10 Regular Conditional Probability / Distribution

(Durrett §5.1.3, p. 193)

Let
\((\Omega,\mathcal{F}_0,P),\qquad \mathcal{F}\subset \mathcal{F}_0.\)

A regular conditional probability is a function
\(\mu:\Omega\times \mathcal{F}_0 \to [0,1]\) such that:

  1. Version property:
    \(\mu(\omega,A)=P_{\mathcal{F}}(A)(\omega)\quad\text{a.s., for all }A\in \mathcal{F}_0.\)

  2. Probability–measure property: For each fixed $\omega$,
    \(A\mapsto \mu(\omega,A)\) is a probability measure on $\mathcal{F}_0$.

Application:
\(E_{\mathcal{F}}(f(X))(\omega)=\int_{-\infty}^\infty f(x)\, d\mu(\omega,x).\)

In general, regular conditional probabilities do not always exist.
They do exist in important cases, such as when: \((\Omega,\mathcal{F}_0,P) = (\mathbb{R},\mathcal{B},P), \quad \mathcal{F}\subset\mathcal{B}.\)


Construction in $\mathbb{R}$

Let $Q = \mathbb{Q}$ (rationals).

Step 1:

Define for all $q\in Q$, \(M_q(\omega) = P_{\mathcal{F}}((-\infty,q])(\omega).\)

Step 2:

Restrict to a full-measure subset $\Omega_0\subseteq \Omega$ with $P(\Omega_0^c)=0$ such that for all $\omega\in \Omega_0$:

(a) Monotonicity:
\(q_2>q_1 \quad \Rightarrow \quad M_{q_2}(\omega)\ge M_{q_1}(\omega).\)

(b) Right-continuity:
\(M_q(\omega) = \lim_{q_n \downarrow q} M_{q_n}(\omega).\)

(c) Limits at ±∞:
\(\lim_{q\to\infty} M_q(\omega)=1, \qquad \lim_{q\to -\infty} M_q(\omega)=0.\)

Step 3:

Define the full CDF: \(M_x(\omega)=\inf_{q\ge x} M_q(\omega),\qquad \omega\in\Omega_0.\)

Then
\(P_{\mathcal{F}}((-\infty,x])(\omega)= M_x(\omega).\)

Step 4:

For any Borel set $(a,b]$, \(\mu(\omega,(a,b]) = M_b(\omega)-M_a(\omega).\)

Thus, \(P_{\mathcal{F}}(A)(\omega)=\mu(\omega,A) \qquad (\forall A\in\mathcal{B}).\)


Regular Conditional Probability for Random Variables

Given a random variable $X:\Omega\to\mathbb{R}$, \(\sigma(X)=\{X\in B : B\in\mathcal{B}(\mathbb{R})\}.\)

A regular conditional distribution of $X$ given $\mathcal{F}$ is: \(\mu(\omega,B)=P_{\mathcal{F}}(X\in B)(\omega).\)


Example: Decomposing Borel sets

Let
\(\mathbb{R}^+=[0,\infty),\qquad \mathbb{R}^- = (-\infty,0).\)

Given a Borel set $A\subset\mathbb{R}$, \(A=A^+ \cup A^-, \qquad A^+=A\cap \mathbb{R}^+, \quad A^- = A\cap\mathbb{R}^-.\)

Define a σ-field: \(\mathcal{F} = \{B,\; B\cup\mathbb{R}^- : B\in\mathcal{B}(\mathbb{R}^+)\}.\)

Then $$ P_{\mathcal{F}}(A)(\omega) = P_{\mathcal{F}}(A^+)(\omega)

  • P_{\mathcal{F}}(A^-)(\omega). $$

If $X\in\mathcal{F}$:

\[\mu(\omega,A)=\mathbf{1}_{A^+}(\omega)+ \frac{P(A^-)}{P(\mathbb{R}^-)}\mathbf{1}_{\mathbb{R}^-}(\omega).\]

Example: Conditional Distribution of $(X,Y)$

Suppose $(X,Y)$ has a joint density $f_{XY}(x,y)$.

Conditional density: \(f_{Y\mid X=x_0}(y)= \frac{f_{XY}(x_0,y)}{f_X(x_0)}, \qquad f_X(x_0)=\int_{-\infty}^\infty f_{XY}(x_0,y)\,dy.\)

Thus for a Borel set $D\subset \mathbb{R}^2$, \(P_{\sigma(X)}((X,Y)\in D)(\omega) = \int_{\{y:(x_0,y)\in D\}} f_{Y\mid X=x_0}(y)\,dy, \quad x_0=X(\omega).\)


Comments