Singular Value Decomposition

M=UΣVTM = U \Sigma V^T

Here, MM is an m×nm \times n matrix, UU is an m×mm \times m orthogonal matrix, Σ\Sigma is a diagonal matrix and VV is an n×nn \times n orthogonal matrix.

The entries in the diagonal matrix Σ\Sigma consist of the square roots of the non-zero eigenvalues of the matrix MTMM^T M, ordered from greatest to least. They are called the Singular Values of M, denoted σ1,...,σn\sigma_1, ..., \sigma_n.

The columns of VV are the normalized eigenvectors belonging to the previously mentioned eigenvalues, also ordered from greatest to least. These column vectors are called the right singular vectors of M Note: VV should be an orthogonal matrix, so you may need to apply the Gram-Schmidt process to orthogonalize the eigenvectors.

The columns of UU, called the left singular vectors of M, are calculated with the following formula, where v1,vn\vec{v_1}, \vec{v_n} are the column vectors of VV:

1σ1Mv1,...1σnMv2\frac{1}{\sigma_1} M \vec{v_1}, ... \frac{1}{\sigma_n} M \vec{v_2}

The columns of UU should add up to be an orthonormal basis for RmR^m. If mm > nn, you will need to apply the Gram-Schmidt process (or insight) to complete the basis.

The SVD splits the transformation MM into three distinct consecutive moves: a rotation (VV), a scaling (Σ\Sigma) and a final rotation (UU).

Last updated