MODAL ANALYSIS OF LINEAR ASYMMETRIC

NONCONSERVATIVE SYSTEMS
By Sondipon Adhikari,1 Member, ASCE

ABSTRACT: In this work, classical modal analysis has been extended to treat lumped parameter asymmetric
linear dynamic systems. In the presence of general nonconservative forces, the damping matrix is not simulta-
neously diagonalizable with the mass and stiffness matrices. The proposed method utilizes left and right eigen-
vectors of the second-order system and does not require conversion of the equations of motion into the first-
order form. Left and right eigenvectors of the nonconservative system are derived in terms of the left and right
eigenvectors of the corresponding conservative system using a Galerkin error minimization approach in con-
junction with a Neumann expansion method. Transfer functions for the asymmetric nonconservative system are
derived in terms of the left and right eigenvectors of the nonconservative system. Suitable numerical examples
Downloaded from ascelibrary.org by Gazi Universitesi on 02/02/15. Copyright ASCE. For personal use only; all rights reserved.

are given to illustrate the proposed method.

INTRODUCTION modal analysis procedure to systems whose inertia, stiffness,

Modal analysis plays a central role in the vibrational studies
of linear engineering structures. Since the publication of Ray-
leigh’s classic monograph (1945, originally 1897), extensive
work has been done in this area over the last 11 decades. The
method was originally proposed for undamped structures
whose inertia and stiffness properties can be represented by
symmetric matrices or self-adjoint differential operators. How-
ever, Rayleigh himself noted that the real-life structural sys-
tems are not undamped, but that they possess some kind of
energy dissipation mechanism or damping. While there are
excellent theories in classical mechanics about why inertia and
stiffness properties can be represented by symmetric matrices
or self-joint operators when the system executes a small os-
cillation around a stable equilibrium, no such theory is avail-
able for energy dissipation. To solve this problem, in analogy
with ‘‘potential energy’’ and ‘‘kinetic energy,’’ Rayleigh as-
sumed a ‘‘dissipation function,’’ which is a nonnegative def-
inite functional of instantaneous generalized velocities. This
kind of damping model is known as viscous damping. So that
the modal analysis of undamped systems is applicable to
damped systems, Rayleigh made one more assumption by tak-
ing the viscous damping forces to be proportional to the inertia
and stiffness forces. Since its introduction, this model has been
used extensively and is known as ‘‘classical damping’’ or
‘‘proportional damping.’’
Rayleigh’s argument behind proportional damping was in-
tuitive rather than theoretical, and it was formulated for math-
ematical convenience only. Until Caughey and O’Kelly (1965)
gave necessary and sufficient conditions for a damped system
to have classical normal modes, the nature of proportional
damping was not very clear. This topic is well understood now,
but unfortunately, there is no mathematical theory or sufficient
experimental evidence to show why a physical system should
obey Caughey and O’Kelly’s criterion. In fact, practical ex-
periences show that most of the real-life structures do not sat-
isfy this criterion as they possess complex modes instead of
real normal modes. This leads to the notion of nonproportional
or nonclassical damping. Extending modal analysis to nonpro-
portionally damped systems is still an active area of research.
Currently, there have been efforts to extend the classical
1
Dept. of Engrg., Univ. of Cambridge, Trumpington St., Cambridge
CB2 1PZ, England, U.K. E-mail: sa225@eng.cam.ac.uk
Note. Associate Editor: Apostolos Papageorgiou. Discussion open until
May 1, 2000. To extend the closing date one month, a written request
must be filed with the ASCE Manager of Journals. The manuscript for
this paper was submitted for review and possible publication on May 26,
1999. This paper is part of the Journal of Engineering Mechanics, Vol.
125, No. 12, December, 1999. qASCE, ISSN 0733-9399/99/0012-1372–
1379/$8.00 1 $.50 per page. Paper No. 21056.
1372 / JOURNAL OF ENGINEERING MECHANICS / DECEMBER 1999
and damping properties cannot be represented by symmetric
matrices or self-adjoint differential operators. These kinds of
problems arise in the dynamics of actively controlled struc-
tures and in many general nonconservative dynamic systems,
for example, a moving vehicle on the road, a missile on its
trajectory, a ship’s motion in seawater, or the study of aircraft
flutter. The asymmetry of damping and stiffness terms are of-
ten addressed in the context of gyroscopic and follower forces
and asymmetry in mass (Soom and Kim 1983). Many authors
have considered this kind of general nonconservative linear
system. Fawzy and Bishop (1976) presented several relation-
ships satisfied by eigenvectors and eigenrows of the second-
order system and they also presented a method to normalize
them. Inman (1983) considered a class of asymmetric systems
that can be transformed into symmetric systems using a linear
transformation. Conditions for the existence of classical nor-
mal modes in this kind of asymmetric systems were given by
Ahmadian and Inman (1984). Later, Caughey and Ma (1993)
gave conditions under which a general asymmetric system can
be decoupled. In a subsequent paper, Ma and Caughey (1995)
used equivalence transform to analyze asymmetric systems.
However, their work was restricted to conservative systems or
to special kinds of nonconservative systems that can be de-
coupled by an equivalence transformation.
In this paper general asymmetric discrete nonconservative
dynamic systems are considered. The method presented does
not require conversion of the equations of motion to the first-
order form. In Section 2, left and right eigenvectors of the
nonconservative system are derived in terms of left and right
eigenvectors of the corresponding conservative system. A
Galerkin error minimization approach in conjunction with a
Neumann expansion method have been used for this purpose.
Transfer functions for the asymmetric system are derived in
Section 3. The derivation method uses the state-space repre-
sentation of equations of motion at the intermediate steps, but
finally relates the transfer functions to the left and right eigen-
vectors of the second-order system so that the first-order eigen-
solutions are not required. Applications of the proposed
method and the related numerical issues are discussed in Sec-
tion 4 using a three-degree-of-freedom asymmetric system. Fi-
nally, Section 5 summarizes the main results of the study re-
ported in this paper.
EIGENVALUES AND EIGENVECTORS OF
NONCONSERVATIVE DYNAMIC SYSTEMS
Background
The equations of motion of a linear damped discrete system
with N degrees of freedom can be written as

