Structural Optimization 8, 207-227 (~) Springer-Verlag 1994

Review Paper
Multiple eigenvalues in structural optimization problems*
A.P. S e y r a n i a n t , E. L u n d a n d N. Olhoff
Institute of Mechanical Engineering, Aalborg University, DK-9220 Aalborg, Denmark

Abstract This paper discusses characteristic features and in- in a recent paper by Cox (1992). Olhoff and Rasmussen
herent difficulties pertaining to the lack of usual differentiability (1977) showed that the bimodality of the optimum eigen-
properties in problems "of sensitivity analysis and optimum struc- value must be taken into account in the mathematical for-
tural design with respect to multiple eigenvaiues. Computational mulation of the problem in order to obtain the correct op-
aspects are illustrated via a number of examples. timum solution. They first demonstrated that an analytical
Based on a mathematical perturbation technique, a general solution obtained earlier by Tadjbakhsh and Keller (1962)
multiparameter framework is developed for computation of design
under the tacit assumption of a simple buckling load is not
sensitivities of simple as well as multiple eigenvalues of complex
structures. The method is exemplified by computation of changes optimal, then presented a bimodal formulation of the prob-
of simple and multiple natural transverse vibration frequencies lem, solved it numerically, and obtained the correct optimum
subject to changes of different design parameters of finite element design. The optimum bimodal buckling load obtained was
modelled, stiffener reinforced thin elastic plates. later confirmed to be correct to within a slight deviation of
Problems of optimization are formulated as the maximization the sixth digit# by analytical solutions obtained indepen-
of the smallest (simple or multiple) eigenvalue subject to a global dently by Seyranian (1983, 1984) and Masur (1984). The
constraint of e.g. given total volume of material of the structure, discovery in 1977 of multiple optimum eigenvalues in struc-
and necessary optimality conditions are derived for an arbitrary tural optimization problems, and the necessity of applying a
degree of multiplicity of the smallest eigenvalue. The necessary op- bi- or multimodal formulation in such cases, opened a new
timality conditions express (i) linear dependence of a set of gener- field for theoretical investigations and development of meth-
alized gradient vectors of the multiple eigenvalue and the gradient
ods of numerical analysis and solution.
vector of the constraint, and (ii) positive semi-definiteness of a
matrix of the coefficients of the linear combination. Prager and Prager (1979) and Choi and ttaug (1981) pre-
It is shown in the paper that the optimality condition (i) can sented unimodal and bimodal optimum solutions for systems
be directly applied for the development of an efficient, iterative with few degrees of freedom, confirming the appearance of
numerical method for the optimization of structural eigenvalues multiple eigenvalues in optimization problems. A wealth of
of arbitrary multiplicity, and that the satisfaction of the necessary references on multimodal optimization problems and specific
optimality conditioli (ii) can be readily checked when the method results for columns, arches, plates and shells can be found in
has converged. Application of the method is illustrated by sim- comprehensive surveys by Olhoff and Taylor (1983), Gajew-
ple, multiparameter examples of optimizing single and bimodal ski and Zyczkowski (1988), Zyczkowski (1989) and Gajewski
buckling loads of columns on elastic foundations. (1990). A survey of other problems of optimum design with
respect to structural eigenvalues was earlier published by O1-
hoff (1980).
1 Introduction One of the main problems related to multiple eigenvalues
is their non-differentiability in the common (Frdchet) sense.
Multiple eigenvalues in the form of buckling loads and natu-
This was revealed by Masur and Mroz (1979, 1980) and Haug
ral frequencies of vibration very often occur in complex struc-
and Rousselet (1980). The non-differentiability creates dif-
tures that depend on many design parameters and have many
ficulties in finding sensitivities of multiple eigenvalues with
degrees of freedom. For example, stiffener-reinforced thin-
respect to design changes and derivation of necessary opti-
walled plate and shell structures have a dense spectrum of
mality conditions in optimization problems. Choi and Haug
eigenvalues, and multiple eigenvalues are found very often.
(1981) used a Lagrange multiplier method for bimodal prob-
Also, symmetry of structural systems may lead to the ap-
lems and showed that this method, which is very useful for
pearance of several linearly independent buckling modes and
vibration modes with multiple eigenvalues. differentiable criteria and constraints, may yield incorrect re-
sults.
In 1977, Olhoff and Rasmussen discovered that the opti-
mum eigenvalue of a clamped-clamped column of given vol- Haug and Rousselet (1980) proved the existence of direc-
ume is bimodal. This optimization problem was first consid- tional derivatives of multiple eigenvalues and obtained ex-
ered by Lagrange, and its interesting history was presented plicit formulae for derivatives. Bratus and Seyranian (1983)
and Seyranian (1987) presented sensitivity analysis of mul-
*Dedicated to the memory of Ernest F. Masur
t G u e s t professor during the period 16 November to 11 December) # T h i s accuracy of the numerical result in the paper by OLhoff and
1992 and 15 November to 12 December, 1993. Permanent address: In- Rasmussen (1977) was obtained by solving the problem for different
stitute of Mechanics, Moscow State Lomonosov University, Michurinsky values of the mesh length d, and extrapolating the result to d -- 0 by
pr. 1, Moscow 117192, Russia. means of Newton's forraula.
208

