Convergence of block coordinate descent with diminishing radius for nonconvex optimization

Block coordinate descent (BCD), also known as nonlinear Gauss-Seidel, is a simple iterative algorithm for nonconvex optimization that sequentially minimizes the objective function in each block coordinate while the other coordinates are held fixed. We propose a version of BCD that is guaranteed to converge to the stationary points of block-wise convex and differentiable objective functions under constraints. Furthermore, we obtain a best-case rate of convergence of order $\log n/\sqrt{n}$, where $n$ denotes the number of iterations. A key idea is to restrict the parameter search within a diminishing radius to promote stability of iterates, and then to show that such auxiliary constraints vanish in the limit. As an application, we provide a modified alternating least squares algorithm for nonnegative CP tensor factorization that converges to the stationary points of the reconstruction error with the same bound on the best-case rate of convergence. We also experimentally validate our results with both synthetic and real-world data.