REVIEW 3 major objections 3 minor 12 references
Gaffke's lower confidence bound is optimal among bounds that share its sample order.
Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →
T0 review · deepseek-v4-flash
2026-08-01 04:01 UTC pith:HT32RZXJ
load-bearing objection Solid framework paper whose central optimality result for Gaffke's bound is checkable only if the authors supply the Vlassis–Thomas inequality they currently cite as a black box. the 3 major comments →
On the Order-Conditional Optimality of Gaffke's Bound
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The central claim is Theorem 7. For every α∈(0,1) and every sample x, the nested class bound N^α_{μ_max}(x; Q_ind, D*_α) equals G_α(x), where Q_ind is the class of all laws with nonnegative independent marginals and D*_α is the total preorder induced by G_α itself. It follows that G_α is conditionally optimal for D*_α: every Q_ind-valid α-level lower confidence bound consistent with the same sample ordering satisfies ψ(x) ≤ G_α(x) for all x, and no such bound can strictly improve on G_α anywhere. The argument leans on the identity G_α(s) = s α^{1/n} for constant samples and on a construction of i.i.d. laws with marginals supported on {0,s} that approach the infimum from above.
What carries the argument
The load-bearing construction is the nested Q-class bound N^α_π(x;Q,D) = inf{π[Q] : Q∈Q, Q[Ω(x,D)] > α}, the infimum of the parameter over laws that give the upper set of x more than α mass. To connect this to Gaffke's bound, the paper uses the conservative completion c(x,u) = Σ_i x_(i)(u_{i+1}-u_i) over the simplex of uniform order statistics u, which defines Gaffke's sublevel regions; G_α(x) is the infimum of c(x,u) over u whose sublevel region has Lebesgue measure exceeding α. The theorem identifies these two infima and uses two-point i.i.d. laws to make the Gaffke infimum approachable from above.
Load-bearing premise
The load-bearing premise is an external theorem, cited in the proof as Inequality (10), that for any independent nonnegative variables with means at most 1, the probability that a uniform-Dirichlet-weighted sum of them is at most 1 is bounded by α; the paper neither states nor proves this theorem, and Theorems 6 and 7 stand or fall with it.
What would settle it
Verify Inequality (10) numerically for small n and α: take Y_i independent with two-point distributions, say Y_i = 0 with probability 1/2 and Y_i = 2 with probability 1/2 so E[Y_i]=1, and check whether Q[∑ D_i Y_i ≤ 1] ≤ α for α=0.05. A single violating product law would invalidate G_α as an α-LCB and destroy the optimality theorem. Alternatively, a direct refutation of conditional optimality would be a Q_ind-valid α-LCB that orders samples exactly as G_α does yet exceeds G_α(x) for some x; no such bound should exist.
If this is right
- No α-level lower bound that respects the Gaffke ordering can strictly dominate Gaffke's bound at any sample; any improvement must come from a different ordering of samples.
- The same conditional optimality transfers to the i.i.d. and independent common-mean models, where μ_max is the common mean.
- For every sample, the Gaffke bound is the limit of parameters of simple i.i.d. two-point laws, so the bound's value is controlled by elementary models.
- A bound that dominates G_α but is valid for product laws must itself be approximable by such two-point families, per Corollary 1.
Where Pith is reading between the lines
- This result leaves open whether another valid bound with a different ordering could be better in a global sense; conditional optimality is with respect to the order Gaffke's bound itself creates.
- The two-point approximation hints at a practical computation: for a given x and α, the value G_α(x) could be approximated by solving for the smallest p such that p^n > α and p·s equals the candidate value, potentially enabling numerical routines for G_α.
- The paper's framework defines 'active' distributions via rejection regions, so the same machinery might be reused for vector-valued parameters or for upper confidence bounds, though the paper only treats scalar lower bounds.
- Two support points deserve checking: the external theorem used as Inequality (10) is neither stated nor proved, and measurability of G_α and of the order graph is deferred to 'standard arguments'; if the former fails, Theorems 6 and 7 collapse.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper develops a general, purely probabilistic framework for order-conditional lower confidence bounds (LCBs), then specializes it to product laws and the maximum marginal mean parameter. The main theorem (Theorem 7) states that Gaffke's bound G_α equals the nested Buehler bound N^α_{μ_max}(·; Q_ind, D^*_α) for the total preorder D^*_α induced by G_α itself, and is therefore conditionally optimal among all Q_ind-valid α-LCBs consistent with that order. The proof combines a general dominance theorem (Theorems 4 and 5) with two special properties of G_α: its homogenous-sample value G_α(s,...,s)=sα^{1/n} (Proposition 11) and its finite-sample validity as an α-LCB for μ_max (Theorem 6). The latter is imported from an external Vlassis–Thomas theorem cited as [11], and this is the main point on which the result depends.