J. Eng. Mech. 1999.125:1372-1379.

 ¨ 1 Cu(t)
Mu(t) ¨ 1 Ku(t)
¨ = f(t)
where M, C, and K [ R N3N are mass, damping, and stiffness
matrices, respectively. u(t) [ R N is the response vector, t [
R1 denotes time, and f(t) [ R N is the forcing vector. Tradi-
tional restrictions of symmetry and positive definiteness are
not imposed on M, C, and K; however, it is assumed that M21
exist, that is the system is not defective or degenerate. It is
well known that for any linear system, if the forcing function
is harmonic, that is f(t) = f exp[st] with s = iv and amplitude
vector f [ R N, in steady state the response will also be har-
monic at frequency v [ R1. Therefore, we seek a solution of
(1) mined by such simple procedures. In this paper we try to de-
termine these quantities in terms of left and right eigenvectors
of the associated asymmetric conservative system. The eigen-
problem of asymmetric conservative systems has been well
studied in the literature (Huseyin 1978; Ma and Caughey
1995). Here we briefly outline the main features for further
reference.
Consider the undamped right eigenvalue problem
Kx ui = v2i Mx u i, ;i = 1, . . . , N (6)
where vi [ R is the ith natural frequency and x u i [ R is the N

the form u(t) = u˜ exp[st], where u˜ [ C N is the response vector ith undamped right eigenvector (mode shape). Similarly, the
in the frequency domain. Substitution of u(t) and f(t) in (1) undamped left eigenvalue problem can be defined by
results in x Tvi K = v2i x Tvi M, ;i = 1, . . . , N (7)
Downloaded from ascelibrary.org by Gazi Universitesi on 02/02/15. Copyright ASCE. For personal use only; all rights reserved.

s Mu˜ 1 sCu˜ 1 Ku˜ = f

2
(2) where x vi [ R is the ith undamped left eigenvector. For dis-
N

tinct eigenvalues it is easy to show that the left and right ei-
The right eigenvalue problem associated with the above
genvectors satisfy the biorthogonality relationship with respect
equation can be represented by the l-matrix problem (Lan-
to M and K (Huseyin 1978). We also normalize the eigenvec-
caster 1966)
tors such that
s 2i Mu i 1 si Cu i 1 Ku i = 0, ;i = 1, . . . , N (3)
x Tvj Mx ui = dji , x vTj Kx ui = v2i dji , ;j, i = 1, . . . , N (8)
where si [ C is the ith latent root (eigenvalue); and u i [ C N

is the ith right latent vector (right eigenvector). Similarly, the
left eigenvalue problem can be represented by
s 2i viT M 1 siviTC 1 viT K = 0, ;i = 1, . . . , N
where vi [ C is the ith left latent vector (left eigenvector);
N
and (?)T denotes the matrix transpose.
where dji is the Kroneker’s delta function. Because all of the
undamped eigenvalues (v2i ) are assumed to be distinct, xui, ;i
= 1, . . . , N form a complete set of vectors, so that ui can be
(4) expanded as a complex combination of xui. Similarly, vi can
also be expanded in terms of xvi. Thus, an expansion of the
form

O O
N N
When M, C, and K are general asymmetric matrices, the
left and right eigenvectors can easily be obtained from the ui = a (i)
j xu j; vi = bj(i) xvj (9a,b)
j =1 j =1
first-order formulations, for example, state-space method
(Newland 1989), Duncan forms (Meirovitch 1980), etc. Al- may be considered. Now, without any loss of generality, we
though exact in nature, the first-order methods require signif- can assume a (i) i = 1 and bi = 1 (normalization), which leaves
(i)

icant numerical efforts for obtaining the eigensolutions as the
size of the problem doubles. Moreover, these methods also
lack some of the intuitive simplicity of the analysis based on
N-space. For these reasons the determination of eigenvalues
and eigenvectors in N-space for asymmetric nonconservative
systems is very desirable. Ma and Caughey (1995, Theorem
3) have shown that in the special case, when M21C and M21K
us to determine a (i)
j , b(i)
j , ;j ≠ i. A Galerkin type of error mini-
mization approach combined with a complex Neumann ex-
pansion method is adopted for this purpose.
Substituting the expansion of u i in (3) and using the usual
definition of natural frequencies in structural dynamics, i.e., li
= si /i, the error vector for the ith mode can be expressed as

O
N
commute in product, the linear asymmetric nonconservative
system (1) can be decoupled by an equivalence transforma- D(i) = 2 li2 a j(i) Mx u j 1 ili a (i)
j Cx u j 1 a j Kx u j [ C
(i) N
(10)
j =1
tion, and hence, the N-space method can be used. But in gen-
eral, linear nonconservative systems do not satisfy this con- Consider the undamped left eigenvectors xvk, ;k = 1, . . . , N
dition and some kind of approximate methods must be used as ‘‘weighting functions,’’ and following the Galerkin method
for further analysis. Meirovitch and Ryland (1985) and Malone we have (xvk, D(i)) = 0 or x vTkD(i) = 0, where (?, ?) denotes the
et al. (1997) have used a perturbation method to determine the standard inner product norm in C N. Using the biorthogonality
eigensolutions of gyroscopic systems. The difficulty with these property of the undamped left and right eigenvectors described
kinds of perturbation methods is that any kth order (k > 2) in (8), one obtains

O
term for eigenvalues or eigenvectors requires the determina- N
tion of all (k 2 1) terms of eigenvalues and eigenvectors, and
they are correlated among themselves. It may be noted that 2l2i a (i)
k 1 ili a (i)
j C9kj 1 vk a k = 0,
2 (i)
;k = 1, . . . , N (11)
j =1
eigenvalue determination is essentially a numerical method
and leads to the solution of where C9kj = x vTkCx u j. The ith equation of this set obtained by
setting k = i is a trivial case because a (i)
i = 1 has already been
det(s 2i M 1 si C 1 K) = 0 (5) assumed. From the above set of equations, excluding this triv-
The above equation is a polynomial of order 2N and yields ial case, one has

