Theorem 5 β ΞΎ^K(Y, X|π) β€ π^K[Y], with equality iff Y = f(X) a.s. for an π-measurable f
ProvedOTSensitivity.MaxFun.theorem_5Let be a probability space, and Polish spaces, and , measurable, with laws and . Let be continuous with , and assume
Let be any sub--algebra of the Borel sets of and any version of the conditional distribution of given , and let . Then
- (equation (20)); in particular is finite;
- if and only if is functionally dependent on through , that is, -almost surely for some -measurable map .
This is the max-functionality property (RΓ©nyi's Postulate E) of OT-based sensitivity measures: the measure is bounded by an explicit constant of the output's law, and reaches it exactly when the information carried by determines the output.
Formalization Note and are Polish (the paper's supports in , are such spaces); all values lie in . "The assumptions of Lemma 4" are the continuity of , the standing assumption of Β§3.1, and ; the finiteness is assumed for the whole theorem, as on the page. The statement holds for every sub--algebra and every version of the conditional distribution (any kernel satisfying IsCondDistrib). is the published RWPI.SqrtLasso.transportCost.
import Mathlib import Definitions.Def_RWPI_SqrtLasso_transportCost import Definitions.Def_OTSensitivity_MaxFun_Setting open MeasureTheory ProbabilityTheory open scoped ENNReal
namespace OTSensitivity.MaxFun
/-- Theorem 5, p. 10 (max-functionality): under the assumptions of Lemma 4 (`k` continuous,
vanishing exactly on the diagonal, `π^K[Y] < β`), for every sub-Ο-algebra `π` of `π(π³)` and
every version `ΞΊ` of the conditional distribution of `Y` given `(X, π)`,
`ΞΎ^K(Y, X | π) β€ π^K[Y]` (20), and equality holds if and only if `Y = f(X)` `P`-a.e. for some
`π`-measurable `f : π³ β π΄`. -/
theorem theorem_5 {Ξ© π³ π΄ : Type*} [MeasurableSpace Ξ©]
[TopologicalSpace π³] [PolishSpace π³] [mπ³ : MeasurableSpace π³] [BorelSpace π³]
[TopologicalSpace π΄] [PolishSpace π΄] [MeasurableSpace π΄] [BorelSpace π΄]
(P : Measure Ξ©) [IsProbabilityMeasure P]
(X : Ξ© β π³) (Y : Ξ© β π΄) (hX : Measurable X) (hY : Measurable Y)
(k : π΄ β π΄ β ββ₯0β) (hk : IsCost k)
(hMK : MK k (P.map Y) < β€)
(F : MeasurableSpace π³) (hF : F β€ mπ³) (ΞΊ : Kernel π³ π΄) (hΞΊ : IsCondDistrib P X Y F ΞΊ) :
xi k P X Y ΞΊ β€ MK k (P.map Y) β§
(xi k P X Y ΞΊ = MK k (P.map Y) β IsFunctionallyDependent P X Y F) := by sorry
end OTSensitivity.MaxFun
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.