{"id":"e6dcd6fa-4344-4779-b393-d63406467e4c","arxiv_id":"2412.03308","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"The claimed existence of relaxed mean field control equilibria for non-convex first-order problems is invalid as stated.","lead":"Mean field control problems with non-convex costs are attacked by relaxing controls to probability measures and proving existence of a new equilibrium. The main existence theorem fails under the paper's own assumptions, and a counterexample with a simple Lagrangian shows the relaxed minimization can have no solution.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reader's negative-L counterexample does not land; the load-bearing defect is Proposition 2.4's use of optimality of P_i against a limit-feasible measure that need not lie in P_R(m_i).","rationale":"I agree that the manuscript is not acceptable as written, but not for the precise reason given by the reader. The stated negative-L counterexample does not falsify Theorem 1: Lemma 2.4(2)'s factor-dropping step is wrong when L is negative, yet the lower semicontinuity conclusion is true by continuity, polynomial growth, and the uniform q-th moment bound on controls. The genuinely load-bearing gap is in Proposition 2.4, where optimality of P_i in P_R(m_i) is applied to a measure hat P in P_R(m) that is not generally feasible for m_i. This breaks the closed-graph proof on which Kakutani's theorem depends. The gap appears repairable by explicitly approximating each limit-feasible hat P with m_i-feasible measures obtained by re-solving the state dynamics with m_i, and by checking cost convergence; I would therefore not assert the theorem is false, but the current proof is incomplete. This moves the verdict from an outright REJECT based on a faulty counterexample to a CONDITIONAL stance requiring a corrected Proposition 2.4. Secondary issues, such as passing integrals over open sets under narrow convergence, are less central than this missing Mosco-type comparison of the feasible sets P_R(m_i).","tokens_in":20620,"tokens_out":25160,"duration_ms":268425,"concrete_test":"Fix data satisfying (L1)-(L4) and (F1)-(F3), take m_i -> m uniformly, and take hat P in P_R(m). For each (gamma, mu) in the support of hat P, solve gamma_i(t) = gamma(0) + integral_0^t integral f(gamma_i(s), u, m_i(s)) mu(ds,du), and set hat P_i to be the pushforward of hat P under (gamma, mu) -> (gamma_i, mu). Verify that hat P_i in P_R(m_i), that e0-sharp pi1-sharp hat P_i = m0, and that J(m_i, hat P_i) -> J(m, hat P). If these three properties hold, the invalid comparison in Proposition 2.4 can be repaired; if any of them fails for some admissible data, Theorem 1 is unproved as stated, and the reader's negative-L counterexample should still be withdrawn.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Theorem 1 rests on Kakutani applied to E(eta) = pi1-sharp of R*((et-sharp eta)_t), so Proposition 2.4 must establish the closed graph of E. In the final comparison, the proof takes an arbitrary hat P in P_R(m) and uses the optimality of P_i in R*(m_i) to assert J(m_i, P_i) <= J(m_i, hat P). This is valid only if hat P belongs to P_R(m_i). But hat P is only known to satisfy the consistency condition (1.7) for the limit distribution m; when f depends on m, hat P generally fails condition (1.7) at the level m_i. The inequality is therefore unsupported, and without it the proof does not rule out hat eta notin E(eta). This is the load-bearing step of the fixed-point argument. The reader's alternative objection is not a valid counterexample: Lemma 2.4(2)'s displayed inequality does implicitly require L >= 0, but the claimed lower semicontinuity is independently true because L is continuous, |L| <= C(1+|u|^q), and measures in P_R_U have uniformly bounded q-th moments, so tightness gives convergence. Hence the negative-L example does not make R*(m) empty.