S O D
2N values of si, which appear in complex conjugate pairs. In N
this paper it is assumed that all of the eigenvalues are distinct. 2l2i a (i)
k 1 ili C9ki 1 a (i)
k C9kk 1 a (i)
j C9kj 1 v2k a (i)
k = 0,
Several efficient numerical methods are available to solve (5), j≠k≠i
and so for eigenvalue determination perturbation methods are
most likely not necessary. ;k = 1, . . . , N ≠ i (12)
which can be written further in matrix form as
Determination of Eigenvectors
[P(i) 2 Q(i)]â (i) = gu(i) (13)
Unlike the eigenvalues, the left and right eigenvectors for
general asymmetric nonconservative systems cannot be deter- In the above equation
JOURNAL OF ENGINEERING MECHANICS / DECEMBER 1999 / 1373

J. Eng. Mech. 1999.125:1372-1379.

 P(i) = diag F v21 2 l2i 1 ili C911
2ili
, . . . , {ith term deleted}, . . . , b̂(i) = {b1(i), b2(i), . . . , {ith term deleted}, . . . , bN(i)}T [ C (N21)

k , ;k = 1, . . . , N, ≠ i. Now, using
is the vector of unknown b(i)
v 2 l 1 ili C9NN
2
N

2ili
2
i
G [ C(N21)3(N21)
(14)
the Neumann expansion method and defining
R (i)
v = P
(i) 21 T
Q(i) [ C (N21)3(N21); b (i)
0 = P
(i) 21
gv(i) [ C (N21)

F G
and the traceless matrix (24a,b)
from (22) one obtains b̂(i) as a series
0 C921 ??? {ith term deleted} ??? C91N
C921 0 ?? ?? ?? C92N
2 3
b̂(i) = [I N21 1 R v(i) 1 R v(i) 1 R v(i) 1 ? ? ?]b 0(i)
? ? ?
? ? ?? ??? ???
?? ??
(i)
Q = ? {ith term deleted}
??? ??? ??? ??? ??? ??? = b 0(i) 1 b1(i) 1 b (i)
2 1 ? ? ? bk 1 ? ? ?
(i)
(25)
C9N1 C9N2 ??? {ith term deleted} ??? 0
Here
[R (N21)3(N21)
(15) b(i) (i) (i) (i) (i) (i) (i) (i) (i)
1 = R v b 0 , b 2 = R v b1 , . . . , b k = R v b k21 (26)
Downloaded from ascelibrary.org by Gazi Universitesi on 02/02/15. Copyright ASCE. For personal use only; all rights reserved.

g = {C91i, C92i, . . . , {ith term deleted}, . . . , C9Ni} [ R

(i)
u
T (N21)
The ith eigenvector of the nonconservative system vi can be
(16) obtained by substituting b̂(i) in the second equation of (9). This
and method does not require much computational time as closed-
form expressions for R (i) (i)
v and b 0 are available (see Section 4).
â(i) = {a 1(i), a 2(i), . . . , {ith term deleted}, . . . , a (i)
N} [ C
T (N21)
It may be noted that by taking more terms in the series (25),
(17) one can obtain vi to arbitrary precession if the complex matrix
2 3
power series I N21 1 R v(i) 1 R v(i) 1 R v(i) 1 ? ? ? is convergent.
k , ;k = 1, . . . , N, ≠ i. From (13),
is the vector of unknown a (i)
â (i) must be determined by performing the associated matrix
inversion, and this is achieved by using the Neumann expan- Convergence
sion method. A similar procedure was used by Adhikari and A necessary condition for validity of the series expression
Manohar (1999) in the context of inversion of the dynamic of â (i) and bˆ (i) in (20) and (25), required for the determination
stiffness matrix of structures with stochastic properties. Now, of the right and left eigenvectors of the nonconservative sys-
using the Neumann expansion we have tem (1), is that the complex matrix power series
21 21 21
â (i) = [I N21 2 P(i) Q(i)] {P (i) g(i)} Su = I N21 1 R (i)
2 3
u 1 Ru 1 R (i) 1 ???
(i)
u (27)
2 3
= [I N21 1 R (i)
u 1 Ru
(i)
1 R (i)
u 1 ? ? ?]a (i)
0 (18) and
where I N21 is (N 2 1) 3 (N 2 1) identity matrix Sv = I N21 1 R (i)
2 3
v 1 Rv 1 R (i) 1 ???
(i)
v (28)
21 21
(i)
R =P (i)
Q [C (i) (N21)3(N21)
; a =P(i) (i)
g [C
(i) (N21)
(19a,b)
u 0 u
are convergent. Looking at the expression of R (i) (i)
u and R v in
(i)
Because P is a diagonal matrix, its inversion can be carried (19) and (24), it may be revealed that they are quite similar,
out analytically, and subsequently, the closed-form expressions and it is sufficient to study the convergence property of any
of R (i) (i)
u and a 0 can be obtained (see Section 4). This makes
one of the series. Here the series in (27) will be studied.
further calculations involving these quantities simpler. From
(18), â (i) can be calculated in an efficient way, as one can write Condition 1
â (i)
=a (i)
0 1a (i)
1 1a (i)
2 1 ??? 1 a (i)
k 1 ??? (20) The complex matrix power series S u converges if, and only
if, for all of the eigenvalues s(i)
j of the matrix R (i)
u , the in-
where equality usj u < 1 holds.
(i)

