I saw an example in a book where $X_n \sim^{iid} Bern(\frac{1}{n})$. The book claims that since $\sum_{n=1}^{\infty}P(\{X_n = 1\}) = \infty$, the event $\{X_n = 1\}$ happens infinitely often with probability 1 (by Borel-Cantelli). Why can't $X_n \overset{a.s.}\to 1$? It seems very obvious to me but I cannot get it to work by the book's definition of almost sure convergence, that $X_n \overset{a.s.}\to X$ means:

$$ P(\{\omega \in \Omega: \lim_{n \to \infty}X_n(\omega) = X(\omega)\})=1. $$

Is there a way to show this using this very definition of almost sure convergence? Thanks!

  • 1
    $\begingroup$ Are we assuming the $X_n$ are independent? $\endgroup$ – Math1000 Nov 29 '15 at 4:02
  • 2
    $\begingroup$ @Math1000 Yes, sorry to not have mentioned that! $\endgroup$ – user136503 Nov 29 '15 at 4:15
  • 1
    $\begingroup$ Perhaps the simplest approach: the series $\sum P(X_n=0)$ and $\sum P(X_n=1)$ both diverges and $(X_n)$ is independent hence $\limsup\{X_n=0\}$ and $\limsup\{X_n=1\}$ are both almost sure, that is, almost surely $X_n=0$ for infinitely many $n$ and $X_n=1$ for infinitely many $n$. Thus $(X_n)$ diverges almost surely. $\endgroup$ – Did Nov 29 '15 at 8:12
  • 2
    $\begingroup$ Or you could notice that $X_n\overset{p}\to 0$ and hence $X_n$ can't converge a.s. to anything other than $0$ - which it doesn't. $\endgroup$ – A.S. Nov 29 '15 at 9:58

Assuming the $X_n$ are independent, then it follows from the second Borel-Cantelli lemma that $$\mathbb P\left(\limsup_{n\to\infty} \{X_n=1\}\right)=1. $$ (See for example here for a proof of the Borel-Canelli lemmas.) However, $$\mathbb P\left(\liminf_{n\to\infty}\{X_n=1\} \right) = \mathbb P\left(\bigcup_{n=1}^\infty\bigcap_{k=n}^\infty \{X_k=1\}\right)=0, $$ since for any $n$, $$\mathbb P\left(\bigcap_{k=n}^\infty \{X_k=1\} \right)=\prod_{k=n}^\infty\mathbb P(X_k=1)=\prod_{k=n}^\infty\frac1k=0 $$ and hence $$P\left(\bigcup_{n=1}^\infty\bigcap_{k=n}^\infty \{X_k=1\}\right)\leqslant \sum_{n=1}^\infty\mathbb P\left(\bigcap_{k=n}^\infty \{X_k=1\}\right)=0. $$ Since $$\mathbb P\left(\liminf_{n\to\infty} \{X_n=1\}\right) \ne \mathbb P\left(\limsup_{n\to\infty}\{X_n=1\}\right), $$ it is clear that $$\mathbb P\left(\lim_{n\to\infty}\{X_n=1\}\right) $$ does not exist, and so $X_n$ does not converge almost surely to $1$.

By similar computations we find that $$\mathbb P\left(\liminf_{n\to\infty}\{X_n=0\}\right)=0<1=P\left(\limsup_{n\to\infty}\{X_n=0\}\right), $$

so $X_n$ does not converge almost surely to $0$.

Footnote: Recall that

$$X_n\stackrel{\mathrm{a.s.}}\longrightarrow X\iff \mathbb P\left(\liminf_{n\to\infty} \{|X_n-X|<\varepsilon\}\right)=1$$

for all $\varepsilon>0$.

Since in this case $\mathbb P(X_n\in\{0,1\})=1$ for all $n$, taking $\varepsilon<\frac12$ justifies the use of e.g. $\limsup_{n\to\infty}\{X_n=1\}$ as opposed to $\{\limsup_{n\to\infty} X_n\}=1$.

