Incorporation of prior knowledge into the learning process can significantly improve low-sample classification accuracy. We show how to introduce prior knowledge into linear support vector machines in form of constraints on the rotation of the normal to the separating hyperplane. Such knowledge frequently arises naturally, e.g., as inhibitory and excitatory influences of input variables. We demonstrate that the generalization ability of rotationally-constrained classifiers is improved by analyzing their VC and fat-shattering dimensions. Interestingly, the analysis shows that large-margin classification framework justifies the use of stronger prior knowledge than the traditional VC framework. Empirical experiments with text categorization and political party affiliation prediction confirm the usefulness of rotational prior knowledge.