1 = R u a 0 , a 2 = R u a 1 , ? ? ? a k = R u a k21
a (i) (i) (i) (i) (i) (i) (i)
This implies that all of the a can be obtained using succes-
sive matrix-vector multiplications only. Now noting that â (i) is
k , ;k = 1, . . . , N, ≠ i, the substitution of it in
the vector of a (i)
(9a) will give the right eigenvectors associated with the non-
conservative system (1). It is easy to see that by taking more
terms in the series represented by (20), one can obtain the right
eigenvectors to any arbitrary precession,2 provided 3
the complex
matrix power series I N21 1 R u(i) 1 R u(i) 1 R (i)
u 1 ? ? ? is con-
vergent. The convergence issue of this series is addressed in
the next subsection.
Similarly, the left eigenvectors can be obtained by substi-
tuting the expansion of vi in (4), and letting x u j as weighting
functions while applying the Galerkin method. Following the
procedure employed for the right eigenvectors, one can write
T
[P (i) 2 Q(i) ]bˆ (i) = gv(i)
where
iN} [ R
(i) (i)
(21) Proof of this directly follows from Wilkinson (1965). Al-
(i)
though the above condition is both necessary and sufficient,
k checking convergence for all i = 1, . . . , N is often not feasi-
ble. Therefore, we look for a sufficient condition that is rela-
tively easy to check and that ensures convergence for all i =
1, . . . , N.
Condition 2
The complex matrix power series S u converges for any li,
vi if C9 is a diagonally dominant matrix.
Proof
Because a matrix norm is always greater than or equal to
its maximum eigenvalue, it follows from Condition 1 that,
convergence of the series is guaranteed if iR (i)
u i < 1. Writing
(22) the sum of absolute values of entries of R (i)
u (see Section 4)
results in the following inequality as the required sufficient
condition for convergence:

OOU U
g(i)
v = {C9
i1, C9
i2, . . . , {ith term deleted}, . . . , C9
T (N21)
N N
liC9kj
(23) (1 2 djk) < 1 (29)
k =1 j =1 v 2 l2i 1 iliC9kk
2
k
and k≠i j≠i

1374 / JOURNAL OF ENGINEERING MECHANICS / DECEMBER 1999

J. Eng. Mech. 1999.125:1372-1379.

 Dividing the numerator and denominator by li, the above in- to be useful to perform the calculations in state-space and then
equality can be written as relate the result to the left and right eigenvectors of the second-

OO
N N order system determined in the last section.
uC9u
kj
Eq. (1) can be transformed into first-order (state-space)
<1 (30)
k =1 j =1 u1/li(v 2 l2i ) 1 iC9kku
2
k form as
k≠i j≠i≠k

ż(t) = Az(t) 1 3(t) (34)

Taking the maximum for all k ≠ i, this condition can further
be represented as where A [ R 2N32N = system matrix; 3(t) [ R 2N = forcing
vector; and z(t) [ R 2N = response vector in the state-space
O
N

uC9u given by

F G H J
kj

u(t)
j =1
0 IN
j ≠ i,k
z(t) =

F G
max <1 (31) A= ; (35a,b)
k≠i
1/2 2M 21 K 2M 21 C u(t)
˙

H J
1
(uliu2 2 v2k )2 1 C9kk2
uliu2 0
3(t) = (35c)
M 21f(t)
Downloaded from ascelibrary.org by Gazi Universitesi on 02/02/15. Copyright ASCE. For personal use only; all rights reserved.

It is clear that (31) always holds if

In the above equation 0 is the N 3 N null matrix and I N is
O
N

uC9u ;k ≠ i the N 3 N identity matrix. For periodic forcing 3(t) = P

kj < uC9
kku, (32)
j =1 exp[st], assume periodic solution of the form z(t) = z̃ exp[st],
j≠i≠k and substituting it in (34) we obtain
which, in turn, implies that for all i = 1, . . . , N, the inequality sz˜ = Az˜ 1 P (36)
iR (i)
u i < 1 holds if C9 is diagonally dominant. It is important
to note that the diagonal dominance of C9 is only a sufficient In the above equation z̃ [ C is the frequency domain re-
2N

condition, and that the lack of it does not necessarily preclude sponse vector in state-space and can be related to the corre-
the convergence of S u. sponding vector of the second-order system as

Discussion z̃ = HJ ũ
su˜
(37)
Following the procedure outlined in the previous subsection
one can obtain right and left eigenvectors of the nonconser- and the amplitude of the forcing vector is P = {0, M 21 f}T [
vative systems up to any desired level of accuracy without R 2N. The right eigenvalue problem associated with (36) can be
using first-order formalizms. It may be noted that the expres- expressed as
sions developed here are somewhat different from perturbation Az i = si z i, ;i = 1, . . . , 2N (38)
solutions [for example, Cronin (1990) for symmetric systems],
because higher-order terms of eigenvectors do not depend on where si [ C is the ith eigenvalue and z i [ C is the ith 2N

the higher-order terms of the eigenvalues. The eigenvalues are right eigenvector which is related to the ith right eigenvector
determined exactly by solving the polynomial equation (5), of the second-order system as

H J
and subsequently, the right and left eigenvectors of the non-
conservative system can be obtained from the series expression ui
zi = (39)
(9). The vector of complex constants a (i) and b(i) si u i
j j appearing in
(9) are further obtained from (20) and (25), respectively. For The left eigenvector yi [ C 2N associated with si is defined
many engineering problems it is often observed that entries in by the equation
the C matrix are not very ‘‘big,’’ and that by retaining only a
few terms in the series, (20) and (25) will result in an ac- yTi A = siyiT (40)
ceptable accuracy. Closed-form approximate expressions of
right and left eigenvectors by retaining one and two terms in For distinct eigenvalues it is easy to show that the right and
these series are given in Appendix I. These expressions might left eigenvectors satisfy an orthogonality relationship, that is
be useful whenever we find that the entries of the C matrix yTi z i = 0, ;j ≠ i (41)
are small compared with that of M and K.
and we also normalize the eigenvectors so that
TRANSFER FUNCTIONS
yTi z i = 1 (42)
Transfer functions of a system completely define its input-
output relationship. Suppose the forcing vector f in (2) is zero The above two equations imply that the dynamic system de-
for all of the entries except the mth, which has an entry p. fined by (34) possesses a set of biorthogonal eigenvectors with
Due to this force, if the response at some nth degree of free- respect to the system matrix A.
dom is ũn, then the transfer function Hnm(v) can be de- Because it has been already assumed that A has distinct
fined by eigenvalues, the right eigenvectors z i, ;i = 1, . . . , 2N, form
a complete set of vectors. Thus, the solution of (36) can be
ũn expanded as
Hnm(v) = (33)

