Observation 3 — per-machine tail Pr[Σ_{j∈J_i} X′_ij > b + 1 + α] ≤ ℓ_i^(−α)
ProvedStochLoadBal.Makespan.observation_3effective-sizep2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1stochastic-schedulingtail-bound
Let , let the sizes be measurable and integrable with the jobs independent, let be a feasible solution of the LP (5)–(9) with parameter , and let be an assignment with properties (13)–(14), with job sets and machine classes . Then for every machine and every ,
Machines of a high class have rapidly decaying truncated loads; Lemma 7 sums these bounds over the machines of class at least .
Formalization Note LP feasibility is assumed so that every class satisfies (Observation 1); with the junk class the right side would be . Only the independence of the jobs on machine is used, which follows from the independence of the jobs.
Preamble
import Mathlib import Definitions.Def_StochLoadBal_Makespan_Rounding
Formal statement
namespace StochLoadBal.Makespan
open MeasureTheory ProbabilityTheory
/-- Observation 3 (p. 13): for a feasible LP solution `y` and an assignment `σ` with
(13)–(14), every machine `i` of class `ℓ_i` satisfies
`Pr[Σ_{j ∈ J_i} X'_{ij} > b + 1 + α] ≤ ℓ_i^{-α}` for all `α ≥ 0`. -/
theorem observation_3 {m n : ℕ} (hm : 1 ≤ m) {Ω : Type} [MeasurableSpace Ω] (P : Measure Ω)
[IsProbabilityMeasure P] (X : Fin m → Fin n → Ω → ℝ)
(hmeas : ∀ i j, Measurable (X i j)) (hnn : ∀ i j ω, 0 ≤ X i j ω)
(hint : ∀ i j, Integrable (X i j) P) (hind : iIndepFun (fun j ω i => X i j ω) P)
(b : ℝ) (y : Fin m → Fin n → ℝ) (hy : LPFeasible P X b y)
(σ : Fin n → Fin m) (hσ : IsRoundingOutput P X b y σ) (i : Fin m) (α : ℝ) (hα : 0 ≤ α) :
(P {ω | b + 1 + α < ∑ j ∈ Finset.univ.filter (fun j => σ j = i), trunc (X i j ω)}).toReal
≤ (machineClass P X b y i : ℝ) ^ (-α) := by sorry
end StochLoadBal.Makespan
Source
Gupta, Kumar, Nagarajan & Shen, Stochastic Load Balancing on Unrelated Machines, arXiv:1904.07271v1, p. 13, Observation 3
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.