Suppose we obtain n labeled samples {(xi , yi)} n i=1 from our underlying distribution d. suppose we break this into ntrain and ntest samples for our training and test set. recall our definition of the true least squares error (f) = e(x, y)∼d[(f(x) − y) 2 ] (the subscript (x, y) ∼ d makes clear that our input-output pairs are sampled according to d). our training and test losses are defined as: btrain(f) = 1 ntrain x (x, y)∈training set (f(x) − y) 2 btest(f) = 1 ntest x (x, y)∈test set (f(x) − y)