Widest Street Wins

Two linear classifiers perfectly separate the same training set with zero training error, but classifier A has a much larger margin than classifier B. Why does the SVM formulation explicitly choose to maximize the margin instead of just finding any separating hyperplane like classifier B?

A wider margin guarantees zero training error even on non-separable data, which narrow-margin hyperplanes cannot achieve

Maximizing the margin minimizes the number of support vectors required, which is the SVM's primary training objective

A wider margin means the decision boundary sits farther from every training point, which tends to reduce generalization error and makes the classifier more robust to small perturbations in new data

Maximizing the margin is only a computational convenience that speeds up gradient descent and has no effect on generalization

The margin size directly determines the number of features the model can use, so a wider margin allows more informative features