REVIEW 3 major objections 3 minor 1 cited by
On Samuels' Conjecture
T0 review · 3 major / 3 minor · reviewed 2026-08-27 · deepseek-v4-flash
Pith's one-line read Samuels' conjecture proved: shortfall chance has an exact minimum
desk verdict A credible, largely self-contained proof of Samuels' conjecture and the sharp Feige bound, with two addressable gaps: an unproved calculus lemma and an underproved corollary. 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 argument runs through a reduction to weighted Bernoulli sums. First, every mean-one nonnegative law is decomposed as a mixture of two-point distributions (Proposition 2.3), so it suffices to prove a uniform lower bound for products of Bernoulli-style two-point laws. The load-bearing object is Theorem 2.1, a lower-tail bound for $T = \sum w_i B_i$ with $B_i \sim \mathrm{Bernoulli}(p_i)$, $0\le w_i\le 1$: if $w_i p_i \le a_i$, $a_1\le\cdots\le a_n$, and $d = 1-\mathbb{E}T>0$, then $\mathbb{P}(T<1) \ge \min_i \prod_{j=i}^n (1-a_j/(d+C_i))$ with $C_i = \sum_{j=i}^n a_j$. Theorem 2.1 is proved by a Bellman-type induction on the dimension, in which a conditioning argument produces an exact recursion for the deletion cost $L$ and the proof reduces to a one-variable minimization of the function $F_i(L)$ on intervals $[C_{i+1},C_i]$; Claim 3.4 asserts this function has no interior minimum.
What would settle it
Take $d=0.1$ and $a=(1,2,3)$; if the function $F_2(L)$ on $[3,5]$ has any interior point with $F_2(L) < \min\{F_2(3), F_2(5)\}$, then Claim 3.4 is false and the induction behind Theorem 2.1 collapses.
Extended reading notes
Core claim
The central claim is Theorem 1.1: with $D_i := \lambda - \sum_{k=1}^{i-1}\mu_k$, the infimum of $\mathbb{P}(\sum_{i=1}^n X_i < \lambda)$ over independent laws with $\mathbb{E}X_i = \mu_i$ equals $\min_{1\le i\le n} q_i(\mu,\lambda)$, where $q_i(\mu,\lambda) := \prod_{j=i}^n (1-\mu_j/D_i)$. The theorem states that this bound is sharp and is attained, meaning there are explicit extremal laws (all coordinates below $i$ deterministic, coordinates from $i$ onward scaled Bernoulli variables) that achieve the minimum. The paper also gives the equivalent upper-tail formulation and derives Corollary 1.3: for equal means $\mathbb{E}X_i \le 1$, $\inf \mathbb{P}(S < \mathbb{E}S + \delta) = \min\{\delta/(1+\delta), (1-1/(n+\delta))^n\}$, and the dimension-free infimum over $n$ is $\min\{\delta/(1+\delta), e^{-1}\}$.
Load-bearing premise
The lower-bound direction depends on Claim 3.4, which says a certain one-variable function $F_i(L)$ attains its minimum on each interval $[C_{i+1},C_i]$ at an endpoint, justified by a single sentence asserting 'elementary calculus shows' the second derivative is negative at interior stationary points, with no calculation shown.
Editorial extensions
If this is right
- Feige's conjecture is true in its sharp dimension-free form: for every $\delta>0$, the least possible value of $\mathbb{P}(S<\mathbb{E}S+\delta)$ across $n$ is $\min\{\delta/(1+\delta), e^{-1}\}$.
- For each finite $n$ and $\delta$, the exact constant is $\min\{\delta/(1+\delta), (1-1/(n+\delta))^n\}$, and it is attained.
- The upper-tail analogue follows immediately, so the same result controls $\mathbb{P}(\sum X_i \ge \lambda)$.
- The proof constructs extremal laws from the means alone, so knowing only the marginal means is enough to certify the sharp lower bound.
- The extremal construction shows the worst case is achieved by two-point (scaled Bernoulli) distributions, with an initial deterministic block.
Reading between the lines
- One practical reading: for a sum of nonnegative risks with known marginal means, the formula gives a guaranteed lower bound on the chance of a shortfall without any distributional fitting.
- The induction may also adapt to thresholds not normalized to 1 and to Bernoulli weights not capped at 1, because the cap is used only to preserve the event in the reduction.
