Hei guys,

I would like to check with you a prove I have

Let the matrix $X_1 = \begin{bmatrix} A_1 & B_1\\ B_1 ^ T & A_1\\ \end{bmatrix}$ be positive definite. Based on this I would like to show that the matrix $X_2 = \begin{bmatrix} X_1 & B_2\\ B_2 ^ T & A_1\\ \end{bmatrix}$, where $B_2 = \begin{bmatrix} 0\\ B_1\\ \end{bmatrix}$ is also positive definite.

The prove I have looks like this:

  1. Based on the assumptions on $X_1$ and the Schur Lemma we get that $A_1, A_1 - B_1 ^ T \times A_1 ^ {-1} \times B_1$ and $A_1 - B_1 \times A_1 ^ {-1} \times B_1 ^ T$ are positive definite.
  2. In order for $X_2$ to be positive definite one has to show (based on the same lemma) that $X_1 - B_2 \times A_1 ^ {-1} \times B_2 ^ T$ is positive definite. This is equivalent with showing that the matrix:

$\begin{bmatrix} A_1 & B_1\\ B_1 ^ T & A_1 - B_1 \times A_1 ^ {-1} \times B_1 ^ T\\ \end{bmatrix}$ is positive definite.

For me this is true on if the matrix $- B_1 \times A_1 ^ {-1} \times B_1 ^ T$ is positive definite.

I would like to ask you if is possible $X_2$ to be positive definite with the $- B_1 \times A_1 ^ {-1} \times B_1 ^ T$ being positive definite?

Thanks, Bogdan.


The result seems to follow from the Sylvester criterion for positive definiteness and determinant identity for Schur complement. You need $A_1-B_1 A^{-1} B^T_1$ to be positive definite (which is positive definite, because $X_1$ is, as you say), but $-B_1 A^{-1}_1B^T_1$ is not positive definite if $X_1$ is positive definite. In other words, $-B_1 A^{-1}_1B^T_1$ is never positive definite if $A_1$ is.

  • $\begingroup$ So $X_2$ is positive definite by construction ? $\endgroup$ – Bogdan May 13 '13 at 9:24
  • $\begingroup$ You were doing fine, I just added details, except the matrix in your last question is negative definite. $\endgroup$ – DVD May 13 '13 at 9:37
  • $\begingroup$ Just to make sure I understood correctly: $X_2$ is positive definite only if $−B_1 \times A_1 ^ {-1}×B_1 ^ T$ is also positive definite? And there is no other way for $X_2$ to be so. $\endgroup$ – Bogdan May 13 '13 at 9:46
  • $\begingroup$ But $A_1 - B_1 A_1 ^ {-1} B_1 ^T$ is positive definite because $X_1$ is. I dont understand your point $\endgroup$ – Bogdan May 13 '13 at 11:01
  • $\begingroup$ Can you give more intuition why this is not possible ... i simply do not see it. $\endgroup$ – Bogdan May 13 '13 at 12:08

I have to say I got lost somewhere on the way, but was thinking about the following alternative proof:

  1. Assuming that $X_1$ is positive definite we would get that $\begin{bmatrix} x & y \end{bmatrix} \times \begin{bmatrix} A & B\\ B^T & A\\ \end{bmatrix} \times \begin{bmatrix} x ^ T\\ y ^ T \end{bmatrix} \geq 0$. This is equivalent with: $x A x^T + 2x B y^T + y A y^T \geq 0, \forall x, y$ (vectors with proper sizes).

  2. For $X_2$ to be positive definite (based on the same definition) would imply that: $u X_1 u^T + 2u B_2 v^T + v A v^T \geq 0, \forall u, v$. Choosing $u = (x, y), v = z$ we get that $x A x^T + 2x B y^T + y A y^T + 2y B z^T + z A z^T \geq 0$.

I fail to see why the second inequality is true $\forall x, y, z$ assuming that the first one is true $\forall x, y$.

This makes me to think that $X_2$ in general is not positive definite, but rather for some specific B with some specific properties.


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.