You are on page 1of 32

INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING

Int. J. Numer. Meth. Engng 2006; 00:1–6 Prepared using nmeauth.cls [Version: 2002/09/18 v2.02]

On the solution of generalized non-linear complex-symmetric


eigenvalue problems

N. A. Dumont∗, †

Departamento de Engenharia Civil


Pontifı́cia Universidade Católica do Rio de Janeiro – PUC-Rio
22451-900 Rio de Janeiro, Brazil

SUMMARY

This paper brings an attempt toward the systematic solution of the generalized non-linear, complex-
symmetric eigenproblem (K0 − iωC1 − ω 2 M1 − iω 3 C2 − ω 4 M2 − · · · )φ = 0, with real, symmetric
matrices K0 , Cj , Mj ∈ Rn×n , which are associated to the dynamic governing equations of a structure
submitted to viscous damping, as laid out in the frame of an advanced mode superposition technique.
The problem can be restated as (K(ω) − ωM(ω) )φ = 0, where K(ω) = KT T
(ω) and M(ω) = M(ω) are
complex-symmetric matrices given as power series of the complex eigenfrequencies ω, such that, if
(ω, φ) is a solution eigenpair, φT M(ω) φ = 1 and φT K(ω) φ = ω. The traditional Rayleigh quotient
iteration and the more recent Jacobi-Davidson method are outlined for complex-symmetric linear
problems and shown to be mathematically equivalent, both with asymptotically cubic convergence.
The Jacobi-Davidson method is more robust and adequate for the solution of a set of eigenpairs.
The non-linear eigenproblem subject of this paper can be dealt with in the exact frame of the linear
analysis, thus also presenting cubic convergence. Two examples help to visualize some of the basic
concepts developed. Three more examples illustrate the applicability of the proposed algorithm to
solve non-linear problems, in the general case of underdamping, but also for overdamping combined
with multiple and close eigenvalues. Copyright ° c 2006 John Wiley & Sons, Ltd.

key words: non-linear eigenproblems, advanced modal analysis, Rayleigh quotient iteration, Jacobi-
Davidson method, complex-symmetric matrices

1. INTRODUCTION

1.1. Problem justification


A variationally-based finite/boundary element formulation is in the background of the present
developments. It was originally conceived for static problems [1, 2] as an extension of Pian’s

∗ Correspondence to: N. A. Dumont, Departamento de Engenharia Civil, Pontifı́cia Universidade Católica do


Rio de Janeiro – PUC-Rio, 22451-900 Rio de Janeiro, Brazil.
† E-mail: dumont@civ.puc-rio.br

Contract/grant sponsor: Brazilian agency CNPq (Projects Nr. 475153/2003-0 and 301227/2003-9).

Received May 2006


c 2006 John Wiley & Sons, Ltd.
Copyright ° Revised
2 N. A. DUMONT

hybrid finite element method [3, 4] and as a counterpart to the traditional boundary element
method [5]. Later on, the formulation was extended to the analysis of 2D and 3D transient
problems [6, 7, 8] built up on a proposition by Przemieniecki [9] intended for the free-vibration
analysis of truss and beam structures.
The resulting advanced mode superposition technique [6, 7] starts with a frequency-domain
formulation in terms of a power series expansion and requires the solution of a non-linear
eigenvalue problem. General domain actions as well as general boundary and initial conditions
are almost as straightforward to deal with [6, 7, 10, 11, 12] as in classical dynamics [13, 14, 15].
The inclusion of viscous damping makes the eigenproblem complex symmetric [10]. The non-
linear eigenvalue problem was firstly solved using linearization, which leads to large, although
sparse, matrices. The present paper is together with References [16, 11] an attempt to lay out
the theoretical basis of the proposed dynamic formulation.

1.2. Paper organization


A systematic review of the technical literature on the solution of linear and non-linear
eigenproblems is beyond the scope of this paper. A general mathematical outline of the non-
linear problem has been accomplished by Hadeler as early as in the year 1967 [17, 18]. See
Voss [19] for an illuminating account of the subject including the historically most important
contributions. Some basic references on eigenproblems are [13, 20, 21, 22, 23, 24].
Section 2 deals with the solution of linear, complex-symmetric eigenproblems. The Rayleigh
quotient iteration and the Jacobi-Davidson method are presented and compared with each
other, to make evident that they can be applied almost unchanged to the class of non-linear
problems one is actually concerned with.
This is the subject of Section 3, for an effective stiffness matrix, Equation (29), formulated
in terms of generalized damping and mass matrices. One firstly outlines the real-symmetric
eigenproblem, in Section 3.2, to which the results of Section 2 can be directly extended. The
outline makes use of sets of linearized eigenvalue problems in terms of augmented matrices with
the purpose of deriving generalized linear-algebra properties of the original non-linear problem.
Some theorems are presented in order to (a) establish the equivalence between the augmented,
linearized system and the original non-linear problem, (b) assess the positivity of the stiffness
and mass polynomial matrices, (c) determine the number of real solutions comprised in the
formulation and (d) assess the existence range of the real eigenvalues.
The non-linear, complex-symmetric problem is formulated in Section 3.3. It is shown that
the eigensolutions are given in complex-associated pairs, in the general case, as corresponding
to underdamping. The eigenproblem related to overdamping (and to the unlikely case of
critical damping) is also discussed. Finally, a perturbation analysis is presented to provide
the mathematical justification for finding the complex solutions of interest starting from the
solutions of the underlying real-symmetric eigenproblem.
Five examples are given. The first example shows the stiffness and mass polynomial matrices
of the simplest mathematical model one may conceive, namely a truss element with two degrees
of freedom. The second example, for the heat conduction in a square plate, helps to visualize
the eigenvalue properties of a real-symmetric problem, as outlined in Section 3.2. The third
example makes use of the matrices of the first example applied to a fixed-free damped bar to
illustrate the iterative eigenvalue solution with the Jacobi-Davidson algorithm of Section 3.5.
The fourth example resources to the matrices of the second example to create an artificial,

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
NON-LINEAR COMPLEX-SYMMETRIC EIGENVALUE PROBLEMS 3

mathematically challenging, problem with overdamping and multiple eigenvalues. The last
example deals with a rail-pad-sleeper-ballast dynamic interaction model of a railway track, in
which several clusters of close eigenvalues have to be evaluated.

2. SOLUTION OF LINEAR, COMPLEX-SYMMETRIC EIGENPROBLEMS

2.1. Introduction
This Section deals with the linear eigenvalue problem
(K − λM) φ = 0 (1)
m×m T T
where K, M ∈ C are complex-symmetric matrices, K = K, M = M, and (λ, φ) is a
generic, complex eigenvalue and eigenvector solution pair. Bold, lower case is used for vectors;
bold, upper case is used for matrices. An eigenvalue is generally characterized by the letter
λ, thus encompassing both the real-symmetric formulation of Section 3.2, with real, positive
λ ≡ ω 2 , and the complex-symmetric formulation of Section 3.3, which leads to complex λ ≡ ω.
Stiffness and mass matrix polynomials of a variable θ are denoted by K(θ) and M(θ) .
Horn and Johnston [20] are the best reference on complex-symmetric matrices, whose
properties are also explored by Arbenz and Hochstenbach [25] in their development of the
Jacobi-Davidson method. These two references, together with the classical book by Bathe and
Wilson for engineers [13], were of invaluable help for the outline of the following sections.
Definition 1 (Inner product for complex-symmetric eigenproblems) Given two com-
plex vectors x, y ∈ C m , and given a complex-symmetric matrix M ∈ C m×m , one defines the
inner product for the complex-symmetric eigenproblem as the indefinite bilinear form
hx, yi = xT My (2)
Strictly speaking, this is not an inner product definition, as it cannot be guaranteed that
hx, xi ≥ 0 for all x and that hx, xi = 0 only if x = 0, that is, x may be a quasi-null or isotropic
vector (see Definition 5.1.3 in Reference [20]).
The eigenproblem of interest given by Equation (1) – as applied to structural dynamics –
requires that K and M be simultaneously diagonalizable. In general, a matrix M ∈ C m×m
is diagonalizable if it has m linearly independent eigenvectors. As complex symmetric, M
is diagonalizable if and only if M = QDQT , where D ∈ C m×m is a diagonal matrix and
Q ∈ C m×m satisfies QT Q = I, as given in Theorem 4.4.13 of Reference [20]. (K and M are
not necessarily normal, whence Q is generally complex.) Accordingly, all eigenpair solutions
(λ, φ) of Equation (1) may be gathered in the pair of matrices (Λ, Φ), where Λ is diagonal
and Φ is a non-singular matrix that can be normalized such that
ΦT MΦ = I, ΦT KΦ = Λ, ΦT Φ = D−1 (3)
The use of the inner product Definition 1 is justified in the sense that, if convergence is
occurring in an iterative procedure for the solution of the eigenproblem of Equation (1), then
all approximations u of an eigenvector φ lay in a subspace of C m for which hu, ui 6= 0 as well
as uT u 6= 0. Even in the case of a real eigenproblem, for which the inner product of Definition
1 has no restrictions, convergence of an iterative process is only asymptotical [13].
The following projector definitions will be helpful in the solution of the eigenvalue problem,
as outlined in Section 2.2. The notation is in accord with Reference [26].

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
4 N. A. DUMONT

Definition 2 (Orthogonal projectors)


MφφT M
PM φ = (4)
φT MMφ
is the orthogonal projector onto the subspace of C m spanned by the non-isotropic vector Mφ:
PM φ Mφ = Mφ (5)
m
This is consistent with the inner product Definition 1, since, given two vectors x, y ∈ C ,
hx, yi = 0 ⇔ PM x y = 0 ⇔ PM y x = 0 (6)
As an orthogonal projector, PM φ is both symmetric and idempotent. Moreover,
P(M φ)⊥ = I − PM φ (7)
Definition 3 (Oblique projectors)
φφT M
Pφ,(M φ)⊥ = (8)
φT Mφ
is the oblique projector onto the subspace spanned by φ along the subspace spanned by (Mφ)⊥ :
Pφ,(M φ)⊥ φ = φ ⇔ PT
φ,(M φ)⊥ Mφ = Mφ (9)
Pφ,(M φ)⊥ is idempotent and generally non-symmetric (for M 6= I). This definition is also
consistent with the inner product Definition 1, as, given two vectors x, y ∈ C m ,
hx, yi = 0 ⇔ Px,(M x)⊥ y = 0 ⇔ Py,(M y)⊥ x = 0 (10)
Observe that PT
φ,(M φ)⊥ ≡ P(M φ),φ⊥ . Moreover,

PM φ Pφ,(M φ)⊥ = PM φ , Pφ,(M φ)⊥ PM φ = Pφ,(M φ)⊥ (11)


The clause on the normalization of u is not strictly required in the following definition, but is
convenient to simplify notation in the subsequent developments.
Definition 4 (Rayleigh quotient) Let u be a vector normalized according to the inner
product Definition 1, provided that u is not quasi-null. The Rayleigh quotient is defined as
uT Ku
θ= ≡ uT Ku (12)
uT Mu
The following Rayleigh quotient iteration [13, 21, 22] is referred to in the outline of the Jacobi-
Davidson method of Section 2.2.
Algorithm 1 (Rayleigh quotient iteration) The procedure starts with an estimated
eigenvalue θ1 and an estimated iteration vector u1 , normalized according to Definition 1. Define
b1 := Mu1 . Then, for k = 1, 2, · · · :
(1) solve f or ũk+1 in (K − θk M) ũk+1 = bk
(2) evaluate uk+1 by normalizing ũk+1 according to Def inition 1
(3) evaluate bk+1 := Muk+1
(4) evaluate θk+1 := uT T
k+1 bk + θk ≡ uk+1 Kuk+1 as in Def inition 4

