{"id":"47bdd65f-f652-48e3-841f-7fdb3dca7e27","arxiv_id":"2412.21125","paper_version":3,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For a hypothesis given by finitely many continuous moment constraints with the origin in the relative interior of the feasible convex hull, the optimal e-variable set is the dual class {1 - λ·Φ : λ ∈ Λ_Φ}.","lead":"A statistics paper characterizes, for hypotheses defined by finitely many moment constraints, the optimal pool of e-variables as the dual class 1 minus a linear combination of the constraints. This gives a principled way to simplify e-value tests and yields new confidence sequences for heavy-tailed mean estimation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 4's claim that every constraint of a properly constrained hypothesis is proper is false; this breaks the proof of Lemma 16 as written.","rationale":"The reader correctly identified Lemma 16 as a load-bearing step, but the concrete failure I found is in the lemma on which Lemma 16's proof depends: Lemma 4's assertion that every constraint of a properly constrained hypothesis is proper is false. Adding a zero (or otherwise redundant) component to a proper constraint yields a valid constraint that is not proper, since the convex hull of its image lies in a lower-dimensional subspace. This invalidates the proof of Lemma 16's equivalence, and therefore the compactness and diagonal arguments used in Theorem 1, as written. However, the central claim itself may still be true if the proof is repaired by either restricting Lemma 16 to minimal constraints (which are proper by definition) or first establishing that compactness of Λ_Φ forces Φ to be proper. The paper's conclusion is plausible and independently supported, but the present proof contains a genuine, checkable flaw. A conditional acceptance, requiring correction of Lemma 4 and a re-derivation of Lemma 16, is therefore the appropriate verdict.","tokens_in":27233,"tokens_out":38774,"duration_ms":355727,"concrete_test":"Verify the counterexample: take X={0,1}, Ψ(x)=x−1/2, and Φ=(Ψ,0). Confirm that H={P∈P_X:E[X]=1/2} is properly constrained, that Φ satisfies Definition 5, and that 0∉int conv Φ(X). Then attempt to repair the proof of Lemma 16 by showing directly that Λ_Φ compact implies Φ is proper (via the supporting-hyperplane argument), without invoking the false 'every constraint is proper' claim. If the repair succeeds for all constraints of properly constrained H, the main theorem stands but Lemma 4 must be corrected; if not, the central claim is unproven.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Lemma 4 states: 'if H is properly constrained, every constraint Φ for H is a proper constraint.' This is false. Let H be properly constrained with a proper constraint Ψ (e.g., X={0,1}, Ψ(x)=x−1/2, so H fixes the mean at 1/2). Define Φ=(Ψ,0): X→R^2. Then H={P:⟨P,Φ⟩=0} and Φ is continuous, so Φ is a constraint under Definition 5. But conv Φ(X) is the line segment from (−1/2,0) to (1/2,0), whose interior in R^2 is empty; hence 0∉int conv Φ(X) and Φ is not proper. Thus Lemma 4's 'Moreover' clause is directly contradicted. This matters because Lemma 16's proof applies Lemma 6—which is explicitly stated only for proper constraints—to an arbitrary constraint Φ of a properly constrained H. As written, the proof of Lemma 16, and hence the compactness and diagonal arguments in Corollary 2, Lemma 17, and Theorem 1, rests on an invalid application. The theorem may be repairable by first proving that compactness of Λ_Φ forces Φ to be proper, or by stating Lemma 16 only for minimal constraints, but the current proof is incomplete.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies e-variable testing in a sequential betting framework and characterizes the optimal e-class (the minimal complete class of e-variables) for hypotheses defined by finitely many continuous constraints. The main theorem (Theorem 1) states that for any properly constrained hypothesis H, the dual e-class E^∨_H = {1−λ·Φ : λ∈Λ_Φ} is the optimal e-class, meaning that every e-variable for H is pointwise majorized by some member of the dual class and every member of the dual class is maximal. The proof first establishes the finite-support case, then extends to compact X via a compactness-and-density argument, and finally to general closed X using sigma-compactness. The paper also treats non-properly constrained hypotheses by restricting to a closed support set, and extends the characterization to loosely constrained hypotheses with slack (inequality) constraints. Applications to confidence sequences for mean estimation of bounded and heavy-tailed random variables are provided.","tokens_in":27454,"tokens_out":17433,"duration_ms":153076,"significance":"If Theorem 1 is correct, it gives a complete and explicit