Modifying a decomposition to obtain a singular value decomposition of matrix

32 Views Asked by At

I've encountered the following question:

Assume that we have \begin{equation*} A = \begin{pmatrix} 1 & -1 & 1\\ -1 & 1 & 1 \\ 2 & 1 & 0 \end{pmatrix} \begin{pmatrix} 1 & 0 & 0\\ 0 & -2 & 0 \\ 0 & 0 & 2 \end{pmatrix} \begin{pmatrix} 2 & -1 & 1\\ 1 & 2 & 0 \\ -2 & 1 & 5 \end{pmatrix} \end{equation*}

Now, label the columns of leftmost matrix as $a_{1}, a_{2}$ and $a_{3}$, the diagonal entries of the middle matrix by $t_{1}, t_{2}$ and $t_{3}$, and finally the rows of the rightmost matrix by $b_{1}^{T}, b_{2}^{T}$ and $b_{3}^{T}$. The question is:

Question: Show how to modify the $a_{i}, b_{i}$ and $t_{i}$'s to yield a singular value decomposition for $A$.

Here, I can write $A$ as sum of rank-1 matrices using $a_{i}, b_{i}$ and $t_{i}$'s (which was asked in the latter part of this question). But I don't know what the question mean by modifying $a_{i}, b_{i}$ and $t_{i}$'s? Can anyone help?

Thanks in advance.

1

There are 1 best solutions below

1
On BEST ANSWER

Let us take the following notations:

$$M= \underbrace{\begin{pmatrix} 1 & -1 & 1\\ -1 & 1 & 1 \\ 2 & 1 & 0 \end{pmatrix}}_A \underbrace{\begin{pmatrix} 1 & 0 & 0\\ 0 & -2 & 0 \\ 0 & 0 & 2 \end{pmatrix}}_T \underbrace{\begin{pmatrix} 2 & -1 & 1\\ 1 & 2 & 0 \\ -2 & 1 & 5 \end{pmatrix}}_B$$

$A$ has orthogonal (but not normalized) columns.

$B$ has orthogonal (but not normalized) lines.

$T$ is diagonal but one its entries is negative (instead of all positive).

These deficiencies can be compensated by inserting diagonal matrices in the following way:

$$M=A\underbrace{\begin{pmatrix} 1/\sqrt{6} & 0 & 0\\ 0 & 1/\sqrt{3} & 0 \\ 0 & 0 & 1/\sqrt{2} \end{pmatrix}}_{D_1} D_2D_2 T \underbrace{\begin{pmatrix} 1/\sqrt{6} & 0 & 0\\ 0 & 1/\sqrt{3} & 0 \\ 0 & 0 & 1/\sqrt{30} \end{pmatrix}}_{D_3}B$$

with $$D_2:=\begin{pmatrix} 1 & 0 & 0\\ 0 & -1 & 0 \\ 0 & 0 & 1 \end{pmatrix}$$

Setting $$U=AD_1D_2, \ \ \ \Sigma=D_2T, \ \ \ V^T=D_3B,$$ we recover the SVD of $M$:

$$M=U\Sigma V^T$$