Minimize a differentiable convex function over a compact convex set \(\mathcal X\). Start with \(\mathbf x_0\in\mathcal X\).

1. Minimize the linear approximation:

\[\mathbf s_t\in\arg\min_{\mathbf u\in\mathcal X}\langle\nabla f(\mathbf x_t),\mathbf u\rangle.\]

2. Move toward that feasible point:

\[\mathbf x_{t+1}=(1-\eta_t)\mathbf x_t+\eta_t\mathbf s_t,\qquad 0\leq\eta_t\leq1.\]

Standard schedule: \(\eta_t=2/(t+2)\). Alternatively, minimize \(f(\mathbf x_t+\eta(\mathbf s_t-\mathbf x_t))\) over \(\eta\in[0,1]\).

For a box: choose the lower coordinate bound when \([\nabla f(\mathbf x_t)]_j\geq0\), and the upper bound otherwise. The update is feasible by convexity.

Frank–Wolfe / conditional gradient · Jaggi (2013), Revisiting Frank-Wolfe