Look into other/better equilibration strategies

[lukevolpatti]

QSS currently has a hard time with big l1 trend-filtering problems. I think this is due to the fact that the constraint matrix in these problems is extremely poorly conditioned - some of its columns are of much larger norm than others (this is due to the fact that a large regularization penalty is needed to get a solution with sparse second difference).

OSQP also seems to do poorly on such problems (though still better than QSS). However, Mosek is able to quickly get a solution with half of the objective value of OSQP.

Perhaps other matrix equilibration strategies could help with performance on this kind of problem? This may also be a question of solution polishing, though.