App. A.3, p.22 — the density is at most
ProvedLatentRobust.gaussian_pdf_leadversarial-robustnessclassificationgaussian-mixturemachine-learningp2o-gran-per-chapterp2o-plan-paperp2o-v1
For every mean , every variance and every , the density of the normal law satisfies
It bounds the probability that a normal variable falls into a short interval by the interval's length times , which turns the slab probabilities of the closed-form step into (19).
Differences from the source:
- reading: N(ν, σ) with density bound 1/√(2πσ) → σ is the variance (Mathlib's gaussianPDFReal ν σ); reason: the page applies the bound to N(a, a) with a a variance, and 1/√(2πσ) is the density's maximum exactly when σ is the variance
Preamble
import Mathlib
Formal statement
namespace LatentRobust
open ProbabilityTheory
/-- **Gaussian density bound** (Javanmard, Mehrabi, arXiv:2110.11950v2, App. A.3, p.22): "the pdf of normal `N(ν, σ)` is
bounded by `1/√(2πσ)`, for any values of `σ > 0` and `ν`".
**Formalization Note** As on the page (which applies it to `N(a, a)`, `a` a variance), the
second parameter `σ` is the variance; Mathlib's `gaussianPDFReal ν v` is the density of the
normal law with mean `ν` and variance `v`. -/
theorem gaussian_pdf_le (ν : ℝ) (v : NNReal) (hv : 0 < v) (x : ℝ) :
gaussianPDFReal ν v x ≤ 1 / Real.sqrt (2 * Real.pi * v) := by sorry
end LatentRobust
Source
Javanmard, Mehrabi, Adversarial robustness for latent models: Revisiting the robust-standard accuracies tradeoff, arXiv:2110.11950v2, App. A.3 (proof of Theorem 3.1), p.22, "the pdf of normal N(ν, σ) is bounded by 1/√2πσ"
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.