3
$\begingroup$

Is it possible to create multivariate normal from the same normal variable? That is i have a normal variable $x_1 \sim N(0,1)$, and $p$ normal variable $y_1$ which their marginal distribution is also $N(0,1)$ they are independent and distributed can be written together as $Y\sim MVN(\underline0,I) $ fbut have some correlation with $x_1$.

I know i can write each $Y_i = c_i * x_1 + z_i * b_i$ , where $z_i$ is also $N(0,1)$, is there a way to write the multivariate normal distribution of $Y$ as a linear combination of $x_1$, while making sure they are independent from one another but keep the correlation with $x_1$ ? i don't need the specific answer, but to know if it is possible, a reference would also be great.

For clarification: Between the $z$'s there could also be correlation, just not correlated to $x_1$

$\endgroup$

1 Answer 1

0
$\begingroup$

Apparently it can't be done, ill give a counter example in 2*2. Here our constraints: $Cov(X_1,Y_1) = \rho $, $Cov(X_1,Y_2) = \rho $, $Cov(Y_1,Y_2) = 0 $, $E(Y_1) = 0$ , $(Y_2) = 0)$ , $Var(Y_1) = 1$ , $Var(Y_2) = 1$.

The means are easily confirmed as we choose the mean of $Z_i$ as zero. $Cov(X_1,Y_1) = \rho \rightarrow E(c_1 * X_1^2 + X_1 * Z_1) = \rho \rightarrow c_1 = \rho$ for the same reason $c_2 = \rho$ . $Var(Y_1) = 1 \rightarrow c_1^2 * var(X_1) + b_1^2 * var(Z_1) \rightarrow b_1 = \sqrt{(1-\rho^2)}$. Now we cant demand that $Cov(Y_1,Y_2) = 0 $ since $0 = E(\rho^2 * X_1^2 + \sqrt{1-\rho^2)}*\rho*Z_1*X_1 + \sqrt{1-\rho^2)}*rho*Z_2*X_1 + (1-\rho^2)* Z_1*Z_2) \rightarrow 0 = \rho^2 + (1-\rho^2) * \rho_z \rightarrow \rho_z = - \frac{\rho^2}{1-\rho^2}$. For values of $\rho$ this would result in a non semi positive definite matrix.

$\endgroup$

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.