Linear Algebra and its Applications

Transcription

1 Linear Algebra and its Applications 437 (2012) Contents lists available at SciVerse ScienceDirect Linear Algebra and its Applications journal homepage: Perturbation of multiple eigenvalues of Hermitian matrices Ren-Cang Li a,1, Yuji Nakatsukasa b,2, Ninoslav Truhar c,,3, Wei-guo Wang d,4 a Department of Mathematics, University of Texas at Arlington, P.O. Box 19408, Arlington, TX , USA b School of Mathematics, The University of Manchester, Manchester M13 9PL, UK c Department of Mathematics, J.J. Strossmayer University of Osijek, Trg Lj. Gaja 6, Osijek, Croatia d School of Mathematical Sciences, Ocean University of China, 238 Songling Road, Qingdao , PR China ARTICLE INFO Article history: Received 25 September 2011 Accepted 26 January 2012 Available online 29 February 2012 Submitted by V. Mehrmann AMS classification: 15A22 15A42 65F15 Keywords: Graded perturbation Multiple eigenvalue Generalized eigenvalue problem ABSTRACT This paper is concerned with the perturbation of a multiple eigenvalue μ of the Hermitian matrix A = diag(μi, A 22 ) when it undergoes an off-diagonal perturbation E whose columns have widely varying magnitudes. When some of E s columns are much smaller than the others, some copies of μ are much less sensitive than any existing bound suggests. We explain this phenomenon by establishing individual perturbation bounds for different copies of μ. They show that when A 22 μi is definite the ith bound scales quadratically with the norm of the ith column, and in the indefinite case the bound is necessarily proportional to the product of E s ith column norm and E s norm. An extension to the generalized Hermitian eigenvalue problem is also presented Elsevier Inc. All rights reserved. 1. Introduction Consider the eigenvalue problem for Hermitian matrix Ã: ( m n m A Ã = 11 E n ), A 11 = μi m, (1.1) Corresponding author. addresses: rcli@uta.edu (R.-C. Li), yuji.nakatsukasa@manchester.ac.uk (Y. Nakatsukasa), ntruhar@mathos.hr (N. Truhar), wgwang@ouc.edu.cn (W.-g. Wang). 1 Supported in part by National Science Foundation Grants DMS and DMS Supported in part by Engineering and Physical Sciences Research Council Grant EP/I005293/1. 3 Supported in part by the Passive control of mechanical models", Grant No of the Croatian MZOS. 4 Supported in part by the National Natural Science Foundation of China under grant and , China Scholarship Council, Shandong Province Natural Science Foundation (Y2008A07). This work was initiated while this author was visiting University of Texas at Arlington from September 2010 to August /$ - see front matter 2012 Elsevier Inc. All rights reserved.

