3
$\begingroup$

Following the (Vasicek 2002) paper, on page 2, it is specified that if the variables $X^i$ are jointly standard normal with equal pairwise correlations $\rho$ they can be represented as:

$X^i = Y\sqrt{\rho} + Z^i\sqrt{1 - \rho}$

where $Y$ and $Z^1...Z^n$ are mutually independent standard normal variables. The paper states that

"This is a property of the equicorrelated normal distribution."

How can I derive this property?

I tried finding the correlation of two variables $X^i$ and $X^j$ constructed in that way, but I don't get $\rho$ as the result.

$\endgroup$
1
  • $\begingroup$ If you would like a very detailed derivation of this result from first principles, then please see stats.stackexchange.com/a/71303/919. Among other things, it shows that this representation of the $X^i$ does not require joint normality: that assumption is needed only to conclude the $Z_i$ are independent. (They will, in any event, be uncorrelated, which often is all one needs.) $\endgroup$
    – whuber
    May 8, 2017 at 20:48

1 Answer 1

4
$\begingroup$

Which property is being called out as a property enjoyed by jointly normal standard normal random variables? That they have a representation as $$X_i = \sqrt \rho\, Y + \sqrt{1-\rho}\,Z_i \tag{1}$$ where $Y$ and the $Z_i$ are independent standard normal random variables? The joint normality of the $X_i$ follows from the fact that they are obtained as linear combinations of jointly normal random variables. That the $X_i$ have expected value $0$ follows straightforwardly from the linearity of expectation applied to $(1)$ while bilinearity of covariance and independence gives \begin{align} \operatorname{cov}(X_i,X_j) &= \operatorname{cov}\left(\sqrt \rho\, Y + \sqrt{1-\rho}\,Z_i,\sqrt \rho\, Y + \sqrt{1-\rho}\,Z_j\right)\\ &= \rho\operatorname{cov}(Y,Y) + 0 + 0 + (1-\rho)\operatorname{cov}(Z_i,Z_j)\\ &= \rho + \begin{cases}1-\rho, & i=j,\\ 0, &i \neq j, \end{cases}\\ &= \begin{cases}1, & i=j,\\ \rho, &i \neq j, \end{cases} \end{align} showing that the $X_i$ have unit variance and correlation coefficient as desired.

Note that all this works only when $\rho \geq 0$. The common correlation $\rho$ satisfies $-\frac{1}{n-1}\leq \rho \leq 1$ but the construction above does not work for $-\frac{1}{n-1}\leq \rho < 0$. For a construction of unit random variables with common correlation coefficient $-\frac{1}{n-1}$, see this answer of mine.

$\endgroup$

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.