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
NON-LINEAR COMPLEX-SYMMETRIC EIGENVALUE PROBLEMS 5

If convergence occurs, bk+1 → Mφ and θk+1 → λ as k → ∞. The iteration is repeated until

| (θk+1 − θk ) /θk+1 |≤ tol (13)

provided that θk+1 > 0. If the system matrix in step (1) is singular, then θk is already the
eigenvalue one is looking for and uk is the corresponding eigenvector.
Algorithm 1 follows a proposition in Reference [13] for a computationally efficient sequence
of calculations. The critical aspect of the algorithm – namely the solution of the equation
system of step (1) – is not addressed here. See References [27, 28, 29, 30, 31, 23], for instance,
which are mostly devoted to non-linear problems, but also present implementations in terms
of inexact algorithms that might increase the overall efficiency of the method.
A right eigenvector of a complex-symmetric eigenproblem is also a left eigenvector. Then,
provided that Definition 1 holds, the following theorem may be proved [25].
Theorem 1 (Quadratic approximation of the Rayleigh quotient) If a vector uk ap-
proaches the eigenvector φ with error O (δk ), then the¡ corresponding
¢ Rayleigh quotient θk of
Definition 4 approaches the eigenvalue λ with error O δk2 .
In the case of a Hermitian eigenproblem, approximation of the Rayleigh quotient is linear.
Convergence of the iterative process can be assessed as in the case of real-symmetric
eigenproblems [13, 25] and stated, for convenience, in terms of the following theorem.
Theorem 2 (Cubic convergence of the Rayleigh quotient iteration) If a vector uk
approaches the eigenvector φ with error O (δk ), then, as obtained in the frame of the Rayleigh
quotient iteration
¡ ¢ of Algorithm 1 and according to Theorem 1, uk+1 for k > 1 approaches φ
with error O δk3 .
As convergence is attained, K − θk M tends to become singular and the elements of the
non-normalized vector ũk+1 in Algorithm 1 tend to become very large numbers. However, if
the eigenvector φ is well conditioned, the error resulting from the solution of ũk+1 is mainly in
the direction engendered by φ, which is the searched direction [32]. In such a case, machine-
precision accuracy may be obtained for the numerical results. The process of proving the
latter theorem [13, 25] enables to generalize (for Hermitian eigenproblems and also for k = 1)
that, if a vector uk approaches the eigenvector φ with error O (δk ) and a value θk approaches
the eigenvalue λ with error O (²k ), then uk+1 approaches φ with error O (²k δk ). As a result,
starting with θ1 equal to an eigenvalue λ within machine precision and with a vector u1 that
is not orthogonal to the searched eigenvector φ, only one iteration step is needed in Algorithm
1 to achieve convergence within machine precision [32].
As shown in the following, the equation system given in Algorithm 1 may be rearranged, so
the increasing ill-conditioning of K − θk M is adequately dealt with.

2.2. Jacobi-Davidson method for complex-symmetric eigenproblems


2.2.1. Problem formulation. The Jacobi-Davidson method was introduced in 1994 by Sleijpen
and van der Vorst [25, 33, 34] as an extension of Davidson’s method [35]. It is a member of a
large family of recently developed projection algorithms, as the Lanczo’s method, the Arnoldi’s
method and the rational Krylov method, which may be extended to non-linear problems. A
few references are [27, 28, 30, 31, 36, 37, 38]. Comprehensive problem formulation, literature
review and algorithm outlines are found in References [29, 23, 39].

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
6 N. A. DUMONT

In the following, the Jacobi-Davidson method will be derived directly from – and will
be shown to be mathematically equivalent to – the Rayleigh quotient iteration, therefore
with asymptotically cubic convergence. Although not strictly original (see Reference [25],
for instance), some didactic merit may be claimed for the present outline, particularly as
a connection will be established to the Bott-Duffin inverse [26, 40].
Given an approximate eigenpair solution (θk , uk ) of the problem introduced in Equation (1),
where θk is the Rayleigh quotient of Definition 4, one may write the improved, non-normalized,
solution of ũk+1 in step (1) of Algorithm 1 as
ũk+1 = (uk + δuk ) βk (14)
where βk is a scaling factor – actually, an increasingly large number as convergence is attained.
The vector increment δuk is necessarily orthogonal to uk in terms of Definition 1,
huk , δuk i = 0 (15)
and presents cubic convergence to zero in the frame of Theorem 2.
The solution for ũk+1 in Algorithm 1 is equivalent to solving the following restricted system
of equations for δuk , given Equation (15) and the orthogonal projector of Definition 2:

 (K − θk M) δuk − Muk /βk = − (K − θk M) uk
restricted to PM uk δuk = 0 (16)

given that PM uk Muk = Muk
with subsequent evaluation of ũk+1 from Equation (14).

2.2.2. Solution in terms of orthogonal projector. Assuming that (K − θk M) P(M uk )⊥ +PM uk


is non-singular, the solution of Equation (16) may be uniquely expressed as the solution of the
following equation for the auxiliary vector δyk :
£ ¤
(K − θk M) P(M uk )⊥ + PM uk δyk = − (K − θk M) uk (17)
with subsequent evaluation of δuk and βk as
¡ ¢−1
δuk = P(M uk )⊥ δyk , βk = − uT
k PM uk δyk (18)
In this procedure, one recognizes
(−1) £ ¤−1
A(M uk ) = P(M uk )⊥ (K − θk M) P(M uk )⊥ + PM uk (19)

as the Bott-Duffin inverse of the restricted Equation (16) [26, 40].


The system matrix of Equation (17) is well conditioned for θk converging to a simple
eigenvalue λ, in which case (K − θk M) P(M uk )⊥ and PM uk are by construction complementary
matrices. In the case of multiple eigenvalues, the system matrix of Equation (17) tends to
become ill conditioned. Nevertheless, δuk still tends cubically to zero, as the error resulting
from the ill-conditioning is mainly in the direction engendered by φ, as one obtains from
Chatelin’s proof for the Rayleigh quotient iteration [32]. In particular, if the starting value θ1
is equal (within machine precision) to a multiple eigenvalue λ, only one iteration is required
for convergence, provided that u1 is not orthogonal to the searched eigenvector φ.
One-step convergence is illustrated in Table I for the non-linear eigenproblem of Example
2, as solved according to Algorithm 2, which is a generalization of the scheme proposed in

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
NON-LINEAR COMPLEX-SYMMETRIC EIGENVALUE PROBLEMS 7

this Section. For large eigenproblems, round-off errors in the solution of the equation system
make the occurrence of multiple eigenvalues within machine precision unlikely. However, the
existence of close eigenvalues generally contributes to the decrease of the number of iterations
required in the sequential solution of an eigenvalue problem. This is illustrated both in the
modified case of Example 2 and in Example 5.
The development above was derived directly from the Rayleigh quotient iteration. However,
this is not the only way of dealing with the present linear algebra problem.

2.2.3. Solution in terms of oblique projectors. Arbenz and Hochstenbach [25] outline a
solution of δuk in the frame of a formulation that is conceptually equivalent to Equation
(16), although contextually different, as it relies on the residual
rk = (K − θk M) uk (20)
which has the property
uT
k rk = 0 (21)
for the Rayleigh quotient θk given as in Definition 4. This formulation is integrated into the
following elaboration of the results of Section 2.2.2, which makes use of the oblique projector
Puk ,(M uk )⊥ ≡ uk uT
k M introduced in Definition 3, for huk , uk i = 1.
¡ ¢T
Pre-multiplying both sides of the equation in step (1) of Algorithm 1 by I − Puk ,(M uk )⊥
one obtains the consistency equation
¡ ¢T
I − Puk ,(M uk )⊥ (K − θk M) ũk+1 = 0 (22)
Since
PM uk δuk = 0 ⇔ Puk ,(M uk )⊥ δuk = 0 (23)
and in view of Equations (20) and (21), an expression equivalent to Equation (16) is
( ¡ ¢T
I − Puk ,(M uk )⊥ (K − θk M) δuk = − (K − θk M) uk
(24)
restricted to Puk ,(M uk )⊥ δuk = 0
and δuk may be solved directly from
h¡ ¢T ¡ ¢ i
I − Puk ,(M uk )⊥ (K − θk M) I − Puk ,(M uk )⊥ + PM uk δuk = − (K − θk M) uk (25)
as an alternative to Equation (17).
Another expression alternative to Equations (17) and (25) is
( £ ¡ ¢ ¤
(K − θk M) I − Puk ,(M uk )⊥ + PM uk δyk = − (K − θk M) uk
¡ ¢T (26)
δuk = I − Puk ,(M uk )⊥ δyk
It is worth observing that the residual rk and the increment δuk are related by
δuT
k rk = 1/βk (27)
This is obtained by first pre-multiplying both sides of Equation (20) by δuT
k,
with the result
δuT r
k k = δuT
k Ku k , according to Equation (15). On the other hand, pre-multiplying both sides
T
of Equation (14) by uT k (K − θ k M) gives δu k Ku β
k k = 1 after making use of the expression
of ũk+1 in Algorithm 1 and of Equations (21) and (15). Since both rk and δuk have cubic
convergence to a zero vector as uk → φ, 1/βk converges cubically to zero with approximately
double the number of exact digits – see remark after Theorem 2. As a result, |1/βk | might be
used as a convergence threshold, instead of Equation (13), which only applies for λ 6= 0.

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
8 N. A. DUMONT

2.3. On the numerical implementation of the Jacobi-Davidson method


The vector uk+1 obtained by normalizing uk + δuk , according to Equation (14), is unique
(except for a plus/minus sign), whether δuk has been evaluated from Equations (17), (25)
or (26), which is in either case mathematically equivalent to using the Rayleigh quotient
iteration, Algorithm 1, as convergence is asymptotically cubic. The Jacobi-Davidson method,
as formulated in terms of the vector increment δuk , is preferable for two main reasons: (a)
the equation system is better conditioned; (b) the search subspace for δuk represented by
I − P(M uk )⊥ may be further restricted with no lack of mathematical rigor, thus enabling either
sequential or simultaneous evaluation of a set of eigenpairs. The designation Jacobi refers to
subspace iteration [25, 33].
The projector PM uk in Equations (17), (25) or (26) may be multiplied by any scalar γ 6= 0,
with no influence in the mathematical results. In fact, such a multiplication is advisable, as
done in step (4.1) of Algorithm 2, so the summands have approximately the same order of
magnitude and round-off errors is minimized.
Once evaluated within the required tolerance, an eigenpair might in principle be removed
from the original eigenproblem by matrix deflation [13, 20]. However, this introduces numerical
errors in the original system and is hardly applicable to non-linear eigenproblems. An
alternative is to solve the problem sequentially, for the eigenpairs of a subspace of interest.
The algorithm of Section 3.5 outlines the numerical implementation of the Jacobi-Davidson
method in terms of the orthogonal projector of Section 2.2.2, for the evaluation of the complete
set (Λ, Φ). The search of each eigenpair (λ, φ) starts with a first estimate u1 orthogonal to
all previously evaluated eigenvectors. The incremental vector δuk is then evaluated iteratively,
according to Equations (17) and (18), but using in place of PM uk an orthogonal projector
that encompasses all eigenvectors that have been already evaluated, besides uk . Algorithm 2
is intended for non-linear problems, with mass and stiffness matrices given as functions of the
eigenvalues, as detailed in Sections 3.2 and 3.3.

3. FORMULATION OF NON-LINEAR, COMPLEX-SYMMETRIC EIGENPROBLEMS


OF THE STRUCTURAL ANALYSIS

3.1. Introduction
The dynamics problem subject of this paper is formulated in the frequency domain in terms
of a generalized effective stiffness matrix Kef f (ω) given as the frequency power series
¡ ¢
Kef f (ω) = K0 −iωC1 −ω 2 M1 −iω 3 C2 −ω 4 M2 −· · ·−iω 2n−1 Cn −ω 2n Mn + O ω 2n+1 (28)
truncated after 2n + 1 terms, compactly expressed as
n
X ¡ 2j−1 ¢ ¡ ¢
Kef f (ω) = K0 − iω Cj + ω 2j Mj + O ω 2n+1 (29)
j=1

where K0 ∈ Rm×m is the stiffness matrix √ of the static case, Cj , Mj ∈ R


m×m
are generalized
damping and mass matrices, and i = −1 is the imaginary number. Thus, Kef f ∈ C m×m
in Equation (29) is a complex-symmetric matrix for a problem with m degrees of freedom.
Basis for this formulation is a frequency-dependent development of stiffness and mass matrices