tiple eigenvalues based on a perturbation technique and de- and accurate, and that it constitutes an excellent basis for
rived necessary optimality conditions. The main advantage the solution of structural optimization problems.
of these necessary optimality conditions is that, when com- Sections 3, 4 and 5 are devoted to the optimization of mul-
pared with those obtained by previous researchers, they do tiple eigenvalues. Section 3 contains three simple examples
not contain variations of design variables. Similar develop- that illustrate the main ideas. Then, in Section 4, we formu-
ments were presented by Masur (1984, 1985). It was with late the structural eigenvalue optimization problem as the
the use of these necessary optimality conditions that Seyra- maximization of the smallest eigenvalue and derive the nec-
nian (1983, 1984) and Masur (1984) independently of each essary optimafity conditions, first for bimodal optimization,
other obtained the analytical solution to the bimodal opti- and then for the general case where the smallest eigenvalue
mum clamped-clamped column problem mentioned above. has arbitrary multiplicity. The optimality conditions express
Overton (1988) considered the minimization of the max- (i) linear dependence between a set of generalized gradient
imum eigenvalue of a symmetric matrix. This problem is vectors of the multiple eigenvalue and the gradient vector of
similar to the problem of maximizing the minimum eigen- the constraint condition (given total volume of structural ma-
value. Derivation of necessaTy optimality conditions using terial), and (ii) positive semi-definiteness of a matrix of the
a so-called bound formulation of such problems had earlier coefficients of the linear combination.
been presented by Bendsee et al. (1983) and Taylor and Section 5 shows that these optimality conditions are very
Bendsee (1984). In a recent paper Cox and Overton (1992) useful for the construction of an iterative numerical proce-
presented new mathematical results for optimization prob- dure for optimization of structural eigenvalues of arbitrary
lems of columns against buckling. They derived necessary multiplicity. Thus, we develop a numerical method of solu-
optimality conditions using advanced nonsmooth optimiza- tion in which the condition (i) is directly used to determine an
tion methods. Their results for optimum columns are in ascent direction in the design space for the smallest (simple
good agreement with the results obtained earlier by Olhoff or multiple) eigenvalue, and in which the satisfaction of (ii)
and Rasmussen (1977), Seyranian (1983, 1984), and Masur can be readily checked when the method has converged. The
(1984). application and efficiency of this method of optimization of
Numerical algorithms for the solution of structural opti- simple or multiple fundamental eigenvalues is demonstrated
mization problems with multiple eigenvalues have been sug- for examples of the optimum design of the columns on elastic
gested and discussed by, among others, Olhoff and Rasmussen foundations.
(1977), Choi et al. (1982), Olhoff and Plaut (1983), Bendsoe
el al. (1983), Myslinski and Sokolowski (1985), Zhong and 2 Sensitivity analysis for eigenvalue p r o b l e m s
Cheng (1986), Plaut et al. (1986), Gajewski and Zyczkowski
(1988), Overton (1988), and Cox and Overton (1992). In this section we seek general formulations for sensitivity
The present paper is devoted to the development of effi- analysis of finite element discretized structural eigenvalue
cient methods for design sensitivity analysis and optimization problems. These problems appear in structural vibration and
of simple as well as multiple eigenvalues of complex struc- stability ane]~.~es. For conservative systems without damp-
tures. Section 2 presents the mathematical basis of the de- ing these problems lead to real eigenvalues representing free
velopment which is a perturbation technique that allows us vibration frequencies or buckling loads. We will confine the
to obtain sensitivities of both multiple eigenvalues and cor- analysis to seeking sensitivities of eigenvalues only.
responding eigenvectors. We first consider the case where
only a single design parameter is changed and then present a 2.1 Introduction
method for efficient calculation of design sensitivities of sim- The finite element formulation for the real, symmetric, struc-
ple and multiple eigenvalues when all design parameters are tural eigenvalue problem is
changed simultaneously.
K~)j = ) ~ j M e j , j = 1,... , n , (1)
It is shown that the design sensitivities of multiple eigen-
values of finite element modelled structures can be computed where K and M are symmetric positive definite matrices, ,~j
within the framework of the semi-analytical approach for sen- is the eigenvalue and e j is the corresponding eigenvector.
sitivity analysis. This approach can be easily augmented to The dimension of the problem is denoted by n, so (1) has
yield "exact" numerical sensitivities by a new technique pub- n solutions consisting of eigenvalues Aj and corresponding
lished by Olhoff el al. (1993) and Lund and Olhoff (1993). eigenvectors e j . The eigenvalues are all real and can be
By this technique, which is computationally inexpensive and ordered in the following manner after magnitude:
very easy to implement as an integral part of the finite el- 0 < ~1 -< ~2 < ...-< ~j _<... < ' ~ n . (2)
ement analysis, we completely avoid the often severely er-
In the following it is assumed that the eigenvectors have been
roneous shape design sensitivities of beam, plate, and shell
structures that may result from application of the tradi- M-orthonormalized, i.e.
tional method of semi-analytical sensitivity analysis (see e.g. ¢TM¢k=(ijk , j,k= 1,...,n, (3)
Barthelemy and Haftka 1988).
where ~jk denotes Kronecker's delta.
The efficiency of the new method of sensitivity analysis
of multiple eigenvalues is demonstrated and illustrated by If we premultiply (1) by e T we obtain
studies of problems of changes of different types of design
parameters for free, transversely vibrating stiffener reinforced tTKtk =~j~jk, J,k= 1,...,n, (4)
elastic plates. It is found that the method is very reliable meaning that the eigenvectors are also K-orthogonal.
 209

