Let $A$ be an $n\times n$ real matrix with entries in a fixed interval $[a_\min,a_\max]$, with $a_\min$, $a_\max>0$.

Question: Are there any upper bounds on the condition number of the eigenvector matrix associated to $A$ as $n$ increases?

The motivation of this question comes from the $n\times n$ random dense case. Indeed, with reference to this case, in [1, Page 338] it is stated that: "[...] experiments indicate that the condition number of the associated eigenvector matrix also grows linearly [as $n$ increases]".

Hence, I wonder whether any similar result can be/has been proved for the deterministic (positive) dense case.

[1] L. N. Trefethen and M. Embree. Spectra and pseudospectra: the behavior of nonnormal matrices and operators. Princeton University Press, 2005.


No, there are no such bounds.

Let $a>0$ be a fixed real and $\varepsilon\ll a$. Take the non-diagonalizable matrix $A=\begin{bmatrix}3a & 0 & 0\\ 0 & 0 & \varepsilon \\ 0 & 0 & 0 \end{bmatrix}$. Let $V$ be a symmetric, orthogonal matrix that maps $\begin{bmatrix}\sqrt {3}\\0\\0\end{bmatrix}$ to $\begin{bmatrix}1\\1\\1\end{bmatrix}$ (for instance, a Householder reflector).

Now $M=VAV^{-1}$ has entries in $[a-O(\varepsilon),a+O(\varepsilon)]$ and is not diagonalizable. So the eigenvector condition number of any sequence of matrices $M_k$ that converges to $M$ blows up.

  • $\begingroup$ I think you need a orthogonal V, at least the first row of V^-1 should be (1,1,1). Furthermore as there is no orthogonal matrix mapping vectors of different lengths you probably mean a scaled version and then multiply the result. $\endgroup$ – user35593 Jan 28 '17 at 1:27
  • $\begingroup$ @user35593 You are correct; one needs to be more careful than I was with the choice of $V$. Fortunately, Householder reflectors (which are the standard tool used to construct orthogonal decompositions) work. I have fixed my answer now. $\endgroup$ – Federico Poloni Jan 28 '17 at 7:06
  • $\begingroup$ @user35593 Thanks a lot for the fixes! $\endgroup$ – Federico Poloni Jan 28 '17 at 21:15

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.