The Chosen Few

In a trained soft-margin SVM, which training points end up as support vectors, i.e., the only points with nonzero dual coefficients that actually determine the decision boundary?

All training points contribute equally regardless of their position relative to the boundary

Only points the model misclassifies; correctly classified points never count as support vectors

Points lying exactly on the margin boundary, plus points that violate the margin (inside it or misclassified, $\xi_i > 0$ )

Only points strictly outside the margin on the correct side of the boundary

Only the points closest to each class's centroid (mean feature vector)