2 R.-C. Li et al. / Linear Algebra and its Applications 437 (2012) where the superscript takes the complex conjugate transpose of a matrix or a vector, and I m (or simply I later if its dimension is clear from the context) is the m m identity matrix. If E is a zero block, then μ is a multiple eigenvalue with multiplicity m. In general, if E is small then à has m eigenvalues close to μ. In fact more can be said qualitatively. Let η be the eigenvalue gap between A 11 = μi and A 22 defined as η = min μ ν, ν eig(a 22 ) (1.2) where eig(a 22 ) is the set of the eigenvalues of A 22, and let ε = E 2, (1.3) where 2 is either the spectral norm of a matrix or the l 2 -norm of a vector. The main result in [1] says à has m eigenvalues θ 1,...,θ m such that 2ε 2 μ θ j for 1 j m. (1.4) η + η 2 + 4ε 2 The right-hand side of (1.4) is of second order in ε if η>0and is never larger than ε. As confirmed by the 2-by-2 example in [1], in general these inequalities cannot be improved without knowing more information on E than just ε = E 2. Suppose now that we do have additional information on E. For example, consider the case where one of the columns of E is zero for which θ i = μ for some i. Can we derive bounds that reflect this a zero column leads to some θ i being μ? A possible and quick answer can be given as follows: first zero out the jth column of E, and then use the well-known perturbation theorem (attributed to Lidskii, Weyl, Wiedlandt and Mirsky in various forms [2, pp ]) to conclude that à has an eigenvalue θ that differs from μ by no more than E (:,j) 2, where E (:,j) denotes the jth column of E. Itobviously implies that if E s jth column is a zero column, then μ must be an eigenvalue of Ã. But there are two drawbacks to this quick answer: 1. E (:,j) 2 can be potentially (much) larger than the right-hand side of (1.4), making the estimate less favorable to (1.4). 2. This does not imply that à has m eigenvalues θ j such that μ θ j E (:,j) 2 because some of the θ by this argument could be the same eigenvalues of Ã, asmentionedin[3, Section 11.5]. The purpose of this article is to develop a theory that will reflect the effect of disparity in the magnitudes of the columns of E on the eigenvalues of Ã,unlike(1.4), through establishing different bounds for the m eigenvalues of à closest to μ. For the sake of convenience, throughout this paper η and ε are always defined by (1.2) and (1.3), respectively, and set ɛ j = E (:,ij ) 2 for 1 j m, (1.5) where {i 1, i 2,...,i m } is the permutation of {1, 2,...,m} such that ɛ 1 ɛ 2 ɛ m. (1.6) It is well-known that ɛ m ascending order: ε m ɛ m. The eigenvalues of E E are τ 1,τ 2,...,τ m, arranged in 0 τ 1 τ 2 τ m. (1.7) We will also use X Y (X Y) for two Hermitian matrices of the same size to mean Y X is positive definite (semi-definite), and X Y (X Y) tomeany X (Y X). In particular, X 0(X 0) means that X is positive definite (semi-definite).

3 204 R.-C. Li et al. / Linear Algebra and its Applications 437 (2012) Our perturbation bounds are actually presented in terms of τ j, the eigenvalues of E E. They can be easily turned into bounds in terms of ɛ j, because of Lemma 3.1 below, in order to serve our purpose of developing a theory that reflects the effect of disparity in the magnitudes of the columns of E. The rest of this paper is organized as follows. We first investigate specific examples in Section 2, which provide insights into possible bounds that could be expected. In Section 3 we give our main results, in which we separately deal with the cases where A 22 μi is definite or indefinite. For the indefinite case, we give asymptotic estimates that are correct up to fourth-order terms, as well as strict bounds that are slightly larger than the asymptotic estimates. In Section 4 we describe how our bounds can be extended to the generalized eigenvalue problem. Finally we summarize our conclusions in Section Motivational examples The examples below will shape our expectation on possible effects of different magnitudes of the columns of E on the eigenvalues of à nearest 0. Example 2.1. Consider the 4-by-4 matrix à given by E = , A 22 = 1 0, à = 0 E In this case A 11 approximately = 0, i.e., μ = 0in(1.1), and η = 1. The two eigenvalues of à closest to 0 are and , (2.1) which are about ɛ1 2 = E (:,1) 2 = and ɛ2 2 = E (:,2) 2 = , respectively. The inequality (1.4)saysà has two eigenvalues that differ from 0 by no more than This estimate is very good for the second eigenvalue in (2.1) but not so for the first one which is about less than the square of the estimate. The quick answer, on the other hand, says à has an eigenvalue that differs from 0 by no more than ɛ 1 = and an eigenvalue from 0 by no more than ɛ 2 = , providing even worse estimates than by (1.4). Example 2.1 may lead us to believe that there are m properly ordered eigenvalues θ 1,...,θ m of à with each difference μ θ j being of second order in ɛ j = E (:,j) 2 if η>0. Later we will show this is indeed true if A 22 μi is definite, but not so in the general case as we can see by the next example. Example 2.2. Consider the 4-by-4 matrix à given by E = δ 1 0, A 22 = 01, à = 0 E, 0 δ 2 10 where both δ i are real numbers and δ i 1. The characteristic equation of à is λ 4 (δ δ )λ2 + δ 2 1 δ2 2 = 0, whose two smallest eigenvalues in magnitude satisfy 2 λ = δ1 2 + δ [1 + (δ 1 + δ 2 ) 2 ][1 + (δ 1 δ 2 ) 2 ] δ 1δ 2. Thus λ / δ 1 δ 2 =1 + O(δ1 2 + δ2 2 ). It follows that the smaller λ can be made arbitrarily larger than O(min{δ1 2,δ2 2 }).

