OctoInform Beta

OpenAI preprint, September 2026 · §2. A finite-alphabet witness of zero entropy · §2

Proposition 2.1 [Zero-entropy witness]: For a mixing TT and layouts whose proper sublayouts mix, the correlations of a finite-alphabet witness tend to 00

No finite-alphabet witness has a function of mean zero and zero entropy.

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

Rokhlin’s problem asks whether every mixing transformation is mixing of all orders. Let TT be an invertible measure-preserving transformation of a probability space (Ω,ℱ,μ)(\Omega,\mathcal F,\mu), with integer powers TqT^q, q∈ℤq\in\mathbb Z. A measure-preserving transformation TT of (Ω,ℱ,μ)(\Omega,\mathcal F,\mu) is mixing if limn→∞μ(S∩T−nS′)=μ(S)μ(S′)\lim_{n\to\infty}\mu\bigl(S\cap T^{-n}S'\bigr)=\mu(S)\,\mu(S')(1) for all measurable sets SS and S′S'.

Higher orders are tested along layouts. Let II be a finite set with |I|≥3|I|\ge 3. A layout sequence dd on II is a sequence of maps dr:I→ℕd_r:I\to\mathbb N, r∈ℕr\in\mathbb N, and TT mixes along the proper sublayouts of dd if limr→∞μ(⋂j∈JT−dr(j)Sj)=∏j∈Jμ(Sj)\lim_{r\to\infty}\mu\Bigl(\bigcap_{j\in J}T^{-d_r(j)}S_j\Bigr)=\prod_{j\in J}\mu(S_j)(2) for every proper subset JJ of II and all measurable sets SjS_j, j∈Jj\in J. An index i∈Ii\in I is remote in dd if dr(i)>dr(j)+rd_r(i)>d_r(j)+r for every rr and every j≠ij\neq i.

A witness of the failure of mixing of order |I||I| along a layout sequence dd on II is a family (fj)j∈I(f_j)_{j\in I} of functions bounded by 11, one of them of mean zero, whose correlations along dd, the integrals of the products ∏j∈Ifj∘Tdr(j)\prod_{j\in I}f_j\circ T^{d_r(j)} over Ω\Omega, stay above a constant δ>0\delta>0. On a finite alphabet the functions are fj∘af_j\circ a, j∈Ij\in I, for one measurable observable a:Ω→Aa:\Omega\to A with values in a finite set AA and functions fj:A→ℝf_j:A\to\mathbb R, and the correlations are Cr(f)=∫Ω∏j∈Ifj∘a∘Tdr(j)dμ.C_r(f)=\int_\Omega\prod_{j\in I}f_j\circ a\circ T^{d_r(j)}\,d\mu.(3)

Coding and Word Entropy

The symbolic coding of Ω\Omega by aa is the map Φa:Ω→Aℤ\Phi_a:\Omega\to A^{\mathbb Z} given by Φa(x)=(a(Tqx))q∈ℤ.\Phi_a(x)=\bigl(a(T^qx)\bigr)_{q\in\mathbb Z}.(4)✦ The process of aa is the sequence (a∘Tq)q∈ℤ(a\circ T^q)_{q\in\mathbb Z} of the coordinates of Φa\Phi_a.✦

The entropy Hμ(X)H_\mu(X) of a random variable XX with values in a finite set EE is the Shannon entropy of its law, Hμ(X)=−∑e∈Eμ(X=e)log⁡μ(X=e),H_\mu(X)=-\sum_{e\in E}\mu(X=e)\log\mu(X=e),(5)✦ with 0log⁡0=00\log 0=0.

For a process F=(Fq)F=(F_q) with values in a finite set BB, indexed by ℕ\mathbb N or by ℤ\mathbb Z, the word of FF on [t,t+n)[t,t+n) is the random variable F[t,t+n)=(Ft,…,Ft+n−1)F_{[t,t+n)}=(F_t,\dots,F_{t+n-1}) with values in BnB^n.✦ The word entropy Hn(F)H_n(F) of a process F=(Fq)q≥0F=(F_q)_{q\ge 0} with values in a finite set is the entropy of its word of length nn from time 00, Hn(F)=Hμ(F[0,n)),H_n(F)=H_\mu\bigl(F_{[0,n)}\bigr),(6)✦ and FF has word-entropy rate zero if Hn(F)/n→0H_n(F)/n\to 0.

