{"id":"dd21e3f7-a58c-49e8-9f04-4db54e73c1ac","arxiv_id":"2506.16253","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"The optimal worst-case bookmaking loss for K outcomes and T rounds is the largest root of an explicit polynomial, with regret scaling as sqrt(T) times the largest Hermite root.","lead":"This paper solves the online bookmaking problem for any number of possible outcomes, giving an exact formula for the bookmaker's worst-case loss. It also shows that bookmakers can set near-fair odds and provides an efficient adaptive strategy.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem B's proof normalizes the wrong polynomial: the stated limit does not exist, so the Hermite-root conclusion is not established by the written argument.","rationale":"I read the paper in good faith and traced the main proof chain. The core characterization for the finite-horizon loss (Theorem A) is supported by a substantial dynamic-programming argument: Theorem 9's decisive-gambler reduction and equalization property rely on the convexity of the value function (Lemma 31), the compact correspondence Z_H (Lemma 27), and strict monotonicity (Lemma 29). I checked the logic of these lemmas, including the induction in Lemma 31 and the use of Jensen's inequality in Lemma 28; I found no counterexample or circularity, and the small cases K=2,3 and T=1,2 match the polynomial roots. The Bellman-Pareto frontier characterization (Theorem 12, Lemma 14) also appears internally consistent, with the base case and inductive step checking out for H=1,2 and K=2. The one substantive flaw I found is in the proof of Theorem B: the polynomial eP_{T,K}(x) is defined with an explicit T^{K/2} factor, so it does not converge to He_K(x) as written; the proof silently treats the normalized polynomial as the object of study. The theorem is almost certainly true because root locations are unchanged by scalar rescaling, but the argument as printed is missing the normalization step and the root-convergence justification. Because Theorem B is a headline contribution highlighted in the abstract, this gap warrants a conditional acceptance rather than unconditional acceptance. The reader's identified weakest assumption (Theorem 9) is not where I found the problem, hence I disagree with the reader's diagnosis while agreeing that the paper is otherwise strong.","tokens_in":37502,"tokens_out":35902,"duration_ms":361357,"concrete_test":"For K=4, compute β_{T,4}=R_{T,4}/√T from the largest real root of P_{T,4} in Theorem A for T=10^4,10^6,10^8 using high-precision arithmetic, and compare against the largest root of He_4(x)=x^4-6x^2+3, i.e., sqrt(3+√6)≈2.334. Also evaluate the normalized polynomial Q_T at a fixed non-root point (e.g., x=1) and check Q_T(1)→He_4(1)=1-6+3=-2. If the ratios converge to sqrt(3+√6) and Q_T(1)→-2, the normalization error is purely presentational; if not, Theorem B's statement is in doubt.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Theorem B is a central contribution, but its proof contains a normalization error. From (D.5), β_{T,K} is the largest root of eP_{T,K}(x)=bP_{T,K}(√T x). Equation (D.6) gives eP_{T,K}(x)=T^{K/2} ∑_{m=0}^K x^{K-m} C(K,m) ec_{T,m}, i.e., eP_{T,K} carries an explicit factor T^{K/2}. The proof then claims 'lim_{T→∞} eP_{T,K}(x) = He_K(x)'. This limit is false whenever He_K(x)≠0: the polynomial diverges like T^{K/2}He_K(x). What actually converges is the normalized polynomial Q_T(x)=T^{-K/2}eP_{T,K}(x)=∑_{m=0}^K C(K,m)ec_{T,m}x^{K-m}, whose roots coincide with those of eP_{T,K}. The written proof never defines Q_T or proves that its largest root converges to that of He_K; it simply extracts the coefficient of T^{K/2} and calls it the limit. Root-convergence for degree-K polynomials with convergent coefficients is a standard fact, so the claim is very likely true, but the argument as printed is incomplete/incorrect at this step. Since Theorem B is highlighted in the abstract (Hermite polynomial relation), this is a load-bearing gap in the central proof chain.