{"id":"54b5568b-cb9c-4427-bafe-786776f9b9af","arxiv_id":"1908.06485","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For first-order stationary mean-field games with quadratic Hamiltonian, the vanishing-discount limit is characterized by a Mather-type selection criterion and shown to converge to a unique solution under stated conditions.","lead":"The paper studies the vanishing-discount limit of a first-order stationary mean-field game and proves that, under growth conditions, the limit is unique and is selected by a Mather-measure criterion. It matters because this is the first selection rule for first-order stationary mean-field games, a setting where multiple limiting solutions can exist.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Prop. 3.4 is the load-bearing weak point: inequality (3.8) requires β+1>0 although Assumption 2 allows β≤−1, so Theorem 1.1 and Corollary 1.2 are not proved in their stated range; the Section 7 selection argument is independent of this step.","rationale":"The reader's weakest-assumption identification is correct and is the most load-bearing issue for the paper's broad claims. I independently checked the Section 7 selection proof and found that its main inequalities are internally consistent: the discounted Mather-measure identity (7.3) follows from the discounted equations, the term ε(∫⟨uε⟩m−∫⟨u⟩mε) in (7.9) is O(ε) because of the uniform L^{2*} and L^{(2*/2)(α+1)} bounds, and the passage from (7.9) to (1.12)–(1.13) is justified using the strong convergence of mε obtained from the monotonicity of g. Thus Theorem 1.9 itself does not appear to have a comparable gap. The concern about Proposition 6.5 relying on unproved uniqueness of Problem 3 is also real but secondary; it affects the refined asymptotics of Theorem 1.4, not the selection criterion. The correct disposition is therefore to retain the conditional verdict: the paper needs a repair or an explicit β+1>0 hypothesis in the classical existence/convergence part, and a uniqueness argument for the linearized problem in Section 6.","tokens_in":24648,"tokens_out":49885,"duration_ms":485666,"concrete_test":"Re-derive Proposition 3.4 for a representative β=−2, for example g(m)=1−1/m with V chosen small enough to satisfy Assumption 1. Check whether the estimate preceding (3.8) can control the leftover ∫m^{p−β} term and whether ∫m^{p+1} can be bounded by ∫m^{p+β+2} in the absence of an upper bound for m. If the chain (3.8) cannot be closed or the iteration cannot be initialized from a positive integrability exponent, then Assumption 2 must be restricted to β+1>0, or the proof of Theorem 1.1 must supply a different argument for β≤−1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 3.4 derives the uniform L∞ bound for m through the chain ∫m^{p+β}|Dm|² ≤ C∫m^{p+1} ≤ C∫m^{p+β+2}, with the last step justified by 'p+β+2>p+1'. This only gives the desired ordering when β+1>0. Assumption 2 explicitly permits any β∈R. For β≤−1, the exponent p+1 is larger than p+β+2, so the displayed estimate goes in the wrong direction; the positive lower bound m≥m0 supplied by Assumption 1 does not repair it because the ratio m^{p+1}/m^{p+β+2}=m^{−β−1} is unbounded from above when β<−1 and m is not yet known to be bounded. Moreover, the Moser iteration is initialized at r0=β+1, which is non-positive for β≤−1 and is not a norm controlled by Lemma 3.3. Consequently, the stated β-range in Assumption 2 is not covered by the proof, and Theorem 1.1 and Corollary 1.2 are not established as written. This is a genuine gap affecting the classical-solution part of the paper, although the weak-solution selection theorem 1.9 is proved under Assumption 3 and does not rely on this Moser step.