The Tail Factor

The completion ℳ¯\overline{\mathcal M} of a sub-σ\sigma-algebra ℳ\mathcal M of ℱ\mathcal F is the σ\sigma-algebra of sets in ℱ\mathcal F that agree μ\mu-almost everywhere with a set in ℳ\mathcal M.✦ The past 𝒫t(F)\mathcal P_t(F) of a process F=(Fq)q∈ℤF=(F_q)_{q\in\mathbb Z} with values in a finite set, at a time t∈ℤt\in\mathbb Z, is the σ\sigma-algebra generated by the words F[t−n,t)F_{[t-n,t)}, n∈ℕn\in\mathbb N.✦ The tail of FF is the σ\sigma-algebra 𝒯(F)=⋂n∈ℕ𝒫−n(F)¯.\mathcal T(F)=\bigcap_{n\in\mathbb N}\overline{\mathcal P_{-n}(F)}.(7)✦ The tail factor 𝒯a\mathcal T_a of the observable aa is the tail of the process of aa.✦

The clipped conditional expectation 𝔼[0,1][h∣ℳ]\mathbb E^{[0,1]}[h\mid\mathcal M] of a real function hh given a sub-σ\sigma-algebra ℳ\mathcal M is its conditional expectation cut off at 00 and 11, 𝔼[0,1][h∣ℳ]=max⁡(0,min⁡(1,𝔼[h∣ℳ])).\mathbb E^{[0,1]}[h\mid\mathcal M]=\max\bigl(0,\min(1,\mathbb E[h\mid\mathcal M])\bigr).(8)✦

The Claim

Theorem 1. Let TT be a mixing invertible measure-preserving transformation of (Ω,ℱ,μ)(\Omega,\mathcal F,\mu), let dd be a layout sequence on a finite set II with |I|≥3|I|\ge 3 such that TT mixes along the proper sublayouts of dd, let a:Ω→Aa:\Omega\to A be a measurable observable with values in a finite set AA, and let i∈Ii\in I.

  1. If ii is remote in dd and fj:A→[0,1]f_j:A\to[0,1] for j∈Ij\in I, then limr→∞∫Ω(fi∘a−∫Ωfi∘adμ)∘Tdr(i)∏j≠ifj∘a∘Tdr(j)dμ=0.\lim_{r\to\infty}\int_\Omega\Bigl(f_i\circ a-\int_\Omega f_i\circ a\,d\mu\Bigr)\circ T^{d_r(i)}\,\prod_{j\neq i}f_j\circ a\circ T^{d_r(j)}\,d\mu=0.(9)✦

  2. If g:Ω→[0,1]g:\Omega\to[0,1] is 𝒯a\mathcal T_a-measurable and fj:A→[−1,1]f_j:A\to[-1,1] for j∈Ij\in I, then limr→∞∫Ω(g−∫Ωgdμ)∘Tdr(i)∏j≠ifj∘a∘Tdr(j)dμ=0.\lim_{r\to\infty}\int_\Omega\Bigl(g-\int_\Omega g\,d\mu\Bigr)\circ T^{d_r(i)}\,\prod_{j\neq i}f_j\circ a\circ T^{d_r(j)}\,d\mu=0.(10)✦

  3. If b:Ω→Bb:\Omega\to B is 𝒯a\mathcal T_a-measurable with values in a finite set BB, v:B→[−1,1]v:B\to[-1,1] satisfies ∫Ωv∘bdμ=0\int_\Omega v\circ b\,d\mu=0, and fj:A→[−1,1]f_j:A\to[-1,1] for j∈Ij\in I, then limr→∞∫Ωv∘b∘Tdr(i)∏j≠ifj∘a∘Tdr(j)dμ=0.\lim_{r\to\infty}\int_\Omega v\circ b\circ T^{d_r(i)}\,\prod_{j\neq i}f_j\circ a\circ T^{d_r(j)}\,d\mu=0.(11)✦

  4. If c:A→Bc:A\to B maps into a finite set BB, v:B→ℝv:B\to\mathbb R satisfies ∫Ωv∘c∘adμ=0\int_\Omega v\circ c\circ a\,d\mu=0, the process (c∘a∘Tq)q≥0(c\circ a\circ T^q)_{q\ge 0} has word-entropy rate zero, and fj:A→[−1,1]f_j:A\to[-1,1] for j∈Ij\in I with fi=v∘cf_i=v\circ c, then the correlations (3) satisfy limr→∞Cr(f)=0,\lim_{r\to\infty}C_r(f)=0,(12)✦ and there is no δ>0\delta>0 with Cr(f)≥δfor every r∈ℕ.C_r(f)\ge\delta\quad\text{for every }r\in\mathbb N.(13)✦

