l2 Perturbation Theory

Lemma

The following identities are true:

\[\begin{split} \begin{gathered} \left\|\boldsymbol{U} \boldsymbol{U}^{\top}-\boldsymbol{U}^{\star} \boldsymbol{U}^{\star \top}\right\|=\|\sin \boldsymbol{\Theta}\|=\left\|\boldsymbol{U}_{\perp}^{\top} \boldsymbol{U}^{\star}\right\|=\left\|\boldsymbol{U}^{\top} \boldsymbol{U}_{\perp}^{\star}\right\| \\ \frac{1}{\sqrt{2}}\left\|\boldsymbol{U} \boldsymbol{U}^{\top}-\boldsymbol{U}^{\star} \boldsymbol{U}^{\star \top}\right\|_{\mathrm{F}}=\|\sin \boldsymbol{\Theta}\|_{\mathrm{F}}=\left\|\boldsymbol{U}_{\perp}^{\top} \boldsymbol{U}^{\star}\right\|_{\mathrm{F}}=\left\|\boldsymbol{U}^{\top} \boldsymbol{U}_{\perp}^{\star}\right\|_{\mathrm{F}} \end{gathered} \end{split}\]
\[\begin{split} \begin{gathered} \left\|\boldsymbol{U} \boldsymbol{U}^{\top}-\boldsymbol{U}^{\star} \boldsymbol{U}^{\star \top}\right\| \leq \min _{\boldsymbol{m} \in \mathcal{O}^{r \times r}}\left\|\boldsymbol{U} \boldsymbol{R}-\boldsymbol{U}^{\star}\right\| \leq \sqrt{2}\left\|\boldsymbol{U} \boldsymbol{U}^{\top}-\boldsymbol{U}^{\star} \boldsymbol{U}^{\star T}\right\| ; \\ \frac{1}{\sqrt{2}}\left\|\boldsymbol{U} \boldsymbol{U}^{\top}-\boldsymbol{U}^{\star} \boldsymbol{U}^{\star \top}\right\|_{\mathrm{F}} \leq \min _{\boldsymbol{R} \in \mathcal{O}^{\star} \times \boldsymbol{r}}\left\|\boldsymbol{U} \boldsymbol{R}-\boldsymbol{U}^{\star}\right\|_{\mathrm{F}} \leq\left\|\boldsymbol{U} \boldsymbol{U}^{\top}-\boldsymbol{U}^{\star} \boldsymbol{U}^{\star \top}\right\|_{\mathrm{F}} . \end{gathered} \end{split}\]

Theorems

We use the Distance modulo optimal rotation

\[\begin{split} \begin{aligned} \operatorname{dist}\left(\boldsymbol{U}, \boldsymbol{U}^{\star}\right) &:=\min _{\boldsymbol{R} \in \mathcal{O}^{r} \times r}\left\|\boldsymbol{U} \boldsymbol{R}-\boldsymbol{U}^{\star}\right\| \\ \operatorname{dist}_{\mathrm{F}}\left(\boldsymbol{U}, \boldsymbol{U}^{\star}\right) &:=\min _{\boldsymbol{R} \in \mathcal{O}^{r \times r}}\left\|\boldsymbol{U} \boldsymbol{R}-\boldsymbol{U}^{\star}\right\|_{\mathrm{F}} \end{aligned} \end{split}\]

Davis-Kahan’s \(sin \boldsymbol{\Theta}\) theorem(simple version).