4 R.-C. Li et al. / Linear Algebra and its Applications 437 (2012) Main results Throughout this section, à is Hermitian and given by (1.1). Without loss of generality, we assume μ = 0. Since by assumption μ is not an eigenvalue of A 22, A 22 is non-singular as a result of assuming μ = 0, and the gap η as defined by (1.2)nowis η = 1/ A For any λ eig(a 22 ),set X = I E (A 22 λi) 1. 0 I Then X(à λi)x = ( λ)i E (A 22 λi) 1 E, (3.1) A 22 λi and thus det(ã λi) = det ( E (A 22 λi) 1 E λi) det(a22 λi). (3.2) From this we see that any eigenvalue λ of à not in eig(a 22) is a root of det( E (A 22 λi) 1 E + ( λ)i). (3.3) Recall from (1.1) that for E sufficiently small in magnitude, the eigenvalues of à consist of two subsets: one spawned from m copies of μ and another from the eigenvalues in eig(a 22 ) upon being moved by E. Henceà has m eigenvalues close to 0 and these m eigenvalues are zeros of (3.3) near0.notethat for λ A = λ /η < 1 we can write (A 22 λi) 1 = j=0 λj A j 1 22,soforsuchλ we have E (A 22 λi) 1 E + ( λ)i = λ j E A j 1 22 E + ( λ)i. (3.4) j=0 Theorem 3.1. Let à be a Hermitian matrix of form (1.1) with μ = Assume ε< 3/4 η.then (a) à has exactly m eigenvalues θ j in the open interval ( η/2,η/2), andmoreover 2ε 2 θ j, (3.5) η + η 2 + 4ε 2 for 1 j m; (b) The function (3.3) has exactly m zeros in ( η/2, η/2) and these zeros are precisely the eigenvalues θ j of Ã. 2. Ãhasmeigenvaluesθ j = ϑ j + O(ε 4 /η 2 ),whereϑ j for 1 j m are the eigenvalues of E A 1 22 E. In particular, if η = O(1), thenθ j = ϑ j + O(ε 4 ). Proof. Since 4t 2 /( t 2 )<1ift 2 < 3/4, we have 2ε 2 < η ε η + η 2 + 4ε 2 2 if η < 3. 4

