{"id":"arxiv/2605.00740","url":"https://arxiv.org/abs/2605.00740","tags":["math.OC","cs.LG","stat.ML"],"score":0,"title":"Randomized Subspace Nesterov Accelerated Gradient","domain":"math.OC","source":"arxiv","authors":["Gaku Omiya","Pierre-Louis Poirion","Akiko Takeda"],"reviews":0,"subject":"math.OC","summary":"Randomized-subspace methods reduce the cost of first-order optimization by using only low-dimensional projected-gradient information, a feature that is attractive in forward-mode automatic differentiation and communication-limited settings. While Nesterov acceleration is well understood for full-gradient and coordinate-based methods, obtaining accelerated methods for general subspace sketches that use only projected-gradient information and can improve over full-dimensional Nesterov acceleration in oracle complexity is technically nontrivial. We develop randomized-subspace Nesterov…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-05-01","reactions":{}}