{"id":"92b72134-1607-4b73-90e3-e68812f10771","arxiv_id":"2412.01995","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The most exciting win-martingale among d+1 players is characterized by a Monge-Ampère equation, and its diffusion coefficient is sqrt(2(∇²g)^{-1}/(1-s)).","lead":"This paper finds the 'most exciting game' for any number of players by showing the optimal win-probability process is driven by the Hessian of a solution to a Monge-Ampère equation on the simplex. The result connects game theory, martingale optimal transport, and fully nonlinear PDEs, and gives the first multi-player solution to Aldous's problem.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uniqueness of the optimizer is asserted in Theorem 1.2(iii) but never proved; the text defers to [4] without a d-dimensional equality argument.","rationale":"The paper makes a strong mathematical contribution: the boundary estimates for the infinite-boundary Monge-Ampère equation (Theorem 1.1) are original and appear technically sound, and the construction of the Aldous martingale via (AM) is convincing given those bounds. My concern is not with the existence or optimality of the Aldous martingale per se, but with the proof of uniqueness in law, which is explicitly claimed in Theorem 1.2(iii) yet entirely deferred. The one-dimensional argument in [4] relies on explicit formulas that are not available here; a full proof in d dimensions requires identifying the equality condition in the Legendre transform used to derive the HJB. This is absent. I also note the unjustified modification of the supersolution in Proposition 4.2 Step 3b, which undermines the paper's stated route to identifying the value function; however, the value function identity can likely be recovered by a direct verification argument, so the more fundamental gap is the uniqueness proof. The reader's weakest_assumption (the modeling connection to Shannon entropy) is a different issue: it affects the interpretation but not the validity of Theorem 1.2 for the chosen cost. Thus I disagree with the reader on what is load-bearing, though my recommended verdict (conditional) matches theirs.","tokens_in":33743,"tokens_out":26010,"duration_ms":224329,"concrete_test":"Write out the equality case in the variational inequality for a generic win-martingale Q: after applying Itô to f(s)+(1-s)g(M_s) and using the Legendre inequality, prove that E∫[-log det Σ_s - (d+d log(1-s)+g(M_s))] ds = 0 implies Σ_s = 2(∇²g(M_s))^{-1}/(1-s) for a.e. s, Q-a.s. Then invoke the strong uniqueness from Lemma 6.1. If this derivation cannot be completed, Theorem 1.2(iii) is unproven.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim includes uniqueness in law of the optimizer (Theorem 1.2(iii)). The proof ends with 'For the uniqueness of the optimizer we can argue just as in [4]' without supplying the argument. In the d-dimensional setting the candidate SDE (AM) has a non-explicit coefficient built from ∇²g, so the one-dimensional proof in [4] (which uses the explicit solution g(x)=log(π²/sin²(πx))) does not carry over verbatim. A rigorous uniqueness proof requires showing that any optimizer Q satisfies Σ_s = 2(∇²g(M_s))^{-1}/(1-s) for a.e. s and Q-a.s., i.e., the equality case in the Legendre inequality -log det Σ + Tr(((1-s)/2)∇²g Σ) ≥ d + d log(1-s) + g. This equality case is neither stated nor proved in the paper. Without it, the uniqueness assertion is unsupported. A secondary gap is in Proposition 4.2 Step 3b, where the supersolution û2 is modified to be +∞ outside a small ball; this is outside the class of finite-valued functions in Definition 4.1 and the preservation of the super-solution property is not shown. Since Corollary 5.1 relies on Proposition 4.2, the identification v=(1-t)g+f is not fully justified as written. Both gaps are fixable, but they currently weaken the proof of Theorem 1.2.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the continuous-time \"most exciting game\" problem for d+1 players. A win-martingale is a martingale on the d-dimensional subprobability simplex that terminates at a vertex, and the cost is E∫_t^1 −log det(Σ_s) ds. The authors prove that the value function has the structure (1−t)g(x)+d(1−t)log(1−t), with g solving the Monge-Ampère equation g=log det(1/2 ∇²g) on the simplex with infinite boundary data. They establish existence, uniqueness, regularity and boundary estimates for g (Theorem 1.1), use these to construct the \"Aldous martingale\" as the strong solution to dM_s = (2(∇²g(M_s))^{-1}/(1−s))^{1/2} dB_s, prove that it is a win-martingale and attains the