2
$\begingroup$

If A is any real mxn-matrix consider the block matrix $\begin{pmatrix} E&A^T \\ A&0\end{pmatrix}$. This matrix seems to have close connections with pseudo inverse, svd etc. which are probably well known. Does anybody know a name for this block matrix and/or have hints where to find more details?

$\endgroup$
2
  • $\begingroup$ Is $E$ the identity? $\endgroup$ Aug 2, 2014 at 8:32
  • $\begingroup$ Yes, E means identity. Your reference is very helpful! Thx! Actually I came across this type of matrix by one of the applications mentioned there. $\endgroup$ Aug 3, 2014 at 8:13

2 Answers 2

2
$\begingroup$

If I had to pick a name, I'd go with saddle-point matrix. It's not a completely fitting name, but it is a matrix that I associate immediately with saddle-point problems in optimization. You can start from the review Numerical solution of saddle-point problems, Benzi, Golub, and Liesen, which focuses on solving large-scale linear systems with similar matrices.

$\endgroup$
0
$\begingroup$

I did not know the reference given by Federico. It is a great paper. Yet, it is not necessary to read it to see that this question is linked to the SVD decomposition.

Let $U=\begin{pmatrix}I_n&A^T\\A&0\end{pmatrix}$ which is a symmetric matrix. Clearly $\det(U-\lambda I_{m+n})=\lambda^{m-n}\det((\lambda^2-\lambda)I_n-A^TA)$ (even if $m-n<0$). Then an eigenvalue $\lambda$ of $U$ is $0$ or satisfies $\lambda^2-\lambda=\sigma$ where $\sigma$ is a singular value of $A$. Moreover if $[x,y]^T$ is an eigenvector associated to $\lambda$, then $AA^Ty=\sigma y,A^TAx=\sigma x$ and $x,y$ are eigenvectors of $A^TA,AA^T$. Then the SVD decomposition of $A$ gives explicitly the diagonalization of $U$.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.