I have a question like this:

Let $(X_n)$ be an i.i.d sequence of random variables with values in $\{-1,1\}$, and define $Y_n:= \sum_{i \leq n} X_i/n$. Show that $(Y_n)$ converges almost surely and in $L_1$. Do not appeal to the law of large numbers.

The question is under the classification of martingale. I know we should be able to do this using backwards martingales. but we never learned the backwards martingales in class and can we do it using regular martingales? I mean, can we define a suitable filtration so that the $(Y_n)$ is indeed a martingale?

  • $\begingroup$ $(nY_n)$ is definitely a martingale $\endgroup$ – gt6989b May 5 '16 at 1:29
  • $\begingroup$ @gt6989b If $\mathbb P(X_1=1)=\mathbb P(X_1=-1)$, yes. Otherwise it is slightly more involved. $\endgroup$ – Math1000 May 5 '16 at 1:41
  • $\begingroup$ @Math1000 yes, otherwise use $Z_n = nY_n - n\mu_X$ where $\mu_X = \mathbb{E}[X]$... $\endgroup$ – gt6989b May 5 '16 at 1:52
  • $\begingroup$ @gt6989b Hence the inclusion of "slightly" :) $\endgroup$ – Math1000 May 5 '16 at 1:55
  • $\begingroup$ You can use Kronecker's lemma applied to the sequence $\sum_{i=1}^n \frac{X_i-\mu}{i}$. $\endgroup$ – Kolmo May 14 '16 at 0:28

I do not know how one can use the particular Rademacher distribution, but there is a useful reverse martingale trick: it is easy to see that $$ E[X_1|S_n, S_{n+1},\dots] = \frac{S_n}{n}. $$ Denoting $\mathcal{F_n} = \sigma\{S_n, S_{n+1},\dots\}$, we thus have a (reverse) Levy martingale $Y_n = E[X_1|\mathcal{F_n}]$. By the reverse martingale convergence, $Y_n\to Y_\infty = E[X_1|\mathcal{F_\infty}]$, $n\to\infty$, where $\mathcal{F}_\infty = \cap_{n\ge 1} \mathcal F_n$. But $\mathcal{F}_\infty$ consists of exchangeable events, which by the Hewitt-Savage 0-1 law have probability $0$ and $1$. Therefore, $Y_\infty = E[X_1]$ a.s.


If the sequence $\left(Y_l-c_l\right)_{l\geqslant 1}$ is a martingale with respect to $\left(\mathcal F_n\right)_{l\geqslant 1}$ (where $c_l$ is a deterministic constant), then for each $l$, the random variables $X_1,\dots,X_l$ are measurable with respect to $\mathcal F_n$. Since we should have that $\mathbb E\left[Y_{n+1}-c_{n+1}+c_n-Y_n\mid\mathcal F_n\right]=0$, we would obtain that $$\left(\frac 1{n+1}-\frac 1n\right)\sum_{j=1}^nX_j+\frac{\mathbb E\left[X_1\right]}{n+1}+c_n-c_{n+1}=0 \mbox{ a.s.}$$ which is not possible.

The idea is to introduce a martingale which is bounded in $\mathbb L^2$. Define for $l\geqslant 1$ the random variable $$Z_l:=\sum_{i=1}^l\frac{X_i-\mathbb E\left[X_1\right]}i.$$ Then the sequence $\left(Z_l\right)_{l\geqslant 1}$ is a martingale with respect to its natural filtration. Now, we can get the wanted result using the following facts:

  • the sequence $\left(\mathbb E\left[Z_l^2\right]\right)_{l\geqslant 1}$ is bounded.
  • The sequence $\left(Z_l\right)_{l\geqslant 1}$ is thus convergent to some $Y$ in $\mathbb L^1$ and almost surely. Now, $$\frac 1n\sum_{i=1}^nX_i=Z_n-\frac 1n\sum_{l=1}^{n-1}Z_l+\mathbb E\left[X_1\right].$$

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.