Shahzaib-07/Low-rank-Sparse-GPU-implementation: v1.0.0

This Python toolbox implements an optimized, GPU-oriented fused low-rank plus sparse decomposition (fusedLS) from paper for multi-subject functional connectivity (FC) data. It performs decomposition by solving a fused variant of principal component pursuit (PCP), which adds a fusion-style penalty on differences between columns of the low-rank matrix to encourage consistent structure across subjects. The resulting optimization problem is solved using a linearized alternating direction method of multipliers (ADMM). The approach is designed to separate shared (correlated) FC structure and subject-specific FC structure from observed multi-subject FC matrices, represented by the low-rank and sparse components, respectively. An example application is the separation of stimulus-driven and background components of dynamic functional connectivity in naturalistic fMRI experiments.