Tightening the Sample Complexity of Empirical Risk Minimization via Preconditioned Stability

We tighten the sample complexity of empirical risk minimization (ERM) associated with a class of generalized linear models that include linear and logistic regression. In particular, we conclude that ERM attains the optimal sample complexity for linear regression. Our analysis relies on a new notion of stability, called preconditioned stability, which may be of independent interest.