Given any matrix , the singular value decomposition (SVD) is where is an unitary matrix, is an unitary matrix, and is an diagonal matrix where all off-diagonal entries are 0 and the diagonal entries are all non-negative real values. The diagonal entries of are referred to as "singular values".
As an example, consider the shear transformation . The singular value decomposition of is:
The set of all unit length vectors such that form a sphere of radius 1 around the origin. When is applied to this sphere, it becomes an ellipsoid. The principal radii of this ellipsoid are the singular values, and their directions form the columns of .
One fact that is not immediately obvious is that the singular value decomposition always exists:
In essence, any linear transformation is a rotation, followed by stretching or shrinking parallel to each axis (with some dimensions added or zeroed out of existence), followed by another rotation.
The following proof will demonstrate that the singular value decomposition always exists. An outline of the proof will be given first:
Proof outline
We need to prove that an arbitrary linear transform is a rotation: , followed by scaling parallel to each axis: , and lastly followed by another rotation: , giving .
If the columns of are already mutually orthogonal, then the first rotation is not necessary: . The entries of are the lengths of the vectors formed by the columns of , and is a rotation that rotates the elementary basis vectors of to be parallel with the columns of .
In most cases however, the columns of are not mutually orthogonal. In this case, the rotation is non-trivial. , so must be chosen so that the columns of are mutually orthogonal. Let . We need to choose orthonormal vectors so that are all mutually orthogonal. This can be done iteratively. Imagine that we have chosen so that when given any vector that is orthogonal to , that is orthogonal to . The effort now switches to finding an orthonormal set of vectors confined to the space of vectors that are perpendicular to such that are mutually orthogonal.
Let be a unitary matrix with as the first column. Factoring from the left side of to get results in a new set of orthonormal vectors that are the columns of . The goal of having the columns of be mutually orthogonal is converted to having the columns of be mutually orthogonal with effectively replacing . transforms to , and the space of vectors orthogonal to transforms to the space spanned by the standard basis vectors . The first column of is and so is orthogonal to all other columns.
If is a unitary matrix where the first column of is normalized to unit length, then factoring from the left side of to get results in a matrix in which the first column is parallel to the standard basis vector . The first column of is orthogonal to all other columns, so the first column of is orthogonal to all other columns, so hence the first row of contains all 0s except for the first column.
can now be determined recursively with the dimension reduced to , and is replaced with with the first row and column removed. This forms the inductive component of the coming proof.
Lastly, how do we know that there exists so that when given any vector that is orthogonal to , that is orthogonal to ? The answer will be that the unit length that maximizes is a valid .
We are now ready to give the proof in full detail:
- Proof of the existence of the singular value decomposition
This proof will proceed using induction on both and .
- Base Case
has a single row, and therefore has the form where is an arbitrary unit length vector, and is an arbitrary non-negative real number. Note that and exist for any single row matrix .
Let , , and where together form an mutually orthogonal set of unit length vectors. can be determined via Gram-Schmidt orthogonalization. It is clear that: .
- Base Case
has a single column, and therefore has the form where is an arbitrary unit length vector, and is an arbitrary non-negative real number. Note that and exist for any single column matrix .
Let , , and where together form an mutually orthogonal set of unit length vectors. can be determined via Gram-Schmidt orthogonalization. It is clear that: .
- Inductive Case
Let denote the standard basis vector of . Let denote a matrix of 0s.
Maximize subject to the constraint . Let be a unit length vector that maximizes , and let . Let (if , then is an arbitrary unit length vector).
Using Gram-Schmidt orthogonalization, unitary matrices and can be determined such that the first columns of and respectively are and : and . where . It will now be proven that the first column and row of contain all 0s except for the (1,1) entry which contains : .
. This means that the first column of is .
To show that the first row of is , we will show that the first column of is orthogonal to all of the other columns of . This will require exploiting the fact that maximizes subject to the constraint .
Let be a parameterized unit length vector. Let .
Taking the derivative of the constraint gives
being maximized at gives:
( and denote the real and imaginary components of a complex number respectively.)
Let be arbitrary. Let . and are orthogonal.
This gives: .
Now let where the outside of the subscript is the imaginary constant. and are orthogonal.
This gives: .
Hence: .
Therefore, the first column of is orthogonal to all of the other columns of , and has the form: .
is an matrix, and therefore by an inductive argument, . Finally,
where
,
, and
.