Lemma 2.1, p. 8 — E[⟨x,u⟩u] = x and E⟨x,u⟩² = ‖x‖² for u the projection of a standard Gaussian onto a subspace containing x
ProvedZORiem.RGD.lemma_2_1gaussian-identitiesgaussian-smoothingp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1zeroth-order-optimization
Let be a finite-dimensional real inner product space (the ambient ), let be a linear subspace of dimension with orthogonal projection , let be a standard Gaussian vector of and . Then for every the random vector is integrable and
In the paper both expectations are written as integrals against the Gaussian density , . Applied with and , the identities say that the projected Gaussian direction is isotropic on the tangent space, which is the basis of every bias estimate for the estimator .
Formalization Note The standard Gaussian is Mathlib's stdGaussian E, so is the integral against it. The dimension hypothesis is the page's "-dimensional" and is not needed for the conclusion.
Preamble
import Mathlib open MeasureTheory ProbabilityTheory open scoped InnerProductSpace
Formal statement
namespace ZORiem.RGD
/-- Lemma 2.1, p. 8: for a `d`-dimensional subspace `X` of `E` with orthogonal projection `P`, a
standard Gaussian `u₀ ∼ N(0, I_n)` and `u = P u₀`, every `x ∈ X` satisfies `E[⟨x, u⟩ u] = x` and
`E[⟨x, u⟩²] = ‖x‖²` (the vector-valued integrand is integrable). -/
theorem lemma_2_1 {E : Type*} [NormedAddCommGroup E] [InnerProductSpace ℝ E]
[FiniteDimensional ℝ E] [MeasurableSpace E] [BorelSpace E]
(X : Submodule ℝ E) (d : ℕ) (hd : Module.finrank ℝ X = d) (x : E) (hx : x ∈ X) :
Integrable (fun u₀ : E => ⟪x, X.starProjection u₀⟫_ℝ • X.starProjection u₀) (stdGaussian E) ∧
∫ u₀, ⟪x, X.starProjection u₀⟫_ℝ • X.starProjection u₀ ∂(stdGaussian E) = x ∧
∫ u₀, ⟪x, X.starProjection u₀⟫_ℝ ^ 2 ∂(stdGaussian E) = ‖x‖ ^ 2 := by sorry
end ZORiem.RGD
Source
Li, Balasubramanian and Ma, arXiv:2003.11238v3, Lemma 2.1, (2.6), p. 8
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.