This repository has been archived on 2023-10-20.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_18.tex

247 lines
8.3 KiB
TeX
Raw Normal View History

2023-06-20 17:56:38 +02:00
\lecture{18}{2023-06-20}{}
2023-06-29 22:18:23 +02:00
Recall our key lemma \ref{lec17l3} for supermartingales from last time:
2023-06-20 17:56:38 +02:00
\[
(b-a) \bE[U_N([a,b])] \le \bE[(X_n - a)^-].
2023-07-06 00:36:26 +02:00
\]
2023-06-20 17:56:38 +02:00
What happens for submartingales?
If $(X_n)_{n \in \N}$ is a submartingale, then $(-X_n)_{n \in \N}$ is a supermartingale.
Hence the same holds for submartingales, i.e.
\begin{lemma}
A (sub-/super-) martingale bounded in $L^1$ converges
a.s.~to a finite limit, which is a.s.~finite.
\end{lemma}
2023-07-05 17:53:41 +02:00
\subsection{Doob's $L^p$ Inequality}
2023-06-29 22:18:23 +02:00
2023-06-20 17:56:38 +02:00
\begin{question}
What about $L^p$ convergence of martingales?
\end{question}
2023-07-17 18:39:05 +02:00
\begin{example}[A martingale not converging in $L^1$ ]
2023-07-18 23:58:27 +02:00
\label{ex:martingale-not-converging-in-l1}
2023-07-17 18:39:05 +02:00
Fix $u > 1$ and let $p = \frac{1}{1+u}$.
2023-06-20 17:56:38 +02:00
Let $ (Z_n)_{n \ge 1}$ be i.i.d.~$\pm 1$ with
2023-07-17 16:11:03 +02:00
$\bP[Z_n = 1] = p$.
Let $X_0 = x > 0$ and
define $X_{n+1} \coloneqq u^{Z_{n+1}} X_{n}$.
Then $(X_n)_n$ is a martingale,
since
\begin{IEEEeqnarray*}{rCl}
\bE[X_{n+1} |\cF_n] &=& X_n \bE[u^{Z_{n+1}}]\\
&=& X_n \left(p \cdot u + (1-p)\cdot\frac{1}{u}\right)\\
&=& X_n \left(\frac{p (u^2-1) + 1}{u}\right)\\
&=& X_n.
\end{IEEEeqnarray*}
By \autoref{doobmartingaleconvergence},
there exists an a.s.~limit $X_\infty$.
2023-07-17 18:39:05 +02:00
By the SLLN, we have almost surely
2023-06-20 17:56:38 +02:00
\[
2023-07-17 16:11:03 +02:00
\frac{1}{n} \sum_{k=1}^{n} Z_k \xrightarrow{a.s.} \bE[Z_1] = 2p - 1.
2023-07-06 00:36:26 +02:00
\]
2023-06-20 17:56:38 +02:00
Hence
\[
\left(\frac{X_n}{x}\right)^{\frac{1}{n}} = u^{\frac{1}{n} \sum_{k=1}^n Z_k}
2023-07-17 18:39:05 +02:00
\xrightarrow{\text{a.s.}} u^{2p -1}.
2023-07-06 00:36:26 +02:00
\]
2023-07-17 18:39:05 +02:00
Since $(X_n)_{n \ge 0}$ is a martingale, we have $\bE[u^{Z_1}] = 1$.
2023-06-20 17:56:38 +02:00
Hence $2p - 1 < 0$, because $u > 1$.
2023-07-17 18:39:05 +02:00
Choose $\epsilon > 0$ small enough such that $u^{2p - 1}(1 + \epsilon) < 1$.
Then there exists $N_0(\epsilon)$ (possibly random)
such that for all $n > N_0(\epsilon)$ almost
2023-06-20 17:56:38 +02:00
\[
2023-07-17 18:39:05 +02:00
\left( \frac{X_n}{x} \right)^{\frac{1}{n}} \overset{\text{a.s.}}{\le}
u^{2p - 1}(1 + \epsilon) %
\implies x [\underbrace{u^{2p - 1} (1+\epsilon)}_{<1}]^n \xrightarrow[n \to \infty]{\bP} 0.
