Theorem 1 — the Allocation Algorithm is -competitive
ProvedBJNAdAuctions.Basic.allocation_algorithm_competitiveConsider an instance of the online ad-auctions problem with a nonempty finite set of buyers, budgets and bids on products. Let bound the bid-to-budget ratios, for all , and let
Run the Allocation Algorithm with parameter and any tie-breaking rule that picks a buyer maximizing . Then for every feasible solution of the packing LP of Fig. 2, the revenue collected by the algorithm satisfies
Taking (when positive) gives the paper's Theorem 1: the Allocation Algorithm is -competitive with . The bound is against the optimum of the fractional LP, which is at least the revenue of any integral offline allocation, so it implies the competitive ratio against the integral offline optimum. Since the algorithm is deterministic and the statement holds for every instance, it holds against every adversary.
Formalization Note The revenue is revenue inst c sel, the sum over buyers of the amounts charged by the run that AllocationAlgorithm computes from the instance; the tie-breaking rule sel is universally quantified subject to the argmax property. The statement is for any bound on the ratios, not only their exact maximum; the exact maximum is one admissible choice.
import Mathlib import Definitions.Def_BJNAdAuctions_Basic_Instance import Definitions.Def_BJNAdAuctions_Basic_AllocationAlgorithm
namespace BJNAdAuctions.Basic
/-- **Theorem 1** (p. 7), first sentence: the Allocation Algorithm is
`(1 - 1/c)(1 - R_max)`-competitive, where `c = (1 + R_max)^(1/R_max)`. Stated for any bound
`R > 0` with `b(i, j) ≤ R · B(i)` for all `i, j` (the exact `R_max` is one such bound), for every
tie-breaking of the argmax, against every feasible solution `y'` of the fractional packing LP of
Fig. 2 (whose optimum bounds the integral offline optimum from above). -/
theorem allocation_algorithm_competitive {I : Type*} [Fintype I] [Nonempty I] {m : ℕ}
(inst : Instance I m) (R : ℝ) (hR : 0 < R) (hbR : ∀ i j, inst.b i j ≤ R * inst.B i)
(sel : (I → ℝ) → Fin m → I) (hsel : IsArgmaxRule inst sel)
(y' : I → Fin m → ℝ) (hy' : PackingFeasible inst y') :
(1 - 1 / (1 + R) ^ (1 / R)) * (1 - R) * packingValue inst y' ≤
revenue inst ((1 + R) ^ (1 / R)) sel := by sorry
end BJNAdAuctions.Basic
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.