(* Title: HOL/Probability/Conditional_Probability.thy
Author: Johannes Hölzl, TU München
*)
header {*Conditional probability*}
theory Conditional_Probability
imports Probability_Measure Radon_Nikodym
begin
section "Conditional Expectation and Probability"
definition (in prob_space)
"conditional_expectation N X = (SOME Y. Y\<in>borel_measurable N \<and> (\<forall>x. 0 \<le> Y x)
\<and> (\<forall>C\<in>sets N. (\<integral>\<^isup>+x. Y x * indicator C x\<partial>M) = (\<integral>\<^isup>+x. X x * indicator C x\<partial>M)))"
lemma (in prob_space) conditional_expectation_exists:
fixes X :: "'a \<Rightarrow> ereal" and N :: "('a, 'b) measure_space_scheme"
assumes borel: "X \<in> borel_measurable M" "AE x. 0 \<le> X x"
and N: "sigma_algebra N" "sets N \<subseteq> sets M" "space N = space M" "\<And>A. measure N A = \<mu> A"
shows "\<exists>Y\<in>borel_measurable N. (\<forall>x. 0 \<le> Y x) \<and> (\<forall>C\<in>sets N.
(\<integral>\<^isup>+x. Y x * indicator C x \<partial>M) = (\<integral>\<^isup>+x. X x * indicator C x \<partial>M))"
proof -
note N(4)[simp]
interpret P: prob_space N
using prob_space_subalgebra[OF N] .
let "?f A" = "\<lambda>x. X x * indicator A x"
let "?Q A" = "integral\<^isup>P M (?f A)"
from measure_space_density[OF borel]
have Q: "measure_space (N\<lparr> measure := ?Q \<rparr>)"
apply (rule measure_space.measure_space_subalgebra[of "M\<lparr> measure := ?Q \<rparr>"])
using N by (auto intro!: P.sigma_algebra_cong)
then interpret Q: measure_space "N\<lparr> measure := ?Q \<rparr>" .
have "P.absolutely_continuous ?Q"
unfolding P.absolutely_continuous_def
proof safe
fix A assume "A \<in> sets N" "P.\<mu> A = 0"
then have f_borel: "?f A \<in> borel_measurable M" "AE x. x \<notin> A"
using borel N by (auto intro!: borel_measurable_indicator AE_not_in)
then show "?Q A = 0"
by (auto simp add: positive_integral_0_iff_AE)
qed
from P.Radon_Nikodym[OF Q this]
obtain Y where Y: "Y \<in> borel_measurable N" "\<And>x. 0 \<le> Y x"
"\<And>A. A \<in> sets N \<Longrightarrow> ?Q A =(\<integral>\<^isup>+x. Y x * indicator A x \<partial>N)"
by blast
with N(2) show ?thesis
by (auto intro!: bexI[OF _ Y(1)] simp: positive_integral_subalgebra[OF _ _ N(2,3,4,1)])
qed
lemma (in prob_space)
fixes X :: "'a \<Rightarrow> ereal" and N :: "('a, 'b) measure_space_scheme"
assumes borel: "X \<in> borel_measurable M" "AE x. 0 \<le> X x"
and N: "sigma_algebra N" "sets N \<subseteq> sets M" "space N = space M" "\<And>A. measure N A = \<mu> A"
shows borel_measurable_conditional_expectation:
"conditional_expectation N X \<in> borel_measurable N"
and conditional_expectation: "\<And>C. C \<in> sets N \<Longrightarrow>
(\<integral>\<^isup>+x. conditional_expectation N X x * indicator C x \<partial>M) =
(\<integral>\<^isup>+x. X x * indicator C x \<partial>M)"
(is "\<And>C. C \<in> sets N \<Longrightarrow> ?eq C")
proof -
note CE = conditional_expectation_exists[OF assms, unfolded Bex_def]
then show "conditional_expectation N X \<in> borel_measurable N"
unfolding conditional_expectation_def by (rule someI2_ex) blast
from CE show "\<And>C. C \<in> sets N \<Longrightarrow> ?eq C"
unfolding conditional_expectation_def by (rule someI2_ex) blast
qed
lemma (in sigma_algebra) factorize_measurable_function_pos:
fixes Z :: "'a \<Rightarrow> ereal" and Y :: "'a \<Rightarrow> 'c"
assumes "sigma_algebra M'" and "Y \<in> measurable M M'" "Z \<in> borel_measurable M"
assumes Z: "Z \<in> borel_measurable (sigma_algebra.vimage_algebra M' (space M) Y)"
shows "\<exists>g\<in>borel_measurable M'. \<forall>x\<in>space M. max 0 (Z x) = g (Y x)"
proof -
interpret M': sigma_algebra M' by fact
have Y: "Y \<in> space M \<rightarrow> space M'" using assms unfolding measurable_def by auto
from M'.sigma_algebra_vimage[OF this]
interpret va: sigma_algebra "M'.vimage_algebra (space M) Y" .
