OctoInformBeta

OpenAI preprint, September 2026 · §7. The range 2 < p < ∞

Lemma 7.1: Nearly additive bounds for q>2q>2, from scalar averages to fixed convolution powers

Bounds for moments of symmetric sums.

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

For q>2q>2, Lemma 7.1 concerns the qqth moment of a fixed finite sum of bounded independent symmetric real samples. The library states moment bounds through scalar estimates and convolution of probability measures.

A bounded probability measure μ\mu on ℝ\mathbb R is a probability measure for which there is B∈ℝB\in\mathbb R such that |x|≤B|x|\le B for μ\mu-almost every xx. A symmetric probability measure μ\mu on ℝ\mathbb R is a probability measure invariant under reflection x↦−xx\mapsto -x. For a probability measure μ\mu on ℝ\mathbb R and r≥0r\ge0, the rrth absolute moment Mr(μ)M_r(\mu) of μ\mu is the integral in (1): Mr(μ)=∫ℝ|x|rdμ(x).M_r(\mu)=\int_{\mathbb R}|x|^r\,d\mu(x).(1)

The convolution μ*ν\mu*\nu of two measures on a real space EE is the image of their product measure under addition. Writing δ0\delta_0 for the Dirac measure at zero, the convolution powers μ*n\mu^{*n} of a measure μ\mu on EE, for nonnegative integers nn, are the measures defined by (2): μ*0=δ0,μ*(n+1)=μ*n*μ(n≥0).\mu^{*0}=\delta_0, \qquad \mu^{*(n+1)}=\mu^{*n}*\mu \quad(n\ge0).(2)✦

The Claim

Theorem 1. Let q≥2q\ge2. For all a,b∈ℝa,b\in\mathbb R, the lower estimate (3) holds: |a|q+|b|q≤|a+b|q+|a−b|q2.|a|^q+|b|^q \le \frac{|a+b|^q+|a-b|^q}{2}.(3)✦

Let q>2q>2 and ε>0\varepsilon>0. There is a constant C≥0C\ge0 such that (4) holds for all a,b∈ℝa,b\in\mathbb R: |a+b|q+|a−b|q2≤|a|q+(1+ε)|b|q+C|a|q−2b2.\frac{|a+b|^q+|a-b|^q}{2} \le |a|^q+(1+\varepsilon)|b|^q+C|a|^{q-2}b^2.(4)✦

Let q>2q>2, ε∈ℝ\varepsilon\in\mathbb R, and C≥0C\ge0 satisfy (4) for every a,b∈ℝa,b\in\mathbb R. If μ\mu and ν\nu are bounded probability measures on ℝ\mathbb R and ν\nu is symmetric, then (5) holds: Mq(μ)+Mq(ν)≤Mq(μ*ν)≤Mq(μ)+(1+ε)Mq(ν)+CMq(μ)(q−2)/qM2(ν).\begin{aligned} M_q(\mu)+M_q(\nu) &\le M_q(\mu*\nu)\\ &\le M_q(\mu)+(1+\varepsilon)M_q(\nu) +C M_q(\mu)^{(q-2)/q}M_2(\nu). \end{aligned}(5)✦

Let q>2q>2 and ε>0\varepsilon>0. There is a constant K≥0K\ge0 such that every bounded symmetric probability measure μ\mu on ℝ\mathbb R and every nonnegative integer nn satisfy (6): nMq(μ)≤Mq(μ*n)≤(1+ε)2nMq(μ)+K(nM2(μ))q/2.nM_q(\mu) \le M_q(\mu^{*n}) \le (1+\varepsilon)^2 nM_q(\mu) +K\bigl(nM_2(\mu)\bigr)^{q/2}.(6)✦

In the paper, the result reads: if ν\nu is a measure and hh is a bounded real function supported on a set of finite ν\nu-measure, with signed Poisson sum SS, then for q>2q>2 and ε>0\varepsilon>0 there is a constant Cq,εC_{q,\varepsilon} such that (7) holds: ∫|h|qdν≤𝔼|S|q≤(1+ε)∫|h|qdν+Cq,ε(∫h2dν)q/2.\int |h|^q\,d\nu \le \mathbb E|S|^q \le (1+\varepsilon)\int |h|^q\,d\nu +C_{q,\varepsilon}\left(\int h^2\,d\nu\right)^{q/2}.(7)

What Is Checked

The scalar estimates are stated over real a,ba,b, and the measure estimates over bounded probability measures on ℝ\mathbb R, with reflection symmetry of ν\nu in (5) and of μ\mu in (6).