Significance. If the main result holds, it is a clean and useful contribution: it gives a rigorous, order-theoretic account of Buehler optimality for Gaffke's bound, extends the class of models beyond i.i.d. to arbitrary product laws, and provides a simple two-point approximation of the optimal bound at each sample. The internal framework is developed carefully: Proposition 5's size argument is sound, Theorem 5's two-point construction works, Proposition 11 is proved directly, and Corollary 2 correctly extends the conclusion to common-mean and i.i.d. subclasses. The paper is also honest about the scope of the optimality: it is order-conditional, relative only to LCBs that order samples exactly as G_α does. The decisive caveat is that the validity of G_α as a Q_ind-valid α-LCB (Theorem 6) rests entirely on an unstated external theorem, so the central contribution is currently conditional on a citation that the reader cannot verify from the manuscript itself.
major comments (3)
- [§6, Inequality (10) in Theorem 6] This inequality is the only step establishing that G_α is a Q_ind-valid α-LCB on μ_max. It is attributed to the Vlassis–Thomas preprint [11], but the theorem is neither stated nor proved. The normalization Y_i = X_i/μ_max only gives E[Y_i] ≤ 1; the reader cannot check whether the theorem in [11] requires a stronger condition, such as a bound on the sum of the means, or a different Dirichlet parameter. If Inequality (10) is false or misquoted, Theorem 6 and hence Theorem 7 collapse. Please include the exact statement of the Vlassis–Thomas theorem with all hypotheses (including the value of α and the Dirichlet parameter) and either a proof or a precise pointer to the theorem number in [11].
- [Remark 1; Definitions 9 and 14] The measurability of G_α, the graph of D^*_α, and the map x ↦ N^α_{μ_max}(x; Q_ind, D^*_α) is essential: Theorem 3 requires measurability to make the nested bound a valid LCB, and Theorem 7 identifies the nested bound with G_α. The manuscript says the first two can be established by 'standard arguments' but omits them, and the measurability of the nested bound is merely assumed after Definition 9. Since the nested bound is an infimum over a nonparametric class of distributions, this is not purely cosmetic. Please supply the arguments or state explicit regularity conditions under which Theorem 7 holds.
- [Equation (12)] The proof of Theorem 6 passes from Definition 19 to G_α(x) = inf{t ≥ 0 : K_t(x) > α} by noting that u = (r,...,r) makes c(x,u) sweep out [0, x_(n)]. The step is plausible, but it deserves a sentence: the set {t : K_t(x) > α} is an upper interval because K_t is monotone, and the one-parameter family u=(r,...,r) realizes every t in [0,x_(n)], so the two infima coincide. As written, the equivalence is asserted too quickly for a reader to verify that no value of t outside the range of c(x,·) can lower the infimum.
minor comments (3)
- [Theorem 5] The phrase 'for any s ∈ Ω, homogeneous in s ≥ 0' is confusing; suggest writing s = (s,...,s) for s ≥ 0.
- [Definition 16] The family of distributions Q_η is denoted with the same letter Q as the model class Q. Consider using P_η to avoid ambiguity.
- [Definition 19] G_1(x) := ∞ is defined but never used in the proofs; either remove it or mention its role in the extension to α=1.
Circularity Check
No significant circularity: the equality N^α=G_α is derived via an external validity theorem and a two-point approximation, not assumed by definition; the order-induced optimality is standard Buehler optimality.
full rationale
The derivation is not circular. Gaffke's bound is defined independently of the nested-bound construction (Definitions 17-20), and Theorem 7's equality is proved by applying Theorem 5 after establishing (i) Q_ind-validity via the external Vlassis-Thomas inequality (10) and (ii) the homogeneous-sample formula G_α(s)=sα^(1/n) (Proposition 11, proved in the text). The upper bound G_α≤N comes from general validity (Proposition 6); the lower bound N≤G_α comes from an explicit family of two-point i.i.d. laws whose μ_max approaches G_α(x). The order D*_α is defined by G_α (Definition 20), so the optimality statement is self-referential in its wording, but that is the standard definition of Buehler/order-conditional optimality; the equality N=G_α is not an identity. The main load-bearing external input, Inequality (10), is cited to Vlassis-Thomas [11] without being stated or proved in the paper; that is a support gap/correctness risk, not circularity, because the authors of [11] do not overlap with the present paper and the inequality is not derived from the paper's own framework. Minor self-citations ([5], [7], [10]) are present but non-load-bearing: [5]'s result is re-proved in Proposition 11, and [7]/[10] are contextual framework citations. Remark 1's omitted measurability proof is a technical omission, also not circularity.