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a relaxed Lagrangian formulation for first-order mean-field-type control problems with non-convex Lagrangians. For a distribution m, the authors define P_R(m), the set of probability measures on curve–relaxed-control pairs whose support satisfies the state equation driven by m, and R*(m), the set of minimizers of the total cost J(m,P) over P_R(m). A relaxed MFC equilibrium is a measure P satisfying P∈R*((e_t#π1#P)_t). The main result, Theorem 1, asserts the existence of such an equilibrium under assumptions (L1)–(L4) and (F1)–(F3), proved by applying Kakutani's fixed-point theorem to the set-valued map E(η)=π1#R*((e_t#η)_t). Theorem 2 states that under a pointwise convexity condition there exists a strict relaxed equilibrium, recovering in particular the classical MFC existence result. The paper also motivates the framework through residual neural network training.","tokens_in":20819,"tokens_out":24379,"duration_ms":227477,"significance":"If the arguments are correct, Theorem 1 would be a useful extension of mean-field-type control existence theory to non-convex Lagrangians, a setting for which the authors note no direct first-order results are available. The relaxed-control construction on the Wasserstein space is natural, and the compactness apparatus for P_R_U and Γ_R^T is mostly well chosen. The paper is self-contained, uses standard measure-theoretic tools, and does not appear to rely on circular reasoning or fitted parameters. However, the proof as written contains a load-bearing gap in Proposition 2.4: the comparison that establishes the closed graph of E is only valid for competitors feasible for m_i, while the proof applies it to competitors feasible only for the limit m. This gap is local and, in my judgment, repairable as described in the major comments. The consequence is that the manuscript is not acceptable in its present form, but the central claim is defensible.","major_comments":[{"comment":"The final displayed chain in the proof of Proposition 2.4 uses the inequality J(m_i,P_i) ≤ J(m_i,\\hat P) for an arbitrary \\hat P ∈ P_R(m). Since P_i is only known to minimize over P_R(m_i), this comparison requires \\hat P ∈ P_R(m_i). The consistency condition (1.7) is formulated with the distribution m, and \\hat P ∈ P_R(m) does not imply \\hat P ∈ P_R(m_i) when f depends on the measure argument; convergence m_i → m does not make the two feasible sets coincide. This step is load-bearing because it is exactly what yields the closed graph of E, which is needed for the Kakutani fixed-point argument in Theorem 1. The gap is repairable: for a fixed \\hat P ∈ P_R(m), define T_i(γ,μ)=(γ_i,μ), where γ_i solves the state equation (1.4) with m_i and γ_i(0)=γ(0). Then \\hat P_i = T_i#\\hat P ∈ P_R(m_i), \\hat P_i → \\hat P, and J(m_i,\\hat P_i) → J(m,\\hat P), so the argument can be completed with \\hat P replaced by \\hat P_i.","section":"Section 2.2, Proposition 2.4"},{"comment":"The measure h#η̃ is not well-defined as written. The map h is introduced only on the set {γ̃_x : x∈T^d} by γ̃_x ↦ (γ̃_x, μ_x), but no measurable selection of the associated optimal controls μ_x is specified. Without such a selection, h need not be a Borel map and h#η̃ may not exist as a push-forward. Since Γ*_m(x) is compact-valued and has closed graph, a measurable selection of optimal controls can be obtained by standard arguments, so this appears to be a technical gap rather than a fatal flaw.","section":"Section 2.2, Proposition 2.3"}],"minor_comments":[{"comment":"The proof's displayed inequality L = L(1+ε|u|^q)/(1+ε|u|^q) ≥ L/(1+ε|u|^q) requires L ≥ 0, which is not among assumptions (L1)–(L4). The claimed lower semicontinuity is nevertheless true: L is continuous, |L| ≤ C(1+|u|^q), and measures in P_R_U have uniformly bounded q-th moments, so Proposition B.2(3) gives ∫ L dμ_i → ∫ L dμ. The proof should be corrected; in particular, the negative-L example sometimes cited against this lemma does not make R*(m) empty.","section":"Lemma 2.4(2)"},{"comment":"Proposition 2.1 has only parts (1)–(3), but Proposition 2.3 and the proof of Theorem 2 refer to Proposition 2.1(4). These references should be to Proposition 2.1(3).","section":"Sections 2 and 3"},{"comment":"The proof establishes only m0-a.e. optimality of the family {u_x^*} from the aggregate inequality ∫[...]m0(dx), not pointwise optimality for every x. If the statement 'for any x ∈ T^d' is intended to mean individual optimality at every x, an additional argument is needed; as written, the proof does not support that stronger reading.","section":"Theorem 2(1)"},{"comment":"The non-emptiness of Γ*_m(x) is dismissed with 'easily proved by convexity'; since J^m is linear in μ, the existence of a minimizer follows from compactness of P_R_U and continuity of μ ↦ J^m(γ(·;x,μ,m),μ). The argument