# Review: posterior escape to unbounded hidden-factor counts under null measurements ## 1. Precise restatement **Family.** For n = 1, 2, …, theory T_n says Y = 3^{n/2} ∏_{i=1}^n U_i, with the U_i i.i.d. Uniform[−1,1]. - Each theory has E[Y] = 0. - Each theory has Var(Y) = 3^n · (1/3)^n = 1. - No theory has free parameters. **Data.** A single realisation of Y is measured with nested intervals. The k-th report is "|Y| ≤ ε_k", with ε_k ↓ 0. The hidden modelling assumption is that the likelihood of a report is exactly P(|Y| ≤ ε | T_n). This is noiseless interval censoring. Because the reports are nested, conditioning on all of them equals conditioning on the finest one. **Prior.** π_n > 0 for every n. Strictly, it is enough that π_n > 0 for infinitely many n. The prior puts no mass on anything outside the family, such as a theory with an atom at 0. **Claim.** For every finite K, P(n ≤ K | |Y| ≤ ε) → 0 as ε → 0. In other words, the posterior is not tight and its mass escapes to infinity. Two points are not hidden assumptions but are worth stating: - The equal means and variances play no role in the mathematics; they are rhetorical. - The theories are not observationally equivalent. For example, E[Y⁴] = (9/5)^n, so higher moments differ. ## 2. Checking the steps, and an independent derivation **The distribution formula.** Write |U_i| ~ Uniform[0,1]. Then −log|U_i| ~ Exp(1), so S = −log ∏|U_i| ~ Gamma(n, 1). For 0 < t ≤ 1: P(∏|U_i| ≤ t) = P(S ≥ log(1/t)) = t Σ_{k=0}^{n−1} (log(1/t))^k / k!. With t_n = ε/3^{n/2} this gives L_n(ε) = P(|Y| ≤ ε | T_n). The condition t_n ≤ 1 holds for every n once ε ≤ 1. The text's formula is correct. **Leading order.** L_n ≈ t_n (log 1/t_n)^{n−1}/(n−1)! ~ 3^{−n/2} ε (log 1/ε)^{n−1}/(n−1)!. This is correct. **Likelihood ratio.** Use t_{n+1} = t_n/√3 and log(1/t_{n+1}) > log(1/t_n). Then L_{n+1}/L_n ≥ (1/√3) · [1 + (log 1/t_n)^n/n! ÷ Σ_{k 0 is needed. This works for any prior tail, even an extremely fast-decaying one, because K is fixed before taking the limit. **The argument is rigorous.** ## 3. Attempts to break it **Numerical check.** | ε | L₁ | L₂ | L₂/L₁ | |---|---|---|---| | 10⁻⁶ | 5.77×10⁻⁷ | 3.33×10⁻⁷ × (1 + ln 3×10⁶) ≈ 5.30×10⁻⁶ | ≈ 9.2 | | 10⁻¹² | 5.77×10⁻¹³ | ≈ 9.91×10⁻¹² | ≈ 17.2 | The ratio diverges, but only logarithmically in the precision. In practice the escape to large n is very slow. With a fast-decaying prior, astronomically fine precision is needed before the posterior moves appreciably. **Gaussian measurement noise instead of interval censoring.** Near 0 the density of Y under T_n behaves like (log 1/|y|)^{n−1}. Smoothing with noise of scale σ therefore gives likelihood ratios of order log(1/σ)/(√3 n), which still diverge. The result is robust to this change. **The observed signal is not exactly null.** If the signal is some y ≠ 0, then L_n ≈ 2ε f_n(y). Likelihood ratios converge to finite limits and the posterior converges to a proper, tight distribution. The phenomenon genuinely requires the null value, which is the point where every T_n with n ≥ 2 has an infinite density. **Misspecification.** If the prior includes any theory with an atom or a stronger singularity at 0, that theory dominates instead. The text correctly frames the result as a misspecification warning. **Prior condition.** "Every finite count positive" is sufficient but not necessary; infinitely many positive π_n suffice. If π_n = 0 for all n beyond some N, the claim fails, which is consistent with the stated hypothesis. I found no counterexample within the stated hypotheses. ## 4. Novelty The ingredients are classical: - The product-of-uniforms law, via the Gamma/log-gamma representation. - The fact that, under shrinking-window data, Bayes factors are governed by the densities at the observed point, with density singularities dominating. - Known phenomena of posterior non-tightness or escape of mass under misspecification. The specific construction, aimed at an argument about ontological commitment in the realism debate, appears to be a neat illustrative variant rather than a new mathematical result. I cannot name a specific prior source for this exact example. ## 5. Fairness of the one-sentence summary The summary is essentially accurate: - It states the parameter-free family, the identical mean and variance, the null signal, and the prior condition. - "Beyond every finite bound" correctly means P(n ≤ K) → 0 for every K. Minor omissions: - The noiseless interval-likelihood model is left implicit, though the result survives Gaussian noise. - The divergence is only logarithmic in precision. - The equal mean and variance do not mean the theories are observationally indistinguishable. None of these is a material overstatement. ```json {"verdict": "correct", "novelty": "special case or variant of known results", "fair_summary": "yes", "confidence": 4, "one_line": "The formula, the divergence of likelihood ratios, and the posterior escape past every finite K are all rigorously correct under nested interval data (and survive Gaussian noise), though the effect is only logarithmic in precision and hinges on the density singularity at an exactly null signal."} ```