value function, and assert that it is the unique optimizer in law (Theorem 1.2). The candidate is parameter-free, determined by a PDE with no free constants, and the verification is via the martingale property of log det of the optimal covariance (Lemma 6.4).","tokens_in":34013,"tokens_out":7503,"duration_ms":84731,"significance":"If correct, the paper gives a d-dimensional solution to a natural martingale optimization problem connected with Aldous's question, and it establishes a striking link to Monge-Ampère equations with infinite boundary data on a non-smooth domain. The main mathematical contributions are the boundary estimates for ∇²g on the simplex and the explicit construction of the optimal SDE without fitted parameters. The verification argument is clean and specific. I also note that the manuscript is transparent that the modeling link to a scaling limit of Shannon entropies is taken from [4]; the variational optimality result stands independently of that interpretation, but the \"most exciting game\" reading is conditional on that identification. However, the claimed uniqueness in law of the optimizer is not established in the paper, and the comparison proof has a gap; these are load-bearing and require repair before the main theorem is fully proven.","major_comments":[{"comment":"Statement (iii), uniqueness in law of the optimizer, is asserted with the sentence \"For the uniqueness of the optimizer we can argue just as in [4]\" and is not proved. The one-dimensional argument in [4] uses the explicit solution g(x)=log(π²/sin²(πx)) and the corresponding explicit SDE, so it does not transfer verbatim to the d-dimensional setting, where the coefficient is built from an implicitly defined ∇²g. To justify uniqueness one must show that any optimizer Q satisfies Σ_s = 2(∇²g(M_s))^{-1}/(1−s) for a.e. s and Q-a.s., i.e. one must characterize the equality case in the Legendre inequality −log det Σ + (1/2)Tr((1−s)∇²g(M_s)Σ) ≥ d + d log(1−s) + g(M_s). That equality condition is neither stated nor proved. Please provide a self-contained argument, or clearly state the missing equality-case result and prove it.","section":"Section 6, proof of Theorem 1.2, final paragraph"},{"comment":"The modification of û2 to be +∞ outside a closed small ball around (t0,x0) takes the function outside the class of finite-valued convex functions allowed in Definition 4.1, and the assertion that this modification preserves the super-solution property is not justified. In particular, test functions with local minima at the boundary of the ball are not covered by the condition in Definition 4.1(iii). This modification is essential to force the maximizer of Φα into the interior, so the comparison principle is incomplete as written. Since Corollary 5.1 relies on Proposition 4.2 to identify v with (1−t)g+f, this gap affects the proof of Theorem 1.2 and should be repaired with a rigorous approximation or localization argument.","section":"Section 4, Proposition 4.2, Step 3b"}],"minor_comments":[{"comment":"The text says \"Lemma 5.6 yields\" the C^{2,α} estimate after Evans–Krylov, but Lemma 5.6 is a first-order estimate for the scaled functions defined later in Section 5.2; the intended reference appears to be Lemma 5.3 (or a direct Evans–Krylov application to the C² bounds).","section":"Section 5.1, proof of Proposition 5.1"},{"comment":"The time-change in Section 2 is defined as Y_t := M_{1−e^{−t}}, whereas Section 6, before Lemma 6.1, uses Y_t := M_{1−e^{−t/2}}; the two conventions should be reconciled or explicitly distinguished.","section":"Section 2 vs. Section 6"},{"comment":"The notation in the convexity proof is dense: the vector of ones is denoted by 1 in the same display as scalar quantities, and the coupling of the Brownian motions is described only in words. A brief notational clarification would improve readability.","section":"Proposition 3.2"},{"comment":"The abstract says the winning-probability \"is described by\" the SDE, while Theorem 1.2 more precisely asserts strong existence, the win-martingale property, and unique optimality in law. The abstract could reflect that the uniqueness part is part of the theorem.","section":"Abstract and Theorem 1.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a strong contribution if the two gaps can be closed. The uniqueness claim is currently deferred to [4] in a way that is not justified for the d-dimensional implicit coefficient, and the comparison-principle proof contains a localization step that is outside the stated solution class. Both are repairable in principle, but they are central to the main theorem. I would also suggest the authors state explicitly in the introduction or in Theorem 1.2 that the interpretation as the \"most exciting game\" inherits the modeling identification from [4]."