2023-06-22 17:50:06 +02:00
\]
2023-07-17 18:39:05 +02:00
However, $X_n$ cannot converge to $0$ in $L^1$,
as $\bE[X_n] = \bE[X_0] = x > 0$.
2023-06-20 17:56:38 +02:00
\end{example}
$L^2$ is nice, since it is a Hilbert space. So we will first
consider $L^2$.
\begin{fact}[Martingale increments are orthogonal in $L^2$ ]
\label{martingaleincrementsorthogonal}
2023-07-17 18:39:05 +02:00
Let $(X_n)_n$ be a martingale with $X_n \in L^2$ for all $n$
2023-07-06 00:36:26 +02:00
and let $Y_n \coloneqq X_n - X_{n-1}$
2023-06-20 17:56:38 +02:00
denote the \vocab{martingale increments}.
Then for all $m \neq n$ we have that
\[
\langle Y_m | Y_n\rangle_{L^2} = \bE[Y_n Y_m] = 0.
\]
\end{fact}
\begin{proof}
2023-07-17 18:39:05 +02:00
As $\bE[Y_n^2] = \bE[X_n^2] - 2\bE[X_nX_{n-1}] + \bE[X_{n-1}^2] < \infty$,
we have $Y_n \in L^2$.
2023-06-20 17:56:38 +02:00
Since $\bE[X_n | \cF_{n-1}] = X_{n-1}$ a.s.,
by induction $\bE[X_n | \cF_{k}] = X_k$ a.s.~for all $k \le n$.
2023-07-17 18:39:05 +02:00
In particular $\bE[Y_n | \cF_k] = 0$ for $k < n$.
Suppose that $m < n$.
Then
\begin{IEEEeqnarray*}{rCl}
\bE[Y_n Y_m] &=& \bE[\bE[Y_n Y_m | \cF_m]]\\
&=& \bE[Y_m \bE[Y_n | \cF_m]]\\
&=& 0
\end{IEEEeqnarray*}
2023-06-20 17:56:38 +02:00
\end{proof}
2023-06-29 22:18:23 +02:00
\begin{fact}[\vocab{Parallelogram identity}]
Let $X, Y \in L^2$.
Then
\[
2 \bE[X^2] + 2 \bE[Y^2] = \bE[(X+Y)^2] + \bE[(X-Y)^2].
2023-07-06 00:36:26 +02:00
\]
2023-06-20 17:56:38 +02:00
\end{fact}
2023-06-29 22:18:23 +02:00
\begin{theorem}\label{martingaleconvergencel2}
2023-06-20 17:56:38 +02:00
Suppose that $(X_n)_n$ is a martingale bounded in
2023-06-29 22:18:23 +02:00
$L^2$,\\
i.e.~$\sup_n \bE[X_n^2] < \infty$.
2023-06-20 17:56:38 +02:00
Then there is a random variable $X_\infty$ such that
\[
X_n \xrightarrow{L^2} X_\infty.
2023-07-06 00:36:26 +02:00
\]
2023-06-20 17:56:38 +02:00
\end{theorem}
\begin{proof}
Let $Y_n \coloneqq X_n - X_{n-1}$ and write
\[
X_n = \sum_{j=1}^{n} Y_j.
2023-07-06 00:36:26 +02:00
\]
2023-06-20 17:56:38 +02:00
We have
\[
\bE[X_n^2] = \bE[X_0^2] + \sum_{j=1}^{n} \bE[Y_j^2]
2023-07-06 00:36:26 +02:00
\]
2023-07-17 18:39:05 +02:00
by \autoref{martingaleincrementsorthogonal}.
2023-06-20 17:56:38 +02:00
In particular,
\[
\sup_n \bE[X_n^2] < \infty \iff \sum_{j=1}^{\infty} \bE[Y_j^2] < \infty.
