Eigenvalue coalescence for parameter dependent matrices
Transcript
Eigenvalue coalescence for parameter dependent matrices
Eigenvalue coalescence for parameter dependent matrices Alessandro Pugliese Università degli Studi di Bari Joint work with Luca Dieci (Georgia Inst. of Tech.) Seminar @ BCAM, Bilbao – February 14, 2011 Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 1 / 49 Based on: L. Dieci, A. P., “Two-parameter SVD: Coalescing singular values and periodicity”, SIAM J. Matrix Analysis (2009). L. Dieci, A. P., “Singular values of two-parameter matrices: An algorithm to accurately find their intersections”, Mathematics and Computers in Simulation (2008). L. Dieci, M.G. Gasparo, A. Papini, A. P., “Locating coalescing singular values of large two-parameter matrices”, Mathematics and Computers in Simulation (2011). L. Dieci, A. P. “Hermitian matrices depending on three parameters: Coalescing eigenvalues” (submitted). L. Dieci, M.G. Gasparo, A. Papini, A. P., “Continuation of smooth decompositions for Hermitian functions, with application to localization of conical intersections for 3-parameter functions” (in preparation). Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 2 / 49 Motivation Matrices that depend on parameters appear all the time in applied sciences and engineering. For instance, they are at the heart of dynamical systems studies, and spectral properties of these matrices are key in determining stability and bifurcations of dynamical invariants (say, equilibria and periodic orbits). Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 3 / 49 Motivation Of particular interest to us is the phenomenon of coalescing of eigenvalues of real symmetric or complex Hermitian matrices (as well as coalescing singular values for real/complex general matrices). This also has received a good deal of attention in several, seemingly disparate, fields. For example: Chemical Physics, conical intersections (e.g., Yarcony [2001]) Quantum Physics, diabolical points (Berry [1984]) Continuation techniques for 1D paths of decompositions (O’Neil, Bunse-Gerstner et alia, ... [from ’90s on]) Best approximation, model reduction, data compression, ... (e.g., Kock-Lubich [2007], Simon-Zha [1999]). (Coalescing leads to singularity!) Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 4 / 49 Motivation Structural Dynamics: Free vibration study using finite element models of beams, plates, ... (e.g., Srikantha Phani et al. [2006]) leads to an Hermitian eigenproblem, where: Eigenvalues are frequencies, eigenvectors are modes of vibration. Veering (a.k.a. “avoided crossing”) of eigenvalue curves is ubiquitous. It has been associated to the onset of (spatial) mode localization and hypersensitive behavior. Mode shape exchange: two distinct modes of vibration interchange upon small modifications of a parameter. The structure responds with a different mode when in internal/external parameter is slightly perturbed. (e.g., Pierre [1988], Ouisse-Guyader [2003]). Originally, it has been debated whether veering was induced by approximation or physically observable (Leissa [1974], Perkins-Mote [1986]). Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 5 / 49 Mathematical Framework We consider the following problems: Problem 1: Let Ω be a bounded region of R2 . Given a smooth symmetric matrix function A : Ω → Rn×n , find values x ∈ Ω where the eigenvalues of A coalesce. Problem 2: Let Ω be a bounded region of R3 . Given a smooth Hermitian matrix function A : Ω → Cn×n , find values x ∈ Ω where the eigenvalues of A coalesce. The singular values decomposition (SVD) of A is intimately related to several symmetric/Hermitian eigenproblems, such as those for AT A and AAT . So, we also consider the analogous problems for the singular values. Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 6 / 49 Background (1 parameter) Differentiability of factors for 1 parameter functions is well established: A analytic =⇒ Schur factors U, Λ are analytic Kato [analytic Schur for Hermitian case, 1976] Bunse-Gerstner, Byers, Mehrmann & Nichols [analytic SVD, 1991] A of class C k =⇒ eigenvalues C k , possible loss of differentiability of orthogonal factor if there are degenerate eigenvalues Dieci & Eirola [Schur and SVD for C k case, 1999] Chern & Dieci [Schur and SVD for periodic C k case, 2000. A of period 1 with distinct eigenvalues: Factors of period 1 or 2?] Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 7 / 49 Background (several parameters) Example (Kato) Complete loss of smoothness even for analytic A: q x1 x2 ; λ± (x1 , x2 ) = ± x21 + x22 A(x1 , x2 ) = x2 −x1 The coalescing causes loss of differentiability in the eigenvalues and loss of continuity in the eigenvectors. Contrast with 1-parameter case! Little hope to detect a coalescing point by simply continuing curves of eigenvalues (or singular values) by freezing one parameter at the time. A different approach is needed. Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 8 / 49 Background (several parameters) Important: If the eigenvalues stay distinct, it is possible to maintain differentiability of the factors also when A depends on several parameters. In this case, U (as well as V for the SVD) is unique except for changes of sign of its columns (real symmetric case) or multiplication of its columns by phase factors eiθ (complex Hermitian case). Alternatively, one can smoothly obtain block reduction in disjoint groups of eigenvalues/singular values. [Hsieh-Sibuya, Gingold] Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 9 / 49 Genericity and Codimension Coalescing of eigenvalues of real symmetric matrices is a codimension 2 phenomenon. This fact had first been observed by von Neumann and Wigner more than 80 years ago (noncrossing rule). Informally: 2 × 2 real symmetric matrices: a b b d 2 × 2 real symmetric matrices with coalescing eigenvalues: a 0 0 a (Triple coalescing has codimension 5, “two-pairs” coalescing has codimension 4, and so on) Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 10 / 49 Genericity and Codimension Unlike the real case, coalescing of eigenvalues of complex Hermitian matrices is a codimension 3 phenomenon. Informally: 2 × 2 complex Hermitian matrices: a b + ic b − ic d 2 × 2 complex Hermitian matrices with coalescing eigenvalues: a 0 0 a formal justification of the n × n case goes through reduction to block-diagonal form Key fact: The codimension tells us how many parameters we need in order to expect the phenomenon to occur. Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 11 / 49 Genericity and Codimension Consequences for real symmetric matrix functions. Generically: one parameter we expect a smooth matrix valued function to have distinct eigenvalues (and consequently a smooth Schur decomposition) when it occurs, coalescing of eigenvalues does not persist under perturbation two parameters coalescing of eigenvalues has to be expected it occurs at isolated points it persists under small perturbations three parameters coalescing of eigenvalues is expected to occur along curves Similar considerations apply to the complex Hermitian case. Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 12 / 49 Genericity and Codimension 1.25 1.2 1 1 0.5 0.5 0 0 −0.5 −0.5 −1 −1 1.15 0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1 0 0.5 1 1.1 1 1 0.5 1.05 1 0 0.2 0.4 0.6 0.8 1 0.5 0 0 −0.5 −0.5 −1 −1 0 0.5 1 1 parameter: unperturbed matrix 1.25 1.2 1 1 0.5 0.5 0 0 −0.5 −0.5 −1 −1 1.15 0 0.5 1 1.1 1 1 0.5 1.05 1 0 0.2 0.4 0.6 0.8 1 0.5 0 0 −0.5 −0.5 −1 −1 0 0.5 1 1 parameter: perturbed matrix Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 13 / 49 Genericity and Codimension 2 parameters: unperturbed matrix Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 14 / 49 Genericity and Codimension 2 parameters: perturbed matrix Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 15 / 49 Genericity and Codimension 2 parameters: zoom in on a point where eigenvalues coalesce Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 16 / 49 Reminder “Avoided crossing” of eigenvalue curves, veering, mode shape exchange ... avoided crossing: a slice of the “double-cone” near a coalescing point? This picture is ubiquitous whenever only 1 parameter is varied. It suggests that a conical intersection could be “nearby”, i.e. that it could be obtained by “freeing one more parameter”. Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 17 / 49 Reminder 2 parameters: Slicing near a conical intersections Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 18 / 49 Theoretical Results, 2 parameters Theorem (real symmetric (2 × 2) case) Consider P ∈ C k (Ω, R2×2 ), k ≥ 1, symmetric. For all x ∈ Ω, write a(x) b(x) P (x) = , b(x) d(x) and let λ1 (x) ≥ λ2 (x) be its two continuous eigenvalues, for all x in Ω. Assume that there exists a unique point ξ0 ∈ Ω where the eigenvalues coalesce: λ1 (ξ0 ) = λ2 (ξ0 ). Consider the C k function F : Ω → R2 given by: a(x) − d(x) F (x) = , b(x) and assume that 0 is a regular value for both functions a − d and b. Then, consider the two C k curves Γ1 and Γ2 through ξ0 , given by: Γ1 = {x ∈ Ω : a(x) − d(x) = 0}, Γ2 = {x ∈ Ω : b(x) = 0}. Assume that Γ1 and Γ2 intersect transversally at ξ0 . .. . Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 19 / 49 Theoretical Results, 2 parameters Theorem (real symmetric (2 × 2) case cont’d) .. . Let Γ be a simple closed curve enclosing the point ξ0 , parametrized as a C p (p ≥ 0) 1-periodic function γ in the variable t: γ : t ∈ R → Ω. Let m = min(k, p), and let Pγ be the C m function P (γ(t)), t ∈ R. Then, for all t ∈ R, Pγ (t) has the eigendecomposition Pγ (t) = Uγ (t)Λγ (t)UγT (t) such that: (i) Λγ ∈ C1m (R, R2×2 ) and diagonal: Λγ (t) = λ1 (γ(t)) 0 0 for all λ2 (γ(t)) t ∈ R; (ii) Uγ ∈ C2m (R, R2×2 ) real orthogonal, and Uγ (t + 1) = −Uγ (t) for all t ∈ R. Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 20 / 49 Theoretical Results, 2 parameters a(x) − d(x) = 0 b(x) = 0 Γ C Ω ξ0 “Generic coalescing point”: Transversal Intersection at ξ0 Key idea: the first (respectively, second) component of the eigenvector associated to λ1 changes sign exactly when C crosses the curve b(x) = 0 and a(x) − d(x) < 0 (respectively, a(x) − d(x) > 0). Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 21 / 49 Theoretical Results, 2 parameters In other words, the eigenvectors have changed sign, they flipped over (cfr. Herzberg & Longuet-Higgins [1963]). Bulding on the previous result, repeatedly using block reduction techniques, and more ... Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 22 / 49 Theoretical Results, 2 parameters Theorem (real symmetric n × n case) Let A : Ω → Rn×n be C k and symmetric on Ω. Let γ : [0, 1] → Ω be a (smooth) simple closed curve, s.t. A(γ(t)) has distinct eigenvalues for all t. Let A(γ(t)) = T λ (t) 1 | | | | .. = u1 (t) · · · un (t) u1 (t) · · · un (t) . | | | λn (t) | | | {z }| {z } {z } U(t) Λ(t) U(t)T be the (smooth) Schur decomposition of A along γ. The following holds: If γ encloses an odd number of generic coalescing points for λi , then ui (t) changes sign along the circuit: ui (0)T ui (1) = −1 . Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 23 / 49 Theoretical Results, 2 parameters Theorem (Converse result) Let A : Ω → Rn×n be C k and symmetric on Ω. Let γ : [0, 1] → Ω be a (smooth) simple closed curve s.t. A(γ(t)) has distinct eigenvalues for all t. Let A(γ(t)) = U (t)Λ(t)U (t)T be the (smooth) Schur decomposition of A along γ. Consider D = U (0)T U (1). Let i1 < i2 < · · · < i2q be the indices for which Dii = −1. Pair them together (i1 , i2 ), . . . , (i2q−1 , i2q ). If i2h−1 ≤ i < i2h for some 1 ≤ h ≤ q, then λi = λi+1 at least once inside γ, and generically at an odd number of points inside γ. Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 24 / 49 Theoretical Results, 2 parameters Remark: We expect all coalescing points to have multiplicity 1 (this is the generic property). To illustrate the previous Theorem, suppose we have n ≥ 4, and −1 1 1 D= −1 .. . Then, we expect that inside the region encircled by Γ, the pairs (λ1 , λ2 ), (λ2 , λ3 ), and (λ3 , λ4 ), have coalesced. Notice: If λi coalesces at an even number of (generic) points inside γ, the corresponding eigenvector does NOT change sign. Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 25 / 49 Extension, 2 parameters For large dimensional problems one may want to monitor only a small number of eigenvalues, say just q n dominant eigenvalues/eigenvectors of A. We work with the reduced Schur decomposition along γ: T U (q) (t) A(γ(t)) U (q) (t) = Λ(q) (t) , and we may have an odd number of −1 on the diagonal of T D(q) = U (q) (0) U (q) (1). Example (q=4) If we have D(q) 1 = 1 −1 1 then (λ3 , λ4 ) and (λ4 , λ5 ) coincide at an odd number of points inside γ. Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 26 / 49 Algorithm, 2 parameters The previous results lend naturally to find regions of Ω inside which some eigenvalues of A coalesce: change of sign in the matrix of eigenvectors along a loop Γ ⇓ coalescing pair of eigenvalues inside Γ. Observations. Period 2 implies coalescing: Detection phase Γ only needs to be continuous and does not have to be near the coalescing point (not a local result) Topological result, analogous of the I.V.T. of Calculus: “Change of sign of a continuous function implies a root” Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 27 / 49 Algorithm, 2 parameters 1) given A(xi,j ) = U0 Λ0 U0T xi+1,j compute continuous Schur of A along red and blue paths xi+1,j+1 2) save orthogonal factors at xi+1,j+1 : x i,j U1 , U2 xi,j+1 3) compare U1 and U2 : D= (U1 )T U2 Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 28 / 49 Algorithm, 2 parameters A 1D solver for computing smooth Schur decomposition paths is required (1D continuation is not trivial) Along the 1D paths, step-size is chosen adaptively Robust: handle the case when eigenvalues coalesce and distinguish it from the (computationally harder) case of near-coalescing This is the workhorse of the algorithm 1.25 1.2 1 1 0.5 0.5 0 0 −0.5 −0.5 −1 1.15 −1 0 0.5 1 0 0.5 1 0 0.5 1 1.1 1.05 1 0 0.2 0.4 0.6 0.8 1 1 1 0.5 0.5 0 0 −0.5 −0.5 −1 −1 0 Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence 0.5 1 February 14, 2011 29 / 49 Algorithm, 2 parameters Grid resolution is very important Detection phase is highly parallelizable Once we have a region where the i-th and (i + 1)-st eigenvalues coalesce, we may proceed in two ways to accurately locate the parameter value where coalescing occurs: 1. progressive refinement of the region 2. Newton (modified) to find a root of the gradient of f (x) = λi (x) − λi+1 (x) Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence 2 February 14, 2011 30 / 49 Examples, 2 parameters 1 0.8 0.6 - Ω = [−1, 1] × [−1, 1] - 2 × 2 grid 0.4 0.2 - λ1 = λ2 , λ2 = λ3 0 box B1 1 B2 1 −0.2 −0.4 pair 2-3 2-3 iterates 7 6 −0.6 −0.8 −1 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 Alessandro Pugliese (Università degli Studi di Bari) 0.4 0.6 0.8 1 Eigenvalue coalescence February 14, 2011 31 / 49 Examples, 2 parameters - 10 × 10 grid 1 - λ1 = λ2 , λ2 = λ3 0.8 0.6 0.4 0.2 0 −0.2 −0.4 −0.6 −0.8 −1 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 Alessandro Pugliese (Università degli Studi di Bari) 0.4 0.6 0.8 1 Eigenvalue coalescence box B4 1 B5 1 B7 2 B9 2 B4 3 B8 3 B3 8 B4 8 B2 9 B4 10 pair 1-2 1-2 1-2 1-2 2-3 2-3 2-3 2-3 1-2 1-2 iterates 8 7 – 7 6 6 7 – 7 8 February 14, 2011 32 / 49 Examples, 2 parameters - 100 × 100 boxes 1 - λ1 = λ2 , λ2 = λ3 0.8 0.6 0.4 0.2 0 −0.2 −0.4 −0.6 −0.8 −1 −1 −0.8 −0.6 −0.4 −0.2 0 0.2 