s23-probability-theory/inputs/lecture_14.tex

\lecture{14}{2023-05-25}{Conditional expectation}

\section{Conditional Expectation}

\subsection{Introduction}

Consider a probability space $(\Omega, \cF, \bP)$
and two events $A, B \in \cF$ with $\bP(B) > 0$.

\begin{definition}
    The \vocab{conditional probability}  of $A$ given $B$ is defined as
    \[
    \bP(A | B) \coloneqq \frac{\bP(A \cap B)}{\bP(B)}.
    \]
\end{definition}

Suppose we have two random variables $X$ and $Y$ on $\Omega$,
such that $X$ takes distinct values $x_1, x_2,\ldots, x_{m}$
and $Y$ takes distinct values $y_1,\ldots, y_n$.
Then for this case, define the \vocab{conditional expectation}
of $X$ given  $Y = y_j$ as
\[
\bE[X | Y = y_j] \coloneqq  \sum_{i=1}^m x_i \bP[X=x_i | Y = y_j].
\]

The random variable $Z = \bE[X | Y]$
is defined as follows:
If $Y(\omega) = y_j$ then
\[
Z(\omega) \coloneqq  \underbrace{\bE[X | Y = y_j]}_{\text{\reflectbox{$\coloneqq$}} z_j}.
\]
Note that $\Omega_j \coloneqq  \{\omega : Y(\omega) = y_j\}$
defines a partition of $\Omega$ and on each $\Omega_j$
(``the  $j^{\text{th}}$ $Y$-atom'')
$ Z$ is constant.


Let $\cG \coloneqq  \sigma(Y)$.
Then $Z$ is measurable with respect to $\cG$.
Furthermore
\begin{IEEEeqnarray*}{rCl}
    \int_{\{Y = y_j\} } Z  \dif \bP &=& z_j \int_{\{Y = y_j\}} \dif \bP\\
                                    &=& z_j \bP[Y=y_j]\\
                                    &=&\sum_{i=1}^m  x_i \bP[X = x_i | Y = y_j] \bP[Y = y_j]\\
                                    &=&\sum_{i=1}^m x_i \bP[X = x_i, Y = y_j]\\
                                    &=& \int_{\{Y = y_j\}} X \dif \bP.
\end{IEEEeqnarray*}
Hence
\[
\int_{G} Z \dif \bP = \int_{G} X \dif \bP
\]
for all $G \in \cG$.

We now want to generalize this to arbitrary random variables.
\begin{theorem}
    \label{conditionalexpectation}
    Let $(\Omega, \cF, \bP)$ be a probability space, $X \in L^1(\bP)$
    and $\cG \subseteq  \cF$ a sub-$\sigma$-algebra.
    Then there exists a random variable $Z$
    such that
    \begin{enumerate}[(a)]
        \item $Z$ is $\cG$-measurable and $Z \in L^1(\bP)$,
        \item $\int_G Z \dif \bP = \int_G X \dif \bP$
            for all $G \in \cG$.
    \end{enumerate}

    Such a $Z$ is unique up to sets of measure $0$ and is
    called the \vocab{conditional expectation}  of $X$ given
    the $\sigma$-algebra $\cG$ and written
    $Z = \bE[X | \cG]$.
\end{theorem}
\begin{remark}
     Suppose $\cG = \{\emptyset, \Omega\}$,
         then
         \[
             \bE[X | \cG] = (\omega \mapsto \bE[X])
         \]
         is a constant random variable.
\end{remark}

\begin{definition}[Conditional probability]
    Let $A \subseteq \Omega$ be an event and $\cG \subseteq  \cF$
    a sub-$\sigma$-algebra.
    We define the \vocab{conditional probability} of $A$ given $\cG$ by
    \[
    \bP[A | \cG] \coloneqq \bE[\One_A | \cG].
    \]
\end{definition}

\subsection{Existence of Conditional Probability}

We will give two different proves of \yaref{conditionalexpectation}.
The first one will use orthogonal projections.
The second will use the Radon-Nikodym theorem.
We'll first do the easy proof, derive some properties
and then do the harder proof.

\begin{lemma}
    \label{orthproj}
    Suppose $H$ is a \vocab{Hilbert space},
    i.e.~$H$ is a vector space with an inner product $\langle \cdot, \cdot \rangle_H$ which defines a norm by $\|x\|_H^2 = \langle x, x\rangle_H$
    making $H$ a complete metric space.

