Required dataset Dmin = (ε₀/εt)1/α
21,544
Break-even N* ≈ Dmin + overhead
21,584
Cost of halving the error
× 10.1
Scaling law ε(D) = ε₀ · D−α Target error εt
Test error falls as a power law in dataset size; the dataset where the curve crosses the target error is the required dataset size.

Both axes are logarithmic, so the power law is a straight line of slope −α. Because 1/α ≈ 2–5, halving the target error multiplies Dmin — and the break-even point — by 21/α.