This repository has been archived on 2023-10-20. You can view files and clone it, but cannot push or open issues or pull requests.
probability-theory/inputs/lecture_15.tex

268 lines
7.9 KiB
TeX
Raw Normal View History

2023-06-07 02:50:56 +02:00
\lecture{15}{2023-06-06}{}
2023-07-05 17:53:41 +02:00
\subsection{Properties of Conditional Expectation}
2023-06-07 02:50:56 +02:00
We want to derive some properties of conditional expectation.
2023-07-05 17:53:41 +02:00
\begin{theorem}[Law of total expectation]
2023-06-07 02:50:56 +02:00
\label{ceprop1}
\label{totalexpectation}
\[
\bE[\bE[X | \cG ]] = \bE[X].
\]
\end{theorem}
\begin{proof}
Apply (b) from the definition for $G = \Omega \in \cG$.
\end{proof}
2023-06-29 22:18:23 +02:00
\begin{theorem}
2023-06-07 02:50:56 +02:00
\label{ceprop2}
2023-06-29 22:18:23 +02:00
If $X$ is $\cG$-measurable, then $X \overset{\text{a.s.}}{=}\bE[X | \cG]$.
2023-06-07 02:50:56 +02:00
\end{theorem}
\begin{proof}
Suppose $\bP[X \neq Y] > 0$.
Without loss of generality $\bP[X > Y] > 0$.
Hence $\bP[ X > Y + \frac{1}{n}]> 0$ for some $n \in \N$.
Let $A \coloneqq \{X > Y + \frac{1}{n}\}$.
2023-06-29 22:18:23 +02:00
Then
\[
\int_A X \dif \bP \ge \frac{1}{n}\bP(A) + \int_A Y \dif \bP,
2023-07-06 00:36:26 +02:00
\]
2023-06-29 22:18:23 +02:00
contradicting property (b) from \autoref{conditionalexpectation}.
2023-06-07 02:50:56 +02:00
\end{proof}
\begin{example}
Suppose $X \in L^1(\bP)$, $\cG \coloneqq \sigma(X)$.
Then $X$ is measurable with respect to $\cG$.
Hence $\bE[X | \cG] = X$.
\end{example}
\begin{theorem}[Linearity]
\label{ceprop3}
\label{celinearity}
For all $a,b \in \R$
we have
\[
\bE[a X_1 + bX_2 | \cG] = a \bE[X_1 | \cG] + b \bE[X_2|\cG].
\]
\end{theorem}
\begin{proof}
2023-06-29 22:18:23 +02:00
trivial\todo{add details}
2023-06-07 02:50:56 +02:00
\end{proof}
\begin{theorem}[Positivity]
\label{ceprop4}
\label{cpositivity}
If $X \ge 0$, then $\bE[X | \cG] \ge 0$ a.s.
\end{theorem}
\begin{proof}
2023-06-07 03:35:56 +02:00
Let $W $ be a version of $\bE[X | \cG]$.
2023-06-07 02:50:56 +02:00
Suppose $\bP[ W < 0] > 0$.
2023-06-29 22:18:23 +02:00
Then \[G \coloneqq \{W < -\frac{1}{n}\} \in \cG.\]
2023-06-07 02:50:56 +02:00
For some $n \in \N$, we have $\bP[G] > 0$.
However it follows that
\[
\int_G \bP[X | \cG] \dif \bP \le -\frac{1}{n} \bP[G] < 0 \le \int_G X \dif \bP.
2023-07-06 00:36:26 +02:00
\]
2023-06-07 02:50:56 +02:00
\end{proof}
\begin{theorem}[Conditional monotone convergence theorem]
\label{ceprop5}
\label{mcmt}
Let $X_n,X \in L^1(\Omega, \cF, \bP)$.
Suppose $X_n \ge 0$ with $X_n \uparrow X$.
Then $\bE[X_n|\cG] \uparrow \bE[X|\cG]$.
\end{theorem}
\begin{proof}
Let $Z_n$ be a version of $\bE[X_n | Y]$.
Since $X_n \ge 0$ and $X_n \uparrow$,
by \autoref{cpositivity},
we have
\[
\bE[X_n | \cG] \overset{\text{a.s.}}{\ge } 0
2023-07-06 00:36:26 +02:00
\]
2023-06-07 02:50:56 +02:00
and
\[
\bE[X_n | \cG] \uparrow \text{a.s.}
2023-07-06 00:36:26 +02:00
\]
2023-06-07 02:50:56 +02:00
(consider $X_{n+1} - X_n$ ).
Define $Z \coloneqq \limsup_{n \to \infty} Z_n$.
Then $Z$ is $\cG$-measurable
and $Z_n \uparrow Z$ a.s.
Take some $G \in \cG$.
We know by (b) % TODO REF
2023-06-07 03:35:56 +02:00
that $\bE[Z_n \One_G] = \bE[X_n \One_G]$.
