This is a multipart question; if there's a strong preference for breaking this into separate questions I'll do that.

Imagine a game between a gambler and a croupier.

Total capital in the game is $1$.

After $n^{\text{th}}$ hand of the game, the fraction of capital held by the gambler is $X_{n}\in[0,1]$, and capital held by croupier is $1-X_{n}$.

Assume $X_{0}=p\in(0,1)$.

Rules of the game: After $n$ hands, probability for gambler to win $(n+1)^{\text{th}}$ hand is $X_{n}$. If he wins he gets half the capital the croupier held after the $n^{\text{th}}$ hand. If he loses, he gives the croupier half of his own capital. Let $$\mathcal{F}_{n}=\sigma\left(X_{i},1\leq i \leq n\right)$$


How do I show that $\left(X_{n}\right)_{n\geq 0}$ is a $\left(\mathcal{F_{n}}\right)_{n\geq 0}$ martingale, converging to a limit $Z$ both a.s and in $L^2$?

Also I need to show that: $$\mathbb E[X_{n+1}^{2}]=\frac{\mathbb E[3X_{n}^2+X_n]}{4}$$

and $$\mathbb E[Z^2]=\mathbb E[Z]=p$$

For any $n\geq 0$, let $Y_n=2X_{n+1}-X_{n}$ Find conditional law of $X_{n+1}$ knowing $\mathcal{F}_{n}$, and prove that $$\mathbb P(Y_n=0|\mathcal{F_n})=1-X_n$$

$$\mathbb P(Y_n=1|\mathcal{F}_n)=X_n$$

And express the law of $Y_n$.

Let $G_n=\{Y_n=1\}$, $P_n=\{Y_n=0\}$ prove $Y_n\rightarrow Z$ a.s., and deduce that $$\mathbb P(\liminf_{n\rightarrow\infty}G_n)=p$$

$$\mathbb P(\liminf_{n\rightarrow\infty}P_n)=1-p$$

Lastly, are the variables $\{Y_n,n\geq 1\}$ independent? I would appreciate any and all solutions or hints.


You've asked for an exhaustive treatment of this martingale. I've attempted below to get it all.

There is a preliminary point to make before pursuing any properties. Let $A_n$ denote the event that the gambler wins the $n$-th hand. The way in which I interpret your rules is $$ X_{n+1} = \frac{1}{2}(1 + X_n) ~~ \text{on }A_{n+1}\\ = \frac{1}{2} X_n ~~ \text{on }A_{n+1}^c $$ and $P(A_{n+1} \mid \mathcal{F}_n) = X_n$. This last line is important for obvious reasons, and says literally that the gambler wins on hand $n+1$ with likelihood $X_n$, when $X_n$ is known.

  1. $X_n$ is a martingale. We compute $$ E[X_{n+1} \mid \mathcal{F}_n] = E[X_{n+1} I_{A_{n+1}} \mid \mathcal{F}_n] + E[X_{n+1} I_{A_{n+1}^c} \mid \mathcal{F}_n] \\ = E[\frac{1}{2}(1 + X_n) I_{A_{n+1}} \mid \mathcal{F}_n] + E[\frac{1}{2}X_n I_{A_{n+1}^c} \mid \mathcal{F}_n] \\ = \frac{1}{2}(1 + X_n) P(A_{n+1} \mid \mathcal{F}_n) + \frac{1}{2} X_n P(A_{n+1}^c \mid \mathcal{F}_n) \\ = \frac{1}{2}(1 + X_n) X_n + \frac{1}{2} X_n(1 - X_n) = X_n $$ where from the second to third lines I used the rule 'taking out what is known'. This is the martingale property, as desired.

2.$E[X_{n+1}^2] = \frac{E[3 X_n^2 + X_n]}{4}$.

