OctoInformBeta

OpenAI preprint, September 2026 · §5. Homogeneous increments and signed Poisson sampling

Lemma 5.1: Exponential integrals multiply under convolution and average under symmetrization

Poisson mixing and symmetrization give explicit transforms.

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

Random sums of real coordinates enter the study of finite subsets of LpL_p through probability laws on the line. Convolution describes addition of independent coordinates, while symmetrization introduces a random sign. The library states characteristic-function identities for Poisson mixtures and symmetrization, together with identities for exponential integrals of fixed convolution sums.

For a measure ρ\rho on a real inner product space EE, its characteristic function ρˆ\widehat\rho is the complex-valued function ρˆ(t)=∫Eei⟨t,x⟩dρ(x),t∈E.\widehat\rho(t)=\int_E e^{i\langle t,x\rangle}\,d\rho(x), \qquad t\in E.(1) On the real line, the inner product in (1) is txtx.

For a measure μ\mu on ℝ\mathbb R, write Mμ(t)M_\mu(t) for its exponential integral, when the integrand is integrable: Mμ(t)=∫ℝetxdμ(x),t∈ℝ.M_\mu(t)=\int_{\mathbb R}e^{tx}\,d\mu(x), \qquad t\in\mathbb R.(2)

Let ρ\rho be a measure on EE, with ** denoting convolution under addition and δ0\delta_0 denoting unit mass at zero. The convolution powers ρ*n\rho^{*n} of ρ\rho are the measures specified recursively by ρ*0=δ0,ρ*(n+1)=ρ*n*ρ,n∈ℕ.\rho^{*0}=\delta_0, \qquad \rho^{*(n+1)}=\rho^{*n}*\rho, \qquad n\in\mathbb N.(3)✦

For a measure ρ\rho on EE and a rate r≥0r\geq0, the compound Poisson measure CP⁡r(ρ)\operatorname{CP}_r(\rho) on EE is the Poisson mixture of the convolution powers in (3): CP⁡r(ρ)=∑n=0∞e−rrnn!ρ*n.\operatorname{CP}_r(\rho) =\sum_{n=0}^{\infty}e^{-r}\frac{r^n}{n!}\,\rho^{*n}.(4)✦

The symmetrization μsym\mu^{\mathrm{sym}} of a measure μ\mu on ℝ\mathbb R is the average of μ\mu and its reflection: μsym=12μ+12(−id)#μ.\mu^{\mathrm{sym}} =\tfrac12\mu+\tfrac12(-\mathrm{id})_\#\mu.(5)✦

The Claim

Theorem 1. Let ρ\rho be a probability measure on EE, let r≥0r\geq0, and let t∈Et\in E. For the compound Poisson measure in (4), CP⁡r(ρ)ˆ(t)=exp⁡(r(ρˆ(t)−1)).\widehat{\operatorname{CP}_r(\rho)}(t) =\exp\!\bigl(r(\widehat\rho(t)-1)\bigr).(6)✦ Let μ\mu be a probability measure on ℝ\mathbb R and t∈ℝt\in\mathbb R. Its symmetrization satisfies μsymˆ(t)=∫ℝcos⁡(tx)dμ(x).\widehat{\mu^{\mathrm{sym}}}(t) =\int_{\mathbb R}\cos(tx)\,d\mu(x).(7)✦ Let μ,ν\mu,\nu be probability measures on ℝ\mathbb R, and suppose that x↦etxx\mapsto e^{tx} is integrable for both measures. With the notation of (2), Mμ*ν(t)=Mμ(t)Mν(t).M_{\mu*\nu}(t)=M_\mu(t)M_\nu(t).(8)✦ For a probability measure μ\mu on ℝ\mathbb R, suppose that x↦etxx\mapsto e^{tx} is μ\mu-integrable, and let n∈ℕn\in\mathbb N. Then et(⋅)∈L1(μ*n),Mμ*n(t)=Mμ(t)n.e^{t(\cdot)}\in L_1(\mu^{*n}), \qquad M_{\mu^{*n}}(t)=M_\mu(t)^n.(9)✦ Let σ\sigma be a measure on ℝ\mathbb R and FF a real normed vector space, and suppose that both f:ℝ→Ff:\mathbb R\to F and x↦f(−x)x\mapsto f(-x) are σ\sigma-integrable. Then ∫ℝf(x)dσsym(x)=12(∫ℝf(x)dσ(x)+∫ℝf(−x)dσ(x)).\int_{\mathbb R} f(x)\,d\sigma^{\mathrm{sym}}(x) =\frac12\left( \int_{\mathbb R}f(x)\,d\sigma(x) +\int_{\mathbb R}f(-x)\,d\sigma(x)\right).(10)✦

In the paper, the result reads: for bounded real hh supported on a region AA of finite η\eta-measure, its signed Poisson sum SS satisfies 𝔼eitS=exp⁡(∫(cos⁡(th)−1)dη)\mathbb E e^{itS}=\exp(\int(\cos(th)-1)\,d\eta), 𝔼etS=exp⁡(∫(cosh⁡(th)−1)dη)\mathbb E e^{tS}=\exp(\int(\cosh(th)-1)\,d\eta), 𝔼S=0\mathbb ES=0 and 𝔼S2=∫h2dη\mathbb ES^2=\int h^2\,d\eta, with the characteristic-function identity also valid for unbounded hh on AA.

What Is Checked

The characteristic-function identities (6) and (7) are stated over probability measures, as are the exponential integral identities (8) and (9), with their integrability assumptions; (10) is stated over an arbitrary measure on the real line and an integrable function into a real normed vector space whose reflection is also integrable.