Proof of Theorem 3.15 — on
ProvedSQMeanEstimation.ring_l1_boundlearning-theorymean-estimationp2o-gran-per-chapterp2o-plan-paperp2o-v1statistical-queries
Let , and with . Then
Since the nonzero coordinates of exceed , its mass is controlled by its mass, which is at most that of .
Differences from the source:
- encoding: ‖·‖₁ and ‖·‖_q^q → ∑ i, |x i| and ∑ i, |x i| ^ q; reason: the definitions of these quantities
- encoding: w ∈ B_q^d → lqNorm q w ≤ 1; reason: the unit ball of ‖·‖_q
Preamble
import Mathlib import Definitions.Def_SQMeanEstimation_lqNorm import Definitions.Def_SQMeanEstimation_posPart import Definitions.Def_SQMeanEstimation_ringPart
Formal statement
namespace SQMeanEstimation
/-- **ℓ₁ bound on a ring of the positive part.**
Feldman, Guzmán, Vempala, arXiv:1512.09170v2, p.18: "We next observe that by Lemma 3.12, for every
w ∈ Bᵈ_q, ‖Rⱼ(w⁺)‖₁ ≤ (2^{−j−1})^{1−q}‖Rⱼ(w⁺)‖_q^q ≤ (2^{−j−1})^{1−q}."
For `q ∈ (1, 2)`, `j ∈ ℕ` and `w` with `‖w‖_q ≤ 1`, both inequalities of the chain hold.
**Formalization Note.** `‖·‖₁` is `Σ |·|` and `‖·‖_q^q` is `Σ |·|^q`. Powers are `Real.rpow`. -/
theorem ring_l1_bound {d : ℕ} (q : ℝ) (hq1 : 1 < q) (hq2 : q < 2) (j : ℕ) (w : Fin d → ℝ)
(hw : lqNorm q w ≤ 1) :
∑ i, |ringPart j (posPart w) i| ≤
((2 : ℝ) ^ (-(j : ℝ) - 1)) ^ (1 - q) * ∑ i, |ringPart j (posPart w) i| ^ q ∧
((2 : ℝ) ^ (-(j : ℝ) - 1)) ^ (1 - q) * ∑ i, |ringPart j (posPart w) i| ^ q ≤
((2 : ℝ) ^ (-(j : ℝ) - 1)) ^ (1 - q) := by sorry
end SQMeanEstimation
Source
Feldman, Guzmán, Vempala, Statistical Query Algorithms for Mean Vector Estimation and Stochastic Convex Optimization, arXiv:1512.09170v2, proof of Theorem 3.15, p.18, display after "We next observe that by Lemma 3.12"
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.