Singular Value Decomposition

August 26, 2021 7 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=( = \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.

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\sigma_i are singular values of matrix UU (). Indeed:

AATui=AATAviσi=A(ATAvi)1σi=Aσi2vi1σi=σi2uiAA^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

If we re-write ui=Aviλiu_i = \frac{Av_i}{\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.

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

TODO: connection to L1-norm, compressed sensing, affine rank minimization problem, sparse identification of nonlinear dynamical systems.

A new perspective on Frobenius norm

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

TODO: prove this fact!

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 minm,nmin{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).


Boris Burkov

Written by Boris Burkov who lives in Moscow, Russia and Cambridge, UK, 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