Let $$X_n=\begin{cases} 1/n, &\text{ with prob } 1-1/n\\ n^2, &\text{ with prob } 1/n.\end{cases}.$$ $$Y_{n}=\begin{cases} 0, &\text{ with prob } 1-1/n\\ n, &\text{ with prob } 1/n.\end{cases}.$$

We want to show that $X_n \rightarrow 0$ in probability, which means showing that $\lim_{n\rightarrow\infty}P(|X_n-0|>\epsilon)=0$ for all $\epsilon>0$. And similarly for showing $Y_n \rightarrow 0$ in probability.

The solutions I've seen to these questions go as follows:

Whenever $\epsilon>1/n$, we have $0\leq P(|X_n-0|>\epsilon)=P(X_n>1/n)=P(X_n=n^2)=1/n$. The result follows by letting $n\rightarrow \infty$ and using the Sandwich theorem.

And for the second one (https://youtu.be/x4q6H6lxFFE?t=594):

Whenever $\epsilon<n$, we have $0\leq P(|Y_n-0|\geq \epsilon)=1/n$. The result follows by letting $n\rightarrow \infty$ and using the Sandwich theorem.

Simple enough, but why are we allowed to put conditions on $\epsilon$? I thought $\epsilon$ was just any positive number. Also even though the examples are really similar, in one case it's $\epsilon>1/n$ and in the other it's $\epsilon<n$. Is that just because one definition allows for $> \epsilon$ and the other is $\geq\epsilon$?

Edit: On second glance I'm confused as to why the second proof has $\epsilon<n$? Shouldn't it just be $\epsilon>0$ as then $P(|Y_n|>\epsilon)=P(Y_n>0)=P(Y_n=n)=1/n$?


This is not a condition to $\epsilon$. This is a condition to $n$. Read it as follows:

Whenever $n>1/\epsilon$, we have $$0\leq P(|X_n−0|>\epsilon)=P(X_n>1/n)=P(X_n=n^2)=1/n$$

For any $\epsilon>0$ and for $n\to\infty$ the moment when $n$ became greater than $1/\epsilon$ will definitely come. Moreover, since we find limits, we are not interested in the behaviour of first elements in the sequence of $P(|X_n−0|>\epsilon)$. We can only check how this probability looks like for sufficiently large $n$.

For the second question: Sure, $\epsilon>0$ by definition of convergence in probability, and for the case when the values of $Y_n$ are either zero or $n$, to have $Y_n>\epsilon$ iff $Y_n=n$ you should have $n>\epsilon$.

  • $\begingroup$ Thanks a lot! And ah, I had just edited my post before your reply when I realised $\epsilon>n$ would make no sense. Can we not just have the $\epsilon>0$ condition though as in my last line? $\endgroup$ – Alex.F Mar 26 '17 at 1:16
  • $\begingroup$ I edited the answer correspondingly. $\endgroup$ – NCh Mar 26 '17 at 1:18
  • $\begingroup$ Am I right in thinking $n>\epsilon$ and $\epsilon>0$ are equivalent conditions in the $Y_n$ case then? $\endgroup$ – Alex.F Mar 26 '17 at 1:28
  • $\begingroup$ None. Look at answer: $P(Y_n>\epsilon)=P(Y_n=n)$ iff $n > \epsilon$. $\endgroup$ – NCh Mar 26 '17 at 1:33
  • 1
    $\begingroup$ You are right. The author of the solution subconsciously assumed that $\epsilon$ is small, and for this case $n>1/\epsilon$ is autometically greater than $\sqrt{\epsilon}$. The reason to do so exists: if you prove that $P(|X_n-0|>\epsilon)\to 0$ for small $\epsilon$, it then holds for large $\epsilon$ due to monotonicity of probability. $\endgroup$ – NCh Mar 26 '17 at 2:38

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.