OctoInformBeta

OpenAI preprint, September 2026 · §5. Contractions into ℓ₁

Lemma 5.2: For C1C^1 functions on ℝm\mathbb{R}^m with factor π/2\pi/2, bounded value and gradient in the Gaussian case, rotation invariance, and endpoint bounds

Cube and Gaussian measures satisfy first-moment Poincaré inequalities.

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 a nonnegative integer mm, let E=ℝmE = \mathbb{R}^m be mm-dimensional Euclidean space with norm ∥⋅∥\|\cdot\| and inner product ⟨⋅,⋅⟩\langle \cdot, \cdot \rangle, and let T=(−1,1)mT = (-1,1)^m be the open cube in EE. For a real-valued C1C^1 function hh with gradient ∇h\nabla h on a neighborhood of TT, the question is how the expected displacement 𝔼|h(θ)−h(η)|\mathbb{E}|h(\theta) - h(\eta)| between independent uniform draws θ,η∈T\theta, \eta \in T is bounded in terms of 𝔼∥∇h(θ)∥\mathbb{E}\|\nabla h(\theta)\|, and how the corresponding inequality behaves under rotations of standard Gaussian vectors in EE.

Measures and Rotation Paths

The uniform measure ν\nu on (−1,1)(-1,1) is the probability measure on ℝ\mathbb{R} obtained by restricting Lebesgue measure to the open interval (−1,1)(-1,1) and scaling by 1/21/2.✦ On EE, the cube uniform measure μ\mu is the product probability measure of mm copies of ν\nu indexed by the coordinates of EE.✦ The standard Gaussian measure γ\gamma on EE is the distribution on EE of a vector with mm independent standard normal coordinates.

For t∈ℝt \in \mathbb{R} and z=(z1,z2)∈E×Ez = (z_1, z_2) \in E \times E, the rotation path pt(z)∈Ep_t(z) \in E is defined by pt(z)=(cos⁡t)z1+(sin⁡t)z2.p_t(z) = (\cos t) z_1 + (\sin t) z_2.(1)✦ The rotation path tangent qt(z)∈Eq_t(z) \in E at t∈ℝt \in \mathbb{R} and z=(z1,z2)∈E×Ez = (z_1, z_2) \in E \times E is defined by qt(z)=−(sin⁡t)z1+(cos⁡t)z2.q_t(z) = -(\sin t) z_1 + (\cos t) z_2.(2)✦

The Claim

Theorem 1. For any C1C^1 function h:E→ℝh : E \to \mathbb{R}, the cube Poincaré inequality (3) states ∫E×E|h(z1)−h(z2)|d(μ×μ)(z1,z2)≤π2∫E∥∇h(x)∥dμ(x).\int_{E \times E} |h(z_1) - h(z_2)| \, d(\mu \times \mu)(z_1, z_2) \le \frac{\pi}{2} \int_E \|\nabla h(x)\| \, d\mu(x).(3)✦ Suppose h:E→ℝh : E \to \mathbb{R} is C1C^1 and constants B∈ℝB \in \mathbb{R} and M≥0M \ge 0 satisfy |h(x)|≤B|h(x)| \le B and ∥∇h(x)∥≤M\|\nabla h(x)\| \le M for all x∈Ex \in E; then the rotation Poincaré inequality (4) gives ∫E×E|h(z1)−h(z2)|d(γ×γ)(z1,z2)≤π2∫E∥∇h(x)∥dγ(x).\int_{E \times E} |h(z_1) - h(z_2)| \, d(\gamma \times \gamma)(z_1, z_2) \le \frac{\pi}{2} \int_E \|\nabla h(x)\| \, d\gamma(x).(4)✦ If f:E×E→ℝf : E \times E \to \mathbb{R} is (γ×γ)(\gamma \times \gamma)-measurable and t∈ℝt \in \mathbb{R}, then integral rotation invariance (5) reads ∫E×Ef(pt(z),qt(z))d(γ×γ)(z)=∫E×Ef(z)d(γ×γ)(z).\int_{E \times E} f(p_t(z), q_t(z)) \, d(\gamma \times \gamma)(z) = \int_{E \times E} f(z) \, d(\gamma \times \gamma)(z).(5)✦ Whenever h:E→ℝh : E \to \mathbb{R} is C1C^1 and z=(z1,z2)∈E×Ez = (z_1, z_2) \in E \times E, the rotation path endpoint bound (6) yields |h(z1)−h(z2)|≤∫0π/2|⟨∇h(pt(z)),qt(z)⟩|dt.|h(z_1) - h(z_2)| \le \int_0^{\pi/2} |\langle \nabla h(p_t(z)), q_t(z) \rangle| \, dt.(6)✦

In the paper, the result reads: for a real-valued C1C^1 function hh on a neighborhood of T=(−1,1)mT = (-1,1)^m and independent uniform points θ,η∈T\theta, \eta \in T, 𝔼|h(θ)−h(η)|≤𝔼∥∇h(θ)∥\mathbb{E}|h(\theta) - h(\eta)| \le \mathbb{E}\|\nabla h(\theta)\|. The bound (3) is stated for hh on all of EE under the cube uniform measure μ\mu and carries the factor π/2\pi/2 on the right-hand side.

What Is Checked

The claims (3), (4), (5), and (6) are stated over mm-dimensional Euclidean space EE, the cube uniform measure μ\mu built from the uniform measure ν\nu on (−1,1)(-1,1), the standard Gaussian measure γ\gamma on EE, and the rotation path pt(z)p_t(z) and rotation path tangent qt(z)q_t(z), for C1C^1 functions h:E→ℝh : E \to \mathbb{R} (with bounded values and bounded gradient in (4), and a pair z∈E×Ez \in E \times E in (6)) and (γ×γ)(\gamma \times \gamma)-measurable functions f:E×E→ℝf : E \times E \to \mathbb{R} at t∈ℝt \in \mathbb{R} in (5).