REVIEW 3 minor 1 references
Equi-dependence implying independence
T0 review · 0 major / 3 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read If an event has the same conditional probability given each trial in an infinite Bernoulli sequence, then it is independent of every trial.
desk verdict A small, clean, correct probability note whose only real defect is an overbroad sentence in Example 1; the central theorems and proofs hold 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 load-bearing object is the empirical average $M_n$ of the indicators of the conditioning events, $M_n=\frac{1}{N_{n,s,t}}\sum_{(S,T)\in\mathcal{P}_{n,s,t}}\mathbf{1}_{B_{S,T}}$. Its expectation is exactly $r=p^s(1-p)^t$, and its variance tends to zero because the covariance of two indicators $\mathbf{1}_{B_{S,T}}$ and $\mathbf{1}_{B_{Q,R}}$ vanishes unless the index sets $(S\cup T)$ and $(Q\cup R)$ overlap; the probability of such an overlap is at most $uv/n$ by a union bound, for the relevant sizes $u,v\in\{s,t\}$. Thus $M_n\to r$ in $L^1$, so $E[\mathbf{1}_A M_n]\to P(A)r$. But the assumed conditional-average condition gives $E[\mathbf{1}_A M_n]\to qr$, forcing $P(A)=q$.
What would settle it
The theorem asserts that no event $A$ and independent Bernoulli sequence can satisfy (4) with $P(A)\neq q$. A direct numerical search over small $n$ and simple events $A$ (for example, Boolean combinations of the first few $B_i$) would find no violation; the parity example is the closest competitor and it satisfies the condition only for $s+t<n$, which is exactly the boundary the proof's variance estimate controls.
Extended reading notes
Core claim
On the paper's own terms, the central discovery is Theorem 3: for fixed nonnegative integers $s$ and $t$, if the $B_i$ are independent with common probability $p$, and the average of $P(A\mid B_{S,T})$ over all disjoint pairs $(S,T)$ with $|S|=s, |T|=t$ within $[n]$ converges to $q$ as $n\to\infty$, then $P(A)=q$. Here $B_{S,T}=\bigcap_{k\in S}B_k\cap\bigcap_{l\in T}B_l^c$. Theorems 1 and 2 are special cases, and Corollary 4 draws the independence conclusion when the conditional probabilities are exactly constant. The paper also gives an example showing that the conclusion cannot be strengthened to independence of $A$ from the entire sequence of trials: if $p=q=1/2$ and $A$ is the event that an even number of the first $n$ trials occur, then $P(A\mid B_{S,T})=1/2$ for every fixed $(s,t)$ with $s+t<n$ and all disjoint $(S,T)$, yet $A$ is not independent of $B_1,\ldots,B_n$.
Load-bearing premise
The theorem rests on the trial events $B_i$ being independent of one another; if they are correlated, the averaging argument loses its grip and the conclusion may be false.
Editorial extensions
If this is right
- Theorem 1 (the $s=1,t=0$ case) says that if $P(A\mid B_i)=q$ for every $i$, then $P(A)=q$; in particular, $A$ is independent of each individual trial event $B_i$.
- Theorem 2 relaxes the pointwise equality to $n^{-1}\sum_{i=1}^n P(A\mid B_i)\to q$ and still concludes $P(A)=q$.
- Theorem 3 extends this to averages over all disjoint pairs of blocks of fixed sizes $s$ and $t$; the limiting conditional probability $q$ is again the unconditional probability.
- Corollary 4 makes the independence conclusion explicit: if $P(A\mid B_{S,T})=q$ for every disjoint pair with $|S|=s,|T|=t$, then $A$ is independent of every such $B_{S,T}$.
- The paper's Example 1 shows the limit of the method: $A$ need not be independent of the whole sequence $(B_i)$, so the conclusion is about the specified conditioning events, not full independence.
Reading between the lines
- A natural next step, not taken in the paper, is to replace the uniform average over disjoint pairs with a weighted average; the same variance argument should work whenever the weights are spread evenly enough that the variance of the weighted empirical average still vanishes.
- The proof is quantitative: the variance bound $\mathrm{Var}(M_n)\le C/n$ yields explicit finite-$n$ error bounds for Theorems 2 and 3 via Chebyshev's inequality, though the paper does not state them.
- The family of block events $\{B_{S,T}\}$ has an almost-orthogonal structure in $L^2$ as $n$ grows, so the theorem can be read as a Hilbert-space principle: asymptotic knowledge of inner products against an almost-orthogonal family determines the mean of the indicator.