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this paper does something genuinely new—it solves the multi-player version of Aldous's most-exciting-game problem and connects it to a Monge-Ampère equation with infinite boundary on the simplex. The main mathematical novelty is the boundary asymptotics of ∇²g, Theorem 1.1(i)-(ii), and the construction of the Aldous martingale in arbitrary dimension. That part is substantial and, as far as I can tell, correct. There is no fitting or tuning; g is pinned down by a PDE and the verification argument in Lemma 6.4 is clean. If the result survives scrutiny, it is a significant bridge between martingale optimal transport and fully nonlinear PDE.\n\nThe soft spots are real but localized. Theorem 1.2(iii) asserts uniqueness in law of the optimizer, and the proof ends with \"argue just as in [4].\" That is not enough. The one-dimensional proof uses explicit g, and the d-dimensional equality case in the Legendre inequality is neither stated nor proved. Without it, uniqueness is unsupported as written. I would guess it is true—the value function is strictly convex, and the HJB machinery should give equality conditions—but the authors need to supply the argument. This is a fixable gap, not a fatal one.\n\nThe second gap is in Proposition 4.2, Step 3b, where they modify the supersolution to be +∞ outside a small ball. That puts it outside the class of finite-valued functions in Definition 4.1, and they don't show the modification preserves the supersolution property. Since Corollary 5.1 (v=(1-t)g+f) depends on Proposition 4.2, the identification of the value function with the PDE solution is not fully justified as written. Again fixable—probably by redoing the doubling variables argument with a smooth cut-off or a penalization—but it needs to be written.\n\nOne more thing, not a flaw in the math: the identification of the cost with Aldous's \"excitement\" comes from earlier work [4] and is not re-derived. If that modeling premise is wrong, the interpretation changes, but the optimality result for this specific cost remains.\n\nWho is this for: people working in martingale optimal transport, stochastic control, or Monge-Ampère equations. It deserves a serious referee. I would send it out, with a request to fill the uniqueness gap and clean up Step 3b before acceptance.","headline":"Real multidimensional progress on the most exciting game, with a genuine Monge-Ampère connection; the uniqueness claim is asserted but not proved, and the comparison principle has a gap—both fixable.","tokens_in":34580,"tokens_out":1935,"would_cite":true,"duration_ms":173198,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60G44","35J96","49L25","60H10"],"pacs":[],"model":"deepseek-v4-flash","headline":"Among all win-martingales on the $d$-dimensional simplex, the unique most-exciting game is the diffusion $dM_s = \\sqrt{2(\\nabla^2g(M_s))^{-1}/(1-s)}\\,dB_s$, where $g$ is the unique smooth solution of the Monge-Amp\\`ere equation with…","keywords":["most exciting game","win-martingales","Monge-Ampère equation","specific relative entropy","stochastic optimal control","viscosity solutions","Aldous martingale","Shannon entropy scaling limit"],"falsifier":"Starting at any fixed $x\\in\\Delta$ for $d\\ge 2$, simulate the diffusion $dM_s = \\sqrt{2(\\nabla^2g(M_s))^{-1}/(1-s)}\\,dB_s$ with a high-precision numerical solution of the Monge-Amp\\`ere equation; if for some $x$ the terminal law is not concentrated on the vertices, meaning positive probability is observed on an open face of the simplex, Theorem 1.2 is false. Equivalently, exhibit any win-martingale with strictly smaller expected cost $\\mathbb{E}\\int -\\log\\det\\Sigma\\,ds$ than the diffusion achieves; the claimed uniqueness would fail.","tokens_in":33531,"feed_emoji":"🎲","tokens_out":10263,"duration_ms":92901,"temperature":0.7,"pith_summary":"What is the most suspenseful way to reveal the winner of a contest with $d+1$ players? This paper treats win probabilities as a martingale on the simplex and measures excitement by the time-integrated log-determinant of the volatility, the continuous-time scaling limit of Shannon entropy. It claims that the unique optimal martingale is an explicit diffusion whose volatility is the inverse Hessian of $g$, where $g$ solves the Monge-Amp\\`ere equation $g = \\log \\det(\\tfrac12\\nabla^2 g)$ on the simplex and blows up on the boundary. Establishing this requires a fine analysis of the degenerate boundary behavior of that PDE, which is the paper's main technical work. If the identification of excitement with this entropy limit is accepted, the result describes exactly how a bookmaker or designer should update odds to keep an audience watching.","feed_headline":"The most exciting game obeys a Monge-Ampère equation","feed_subtitle":"A single PDE with infinite boundary data describes the unique most-exciting martingale for any number of players.","key_machinery":"The load-bearing object is the Monge-Amp\\`ere equation with infinite boundary condition on the simplex: $g = \\log \\det(\\tfrac12\\nabla^2 g)$ for $x\\in\\Delta$ and $g(x)=\\infty$ on $\\partial\\Delta$. The paper constructs $g$ by approximating the simplex with strictly convex sublevel sets of the explicit barrier $w(x)=-2\\sum_i \\log x_i -2\\log(1-\\sum_i x_i)$, applying comparison and Pogorelov-type estimates to obtain uniform $C^2$ bounds, and then sending the approximation to the limit. The crucial boundary analysis rescales coordinates near each face and uses the scaling invariance of the equation to show that $\\nabla^2g$ behaves like $\\mathrm{diag}(1/x_1^2,\\dots,1/x_k^2,1,\\dots,1)$; this controls the diffusion coefficient of the Aldous martingale and guarantees both non-exit and convergence to a vertex. The final identification of the optimizer is completed by showing that $\\log\\det\\Sigma_*(t,M_t)$ is a true martingale via the global bound on $\\nabla g^{\\top}(\\nabla^2g)^{-1}\\nabla g$.","core_discovery":"The paper's central discovery is that the Aldous martingale, the diffusion $dM_s = \\sqrt{2(\\nabla^2g(M_s))^{-1}/(1-s)}\\,dB_s$ started anywhere in the simplex, is the unique (in law) optimizer of the cost $\\mathbb{E}\\int_t^1 -\\log\\det(\\Sigma_s)\\,ds$ among all win-martingales. Here $g$ is the unique smooth solution of the Monge-Amp\\`ere equation $g = \\log \\det(\\tfrac12\\nabla^2 g)$ on the interior of the simplex and $g = \\infty$ on its boundary. Along the way, the paper proves new boundary asymptotics: the minimal eigenvalue of $\\nabla^2g$ stays bounded as the state approaches a face of the simplex but not a vertex, and the quadratic form $\\nabla g^{\\top}(\\nabla^2g)^{-1}\\nabla g$ is bounded globally. These estimates are what let the authors construct the strong solution, show it never exits the simplex before time $1$, and prove that it lands on a vertex almost surely. The value function splits as $v(t,x)=(1-t)g(x)+d(1-t)\\log(1-t)$, which reduces the parabolic Hamilton–Jacobi–Bellman equation to the elliptic Monge-Amp\\`ere problem.","pith_inferences":["Beyond the paper: the time-changed Aldous martingale has an invariant sigma-finite measure $e^g\\,dx$ on the simplex, which suggests a mirror-Langevin sampler for densities supported on the simplex; the paper notes the moment-measure link but does not pursue sampling applications.","Beyond the paper: if the Shannon-type entropy criterion is replaced by another notion of suspense, the same HJB-to-elliptic reduction may survive in modified form, but the singular boundary analysis would have to be reworked for each new cost.","Beyond the paper: for $d=2$, a high-precision numerical solution of the Monge-Amp\\`ere equation could be used to simulate the Aldous martingale and directly test the prediction that terminal mass concentrates on vertices rather than on open faces of the simplex."],"forward_implications":["For $d=1$, the construction reduces to the known explicit diffusion $dM_s = \\frac{\\sin(\\pi M_s)}{\\pi\\sqrt{1-s}}\\,dB_s$, recovering the earlier solution of the most-exciting-game problem.","For any finite number of players, there now exists a unique optimal win-martingale, so the question of the most exciting game has a complete answer within this entropy-limit model.","The boundary estimates identify how the optimal volatility diverges near the boundary: as some coordinate $x_i\\to 0$, the diffusion coefficient grows like $1/x_i$, which is exactly the singular strength that forces an eventual arrival at a vertex.","The same analysis shows that the value function is finite, convex, and the unique viscosity solution of the associated Hamilton–Jacobi–Bellman equation, so no second optimizer can beat the Aldous martingale."],"supporting_citations":[{"why":"Supplies the original 'most exciting game' question and motivates the entropy-scaling-limit model.","marker":"[1]"},{"why":"States the max-entropy win-probability martingale problem that the paper aims to