- One could test whether the same formula governs $\mathbb{P}(\sum X_i \le \lambda)$; the proof treats the strict inequality, and boundary cases may require separate treatment.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper claims to settle Samuels' conjecture on the infimum of P(sum X_i < lambda) over independent nonnegative variables with prescribed ordered means. The main result, Theorem 1.1, states that this infimum equals min_i prod_{j=i}^n (1 - mu_j/D_i), where D_i = lambda - sum_{k<i} mu_k. The proof is structured as follows: a reduction to two-point distributions via a decomposition lemma; a lower-bound theorem (Theorem 2.1) for weighted Bernoulli sums, proved by a Bellman-type induction; a sharpness construction using truncated Bernoulli variables; and a boundary regularization argument. The paper also derives Feige's conjecture as Corollary 1.3. The lower-bound direction is the heart of the argument and depends on a sequence of claims about the Bellman recursion, the last of which, Claim 3.4, is asserted with a one-sentence 'elementary calculus' justification. The paper is self-contained and does not assume either conjecture.
Significance. If the proof is correct, this resolves a long-standing open problem in probability inequalities: it gives the exact worst-case small-deviation probability using only nonnegativity, independence, and first moments. It would also settle Feige's conjecture with sharp dimension-dependent constants. Notable strengths are the explicit two-point decomposition, the self-contained nature of the argument, the boundary approximation step, and the explicit sharpness construction. However, the load-bearing Claim 3.4 is not actually proved in the manuscript, and the derivation of Corollary 1.3 from the main theorem is only sketched. The paper is therefore significant but not yet in a form that supports its claims.
major comments (3)
- [Section 3.1, Claim 3.4 and Eq. (57)] Claim 3.4 is the terminal step of the Bellman induction that proves Theorem 2.1 and hence the lower bound in Theorem 1.1. Its entire proof is the sentence 'Elementary calculus shows that g_i''(L)<0 for every interior stationary point, so g_i, and hence F_i, has no interior minimum.' No computation is shown. This is not a routine convexity assertion: g_i is not concave on the whole interval, and the sign of g_i'' at stationary points is a genuine algebraic fact that requires an explicit calculation or a rigorous bound. If this claim failed for some ordered a_i, d, and L in [C_{i+1}, C_i], the lower-bound direction would collapse. Please supply the complete derivative computation and sign analysis.
- [Section 1, Corollary 1.3 and Eq. (5)] The formula for c_{n,delta} is stated as 'immediate from the equal-means case', but this reduction is not actually shown. The equal-means case of Theorem 1.1 with mu_i=1 gives only the particular constructions leading to (1-1/(n+delta))^n, and the n=1 construction gives delta/(1+delta). To identify the infimum over all mean vectors with 0<=mu_i<=1, one must prove that min_{0<=mu_i<=1} min_i q_i(mu, sum mu + delta) equals the stated constant. No such argument appears. Since the abstract and introduction claim that Feige's conjecture is resolved, this missing step is load-bearing for that claim.
- [Section 2.3, Sharpness] The sharpness construction is explicit and demonstrates that each q_i(mu,lambda) is attained, which gives the upper bound c(mu,lambda) <= min_i q_i(mu,lambda). However, the lower bound c(mu,lambda) >= min_i q_i(mu,lambda) relies entirely on Theorem 2.1. The current manuscript does not establish Theorem 2.1 because of the gap in Claim 3.4. I am not claiming the theorem is false; I am saying that the proof as written is incomplete at this point.
minor comments (3)
- [Throughout] The phrase 'elementary calculus' should be replaced by a full derivation; if the calculation is short, it can be included in the text, and if it is long, it should be placed in an appendix.
- [Section 2.2, Eq. (20)-(25)] The notation w_i is used both for the unscaled weight mu_i(y_i-x_i)/Lambda and for its capped version min{w_i,1}. The text explains the capping, but the double use of the same symbol makes the display in Eq. (22) momentarily ambiguous. A separate symbol for the capped weights would improve readability.
- [References] Several references are to arXiv preprints with 2026 dates. This is acceptable for a quickly evolving area, but the paper should clearly indicate which results are published and which are preprints, especially those used for context.