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the vanishing-discount limit of a first-order stationary mean-field game with quadratic Hamiltonian. Under Assumptions 1 and 2 it claims existence and uniqueness of classical solutions of the discounted problem and convergence to a unique classical solution of the ergodic problem (Theorem 1.1, Corollary 1.2), together with refined asymptotics via a linearized problem (Theorem 1.4). Under a different growth range, Assumption 3, it develops a weak-solution framework, constructs discounted and limit Mather measures, and proves a selection criterion (Theorem 1.9): any weak limit of regular weak discounted solutions is the regular weak solution of the limit problem that minimizes the Mather-type functional ∫⟨u⟩m. An explicit one-dimensional example with non-unique weak solutions is used to illustrate the selection rule.","tokens_in":24944,"tokens_out":20759,"duration_ms":194043,"significance":"If the selection theorem is correct, it is a genuinely interesting application of Aubry-Mather ideas to mean-field games and gives a sharp variational selection principle for the vanishing-discount limit. The paper also provides a welcome explicit example of non-uniqueness and a workable notion of regular weak solutions. The classical existence and refined-asymptotics part is also potentially valuable, and the Mather-measure construction in Section 7 is clearly the strongest and most original part of the paper. However, the classical part contains a real gap in the stated parameter range, and one load-bearing uniqueness assertion is not proved; these issues need to be addressed before the paper can be accepted.","major_comments":[{"comment":"The proof of the uniform L∞ bound for m uses the chain ∫m^{p+β}|Dm|² ≤ C∫m^{p+1} ≤ C∫m^{p+β+2}, with the last step justified by p+β+2>p+1. This ordering only holds when β+1>0, whereas Assumption 2 explicitly permits every β∈R. For β≤−1 the exponent ordering is reversed; the lower bound m≥m0 from Proposition 3.2 does not repair the estimate because m^{p+1}/m^{p+β+2}=m^{−β−1} is unbounded above when β<−1. In addition, the Moser iteration is initialized at r0=β+1, which is not a norm controlled by Lemma 3.3 when β+1≤0. Consequently Theorem 1.1 and Corollary 1.2 are not proved in the full stated range of Assumption 2. This does not affect the Section 7 selection argument, which is run under Assumption 3, but it is a genuine gap in the classical part of the paper.","section":"§3, Proposition 3.4, inequality (3.8)"},{"comment":"The proof of Proposition 6.5 asserts 'Because the solution to (1.5) is unique' and uses that assertion to pass from subsequential convergence of (vϵ,θϵ) to full convergence. No proof or reference for uniqueness of Problem 3 is given. Uniqueness for the linear first-order system (1.5), including uniqueness of the constant λ under the normalization of u, is not evident and is load-bearing for Theorem 1.4; without it the refined asymptotics are only subsequential. The authors should either prove the uniqueness statement or replace it with a direct argument that does not require it.","section":"§6, Proposition 6.5"},{"comment":"The derivation of (7.8) is summarized as 'proceeding in a similar manner,' but the symmetric argument is not identical to the one leading to (7.7). It requires an estimate on εuε+1/2|D(ηδ*uε)|²+Wε integrated against the arbitrary regular weak solution m, whereas (1.6) is stated only in the sense of distributions. The mollification step introduces ηδ*m in place of m, so an additional argument is needed, for instance using the pointwise identity forced by (1.6)–(1.7) together with uniform integrability, or a different mollification procedure. Since the selection inequality (1.13) rests on (7.8), this step should be spelled out.","section":"§7, proof of Theorem 1.9, inequality (7.8)"}],"minor_comments":[{"comment":"The display on the left-hand side of (7.5) reads ∫ v·D(ηδ*u)−Lϵ+W−Wϵ dμϵ, but the left-hand side of (7.4) has −v·D(ηδ*u)−Lϵ+W−Wϵ; the minus sign is missing in the display, even though the subsequent computation appears to use it.","section":"§7, equation (7.5)"},{"comment":"The second condition in Assumption 2 is printed as g(z)≤C2+zg(z); the way this condition is used in Lemma 3.3 and later suggests that the intended inequality is g(z)≤C2+zg'(z) or a similar growth condition. Please correct or clarify the statement.","section":"Assumption 2"},{"comment":"In the integration-by-parts step after equation (6.5) the displayed formula contains the fragment '−hzxi' with an undefined h; it should involve the function ψ_i defined in (6.5).","section":"§6, Proposition 6.3"},{"comment":"The proof states strict inequalities comparing ˜u and u on the two halves of the interval, but the preceding bounds (7.12)–(7.13) appear to justify only ≥/≤, with possible equality on parts of the support of m; the strict inequalities should be relaxed or justified.","section":"§7.4, Proposition 7.9"}],"recommendation":"major_revision","confidential_remarks":"The selection principle and the Mather-measure construction are the strongest contribution and appear to be independent of the classical-existence gap. The β≤−1 issue in Proposition 3.4 and the unproved uniqueness in Proposition 6.5 are fixable in principle, but the stated theorems currently exceed what is proved; restricting Assumption 2 to β>−1 would be an acceptable fallback if the selection results remain unchanged."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this paper. First, the selection criterion in Section 7 is new and, as far as I can tell, correct: for regular weak solutions, the vanishing-discount limit is pinned down by a Mather-type minimization, and the inequality (1.13) gives a clean organizing principle for the non-unique limits. That is a real step forward for first-order stationary mean-field games. Second, the classical-solution part has a gap in the stated generality. Proposition 3.4 claims uniform L∞ bounds for m using the chain ∫m^{p+1} ≤ C∫m^{p+β+2}, justified by p+β+2>p+1. That inequality only holds when β+1>0, and Assumption 2 allows any real β. For β≤−1 the displayed estimate goes the wrong way, and the Moser iteration starts at r0=β+1, which is not even a valid exponent. So Theorems 1.1 and Corollary 1.2 are not established as written. This looks fixable by adding β+1>0 to Assumption 2 (or by a different estimate), but it is a real restriction, not a typo.\n\nWhat the paper does well: the weak-solution framework in Section 7 is careful, the Mather-measure construction is sound, and the proof of Theorem 1.9 is reconstructible. The explicit example in 7.4 is informative and consistent. The refined asymptotics in Theorem 1.4 are a useful complement, though they inherit the same classical-solution assumptions. There are also minor issues: Proposition 6.5 invokes uniqueness of the linearized problem without proof, and a few displayed inequalities in Section 7 (around (7.5)) have sign or indexing errors that obscure but do not destroy the argument.\n\nBottom line: the selection theorem deserves serious attention, and the paper is worth refereeing. The authors should restrict or repair the classical-solution assumptions before publication. The citation pattern is fine; the dependence on Gomes [19] and Ferreira–Gomes [16] is explicit and appropriate.\n\nI would not desk-reject this. I would send it to a careful referee and ask them to check the Moser iteration range.","headline":"The weak-solution selection theorem is a genuine advance and looks right, but the classical existence results are stated more broadly than the proofs support: Proposition 3.4 needs β+1>0, which Assumption 2 does not guarantee.","tokens_in":25477,"tokens_out":1327,"would_cite":true,"duration_ms":15549,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["35A01","49L25","91A13"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves that, as the discount rate vanishes, the limit of a first-order stationary mean-field game is the regular weak solution minimizing the weighted action functional $\\int\\langle u\\rangle m\\,dx$ among all regular weak…","keywords":["mean-field games","Hamilton-Jacobi equations","selection problem","vanishing discount","weak solutions","regularity estimates","stationary mean-field games"],"falsifier":"Solve the explicit example $g(m)=m$, $V(x)=\\pi\\cos(2\\pi x)$ numerically for a sequence $\\epsilon\\to0$ and test whether $u_\\epsilon-\\int u_\\epsilon$ converges weakly in $H^1(\\mathbb T)$ to the minimizer $\\tilde u$ of $\\int\\langle u\\rangle m\\,dx$; convergence to the other regular weak solution $\\hat u$ would disprove Theorem 1.9.","tokens_in":24418,"feed_emoji":"🎯","tokens_out":8031,"duration_ms":71613,"temperature":0.7,"pith_summary":"The paper studies