2023-07-06 00:36:26 +02:00
The LHS increases to $\bE[Z \One_G]$ by the monotone
2023-06-07 02:50:56 +02:00
convergence theorem.
Again by MCT, $\bE[X_n \One_G]$ increases to
$\bE[X \One_G]$.
Hence $Z$ is a version of $\bE[X | \cG]$.
\end{proof}
\begin{theorem}[Conditional Fatou]
\label{ceprop6}
\label{cfatou}
Let $X_n \in L^1(\Omega, \cF, \bP)$, $X_n \ge 0$.
Then
\[
\bE[ \liminf_{n \to \infty} X_n | \cG] \le \liminf_{n \to \infty} \bE[X_n | \cG].
2023-07-06 00:36:26 +02:00
\]
2023-06-07 02:50:56 +02:00
\end{theorem}
\begin{proof}
2023-06-29 22:18:23 +02:00
\notes
2023-06-07 02:50:56 +02:00
\end{proof}
\begin{theorem}[Conditional dominated convergence theorem]
\label{ceprop7}
\label{cdct}
2023-07-19 20:59:36 +02:00
Let $X_n,Y \in L^1(\Omega, \cF, \bP)$.
Suppose that $\sup_n |X_n(\omega)| < Y(\omega)$ a.e.~
and that $X_n$ converges to a pointwise limit $X$.
Then $\bE[ X_n | \cG] \to \bE[X | \cG]$ a.e.
2023-06-07 02:50:56 +02:00
\end{theorem}
\begin{proof}
2023-06-29 22:18:23 +02:00
\notes
2023-06-07 02:50:56 +02:00
\end{proof}
Recall
2023-06-29 22:18:23 +02:00
\begin{fact}[Jensen's inequality]
2023-07-14 22:07:36 +02:00
\label{jensen}
2023-06-07 02:50:56 +02:00
If $c : \R \to \R$ is convex and $\bE[|c \circ X|] < \infty$,
2023-06-29 22:18:23 +02:00
then $\bE[c \circ X] \overset{\text{a.s.}}{\ge} c(\bE[X])$.
\end{fact}
2023-06-07 02:50:56 +02:00
For conditional expectation, we have
\begin{theorem}[Conditional Jensen's inequality]
\label{ceprop8}
\label{cjensen}
Let $X \in L^1(\Omega, \cF, \bP)$.
If $c : \R \to \R$ is convex and $\bE[|c \circ X|] < \infty$,
then $\bE[c \circ X | \cG] \ge c(\bE[X | \cG])$ a.s.
\end{theorem}
\begin{fact}
\label{convapprox}
If $c$ is convex, then there are two sequences of real numbers
$a_n, b_n \in \R$
such that
\[
c(x) = \sup_n(a_n x + b_n).
2023-07-06 00:36:26 +02:00
\]
2023-06-07 02:50:56 +02:00
\end{fact}
\begin{refproof}{cjensen}
By \autoref{convapprox}, $c(x) \ge a_n X + b_n$
for all $n$.
Hence
\[
\bE[c(X) | \cG] \ge a_n \bE[X | \cG] + \bE[b_n | \cG]
2023-06-29 22:18:23 +02:00
= a_n \bE[X | \cG] + b_n \text{ a.s.}
2023-07-06 00:36:26 +02:00
\]
2023-06-07 02:50:56 +02:00
for all $n$.
Using that a countable union of sets o f measure zero has measure zero,
we conclude that a.s~this happens simultaneously for all $n$.
Hence
\[
\bE[c(X) | \cG] \ge \sup_n (a_n \bE[X | \cG] + b_n) \overset{\text{\autoref{convapprox}}}{=} c(\bE(X | \cG)).
2023-07-06 00:36:26 +02:00
\]
2023-06-07 02:50:56 +02:00
\end{refproof}
Recall
2023-06-29 22:18:23 +02:00
\begin{fact}[Hölder's inequality]
2023-06-07 02:50:56 +02:00
Let $p,q \ge 1$ such that $\frac{1}{p} + \frac{1}{q} = 1$.
Suppose $X \in L^p(\bP)$ and $Y \in L^q(\bP)$.
Then
\[
\bE(X Y) \le \underbrace{\bE(|X|^p)^{\frac{1}{p}}}_{\text{\reflectbox{$\coloneqq$}} \|X\|_{L^p}} \bE(|Y|^q)^{\frac{1}{q}}.
2023-07-06 00:36:26 +02:00
\]
2023-06-29 22:18:23 +02:00
\end{fact}
2023-06-07 02:50:56 +02:00
\begin{theorem}[Conditional Hölder's inequality]
\label{ceprop9}
\label{choelder}
Let $p,q \ge 1$ such that $\frac{1}{p} + \frac{1}{q} = 1$.
Suppose $X \in L^p(\bP)$ and $Y \in L^q(\bP)$.