should be stated explicitly rather than deferred.","section":"Proposition 2.2(1)"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of the journal and addresses an interesting problem. The central existence theorem is plausible, and the gap in Proposition 2.4 appears repairable by the approximation construction described in the major comments. The negative-L objection to Lemma 2.4(2) does not actually invalidate the lemma, because the lower semicontinuity follows from uniform q-th moment bounds rather than from the displayed (incorrect) inequality. I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core new content is solid: the definition of relaxed MFC equilibrium for first-order non-convex mean field type control, the strict version under a convexity condition, and the connection to residual neural network training. The paper is clearly organized, the assumptions are standard, and the proof strategy is coherent: compactness from moment bounds, a fixed point on the space of state-trajectory laws, and a convexity-enforced strictification step in Theorem 2. That part deserves credit and could be a useful contribution to the literature.\n\nThe soft spot is in Proposition 2.4, the closed-graph argument for the set-valued map E. In the final comparison, the proof takes an arbitrary hat-P in P_R(m), the limit feasibility set, and uses optimality of P_i in R*(m_i) to claim J(m_i, P_i) <= J(m_i, hat-P). That only works if hat-P belongs to P_R(m_i). It generally does not: hat-P satisfies the state consistency condition for m, not for m_i, and when f depends on the measure the condition fails at the level of m_i. So that inequality is unsupported, and without it the closed graph is not established. This is load-bearing for Theorem 1.\n\nThe reader's proposed counterexample, however, does not land. The inequality in Lemma 2.4(2) is written with a sign error that implicitly assumes L >= 0, but the claimed lower semicontinuity is independently true because L is continuous with |L| bounded by C(1+|u|^q) and the measures in P_R_U have uniformly bounded q-th moments; tightness gives the convergence. The reader's Lagrangian L(u) = -|u|^2 + |u|^2/(1+|u|) with f = u does not make R*(m) empty; the functional is bounded below on P_R_U and the infimum is attained by compactness. So that specific objection should be withdrawn.\n\nThe Proposition 2.4 gap looks repairable: one can likely push a limit-feasible measure hat-P forward through the map that replaces each state curve by the solution of the state equation with the same initial point and control but with m_i in place of m. That would produce hat-P_i in P_R(m_i) with hat-P_i -> hat-P. The paper does not do this, so as written the proof is incomplete.\n\nBottom line: this is a serious paper with a significant but fixable gap in the main fixed-point argument. It deserves a serious referee, not a desk reject. I would send it out and ask for a corrected proof of Proposition 2.4. If the authors can make that step work, the paper is likely acceptable.","headline":"A genuinely interesting relaxed-MFC existence paper with a real gap in the final fixed-point argument, though the negative-L counterexample in the reader's report does not stand up.","tokens_in":21395,"tokens_out":7269,"would_cite":false,"duration_ms":69104,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49N80","49N90","68T07","93C10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Relaxed controls give non-convex mean field control a fixed-point equilibrium","keywords":["Mean Field Control","Relaxed Lagrangian Approach","Relaxed controls","Residual Neural Networks","Non-convex optimal control","Wasserstein space","Existence of equilibria","Kakutani fixed point"],"falsifier":"Take $L(x,u,\\nu) = -1/(1+|u|^q)$ (which satisfies (L1)–(L4)) with a drift $f$ satisfying (F1)–(F3), and compute whether the inequality $\\liminf_i J^m(\\gamma,\\mu_i) \\ge J^m(\\gamma,\\mu)$ in Lemma 2.4(2) still holds; if for some $m\\in M_r$ the argmin set $R^*(m)$ is empty, then the existence theorem as stated is not valid.","tokens_in":20339,"feed_emoji":"🎯","tokens_out":9522,"duration_ms":79710,"temperature":0.7,"pith_summary":"This paper aims to prove that first-order mean field type control (MFC) problems with non-convex running costs still admit equilibria, provided controls are allowed to be randomized. The proposed 