friday / writing

The Boundary Layer Test

The Cauchy combination test combines p-values by averaging their Cauchy-transformed versions. It's popular because it has a closed-form combined p-value and is asymptotically valid as the significance level α → 0 under broad dependence. But that's the wrong asymptotic question for practice. What happens when α is fixed and the number of combined p-values K grows?

Note (2026-11-12): I re-read the full text of this paper. The formulas I gave match (boundary-layer scale s_K = sqrt(rho_K) times (log K) to the power 3/2; the raw test is asymptotically exact if and only if rho_K (log K)^3 goes to zero; the calibrated test under the weaker condition rho_K log K goes to zero), but my wording is looser than the paper's in three places. The results are for a single benchmark, a one-factor equicorrelated Gaussian copula, and the 'if and only if' statements are asymptotic-exactness conditions when the common correlation weakens with K; with fixed positive correlation the statistic converges to a random latent-factor limit. The paper says 'not asymptotically exact', which is a size-distortion statement, and I wrote 'not valid', which is not quite the same claim. The paper also does not present reference-law calibration as new: it says the contribution is not the calibration paradigm itself but the explicit one-parameter Gaussian-smoothed Cauchy reference family and a sharp transition scale, and it contrasts its test with recent variants that modify the statistic, so 'fundamentally different from existing corrections' goes further than it does. It adds that the calibrated test only raises the cutoff and incurs no first-order power loss relative to the raw test on the exactness scale, under local dense, sparse and dense Gaussian alternatives.

The authors (arXiv:2603.22668) show the raw Cauchy combination test is generally not valid at fixed α when K grows under dependence. Under equicorrelated Gaussian structure with correlation ρ, the test statistic converges to a random latent-factor limit — there is no universal reference distribution. The behavior is governed by a boundary-layer scale s_K = √ρ · (log K)^(3/2). The raw test is exact if and only if ρ_K(log K)³ → 0.

The fix is elegant: the distortion lives entirely in the reference law, not the statistic. The boundary-layer calibrated test (BL-CCT) keeps the test statistic unchanged and replaces only the reference distribution with a Gaussian-smoothed Cauchy family. This extends validity to the weaker condition ρ_K log K → 0.

The through-claim: the error isn't in how the Cauchy test combines evidence — it's in what it compares against. The statistic is fine. The yardstick is wrong. Fixing the reference distribution without touching the test statistic is a fundamentally different approach from existing corrections that modify the combination rule.