OctoInform Beta

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

Lemma 6.5 [Polynomial cancellation]: For bounded FF, projective characters χ\chi and a difference polynomial PP, the Ces`aro means of |∫F(tmx)χP(m)¯dη||\int F(t_m x)\,\overline{\chi_{P(m)}}\,d\eta| tend to 00

Mixing cancels centered correlations twisted along a polynomial.

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 t=(tm)m∈ℕt=(t_m)_{m\in\mathbb N} a measure-preserving action of the additive monoid ℕ\mathbb N on it, so that ta+b=tb∘tat_{a+b}=t_b\circ t_a; for example tm=Tpmt_m=T^{pm} for a measure-preserving transformation TT of XX and an integer p≥1p\ge1. A second probability space (Y,η)(Y,\eta) carries a random point xx of XX with law μ\mu and a family (χl)l∈L(\chi_l)_{l\in L} of unimodular functions indexed by an abelian group LL. A complex function FF on XX and a sequence P:ℕ→LP\colon\mathbb N\to L give at each time mm the correlation ∫YF(tm(x(y)))χP(m)(y)¯dη(y)\int_Y F(t_m(x(y)))\,\overline{\chi_{P(m)}(y)}\,d\eta(y) of the observable FF, read along the orbit of xx, with χP(m)\chi_{P(m)}. The question is whether these correlations are small on average in mm when tt is mixing, FF has mean zero and PP is polynomial.

The library states the answer when PP is a difference polynomial, a sequence whose iterated differences are constant, and χ\chi is a family of projective characters, characters that are multiplicative up to constant phases; beside it, it states van der Corput’s lemma for bounded sequences in a real inner product space.

Difference Polynomials and Projective Characters

For a sequence P:ℕ→LP\colon\mathbb N\to L in an abelian group LL and a step h≥1h\ge1, the difference sequence ΔhP\Delta_hP is given by ΔhP(m)=P(m+h)−P(m)\Delta_hP(m)=P(m+h)-P(m). A difference polynomial of degree dd in LL is a sequence P:ℕ→LP\colon\mathbb N\to L whose dd-fold differences with positive steps are constant: for all steps h1,…,hd≥1h_1,\dots,h_d\ge1 and all m∈ℕm\in\mathbb N, Δh1Δh2⋯ΔhdP(m)=Δh1Δh2⋯ΔhdP(0).\Delta_{h_1}\Delta_{h_2}\cdots\Delta_{h_d}P(m)=\Delta_{h_1}\Delta_{h_2}\cdots\Delta_{h_d}P(0).(1)✦

A family of projective characters of an abelian group LL on a measure space (Y,η)(Y,\eta) is a family χ=(χl)l∈L\chi=(\chi_l)_{l\in L} of η\eta-measurable complex functions on YY with |χl|=1|\chi_l|=1 and χ0=1\chi_0=1 almost everywhere and ∫Yχldη=0\int_Y\chi_l\,d\eta=0 for l≠0l\ne0, which is multiplicative up to constant phases: for all l,l′∈Ll,l'\in L there is a complex number cc with |c|=1|c|=1 and χlχl′=cχl+l′η-almost everywhere.\chi_l\,\chi_{l'}=c\,\chi_{l+l'}\quad\text{$\eta$-almost everywhere.}(2)✦

Cesàro Averages

The Cesàro averages of a sequence v=(vn)n∈ℕv=(v_n)_{n\in\mathbb N} in a real vector space EE are the vectors ANv=1N∑n=0N−1vn,N≥1.A_Nv=\frac1N\sum_{n=0}^{N-1}v_n,\qquad N\ge1.(3)✦

A sequence vv in a real inner product space EE has lag Cesàro bound cc, for a real sequence c=(ch)h∈ℕc=(c_h)_{h\in\mathbb N}, if for every lag h≥1h\ge1 limsupN→∞1N∑n=0N−1|⟨vn,vn+h⟩|≤ch.\limsup_{N\to\infty}\frac1N\sum_{n=0}^{N-1}\bigl|\langle v_n,v_{n+h}\rangle\bigr|\le c_h.(4)✦

The Claim

Theorem 1 (Van der Corput’s lemma). Let vv be a bounded sequence in a real inner product space EE and c=(ch)c=(c_h) a real sequence with ch→0c_h\to0 as h→∞h\to\infty. If vv has lag Cesàro bound cc, then its Cesàro averages tend to zero in EE: limN→∞ANv=0.\lim_{N\to\infty}A_Nv=0.(5)✦

Theorem 2 (Polynomial cancellation). Let tt be a measure-preserving action of ℕ\mathbb N on a probability space (X,μ)(X,\mu), with ta+b=tb∘tat_{a+b}=t_b\circ t_a, which is mixing: μ(A∩tm−1B)→μ(A)μ(B)\mu(A\cap t_m^{-1}B)\to\mu(A)\mu(B) as m→∞m\to\infty for all measurable A,B⊆XA,B\subseteq X. Let xx be a random point of XX with law μ\mu on a probability space (Y,η)(Y,\eta), let χ=(χl)l∈L\chi=(\chi_l)_{l\in L} be a family of projective characters of an abelian group LL on (Y,η)(Y,\eta), and let P:ℕ→LP\colon\mathbb N\to L be a difference polynomial of degree dd. Let FF be a μ\mu-measurable, essentially bounded complex function on XX with ∫XFdμ=0\int_XF\,d\mu=0. Then limN→∞1N∑m=0N−1|∫YF(tm(x(y)))χP(m)(y)¯dη(y)|=0.\lim_{N\to\infty}\frac1N\sum_{m=0}^{N-1} \Bigl|\int_YF\bigl(t_m(x(y))\bigr)\,\overline{\chi_{P(m)}(y)}\,d\eta(y)\Bigr|=0.(6)✦

In the paper, the result reads: if TXT_X is a mixing transformation of (X,μ)(X,\mu), with μ\mu the paper’s μX\mu_X, xx has law μ\mu, the χl\chi_l are projective characters, FF is bounded and centered, p≥1p\ge1 and PP is a group-valued polynomial, then 1N∑m=0N−1|∫YF(TXpmx)χP(m)¯dη|→0\frac1N\sum_{m=0}^{N-1}\bigl|\int_YF(T_X^{pm}x)\,\overline{\chi_{P(m)}}\,d\eta\bigr|\to0 as N→∞N\to\infty.

What Is Checked

Van der Corput’s lemma (5) is stated over the Cesàro averages (3) and the lag Cesàro bound (4) of a bounded sequence in a real inner product space, and polynomial cancellation (6) over a mixing measure-preserving action of ℕ\mathbb N, a random point of law μ\mu, a family of projective characters (2), a difference polynomial (1) of degree dd and an essentially bounded centered function.