Documentation

Linglib.Core.Probability.Exchangeable.Basic

Exchangeable laws #

A family of random variables indexed by ι with values in α is exchangeable when its law, a measure on ι → α, is unchanged by permuting the indices. The permutations act on ι → α by precomposition, as the domain action (Equiv.Perm ι)ᵈᵐᵃ ((DomMulAct.mk σ • x) i = x (σ i)), and Exchangeable μ is invariance under that action, SMulInvariantMeasure. Random partitions are exchangeable in the same sense, for the action of the permutations on partitions. On a countable space with measurable singletons, invariance is invariance of the point masses. Independent and identically distributed values are exchangeable, and under an exchangeable law of finitely many nonnegative quantities each takes in expectation the same share of their total.

Main definitions #

Main results #

Implementation notes #

For infinite ι the usual definition asks for invariance only under the permutations that move finitely many indices; Exchangeable asks for all of them, which de Finetti's theorem shows equivalent on standard Borel spaces but which is not proved here. [UPSTREAM] candidate: mathlib has no notion of exchangeability.

References #

theorem ProbabilityTheory.smulInvariantMeasure_of_measure_singleton {G : Type u_1} {X : Type u_2} [Group G] [MulAction G X] [MeasurableSpace X] [Countable X] [MeasurableSingletonClass X] {ν : MeasureTheory.Measure X} (h : ∀ (g : G) (x : X), ν {g • x} = ν {x}) :
MeasureTheory.SMulInvariantMeasure G X ν

On a countable space with measurable singletons, a measure is invariant under a group action as soon as moving a point never changes its mass.

@[reducible, inline]
abbrev ProbabilityTheory.Exchangeable {ι : Type u_1} {α : Type u_2} [MeasurableSpace α] (μ : MeasureTheory.Measure (ι → α)) :

The law μ of an ι-indexed family is exchangeable when every permutation of the indices, acting by precomposition, leaves it unchanged.

Equations
Instances For
    theorem ProbabilityTheory.exchangeable_of_measure_singleton {ι : Type u_1} {α : Type u_2} [MeasurableSpace α] {μ : MeasureTheory.Measure (ι → α)} [Countable (ι → α)] [MeasurableSingletonClass (ι → α)] (h : ∀ (x : ι → α) (σ : Equiv.Perm ι), μ {x ∘ ⇑σ} = μ {x}) :

    On a countable space with measurable singletons, a law is exchangeable as soon as permuting a point never changes its mass.

    theorem ProbabilityTheory.Exchangeable.map_comp_perm {ι : Type u_1} {α : Type u_2} [MeasurableSpace α] {μ : MeasureTheory.Measure (ι → α)} (hμ : Exchangeable μ) (σ : Equiv.Perm ι) :
    MeasureTheory.Measure.map (fun (x : ι → α) => x ∘ ⇑σ) μ = μ

    Permuting the indices preserves an exchangeable law.

    theorem ProbabilityTheory.exchangeable_of_map_comp_perm {ι : Type u_1} {α : Type u_2} [MeasurableSpace α] {μ : MeasureTheory.Measure (ι → α)} (h : ∀ (σ : Equiv.Perm ι), MeasureTheory.Measure.map (fun (x : ι → α) => x ∘ ⇑σ) μ = μ) :
    theorem ProbabilityTheory.exchangeable_iff_map_comp_perm {ι : Type u_1} {α : Type u_2} [MeasurableSpace α] {μ : MeasureTheory.Measure (ι → α)} :
    Exchangeable μ ↔ ∀ (σ : Equiv.Perm ι), MeasureTheory.Measure.map (fun (x : ι → α) => x ∘ ⇑σ) μ = μ
    theorem ProbabilityTheory.exchangeable_pi {ι : Type u_1} {α : Type u_2} [MeasurableSpace α] [Fintype ι] (ν : MeasureTheory.Measure α) [MeasureTheory.SigmaFinite ν] :
    Exchangeable (MeasureTheory.Measure.pi fun (x : ι) => ν)

    Independent and identically distributed values are exchangeable.

    theorem ProbabilityTheory.Exchangeable.integral_sum_div_sum {ι : Type u_1} {α : Type u_2} [MeasurableSpace α] {μ : MeasureTheory.Measure (ι → α)} [Fintype ι] {f : α → ℝ} [MeasureTheory.IsFiniteMeasure μ] (hμ : Exchangeable μ) (hf : Measurable f) (hf0 : ∀ (a : α), 0 ≤ f a) (A : Finset ι) :
    ∫ (x : ι → α), (∑ i ∈ A, f (x i)) / ∑ i : ι, f (x i) ∂μ = ↑A.card / ↑(Fintype.card ι) * μ.real {x : ι → α | ∑ i : ι, f (x i) ≠ 0}

    Under an exchangeable law of finitely many nonnegative quantities, each one's share of their total has the same expectation, so the members of A take together the fraction #A / card ι of the probability that the total is nonzero. Where the total is zero, the share is 0 / 0 = 0, so the event on which the total vanishes contributes nothing.