How to prove the following equality with matrix norm?

72 Views Asked by At

$\|(H+\phi I)^{-1}(H-\varphi I)\|_2^2=\max_{\|x\|_2=1}\frac{\|(H-\varphi I)x\|_2^2}{\|(H+\phi I)x\|_2^2}$, where both H is a nonsymmetric matrix and $\phi$ and $\varphi$ are two positive parameters. I could not derive it from the definition.

Is it also valid for $\|(H+\phi I)^{-1}(V-\varphi I)\|_2^2=\max_{\|x\|_2=1}\frac{\|(V-\varphi I)x\|_2^2}{\|(H+\phi I)x\|_2^2}$, where $H$ and $V$ are not commute (HV=VH does not hold)?

1

There are 1 best solutions below

3
On BEST ANSWER

You have \begin{align} \sup\{\|(H+\phi I)^{-1}(H-\varphi I)x\|:\ \|x\|=1\} &=\sup\left\{\frac{\|(H-\varphi I)(H+\phi I)^{-1}x\|}{\|x\|}:\ x\ne0\right\}\\ \ \\ &=\sup\left\{\frac{\|(H-\varphi I)x\|}{\|(H+\phi I)x\|}:\ x\ne0\right\}\\ \ \\ &=\sup\left\{\frac{\|(H-\varphi I)x\|}{\|(H+\phi I)x\|}:\ \|x|=1\right\},\\ \ \\ \end{align} where

  • the first equality is due to $H+\phi I$ and $H-\varphi I$ commuting;

  • the second equality is due to $H-\phi I$ invertible, and so $(H-\phi I)x$ runs over all vectors when $x$ runs over all vectors;

  • the third equality can be achieved by dividing numerator and denominator by $x$.

You cannot replace one of the $H$ for an arbitrary $V$. For instance take $\phi=\varphi=0$, $$ H=\begin{bmatrix}1&0\\0&2 \end{bmatrix} ,\ \ \ \ V=\begin{bmatrix} 2&2\\0&0\end{bmatrix}. $$ Then $$ \|H^{-1}V\|=\left\|\begin{bmatrix} 2&2\\0&0\end{bmatrix}\right\| =\left\|\begin{bmatrix} 2&2\\0&0\end{bmatrix}\begin{bmatrix} 2&0\\2&0\end{bmatrix}\right\|^{1/2}=\sqrt8=2\sqrt2. $$ And, with $x_1^2+x_2^2=1$, $t=x_1$, $$ \frac{\|Vx\|}{\|Hx\|}=\frac{2|x_1+x_2|}{\sqrt{x_1^2+2x_2^2}} =\frac{2|x_1+\sqrt{1-x_1^2}|}{\sqrt{1+x_2^2}} =\frac{2(t+\sqrt{1-t^2})}{2-t^2}\leq\sqrt6<\sqrt8=\|H^{-1}V\|. $$