In a hard-margin support vector machine (SVM), convex optimization selects the separating boundary with the widest margin. The objective is convex, and the constraints are linear, so an optimal solution is global rather than an inferior local solution. section 8.2.3
This formulation minimizes the squared weight norm while requiring every training point to be correctly separated by the margin. section 8.2.3
Training observations that meet the constraint with equality are support vectors: they lie on the margin boundaries and constrain their maximum separation. Data that is not perfectly separable requires a more flexible soft-margin formulation. section 8.2.3

