Using the sample covariance to estimate the principal components of a hidden covariance matrix is a central problem in statistics and data science. We view this as a perturbation problem where the noise is a zero mean random matrix which is the difference between the sample covariance matrix and the truth. In certain special settings, it can also be viewed as a deformed random matrix model. In this talk, using recent progress on matrices with random perturbation, we derive new estimates, which extend and improve several well known results in the field. (join work with V. Vu)