Observation 1, p. 10 β π£_d is jointly convex, and the resulting Jensen bound on E[π*]
ProvedLostSalesExpGap.ConstOrder.observation_1convexityjensenlost-salesp2o-batch-pfp2bp2o-gran-per-chapterp2o-plan-paperp2o-v1
Fix . For and let
- For each fixed , is jointly convex on .
- Let be i.i.d. nonnegative demands with finite mean, and let be a random vector, independent of the demands, such that (i) it is nonnegative with probability one and has finite mean; (ii) ; (iii) for all . Then
Conditions (i)β(iii) are those of the paper's Theorem 2, which constructs such a vector from a stationary optimal policy; this observation is how the optimal policy is compared with the constant-order policy ordering .
Formalization Note The law of is an arbitrary probability measure on satisfying (i)β(iii) as hypotheses; independence from the demands is the product measure . Theorem 2 itself is not part of this mission.
Preamble
import Mathlib import Definitions.Def_LostSalesExpGap_ConstOrder_ConstantOrder open MeasureTheory ProbabilityTheory open scoped ENNReal NNReal BigOperators
Formal statement
namespace LostSalesExpGap.ConstOrder
/-- Observation 1, p. 10. (a) For each `d β β^L`, `π£_d(Ο_1, β¦, Ο_L, I)` is jointly convex on
`β^{L+1}`. (b) For every law `Ξ½` of `(Ο*, π*)` with Theorem 2.(i)β(iii) (nonnegative, finite mean,
independent of the demands, `E[π*] = E[π£_{(D_1,β¦,D_L)}(Ο*, π*)]`, `E[Ο*_i] = E[Ο*_1] β€ E[D]`),
`E[π*] β₯ E[max_{j=0,β¦,L} (j E[Ο*_1] β Ξ£_{i=1}^j D_i + Ξ΄_{j,L} E[π*])]`. -/
theorem observation_1 (P : CappedBaseStock.DemandLaw) (L : β) (hL : 1 β€ L) :
(β d : Fin L β β, ConvexOn β Set.univ (fun v : (Fin L β β) Γ β => fObs L d v.1 v.2)) β§
(β Ξ½ : Measure ((Fin L β β) Γ β), IsProbabilityMeasure Ξ½ β
(βα΅ v βΞ½, (β i, 0 β€ v.1 i) β§ 0 β€ v.2) β
(β i, Integrable (fun v : (Fin L β β) Γ β => v.1 i) Ξ½) β
Integrable (fun v : (Fin L β β) Γ β => v.2) Ξ½ β
β« v, v.2 βΞ½ =
β« w, fObs L (firstDemands L w.2) w.1.1 w.1.2
β(Ξ½.prod (CappedBaseStock.demandPathLaw P)) β
(β i, β« v, v.1 i βΞ½ = β« v, v.1 β¨0, hLβ© βΞ½) β
β« v, v.1 β¨0, hLβ© βΞ½ β€ CappedBaseStock.mean P β
β« d, (Finset.range (L + 1)).sup' Finset.nonempty_range_add_one (fun j =>
(j : β) * β« v, v.1 β¨0, hLβ© βΞ½ - β i β Finset.Icc 1 j, (d (i - 1) : β) +
if j = L then β« v, v.2 βΞ½ else 0) βCappedBaseStock.demandPathLaw P
β€ β« v, v.2 βΞ½) := by sorry
end LostSalesExpGap.ConstOrder
Source
Xin, Goldberg, Optimality gap of constant-order policies decays exponentially in the lead time for lost sales models, arXiv:1409.1499v2, Observation 1, p. 10 (with Theorem 2.(i)β(iii), pp. 9β10)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.