0
$\begingroup$

Problem

Show $\xi = \frac{(y - Mx)x^\ast}{x^\ast x}$ is the solution of the following

$$ (M + \xi)x = y $$

with the minimum $\Vert \xi\Vert_2$, where $M \in \mathbb{C}^{m \times n}, x \neq 0 \in \mathbb{C}^n, y \in \mathbb{C}$.


Try

We have the form as follows:

$$ M + \xi = \frac{M x^\ast x}{x^\ast x} + \frac{(y - Mx)x^\ast}{x^\ast x} $$

I have noticed the form's similarity with the operator 2-norm:

$$ \sup_{x\neq0} \frac{\Vert Ax \Vert_2}{\Vert x \Vert_2} = \sup_{x\neq0} \frac{x^\ast A^\ast A x}{x^\ast x} $$

But here I cannot proceed. Any help will be appreciated.

$\endgroup$
2
$\begingroup$

From $\xi x = y - Mx$ we have $$ \|\xi\|_2\geq\frac{\|\xi x\|_2}{\|x\|_2}=\frac{\|y-Mx\|_2}{\|x\|_2}. $$ It means that whatever $\xi$ such that $(M+\xi)x=y$ we have, its 2-norm cannot be smaller than $\|y-Mx\|_2/\|x\|_2$.

Now just verify we get equality above with $\xi_\star=(y-Mx)x^*/\|x\|_2^2$.

Note that this solution is also unique. Any other solution $\xi$ can be expressed as $$ \xi=\xi_\star+\eta X_\perp^*, $$ where the columns of $X_\perp\in\mathbb{C}^{n\times (n-1)}$ are linearly independent and orthogonal to $x$ and $\eta\in\mathbb{C}^{m\times (n-1)}$ is a "parameter" matrix. If we also assume that the columns of $X_\perp$ are orthonormal then $$ \xi=\left[\frac{y-Mx}{\|x\|_2},\eta\right]\left[\frac{x}{\|x\|_2},X_\perp\right]^* $$ Since the matrix on the right side is unitary, we have $$ \|\xi\|_2=\left\|\left[\frac{y-Mx}{\|x\|_2},\eta\right]\right\|_2>\frac{\|y-Mx\|_2}{\|x\|_2}=\|\xi_\star\| $$ if $\eta\neq 0$.

$\endgroup$

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.