Sebastian Pokutta

that's a very good idea!

i thinkt that would be a cool example. sometimes fix bit precision is enough also.

would be great to interface with jump!

especially the projections we might need for bcg or if we want to add fcfw

so there is two ways actually: 1. use jump to formulate and solve the lmo. this will add some overhead but better than interfacing scip, glop, etc all separately. 2....

but also the other around is interesting: using this package as a solver within jump so i guess it goes both ways

wait until nonlinear interface in jump is stable.

you mean the barron space one? F. Bach: Breaking the Curse of Dimensionality with Convex Neural Networks: [link](https://arxiv.org/abs/1412.8690) see page 9ff

will fix a few more things and then this can be merged. we need to make alpha and M parameters though and propagate them through

basically implements https://arxiv.org/pdf/2102.02029.pdf -> not done yet. need to change theory also to be more useful