Popoviciu's inequality: `Var_p(r) ≤ range(r)²/4`.
ProvedRLHF.variance_le_range_sqaether-catalogalgebra
Popoviciu's inequality: Var_p(r) ≤ range(r)²/4.
theorem RLHF.variance_le_range_sq{p r : Ω → ℝ} (hp : IsDist p) :
variance p r ≤ rewardRange r ^ 2 / 4 := by sorry
Formalization Note Transplanted verbatim from the Aether Catalog source Algebra/RLHFDriftCore.lean; the statement is byte-identical to the source declaration, elaborated with autoImplicit disabled in the platform environment.
Preamble
-- Thm stub generated from Algebra/RLHFDriftCore.lean
import Mathlib
import Definitions.Def_Algebra_RLHFDriftCore
/-!
# Core objects of the RLHF alignment-drift catalogue
Domain: Algebra (convex analysis × information theory × alignment theory).
This file collects the base objects that the RLHF drift thread of the catalogue is
phrased in — finite probability vectors, the Gibbs (KL-regularised) policy, the
Kullback–Leibler divergence, the total-variation (`ℓ¹`) distance, and the first two
moment functionals of a reward — together with the elementary API used downstream.
The names and definitions follow the conventions of the RLHF drift files of the
catalogue (`RLHF.IsDist`, `RLHF.gibbsPolicy`, `RLHF.klDiv`, `RLHF.l1Dist`,
`RLHF.rewardRange`, `RLHF.mean`, `RLHF.variance`, ...), so the results proved here and
in `Algebra.RLHFMeanAbsoluteDeviation` / `Algebra.RLHFKLSecondOrder` speak about
exactly the same objects as `RLHF.kl_gibbs_le_variance` and
`RLHF.gibbs_l1_le_variance`.
Nothing in this file is deep; it exists so that the two research files that follow are
self-contained and compile.
-/
open RLHF
open Finset
variable {Ω : Type*} [Fintype Ω]
/-! ## 1. Finite probability vectors -/
/-! ## 2. Moment functionals -/
/-! ## 3. The Gibbs policy -/
variable [Nonempty Ω]Formal statement
theorem RLHF.variance_le_range_sq{p r : Ω → ℝ} (hp : IsDist p) :
variance p r ≤ rewardRange r ^ 2 / 4 := by sorrySource