0
$\begingroup$

Let $X_n$ be independent Poisson random variables with $E[X_i] = \mu_i$, and let $Y_n = X_1+...+X_n$. I want to show that if $\sum_n \mu_n = \infty $ then $Y_n/E[Y_n] \rightarrow 1$ almost surly.
What I do know is that if $X_1,...$ are independent, and $E\left[X^4\right] < \infty$, then $Y_n/n \rightarrow E[X]$ a.s.* So What I tried is to center the random variables: let $C_n := X_n-\mu_n$. The $C$'s are independent, $E[C_n^4]< \infty, E[C_n] = 0$, which means:
$\frac{C_1 +...+C_n}{n} \rightarrow 0$ a.s., or $\frac{\sum X_i - \sum \mu_i}{n} \rightarrow 0$ a.s. which, I think, completes the proof.

However I am not sure why the requirement $\sum_n \mu_n = \infty $ is necessary, so I suspect this proof is incorrect. Is it?

Edit: This argument is incorrect since the theorem requires $E[C^4]$ to be uniformly bounded, not just bounded. Any ideas?

$\endgroup$
  • 1
    $\begingroup$ You cannot apply the result you recall since it requires the increments are identically distributed -- which is not true in your case (except if $\mu_n$ does not depend on $\mu$). $\endgroup$ – Did Dec 16 '18 at 13:04
  • $\begingroup$ To see why the $\sum_n\mu_n=\infty$ requirement is needed consider what happens if $\mu_1=1$ and $\mu_n=0$ for $n>1$. $\endgroup$ – kimchi lover Dec 16 '18 at 13:07
  • $\begingroup$ @Did - The result I cited does not require i.d. See "Probability with martingales" Chapter 7 $\endgroup$ – SomeoneHAHA Dec 16 '18 at 13:15
  • $\begingroup$ Except you misquoted heavily. $\endgroup$ – Did Dec 16 '18 at 13:17
  • $\begingroup$ @kimchilover If $\mu_n = 0$ it would mean that $X_n = 0$, so I am not sure where is the problem $\endgroup$ – SomeoneHAHA Dec 16 '18 at 13:17
0
$\begingroup$

If, in your argument, $M=\sum_n \mu_n<\infty$ then your $Y_n$ converge in distribution to a Poisson rv. with expectation $M$, and your ratio $Y_n/EY_n$ has a non trivial limit distribution, violating a.s. convergence to a constant.

And, the theorem you quote needs a uniform 4th moment bound $EX_k^4\le K$, which need not hold given your hypotheses.

A better way to prove your result might be to write your $Y_n$ as $Y_n=Y(M_n)$, where $M_n=\sum_{t\le n} \mu_t$ and where $Y(t)$ is a Poisson process, as described in the wikipedia article. You might be able to show $Y(t)/t\to 1$ almost surely.

Alternatively, you could use your argument with the added hypothesis that $\mu_n\le1$ for all $n$ and then use that to prove the general result by representing each $\mu_n$ as a sum of quantities that are individually bounded by $1$, and your corresponding $Y_n$ as sums of independent Poissons whose expectations are bounded and add up to the right thing.

$\endgroup$
  • $\begingroup$ I see. So: the overall argument is correct, except the punch: $C_1+...+C_n /n \rightarrow 0$ implies the claim only if $M = \infty$, or should I change the entire argument? The comments in the original question made me completely not sure about is correctness. Thanks! $\endgroup$ – SomeoneHAHA Dec 16 '18 at 13:38
  • $\begingroup$ Re the edit: I see now that my argument is not true due to the uniform bounded requirement. Do you think that there is a proof which is not going through Poisson process? $\endgroup$ – SomeoneHAHA Dec 16 '18 at 13:49
  • $\begingroup$ I suppose there could be a non-Poisson process proof, but it would probably be ugly, like an apparatus designed by a lawyer to not infringe on a particular patent. $\endgroup$ – kimchi lover Dec 16 '18 at 13:58
  • $\begingroup$ Thanks. May you elaborate more on this way of proof? I am not familiar with Poisson processes, so I read about it but still have no idea how to use it here $\endgroup$ – SomeoneHAHA Dec 16 '18 at 14:55

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.