Fig. 1.

Overview of the SMSSVD algorithm. It starts from a P × N data matrix with P variables and N samples (the leftmost matrix in the figure). 1. A subset of the variables are selected, creating a smaller data matrix. 2. A low-rank representation of the new matrix is computed using SVD. 3. The representation is expanded to the full set of variables, producing a low-rank representation of the strongest signal in the dataset. 4. (Not depicted.) The signal is removed from the original data matrix and the process is repeated to find more signals