Equality in the superadditivity of iff proportional
ProvedBanditAlgorithm.pair_weight_add_eq_iff_proportionalbanditssequential-analysis
For positive reals, equality holds in the superadditivity of ,
if and only if , that is, if and only if and lie on a common ray through the origin.
Both directions follow from the polynomial identity
after clearing the (positive) denominators.
Since is homogeneous of degree one it is genuinely linear along each ray, so no concavity statement about it can be strict in every direction; proportionality is exactly the degenerate direction, and the statement above says it is the only one. That dichotomy is what upgrades convexity of the set of maximisers of to uniqueness: two maximisers would have to satisfy for every , hence be proportional, and two proportional probability vectors are equal.
Preamble
import Mathlib.Analysis.SpecialFunctions.Log.Basic
Formal statement
theorem BanditAlgorithm.pair_weight_add_eq_iff_proportional {x₁ y₁ x₂ y₂ : ℝ}
(hx₁ : 0 < x₁) (hy₁ : 0 < y₁) (hx₂ : 0 < x₂) (hy₂ : 0 < y₂) :
(x₁ * y₁ / (x₁ + y₁) + x₂ * y₂ / (x₂ + y₂)
= (x₁ + x₂) * (y₁ + y₂) / ((x₁ + x₂) + (y₁ + y₂)))
↔ x₁ * y₂ = x₂ * y₁ := by
sorrySource
Equality case of the superadditivity of the parallel sum; the strict-concavity-transverse-to-rays statement underlying uniqueness of the optimal allocation in Garivier & Kaufmann, Optimal Best Arm Identification with Fixed Confidence, COLT 2016, Lemma 4.