Theorem 6.2 (proof), p. 106 — E_{M∼µ}E_{M̃∼µ}[(f^M(π) − f^{M̃}(π))²] ≤ 4·E_{M∼µ}[(f^M(π) − f^{M̄}(π))²]
ProvedStatComplexityDM.Disagreement.sq_shift_le_fourdecision-estimation-coefficientp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-bookp2o-v1posterior-sampling
Let be a finitely supported prior over models with mean-reward functions , and let be the mean reward of a reference model . Then for every decision ,
where and are independent draws from .
This inequality lets the proof of Theorem 6.2 replace the decoupled error against a second draw from the prior by the square-loss error against the reference model.
Formalization Note The prior is a finite weighted family of mean-reward functions.
Preamble
import Mathlib import Definitions.Def_StatComplexityDM_Disagreement_Coefficient
Formal statement
namespace StatComplexityDM.Disagreement
/-- Proof of Theorem 6.2, p. 106 (arXiv:2112.13487v3): for a finitely supported prior
`µ = Σ_i w i · δ_{M_i}` with mean rewards `fM i = f^{M_i}`, a reference mean reward `fbar = f^{M̄}`,
and every decision `π`,
`E_{M∼µ} E_{M̃∼µ}[(f^M(π) − f^{M̃}(π))²] ≤ 4 E_{M∼µ}[(f^M(π) − f^{M̄}(π))²]`. -/
theorem sq_shift_le_four {ι Act : Type*} [Fintype ι] (w : ι → ℝ) (hw : StatComplexityDM.LowerBound.IsDist w)
(fM : ι → Act → ℝ) (fbar : Act → ℝ) (π : Act) :
∑ i, w i * ∑ k, w k * (fM i π - fM k π) ^ 2 ≤ 4 * ∑ i, w i * (fM i π - fbar π) ^ 2 := by sorry
end StatComplexityDM.Disagreement
Source
arXiv:2112.13487v3, App. E.2.2, proof of Theorem 6.2, p. 106 ("Observe that for all π ∈ Π")
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.