Let $\xi_1,\xi_2,\ldots$ be i.i.d. positive random variables with infinite mean $\mathbb E[\xi_i]=\infty$, and consider the random walk $$T_n=\sum_{i=1}^n\xi_i,\qquad n\in\mathbb N.$$ Here's an example that illustrates why I'm interested in such questions.

Example. If $(S_n)_{n\in\mathbb N}$ is a simple symmetric random walk on $\mathbb Z$, then $\xi_1,\xi_2,\ldots$ could represent the time intervals between consecutive returns to zero of $S_n$.

Q. Does there exist general local limit theorems for the random walk $T_n$?

In the case of the example I have above, everything can be computed explicitly quite nicely, and it can be argued that $$n\cdot \mathbb P\{T_{\sqrt{n}x}=n\}\asymp xe^{-x^2/2}.$$ I'm interested in more general statements of this form.

  • 1
    $\begingroup$ browse petrov "sums of independent random variables". I don't know what you will find there but it is a good source for local limit theorems. $\endgroup$
    – user83457
    Commented Jul 21, 2017 at 14:41

1 Answer 1


Yes, there is a general result, see Chapter 9 of Gnedenko-Kolmogorov book.

The theorem says that if $\xi_i$ are i.i.d with values in $\mathbb{Z}$ such that $$\text{gcd}\{s-s':\mathbb{P}(\xi_1=s)>0,\;\mathbb{P}(\xi_1=s')>0\}=1,$$ and $b_n^{-1}(T_n-a_n)$ converges in distribution to a stable law, then also the local limit theorem holds, namely, $$ b_n\mathbb{P}(T_n=k)-g(b_n^{-1}(k-a_n))\to 0 $$uniformly in $k$, where $g$ is the density of the stable law.

There is also a necessary and sufficient condition for the existence of $b_n$ and $a_n$ such that $b_n^{-1}(T_n-a_n)$ converges in distribution, see e. g. Feller's or Durrett's book.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.