Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

Suppose $w_1$ and $w_2$ are zero-mean jointly Gaussian random vectors. Further suppose that they have a covariance matrix given by $$ \mathbf{cov}\begin{bmatrix}w_1 \\ w_2\end{bmatrix} = \begin{bmatrix}I & X \\ X^T & I\end{bmatrix} $$ Let $X = \begin{bmatrix}U_1 & \bar U_1\end{bmatrix} \begin{bmatrix}\Sigma & 0 \\ 0 & 0\end{bmatrix} \begin{bmatrix}U_2 & \bar U_2\end{bmatrix}^T$ be the full SVD of $X$. Now decompose $w_i$ as follows: \begin{align*} w_1 &= U_1 \eta_1 + \bar U_1 \nu_1 \\ w_2 &= U_2 \eta_2 + \bar U_2 \nu_2 \end{align*} This is done by projecting, $\eta_1 = U_1^T w_1$, etc. It's easy to verify that the random vectors $\begin{bmatrix}\eta_1 \\\eta_2\end{bmatrix}$, $\nu_1$, $\nu_2$ are mutually independent. Furthermore, they have covariances: \begin{align*} \mathbf{cov}\begin{bmatrix}\eta_1 \\ \eta_2\end{bmatrix} &= \begin{bmatrix}I & \Sigma \\ \Sigma & I\end{bmatrix}, & \mathbf{cov}\nu_1 &= I, & \mathbf{cov}\nu_2 &= I \end{align*} In a sense, this decomposition provides a new set of coordinates such that the associated covariance matrices now have diagonal blocks. I am looking for a similar result when I have more than two vectors. Say we have $w_1$, $w_2$, and $w_3$ jointly Gaussian, now with $$ \mathbf{cov}\begin{bmatrix}w_1 \\ w_2 \\ w_3\end{bmatrix} = \begin{bmatrix}I & X & Y\\ X^T & I & Z \\ Y^T & Z^T & I\end{bmatrix} $$ I have tried to extend the same technique as in the two-vector case, but run into problems because coordinates that "diagonalize" $w_1$ and $w_2$ don't also diagonalize with respect to $w_3$. I suspect one might need to take into account every possible subset of the vectors... This leads me to conjecture a decomposition of the form: \begin{align*} w_1 &= M_1 \eta_1 + P_1 \alpha_1 + Q_1 \beta_1 \phantom{+R_1\alpha_3}+ \bar U_1 \nu_1 \\ w_2 &= M_2 \eta_2 + P_2 \alpha_2 \phantom{+Q_2\alpha_3}+ R_2 \gamma_2 + \bar U_2 \nu_2 \\ w_3 &= M_3 \eta_3 \phantom{+P_3\alpha_3} + Q_3 \beta_3 + R_3\gamma_3 + \bar U_3 \nu_3 \end{align*} where $\begin{bmatrix}M_1 & P_1 & Q_1 & \bar U_1\end{bmatrix}$ is orthogonal and similarly for the other two equations, and where $\begin{bmatrix}\eta_1 \\\eta_2\\\eta_3\end{bmatrix}$, $\begin{bmatrix}\alpha_1 \\\alpha_2\end{bmatrix}$, $\begin{bmatrix}\beta_1 \\\beta_3\end{bmatrix}$, $\begin{bmatrix}\gamma_2 \\\gamma_3\end{bmatrix}$, $\nu_1$, $\nu_2$, $\nu_3$ are mutually independent, again with covariances with diagonal blocks, as in the two-vector case.

Any thoughts on how I might go about proving whether such a decomposition exists for $n\ge 3$, and computing it?

share|improve this question
add comment

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

Browse other questions tagged or ask your own question.