I'll prove this in a stronger form: $$ E[X_{n+1}^2 \mid \mathcal{F}_n] = E[X_{n+1}^2 I_{A_n} \mid \mathcal{F}_n] + E[X_{n+1}^2 I_{A_n^c} \mid \mathcal{F}_n] \\ = E[\frac{1}{4} (1 + X_n)^2 I_{A_n} \mid \mathcal{F}_n] + E[\frac{1}{4} X_n^2 I_{A_n^c} \mid \mathcal{F}_n] \\ = \frac{1}{4} (1 + X_n)^2 X_n + \frac{1}{4} X_n^2(1 - X_n) \\ = \frac{3 X_n^2 + X_n}{4} $$ From the second to the third lines, I took out what is known and plugged in $P(A_{n+1} \mid \mathcal{F}_n) = X_n$ in one shot. The result you want now follows from taking expectations.

3.$X_n$ converges almost surely and in $L^2$ to a random variable $Z$ for which $E[Z] = E[Z^2] = p$.

Note that your martingales are nonnegative and bounded from above by $1$. Therefore the MG convergence theorem says that $X_n$ converges in $L^2$ and almost surely. This is well-known and general and I won't discuss this theorem any further, except to point you to wikipedia.

Now it is also the case that $X_n$ is uniformly integrable, being bounded uniformly in $n$ by the constant function $1$. Therefore we use the MCT and the bounded convergence theorem to get $$ E[X_n] \rightarrow E[Z] $$ However $X_n$ is an MG, hence $E[X_n] = p$ for all $n$, and $E[Z] = p$ as desired.

To show that $E[Z^2] = p$ we'll have to be more clever. Iterating the identity I proved in the answer to 2, it is possible to show that for all $m < n$, $$ E[X_n^2 \mid \mathcal{F}_m] = (\frac{3}{4})^{n-m} X_m^2 + \frac{1}{4} \left( 1 + \frac{3}{4} + \cdots + \left( \frac{3}{4}\right)^{n-m-1}\right) X_m $$ First, let $m = 0$, and second, take the limit as $n \rightarrow \infty$. We obtain that $$ E[Z^2] = \lim_n E[X_n^2] = 0 + \frac{p}{4} \sum_{k = 0}^{\infty} \left(\frac{3}{4} \right)^k = \frac{p}{4} \frac{4}{1} = p $$ using the $L^2$ convergence of the MCT (or bounded convergence, depending on your mood today).

4.With $Y_n = 2 X_{n+1} - X_n$, find the law of $Y_n$, show that $Y_n \rightarrow Z$ almost surely, and compute $P(\liminf_n G_n) = p, P(\liminf_n P_n) = 1-p$.

You should realize by now that $\{Y_n = 1\} = A_{n+1}$ is precisely the event that the Gambler has won the $n+1$-th hand. To find the law, recall that $$ P(A_{n+1} \mid \mathcal{F}_n) = X_n \\ P(A_{n+1}) = E[X_n] = p $$ by martingality. Thus is the law of $Y_n$: a biased coin toss with likelihood $p$.

That $Y_n \rightarrow Z$ almost surely follows immediately from $X_n \rightarrow Z$.

Notice now that $Y_n$ is $\{0,1\}$-valued, and so $Z$ is almost surely $\{0,1\}$ valued as well. Realize also that $\liminf_n G_n$ is the event "always winning after a certain time" and $\liminf_n P_n$ is the event "always losing after a certain time", so we see that $$ \liminf_n G_n \subset \{Z = 1\} \\ \liminf_n P_n \subset \{Z = 0\} $$ Now we'll do something cheap: $$ E[Z] = p = 1 \cdot P(Z = 1) + 0 \cdot P(Z = 0) = P(Z = 1) $$ Now what remains is to argue that $\liminf_n G_n$ has full measure inside $\{Z = 1\}$. To see this, think about what it would mean if on a positive measure set inside $\{Z = 1\}$ we had $Y_n = 0$ infinitely often. This last part will finish item 4.

5.Are the $Y_n$ independent?

No, I don't think so. This is a bit counterintuitive, but that's life. You can compute $E[Y_0Y_1]$ by hand and see that it's not equal to $p^2$ by considering $E[Y_0Y_1 \mid \mathcal{F}_1]$ and then computing an expectation.

  • $\begingroup$ Thank you so much for the incredibly exhaustive and well thought out response! $\endgroup$ – Chris May 10 '13 at 0:27

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.