Skip to content

test SVD & PCA

Let $A \in R^{m * n}$, wlog assume $m \geq n, Rank(A) = r$ $$ A'A = U DU'

$$

$U = [u_1,\cdots, u_n] \in R^{n * n}$ is orthogonal. $D = diag(d_r, d_{r-1},\cdots, d_1, 0,\cdots, 0)$

$(Au_i)'(Au_j) = u_i'A'Au_j = u_i' (d_ju_j)= d_j 1(i=j)$ , it also shows, for $d_j =0, Au_j = 0$ since $(Au_j)'(Au_j) = d_j *1 = 0$.

define $v_i = \frac{1}{\sqrt{d_i}} Au_i, i=1,\cdots, r; V = [v_r, \cdots, v_1,0,\cdots,0]\in R^{m*n}$

$\implies AU = Vdiag(\sqrt{d_r},\cdots, \sqrt{d_1}, 0,\cdots,0)\implies$ $$ A = Vdiag(\sqrt{d_r},\cdots, \sqrt{d_1}, 0,\cdots,0)U' $$

also,

$$ \begin{align} A &= \underbracket{[V, \mathbf{0}] }_

\underbracket{\left[\begin{matrix}diag(\sqrt{d_r},\cdots, \sqrt{d_1}, 0,\cdots,0)\\mathbf{0} \end{matrix}\right]}{m*n} \underbracket{U'} \ & = \hat V \left[\begin{matrix}diag(\sqrt{d_r},\cdots, \sqrt{d_1}, 0,\cdots,0)\\mathbf{0} \end{matrix}\right] U' \end{align} $$ also,

$$ \begin{align} A'A &= Udiag({d_r},\cdots, {d_1}, 0,\cdots,0)U'\ AA' &= Vdiag({d_r},\cdots, {d_1}, \underbracket{0,\cdots,0}{n-r})V' \ &= \hat V diag({d_r},\cdots, {d_1}, \underbracket{0,\cdots,0})\hat V' \end{align} $$ s.t. $\hat V = [v_r, \cdots, v_1,\hat v_1,\cdots,\hat v_{m-r}]$, is an orthogonal basis in $R^m$ by extending $v_1,\cdots, v_r$

also, it shows $A'A$ and $AA'$ have the same set of non-zero spectral diagonal elements