Assume $a\in {{R}^{d}}$ and $B\in {{R}^{d\times d}}$. Consider minimizing an objective function of the form
$\text{sgn}{{\left( a \right)}^{T}}B~\text{sgn}(a)$
wrt to $a$. Assume it is bounded below.
Is replacing the sign functions with $f(x)=2/(1+e(-x)) - 1$ (sigmoid smoothing) and doing gradient descent a good idea? Any other suggestions?