To describe the initialization in detail, let J˜ denote the number of “strongest effects” selected above, and let Z˜ denote the column-centered J˜×R matrix of Z scores for these “strong effects”. To extract the main patterns in Z˜, we perform dimension reduction on Z˜; specifically, we apply principal components analysis (through a singular value decomposition, or SVD) and sparse factor analysis21 (SFA) to Z˜. SVD yields a set of singular values and singular vectors of Z˜. Let λp, vp denote the pth singular value and corresponding right singular vector. SFA yields matrix factorization Z˜=LF+E,
