2
$\begingroup$

In Williams' "Probability with Martingales" he defines a regular conditional probability given $\mathcal{G} \subset \mathcal{F}$ as the map $$ \mathbf{P}(\cdot, \cdot) : \Omega \times \mathcal{F} \to [0,1) $$ such that \begin{align*} &1.\quad \forall F \in \mathcal{F}, \omega \mapsto \mathbf{P}(\omega,F) \text{ is a version of } \mathbf{P}(F \mid \mathcal{G}),\\ &2.\quad\text{ for almost every } \omega \in \Omega, F \mapsto \mathbf{P}(\omega,F) \text{ is a probability measure on } (\Omega,\mathcal{F}). \end{align*} Property 1 tells us that for any $F \in \mathcal{F}$ and any $G \in \mathcal{G}$, \begin{align*} \int_G \mathbf{P}(\omega, F) \, \mathrm{d}P(\omega) & = \int_G \mathbf{P}(F \mid \mathcal{G})(\omega) \, \mathrm{d}P(\omega) \\ & = \int_G E(\mathbb{I}_F \mid \mathcal{G})(\omega) \, \mathrm{d}P(\omega) \\ & = \int_G \mathbb{I}_F(\omega) \, \mathrm{d}P(\omega) \\ & = P(F \cap G). \end{align*}

Fair enough, but I'm trying to get a handle on the regular conditional probability given a random variable $X: \Omega \to \mathbb{R}$. It seems like we should define it as the map $$ \mathbf{P}(\cdot, \cdot) : \mathbb{R} \times \mathcal{F} \to [0,1) $$ such that \begin{align*} &1.\quad \forall F \in \mathcal{F}, x \mapsto \mathbf{P}(x,F) \text{ is a version of } \mathbf{P}(F \mid \sigma(X)),\\ &2.\quad\text{ for almost every } x \in \mathbb{R}, F \mapsto \mathbf{P}(x,F) \text{ is a probability measure on } (\Omega, \mathcal{F}). \end{align*} However, I'm not sure how to get the "defining equation" as above, but here's my shot. Let $F \in \mathcal{F}$ and $C \in \sigma(X) \subset \mathcal{F}$. Note $C = X^{-1}(B)$ for some $B \in \mathcal{B}$. Letting $\Lambda_X$ be the distribution of $X$, by property 1 we get \begin{align*} \int_B \mathbf{P}(x,F) \, \mathrm{d}\Lambda_X(x) & = \int_C \mathbf{P}(X(\omega), F) \, \mathrm{d} P(\omega) \\ & = \int_C \mathbf{P}(F \mid \sigma(X))(\omega) \, \mathrm{d} P(\omega) \\ & = \int_C E(\mathbb{I}_F \mid \sigma(X))(\omega) \, \mathrm{d} P(\omega) \\ & = \int_C \mathbb{I}_F(\omega) \, \mathrm{d} P(\omega) \\ & = P(F \cap \{X \in B\}). \end{align*} Does this seem correct?

Update Based on Conrado Costa's answer, I think the proper definition for the regular conditional probability given $X$ should be as follows. Below, let $X: (\Omega, \mathcal{F}) \to (E, \mathcal{E})$.

Assume the regular conditional probability given $\sigma(X)$ exists. That is, there exists a function $\mathbf{P}^2 : \Omega \times \mathcal{F} \to [0,1]$ such that

  1. $\forall F \in \mathcal{F}$, $\omega \mapsto \mathbf{P}^2(\omega,F)$ is a version of $\mathbf{P}(F \mid \sigma({X}))$,
  2. $\forall \omega \in \Omega$, $F \mapsto \mathbf{P}^2(\omega, F)$ is a probability measure on $(\Omega, \mathcal{F})$.

