1
$\begingroup$

Let $X\in\mathbb{R}^{n}$ be a multivariate normal variable with the mean vector $\mu$ and the covariance matrix $\Sigma$. It is well known that if the matrix $\Sigma$ is positive-definite the following cumulative distribution \begin{equation} F\left(a_{1}, a_{2}, \cdots, a_{n}\right) = \mathbb{P}\left(X_{1}<a_{1},X_{2}<a_{2},\cdots,X_{n}<a_{n}\right) \end{equation} is monotonically increasing with respect to $a_{1},a_{2},\cdots,a_{n}$.

Is the above property still valid if the covariance matrix $\Sigma$ is not positive-definite (the degenerate case)?

$\endgroup$
1
$\begingroup$

The definition of a probability (and of a sigma algebra) necessarily implies that the CDF is always monotonically increasing with respect to any of its variables.

So, yes, it is also the case if Σ is degenerate.

Edit : A less theoretical answer consists of stating that if this was not the case, then you could find strictly negative probabilities. If $a_{1} < a_{1}^{'}$ and

\begin{equation} F\left(a_{1}, a_{2}, \cdots, a_{n}\right) > F\left(a_{1}^{'}, a_{2}, \cdots, a_{n}\right) \end{equation}

Then this means that

\begin{equation} \mathbb{P}\left(a_{1}\leq X_{1}<a_{1}^{'},X_{2}<a_{2},\cdots,X_{n}<a_{n}\right) < 0 \end{equation}

Edit : It seems I got tricked by the wording (I am not a native English speaker). The CDF is always nondecreasing, no matter what Σ is. If Σ is degenerate, then the PDF can be "not increasing". The simplest example is $ X=[X_{1}, X_{2}]$ where $X_1$ is a normal distribution and $X_2$ is $0$.

$\endgroup$
  • $\begingroup$ This answer uses "increasing" where one often uses "nondecreasing". $\endgroup$ – Did Jun 19 '14 at 19:26
  • $\begingroup$ I wanted to use the same notation as the OP. $\endgroup$ – Fezvez Jun 19 '14 at 22:14
  • $\begingroup$ @Did: Could you please give me an example for the case that $F\left(X_{1}<a_{1},X_{2}<a_{2},\cdots,X_{n}<a_{n}\right) = F\left(X_1<a^{'}_{1},X_{2}<a_{2},\cdots,X_{n}<a_{n}\right)$ for $a^{'}_{1}>a_{1}$? (In sigma algebra and $-\infty<a_{j}<\infty$) $\endgroup$ – Van Long DO Jun 20 '14 at 8:46
  • $\begingroup$ You just need to have $0$ probability between $a_{1}$ and $a_{1}^{'}$. For example because F is the uniform distribution in the unit cube, and $a_{1} = 2$ (and $a_{1}^{'} = 3$). $\endgroup$ – Fezvez Jun 20 '14 at 16:47
  • $\begingroup$ I missed that $X$ is a multivariate normal distribution in this case. Then, it is not possible $\endgroup$ – Fezvez Jun 20 '14 at 16:50

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.