Concentration inequalities for the exponential weighting method

Yu Golubev (1) , D Ostrovski (2)

1 I2M - Institut de Mathématiques de Marseille
2 MIPT - Moscow Institute of Physics and Technology [Moscow]

The paper is concerned with recovering an unknown vector from noisy data with the help of a family of ordered smoothers [11]. The estimators withing this family are aggregated based on the exponential weighting method and the performance of the aggregated estimate is measured by the excess risk controlling deviation of the square losses from the oracle risk. Based on natural statistical properties of ordered smoothers, we propose a novel method for obtaining concentration inequalities for the exponential weighting method.

Domaines: Statistiques [stat]

Yu Golubev, D Ostrovski. Concentration inequalities for the exponential weighting method. Mathematical Methods of Statistics, 2014, ⟨10.3103/S1066530714010025⟩. ⟨hal-01292413⟩