Maximizing a convex function with a linear constraint

Question

The problem is $$\max f(\mathbf{x}) \text{ subject to } \mathbf{Ax} = \mathbf{b}$$

where $f(\mathbf{x}) = \sum_{i=1}^N\sqrt{1+\frac{x_i^4}{(\sum_{i=1}^{N}x_i^2)^2}}$,
$\mathbf{x} = [x_1,x_2,...,x_N]^T \in \mathbb{R}^{N\times 1}$, and
$\mathbf{A} \in \mathbb{R}^{M\times N} $

We can see that $f(.)$ is in the form of $\sqrt{1+y^2}$ and is a convex function.
It can be also shown that f(.) is bounded in $[\sqrt{2}, 2]$.

I know that a convex maximization problem is NP-hard, in general.

But using the specific nature of the problem, is it possible to solve it using any standard convex optimization software/package?

Answer 1

Yes, Convex Optimization with equality constraint is NP-Hard in general. However, there exist mature techniques that find very nice approximate solutions to convex optimization problems, like Coordinate Descent.

Suppose you use coordinate descent and the matrix A has rank $k$. You can fix n-k-1 coordinates of your feasible solution $x = (x_1, x_2, x_3, ..., x_n)$ and then the solution vectors in the solution space are uniquely determined by one coordinate, for example $x_i$. In that case, you can just take the derivative of $f(\cdot)$ with respect to $x_i$ to find the maximum in this iteration.

Then we iteratively fix n-k-1 coordinate and improve the solution until a approximately optimal one is found.