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
NON-LINEAR COMPLEX-SYMMETRIC EIGENVALUE PROBLEMS 9

proposed by Przemieniecki [9] in terms of a displacement approach for the free-vibration


analysis of damping-free truss and beam structures, thus without the imaginary terms. He
obtained the real-symmetric, effective stiffness matrix Kef f (ω2 ) of an element in the shape
Kef f (ω2 ) = K0 − ω 2 M0 − ω 4 (M2 − K4 ) − ω 6 (M4 − K6 ) + O(ω 8 ) (30)
which gave rise to a dynamic finite element formulation [27, 39, 41, 42], a not quite adequate
denomination, as it was only applied to free-vibration problems. It is worth remarking that
Equation (30) is sometimes represented as [27, 39]
Kef f (ω2 ) = K0 − ω 2 (M0 − K2 ) − ω 4 (M2 − K4 ) − ω 6 (M4 − K6 ) + O(ω 8 ) (31)
2
with a coefficient matrix K2 in the term that multiplies ω , probably as a consequence
of having been obtained in the frame of a displacement formulation that does not seek
dynamic equilibrium satisfaction inside the finite element. However, this term is void in any
variationally-based finite/boundary element expansion one may obtain [6, 7, 10, 11, 12, 16],
which is coherent with the developments in the classical books on dynamics that retain terms
up to ω 2 : Kef f (ω2 ) = K0 −ω 2 M0 +O(ω 4 ) [9, 13, 14, 43, 44]. Except for M0 , which corresponds
to M1 in the present paper, the mass matrix terms given in Equations (30) and (31) differ in
physical meaning from the ones introduced in Equation (29).
The simplest illustration of this problem is the stiffness matrix of a truss element with two
degrees of freedom, as given in Example 1, which shows that the matrix terms Cj , Mj , j > 1,
when consistently obtained, contain damping, mass and stiffness contributions – in fact,
the first mass matrix, M1 , is already affected by viscous damping. In spite of the mixed
nature of Cj and Mj , they will be referred to as generalized damping and mass matrices.
Equation (88) illustrates that, if the variational formulation makes use of fundamental solutions
given as real functions of complex arguments, the effective matrix Kef f (ω) turns out to be
complex symmetric. Inherently complex formulations such as in References [45, 46] (the latter
formulation is also non-variational) do not lead directly to complex-symmetric problems.
The linear algebra properties of these matrix terms are outlined in the following
developments, which omit, for simplicity, the truncation error orders O(ω 2n+1 ) and O(ω 2n+2 )
(for complex and real problems). Observe the notation Kef f (ω) as in Equation (29) for the
complex-symmetric problem, and Kef f (ω2 ) as in Equation (30) for the real-symmetric problem.
One can always infer from context whether real or complex matrices are being dealt with.
This paper is concerned with the non-linear eigenvalue problem associated to Equation (29):
 
Xn
¡ 2j−1 ¢
Kef f (ω) φ ≡ K0 − iω Cj + ω 2j Mj  φ = 0 (32)
j=1

In order to adequately solve the complex eigenproblem of practical interest, it is advisable


to start with the outline of the underlying real problem.

3.2. Non-linear, real-symmetric eigenproblems


3.2.1. Formulation. Equation (32), particularized to damping-free problems and written
with λ in place of ω 2 , for notation simplicity,
 
n
X
Kef f (λ) φ ≡ K0 − λj M j  φ = 0 (33)
j=1

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
10 N. A. DUMONT

may be restated as the linearized, augmented set of equations [6],


     
K0 0 0 ··· 0 M1 M2 M3 ··· Mn  φ0 
  0   

 0 M2 M3 · · · Mn   M2 M3 ··· ···  

 φ1  

  .. .. 
 0 M3 M4 ··· 0 − λ  M3 . . ··· 0  φ 2 =0 (34)
 . .. .. .. ..    .. 
 . .   . .. .. .. ..  
 

 . . . .  .. . . . .  
 . 
 
0 Mn 0 ··· 0 Mn 0 0 ··· 0 φ n−1

where
φj = φ λj , j = 0, · · · , n − 1 (35)
The matrix product corresponding to the first row in Equation (34) is exactly the power series
expansion of Equation (33), with the remaining equations vanishing identically.
Equation (34) is a generalization of Duncan’s original proposition [47], as it is valid for any
number n of matrix terms and is disposed in such a way that the augmented stiffness and mass
matrices grow from left to right and from top down, while preserving symmetry. A similar
structure for complex-symmetric matrices is proposed in Section 3.3. There is an extensive
literature on matrix arrangements similar to Equation (34), called matrix pencils [20, 48]. As
outlined in the following, Equation (34) is key to the theoretical argumentation, but is not
required for an algorithm implementation.
The linearized, augmented eigenproblem of Equation (34), compactly expressed as

(Kaug − λMaug ) φaug = 0 (36)

leads to an enlarged set with m × n solution eigenpairs, which are in part complex even
in the case of a damping-free problem, as the augmented ¡stiffness¢ and mass matrices are
not positive definite. However, only m real eigensolutions λ, φaug , corresponding to real
¡ ¢
eigenpairs (λ, φ) ≡ ω 2 , φ , will be solutions of Equation (33) of practical interest in a vibration
or transient analysis. This is discussed and justified in the next Sections.
Given a non-normalized eigenvector φ̃aug , the corresponding normalized eigenvector φaug
of the augmented, linearized system is obtained as
³ T ´−1/2
φaug = φ̃aug φ̃aug Maug φ̃aug (37)

in such a way that


φT
aug Maug φaug = 1 φT
aug Kaug φaug = λ (38)
assuming that Maug works as positive definite for φaug , according to Definition 1.
From the first matrix row in Equation (34), which is actually the only one of interest, one
infers that Equation (33) may be alternatively expressed as
¡ ¢
Kef f (λ) φ ≡ K(λ) − λM(λ) φ = 0 (39)

in terms of frequency-dependent mass and stiffness matrices M(λ) and K(λ) defined as
n
X
M(λ) = jλj−1 Mj = M1 + 2λM2 + 3λ2 M3 + · · · (40)
j=1

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
NON-LINEAR COMPLEX-SYMMETRIC EIGENVALUE PROBLEMS 11

n
X
K(λ) = K0 + (j − 1)λj Mj = K0 + λ2 M2 + 2λ3 M3 + · · · (41)
j=2

According to Equation (37) for the augmented, linearized eigenproblem, one obtains for the
eigenvectors of interest φ0 ≡ φ in Equation (39) that, given a non-normalized eigenvector φ̃,
the corresponding normalized eigenvector φ is obtained as
³ T ´−1/2
φ = φ̃ φ̃ M(λ) φ̃ (42)

in such a way that


φT M(λ) φ = 1 φT K(λ) φ = λ (43)
M(λ) and K(λ) must be positive definite and semidefinite, respectively, which depends on the
physical meaningfulness of the mathematical model, as outlined in Section 3.2.2.
Given two normalized eigenvectors φr and φs ,
φT
r M(λr ,λs ) φs = δrs (44)
where M(λr ,λs ) is the generalization of the frequency-dependent mass matrix of Equation (40)
as referred to the eigenvalues λr , λs :
P
n P
j
M(λr ,λs ) = λk−1
r λj−k
s Mj
j=1 k=1 (45)
¡ ¢
≡ M1 + (λr + λs ) M2 + λ2r + λr λs + λ2s M3 + · · ·
If all eigenpair solutions are gathered in the pair of matrices (Λ, Φ), where Λ is diagonal and
Φ is the set of eigenvectors normalized according to Equation (42), one obtains for Equations
(43) and (44), as a generalization of Equation (3) for the present non-linear problem [10]:
X j
n X j−1
n X
X
Λk−1 ΦT Mj ΦΛj−k = I, ΦT K0 Φ + Λk ΦT Mj ΦΛj−k = Λ (46)
j=1 k=1 j=2 k=1

3.2.2. Consistency of the non-linear, real-symmetric formulation. Equation (39) is


equivalent to the original Equation (33), only rearranged in a convenient way in terms of
eigenvalue-dependent stiffness and mass matrices. Equation (34) – for an augmented, linearized
problem – was laid out as an intermediary step between Equations (33) and (39). All three
equations lead to the same set of m × n eigensolutions, from which only a subset of m real
eigensolutions are of interest in a practical application. The orthogonality statement hφr , φs i
arrived at in Equation (44) is consistent with both Definition 1 and the generalized inner
product definition of Reference [17] for non-linear problems.
There are some key issues to be discussed about the present outline. The first one is
formulated as the following Theorem.
Theorem 3 (Equivalence of Equations (34) and (39)) Equations (34) and (39) have
the same set of m × n solution pairs (λ, φ0 ) ≡ (λ, φ).
Proof: Equation (39), a non-linear system of degree n in λ and with coefficient matrices
of order m, has exactly m × n¡ generally¢ complex solution eigenpairs (λ, φ), since the
characteristic polynomial det Kef f (λ) has m × n roots λ. Every solution of Equation (39)
satisfies Equation (34) identically, and this is a linear eigensystem in λ of order m × n.

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
12 N. A. DUMONT

The following Theorem, adapted from Observation 7.1.3 by Horn and Johnson [20], is of
interest for the forthcoming developments.
Theorem 4 (Positivity of mass and stiffness matrices) M(λ) in Equation (40) is
positive definite for all nonnegative λ if and only if all matrices Mj , j = 1, · · · , n are positive
definite. Moreover, given all Mj , j = 1, · · · , n as positive definite, K(λ) in Equation (41) is
positive definite or semidefinite if and only if K0 is positive definite or semidefinite.
A last concern are number and properties of the real solution eigenpairs one is able to obtain
in any of the equivalent Equations (34) or (39).

3.2.3. Eigenvalue properties of the non-linear, real-symmetric formulation. The conclusions


of interest might be drawn from the minmax characterization of eigenvalues laid out by Hadeler
for non-linear eigenproblems [17, 18, 19]. The following independent development addresses
the non-linear eigenproblem as formulated in Equations (39-41), for n > 1. One starts by
proving the following Theorem, in which K ← M{k} defines a matrix whose kth column is the
k
column M{k} of M and whose remaining columns coincide with those of K [20].
Theorem 5 (Determinant derivatives and positive definite matrices) Let K, M ∈
C m×m be Hermitian matrices with constant elements, and suppose that K is positive
semidefinite and M is positive definite. Then,
Xm µ ¶
det K ← M{k} > 0 (47)
k
k=1

Proof: K + Mx is positive semidefinite for all x ≥ 0, according to Theorem 4. Also,


K + Mx + M∆x is positive definite for all x ≥ 0 and a positive increment ∆x. Since M∆x =
(K + Mx + M∆x) – (K + Mx) is positive definite, it follows from Corollary 7.7.4(b) by
Horn and Johnson [20] that

det (K + Mx + M∆x) > det (K + Mx) (48)

As a result, the following determinant derivative [48] is positive for all x ≥ 0:


Xm µ ¶
d
det (K + Mx) = det K + Mx ← M{k}
dx k
k=1

det (K + Mx + M∆x) − det (K + Mx)


= lim >0 (49)
∆x→0 ∆x
which proves the theorem when evaluated for x = 0.

The developments above apply to a negative matrix K, as well, by just replacing


m
K with −K. Sinceµ det (−K)¶ = (−1) det (K), Equation (47) of Theorem 5 reads
Pm
(−1)m−1 k=1 det K ← M{k} > 0 in the case of a negative K (as there are m − 1 columns
k
of K in the summand). The proof given in Reference [20] for the item (b) of Corollary 7.7.4
is also meant to apply to the case of the difference of two positive definite matrices A and B
being positive semidefinite, leading to det (A) ≥ det (B). However, this apparently involves a

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
NON-LINEAR COMPLEX-SYMMETRIC EIGENVALUE PROBLEMS 13

mistake and the corollary is actually only applicable, except for the trivial case A ≡ B, in the
case of a positive definite matrix difference, as needed for arriving at Equation (48).
The eigenvalues λ of Equation (33) are the roots of
¡ ¢
det K(λ) − λM(λ) = 0 (50)
which has m × n complex solutions. However, if one is only interested in the subset of the real
eigenvalue solutions, Equation (50) is more conveniently stated as the restricted linear system
½ ¡ ¢
det K(θ) − λM(θ) = 0 for all θ ≥ 0
(51)
such that λ = θ
On the other hand, the set of isocurves
¡ ¢ ¡ ¢
det Kef f (θ,λ) ≡ det K(θ) − λM(θ) = C (52)
(with the simplifying notation Kef f (θ,λ) used only in the following outline) may be stated in
the Cartesian system (θ, λ) as
∂ ¡ ¢ ∂ ¡ ¢ dλ
det Kef f (θ,λ) + det Kef f (θ,λ) =0 (53)
∂θ ∂λ dθ
The determinant derivatives are conveniently expressed as
m µ ¶
∂ ¡ ¢ X ∂
det Kef f (θ,λ) = det Kef f (θ,λ) ← Kef f (θ,λ){k} (54)
∂θ k ∂θ
k=1
m µ ¶
∂ ¡ ¢ X ∂
det Kef f (θ,λ) = det Kef f (θ,λ) ← Kef f (θ,λ){k} (55)
∂λ k ∂λ
k=1

according to the notation introduced in Equation (47). The matrix derivatives needed in
Equations (54) and (55) are obtained from Equations (40) and (41):
Xn

