OctoInform Beta

OpenAI preprint, September 2026 · §6. Cancellation in one time dimension · §6

Proposition 6.1 [Orbit cancellation]: For a mixing system and any coupling, the correlations with χGml\chi_{G^m l} along the orbit are small in modulus on average over mm

Centered zero-entropy observables cancel against projective characters on average.

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

Let (X,μ)(X,\mu) be a probability space and let t0,t1,…t_0,t_1,\dots be measure-preserving maps of XX with tm+n=tn∘tmt_{m+n}=t_n\circ t_m for all m,n≥0m,n\ge 0. The family is mixing when, for all measurable sets B,C⊆XB,C\subseteq X, μ(B∩tm−1C)→μ(B)μ(C)(m→∞).\mu\bigl(B\cap t_m^{-1}C\bigr)\longrightarrow\mu(B)\,\mu(C)\qquad(m\to\infty).(1)

A measurable map a:X→𝒜a\colon X\to\mathcal{A} to a finite alphabet 𝒜\mathcal{A} reads the system as the finite-valued process a∘t0,a∘t1,…a\circ t_0,a\circ t_1,\dots. For a function v:𝒜→ℂv\colon\mathcal{A}\to\mathbb{C}, the complex observable v∘av\circ a on XX is centered when ∫Xv∘adμ=0\int_X v\circ a\,d\mu=0.

The system is observed from a second probability space (Y,η)(Y,\eta) through a measure-preserving map x:Y→Xx\colon Y\to X, a coupling under which xx is a random point of XX with law μ\mu. On YY lives a family of projective characters χl\chi_l, indexed by an abelian group LL, and an automorphism GG of LL moves the index along the orbit l,Gl,G2l,…l,Gl,G^2l,\dots of a point l∈Ll\in L. At time mm the observable seen through the coupling, v∘a∘tm∘xv\circ a\circ t_m\circ x, is paired with the character χGml\chi_{G^m l} by the inner product of L2(Y,η)L^2(Y,\eta). The question is whether these correlations are small in modulus on average over mm when the process a∘t0,a∘t1,…a\circ t_0,a\circ t_1,\dots has entropy rate zero.

Projective Characters

A family of projective characters χ=(χl)l∈L\chi=(\chi_l)_{l\in L} of an abelian group LL on a measure space (Y,η)(Y,\eta) is a family of η\eta-measurable functions χl:Y→ℂ\chi_l\colon Y\to\mathbb{C} with |χl|=1|\chi_l|=1 almost everywhere for every ll, with χ0=1\chi_0=1 almost everywhere, with mean ∫Yχldη=0\int_Y\chi_l\,d\eta=0 for every l≠0l\neq 0, and multiplicative up to constant phases: for all l,l′∈Ll,l'\in L there is a complex number cl,l′c_{l,l'} of modulus one with χlχl′=cl,l′χl+l′η-almost everywhere.\chi_l\,\chi_{l'}=c_{l,l'}\,\chi_{l+l'}\qquad\eta\text{-almost everywhere}.(2)✦

Word Entropy

The Shannon entropy of a function p:S→ℝp\colon S\to\mathbb{R} on a finite set SS is H(p)=−∑s∈Sp(s)log⁡p(s),H(p)=-\sum_{s\in S}p(s)\log p(s),(3)✦ with the natural logarithm and the convention 0log⁡0=00\log 0=0.

For maps F0,F1,…F_0,F_1,\dots from a set ZZ to an alphabet 𝒜\mathcal{A}, the length-nn word map sends a point zz of ZZ to the word (F0(z),…,Fn−1(z))(F_0(z),\dots,F_{n-1}(z)) in 𝒜n\mathcal{A}^n.✦ When 𝒜\mathcal{A} is finite and ZZ carries a measure ν\nu, the word entropy Hn(F)H_n(F) at length nn is the Shannon entropy of the law of the word map: Hn(F)=H(pn),pn(w)=ν{z∈Z:(F0(z),…,Fn−1(z))=w}(w∈𝒜n).H_n(F)=H(p_n),\qquad p_n(w)=\nu\bigl\{z\in Z:(F_0(z),\dots,F_{n-1}(z))=w\bigr\} \quad(w\in\mathcal{A}^n).(4)✦ The process a∘t0,a∘t1,…a\circ t_0,a\circ t_1,\dots on (X,μ)(X,\mu) has entropy rate zero when its word entropies satisfy Hn/n→0H_n/n\to 0 as n→∞n\to\infty.

The Claim

Theorem 1 (Orbit cancellation). Let (X,μ)(X,\mu) and (Y,η)(Y,\eta) be probability spaces. Let t0,t1,…t_0,t_1,\dots be measure-preserving maps of XX with tm+n=tn∘tmt_{m+n}=t_n\circ t_m, mixing in the sense of (1), and let x:Y→Xx\colon Y\to X be measure-preserving. Let a:X→𝒜a\colon X\to\mathcal{A} be a measurable map to a finite set whose process a∘t0,a∘t1,…a\circ t_0,a\circ t_1,\dots has entropy rate zero for μ\mu, and let v:𝒜→ℂv\colon\mathcal{A}\to\mathbb{C} satisfy ∫Xv∘adμ=0\int_X v\circ a\,d\mu=0. Let χ\chi be a family of projective characters of an abelian group LL on (Y,η)(Y,\eta), let GG be an automorphism of LL, and let l∈Ll\in L. Then 1N∑m=0N−1|∫Yv(a(tm(x(y))))χGml(y)¯dη(y)|→0(N→∞).\frac{1}{N}\sum_{m=0}^{N-1}\Bigl|\int_Y v\bigl(a(t_m(x(y)))\bigr)\, \overline{\chi_{G^m l}(y)}\,d\eta(y)\Bigr|\longrightarrow 0\qquad(N\to\infty).(5)✦ If moreover |v|≤1|v|\le 1 on 𝒜\mathcal{A}, then ∑m=0N−1|∫Yv(a(tm(x(y))))χGml(y)¯dη(y)|N→0(N→∞).\frac{\sum_{m=0}^{N-1}\bigl|\int_Y v\bigl(a(t_m(x(y)))\bigr)\, \overline{\chi_{G^m l}(y)}\,d\eta(y)\bigr|}{N}\longrightarrow 0\qquad(N\to\infty).(6)✦

In the paper, the result reads: for a mixing two-sided stationary process XX on a finite alphabet, of entropy rate zero, with shift TXT_X, a centered finite-valued one-site function ff, any coupling x:Y→Xx\colon Y\to X, any family of projective characters χ\chi, any automorphism GG and any l∈Ll\in L, N−1∑r<N|∫Yf(TXrx)χGrl¯dη|→0N^{-1}\sum_{r<N}\bigl|\int_Y f(T_X^r x)\,\overline{\chi_{G^r l}}\,d\eta\bigr|\to 0, with ff and TXrT_X^r standing for v∘av\circ a and trt_r.

What Is Checked

Both claims are stated over the library’s projective characters (2) and its word entropy (4), built from the length-nn word map and the Shannon entropy (3), with measure preservation, the rule tm+n=tn∘tmt_{m+n}=t_n\circ t_m and mixing (1) as hypotheses.