Pca done on correlation matrix

The PCA done on the matrix reals_U, of dimension n_samples x n_coeffs is too heavy. To make it lighter, it is possible to take only a subset of the samples, which should not deteriorate the results significantly. To do that, I propose the introduction of the parameter fraction_u, so that if fraction_u = 0 .66, only 2/3 of the samples are considered.