Mathematically, Principal components are a linear combination of eigenvectors and all observations. So in some sense, by multiplying eigenvectors with observations of predictors we implicitly capture the variance.

How can you sample data from a large dataset?