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies a repeated zero-sum game between a bookmaker and a gambler with K possible outcomes and T betting rounds. The bookmaker chooses odds, the gambler chooses bets, and the bookmaker's worst-case loss is the maximal committed payout across outcomes. The main result, Theorem A, states that the optimal loss is the largest root of the explicit polynomial P_{T,K}(x)=∑_{m=0}^K C(K,m)(-T)^{K-m}x^m, where the superscript denotes the rising factorial. Theorem B claims that the regret R_{T,K}=L*_{T,K}-T grows like β_K√T, with β_K equal to the largest root of the K-th probabilist's Hermite polynomial He_K. The paper also introduces the Bellman-Pareto frontier, proves Theorem C characterizing the optimal opportunistic bookmaking loss and an O(TK^2) algorithm achieving it, and shows that bookmakers can make the overround arbitrarily close to 1 for large T. The proofs are carried out through a dynamic-programming value function, a Nash-equilibrium theorem for decisive gamblers, and a polynomial characterization of achievable residual loss vectors.","tokens_in":37724,"tokens_out":13586,"duration_ms":155562,"significance":"If the results hold, this is a substantial contribution: it gives an exact, closed-form characterization of the optimal loss for all K and T, extends the binary result of Bhatt et al. (2025) to arbitrary numbers of outcomes, provides a tractable algorithm with per-round complexity O(K^2), and establishes a surprising connection between bookmaking regret and Hermite polynomials. The paper is largely self-contained and parameter-free: no constants are fitted to data, and the central polynomial characterization is derived from first principles. I verified the K=2 and K=3 special cases and the structure of the Appendix D.1 coefficient extraction; the main proof architecture is coherent. However, the proof of Theorem B as printed contains a load-bearing normalization error: it claims a pointwise limit of polynomials that actually diverge by a factor of T^{K/2}. The claim is very likely true after an easy rescaling, but the written argument does not establish it. For this reason the appropriate verdict is major revision rather than acceptance in the current form.","major_comments":[{"comment":"The proof of Theorem B does not establish the stated pointwise limit. Equation (D.6) gives eP_{T,K}(x) = T^{K/2} ∑_{m=0}^K C(K,m) ec_{T,m} x^{K-m} with ec_{T,m} having powers of T at most 0 by Lemma 40. Consequently, for a fixed x with He_K(x) ≠ 0, eP_{T,K}(x) diverges in magnitude like T^{K/2}He_K(x); it does not converge to He_K(x). What is true is that the roots of eP_{T,K} are the same as the roots of Q_{T,K}(x) := T^{-K/2}eP_{T,K}(x). The proof should define Q_{T,K}, prove its coefficients converge to those of He_K (using the coefficient identities already present, including the m=2n terms and the Gould et al. identity), and then use the standard fact that roots of a monic polynomial depend continuously on its coefficients, or Hurwitz's theorem, to conclude β_K is the largest root of He_K. As printed, the limit step in Theorem B's proof is incorrect, and the Hermite relation asserted in the abstract is not justified by the written argument.","section":"Appendix D.1, Eq. (D.6) and proof of Theorem B"}],"minor_comments":[{"comment":"In the forward direction, after setting s^ = -u and x = V_H(s^), the proof rules out x < 0 using Lemma 14.3 but does not explicitly rule out x > 0. This case should be excluded: if x > 0, then the vector x·1 - s^ = u + x·1 strictly dominates u, so by Lemma 14.3 it cannot be H-achievable, contradicting Theorem 9. An added sentence would make the argument complete.","section":"Section 4.2, proof of Theorem 12"},{"comment":"The notation ec_{T,m} is used for a quantity that depends on T, but the subscript notation suggests a constant independent of T. Renaming this coefficient (for example c_{m,T}) would improve readability, especially since Theorem B takes limits in T.","section":"Appendix