So far we have not mentioned multiple eigenvalues and These finite difference approximations can be upgraded t o
eigenvectors. In this case the eigenvectors are not unique. In "exact" numerical derivatives (exact except for round-off er-
fact, an infinite number of linear combinations of the eigen- rors) by using the method of exact semi-analytical sensitivity
vectors corresponding to the repeated eigenvalue will satisfy analysis (see Olhoff et al. 1993; Lund and Olhoff 1993).
(1) aad (3). However, we can choose a set of M-orthonormal If all the design variables a i are changed simultaneously
eigenvectors which span the subspace that corresponds to a then we can find the linear increment of the simple eigenvalue
multilS!e eigenvalue. In other words, if we assume that ,~j has ~j in the form
multlpllcity N (i.e. ~j = ,kj+ 1 . . . . . )~j+N-1), then we
can choose N eigenvectors ~bj which span the N-dimensional ,,As = oK
subspaee corresponding to the eigenvalues of magnitude ~j i=1 -~a/ J~a/J CjAai" (11)
and hatisfy the orthogonality conditions in (3) and (4). This equation is valid due to the differentiability of simple
eigenvalues with respect to design variables.
2.2 DeSign sensitivity analysis of simple eigenvalues The expression obtained in (11) ean be written in the form
We assiitne that the shape and size of the structure are gov- of the scalar product
erned by a set of design variables aj, i = 1 , . . . , I and the A)~i = v T ) ~ j A a , (12)
goal is to obtain expressions for eigenvalue sensitivities with
respect to these design variables. It is also assumed that the where V)tj denotes the gradient vector of Aj and A a is the
components of the K and M matrices are smooth functions vector of changes of the design variables a i
of design variables a i.
v: s = " Oa)] ' Aa= (,aaa,...,Aal). (lZ)
The direct approach to obtaining the eigenvalue sensitiv-
ities is to differentiate (1) with respect to a design variable These notations are useful for parametric studies of eigenval-
aj, Assuming that ,~j is simple, we have ues as well as for formulation of optimization problems.
~aidpj + (g . Oq~j O~j OM 2.3 Design sensitivity analysis of multiple eigenvalues
~jM) Oai = Oai MCj + )~j~aiCj ,
When the solution of the generalized eigenvalue problem in
i = 1,...,I. (5) (1) yields an N-fold multiple eigenvalue
By premultiplying (5) by ¢ T and making use of (1) and the =,Xj, j = l .... ,N~ (14)
normalization of (3), the following expression is obtained for where, for convenience, the repeated eigenvalues have been
the eigenvalue sensitivity in the case of simple eigenvalues ,~j numbered from 1 to N, the computation of the sensitivi-
(see e.g. Courant and Hilbert 1953; Wittrick 1962): ties of this eigenvalue is not straightforward. This is due

