REVIEW 4 minor 3 cited by
Optimal e-value testing for properly constrained hypotheses
T0 review · 0 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read For any properly constrained hypothesis, the optimal pool of e-variables is exactly the dual class built from the constraint map.
desk verdict 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. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
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.
What would settle it
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.
Extended reading notes
Core claim
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.
Load-bearing premise
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.
Editorial extensions
If this is right
- 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 $X_0$.
Reading between the lines
- 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
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
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.
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.
minor comments (4)
- [§9.1] The text says 'with δ∈(0,1) the type II confidence level'; this should be 'type I confidence level'.
- [Introduction and §4] 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'.
- [Lemma 16 proof] 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.
- [Lemma 26] 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'.
Circularity Check
No circular reduction found: optimality of the dual e-class is derived from independent definitions; self-citations are minor and non-load-bearing.
full rationale
The central claim (Theorem 1: 'Let H be a properly constrained hypothesis. Then, the dual e-class E^∨_H is optimal.') is genuinely derived, not assumed. The dual e-class of Definition 7 builds only on the constraint Φ and the set Λ_Φ = {λ : sup_x λ·Φ(x) ≤ 1}, independently of the optimality notion of Definition 3, and both directions are argued in-paper: every dual element is maximal (Lemma 9, via the atom-charging bound of Lemma 4), and every e-variable is pointwise majorised by a dual element (Proposition 4 for finite X; Proposition 5 for compact X via matching-set compactness from Lemmas 16–17; the sigma-compact diagonal argument in the proof of Theorem 1 for closed X). No quantity is fitted and no empirical prediction is made; the Section 10 applications instantiate the theorem to construct confidence sequences. The self-citations to Clerico (2024) are limited to the game terminology (Definition 2), the majorising/optimal concepts (Definition 3), Proposition 1 (a standard Ville-inequality bound), and Lemma 1, whose proof is imported ('For the proof, see Lemmas 3 and 4 in Clerico (2024)'). Lemma 1 is an elementary admissible/complete-class criterion whose stated assumptions do not include the target result, so it is a minor, non-load-bearing self-citation rather than a circular reduction, placing the score at the 1–2 boundary. The identification of the [0,1] case with the 'coin-betting e-class studied in Clerico (2024)' is an honest attribution of a known special case. The skeptical attack on Lemma 4 fails as stated: for Φ = (Ψ,0) over X = {0,1}, 0 lies in the relative interior of the segment conv Φ(X), so Φ is proper; the attack conflates properness with minimality. However, the appendix proof of Lemma 4 does contain an invalid step ('there is x0 ∈ X such that Φ(x0) ∈ rel int C', contradicted by any Φ whose image avoids the relative interior), leaving a soundness gap in the chain that applies Lemma 6 inside Lemma 16; this is a correctness risk, not a circularity. The main characterization has also been independently re-derived by Larsson et al. (2025), supporting that it is not forced by definition.
Assumptions & free parameters
assumptions (5)
- domain assumption X is a closed subset of R^n and Φ is continuous.
- domain assumption H is properly constrained: 0 ∈ rel int conv Φ(X).
- domain assumption Data are i.i.d. draws from P ∈ H; only single-round e-variables are considered.
- standard math Standard finite-dimensional convex analysis results (Carathéodory, separating hyperplane, relative interior geometry).
- domain assumption Lemma 27 (reverse information projection / KL minimizer) from Gruenwald et al. (2024) and Larsson et al. (2024).
Cite this review
Pith. "Pith review of Optimal e-value testing for properly constrained hypotheses." pith.science (2026). https://pith.science/paper/32UG5DHV
@misc{pith2026241221125,
author = {Pith},
title = {Pith review of: Optimal e-value testing for properly constrained hypotheses},
year = {2026},
howpublished = {\url{https://pith.science/paper/32UG5DHV}},
note = {Machine review of arXiv:2412.21125}
}
read the original abstract
Hypothesis testing via e-variables can be framed as a sequential betting game, where a player each round picks an e-variable. A good player's strategy results in an effective statistical test that rejects the null hypothesis as soon as sufficient evidence arises. Building on recent advances, we address the question of restricting the pool of e-variables to simplify strategy design without compromising effectiveness. We extend the results of Clerico(2024), by characterising optimal sets of e-variables for a broad class of non-parametric hypothesis tests, defined by finitely many regular constraints. As an application, we discuss this notion of optimality in algorithmic mean estimation, including for heavy-tailed random variables.
