Theorem 8.1.1 — Gaussian concentration of a linear supremum
ProvedTalagrandConc.LinearSuprema.theorem_8_1_1concentration-of-measurep2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-bookp2o-v1probability
Let be a nonempty family of real -tuples with finite, positive . Let be independent real random variables, with almost surely for each , and put
Assume is measurable, and let be any median of . For every ,
The bound gives Gaussian-scale concentration for an arbitrary, possibly infinite family of linear forms and does not require identical coordinate laws.
Formalization Note The independent variables are represented on their canonical product space. The assumptions that is nonempty, is finite and positive, and is measurable make the paper's implicit real-valued probability statement precise.
Preamble
import Mathlib import Definitions.Def_TalagrandConc_LinearSuprema_Basic
Formal statement
namespace TalagrandConc.LinearSuprema
theorem theorem_8_1_1 {N : ℕ} (F : Set (Fin N → ℝ))
(hF : F.Nonempty) (hσfinite : BddAbove (coeffNorm '' F))
(hσ : 0 < sigma F) (r : Fin N → ℝ)
(μ : Fin N → MeasureTheory.Measure ℝ)
[∀ i, MeasureTheory.IsProbabilityMeasure (μ i)]
(hμ : ∀ i, μ i (Set.Icc (r i) (r i + 1)) = 1)
(hZmeas : Measurable (linearSupremum F))
(M : ℝ)
(hM : TalagrandConc.BinPacking.IsMedian (MeasureTheory.Measure.pi μ) (linearSupremum F) M)
(u : ℝ) (hu : 0 < u) :
(MeasureTheory.Measure.pi μ) {x | |linearSupremum F x - M| ≥ u} ≤
ENNReal.ofReal (4 * Real.exp (-(u ^ 2) / (4 * (sigma F) ^ 2))) := by sorry
end TalagrandConc.LinearSuprema
Source
Talagrand, Concentration of measure and isoperimetric inequalities in product spaces, Publ. Math. IHÉS 81 (1995), p. 156, Theorem 8.1.1, Eq. (8.1.1)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.