Proposition 7.3 — expectation, the worst case and CVaR are coherent and convex risk measures
ProvedKallMayerSLP.RiskMeasures.proposition_7_3Let be a probability space. Each of the following risk measures is positive homogeneous, subadditive, monotonically increasing and translation invariant; consequently it is a convex risk measure in the sense of the axioms (7.3) and, being positively homogeneous, a coherent risk measure in the sense of the axioms (7.2):
- (A) the expectation on ;
- (B) the worst case on , where is the support of ;
- (C) for every , the conditional value-at-risk on ,
By Proposition 7.2, each of them therefore turns a linear-affine random variable into a convex evaluation function.
Formalization Note Coherence bundles (A1)–(A4) and so contains the four listed properties. The monotonicity hypothesis is almost sure. The maximum in (B) and the minimum in (C) are a supremum and an infimum (see the definitions of and ); for the random variables in question both are attained.
Differences from the source:
- generalization: ϑ₁ ≤ ϑ₂ in (A3), (F2): pointwise, "ϑ₁(ω) ≤ ϑ₂(ω) for all ω" in the proof of (B), p.185 → ϑ₁ ≤ ϑ₂ almost surely; reason: a weaker hypothesis than the pointwise order of the book's proof ("ϑ₁(ω) ≤ ϑ₂(ω) for all ω"), so the monotonicity clause implies the book's
- encoding: 𝓛^p₁ (p = 1, 2, ∞) of (1.6) → Lspace ℝ p P: functions Ω → ℝ with MemLp · p P, a submodule of Ω → ℝ; reason: random variables with finite p-th moment (resp. bounded support); same instances
- encoding: max over the support Θ in (B); min over z in (C) → sSup of the topological support of the law P.map ϑ; real infimum ⨅ z; reason: both are attained for ϑ ∈ 𝓛^∞₁, resp. ϑ ∈ 𝓛¹₁ and 0 < α < 1 (§4.3), so the values agree
import Mathlib import Definitions.Def_KallMayerSLP_RiskMeasures_Lspace import Definitions.Def_KallMayerSLP_RiskMeasures_IsCoherentRiskMeasure import Definitions.Def_KallMayerSLP_RiskMeasures_IsConvexRiskMeasure import Definitions.Def_KallMayerSLP_RiskMeasures_rhoE import Definitions.Def_KallMayerSLP_RiskMeasures_rhoFat import Definitions.Def_KallMayerSLP_RiskMeasures_cvar open MeasureTheory open scoped ENNReal
namespace KallMayerSLP.RiskMeasures
/-- **Proposition 7.3** (expectation, worst case and CVaR are coherent and convex risk measures).
Kall & Mayer, *Stochastic Linear Programming* (2005), Chapter 2, §7.2, Prop. 7.3, p.184 (PDF 194). The following risk measures are positive homogeneous and
subadditive, monotonically increasing and translation invariant; consequently they are convex
risk measures (axioms (7.3)) and coherent risk measures (axioms (7.2)):
(A) `ρ_E(ϑ) := 𝔼[ϑ]`, `ϑ ∈ 𝓛¹₁`;
(B) `ρ_fat(ϑ) := max_{ϑ̂ ∈ Θ} ϑ̂`, `ϑ ∈ 𝓛^∞₁`, `Θ` the support of `ϑ`;
(C) `ρ^α_CVaR(ϑ) := min_z [z + (1/(1 − α)) 𝔼[(ϑ − z)⁺]]`, `ϑ ∈ 𝓛¹₁`, `0 < α < 1`.
**Formalization Note.** `(Ω, 𝓕, P)` is an arbitrary probability space; `𝓛^p₁` is `Lspace ℝ p P`
(random variables `Ω → ℝ` with `MemLp · p P`). Coherence bundles (A1)–(A4), so it contains the
four properties listed first. The hypothesis of monotonicity is `ϑ₁ ≤ ϑ₂` almost surely, weaker
than the pointwise order used in the book's proof, so the statement is stronger. The `max` in (B) and the
`min` in (C) are `sSup` and `⨅` (see the definitions). -/
theorem proposition_7_3 {Ω : Type*} [MeasurableSpace Ω] (P : Measure Ω) [IsProbabilityMeasure P] :
(IsCoherentRiskMeasure P (Lspace ℝ 1 P) (rhoE P) ∧
IsConvexRiskMeasure P (Lspace ℝ 1 P) (rhoE P)) ∧
(IsCoherentRiskMeasure P (Lspace ℝ ⊤ P) (rhoFat P) ∧
IsConvexRiskMeasure P (Lspace ℝ ⊤ P) (rhoFat P)) ∧
(∀ α : ℝ, 0 < α → α < 1 →
IsCoherentRiskMeasure P (Lspace ℝ 1 P) (cvar P α) ∧
IsConvexRiskMeasure P (Lspace ℝ 1 P) (cvar P α)) := by sorry
end KallMayerSLP.RiskMeasures
Confirmed by the mission captain (proposal self-audit).
Confirmed by the moderator at approval.