Press release · 16 November 2026 · version 0.1.0-candidate
Resampling heterogeneous experiments: sharp amplification constants and high-signal error asymptotics
Resampling n different sources with replacement can raise a test's advantage over guessing, but in a natural class by at most 14/9 (at three sources); in general the homogenisation constant is at most 9/14, below the earlier bound 8/9.
Summary
In one sentence. Looking at $n$ different sources of evidence through $n$ random draws with replacement can make a test between two explanations slightly better than looking once at each source; this candidate finds exact worst cases, showing that in a natural class the advantage over guessing can grow by at most $14/9$, and that in general the best universal homogenisation constant is at most $9/14$, below the earlier bound $8/9$.
The everyday problem. You want to decide between two explanations using several different sources of evidence, say $n$ different sensors or $n$ different sites. Design A looks at every source once. Design B makes $n$ draws at random with replacement, so some sources are seen twice and some not at all, as a bootstrap or a random-design study does. You would expect B to be worse, since it may miss a source altogether. Surprisingly, it can be slightly better. The question is by how much.
How "better" is measured. The quality of a design is the total-variation distance between what the data look like under the two explanations. It is the advantage over guessing of the best possible test: with equal prior odds the best test is right with probability $(1+\text{distance})/2$. Write $D$ for the distance under design A (the fixed design) and $H$ for the distance under design B (the resampled, or homogenised, design).
What was known. Aryeh Kontorovich proved in a series of preprints that resampling can never help by more than a constant factor: $D\ge cH$ for a universal constant $c>0.1489$, whatever the sources and however many there are. In the other direction, his yes/no (Bernoulli) example showed $c\le8/9$, and $8/9$ was conjectured to be optimal for Bernoulli data.
What is new.
| What | In plain terms | Where |
|---|---|---|
| A better bound for the general constant | $c\le9/14\approx0.643$; two sources already give $2/3$. So $8/9$ is not the answer for general data. The Bernoulli conjecture is not addressed | Theorem 4.1 |
| Exact worst cases for erasure experiments | Each observation either rules one explanation out or lands in a shared zone where both behave alike. For every random sampling rule chosen independently of the data, the worst-case factor $\sup H/D$ is found; it depends only on how often pairs of sources are drawn | Theorem 3.1 |
| Uniform resampling | The factor is $3/2$ for two sources, $14/9$ for three (its maximum), and decreases towards about $1.524$ for many sources | Theorem 4.1, Proposition 4.2 |
| Two sources, all data | For arbitrary experiments with two sources the largest possible $H$ is $\tfrac32D-\tfrac12D^2$; more generally the exact bound is known whenever only two sources differ, and the exact optimum for three erasure sources is found | Theorem 5.1, Corollary 5.2, Theorem 6.1 |
| Strong evidence | Resampling with $n$ draws always leaves at least $n!/n^n$ of the fixed design's error probability (for three sources, $2/9$ of it), the chance that every source is drawn exactly once; this factor is attained when the evidence is already strong | Theorem 7.1, Corollary 7.3 |
A concrete example. Take three sources: two carry weak opposing "rules out" signals and one carries nothing. Looking once at each gives advantage $D$. Three random draws give exactly
$$H=\tfrac{14}{9}D-\tfrac23D^2+\tfrac{2}{27}D^3,$$
about $14/9$ times $D$ when $D$ is small. That is $14/9$ times the advantage over guessing, not $55.6\%$ more accuracy: accuracy is $(1+H)/2$, so only the excess over one half is multiplied.
The idea of the proof. In the erasure class the information carried by any sampling outcome is a maximum of two linear budgets, one for each explanation. The worst case puts each explanation's whole budget on a single source, so only pairs of counts matter, and the factor becomes the expected larger of two counts. For strong evidence, a base-$(k+1)$ "digit" construction makes every unbalanced draw much less informative than the balanced one, so the chance of drawing every source equally often is exactly what survives.
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 data. The high-signal limit can be approached slowly. The Bernoulli conjecture $8/9$ is not addressed. This is an unrefereed candidate; its assurance is internal AI-assisted review and computational checks, not peer review. The homogenisation problem, the bounds $c>0.1489$ and $c\le8/9$ and the labelled-lift argument used here are Kontorovich's.