from va.borel_measurable_implies_simple_function_sequence'[OF Z] guess f . note f = this
have "\<forall>i. \<exists>g. simple_function M' g \<and> (\<forall>x\<in>space M. f i x = g (Y x))"
proof
fix i
from f(1)[of i] have "finite (f i`space M)" and B_ex:
"\<forall>z\<in>(f i)`space M. \<exists>B. B \<in> sets M' \<and> (f i) -` {z} \<inter> space M = Y -` B \<inter> space M"
unfolding simple_function_def by auto
from B_ex[THEN bchoice] guess B .. note B = this
let ?g = "\<lambda>x. \<Sum>z\<in>f i`space M. z * indicator (B z) x"
show "\<exists>g. simple_function M' g \<and> (\<forall>x\<in>space M. f i x = g (Y x))"
proof (intro exI[of _ ?g] conjI ballI)
show "simple_function M' ?g" using B by auto
fix x assume "x \<in> space M"
then have "\<And>z. z \<in> f i`space M \<Longrightarrow> indicator (B z) (Y x) = (indicator (f i -` {z} \<inter> space M) x::ereal)"
unfolding indicator_def using B by auto
then show "f i x = ?g (Y x)" using `x \<in> space M` f(1)[of i]
by (subst va.simple_function_indicator_representation) auto
qed
qed
from choice[OF this] guess g .. note g = this
show ?thesis
proof (intro ballI bexI)
show "(\<lambda>x. SUP i. g i x) \<in> borel_measurable M'"
using g by (auto intro: M'.borel_measurable_simple_function)
fix x assume "x \<in> space M"
have "max 0 (Z x) = (SUP i. f i x)" using f by simp
also have "\<dots> = (SUP i. g i (Y x))"
using g `x \<in> space M` by simp
finally show "max 0 (Z x) = (SUP i. g i (Y x))" .
qed
qed
lemma (in sigma_algebra) factorize_measurable_function:
fixes Z :: "'a \<Rightarrow> ereal" and Y :: "'a \<Rightarrow> 'c"
assumes "sigma_algebra M'" and "Y \<in> measurable M M'" "Z \<in> borel_measurable M"
shows "Z \<in> borel_measurable (sigma_algebra.vimage_algebra M' (space M) Y)
\<longleftrightarrow> (\<exists>g\<in>borel_measurable M'. \<forall>x\<in>space M. Z x = g (Y x))"
proof safe
interpret M': sigma_algebra M' by fact
have Y: "Y \<in> space M \<rightarrow> space M'" using assms unfolding measurable_def by auto
from M'.sigma_algebra_vimage[OF this]
interpret va: sigma_algebra "M'.vimage_algebra (space M) Y" .
{ fix g :: "'c \<Rightarrow> ereal" assume "g \<in> borel_measurable M'"
with M'.measurable_vimage_algebra[OF Y]
have "g \<circ> Y \<in> borel_measurable (M'.vimage_algebra (space M) Y)"
by (rule measurable_comp)
moreover assume "\<forall>x\<in>space M. Z x = g (Y x)"
then have "Z \<in> borel_measurable (M'.vimage_algebra (space M) Y) \<longleftrightarrow>
g \<circ> Y \<in> borel_measurable (M'.vimage_algebra (space M) Y)"
by (auto intro!: measurable_cong)
ultimately show "Z \<in> borel_measurable (M'.vimage_algebra (space M) Y)"
by simp }
assume Z: "Z \<in> borel_measurable (M'.vimage_algebra (space M) Y)"
with assms have "(\<lambda>x. - Z x) \<in> borel_measurable M"
"(\<lambda>x. - Z x) \<in> borel_measurable (M'.vimage_algebra (space M) Y)"
by auto
from factorize_measurable_function_pos[OF assms(1,2) this] guess n .. note n = this
from factorize_measurable_function_pos[OF assms Z] guess p .. note p = this
let "?g x" = "p x - n x"
show "\<exists>g\<in>borel_measurable M'. \<forall>x\<in>space M. Z x = g (Y x)"
proof (intro bexI ballI)
show "?g \<in> borel_measurable M'" using p n by auto
fix x assume "x \<in> space M"
then have "p (Y x) = max 0 (Z x)" "n (Y x) = max 0 (- Z x)"
using p n by auto
then show "Z x = ?g (Y x)"
by (auto split: split_max)
qed
qed
end