2
$\begingroup$

We have the following theorems!

Corollary VI.1.6 [Bhatia: Matrix Analysis]: Let $a_j(t)$, where $1\leq j \leq n$ be continuous complex valued functions defined on an interval $I$. Then there exists continuous functions $\alpha_1(t), \ldots, \alpha_n(t)$ such that for each $t \in I$, these constitute the roots of the monic polynomial $z^n-a_1(t)z^{n-1}+\cdots+(-1)^na_n(t)$.

Theorem [Kato: Perturbation theory for linear operators]: Suppose $D \subset \mathbb{C}$, and that $A: D \rightarrow M_n$ is a continuous function. If (1)- D is a real interval, or (2)- A(t) has only real eigenvalues then there exist n eigenvalues (counted with algebraic multiplicities) of A(t) that can be parameterized as continuous functions $\lambda_1(t), \ldots, \lambda_n(t)$ from D to $\mathbb{C}$. In the second case, one can set $\lambda_1(t) \geq \cdots \geq \lambda_n(t)$.

These two theorems imply that the eigenvalues of a matrix (whose entries are continuous functions of a real parameter) are continuous functions of the parameter. Also, the second theorem says that this continuity of eigenvalues is still there if the parameter is a complex number, and the matrix is Hermitian.

Let $A = \left[x_i+x_j\right]$, where $0<x_1<\ldots<x_n$ are positive real numbers. Then the entries of $A$ are the continuous functions of $x_i$'s. Since A is Hermitian, all its eigenvalues are real. The characteristic polynomial of A is a poly with coefficients as functions of $x_i$'s.

Que: Is there any extensions of the above theorems, which can imply that the eigenvalues of $A$ can be parametrized as continuous functions of $x_i$'s?

What I mean is that: Is there any analogous result to Corollary VI.1.6, where $a_j$ and $\alpha_j$ are functions of $x_i$'s?

Is there any analogous result to Kato theorem, where $D \subset \mathbb{R}^n$?

I couldn't find any such result. Any help is really appreciated.

$\endgroup$
6
  • $\begingroup$ The "set" of eigenvalues depends continuously on the martrix coeffients, see Kato Theorem 5.14. The problem arises when one wants to choose continuously every single eigenvalue. However, this is possible when they are real, by ordering them. $\endgroup$ Commented Sep 3, 2023 at 22:10
  • $\begingroup$ I know that. Is it true if the eigenvalues are real but the matrix entries are continuous functions of n variables $0<x_1<\cdots<x_n$, instead of a single real parameter? $\endgroup$ Commented Sep 4, 2023 at 8:47
  • $\begingroup$ Yes. Once the "set" is continuous in the sense of Theorem 5.14, you can list them in increasing order and each eigenvalue will be continuous. $\endgroup$ Commented Sep 4, 2023 at 8:58
  • $\begingroup$ Thanks for the reply, but can you please explain, why Theorem 5.14 implies the continuity of eigenvalues (individually) in case of more than one parameter? $\endgroup$ Commented Sep 4, 2023 at 10:02
  • $\begingroup$ The argument I have in mind is the following. assume Theorem 5.14, so that the distance $d$ between the sets $\Sigma (T), \Sigma (T+A)$ tends to zero as $A\to 0$ (I use the notation in the book, so $T$ is a fixed matrix and $A$ is a perturbation). For the definition of distance, see (5.3) on page 108, beginning of II.2, "the numbering of the eigenvlaues". If all eigenvalues are real, list them in increasing order (repeated according to algebraic multiplicities ): $\lambda_1(T+A) \cdots \leq \lambda_n(T+A)$. It turns out that $d=\max_i \{|\lambda_i(T+A)-\lambda_i(T)|$, giving the continuity. $\endgroup$ Commented Sep 4, 2023 at 11:03

0

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.