Im trying to undestand the proof of the next: Suppose $X_n$ are iid Poisson random variable with rate $\lambda>0.$ Prove that $$\displaystyle\limsup\frac{X_n\log(\log n)}{\log(n)}=1\space a.s.$$

The proof given is following: We have $P(X\geq n)\leq e^{\lambda}P(X=n).$

Let $a_n$ be the integer part of ${\delta \log n \over \log\log n}$ for a $\delta >0$. Then

$$P(X=a_n)={e^{-\lambda}\lambda^{a_n}\over a_n!}=e^{-\lambda}e^{a_n \log \lambda}e^{-\sum_{j=1}^{a_n}\log j}=e^{-a_n\log a_n (1+o(1))}=n^{-\delta +o(1)}$$

Then, $\sum_{n=1}^{\infty} P(X_n\geq a_n)<\infty$ or $=\infty$ depending upon $\delta >1$ or $\delta<1$. Borel-Cantelli lemma finishes the proof.

I'm stuck in this: $$e^{-\lambda}e^{a_n \log \lambda}e^{-\sum_{j=1}^{a_n}\log j}=e^{-a_n\log a_n (1+o(1))}=n^{-\delta +o(1)}.$$ I don't get such expressions; I have understood that little o-notation means a quotient between two functions have zero limit but I don't know how this was used here to get such expressions.

Any kind of help is thanked in advanced.


The notation $b_n=o(c_n)$ means $b_n/c_n\to0$ as $n\to\infty$. Therefore, $b_n=o(1)$ simply means $b_n\to0$. Thus, the first equality is asserting that $$ -\lambda + a_n\log\lambda - \sum_{j=1}^{a_n}\log j = -a_n\log a_n(1 + b_n), $$ for some sequence $\{b_n\}$ satisfying $b_n\to0$. This is equivalent to $$ \frac{\lambda - a_n\log\lambda + \sum_{j=1}^{a_n}\log j}{a_n\log a_n} \to 1. $$ The main thing needed to prove this is to show that $\sum_{j=1}^n \log j\sim n\log n$, meaning their ratio tends to $1$ as $n\to\infty$. You can guess at this result by comparing the sum to the corresponding integral $\int_1^n\log x\,dx$. To check it rigorously, you could use the Stolz–Cesàro theorem.

The second equality asserts that if $b_n\to0$, then $$ -a_n\log a_n(1 + b_n) = (-\delta + c_n)\log n $$ for some sequence $\{c_n\}$ satisfying $c_n\to0$. This is equivalent to \begin{equation}\label{1} \frac{a_n\log a_n}{\log n} \to \delta.\tag{1} \end{equation} A complete, rigorous proof of this would need to account for the fact that $a_n$ is an integer. But in this sketch, I will work as though $$ a_n = \frac{\delta\log n}{\log\log n}. $$ In this case, $$ \log a_n = \log\delta + \log\log n - \log\log\log n. $$ The leading term here is $\log\log n$ and $$ \frac{a_n\log\log n}{\log n} = \delta, $$ suggesting that \eqref{1} holds and verifying the second equality.

  • 1
    $\begingroup$ I really appreciate your help; that was very clear. $\endgroup$ – Suiz96 May 2 at 3:42

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.