next up previous
Next: Dimensionality Reduction Up: tr01mj2 Previous: Simple Problem Formulation

The Adventures of PCA in the Decorrelation Manifold1

A PCA decomposition of the original data can be written as

$\displaystyle Y = A_P S_P$ (2)

where the spatial maps, $ S_P$ are uncorrelated and of unit variance. That is, the principle basis vectors (spatial maps) are orthonormal (i.e. orthogonal and normalised). In matrix notation this is written as $ S_P S_P^T = I$.

Note that this is the same as the ICA decomposition, but uses a different function to minimise -- in this case, one that measures correlation.

The PCA decomposition is easily found using SVD. That is $ Y = U D
V^T$, with $ U$ and $ V$ being orthogonal matrices (i.e. $ U U^T = V V^T =
I$). Hence the PCA decomposition is given by $ S_P = V^T$ and $ A_P = U
D$.



Subsections

Stephen Smith 2001-11-29