Gambling system theorem given by Doob

248 Views Asked by At

Let$\{X_n\}_{n=1}^{\infty}$ be a sequence of i.i.d. random variable. Let $\{\alpha_k\}_{k=1}^{\infty} $be a sequence of strictly increasing finite stopping times.
Then $\{X_{\alpha_k+1}\}_{n=1}^{\infty}$ is also a sequence of i.i.d. random variable.
I can prove this theorem under the assumption that the family $\{X_n,\alpha_k\}_{n,k\in\mathbb{N}}$ is independent. Actually, to prove the independence I just prove the case of finite intersection and the independence of infinite intersection follows by taking the limit. Here is my proof:
Let $\Lambda_n=\bigcap\limits_{k=1}^{n} \{X_{\alpha_k+1} \in B_k\} $ for $ n\in\mathbb{N}$
Claim: $P(\Lambda_n)=\prod\limits_{k=1}^{n}P(X_{k} \in B_k)$ $$P(\Lambda_1)=\sum\limits_{m=0}^{\infty}P(X_{m+1} \in B_1,\alpha_1=m) \\ =\sum\limits_{m=0}^{\infty}P(X_{m+1} \in B_1)P(\alpha_1=m) \\ =\sum\limits_{m=0}^{\infty}P(X_{1} \in B_1)P(\alpha_1=m)\\=P(X_{1} \in B_1)$$ Assume it is true for $n=N-1$.
$$P(\Lambda_N)=\sum\limits_{m=N}^{\infty}P(X_{m+1} \in B_N,\alpha_N=m,\Lambda_{N-1})\\ =\sum\limits_{m=N}^{\infty}\sum\limits_{0\leq i_1 < i_2<...<i_{N-1}<m}P(X_{m+1} \in B_N,\alpha_N=m,\alpha_k=i_k,X_{i_k+1}\in B_k , k=1,2,...,N-1) \\ =\sum\limits_{m=N}^{\infty}\sum\limits_{0\leq i_1 < i_2<...<i_{N-1}<m}P(X_{m+1} \in B_N)P(\alpha_N=m,\alpha_k=i_k,X_{i_k+1}\in B_k , k=1,2,...,N-1) \\ =P(X_N\in B_N)P(\Lambda_{N-1})$$

As $P(X_{\alpha_k+1}\in B)=P(X_{k} \in B)$ (from the proof for n=1), the independence is proved.
And by the same reason, $P(X_{\alpha_k+1}\in B)=P(X_{k} \in B)=P(X_{1} \in B)=P(X_{\alpha_1+1}\in B)$ for all $k\in \mathbb{N}$, it is identically distributed.

I am not sure whether my proof is correct.
The main problem is whether assumption "The family $\{X_n,\alpha_k\}_{n,k\in\mathbb{N}}$ is independent." can be dropped. Since one can take $\alpha_k $ to be the first hitting time of ${X_k} $ with some set, it is not suitable to assume this family is independent.
If yes, how to prove the theorem without this assumption?
If no, is there any counter example?

1

There are 1 best solutions below

0
On BEST ANSWER

If the sequence $\{X_n\}$ is i.i.d., and if the $\alpha_k$ are stopping times of the filtration $\mathcal F_n:=\sigma\{ X_k: k=1,2,\ldots,n\}$, $n\ge 1$, then your argument as it stands shows that $\{X_{\alpha_k+1}\}_{k=1}^\infty$ is i.i.d. For example, in considering $$ P(X_{m+1} \in B_N,\alpha_N=m,\Lambda_{N-1}) $$ in your calculation, for $m\ge N$, the event $\{\alpha_N=m\}\cap \Lambda_{N-1}$ is an element of $\mathcal F_m$, hence independent of $X_{m+1}$. Therefore $$ P(X_{m+1} \in B_N,\alpha_N=m,\Lambda_{N-1})=P(X_{m+1} \in B_N)\cdot P(\alpha_N=m,\Lambda_{N-1}), $$ as required.