Forward citations
Cited by 3 Pith papers
-
Optimal e-values for testing the mean of a bounded random variable against a composite alternative
For bounded-mean testing without absolute continuity, GROW and REGROW optimal e-variables exist and are exactly coin-betting e-values Eα(x)=1+α(x−µ0), with α given by explicit formulas.
-
On Stopping Times of Power-one Sequential Tests: Tight Lower and Upper Bounds
For arbitrary composite nulls and alternatives, any power-one sequential test needs at least log(1/alpha)/KL_inf samples when alpha is small, and at least a law-of-iterated-logarithm scale when the alternative is clos...
-
Uniform mean estimation for monotonic processes
Coin-betting plus a monotonicity-based continuous union bound yields uniform, anytime-valid, variance-adaptive confidence bands for monotonic mean functions such as CDFs.
Reference graph
Works this paper leans on
-
[1]
Then, the relative interior of PX (which can be seen as a subset of Rd) is the set {P ∈ PX : Supp P = X }. Proof. Let Rd ≥0 denote the subset of Rd of vectors whose components are all non-negative. Its relative interior is Rd >0, the set of vectors with only strictly positive components. Let 1 ∈ Rd denote the vector with all components equal to one and V ...
work page 2004
-
[3]
If X is countable and supP ∈H P ({x}) > 0 for all x ∈ X, the optimal e-class exists. Proof. This proof makes use of elementary properties of ordinals. For an accessible reference, see Aliprantis and Border (2006). By Lemma 1, it is enough to show that every e-variable is majorised by a maximal e-variable, which implies that the set of all maximal e-variab...
work page 2006
- [6]
-
[7]
13 R. Wang. The only admissible way of merging e-values. arXiv:2409.19888,
-
[12]
By Lemma 27, we have that ˆE ∈ arg maxE∈EH ⟨ ˆQ, log E⟩
Define ˆQ ∈ PX with mass ˆQ({x}) = ˆE(x) ˆP ({x}). By Lemma 27, we have that ˆE ∈ arg maxE∈EH ⟨ ˆQ, log E⟩. Moreover, Supp ˆQ = Supp ˆE = {x : E(x) > 0}. Writing the dual problem for the maximisation of E 7→ ⟨ˆQ, log E⟩, a solution must be in the form Eλ (on Supp ˆQ), with λ ∈ ˜ΛΦ′,Φ′′ . (This follows from usual dual Lagrangian arguments, and can for inst...
work page 2020
-
[20]
, vr} ⊆Rm be a family of vectors that spans Rm
Let {v1, . . . , vr} ⊆Rm be a family of vectors that spans Rm. Let {(Ai, Bi)}r i=1 be a family of non-empty bounded intervals. Then, the set Λ = {λ ∈ Rm : λ · vi ∈ [Ai, Bi] , ∀i = 1 . . . r} is bounded. 21 Proof. Since {v1, . . . , vr} spans Rm, each element ei of the standard basis of Rm can be written as a linear combination ei = Pr j=1 αijvj. Then, if ...
work page 2004
-
[26]
Then, for any Q ∈ PX , the set arg maxE∈EH ⟨Q, log E⟩ is non-empty
Let H be a properly loosely constrained hypothesis on X and X have finite cardi- nality. Then, for any Q ∈ PX , the set arg maxE∈EH ⟨Q, log E⟩ is non-empty. Moreover, all its elements coincide on Supp Q. Proof. Fix Q ∈ PX . Since EH is compact and Φ : E 7→ ⟨Q, log E⟩ = P x∈Supp Q Q({x}) logE(x) is, upper semi-continuous, concave, and valued on [−∞, +∞), w...
work page 2024
-
[1976]
1 M. Larsson, A. Ramdas, and J. Ruf. The numeraire e-variable and reverse information projection. arXiv:2402.18810,
Show all 12 references
-
[2006]
1 E. Clerico. On the optimality of coin-betting for mean estimation. arXiv:2412.02640,
-
[2022]
14 F. Orabona. A modern introduction to online learning. arXiv:1912.13213,
1912 arXiv
- [2023]
-
[2024]
Larsson, A
2, 23 M. Larsson, A. Ramdas, and J. Ruf. E-variables for hypotheses generated by constraints. arxiv:2504.02974,
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.