  • $\begingroup$ What does $\lim\{X_n=1\}$ have to do w/ $[\lim X_n]=1$? $\endgroup$ – BCLC Nov 29 '15 at 4:27
  • $\begingroup$ @BCLC What exactly do you mean by $\lim_{n\to\infty} X_n$? $\endgroup$ – Math1000 Nov 29 '15 at 4:37
  • $\begingroup$ Math1000, $\lim X_n = \limsup X_n = \liminf X_n$...we had this argument w/ Did remember? $\endgroup$ – BCLC Nov 29 '15 at 4:37
  • 2
    $\begingroup$ $$X_n\stackrel{\mathrm{a.s.}}\longrightarrow X\iff \mathbb P\left(\liminf_{n\to\infty} \{|X_n-X|<\varepsilon\}\right)=1 $$ for all $\varepsilon>0$. Since in this case $\mathbb P(X_n\in\{0,1\})=1$ there is no distinction to be made; take e.g. $\varepsilon<\frac12$. $\endgroup$ – Math1000 Nov 29 '15 at 4:57
  • $\begingroup$ Math1000, uggghhhh so what's wrong w/ my argument? That seems to be consistent contrapositive of #4 there, w/c says merely $\ge$... $\endgroup$ – BCLC Nov 29 '15 at 4:59

By the second Borel-Cantelli lemma we have that $X_n=1$ infinitely often with probability 1, as the book states.

The definition given says that convergence occurs a.s. if the probability that the limit equals $0$ is $1$. This means that the for every $\epsilon$, there exists an $N$ such that $n>N$ implies $|X_n(\omega)|<\epsilon$ for a.e. $\omega$.

But we know that $X_n(\omega)=1$ infinitely often for a.e. $\omega$, and hence you cannot find any such $N$. So in fact the relevant probability is $0$ and not $1$, and so the a.s. convergence to $0$ does not occur.

Since $\sum_{n=1}^\infty(1-\frac1n) = \infty$ also, this means that $X_n=0$ infinitely often as well. Then you can repeat the above argument to conclude that $X_n$ does not converge a.s. to $1$.

  • $\begingroup$ This is similar to the argument of Math1000. What is the relationship b/w $\lim [X_n = 1]$ and $[\lim X_n] = 1$? $\endgroup$ – BCLC Nov 29 '15 at 4:52
  • $\begingroup$ What is the meaning of $\lim [X_n=1]$? Is it the limit of $\mathrm{Pr}(X_n=1)$? $\endgroup$ – Milind Nov 29 '15 at 4:54
  • $\begingroup$ Milind, $\lim[X_n = 1] = \liminf[X_n = 1] = \limsup[X_n = 1]$. However, $\lim X_n = \liminf X_n = \limsup X_n$ $\endgroup$ – BCLC Nov 29 '15 at 4:57
  • $\begingroup$ But could you explain the notation $\lim [X_n=1]$ (or $\liminf, \limsup$ of the same)? I have not seen it before. $\endgroup$ – Milind Nov 29 '15 at 4:59
  • $\begingroup$ Milind, see my answer. Essentially, you and Math1000 seem to be arguing that $\omega \in \lim[X_n = 1] \to \lim X_n(\omega) = 1$ $\endgroup$ – BCLC Nov 29 '15 at 5:07

$$\liminf X_n$$

$$ = \sup [\inf_{n \ge 1} X_n, \inf_{n \ge 2} X_n, ... ]$$

$$ = \sup [0, 0, ... ] = 0$$

$$\limsup X_n$$

$$ = \inf [\sup_{n \ge 1} X_n, \sup_{n \ge 2} X_n, ... ]$$

$$ = \inf [1, 1, ... ] = 1$$


$$\limsup X_n \ne \liminf X_n,$$

$\lim X_n$ does not exist.

Note that the reason we say that $\inf_{n \ge m} X_n = 0 \ \forall m \ge 1$ is $X_n = 0 \ \text{i.o.}$ because $X_n = 0 \ \text{i.o.}$ means that $\forall m \ge 1, \exists n \ge m$ s.t. $X_n = 0$.

///ly, the reason we say that $\sup_{n \ge m} X_n = 1 \ \forall m \ge 1$ is $X_n = 1 \ \text{i.o.}$ because $X_n = 1 \ \text{i.o.}$ means that $\forall m \ge 1, \exists n \ge m$ s.t. $X_n = 1$.

  • $\begingroup$ Not an answer. This takes for granted that $X_n=0$ infinitely often, the fact the OP is confused with. $\endgroup$ – Did Nov 30 '15 at 6:46
  • $\begingroup$ @Did What do you mean by 'takes for granted'? user136503 is asking why it is not the case that almost surely $X_n \to 1$. My answer is that $\lim X_n$ does not exist based on facts 1. $X_n = 0$ i.o. 2. $X_n = 1$ i.o. How is this not an answer? $\endgroup$ – BCLC Nov 30 '15 at 11:51

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.