Megadose AI progress, ranked and analyzed.

Improved Gradient Descent Lower Bounds Beyond Nesterov

· ArXiv · AI/CL/LG ·
Predetermined stepsizes cannot push gradient descent as far as earlier bounds left open.

Ye and Liu prove new lower bounds for smooth convex optimization: $\Omega(n^{-1.6342})$ in the non-anytime setting and $\Omega(n^{-1.2408})$ in the anytime setting. The paper says these improve prior lower bounds from Ma and Chen and from Tsai et al. It also claims the anytime result, compared with known silver-schedule rates, separates what is achievable in anytime versus non-anytime regimes. ArXiv · AI/CL/LG's note

score 5

Categories: Research