- The parity example suggests that to force independence from the whole sequence one would need the conditional-probability condition to hold for infinitely many sizes $(s,t)$ simultaneously, since a fixed pair is satisfied by a finite-dimensional parity event that is not independent of the whole sequence.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper studies an event A in a probability space with an infinite sequence of independent Bernoulli events B_i, each of probability p. Theorem 1 states that if P(A|B_i)=q for every i, then P(A)=q, so A is independent of each B_i. Theorem 2 relaxes the exact equality to Cesàro convergence of the conditional probabilities. Theorem 3 generalizes this to averages, over all disjoint pairs (S,T) of subsets of [n] with |S|=s and |T|=t, of P(A|B_{S,T}); convergence of this average to q again forces P(A)=q. Corollary 4 is the exact version for infinite index pairs. Example 1 constructs, for p=1/2, an event A=D_n that satisfies the hypotheses of Corollary 4 for certain (s,t) yet is not independent of the entire sequence of B_i, thereby negating a natural stronger conjecture.
Significance. The main theorems are correct and the proofs are genuinely self-contained and elementary. The variance bound in the proof of Theorem 3 is clean: the covariance of the two block indicators vanishes unless the index sets overlap, and the union bound correctly gives an O(1/n) bound for each overlap probability, yielding L1 convergence of the empirical average. The paper also gives a useful stability statement (Theorem 3) and a neat counterexample to a natural conjecture. The contribution is modest but well suited to a general mathematical journal. The only substantive flaw is an overstatement in Example 1 that does not affect the central theorems.
minor comments (3)
- [Example 1] The final sentence of Example 1 claims that condition (5) holds for all nonnegative integers s and t and all (S,T) in P∞,s,t, but the derivation immediately preceding it is explicitly restricted to s+t<n. The claim is false without this restriction: for instance, with n=1, s=1, t=0, A=D_1=B_1, and P(A|B_1)=1, not 1/2. The counterexample to the conjecture still works by fixing any pair (s,t) with s+t<n (e.g., s=1, t=0 and n≥2), so the overstatement should be corrected by adding the restriction s+t<n to the stated validity of condition (5) in this example.
- [Theorem 3] In the sentence defining Pn,s,t, the notation is introduced for n≥s+t, but Theorem 3 and equation (4) use the limit as n→∞ without repeating this restriction. It would be clearer to state that convergence in (4) is taken over n≥s+t, although this is presumably intended.
- [Remark 1] The proof of Theorem 1 is written for q in (0,1), but the argument also works for q=0 or q=1; if the author prefers to keep the stated range, a brief note of this extension would prevent a reader from wondering whether the endpoint cases are excluded for a reason.
Circularity Check
No circularity: The derivation is self-contained; the central theorems follow from the stated assumptions by standard convergence arguments.
full rationale
No circularity is present in the paper's central argument. The proofs of Theorems 1 and 3 derive P(A)=q directly from the stated hypotheses: the event probabilities P(B_i)=p, the independence of the B_i's, and the assumed conditional-probability limits. In Theorem 1, the empirical average M_n of indicators 1_{B_i} converges to p in L1, and the assumption P(A|B_i)=q gives E1_A M_n = pq, so P(A)p = pq and hence P(A)=q. Theorem 3 uses the same structure with M_n averaging indicators of the events B_{S,T}; equations (7), (9), and (10) bound the variance by O(1/n), forcing E1_A M_n to converge to both P(A)r and rq, yielding P(A)=q. No fitted parameters, no hidden prior result by the author, and no input is equivalent by construction to the target conclusion. The Hewitt-Savage reference is contextual only and is not load-bearing. I do flag a non-circular correctness issue in Example 1: the text asserts that condition (5) holds for all nonnegative integers s and t, but the extension argument is explicitly restricted to s+t<n. For example, with n=1, s=1, t=0, A=D_1 is the complement of B_1, so P(A|B_1)=0, not 1/2. This overclaim does not affect the self-contained proofs of Theorems 1-3.
Assumptions & free parameters
assumptions (3)
- standard math Kolmogorov probability axioms and the definition of conditional probability P(A|B)=P(A∩B)/P(B) for events with P(B)>0.
- domain assumption The events B_i are independent with common probability p∈(0,1).
- standard math Standard convergence facts: if E|M_n-r|^2 tends to 0 then E|M_n-r| tends to 0, and the Cesaro average in (4) exists.
Cite this review
Pith. "Pith review of Equi-dependence implying independence." pith.science (2026). https://pith.science/paper/LBXPMYM2
@misc{pith2026260813559,
author = {Pith},
title = {Pith review of: Equi-dependence implying independence},
year = {2026},
howpublished = {\url{https://pith.science/paper/LBXPMYM2}},
note = {Machine review of arXiv:2608.13559}
}
abstract
It is shown that, if an event $A$ has the same conditional probability in each trial in an infinite sequence of Bernoulli trials, then $A$ is independent of each trial. More general results are actually established.
Reference graph
Works this paper leans on
-
[1]
1. Hewitt E, Savage LJ. Symmetric measures on Cartesian products. Trans Amer Math Soc. 1955;80:470- 501. January 2014] 5
work page 1955
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.