"Best" Submultiplicative / Subordinate norm?

33 Views Asked by At

I have $y = Ax$ where x, y are vectors and $A$ is a matrix. I want to get the best $K$ such that $||y|| \leq K||x||$. Ideally, $K$ is a matrix norm. Especially, $K$ can be a subordinate matrix norm. I tried with Spectral norm and Frobenius norms and they provide a very loose bound for $K$. Are there studies in this direction?

I have also tried the largest sigular value of A and that is a worse bound too.

If that helps in my case A can be expressed as $(I-BC)^{-1}$ where $B$ and $C$ are symmetric metrices.

1

There are 1 best solutions below

0
On

Let $\lVert A \rVert_\epsilon := \lVert D^{-1} T^{-1} A T D \rVert_\infty$ where $D := \operatorname{diag}\{\delta, \delta^2, \dots, \delta^n\}$ for some $1 > \delta > 0$ and $T$. It is easy to check that $\lVert \cdot \rVert_\epsilon$ is a submultiplicative norm.

Let $T$ be such that $T^{-1} A T = \Lambda + U$ where $\Lambda$ is diagonal and $U$ is strictly upper triangular. Therefore, $$ \lVert A \rVert_\epsilon = \lVert \Lambda + D^{-1} U D \rVert_\infty \leq \rho(A) + \lVert D^{-1} U D \rVert_\infty $$

where $\rho(\cdot)$ is the spectral radius. Note that $(D^{-1} U D)_{ij} = \delta^{j-i} U_{ij}, \forall j > i$. Therefore, $$ \lVert D^{-1} U D \rVert_\infty = \max_i \sum_{j=1}^n \lvert \delta^{j-i} U_{ij} \rvert \leq \delta \max_i \sum_{j=1}^n \lvert U_{ij} \rvert = \delta \lVert U \rVert_\infty $$

By choosing $\delta := \epsilon / \lVert U \rVert_\infty$ for some $\epsilon > 0$, we can obtain $$ \lVert A \rVert_\epsilon \leq \rho(A) + \epsilon $$

Now you can make $\epsilon$ as small as you like.