Sketch Placement Matters in Langevin Sampling
I was messing around with Langevin samplers because I worked with underdamped Langevin diffusion a little bit during my graduate years.
Langevin samplers often swap the true gradient for a cheaper random estimate that is correct on average. The question I wanted to answer: if two estimates are both unbiased, does it matter where you put the randomness? It turns out it does, and the answer is a bit weird. Read the paper.