On the Nystr\"om and Column-Sampling Methods for the Approximate Principal Components Analysis of Large Data Sets