Our goal is to find the maximum-margin linear classifier for this data. In easy cases, the shortest line between a positive and negative point has a perpendicular bisector that separates the points. If so, the perpendicular bisector is surely the maximum-margin separator. Alas, in this case, the closest pair of positive and negative points, x2 and x3, have a perpendicular bisector that misclassifies x1 as negative, so that won't work.