Abstract
Subspace segmentation or subspace learning is a challenging and complicatedtask in machine learning. This paper builds a primary frame and solidtheoretical bases for the minimal subspace segmentation (MSS) of finitesamples. Existence and conditional uniqueness of MSS are discussed withconditions generally satisfied in applications. Utilizing weak priorinformation of MSS, the minimality inspection of segments is further simplifiedto the prior detection of partitions. The MSS problem is then modeled as acomputable optimization problem via self-expressiveness of samples. A closedform of representation matrices is first given for the self-expressiveness, andthe connection of diagonal blocks is then addressed. The MSS model uses a rankrestriction on the sum of segment ranks. Theoretically, it can retrieve theminimal sample subspaces that could be heavily intersected. The optimizationproblem is solved via a basic manifold conjugate gradient algorithm,alternative optimization and hybrid optimization, taking into account ofsolving both the primal MSS problem and its pseudo-dual problem. The MSS modelis further modified for handling noisy data, and solved by an ADMM algorithm.The reported experiments show the strong ability of the MSS method onretrieving minimal sample subspaces that are heavily intersected.