D.6, Eq. (D.6)"},{"comment":"The last displayed equation uses the symbol m in binom(m,d) even though the summation has been reindexed to n; this appears to be a typo for binom(n,d). Please correct it.","section":"Appendix E.4.1, final display of the derivation of (D.4)"},{"comment":"The notation i_1 for the smallest positive root of the Airy function is easy to confuse with the imaginary unit; a name such as a_1 would be clearer.","section":"Equation (13)"},{"comment":"The sentence about UK football betting reads 'the overround typically 1 remains below 1.1'; the stray numeral '1' should be removed.","section":"Section 1, first paragraph"}],"recommendation":"major_revision","confidential_remarks":"I want to stress that the central polynomial characterization in Theorem A and the algorithmic results in Theorem C appear sound and well supported by the appendix proofs. The only load-bearing issue I found is the normalization error in Theorem B's proof, and it is readily fixable by rescaling eP_{T,K} by T^{-K/2} before taking limits. If the authors repair this step and add the missing one-sentence argument in the proof of Theorem 12, the paper should be acceptable. I recommend major revision because Theorem B is highlighted in the abstract and its proof, as currently written, does not establish the stated Hermite limit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Good paper. The exact minimax loss for any K as the largest root of a degree-K polynomial is genuinely new, and the Bellman-Pareto frontier framework is a nice technical contribution. The special cases check out, the reduction to decisive gamblers is proven carefully, and the algorithm is clean. This deserves a serious referee.\n\nThe soft spot is in Appendix D.1, proof of Theorem B. Equation (D.6) defines eP_{T,K}(x) with an explicit factor T^{K/2}. The proof then claims lim_{T→∞} eP_{T,K}(x) = He_K(x). That limit is false for any x where He_K(x) ≠ 0; the polynomial diverges. What actually converges is the normalized polynomial Q_T(x)=T^{-K/2}eP_{T,K}(x). The roots of Q_T are identical to those of eP_{T,K}, so the argument can be repaired by defining Q_T and invoking standard root-continuity. But as printed, the step is incorrect and the Hermite-root conclusion is not established by the written argument. This is a load-bearing gap because Theorem B is one of the two headline results. The fix is short, but it must be made.\n\nMinor: the text is dense and some displayed formulas are garbled in the arXiv version, which slows reading. No code is shipped, but the algorithm is simple enough that this is not a serious issue. The reliance on Bhatt et al. is limited and properly attributed.\n\nIf I were refereeing, I'd recommend accept after this revision. The main polynomial characterization (Theorem A) and the opportunistic algorithm stand independently of the Hermite asymptotics, and I see no issue with the core proof chain there.","headline":"Genuinely new minimax characterization for all K, but the proof of Theorem B has a real normalization slip that needs fixing before the Hermite connection is established.","tokens_in":38285,"tokens_out":2314,"would_cite":true,"duration_ms":26931,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["91A05","91A20","91A50","05E05","33C45"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper derives the exact worst-case loss of an online bookmaker facing an adversarial gambler, for any number of event outcomes and any betting horizon, and shows it equals the largest root of a degree-K polynomial built from rising…","keywords":["online bookmaking","minimax value","Bellman-Pareto frontier","elementary symmetric polynomials","Hermite polynomials","regret","vector repeated games","opportunistic strategy"],"falsifier":"Fix $T=3$, $K=2$ and compute $V_3(0)$ by exact backward induction over a fine discretization of the simplex of bookmaker actions. The theorem implies the value equals $3+\\sqrt{3}\\approx 4.732$, the largest root of $x^2-6x+6$. A brute-force