Regular conditional probability distribution #
We define the regular conditional probability distribution of Y : α → Ω
given X : α → β
, where
Ω
is a standard Borel space. This is a kernel β Ω
such that for almost all a
, condDistrib
evaluated at X a
and a measurable set s
is equal to the conditional expectation
μ⟦Y ⁻¹' s | mβ.comap X⟧
evaluated at a
.
μ⟦Y ⁻¹' s | mβ.comap X⟧
maps a measurable set s
to a function α → ℝ≥0∞
, and for all s
that
map is unique up to a μ
-null set. For all a
, the map from sets to ℝ≥0∞
that we obtain that way
verifies some of the properties of a measure, but in general the fact that the μ
-null set depends
on s
can prevent us from finding versions of the conditional expectation that combine into a true
measure. The standard Borel space assumption on Ω
allows us to do so.
The case Y = X = id
is developed in more detail in Probability/Kernel/Condexp.lean
: here X
is
understood as a map from Ω
with a sub-σ-algebra m
to Ω
with its default σ-algebra and the
conditional distribution defines a kernel associated with the conditional expectation with respect
to m
.
Main definitions #
condDistrib Y X μ
: regular conditional probability distribution ofY : α → Ω
givenX : α → β
, whereΩ
is a standard Borel space.
Main statements #
condDistrib_ae_eq_condexp
: for almost alla
,condDistrib
evaluated atX a
and a measurable sets
is equal to the conditional expectationμ⟦Y ⁻¹' s | mβ.comap X⟧ a
.condexp_prod_ae_eq_integral_condDistrib
: the conditional expectationμ[(fun a => f (X a, Y a)) | X; mβ]
is almost everywhere equal to the integral∫ y, f (X a, y) ∂(condDistrib Y X μ (X a))
.
Regular conditional probability distribution: kernel associated with the conditional
expectation of Y
given X
.
For almost all a
, condDistrib Y X μ
evaluated at X a
and a measurable set s
is equal to
the conditional expectation μ⟦Y ⁻¹' s | mβ.comap X⟧ a
. It also satisfies the equality
μ[(fun a => f (X a, Y a)) | mβ.comap X] =ᵐ[μ] fun a => ∫ y, f (X a, y) ∂(condDistrib Y X μ (X a))
for all integrable functions f
.
Instances For
Equations
- (_ : ProbabilityTheory.IsMarkovKernel (ProbabilityTheory.condDistrib Y X μ)) = (_ : ProbabilityTheory.IsMarkovKernel (ProbabilityTheory.condDistrib Y X μ))
If the singleton {x}
has non-zero mass for μ.map X
, then for all s : Set Ω
,
condDistrib Y X μ x s = (μ.map X {x})⁻¹ * μ.map (fun a => (X a, Y a)) ({x} ×ˢ s)
.
condDistrib
is a.e. uniquely defined as the kernel satisfying the defining property of
condKernel
.
For almost every a : α
, the condDistrib Y X μ
kernel applied to X a
and a measurable set
s
is equal to the conditional expectation of the indicator of Y ⁻¹' s
.
The conditional expectation of a function f
of the product (X, Y)
is almost everywhere equal
to the integral of y ↦ f(X, y)
against the condDistrib
kernel.
The conditional expectation of a function f
of the product (X, Y)
is almost everywhere equal
to the integral of y ↦ f(X, y)
against the condDistrib
kernel.
The conditional expectation of a function f
of the product (X, Y)
is almost everywhere equal
to the integral of y ↦ f(X, y)
against the condDistrib
kernel.
The conditional expectation of Y
given X
is almost everywhere equal to the integral
∫ y, y ∂(condDistrib Y X μ (X a))
.