4
$\begingroup$

Let $X \sim N(0, \Sigma)$ be a $d$-dimensional Gaussian random vector. Suppose I have $n$ samples $X_1, \ldots, X_n$.

Consider the large sample regime where $d$ is fixed and $n$ goes to infinity, the MLE estimator of $\Sigma^{-1}$ is $\hat{\Sigma}^{-1} = (\frac{1}{n}\sum_{i=1}^nX_iX_i^\top)^{-1}$.

My question is how to establish the consistency result that $\hat{\Sigma}^{-1} \rightarrow \Sigma^{-1}$ and the asymptotic distribution of $\hat{\Sigma}^{-1} - \Sigma^{-1}$.

Thank you!

| cite | improve this question | | | | |
$\endgroup$
1
$\begingroup$

Hint: Look at each entry of $X_i X_i^T$, then, look at their average. Then, use the Law of large numbers. Finally, use the continuous mapping theorem.

| cite | improve this answer | | | | |
$\endgroup$
  • $\begingroup$ Thank you! But how about the asymptotic normality part? $\endgroup$ – Wuchen Aug 18 '16 at 21:35
  • $\begingroup$ @Wuchen Asymptotic normality is trickier. You need to prove a couple of additional things. For instance, you can use some general results from M-estimation theory (see Theorem 5.23 from van der Vaart's Asymptotics Statistics book) $\endgroup$ – Luck Aug 18 '16 at 23:30
  • $\begingroup$ Thank you for the pointer! I will check that. So there is no exsiting results on the asymptotic normality of the inverse covariance matrix in the large sample regime? A bunch of research has been working on the high dimensional regime. $\endgroup$ – Wuchen Aug 19 '16 at 1:47

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.