{"id":"719c0865-5bd0-43c3-bb1b-d10e55a68b39","arxiv_id":"1908.03330","paper_version":3,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For mean field games with double-integrator dynamics, existence of weak solutions and a flow representation of the population density are proved despite a non-coercive Hamiltonian.","lead":"This paper proves that a class of multi-agent models, where each agent controls its acceleration rather than its velocity, has a well-defined solution when the number of agents is infinite. It also shows the crowd density evolves exactly by transporting the initial population along optimal trajectories.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: the existence and representation argument is internally coherent; the strong coupling regularity in (H3) is an explicit scope restriction, not a hidden flaw.","rationale":"The reader's weakest_assumption correctly identifies H3 as the point where the proof leans hardest. This is an explicit assumption, so it is a scope limitation rather than a correctness risk. I checked the potentially fragile steps: the stochastic estimates in Lemma 4.1 only need the L2 control bound, not deterministic Corollary 3.1; noise terms cancel in the semiconcavity computation; the superposition-principle representation is justified by optimal synthesis; the Schauder fixed point is standard, and the uniformity in sigma comes from H3 and from the compact support of m0. The only imprecision I noticed is that Definition 2.1 does not require u to be the value function, while Theorem 4.1 is stated for the value function; however the solution produced by the fixed point is the value-function solution, so the central existence-representation claim is unaffected. I therefore see no reason to change the reader's ACCEPT verdict.","tokens_in":26976,"tokens_out":47651,"duration_ms":503301,"concrete_test":"Worth verifying: re-derive Proposition 4.2's claim that the set (4.13) is a singleton for a.e. initial, explicitly combining Lemma 3.5.1 with Lemma 3.4.3 and checking that every absolutely continuous solution of (3.38)-(3.40) satisfies condition (3.39); if a solution not satisfying (3.39) existed, the representation formula would need an additional argument.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I read the proof as a chain: fixed m, value function regularity (Lemmas 3.2-3.3), vanishing-viscosity estimates uniform in sigma (Lemma 4.1), Fokker-Planck well-posedness and superposition representation (Theorem 4.1), then Schauder fixed point. The most load-bearing condition is (H3): Lipschitz maps from P1 to C^2(R^{2N}) and C^2-bounded l are what produce semiconcavity and the sigma-uniform bounds in Lemma 4.1; with local couplings the argument would not follow. This is exactly what the paper assumes, so it restricts scope without making the theorem wrong. The delicate uniqueness/representation step in Proposition 4.2 is supported by Lemma 3.5.1 (any ODE solution is optimal) together with Lemma 3.4.3 (unique optimal trajectory for differentiable initial data); the text compresses this but the ingredients are present. I found no circularity: Lemma 3.2's use of (3.33) from Corollary 3.1 is legitimate because Corollary 3.1 is proved from Pontryagin conditions independently. No fitted parameters, no hidden selection, and no missing proof that affects the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies deterministic mean field games with finite horizon in which each agent controls acceleration, so that the state is the pair (x,v) of position and velocity and the dynamics is the double integrator. The MFG system couples a backward Hamilton-Jacobi equation for the value function u with a forward continuity equation for the density m. The Hamiltonian is H(x,v,D_vu)=1/2|D_vu|^2-1/2|v|^2-l(x,v), which is neither strictly convex nor coercive in the full momentum variable. Under assumptions (H), in particular that the couplings F and G map P_1 into C^2(R^{2N}) with uniformly bounded C^2 norms and that m_0 is a compactly supported Hölder density, the authors prove the existence of a weak solution (Theorem 2.1) and characterize m(t) as the image of m_0 under the flow generated by x'=v, v'=-D_vu(x,v,t). They also prove uniqueness under a monotonicity condition on F and G (Proposition 2.1) and, in the appendix, existence and uniqueness for the corresponding second-order MFG system.","tokens_in":27189,"tokens_out":22128,"duration_ms":217969,"significance":"If correct, this is the first existence theory for deterministic MFGs with acceleration control and a non-coercive Hamiltonian, and the representation