2
$\begingroup$

Consider a matrix function $A(x)$, analytically depending on single parameter $x$.

Consider the eigenvalue/eigenvector pair of $A(0)$, namely $\lambda(0)$ and $w(0)$.

The question is whether we can construct the analytic functions $\lambda(x)$ and $w(x)$ that are the eigenvalue/eigenvector of $A(x)$. (I assume that both $\lambda(x)$ and $w(x)$ could have poles or branch points, hence being multi-valued).

In books I found this statement for the case of Hermitian matrix $A(t)$ only. Is it true for non-Hermitian comlex matrices as well?

$\endgroup$

3 Answers 3

4
$\begingroup$

This is completely answered in Kato's Perturbation theory for linear operators, chapter 1 (he treats the non-hermitian case). Eigenvalues are generally analytic, eigenvectors not so much (in the neighborhood of values of $A(x)$ with multiple eigenvalues.

$\endgroup$
1
  • $\begingroup$ Can you be a little bit more precise? As far as I know eigenvalues are in general continuous (but not necessarily analytic), in fact they can have root singularities. Eigenvectors instead can also have poles $\endgroup$
    – lcv
    Jun 3, 2015 at 19:22
2
$\begingroup$

This isn't a complete answer but may help a little. One can begin by considering the function of two complex variables $$ f(z,w)=\det (A(z)-wI) $$ and its zero variety $$ V=\{(z,w):f(z,w)=0\}. $$ So, a point $(z,w)$ belongs to $V$ if and only if $w$ is an eigenvalue of $A(z)$. Now away from singular points of $V$, one should be able to apply the implicit function theorem to get an analytic family of eigenvalues; one can then get the corresponding eigenvectors via Cramer's rule. On the other hand, from this point of view one can see that in some cases branching is inevitable: for example consider $$ A(z)=\begin{pmatrix} 0 & z^2 \\ z & 0\end{pmatrix}. $$ Then $f(z,w)=w^2-z^3$ and $(0,0)$ is the only singular point of $V$. Away from $0$ we can then take $w(z)$ to be any branch of $z^{3/2}$, but this also makes clear that one can't have analyticity near $0$, since $|w(z)|=|z|^{3/2}$ near $0$ and no analytic function can do this.

$\endgroup$
0
$\begingroup$

The functions $\lambda(x)$ and $w(x)$ obey a set of coupled first-order differential equations in $x$, derived in Structure of trajectories of complex matrix eigenvalues (Bohigas, De Carvalho, and Pato, 2012). For the case considered in that paper (a simple linear dependence of $A$ on $x$), no singularities in the $x$-dependence of the eigenvalues appear, and I would think that this holds more generally.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.