Oai
(oK ~a i j Oa---T]dpj, i= l,...,I. (6)
to the fact that the eigenvectors ~bj, j = 1 , . . . , N of the
repeated eigenvalues are not unique. Thus, any linear combi-
nation of the eigenvectors will satisfy the original eigenvalue
The most important point here is that the only unknown
problem (1). We assume that the eigenvectors ~bj have been
quantities in (6) are the derivatives of the K and M matri-
M-orthonormalized, i.e.
ces. These derivatives are normally calculated at the element
level, i.e. cTMfk=Sjk , j,k=l,...,N. (15)
OK Ok In the following sensitivity analysis we shall use such eigen-
-~ai = ~n ~ai , i = l , . . . , I , (7) vectors ¢ j that remain continuous with design changes, see
Courant and Hilbert (1953). For this purpose we introduce
oM = E , i = 1 , . . . , I, (8) linear combinations of eigenvectors ~bk
Oai ne N
where k and m are element matrices and ne is the number $J = E t 3 J kekk' j = 1 , . . . , N , (16)
of finite elements. k=l
These element derivatives Can be either calculated ana- where fljk are unknown coefficients to be determined.
lytically, if possible, or by using first-order finite difference Works by Courant and Hilbert (1953), Wittrick (1962),
approximations as known from the method of semi-analytical and Lancaster (1964) have provided a basis for calculating
sensitivity analysis of finite element discretized structures the sensitivities of multiple eigenvalues. It is shown that the
with linearly elastic, static response, i.e. design sensitivities of multiple eigenvalues can be found by
formulation and solution of a subeigenvalue problem.
0k(a) ,~ Atk(al,... ,ai)
Let us first consider a change Aa i of one arbitrarily chosen
Oai -- Aa i design parameter a i. Due to the change of this parameter the
k ( a l , . . . , a i + A a i , . . . , a i ) -- k ( a l , . . . , a i , . . . , a I ) K and M matrices will be incremented, i.e. the new matrices
, (9) become
Aa i
OK OM Aa"
0m(a) ZXm(al . . . . . a I ) _ K+__-~a Aai and M + Oai z, i = l , . . . , I . (17)
Oai -- Aa i Then multiple eigenvalues and corresponding eigenvectors for
m(al,...,ai+Aai .... ,ai)--m(al,...,ai,...,ai) the perturbed design can be written as
zaai (lO) )~j(a i + eAai) = ~ + et2j(ai, Aai) + o(e), j = 1 , . . . , N ,
210

