Two slides: Frank-Wolfe is an approximation of BroxGD, itself approximating BPM
peter_richtarik · x · 2026-11-05
Optimization researcher Peter Richtárik distills the Frank-Wolfe (FW) method into two slides:
- FW is an approximation of Broximal Gradient Descent (BroxGD), a recently rediscovered projection-free sister method to Proximal/Projected Gradient Descent;
- BroxGD is itself an approximation of the Broximal Point Method (BPM).
The accompanying slide lays out a unified derivation chain: start from BPM, linearize f at xk to get BroxGD, then replace the radius-tk ball around xk with an inner approximation to obtain Frank-Wolfe. One-line summary: BroxGD linearizes f; Frank-Wolfe also shrinks the ball to an inner approximation.
Related event: KAUST's BroxGD Paper Clarifies Frank-Wolfe as Projection-Free Approximation(3 posts)→
More from Research
- New paper: LLM-based simulators for policy design must generalize beyond frozen in-distribution data — soumitrashukla9 · 2026-11-05
- LLMs learn to detect activation monitoring and hide info from the exact layer being watched — Sauers_ · 2026-11-05
- Agent grid eval: center agent scores 1.00, corners drop to 0.46 on map reconstruction — cephaloform · 2026-11-05
- LLM agent grid develops ASCII dialect: "ARE YOU LLM?" emerges at tick 30 — cephaloform · 2026-11-05
- Nine LLM agents in a grid invent their own 128-bit protocol to talk via neighbor buffers — cephaloform · 2026-11-05
- Datalab finds judging errors on 68 of 200 pages in olmOCR-bench, shares 5 checks — VikParuchuri · 2026-11-05