O
p 2N

In principle, one can carry out the associated matrix inver- z̃ = ej z j (43)
j =1
sion with (2) and subsequently calculate u˜ to obtain u˜ n. How-
ever, this would be a somewhat expensive numerical exercise where ej [ C, ;j = 1, . . . , 2N are set of constants to be
and may not offer much physical insight to the analyst. In- determined. Note that the above equation is a vector equation
stead, we seek a solution analogous to the classical modal with 2N rows: The first N rows are a solution of the second-
series solution of the undamped or proportionally damped order system (2), which we aim to obtain. Substituting z̃ in
symmetric system (Meirovitch 1967). At this stage it turns out (36), and premultiplying by the left eigenvector yTk , one has
JOURNAL OF ENGINEERING MECHANICS / DECEMBER 1999 / 1375

J. Eng. Mech. 1999.125:1372-1379.

 O case M = M T, K = K T, and C = 0 results in vj = u j =
2N

y (s 2 A)ej z j = y P
T
k
T
k (44) x u j , and mj = 1/2vj , which reduces expression (53) to

O
j =1
N
x u( j)m x (uj)n
Using the biorthogonality relationship of the left and right Hnm(v) = (54)
eigenvectors of A, the above equation results in j =1 v2j 2 v2

yTk P 2. Asymmetric conservation system (Huseyin 1978): In this

ek = , ;k = 1, . . . , 2N (45) case C = 0 results in u j = x u j , vj = x v j , and mj = 1/2vj ,
s 2 sk
which reduces expression (53) to

O
The ek expressed above is not very useful because it is in terms N
of left and right eigenvectors of the first-order system. To ob- x v( mj) x u( j)n
Hnm(v) = (55)
tain a relationship with the eigenvectors of the second-order j =1 v2j 2 v2
system, assume

HJ
3. Symmetric nonconservative system (Vigneron 1986): In
y1 i this case M = M T, K = K T, and C = C T results in vj =
yi = (46) u j and mj = 1/(u Tj [2lj M 2 iC]u j), which reduces expres-
y2 i
Downloaded from ascelibrary.org by Gazi Universitesi on 02/02/15. Copyright ASCE. For personal use only; all rights reserved.

sion (53) to
where y1i, y2i [ C N. Substituting yi in (40) and taking the

OH J
N
transpose, one obtains mj u (mj) u (nj) m* ( j)* ( j)*
j um un
Hnm(v) = 2 1 (56)
siy1i = 2K T M 21 y2i
T
(47a) j =1 v 2 lj v 1 l* j

siy2i = y1i 2 CT M 21 y2i

T
or y1i = [si I N 1 CT M 21 ]y2i
T
(47b) From the transfer function expression [(53)], the steady-state
response due to harmonic load or response due to broadband
Elimination of y1i from the above two equations yield random excitation can be obtained directly. However, response
T T due to transient loads or that due to initial conditions can also
si(siy2i 1 C T M 21 y2i) = 2K T M 21 y2i or
be obtained by familiar methods using convolution integrals
( yT2i M 21) [s i2 M 1 si C 1 K] = 0 (48) in the time domain after obtaining the impulse response func-
tion by taking the Fourier transform of (53) (Meirovitch 1967,
By comparing this equation with (4), it can be observed that section 7.6).
the vector yT2i M 21 is parallel to v Ti , that is, there exists a non-
zero gi [ C, such that APPLICATIONS AND EXAMPLES
21
y M T
2i = gv T
i i or y2i = gi M vi T
(49) Following the procedure outlined before the frequency do-
main response of an asymmetric nonconservative system can
Substituting y1i, y2i, and also z i from (39) into the normali-
be obtained using the right and left eigenvectors. The method
zation condition (42), we have the scalar equation
presented here is very similar to that of the classical modal
yT1i u i 1 siy2Ti u i = 1 or giv iT [C 1 si M]u i 1 gi siv iT Mu i = 1 analysis because it appears that only the undamped modes and
(50) frequencies have to be replaced appropriately by right and left
eigenvectors and complex eigenvalues of the asymmetric non-
From the above equation, the scalar constant gi can be ob- conservative system. The left and right eigenvectors can be
tained as obtained from that of the corresponding conservative system
1 using a series expansion in N-space and the eigenvalues can
gi = (51) be obtained directly from the characteristic equation. Transfer
v iT [2si M 1 C]u i
functions are further expressed in terms of these left and right
Now taking the first N rows of (43) and using (49), one eigenvectors and complex eigenvalues. Here we briefly sum-
obtains marize the steps to be followed:

O O
2N
y2Tj M 21 f
2N
pgj v m( j) 1. Obtain the eigenvalues vi, right eigenvectors x u i, and left
ũ = uj = uj (52)
j =1 s 2 sj j =1 s 2 sj eigenvectors x v i of the conservative system from Kx ui =
v2i Mx ui and x Tvi K = v2i x TviM for all i = 1, . . . , N. Nor-
where v (mj) represents the mth element of the jth left eigenvector malize x u i and x v i so that x vTj Mx ui = dji.
of the second-order system vj . 2. Determine the eigenvalues li of the nonconservative sys-
It is well known that for any system with real coefficient tem by solving the polynomial equation det(s 2i M 1 isi C
matrices, the eigenvalues and eigenvectors appear in complex 1 K) = 0. Obtain the complex natural frequencies li
conjugate pairs. This implies that with the usual definition of = si /i.
natural frequencies, sj = ilj and s j* = 2ilj* occurs with u j , 3. Set up the C9 matrix using C9kj = x vTkCx u j, ;j, k =
vj , and u *
j , v j* , respectively. Here (?)* denotes complex con-
1, . . . , N. Calculate the matrices R (i) (i)
u and R v [obtained
jugation. Now from the definition of the transfer function in by simplifying (19) and (24), respectively], using
(33) and from (52), writing the terms corresponding to si and
s i* separately, we finally obtain 2ili C9kj (1 2 dk1 j1)

