§6, p. 387 — Brier regret as a difference of squared calibration errors
ProvedAsympCalib.Main.S_eq_sq_sub_sqbrier-scorecalibrationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1regret
For grid points and , any real outcome sequence , and any sequence of probability vectors on the grid, the signed regret after rounds obeys
This algebraic identity connects pairwise Brier-score regret to the weighted calibration score in Theorem 3.
Formalization Note The identity holds for real ; the surrounding calibration theorems restrict outcomes to or . The page's middle line has where expansion gives ; the final identity, used here, has the correct sign and indices. Rounds are indexed from , and is a probability vector so zero weighted count also gives zero weighted outcome count.
Preamble
import Mathlib import Definitions.Def_AsympCalib_Main_Setting noncomputable section
Formal statement
namespace AsympCalib.Main
theorem S_eq_sq_sub_sq (k : ℕ) (hk : 0 < k) (X : ℕ → ℝ)
(μ : ℕ → Fin (k + 1) → ℝ)
(hμ : ∀ s, CalibratedCE.Forecast.IsDist (μ s))
(t : ℕ) (i j : Fin (k + 1)) :
CalibratedCE.Forecast.S (brierLoss k X) μ t i j =
nTilde k μ t i * (rhoTilde k X μ t i - gridPt k i) ^ 2 -
nTilde k μ t i * (rhoTilde k X μ t i - gridPt k j) ^ 2 := by sorry
end AsympCalib.Main
Source
Foster and Vohra, Asymptotic calibration, Biometrika 85 (1998), p. 387, §6, proof of Theorem 3, first display; https://doi.org/10.1093/biomet/85.2.379
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.