PCA:
1) A set of orthogonal e vectors that are ranked by the projected variance
2) Top k e's give the best k-th order approximation of the v's in terms of variance explained
3) The e's are actually the eigenvectors of the covariance matrix of the v's.
ICA:
1) Similar idea but the e's do not have to be orthogonal
2) Don't have any nice "best k-th-order approximation" property