SVD and low-rank approximation

The singular value decomposition (SVD) factorizes a matrix where are unitary and is diagonal with nonnegative entries. Unlike the eigenvalue decomposition, every (real or complex) matrix is guaranteed to have a SVD. Geometrically, the SVD implies that every linear transformation is equivalent to a rotation, a scaling, then another rotation. In data analysis, the SVD can be computed to perform principal component analysis (PCA), where an ellipsoid is used to “fit” a data matrix. In other words, once the SVD is computed, a partial SVD, can be used to approximate the matrix, where and are the left and right singular vectors, and are the singular values in decreasing order. Equivalently, a partial […]

More