computation that yields a strictly larger quantity for some gambler strategy, or a simulation of Algorithm 1 against a decisive gambler whose final loss exceeds $3+\\sqrt{3}$, would refute the central claim.","tokens_in":37239,"feed_emoji":"🎲","tokens_out":6460,"duration_ms":69894,"temperature":0.7,"pith_summary":"The paper asks how much an online bookmaker must be willing to lose when a gambler bets one unit per round on one of K mutually exclusive outcomes and the bookmaker's odds are the only tool. For any horizon $T$ and any $K$, the worst-case loss over all gamblers and outcome realizations is shown to be exactly equal to the largest real root of a degree-$K$ polynomial whose coefficients are built from rising factorials. The same characterization yields an efficient, forward-running strategy that achieves this loss, and shows that the regret — the loss in excess of the $T$ units collected — grows like $\\sqrt{T}$ times the largest root of the $K$-th Hermite polynomial. Because the per-round loss approaches $1$, the bookmaker can set the overround arbitrarily close to $1$ (nearly fair odds) while still guaranteeing a profit, provided $T$ is large enough.","feed_headline":"Optimal bookmaking loss is a polynomial's largest root","feed_subtitle":"Exact minimax value and an efficient strategy for any number of outcomes; regret scales like sqrt(T).","key_machinery":"The Bellman-Pareto frontier $\\mathcal{P}_{H,K}$, the set of all residual loss vectors that a bookmaker can force against a decisive gambler over $H$ rounds. It carries the argument by reducing the dynamic program $$V_H(s)=\\inf_r\\max_q V_{H-1}(s+q\\oslash r)$$ to a static polynomial identity: the residual vector $v=L^\\star_{H,K}(s)\\mathbf{1}_K - s$ satisfies $D_{H,K}(v)=0$ with $$D_{H,K}(v)=\\sum_{m=0}^K (-H)^{K-m}\\sigma_m(v),$$ and the optimal odds are $r(k)=D_{H-1,K-1}(v^{\\setminus k})/D_{H-1,K}(v)$. The recurrence $D_{H,K}(v)=v(k)D_{H,K-1}(v^{\\setminus k}) - H\\,D_{H-1,K-1}(v^{\\setminus k})$ is the engine that expresses everything through elementary symmetric polynomials.","core_discovery":"The central discovery is that the minimax value of the online bookmaking game has a closed form: $L^\\star_{T,K}$ is the largest root of $$P_{T,K}(x)=\\sum_{m=0}^K \\binom{K}{m}(-T)^{K-m}x^m,$$ where $(-T)^{K-m}$ is the rising factorial. Equivalently, the asymptotic regret factor $\\beta_K = \\lim_T (L^\\star_{T,K}-T)/\\sqrt{T}$ is the largest root of the $K$-th probabilist's Hermite polynomial. The proof proceeds through a Bellman-Pareto frontier: the set of residual payout vectors achievable against a decisive gambler is exactly $$\\mathcal{P}_{H,K}=\\{L^\\star_{H,K}(s)\\mathbf{1}_K - s : s\\in\\mathbb{R}^K,\\ \\min_k s(k)=0\\},$$ and this frontier is characterized by a polynomial $D_{H,K}$. The frontier yields a Nash equilibrium — the worst-case gambler can be taken to bet one-hot each round, and the optimal bookmaker balances losses so that every decisive betting sequence ends at the same total — together with an algorithm whose per-round complexity is $O(K^2)$ and whose odds are given by a ratio of two evaluations of $D$.","pith_inferences":["Editorial extension: the polynomial-root formula suggests that the value function is an algebraic function of the state for finite horizons; introducing noise or asymmetric outcomes might yield a perturbed polynomial whose roots still determine the value — a testable conjecture.","Editorial extension: the water-filling optimal strategy implies that the bookmaker's worst-case loss is insensitive to how the gambler mixes across outcomes, so the same policy could serve as a risk-parity rule in continuous prediction markets.","Editorial extension: for large $K$, the largest Hermite root behaves like $\\sqrt{2K}$ with Airy-type corrections, so one could test whether finite-$T$ regret approaches $\\sqrt{T}\\beta_K$ from above or below as a function of $K$, a question not settled by the asymptotic theorem."],"forward_implications":["For any fixed $K$ and $T$, the