of the population density by the optimal flow is a genuinely useful structural result. The proof is a long but coherent chain: Pontryagin maximum principle and optimal synthesis (Section 3), semiconcavity, vanishing viscosity and Fokker-Planck estimates (Section 4), and a Schauder fixed point argument (Section 5). I checked the potentially delicate steps: the use of estimate (3.33) in Lemma 3.2 is not circular because Corollary 3.1 is proved independently from the Pontryagin conditions; the uniqueness step in Proposition 4.2 is supported by Lemma 3.5(1) together with Lemma 3.4(3); and the strong regularity of F and G in (H3) is an explicit scope restriction rather than a hidden flaw. The paper contains no fitted parameters or circular normalizations, and the claims are falsifiable through the stated assumptions. The main limitation is that local couplings such as F[m]=f(m) are excluded, but the authors state this clearly through (H3).","major_comments":[],"minor_comments":[{"comment":"The proof of Lemma 3.2 uses estimate (3.33), which is proved later in Corollary 3.1. This is not circular, since Corollary 3.1 follows from the Pontryagin maximum principle independently of Lemma 3.2, but the forward reference should be flagged explicitly so the reader does not perceive a circularity.","section":"Section 3, Lemma 3.2"},{"comment":"The sentence 'since for all t, u(·,·,t) is Lipschitz continuous' overstates the regularity proved in Lemma 3.2: u is Lipschitz in x and only locally Lipschitz in v. For the conclusion that the set in (4.13) is a singleton for almost every initial condition, one should explicitly invoke the semiconcavity of u from Lemma 3.3 together with Rademacher's theorem to obtain a.e. differentiability of u(·,·,0).","section":"Section 4.2, Proposition 4.2"},{"comment":"The claim that u=u_1-u_2 is an admissible test function for the continuity equation should be justified by a brief density argument: u is only locally Lipschitz, while the weak formulation in Definition 2.1 uses smooth compactly supported test functions. The compact support of m_i together with the local Lipschitz regularity of u makes this standard, but the step is currently implicit.","section":"Section 5, Proposition 2.1"},{"comment":"Equation (6.3) is hard to parse as displayed: the drift term in the first weak formulation should involve D_vu_i, the coefficient corresponding to the equation for m_i, rather than D_vu without an index; the passage to the second equality and the subsequent subtraction step should be written out with explicit indices.","section":"Section 6, equation (6.3)"},{"comment":"There is a typo in the display preceding (3.41): the term 'u(x(s), x(s), s)' should read 'u(x(s), v(s), s)'.","section":"Section 3, Lemma 3.5"}],"recommendation":"minor_revision","confidential_remarks":"I agree with the reader's positive assessment: the central existence and representation theorem is sound within the stated assumptions, and I see no grounds for rejection. The requested changes are clarifications and presentational fixes. The paper is a good fit for math.AP, and the authors are appropriately credited in the acknowledgments and references for the independent related work."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is the first existence theorem for deterministic MFGs where agents control acceleration (the double integrator), and the Hamiltonian is neither strictly convex nor coercive. The proof chain is long but coherent; I could not find a genuine gap. It deserves a serious referee.\n\nWhat is actually new: previous MFG theory couldn't touch this setting because the Hamiltonian has flat directions and the running cost is unbounded in velocity. The value function is only locally Lipschitz in v, so the drift in the continuity equation is not Lipschitz and standard flow arguments fail. The authors get around that with a vanishing viscosity limit, sigma-uniform estimates, the Ambrosio-Gigli-Savare superposition principle, and a Schauder fixed point. On top of existence, they prove that m is the image of m0 under the flow of the optimal feedback. That representation result is a useful byproduct, not a throwaway.\n\nCredit where it is due: each step is anchored in cited theorems—Da Lio-Ley comparison for quadratic-growth HJ, Pontryagin, the non-coercive control synthesis from their own [29], and the superposition principle. The use of [29] is legitimate; it supplies lemmas, not the main result. The independent simultaneous work by Cannarsa and Mendico is acknowledged. The proof is careful, and the appendix gets the second-order system for free from the same estimates.