Condition for negative semidefiniteness of matrix

70 Views Asked by At

Suppose I have matrices $A, B, D \in \mathbb{R}^{n \times n}$, where $D$ is symmetric and positive semidefinite. Furthermore, let $\alpha_1 \in [0,1], \alpha_2 \in [0,2]$ be known scalars. Is there any condition I could put on $A$ and $B$ to ensure the following?

\begin{equation} \begin{bmatrix} \alpha_1 D & - \alpha_1 D \\ -\alpha_1 D & \alpha_2 D \end{bmatrix} - \begin{bmatrix} A & 0 \\ 0 & B \end{bmatrix}^T \begin{bmatrix} D & D \\ D & D \end{bmatrix} \begin{bmatrix} A & 0 \\ 0 & B \end{bmatrix} \leq 0 \end{equation}

I tried to explore the Schur complement which works quite fine for the positive semidefinite counterparts, but for negative semidefiniteness, it seems to be quite hard.

1

There are 1 best solutions below

3
On

You could write this in the form $P - Q^TDQ \leq 0$, where $$ P = \begin{bmatrix} \alpha_1 D & - \alpha_1 D \\ -\alpha_1 D & \alpha_2 D \end{bmatrix}, \quad Q =\begin{bmatrix} A & B \end{bmatrix}. $$ If $D$ is known to be invertible, then this is the Schur complement of $D^{-1}$ within the matrix $$ M = \begin{bmatrix} D^{-1} & Q\\ Q^T & P \end{bmatrix} = \begin{bmatrix} D^{-1} & A & B\\ A^T & \alpha_1 D & -\alpha_1 D\\ B^T & -\alpha_1 D & \alpha_2 D \end{bmatrix} $$ Thus, your condition is equivalent to $M$ having $n$ positive eigenvalues and $2n$ non-positive eigenvalues.

Note that the block-matrix $P$ can be written as the Kronecker product $$ P = \begin{bmatrix} \alpha_1 & -\alpha_1\\ -\alpha_1 & \alpha_2 \end{bmatrix}\otimes D. $$ As such, we can see that $P$ will have the maximal possible number of non-negative eigenvalues (excluding the case where $\alpha_1 = \alpha_2 = 0$) when $\alpha_1 \alpha_2 - \alpha_1^2 \leq 0 \implies \alpha_2 \leq \alpha_1$; the original condition cannot hold otherwise. In this case, and if $P$ is inveritble, $M$ will satisfy the desired condition if and only if the Schur complement of $P$ is negative definite.

Putting it all together, we can say that if $D$ is invertible and $\alpha_2 < \alpha_1$, then your condition is equivalent to $$ D^{-1} - QPQ^T \leq 0\implies\\ D^{-1} - \alpha_1(ADA^T) + \alpha_1(ADB^T + BDA^T) - \alpha_2 BDB^T \leq 0. $$