solve in higher dimension.","marker":"[2]"},{"why":"Derives the one-dimensional case and identifies the cost $\\mathbb{E}\\int -\\log\\det\\Sigma\\,ds$ with the relevant scaling limit of Shannon entropies, which the present paper adopts.","marker":"[4]"},{"why":"Provides the classical Dirichlet regularity theory for Monge-Amp\\`ere equations used to obtain smooth approximating solutions on convex subdomains.","marker":"[19]"},{"why":"Establishes existence and uniqueness for the infinite-boundary Monge-Amp\\`ere equation, supplying the baseline that Theorem 1.1 refines with boundary asymptotics.","marker":"[43]"},{"why":"Provides the Lyapunov-type stochastic stability results invoked to prove that the Aldous martingale never exits the simplex before the terminal time.","marker":"[51]"}],"fun_headline_variants":["Monge-Ampère equation determines the most exciting game","Most exciting game obeys a single PDE with infinite boundary data","Surprising connection: games and Monge-Ampère equations","Unique optimal martingale emerges from a Monge-Ampère PDE","The most exciting game: solved by a single Monge-Ampère equation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that Aldous's 'excitement' is quantified by the expected integral of $-\\log\\det(\\Sigma_s)$ over time, coming from a scaling limit of Shannon entropies; that equivalence is imported from the earlier one-dimensional treatment rather than re-derived, so if a different notion of suspense is intended, the interpretation of the optimizer as the most exciting game would not follow even though the mathematical optimality result for this cost still stands.","fun_headline_variants_meta":{"raw":{"variants":["Monge-Ampère equation determines the most exciting game","Most exciting game obeys a single PDE with infinite boundary data","Surprising connection: games and Monge-Ampère equations","Unique optimal martingale emerges from a Monge-Ampère PDE","The most exciting game: solved by a single Monge-Ampère equation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000657,"raw_usage":{"total_tokens":3084,"prompt_tokens":1099,"completion_tokens":1985,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":715,"completion_tokens_details":{"reasoning_tokens":1896}},"tokens_in":715,"tokens_out":1985,"duration_ms":12710,"temperature":1.0,"reasoning_tokens":1896,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:57:30.513791+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Starting at any fixed $x\\in\\Delta$ for $d\\ge 2$, simulate the diffusion $dM_s = \\sqrt{2(\\nabla^2g(M_s))^{-1}/(1-s)}\\,dB_s$ with a high-precision numerical solution of the Monge-Amp\\`ere equation; if for some $x$ the terminal law is not concentrated on the vertices, meaning positive probability is observed on an open face of the simplex, Theorem 1.2 is false. Equivalently, exhibit any win-martingale with strictly smaller expected cost $\\mathbb{E}\\int -\\log\\det\\Sigma\\,ds$ than the diffusion achieves; the claimed uniqueness would fail.","supporting_citations":[{"cited_title":"A ldous, notes, https://www.stat.berkeley.edu/ aldous/Research/OP/ent-MG.pdf","cited_arxiv_id":null,"evidence_quote":"Supplies the original 'most exciting game' question and motivates the entropy-scaling-limit model."},{"cited_title":"A ldous, What is the max-entropy win-probability martingale? https://www.stat.berkeley.edu/ al- dous/Research/OP/max ent mg.html","cited_arxiv_id":null,"evidence_quote":"States the max-entropy win-probability martingale problem that the paper aims to solve in higher dimension."},{"cited_title":"B ackhoff-Veraguas and M","cited_arxiv_id":null,"evidence_quote":"Derives the one-dimensional case and identifies the cost $\\mathbb{E}\\int -\\log\\det\\Sigma\\,ds$ with the relevant scaling limit of Shannon entropies, which the present paper adopts."},{"cited_title":"C affarelli, L","cited_arxiv_id":null,"evidence_quote":"Provides the classical Dirichlet regularity theory for Monge-Amp\\`ere equations used to obtain smooth approximating solutions on convex subdomains."},{"cited_title":"G uan and H.-Y","cited_arxiv_id":null,"evidence_quote":"Establishes existence and uniqueness for the infinite-boundary Monge-Amp\\`ere equation, supplying the baseline that Theorem 1.1 refines with boundary asymptotics."},{"cited_title":"K hasminskii, Stochastic stability of di fferential equations, vol","cited_arxiv_id":null,"evidence_quote":"Provides the Lyapunov-type stochastic stability results invoked to prove that the Aldous martingale never exits the simplex before the terminal time."}],"review_version":1}