by a rate that is logarithmic in $N$ when $d=4$ and polynomial otherwise.","The learned feedback control inherits a performance certificate: Remark 6 shows how to bound the difference between its cost in the proxy and the cost of the exact optimal feedback control.","The same three-step decomposition can serve as a template for future machine-learning-based schemes: separate the population error, the network approximation error, and the time-discretization error."],"supporting_citations":[{"why":"Supplies the propagation-of-chaos estimate that yields the finite-population error $\\epsilon_1(N)$ in Proposition 7.","marker":"[19]"},{"why":"Supplies the neural-network degree-of-approximation theorem used to approximate the optimal feedback control and its derivatives simultaneously.","marker":"[34]"},{"why":"Supplies the multivariate trigonometric approximation bound used to control the degree of approximation in Proposition 10.","marker":"[37]"},{"why":"Provides the FBSDE, master-field, and L-derivative machinery that grounds the decoupling field $V$ and the standing assumptions.","marker":"[18]"},{"why":"Identifies low-level conditions under which the decoupling field is Lipschitz and the standing assumptions hold.","marker":"[22]"}],"fun_headline_variants":["Neural feedback control gets provable error bounds","Finite-horizon mean field control: NN convergence analysis","Guaranteed approximation for neural control in mean field games","Rigorous error rates for machine learning in control","Convergence proven for neural network methods in control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof requires the value function's spatial derivative to be smooth, meaning Lipschitz with a Lipschitz second derivative, and, for the time-discretization part, the drift and volatility to be linear in state, mean, and control.","fun_headline_variants_meta":{"raw":{"variants":["Neural feedback control gets provable error bounds","Finite-horizon mean field control: NN convergence analysis","Guaranteed approximation for neural control in mean field games","Rigorous error rates for machine learning in control","Convergence proven for neural network methods in control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000248,"raw_usage":{"total_tokens":1596,"prompt_tokens":1044,"completion_tokens":552,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":660,"completion_tokens_details":{"reasoning_tokens":484}},"tokens_in":660,"tokens_out":552,"duration_ms":6085,"temperature":1.0,"reasoning_tokens":484,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:08:19.737118+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve a smooth linear-quadratic mean field control problem with known optimum, at dimension $d=4$, and measure the deficit $\\inf_\\alpha J(\\alpha) - \\inf_\\phi \\tilde J_N(\\phi)$ as $N$ doubles with $n_{\\mathrm{in}}$ and $\\Delta t$ fixed and large; if the deficit does not decay like $N^{-1/4}\\sqrt{\\ln N}$, the bound is wrong. In the other direction, running the same test with a nonsmooth terminal cost, such as the minimum of distances to two targets, tests whether the smoothness assumption is truly needed for the claimed rate.","supporting_citations":[{"cited_title":"II , volume 84 of Probability Theory and Stochastic Modelling","cited_arxiv_id":null,"evidence_quote":"Supplies the propagation-of-chaos estimate that yields the finite-population error $\\epsilon_1(N)$ in Proposition 7."},{"cited_title":"Mhaskar and Charles A","cited_arxiv_id":null,"evidence_quote":"Supplies the neural-network degree-of-approximation theorem used to approximate the optimal feedback control and its derivatives simultaneously."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the multivariate trigonometric approximation bound used to control the degree of approximation in Proposition 10."},{"cited_title":"I , volume 83 of Probability Theory and Stochastic Modelling","cited_arxiv_id":null,"evidence_quote":"Provides the FBSDE, master-field, and L-derivative machinery that grounds the decoupling field $V$ and the standing assumptions."},{"cited_title":"A Probabilistic approach to classical solutions of the master equation for large population equilibria","cited_arxiv_id":"1411.3009","evidence_quote":"Identifies low-level conditions under which the decoupling field is Lipschitz and the standing assumptions hold."}],"review_version":1}