Minimum norm interpolation
new sample
large eigenvalues k
tail length (small λ)
tail height λ_small
Theorem 4: expected excess risk of the minimum norm solution, isotropic Gaussian inputs
training points n
noise σ²
signal ‖θ*‖²