optimal bookmaking loss can be computed by finding the largest root of a degree-$K$ polynomial, so the computational cost does not scale with the horizon $T$.","The regret $R_{T,K}=L^\\star_{T,K}-T$ scales as $\\sqrt{T}$ for every $K$, with scaling factor given by the largest root of the $K$-th Hermite polynomial; hence bookmakers can be as fair as desired while remaining profitable as $T$ grows.","Algorithm 1 achieves the optimal loss against an optimal (decisive) gambler, and whenever the gambler deviates from optimality the algorithm lowers the 'water level' to the optimal opportunistic loss.","The characterization subsumes the previously known binary case $K=2$, where the loss is $T+\\sqrt{T}$, and extends it to arbitrary $K$.","The Bellman-Pareto frontier provides an explicit, efficiently computable description of all achievable residual payout vectors, which the paper argues may serve as a general tool for vector-valued repeated games."],"supporting_citations":[{"why":"Introduced the online bookmaking problem and solved the binary case $K=2$; supplies the decisive-gambler reduction and the ODG algorithm that Algorithm 1 generalizes.","marker":"Bhatt et al. (2025)"},{"why":"Provides explicit bounds on the largest root of Hermite polynomials, used in Theorem B to state the bounds on the asymptotic regret factor.","marker":"Krasikov (2004)"},{"why":"Supplies the alternating binomial-sum identity with Stirling numbers that converts the polynomial limit into the Hermite polynomial in the proof of Theorem B.","marker":"Gould et al. (2015)"},{"why":"Gives the probabilist's Hermite polynomial definition used to recognize the limiting polynomial as $He_K$.","marker":"Patarroyo (2019)"},{"why":"Standard reference for the elementary symmetric polynomial recurrence relations (Lemma 24) that underlie the Bellman-Pareto polynomial algebra.","marker":"Macdonald (1998)"}],"fun_headline_variants":["Optimal bookmaking loss: root of a Hermite polynomial","Exact minimax value for bookmaking: polynomial's largest root","Bookmaker's worst-case loss is a polynomial's largest root","Hermite polynomial determines optimal bookmaking regret","Optimal odds for any number of outcomes via polynomial roots"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proof rests on the claim (Theorem 9) that the worst-case gambler can always be taken to bet on a single outcome each round and that the optimal bookmaker response is unique and equalizes the loss across all such decisive bets; this is established through convexity of the value function and existence of a compact set of safe odds. If either property fails for some $K$ or $H$, the polynomial-root formula would not give the true worst-case loss.","fun_headline_variants_meta":{"raw":{"variants":["Optimal bookmaking loss: root of a Hermite polynomial","Exact minimax value for bookmaking: polynomial's largest root","Bookmaker's worst-case loss is a polynomial's largest root","Hermite polynomial determines optimal bookmaking regret","Optimal odds for any number of outcomes via polynomial roots"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000809,"raw_usage":{"total_tokens":3591,"prompt_tokens":1030,"completion_tokens":2561,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":646,"completion_tokens_details":{"reasoning_tokens":2479}},"tokens_in":646,"tokens_out":2561,"duration_ms":20672,"temperature":1.0,"reasoning_tokens":2479,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:46:01.291659+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fix $T=3$, $K=2$ and compute $V_3(0)$ by exact backward induction over a fine discretization of the simplex of bookmaker actions. The theorem implies the value equals $3+\\sqrt{3}\\approx 4.732$, the largest root of $x^2-6x+6$. A brute-force computation that yields a strictly larger quantity for some gambler strategy, or a simulation of Algorithm 1 against a decisive gambler whose final loss exceeds $3+\\sqrt{3}$, would refute the central claim.","supporting_citations":[],"review_version":1}