characterization of the optimal e-class for a broad class of non-parametric hypotheses, extending the single-constraint result of Clerico (2024). The proof is elementary and self-contained, relying on convex analysis and finite-dimensional linear algebra rather than advanced functional-analytic duality, which makes the result broadly accessible. The characterization is sharp: all dual e-variables are maximal, and every e-variable is dominated pointwise by a dual one. The applications to mean estimation, including heavy-tailed settings, demonstrate practical value. I have checked the concern raised during review about Lemma 4; the alleged counterexample confuses the relative interior with the interior of conv Φ(X), so the proof of Lemma 16 stands.","major_comments":[],"minor_comments":[{"comment":"The text says 'with δ∈(0,1) the type II confidence level'; this should be 'type I confidence level'.","section":"§9.1"},{"comment":"There are several typos: 'Tor many instantiations' should be 'For many instantiations'; 'presisely' should be 'precisely'; 'it is reach enough' should be 'it is rich enough'; and in the proof of Lemma 16, 'identicalluy null' should be 'identically null'.","section":"Introduction and §4"},{"comment":"The proof applies Lemma 6 to the constraint Φ, but Lemma 6 is stated only for proper constraints. Since H is properly constrained, Lemma 4 guarantees that every constraint of H is proper; explicitly citing Lemma 4 here would make the argument easier to follow.","section":"Lemma 16 proof"},{"comment":"The statement says 'Φ is strictly convex on Π_Q(E_H)' and then concludes uniqueness of the maximizer; since Φ(E)=⟨Q, log E⟩ is concave, the word 'convex' should read 'concave'.","section":"Lemma 26"}],"recommendation":"minor_revision","confidential_remarks":"The stress-test concern about Lemma 4 is not valid: the proposed counterexample Φ=(Ψ,0) with X={0,1} satisfies 0∈rel int conv Φ(X), because properness is defined via the relative interior. Thus Lemma 16 and the main theorem are not endangered. The manuscript is a solid contribution; the requested changes are limited to clarifications and typo fixes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The main result — Theorem 1, the dual e-class is optimal for properly constrained hypotheses — is real and new. It extends Clerico's earlier coin-betting result to general continuous constraints, with an elegant proof that goes through finite, compact, then sigma-compact cases. The heavy-tailed mean application is a nice bonus. I believe the paper deserves a serious referee.\n\nThe stress-test note claims Lemma 4 fails because a zero component makes the constraint non-proper. That's a misreading of the definition: properness requires 0 in the relative interior of the convex hull of the image, and the line segment example has 0 in the relative interior. So the concern dissolves. The appendix proof of Lemma 4 looks right.\n\nThe paper's main proof is sound in outline. The biggest actual gaps are deferred to the appendix, including Lemma 16's proof, which is the linchpin for compactness; the appendix arguments are there but I haven't checked every line. Proposition 6 for loose constraints imports Lemma 27 (a KL minimizer result) from the literature without proof; that's fine if the result is standard, but it's worth verifying. There are minor typos and the exposition is a bit dense in places.\n\nThis is for researchers in e-value testing and anytime-valid inference. It gives a clean structural result and consolidates the optimality notion. I'd cite it.\n\nYes, send to peer review. The core claim is well-supported, independently corroborated by Larsson-Ramdas-Ruf, and the proof strategy is transparent. It's not a blockbuster but it's a solid contribution.","headline":"A clean, credible characterization of optimal e-classes for properly constrained hypotheses; the main theorem holds up, and the stress-test concern about Lemma 4 rests on a rel-int/int mix-up.","tokens_in":27982,"tokens_out":2336,"would_cite":true,"duration_ms":21569,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62L10","62C15","62F03"],"pacs":[],"model":"deepseek-v4-flash","headline":"For any properly constrained hypothesis, the optimal pool of e-variables is exactly the dual class built from the constraint map.","keywords":["e-values","e-variables","testing by betting","optimal e-class","dual e-class","properly constrained hypotheses","confidence sequences","heavy-tailed mean estimation"],"falsifier":"Take $X=\\mathbb{R}^2$ with a full-dimensional convex hull and the properly constrained hypothesis $H=\\{P:\\langle P,x_1\\rangle=\\mu,\\ \\langle P,x_2\\rangle=\\nu\\}$ for $(\\mu,\\nu)$ in the relative interior of $\\operatorname{conv}X$; for a grid of candidate bounded e-variables $E$ (e.g., truncated