1
$\begingroup$

I am familiar with the usual CLT's, but I was wondering whether there would be a corresponding CLT or for the following sort of sum of random variables:

Let $\{u_t\}_{t\in\mathbb{N}}\sim $ iid $(0,\sigma^2)$. Then let us observe that $\forall j=1,...,n-1$ the random variable $u_n^2$ is independent of $u_j^2$ and thus: \begin{equation*} \mathbb{E}\left[ u_nu_j\right] =\mathbb{E}\left[ u_n\right] \mathbb{E}\left[ u_j\right] =0 \qquad \text{ and } \qquad \text{Var}(u_nu_j) =\mathbb{E}\left[ u_n^2u_j^2\right] = \mathbb{E}\left[ u_n^2 \right]\mathbb{E}\left[u_j^2\right] = \sigma^4 \end{equation*} So that the sequence of random variables $\{u_{n+1}u_j\}_{j=1}^n$ for each $n\in\mathbb{N}$ is identically distributed with mean 0 and variance $\sigma^4$ but the sequence of random variables across $j$ is dependent. Then does the sum of random variables: \begin{equation*} \sum_{j=1}^{n} u_{n+1}u_j \quad \overset{\mathcal{D}}\longrightarrow \quad??? \end{equation*}

i.e. After some appropriate normalisation/rescaling converge to something meaningful? (if it does at all...)as $n\rightarrow\infty$)

Thanks for any help in advanced!

$\endgroup$
1
$\begingroup$

A normalization is required. Let $X$ be a random variable independent of the sequence $\left(u_i\right)_{i\in\mathbb N}$ and having the same distribution as $u_0$ (enlarging the probability space if needed). Then for any positive integer $n$, the random variable $Y_n$ and $Z_n$ defined by $$Y_n=u_ {n+1 } \frac 1 { \sqrt n} \sum_{i=1}^nu_i \mbox{ and }Z_n=X \frac 1 { \sqrt n} \sum_{i=1}^nu_i$$ have the same distribution. The sequence $ \left(Z_n\right)_{n\geqslant 1}$ converges in distribution to $X\cdot N$, where $N$ is a zero-mean Gaussian random variable with variance $\sigma^2$.

$\endgroup$
  • $\begingroup$ Hi Davide, Thank you for this, reply. I definitely agree with this, and I am gathering that your answer implies that \begin{equation*} Y_n \overset{\mathcal{D}}\longrightarrow X\cdot N \end{equation*} Do you think though, there is any way of getting a more "exact/known" distribution of what this sum may possibly converge to? Or do you think with this limited/given information this would be the best asymptotic distribution that one may obtain? $\endgroup$ – user405613 Jan 10 '17 at 18:47
  • $\begingroup$ The reason being why I ask is that this "problem" I have asked is but a small part of a larger problem which I am trying to solve (and I eventually need to obtain an (ideally) pivotal asymptotic distribution of an estimator which would rely on not having any unknown parameters in it)! $\endgroup$ – user405613 Jan 10 '17 at 18:47
  • $\begingroup$ You can get the characteristic function of the limiting distribution, given by $t\mapsto \mathbb E\left[\exp\left(-t^2\sigma^2u_0^2/2\right)\right]$. $\endgroup$ – Davide Giraudo Jan 11 '17 at 9:28
0
$\begingroup$

If you rescale with $\sqrt{n},$ you get $u_{n+1}\frac{1}{\sqrt{n}}\sum_{j=1}^nu_j.$

$u_{n+1}$ is from it's distribution and $\frac{1}{\sqrt{n}}\sum u_j$ converges in distribution to an independent normal so their distribution is whatever the product of independent samples from a normal and the parent distribution is.

$\endgroup$
  • $\begingroup$ Unfortunately its not quite that simple...$u_{t+1}$ is only iid $(0,\sigma^2)$ we have no normality assumption on it! $\endgroup$ – user405613 Jan 10 '17 at 6:51
  • $\begingroup$ @user405613 oh ok, misread. Then it's like whatever the distribution is times a normal. $\endgroup$ – spaceisdarkgreen Jan 10 '17 at 15:58
  • $\begingroup$ Thanks for this reply. Yes I agree. $\endgroup$ – user405613 Jan 10 '17 at 18:48

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.