Closed GeoffNN closed 3 years ago
Could you add some tests for this so we don't break it again? thanks!
Not entirely convinced myself -- still working on it.
This method is for the L1 ball, so it always includes the 0 point. The 0 "vertex" is needed because you can always write the iterate as a convex combo of 0 and the +/- e_i vectors. It is the one chosen by the away step when the negative gradient has the same signs coordinate wise as the iterate.
I see, I missed that part. I'm fine with this solution as long as it has some tests then :-)
The previous commit doesn't work as such: the iterate can go out of the constraint set.
An attempt to fix #75