How to define an antidiagonal positive definite matrix with a given structure?

161 Views Asked by At

Let us assume that I have a matrix $D\in\Re^{2N\times 2N}$ with the following structure: $$ D=\begin{bmatrix} 0 & A \\ A^T & 0 \\ \end{bmatrix} \quad $$

where $A \in\Re^{N\times N}$. Is it possible to find an $A$ such that $D > 0$? I would like $A$ to be orthogonal.

2

There are 2 best solutions below

2
On BEST ANSWER

Here are two reasons for which it is not possible :

a) Such a matrix cannot fullfill Sylvester criterion for positive definiteness : all principal minors have to be $>0$.

b) The eigenvalues of $D$ come by pairs of opposite values (i.e., if $\lambda$ is an eigenvalue of $D$, so is $-\lambda$) meaning that there are negative eigenvalues which is not possible for a psd matrix.

More precisely : the spectrum of $D$ is $S \cup (-S)$ where $S$ is the set of (positive !) singular values of $A$.

Let us prove it. Let $A=USV^T$ be the Singular Values Decomposition of $A$. One can write :

$$D=\begin{pmatrix}0&A\\A^T&0\end{pmatrix}=\underbrace{\begin{pmatrix}U&0\\0&V\end{pmatrix}}_{P}\underbrace{\begin{pmatrix}0&S\\S&0\end{pmatrix}}_{\Delta}\underbrace{\begin{pmatrix}U^T&0\\0&V^T\end{pmatrix}}_{P^T=P^{-1}}$$

$D$ and $\Delta$, being similar, have the same spectrum. The spectrum of $\Delta$ is given by $$\det(\Delta-\lambda I_{2n})=0$$ which in fact is equal to

$$\det(\lambda^2I_n-S^2)=\prod_{k=1}^n(\lambda^2-\sigma_k^2)$$

(due to rule $\det\begin{pmatrix}A&B\\C&D\end{pmatrix}=\det(AD-BC)$ valid when, for example, $C$ and $D$ commute ; see theorem 3 in this detailed document)

Very similar proofs here and here.

0
On

Your matrix has trace zero. The only real symmetric positive semidefinite matrix with zero trace, is the zero matrix. (Check eg trace of zero implies all eigenvalues are zero)