We need to assume that the given data is corrupted, and the data we get after dimensionality reduction is the real data.
The least important dimensions, i.e the dimensions in which variance is very less, are actually noise and hence can be discarded.
Kuhu
