Facial recognition/verification [1] is a continuing and growing area of research in the field of biometrics. One of the first approaches to this challenge was Principal Component Analysis (PCA) [2, 3]. Typically PCA is performed in the original spatial domain. However, PCA has a high sensitivity to illumination effects in the original spatial domain. We propose that by using Wavelet Packet Decomposition [4] to create localized space-frequency subspaces of the original data, we can perform PCA in these subspaces which can generalize better across illumination variations. We report results on the CMU PIE database [5] by comparing reconstruction error in the original spatial domain to that of the reconstruction error in the spatial subspaces (keeping same number of eigenvectors). It is seen that the total reconstruction error of the space-frequency subspaces is smaller than that of the original space and the automatically pruned Wavelet packet PCA produced better face recognition performance across illumination.