Lemma 1 (1), p. 15 — the inner product distance satisfies I_Γ(π₁,π₂) ≤ I_Γ(π₁,π₃) + I_Γ(π₃,π₂)
ProvedOfflinePolicyLearning.Regret.lemma1_triangle_inner_prod_distp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1policy-learningpseudometrictriangle-inequality
Let be a policy class, points and weights, and let be the inner product distance of Definition 6 (actions identified with basis vectors, so ). For all ,
Together with symmetry this makes a pseudometric on , which is how the paper's chaining argument uses it.
Formalization Note Only part 1 of Lemma 1 is posed; part 2 () is false as printed (a three-point counterexample is recorded in the mission description). The convention is Lean's division by zero.
Preamble
import Mathlib import Definitions.Def_OfflinePolicyLearning_Regret_Hamming import Definitions.Def_OfflinePolicyLearning_Regret_Setting open MeasureTheory ProbabilityTheory Filter Topology
Formal statement
namespace OfflinePolicyLearning.Regret
theorem lemma1_triangle_inner_prod_dist {X : Type*} {d n : ℕ} (Pol : Set (X → Fin d))
(x : Fin n → X) (γ : Fin n → Fin d → ℝ) (π₁ π₂ π₃ : X → Fin d)
(h₁ : π₁ ∈ Pol) (h₂ : π₂ ∈ Pol) (h₃ : π₃ ∈ Pol) :
innerProdDist Pol x γ π₁ π₂ ≤ innerProdDist Pol x γ π₁ π₃ + innerProdDist Pol x γ π₃ π₂ := by sorry
end OfflinePolicyLearning.Regret
Source
Zhou, Athey & Wager, Offline Multi-Action Policy Learning, arXiv:1810.04778v2, p. 15, Lemma 1 (1)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.