what happens to a stationary mean-field game on a torus when the discount rate tends to zero. It proves that, under polynomial growth and compactness assumptions, the discounted problem has a unique classical solution and that this solution converges to a unique classical solution of the limit problem. When uniqueness fails, the paper gives a selection rule: any weak limit of the discounted solutions must be a solution of the limit problem that minimizes the weighted action functional $\\int\\langle u\\rangle m\\,dx$ among all regular weak solutions. The rule is proven by constructing a phase-space measure from each discounted solution and passing to a limit measure that satisfies a holonomy constraint. An explicit one-dimensional example with non-unique solutions shows what the selection criterion predicts.","feed_headline":"Vanishing discount picks the cost-minimizing mean-field game solution","feed_subtitle":"In first-order stationary mean-field games, the epsilon-to-zero limit is shown to minimize the weighted cost over all regular weak…","key_machinery":"The central object is a phase-space probability measure attached to each solution, the discounted holonomy measure\n$$\\mu_\\epsilon(dx,dv)=m_\\epsilon(x)\\,dx\\otimes\\delta_{-Du_\\epsilon(x)}(dv).$$\nIt encodes the transport equation through the discounted holonomy condition\n$$\\int(-\\epsilon\\phi+v\\cdot D\\phi)\\,d\\mu_\\epsilon=-\\epsilon\\int\\phi\\,dx,$$\nand its Lagrangian action equals $\\epsilon\\int u_\\epsilon\\,dx$. Passing $\\epsilon\\to0$ yields a holonomy measure $\\mu$ satisfying $\\int v\\cdot D\\phi\\,d\\mu=0$. Comparing the two measures through the Legendre transform of the quadratic Hamiltonian produces the monotone inequality whose $\\epsilon\\to0$ limit is (1.12), and hence the selection inequality (1.13). This is the mechanism that converts compactness of the discounted solutions into a variational selection principle.","core_discovery":"The central discovery is a selection criterion for the vanishing-discount limit of first-order stationary mean-field games. If $(u_\\epsilon,m_\\epsilon)$ is a regular weak solution of the discounted problem and, after normalization, $u_\\epsilon$ converges weakly in $H^1$ to $\\bar u$ while $m_\\epsilon$ converges weakly in $L^1$ to $\\bar m$, then for any regular weak solution $(u,m)$ of the limit problem the difference term $\\int(g(m_\\epsilon)-g(m))(m_\\epsilon-m)\\,dx$ tends to zero. Since $g$ is strictly increasing, this forces $\\bar m=m$. The same argument yields the variational inequality $\\int\\langle\\bar u\\rangle m\\,dx\\le\\int\\langle u\\rangle m\\,dx$ for every regular weak solution of the limit problem, so the vanishing-discount limit is the regular weak solution that minimizes the functional $u\\mapsto\\int\\langle u\\rangle m\\,dx$. The paper also proves that when a uniform lower bound on the density and polynomial growth of $g$ hold, the discounted problem has a unique classical solution and the limit is classical and unique, with the refined rate $\\|u_\\epsilon-\\bar H/\\epsilon-u-\\lambda\\|_\\infty+\\|m_\\epsilon-m\\|_\\infty\\to 0$.","pith_inferences":["This selection rule could be used as an a posteriori numerical check: compute any regular weak solution of the limit problem, evaluate $u\\mapsto\\int\\langle u\\rangle m\\,dx$, and the vanishing-discount limit must be the minimizer.","The same holonomy-measure comparison may apply to other singular perturbations—small viscosity, finite horizon, or stochastic noise—provided a discounted holonomy condition with a similar limit exists; the paper does not prove those cases.","If the minimizing solution is not unique, the first-order correction $\\lambda$ in the refined expansion would be the natural tie-breaker; the paper leaves that question open.","The explicit example suggests that non-uniqueness is tied to regions where the limiting density vanishes, so the selection criterion can be interpreted as choosing the solution whose active positive-density region realizes the lowest weighted action."],"forward_implications":["If the selection inequality holds, the vanishing-discount limit is singled out by a