Axiom & Free-Parameter Ledger
axioms (4)
- domain assumption Vlassis-Thomas theorem: for independent nonnegative Y_i with E[Y_i]≤1, Q[K_1(Y)≤α]≤α, where K_1(Y)=P_D[Σ Y_i D_i≤1] with D~Dir(1,...,1) (n+1 components).
- ad hoc to paper Measurability of G_α and of the graph of D^*_α (Remark 1); measurability of x↦N^α_μ_max(x;Q,D) (assumed after Definition 9 and in Theorems 3 and 5).
- domain assumption The model class Q_ind (product laws) is contained in the class Q, and Q contains every i.i.d. two-point law ((1-p)δ_0+pδ_s)^{⊗n}.
- domain assumption Total preorder D is Borel measurable and upper sets Ω(x,D) are measurable.
read the original abstract
Let $X = (X_1, \ldots, X_n)$ be a random vector from any Borel probability law on $\mathbb{R}_+^n$. We revisit the problem of deriving a lower confidence bound (LCB) on a scalar parameter of that law. We recast classical work, beginning with Buehler, in purely probabilistic terms to form a more accessible and extensible framework. We then specialize the framework to the case where the components of $X$ are independent. In this context, we prove that Gaffke's bound is Buehler optimal for the order that it induces with respect to the maximum marginal mean parameter: $max_{i \in [n]} E_Q[X_i]$, which reduces to the common mean when the $X_i$ are independent and identically distributed. That is to say, no other valid LCB that orders samples in the same way as Gaffke's bound can improve on it with respect to this parameter.
Reference graph
Works this paper leans on
-
[1]
Outline of a Theory of Statistical Estimation Based on the Classical Theory of Prob- ability
Jerzy Neyman. “Outline of a Theory of Statistical Estimation Based on the Classical Theory of Prob- ability”. In:Philosophical Transactions of the Royal Society of London. Series A, Mathematical and Physical Sciences236.767 (1937), pp. 333–380.doi:10.1098/rsta.1937.0005
arXiv 1937
-
[2]
Confidence Intervals for the Product of Two Binomial Parameters
Robert J. Buehler. “Confidence Intervals for the Product of Two Binomial Parameters”. In:Journal of the American Statistical Association52.280 (1957), pp. 482–493.doi:10.1080/01621459.1957. 10501404
arXiv 1957
-
[3]
Buehler Confidence Bounds
Margarita F. Guerrero and Herbert T. David. “Buehler Confidence Bounds”. In:The Philippine Statis- tician(1985), pp. 86–106
1985
-
[4]
Three test statistics for a nonparametric one-sided hypothesis on the mean of a nonnegative variable
Norbert Gaffke. “Three test statistics for a nonparametric one-sided hypothesis on the mean of a nonnegative variable”. In:Mathematical Methods of Statistics14.4 (2005), pp. 451–467
2005
-
[5]
A New Confidence Interval for the Mean of a Bounded Random Variable
Erik Learned-Miller and Philip S. Thomas. “A New Confidence Interval for the Mean of a Bounded Random Variable”. In:arXiv preprint arXiv:1905.06208(2019). 12
Pith/arXiv arXiv 1905
-
[6]
The Lexicographic Method in Preference Theory
Michael Mandler. “The Lexicographic Method in Preference Theory”. In:Economic Theory71.2 (Mar. 2021), pp. 553–577.doi:10.1007/s00199-020-01256-2
-
[7]
Erik Learned-Miller.On the admissibility of bounds on the mean of discrete, scalar probability distribu- tions from an iid sample. 2025. arXiv:2502.17223 [math.ST].url:https://arxiv.org/abs/2502. 17223
Pith/arXiv arXiv 2025
-
[8]
George Bissias.Algorithms for Approximating Conditionally Optimal Bounds. 2026. arXiv:2507.15529 [stat.CO].url:https://arxiv.org/abs/2507.15529
arXiv 2026
-
[9]
Jiahao Ming et al.Gaffke’s confidence interval for the mean of bounded data is inadmissible but asymp- totically efficient. 2026. arXiv:2607.18661 [math.ST].url:https://arxiv.org/abs/2607.18661
Pith/arXiv arXiv 2026
-
[10]
My Phan and Erik Learned-Miller.Towards Automated Confidence Bound Provers and Searchers
-
[11]
An Exact Distribution-Free Test for Means of Nonnegative Random Variables
Nikos Vlassis and Philip S. Thomas. “An Exact Distribution-Free Test for Means of Nonnegative Random Variables”. In:arXiv preprint arXiv:2607.08415(2026). 13
Pith/arXiv arXiv 2026
-
[2026]
arXiv:2607.10379 [stat.CO].url:https://arxiv.org/abs/2607.10379
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.