Proof of Theorem 1, p. 32 — Δ = 𝔼_{P_Y}S(P_X, •) − 𝔼_{P_Y}S(P_Y, •) = ∫ (f_X ∗ w − f_Y ∗ w)² h dz ≥ 0
ProvedMultivarScores.L2Rule.divergence_eqdivergencep2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1proper-scoring-rulescoring-rules
Let , and a class of probability distributions on satisfy Assumption 1, and let . Then the divergence of the L² scoring rule (2) is
and this quantity is nonnegative.
This is the computation in the proof of Theorem 1; it shows that the divergence of is precisely the weighted L² distance (1) between the smoothed densities, from which properness follows.
Formalization Note is the integral of against . Assumption 1 is the structure described in the Setting item.
Preamble
import Mathlib import Definitions.Def_MultivarScores_L2Rule_Setting
Formal statement
namespace MultivarScores.L2Rule
open MeasureTheory
/-- Divergence display in the proof of Theorem 1, p. 32: under Assumption 1,
`Δ = 𝔼_{P_Y}[S(P_X, •)] − 𝔼_{P_Y}[S(P_Y, •)] = ∫ (f_X ∗ w − f_Y ∗ w)²(z) h(z) dz ≥ 0`. -/
theorem divergence_eq {d : ℕ} (w h : E d → ℝ) (𝒱 : Set (Measure (E d)))
(hA : Assumption1 w h 𝒱) (PX PY : Measure (E d)) (hPX : PX ∈ 𝒱) (hPY : PY ∈ 𝒱) :
(∫ y, S PX w h y ∂PY) - (∫ y, S PY w h y ∂PY) = L2div PX PY w h ∧
0 ≤ L2div PX PY w h := by sorry
end MultivarScores.L2Rule
Source
Meng, Taylor, Ben Taieb, Li, Scores for Multivariate Distributions and Level Sets, arXiv:2002.09578v6, p. 32, proof of Theorem 1, divergence display
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.