Inequality (3) — conditional loss in the positive-gain case
ProvedDoubleGreedyUSM.Randomized.inequality_3approximation-algorithmsp2o-batch-pfp1ap2o-gran-per-chapterp2o-plan-paperp2o-v1randomized-algorithmssubmodular-functions
Let be submodular. Take , , and any comparison set . Put , , and . If and , then
This bounds the one-step expected loss of the comparison set in Case 3 of Lemma III.1.
Formalization Note The statement covers any nested pair with , a generalization of the reachable states conditioned on in the paper. The positivity of ensures the denominator is nonzero.
Preamble
import Mathlib import Definitions.Def_NonmonotoneSubmod_Shared_Submodular import Definitions.Def_DoubleGreedyUSM_Randomized_Algorithm2
Formal statement
namespace DoubleGreedyUSM.Randomized
/-- Inequality (3), Case 3 of the proof of Lemma III.1 (PDF p. 6). -/
theorem inequality_3 {X : Type} [Fintype X] [DecidableEq X]
(f : Finset X → ℝ) (hf : NonmonotoneSubmod.Shared.Submodular f)
(Xs Ys O : Finset X) (u : X) (hsub : Xs ⊆ Ys)
(huY : u ∈ Ys) (huX : u ∉ Xs)
(ha : 0 ≤ f (insert u Xs) - f Xs)
(hb : 0 < f (Ys.erase u) - f Ys) :
let a := f (insert u Xs) - f Xs
let b := f (Ys.erase u) - f Ys
let P := (O ∪ Xs) ∩ Ys
a / (a + b) * (f P - f (insert u P)) +
b / (a + b) * (f P - f (P.erase u)) ≤ a * b / (a + b) := by sorry
end DoubleGreedyUSM.Randomized
Source
Buchbinder, Feldman, Naor, Schwartz, A Tight Linear Time (1/2)-Approximation for Unconstrained Submodular Maximization, FOCS 2012 version, proof of Lemma III.1, inequality (3) (PDF p. 6)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.