Kef f (θ,λ) = (θ − λ) j (j − 1) θj−2 Mj = (θ − λ) (2M2 + 6θM3 + · · · ) (56)
∂θ j=2


Kef f (θ,λ) = −M(θ) (57)
∂λ
Then,
 
m
X Xn
∂ ¡ ¢
det Kef f (θ,λ) = (θ − λ) det Kef f (θ,λ) ← j (j − 1) θj−2 Mj{k}  (58)
∂θ k
j=2
k=1

m
X µ ¶
∂ ¡ ¢
det Kef f (θ,λ) = − det Kef f (θ,λ) ← M(θ){k} (59)
∂λ k
k=1

Suppose Kef f (θ,λ) is positive semidefinite, that is, C ≥ 0 in Equation (52). Then, for θ ≥ 0
the sums given in Equations (58) and (59) are positive, according to Theorem 5, and it follows
from Equation (53) that, necessarily,
µ ¶

signum = signum (θ − λ) (60)

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
14 N. A. DUMONT

In the case of negative definite Kef f (θ,λ) , that is, C < 0 in Equation (52), Equation (60) still
holds true, according to the remark made after the proof of Theorem 5.
According to Equation (60), the eigensolutions λj (θ), j = 1, · · · , m of Equation (51),
interceptions with the straight line λ = θ, are determined as unique minimum values for
each isocurve λj (θ).
The developments above are illustrated in Figure 3 for Example 2, with a set of m = 12
curves, showing the interception points with the line λ = θ as the eigenvalue solutions of
the non-linear problem of Equation (50). The interceptions of the curves λj (θ) with the axis
θ = 0 are the solutions of the linear problem given in Equation (51). As θ increases, each λj (θ)
decreases continually until a minimum value, and then increases indefinitely again, according to
Equation (60). Multiple eigenvalues cannot be perceived
¡ in a 2D
¢ plot as in Figure 3. Only a 3D
analysis, with a third axis corresponding to det K(θ) − λM(θ) , would enable the visualization
of the effect of multiple eigenvalues, for the surface given by Equation (52) cut by the plane
(θ, λ, 0). All curves λj (θ), j = 1, · · · , m, converge to one single value, as, from Equation (50)
together with Equations (40) and (41),
λj n−1
lim = for all j (61)
θ→∞ θ n
The issues of intercepting or coinciding curves λj (θ) do not deserve a closer examination, as the
implemented iterative procedure follows a path given by the eigenpair (λ, φ) that is uniquely
determined even in the case of multiple eigenvalues. It is worth observing in Figure 3 the case
of curves intercepting in the range between θ = 0 and λ = θ, for eigenvalues λ10 ≡ λ11 and λ12 ,
more clearly seen by comparing the columns for n = 1 and n = 3 in Table I. Bifurcations cannot
occur in the present context. However, it is possible that two curves intercept at θ = 0, for K0
and M1 corresponding to a topological symmetry of the mathematical model, if, owing to some
inappropriateness or numerical error in their evaluation, one or more of the remaining matrices
Mj in Equations (40) and (41), while still positive definite, fail to reflect the symmetry. This
is illustrated in the modified case of Example 2.
The relevant conclusions of these developments are summarized in the following theorem.
Theorem 6 (Consistency of the non-linear symmetric eigenproblem) Given the non-
linear, real-symmetric eigenvalue problem of Equation (50), where M(λ) and K(λ) are defined
as in Equations (40) and (41), for K0 positive semidefinite and Mj , j = 1, · · · , n positive
definite, (a) there are m real, non-negative eigenvalues and n × m − m complex eigenvalues;
(b) the non-zero real eigenvalues of the non-linear problem are smaller than the corresponding
eigenvalues of the underlying linear problem.

Proof: The proof has already been given above. However, it is worth recalling that the
first of Equation (51) admits exactly m sets λj (θ) of curves (including repeated eigenvalues),
each one starting from λj (0) and decreasing until the global minimum value λj (θ) = θj is
reached, which coincides with the restriction given by the second of Equation (51), and
increasing indefinitely again, according to Equation (61). The trivial linear case n = 1 is
included in the theorem.

It may happen that some matrix Mj – whose elements usually decrease in magnitude as j
increases – are actually non-positive, as a consequence of some conceptual inappropriateness
[see the remark after Equation (31)] or of round-off errors. Then, it will be impossible to

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
NON-LINEAR COMPLEX-SYMMETRIC EIGENVALUE PROBLEMS 15

evaluate m eigenpairs as real solutions. In practice, solutions related to higher eigenvalues will
be missing, in such a case, as the relative contribution of the terms affected by the matrices Mj
increases with higher subscripts j. Then, although accuracy is expected to generally increase
with the number of power series terms, the quality of the evaluated eigenpairs is compromised
if inaccurate higher-order matrix terms are included.

3.3. Non-linear, complex-symmetric eigenproblems


3.3.1. Formulation. Although apparently repeated, the following developments are more
than a mere extension of the ones of Section 3.2. Equation (32) may also be expressed as a
linearized, augmented set of equations, for φj = φω j , j = 0, · · · , n − 1 [10],
    
K0 0 0 0 ··· 0 iC1 M1 iC2 M2 · · · Mn  φ0 
 0 M1 iC2 M2 · · · Mn   M1 iC2 M2 iC3 ··· 0  
 

   

 φ1 

 0 iC2 M2 iC3 ··· 0   iC2 M2 iC3 · · · ··· 0   
   
 φ2 
 .  − ω . .. 
 0 M2 iC3 . . ··· 0    M2 iC3 .. . ··· 0   φ3  = 0 (62)
  
 . 
 . .. .. .. .. .   . .. .. .. .. . 

 .. 



 .. . . . . ..   .. . . . . ..  

 
0 Mn 0 0 ··· 0 Mn 0 0 0 ··· 0 φn−1

This augmented complex-symmetric eigenproblem leads to an enlarged set of eigensolutions


with 2 × n × m mathematically possible solutions. However, the subvector solutions φ0 ≡
φ of immediate interest correspond to a basic set of m eigenpairs ¡(ω, φ). ¢Section 3.3.2
extends a generic solution (ω, φ) to its complex-associated eigenpair −ω, iφ , in the case
of underdamping. The case of overdamping is briefly assessed in Section 3.3.3.
From the first matrix row in Equation (62), which is the only one of interest, one infers that
Equation (32) may be alternatively expressed as
¡ ¢
Kef f (ω) φ ≡ K(ω) − ωM(ω) φ = 0 (63)

in terms of frequency-dependent mass and stiffness matrices M(ω) and K(ω) :


n ¡
P ¢
M(ω) = i(2j − 1)ω 2j−2 Cj + 2jω 2j−1 Mj
j=1 (64)
≡ iC1 + 2ωM1 + 3iω 2 C2 + 4ω 3 M2 + · · ·

n ¡
P ¢
K(ω) = K0 + i(2j − 2)ω 2j−1 Cj + (2j − 1)ω 2j Mj
j=1 (65)
≡ K0 + ω 2 M1 + 2iω 3 C2 + 3ω 4 M2 + · · ·

Observe that, while Kef f (λ≡ω2 ) in Equation (39) is a particular case of Kef f (ω) in Equation
(63), the same cannot be said about M(λ≡ω2 ) and K(λ≡ω2 ) in Equations (40) and (41) as
compared with M(ω) and K(ω) in Equations (64) and (65).
According to Equation (62) for the augmented, linearized eigenproblem, one normalizes an
eigenvector φ ≡ φ0 of the non-linear problem, Equation (63), by proceeding as in Equation
(42). However, this is only feasible if φT M(ω) φ 6= 0. For an exact solution eigenpair (ω, φ),
such a condition is a premise that the proposed eigenproblem has a physical meaning.

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
16 N. A. DUMONT

Given two normalized eigenvectors φr and φs , the orthogonality expression of Equation (44)
is valid for eigenvalues ωr , ωs and the frequency-dependent mass matrix of Equation (64):
µ 2j−2 ¶
Pn P k 2j−k−2 P k 2j−k−1
2j−1
M(ωr ,ωs ) = i ωr ωs Cj + ωr ωs Mj
j=1 (66)
k=0 ¡ ¢
k=0 ¡ ¢
= iC1 + (ωr +ωs ) M1 + i ωr2 +ωr ωs +ωs2 C2 + ωr3 +ωr2 ωs +ωr ωs2 +ωs3 M2 + · · ·

Matrix expressions similar to Equation (46) may be obtained for the complex case by
gathering all eigenpair solutions in the pair of matrices (Ω, Φ) [10]:
µ 2j ¶
P
n P k−2 T 2j−k P2j
k−1 T 2j−k
i Ω Φ Cj ΦΩ + Ω Φ Mj ΦΩ =I
j=1 k=2 k=1
µ ¶ (67)
Pn P k−1 T
2j−1 P k T
2j−1
ΦT K0 Φ + i Ω Φ Cj ΦΩ2j−k + Ω Φ Mj ΦΩ2j−k = Ω
j=1 k=2 k=1

3.3.2. Complex-associated eigenpairs in the general case of underdamping. The following


theorem states that to every complex eigenpair (ω, φ) corresponds an eigenpair (−ω, iφ), as
required in the modal analysis, which actually deals with real transient problems [10]. This
theorem, although generally valid, is of practical relevance only in the case of underdamping.
Theorem 7 (Complex-associated solutions) If the complex eigenpair (ω, φ) is a solution
of Equation (32), normalized as in Equation (43) for M(ω) and K(ω) given in Equations (64)
and (65), then the eigenpair (−ω, iφ) is also a normalized solution. Moreover, φr ≡ φ and
φs ≡ iφ are orthogonal eigenvectors in the sense of the inner product of Equation (44) for the
generalized mass matrix M(ωr ,ωs ) of Equation (66).

Proof: Only for the scope of this proof, write ω = r (cos θ + i sin θ), where r = |ω| is the
modulus and θ is the amplitude of the complex number ω. According to Moivre’s theorem,
ω j = rj (cos jθ + i sin jθ). Moreover, ω = r (cos θ − i sin θ) and
j
(−ω) = (−r)j (cos jθ − i sin jθ). Then, from the definitions of M(ω) and K(ω) in Equations
(64) and (65), ωM(ω) ≡ −ωM(−ω) , K(ω) ≡ K(−ω) . As a result,
¡ ¢ ¡ ¢ ¡ ¢
K(ω) − ωM(ω) φ = 0 ⇔ K(ω) − ωM(ω) φ = 0 ⇔ K(−ω) − ωM(−ω) iφ = 0 (68)

Similarly,
T
φT M(ω) φ = 1 ⇔ iφ M(−ω) iφ = 1 (69)
T
φT K(ω) φ = ω ⇔ iφ K(−ω) iφ = −ω (70)
Also, it follows from the expression of M(ωr ,ωs ) in Equation (66) that
T
φT
r M(ωr ,ωs ) φs = δrs ⇔ iφr M(−ωr ,−ωs ) iφs = δrs (71)

Finally, since (ω, φ) and (−ω, iφ) are solution eigenpairs of Equation (63), there must be by
definition corresponding augmented, orthogonal eigenpairs of the augmented, linearized
eigenproblem of Equation (62). This proves that φT M(ω,−ω) iφ = 0.