Suppose \(\boldsymbol{M}^{\star} \succeq \mathbf{0}\) and is rank-r. If \(\|\boldsymbol{E}\|<(1-1 / \sqrt{2}) \lambda_{r}\left(\boldsymbol{M}^{\star}\right)\), then

\[\begin{split} \begin{gathered} \operatorname{dist}\left(\boldsymbol{U}, \boldsymbol{U}^{\star}\right) \leq \sqrt{2}\|\sin \boldsymbol{\Theta}\| \leq \frac{2\left\|\boldsymbol{E} \boldsymbol{U}^{\star}\right\|}{\lambda_{r}\left(\boldsymbol{M}^{\star}\right)} \leq \frac{2\|\boldsymbol{E}\|}{\lambda_{r}\left(\boldsymbol{M}^{\star}\right)} \\ \operatorname{dist}_{F}\left(\boldsymbol{U}, \boldsymbol{U}^{\star}\right) \leq \sqrt{2}\|\sin \boldsymbol{\Theta}\|_{\mathrm{F}} \leq \frac{2\left\|\boldsymbol{E} \boldsymbol{U}^{\star}\right\|_{\mathrm{F}}}{\lambda_{r}\left(\boldsymbol{M}^{\star}\right)} \leq \frac{2 \sqrt{r}\|\boldsymbol{E}\|}{\lambda_{r}\left(\boldsymbol{M}^{\star}\right)} \end{gathered} \end{split}\]

Davis-Kahan’s \(sin \boldsymbol{\Theta}\) theorem(General version)

Assume that

\[\begin{split} \begin{aligned} &\text { eigenvalues }\left(\boldsymbol{\Lambda}^{\star}\right) \subseteq(-\infty, \alpha-\Delta] \cup[\beta+\Delta, \infty) \\ &\operatorname{eigenvalues}\left(\boldsymbol{\Lambda}_{\perp}\right) \subseteq[\alpha, \beta] \end{aligned} \end{split}\]
for some quantities \(\alpha, \beta \in \mathbb{R}\) and eigengap \(\Delta>0\). Then one has
\[\begin{split} \begin{gathered} \operatorname{dist}\left(\boldsymbol{U}, \boldsymbol{U}^{\star}\right) \leq \sqrt{2}\|\sin \boldsymbol{\Theta}\| \leq \frac{\sqrt{2}\left\|\boldsymbol{E} \boldsymbol{U}^{\star}\right\|}{\Delta} \leq \frac{\sqrt{2}\|\boldsymbol{E}\|}{\Delta} \\ \operatorname{dist}_{\mathrm{F}}\left(\boldsymbol{U}, \boldsymbol{U}^{\star}\right) \leq \sqrt{2}\|\sin \boldsymbol{\Theta}\|_{\mathrm{F}} \leq \frac{\sqrt{2}\left\|\boldsymbol{E} \boldsymbol{U}^{\star}\right\|_{\mathrm{F}}}{\Delta} \leq \frac{\sqrt{2 r}\|\boldsymbol{E}\|}{\Delta} \end{gathered} \end{split}\]

Wedin’s sinΘ theorem

If \(\|\boldsymbol{E}\|<\sigma_{r}^{\star}-\sigma_{r+1}^{\star}\), then one has

\[\begin{split} \begin{gathered} \max \left\{\operatorname{dist}\left(\boldsymbol{U}, \boldsymbol{U}^{\star}\right), \operatorname{dist}\left(\boldsymbol{V}, \boldsymbol{V}^{\star}\right)\right\} \leq \frac{\sqrt{2} \max \left\{\left\|\boldsymbol{E}^{\top} \boldsymbol{U}^{\star}\right\|,\left\|\boldsymbol{E} \boldsymbol{V}^{\star}\right\|\right\}}{\sigma_{r}^{\star}-\sigma_{r+1}^{\star}-\|\boldsymbol{E}\|} \\ \max \left\{\operatorname{dist}_{\mathrm{F}}\left(\boldsymbol{U}, \boldsymbol{U}^{\star}\right), \operatorname{dist}_{\mathrm{F}}\left(\boldsymbol{V}, \boldsymbol{V}^{\star}\right)\right\} \leq \frac{\sqrt{2} \max \left\{\left\|\boldsymbol{E}^{\top} \boldsymbol{U}^{\star}\right\|_{\mathrm{F}},\left\|\boldsymbol{E} \boldsymbol{V}^{\star}\right\|_{\mathrm{F}}\right\}}{\sigma_{r}^{\star}-\sigma_{r+1}^{\star}-\|\boldsymbol{E}\|} \end{gathered} \end{split}\]