likelihood ratios), numerically compute $\\sup_{\\lambda\\in\\Lambda_\\Phi}\\inf_{x\\in X}[(1-\\lambda\\cdot\\Phi(x))-E(x)]$. Theorem 1 predicts this value is nonnegative for every $E$; finding a single $E$ with a negative value would disprove the theorem.","tokens_in":27007,"feed_emoji":"🎲","tokens_out":10676,"duration_ms":92847,"temperature":0.7,"pith_summary":"The paper establishes that for a broad class of non-parametric hypotheses defined by finitely many continuous expectation constraints, the optimal set of single-round e-variables is precisely the dual e-class: all functions $1-\\lambda\\cdot\\Phi$ for those $\\lambda$ with $\\sup_{x\\in X}\\lambda\\cdot\\Phi(x)\\le 1$. Optimal here means that every e-variable is pointwise dominated by a member of this class, every member is maximal, and any other class with the domination property must contain it. In betting-game terms, restricting a player to this class loses no statistical power and removes every redundant e-variable. This extends the coin-betting mean-estimation result of the author's earlier work to a much wider family of hypotheses and yields explicit optimal classes for bounded and heavy-tailed mean estimation.","feed_headline":"Optimal e-variables for constrained hypotheses form a dual class","feed_subtitle":"A theorem pins down the minimal complete pool of e-variables for constrained hypotheses.","key_machinery":"The central object is the dual e-class $\\mathcal{E}^\\vee_H=\\{1-\\lambda\\cdot\\Phi:\\lambda\\in\\Lambda_\\Phi\\}$, the set of affine functions of the constraint map $\\Phi$ whose expectation is exactly one under every $P\\in H$. Each member is automatically non-negative precisely because $\\lambda\\cdot\\Phi\\le 1$ everywhere on $X$, so the dual class turns the linear constraint $\\langle P,\\Phi\\rangle=0$ into a family of pointwise-bounded, exact-mean-one e-variables. The proof machinery that carries the result from finite to infinite sample spaces is the notion of a matching set---a closed subset $S\\subseteq X$ on which some $P\\in H$ is supported---together with Lemma 16, which equates minimality of $\\Phi$ (linear independence of its components) with compactness of the parameter sets $\\Lambda_{\\Phi|_S}$ and guarantees a finite matching set $S_\\star$; this compactness drives the diagonal-subsequence arguments in Lemma 17, Proposition 5, and Theorem 1.","core_discovery":"Theorem 1 asserts that if $H$ is a properly constrained hypothesis on a closed set $X\\subseteq \\mathbb{R}^n$---meaning $H=\\{P\\in\\mathcal{P}_\\Phi:\\langle P,\\Phi\\rangle=0\\}$ for a continuous $\\Phi:X\\to\\mathbb{R}^m$ with $0$ in the relative interior of $\\operatorname{conv}\\Phi(X)$---then the dual e-class $\\mathcal{E}^\\vee_H=\\{1-\\lambda\\cdot\\Phi:\\lambda\\in\\Lambda_\\Phi\\}$, with $\\Lambda_\\Phi=\\{\\lambda:\\sup_{x\\in X}\\lambda\\cdot\\Phi(x)\\le 1\\}$, is the optimal e-class. Optimality is meant in the poset-theoretic sense: every e-variable for $H$ is pointwise majorised by some member of $\\mathcal{E}^\\vee_H$; every member of $\\mathcal{E}^\\vee_H$ is maximal, so no larger class can be minimal; and $\\mathcal{E}^\\vee_H$ is contained in every majorising e-class, making it the unique minimal complete class of admissible e-variables. The proof is constructive and elementary: the finite-sample-space case is proved by a vertex-polytope argument, the notion of matching sets transfers the result to compact $X$, and $\\sigma$-compactness of closed $X$ finishes the general case.","pith_inferences":["Since every member of the dual class has expectation exactly one under every null distribution, the optimal e-variables are the non-parametric analogue of likelihood-ratio boundaries; the paper does not draw the connection, but $\\mathcal{E}^\\vee_H$ can be read as the complete class of 'exact' e-variables, paralleling the role of the natural sufficient statistic in exponential families.","A natural conjecture, not stated in the paper: in any sequential testing game over i.i.d. draws, choosing each round's dual-class e-variable with $\\lambda$ adapted to past data attains the optimal capital process, so the closure of products of dual e-variables should coincide with the class of admissible e-processes for these hypotheses.","For the heavy-tailed mean problem with bounded central moment $\\langle|X-\\mu|^{1+\\varepsilon}\\rangle\\le B$, Proposition 6 gives an explicit dual class; benchmarking the widths of the resulting confidence sequences against the Catoni-style sequences cited in the paper would yield a quantitative measure of the improvement from using the optimal class, since the paper notes those sequences do not res"],"forward_implications":["For any properly constrained hypothesis $H$, a testing-by-betting game can be restricted to the dual