Let $S\in\mathbb{R}^{n\times n}$ be such that $\|S\|_2\leq 1$, $P\in\mathbb{R}^{n\times m}$, $m<n$, with orthogonal columns ($P^TP=I$) so that $PP^T$ and $I-PP^T$ are orthogonal projectors, and $F\in\mathbb{R}^{m\times m}$ be symmetric such that $\|F\|_2<1$. Let $$ E=S^T(I-PP^T)S, \quad \tilde{E}=S^T[I-P(I-F)P^T]S. $$ Both $E$ and $\tilde{E}$ are clearly symmetric and in addition $E$ is positive semidefinite. It might be true that $F$ could be positive semidefinite too (hence $\tilde{E}$ would be positive semidefinite as well) but it seems that this assumption might not be needed.

Now let $\|E\|_2<1$. I need to show that $$ \color{blue}{\|\tilde{E}\|_2\leq \|E\|_2+\|F\|_2(1-\|E\|_2)=1-(1-\|E\|_2)(1-\|F\|_2)} $$ and hence $\|\tilde{E}\|<1$ (actually, I would like to show that inequality, not just that $\|\tilde{E}\|<1$).

A simple bound is obvious. Since $\tilde{E}=E+S^TPFP^TS$, clearly $$\tag{1}\|\tilde{E}\|_2\leq\|E\|_2+\|F\|_2\|P^TS\|_2^2.$$ We have $\|P^TS\|_2\leq 1$ since $P$ is orthogonal and $\|S\|_2\leq 1$. This is however useless as it gives me only $\|\tilde{E}\|_2\leq\|E\|_2+\|F\|_2$.

If I showed in (1) that $\|P^TS\|_2^2\leq 1-\|E\|_2$, I would be done then. However, it seems that this does not generally hold under the stated assumptions so that the simple triangle inequality (1) is already an overestimate.

I would highly appreciate any comment or pointer to a possible solution.


I succeeded to prove the weaker inequality. Maybe my computations help you to give insight into the stronger inequality.

First, let's ignore $S$ and let $E'=I-PP^T$ and $\tilde{E}'=E'+PFP^T$. Then $E$ and $\tilde{E}$ arise from $E'$ and $\tilde{E}'$ by conjugating with $S$.

Choose ON-bases of $\mathbb{R}^n,\mathbb{R}^m$ such that $P=\binom{I}{0}$. Then $PP^T=\binom{I0}{00}$ and $E'=\binom{00}{0I}$. Furthermore, $\tilde{E}'=\binom{F0}{0I}$. From this it follows that $\|\tilde{E}\|_2\le\|\tilde{E}'\|_2\le 1$. In fact, $\|\tilde{E}\|_2<1$:

Let $v$ be a vector with $\|v\|_2=1$, and let $Sv=\binom{v_1}{v_2}$. Then $\tilde{E}v=S^T\binom{Fv_1}{v_2}$. Either $v_1=0$, and then $\|\tilde{E}v\|_2=\|Ev\|_2<1$, or $\|Fv_1\|_2<\|v_1\|_2$, and so $\|\binom{Fv_1}{v_2}\|_2<1$; hence $\|\tilde{E}v\|_2<1$.

| cite | improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.