{"id":"arxiv/2603.20929","url":"https://arxiv.org/abs/2603.20929","tags":["stat.ML","cs.LG","math.ST","stat.CO"],"score":0,"title":"Stability of Sequential and Parallel Coordinate Ascent Variational Inference","domain":"stat.ML","source":"arxiv","authors":["Debdeep Pati"],"reviews":0,"subject":"stat.ML","summary":"We highlight a striking difference in behavior between two widely used variants of coordinate ascent variational inference: the sequential and parallel algorithms. While such differences were known in the numerical analysis literature in simpler settings, they remain largely unexplored in the optimization-focused literature on variational inference in more complex models. Focusing on the moderately high-dimensional linear regression problem, we show that the sequential algorithm, although typically slower, enjoys convergence guarantees under more relaxed conditions than the parallel variant,…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-03-21","reactions":{}}