Lemma 1 — the quantile is non-decreasing and left continuous in , and
ProvedQuantileMDP.lemma_1p2o-gran-per-chapterp2o-plan-paperp2o-v1probabilityquantile
Let be a probability law on , the distribution of a random variable , and let be its -quantile (Definition 1, footnote 1), with values in . Then
- is non-decreasing on ;
- is left continuous at every (as a function on );
- for every ,
These are the basic properties of the quantile used throughout the dynamic programming of the paper; in particular the third one says that the -quantile is attained.
Formalization Note No boundedness or finiteness hypothesis is placed on ; the quantile is extended-real valued.
Differences from the source:
- encoding: a random variable → its law , an arbitrary probability measure on ; reason: the quantile and depend on only through its law.
- encoding: , → the essential infimum and supremum of footnote 1, valued in the extended reals; reason: footnote 1 gives the precise form, and occur for unbounded .
- encoding: "left continuous … function of " → continuity within at each ; reason: is the domain on which Definition 1 defines , and left continuity at is vacuous.
Preamble
import Mathlib import Definitions.Def_QuantileMDP_quantile open MeasureTheory
Formal statement
namespace QuantileMDP
/-- Lemma 1: for a random variable `X` with law `μ`, `τ ↦ Q_τ(X)` is non-decreasing and left
continuous on `[0,1]`, and `P(X ≤ Q_τ(X)) ≥ τ`.
Li, Zhong, Brandeau, *Quantile Markov Decision Processes*, arXiv:1711.05788v5, §3.1, Lemma 1, p.9.
**Formalization Note** Stated for every probability measure `μ` on `ℝ` (the law of `X`), with the
`EReal`-valued quantile of Definition 1 and footnote 1. Left continuity is continuity within
`[0, τ]` at every `τ ∈ (0, 1]`; the event `{X ≤ Q_τ(X)}` is `{x | (x : EReal) ≤ Q_τ}`. -/
theorem lemma_1 (μ : Measure ℝ) [IsProbabilityMeasure μ] :
MonotoneOn (quantile μ) (Set.Icc 0 1) ∧
(∀ τ ∈ Set.Ioc (0 : ℝ) 1, ContinuousWithinAt (quantile μ) (Set.Icc 0 τ) τ) ∧
(∀ τ ∈ Set.Icc (0 : ℝ) 1, ENNReal.ofReal τ ≤ μ {x : ℝ | (x : EReal) ≤ quantile μ τ}) := by sorry
end QuantileMDP
Source
Li, Zhong, Brandeau, Quantile Markov Decision Processes, arXiv:1711.05788v5, §3.1, Lemma 1, p.9
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.