I have a symmetric variance/covariance matrix $A$ which is of size (27 x 27). I know that it's rank deficient (rank = 21).
I also know that the units of $A$ are $m^2$. I am trying to use Singular Value Decomposition (SVD) to reduce the dimensionality of the variance/covariance matrix so that it is non-singular and I can use it as a proper variance/covariance matrix to draw random samples from (given a mean $X_0$ of size (27 x 1)).
My plan is to reduce the dimensionality of $A$ to size (21 x 21) ($A^*$) and use this matrix. I know have created a matrix $u$ of size (27 x 1) and $s$ of size (21 x 21) such that $A = u*s*u^T$.
Now, I think that the rectangular matrix $u$ is a mapping of 27 variables in the domain to the 21 variables comprising the subspace. If this is true, I should be able to take $u^T*A*u$ to get $A^*$. This turns into a diagonal (21 x 21) matrix.
My question is:
1.- is this correct until now?
2.- Assuming this is correct, what are the units of $u$ and $s$? I am inclined to believe that $u$ is unit-less (because of my assumption that it is a mapping from one domain to another) and that the units of $s$ are the same as $A$, but I am not sure.