Skip to content
nonconvexcard

The Frank–Wolfe duality gap certificate

ProofIntermediate⏱ 30 minAdvanced Optimizationfrank-wolfeduality gapstopping criterion

Problem

Show that the quantity gapt=⟨∇f(xt),xt−st⟩\mathrm{gap}_t = \langle\nabla f(x_t), x_t - s_t\rangle computed at each Frank–Wolfe step, where st=arg⁡min⁡s∈C⟨∇f(xt),s⟩s_t = \arg\min_{s\in\mathcal C}\langle\nabla f(x_t),s\rangle, is an upper bound on f(xt)−f⋆f(x_t)-f^\star. Explain why this makes Frank–Wolfe self-certifying.

Write a complete argument. Every inequality you use should be one you could state precisely.

Solution

Opening this before you have a real attempt on paper is the fastest way to feel like you understand something you cannot reproduce.

Related problems

All problems
SolvedProblemFieldTypeDifficultyTime
The proximal operator of the l1 normAdvanced OptimizationProofBeginner15 min
A convex problem with a duality gapAdvanced OptimizationProofAdvanced30 min
Projection onto the probability simplexAdvanced OptimizationProofAdvanced1 hour
Mirror descent on the simplex beats Euclidean projection abyssAdvanced OptimizationProofGraduate1 hour