Quadratically constrained quadratic programming

61 Views Asked by At

I tried to find something before but it seems all the answers do not include my kind of problem.

Basically I want to minimize

$\vec{a}^TM\vec{a}+\vec{a}^T\vec{b}$, $M$ is symmetric, positive definite

subject to

$\vec{a}^T\vec{c}=\alpha$

and

$\vec{a}^TZ_i\vec{a}=0$ where $Z_i, i=1, 2, 3$ are general matrices, without assumption on them. The only thing you could assume is that they could be symmetrised in this expression.

I tried with the standard approach of Lagrange multipliers,

$L = \vec{a}^TM\vec{a}+\vec{a}^T\vec{b}-\sum_i \lambda_i (\vec{a}^TZ_i\vec{a})+\lambda (\alpha-\vec{a}^T\vec{c}) $

Taking the gradient of this and looking for stationary points

$2M\vec{a}-2\sum_i \lambda_i Z_i\vec{a} -\lambda \vec{c} -\vec{b}= 0$

Then you have also the constraints. Now, the presence of the $\sum_i \lambda_i Z_i\vec{a}$ makes quite difficult to me to solve the problem, and I am not sure if there is a trickier way.

Also, looking for the Hessian of the problem $M-\sum_i \lambda_i Z_i$. So I guess I should look at all the configuration that make this positive definite? Is this a doable thing analytically speaking?

I am not looking for a solution, although I would appreciate some hint or some resources that I could look into them.

Thanks!