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
NON-LINEAR COMPLEX-SYMMETRIC EIGENVALUE PROBLEMS 17

3.3.3. Overdamping. The concepts of underdamping, critical damping and overdamping in


a transient analysis are briefly introduced by Meirovitch [49]. The subject is dealt with in more
detail by Kolousek [43] and Warburton [44] in the frame of the dynamic analysis of engineering
structures. In mathematics, the term ”overdamping” is used in a different context [50,¡ 19,¢ 51].
An undamped problem of structural dynamics has m real eigenpair solutions ω 2 , φ , as
outlined in Section 3.2, where the positive square root ω is the eigenfrequency of interest.
As damping
¡ ¢ takes place, the solution evolves into m sets of complex eigeinpairs (ω, φ)
and −ω, iφ , according to Theorem 7, with decreasing contribution of the real part and
increasing contribution of the (negative) imaginary part of ω. For a first critical damping
value, the associated complex eigenpairs corresponding to the smallest eigenvalues coalesce
into a single one, when ω becomes negative imaginary and φ may be made real. By increasing
damping the eigensolution branches again, with one of the imaginary eigenvalues tending
to zero and the other one tending to minus infinity. The eigenvectors remain real, but
inassociated. (The denomination inassociated – in the case of overdamping – is used as opposed
to the denomination complex-associated introduced for underdamping in Section 3.3.3.) The
eigenfrequency evolution described above may be followed analytically for the simple truss
element of Example 1 by solving Equation (87) for ω and then plotting the solutions for
increasing values of ζ, given a sequence of wave numbers k = (2j − 1)π/2/`, j = 1, 2, · · · . As
damping increases, the described overdamping effect gradually affects all the eigenfrequencies.
Assume that an eigenvalue solution of Equation (63) is negative imaginary. Then, one may
replace ω with −iλ in Equations (63), (64) and (65), where λ is real positive, obtaining
³ ´
K̃(λ) − λM̃(λ) φ = 0 (72)

with a new definition of real-symmetric, generalized mass and stiffness matrices:


P
n ¡ ¢
M̃(λ) = (−1)j −(2j − 1)λ2j−2 Cj + 2jλ2j−1 Mj
j=1 (73)
≡ C1 − 2λM1 − 3λ2 C2 + 4λ3 M2 + · · ·
P
n ¡ ¢
K̃(λ) = K0 + (−1)j −(2j − 2)λ2j−1 Cj + (2j − 1)λ2j Mj
j=1 (74)
≡ K0 − λ2 M1 − 2λ3 C2 + 3λ4 M2 + · · ·

The assumption λ ≥ 0 is consistent only if M̃(λ) is positive definite and K̃(λ) is positive
semidefinite, in which case φ can be made real. Not coincidentally, the eigenfrequency evolution
described in the second paragraph of this Section resembles, for ω substituted with −iλ,
the reverse of the bifurcation process that takes place in a structural stability problem [52].
However, no matter how interesting these developments may look like from the mathematical
point of view, their applicability in terms of algorithm implementation according to Section 3.2
remains to be demonstrated. In fact, convergence is unlikely to occur in the present framework,
as an eigenvalue corresponding to overdamping may be of larger absolute value than some other
eigenvalues corresponding to underdamped behavior. This is illustrated in Example 4.
On the other hand, the algorithm implemented for non-linear, complex-symmetric
eigenproblems, as outlined in Section 3.5, has proved suited for the automatic evaluation of a
whole set of m eigenpairs (ω, φ) unregarded underdamping, critical damping or overdamping
effects. As illustrated in Example 4 (for a problem with multiple eigenvalues and multiple

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
18 N. A. DUMONT

overdamping), the remaining set of inassociated eigenpairs can be obtained in a second


computer run by entering a new set of estimates. Although this last step may involve some trial
and error, problems with overdamping are quite infrequent and, when existent, usually affect
only a few eigenpairs. In the present theoretical frame, the unlikely case of critical damping,
when two eigenpairs coalesce into a single one (within machine precision), can only be inferred
from the impossibility of finding a complementary (inassociated) solution.

3.4. Perturbation analysis


As given in Section 3.2, among all n × m solution eigenpairs (λ ≡ ω 2 , φ) of the non-linear,
real-symmetric problem, only the subset of m real eigenpairs is of interest in an engineering
application, as there are m degrees of freedom in the discrete mathematical model. In the case
of a complex formulation, on the other hand, since all 2 × n × m solutions (ω, φ) are complex,
a perturbation analysis of the underlying real formulation is required in order to tell which
subset of 2 × m complex eigenpairs is of actual interest, as their mathematical patterns cannot
differ too much from those of a corresponding real eigenproblem.
Although established for non-linear problems, both Equations (39) and (63) can be rewritten,
for use in an iterative procedure, as the stepwise linear eigenproblem
¡ ¢
K(θk ) − λk M(θk ) φk = 0 (75)
where θk is the eigenvalue estimate for the kth step and the eigenpair (λk , φk ) is the
corresponding exact solution. The matrices M(θk ) and K(θk ) do not depend on λk , as they
are evaluated at each step for either λ ≡ θk according to Equations (40) and (41), for the
real problem, or ω ≡ θk according to Equations (64) and (65), for the complex problem. As a
result, all definitions and theorems of Section 2 for the solution of the linear eigenproblem of
Equation (1) apply directly, if one just replaces λ, φ, M and K with λk , φk , M(θk ) and K(θk ) .
It must be shown that the solution of Equation (75) corresponds to one step toward the
actual solution of the non-linear problem given by either Equation (39) or (63):
¡ ¢
K(λ) − λM(λ) φ = 0 (76)
where the eigenpair (λ, φ) is the exact solution one is looking for. The scope of a perturbation
analysis is to investigate in what amount changes K(λ) −K(θk ) and M(λ) −M(θk ) in the stiffness
and mass matrices, as applied to the present problem, affect the eigenvalue solution [20].
Given the mass and stiffness definitions of Equations (40) and (41), for the real-symmetric
problem, or of Equations (64) and (65), for the complex-symmetric problem, one may write
from Equations (75) and (76) the perturbation equation
2
K(λ) − λM(λ) = K(θk ) − λM(θk ) − (λ − θk ) δM(θk ) (77)
where, for real-symmetric and complex-symmetric problems, respectively,
P
n−1 P
j
δM(θk ) = Mj+1 lλj−l θkl−1
j=1 (78)
l=1 ¡ ¢
≡ M2 + (2θk + λ) M3 + 3θk2 + 2θk λ + λ2 M4 + · · ·
µ 2j−2 ¶
Pn P 2j−l−2 l−1 P 2j−l−1 l−1
2j−1
δM(θk ) = i lλ θk Cj + lλ θk M j
j=1 (79)
l=1 ¡ 2 l=1 ¢
≡ M1 + i (2θk + λ) C2 + 3θk + 2θk λ + λ2 M2 + · · ·

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
NON-LINEAR COMPLEX-SYMMETRIC EIGENVALUE PROBLEMS 19

¡ ¢
Next, one investigates the product ΦTk K(λ) − λM(λ) Φk , where K(λ) −λM(λ) from Equation
(76) is singular, and Φk is the m×m non-singular eigenvector matrix of the linear eigenproblem
of Equation (75), normalized according to Equation (3) and with corresponding diagonal
eigenvalue matrix Λk . One obtains from Equation (77):
¡ ¢ 2 T
ΦT
k K(λ) − λM(λ) Φk = Λk − λI − (λ − θk ) Φk δM(θk ) Φk (80)
If Λk − λI is singular, then λ = θk is the solution of the non-linear problem of Equation (76)
and convergence has been achieved. If Λk −λI is non-singular, one may multiply the right-hand
−1
side of Equation (80) by (Λk − λI) , obtaining
³ ´
−1 2
(Λk − λI) Λk − λI − (λ − θk ) ΦT k δM(θk ) Φk
2 −1
(81)
= I − (λ − θk ) (Λk − λI) ΦT k δM(θk ) Φk

Since this is a singular matrix,


2 −1
1 ≤ (λ − θk ) k (Λk − λI) ΦT
k δM(θk ) Φk k (82)
for a matrix norm k · k (Corollary 5.6.16 by Horn and Johnson [20]). Taking k · k such that
−1 −1
k (Λk − λI) k = maximum diagonal value of (Λk − λI) , it results that
2
1 ≤ (λ − θk ) |λk − λ|−1 kΦT
k k kΦk k kδM(θk ) k (83)
or
2
|λk − λ| ≤ (λ − θk ) kΦT
k k kΦk k kδM(θk ) k (84)
where λk is the closest eigenvalue to the actual solution λ of the non-linear problem of Equation
(76) one is attempting to solve. Equation (84) may also be written as
2
|λk − λ| ≤ (λ − θk ) κ (Φk ) kδM(θk ) k (85)
where κ (Φk ) is the condition number of the eigenproblem of Equation (75), referred to matrices
K(θk ) and M(θk ) , with respect to the matrix norm k · k.
This perturbation analysis is mainly based on Horn and Johnson [20]. A closer investigation
of the subject [22, 53, 54, 55] is beyond the scope of the present paper. However, one observes
that, for real-symmetric problems, all matrices are normal; moreover, the difference matrix
δM(θk ) , as given in Equation (78), is positive definite and has elements of much smaller
magnitude than M(θk ) in Equation (40). In the case of a complex-symmetric problem, on
the other hand, not only are the matrices non-normal but also δM(θk ) , as given in Equation
(79), has elements of magnitude comparable to M(θk ) in Equation (64).
The main conclusion from Equation (85) is summarized in the following Theorem.
Theorem 8 (Perturbation Theorem) Let λ be a solution of the non-linear, complex-
symmetric eigenproblem outlined in Equation (75) and let θk be an approximation. ¡ If K(λ) ¢
and M(λ) are perturbed by an error O (|λ − θk |), then λ is perturbed by an error O |λ − θk |2 .
Observe that, for¡the¢ non-linear eigenproblem dealt with in this Section, Theorem 1 is restated
as |λk − θk | = O δk2 . In Equation (85), θk is the Rayleigh quotient corresponding to the kth
step of an iterative procedure, λk is the exact eigenvalue solution of the perturbed problem with
matrices K(θk ) and M(θk ) – only required for the theoretical developments but not actually
evaluated – and λ is the actual eigenvalue of the non-linear problem one is attempting to solve.

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
20 N. A. DUMONT

3.5. Jacobi-Davidson algorithm for non-linear, complex-symmetric eigenproblems


The algorithm outlined in Section 2.2 requires the input of the matrices and control variables:
m := number of degrees of freedom (matrices order)
n := number of summands in either Equation (32) or (33)
K0 , Cj , Mj , j = 1, · · · , n
number eig := number of eigenpairs of interest (as implemented, number eig = m)
max it := maximum number of iterations for each eigenpair
tol := convergence threshold
γ := scaling factor to prevent round-off errors (used in step (4.1))
Output is the eigenpair (Λ, Φ), where, for r = 1, · · · , number eig,
Λ is a vector of the eigenvalues λr and
Φ is a matrix whose columns are the eigenvectors φr
Following auxiliary matrices are needed in the evaluation of the rth eigenvalue problem:
Kef f (λr ) := the effective stiffness matrix
bs := a matrix with s = 1, · · · , r base vectors
M(λr ,λs ) := actually a row matrix, for s = 1, · · · , r
rk and δuk := residual and incremental vectors, which may share the same allocation
For simplicity of writing the algorithm below, one uses λr and φr instead of θk and uk ,
which denote the current eigenvalue and eigenvector estimates of the developments in Sections
2 and 3. The non-linear eigenproblem that can be either real symmetric or complex symmetric,
with λ corresponding to either ω 2 or ω, although two separate subroutines are required in the
FORTRAN code, for the sake of adequately handling real and complex variables. A linear
eigenproblem may be solved as a particular case, with no loss of computational efficiency.
Step (1) produces an eigenvector estimate φr that is orthogonal to the previously evaluated
eigenvectors φs , s = 1, · · · , r − 1, in terms of the orthogonal projector P(M φ)⊥ of Equation (7),
and at the same time creates a matrix with s = 1, · · · , r orthogonal base vectors bs that will be
used in the Bott-Duffin solution for δuk in Step (4), according to Section 2.2.2. The matrices
Kef f (λr ) and M(λr ,λs ) are referred to in Sections 2 and 3 as either Equations (39) and (45),
for the real eigenproblem, or Equations (63) and (66), for the complex eigenproblem. When
searching for an inassociated eigenpair, in the case of overdamping, the eigenvalue estimate
must be different from the one given in the algorithm, as illustrated in Example 4. Moreover,
although not indicated, it is checked at every step whether φT r Mθk φr is positive, for a real
eigenproblem, or non-null, for a complex eigenproblem.
Algorithm 2 (Jacobi-Davidson method) For r = 1, · · · , number eig:
If this is a real-symmetric eigenproblem:
T
Input the eigenvector estimate φr := h1 1 1 · · · i
Input the eigenvalue estimate: If r = 1, then λr := 0, else λr := λr−1
If this is a complex-symmetric eigenproblem:
Eigenvector estimate φr := eigenvector of the underlying real-symmetric problem.
Eigenvalue estimate λr := eigenfrequency of the underlying real-symmetric problem.
Input conv := 1
For k = 1, · · · , max it while conv > tol:
(1) Orthogonalize φr with respect to φs , s = 1, · · · , r − 1:
(1.1) For s = 1, · · · , r
(1.1.1) Define bs := M(λr ,λs ) φs

