Faster SVD-Truncated Least-Squares Regression

Abstract: We develop a fast algorithm for computing the "SVD-truncated" regularized solution to the least-squares problem: $ \min_{\x} \TNorm{\matA \x - \b}. $ Let $\matA_k$ of rank $k$ be the best rank $k$ matrix computed via the SVD of $\matA$. Then, the SVD-truncated regularized solution is: $ \x_k = \pinv{\matA}k \b. $ If $\matA$ is $m \times n$, then, it takes $O(m n \min{m,n})$ time to compute $\x_k $ using the SVD of \math{\matA}. We give an approximation algorithm for \math{\x_k} which constructs a rank-\math{k} approximation $\tilde{\matA}{k}$ and computes $ \tilde{\x}{k} = \pinv{\tilde\matA}{k} \b$ in roughly $O(\nnz(\matA) k \log n)$ time. Our algorithm uses a randomized variant of the subspace iteration. We show that, with high probability: $ \TNorm{\matA \tilde{\x}_{k} - \b} \approx \TNorm{\matA \x_k - \b}$ and $\TNorm{\x_k - \tilde\x_k} \approx 0. $