Eigenvalue stability of Hermitian and normal matrices
Adam Parusiński and Armin Rainer · 2026
Read Eigenvalue stability of Hermitian and normal matrices by Adam Parusiński and Armin Rainer (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-23056.txt.
Introduction
Parusiński and Rainer study how stable the eigenvalues of Hermitian matrices are. When the matrix changes a bit, its sorted eigenvalues change in a Lipschitz way. The authors lift this map to Sobolev spaces of matrix-valued functions. For finite exponents, the lifted map is continuous, though not uniformly so. At the top exponent, it stays well defined and bounded but loses continuity. The same picture extends to normal matrices, read in Almgren's multivalued sense. Uses cover singular values and condition numbers.
Catalog record
| CREATOR | Adam Parusiński and Armin Rainer |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 163913 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.23056
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Eigenvalue stability of Hermitian and normal matrices Thanks: This research was funded in whole or in part by the Austrian Science Fund (FWF) DOI 10.55776/PAT1381823. For open access purposes, the authors have applied a CC BY public copyright license to any author-accepted manuscript version arising from this submission. Adam Parusiński and Armin Rainer Abstract . The ordered eigenvalues define a Lipschitz map on the real vector space of Hermitian d × d d\times d matrices. We prove that this map acts continuously, but not uniformly continuously, by superposition on the Sobolev spaces W 1 , q W^{1,q} , for all 1 ≤ q < ∞ 1\leq q<\infty , on bounded open domains. For q = ∞ q=\infty , the action is still well-defined and bounded but not continuous. We show that this stability result extends to normal matrices, where the eigenvalues are naturally interpreted as multivalued Sobolev functions in the sense of Almgren. Several applications are given, including the stability of singular values, condition numbers of matrices, surface area of eigenvalue graphs, and compact self-adjoint operators in Hilbert space. 2020 Mathematics Subject Classification 15A18, 15A42, 15B57, 46E35, 47A55, 47B15, 47H30 Contents 1. Introduction The perturbation theory of Hermitian and normal operators has a long history. The analytic theory started with Rellich’s work [ Rel37 ] , culminated in Kato’s monograph [ Kat76 ] , and extended to smooth perturbation theory in more recent papers, e.g., [ AKLM98 ] , [ KM03 , KMR11 , KMR12 ] , [ Rai09 , Rai13 , Rai22 ] , and [ PR20 ] . For a more comprehensive account of the history, we refer to the recent survey article [ PR25 ] . For the present paper, Weyl’s perturbation theorem [ Wey12 ] plays a fundamental role. It asserts that the ordered eigenvalues form a Lipschitz map on the space of Hermitian matrices with Lipschitz constant 1 , when the space is equipped with the operator norm and the eigenvalues with the \infty -norm. Together with Löwner’s theorem [ Löw34 ] , which provides the corresponding result for the 2 -norm, and its extension to normal matrices by Hoffman and Wielandt [ HW53 ] , this theorem is foundational for nearly all of the results in this paper. Of course, in the case of normal matrices, the eigenvalues must be equipped with a metric that accounts for the fact that they are general complex numbers not admitting a coherent ordering. See also [ BDM83 ] for a variant with the operator norm (where, however, the Lipschitz constant C satisfies 1<C<3 ). For locally Lipschitz curves of normal matrices, each continuous choice of the eigenvalues is actually locally Lipschitz, as shown by [ Rai13 ] . If the (real) parameter space is at least 2 -dimensional, continuous choices of the eigenvalues of normal matrices may not exist. The Lipschitz continuity of the eigenvalues as functions of the matrices immediately implies that Lipschitz families of Hermitian or normal matrices are assigned to the Lipschitz families of their eigenvalues, in a bounded way, that is, with controlled Lipschitz constants. This defines a map which we call the characteristic map : it is given by superposition with the ordered eigenvalue map, in the Hermitian case, and with the unordered eigenvalue map, in the normal case. In this paper, we are interested in the continuity properties of the characteristic map. The characteristic map is not continuous with respect to the Lipschitz topology (on the eigenvalues), but it is continuous with respect to the Sobolev W^{1,q} topology, for every 1\leq q<\infty . Specifically, we prove that for Hermitian matrices the characteristic map is continuous, but not uniformly continuous, from W^{1,q} matrices to ordered W^{1,q} eigenvalues, while for normal matrices it is continuous from W^{1,q} matrices to unordered W^{1,q} eigenvalues — in both cases, for every 1\leq q<\infty . In the latter setting, the unordered eigenvalues are naturally treated as multivalued Sobolev functions in the sense of Almgren [ Alm00 ] ; see also [ DLS11 ] . These results are considerably stronger than the related continuity properties of the solution map for hyperbolic and general polynomials established, respectively, in the recent papers [ PR24a ] and [ PR24b ] . The solution map is defined in analogy to the characteristic map. Recall that Bronshtein’s theorem [ Bro79 ] implies that the ordered roots of a C^{d-1,1} family of hyperbolic (i.e. monic real-rooted) polynomials of degree d are locally Lipschitz; explicit bounds for the Lipschitz constants of the roots in terms of the C^{d-1,1} norms of the coefficients were obtained by [ PR15 ] . By [ PR24a ] , the solution map for hyperbolic polynomials of degree d is continuous from C^{d} coefficients to C^{0,1} roots with respect to the W^{1,q}_{\operatorname{loc}} topology on the roots, for each 1\leq q<\infty — but not with respect to the C^{0,1} topology. For general complex polynomials, the optimal Sobolev regularity W^{1,q} , for 1\leq q<\frac{d}{d-1} , of the roots was established in [ PR16 , PR18 , PR20a ] , and the existence and boundedness of the solution map rests on these results. As shown in [ PR24b ] , the solution map is then continuous from C^{d} coefficients to unordered W^{1,q} roots, for each 1\leq q<\frac{d}{d-1} . Two further points underline the superiority of the matrix setting. First, unlike the polynomial case, the size of the matrices plays no role. Second, the Sobolev spaces W^{1,q} of Hermitian matrices and their eigenvalues are defined globally, whereas for hyperbolic polynomials the spaces of coefficients and roots are necessarily local. These results also have concrete applications. We discuss several, including singular values, the condition number for matrices, the surface area of eigenvalue graphs, and compact self-adjoint operators in Hilbert space (made possible by independence of dimension). It should be mentioned that the orthonormal eigenvectors do not share the regularity of the eigenvalues and, in general, fail to be even continuous in this setting. In addition to Weyl’s perturbation theorem and its relatives, a key ingredient in the proofs is a splitting lemma for complex triangular matrices due to [ PR20 ] , where it appears in a slightly different context. It provides a local uniform unitary block-diagonalization of Hermitian and normal matrices whenever not all eigenvalues coincide. This suggests an inductive argument on the size of the matrices and has the advantage of working directly with the matrices themselves rather than through their characteristic polynomials. In the following, we present our results in more detail. 1.1. Hermitian matrices Let M_{d}(\mathbb{C}) be the complex vector space of complex d\times d matrices. We endow M_{d}(\mathbb{C}) with the Frobenius norm \|A\|_{2}=(\sum_{i,j=1}^{d}|a_{ij}|^{2})^{1/2} . Let \operatorname{Herm}(d)=\{A\in M_{d}(\mathbb{C}):A^{*}=A\} denote the real vector space of complex Hermitian d\times d matrices. We associate with A\in\operatorname{Herm}(d) its increasingly ordered eigenvalues \lambda_{1}^{\uparrow}(A)\leq\lambda_{2}^{\uparrow}(A)\leq\cdots\leq\lambda_{d}^{\uparrow}(A) and thus obtain a continuous map (1.1) \lambda^{\uparrow}=(\lambda_{1}^{\uparrow},\ldots,\lambda_{d}^{\uparrow}):\operatorname{Herm}(d)\to\mathbb{R}^{d}. In fact, by a result of Löwner [ Löw34 ] (see Proposition 3.2 ), this map is even Lipschitz continuous: for A,B\in\operatorname{Herm}(d) , (1.2) \|\lambda^{\uparrow}(A)-\lambda^{\uparrow}(B)\|_{2}\leq\|A-B\|_{2}. As a consequence (see Proposition 3.5 ), the map ( 1.1 ) induces a bounded map (1.3) \mathcal{E}:=(\lambda^{\uparrow})_{*}:W^{1,q}(U,\operatorname{Herm}(d))\to W^{1,q}(U,\mathbb{R}^{d}),\quad A\mapsto\lambda^{\uparrow}\circ A, for each 1\leq q<\infty , which takes W^{1,q} families of Hermitian matrices to their increasingly ordered W^{1,q} eigenvalues, where U\subseteq\mathbb{R}^{m}