Summary for specialists
Experiments $(P_i,Q_i)$, $i=1,\dots,n$, on one measurable space; $D=\mathrm{TV}(\bigotimes_iP_i,\bigotimes_iQ_i)$ and $H=\mathrm{TV}(\bar P^{\otimes n},\bar Q^{\otimes n})$ with $\bar P=\frac1n\sum_iP_i$. Let $c^*$ be the best constant in $D\ge cH$ over all $n$ and all spaces. An erasure experiment is a pair whose likelihood ratio takes at most one value in $(0,\infty)$. A schedule is a random count vector $N$ independent of the data; $K,L$ denote the numbers of draws of two fixed experiments under $\mathrm{Mult}(n;1/n,\dots,1/n)$.
- Theorem 3.1 (proof): for erasure experiments and any revealed schedule, the sharp constant in $H_\nu\le\kappa_\nu D$ is $\max\{\mathbb P(N\ne0),\ \max_{i,j}\mathbb E\max(N_i,N_j)\}$. The lift to the labelled experiment ($H\le H_\nu$) is Kontorovich's.
- Theorem 4.1, Proposition 4.2 (proof): for uniform resampling of erasure experiments $H\le C_nD$ with $C_n=\mathbb E\max(K,L)$ sharp as $D\to0$; $C_2=3/2$, $C_3=14/9$, $C_4=99/64$, strictly decreasing for $n\ge3$ to $C_\infty=1+e^{-2}(I_0(2)+I_1(2))\approx1.5238$. A five-outcome family gives $c^*\le9/14$ for arbitrary experiments; every $n\ge3$ gives $c^*<0.6563$. With Kontorovich's lower bound, $0.1489<c^*\le9/14$.
- Remark 4.3: the family has labelled ratio $H_\nu/D\to14/9$ even on a two-point space, so any proof through the labelled lift cannot give $c>9/14$, and a proof of the Bernoulli conjecture would have to use the loss of labels.
- Theorem 5.1, Corollary 5.2 (proof): if $P_i=Q_i$ except in two coordinates, the exact bound is $H\le F_n(D)$, attained at every $D$. For $n=2$ the unrestricted problem is solved: $\sup H=\tfrac32D-\tfrac12D^2$. Kontorovich's local-to-global paper had the linear bound $H\le\tfrac32D$.
- Theorem 6.1, Corollary 6.2 (proof): the exact three-component erasure optimum at every distance, with transitions $D_0\approx0.3249$, $D_1\approx0.6972$ and crossover $D_*=(9-\sqrt{33})/8\approx0.4069$; near $D=1$ the optimal overlap is $\tfrac29r+O(r^{6/5})$ with $r=1-D$.
- Theorem 7.1, Corollary 7.3 (proof): for arbitrary experiments and any data-independent schedule with $kn$ draws, $1-H\ge q\,(1-D)^k$ at every distance, where $q$ is the probability that every experiment is used exactly $k$ times; a radix construction with base $B\ge k+1$ attains the leading term as $D\to1$ with an explicit power-saving remainder. For uniform resampling with $k=1$ the optimal ratio of Bayes errors tends to $n!/n^n$, the classical no-collision probability (Darroch–Jirina–Speed).
- Conjecture 8.1: $c^*=9/14$. Evidence: 40 Nelder–Mead restarts per configuration over arbitrary finite experiments found no labelled ratio above $C_n$ (tolerance $10^{-6}$); without labels on a common two-point space the best ratio for $n=2$ was $1.12495$, close to the Bernoulli value $9/8$.
Section 8.3 shows that the erasure class is not a reduction target: garbling preserves mutual absolute continuity, and revealing common parts can never show amplification. Privacy amplification by subsampling (Balle–Barthe–Gaboardi, Theorem 10, at $\varepsilon=0$) is the one-differing-record counterpart of Theorem 3.1.