\n\nSoft spots: the main one is assumption (H3). F and G have to map P1 into C^2(R^{2N}) with uniform C^2 bounds, and l has to be C^2-bounded. That is a strong regularizing assumption and rules out local couplings like F[m]=f(m(x,v)), which is the most common case in MFG applications. It is an explicit scope restriction, not a hidden flaw, but it does mean the theorem is a foundation for nonlocal/regularized couplings, not the final word. A second, smaller issue: the uniqueness part of Proposition 4.2 is compressed. The argument needs that for a.e. initial (x,v) the optimal trajectory is unique, which follows from Lemma 3.5 plus Lemma 3.4(3), but the text doesn't spell that out. A referee should ask for a sentence or two there. Also, Lemma 3.2 refers forward to a bound in Corollary 3.1; that looks odd but is not circular, since Corollary 3.1 is proved independently.\n\nWho it is for: people working on MFG with higher-order dynamics, constrained MFG, or non-coercive Hamiltonians. It is a solid technical foundation and likely to be built on. It won't reshape the field, but it fills a real gap.\n\nRecommendation: send it to peer review. I would sign an accept after minor revisions, mainly asking for clarification in Proposition 4.2.","headline":"First existence theorem for deterministic MFGs with acceleration control; long, coherent proof chain and an honest scope restriction; deserves a serious referee.","tokens_in":27720,"tokens_out":6967,"would_cite":true,"duration_ms":65086,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35F50","35Q91","49K20","49L25"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves existence of weak solutions for acceleration-controlled deterministic mean field games and shows the density is transported by the optimal flow.","keywords":["mean field games","control on acceleration","double integrator","non-coercive Hamiltonian","first-order Hamilton-Jacobi equations","continuity equation","weak solutions","vanishing viscosity"],"falsifier":"For a one-dimensional double integrator with a smooth compactly supported $m_0$ and a convolution coupling $F[m](x,v)=\\int\\rho(x-x',v-v')\\,dm(x',v')$ with smooth compactly supported $\\rho$, compute the value function in (3.5) numerically and check the semiconcavity inequality $\\lambda u(x,v,t)+(1-\\lambda)u(y,w,t)-u(x_\\lambda,v_\\lambda,t)\\le C\\lambda(1-\\lambda)(|x-y|^2+|v-w|^2)$ with $C$ independent of the viscosity parameter. A violation would contradict Lemma 3.3, on which the convergence of $D_v u_\\sigma\\to D_v u$ rests, and would therefore break the proof of Theorem 2.1.","tokens_in":26798,"feed_emoji":"","tokens_out":9702,"duration_ms":90683,"temperature":0.7,"pith_summary":"This paper tackles deterministic mean field games in which each agent controls acceleration rather than velocity, so the state is a position-velocity pair and the dynamics is the double integrator. The associated Hamiltonian is neither strictly convex nor coercive, and the running cost grows quadratically in velocity, so off-the-shelf existence results do not apply. The authors prove that the coupled Hamilton-Jacobi and continuity equations still have a weak solution, and that the population density is the image of the initial density under the flow of the optimal feedback, $x'(s)=v(s)$, $v'(s)=-D_v u(x(s),v(s),s)$. The proof works by adding viscosity to both equations, deriving estimates uniformly in the viscosity parameter, and passing to the limit, then closing the loop with a fixed-point argument. A byproduct is existence and uniqueness for the corresponding second-order (viscous) system.","feed_headline":"Acceleration-controlled crowds have a mean-field solution","feed_subtitle":"Even with a non-coercive Hamiltonian, the density follows the optimal feedback flow.","key_machinery":"The central object is the value function $u$ of a representative agent's optimal control problem with the population density fixed, together with its optimal synthesis: a trajectory $s\\mapsto(x(s),v(s))$ is optimal exactly when it solves the characteristic ODE $x'(s)=v(s)$, $v'(s)=-D_v u(x(s),v(s),s)$, with the optimal acceleration given by the feedback $\\alpha^*(s)=-D_v u(x(s),v(s),s)$. The proof's technical engine is the vanishing viscosity method: adding $\\sigma\\Delta_{x,v}$ to both the Hamilton-Jacobi and the continuity equation makes the system classical, and uniform estimates on $u_\\sigma$ (Lipschitz in $x$, locally Lipschitz in $v$, semiconcave) and on $m_\\sigma$ ($L^\\infty$ bound, $C^{1/2}$ in time, uniform second-moment bound) allow a limit $\\sigma\\to0^+$ that preserves