OH J
R (iu k) j =
N
mj v u
( j) ( j)
m* u( j)* ( j)* 1 1
vk2 2 li2 1 ili C9kk
m n j v m n
Hnm(v) = 2 1 (53)
j =1 v 2 lj v 1 l*
j and
where m j = igj = 1/(v [2lj M 2 iC]u j). This is a generalization
T
j 2ili C9jk (1 2 dk1 j1)
of the known expression of the transfer function for symmetric R (iv k) j = , ;j, k = 1, . . . , N, ≠ i
1 1
vk2 2 li2 1 ili C9kk
conservative systems to asymmetric nonconservative systems.
For keeping the dimension of R (i) u and R v to (N 2 1) 3
(i)
Transfer functions for several interesting special cases may be
obtained from (53): (N 2 1), express k1, and consequently, j1 as k1 = k 2
8 (k,i) and ji = j 2 8 ( j,i). Here the function 8 (k,i) (and
1. Symmetric conservative system (Rayleigh 1897): In this similarly 8 ( j,i)) is defined as
1376 / JOURNAL OF ENGINEERING MECHANICS / DECEMBER 1999

J. Eng. Mech. 1999.125:1372-1379.

 8 (k,i ) = H
0,
1,
not defined
if
if
if
k<i
k>i
k=i
(57) C= F 2.6710
0.4843
0.2875
2.9592
1.2606
20.5808
2.9651
20.6119
1.2272
G (61)

Also obtain the vectors a (i)

F G
(i)
0 and b 0 , using 1.3748 10.9440 25.2975
K= 1.2625 2.8770 217.4195 (62)
2ili C9ji
a (i)
0j =
0.7455 24.1244 0.8625
1
vj2 2 li2 1 ili C9jj
Numerical values for the entries of M and K matrices are
and taken from Ma and Caughey (1995). It may be easily verified
2ili C9ij that all of the above matrices are neither symmetric nor pos-
b (i)
0j = , ;j = 1, . . . , N, ≠ i itive definite. The problem of determining the left and right
1
vj2 2 li2 1 ili C9jj eigenvectors will be solved by following the steps described
4. Select the number of terms, say r, to be retained in the earlier.
expansion of the eigenvectors [see (20) and (25) for ref- Step 1: Solution of the right and left eigenvalue problem of
Downloaded from ascelibrary.org by Gazi Universitesi on 02/02/15. Copyright ASCE. For personal use only; all rights reserved.

erence]. Calculate a (i)k = R (i) (i) (i) (i) (i)

u a k21 and b k = R v b k21 for
the associated conservative system given by (6) and (7) and
all k = 1, . . . , r, and subsequently obtain â (i) = normalization according to (8) yields the natural frequencies
(rk=1a (i) ˆ (i) = (rk=1 b (i)
k21 and b k21.
v1 = 1.3506, v2 = 3.0913, v3 = 4.8527, and the eigenvectors

F G
5. Obtain the right eigenvectors of the nonconservative sys-
1.0741 20.6240 0.4421
tem as
X u = [x u 1, x u 2, x u 3] = 20.0292 0.9635 20.5428 (63)

O 20.0047 20.0290
N
0.8731

F G
ui = xui 1 â (i)
j1 x u j
1
j =1 0.3082 0.4876 0.5167
j≠i
X v = [x v1, x v2, xv3] = 0.5025 0.3000 20.7973 (64)
and left eigenvectors as 0.8983 20.9954 0.5915

O Step 2: By solving the characteristic determinant, det(s2i M

N

vi = x v i 1 bˆ (i)
j1 x v j
1
1 isi C 1 K) = 0, the eigenvalues of the nonconservative sys-
j =1
j≠i
tem are obtained; s1 = 20.7725 1 1.1965i, s2 = 20.7251 1
3.0560i, s3 = 21.3949 1 4.3092i, together with the corre-
for all i = 1, . . . , N. sponding complex conjugate values. The complex natural fre-
6. Finally, calculate the transfer function matrix using these quencies are obtained by li = si /i: l1 = 1.1965 1 0.7725i, l 2
left and right eigenvectors = 3.0560 1 0.7251i, l 3 = 4.3092 1 1.3949i.

OH J
Step 3: The C9 matrix calculated from C9 = X Tv CX u as

F G
N
mj u jv jT m* * T*
j uj vj
H(v) = 2 1 ,
j =1 v 2 lj v 1 l* j
1.3850 0.1101 1.5220
C9 = 1.1774 1.5858 20.7533 (65)
1 1.2321 20.6323 2.8142
with mj = T
v j [2lj M 2 iC]u j (58)
From the C9 matrix we can further obtain R (i)u , i = 1, 2, 3,
or alternatively, in the time domain, obtain the impulse using the closed-form expression described earlier

F G
response function matrix
0 20.0768 1 0.1207i

O
N R (1)
u = (66a)
2ilj*t
20.0209 1 0.0384i 0
{2mjvj u Tj eilj t 1 m j*v j* u j* e
T

F G
h(t) = } (59)
j =1
0 20.1235 1 0.5850i
R (2)
u = (66b)
Compare to the state-space approach, the procedure outlined 20.0244 2 0.2886i 0
above offers a significant reduction in computational effort be-
cause two eigenvalue problems of size N 3 N have to be
solved (Step 1) instead of solving two eigenvalue problems of
R (3)
u = F 0
0.0981 1 0.4920i
0.0010 1 0.0280i
0 G (66c)

size 2N 3 2N. Moreover, this method provides a better phys- Similarly R (iv ), i = 1, 2, 3, can also be calculated
ical insight because only the familiar N-space eigenvectors are
used. Various efficient numerical algorithms are available
(Press et al. 1992) to solve the characteristic polynomial equa- R (1)
v = F 0
20.0250 1 0.0457i
20.0645 1 0.1013i
0 G (67a)

F G
tion (Step 2). Many of these algorithms are iterative in nature
and to increase the efficiency of the solution procedure, the 0 20.1000 1 0.4735i
R (2)
v = (67b)
eigenvalues obtained from a first-order perturbation method 20.0301 2 0.3565i 0

F G
given by (76) in Appendix I can be used as an initial guess.
The procedure to be followed later for obtaining the eigen- 0 0.0102 1 0.2996i
R (3)
v = (67c)
vectors of the nonconservative system, i.e., Step 3 to Step 5, 0.0092 1 0.0460i 0
is quite systematic and straightforward and does not involve
intensive computation. Now obtain a (i) (i)
0 , b 0 , i = 1, 2, 3 using the closed-form expres-

