3
$\begingroup$

Suppose we are given a block $2\times 2$ matrix that is positive-definite, and let's suppose for simplicity that the blocks along the main diagonal are the identity. So $$ \begin{bmatrix} I & X \\\ X^T & I \end{bmatrix} > 0 $$ It turns out that any such matrix admits a factorization of the following form: $$ \begin{bmatrix}I & X \\\ X^T & I\end{bmatrix} = \begin{bmatrix}M_1 & N_1 & 0\\\ M_2 & 0 & N_2\end{bmatrix} \begin{bmatrix}M_1 & N_1 & 0\\\ M_2 & 0 & N_2\end{bmatrix}^T $$ such that the $N_i$ are square and invertible, and the $M_i$ are tall and full-rank. Here is one possible construction. Let $X = U_1 \Sigma U_2^T$ be a compact SVD, so $\Sigma > 0$ is diagonal, and the $U_i$ have orthonormal columns. Let $\bar U_i$ be the orthogonal completion of $U_i$, so that the matrix $\begin{bmatrix}U_i & \bar U_i\end{bmatrix}$ is orthogonal. Now define the matrices as follows: $M_i = U_i\Sigma^{1/2}$, and $$N_i = \begin{bmatrix} U_i(I-\Sigma)^{1/2} & \bar U_i\end{bmatrix}$$ These matrices are always well-defined because we must have $I - XX^T > 0$, and thus $0 < \Sigma < I$ in order to guarantee that the $2\times 2$ block matrix is positive.

Two questions. First, does this factorization have a name? I came up with it, but I figure I'm probably not the first to do so. Second, can one generalize this factorization to the $3\times 3$ case or greater. For example, how does one factor the following matrix? $$ \begin{bmatrix}I & X & Y\\\ X^T & I & Z \\\ Y^T & Z^T & I\end{bmatrix} > 0 $$

Some motivation for the factorization in the $2\times 2$ case: Suppose the matrix is in fact the joint covariance matrix for two Gaussian random zero-mean vectors $w_1$ and $w_2$. Then using the factorization above, we can write: \begin{align*} w_1 &= M_1 e_c + N_1 e_1 \\\ w_2 &= M_2 e_c + N_2 e_2 \end{align*} where $e_c$, $e_1$, and $e_2$ are mutually independent white Gaussian noise. We can think of $e_c$ as the "common" noise driving $w_1$ and $w_2$. It's like an innovations decomposition, but symmetric in the sense that we are not required to choose which variable is written in terms of the other.

$\endgroup$

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.