'relaxed Lagrangian approach' replaces pointwise controls by probability measures over the control space, turning the non-convex minimization into a linear-in-measure problem on the Wasserstein space. Under smoothness and controlled-growth assumptions on the Lagrangian and drift, the authors claim there always exists a relaxed MFC equilibrium — a probability law over state trajectories and control measures that is optimal for its own induced population distribution. They also show that when the data satisfy a classical convexity condition, the relaxed equilibrium can be written back as a classical feedback control, recovering the standard MFC existence result. A sympathetic reader would care because the result would extend mean field control theory to settings where convexity fails, such as mean-field formulations of neural network training.","feed_headline":"Non-convex mean field control gets relaxed equilibria","feed_subtitle":"A measure-valued control trick turns the non-convex problem into a fixed-point search with a guaranteed solution.","key_machinery":"The central object is the set $P^R_U$ of relaxed controls: probability measures $\\mu = dt\\otimes \\mu_t$ on $[0,T]\\times\\mathbb{R}^n$ with $\\int |u|^q \\mu(dt,du) \\le R$. The state equation becomes $\\dot\\gamma(t)=\\int f(\\gamma(t),u,m(t))\\,\\mu_t(du)$, and the cost $J^m(\\gamma,\\mu)=\\int L(\\gamma(t),u,m(t))\\,\\mu(dt,du)$ is linear in $\\mu$, so convexity in the control is no longer needed. The proof is carried by the set-valued map $E(\\eta)=\\{\\pi_{1\\sharp}P : P\\in R^*(m)\\}$ with $m(t)=e_{t\\sharp}\\eta$; the load-bearing results are that $E$ has non-empty, convex, compact values and a closed graph, obtained via the lower semicontinuity of $J^m$ in $\\mu$, and then Kakutani's fixed-point theorem yields $\\bar\\eta\\in E(\\bar\\eta)$, hence the equilibrium $P$.","core_discovery":"The paper's central claim is that under assumptions (L1)–(L4) and (F1)–(F3) — smoothness, controlled growth, and Lipschitz dependence on the measure — there exists at least one relaxed MFC equilibrium for the first-order mean field type control problem, even when the running cost $L(x,u,m)$ is non-convex in the control $u$. The relaxation consists in taking controls to be probability measures $\\mu\\in P^R_U$ on $[0,T]\\times\\mathbb{R}^n$ with bounded $q$-th moment, so the state equation and the cost become linear in the control variable. An equilibrium is a joint law $P$ of state trajectories $\\gamma$ and relaxed controls $\\mu$ that minimizes the total cost $J(m,P)$ for its own induced distribution $m(t)=e_{t\\sharp}\\pi_{1\\sharp}P$. The proof models this as a fixed point of the set-valued map $E(\\eta)=\\{\\pi_{1\\sharp}P : P\\in R^*(m)\\}$ with $m(t)=e_{t\\sharp}\\eta$, shows $E$ is non-empty, convex, compact-valued and has closed graph, and applies Kakutani's theorem. Under an additional convexity condition on the epigraph set $\\mathcal{L}(t,x)$, the same machinery produces a strict relaxed equilibrium and recovers the classical MFC existence theorem.","pith_inferences":["If the existence theorem is correct, the relaxed MFC equilibrium can be interpreted as a mixed-strategy Nash equilibrium of a mean field game, suggesting a connection between the relaxed Lagrangian approach and randomized strategies in multi-agent reinforcement learning.","The nonnegativity used in the lower-semicontinuity step suggests the theorem likely needs an explicit lower bound on $L$, and the relaxation method may still be applicable under such a sign condition even where pointwise convexity fails.","The same relaxation on the Wasserstein space could be extended to second-order or stochastic MFC problems, replacing the pathwise Lagrangian by a relaxed control over the control space in a McKean-Vlasov dynamics.","A numerical test could check whether the relaxed equilibrium coincides with the classical equilibrium in the convex case and whether the gap between them measures the cost of non-convexity."],"forward_implications":["Every first-order MFC problem with $C^2$, controlled-growth, possibly non-convex data admits a relaxed MFC equilibrium in the sense of Definition 1.2.","When the epigraph set $\\mathcal{L}(t,x)$ is convex, the relaxed equilibrium can be refined to a strict relaxed equilibrium, and the relaxed problem reduces to the classical MFC optimal control