Technical account
Erasure overlaps. For an erasure experiment, given the counts $N$, the overlap of the labelled experiment is a product over coordinates. Taking logarithms, the information is $\max(N\cdot a,N\cdot b)$ for two nonnegative budget vectors $a,b$, one per hypothesis, and the fixed-design distance constrains their sums. The expectation of a maximum of two linear forms is convex in $(a,b)$, so the worst case sits at extreme points of a product of simplices: each hypothesis puts its whole budget on one coordinate. That turns the problem into the law of a pair of counts $(N_i,N_j)$, giving $\mathbb E\max(N_i,N_j)$, or a single count, giving $\mathbb P(N\ne0)$.
Uniform resampling. For multinomial counts the pair $(K,L)$ has an explicit law, and $C_n=\mathbb E\max(K,L)$ is computed exactly. As $n\to\infty$, $(K,L)$ tends to two independent Poisson(1) variables, and a telescoping identity for the Bessel functions gives $C_\infty$. The extremal five-outcome family puts weak opposing signals on two coordinates and nothing on the rest; every observation that can occur under both hypotheses reveals its index, so $H$ equals the labelled distance and the ratio $H/D$ tends to $C_n$.
Two differing coordinates. Given counts $k$ and $l$ for the two differing coordinates, the observations group into $\min(k,l)$ copies of the pair and $|k-l|$ single copies, each with distance at most $D$. The product-overlap inequality then bounds the conditional distance by $1-(1-D)^{\max(k,l)}$, and the five-outcome family attains this in every count case; averaging gives $F_n$, and for $n=2$ the quadratic modulus.
High signal. Every count vector other than the balanced one $(k,\dots,k)$ is shown to carry overlap of smaller order. The construction gives coordinate $i$ a weight $B^{i-1}$, so that the total exponent for a count vector is a base-$B$ number; only the balanced vector achieves the minimum, and unbalanced vectors lose at least a fixed power of $1-D$.
Evidence, assurance and limitations
The synthetic audio briefing is explanatory, not additional evidence. Its complete transcript is available for reading.
- Proofs. Every theorem, proposition and corollary has a written proof in the 16-page paper. The paper opens with a plain introduction (the problem, what was known, what is new with the class and regime of each result, relation to other work, and a roadmap) and ends with interpretation, the conjecture with its evidence, why the erasure class is not a reduction, open problems and an AI-use statement. No proof assistant was used.
- Exact checks. Exact rational arithmetic checks the extremal families directly over finite product spaces for $n=2,\dots,5$; the constants $C_n$ for $n\le40$ and the limit; the identities for $F_2$, $F_3$ and $D_*$; the three-component formula against independent constrained optimisation at nine distances and against 4916 rational chamber cases; the high-signal construction for nine choices of $(n,k,B)$ (107 count vectors) and the base-$B$ separation in 72 cases; and the Bernoulli example of Section 8.3. These test the formulas and constructions; the general statements rest on the proofs.
- Numerical search. A fixed-seed search over arbitrary finite experiments supports Conjecture 8.1. It is a finite numerical search, not a proof.
- Tests that can fail. No script uses
assert, so the checks behave identically under Python's optimised mode. Changing $87/400$ to $88/400$ in the Bernoulli check, or doubling the lower bound in the high-signal sandwich check, makes the script exit with failure. - Replay. All nine scripts were rerun from fresh copies before release under Python 3.12 with pinned libraries, the exact checks also under
python -O; every log matched apart from wall-clock timings. GitHub CI reruns the exact checks under-Oagainst the recorded logs, with both negative controls, as a blocking job; the figure and the search run in a separate non-blocking job. - Review. Five internal AI-assisted reviewer roles, an adversarial review by a second AI model and a clean replay found no error in any theorem and led to one revision round. It stated the improvement over $8/9$ prominently, put the class restriction into every headline, credited the labelled lift and the earlier two-source bound, added related work, replaced an unread terminology citation by an intrinsic definition, rewrote the introduction for non-specialists, replaced
assertchecks, and added extensions: the exact two-source problem, the large-$n$ limit, a sharper high-signal construction and the documented conjecture. A re-review, a proof audit and a second-model audit of the new proofs followed. A rating-free summary is in the package. None of this is peer review, independent reproduction or formal verification. - Limitations. The exact general constant is conjectural. The erasure constants are not known to bound general experiments. The high-signal limit keeps $n$ and $k$ fixed and can converge slowly. Finite-distance optima for four or more erasure experiments, budgets that are not multiples of $n$, adaptive schedules and differential-privacy versions are open. Attributions rest on a bounded literature search. This remains an unrefereed candidate.
