Risk functionals, version independence, Value-at-Risk and Average Value-at-Risk
DefinitionMultistageStochastic_RiskFunctionalThis file fixes the risk functionals of Pflug and Pichler's Chapter 3.
Risk functionals. A random variable is read as a loss, and a risk functional assigns it a number to be made small. Definition 3.2 lists the axioms:
- (M) Monotonicity: whenever almost surely;
- (C) Convexity: for ;
- (T) Translation equivariance: for real ;
- (H) Positive homogeneity: for .
The source reserves the unqualified term risk functional for a functional satisfying all four — "here we shall mention explicitly, if a risk functional is not positively homogeneous" — which is what is elsewhere called a coherent risk measure.
Version independence. Definition 3.12: is version independent, or law invariant, when whenever for every real .
Value-at-Risk and its average. The Value-at-Risk at level is the lower inverse of the distribution function,
and the (upper) Average Value-at-Risk averages it over the upper tail,
extended to by the limit, which the source identifies with the essential supremum .
Atomless spaces. A probability space is without atoms when every set of positive measure contains a subset of strictly smaller positive measure.
Formalization Note L^∞ is spelled out as "measurable and almost surely bounded", and every
axiom is quantified over that class rather than over a quotient space, so no L^p quotient
construction is needed and a risk functional is an ordinary function on random variables. The
essential supremum is taken in the source's own form , which avoids
committing to a particular lattice-theoretic essSup. Atomless is stated in the splitting
form, which is the standard meaning of "probability space without atoms" and is strictly stronger
than requiring singletons to be null on a general measurable space.
import Mathlib
open MeasureTheory
open scoped ENNReal
namespace MultistageStochastic
variable {Ω : Type*} [MeasurableSpace Ω]
/-- The random variables the risk functionals of Chapter 3 act on: `L^∞`, the measurable and
essentially bounded functions. Pflug and Pichler, *Multistage Stochastic Optimization*, §3.1,
p. 96. -/
def MemLinfty (P : Measure Ω) (Y : Ω → ℝ) : Prop :=
Measurable Y ∧ ∃ C : ℝ, ∀ᵐ ω ∂P, |Y ω| ≤ C
/-- Definition 3.2, the axioms for a **positively homogeneous risk functional** — what the book
calls a risk functional unless it says otherwise, and what is elsewhere called a coherent risk
measure: monotonicity (M), convexity (C), translation equivariance (T) and positive homogeneity
(H). Pflug and Pichler, §3.1, p. 96. -/
structure IsRiskFunctional (P : Measure Ω) (R : (Ω → ℝ) → ℝ) : Prop where
/-- (M) `R Y₁ ≤ R Y₂` whenever `Y₁ ≤ Y₂` almost surely. -/
mono : ∀ Y₁ Y₂, MemLinfty P Y₁ → MemLinfty P Y₂ → (∀ᵐ ω ∂P, Y₁ ω ≤ Y₂ ω) → R Y₁ ≤ R Y₂
/-- (C) `R` is convex along linear interpolation. -/
convex : ∀ Y₀ Y₁ (t : ℝ), MemLinfty P Y₀ → MemLinfty P Y₁ → 0 ≤ t → t ≤ 1 →
R (fun ω => (1 - t) * Y₀ ω + t * Y₁ ω) ≤ (1 - t) * R Y₀ + t * R Y₁
/-- (T) adding a constant to the loss adds it to the risk. -/
translation : ∀ Y (c : ℝ), MemLinfty P Y → R (fun ω => Y ω + c) = R Y + c
/-- (H) `R` is positively homogeneous. -/
homogeneous : ∀ Y (t : ℝ), MemLinfty P Y → 0 < t → R (fun ω => t * Y ω) = t * R Y
/-- Definition 3.12: a functional is **version independent** (law invariant) when its value
depends on the distribution of its argument only. Pflug and Pichler, §3.3, p. 103. -/
def VersionIndependent (P : Measure Ω) (R : (Ω → ℝ) → ℝ) : Prop :=
∀ Y Y', MemLinfty P Y → MemLinfty P Y' →
(∀ y : ℝ, P {ω | Y ω ≤ y} = P {ω | Y' ω ≤ y}) → R Y = R Y'
/-- The **Value-at-Risk** at level `α`, formula (3.2): `inf {y : P(Y ≤ y) ≥ α}`, the lower
inverse of the distribution function. Pflug and Pichler, §3.2, p. 97. -/
noncomputable def valueAtRisk (P : Measure Ω) (Y : Ω → ℝ) (α : ℝ) : ℝ :=
sInf {y : ℝ | ENNReal.ofReal α ≤ P {ω | Y ω ≤ y}}
/-- The essential supremum in the form the source uses, `ess sup Y = sup {y : G_Y(y) < 1}`.
Pflug and Pichler, §3.2, p. 97. -/
noncomputable def essSupBook (P : Measure Ω) (Y : Ω → ℝ) : ℝ :=
sSup {y : ℝ | P {ω | Y ω ≤ y} < 1}
/-- The (upper) **Average Value-at-Risk** at level `α`, Definition 3.3 (3.1):
`AV@R_α(Y) = (1-α)⁻¹ ∫_α^1 V@R_p(Y) dp` for `α ∈ [0,1)`, extended to `α = 1` by the limit, which
the source identifies with the essential supremum. Pflug and Pichler, §3.2, p. 97. -/
noncomputable def averageValueAtRisk (P : Measure Ω) (Y : Ω → ℝ) (α : ℝ) : ℝ :=
if α = 1 then essSupBook P Y
else (1 - α)⁻¹ * ∫ p in Set.Ioo α 1, valueAtRisk P Y p
/-- A probability space is **without atoms** when every set of positive measure splits into two
sets of smaller, positive measure. The hypothesis of Kusuoka's theorem, Pflug and Pichler,
§3.3.1, p. 103. -/
def Atomless (P : Measure Ω) : Prop :=
∀ s : Set Ω, MeasurableSet s → 0 < P s →
∃ t : Set Ω, t ⊆ s ∧ MeasurableSet t ∧ 0 < P t ∧ P t < P s
/-- The measures appearing in a Kusuoka representation: probability measures carried by the unit
interval. Pflug and Pichler, §3.3.1, p. 103. -/
def IsKusuokaMeasure (μ : Measure ℝ) : Prop :=
IsProbabilityMeasure μ ∧ μ (Set.Icc (0 : ℝ) 1)ᶜ = 0
end MultistageStochastic