5 206 R.-C. Li et al. / Linear Algebra and its Applications 437 (2012) By the main result of [1], we conclude that à has exactly m eigenvalues θ j in the open interval ( η/2,η/2) and (3.5) holds. Item 1(b) is a consequence of Item 1(a), (3.2) and det(a 22 λi) = 0forλ ( η/2,η/2). The expression in (3.4) isequalto E A 1 22 E + ( λ)i, uptoo(ε4 /η 2 ),for λ =O(ε 2 /η). Since by (1.4) à has exactly m eigenvalues no larger than O(ε2 /η) in magnitude, we conclude that θ j = ϑ j + O(ε 4 /η 2 ) for 1 j m. Example 2.1 (revisit). The eigenvalues of E A 1 22 E are , which are extremely close to the exact values given in (2.1). Theorem 3.1 gives asymptotic estimates for θ j in terms of ϑ j. In the subsections that follow, we will establish bounds that reflect the effect of disparity in the magnitudes of the columns of E. To this end, we normalize the columns of E by their l 2 -norms to get E = E 0 D, (3.6) where D = diag( E (:,1) 2, E (:,2) 2,..., E (:,m) 2 ), { E (E 0 ) (:,j) = (:,j) / E (:,j) 2, if E (:,j) = 0, 0, if E (:,j) = 0. (3.7a) (3.7b) Lemma 3.1. Let τ 1,τ 2,...,τ m be the eigenvalues of E E, arranged in ascending order as in (1.7),andlet ɛ j be defined as in (1.5) and (1.6).Then τ j E ɛ2 j m ɛ 2 j. (3.8) Proof. Use 0 E E = DE 0 E 0D E D2 to get τ j E D2 (i j,i j ) = E ɛ2 j. The second inequality is due to E 0 2 m. Next, we separately consider the cases according to whether A 22 is definite or not. All bounds will begivenintermsofτ j. Corresponding bounds in terms of ɛ j can then be easily derived by using (3.8) Positive (negative) definite A 22 Theorem 3.2. For Hermitian matrix Ãasin(1.1) with μ = 0, suppose ε< 3/4 η. IfA 22 is positive (negative) definite, then à has m nonpositive (nonnegative) eigenvalues θ 1,...,θ m arranged in ascending order satisfying 2τ j 0 θ m j+1 η +, if A 22 0, (3.9a) η 2 + 4τ j 2τ j 0 θ j η +, if A 22 0, (3.9b) η 2 + 4τ j for 1 j m. Proof. The case in which A 22 0 can be turned into the case in which A 22 0 by considering à instead. Suppose that A 22 0, i.e., A 22 is positive definite. Set

6 R.-C. Li et al. / Linear Algebra and its Applications 437 (2012) B(t) = E (A 22 ti) 1 E (3.10) for t R. ByTheorem3.1 and the assumption ε< 3/4 η, we know à has exactly m eigenvalues in ( η/2,η/2) and these m eigenvalues are the zeros of det (B(t) ti) in ( η/2,η/2). Sinceforany t ( η/2,η/2), 0 A 22 ti and thus B(t) 0; so B(t) ti 0 for t (0,η/2). Therefore the m eigenvalues of à are in ( η/2, 0]. Denotethemby η/2 <θ 1 θ 2 θ m 0. Also denote by λ 1 (t) λ 2 (t) λ m (t) 0 (3.11) the m eigenvalues of B(t) for t ( η/2, 0]. They are continuous. The fixed points of λ i (t) within t ( η/2, 0] give all the θ j. In fact, we have λ j (θ j ) = θ j.thisisbecauseλ j (t) is a decreasing function for t ( η/2, 0] and thus λ j (t) = t has a unique solution on ( η/2, 0]. Henceθ j is the jth smallest eigenvalue of B(θ j ). This implies that θ j = θ j is the jth largest eigenvalue of B(θ j ).Since we have B(θ j ) = E (A 22 θ j I) 1 E E E η + θ j, θ j τ m j+1 η + θ j 2τ m j+1 implying θ j η + η 2 + 4τ m j+1 which gives (3.9a). Remark 3.1. Since the right-hand sides in (3.9) are increasing as τ j does, replacing τ j by its upper bound in (3.8) yieldsboundson θ j in terms of ɛ j, the norms of E s columns. Example 3.1. Consider the 4-by-4 matrix à given by E = , A 22 = 10, à = 0 E. (3.12) Fig. 1. The log log scale plot of λ i (t) for à in (3.12).