2023-07-06 00:36:26 +02:00
\]
2023-06-20 17:56:38 +02:00
Since $(X_n)_n$ is bounded in $L^2$,
there exists $X_\infty$ such that $X_n \xrightarrow{\text{a.s.}} X_\infty$
by \autoref{doob}.
It remains to show $X_n \xrightarrow{L^2} X_\infty$.
For any $r \in \N$, consider
\[\bE[(X_{n+r} - X_n)^2] = \sum_{j=n+1}^{n+r} \bE[Y_j^2] \xrightarrow{n \to \infty} 0\]
as a tail of a convergent series.
Hence $(X_n)_n$ is Cauchy, thus it converges in $L^2$.
Since $\bE[(X_\infty - X_n)^2]$ converges to the increasing
limit
\[
2023-07-17 18:39:05 +02:00
\sum_{j \ge n + 1} \bE[Y_j^2] \xrightarrow{n\to \infty} 0
2023-07-06 00:36:26 +02:00
\]
2023-06-20 17:56:38 +02:00
we get $\bE[(X_\infty - X_n)^2] \xrightarrow{n\to \infty} 0$.
\end{proof}
Now let $p \ge 1$ be not necessarily $2$.
First, we need a very important inequality:
\begin{theorem}[Doob's $L^p$ inequality]
\label{dooblp}
2023-07-17 20:56:30 +02:00
Suppose that $(X_n)_n$ is a martingale
or a non-negative submartingale.
2023-06-20 17:56:38 +02:00
Let $X_n^\ast \coloneqq \max \{|X_1|, |X_2|, \ldots, |X_n|\}$
denote the \vocab{running maximum}.
\begin{enumerate}[(1)]
2023-07-16 01:15:14 +02:00
\item Then \[ \forall \ell > 0 .~\bP[X_n^\ast \ge \ell] \le \frac{1}{\ell} \int_{\{X_n^\ast \ge \ell\}} |X_n| \dif \bP \le \frac{1}{\ell} \bE[|X_n|]. \]
2023-06-20 17:56:38 +02:00
(Doob's $L^1$ inequality).
\item Fix $p > 1$. Then \[
\bE[(X_n^\ast)^p] \le \left( \frac{p}{p-1} \right)^p \bE[|X_n|^p].
2023-06-22 17:50:06 +02:00
\]
2023-06-20 17:56:38 +02:00
(Doob's $L^p$ inequality).
\end{enumerate}
\end{theorem}
2023-06-29 22:18:23 +02:00
In order to prove \autoref{dooblp}, we first need
2023-06-20 17:56:38 +02:00
\begin{lemma}
\label{dooplplemma}
2023-07-17 20:56:30 +02:00
Let $p > 1$ and $X,Y$ non-negative random variables
2023-06-20 17:56:38 +02:00
such that
\[
2023-07-17 18:39:05 +02:00
\forall \ell > 0 .~ \bP[Y \ge \ell] \le
\frac{1}{\ell} \int_{\{Y \ge \ell\} } X \dif \bP
2023-06-22 17:50:06 +02:00
\]
2023-06-20 17:56:38 +02:00
Then
\[
\bE[Y^p] \le \left( \frac{p}{p-1} \right)^p \bE[X^p].
2023-06-22 17:50:06 +02:00
\]
2023-06-20 17:56:38 +02:00
\end{lemma}
\begin{proof}
First, assume $Y \in L^p$.