Then
\[
\bE(X Y | \cG) \le \bE(|X|^p | \cG)^{\frac{1}{p}} \bE(|Y|^q | \cG)^{\frac{1}{q}}.
2023-07-06 00:36:26 +02:00
\]
2023-06-07 02:50:56 +02:00
\end{theorem}
\begin{proof}
\todo{Exercise}
\end{proof}
\begin{theorem}[Tower property]
\label{ceprop10}
2023-06-27 17:08:59 +02:00
\label{cetower}
2023-06-07 02:50:56 +02:00
Suppose $\cF \supset \cG \supset \cH$ are sub-$\sigma$-algebras.
Then
\[
2023-06-29 22:18:23 +02:00
\bE\left[\bE[X | \cG] \mid \cH\right] \overset{\text{a.s.}}{=} \bE[X | \cH].
2023-07-06 00:36:26 +02:00
\]
2023-06-07 02:50:56 +02:00
\end{theorem}
\begin{proof}
2023-07-05 17:53:41 +02:00
By definition, $\bE[\bE[X | \cG] | \cH]$ is $\cH$-measurable.
For any $H \in \cH$, we have
\begin{IEEEeqnarray*}{rCl}
\int_H \bE[\bE[X | \cG] | \cH] \dif \bP
&=& \int_{H} \bE[X | \cG] \dif \bP\\
&=& \int_H X \dif \bP.
\end{IEEEeqnarray*}
Hence $\bE[\bE[X | \cG] | \cH] \overset{\text{a.s.}}{=} \bE[X | \cH]$.
2023-06-07 02:50:56 +02:00
\end{proof}
\begin{theorem}[Taking out what is known]
\label{ceprop11}
\label{takingoutwhatisknown}
If $Y$ is $\cG$-measurable and bounded, then
\[
\bE[YX| \cG] \overset{\text{a.s.}}{=} Y \bE[X | \cG].
2023-07-06 00:36:26 +02:00
\]
2023-06-07 02:50:56 +02:00
\end{theorem}
\begin{proof}
Assume w.l.o.g.~$X \ge 0$.
Assume $Y = \One_B$, then $Y$ simple, then take the limit (using that $Y$ is bounded).
\todo{Exercise}
\end{proof}
\begin{definition}
Let $\cG$ and $\cH$ be $\sigma$-algebras.
We call $\cG$ and $\cH$ \vocab[$\sigma$-algebra!independent]{independent},
2023-06-29 22:18:23 +02:00
if $\bP(G \cap H) = \bP(G) \bP(H)$
for all events $G \in \cG$, $H \in \cH$.
2023-06-07 02:50:56 +02:00
\end{definition}
\begin{theorem}[Role of independence]
\label{ceprop12}
2023-06-28 23:40:09 +02:00
\label{ceroleofindependence}
Let $X$ be a random variable,
and let $\cG, \cH$ be $\sigma$-algebras.
If $\cH$ is independent of $\sigma\left( \sigma(X), \cG \right)$,
then
\[
\bE[X | \sigma(\cG, \cH)] \overset{\text{a.s.}}{=} \bE[X | \cG].
2023-07-06 00:36:26 +02:00
\]
2023-06-28 23:40:09 +02:00
In particular, if $X$ is independent of $\cG$,
then
\[
\bE[X | \cG] \overset{\text{a.s.}}{=} \bE[X].
2023-07-06 00:36:26 +02:00
\]
2023-06-07 02:50:56 +02:00
\end{theorem}
2023-06-28 23:40:09 +02:00
2023-06-07 02:50:56 +02:00
\begin{example}[Martingale property of the simple random walk]
Suppose $X_1,X_2,\ldots$ are i.i.d.~with $\bP[X_i = 1] = \bP[X_i = -1] = \frac{1}{2}$.
Let $S_n \coloneqq \sum_{i=1}^n X_i$ be the \vocab{simple random walk}.
Let $\cF$ denote the $\sigma$-algebra on the product space.
Define $\cF_n \coloneqq \sigma(X_1,\ldots)$.
Intuitively, $\cF_n$ contains all the information gathered until time $n$.
We have $\cF_1 \subset \cF_2 \subset \cF_3 \subset \ldots$
For $\bE[S_{n+1} | \cF_n]$ we obtain
\begin{IEEEeqnarray*}{rCl}
2023-06-29 22:18:23 +02:00
\bE[S_{n+1} | \cF_n] &\overset{\text{\autoref{celinearity}}}{=}&
2023-06-07 02:50:56 +02:00
\bE[S_n | \cF_n] + \bE[X_{n+1} | \cF_n]\\
&\overset{\text{a.s.}}{=}& S_n + \bE[X_{n+1} | \cF_n]\\
&\overset{\text{\autoref{ceprop12}}}{=}& S_{n} + \bE[X_n]\\
2023-06-13 17:59:57 +02:00
&=& S_n.
2023-06-07 02:50:56 +02:00
\end{IEEEeqnarray*}
\end{example}