Alessandro Pugliese (Università degli Studi di Bari) 0.4 0.6 0.8 1 Eigenvalue coalescence box B34 9 B48 9 B70 13 B82 19 B34 26 B72 29 B29 72 B34 75 B19 82 B34 92 pair 1-2 1-2 1-2 1-2 2-3 2-3 2-3 2-3 1-2 1-2 iterates 5 5 5 5 5 5 5 6 6 6 February 14, 2011 33 / 49 Examples, 2 parameters −0.6 −0.7 −0.8 0.2 0.3 0.4 Local refinement: box B7 2 Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 34 / 49 Examples, 2 parameters 300 250 200 150 100 50 0 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 −0.8 −0.6 −0.4 −0.2 0 0.2 0.4 0.6 0.8 Figure: distribution of the workload over the region Ω Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 35 / 49 Examples, 2 parameters Behavior on a large dimensional problems: A(x1 , x2 ) ∈ Rn×n , Ω = [−1, 1] × [−1, 1], grid 10 × 10, follow q = 6 dominant eigenvalues, pairs of eigenvalues coalesce at ∼ 20 points. Table: Experiments, m = 3200 n 200 400 800 1600 3200 average # of steps 132 129 135 137 151 Alessandro Pugliese (Università degli Studi di Bari) min-max # steps 75-679 72-503 71-642 77-702 71-549 Eigenvalue coalescence rejected steps 1635 1558 1808 1682 1819 February 14, 2011 36 / 49 Theoretical Results, 3 parameters What if A is complex Hermitian? real symmetric complex Hermitian 1D smooth continuation uniquely defined (eigenvectors defined up to sign change) NOT uniquely defined (eigenvectors defined up to phase factor eiθ ) codimension of coalescing 2 3 It is natural to work with 3 parameters, eigenvalue coalescence is expected to occur at isolated points in the 3-dimensional space. New ideas are needed. Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 37 / 49 Theoretical Results, 3 parameters How to resolve the non-uniqueness of smooth continuation along 1D paths? Several different approaches have been proposed: Stone [1976]: Im u∗j (t + dt)uj (t) = O(dt2 ) for all t and j Bunse-Gerstner et alia [1991]: minimize (Minimum Variation Decomposition, MVD) Dieci & Eirola [1999]: Rb U̇ dt a F u∗j (t)u̇j (t) = 0 for all t and j They all lead to the same unitary Schur factor! Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 38 / 49 Theoretical Results, 3 parameters Let us now consider loops in parameter space: γ : t ∈ [0, 1] 7→ γ(t) ∈ R3 , smooth, with γ(0) = γ(1). Suppose, all eigenvalues of A stay distinct for all t, and let A(γ(t)) = U (t)Λ(t)U ∗ (t) be a Minimum Variation Decomposition. How are U (0) and U (1) related? We have: iα e 1 U (1) = U (0) .. . eiαn For each j = 1, . . . , n, αj ∈ (−π, π] is know as the Berry phase (or geometric phase) associated to λj . Note: A real symmetric =⇒ αj = 0, π ⇐⇒ eiαj = ±1 Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 39 / 49 Theoretical Results, 3 parameters Stone insight: variation of Berry phase as a loop covers a sphere-like surface is related to presence (or lack thereof) of coalescing points inside the surface. To fix ideas, let us think of the family of parallels (lines of constant latitude) {γs }s∈[0,1] that covers S2 , where γ0 , γ1 represent (respectively) north pole and south pole. We can define n smooth maps: αj : s ∈ [0, 1] 7→ αj (s) , where αj (s) is the Berry phase associated to λj along γs . αj (s) must be allowed to go beyond (−π, π] (to retain smoothness) γ0 and γ1 are constant; therefore αj (0) can (and will) be chosen to be 0, and αj (1) has to be a multiple of 2π (A constant along γ0,1 ⇒ U(1)=U(0)). Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 40 / 49 Theoretical Results, 3 parameters The same considerations apply to more general surfaces and loops’ coverings. We call S a 2-sphere if it is the embedded images of S2 in R3 under a smooth map f . f − → Definition (Stone) We say that a 2-sphere S is phase rotating if αj (1) 6= 0 for at least one j. Otherwise, we say that S is phase preserving. Remark: Definition doesn’t depend on the choice of loops’ covering. Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 41 / 49 Theoretical Results, 3 parameters Theorem (complex Hermitian 2 × 2 case) Let A ∈ C 1 (Ω, C2×2 ) be Hermitian. Let λ1 (ξ) ≤ λ2 (ξ) be its continuous eigenvalues, for all ξ ∈ Ω. Let ξ0 ∈ Ω be a “generic” coalescing point for A in Ω, and suppose that ξ0 is the only coalescing point in Ω. Let S ⊂ Ω be a 2-sphere. If the interior part of S contains ξ0 , then S is phase-rotating, with α1,2 (1) = ±2π. proof makes use of powerful topological results: show the result is true for S2 and then extend it to S via homotopy arguments. repeatedly using block reduction techniques, and more ... Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 42 / 49 Theoretical Results, 3 parameters Theorem (complex Hermitian n × n case) Let A ∈ C 1 (Ω, Cn×n ) be Hermitian. Let λ1 (ξ), . . . , λn (ξ) be its continuous eigenvalues, labeled in ascending order. Suppose that, for any j = 1, . . . , n − 1, we have: λj = λj+1 solely at dj distinct generic coalescing points in Ω, with a total numer of Pn−1 N = j=1 dj distinct coalescing points. Let S ⊂ Ω be a 2-sphere whose interior part contains all those coalescing points. Then we have that: (i) α1 (1) = 0 (resp. 2π) (mod 4π) (ii) αj (1) = 0 (resp. 2π) (mod 4π) (iii) αn (1) = 0 (resp. 2π) if d1 is even (resp. odd) if dj−1 + dj is even (resp. odd) and j = 2, . . . , n − 1 (mod 4π) if dn is even (resp. odd) with |αj (1)| ≤ 2dj π for j = 1, n and |αj (1)| ≤ 2(dj−1 + dj )π for j = 2, . . . , n − 1. Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 43 / 49 Theoretical Results, 3 parameters Observations: Stone went in the opposite direction, and only considered the case of one coalescing point “phase rotations” add up nicely a truly topological result, that lends nicely to find regions of R3 where eigenvalues coalesce again, an even number of coalescing points involving the same eigenvalue may produce no phase rotation Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 44 / 49 Algorithm, 3 parameters subdivide region in cubes cover each cube with square-shaped loops, compute Berry phase along each loop you have obtained n functions αj (s), where s parametrizes the red C-shaped meridian if a cube is phase rotating, zoom-in to locate coalescing point accurately Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 45 / 49 Algorithm, 3 parameters Algorithmic details: 1D continuation of the MVD Schur is performed via a 2 steps process. At step k, with U (k) ≈ U (tk ): construct “an” orthogonal Schur factor at tk+1 , call it Q(k+1) rotate Q(k+1) to enforce minimum variation w.r.t. U (k) : U (k+1) = Q(k+1) diag([eiθk , k = 1, . . . , n]) trivial predictor is O(h3 ) in enforcing minimum variation, surprising! step-size is chosen adaptively along parallels (square shaped loops) and meridian (C-shaped curve) when a cube is declared phase rotating, multiple seeds are corrected via a Newton-like procedure (zoom-in) to locate the coalescing point accurately in case of zoom-in failure (or if multiple coalescing points are detected) local refinement is performed (cube is divided in 8 sub-cubes) the code is highly parallelizable by nature it has proved to be very robust Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 46 / 49 Examples, 3 parameters 15 10 λ 5 Berry phase 1 λ 2 λ 3 0 −5 −10 0 0.5 1 1.5 2 2.5 s Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 47 / 49 Examples, 3 parameters 15 10 5 λ Berry phase 1 λ 2 λ 3 0 λ 4 λ 5 λ 6 −5 −10 −15 0 0.5 1 1.5 2 2.5 s Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 48 / 49 Work To Do complete the study of an alternative approach for complex Hermitian problem (based on a perturbative idea) analyze problems arising in Structural Engineering generalized eigenvalue problem non-generic coalescence of eigenvalues study eigenvalue coalescence for non symmetric/Hermitian matrix functions Alessandro Pugliese (Università degli Studi di Bari) Eigenvalue coalescence February 14, 2011 49 / 49