To illustrate the proposed method, a numerical example of sion described earlier

a three-degree-of-freedom system is considered. The M, C,
and K matrices appearing in (1) are defined by a (1)
0 = H 0.1201 2 0.1887i
0.0408 2 0.0748i J (68a)

M= F 0.5740
0.7070
0.4620
1.3858
0.7070
20.1914
1.3858
20.7070
20.1914
G (60)
a (2)
0 = H 20.0089 1 0.0423i
0.0125 1 0.1481i J (68b)

JOURNAL OF ENGINEERING MECHANICS / DECEMBER 1999 / 1377

J. Eng. Mech. 1999.125:1372-1379.

 a (3)
0 = H 0.0132 1 0.3872i
20.0627 2 0.3148i J (68c)
have the same numerical value in the first element. It is clear
that the results obtained from the method outlined hereby

H J
0.0112 2 0.0176i
b (1)
0 =
0.0504 2 0.0924i
H J
20.0955 1 0.4525i
b (2)
0 =
0.0149 1 0.1764i
match the exact solutions to an excellent accuracy. An inter-
esting point to be noted is that the matrix C9 is not diagonally
(69a)
dominant as uC911u < uC913u. However, all of the eigenvalues sj(i)
satisfy usj(i)u < 1, as we have obtained
of the matrix R (i)
u
max(s ) = 0.0791, max(s(2)) = 0.4161, and max(s(3)) =
(1)
(69b)
0.1186. This demonstrates that the condition established in

H J
(32) is a sufficient condition, and there may be cases when
0.0107 1 0.3135i one can obtain the eigenvectors of the nonconservative system
b (3)
0 = (69c)
20.0527 2 0.2642i even if this condition is not satisfied.
Step 4: Assume the number of terms to be retained r = 7,
in the series expansion (20) and (25). Computing each of the CONCLUSIONS
terms using (21) and (26) and summing them results

H J
The problem of dynamic analysis of nonconservative linear
Downloaded from ascelibrary.org by Gazi Universitesi on 02/02/15. Copyright ASCE. For personal use only; all rights reserved.

0.1247 2 0.1782i multiple degrees-of-freedom systems has been considered. The

â (1) = (70a)
0.0450 2 0.0663i central theme of the approach adopted in this paper is to utilize

H J
20.1181 1 0.0348i
â (2) =
0.0255 1 0.1812i
the left and right eigenvectors of the second-order system so
that conversion of the equations of motion to the first-order
(70b)
form can be avoided. The analysis is restricted to systems with

H J
0.0205 1 0.3800i
â (3) =
20.2477 2 0.2674i
nonrepetitive eigenvalues and with a nonsingular mass matrix.
It has been assumed that, in general, the mass, damping, and
(70c)
stiffness matrices are neither symmetric nor positive definite

H J
0.0173 2 0.0067i
b̂ (1) =
0.0503 2 0.0914i
and cannot be simultaneously diagonalized by any linear trans-
formation. Complex eigenvalues of the system are obtained
(71a)
from the characteristic equation. The left and right eigenvec-

H J
20.2312 1 0.5273i
b̂ (2) =
0.2099 1 0.2429i
tors of the nonconservative system are expressed as a complex
combination of the left and right eigenvectors of the corre-
(71b)
sponding conservative system. The vectors of these complex

H J
constants for both eigenvectors are further determined from a
0.0872 1 0.2913i series obtained by the Neumann expansion method in con-
b̂ (3) = (71c)
20.0653 2 0.2575i junction with a Galerkin-type error minimization. A useful suf-
Step 5: Using the assumed expansion of u i and vi in (9), ficient condition for convergence of this series is established.
and also noting that â (i) and b¯ (i) are vectors of a (i)
j and bj , we
(i) Transfer functions of the system considered are derived as a
finally obtain the right and left eigenvectors of nonconserva- series involving the left and right eigenvectors and complex
tive system natural frequencies of the nonconservative system. The method
described is quite similar to the classical modal analysis where
U = [u1, u 2, u 3] undamped natural frequencies and modes must be appropri-

F G
1.0162 1 0.0819i 20.7395 1 0.1175i 0.6187 1 0.5750i
= 0.0665 2 0.1357i 0.9531 2 0.0994i 20.7821 2 0.2688i
0.0310 2 0.0527i 20.0062 1 0.1580i 0.8802 1 0.0060i
(72)
V = [v1, v2, v3]
ately replaced by complex natural frequencies and left and
right eigenvectors of the nonconservative system. Compared
with the state-space approach, the procedure outlined herein
offers a reduction in computational effort and provides more
physical insight. The expressions developed for the eigenvec-
tors and transfer functions in this paper are very general in

F G
0.3426 2 0.0505i 0.5248 1 0.2880i 0.5117 2 0.0358i
= 0.4676 1 0.0709i 0.0164 1 0.0714i 20.7731 1 0.0691i
0.9109 2 0.0475i 21.0789 1 0.6174i 0.7347 1 0.5180i
(73)
It is useful to check the accuracy of these quantities against
the exact ones obtained from the state-space method. The exact
right and left eigenvectors of nonconservative system are ob-
tained as
Uexact
nature, and most of the familiar linear dynamic systems, for
example, undamped/classically damped symmetric systems,
nonclassically damped symmetric systems, undamped asym-
metric systems, damped/undamped gyroscopic systems, etc.,
can be treated as special cases.
APPENDIX I. APPROXIMATE EXPRESSIONS OF
RIGHT AND LEFT EIGENVECTORS
The expressions of the right and left eigenvectors obtained
by taking one term in the series (20) and (25) produce similar

F G
1.0162 1 0.0819i 20.7395 1 0.1175i 0.6187 1 0.5750i
= 0.0665 2 0.1357i 0.9530 2 0.0994i 20.7821 2 0.2688i
0.0310 2 0.0527i 20.0063 1 0.1582i 0.8802 1 0.0060i
(74)
expressions as those obtained from the first-order perturbation
analysis. The validity of this analysis relies on the fact that the
entries of the C matrix are not very large. Considering the
ith set of (11), and neglecting the second-order terms involv-
ing a (i) ij ;j ≠ i, and also noting that a i = 1, one obtains
j and C9,
(i)

