Randomized Primal-Dual Coordinate Method for Large-scale Linearly Constrained Nonsmooth Nonconvex Optimization

The large-scale linearly constrained nonsmooth nonconvex optimization finds wide applications in machine learning, including non-PSD Kernel SVM, linearly constrained Lasso with nonsmooth nonconvex penalty, etc. To tackle this class of optimization problems, we propose an efficient algorithm called Nonconvex Randomized Primal-Dual Coordinate (N-RPDC) method. At each iteration, this method only randomly selects a block of primal variables to update rather than updating all the variables, which is suitable for large-scale problems. We provide two types of convergence results for N-RPDC. We first show that any cluster point of the sequence of iterates generated by N-RPDC is almost surely (i.e., with probability 1) a stationary point. In addition, we also provide an almost sure asymptotic convergence rate of $O(1/\sqrt{k})$. Next, we establish the expected $O(\varepsilon^{-2})$ iteration complexity of N-RPDC in order to drive a natural stationarity measure below $\varepsilon$ in expectation. The fundamental aspect to establishing the aforementioned convergence results is a \emph{surrogate stationarity measure} we discovered for analyzing N-RPDC. Finally, we conduct a set of experiments to show the efficacy of N-RPDC.