Evaluate sensitivity to nonconvex optimization choices

Evaluate the sensitivity of the library-learning-assisted robust principal component analysis decomposition to multiple initial learned spatial subspaces and continuation schedules.

Background

The LLA-RPCA optimization is jointly nonconvex in the learned-basis coefficients and modal coefficients. Consequently, convergence does not establish global optimality, and the recovered decomposition may depend on the initial learned subspace and the continuation schedule used in the augmented-Lagrangian iterations. The paper uses one initialization strategy and one continuation schedule, leaving the robustness of the resulting decomposition to these choices unresolved.