Singular value decomposition

Every real matrix has a singular value decomposition (SVD).

The SVD takes a \(m \times n\) matrix \(\textbf{A}\) and decomposes it as follows:

\(\textbf{A} = \textbf{U} \textbf{D} \textbf{V}^T\)

where \(\textbf{U}\) is a \(m \times m\) orthogonal matrix, \(\textbf{D}\) is a \(m \times n\) diagonal matrix and \(\textbf{V}\) is a \(n \times n\) orthogonal matrix.

The columns of \(\textbf{U}\) are called the left-singular vectors, the columns of \(\textbf{V}\) are called the right-singular vectors and the values along the diagonal of \(\textbf{D}\) are called the singular values.

The left-singular vectors are actually the eigenvectors of \(\textbf{A} \textbf{A}^T\), the right-singular vectors are the eigenvectors of \(\textbf{A}^T \textbf{A}\) and the nonzero singular values are the square roots of the eigenvalues of \(\textbf{A}^T \textbf{A}\) (and of \(\textbf{A} \textbf{A}^T\)).

Sources