Broximal Gradient Descent: A Projection-Free Sister of Projected Gradient Descent
The paper introduces Broximal Gradient Descent, a projection-free method that aims to be a 'sister' to Projected Gradient Descent (PGD). It replaces global linear minimization with local minimization over a small ball around the current iterate. This approach yields strong convergence guarantees for constrained optimization without needing the constraint set to be bounded.