Eq. (14), one bit — the error probability of any estimate of a binary parameter is at least the minimum Bayes risk
ProvedFastCLO.LowerBound.eq_14bayes-risklower-boundsp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1
Let a binary parameter and an observation taking finitely many values have joint probabilities and , and let be any estimate of . Then
The left side is , the minimum Bayes risk; the right side is .
This is Eq. (14) of the paper for a single coordinate ; summing over coordinates and multiplying by gives the display. It is the step that turns the regret of an arbitrary algorithm into a quantity that depends only on the data distribution.
Formalization Note The expectations are written as finite sums over the observation space, which covers the finitely supported data of the paper's construction.
Preamble
import Mathlib
Formal statement
namespace FastCLO.LowerBound
/-- Eq. (14) for one bit (Hu, Kallus, Mao, arXiv:2011.03030v3, proof of Theorem 3, p. 21; reused in
the proof of Theorem 7, p. 30): the probability that an estimate `b̂(D)` of a binary parameter `b`
is wrong is at least the minimum Bayes risk `E[min{P(b = 1 | D), 1 − P(b = 1 | D)}]`.
Formalization Note: the data take finitely many values `ω`; `w0 ω = P(b = 0, D = ω)` and
`w1 ω = P(b = 1, D = ω)`. Then `E[min posterior] = Σ_ω min(w0 ω, w1 ω)` and
`P(b̂ ≠ b) = Σ_ω (w0 ω if b̂(ω) = 1 else w1 ω)`. -/
theorem eq_14 {Ω : Type*} [Fintype Ω] (w0 w1 : Ω → ℝ) (h0 : ∀ ω, 0 ≤ w0 ω) (h1 : ∀ ω, 0 ≤ w1 ω)
(bhat : Ω → Bool) :
∑ ω, min (w0 ω) (w1 ω) ≤ ∑ ω, (if bhat ω then w0 ω else w1 ω) := by sorry
end FastCLO.LowerBound
Source
Hu, Kallus, Mao, Fast Rates for Contextual Linear Optimization, arXiv:2011.03030v3, proof of Theorem 3 (A.2), Eq. (14), p. 21; reused in the proof of Theorem 7, p. 30
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.