Sparse spectral estimation with missing and corrupted measurements

• DOI: 10.48550/arxiv.1811.10443
• Publication Date: 2018-01-01
• Authors: Andreas Elsener; Sara van de Geer
• Keywords: Imputation (statistics); Matrix Completion; Rank (graph theory)

Supervised learning methods with missing data have been extensively studied not just due to the techniques related low-rank matrix completion. Also in unsupervised one often relies on imputation methods. As a matter of fact, values induce bias various estimators such as sample covariance matrix. In present paper, convex method for sparse subspace estimation is extended case and corrupted measurements. This done by correcting instead imputing values. The estimator then used an initial value nonconvex procedure improve overall statistical performance. methodological well theoretical frameworks are applied wide range problems. These include Principal Component Analysis different types randomly eigenvectors matrices Finally, performance demonstrated synthetic data.