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
NON-LINEAR COMPLEX-SYMMETRIC EIGENVALUE PROBLEMS 21

¡ ¢
(1.1.2) For l = 1, · · · , s − 1 project bs := bs − bl bT bs
p l
(1.1.3) Normalize bs := bs / bT s bs ¡ ¢
(1.2) For s = 1, · · · , r − 1 project φr := φr − bs bT s φr
(2) Evaluate the effective stiffness matrix Kef f (λr ) := K(λr ) − λr M(λr )
(3) Evaluate the residual vector rk := Kef f (λr ) φr
(4) Solve for δuk in Kef f (λr ) δuk = −r such that PM(λs ) φs δuk = 0, s = 1, · · · , r:
(4.1) For s = 1, · · · , r ¡ ¢
modify Kef f (λr ) := Kef f (λr ) + γbs − Kef f (λr ) bs bT s
(4.2) Solve Kef f (λr ) δuk = −rk ¡ ¢
(4.3) For s = 1, · · · , r project δuk := δuk − bs bT s δuk
(5) Update φr := φr + δuk ³ ´ ³ ´
(6) Evaluate the Rayleigh quotient θ := φT T
r K(λr ) φr / φr M(λr ) φr
(7) Evaluate convergenceq conv := |θ − λr |/θ
(8) Normalize φr := φr / φT
r M (λr ) φr
(9) Update λr := θ

4. EXAMPLES

4.1. Example 1 – Frequency-domain effective stiffness matrix for a truss element


4.1.1. Problem formulation. To illustrate the central developments of this paper, one presents
the simplest academic example conceivable, consisting of a truss element of constant cross-
section A, length `, elasticity modulus E, specific mass density ρ and viscous damping µ = 2ζρ,
submitted to harmonic vibration (Figure 1). The governing differential equation is:
∂ 2 u∗ (x)
+ k 2 u∗ (x) = 0 (86)
∂x2
where
ρ 2
k2 =
(ω + 2iζω) (87)
E
One obtains the effective symmetric stiffness matrix for the truss element of Figure 1, in the

d1, p1 d2 , p 2 G1 W G2
a) b)
h1 h2
x
Figure 1. (a) Coordinate system for the stiffness matrix of a truss element; (b) definition of domain
Ω, boundaries Γ1 and Γ2 and corresponding cosine directors η1 and η2 .

frame of a hybrid finite element formulation [16], as


· ¸
kEA cos k` −1
Kef f (ω) = (88)
sin k` −1 cos k`
In the case of a damping-free structure, this is equivalent to
Kef f (ω) = K(ω) − ω 2 M(ω) (89)

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
22 N. A. DUMONT

where K(ω) and M(ω) (here differently defined than in Section 3) are the split frequency-
dependent stiffness and mass matrices obtained by Przemieniecki [9] in the frame of the
displacement finite element formulation:
· ¸
kEA k` csc k` + cos k` −1 − k` cot k`
K(ω) = (90)
2 sin k` −1 − k` cot k` k` csc k` + cos k`
· ¸
kEA k` csc k` − cos k` 1 − k` cot k`
M(ω) = (91)
2ω 2 sin k` 1 − k` cot k` k` csc k` − cos k`
These analytical expressions, as derived by Przemieniecki, can by no straightforward means
be adapted for viscous damping.
The more compact expression of K as an effective stiffness matrix, Equation (88), is not
only simpler but also easier and more convenient to arrive at in the general frame of a hybrid
finite element formulation including viscous damping. The expression of Kef f (ω) in terms of
transcendental functions of ω is only possible when the finite element boundaries coalesce to
points, as for trusses and beams. The general evaluation of mass and stiffness matrices as
series expansions, according to Section 3.1 and as illustrated in the following, is conceptually
straightforward and applicable to large finite/boundary element families [6, 12, 16].

4.1.2. Series expansions. The frequency power series expansion of the effective stiffness
matrix of Equation (88) is, for a damping-free problem,
" #  2   4 
` `2 ` 7`4
EA  1 −1 ω2  3 6 ω 4
Kef f (ω) = − 2 −  45 360  + O(ω 6 ) (92)
` −1 1 c `2 `2 c4 7`4 `4
6 3 360 45
p
where c = E/ρ is the wave propagation velocity through the elastic medium. The
corresponding frequency-dependent mass and stiffness matrices, as developed in Equations
(40) and (41), coincide with the expansions of Przemieniecki’s Equations (90) and (91):
  2   4 
` `2 2 2` 7`4
EA  1  3 6
+ ω  45 180
 + O(ω 6 )
M(ω) = (93)
` c2 `2 `2 c4 7`4 2`4
6 3 180 45
" #  
`4 7`4
EA  1 −1 4
ω  45 360
K(ω) = +  + O(ω 6 ) (94)
` −1 1 c4 7`4 `4
360 45

In the case of viscous damping, one obtains the frequency power series expansion of the
effective stiffness matrix of Equation (88):
" # " 2 1 #  
15−4α2 15−7α2
1 −1 3 3 2 2 45 90
Kef f (ω) = EA  − iω`α − ωc2`  
` c 1 2 15−7α2 15−4α2
−1 1 3 3 90 45
 2 2
  4 2 4 2
 (95)
4(21−4α ) 147−31α 16α −120α +105 127α −930α +735
3 3 945 1890 4725 37800
− iω c`3 α   − ω44`4  +O(ω 5 )
4(21−4α2 ) c 127α4 −930α2 +735 16α4 −120α2 +105
147−31α2
1890 945 37800 4725

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
NON-LINEAR COMPLEX-SYMMETRIC EIGENVALUE PROBLEMS 23

where ζ is replaced with αc/`, in terms of a nondimensional viscosity parameter α, to


simplify notation. The corresponding expansions of the frequency-dependent mass and stiffness
matrices, as introduced in Equations (64) and (65), are
 " #    
2 1 15−4α2 15−7α2 4(21−4α2 ) 147−31α2
3 3 45 90
M(ω) = EA i`α
2
+ 2ω`2   + iω `3 α  315
2 3 630

` c 1 2 c 15−7α2 15−4α2 c 4(21−4α2 )
147−31α2
3 3 90 45 630 315
  (96)
64α4 −480α2 +420 127α4 −930α2 +735
3 4 4725 9450
+ ωc4`   + O(ω 5 )
127α4 −930α2 +735 64α4 −480α2 +420
9450 4725
" #    
15−4α2 15−7α2 8(21−4α2 ) 147−31α2
1 −1 45 90 945 945
EA  ω 2 `2  + iω 3 `3 α  
K(ω) = ` + c2 c3
−1 1 15−7α2 15−4α2 147−31α2 8(21−4α2 )
90 45 945 945
 4 2
 (97)
16α −120α +105 127α4 −930α2 +735
4 4 1575 12600
+ ωc4`   + O(ω 5 )
127α4 −930α2 +735 16α4 −120α2 +105
12600 1575

All matrix terms Mj in the expansions for the undamped problem are positive definite. In
the case of damping, however, only C1 is unconditionally positive definite – see Example 3.

4.2. Example 2 – Non-linear eigenproblem for a two-dimensional transient heat conduction


in a homogeneous square plate
Figure 2 represents a square domain for a homogeneous heat conduction problem with the
indicated boundary conditions [56]. Isotropic thermal conductivity and specific heat are
assumed as unity. The plate is discretized with 2 × 2 quadratic finite elements. A transient
analysis of this problem for homogeneous initial temperature condition in terms of advanced
mode superposition is given in Reference [12]. At present, one is only concerned with the non-
linear eigenvalue problem, that is, the solution of Equation (33) for heat conduction. Owing
to imposed temperature along the edges x = 1.0 and y = 1.0, there is a total of 12 degrees
of freedom. The first column of Table I presents, in order of magnitude, all 12 eigenvalues
evaluated for the linear case (n = 1), with the two subsequent columns indicating number of
iterations needed to achieve convergence and order of evaluation of the results as obtained
from Algorithm 2. The calculations are in double precision with error tolerance tol = 10−14 ,
although only seven digits are displayed in the table. As a result of the plate’s symmetry about
the line y = x, the 2nd and 3rd, the 5th and 6th, and the 10th and 11th eigenvalues turn out
to be equal within machine precision. A total of 41 iterations were necessary for the solution
of the linear eigenproblem. The next set of three columns in Table I gives the eigenvalues
for the non-linear case of n = 3 generalized ”mass” matrices in Equation (33) together with
number of iterations and order of evaluation, as obtained in the algorithm. A total number of
55 iterations were required for the solution of the complete non-linear eigenproblem. As in the
linear case, once the first of a double eigenvalue is evaluated, only one iteration is necessary
to obtain the second eigenpair. An idea of the relative order of magnitude of the elements of
the matrices involved may be obtained from the Frobenius norms of K0 , M1 , M2 and M3 ,
which are approximately 8.5, 0.11, .73E-3 and .81E-5. In order to assess the performance of
the algorithm in the case of close eigenvalues, the numerical model was also analyzed for a

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
24 N. A. DUMONT

small change in two elements of the matrix M3 , to just break the model symmetry about the
line y = x although still keeping M3 positive definite. The results are shown in the last three
columns of Table I, indicating that, although more iterations have been required, in general,
the evaluation of a neighbor eigenpair is easily taken care of with Algorithm 2. A consistency
check of the calculated eigenpairs was carried out by constructing the matrix on the left-hand
side of Equation (44) and comparing with the identity matrix. In all three cases of Table I, the
global error was of magnitude 10−17 . Repeated evaluations for an error tolerance tol = 10−8
in the algorithm resulted in only one iteration saved for each eigenpair, in average, which is to
be expected for an algorithm with cubic convergence.
Figure 3 plots the curves λj (θ) given by the first of Equation (51) for all 12 eigenvalues of
the present example, to illustrate that the solution of the non-linear eigenproblem corresponds
to a minimum coinciding with the interception with the line λ = θ, as theoretically outlined
in Section 3.2.3. It is worth observing that λ12 is slightly larger than λ10 ≡ λ11 , in the linear
case (θ = 0), but soon becomes smaller.

y
(0, 1) u = 1.0 (1, 1)

u, x = 0.0 u = 1.0

(0, 0) u, y = 0.0 (0, 1)

Figure 2. Scheme for the homogeneous heat conduction in the square plate of Example 2, to illustrate
some features of the solution of linear and non-linear eigenproblems.

4.3. Example 3 – Fixed-free bar with viscous damping


This example assesses the non-linear eigenvalues related to a fixed-free bar modeled with three
truss elements, as given in Figure 4, for the matrices introduced in Example 1. The damping-
free case (µ = 0) is analyzed first, for n = 1, . . . , 4, according to Equation (33). The results for
all three eigenfrequencies ω – the square roots of the eigenvalues – are given in Table II, with
corresponding numbers of iteration required to achieve a tolerance error tol = 10−14 . Although
only seven digits are displayed, double precision was used in all evaluations. As in Example
2, a tolerance error tol = 10−8 would require about one iteration less per eigenvalue. The last
row shows the analytical eigenfrequencies for the fixed-free bar, given from k = (2j − 1)π/2/`,
where k is the wave number introduced in Equation (87) and j = 1, 2, · · · . A second analysis
was run for a viscous damping coefficient µ = 10, which corresponds to a complex-symmetric
eigenproblem with underdamping. An estimate of the relative magnitude of the elements of the
matrices involved in the numerical discretization can be obtained from the Frobenius norms

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
NON-LINEAR COMPLEX-SYMMETRIC EIGENVALUE PROBLEMS 25

Linear case (n = 1) Case with n = 3 Modified case with n = 3


# eval. # eval. # eval.
λ λ λ
iter. order iter. order iter. order
5.149701 4 1 4.935737 5 1 4.935695 5 1
29.84521 4 2 24.98865 6 2 24.96592 3 3
29.84521 1 3 24.98865 1 3 24.97391 7 2
58.58476 5 6 46.04634 7 6 45.83252 7 6
101.5537 7 4 70.85867 6 4 70.06585 6 4
101.5537 1 5 70.85867 1 5 70.35938 4 5
127.2303 4 11 89.37343 5 11 86.70345 6 12
142.8023 5 7 100.9992 6 10 99.96489 6 10
215.1464 3 12 148.4600 7 12 145.8409 5 11
260.7451 2 9 180.4566 4 8 177.9188 8 7
260.7451 1 10 180.4566 1 9 178.6309 5 8
261.6948 4 8 177.8527 6 7 175.6848 4 9

Table I. Eigenvalues, number of iterations required for tol = 10−14 and order in which the eigenvalues
were found for Example 2, covering both linear (n = 1) and non-linear eigenproblems (n = 3), besides
a modified problem for the assessment of Algorithm 2 in the case of close eigenvalues.