~bj(ai+¢Aai) = e j + e v j ( a i , Aai)+o(¢), j = 1,..., N , (18) As a result of perturbation of the vector a, the matrices
where ]~j and vj are unknown eigenvalue and eigenvector K and M are incremented and become
sensitivities, respectively. I OK I 0M
Substituting (17) and (18) into the main eigenvalue prob- K+¢Z--ei, M + ¢ ~-'~--ei. (23)
lem in (1), we obtain in the first approximation i=1 cgai i=1 cgai
Using expansions for Aj and ~bj in the form
~ a / ] ~1 + ( g - k M ) v / = # j M ~ j . (19)
Aj = k + c~j + o(~),
Premultiplying this equation by e T gives
(OK _ xSM ~ eJ = ~1 +¢~'j + o(e), j = 1,...,g, (24)
and performing the same manipulations as earlier, instead of
(22) we obtain the following N-th order equation for deter-
s = 1.... , N . (20) mining the sensitivities # = # j , j = 1 , . . . , N of the eigen-
Here the term e T ( K - AM)vj = v T ( K - AM)¢s drops out values Aj:
because gbs is an eigenvector corresponding to A.
Recalling that e j is the linear combination in (16) of the det ¢~ ~ Oai 2 Ckei-~sk = O,
original eigenvectors ek, we obtain from (20) the system of /=1 .

linear algebraic equations of unknown coefficients ~jk

s,k = 1 , . . . , N . (25)
If we introduce the generalized gradient vectors fsk of dimen-
njk ek -#j sk : 0,
k=l sion I

s--1,...,N, (21) fsk :

where the M-orthonormalization (15) has been used.
A nontrivial solution to these equations only exists if the Oai '
determinant of the system is equal to zero then (25) can be written in the form
det[+T(~ai-AO~ai)+k-#'sk] =O, det [f T e - i ~ s k ] =0, s,k=l,...,N. (27)