In the paper, the result reads: if an invertible mixing transformation is not mixing of every order, then there are k≥3k\ge 3, a finite alphabet 𝒜\mathcal A, a shift-invariant probability measure μX\mu_X on X=𝒜ℤX=\mathcal A^{\mathbb Z} and real functions f1,…,fkf_1,\dots,f_k of the coordinate x0x_0 such that the shift σ\sigma is mixing of every order below kk for μX\mu_X, ∫XfidμX=0\int_X f_i\,d\mu_X=0, ∥fi∥∞≤1\|f_i\|_\infty\le 1, the entropy rate of σ\sigma is 00, and, for some δ>0\delta>0 and layouts 0=d1(r)<…<dk(r)0=d_1(r)<\dots<d_k(r) whose least gap tends to infinity, ∫X∏i=1kfi∘σdi(r)dμX≥δ\int_X\prod_{i=1}^k f_i\circ\sigma^{d_i(r)}\,d\mu_X\ge\delta for every rr, its kk being the page’s |I||I|. The library states the result not as that existence statement but as its contrapositive: part 4 rules out, along layouts whose proper sublayouts mix, a finite-alphabet witness in which the function of mean zero is a function of a process of word-entropy rate zero.

What Is Checked

The theorem is stated over a mixing invertible measure-preserving transformation TT of a probability space, a layout sequence on at least three indices along whose proper sublayouts TT mixes, a remote index, a finite-valued observable aa with its tail factor 𝒯a\mathcal T_a, and the word entropy of a process coded by aa.

About these constructions the library states that the tail factor has zero entropy: for an invertible measure-preserving TT and a finite-valued observable aa, every 𝒯a\mathcal T_a-measurable b:Ω→Bb:\Omega\to B with values in a finite set satisfies limn→∞1nHμ(b,b∘T,…,b∘Tn−1)=0,\lim_{n\to\infty}\frac{1}{n}\,H_\mu\bigl(b,\,b\circ T,\dots,b\circ T^{n-1}\bigr)=0,(14)✦ and, in integer time, the tail process G=(b∘Tq)q∈ℤG=(b\circ T^q)_{q\in\mathbb Z} satisfies limn→∞1nHμ(G[0,n))=0.\lim_{n\to\infty}\frac{1}{n}\,H_\mu\bigl(G_{[0,n)}\bigr)=0.(15)✦

For a decreasing sequence ℳ0⊇ℳ1⊇⋯\mathcal M_0\supseteq\mathcal M_1\supseteq\cdots of sub-σ\sigma-algebras of ℱ\mathcal F with ℳ∞=⋂nℳn¯\mathcal M_\infty=\bigcap_n\overline{\mathcal M_n}, a real h∈L2(μ)h\in L^2(\mu), indices mn→∞m_n\to\infty and ℳmn\mathcal M_{m_n}-measurable functions unu_n with |un|≤1|u_n|\le 1, one has limn→∞∫Ω(h−𝔼[h∣ℳ∞])undμ=0.\lim_{n\to\infty}\int_\Omega\bigl(h-\mathbb E[h\mid\mathcal M_\infty]\bigr)\,u_n\,d\mu=0.(16)✦

For an invertible measure-preserving TT, a layout sequence dd in which ii is remote, a finite-valued observable aa and functions fj:A→[0,1]f_j:A\to[0,1], the factor at the remote index may be replaced in the limit by its clipped conditional expectation given 𝒯a\mathcal T_a: limr→∞∫Ω(fi∘a−𝔼[0,1][fi∘a∣𝒯a])∘Tdr(i)∏j≠ifj∘a∘Tdr(j)dμ=0.\lim_{r\to\infty}\int_\Omega\bigl(f_i\circ a-\mathbb E^{[0,1]}[f_i\circ a\mid\mathcal T_a]\bigr)\circ T^{d_r(i)}\,\prod_{j\neq i}f_j\circ a\circ T^{d_r(j)}\,d\mu=0.(17)✦