I have a Hermitian matrix $H=H^\dagger$, and some complex matrix $B$ (not Hermitian).
I want to do the following: $$minimize\ \ \ \ \xi^\dagger H\ \xi$$ $$s.t.\ \ \ \ \ \xi^\dagger B\xi=0,\ \xi^\dagger\xi=1$$ In order to solve I want to use Lagrange multipliers, and vary with respect to $\xi^\dagger$: $$(H-\mu B)\xi=\lambda\xi$$ Where $\mu,\ \lambda$ are Lagrange multipliers.
The problem is I am not sure about if I can do that because $\xi^\dagger B\xi$ is not a real function outside of the restricted plane.

So, The question is: Can I use Lagrange multipliers like this?
If not, would the answer change if $B$ were a real matrix?

Any help will be greatly appreciated.

btw - I know I can just replace the condition to $\xi^\dagger\ (B+B^\dagger)\xi=0$, but the note in the top answer to this question: Lagrangian multipliers in complex optimization made me think it is not neccesary. If not I wonder whats working there that is not working here


Considering $\xi = \xi_r+j\xi_i$ and $B = B_r+jB_i$

we have

$$ (\xi_r+j\xi_i)^{\dagger}(B_r+jB_i)(\xi_r+j\xi_i)= \xi_r^{\top} B_r \xi_r+\xi_i^{\top} B_r\xi_i + j(\xi_r^{\top} B_i\xi_r+\xi_r^{\top} B_i\xi_r) = 0 $$

which is equivalent to

$$ \xi_r^{\top} B_r \xi_r+\xi_i^{\top} B_r\xi_i = 0\\ \xi_r^{\top} B_i\xi_r+\xi_r^{\top} B_i\xi_r = 0 $$

The stationary conditions are

$$ \xi^{\dagger}H+\lambda_1 \xi^{\dagger}B+\lambda_2\xi^{\dagger} = 0 $$


$$ \xi^{\dagger}(H+\lambda_1 B+\lambda_2 I)=0 $$

Those conditions can be established over $\xi_r,\xi_i$ instead. The multipliers $\lambda_k = \lambda_{kr}+j\lambda_{ki}$

  • $\begingroup$ First you got some $i$'s instead of $r's$ in several places. Second I can not see why your second line is correct - you suppose to have mix terms like $\xi_r^\dagger \ B_r \xi_i$ where did they go? last thing I did not understand how is this answer my question. you care to clarify? thanks for trying to help $\endgroup$ – Reguar May 26 '18 at 17:46
  • $\begingroup$ First, thanks for checking the indexes. Now they are fixed. Second a complex restriction can be developed into two real restrictions. This is the best way to handle those problems when $B$ is not Hermitian. $\endgroup$ – Cesareo May 26 '18 at 19:26

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.