1
$\begingroup$

Consider the block matrix given by $$M = \begin{bmatrix} A_{11}&A_{12}&0\\ A_{12}&A_{22}&A_{23}\\ 0&A_{23}&A_{33}\end{bmatrix}$$

What conditions should I impose on each matrix $A_{ij}$ to ensure that M is positive definite? Each matrix $A_{ij}$ is symmetric. I would like not to use a Schur complement-like condition to avoid deal with inverses. Also, could someone provide a good reference on this subject?

Thank you

$\endgroup$
  • 1
    $\begingroup$ If $M$ is PD, its principal submatrices have to be PD too. So, the first question that you should ask yourself is: how to decide whether $\pmatrix{A_{11}&A_{12}\\ A_{12}&A_{22}}$ is PD without using Schur complements? $\endgroup$ – user1551 Jul 12 '16 at 11:17
  • $\begingroup$ Actually, my question was the opposite. Formulating in terms of the principal submatrices would be like this. What are the conditions over the principal submatrices for which M is PD? Showing that these submatrices are PD is enough? $\endgroup$ – Cybernetician Jul 12 '16 at 11:32
  • $\begingroup$ $\left(\begin{array}{c} A & X \\ X^* & B \end{array}\right)$ is positive iff $X=A^{1/2}K B^{1/2}$ where $K$ is a contraction $\endgroup$ – Hamza Jul 12 '16 at 19:57
  • $\begingroup$ How do you define a contraction? $\endgroup$ – Cybernetician Jul 13 '16 at 0:39
1
$\begingroup$

Based on the post by user1551, I came up with the following proof.

If each matrix $A_{ij}$ is such that the block submatrices

\begin{equation*} M_1=\begin{bmatrix} A_{11}&A_{12}\\ A_{12}&A_{22}/2 \end{bmatrix}, M_2=\begin{bmatrix} A_{22}/2&A_{23}\\ A_{23}&A_{33} \end{bmatrix} \end{equation*} is positive definite, then the matrix $M$ is positive definite.

Proof. By assumption $M_1$ and $M_2$ are positive definite. Consequently, for vectors $(x,y)$ and $(y,z)$ the following inequalities hold \begin{align*} x^\top A_{11}x + x^\top A_{12}y + y^\top A_{12} x + y^\top \dfrac{A_{22}}{2} y&>0\\ y^\top \dfrac{A_{22}}{2}y + y^\top A_{23}z + z^\top A_{23} y + z^\top A_{33} z&>0 \end{align*} Summing both equations yields \begin{equation} x^\top A_{11}x + x^\top A_{12}y + y^\top A_{12} x + y^\top A_{22}y+y^\top A_{23}z + z^\top A_{23} y + z^\top A_{33} z>0. \end{equation} This inequality implies that the matrix $M$ is positive definite.

$\endgroup$

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.