They declare by fiat that retaining only the first few dimensions should lead to low classification error and construct the training loss accordingly. (Equation 1 on page 4 of the PDF.)