Suppose $\{X_i,i\ge 1\}$ is a sequence of i.i.d. random variables of exponential distribution with mean 1. Let $M_n=max_{i=1,\cdots,n}X_i$ and $Z_n=M_n-\ln n$. It is not hard to see $Z_n$ converges to $Z_\infty$ in distribution, where $P(Z_\infty\le x)=e^{-e^{-x}}$. And we need to show whether or not $Z_n$ converges to some limiting r.v. almost surely.

My idea is the following: since the distribution of $Z_\infty$ is continuous, then suppose $Z_n$ converges to some r.v. a.s., then the limiting r.v. should be $Z_\infty$. I want to show actually $Z_n$ does not converge to $Z_\infty$ in probability, then we will get contradiction.
Then pick fixed $x,\epsilon>0$, $P(|Z_n-Z_\infty|>\epsilon)\ge P(Z_\infty>x+\epsilon,Z_n\le x)$. But we don't know $Z_\infty$ is independent of $Z_n$. So is there any other idea to prove it?

  • $\begingroup$ I see... I revised the problem. $\endgroup$
    – Connor
    Jan 28 '17 at 4:59
  • $\begingroup$ In these situations, it's usually easier to show that $Z_n$ is not Cauchy in probability i.e. $P(|Z_n-Z_{m+n}| > \epsilon)$ does not go to zero for all $m\ge 1$ as $n\rightarrow \infty.$ Haven't done this problem before so not sure if it'll work here. (or if no is the right answer) $\endgroup$ Jan 28 '17 at 5:01
  • $\begingroup$ @spaceisdarkgreen Correct. When m is comparable to n, it will be done. $\endgroup$
    – Connor
    Jan 28 '17 at 5:18
  • $\begingroup$ I think I got it (see answer)... not sure its 100% rigorously correct and maybe there's an easier way. (Is that what you had in mind with $m$ comparable to $n$?) $\endgroup$ Jan 28 '17 at 6:07
  • $\begingroup$ Yes. Or we can pick a fixed $x$, then $P(|Z_n-Z_{2n}|>\epsilon)\ge P(Z_n\le x, |Z_{2n}|>x+\epsilon)$ and use the fact that $X_{n+1},...,X_{2n}$ is independent of the former ones to show the limit of the second term in the inequality is positive. $\endgroup$
    – Connor
    Jan 31 '17 at 0:56

Let $\epsilon>0$ be anything. Consider $Z_{2n}-Z_n = -\ln(2) + M_{2n}-M_n$ and look at $$P(|Z_{2n}-Z_n|> \epsilon)$$

We have $|Z_{2n}-Z_n| > \epsilon$ in the case that $M_{2n}-M_n > \ln(2) + \epsilon$ so$$P(|Z_{2n}-Z_n|> \epsilon) \ge P(M_{2n}-M_{n}> \ln(2)+\epsilon).$$

Let $$M_{2n,n} \equiv \max\{X_{n+1},\ldots,X_{2n}\}.$$ The only way we can have $M_{2n} -M_n >0$ is if $M_{2n,n}=M_{2n},$ so $M_{2n}-M_n > \ln(2) + \epsilon$ if and only if $M_{2n,n} - M_n > \ln(n)+\epsilon.$ This means $$ P(M_{2n}-M_{n}> \ln(2)+\epsilon) = P(M_{2n,n}-M_{n}> \ln(2)+\epsilon).$$

If we then define $Z_{2n,n} \equiv M_{2n,n}-\ln(n)$ and recall $Z_n=M_n-\ln(n),$ we can write $$P(|Z_{2n}-Z_n|> \epsilon) \ge P(M_{2n,n}-M_n>\ln(2)+\epsilon) = P(Z_{2n,n} - Z_n > \ln(2) + \epsilon). $$ But now notice that $Z_{2n,n}$ and $Z_n$ are independent by construction and in the limit of large $n$, they are Gumbel-distributed as you've shown. So we have $$\lim_{n\rightarrow\infty}P(|Z_{2n}-Z_n|> \epsilon) \ge P(A_1-A_2>\ln(2) + \epsilon)$$ where $A_1$ and $A_2$ are independent Gumbels, and it's pretty clear that $P(A_1-A_2>\ln(2) + \epsilon)>0$ since it's a continuous distribution with support on all the reals.

So this shows that $Z_n$ is not Cauchy in probability and therefore does not converge in probability.


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.