{"id":"d62e441e-cf59-49a3-9185-9183fdcdabb0","arxiv_id":"2607.26808","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Under convexity, potential MFG equilibria solve mean-field singular control; for the mean-field monotone follower with strategic complementarities the optimum is a free boundary uniquely solving a nonlinear integral equation.","lead":"The paper links mean-field singular control problems to an auxiliary potential mean-field game and, under convexity, shows game equilibria solve the control problem. For a mean-field monotone follower, it fully characterizes the optimal policy by a free boundary solving a nonlinear integral equation.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The paper delivers exactly what it claims: a convexity bridge from potential singular MFG to MFC (Theorem 3.1) and a complete free-boundary characterization of the mean-field monotone follower for \theta continuous and \theta\notin(0,2) (Theorems 4.10, 4.13, 4.14). All steps use classical tools (singular SMP, optimal-stopping connection of Karatzas–Shreve type, topological/lattice fixed-point theorems, Peskir change-of-variable) under hypotheses that are standard and verified in the example. The reader correctly flags Assumption 3.2 as the structural hinge; that hinge is not a soft spot but a transparent modeling condition. No internal inconsistency, no unstated regularity leap, and no circularity appear. Therefore the ACCEPT verdict with high confidence stands; no adjustment is warranted.","tokens_in":29217,"tokens_out":532,"duration_ms":11206,"concrete_test":"Independently re-derive the passage from the BSDE adjoint (3.5) and the singular SMP inequality (3.6) through the L-convexity step (3.10)–(3.11) to the final comparison J(ξ*)\to J(ξ)\to0, confirming that the identification \\partial_\\mu H=\\partial_\nu \theta C and the interchange of derivative/integral under Assumption 3.1(v) hold without additional regularity; if the inequality direction reverses or an extra remainder appears, the bridge fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claims rest on standard, explicitly stated structural hypotheses (L-joint convexity of H and G in Assumption 3.2; strategic complementarity α∈(0,2) for the free-boundary integral equation). The proof of Theorem 3.1 correctly routes the potential-MFG maximum principle through the convexity inequality (3.10)–(3.11) to obtain the MFC comparison; the monotone-follower analysis (optimal-stopping connection, Kakutani–Fan–Glicksberg fixed point on the weakly compact set E, Tarski on the continuous subclass, and uniqueness of the continuous nonincreasing solution of (4.42)) follows classical arguments without hidden gaps. The reader’s weakest-assumption identification is accurate but does not undermine the logical structure: the hypothesis is necessary, stated up front, and verified for the case study. Residual risk is ordinary line-by-line verification, not a load-bearing conceptual flaw.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper studies finite-horizon mean-field control (MFC) problems with singular controls and general measure dependence in the cost. Under growth/regularity (Assumption 3.1) and L-joint convexity of the Hamiltonian and terminal cost (Assumption 3.2), it constructs an auxiliary potential MFG whose costs are built from C, G and their linear derivatives (3.3), and proves that any MFG equilibrium yields an MFC optimum (Theorem 3.1), with uniqueness of the MFG equilibrium when the MFC problem is unique (Corollary 3.2). The result is applied to a mean-field monotone follower problem with scalar interaction parameter α. The associated potential MFG is solved by linking the representative-agent problem to optimal stopping, constructing the best-reply map on a weakly compact convex subset of L², and applying Kakutani–Fan–Glicksberg (Theorem 4.8). For strategic complementarities α∈(0,2), the unique equilibrium (hence the MFC optimum) is reflection at a continuous nonincreasing free boundary that uniquely solves a nonlinear integral equation (4.42) in a suitable class (Theorems 4.10, 4.13, 4.14), with an iterative numerical scheme illustrated in Figure 1.","tokens_in":29425,"tokens_out":1307,"duration_ms":39917,"significance":"The contribution is twofold and genuine. First, the potential-MFG link for singular controls is a useful converse-type companion to the regular-control results of Höfer–Soner and related work: under stated convexity it reduces MFC characterization to a more tractable fixed-point problem. Second, the finite-horizon mean-field monotone follower is given a complete free-boundary characterization (integral equation plus consistency), which the literature review and related ergodic/one-dimensional works do not provide. The case-study analysis is classical but carefully executed (optimal-stopping connection, weak compactness and closed-graph argument, Tarski on the continuous subclass, uniqueness of the continuous nonincreasing solution of (4.42)). The structural hypotheses are stated up front and verified for the quadratic example. This is a solid, publishable contribution in mean-field singular control.","major_comments":[{"comment":"Proof of Theorem 3.1 (pp. 5–6): the appeal to the singular SMP of Bahlali–Djehiche–Mezerdi [2, Thm. 3.6] is extended from bounded state derivatives to linear growth by a one-line dominated-convergence remark. Given that the adjoint BSDE (3.5) and the comparison (3.6) are load-bearing for the whole implication MFG ⇒ MFC, a short self-contained justification (or a precise citation to an extension covering linear growth and the Stieltjes integral up to T) would make the argument fully checkable without external reconstruction.","section":"§3, Theorem 3.1 and display (3.5)–(3.6)"},{"comment":"Existence