300

250
λ=θ

λ12 λ10 ≡ λ11


200
λ9
j

150
λ

λ8

λ7
100
λ5 ≡ λ6
λ ≡λ
2 3
λ4
50

λ1
0
0 50 100 150 200 250 300
θ

Figure 3. Curves λj (θ) given by the first of Equation (51) for all 12 eigenvalues of Example 2, to
illustrate that the solution of the non-linear eigenproblem corresponds to a minimum coinciding with
the interception with the line λ = θ.

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
26 N. A. DUMONT

of K0 , C1 , . . . , C4 and M1 , . . . , M4 , which are respectively 360., 10.5, 0.147E-1, 0.197E-4,


0.232E-7, 0.978, 0.553E-3, 0.337E-6 and 0.103E-9. The numerical results are given in Table
III in blocks that contain the complex eigenvalue and its absolute value, besides the starting
real eigenvalue used in the solution of the complex-symmetric eigenproblem, as described in
the perturbation analysis of Section 3.4 and in Algorithm 2. The analytical eigenfrequency
results, obtained from k = (2j − 1)π/2/`, are displayed in the bottom row of Table III
together with their absolute values. Since the starting eigenpair estimates for the complex-
symmetric case (ω, φ) are obtained in a first run of Algorithm 2 for a real-symmetric problem
resulted from disregarding the generalized damping matrices Cj , these estimates correspond
only approximately to the results of the damping-free analysis, as given in Table II. In fact,
as shown in the developments of Section 4.1.1, the generalized mass matrices are affected by
the damping coefficient µ when obtained in the frame of a consistent formulation. In this
and in other examples analyzed, the use of the damping-free results as starting values for
the complex-symmetric analysis ended up with approximately the same number of iterations
to achieve convergence. In the FORTRAN code implemented, Algorithm 2 is used in two
steps, first by calling a routine for real numbers that disregards Cj and produces the real
eigenpair estimates, and then by calling a version of the same routine for complex numbers
that solves the complete complex-symmetric eigenproblem. In the present example, the first
step has required the same iteration numbers to achieve convergence as given in Table II.
The second step required a total of 20 iterations for each value of n. The present strategy of
running a first step for the corresponding real-symmetric problem obtained by disregarding the
damping matrices Cj only works if the mass Mj are all positive definite even in the presence
of damping. As given in Equation (96) and assembled for three truss elements, this is true
only if α < 1.748474, which corresponds to µ < 34.96949. For larger damping factors in the
present truss problem, the real-symmetric problem of the first step has to correspond to the
actual damping-free case.

Figure 4. Scheme of a fixed-free bar for Example 3, modeled with three truss elements as described
in Example 1. In consistent units: total length ` = 3, elasticity modulus E = 100, cross area A = 1,
inertia ρ = 1 and viscous damping µ = 10.

# # #
ω1 ω2 ω3
iter. iter. iter.
n=1 5.295986 4 17.32050 4 31.42192 2
n=2 5.237565 5 16.00720 6 28.34557 6
n=3 5.236031 5 15.77676 6 27.39721 6
n=4 5.235988 5 15.72472 6 26.92935 6
analytical 5.235987 – 15.70796 – 26.17993 –

Table II. Eigenfrequencies for the fixed-free bar of Example 3 modeled as a real-symmetric
eigenproblem for µ = 0.

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
NON-LINEAR COMPLEX-SYMMETRIC EIGENVALUE PROBLEMS 27

ω1 ω2 ω3
|ω1 | ω1(start) |ω2 | ω2(start) |ω3 | ω3(start)
0.9455901 − 5.445289i 17.10934 − 5.357142i 31.46100 − 5.146581i
n=1
5.526781 5.526781 17.92842 17.92842 31.87918 31.87918
1.556697 − 4.977728i 15.34120 − 5.262882i 28.25366 − 5.245177i
n=2
5.215466 5.475053 16.21883 16.75392 28.73641 29.07573
1.554601 − 5.000691i 14.96868 − 5.111293i 27.18905 − 5.311572i
n=3
5.236764 5.474121 15.81729 16.60472 27.70301 28.44241
1.554187 − 4.999989i 14.89600 − 5.034719i 26.59903 − 5.316633i
n=4
5.235971 5.474112 15.72384 16.59288 27.12518 28.31762
1.554209 − 5.000000i 14.89094 − 5.000000i 25.69803 − 5.000000i
analytical
5.235987 − −− 15.70796 − −− 26.17993 − −−

Table III. Eigenfrequencies for the fixed-free bar of Example 3 modeled as a complex-symmetric
eigenproblem for µ = 10. Each block contains ω, its absolute value and the starting real estimate
obtained by disregarding all generalized damping matrices Cj in a first run of Algorithm 2.

4.4. Example 4 – Two artificial problems with overdamping


The application of Algorithm 2 to overdamping is demonstrated for two artificial problems:
In the first problem, one uses matrices K0 , M1 and M2 from Example 2 as the entries K0 ,
C1 and M1 of a new mathematical model that corresponds to n = 1 in Equation (32), which
is just non-linear; In the second problem, one uses matrix M3 of Example 2 twice, as entries
C2 and M2 , in order to create a generalization (n = 2) of the first problem. Although these
artificial problems correspond to no actual modeling of a mechanical phenomenon, it is worth
investigating them from the mathematical point of view, as the Frobenius norm of the first
damping matrix, C1 , is much larger than that of the first mass matrix, M1 , which are .11 and
.73E-3, respectively, according to the data of Example 2. As in the previous example, Algorithm
2 is run first for the corresponding damping-free problem, with results used subsequently for
the complex case, as outlined in Section 3.5. Tolerance error tol = 10−14 is adopted in both
steps. The eigenvalue results are displayed in Table IV with seven digits and in order of
evaluation. The two problems analyzed, as cases n = 1 and n = 2, are actually unrelated and
are presented side by side only for the sake of brevity. In brackets are indicated the numbers
of iterations required for convergence in the real and complex steps. As the problem topology
of Example 2 has been preserved, three double eigenvalues are also obtained. The problem
turns out to be overdamped for four out of the 12 eigenvalues in the case of n = 1, and for
only one eigenvalue in the case of n = 2. In the case of underdamping, the complex-associated
eigenpairs are given directly as (−ω, iφ). For overdamping, however, Algorithm 2 has to be run
once more in order to evaluate the inassociated eigenpairs (Section 3.3.3). New starting values
for the complex step, in the present examples, were taken as shifts of the negative square roots
of the eigenvalues obtained in the real step, −ω − 100, while keeping the initial eigenvectors
as the same ones of the real step. The results are shown in Table IV side by side with the
corresponding eigenvalues found previously and with numbers of iterations as the second values
in brackets. The eigenvectors obtained in the first complex run are approximately proportional
to the eigenvectors of the initial real step. The eigenvectors corresponding to the inassociated
eigenpairs, on the other hand, are of completely different pattern, which makes them very

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
28 N. A. DUMONT

elusive in terms of numerical evaluation. In the present theoretical frame it is not expected
that a code can run without human interference to obtain all complex eigenpairs, in the case
of overdamping, or else to infer that the unlikely case of critical damping has occurred.

Artificial problem with n = 1 Artificial problem with n = 2


ω(start) ω (# iter.) ω(start) ω (# iter.)
24.89241 (4) −5.390892i,−121.1726i (6,5) 24.86960 (7) 25.56068 − 5.380976i (6)
68.16300 (4) −39.15214i,−231.4377i (7,9) 24.86960 (1) 25.56068 − 5.380976i (6)
68.16300 (1) −39.15214i,−231.4377i (7,9) 13.81117 (7) −5.321045i,−84.61692i (7,11)
112.9713 (5) −96.92565i,−281.6820i (9,8) 34.73903 (10) 35.18777 − 5.165071i (6)
165.2637 (4) 72.19122 − 102.4555i (6) 48.22291 (6) 48.66042 − 7.541742i (6)
165.2637 (1) 72.19122 − 102.4555i (6) 48.22291 (1) 48.66042 − 7.541742i (6)
240.1926 (4) 89.63278 − 137.6027i (7) 91.40394 (6) 91.79706 − 9.275463i (6)
406.6003 (5) 213.1366 − 333.7571i (7) 93.80572 (5) 94.19215 − 9.682346i (6)
395.7893 (2) 224.1970 − 298.9793i (7) 93.80572 (1) 94.19215 − 9.682346i (6)
395.7893 (1) 224.1970 − 298.9793i (7) 65.51588 (6) 66.05254 − 9.764081i (6)
193.2663 (5) 125.7180 − 146.7885i (7) 52.99561 (6) 53.25906 − 5.983618i (6)
316.8612 (3) 214.3764 − 233.3319i (7) 76.40455 (6) 76.68535 − 7.235495i (6)

Table IV. Eigenfrequency results for two overdamped problems of Example 4 obtained artificially by
using the matrices of Example 2.

4.5. Example 5 – A rail-pad-sleeper-ballast model


A railway track comprising 20 sleepers is modeled as in Figure 5 [57]. Owing to longitudinal
symmetry, only half of a complete railway track is represented. Each rail segment between
sleepers is modeled with one Timoshenko beam element. Truss elements simulate the pads
that connect rail and sleepers. A half sleeper consists of two Timoshenko beam elements of
lengths 0.76 m and 0.50 m on a visco-elastic foundation to take into account the ballast (the
distance between rails is 1.52 m) [58]. Table V shows the geometrical and mechanical properties
used in the model: cross-section area A, moment of inertia I, length l, cross-section factor κ
related to shear force, Poisson’s ratio ν, elasticity modulus E, mass m per unit length; the
ballast has viscosity µ and stiffness w, both defined per unit length.
The eigenproblem is assessed for n = 2 generalized damping and mass matrices, according
to Equation (32), thus dealing with a total of five highly sparse matrices of order 140. In the
initial run without the damping matrices, an average of 5.99 iterations was necessary to solve
for each eigenpair with tolerance error tol = 10−14 and using double precision. As shown in
Figure 6, there are several clusters of close eigenvalues (equal within two to six digits), although
no machine-precision multiplicity is obtained (as in the case of Example 2). The structure’s
symmetry is reflected in the eigenvectors, as illustrated schematically in the left plot of Figure
7 for values corresponding to vertical displacements of the rail nodes. Results corresponding to
the four lowest and the two highest eigenvalues (φ1 , φ85 , φ2 , φ84 and φ94 , φ95 ) are displayed.
The symmetry/antisymmetry patterns of these eigenvectors are 12 digits accurate. The plot
on the right of Figure 7 shows real versus imaginary parts of the complex eigenvalues obtained
in the second run of the algorithm, starting from the results without damping. An average of
4.47 iterations was necessary to solve for each complex eigenpair starting from the real results

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
NON-LINEAR COMPLEX-SYMMETRIC EIGENVALUE PROBLEMS 29

A(m2 ) I(m4 ) l(m) κ ν E(N/m2 ) m(kg/m) µ(N s/m) w(N/m2 )


−4
Sleeper 0.05126 2.31×10 1.26 5/6 0.25 2.1×1010 99.603 4.667×104 1.1×108
−5 11
Rail 0.007686 3.217×10 0.545 1 0.25 2.059×10 60.640 – –
Pad 0.04 – 0.02 – – 3.25×108 39.2 3.75 ×106 –

Table V. Geometric and physical properties used in Example 5 [57].

of the first run, also with tolerance error tol = 10−14 . In the consistency check of the calculated
eigenpairs, according to Equation (44), the global errors were of magnitude 10−17 in the first
run and 10−15 for the complex eigenproblem.

Figure 5. Rail-pad-sleeper-ballast model with a total of 20 × 7 degrees of freedom.

9 9
10 10

8 8
10 10
λj

λj

7 7
10 10

6 6
10 10

0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140

Figure 6. Real eigenvalues of Example 5 in the sequence of evaluation (left) and of magnitude (right).

