Lemma 1, p. 56 — variant of Jensen's inequality: 𝔼[max{X + φ(X), 0}] ≥ 𝔼[max{𝔼X + φ(X), 0}]
ProvedInfoRelaxTS.Monotone.lemma_1jensen-inequalityp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1probability
Let be a probability space, a non-decreasing function and a real random variable with . Then
Replacing by its mean inside the first argument can only lower the expected positive part, even though keeps its randomness. This is the one-dimensional engine behind .
Formalization Note Both sides are expectations of nonnegative variables and are taken as lower Lebesgue integrals in , so neither side can be a junk value. "Increasing" is read as non-decreasing (the paper's proof uses only monotonicity).
Preamble
import Mathlib import Definitions.Def_InfoRelaxTS_Monotone_Setting
Formal statement
namespace InfoRelaxTS.Monotone
open MeasureTheory ProbabilityTheory
/-- Lemma 1 (variant of Jensen's inequality), p. 56, (125): for a non-decreasing `φ : ℝ → ℝ` and an
integrable `X`, `𝔼[max{X + φ(X), 0}] ≥ 𝔼[max{𝔼(X) + φ(X), 0}]`. Both sides are expectations of
nonnegative variables, taken in `[0, ∞]`. -/
theorem lemma_1 {Ω : Type*} [MeasurableSpace Ω] (μ : Measure Ω) [IsProbabilityMeasure μ]
(φ : ℝ → ℝ) (hφ : Monotone φ) (X : Ω → ℝ) (hX : Integrable X μ) :
∫⁻ ω, ENNReal.ofReal (max ((∫ ω', X ω' ∂μ) + φ (X ω)) 0) ∂μ
≤ ∫⁻ ω, ENNReal.ofReal (max (X ω + φ (X ω)) 0) ∂μ := by sorry
end InfoRelaxTS.Monotone
Source
Min, Maglaras, Moallemi, Thompson Sampling with Information Relaxation Penalties, arXiv:1902.04251v2, p. 56, Lemma 1, (125)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.