Circularity Check
No circularity: the proof is self-contained; the only flagged issue is an unproved calculus assertion in Claim 3.4, which is a correctness gap rather than a circular reduction.
full rationale
The derivation of Theorem 1.1 does not assume Samuels' or Feige's conjecture. The lower bound c(mu,lambda) >= min_i q_i is obtained by a two-point decomposition (Proposition 2.3), a common-lower-bound transfer (Lemma 2.4), and the internally proved Bernoulli-sum Theorem 2.1. Theorem 2.1 is established by a Bellman induction (Claim 3.2), an interval estimate (Claim 3.3), and an endpoint-minimum assertion (Claim 3.4). No fitted parameter is renamed as a prediction, and no load-bearing result is imported from the author's prior work; the only self-citation, [GHLL20], is background on a previously known constant and plays no role in the proof. The sharpness direction is an explicit construction attaining each q_i, so equality is not forced by definition. Two non-circular weaknesses should be flagged: (i) Claim 3.4 says 'Elementary calculus shows that g_i''(L)<0 for every interior stationary point' without displaying the computation; this is a substantive unproved step on which the lower-bound induction depends, but it is a proof gap, not a circularity, because the claimed algebraic fact is independent of the theorem's conclusion. (ii) Corollary 1.3 asserts that the equal-means result 'immediately yields' the sharp Feige formula for general EX_i<=1, but the reduction from arbitrary bounded means to equal means is not shown; this is omitted support, not circular reasoning. Neither issue makes any equation in the paper equivalent to its inputs by construction, so the circularity score is 0.
Assumptions & free parameters
assumptions (2)
- standard math Standard measure-theoretic probability background (Borel laws, conditioning, Tonelli's theorem).
- domain assumption The feasible class is exactly independent nonnegative random variables with specified means; the infimum is over Borel laws on [0,∞).
Cite this review
Pith. "Pith review of On Samuels' Conjecture." pith.science (2026). https://pith.science/paper/BOQAGY4E
@misc{pith2026260818392,
author = {Pith},
title = {Pith review of: On Samuels' Conjecture},
year = {2026},
howpublished = {\url{https://pith.science/paper/BOQAGY4E}},
note = {Machine review of arXiv:2608.18392}
}
abstract
Let $0\leq\mu_1\leq\cdots\leq\mu_n$ and let $\lambda>\sum_{i=1}^n\mu_i$. Let $X_1,...,X_n$ be independent nonnegative random variables satisfying $\mathbb{E}X_i=\mu_i$, and write $D_i := \lambda-\sum_{k=1}^{i-1}\mu_k$ for $1\leq i\leq n$. We prove that $$ \inf_{X_1,...,X_n} \mathbb{P}\left( \sum_{i=1}^nX_i<\lambda \right) = \min_{1\leq i\leq n} \prod_{j=i}^n \left( 1-\frac{\mu_j}{D_i} \right). $$ The bound is sharp and is attained. This proves Samuels' conjecture. Feige's conjecture is thereby resolved, since it follows immediately from the equal-means case. The proof is self-contained.
Forward citations
Cited by 1 Pith paper
-
A Cap-Move Reformulation of Ling's Proof of Samuels' Conjecture
A cap-move proof establishes Samuels' conjecture for finitely supported distributions, giving an alternative to Ling's Bernoulli-based proof.
Reference graph
Works this paper leans on
- [1]
- [2]
- [3]
-
[4]
SIAM Journal on Computing , volume =
On sums of independent random variables with unbounded variance and estimating the average degree in a graph , author =. SIAM Journal on Computing , volume =
-
[5]
Bounding probability of small deviation:
He, Simai and Zhang, Jiawei and Zhang, Shuzhong , journal =. Bounding probability of small deviation:
-
[6]
Journal of Combinatorial Theory, Series A , volume =
Small deviations of sums of independent random variables , author =. Journal of Combinatorial Theory, Series A , volume =
-
[7]
Alqasem, Abdulmajeed and Aravinda, Heshan and Marsiglietti, Arnaud and Melbourne, James , journal =. On a conjecture of
-
[8]
Bounding probability of small deviation on sum of independent random variables:
Guo, Jiayi and He, Simai and Ling, Zi and Liu, Yicheng , journal =. Bounding probability of small deviation on sum of independent random variables:
Show all 14 references
-
[9]
A short proof of
Egozcue, Mart. A short proof of. arXiv preprint arXiv:2509.19949 , year =
-
[10]
arXiv preprint arXiv:2608.06317 , year =
Sharp tail bounds beyond twice the mean , author =. arXiv preprint arXiv:2608.06317 , year =
-
[11]
On some conjectures of
Paulin, Roland , journal =. On some conjectures of
-
[12]
arXiv preprint arXiv:2607.08415 , year =
An exact distribution-free test for means of nonnegative random variables , author =. arXiv preprint arXiv:2607.08415 , year =
-
[13]
Nie, Zipei and Wei, Jiaye , journal =. On
-
[14]
arXiv preprint arXiv:2607.23980 , year =
Sharp small-deviation inequalities for sums of independent nonnegative random variables , author =. arXiv preprint arXiv:2607.23980 , year =
Reviewed August 27, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.