1
$\begingroup$

There is a lot of work done on low rank updates to inverses or SVDs (or similar decompositions), but I am wondering what can be said about the inverse or SVD

$(A + c \times B)$

in terms of $(A + d \times B)$ for some other $d$. That is, in general full-rank updates, but along very particular dimensions.

$c$ and $d$ are positive scalars, and $A$ and $B$ are symmetric, positive definite matrices. Quantities like this arise in MLE-like computations in linear mixed models (e.g. http://onlinelibrary.wiley.com/doi/10.1111/1467-9868.00092/abstract).

$\endgroup$
1
$\begingroup$

I don't think there is any computational advantage in general. For any pair of symmetric $M$, $N$ and given $c\neq d$, one can find symmetric $A, B$ such that $M=A+cB$, $N=A+dB$, so essentially you are asking "if I already know the eigendecomposition of a symmetric matrix $M$, can I get the one of any other completely unrelated symmetric matrix $N$ for cheaper?"

In your question $A$ and $B$ are constrained to be positive definite, but that restriction doesn't matter because one can always add multiples of the identity, changing the eigendecomposition in a predictable way.

If you want a decomposition that makes sense for the whole matrix pencil $A+\lambda B$, you can use a QZ decomposition $A=QTZ, B=QSZ$, with $T$ and $S$ upper triangular. Those $Q$ and $Z$ are two different orthogonal matrices, though.

| cite | improve this answer | |
$\endgroup$

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.