Proposition 5.1 — the quadratic-deviation evaluation function is convex
ProvedKallMayerSLP.RiskMeasures.proposition_5_1convex-analysisp2o-gran-per-chapterp2o-plan-bookp2o-strengthenedp2o-v1risk-measuresstochastic-programming
Let be an -dimensional random vector on a probability space , with an -dimensional random vector and a random variable, and assume that the second moments of exist. Then the evaluation function
is convex on .
is the quadratic deviation between the random variables and ; its convexity makes the quadratic-deviation SLP models of §5.1 convex programs.
Formalization Note "Second moments exist" is square integrability of every component of and of .
Differences from the source:
- generalization: the second moments of (ηᵀ, ξ) exist and the covariance matrix of (ηᵀ, ξ) is positive definite → second moments exist (MemLp 2 for each component of η and for ξ); no non-degeneracy hypothesis; reason: the book's proof does not use non-degeneracy; dropping a hypothesis gives a statement that specializes to the book's
- encoding: Q(x) := √(𝔼[(ηᵀx − ξ)²]) → rhoQ P (fun ω => η ω ⬝ᵥ x − ξ ω), with rhoQ P ϑ = √(∫ ϑ² ∂P); reason: the same function
- encoding: x ∈ ℝⁿ, ηᵀx → x : Fin n → ℝ, η ω ⬝ᵥ x; reason: Lean convention; same instances
Preamble
import Mathlib import Definitions.Def_KallMayerSLP_RiskMeasures_rhoQ open MeasureTheory open scoped ENNReal
Formal statement
namespace KallMayerSLP.RiskMeasures
/-- **Proposition 5.1** (convexity of the quadratic-deviation evaluation function).
Kall & Mayer, *Stochastic Linear Programming* (2005), Chapter 2, §5.1, (5.1) and Prop. 5.1, p.167 (PDF 177), with the standing assumptions of §5
(p.166, PDF 176): `(η, ξ)` is an `(n+1)`-dimensional random vector on `(Ω, 𝓕, P)` whose second
moments exist. Then
`Q(x) := √(𝔼[(ηᵀx − ξ)²])` is a convex function of `x ∈ ℝⁿ`.
**Formalization Note.** `Q(x) = ρ_Q(ηᵀx − ξ)` with `ρ_Q(ϑ) = √𝔼[ϑ²]`. Vectors are `Fin n → ℝ`,
`ηᵀx` is `η ω ⬝ᵥ x`. "Second moments exist" is `MemLp · 2 P` for each component of `η` and for
`ξ`. The book also assumes the covariance matrix of `(ηᵀ, ξ)` positive definite; that hypothesis
is dropped (its proof does not use it), which makes the statement more general. -/
theorem proposition_5_1 {Ω : Type*} [MeasurableSpace Ω] (P : Measure Ω) [IsProbabilityMeasure P]
{n : ℕ} (η : Ω → Fin n → ℝ) (ξ : Ω → ℝ)
(hη : ∀ i, MemLp (fun ω => η ω i) 2 P) (hξ : MemLp ξ 2 P) :
ConvexOn ℝ Set.univ (fun x : Fin n → ℝ => rhoQ P (fun ω => η ω ⬝ᵥ x - ξ ω)) := by sorry
end KallMayerSLP.RiskMeasures
Source
Kall & Mayer, Stochastic Linear Programming (2005), Chapter 2, §5.1, (5.1) and Proposition 5.1, p.167 (PDF 177); standing assumptions of §5, p.166 (PDF 176)
Human review
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.