7 208 R.-C. Li et al. / Linear Algebra and its Applications 437 (2012) In this case A 11 = 0, i.e., μ = 0in(1.1), and η = 1. The following table displays the eigenvalues θ j of à nearest to 0, the eigenvalues ϑ j of E A 1 22 E, and the upper bounds in (3.9) and the ones after τ j replaced by mɛ 2 j. θ j ϑ 2τ j 2mɛ 2 j j η+ η 2 +4τ j η+ η 2 +4mɛ 2 j Thus our bounds are remarkably sharp. Let λ i (t) be as in the proof of Theorem 3.2 for this example. Fig. 1 plots λ 1 (t) and λ 2 (t) as functions of t. The intersections with the curve for t are the eigenvalues θ 1 and θ 2.NotethatinFig.1 λ 1 (t) and λ 2 (t) appear to be nearly constants. That is because they decrease very slowly, which is a typical behavior of λ i (t) when ε η/2. In fact it can be shown that ε2 η 2 dλ i(t) dt 0fort ( η/2, 0] and 1 i m Indefinite A 22 Consider now that A 22 μi is indefinite. We will use the following result, which is a direct consequence of the proof of [4,Theorem1]. Lemma 3.2. Let W be an l-by-l Hermitian matrix, and let D = diag(δ 1,δ 2,...,δ l ) with δ 1 δ 2 δ l. Denote the eigenvalues of D WD by ω 1,...,ω l arranged such that ω 1 ω 2 ω l.thenfor1 i l ω i min δ l j+1 δ i+j 1 W 2 1 j l i+1 δ l δ i W 2. Two types of bounds will be proven for the eigenvalues θ j of interest of Ã: asymptotical bounds up to O(ε 4 ), and strict bounds at a tradeoff of being slightly larger than the asymptotic bounds if higher order terms O(ε 4 ) are ignored. Lemma 3.3. Let ϑ j for 1 j m be the eigenvalues of E A 1 22 Earrangedsuchthat ϑ 1 ϑ 2 ϑ m. Then ϑ j ζ j def = 1 η min τm+1 k τ j+k 1 1 k m j+1 1 η τm τ j, (3.13a) (3.13b) where τ i (1 i m) are the eigenvalues of E EasinLemma 3.1. Proof. Inequality (3.13b) follows from (3.13a) by simply picking k = 1 without the minimization. We now prove (3.13a). Let E = UΣV be the SVD of E, where U and V are unitary and ( diag( τ1, τ 2,..., ) τ m ), if n m, Σ = ( 0 (n m) m diag( τ m n+1, τ m+n+2,..., τ m ) 0 n (m n) ), if n < m.

8 R.-C. Li et al. / Linear Algebra and its Applications 437 (2012) Note that in the case when n < m, τ 1 = = τ m n = 0. We have E A 1 22 E = VΣ U A 1 which has the same eigenvalues as Σ U A 1 22 UΣ. It can be proven that for either n m or n < m, 22 UΣV Σ U A 1 22 UΣ = DWD for some matrix W satisfying W 2 1/η and D = diag( τ 1,..., τ m ). Now apply Lemma 3.2 to complete the proof. Remark 3.2. When A 22 isdefinite,wecanget ϑ j τ j /η which is stronger than (3.13a) and thus (3.13b). Theorem 3.3. For Hermitian matrix Ãasin(1.1) with μ = 0, suppose ε < 3/4 η. ThenÃhasm eigenvalues θ 1,...,θ m arranged such that θ 1 θ 2 θ m (3.14) satisfying θ j ζ j + O(ε 4 ), (3.15) where ζ j is defined by (3.13a). Proof. It is a consequence of Theorem 3.1 and Lemma 3.3. Next we derive strict bounds, i.e., without the term O(ε 4 ) in (3.15). One difficulty here is that λ i (t) is no longer monotonic. However, the fact remains true that if θ i ( η/2,η/2) is an eigenvalue of B(θ i ) = E (A 22 θ i I) 1 E, then θ i is also an eigenvalue of Ã. Lemma 3.4. Let B(t) be defined as in (3.10) with eigenvalues λ 1 (t) λ 2 (t) λ m (t) (3.16) each of which are piecewise differentiable 5.Ifε<η/2, then db(t) 4ε2 dt 2 η < 2 1 and dλ j (t) 4ε2 < 1 for t ( η/2,η/2). (3.17) dt η2 Proof. We have B(t) B(t + Δt) = E (A 22 ti) 1 E E [A 22 (t + Δt)I] 1 E = E { (A 22 ti) 1 [A 22 (t + Δt)I] 1} E = E (A 22 ti) 1 {I [ I Δt(A 22 ti) 1] 1 } E. Therefore B(t) B(t + Δt) E (A 22 ti) {I 1 [ } I Δt(A 22 ti) 1] 1 E = 2 Δt 2 Δt ε (A 22 ti) 1 2 I [ I Δt(A 22 ti) 1] 1 2 ε. Δt 5 By [5, Theorem 4.8], there are countable points in ( η/2,η/2) such that between any two nearby points, each λ i (t) is differentiable.

