Double Descent at the Interpolation Threshold

We fit forty noisy samples of a linear target that uses two hundred features, keeping only the first p features, with the minimum norm least squares solution.

The dark curve along the front of the surface is the test error as p grows. It rises, explodes at p equal to forty, the number of samples, and then falls again, below where it started. This is double descent.

The rest of the narration is in the book's Companion.