WorksMathematics
A Euclidean Quasi-Newton Method for Matrix-Manifold Optimization via Almost-Global Coordinates
Abstract
We study smooth matrix-manifold optimization using objective values and first derivatives. Almost-global coordinates transform the manifold-constrained problem into an unconstrained pullback problem, to which we apply BFGS with a chart-restart strategy and a limited-memory implementation. We establish correspondence of first- and second-order stationary points and minimizers, together with conditional convergence and local q-superlinear convergence. The local rate follows from Euclidean full-memory BFGS theory once the chart is fixed, without the locking condition required by classical transported-gradient Riemannian BFGS analyses. Tests on 15 benchmark problems from Manopt and Pymanopt, with ten instances per problem, show that our limited-memory method solves all 150 cases, whereas Riemannian BFGS and Riemannian steepest descent solve 132 and 117 cases, respectively. On the 115 cases successfully solved by all three methods, our method takes roughly half the time of either baseline, measured by geometric-mean solve time.