Understanding scikit learn import variants

Scikit learn import statements in their tutorials are on the form

from sklearn.decomposition import PCA

Another versions that works is

import sklearn.decompositionpca = sklearn.decomposition.PCA(n_components = 2)

However

import sklearnpca = sklearn.decomposition.PCA(n_components = 2)

does not, and complains

AttributeError: module 'sklearn' has no attribute 'decomposition'

Why is this, and how can I predict which ones will work and not so i don't have to test around? If the understanding and predictiveness extends to python packages in general that would be the best.