In particular, property 1 tells us that $\forall F \in \mathcal{F}$, $\omega \mapsto \mathbf{P}^2(\omega,F)$ is $(\sigma(X), \mathcal{B}[0,1])$-measurable. Then, by Theorem 18 listed below in Conrado Costa's answer, there exists a function $\mathbf{P}^1: E \times \mathcal{F} \to [0,1]$ that is $(\mathcal{E}, \mathcal{B}[0,1])$-measurable such that $\forall F \in \mathcal{F}$, $\omega \mapsto \mathbf{P}^2(\omega,F) = \mathbf{P}^1(X(\omega),F)$.

So, for any $F \in \mathcal{F}$ and any $C \in \sigma(X)$, by the change of variables theorem we get \begin{align*} \int_{X(C)} \mathbf{P}^1(x,F) \, d\Lambda_X(x) & = \int_C \mathbf{P}^2(\omega,F) \, dP(\omega) \\ & = \int_C \mathbf{P}(F \mid \sigma(X))(\omega) \, dP(\omega) \\ & = \int_C E(\mathbb{I}_F \mid \sigma(X))(\omega) \, dP(\omega) \\ & = \int_C \mathbb{I}_F(\omega) \, dP(\omega) \\ & = P(F \cap \{X \in B\}). \end{align*}

$\endgroup$
2
$\begingroup$

This is a subtle question.

you wrote: \begin{align*} &1.\quad \forall F \in \mathcal{F}, x \mapsto \mathbf{P}(x,F) \text{ is a version of } \mathbf{P}(F \mid \sigma(X)),\\ &2.\quad\text{ for almost every } x \in \mathbb{R}, F \mapsto \mathbf{P}(x,F) \text{ is a probability measure on } (\Omega, \mathcal{F}). \end{align*}

for 1 It is important to note that the first $\mathbf{P}$ the one atributed to $\mathbf{P}(x,F)$ is a function from $\Bbb{R} \times \mathcal{F}$ into $[0,1]$, the second $\mathbf{P}$ (atributed to $\mathbf{P}(F \mid \sigma(X))$ is a function from $\Omega \times \mathcal{F}$ into $[0,1]$.

So Let's give them different names, call the first one $\mathbf{P}^1(x,F)$ and the second one $\mathbf{P}^2(\omega, F)$

The second one is readily defined from your previous consideration as you already know how to deal with regular conditional probability distributions.

The first one needs the following theorem: (taken from Meyers Probability and potentials, page 10)

enter image description here

It tells us that $\mathbf{P}^2(\omega,F) = P^1(X(\omega),F)$ Use this $\mathbf{P}^1(x,F)$ as your object and then you can safely say that

\begin{align*} \int_B \mathbf{P}^1(x,F) \, \mathrm{d}\Lambda_X(x) & = \int_C \mathbf{P}^1(X(\omega), F) \, \mathrm{d} P(\omega) = \int_C\mathbf{P}^2(\omega, F) \, \mathrm{d} P(\omega) \; (\ldots)\\ \end{align*}

$\endgroup$
  • $\begingroup$ Thanks for the great explanation. Do you have a good definition for the regular conditional probability given $X$, then? $\endgroup$ – bcf Sep 15 '15 at 23:54
  • $\begingroup$ :) The best definition I know is on the Book of Karatzas Brownian motion and Stochastoc Calculus on page 307. Should you need more detail just ask. $\endgroup$ – Conrado Costa Sep 15 '15 at 23:57
  • $\begingroup$ I updated by question based on your answer, would you mind seeing if my understanding is correct? $\endgroup$ – bcf Sep 20 '15 at 15:37
  • $\begingroup$ I.e., $\mathbf{P}^1$ is the $h$ in Theorem 18, which is guaranteed to exist as soon as $\mathbf{P}^2$ (which would be $g$ in the theorem) exists? $\endgroup$ – bcf Sep 20 '15 at 15:46
  • $\begingroup$ Your understanding is correct. $\endgroup$ – Conrado Costa Sep 20 '15 at 19:26

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.