2.5 Low-rank approximation

One of the reasons the SVD is so widely used is that it can be used to find the best low rank approximation to a matrix. Before we discuss this, we need to define what it means for some matrix \(\bB\) to be a good approximation to \(\bA\). To do that, we need the concept of a matrix norm.

2.5.1 Matrix norms

In Section 1.3.1 we described norms on vectors. Here will extend this idea to include norms on matrices, so that we can discuss the size of a matrix \(||\bA||\), and the distance between two matrices \(||\bA-\bB||\). There are two particular norms we will focus on. The first is called the Frobenious norm (or sometimes the Hilbert-Schmidt norm).

Definition 2.1 Let \(\bA \in \mathbb{R}^{n\times p}\). The Frobenius norm of \(\bA\) is \[||\bA||_F = \left(\sum_{i=1}^n\sum_{j=1}^p |a_{ij}|^2\right)^{\frac{1}{2}}=(\tr \bA^\top\bA)^{\frac{1}{2}} \] where \(a_{ij}\) are the individual entries of \(\bA\).

Note that the Frobenius norm is invariant to rotation by an orthogonal matrix \(\bU\): \[\begin{align*} ||\bA\bU||_F^2 &= \tr(\bU^\top \bA^\top \bA\bU)\\ &=\tr(\bU\bU^\top \bA^\top \bA)\\ &= \tr(\bA^\top\bA)\\ &= ||\bA||_F^2. \end{align*}\]

Proposition 2.10 \[||\bA||_F = \left(\sum_{i=1}^r \sigma_i^2\right)^{\frac{1}{2}}\] where \(\sigma_i\) are the singular values of \(\bA\), and \(r = \operatorname{rank}(\bA)\).

Proof. Using the (non-compact) SVD \(\bA = \bU \bSigma \bV^\top\) we have \[||\bA||_F=||\bU^\top \bA||_F = ||\bU^\top \bA \bV||_F = ||\bSigma||_F=\tr(\bSigma^\top\bSigma)^\frac{1}{2}=\left(\sum \sigma_i^2 \right)^\frac{1}{2}.\]

We previously defined the p-norms for vectors in \(\mathbb{R}^p\) to be \[||\bx||_p = \left(\sum |x_i|^p\right)^{\frac{1}{p}}.\] These vector norms induce matrix norms, sometimes also called operator norms:

Definition 2.2 The p-norms for matrices are defined by \[||\bA||_p = \sup_{\bx\not=0} \frac{||\bA\bx||_p}{||\bx||_p} = \sup_{\bx: ||\bx||_p=1} ||\bA\bx||_p\]

Proposition 2.11 \[||\bA||_2 = \sigma_1\] where \(\sigma_1\) is the first singular value of \(\bA\).

Proof. By Proposition 2.8.

2.5.2 Eckart-Young-Mirsky Theorem

Now that we have defined a norm (i.e., a distance) on matrices, we can think about approximating a matrix \(\bA\) by a matrix that is easier to work with. We have shown that any matrix can be split into the sum of rank-1 component matrices \[\bA = \sum_{i=1}^r \sigma_i \bu_i \bv_i^\top\] We’ll now consider a family of approximations of the form \[\begin{equation} \bA_k = \sum_{i=1}^k \sigma_i \bu_i \bv_i^\top \tag{2.3} \end{equation}\] where \(k\leq r=\operatorname{rank}(\bA)\). This is a rank-k matrix, and as we’ll now show, it is the best possible rank-k approximation to \(\bA\).

Theorem 2.1 (Eckart-Young-Mirsky) For either the 2-norm \(||\cdot||_2\) or the Frobenious norm \(||\cdot||_F\) \[||\bA-\bA_k|| \leq ||\bA-\bB|| \mbox{ for all rank-k matrices }\bB.\] Moreover, \[||\bA-\bA_k|| =\begin{cases} \sigma_{k+1} &\mbox{for the }||\cdot||_2 \mbox{ norm}\\ \left(\sum_{i=k+1}^r \sigma_{i}^2\right)^{\frac{1}{2}} &\mbox{for the }||\cdot||_F \mbox{ norm.}\\ \end{cases}\]

Proof. The last part follows from Propositions 2.11 and 2.10.

