Given a $N\times N$ matrix $A$ over $\mathbb R$. Let $ \rho\left( A \right) = \max \left\{ {\left| \lambda \right|;\lambda \mbox{ eigenvalue of }A} \right\}$. Someone told me that, the following holds:

$$ \rho\left( A \right) = \inf \left\{\lVert| A \rVert|,\lVert \cdot \rVert\mbox{ matrix norm}\right\} . $$

Here is the definition of a matrix norm Wikipedia.

I'm not sure if this is true, and in that case, how to prove it :S


It's true and the proof is classical, in fact in a standard course about numerical linear algebra . We can see $A$ as a matrix over the field of complex numbers. If $|\lambda|=\rho(A)$, $x$ is an eigenvector associated to $\lambda$ and $N$ a matrix norm then $|\lambda|\leq \sup_{x\neq 0}\frac{N'(Ax)}{N'(x)}=N(A)$ where $N'$ is some norm over $\mathbb C^n$, so $\rho(A)\leq N(A)$ for all matrix norm $N$.

Now we fix $\varepsilon>0$. Let $U$ such that $U^{-1}AU=\pmatrix{\lambda_1&t_{12}&\ldots&t_{1n}\\\ 0&\lambda_2&&t_{2,n}\\\ \vdots&&\ddots&\vdots\\\ 0&0&\ldots&\lambda_n}$ where $\lambda_j$ are the eigenvalues of $A$. For $\delta>0$, denote $D_{\delta}=\operatorname{diag}(1,\delta,\ldots,\delta^{n-1})$, and put $V_{\delta}:=UD_{\delta}$. We have $$V_{\delta}^{—1}AV_{\delta}=\pmatrix{\lambda_1&\delta \cdot t_{12}&\ldots&\delta^{n-1}\cdot t_{1n}\\\ 0&\lambda_2&&\delta^{n-2}\cdot t_{2,n}\\\ \vdots&&\ddots&\vdots\\\ 0&0&\ldots&\lambda_n}.$$ We fix $\delta$ such that for all $k\in\{1,\ldots,n-1\}$, $\sum_{j=k+1}^n|t_{k,j}\delta^{j-k}|<\varepsilon$ and we define $N(M)=\sup_{x\neq 0}\frac{\lVert V_{\delta}^{-1}MV_{\delta}x\rVert_{\infty}}{\lVert V_{\delta}^{-1}x\rVert_{\infty}}$.

Thanks to the characterisation of the norm $\lVert\cdot\rVert_{\infty}$ of a matrix as the maximum of the sum of the absolute values of the coefficients of a column we can check that $N(A)\leq \rho(A)+\varepsilon$.


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.