CONCLUDING REMARKS

The solution Algorithm 2 is based on the Rayleigh quotient iteration and on the Jacobi-
Davidson method. It was implemented in FORTRAN for the complete in-core solution. To
make sure that convergence is attained to the complex solutions of interest, a first analysis
is run for the underlying real-symmetric problem, where K0 is positive semidefinite and
Mj , j = 1, 2, · · · are positive definite. The solution of a problem with (machine-precision)

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
30 N. A. DUMONT

−3 3
x 10 10
4

3
φ95
2 2
10
φ1
1

− Im(ω )
j
0

1
−1
φ φ2 φ84 10
85
−2 φ94

−3

0
−4 10 3 4
10 Re (ωj) 10
2 4 6 8 10 12 14 16 18 20

Figure 7. Real-eigenmode schemes of the four lowest (φ1 , φ85 , φ2 , φ84 ) and the two highest (φ94 , φ95 )
eigenvalues of Example 5 (left). On the right: real versus imaginary parts of the complex eigenvalues.

multiple and close real eigenvalues is illustrated in Example 2. In a built-in second run of
the algorithm, the results of the real analysis are input as starting estimates for the actual
complex problem. The damping matrix C1 must be non-singular, although Mj , j = 1, 2, · · · , as
eventually modified by the presence of viscous damping, are no longer required to be positive
definite (Example 1). In a problem with m degrees of freedom, a maximum of m complex
solutions can be found directly. In the case of underdamping, there is a complementary complex
eigenpair (−ω, iφ) associated to every primary solution (ω, φ) found with the algorithm
(Example 3), as required in the numerical simulation of the transient problem using modal
analysis [10, 11]. If overdamping occurs, some eigenvalues are negative imaginary – with
corresponding eigenvectors that can be made real by properly scaling (although they are
generally multiplied by a complex factor when normalized). In such a case, the remaining
eigenpairs (here called inassociated) must be evaluated in an additional run of the code. As
these eigenpairs are very elusive to obtain in the iterative process, the analyst’s interference
is required to choose new, more adequate starting estimates. As shown in Example 4, the
solution of problems with multiple and close eigenvalues poses no additional difficulties even
in the case of overdamping. Example 5 illustrates the application of the algorithm to a larger
eigenproblem, in which several clusters of close eigenvalues are obtained. Given the small
tolerance error of 10−14 , round-off errors have been kept very low.
The algorithm always finds the smallest eigenvalue as the first solution. However, the
remaining eigenvalues are found in ascending order of magnitude, even in the case of linear
problems. To be applicable to large problems, the code must be modified, so that, after
solution of the complete real eigenproblem, a set of eigenpairs of interest is selected and
than input as estimates of the complex case. Adequate storage allocation of large matrices
and efficient solution for δuk in step (4.2) of Algorithm 2 also should be a concern. Iteration
on an a priori given subspace as well as matrix deflation might be conceived to improve
the algorithm. However, these techniques themselves are not exempt from theoretical and
numerical difficulties in the non-linear context.
The author is indebted to both reviewers for their invaluable contribution by pointing out

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
NON-LINEAR COMPLEX-SYMMETRIC EIGENVALUE PROBLEMS 31

several mistakes in the initial version of the paper and by suggesting some relevant literature.

REFERENCES

1. Dumont NA. The hybrid boundary element method: an alliance between mechanical consistency and
simplicity. Applied Mechanics Reviews 1989; 42(11 Part 2):S54-S63.
2. Dumont NA. Variationally-based, hybrid boundary element methods. Computer Assisted Mechanics and
Engineering Sciences (CAMES) 2003; 10:407-430.
3. Pian THH. Derivation of element stiffness matrices by assumed stress distribution. AIAA Journal 1964;
2:1333-1336.
4. Pian THH. Reflections and remarks on hybrid and mixed finite element methods. In Hybrid and Mixed
Finite Element Methods, Atluri, SN, Gallagher, RH & Zienkiewicz OC (eds). John Wiley & Sons, 1983;
565-570.
5. Brebbia CA, Telles JFC, Wrobel LC. Boundary Element Techniques. Springer-Verlag: Berlin and New
York, 1984.
6. Dumont NA, Oliveira R. From frequency-dependent mass and stiffness matrices to the dynamic response
of elastic systems. International Journal of Solids and Structures 2001; 38(10-13):1813-1830.
7. Dumont NA, Chaves RAP. General time-dependent analysis with the frequency-domain hybrid boundary
element method. Computer Assisted Mechanics and Engineering Sciences (CAMES) 2003; 10:431-452.
8. Dumont NA, Oliveira R. On exact and approximate frequency-domain formulations of the hybrid boundary
element method. In Procs. EURODINAME’99 – Dynamic Problems in Mechanics and Mechatronics,
Günzburg, Germany, 1999;219-224.
9. Przemieniecki J S. Theory of Matrix Structural Analysis. McGraw-Hill Book Company: New York, 1968.
10. Dumont NA. An advanced mode superposition technique for the general analysis of time-dependent
problems. Advances in Boundary Element Techniques VI 2005, eds. Selvadurai APS, Tan CL, Aliabadi
MH; CL Ltd.: England, 333-344.
11. Dumont NA. Advanced mode superposition and hybrid variational technique for the general analysis
of time-dependent problems. International Journal for Numerical Methods in Engineering 2006, to be
submitted.
12. Dumont NA, Prazeres PGC. A family of advanced hybrid finite elements for the general analysis of
time-dependent problems and non-homogeneous materials. In Procs. XXV CILAMCE – XXV Iberian
Latin-American Congress on Computational Methods in Engineering 2004, Recife, Brazil; 15 pp on CD.
13. Bathe K-J, Wilson EL. Numerical Methods in Finite Element Analysis. Prentice-Hall: New Jersey, 1976.
14. Clough RW, Penzien J. Dynamics of Structures. McGraw-Hill: New York, 1975.
15. Petyt M. Introduction to Finite Element Vibration Analysis. Cambridge Univ. Press: Cambridge, 1990.
16. Dumont NA. On the inverse of generalized λ-matrices with singular leading term. International Journal
for Numerical Methods in Engineering 2006; 66(4):571-603.
17. Hadeler KP. Mehrparametrige und nichtlineare Eigenwertaufgaben. Arch. Rat. Mech. Anal 1967; 27:306-
328.
18. Hadeler KP. Variationsprinzipien bei nichtlinearen Eigenwertaufgaben. Arch. Rat. Mech. Anal 1968;
30:297-307.
19. Voss H. Variational characterization of eigenvalues of nonlinear eigenproblems. Proceedings of the
International Conference on Mathematical and Computer Modelling in Science and Engineering 2003,
eds. Kocandrlova M, Kelar V; Czech Technical University in Prague, 379-383.
20. Horn RA, Johnson CR. Matrix Analysis. Cambridge University Press: Cambridge, 1985.
21. Wilkinson JH, The Algebraic Eigenvalue Problem. Oxford University Press: Oxford, 1977.
22. Golub GH, Van Loan CF. Matrix Computations. The Johns Hopkings University Press: Baltimore, MD,
3rd ed., 1996.
23. Mehrmann V, Voss H. Nonlinear Eigenvalue Problems: A Challenge for Modern Eigenvalue Methods.
Report 83, Arbeitsbereich Mathematik, TU Hamburg-Harburg, 2004.
24. Bai Z, Demmel J, Dongarra J, Ruhe A, van der Vorst H, editors. Templates for the Solution of Algebraic
Eigenvalue Problems: A Practical Guide. SIAM, Philadelphia, 2000.
25. Arbenz P, Hochstenbach ME. A Jacobi-Davidson method for solving complex-symmetric eigenvalue
problems. SIAM Journal of Scientific Computation 2004; 25(5): 1655-1673.
26. Ben-Israel A, Greville TNE. Generalized Inverses: Theory and Applications (2nd edn). Springer-Verlag:
New York, 2003.
27. Voss H. A new justification of finite dynamic element methods. International Series of Numerical
Mathematics 1987; 83:232-242.

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls
32 N. A. DUMONT

28. Jarlebring E, Voss H. Rational Krylov for nonlinear eigenproblems, an iterative projection method .
Applications of Mathematics 2005; 50(6):543-554.
29. Betcke T. Efficient Methods for Nonlinear Eigenvalue Problems. Diploma Thesis, Technical University of
Hamburg-Harburg, 2002.
30. Betcke T, Voss H. A Jacobi-Davidson-type projection method for nonlinear eigenvalue problems. Future
Generation Computer Systems 2004; 20(3):363-372.
31. Mehrmann V, Watkins D. Polynomial eigenvalue problems with Hamiltonian structure. Electronic
Transactions on Numerical Analysis 2002; 13:106-118.
32. Chatelin F.Valeurs Propres de Matrices. Masson: Paris, 1988.
33. Sleijpen GLG, Booten AGL, Fokkema DR, Van der Vorst HA. Jacobi-Davidson type methods for
generalized eigenproblems and polynomial eigenproblems: part I. Internal Report Universiteit Utrecht,
1995.
34. Sleijpen GLG, Van der Vorst HA. A Jacobi-Davidson iteration method for linear eigenvalue problems.
SIAM Journal of Matrix Analysis Applications 1996; 17:401-425.
35. Davidson ER. The iterative calculation of a few or the lowest eigenvalues and corresponding eigenvectors
of large real-symmetric matrices. Journal of Computational Physics 1975; 17:87-94.
36. Tisseur F, Higham NJ. Structured Pseudospectra for polynomial eigenvalue problems, with applications.
SIAM Journal of Matrix Analysis 2001; 23(1):187-208.
37. Casciaro R. A fast iterative solver for the nonlinear eigenvalue problem. Report 39, Laboratorio di
Meccanica Computazionale, Universita della Calabria, 2004.
38. Hwang TM, Lin WW, Liu JL, Wang W. Jacobi-Davidson methods for cubic eigenvalue problems.Numerical
Linear Algebra Applicattions 2004; 12(7):605-624
39. Triebsch F. Eigenwertalgorithmen für symmetrische lambda-Matrizen. Ph.D. Thesis, Technische
Universität Chemnitz-Zwickau, Germany, 1995.
40. Bott R, Duffin RJ. On the algebra of networks. Trans. Amer. Math. Soc. 1953; 74:99-109.
41. Gupta KK. On a finite dynamic element method for free vibration analysis of structures. Computational
Methods in Applied Mechanics and Engineering 1976; 9:105-120.
42. Paz M, Dung L. Power series expansion of the general stiffness matrix for beam elements. International
Journal for Numerical Methods in Engineering 1975; 9:449-459.
43. Kolousek V. Dynamics in Engineering Structures. Butterworths: London, 1973.
44. Warburton GB. The Dynamical Behaviour of Structures. Pergamon Press, 2nd edition, 1976.
45. Gaul L, Wagner M, Wenzel W, Dumont NA. On the treatment of acoustical problems with the hybrid
boundary element method. International Journal of Solids and Structures 2001; 38(10-13):1871-1888.
46. Sato K. Complex variable boundary element method for potential flow with thin objects. Computer
Methods in Applied Mechanics and Engineering 2003; 192:1421-1433.
47. Duncan WJ. Some devices for the solution of large sets of simultaneous linear equations. Philosofic
Magazine 1944; 35(7):660-670.
48. Lancaster P. Lambda-Matrices and Vibrating Systems. Pergamon Press: Oxford, 1966.
49. Meirovitch L. Elements of Vibration Analysis. McGraw-Hill: Tokyo, 1975.
50. Rogers EH. A minimax theory for overdamped systems. Arch. Rat. Mech. Anal 1964; 16:89-96.
51. Guo C-H, Lancaster P. Algorithms for hyperbolic quadratic eigenvalue problems. Mathematics of
Computation 2005; 74(252):1777-1791.
52. Croll JGA, Walker AC. Elements of Structural Stability. Butter & Tanner Ltd.: London, 1972.
53. Stewart GW. Perturbation theory for the generalized eigenvalue problem. Recent Advances in Numerical
Analysis 1978, eds. de Boor C, Golub GH, Academic Press: New York
54. Bathia R, Li R-C. On perturbations of matrix pencils with real spectra. II. Mathematics of Computation
1996; 65(214):637-645.
55. Fassbender H, Kressner D. Structured eigenvalue problems. GAMM Mitteilungen, Themenheft Applied
and Numerical Linear Algebra, Part II, 2005.
56. Bruch JC, Zyvoloski G. Transient two dimensional heat conduction problems solved by the finite element
method. International Journal for Numerical Methods in Engineering 1974; 8:481-494.
57. Oliveira AC. Um Modelo de Análise da Interação Dinâmica entre os Elementos de uma Via Férrea. M.Sc.
Thesis, PUC-Rio, Brazil, 2006.
58. Dumont, NA, Oliveira AC. A dynamic interaction model of railway track structural elements. In Procs.
XXVII CILAMCE – XXVII Iberian Latin-American Congress on Computational Methods in Engineering
2006, Belém, Brazil; 16 pp on CD.

Copyright °c 2006 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 00:1–6
Prepared using nmeauth.cls

You might also like