Double descent is an empirical phenomenon where test error first increases then decreases as model size increases beyond the interpolation threshold, contradicting classical statistical intuition that overfitting increases with model capacity.