Non-examinable: this is quite a tricky proof, but I’ve included it as its interesting to see. We’ll just prove it for the 2-norm. Let \(\bB\) be an \(n\times p\) matrix of rank \(k\). The null space \(\mathcal{N}(\bB)\subset\mathbb{R}^p\) must be of dimension \(p-k\) by the rank nullity theorem.

Consider the \(p \times (k+1)\) matrix \(\bV_{k+1}=[\bv_1\; \ldots \;\bv_{k+1}]\). This has rank \(k+1\), and has column space \(\mathcal{C}(\bV_{k+1})\subset \mathbb{R}^{p}\). Because \[\dim \mathcal{N}(\bB)+\dim \mathcal{C}(\bV_{k+1})=p-k+k+1=p+1\] we can see that \(\mathcal{N}(B)\) and \(\mathcal{C}(V_{k+1})\) cannot be disjoint spaces (as they are both subsets of the p-dimensional space \(\mathbb{R}^p\)). Thus we can find \(\bw \in \mathcal{N}(B)\cap\mathcal{C}(V_{k+1})\), and moreover we can choose \(\bw\) so that \(||\bw||_2=1\).

Because \(\bw \in \mathcal{C}(\bV_{k+1})\) we can write \(\bw = \sum_{i=1}^{k+1}w_i \bv_i\) with \(\sum_{i=1}^{k+1}w_i^2=1\).

Then \[\begin{align*} ||\bA-\bB||_2^2 &\geq ||(\bA-\bB)\bw||_{2}^2 \quad \mbox{ by definition of the matrix 2-norm}\\ &=||\bA\bw||_2^2 \quad \mbox{ as } \bw\in \mathcal{N}(\bB)\\ &=\bw^\top \bV \bSigma^2\bV^\top \bw \quad\mbox{ using the SVD} \bA=\bU\bSigma \bV^\top\\ &=\sum_{i=1}^{k+1}\sigma_i^2 w_i^2 \quad\mbox{ by substituting }\bw = \sum_{i=1}^{k+1}w_i \bv_i\\ &\geq \sigma_{k+1}^2 \sum_{i=1}^{k+1} w_i^2\quad\mbox{ as } \sigma_1\geq\sigma_2\geq\ldots\\ &= \sigma_{k+1}^2 \quad\mbox{ as } \sum_{i=1}^{k+1}w_i^2=1\\ &=||\bA-\bA_k||_2^2 \end{align*}\] as required

This best-approximation property is what makes the SVD so useful in applications.

2.5.3 Example: image compression

As an example, lets consider the image of some peppers from the USC-SIPI image database.

library(tiff)
library(rasterImage)
peppers<-readTIFF("figs/Peppers.tiff")
plot(as.raster(peppers))

This is a \(512 \times 512\) colour image, meaning that there are three matrices \(\bR, \bB,\bG\) of dimension \(512\times 512\)) giving the intensity of red, green, and blue for each pixel. Naively storing this matrix requires 5.7Mb.

We can compute the SVD of the three colour intensity matrices, and the view the image that results from using reduced rank versions \(\bB_k, \bG_k, \bR_k\) instead (as in Equation (2.3)). The image below is formed using \(k=5, 30, 100\), and \(300\) basis vectors.

svd_image <- function(im,k){
  s <- svd(im)
  Sigma_k <- diag(s$d[1:k])
  U_k <- s$u[,1:k]
  V_k <- s$v[,1:k]
  im_k <- U_k %*% Sigma_k %*% t(V_k)
   ## the reduced rank SVD produces some intensities <0 and >1. 
  # Let's truncate these
  im_k[im_k>1]=1
  im_k[im_k<0]=0
  return(im_k)
}

par(mfrow=c(2,2), mar=c(1,1,1,1))

pepprssvd<- peppers
for(k in c(4,30,100,300)){
  svds<-list()
  for(ii in 1:3) {
    pepprssvd[,,ii]<-svd_image(peppers[,,ii],k)
  }
  plot(as.raster(pepprssvd))
}

You can see that for \(k=30\) we have a reasonable approximation, but with some errors. With \(k=100\) it is hard to spot the difference with the original. The size of the four compressed images is 45Kb, 345Kb, 1.1Mb and 3.4Mb.

You can see further demonstrations of image compression with the SVD here.

We will see much more of the SVD in later chapters.