problem, so the result generalizes the existing convex theory.","The relaxed equilibrium carries all the information needed to design an optimal neural-network architecture in the mean-field training formulation: the parameters are read off from the second marginal of the equilibrium.","The fixed-point structure gives an algorithmic route: any numerical scheme that approximates the set-valued map $E$ and computes a fixed point would produce an approximate relaxed equilibrium."],"supporting_citations":[{"why":"Supplies the relaxed-control method: replacing non-convex control sets by probability measures over controls to recover existence.","marker":"[4]"},{"why":"Gives the convexity/compactness argument used to prove non-emptiness of the optimal-curve set $\\Gamma^*_m(x)$.","marker":"[10]"},{"why":"Provides the mean field game framework, the Kantorovich-Rubinstein distance $d_1$, and the disintegration theorem used to disintegrate relaxed controls.","marker":"[11]"},{"why":"Provides the measurable selection theorem used to pick an optimal curve $\\tilde\\gamma_x\\in\\Gamma^*_m(x)$ for each initial state.","marker":"[16]"},{"why":"Supplies the convexification result (Theorem A.9) that turns the relaxed control into a classical feedback control in Theorem 2.","marker":"[21]"},{"why":"Used to establish measurability of the set-valued map $F_m$ that underlies the measurable selection in Proposition 2.3.","marker":"[8]"}],"fun_headline_variants":["Relaxed controls crack non-convex mean field games","Non-convex MFC solved via measure-valued controls","Kakutani fixed point guarantees relaxed equilibria","Convexity bypassed: new existence proof for MFC","Relaxed L1 trick: non-convex control gets fixed points"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The existence proof needs the running cost $L$ to be nonnegative at the step where the factor $1+\\varepsilon|u|^q$ is dropped from a lower-semicontinuity estimate; the stated assumptions only bound $|L|$, not its sign, so if $L$ can go negative the argument breaks.","fun_headline_variants_meta":{"raw":{"variants":["Relaxed controls crack non-convex mean field games","Non-convex MFC solved via measure-valued controls","Kakutani fixed point guarantees relaxed equilibria","Convexity bypassed: new existence proof for MFC","Relaxed L1 trick: non-convex control gets fixed points"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000151,"raw_usage":{"total_tokens":1168,"prompt_tokens":881,"completion_tokens":287,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":497,"completion_tokens_details":{"reasoning_tokens":205}},"tokens_in":497,"tokens_out":287,"duration_ms":3451,"temperature":1.0,"reasoning_tokens":205,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:34:25.387607+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $L(x,u,\\nu) = -1/(1+|u|^q)$ (which satisfies (L1)–(L4)) with a drift $f$ satisfying (F1)–(F3), and compute whether the inequality $\\liminf_i J^m(\\gamma,\\mu_i) \\ge J^m(\\gamma,\\mu)$ in Lemma 2.4(2) still holds; if for some $m\\in M_r$ the argmin set $R^*(m)$ is empty, then the existence theorem as stated is not valid.","supporting_citations":[{"cited_title":"Buttazzo","cited_arxiv_id":null,"evidence_quote":"Supplies the relaxed-control method: replacing non-convex control sets by probability measures over controls to recover existence."},{"cited_title":"Cannarsa and C","cited_arxiv_id":null,"evidence_quote":"Gives the convexity/compactness argument used to prove non-emptiness of the optimal-curve set $\\Gamma^*_m(x)$."},{"cited_title":"Cardaliaguet.Notes on mean field game","cited_arxiv_id":null,"evidence_quote":"Provides the mean field game framework, the Kantorovich-Rubinstein distance $d_1$, and the disintegration theorem used to disintegrate relaxed controls."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the measurable selection theorem used to pick an optimal curve $\\tilde\\gamma_x\\in\\Gamma^*_m(x)$ for each initial state."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the convexification result (Theorem A.9) that turns the relaxed control into a classical feedback control in Theorem 2."},{"cited_title":"Cannarsa and T","cited_arxiv_id":null,"evidence_quote":"Used to establish measurability of the set-valued map $F_m$ that underlies the measurable selection in Proposition 2.3."}],"review_version":1}