    For any $x \in H$ and closed, convex subspace $K \subseteq H$,
    there exists a unique $z \in K$ such that the following equivalent conditions hold:
    \begin{enumerate}[(a)]
        \item  $\forall  y \in K : \langle x-z, y\rangle_H = 0$,
        \item $\forall y \in K: \|z-x\|_H \le \|z-x\|_H$.
    \end{enumerate}
\end{lemma}
\begin{proof}
    \notes
\end{proof}

\begin{refproof}{conditionalexpectation}

    Almost sure uniqueness of $Z$:

    Suppose $X \in L^1$ and $Z$ and $Z'$ satisfy (a) and (b).
    We need to show that $\bP[Z \neq Z'] = 0$.
    By (a), we have $Z, Z' \in  L^1(\Omega, \cG, \bP)$.
    By (b), $\bE[(Z - Z') \One_G] = 0$ for all $G \in \cG$.

    Assume that $\bP[Z > Z'] > 0$.
    Since $\{Z > Z' + \frac{1}{n}\} \uparrow \{Z > Z'\}$,
    we see that $\bP[Z > Z' + \frac{1}{n}] > 0$ for some $n$.
    However $\{Z > Z' + \frac{1}{n}\} \in \cG$,
    which is a contradiction, since
    \[
    \bE[(Z - Z') \One_{Z - Z' > \frac{1}{n}}] \ge \frac{1}{n} \bP[ Z - Z' > \frac{1}{n}] > 0.
    \]


    \bigskip
    Existence of $\bE(X | \cG)$ for $X \in L^2$:

    Let $H = L^2(\Omega, \cF, \bP)$
    and $K = L^2(\Omega, \cG, \bP)$.

    $K$ is closed, since a pointwise limit of $\cG$-measurable
    functions is $\cG$ measurable (if it exists).
    By \yaref{orthproj},
    there exists $z \in K$ such that
    \[\bE[(X - Z)^2] = \inf \{ \bE[(X- W)^2] ~|~ W \in L^2(\cG)\}\]
    and
    \begin{equation}
    \forall  Y \in L^2(\cG) : \langle X - Z, Y\rangle = 0.
    \label{lec13_boxcond}
    \end{equation}
    Now, if $G \in \cG$, then $Y \coloneqq  \One_G \in L^2(\cG)$
    and by \eqref{lec13_boxcond} $\bE[Z \One_G] = \bE[X \One_G]$.


    \bigskip
    Existence of  $\bE(X | \cG)$ for $X \in L^1$ :

    Let $X = X^+ - X^-$.
    It suffices to show (a) and (b) for  $X^+$.
    Choose bounded random variables $X_n \ge 0$ such that $X_n \uparrow X$.
    Since each $X_n \in L^2$, we can choose a version $Z_n$ of $\bE(X_n | \cG)$.

    \begin{claim}
        $0 \overset{\text{a.s.}}{\le} Z_n \uparrow$.
    \end{claim}
    \begin{subproof}
        \notes
    \end{subproof}

    Define $Z(\omega) \coloneqq  \limsup_{n \to \infty} Z_n(\omega)$.
    Then $Z$ is $\cG$-measurable and since $Z_n \uparrow Z$,
    by the \yaref{cmct},
    $\bE(Z \One_G) = \bE(X \One_G)$ for all  $G \in \cG$.
\end{refproof}