Suppose that I have "a realization" of random vector $x=(x_1,\cdots,x_N)$ where $N$ is sufficiently large $N>100$. I know that random vector is joint normally distributed $$x \sim N(\mu,\Sigma), $$ where diagonal elements are $\sigma^2$ and off-diagonal elements are $\rho\sigma^2$.

I wonder how I could estimate $\mu,\Sigma$ when I only observe one sample. I guess $\mu$ and $\sigma^2$ can be estimated from a sample mean and variance, following the law of large numbers. I wonder if it is possible to estimate $\rho$.

  • 1
    $\begingroup$ $\mu$ should be written as $\mu\mathbf 1$ to alleviate the confusion $\endgroup$
    – Xi'an
    Aug 16, 2021 at 12:18

1 Answer 1


With only one sample, the maximum likelihood estimate is $\mu=x, \Sigma=0$.

In general, this problem is a estimation problem for an exponential family. Take the expected value of the sufficient statistics of your samples—these are the expectation parameters of the maximum likelihood distribution.

  • $\begingroup$ Yes, I understand that I cannot do much with one sample in general. But here, I assume that $\mathbb{E}x_i = \mu$ for all $i$, and $var(x_i) = \sigma^2$ for all $i$, and $cov(x_i,x_j) = \rho \sigma^2$ for all $i\ne j$. By this, I said $\mu$ can be estimated by sample mean and law of large number. Sorry that I was not clear enough. $\endgroup$ Aug 16, 2021 at 6:45
  • $\begingroup$ @user1292919 Oh! I thought you were considering a curved exponential family. I'm not sure then. I'd have to think about it more. $\endgroup$
    – Neil G
    Aug 16, 2021 at 14:45

Your Answer

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

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