cover

Singular Value Decomposition

August 26, 2021 12 min read

Singular value decomposition is a way of understanding a rectangular (i.e. not necessarily square) matrix from the operator norm standpoint. It is complementary perspective to eigenvalue decomposition that finds numerous application in statistics, machine learning, bioinformatics, quantum computers etc. This post explains its nature and connections to operator norm, least squares fitting, PCA, condition numbers, regularization problems etc.

Intuition for SVD from operator norm perspective

Suppose that you have a rectangular matrix AA, where its columns are people and each element of a column is how much they love horror movies and drama movies (I’m obviously drawing some inspiration from Netflix challenge).

A=(a1,horrora2,horrora3,horrora1,dramaa2,dramaa3,drama)A = \begin{pmatrix} a_{1, \text{horror}} && a_{2, \text{horror}} && a_{3, \text{horror}} \\ a_{1, \text{drama}} && a_{2, \text{drama}} && a_{3, \text{drama}} \\ \end{pmatrix}

Suppose that you are going to estimate the average fractions of your audience that prefer horror or drama movies. So you might be interested in multiplying this matrix by a vector of weights. For instance, if you are Netflix and 0.5 of your income is created by person 1, 0.2 - by person 2 and 0.3 - by person 3, your vector x=(0.50.20.3)x = \begin{pmatrix} 0.5 \\ 0.2 \\ 0.3 \end{pmatrix} and AXAX is a 2-vector that reflects, what fractions of your total income horror movies and drama movies produce.

In the spirit of matrix/operator norms you might ask: if I applied this matrix to a vector of length 1, what is the largest length of vector I might receive on output?

The answer is simple: you need to calculate the following dot product (and then take square root of it to get the length):

d(Ax)2=(Ax,Ax)=xTATAxd(Ax)^2 = (Ax, Ax) = x^T A^T A x

As you can see, in the middle of this dot product we have a square 3-by-3 matrix ATAA^T A, which is symmetric. That means that its eigenvectors are orthogonal/unitary. So we can represent it as an eigen decomposition: ATA=VDVTA^TA = V D V^T, where VV is an orthogonal matrix of eigenvectors.

Then the squared length of vector AxAx takes the form d(Ax)2=(Ax,Ax)=xTATAx=(xTV)D(VTx)d(Ax)^2 = (Ax, Ax) = x^T A^T A x = (x^T V) D (V^T x) and the answer to our question becomes straightforward: the largest length of vector AxAx is achieved when x is the eigenvector uiu_i corresponding to the largest eigenvalue λi\lambda_i. The square roots of elements of matrix D of the eigenvectors of matrix ATAA^TA, are known as singular values and denoted σi=λi\sigma_i = \sqrt{\lambda_i}. We’ve already run into them previously, while exploring the condition numbers.

On the other hand, we can consider a matrix AATAA^T instead of ATAA^TA. It is a 2-by-2 matrix, which, again, is symmetric. Again, it has orthogonal/unitary eigenvectors matrix, which we will denote UU.

Note that matrices ATAA^TA and AATAA^T are known as Gram matrices. As we’ve seen above that each Gram matrix is a square of a vector length by design, Gram matrices are positive semi-definite and, thus, their eigenvalues are non-negative.

Left Gram matrix and right Gram matrix have identical eigenvalues

Consider a rectangular matrix A=(142536)A = \begin{pmatrix} 1 && 4 \\ 2 && 5 \\ 3 && 6 \\ \end{pmatrix}. We can construct two different Gram matrices from it:

AAT=(142536)(123456)A A^T = \begin{pmatrix} 1 && 4 \\ 2 && 5 \\ 3 && 6 \\ \end{pmatrix} \cdot \begin{pmatrix} 1 && 2 && 3 \\ 4 && 5 && 6 \\ \end{pmatrix} of dimensionality 33 x 33 and non-full rank 2

ATA=(123456)(142536)A^T A = \begin{pmatrix} 1 && 2 && 3 \\ 4 && 5 && 6 \\ \end{pmatrix} \cdot \begin{pmatrix} 1 && 4 \\ 2 && 5 \\ 3 && 6 \\ \end{pmatrix} of dimensionality 22 x 22 and full rank 2.

However, both of these matrices have identical eigenvalues. It is surprisingly easy to see this.

