The cost of exact calibration for growing Gaussian null families
Unrefereed research candidate. In one sentence: when a statistical test must work against a growing grid of Gaussian null hypotheses, insisting that its e-values be exactly calibrated (average exactly one under every null) can destroy all of their power once the grid is fine enough, while ordinary e-values (average at most one) keep theirs; this paper gives the exact phase diagram. The question. An e-value is a number that measures evidence against a hypothesis, like the payoff of a fair bet against it: if the hypothesis is true, its average is at most one. Its quality is the growth rate of evidence when the hypothesis is false, the expected logarithm (e-power). Sometimes one asks for exact e-values, whose average is exactly one under every null in a family, for example to waste no evidence or to match exactly uniform p-values. Zhang, Ramdas and Wang (2024) showed that for the whole continuum of Gaussian means no exact e-value has positive e-power; He and Cheng (2026) showed that for a fixed finite null family exactness costs nothing in the long run. This paper measures what happens in between, when the number of null means grows with the amount of information. Setting. Nulls N(jε,1), |j| ≤ m (2m+1 equally spaced means); alternative N(−t,1), where t plays the role of b√n for n observations. G is the optimal e-power of exact e-variables; the optimum for ordinary (valid) e-variables is (t−mε)2/2. Results. (1) Confluent nulls (εt → 0, m/t2 → c): G/t2 converges to an explicit function F(c), the Hermite exterior action, which falls from 1/2 at c = 0 to 0 at c = 1/4 and vanishes beyond, while the valid rate per t2 stays 1/2. The transition comes when the number of nulls is about t2/2. In the critical window m = t2/4 + at + o(t) (with εt5/2 → 0), G grows like √t with an explicit limit Ψ(a); Ψ(0) ≈ 0.405. (2) Spacing comparable to 1/t (εt → κ > 0): G/t2 converges to an explicit Stieltjes–Wigert exterior action F(c,κ), which vanishes exactly when cosh(κ/2) ≤ ecκ2/2. (3) In sample terms: if a nuisance interval [b−w, b+w], 0 ≤ w < b, is discretised at spacing s/n, exact e-variables have zero growth rate if and only if w ≥ (2/s) log cosh(bs/2), whereas valid e-variables keep the rate (b−w)2/2. For b = 1 and spacing 1/n the threshold half-width is about 0.240, where the valid rate is still about 0.289 per observation. (4) A finite kernel certificate determines G to within log(2m+1) for every spacing; for n = 6400 observations and 1281 null means from 0.9 to 1.1 it places the exact optimum between 1121 and 1128 nats, against 2592 nats for valid e-variables. (5) Exact pivotal e-variables are bounded by the same exact rate along these growing families, although for each fixed finite subfamily they attain the full rate. (6) The loss is a property of exact equality: allowing each null expectation to fall short of one by 1−e−2mε(t−mε) restores the valid rate. How. By finite entropy duality, G is the Kullback–Leibler distance from the alternative to the probability densities in the linear span of the null densities, whereas the valid optimum is the distance to their convex hull. Signed combinations of nearby Gaussian densities build Hermite polynomials (Stieltjes–Wigert polynomials at spacing of order 1/t), so the span comes very close to the alternative; Christoffel functions and classical Plancherel–Rotach asymptotics evaluate the distance. Limits. Unit-variance Gaussian location families only. The spacing conditions for confluent nulls are sufficient, not shown necessary; the critical window is treated only for confluent nulls; above the threshold, exponential smallness is proved only for the auxiliary moment problem (for grids, loss of the leading rate). The exact-and-pivotal rate below the threshold is bounded, not determined. Displayed numbers are ordinary floating-point and mpmath quadrature of proved inequalities, not interval-arithmetic certificates. Contents. The paper (LaTeX source and PDF, 31 pages) with its two figures; three Python verification scripts (numerical checks of the phase functions, kernel brackets and critical constant; a reproduction of the finite-grid example; Stieltjes–Wigert recurrence, kernel-formula, phase-diagram and certificate checks, about 6 minutes) and the phase-diagram figure script, with recorded outputs; the bibliography from registry records; a dated prior-art record; a machine-readable index; a research-gate record; a rating-free review summary; a replay receipt and SHA-256 sums. Assurance. Written proofs. Internal AI-assisted review only (five reviewer roles, an adversarial review and audit by a second AI model, clean replays, one revision round with an extension, a re-review and a proof audit of the new proofs): not peer review, independent reproduction or formal verification. Creator: Anonymous. Text and data CC0-1.0; code MIT.
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-15
- DOI
- https://doi.org/10.5281/zenodo.24252836
- Primary Topic
- Statistical Methods and Inference
- Type
- preprint