Published November 16, 2026 | Version 0.1.0-candidate

Resampling heterogeneous experiments: sharp amplification constants and high-signal error asymptotics

Authors/Creators

Description

Unrefereed research candidate. In one sentence: resampling a set of different experiments (n draws with replacement instead of one look at each) can make a test between two hypotheses more informative than looking once at each, and this paper finds exact worst-case factors: in a natural class of experiments the advantage over guessing can grow by at most 14/9 (at three experiments), and in general the best universal homogenisation constant is at most 9/14, below the previously best bound 8/9.

Context. Suppose a decision between two explanations uses n different sources of evidence. Design A looks at every source once. Design B makes n random draws with replacement, as a bootstrap or a random-design study does, so some sources are seen twice and some not at all. One would expect B to be worse, but it can be slightly better. Kontorovich (arXiv preprints, 2026) proved that B is never more than a constant factor better in total variation, the quantity that sets the best possible error rate of a test: D ≥ cH with c > 0.1489, where D and H are the total-variation distances under the two designs. The best upper bound was c ≤ 8/9, conjectured optimal for Bernoulli (yes/no) data.

Results. (1) For experiments on general outcome spaces c ≤ 9/14 (about 0.643); two sources already give 2/3. So 8/9 is not the general optimum; the Bernoulli conjecture is not addressed. (2) For erasure experiments (each observation either rules one hypothesis out or lands in a shared region whose internal distribution is the same under both hypotheses) the exact worst-case amplification sup H/D is found for every resampling schedule chosen independently of the data; it depends only on how often pairs of sources are drawn. For uniform resampling it is largest at three sources, with value 14/9, and decreases to 1 + e−2(I0(2) + I1(2)) ≈ 1.524 for many sources. (3) With two sources the problem is solved exactly for all experiments, and the exact bound is given at every distance when only two experiments differ; the exact optimum for three erasure experiments is also found. (4) For arbitrary experiments, resampling with n draws always leaves an overlap 1 − H (twice the Bayes error) of at least n!/nn times the fixed-design overlap (2/9 for three sources), and when the evidence is already strong this factor is attained to leading order; a version with kn draws is also proved.

Limits. Whether 9/14 is the exact general constant is open: it is conjectured, and a numerical search found nothing larger, but that is not a proof. The erasure-class results are not known to bound general experiments. The high-signal limit can be approached slowly. The Bernoulli conjecture 8/9 is not addressed. The factor 14/9 multiplies the advantage over guessing (the total-variation distance), not accuracy.

Contents. The paper (LaTeX source and PDF); nine Python scripts (exact rational and symbolic checks, the occupancy constants, the two-coordinate modulus, the three-component reduction, the high-signal construction and its radix family, a bounded-search script for the conjecture, and the figure script) with their recorded outputs; the figure; the bibliography from registry records; a machine-readable index; a research-gate record; a review summary; a replay receipt and SHA-256 sums.

Assurance. Written proofs; exact computations check the stated constants, identities and finite constructions, and a fixed-seed numerical search supports the conjecture without proving it. Internal AI-assisted review only (five reviewer roles, an adversarial review and an audit of new proofs by a second AI model, clean replays, one revision round, a re-review and a proof audit): not peer review, independent reproduction or formal verification. The homogenisation problem, the bounds c > 0.1489 and c ≤ 8/9 and the labelled-lift argument are due to Kontorovich. Creator: Anonymous. Text and data CC0-1.0; code MIT.

Notes

Subtitle: In a natural class, resampling raises a test's advantage by at most 14/9; in general the homogenisation constant is at most 9/14. Code (code/*.py and the CI workflows) is MIT; prose, manuscript, figure and data are CC0-1.0.

Files

resampling-homogenisation-constants-0.1.0-candidate.pdf

Files (1.2 MB)

Name Size
md5:b07579d49d8e11103447a50328bb30fa
244 Bytes Download
md5:f278726bc13d949a0a033db9fd4986cf
440.1 kB Preview Download
md5:d2b112a26b00ee985a8bea5ca281d66a
739.3 kB Preview Download

Additional details