Vexact
2l2i 1 ili C9ii 1 v2i = 0, li ' 6vi 1 iC9/2

F G
0.3426 2 0.0505i 0.5248 1 0.2880i 0.5117 2 0.0358i
= 0.4676 1 0.0709i 0.0162 1 0.0713i 20.7731 1 0.0691i
0.9109 2 0.0475i 21.0789 1 0.6182i 0.7347 1 0.5180i
(75)
For comparison, the exact eigenvectors are normalized to
1378 / JOURNAL OF ENGINEERING MECHANICS / DECEMBER 1999
or ii (76)
This is the first-order approximate expression for the complex
natural frequencies of system (1). To obtain an approximate
expression of the right and left eigenvectors, one simply con-
siders only the first term of the series (20) and (25) and sub-
stitutes a (i) and b(i) in (9) to obtain

J. Eng. Mech. 1999.125:1372-1379.

 O
N
ili C9ki x u k Ahmadian, M., and Inman, D. J. (1984). ‘‘Classical normal modes in
ui ' xui 1 (77a) asymmetric nonconservative dynamic system.’’ AIAA, 22, 1012–1015.
k =1 (li 2 lk)(li 1 l*
k ) Caughey, T. K., and Ma, F. (1993). ‘‘Complex modes and solvability of
k≠i
non-classical linear systems.’’ J. Appl. Mech., 60, 26–28.

O
N Caughey, T. K., and O’Kelly, M. E. J. (1965). ‘‘Classical normal modes
ili C9ik x v k
vi ' x v i 1 (77b) in damped linear dynamic system. J. Appl. Mech., 32, 583–588.
k =1 (li 2 lk)(li 1 l*
k ) Cronin, D. L. (1990). ‘‘Eigenvalue and eigenvector determination for non-
k≠i classically damped dynamic systems.’’ Comp. and Struct., 36(1), 133–
138.
In deriving the above expression, using an approximate ex- Fawzy, I., and Bishop, R. E. D. (1976). ‘‘On the dynamics of linear non-
pression of the complex frequencies in (77), we have factor- conservative systems.’’ Proc., Royal Soc., London, Series A, 352, 25–
ized the denominator as v2k 2 l2i 1 ili C9kk ' 2(li 2 lk)(li 1 40.
lk*). Using this factorization and retaining the first two terms Huseyin, K. (1978). Vibration and stability of multiple parameter systems.
of the series expression (20) and (25), and substituting a (i) and Sithoff and Noordhoff International Publishers, Alpen aan den Rijn,
b(i) in (9), one obtains The Netherlands.
Inman, D. J. (1983). ‘‘Dynamics of asymmetric non-conservative system.

O
N
ili C9ki x u k J. Appl. Mech., 50, 199–203.
ui ' xui 1
Downloaded from ascelibrary.org by Gazi Universitesi on 02/02/15. Copyright ASCE. For personal use only; all rights reserved.

Lancaster, P. (1966). Lambda-matrices and vibrating system. Pergamon,

k =1 (li 2 lk)(li 1 l*
k ) Tarrytown, N.Y.
k≠i
Ma, F., and Caughey, T. K. (1995). ‘‘Analysis of linear conservative vi-

OO
N N bration.’’ J. Appl. Mech., 62, 685–691.
li2 C9kj C9x
ji u k
2 Malone, D. P., Cronin, D. L., and Randolph, T. W. (1997). ‘‘Eigenvalue
k =1 j =1 (li 2 lk)(li 1 l* k )(l i 2 lj)(li 1 l*
j ) and eigenvector determination for damped gyroscopic system.’’ J.
k≠i j≠i≠k (78a) Appl. Mech., 64, 710–712.

O
N Meirovitch, L. (1967). Analytical methods in vibrations. Macmillan, New
ili C9ik x v k York.
vi ' x v i 1
k =1 (li 2 lk)(li 1 l* k )
Meirovitch, L. (1980). Computational methods in structural dynamics.
k≠i Sithoff and Noordhoff International Publishers, Alpen aan den Rijn,

OO
The Netherlands.
N N
li2 C9jk C9x
ij u k Meirovitch, L., and Ryland, G. (1985). ‘‘A perturbation technique for
2 gyroscopic systems with small internal and external damping.’’ J.
k =1 j =1 (li 2 lk)(li 1 l* k )(li 2 lj)(li 1 l*
j ) Sound and Vibration, 100(3), 393–408.
k≠i j≠i≠k (78b) Newland, D. E. (1989). Mechanical vibration analysis and computation.
The above are second-order approximate expressions for the Longman, Harlow and Wiley, New York.
Press, W. H., Teukolsky, S. A., Vetterling, W. T., and Flannery, B. P.
right and left eigenvectors of the nonconservative system (1).
(1992). Numerical recipes in C. Cambridge University Press, Cam-
bridge, England.
ACKNOWLEDGMENTS Rayleigh. (1897). Theory of Sound, 2nd Ed., reissued 1945, Dover, New
The writer is indebted to Dr. Atanas Popov for his careful reading of York.
the manuscript and valuable comments on this problem. Soom, A., and Kim, C. (1983). ‘‘Roughness-induced dynamics loading
at dry boundary-lubricated sliding contacts.’’ J. Lubrication Technol.,
105, 514–517.
APPENDIX II. REFERENCES Vigneron, F. R. (1986). ‘‘A natural modes model and modal identities for
Adhikari, S., and Manohar, C. S. (1999). ‘‘Dynamic analysis of framed damped linear structures.’’ J. Appl. Mech., 53, 33–38.
structures with statistical uncertainties.’’ Int. J. Numer. Methods in Wilkinson, J. H. (1965). The algebraic eigenvalue problem. Clarendon,
Engrg., 44(8), 1157–1178. Oxford, England.

JOURNAL OF ENGINEERING MECHANICS / DECEMBER 1999 / 1379

J. Eng. Mech. 1999.125:1372-1379.