Proving existence of SVD

516 Views Asked by At

I'm having problems following the proof shown on page 5 here from the note of MIT OCW $18.335J$ introduction to numerical methods class.

I can't see how you get from eqn $1.28$ to $1.29$?

Can someone please break it down step by step?

I seem to be totally confused I don't know where the 0 vector comes from? I don't know where the $\omega$ comes from or where the B comes from?

The best I could come up with is that $u_1 = \frac{A v_1}{\sigma_1}$

Then $U^*AV = \frac{v_1^*A^*Av_1}{\sigma_1}$

This is clearly not the full solution for the upper left entry.

Thanks

Baz

1

There are 1 best solutions below

2
On BEST ANSWER

There is a typo.

Rather than $u_1^*Av_1$, the left hand side should be $U_1^*AV_1$.

They define $S=U_1^*AV_1$ and equation $(1.29)$ claims that the first column has at most one non-zero entry at the very beginning followed by zero.

To see that the $(1,1)$-entry of $S$ is $\sigma_1$:

$$u_1^*Av_1=\frac{\|Av_1\|^2}{\sigma_1}=\frac{\sigma_1^2}{\sigma_1}=\sigma_1$$

Remark: if $\sigma_1=0$, the matrix is the zero matrix and existence of SVD for the zero matrix is trivial.

Now we show that the $(j,1)$-entry of $S$ is $0$ when $j>1$:

$$u_j^*Av_1=u_j^*(\sigma_1u_1)=\sigma_1(u_j^*u_1)=0$$

since $u_j$ and $u_1$ are orthogonal.

$w^*$ and $B$ are meant to denote the other part of the matrix of $S$.