{"id":"arxiv/2610.15723","url":"https://arxiv.org/abs/2610.15723","tags":["math.ST","cs.DS","cs.LG","stat.ME","stat.ML"],"score":0,"title":"Adaptive Robust Confidence Sets in High Dimensions","domain":"math.ST","source":"arxiv","authors":["Ilias Diakonikolas","Chao Gao","Daniel M. Kane","Yuetian Luo"],"reviews":0,"subject":"math.ST","summary":"We study adaptive confidence sets for mean vectors of multivariate Gaussians with identity covariance under Huber contamination when the contamination proportion is unknown. We show that an information-theoretically optimal adaptive confidence set can be constructed by simultaneously quantifying uncertainty across all one-dimensional projections. The resulting confidence set achieves the optimal diameter and reveals a substantial adaptation cost compared with the minimax estimation error. Although statistically optimal, this construction is computationally infeasible. We establish a…","category":"research","comments":0,"disputed":false,"division":"math","mentions":0,"discussed":0,"published":"2026-10-11","reactions":{}}