Abstract:  

In this talk, asymptotic properties of  inverted sample covariance matrices are introducted. Moreover, we find that the occurrence of the double descent phenomenon is linked to the inverted sample covariance matrix. From random matrix theory, we give an theoretical explanation of the double descent/ascent phenomenon from the inverted sample covariance matrix.  Based on this, a regularization method is proposed to eliminate the double ascent phenomenon. Simulation studies are demonstrated to show the good performance of the regularization method.

Bio:

郑术蓉， 东北师范大学教授，主要研究方向为：大维随机矩阵理论及其应用。相关研究成果发表在Annals of Statistics、Journal of the American Statistical Association、Biometrika、Bernoulli等学术期刊上。现任 Statistica Sinica, Journal of Business & Economic Statistics, Journal of Multivariate Analysis 等多个学术期刊编委。