They declare by fiat that retaining only the first few dimensions should lead to low classification error and construct the training loss accordingly. (Equation

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.)