Smoothly Bounding User Contributions in Differential Privacy

In many applications of differential privacy, a single user might contribute more than one data point. A prominent example, which is the focus of this paper, is private machine learning, where a user often provides several points in the training data set.