9 210 R.-C. Li et al. / Linear Algebra and its Applications 437 (2012) Noting that for t ( η/2,η/2), wehave (A 22 ti) 1 2 < 2 η, I [ I Δt(A 22 ti) 1] < 1 Δt 2/η 1 = Δt 2/η 1 Δt 2/η, and thus B(t) B(t + Δt) ε2 Δt 2/η 2/η 1 Δt 2/η 4ε 2 = Δt 2 Δt η 2 (1 Δt 2/η). Let Δt 0toget db(t) 4ε2 dt 2 η < 1, 2 since ε < η/2. Finally, we use the well-known perturbation theorem (attributed to Lidskii, Weyl, Wiedlandt and Mirsky in various forms [2, pp ]) to conclude that dλ j (t) dt db(t) 4ε2 dt 2 η < 1, 2 as expected. Theorem 3.4. For Hermitian Ãasin(1.1),ifε<η/2,thenÃhasmeigenvaluesθ 1,...,θ m (arranged as in (3.14)) satisfying ζ j θ j 1 4ρ, 2 for 1 j m, where ρ = ε/η < 1/2 and ζ j is defined by (3.13a). (3.18) Proof. Instead of proving (3.18) directly, we shall prove that for any given j {1,...,m} there are j of θ i s satisfying θ i ζ j /(1 4ρ 2 ). Thus (3.18)musthold. Adopt the notations in Lemmas 3.3 and 3.4. By(3.17), for any t ( η/2,η/2), wehave λ i (t) λ i (0) for 1 i m. Letδ j = t 0 dλ i (τ) dτ dτ 4ε2 t = 4ρ 2 t η 2 (3.19) ζ j 1 4ρ 2. We claim that there are at least j of λ i (t) such that λ i (t) [ δ j,δ j ] for all t [ δ j,δ j ]. (3.20) This means that each function λ i (t) maps the interval t [ δ j,δ j ] into itself. By Brouwer s fixed point theorem, each of such λ i (t) has a fixed point t i [ δ j,δ j ] such that λ i (t i ) = t i. Hence, recalling (3.2) we see that t i is an eigenvalue of Ã. Note that the second inequality in (3.17) implies that t i is a unique fixed point of λ i (t) in ( η/2,η/2). Therefore all counted, Ã has at least j eigenvalues in [ δ j,δ j ]. It remains to show that there are at least j of λ i (t) satisfying (3.20). To see this, we notice ϑ k [ ζ k,ζ k ] [ ζ j,ζ j ] [ δ j,δ j ] for 1 k j. These ϑ k for 1 k j are taken by j different λ i (t) at t = 0, i.e., ϑ k = λ lk (0), where l k {1,...,m} are distinct for k {1,...,j}.Wenowprovethatλ lk (t) for k {1,...,j} are the j of λ i (t) satisfying (3.20). In fact, for t [ δ j,δ j ] and k {1,...,j}

