Sample-Wise Double Descent in Minimum-Norm Linear Regression

Asked at Anthropic · Research & Paper Understanding · Difficulty: Hard · Last asked Apr 19, 2026

In linear regression with a fixed number of features, adding training samples usually lowers the test error. For the minimum-norm least-squares fit this fails: as the number of samples `n` grows toward the number of features `d`, the test error rises sharply, peaks at `n = d`, and only then falls again. Explain this curve, derive it for a simple model, and say how to remove the peak.

Loading the problem…