the viscosity and distributional equations. The representation formula then follows from the superposition principle, which writes $m$ as the image of a measure on continuous curves and identifies the curves with the unique optimal trajectories of the control problem. The fixed-point map $m\\mapsto T(m)$—solve the Hamilton-Jacobi equation with $m$ frozen, then solve the continuity equation with the resulting $D_v u$—is continuous and compact, so existence follows by a fixed-point argument.","core_discovery":"On the paper's own terms, the core claim is Theorem 2.1: under assumptions (H), the system (2.1) has a weak solution $(u,m)$ in the sense of Definition 2.1, with $u$ a viscosity solution of the backward Hamilton-Jacobi equation and $m$ a distributional solution of the forward continuity equation. The velocity variable $v$ appears in the state, and the Hamiltonian $H(x,v,p_v)=\\frac12|p_v|^2-\\frac12|v|^2-l(x,v)$ depends on the momentum only through $p_v$, which is why it is not coercive or strictly convex in the full momentum. The second part of the theorem is the representation formula: $m$ is the image of the initial distribution $m_0$ under the flow $x'=v$, $v'=-D_v u(x,v,t)$, so the aggregate density is exactly what is obtained by letting every agent follow the optimal feedback. The proof of this representation combines the optimal synthesis of the control problem with the superposition principle for the continuity equation; a fixed-point argument on the map that sends a density to the continuity-equation solution generated by its optimal value function gives existence.","pith_inferences":["If Theorem 2.1 is right, a numerical scheme built on the characteristic flow should converge to the weak solution even when individual optimal trajectories are not unique for every initial condition, because the representation needs the flow only $m_0$-almost everywhere.","The same proof route is likely to work for other non-coercive dynamics, such as controls entering only some velocity components or forbidden directions, as long as the value function remains semiconcave; the proof relies on that structure rather than on the specific double-integrator form.","If one replaces the global $C^2$ couplings by local ones such as $F[m]=f(m)$, the uniform semiconcavity estimates fail and the vanishing viscosity limit may select a different or no solution; testing this numerically in one dimension is a quick way to map the boundary of the method.","The representation formula suggests that macroscopic quantities such as moments of $m$ are determined by the value function's gradient, so a natural next step is to prove quantitative stability of these moments with respect to perturbations of the coupling operators $F$ and $G$."],"forward_implications":["The representation formula gives a Lagrangian description of the crowd: the macroscopic density is obtained by pushing the initial density forward along the optimal feedback flow, so particle simulations along $x'=v$, $v'=-D_v u$ are faithful at the mean-field limit.","The solution $m$ has a density with uniform $L^\\infty$ and second-moment bounds and is $C^{1/2}$ in time with values in the space of probability measures with finite first moment, so the aggregate state distribution is a well-behaved probability measure at all times.","Under the monotonicity conditions (2.3), the weak solution of the first-order system is unique; this is the classical monotonicity uniqueness mechanism adapted to the acceleration-controlled setting.","The vanishing viscosity construction also yields a classical solution of the corresponding second-order (viscous) mean field game system, and that solution is unique when the same monotonicity conditions hold.","The regularity results imply that along optimal trajectories the optimal control is a $C^1$ feedback, $\\alpha^*(s)=-D_v u(x(s),v(s),s)$, so the characteristic flow is well defined for almost every initial condition."],"supporting_citations":[{"why":"Supplies the superposition principle that represents the solution of the continuity equation as the image of a measure on continuous curves.","marker":"[2]"},{"why":"Provides the semiconcavity and reachable-gradient machinery used to extract uniform gradient convergence from the value-function estimates.","marker":"[11]"},{"why":"Supplies the optimal-synthesis method and the estimates later adapted to the non-coercive setting of the double integrator.","marker":"[12]"},{"why":"Provides the vanishing-viscosity compactness argument used to pass to the limit in the continuity equation.","marker":"[13]"},{"why":"Shows uniqueness of the