Singular Value Decomposition (SVD)

if 𝐴𝑥 = 𝜆𝑥 & ||𝑥|| = 1, then:

  • ||𝐴𝑥|| = ||𝜆𝑥||
  • ||𝐴𝑥|| = |𝜆| ||𝑥||
  • ||𝐴𝑥|| = |𝜆|

SVD - Definition (Real and/or Complex Matrix)

The SVD of an 𝑚✕𝑛 real and/or complex matrix 𝐴 with rank 𝑟 is:

  • 𝐴 = 𝑈𝛴𝑉*

where:

  • 𝑈 - 𝑚✕𝑚 real or complex unitary matrix
  • 𝛴 - 𝑚✕𝑛rectangular matrix with diagonal entries of the first 𝑟 singular values of 𝐴 (𝜎1 ≥ 𝜎2 ≥ … ≥ 𝜎𝑟)
  • 𝑉 - 𝑛✕𝑛 real or complex unitary matrix (𝑉* - complex conjugate of 𝑉)

SVD - Definition (Real Matrix)

If 𝐴 is real, 𝑈 and 𝑉*=𝑉𝑇 are real orthogonal matrices (where 𝑉𝑇 is the transpose)

  • 𝐴 = 𝑈𝛴𝑉𝑇

where:

  • 𝑈 - orthogonal 𝑚✕𝑚unitary matrix. normalized eigenvectors of the symmetric matrix 𝐴𝐴𝑇 (𝑖.𝑒. 𝑈𝑇𝑈=𝐼) columns of 𝑈 are called the left singular vectors of 𝐴
  • 𝑉 - orthogonal 𝑛✕𝑛 unitary matrix. normalized eigenvectors of the symmetric matrix 𝐴𝑇𝐴 (𝑖.𝑒. 𝑉𝑇𝑉=𝐼) columns of 𝑉 are called the right singular vectors of 𝐴
  • 𝛴 - diagonal matrix of singular values
  • 𝛴 = 𝐷(1/2) where 𝐷 is the diagonal matrix of the eigenvalues of matrix 𝐴𝐴𝑇and 𝐴𝑇𝐴

derivation:

The Invertible Matrix Theorem

let 𝐴 be 𝑛✕𝑛 matrix. then the following statements are equivalent to the statement that 𝐴 is an invertible matrix:

  • (𝐶𝑜𝑙𝐴)= {0}
  • (N𝑢𝑙𝑙𝐴) = ℝ𝑛
  • 𝑅𝑜𝑤𝐴 = ℝ𝑛
  • 𝐴 has 𝑛 non-zero singular values

SVD - Reduced SVD & Pseudo Inverse (Moore-Penrose Inverse)

Indent

𝑈𝑟𝐷𝑉𝑟𝑇is called reduced SVD

pseudoinverse (Moore-Penrose Inverse) of 𝐴:

  • 𝐴= 𝑉𝑟𝐷-1𝑈𝑟𝑇
application to least-squares solution

given the equation 𝐴𝑥 = 𝑏, use pseudoinverse of 𝐴

  • 𝑥̂ = 𝐴𝑏
  • 𝑥̂ = 𝑉𝑟𝐷-1𝑈𝑟𝑇𝑏

then from SVD:

  • 𝐴𝑥̂ = 𝑈𝑟𝐷𝑉𝑟𝑇𝑉𝑟𝐷-1𝑈𝑟𝑇𝑏
  • 𝐴𝑥̂ = 𝑈𝑟𝑈𝑟𝑇𝑏

SVD - Intuition (As Linear Transformations)

When 𝐴 is a Square 𝑚✕𝑚 Real Matrix
  • 𝐴 is a linear transformation of space ℝ𝑚
  • SVD breaks down any invertible 𝑀 into 3 geometric transformations:
    • 𝑉* - rotation or reflection
    • 𝐷 - scaling
    • 𝑈 - rotation or reflection
  • 𝑈 & 𝑉* can be chosen as 𝑚✕𝑚 real matrices. Therefore, unitary becomes orthogonal and therefore both 𝑈 & 𝑉* can be thought as rotations and/or reflections
  • if the determinant of 𝐴 is:
    • positive - 𝑈 & 𝑉* is either both reflections or both rotations
    • negative - exactly one is reflection
    • zero - anything goes
When 𝐴 is a Rectangular 𝑚✕𝑛 Real Matrix
  • 𝐴 is a linear transformation of space ℝ𝑛to ℝ𝑚
  • SVD breaks down any invertible 𝐴 into 3 geometric transformations:
    • 𝑉* - rotation or reflection in space ℝ𝑛
    • 𝐷 - scaling from ℝ𝑛to ℝ𝑚
    • 𝑈 - rotation or reflection in space ℝ𝑚

SVD - Subpages

Resources