Then
\begin{IEEEeqnarray}{rCl}
2023-07-17 18:39:05 +02:00
\|Y\|_{L^p}^p
&=& \bE[Y^p]\\
&=& \int Y(\omega)^p \dif \bP(\omega)\\
&=&\int_{\Omega} \left( \int_0^{Y(\omega)} p \ell^{p-1} \dif \ell
\right) \dif \bP(\omega)\\
&\overset{\text{Fubini}}{=}&
\int_0^\infty p \ell^{p-1}\underbrace{\int_\Omega \One_{Y \ge \ell}\dif \bP}_%
{\bP[Y \ge \ell]} \dif\ell. \label{l18star}
2023-06-20 17:56:38 +02:00
\end{IEEEeqnarray}
2023-07-17 18:39:05 +02:00
By the assumption it follows that
2023-06-20 17:56:38 +02:00
\begin{IEEEeqnarray*}{rCl}
2023-07-17 18:39:05 +02:00
\eqref{l18star}
&\le& \int_0^\infty p \ell^{p-2}
\int_{\{Y(\omega) \ge \ell\}} X(\omega) \bP(\dif \omega)\dif \ell\\
&\overset{\text{Fubini}}{=}&
\int_\Omega X(\omega) \int_{0}^{Y(\omega)} p \ell^{p-2} \dif \ell\bP(\dif \omega)\\
&=& \frac{p}{p-1} \int_{\omega} X(\omega) Y (\omega)^{p-1} \bP(\dif \omega)\\
&\overset{\text{Hölder}}{\le}& \frac{p}{p-1} \|X\|_{L^p} \|Y\|_{p}^{p-1},
2023-06-20 17:56:38 +02:00
\end{IEEEeqnarray*}
where the assumption was used to apply Hölder.
Suppose now $Y \not\in L^p$.
Then look at $Y_M = Y \wedge M$.
2023-06-29 22:18:23 +02:00
Apply the above to $Y_M \in L^p$ and use the monotone convergence theorem.
2023-06-20 17:56:38 +02:00
\end{proof}
\begin{refproof}{dooblp}
Let $E \coloneqq \{X_n^\ast \ge \ell\} = E_1 \sqcup \ldots \sqcup E_n$
where
\[
E_j = \{|X_1| \le \ell, |X_2| \le \ell, \ldots, |X_{j-1}| \le \ell, |X_j| \ge \ell\}.
2023-07-06 00:36:26 +02:00
\]
Then
2023-06-29 22:18:23 +02:00
\begin{equation}
\bP[E_j] \overset{\text{Markov}}{\le } \frac{1}{\ell} \int_{E_j} |X_j| \dif \bP
\label{lec18eq2star}
\end{equation}
2023-07-17 20:56:30 +02:00
We have that $(|X_n|)_n$ is a submartingale,
by \autoref{cor:convexmartingale}
in the case of $X_n$ being a martingale
and trivially if $X_n$ is non-negative.
2023-06-20 17:56:38 +02:00
Hence
\begin{IEEEeqnarray*}{rCl}
\bE[\One_{E_j}(|X_n| - |X_{j}|) | \cF_j]
&=& \One_{E_j} \bE[(|X_n| - |X_{j}|)|\cF_j]\\
&\overset{\text{a.s.}}{\ge }& 0.
\end{IEEEeqnarray*}
By the law of total expectation, \autoref{totalexpectation},
it follows that
2023-06-29 22:18:23 +02:00
\begin{equation}
\bE[\One_{E_j} (|X_n| - |X_j|)] \ge 0. \label{lec18eq3star}
\end{equation}
2023-07-06 00:36:26 +02:00
2023-06-20 17:56:38 +02:00
Now
\begin{IEEEeqnarray*}{rCl}
\bP(E) &=& \sum_{j=1}^n \bP(E_j)\\
2023-06-29 22:18:23 +02:00
&\overset{\eqref{lec18eq2star}, \eqref{lec18eq3star}}{\le }& \frac{1}{\ell} \left( \int_{E_1} |X_n| \dif \bP + \ldots + \int_{E_n} |X_n| \dif \bP \right)\\
2023-06-20 17:56:38 +02:00
&=& \frac{1}{\ell} \int_E |X_n| \dif \bP
\end{IEEEeqnarray*}
This proves the first part.
2023-07-06 00:36:26 +02:00
For the second part, we apply the first part and
2023-06-20 17:56:38 +02:00
\autoref{dooplplemma} (choose $Y \coloneqq X_n^\ast$).
\end{refproof}
2023-07-17 18:39:05 +02:00
\todo{Branching process}