OctoInformBeta

OpenAI preprint, September 2026 · §6. The range 1 < p < 2

§6 (proof of Theorem 1.1): common stable marginals and their tails: Tail estimates and capped second moments from the characteristic function for 0<p<20<p<2

Stable laws yield common gradient marginals.

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 1<p<21<p<2, the probabilistic construction for LpL_p dimension reduction asks for a random vector in a normalized-gradient subspace with prescribed scalar marginals. The scalar variable is denoted by x∈ℝx\in\mathbb R, and the gradient space by V⊆ℝEV\subseteq\mathbb R^E, where EE is a finite coordinate set.

The characteristic function of a probability measure μ\mu on ℝ\mathbb R is the function μˆ\widehat\mu defined by (1): μˆ(t)=∫ℝeitxdμ(x),t∈ℝ.\widehat\mu(t)=\int_{\mathbb R}e^{\mathrm i t x}\,d\mu(x), \qquad t\in\mathbb R.(1) For a real exponent pp, the stable kernel Kp:ℝ2→ℝK_p:\mathbb R^2\to\mathbb R is the function defined by (2) for all real t,rt,r, and the stable scale constant cp∈ℝc_p\in\mathbb R is defined by (3). Kp(t,r)=(cos⁡(tr)−1)r−p−1.K_p(t,r)=(\cos(tr)-1)r^{-p-1}.(2)✦ cp=−p∫0∞Kp(1,r)dr.c_p=-p\int_0^\infty K_p(1,r)\,dr.(3)✦

For finite index sets II and EE and a real linear subspace V⊆ℝEV\subseteq\mathbb R^E, a coordinatewise pp-normalized family in VV is a family (gi)i∈I(g_i)_{i\in I} of vectors in VV satisfying (4): ∑i∈I|gi(e)|p=1,e∈E.\sum_{i\in I}|g_i(e)|^p=1, \qquad e\in E.(4) Write μI\mu^I for the product probability measure on ℝI\mathbb R^I with each factor equal to μ\mu. For a map Y:ℝI→VY:\mathbb R^I\to V and a vector x∈ℝIx\in\mathbb R^I, write Ye(x)=Y(x)(e)Y_e(x)=Y(x)(e) for its eeth coordinate.

The Claim

Theorem 1. Let 0<p<20<p<2. There is a probability measure μp\mu_p on ℝ\mathbb R with characteristic function (5), and the stable scale constant satisfies (6): μpˆ(t)=exp⁡(−cp|t|p),t∈ℝ.\widehat{\mu_p}(t)=\exp(-c_p|t|^p), \qquad t\in\mathbb R.(5)✦ cp>0.c_p>0.(6)✦

Let p>0p>0 and c∈ℝc\in\mathbb R, and let μ\mu be a probability measure on ℝ\mathbb R with μˆ(t)=exp⁡(−c|t|p)\widehat\mu(t)=\exp(-c|t|^p) for every t∈ℝt\in\mathbb R. For finite sets I,EI,E, a real linear subspace V⊆ℝEV\subseteq\mathbb R^E, and a coordinatewise pp-normalized family (gi)i∈I(g_i)_{i\in I} in VV, there exists a map YY satisfying (7): Y:ℝI→Vis measurable,(Ye)*μI=μ(e∈E).Y:\mathbb R^I\longrightarrow V\ \text{is measurable}, \qquad (Y_e)_*\mu^I=\mu\quad(e\in E).(7)✦

Let p>0p>0 and c≥0c\geq0, and let μ\mu be a probability measure on ℝ\mathbb R such that Re⁡μˆ(t)=exp⁡(−c|t|p)\operatorname{Re}\widehat\mu(t)=\exp(-c|t|^p) for every real tt. If p<2p<2 and c>0c>0, there is a constant k>0k>0 for which (8) holds: ka−p≤μ{x∈ℝ:|x|>a},a≥1.k a^{-p}\leq\mu\{x\in\mathbb R:|x|>a\}, \qquad a\geq1.(8)✦ For every a>0a>0, the upper tail and capped second moment satisfy (9) and (10): μ{x∈ℝ:|x|>a}≤2c2pa−p.\mu\{x\in\mathbb R:|x|>a\} \leq 2c\,2^p a^{-p}.(9)✦ ∫ℝmin⁡{x2,a2}dμ(x)≤24ca2−p.\int_{\mathbb R}\min\{x^2,a^2\}\,d\mu(x) \leq24c\,a^{2-p}.(10)✦

In the paper, the result reads: the symmetric intensity construction gives a VV-valued random vector YY whose coordinates have characteristic function exp⁡(p∫0∞(cos⁡(θu)−1)u−p−1du)\exp\bigl(p\int_0^\infty(\cos(\theta u)-1)u^{-p-1}\,du\bigr) for every real θ\theta and satisfy kpa−p≤ℙ(|Ye|>a)≤Cpa−pk_p a^{-p}\leq\mathbb P(|Y_e|>a)\leq C_p a^{-p} for a≥1a\geq1, with constants kp,Cp>0k_p,C_p>0.

What Is Checked

The scalar claims are stated over probability measures on ℝ\mathbb R, and the joint claim is stated over finite families in a real linear subspace with the normalization (4).