continuity equation and the push-forward representation via the superposition principle.","marker":"[15]"},{"why":"Supplies the maximum-principle form for optimal control problems with unbounded control values, used for the feedback formula.","marker":"[18]"},{"why":"Provides the comparison principle for Hamilton-Jacobi equations with quadratic growth, used for uniqueness of the value function.","marker":"[19]"},{"why":"Gives the monotonicity argument that yields uniqueness of the mean field game solution.","marker":"[26]"},{"why":"Treats a related non-coercive first-order mean field game and supplies the dynamic programming arguments adapted here.","marker":"[29]"}],"fun_headline_variants":["Mean-field games with acceleration control: existence via vanishing viscosity","Non-coercive Hamiltonians still yield weak MFG solutions","Acceleration-driven crowds: density follows optimal flow","Existence for deterministic MFGs with acceleration and unbounded velocity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument needs the coupling terms to smooth the population density into globally $C^2$ functions with bounded second derivatives, uniformly in the density; if the coupling only depends on the density pointwise, say $F[m]=f(m)$, the uniform estimates that drive the proof are no longer available.","fun_headline_variants_meta":{"raw":{"variants":["Mean-field games with acceleration control: existence via vanishing viscosity","Non-coercive Hamiltonians still yield weak MFG solutions","Acceleration-driven crowds: density follows optimal flow","Existence for deterministic MFGs with acceleration and unbounded velocity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1369,"prompt_tokens":971,"completion_tokens":398,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":331}},"tokens_in":587,"tokens_out":398,"duration_ms":4999,"temperature":1.0,"reasoning_tokens":331,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:17:29.925438+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a one-dimensional double integrator with a smooth compactly supported $m_0$ and a convolution coupling $F[m](x,v)=\\int\\rho(x-x',v-v')\\,dm(x',v')$ with smooth compactly supported $\\rho$, compute the value function in (3.5) numerically and check the semiconcavity inequality $\\lambda u(x,v,t)+(1-\\lambda)u(y,w,t)-u(x_\\lambda,v_\\lambda,t)\\le C\\lambda(1-\\lambda)(|x-y|^2+|v-w|^2)$ with $C$ independent of the viscosity parameter. A violation would contradict Lemma 3.3, on which the convergence of $D_v u_\\sigma\\to D_v u$ rests, and would therefore break the proof of Theorem 2.1.","supporting_citations":[{"cited_title":"Ambrosio, N","cited_arxiv_id":null,"evidence_quote":"Supplies the superposition principle that represents the solution of the continuity equation as the image of a measure on continuous curves."},{"cited_title":"Cannarsa, C","cited_arxiv_id":null,"evidence_quote":"Provides the semiconcavity and reachable-gradient machinery used to extract uniform gradient convergence from the value-function estimates."},{"cited_title":"Cardaliaguet, Notes on Mean Field Games , from P.L","cited_arxiv_id":null,"evidence_quote":"Supplies the optimal-synthesis method and the estimates later adapted to the non-coercive setting of the double integrator."},{"cited_title":"Cardaliaguet, Long time average of first order mean field games and weak KAM theory, Dyn","cited_arxiv_id":null,"evidence_quote":"Provides the vanishing-viscosity compactness argument used to pass to the limit in the continuity equation."},{"cited_title":"Cardaliaguet, S","cited_arxiv_id":null,"evidence_quote":"Shows uniqueness of the continuity equation and the push-forward representation via the superposition principle."},{"cited_title":"Clarke, Functional Analysis, Calculus of Variations and Optimal Co ntrol , Graduate Text in Mathematics 264, Springer-Verlag, London 2013","cited_arxiv_id":null,"evidence_quote":"Supplies the maximum-principle form for optimal control problems with unbounded control values, used for the feedback formula."},{"cited_title":"Da Lio, O","cited_arxiv_id":null,"evidence_quote":"Provides the comparison principle for Hamilton-Jacobi equations with quadratic growth, used for uniqueness of the value function."},{"cited_title":"Lasry and P-L","cited_arxiv_id":null,"evidence_quote":"Gives the monotonicity argument that yields uniqueness of the mean field game solution."},{"cited_title":"Mannucci, C","cited_arxiv_id":null,"evidence_quote":"Treats a related non-coercive first-order mean field game and supplies the dynamic programming arguments adapted here."}],"review_version":1}