A theoretical guarantee on classification error based on how well-separated different classes are in the learned representation.