3
$\begingroup$

The univariate version of the Lyapunov CLT (a version of the CLT for independent but not necessarily iid random variables) is as follows according to Wikipedia:

Lyapunov CLT from https://en.wikipedia.org/wiki/Central_limit_theorem#Multidimensional_CLT

Is there a multivariate version of the Lyapunov CLT (ideally one that doesn't require the means of the random variables to all be zero vectors)? I'm having a very hard time finding any information on this, even on Google Scholar.

$\endgroup$
4
  • $\begingroup$ Your last paragraph is a little puzzling, because it suggests you are interested in a sequence of random matrices--that, after all, is what "means ... be[ing] zero matrices" implies. Indeed, even in the univariate case what version of the CLT requires all variables to have zero means?? $\endgroup$
    – whuber
    Commented May 18, 2022 at 21:55
  • $\begingroup$ Sorry I meant zero vectors. I'll fix it. $\endgroup$
    – bob
    Commented May 18, 2022 at 22:23
  • 1
    $\begingroup$ I guess you are searching for this theorem: en.wikipedia.org/wiki/Berry%E2%80%93Esseen_theorem See the Multidimensional version. $\endgroup$ Commented Jan 8 at 8:53
  • $\begingroup$ I get a lot of results when I search for multivariate Lindeberg CLT. The univariate Lindberg condition is implied by the Lyapunov condition. $\endgroup$
    – Taylor
    Commented Jan 8 at 14:22

1 Answer 1

1
$\begingroup$

It seems to me that the multivariate version may follow from the Cramér-Wold theorem. You have to check that the conditions stated in your question are satisfied on every possible linear combination of components of $\mathbf X_i$, so this seems to require suitable behaviour of the cumulative variances $s_n(\mathbf a)$ defined as $s_{n}(\mathbf a) = \sum_{i=1}^n \mathrm{Var} (\mathbf a^T \mathbf X_i)$.

$\endgroup$

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.