The standard objective in machine learning is to train a single model for all users. However, in many learning scenarios, such as cloud computing and federated learning, it is possible to learn one personalized model per user. In this work, we present a systematic learning-theoretic st