The Sample Complexity of Subspace Learning with Partial Information

The goal of subspace learning is to find a $k$-dimensional subspace of $\mathbb{R}^d$, such that the expected squared distance between instance vectors and the subspace is as small as possible. In this paper we study the sample complexity of subspace learning in a \emph{partial information} setting, in which the learner can only observe $r \le d$ attributes from each instance vector. We derive upper and lower bounds on the sample complexity in different scenarios. In particular, our upper bounds involve a generalization of vector sampling techniques, which are often used in bandit problems, to matrices.