When can we test functions more efficiently than we can learn them? We will consider this fundamental problem in the standard PAC learning setting, which corresponds to the sample-based distribution-free property testing model. The VC dimension of a class of functions characterizes the number of samples needed to learn the class. But since some properties can indeed be tested more efficiently than they can be learned, VC dimension by itself does not characterize the number of samples needed to test it. Nonetheless, we will see that VC dimension combined with a closely-related “lower VC dimension" variant does give strong lower bounds on the number of samples required to test many properties of functions, including halfspaces, intersection of halfspaces, polynomial threshold functions, and decision trees.

This is joint work with Renato Ferreira Pinto Jr. and Nathan Harms.

Video Recording