0
$\begingroup$

Problem

I am wondering if there is a way to efficiently compute the maximum singular value of \begin{align} C(\phi)=A \cos\phi+B \sin\phi, \end{align} where A, B and C are real 2x2 matrices.

Numerical Approach (edited)

So far, I could only solve the problem 'numerically' by

  • Direct optimization (e.g. using matlabs's fmincon), which is sensitive to the choice of the intial value of $\phi$. Other (global) optimizers are more reliable but need far more evaluations. Maybe an appropriate bisection algorithm would help here? Suggestions?
  • Rank-1 tensor approximation (e.g. using 'tenser toolbox' for matlab), which also needs quite some iterations. I am not very familiar with tensor algebra so I am not sure if this approach is even valid in general.

Notes (edited)

  • It can be assumed that either $A$ or $B$ is positive semidefinite.
  • QZ-decompositon may be applied to make $A$ and $B$ (and thus also $C$) triangular. But I am not sure if this helps.
$\endgroup$
0
$\begingroup$

In the special case of $\phi=\pi/4$, you're essentially asking for the maximum singular value of $A+B$. So without further structural assumptions imposed on $A$ and $B$, the question is basically unanswerable, as the eigenvalues of $A+B$ can basically be anything within the spectrum of $A$ and $B$.

$\endgroup$
  • $\begingroup$ Thanks for your answer. If it helps, it can be assumed that either A or B is positive semidefinite. Furthermore, QZ-decompositon may be applied to make A and B (and thus also C) triangular . So far, I could solve the problem numerically by a) direct optimization (e.g. using matlabs's fmincon) which is sensitive to the choice of the intial value of $\phi$. Other (global) optimizers are more reliable but need far more evaluations. An appropriate bisection method may help?! OR b) rank-1 tensor approximation (e.g. using tenser toolbox for matlab), which also needs quite some iterations. $\endgroup$ – puman Jul 3 '17 at 11:03
  • $\begingroup$ The additional constraint makes no difference. You can easily come up with $A$ and $B$ that satisfy all your assumptions, yet $A+B$ can be literally anything. $\endgroup$ – Yining Wang Jul 8 '17 at 1:39

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.