e-class without loss: every strategy that uses arbitrary e-variables is pointwise dominated by a strategy that uses only members of $\\mathcal{E}^\\vee_H$.","Since every member of $\\mathcal{E}^\\vee_H$ has expectation exactly one under every null distribution, each is a maximal e-variable; consequently the optimal e-class coincides with the full set of maximal e-variables whenever it exists.","For mean estimation with $X$ compact and $\\mu$ in the relative interior of $\\operatorname{conv}X$, the optimal e-class for $H_\\mu=\\{P:\\langle P,X\\rangle=\\mu\\}$ is $\\{1+\\lambda\\cdot(x-\\mu):\\sup_{x\\in X}\\lambda\\cdot(x-\\mu)\\le 1\\}$, subsuming the coin-betting class for $X=[0,1]$.","For heavy-tailed mean estimation under $\\langle P,X^2\\rangle\\le 1$, the optimal e-class is $\\{1+\\alpha(x-\\mu)+\\beta(x^2-1):(\\alpha,\\beta)\\in \\widetilde\\Lambda_{\\Phi'_\\mu,\\Phi''}\\}$ with an explicit elliptic parameter region, giving confidence sequences with an optimal single-round betting class.","For finitely constrained hypotheses that are not properly constrained, no optimal e-class exists; the correct reduction is to test on the closed support $X_0=\\bigcup_{P\\in H}\\operatorname{Supp}P$, where the hypothesis becomes properly constrained, and reject immediately if data fall outside $X_0$."],"supporting_citations":[{"why":"Defines majorising and optimal e-classes and proves the base coin-betting case that this paper generalises.","marker":"Clerico (2024)"},{"why":"Supplies the betting-game formulation and Ville's-inequality guarantee (Proposition 1) that motivate restricting the pool of e-variables.","marker":"Ramdas et al. (2023)"},{"why":"Provides the e-value testing framework and the construction of confidence sequences used in Section 10.","marker":"Ramdas and Wang (2024)"},{"why":"Shows that the coin-betting class for mean estimation yields tight confidence sequences via universal-portfolio regret bounds, a baseline the dual class matches.","marker":"Orabona and Jun (2023)"},{"why":"Uses the coin-betting class for bounded-mean estimation, which the paper's dual class reproduces for $X=[0,1]$.","marker":"Waudby-Smith and Ramdas (2023)"},{"why":"Supplies the convex-analysis results (relative interiors, affine hulls, polytope interior properties) used in Lemmas 2, 10, and 24.","marker":"Hiriart-Urrut and Lemaréchal (2004)"},{"why":"Provides the reverse information projection and log-optimal e-variable results used in Appendix C to handle loosely constrained hypotheses.","marker":"Grünwald et al. (2024)"},{"why":"Together with Grünwald et al., supplies the existence and uniqueness of the log-optimal e-variable used to prove Proposition 6.","marker":"Larsson et al. (2024)"}],"fun_headline_variants":["Dual e-class is optimal for properly constrained hypotheses","Theorem: dual e-class is the unique minimal complete pool","Optimal e-variables: dual class for constrained hypotheses","Dual e-class characterizes optimal e-variables"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The theorem is conditional on $H$ being properly constrained, meaning $0$ lies in the relative interior of $\\operatorname{conv}\\Phi(X)$, and on the pointwise-majorisation notion of optimality; inside the proof, the load-bearing premise is Lemma 16, which equates minimality of $\\Phi$ with compactness of $\\Lambda_\\Phi$ and existence of a finite matching set, since that compactness is what allows the diagonal arguments to pass from finite to general closed sample spaces.","fun_headline_variants_meta":{"raw":{"variants":["Dual e-class is optimal for properly constrained hypotheses","Theorem: dual e-class is the unique minimal complete pool","Optimal e-variables: dual class for constrained hypotheses","Dual e-class characterizes optimal e-variables"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000392,"raw_usage":{"total_tokens":2047,"prompt_tokens":920,"completion_tokens":1127,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":1063}},"tokens_in":536,"tokens_out":1127,"duration_ms":8951,"temperature":1.0,"reasoning_tokens":1063,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:02:39.210016+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take $X=\\mathbb{R}^2$ with a full-dimensional convex hull and the properly constrained hypothesis $H=\\{P:\\langle P,x_1\\rangle=\\mu,\\ \\langle P,x_2\\rangle=\\nu\\}$ for $(\\mu,\\nu)$ in the relative interior of $\\operatorname{conv}X$; for a grid of candidate bounded e-variables $E$ (e.g., truncated likelihood ratios), numerically compute $\\sup_{\\lambda\\in\\Lambda_\\Phi}\\inf_{x\\in X}[(1-\\lambda\\cdot\\Phi(x))-E(x)]$. Theorem 1 predicts this value is nonnegative for every $E$; finding a single $E$ with a negative value would disprove the theorem.","supporting_citations":[],"review_version":1}