Who should care, and why
| Audience | Potential use | Required caution |
|---|---|---|
| Statisticians comparing fixed and random designs | Exact worst-case factors by which resampling can raise a test's power in a natural class, and the exact two-source answer | The general constant $9/14$ is an upper bound; its exactness is conjectural |
| Information theorists | Sharp tensorisation-type constants for mixtures of product experiments | Erasure results do not transfer to general experiments |
| Differential-privacy researchers | Two-record analogues of amplification by subsampling at $\varepsilon=0$ | Hockey-stick versions are not proved here |
| Research agents | Machine-readable claims, an agent index and replayable exact checks | The checks test formulas; they are not part of the proofs |
Why the problem matters
Resampling is everywhere: the bootstrap, random-design experiments, randomised data collection and subsampling for privacy all replace "look once at each source" by "draw at random". Knowing how much such a design can gain or lose against the fixed design, in the quantity that controls the best possible error rate, tells a designer how much randomness costs or buys. Kontorovich showed the gain is bounded by a universal constant; this candidate pins down exact constants in a natural class, improves the general bound, and identifies exactly what survives when the evidence is strong.
How to inspect or reproduce the recorded checks
From the GitHub release or the Zenodo archive, on a copy of the package (the scripts write into logs/ and figures/), with uv:
shasum -a 256 -c MANIFEST.sha256
UV="uv run --no-project --python 3.12 --with numpy==2.3.5 \
--with scipy==1.17.0 --with sympy==1.14.0 --with mpmath==1.3.0 \
--with matplotlib==3.10.8"
$UV python -I -B -O code/verify.py # about 2 s
$UV python -I -B -O code/check_extensions.py
$UV python -I -B code/search_unrestricted.py # about 2.5 min
ENVIRONMENT.txt lists the software used. AI_INDEX.md maps each claim to its proof and checks, RESEARCH_GATES.json records the prior-work boundary and the sources read, and REVIEW_SUMMARY.md lists the review objections and how each was handled.
The most valuable next projects
- Prove or refute $c^*=9/14$, even for three sources or in the weak-signal limit $D\to0$ (Conjecture 8.1).
- Determine the Bernoulli constant, conjecturally $8/9$.
- Find the exact finite-distance optimum for four or more erasure experiments, and for three or more arbitrary experiments.
- Extend the schedule constant to the hockey-stick divergences of differential privacy, which would give exact amplification factors for subsampling with several differing records, and to adaptive schedules.
What is in the evidence package
The paper (LaTeX source and PDF); nine scripts in code/ with their recorded logs in logs/; Figure 1 in figures/; requirements.txt; refs.bib from Crossref and arXiv records; AI_INDEX.md; REVIEW_SUMMARY.md; RESEARCH_GATES.json; REPLAY_RECEIPT.md and ENVIRONMENT.txt; citation metadata, licences (CC0-1.0 for text, figure and data, MIT for code), the CI workflows and SHA-256 sums.
Verification status
Unrefereed AI-assisted mathematics candidate with written proofs and exact computational checks. Internal AI-assisted review in five reviewer roles and an adversarial review by a second AI model, a clean replay, one revision round with extensions (the exact two-source problem, the large-n limit, a sharper high-signal construction and a documented conjecture), then a re-review, a proof audit and a second-model audit of the new proofs; no error was found in any theorem. The computations check formulas and constructions and are not part of the proofs. External specialist review, independent reproduction and formal verification have not been established.
Cite
BibTeX
@misc{resamplinghomogenisationconstants2026,
title = {Resampling heterogeneous experiments: sharp amplification constants and high-signal error asymptotics},
author = {Anonymous},
year = {2026},
doi = {10.5281/zenodo.24267748},
url = {https://doi.org/10.5281/zenodo.24267748},
version = {0.1.0-candidate},
howpublished = {Zenodo},
note = {Unrefereed; internally replayed evidence package. Press page: https://evidencepress.org/releases/resampling-homogenisation-constants/}
}Also: cite.bib · paper.json · this page as Markdown