Proof of Theorem 5, p. 40 ā every š«ā² distribution is the zĢ-marginal of its lift by Å© = g(zĢ) ā E[g(zĢ)] + f
ProvedDRConvexOpt.Lifting.momentSet_subset_marginalambiguity-setdistributionally-robust-optimizationp2o-batch-pfp2ap2o-gran-per-chapterp2o-plan-paperp2o-v1pushforward-measure
Let be measurable, , a proper cone, and , , measurable sets with bounds . Let be a distribution in the moment set , that is, is -integrable, and for all . Let be the joint law of where and
i.e. the pushforward of under . Then belongs to the lifted set and .
This is the inclusion of the Lifting Theorem, stated together with the explicit lift used in the paper.
Formalization Note The lift is written as ν.map (fun z => (z, g z - ā« z', g z' āν + f)). Measurability of makes this map measurable, so the pushforward is a probability measure; measurability of the makes their probabilities transfer. The cone hypothesis is carried as on the page.
Preamble
import Mathlib import Definitions.Def_DRConvexOpt_Lifting_Setting
Formal statement
namespace DRConvexOpt.Lifting
open MeasureTheory
/-- Proof of Theorem 5, p. 40: every ν ā š«ā² is the zĢ-marginal of the pushforward of ν under
z ⦠(z, g(z) ā E_ν[g(zĢ)] + f), and that pushforward lies in the lifted set š«. -/
theorem momentSet_subset_marginal {nP nM nI : ā} (g : (Fin nP ā ā) ā (Fin nM ā ā))
(hg : Measurable g) (f : Fin nM ā ā) (K : Set (Fin nM ā ā)) (hK : DRConvexOpt.Reform.IsProperCone K)
(C : Fin nI ā Set (Fin nP ā ā)) (hC : ā i, MeasurableSet (C i)) (plo phi : Fin nI ā ā)
(ν : Measure (Fin nP ā ā)) (hν : ν ā momentSet g f K C plo phi) :
ν.map (fun z => (z, g z - ā« z', g z' āν + f)) ā liftedSet g f K C plo phi ā§
(ν.map (fun z => (z, g z - ā« z', g z' āν + f))).map Prod.fst = ν := by sorry
end DRConvexOpt.Lifting
Source
Wiesemann, Kuhn & Sim, Distributionally Robust Convex Optimization, Optimization Online preprint 3757 (version of September 22, 2013), p. 40, proof of Theorem 5, first inclusion š«ā² ā Ī _zĢš«
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.