Why are the singular values non-negative? I am working with SVD and trying to understand it. Through some thinking and help here I have a better understanding of it, but I would like some feedback.
Thought process:
$AA^T$ and $A^TA$ share eigenvalues and if they have different dimensions the bigger one has 0-eigenvalues for the eigenvalues not present in the lesser dimension one.
$A^TA\vec{v}=\lambda\vec{v} \Leftrightarrow AA^T(A\vec{v})=\lambda(A\vec{v})$
This means the $A$ projects the orthogonal eigenvectors of $A^TA$ on to the orthogonal eigenvectors of $AA^T$.
If we express vectors in the eigenvectors of $A^TA$ with the matrix $V^T$, scale the components, express the answer in the standard-base eigenvectors of $AA^T$ with matrix $U$ we now have expressed $A=U\Sigma V^T$.
The problem I have is that I don't see why the scaling factors in $\Sigma$ can't be negativ.
$AA^T=U \Sigma \Sigma^TU^T$. The eigenvalues of $AA^T$ are the scaling factors of $\Sigma$, $\sigma_k^2$
The same can be said for $A^TA$
My question is why we assume that all singular values $\sigma_k$ are positive?
 A: If $v$ is an eigenvector of $A^T A$ with corresponding eigenvalue $\lambda$ then $$
\lambda \|v\|^2 = \lambda \langle v, v \rangle = \langle \lambda v, v \rangle = \langle A^T A v, v \rangle = \langle Av, Av \rangle = \|Av\|^2.
$$
Because $\|v\|^2$ and $\|Av\|^2$ are nonnegative, so must $\lambda$ be.  The same argument shows that the eigenvalues of $AA^T$ are nonnegative.  So those particular eigenvalues are going to be nonnegative.
It's certainly also possible to express matrices in the form $U \Sigma V^T$ with $U$ and $V$ real orthogonal and $\Sigma$ diagonal with entries that are not all negative.  For example in the $1 \times 1$ case we have $1 = (-1) \cdot (-1) \cdot 1$ and in the $2 \times 2$ case we have $$
\begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix} = \begin{pmatrix} -1 & 0 \\ 0 & 1 \end{pmatrix} \begin{pmatrix} -1 & 0 \\ 0 & 1 \end{pmatrix} \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix}^T$$ and there are many other variations and generalizations of this theme.  Depending on your application, these sorts of matrix factorizations might be just as useful for some specific purpose as the SVD.  But the SVD does show that it's possible to have a factorization where the diagonal matrix $\Sigma$ has nonnegative entries.  Speaking informally, if you have such a factorization where $\Sigma$ is diagonal but does not have all nonnegative entries, you can "absorb" any negative signs into the choice of either $U$ or $V$.  And for some applications it is useful to be able to do so.  I hope this helps.
