backtracing / sight /sources /annotated_chunks_window20 /0Qws8BuK3RQ_1_chunk9.json
rewang's picture
Upload
277287c
[
{
"text": "Don't change."
},
{
"text": "That's the key point about singular values."
},
{
"text": "If I multiply, so A has an SVD U sigma V transpose."
},
{
"text": "And QA will have the SVD QU sigma V transpose."
},
{
"text": "So all I've changed when I multiply by Q, all I changed was the first factor, the first orthogonal factor in the SVD."
},
{
"text": "I didn't change the sigmas."
},
{
"text": "They're still sitting there."
},
{
"text": "And of course, I could do also Q on the other side."
},
{
"text": "Different Q, same Q or a different Q on the other side would show up here."
},
{
"text": "And would not change the sigmas, and therefore would not change the Frobenius norm."
},
{
"text": "So these are important properties of this Frobenius norm."
},
{
"text": "It looks messy to write down in that form, but it's much nicer in these forms and in that form."
},
{
"text": "So now if I can just, then we saw that it involves traces."
},
{
"text": "So let me make a few observations about traces."
},
{
"text": "So I'll just, we want to be able to play with traces, and that's something we really haven't done."
},
{
"text": "Here's a fact, that the trace of A transpose B is equal to the trace of B transpose A."
},
{
"text": "Of course, if B is A, it's clear."
},
{
"text": "And it's equal to the trace of B A transpose."
},
{
"text": "So you can do little changes in your matrix without changing the trace."
},
{
"text": "Let's see why one of these is true."
}
]