OctoInform Beta

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

Lemma 6.3 [Entropy cancellation]: If FmF_m takes values in a finite set and Hn(F)/n→0H_n(F)/n\to0, the averages over m<nm<n of |∫v(Fm)χlm¯dη||\int v(F_m)\,\overline{\chi_{l_m}}\,d\eta| tend to 00

Zero-entropy processes decorrelate on average from dissociated projective characters.

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

Take a measure space (Y,η)(Y,\eta), an abelian group LL written additively, a family of projective characters (χl)l∈L(\chi_l)_{l\in L} of LL on (Y,η)(Y,\eta) and a dissociated sequence l0,l1,…l_0, l_1, \dots in LL. Against the characters χlm\chi_{l_m} put the observables v(Fm)v(F_m), where F0,F1,…F_0, F_1, \dots are measurable maps from YY to a fixed finite set AA and vv is a complex function on AA with |v|≤1|v| \le 1. The question is how large the correlations ∫Yv(Fm)χlm¯dη\int_Y v(F_m)\,\overline{\chi_{l_m}}\,d\eta can be on average over mm when the process (Fm)(F_m) has entropy rate zero, the information carried by its first nn letters growing sublinearly in nn.

Projective Characters and Dissociated Sequences

A family of projective characters of an abelian group LL on a measure space (Y,η)(Y,\eta) is a family (χl)l∈L(\chi_l)_{l\in L} of η\eta-measurable complex functions on YY with |χl|=1|\chi_l| = 1 almost everywhere for every ll, with χ0=1\chi_0 = 1 almost everywhere and ∫Yχldη=0\int_Y \chi_l\,d\eta = 0 for every l≠0l \ne 0, which is multiplicative up to constant phases: for all l,l′∈Ll, l' \in L there is a complex number cc of modulus one with χlχl′=cχl+l′η-almost everywhere.\chi_l\,\chi_{l'} = c\,\chi_{l+l'} \quad \eta\text{-almost everywhere}.(1)✦

A sequence (lm)m≥0(l_m)_{m\ge0} in an abelian group LL is dissociated if no signed sum of finitely many of its terms with distinct indices vanishes: for every nonempty finite set SS of indices and every choice of signs εm∈{1,−1}\varepsilon_m \in \{1,-1\}, ∑m∈Sεmlm≠0.\sum_{m\in S} \varepsilon_m\, l_m \ne 0.(2)✦

Word Entropy

The Shannon entropy of a function pp from a finite set BB to the real numbers is H(p)=−∑b∈Bp(b)log⁡p(b),H(p) = -\sum_{b\in B} p(b)\log p(b),(3)✦ with the natural logarithm and the convention 0log⁡0=00\log 0 = 0.

The word map of length nn of a sequence F=(Fm)m≥0F = (F_m)_{m\ge0} of maps from YY to a set AA is the map F(n):Y→AnF^{(n)}\colon Y \to A^n that sends yy to the word (F0(y),…,Fn−1(y))(F_0(y), \dots, F_{n-1}(y)).✦

The word entropy Hn(F)H_n(F) at length nn of a sequence F=(Fm)m≥0F = (F_m)_{m\ge0} of maps from YY to a finite set AA is the Shannon entropy of the law of its word map of length nn under η\eta, which gives each word the measure of the set of points sent to it: Hn(F)=H(pn),pn(w)=η({y∈Y:F(n)(y)=w})(w∈An).H_n(F) = H(p_n), \qquad p_n(w) = \eta\bigl(\{\, y \in Y : F^{(n)}(y) = w \,\}\bigr) \quad (w \in A^n).(4)✦

A sequence F=(Fm)m≥0F = (F_m)_{m\ge0} of maps from YY to a finite set AA has entropy rate zero if Hn(F)/n→0H_n(F)/n \to 0 as n→∞n \to \infty.

The Claim

Theorem 1. Let (Y,η)(Y,\eta) be a probability space, LL an abelian group and (χl)l∈L(\chi_l)_{l\in L} a family of projective characters of LL on (Y,η)(Y,\eta). Let (lm)m≥0(l_m)_{m\ge0} be a dissociated sequence in LL, let F0,F1,…F_0, F_1, \dots be measurable maps from YY to a finite set AA of entropy rate zero, and let vv be a complex function on AA with |v|≤1|v| \le 1. Then limn→∞1n∑m=0n−1|∫Yv(Fm(y))χlm(y)¯dη(y)|=0.\lim_{n\to\infty} \frac1n \sum_{m=0}^{n-1} \Bigl| \int_Y v\bigl(F_m(y)\bigr)\, \overline{\chi_{l_m}(y)}\, d\eta(y) \Bigr| = 0.(5)✦

In the paper, the result reads: if the χl\chi_l are projective characters, (lm)(l_m) is dissociated and functions fmf_m, written FmF_m in the paper, take values in a fixed finite subset of the closed unit disk with H(f0,…,fn−1)=o(n)H(f_0, \dots, f_{n-1}) = o(n), then 1n∑m<n|∫fmχlm¯dη|→0\frac1n \sum_{m<n} \bigl| \int f_m\, \overline{\chi_{l_m}}\, d\eta \bigr| \to 0.

What Is Checked

The claim (5) is stated over a family of projective characters (1), a dissociated sequence (2) and the word entropy (4), the Shannon entropy (3) of the law of the word map of length nn.