10 R.-C. Li et al. / Linear Algebra and its Applications 437 (2012) λ lk (t) λ lk (0) + λ lk (t) λ lk (0) = ϑ k + λ lk (t) λ lk (0) ζ j + 4ρ 2 δ j = δ j, as expected. Remark 3.3. Compared with (3.15)inTheorem3.3,theboundin(3.18) removes the term O(ε 4 ) at the expense of the factor (1 4ρ 2 ) 1. Example 2.1 (revisit). The following table displays the eigenvalues θ j of à nearest to 0, the eigenvalues ϑ j of E A 1 22 E, and the upper bounds in (3.18) and the ones after τ j replaced by mɛ 2 j. θ j ϑ j ζ j 1 4ρ 2 mɛ j ɛ m 1 4ρ The bounds are rather sharp. 4. Possible extensions to the generalized eigenvalue problem So far we have focused on the Hermitian eigenvalue problem (1.1). We now consider the following Hermitian definite generalized eigenvalue problem à = μb 11 E, B = B 11 F, (4.1) F B 22 where B ii 0, and F 2 is sufficiently small 6 so that B 0also. If E = F = 0, then μ is an eigenvalue of the pencil à λ B of multiplicity m. Inthissectionwe outline how to develop perturbation bounds using what we have gotten in Section Special case: B ii = Iandμ = 0 In this case, à = 0 E, B = I m F. (4.2) F I n Assume that F 2 < 1. A similar approach to the one in [6, Section 2.1] can be applied as follows. We first let X = I m F, W = I m 0 0 I n 0 [I FF, (4.3) ] 1/2 and then let def B = X BX = I m 0 0 I FF = W 2, (4.4a) 6 For example, F 2 < min i {σ min (B ii )} guarantees B 0, where σ min (B ii ) is the smallest singular value of B ii.

11 212 R.-C. Li et al. / Linear Algebra and its Applications 437 (2012)  def = X ÃX = 0 E E Â22, (4.4b) where W is the unique Hermitian definite square root [7, Chapter6]of B, and  22 = A 22 EF FE. à λ B has the same eigenvalues as W 1 ÂW 1 λi N.SinceW 1 ÂW 1 takes the form of (1.1), our theory in Section 3 applies to W 1 ÂW 1, leading to various bounds General case Now we consider the general case (4.1). Assume μ = 0; otherwise we shall consider (à μ B) λ B instead. Suppose à = 0 E, B = B 11 F. (4.5) F B 22 Set Y = diag(b 1/2 11, B 1/2 22 ) to get Y ÃY = 0 Ê, Y BY = I m F, (4.6) Ê Â22 F In which reduces to the case in Section 4.1, where  22 = B 1/2 22 A 22 B 1/2 22, F = B 1/2 22 FB 1/2 11, Ê = B 1/2 22 EB 1/2 11. (4.7) 5. Conclusion We established perturbation bounds for the multiple eigenvalue μ of Hermitian matrix A under a perturbation in the off-diagonal block: A = μi m 0 perturbed to à = μi m E, 0 A 22 with an emphasis on the case where the magnitudes of the columns of E vary widely. We show that whether A 22 μi m is definite or not plays a major role: if it is (positive or negative) definite, then à has m eigenvalues θ i (1 i m) such that the ith difference θ i μ is bounded by a quantity that is proportional to the square of the norm of E s ith column, but when A 22 μi m is indefinite the quantity is proportional to the product of the ith column norm and the norm of E. We also outline a possible extension to the Hermitian definite generalized eigenvalue problem. Acknowledgment We would like to thank the anonymous reviewers for their valuable comments and suggestions as thesehaveimprovedthepaper. References [1] C.-K. Li, R.-C. Li, A note on eigenvalues of perturbed Hermitian matrices, Linear Algebra Appl. 395 (2005) [2] G.W. Stewart, J.-G. Sun, Matrix Perturbation Theory, Academic Press, Boston, 1990.

12 R.-C. Li et al. / Linear Algebra and its Applications 437 (2012) [3] B.N. Parlett, The Symmetric Eigenvalue Problem, SIAM, Philadelphia, [4] Y. Nakatsukasa, On the condition numbers of a multiple eigenvalue of a generalized eigenvalue problem, Numer. Math., in press. arxiv: [5] K.R. Parthasarathy, Eigenvalues of matrix-valued analytic maps, J. Austral. Math. Soc. Ser. A 26 (1978) [6] R.-C. Li, Y. Nakatsukasa, N. Truhar, S.-F. Xu, Perturbation of partitioned hermitian generalized eigenvalue problem, SIAM J. Matrix Anal. Appl. 32 (2) (2011) [7] N.J. Higham, Functions of Matrices: Theory and Computation, SIAM, Philadelphia, PA, USA, 2008.