There has been much research about using Gabor wavelet for face recognition. Other multiscale geometrical tools, such as curvelet and contourlet, have also been used for face recognition, thus it is interesting to know which method performs best, especially under illumination and expression changes. In this paper, we make a systematic comparison of wavelet, Gabor and curvelet for recognition, and find the best subband irrelevant to expression and illumination changes. We combine the multiscale analysis with subspace decomposition as our algorithm. Experiments show that for expression changes, the properties of the coarse layer of curvelet and wavelet are very good. Whilst for illumination changes, the low frequency parts of the two methods are similarly influenced, but the detail coefficients of curvelet and the high frequency of wavelet work fine with PCA, with the former outperforming the latter. When these two factors change simultaneously, the detail layer of curvelet is better relative to the others.