variational principle: among all regular weak solutions of the stationary problem, it minimizes $u\\mapsto\\int\\langle u\\rangle m\\,dx$.","In the classical regime, the limit is unique up to constants and smooth, with the refined expansion $u_\\epsilon=\\bar H/\\epsilon+u+\\lambda+o(1)$ and $m_\\epsilon=m+o(1)$.","The monotone term $\\int(g(m_\\epsilon)-g(m))(m_\\epsilon-m)\\,dx\\to0$ forces the limiting density to match the density of any regular weak solution, so $\\bar m$ is unique.","The explicit one-dimensional example shows that non-uniqueness arises where the limiting density vanishes, and the selection criterion identifies the minimizer among the possible solutions."],"supporting_citations":[{"why":"Supplies the existence of weak solutions to the discounted problem and the uniform estimates that define regular weak solutions.","marker":"[16]"},{"why":"Introduces the generalized discounted-holonomy viewpoint used to prove the selection inequality.","marker":"[19]"},{"why":"Establishes convergence of discounted Hamilton-Jacobi equations, the model for the vanishing-discount argument.","marker":"[5]"},{"why":"Studies the vanishing-discount limit for second-order mean-field games and inspires the refined asymptotic expansion.","marker":"[4]"},{"why":"Provides the explicit non-uniqueness example that motivates the selection criterion.","marker":"[23]"},{"why":"Original vanishing-discount method for Hamilton-Jacobi equations, used to justify the limit process.","marker":"[31]"}],"fun_headline_variants":["Vanishing discount picks cost-minimizing mean-field game limit","Mean-field games: discount limit selects minimal-cost solution","Discount limit picks cost-minimizing MFG solution via Aubry-Mather"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The selection mechanism relies on the discounted densities $m_\\epsilon$ converging strongly in some $L^p$ space (not merely weakly) as $\\epsilon\\to0$; if only weak convergence is available, the monotonicity argument cannot force the vanishing-discount density to coincide with the density of the limit solution.","fun_headline_variants_meta":{"raw":{"variants":["Vanishing discount picks cost-minimizing mean-field game limit","Mean-field games: discount limit selects minimal-cost solution","Discount limit picks cost-minimizing MFG solution via Aubry-Mather"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000525,"raw_usage":{"total_tokens":2520,"prompt_tokens":911,"completion_tokens":1609,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":1552}},"tokens_in":527,"tokens_out":1609,"duration_ms":10650,"temperature":1.0,"reasoning_tokens":1552,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:47:08.358060+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Solve the explicit example $g(m)=m$, $V(x)=\\pi\\cos(2\\pi x)$ numerically for a sequence $\\epsilon\\to0$ and test whether $u_\\epsilon-\\int u_\\epsilon$ converges weakly in $H^1(\\mathbb T)$ to the minimizer $\\tilde u$ of $\\int\\langle u\\rangle m\\,dx$; convergence to the other regular weak solution $\\hat u$ would disprove Theorem 1.9.","supporting_citations":[{"cited_title":"Ferreira and D","cited_arxiv_id":null,"evidence_quote":"Supplies the existence of weak solutions to the discounted problem and the uniform estimates that define regular weak solutions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces the generalized discounted-holonomy viewpoint used to prove the selection inequality."},{"cited_title":"Davini, A","cited_arxiv_id":null,"evidence_quote":"Establishes convergence of discounted Hamilton-Jacobi equations, the model for the vanishing-discount argument."},{"cited_title":"Cardaliaguet and A","cited_arxiv_id":null,"evidence_quote":"Studies the vanishing-discount limit for second-order mean-field games and inspires the refined asymptotic expansion."},{"cited_title":"Gomes, L","cited_arxiv_id":null,"evidence_quote":"Provides the explicit non-uniqueness example that motivates the selection criterion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Original vanishing-discount method for Hamilton-Jacobi equations, used to justify the limit process."}],"review_version":1}