for the general MFC/potential MFG is not claimed outside the case study; only the implication “MFG equilibrium ⇒ MFC optimum” is proved under Assumptions 3.1–3.2. That is consistent with the abstract, but the introduction’s framing (“we derive an auxiliary MFG… and show that any solution yields…”) could briefly flag that existence of the potential MFG is left open in the abstract setting and is obtained only for the monotone follower via Kakutani–Fan–Glicksberg. This is a scope clarification, not a gap in the proved theorems.","section":"§1 and §3"}],"minor_comments":[{"comment":"Notation: the same letter K is used for the control-cost process K(t) in the general problem and for the constant K in the monotone-follower cost; a local rename in §4 would avoid confusion.","section":"§3–§4"},{"comment":"In (3.3) and the subsequent Hamiltonian H^μ, the dependence of c on the full measure flow versus the scalar mean is clear in §4 but could be signposted once when specializing from μ to θ.","section":"§4, after (4.3)"},{"comment":"Lemma 4.3 / (4.15): the inclusion S ⊆ {x ≥ Kρ + α(2−α)θ_t} is standard; a one-line reminder that the same lower bound is reused in Lemma 4.11(iii) and in the uniqueness class for (4.42) would help the reader track the a-priori bound.","section":"§4.1–§4.3"},{"comment":"Figure 1 caption: state the Monte Carlo sample size and the numerical solver used for the integral equation (4.47) so the plot is reproducible at the level claimed by the iterative scheme.","section":"§4.3.1, Figure 1"},{"comment":"Typos / style: “vice versa result to that achieved” (p. 1); “somewat” is not present but several long sentences in the introduction could be split; arXiv ID and date line are fine.","section":"§1"},{"comment":"References [8] and the ergodic companion works are appropriately cited; ensure the final version updates “To appear” items consistently.","section":"References"}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is a clean, technically competent contribution; the reader’s and skeptic’s assessments align with mine. Self-citations to overlapping ergodic work are related and not excessive. Fit for a strong math.OC / applied-probability journal is good. I see no integrity or novelty-disclosure issues. Minor revision is appropriate mainly to tighten the SMP-extension remark and a few presentation points; I would not require new theorems."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The two things worth knowing: they reverse the usual potential direction (MFG solution implies MFC optimum under L-joint convexity) for singular controls, and they give a complete free-boundary-plus-integral-equation description of the finite-horizon mean-field monotone follower when α∈(0,2). Both claims check out on the manuscript.\n\nWhat is new is real and cleanly scoped. Theorem 3.1 takes the potential costs built from the linear derivatives, runs the singular-control maximum principle, and feeds the L-convexity comparison to recover the MFC inequality. That is the opposite direction from Höfer–Soner for regular controls, and it is new for singular controls. The case study then does the classical work carefully: optimal-stopping connection, free boundary, Kakutani–Fan–Glicksberg on the weakly compact set E in L^{2}, uniqueness from strict convexity of the original MFC, and, under strategic complementarity, continuity of the boundary plus uniqueness of the continuous nonincreasing solution of the nonlinear integral equation (4.42). The iterative scheme and the plot are honest illustrations, not data fitting. Citations sit in the right places; the self-cites are related ergodic work, not load-bearing.\n\nSoft spots are ordinary and stated up front. Assumption 3.2 (L-joint convexity of H and G) is necessary for the comparison step; without it the implication fails. The sharp free-boundary uniqueness is restricted to α∈(0,2). Growth conditions need the usual dominated-convergence extension of the maximum principle in [2]. None of these is hidden or fatal; they are the standard structural hypotheses for this style of argument. Residual risk is line-by-line verification, not conceptual opacity.\n\nThis is for people who work on singular MFGs/MFC or free-boundary problems and want a transferable reduction plus an explicit finite-horizon example. It deserves a serious referee. I would cite the characterization and the potential link, and I would bring it to reading group.","headline":"Solid reverse potential-MFG link for singular controls plus the first full free-boundary characterization of a finite-horizon mean-field monotone follower.","tokens_in":30049,"tokens_out":524,"would_cite":true,"duration_ms":9880,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49N80","91A16","93E20","65D15"],"pacs":[],"model":"grok-4.5","headline":"Any equilibrium of a derived potential mean-field game solves the original mean-field singular control problem, and for the monotone follower the optimum is reflection at a free boundary that solves a nonlinear integral equation.","keywords":["singular stochastic control","mean-field control","mean-field games","potential games","monotone follower","free boundary","optimal stopping"],"falsifier":"Construct a concrete mean-field singular-control cost that violates L-joint convexity yet still admits a unique potential-MFG equilibrium, and check whether that equilibrium fails to minimise the original mean-field cost; or, for the follower problem with α∈(0,2), exhibit two distinct continuous nonincreasing free boundaries both solving the integral equation (4.42).","tokens_in":30050,"feed_emoji":"📈","tokens_out":1024,"duration_ms":17399,"temperature":0.7,"pith_summary":"Mean-field control problems with singular controls are hard to solve directly because