The Fourth Quadrant: A Stylized View of Benign Misfitting
The research studies linear regression in a specific geometric setup where training error and test error diverge unexpectedly. It shows that for certain sample sizes, a predictor can achieve good performance on new data (low test error) even if it performs poorly on the training data (high training error). This regime, called the 'fourth quadrant,' is driven by how shared signals interact with unique noise components in the data structure.