Suppose that uiu_i is an eigenvector of ATAA^T A and λi\lambda_i is its corresponding eigenvalue:

ATAui=λiuiA^T A u_i = \lambda_i u_i

Multiply both left and right sides of this expression by AA:

AATAui=λiAuiA A^T A u_i = \lambda_i A u_i

Now from this formula by definition the vector AuiA u_i is an eigenvector for AATA A^T, and λi\lambda_i is the corresponding eigenvalue.

Hence, the eigenvalues Σ\Sigma are identical for AATA A^T and ATAA^T A, and their eigenvectors are also connected (see next part).

Connection between left and right singular vectors

Here comes the key point of SVD: we can express vectors uiu_i through viv_i. I will show how, but first take note that uiu_i is a 2-vector, and viv_i is a 3-vector. Matrix U has only 2 eigenvalues, while the matrix v has 3. Thus, in this example every uiu_i can be expressed through uiu_i, but not the other way round.

Now let us show that ui=Aviσiu_i = \frac{Av_i}{\sigma_i}, where σi=λi\sigma_i = \sqrt{\lambda_i} are singular values of matrix AA. Indeed:

AATui=AATAviσi=A(ATAvi)1σi=Aσi2vi1σi=σi2ui=λiuiAA^T u_i = A A^T \frac{A v_i}{\sigma_i} = A (A^TA v_i) \frac{1}{\sigma_i} = A \sigma_i^2 v_i \frac{1}{\sigma_i} = \sigma_i^2 u_i = \lambda_i u_i

If we re-write ui=Aviλiu_i = \frac{Av_i}{\sqrt{\lambda_i}} in matrix form, we get:

U=AVΣ1U = A V \Sigma^{-1} or, equivalently, UΣ=AVU \Sigma = A V, or A=UΣVTA = U \Sigma V^T.

This proves that singular value decomposition exists if matrices ATAA^TA and AATAA^T have eigenvalue decomposition.

Clarification of notation for eigenvalues matrix

Let us assume that we can always find a decomposition:

A=UΣVT=(u1,1u2,1u3,1u1,2u2,2u2,3u3,1u3,2u3,3)(σ100σ200)(v1,1v1,2v2,1v2,2)A = U \Sigma V^T = \begin{pmatrix} u_{1,1} && u_{2,1} && u_{3,1} \\ u_{1,2} && u_{2,2} && u_{2,3} \\ u_{3,1} && u_{3,2} && u_{3,3} \end{pmatrix} \cdot \begin{pmatrix} \sigma_1 && 0 \\ 0 && \sigma_2 \\ 0 && 0 \\ \end{pmatrix} \cdot \begin{pmatrix} v_{1,1} && v_{1,2} \\ v_{2,1} && v_{2,2} \end{pmatrix}

Then AT=VΣUTA^T = V \Sigma U^T and:

ATA=VΣUTUΣVT=VΣ2VTA^T A = V \Sigma U^T U \Sigma V^T = V \Sigma^2 V^T

AAT=UΣVTVΣUT=UΣ2UTA A^T = U \Sigma V^T V \Sigma U^T = U \Sigma^2 U^T

In both cases of ATAA^T A and AATAA^T this decomposition is consistent with properties of a Gram matrix being symmetric and positive-semidefinite: the eigenvectors of both matrices are orthogonal and eigenvalues are non-negative.

When I write Σ2\Sigma^2 note a notation abuse here: in reality we are multiplying rectangular matrices and resulting matrices Σ2\Sigma^2 are of different dimensionality. In ATAA^TA we call Σ2\Sigma^2 a 33x33 matrix:

Σ2=(σ100σ200)(σ1000σ20)=(σ12000σ220000)\Sigma^2 = \begin{pmatrix} \sigma_1 && 0 \\ 0 && \sigma_2 \\ 0 && 0 \\ \end{pmatrix} \cdot \begin{pmatrix} \sigma_1 && 0 && 0 \\ 0 && \sigma_2 && 0 \\ \end{pmatrix} = \begin{pmatrix} \sigma_1^2 && 0 && 0 \\ 0 && \sigma_2^2 && 0 \\ 0 && 0 && 0 \\ \end{pmatrix}.