the value function lives on the space of probability measures. This paper shows that, under convexity, one can instead solve an auxiliary mean-field game (the potential MFG) whose equilibria automatically give optimal controls for the original problem. The construction is applied to a mean-field version of the classical monotone follower problem: a controller tracks a Brownian path while paying a quadratic penalty that depends on the distance from a multiple of the population mean. Existence of equilibrium follows from a fixed-point argument; uniqueness follows from strict convexity of the original cost. When the interaction strength lies in (0,2), strategic complementarities appear and the unique equilibrium (hence the optimal mean-field control) is completely described by continuous reflection at a free boundary that is the unique continuous nonincreasing solution of a nonlinear integral equation. An elementary iteration then produces numerical plots of the boundary, the mean trajectory and sample paths of the optimal control.","feed_headline":"Mean-field singular control solved via a free-boundary game","feed_subtitle":"Convexity turns potential-game equilibria into optimal policies; the follower problem yields an explicit integral equation","key_machinery":"The potential MFG: the representative player’s running and terminal costs are obtained from the original mean-field costs by adding their linear derivatives with respect to the measure. Under L-joint convexity this game’s equilibria solve the original control problem; in the follower example the game reduces to optimal stopping plus a Kakutani–Fan–Glicksberg (or Tarski) fixed point, yielding the free-boundary integral equation.","core_discovery":"Under linear growth, continuity and L-joint convexity of the Hamiltonian and terminal cost, every solution of the auxiliary potential mean-field game with singular controls is optimal for the original mean-field singular control problem. In the mean-field monotone follower with interaction parameter α∈(0,2) the unique equilibrium (and therefore the unique optimal policy) is the Skorokhod reflection of Brownian motion at a continuous nonincreasing free boundary that uniquely solves a nonlinear integral equation within the class of continuous functions lying above the natural obstacle.","pith_inferences":["The same convexity bridge should extend, with only technical changes, to mean-field problems that mix singular and regular controls or that include common noise.","Once the free boundary is known to solve a scalar integral equation, standard comparative-statics arguments become available: how the boundary moves with discount rate, volatility or interaction strength can be read off by differentiation under the integral.","The construction suggests a practical route to reinforcement-learning algorithms for mean-field singular control: learn the best-reply free boundary for frozen mean-field paths and then iterate the consistency map."],"forward_implications":["Any strictly convex mean-field singular-control problem automatically inherits uniqueness of its associated potential-MFG equilibrium.","The free-boundary integral equation supplies a practical numerical scheme (fixed-point iteration of optimal stopping plus Monte-Carlo expectation) that converges to the unique mean-field optimum when strategic complementarities hold.","The same potential-game reduction can be tried on other finite-horizon singular-control models (irreversible investment, capacity expansion, dividend problems) once their Hamiltonians satisfy the convexity hypothesis.","When α lies outside (0,2) the best-reply map reverses monotonicity, so existence still holds but the free-boundary characterisation and the monotone iteration are lost."],"fun_headline_variants":["Potential MFG equilibria solve mean-field singular control under convexity","Monotone follower MFC: free boundary solves nonlinear integral equation","Singular MFC optimal via Skorokhod reflection at continuous free boundary","Convex Hamiltonian turns potential MFG solutions into MFC optima","Mean-field monotone follower yields unique free-boundary equilibrium"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The Hamiltonian and terminal cost must be jointly convex in the state and the measure; without that convexity an equilibrium of the auxiliary game need not be optimal for the original control problem.","fun_headline_variants_meta":{"raw":{"variants":["Potential MFG equilibria solve mean-field singular control under convexity","Monotone follower MFC: free boundary solves nonlinear integral equation","Singular MFC optimal via Skorokhod reflection at continuous free boundary","Convex Hamiltonian turns potential MFG solutions into MFC optima","Mean-field monotone follower yields unique free-boundary equilibrium"]},"model":"grok-4.5","effort":"low","cost_usd":0.00367,"raw_usage":{"total_tokens":1192,"prompt_tokens":816,"num_sources_used":0,"completion_tokens":70,"cost_in_usd_ticks":36704000,"prompt_tokens_details":{"text_tokens":816,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":306,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":816,"tokens_out":70,"duration_ms":5707,"temperature":1.0,"reasoning_tokens":306,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-30T20:27:29.643339+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Construct a concrete mean-field singular-control cost that violates L-joint convexity yet still admits a unique potential-MFG equilibrium, and check whether that equilibrium fails to minimise the original mean-field cost; or, for the follower problem with α∈(0,2), exhibit two distinct continuous nonincreasing free boundaries both solving the integral equation (4.42).","supporting_citations":[],"review_version":1}