This shows you the differences between two versions of the page.
world:martingale [2021/03/30 14:08] rdouc |
world:martingale [2022/03/16 07:40] |
||
---|---|---|---|
Line 1: | Line 1: | ||
- | {{page>:defs}} | ||
- | |||
- | ===== Supermartingale convergence results ===== | ||
- | --- //[[[email protected]|randal]] 2021/03/30 14:04// | ||
- | |||
- | <WRAP center round tip 80%> | ||
- | **__Theorem__**. | ||
- | Let $\mcf=(\mcf_n)_{n\in\nset}$ be a filtration and let $\seq{X_n}{n\in\nset}$ be an $\mcf$-adapted sequence of $\lone$-random variables such that | ||
- | * $M=\sup_{n\in\nset} \PE[(X_n)^-]<\infty$ | ||
- | * for all $n\geq 1$, we have $\PE[X_{n}|\mcf_{n-1}]\leq X_{n-1}$. | ||
- | that is, $\seq{X_n}{n\in\nset}$ is a $(\mcf_n)_{n\in\nset}$-supermartingale, with negative part bounded in $\lone$. | ||
- | Then, almost surely, $X_\infty=\lim_{n\to\infty} X_n$ exists and is in $\lone$. | ||
- | |||
- | </WRAP> | ||
- | |||
- | |||
- | |||
- | ==== Proof ==== | ||
- | |||
- | Let $a<b$ and define $C_1=\indiacc{X_0<a}$ and for $n\geq 2$, | ||
- | $$ | ||
- | C_n=\indiacc{C_{n-1}=1,X_{n-1} \leq b}+\indiacc{C_{n-1}=0,X_{n-1} < a} | ||
- | $$ | ||
- | In words, the first time $C_n $ flags $1$ is when $X_{n-1}<a$. Then it flags $1$ until $X_{n-1}$ goes above $b$. Then it flags $0$ until $X_n$ goes below $a$. So consecutive sequences of $C_n=1$ are linked with upcrossings of $[a,b] $ for $(X_n) $. Now, define | ||
- | $$ | ||
- | Y_n=\sum_{k=1}^n C_k(X_k-X_{k-1}). | ||
- | $$ | ||
- | |||
- | Define $U_N[a,b]$ the number of upcrossings of $[a,b]$ for $(X_n)_{0\leq n \leq N } $. Then, | ||
- | $$ | ||
- | Y_N=\sum_{k=1}^N C_k(X_k-X_{k-1}) \geq (b-a) U_N[a,b]-(X_N-a)^- | ||
- | $$ | ||
- | From the fact that $\seq{X_n}{n\in\nset}$ is a $\mcf$-supermartingale and $(C_n)_{n\in\nset}$ is $\mcf$-previsible, we deduce that $\seq{Y_n}{n\in\nset}$ is also a $\mcf$-supermartingale so that $\PE[Y_N]\leq 0$. Then, | ||
- | $$ | ||
- | (b-a) \PE[U_N[a,b]] \leq \PE[(X_N-a)^-]=\PE[(a-X_N)^+] \leq a^+ + \PE[(-X_N)^+]\leq a^+ + M | ||
- | $$ | ||
- | Letting $N$ goes to infinity, the monotone convergence theorem yields: | ||
- | $$ | ||
- | (b-a) \PE[U_\infty[a,b]] \leq a^+ + M | ||
- | $$ | ||
- | and thus, $\PP(U_\infty[a,b]<\infty)=1$ for all $a<b$. Now, | ||
- | \begin{align*} | ||
- | \PP(\liminf_n X_n<\limsup_n X_n) &\leq \sum_{a,b\in \mathbb{Q}, a<b} \PP\lr{\liminf_n X_n <a<b<\limsup_n X_n}\\ | ||
- | & \leq \sum_{a,b\in \mathbb{Q}, a<b} \PP(U_\infty[a,b]=\infty)=0 | ||
- | \end{align*} | ||
- | which shows that $X_\infty=\lim_{n \to \infty} X_n$ exits almost surely. | ||
- | |||
- | Moreover, $\PE[X_0] \geq \PE[X_n]=\PE[X^+_n]-\PE[X^-_n]$ so that | ||
- | $$ | ||
- | \sup_n \PE[|X_n|]=\sup_n \lr{\PE[X^+_n]+\PE[X^-_n]} \leq \PE[X_0] +2 \sup_n \PE[X^-_n] \leq \PE[X_0] +2 M<\infty | ||
- | $$ | ||
- | which implies by Fatou's lemma that $\PE[|X_\infty|]=\PE[\liminf_{n}|X_n|] \leq \liminf_{n} \PE[|X_n|] \leq \sup_n \PE[|X_n|]<\infty$. The proof is completed. | ||
- | ===== Corollary: Submartingale convergence results ===== | ||
- | |||
- | {{anchor:submartingale:}} | ||
- | As a consequence, the conclusion also holds if $\seq{X_n}{n\in\nset}$ is a $\mcf$-submartingale, with positive part bounded in $\lone$ : indeed, we only need to apply the previous result to $-X_n$. | ||
- | |||
- | <WRAP center round tip 80%> | ||
- | **__Corollary__** Assume that $\mcf=(\mcf_n)_{n\in\nset}$ is a filtration and let $\seq{X_n}{n\in\nset}$ be an $\mcf$-adapted sequence of $\lone$ random variables such that | ||
- | * $M=\sup_{n\in\nset} \PE[(X_n)^+]<\infty$ | ||
- | * for all $n\geq 1$, we have $\PE[X_{n}|\mcf_{n-1}]\geq X_{n-1}$. | ||
- | that is, $\seq{X_n}{n\in\nset}$ is a $\mcf$-submartingale, with positive part bounded in $\lone$. | ||
- | Then, almost surely, $X_\infty=\lim_{n\to\infty} X_n$ exists and is in $\lone$. | ||
- | |||
- | </WRAP> | ||
- | |||