In AATA A^T we call Σ2\Sigma^2 a 22x22 matrix:

Σ2=(σ1000σ20)(σ100σ200)=(σ1200σ22)\Sigma^2 = \begin{pmatrix} \sigma_1 && 0 && 0 \\ 0 && \sigma_2 && 0 \\ \end{pmatrix} \cdot \begin{pmatrix} \sigma_1 && 0 \\ 0 && \sigma_2 \\ 0 && 0 \\ \end{pmatrix} = \begin{pmatrix} \sigma_1^2 && 0 \\ 0 && \sigma_2^2 \end{pmatrix}

Matrix norms from singular values perspective

Nuclear norm and Schatten norm

In the condition numbers post we considered two kinds of matrix norms: operator norm and Frobenius norm.

In this post it would be appropriate to mention another family of norms: Schatten norms and their important special case, [Nuclear norm/Trace norm](Mirsky’s inequality)

Trace norm is awesome: if you take a rectangular matrix, apply L1 regularization to it, and apply a Trace lasso to it, you will effectively find low-rank approximations of your matrix, giving you a nice alternative to non-negative matrix approximation! See Trace Lasso paper. Hence, amazingly, the problems of sparse and low-rank approximations are interconnected.

Read more on affine rank minimization problem and sparse identification of nonlinear dynamical systems, if you find this topic interesting.

A new perspective on Frobenius norm

Singular values also provide another perspective on Frobenius norm: it is a square root of sum of squares of singular values:

AF=i=1mj=1nai,j2=k=1min(m,n)σk2\Vert A \Vert_F = \sqrt{\sum \limits_{i=1}^{m} \sum \limits_{j=1}^{n} |a_{i,j}^2|} = \sqrt{\sum \limits_{k=1}^{min(m,n)}\sigma_k^2}

To see this, consider a Gram matrix ATA=(123456)(142536)A^T A = \begin{pmatrix} 1 && 2 && 3 \\ 4 && 5 && 6 \\ \end{pmatrix} \cdot \begin{pmatrix} 1 && 4 \\ 2 && 5 \\ 3 && 6 \\ \end{pmatrix}.

Obviously, the diagonal elements of Gram matrix are sums of squares of the rows of matrix AA.

Hence, the trace of Gram matrix equals to Frobenius norm of matrix AA. At the same time, as we know, trace of Gram matrix equals to the sum of eigenvalues of Gram matrix or sum of squares of singular values of the original matrix AA.

This fact, I believe, gives rise to an uncannily useful family of statements, called Weyl’s perturbation theorem or a similar Hoffman–Wielandt inequality for Hermitian matrices in matrix analysis. Both results are generalized by Mirsky’s inequality.

Informally speaking, they state that if two matrices differ in Frobenius norm by less than cc, their respective eigenvalues (also differ by less that cc. This theorem, for instance, lets us prove that eigenvalues of a Toeplitz matrix converge to the eigenvalues of a circulant matrix.

Operator/spectral norm

Matrix spectral norm is simply its largest singular value.

Application: low-rank matrix approximation

As we’ve seen singular values provide a convenient representation of a matrix as a sum of outer products of column and row-vectors (each outer product, thus, results in a matrix of rank 1):

A=i=1min(m,n)σiuiviT=σ1(u1,1u1,2...u1,m)(v1,1v1,2...v1,n)+σ2(u2,1u2,2...u2,m)(v2,1v2,2...v2,n)+...A = \sum \limits_{i=1}^{min(m,n)} \sigma_i u_i v_i^T = \sigma_1 \cdot \begin{pmatrix} u_{1,1} && u_{1,2} && ... && u_{1,m} \\ \end{pmatrix} \cdot \begin{pmatrix} v_{1,1} \\ v_{1,2} \\ ... \\ v_{1,n} \\ \end{pmatrix} + \sigma_2 \cdot \begin{pmatrix} u_{2,1} && u_{2,2} && ... && u_{2,m} \\ \end{pmatrix} \cdot \begin{pmatrix} v_{2,1} \\ v_{2,2} \\ ... \\ v_{2,n} \\ \end{pmatrix} + ...

Here’s the catch: as singular values are ordered in decreasing order, we can use SVD as a means of compression of our data.

If we take only a subset of first kk elements of our SVD sum, instead of the full min(m,n)min(m,n) elements, it is very likely that we would preserve most of the information, contained in our data, but represent it with only a limited number of eigenvectors. This feels very similar to Fourier series. This is also a reason, why PCA works (it is basically a special case of SVD).

References


Boris Burkov

Written by Boris Burkov who lives in Moscow, Russia, loves to take part in development of cutting-edge technologies, reflects on how the world works and admires the giants of the past. You can follow me in Telegram