Proof of Theorem 1, pp. 36–37 — with x_i, y_i, w_ij the loss is the ratio (12) and (11) holds: (x, y, w) is feasible for (13)
ProvedProviderGuarantee.Bound.reduction_eq_11_12allocationlinear-fractional-programp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1provider-guarantees
Throughout, is a finite set of jobs with real values and ; there are providers ; and ; is a set of allocations satisfying Assumption 1, and satisfies Assumption 2. The optimal sets are and .
Let and , and define
Then:
- the inequalities (11) hold: for all ;
- for every , and for every ;
- , and for all ;
- the value loss equals the ratio (12):
Together, (1)–(4) say that is a feasible point of the fractional linear program (13) whose objective value is the loss. This turns the combinatorial question into a linear-fractional one.
Differences from the source:
- reading: Lemma 1 is stated 'for any fixed F and F_G' → F satisfies Assumption 1, F_G ⊆ F satisfies Assumption 2, and γ ∈ [γmin, γmax]^n with 0 < γmin ≤ γmax; reason: these are the standing assumptions of §2 (pp. 7–12) under which the appendix works; the proof on pp. 35–36 uses Assumption 1-(ii), (iii), (iv), Assumption 2 and γ_j > 0, and the lemma is false without them
- reading: the page uses (11) and the decompositions of v(A_i), v(B_j) to call (13), which imposes x, y, w ≥ 0, an upper bound on L_γ → the non-negativity x_i, y_i, w_ij ≥ 0 is stated as part of the conclusion; reason: the page's step needs (x, y, w) feasible for (13); non-negativity holds because optimal allocations hold no job of negative value (Assumption 1-(iii), Assumption 2, γ > 0) and is a conclusion, not an added hypothesis on v
- encoding: (12) equals L_γ(F, F_G), both sides being 0/0 when max_F Σγ_i v(A_i) = 0 → equality of Lean's quotients, both 0 in that case; reason: Lean's division by zero; the denominator of (12) is Σγ_i v(A_i) = max_F, so the two sides are undefined on the page exactly together
- reading: a set of jobs D (never said to be finite; v(S) is a finite sum and maxima over F are taken) → D a finite type; reason: the page writes v(S) = Σ_{d∈S} v(d) and max over F, which presuppose finitely many jobs; the paper calls D a discrete set of jobs
- encoding: providers N = {1, …, n} → indices i : Fin n (0, …, n−1); reason: Lean convention; a relabelling, same instances
Preamble
import Mathlib import Definitions.Def_ProviderGuarantee_Bound_Model
Formal statement
namespace ProviderGuarantee.Bound
theorem reduction_eq_11_12 {D : Type*} [Fintype D] [DecidableEq D] {n : ℕ} (v : D → ℝ)
(γmin γmax : ℝ) (hγmin : 0 < γmin) (hγ : γmin ≤ γmax)
(γ : Fin n → ℝ) (hγmem : ∀ i, γ i ∈ Set.Icc γmin γmax)
(F FG : Set (Fin n → Finset D)) (h1 : Assumption1 F) (h2 : Assumption2 v F FG)
(A B : Fin n → Finset D) (hA : IsOptimal γ v F A) (hB : IsOptimal γ v FG B) :
let x : Fin n → ℝ := fun i => val v (A i \ Finset.univ.biUnion B)
let y : Fin n → ℝ := fun i => val v (B i \ Finset.univ.biUnion A)
let w : Fin n → Fin n → ℝ := fun i j => val v (A i ∩ B j)
(∀ i j, x i ≤ ∑ k, w k j - w i j + y j) ∧
(∀ i, val v (A i) = x i + ∑ j, w i j) ∧
(∀ j, val v (B j) = y j + ∑ i, w i j) ∧
(∀ i, 0 ≤ x i) ∧ (∀ i, 0 ≤ y i) ∧ (∀ i j, 0 ≤ w i j) ∧
valueLoss γ v F FG =
(∑ i, γ i * x i - ∑ i, γ i * y i + ∑ i, ∑ j, w i j * (γ i - γ j)) /
(∑ i, γ i * x i + ∑ i, ∑ j, w i j * γ i) := by sorry
end ProviderGuarantee.Bound
Source
Gur, Iancu & Warnes, Value Loss in Allocation Systems with Provider Guarantees, SSRN 3351509 (version of March 24, 2020), pp. 36–37, proof of Theorem 1, (11) and (12)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.