Query Lower Bounds for Log-Concave Sampling
Explore the fundamental challenges in generative machine learning through this 28-minute lecture on query lower bounds for log-concave sampling. Delve into the central problem of sampling from probability distributions with prescribed densities, examining both simple and sophisticated algorithms like rejection sampling and Langevin-based models. Focus on the converse question of finding universal complexity lower bounds, specifically for cases where the log-density is a strictly concave smooth function. Discover how tight bounds in low dimensions are constructed using a modified version of Perron's sprouting technique for Kakeya sets. Gain insights from joint research conducted with experts from institutions such as IAS, Microsoft Research, and MIT.