Pareto optimality and linear scalarization #
Two objectives to be minimized are compared in the product order on pairs, so the Pareto optimal
points among those satisfying P under the objectives f are those with MinimalFor P f. The
linear scalarization c.1 + β * c.2 is strictly monotone for β > 0, so its minimizers are Pareto
optimal (MinimalFor.of_fst_add_mul_snd_le). The converse fails without convexity: a Pareto
optimal point need not minimize the scalarization at any weight. As the weight β grows, the
minimizers trade the second objective for the first.
Main statements #
strictMono_fst_add_mul_snd: the scalarization is strictly monotone for a positive weight.MinimalFor.of_fst_add_mul_snd_le: a minimizer of the scalarization at a positive weight is Pareto optimal.exists_minimal_forall_exists_fst_add_mul_snd_lt: three points of which one is Pareto optimal yet beaten under the scalarization at every weight.snd_le_of_fst_add_mul_snd_le,fst_le_of_fst_add_mul_snd_le: of two points, each no worse than the other at one of two weights, the one preferred at the larger weight has the smaller second objective and, for nonnegative weights, the larger first objective.antitone_snd_of_fst_add_mul_snd_le,monotoneOn_fst_of_fst_add_mul_snd_le: along a curve of minimizers the second objective falls and the first rises with the weight.
References #
- [BV04], the scalarization of vector optimization problems.
A minimizer of the scalarization at a positive weight is Pareto optimal.
The converse fails without convexity. Of the points (0, 4), (4, 0) and (2, 3) the last
is Pareto optimal, yet at every weight one of the others has a smaller scalarization.