Current browse context:
math.ST
Change to browse by:
References & Citations
Mathematics > Statistics Theory
Title: A concentration inequality for the excess risk in least-squares regression with random design and heteroscedastic noise
(Submitted on 16 Feb 2017 (v1), last revised 9 Mar 2018 (this version, v2))
Abstract: We prove a new and general concentration inequality for the excess risk in least-squares regression with random design and heteroscedastic noise. No specific structure is required on the model, except the existence of a suitable function that controls the local suprema of the empirical process. So far, only the case of linear contrast estimation was tackled in the literature with this level of generality on the model. We solve here the case of a quadratic contrast, by separating the behavior of a linearized empirical process and the empirical process driven by the squares of functions of models.
Submission history
From: Adrien Saumard [view email][v1] Thu, 16 Feb 2017 17:35:06 GMT (12kb)
[v2] Fri, 9 Mar 2018 14:25:00 GMT (13kb)
Link back to: arXiv, form interface, contact.