5
$\begingroup$

Let $A$ and $B$ be two matrices of eigenvalues $\lambda_i$ and $\mu_i$, respectively.

The spectral variation of $B$ w.r.t. $A$ and the eigenvalue variation of $B$ and $A$ are, respectively, \begin{align} s_B(A)&=\max_i\min_j\vert\lambda_i-\mu_j\vert, \\ v(A,B)&=\min_{\pi}\max_i\vert\lambda_i-\mu_{\pi(i)}\vert;\end{align} where in the latter the minimum is to be taken over all permutations $\pi$ of the indices.

Question 1. If $A$ and $B$ are Hermitian matrices, then for which norms is this true? $$s_B(A)\leq\Vert A-B\Vert.$$

Question 2. If $A$ and $B$ are normal matrices (more generally for fully symmetric operators), then for which norms is this true? $$v(A,B)\leq\Vert A-B\Vert.$$

I would appreciate any reference to the state-of-the-art in this matter.

$\endgroup$
1
  • 3
    $\begingroup$ Why the downvote? $\endgroup$ May 17, 2017 at 2:20

4 Answers 4

2
$\begingroup$

The Hermitian case is more like a state-of-art answer. A good review of results in given in [Holbrook].

$\nu(A,B)\leq\|A-B\|$ for the operator norm. This is a direct consequence from Weyl's inequality.

This problem about spectral variation bound is fully discussed in [Bhatia] Chap 3&4(with a supplement in Chap7&8 if you got the 2006 ed.)

“The spectral variation problem for the class of Hermitian matrices has been completely solved in the following sense. For any two Hermitian matrices a tight upper bound for the distance between their eigenvalues is known. Such bounds are known when the distance is measured in any unitarily-invariant norm.”[Bhatia]p.34

Later development including marjorization inequalities used in controlling covariance matrices in statistics as motivation [Marshall&Olkin], as described in[Bhatia] 3.9

(This is quite clear once you know the reference, probably that is why it gets downvotes.)

Reference

[Holbrook]Holbrook, John A. "Spectral variation of normal matrices." Linear algebra and its applications 174 (1992): 131-144.

[Bhatia]Bhatia, Rajendra. Perturbation bounds for matrix eigenvalues. Society for Industrial and Applied Mathematics, 2007.

[Marshall&Olkin]Marshall, Albert W., Ingram Olkin, and Barry C. Arnold. Inequalities: theory of majorization and its applications. Vol. 143. New York: Academic press, 1979.

$\endgroup$
2
  • $\begingroup$ Is your answer to Question 2 really true for normal matrices? $\endgroup$ May 17, 2017 at 14:08
  • $\begingroup$ @T.Amdeberhan I think it only works for hermitians, corrected. $\endgroup$
    – Henry.L
    May 17, 2017 at 15:21
1
$\begingroup$

The most elementary case of Weyl's inequality says that, if $\lambda_i(S)$ denote the $i$th eigenvaue of the Hermitian matrix $S$ (increasing order), then $\lambda_i(S)+\lambda_1(T)\le\lambda_i(S+T)\le\lambda_i(S)+\lambda_n(T)$. Since the standard operator norm equals, for Hermitian matrices $T$, $\max\{-\lambda_1(T),\lambda_n(T)\}$, you infer immediately $$\nu(A,B)\le\|B-A\|.$$

$\endgroup$
1
$\begingroup$

This is true for (EDIT: only the Euclidean norm) by the Bauer-Fike theorem, since for a symmetric / normal matrix the eigenvalue matrix is unitary and hence has condition number 1.

$\endgroup$
4
  • $\begingroup$ The question is not a yes/no question, but a "for which" question. $\endgroup$ May 17, 2017 at 8:35
  • $\begingroup$ @BenoîtKloeckner Thanks, fixed to specify for which norms it holds. $\endgroup$ May 17, 2017 at 9:02
  • $\begingroup$ The condition number in the $p$-norm of the eigenvalue matrix is $1$ if $p=2$, and only in that case. $\endgroup$ May 17, 2017 at 11:39
  • $\begingroup$ @DenisSerre Oops you are right -- I have edited and changed my claim. $\endgroup$ May 17, 2017 at 12:02
0
$\begingroup$

For comparison of spectral variations of Hermitian and non-Hermitian matrices and operators I suggest the following book

Gil’, Michael I., Operator functions and operator equations, ZBL06793919.

$\endgroup$

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.