Take the 2-minute tour ×
Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. It's 100% free, no registration required.

Let $v,x,g$ be three vectors and $\alpha$ be a constant. The problem is

$$\min\limits_v \{\alpha\|v\|^2-\|x^Tv\|^2+\|g^Tv\|^2\}$$

where $\|v\|^2=\sum\limits_{i=1}^{|v|}v_i^2$ and $|v|$ is the cardinality of $v$.

Its gradient is

$$\alpha v-x^Tvx+g^Tvg.$$

How should I get all vector $v^*$ such that $\alpha v^*-x^Tv^*x+g^Tv^*g=0$?

share|improve this question
1  
Your equation is equivalent to : $$ (\alpha I - xx^T+gg^T)v=0 $$ So $v$ should be in the kernel of matrix ($\alpha I -xx^T+gg^T)$ (which might not exist). So it must be a linear combination of vectors $x$ and $g$. Define $v= ax+bg$. Then plugging this parametrization in the equation you will get equation for $a,b$ which are defined up to multiplier. But ...taking the gradient it is not a correct way to solve the problem. Without the constraint on $v$ it might be not well defined. It depends on $\alpha,x,g$. –  Alexander Vigodner Jul 2 at 20:38
1  
I meant the problem is not well defined. Gradient of course can be calculated anyway. Assume that $v^*$ is an optimal solution. Then if $\min$ is positive it can be decreased to zero setting $v=0$. So $v^*$ is not optimal. If $\min$ is negative then it can be set to $-\infty$ increasing the vector. So you must set a constraint on $v$. –  Alexander Vigodner Jul 2 at 21:12
add comment

1 Answer 1

up vote 2 down vote accepted

Obviously, there is no closed form solution. But what you can do is to solve the system $v=1/\alpha*(x^Tvx-g^Tvx)$. One way to solve is to iterative update $v$ from a initial point $v^0$. It is guarantee to converge but the result is only one local solution. If you want to obtain all $v$, it seems impossible.

share|improve this answer
add comment

Your Answer

 
discard

By posting your answer, you agree to the privacy policy and terms of service.

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