Linear constraint in convex optimization

366 Views Asked by At

Is it true that the solution to a linearly constrained convex minimization problem can only be placed on the boundary of the constraint set, for any nonlinear convex objective, e.g.

$$ \min_x f(x)$$ s.t. $$Ax \le b$$

when $f$ is convex?

If yes, why?

Update: I guess the answer is no. $\min (x-2)^2$ s.t. $x \le 100$

But then why adding a $L_1$ norm to convex objectives leads to exact sparsity?

By exact sparsity I mean all elements of the solution vector are either zero or one.(As opposed to $L_2$ norm regularization that only shrinks value of solution.) The reason I heard was that solution is guaranteed to happen on vertices of $L_1$ ball and on the vertices all elements are either zero or one.

1

There are 1 best solutions below

2
On

Of course this is not correct, you are confused. Convex function might have an optimum without any constraints like you showed yourself. I am not sure I understood your last question regarding $L_1$ norm but coming back to the original problem actuall the opposite is true - a linear (or concave) function on convex sets always has an optimum on the boundary.