dmeoli / optiml / 28231681366

84%

Build Type

push

github

Committed by dmeoli

Commit Message

Support a generic lower bound lb <= x <= ub in the box-constrained optimizers

Generalize ProjectedGradient, FrankWolfe, ActiveSet and InteriorPoint (and the
BoxConstrainedQuadraticOptimizer base) from 0 <= x <= ub to a generic lb <= x <= ub
via a new optional `lb` parameter (default all-zeros, fully backward compatible):

- the projection, the maximum-feasible-step computation and the active-set lower
  set now use lb instead of 0; the Frank-Wolfe vertex picks lb/ub per coordinate
  (and the stabilized variant is fixed to use np.clip over the lb..ub box)
- the Interior Point barrier/KKT use the lower slack (x - lb): H, the rhs w, the
  dual increment and the lower dual bound (+ lambda^- lb) are generalized accordingly
- x_star and the starting point use lb

This is a pure change of variables, so it adds no per-iteration cost. Verified
against the cvxopt QP reference with a nonzero lb for all four optimizers, and
covered by a new parametrized test.

Coverage Stats

21 of 23 new or added lines in 5 files covered. (91.3%)

4 existing lines in 1 file now uncovered.

2679 of 3171 relevant lines covered (84.48%)

0.84 hits per line