Distortion risk functionals: distortion densities, R_σ, the dual constraint Z ≼ σ, uniform variables and conjugates
DefinitionMultistageStochastic_DistortionThis file adds to the risk-functional layer of mission II (risk functionals, Value-at-Risk, Average Value-at-Risk, membership, atomless spaces, Kusuoka measures) the notions Sections 3.2 to 3.4 need for distortion risk functionals.
Distortion functions (Definition 3.6). A distortion function, or distortion density, is a nonnegative, nondecreasing function with . The distortion risk functional with density is
with the Value-at-Risk (lower quantile) of Definition 3.3. The Average Value-at-Risk is the case , (3.10).
The dual constraint (3.16). A random variable is dominated by , written , when , and
Uniform variables (footnote 2, p. 99). is uniformly distributed on when for all .
Conjugates (footnote 6, p. 111). For , , valued in . A measurable is admissible for when is finite for almost every level, is integrable on , and , which is the constraint of (3.24).
Formalization Note is carried as a function on whose values outside play no role; its integrability on is part of the definition so that the normalisation is a genuine integral rather than the value that Lean assigns to a non-integrable one. Both integrals in (3.7) and in the normalisation run over the open interval, which removes the level , where the source's formula for the Value-at-Risk is not meaningful, without changing any value. The source writes the constraint of (3.16) for ; at the right-hand side is and is meant as the limit , and since both sides are the limits of the case, the constraint at is implied and is not restated. The conjugate takes values in the extended reals so that no supremum is replaced by a junk value; admissibility then requires it to be finite where it is integrated.
import Mathlib
import Definitions.Def_MultistageStochastic_RiskFunctional
open MeasureTheory
open scoped ENNReal
namespace MultistageStochastic
variable {Ω : Type*} [MeasurableSpace Ω]
/-- Definition 3.6: a **distortion function** (distortion density) is a nonnegative, nondecreasing
function `σ : [0, 1) → [0, ∞)` with `∫₀¹ σ(u) du = 1`. It is carried here as a function on `ℝ`
whose values outside `[0, 1)` are irrelevant; integrability on `(0, 1)` is stated so that the
normalisation is a genuine integral. Pflug and Pichler, *Multistage Stochastic Optimization*,
§3.2, p. 99. -/
def IsDistortionFunction (σ : ℝ → ℝ) : Prop :=
(∀ u ∈ Set.Ico (0 : ℝ) 1, 0 ≤ σ u) ∧ MonotoneOn σ (Set.Ico 0 1) ∧
IntegrableOn σ (Set.Ioo 0 1) ∧ ∫ u in Set.Ioo (0 : ℝ) 1, σ u = 1
/-- The **distortion risk functional** (3.7), `R_σ(Y) := ∫₀¹ σ(u) G_Y⁻¹(u) du = ∫₀¹ σ(u) V@R_u(Y) du`,
with the Value-at-Risk of Definition 3.3. The integral is over the open interval, which excludes
the level `0`, where the Value-at-Risk is not given by the source's formula. Pflug and Pichler,
§3.2, p. 99. -/
noncomputable def distortionFunctional (P : Measure Ω) (σ : ℝ → ℝ) (Y : Ω → ℝ) : ℝ :=
∫ u in Set.Ioo (0 : ℝ) 1, σ u * valueAtRisk P Y u
/-- The binary relation `Z ≼ σ` of (3.16), the constraint set of the dual representation (3.15):
`Z` is integrable, `E(Z) = 1`, and `AV@R_α(Z) ≤ (1-α)⁻¹ ∫_α¹ σ(u) du` for every level `α ∈ [0, 1)`.
The source writes the last condition for `α ∈ [0, 1]`; at `α = 1` both sides are limits of the
`α < 1` case, so the constraint there is implied. Pflug and Pichler, §3.3.2, p. 106. -/
def DominatedByDistortion (P : Measure Ω) (σ : ℝ → ℝ) (Z : Ω → ℝ) : Prop :=
Integrable Z P ∧ ∫ ω, Z ω ∂P = 1 ∧
∀ α : ℝ, 0 ≤ α → α < 1 →
averageValueAtRisk P Z α ≤ (1 - α)⁻¹ * ∫ u in Set.Ioo α 1, σ u
/-- A random variable `U : Ω → ℝ` is **uniformly distributed** on `[0, 1]` when
`P(U ≤ u) = u` for every `u ∈ [0, 1]`. Pflug and Pichler, §3.2, p. 99, footnote 2. -/
def IsUniform (P : Measure Ω) (U : Ω → ℝ) : Prop :=
Measurable U ∧ ∀ u ∈ Set.Icc (0 : ℝ) 1, P {ω | U ω ≤ u} = ENNReal.ofReal u
/-- The **conjugate** (Legendre transform) `h*(s) = sup_y (s·y − h(y))` of a function `h : ℝ → ℝ`,
valued in the extended reals since the supremum may be `+∞`. Pflug and Pichler, §3.4, p. 111,
footnote 6. -/
noncomputable def conjugate (h : ℝ → ℝ) (s : ℝ) : EReal :=
⨆ y : ℝ, ((s * y - h y : ℝ) : EReal)
/-- The constraint set of Theorem 3.22 (3.24): measurable `h : ℝ → ℝ` whose conjugate is finite
along `σ` for almost every level, with `u ↦ h*(σ(u))` integrable on `(0, 1)` and
`∫₀¹ h*(σ(u)) du ≤ 0`. Pflug and Pichler, §3.4, p. 111. -/
def IsAdmissibleConjugate (σ : ℝ → ℝ) (h : ℝ → ℝ) : Prop :=
Measurable h ∧
(∀ᵐ u ∂(volume.restrict (Set.Ioo (0 : ℝ) 1)), conjugate h (σ u) ≠ ⊤) ∧
IntegrableOn (fun u => (conjugate h (σ u)).toReal) (Set.Ioo 0 1) ∧
∫ u in Set.Ioo (0 : ℝ) 1, (conjugate h (σ u)).toReal ≤ 0
end MultistageStochastic