OctoInformBeta

OpenAI preprint, September 2026 · §5. Choice of parameters and failure of duality · §5

§5 (setting): the parameters: For a≥1a \ge 1 and θ=1/(1000a)\theta = 1/(1000a), the radius hh meets log⁡3/log⁡(h+1)≤θ\log 3/\log(h+1) \le \theta, a rank and then a prime exist, and 38θ<1/a38\theta < 1/a

The counterexample’s parameters can be chosen, the prime last.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

A counterexample to the duality conjecture for metric entropy with constants a≥1a \ge 1 and bb is a pair of convex bodies whose covering numbers, and those of their polars, break the comparison the conjecture asserts with these constants. Its construction works with forms in rr variables over a prime field 𝔽p\mathbb{F}_p and depends on numbers fixed in advance: an accuracy θ\theta, a number ss of pivot slots, a compression radius hh, a rank rr, a number uu of directions, an encoding cost CC and the prime pp. Each number is chosen after those it depends on, and pp comes last, larger than quantities built from all the others.

The library defines θ\theta, ss, hh, uu and CC by formulas and states that a suitable rank rr and then a suitable prime pp exist.

The Parameters

The accuracy parameter θ\theta of a real number aa is the real number θ=11000a.\theta = \frac{1}{1000\,a}.(1)✦

For a real number θ\theta, the number of pivot slots of θ\theta is the least natural number ss that is at least 1/θ1/\theta: s=⌈1/θ⌉.s = \lceil 1/\theta \rceil.(2)✦

The compression radius hh of a real number aa with accuracy parameter θ\theta is the natural number h=⌈exp⁡(log⁡3/θ)⌉+1,h = \bigl\lceil \exp(\log 3/\theta) \bigr\rceil + 1,(3)✦ with log\log the natural logarithm.

Given natural numbers rr and jj, the form dimension DjD_j is the binomial coefficient Dj=(r+j−1j),D_j = \binom{r+j-1}{j},(4)✦ the number of multisets of size jj drawn from rr indices, which is the dimension of the space of forms of degree jj on E=𝔽prE = \mathbb{F}_p^r, denoted Sym⁡j(E*)\operatorname{Sym}^j(E^*).

The direction count uu of natural numbers rr and hh and a real number θ\theta is the natural number u=⌈2hDhθ⌉.u = \Bigl\lceil \frac{2hD_h}{\theta} \Bigr\rceil.(5)✦

The encoding cost CC of a natural number hh, a real number θ\theta and a natural number uu is the real number C=log⁡u+hs2ulog⁡(1+s2uθ),C = \log u + h\,s\,2^u \log\Bigl(1 + \frac{s\,2^u}{\theta}\Bigr),(6)✦ where ss is the number of pivot slots (2) of θ\theta.

The Claim

Theorem 1. Let bb be a real number. For every positive integer hh and every real number θ\theta, with ss the number of pivot slots of θ\theta, there is a positive integer rr with 8bh2s<r+h−1.8b\,h^2 s < r + h - 1.(7)✦ Let now a≥1a \ge 1 be a real number, θ\theta its accuracy parameter, ss the number of pivot slots of θ\theta and hh the compression radius of aa. Then log⁡3log⁡(h+1)≤θ.\frac{\log 3}{\log(h+1)} \le \theta.(8)✦ There is a positive integer rr satisfying (7) for these hh and ss for which, with DhD_h, uu and CC formed from rr, hh and θ\theta, h>0,Dh>0,u>0,2hDh≤θu,C>0.h > 0, \qquad D_h > 0, \qquad u > 0, \qquad 2hD_h \le \theta u, \qquad C > 0.(9)✦ For every positive integer rr, every real number B>0B > 0 and every real number MM, with DhD_h, uu and CC formed from rr, hh and θ\theta, there is a prime pp with p>max⁡{M,h,h2u2h,exp⁡(BC/Dh)}.p > \max\bigl\{M,\ h,\ h^2 u^{2h},\ \exp(BC/D_h)\bigr\}.(10)✦ Finally, 38θ<1a.38\,\theta < \frac{1}{a}.(11)✦

In the paper, the result reads: with θ=1/(1000a)\theta = 1/(1000a) and s=⌈1/θ⌉s = \lceil 1/\theta \rceil, take a radius hh with log⁡3/log⁡(h+1)≤θ\log 3/\log(h+1) \le \theta, a rank rr with r+h−1>8bh2sr+h-1 > 8bh^2 s as in its (5.1), uu and CC as in its (5.2), and last a prime p>max⁡{h,h2u2h,exp⁡(8bC/Dh)}p > \max\{h, h^2u^{2h}, \exp(8bC/D_h)\} as in its (5.3); then 38θ<1/a38\theta < 1/a.

For b>0b > 0, the rank of (9) and the prime of (10) with B=8bB = 8b are such choices, with the radius fixed by (3).

What Is Checked

The claim is stated over the accuracy parameter (1), the number of pivot slots (2), the compression radius (3), the form dimension (4), the direction count (5) and the encoding cost (6), for a real a≥1a \ge 1, a real bb, a positive rank rr, a real B>0B > 0 and a real bound MM, and in (7) for any positive integer hh and real θ\theta.