0
$\begingroup$

On Wikipedia, it states that the covariance matrix of the conditional Gaussian random variable, $f_{X_1|X_2}(x_1,x_2)$ is given by:

$${\displaystyle {\overline {\boldsymbol {\Sigma }}}={\boldsymbol {\Sigma }}_{11}-{\boldsymbol {\Sigma }}_{12}{\boldsymbol {\Sigma }}_{22}^{-1}{\boldsymbol {\Sigma }}_{21}.}$$

Does the conditional covariance matrix depend on $X_2$? I believe this is the case because, for instance, $\mathbf{\Sigma}_{22} = \mathbb{E}((X_2 - m_{X_2})(X_2 - m_{X_2}))$

However, I am not sure if I fully understood the meaning of "depending on $X_2$". Can someone chime in?

$\endgroup$
  • 1
    $\begingroup$ The conditional covariance will not depend on $X_2$ if $X_1$ and $X_2$ are uncorrelated. That's why you have the cross-terms $\Sigma_{12/21}$ in that expression. Otherwise, in general, fixing $X_2$ to different values will produce different distributions of $X_1$ and hence different covariances. $\endgroup$ – Moss Murderer Feb 4 '18 at 7:33
  • $\begingroup$ Do you mean that the expression $\overline \Sigma$ will not depend on $X_2$?. Is it because the product term $\Sigma_{12}....$ cancels out the $X_2$? Doesn't seem to be obvious $\endgroup$ – Detective Mooch Feb 4 '18 at 7:37
  • $\begingroup$ I was specifically talking about the case of $\Sigma_{12}=\Sigma_{21}=0$, so that must be obvious. There may be other cases where those terms are non-zero, and yet there is no dependence on $X_2$. $\endgroup$ – Moss Murderer Feb 4 '18 at 7:42
1
$\begingroup$

Given that the variance-covariance matrix of the distribution of $X_1$ conditional on $X_2$ is $${\displaystyle {\overline {\boldsymbol {\Sigma }}}={\boldsymbol {\Sigma }}_{11}-{\boldsymbol {\Sigma }}_{12}{\boldsymbol {\Sigma }}_{22}^{-1}{\boldsymbol {\Sigma }}_{21}}$$ there are only constant terms in this expression and no function of the random variable $X_2$. The covariance matrix therefore does not depend on the realisation of $X_2$. Dependence can only be understood in this functional sense and the covariance matrix is not a random variable in this Gaussian case.

$\endgroup$
1
$\begingroup$

What @Xi'an wrote is true, but let me add further, which hopefully resolves the OP's confusion.

The conditional covariance of $X_1$ given $X_2$ does not depend on $X_2$. However, the conditional mean, i.e., the mean of $X_1$ conditional on $X_2$, let's call it $\overline{\mu}$, does depend on $X_2$. Therefore, the conditional distribution of $X_1$ given $X_2$ does depend on $X_2$.

Specifically, as stated at the same Wikipedia link, $\overline{\mu} = \mu_1 + \Sigma_{12}\Sigma_{22}^{-1}(a-\mu_2)$, where $\mu_1$ and $\mu_2$ are the unconditional means and $a$ is the value of $X_2$ being conditioned on.

Per the formulas, if that value $a$ being conditioned on happens to equal the unconditional mean of $X_2$, or if $X_1$ and $X_2$ are uncorrelated ($\Sigma_{12}$ = zero matrix), then the conditional mean of $X_1$ equals the unconditional mean of $X_1$, i.e., there is no adjustment to the unconditional mean to get the conditional mean. Otherwise, the conditional mean of $X_1$ is adjusted from its unconditional mean based on the realized value of $X_2$ per the formula shown.

$\endgroup$

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.