Decentralized Riemannian Gradient Descent on the Stiefel Manifold
This addresses distributed optimization problems for networks of agents, with the gradient tracking algorithm being the first to offer exact convergence on the Stiefel manifold, though it is incremental in method.
The authors tackled decentralized non-convex optimization over the Stiefel manifold by proposing two algorithms: a decentralized Riemannian stochastic gradient method with O(1/√K) convergence and a decentralized Riemannian gradient tracking algorithm with O(1/K) convergence, achieving exact convergence with constant stepsize.
We consider a distributed non-convex optimization where a network of agents aims at minimizing a global function over the Stiefel manifold. The global function is represented as a finite sum of smooth local functions, where each local function is associated with one agent and agents communicate with each other over an undirected connected graph. The problem is non-convex as local functions are possibly non-convex (but smooth) and the Steifel manifold is a non-convex set. We present a decentralized Riemannian stochastic gradient method (DRSGD) with the convergence rate of $\mathcal{O}(1/\sqrt{K})$ to a stationary point. To have exact convergence with constant stepsize, we also propose a decentralized Riemannian gradient tracking algorithm (DRGTA) with the convergence rate of $\mathcal{O}(1/K)$ to a stationary point. We use multi-step consensus to preserve the iteration in the local (consensus) region. DRGTA is the first decentralized algorithm with exact convergence for distributed optimization on Stiefel manifold.