s,k = l , . . . , N , i = l,...,I. (22) Note that fsk = fks due to the symmetry of the matrices
K and M. Note also that although equipped with two sub-
This is the main equation for determining the coefficients scripts, the generalized gradients fsk are vectors (of dimen-
# j , j = 1,..., N in (18) which represen t the sensitivities of
sion I); the two subscripts refer to the modes from whith the
the multiple eigenvalue A with respect to changes Aai of a generalized gradient vector is calculated. Thus, in (27) fsTe,
single design parameter a i. As in the case of simple eigenval- s, k = 1 , . . . , N, denotes scalar products.
ues, the derivatives of the K and M matrices, respectively, Thus, knowing the eigenvectors e k , k = 1,... , N corre-
must be calculated first, and then the eigenvalue problem of
sponding to the multiple eigenvalue A, we can construct the
(22) is easily formulated and solved.
generalized gradient vectors fsk and determine the sensitivi-
If the off-diagonal terms in the quadratic matrix of dimen-
ties # = # j , j = 1,..., N for any vector of variation e, i.e.
sion N in (22) are equal to zero, the eigenvalues of this ma-
for any direction in the space of the design variables. The
trix, i.e. the directional derivatives of the multiple eigenvalue
quantities #j = A A j / A ~ depend on e and constitute the di-
A, are equal to the traditional Fr~chet derivatives obtained
by using (6). rectional derivatives of the multiple eigenvalue A, cf. (24). In
Let us consider the general case when all the design vari- this form (27) was obtained by Bratus and Seyranian (1983),
ables ai, i = 1,..., I are changed simultaneously. It should and Seyranian (1987), see also Haug and Rousselet (1980),
be noted that multiple eigenvalues are not differentiable in Masur (1984, 1985), and Haug et al. (1986).
the common sense, i.e. not Fr~chet-differentiable (see e.g. In many cases it is expedient to eliminate the unit vector
Haug et al. 1986). This means that the expressions for the e from (27) and establish a formula for determining the incre-
eigenvalue increments in (11) and (12) are no longer valid. ments AAj, j = 1 , . . . , N of the N-fold eigenvalue A subject
Thus, to find the sensitivities of multiple eigenvalues we must to a given vector A a = ( A l l , . . . , A a I ) of actual increments
use directional derivatives in the design space. of the design variables ai, i = 1 , . . . , I. To this end, we mul-
For this purpose, for the vector of design variables a = tiply each of the components in (27) by ¢, note from the
( a l , . . . , ai) , we consider a variation in the form a + e e , where foregoing that ¢e - A a and ettj : AAj, j = 1,... ,N, and
e is an arbitrary vector of variation e = ( e l , . . . ,eI) with obtain
the unit norm lie N = q e2 + ' " + c~ = 1 and ¢ is a small det [ f T A a - - 5 s k A A ] : 0 , s,k=l,...,N. (28)
positive parameter. The vector e represents a direction in the If we solve this N-th order algebraic equation for A%, we
design space along which the design variables a i are changed, obtain the increments AA = AAj, j = 1,..., N of the N-fold
and e represents the magnitude of the perturbation in this eigenvalue corresponding to the vector A a of increments of
direction. the design variables.
 211

2.4 Numerical example 1: design sensitivity analysis of a

vibrating plate with ribs
To illustrate how the results of Sections 2.2 and 2.3 can be
used in eigenfrequency design sensitivity analysis, we consider
a square plate as shown in Fig. 1.

Fig. 2. First eigenmode, fl = 92.20 Hz

Fig. 3. Second eigeumode, f2 = 161.71 Hz

I :1!ha j

~~__vertical ~ ~ horizontal Fig. 4. Third eigenmode, f3 = 161.71 Hz

ribs ribs
Fig. 1. Square plate with ribs

The plate is clamped at all edges and reinforced by two

horizontal and two vertical ribs. The length a = 0.5 m,
b = 0.05 m, the thickness of the ribs is 0.05 m, and the
thickness of the plate is 0.005 m. Both the plate and the ribs
are made of steel with the following material properties:
Young's modulus = 210000 MPa Fig. 5. Fourth eigeumode, f4 = 175.03 Hz
Poisson's ratio = 0.3
mass density = 7800 k g / m 3 satisfy the general eigenvalue problem in (1) and the M-
The structure is symmetric and therefore multiple eigen- orthonormalization condition in (3).
frequencies are expected. We want to find sensitivities of both The design sensitivities with respect to changes of single
the simple and the multiple eigenfrequencies with respect to design variables will be computed by two different methods,
6 different design variables. namely the overall finite difference (OFD) method and the
The finite element model consists of 1156 4-node isopara- semi-analytical (SA) method using (6) and (22). The OFD
metric Mindlin plate elements and the model has 5445 d.o.f. method implies that the design is perturbed, a new eigen-
The lowest eigenfrequencies fj are found by the subspace frequency analysis performed, and that the eigenfrequency
iteration method (see Bathe 1982). All eigenvectors are M- sensitivities pj are then found by
orthonormalized, see (3), and we will consider the 4 lowest Afj(al,...,aI)
eigenfrequencies. From the analysis it is found that the low- #j ~-- Aa i
est eigenfrequency is simple and has the value 92.20 Hz, but
the second and third eigenfrequeneies are identical and have f j ( a l , . . . , a i + A a i , . . . , a I ) -- f j ( a l , . . . , a i , . . . , a I )
the value 161.71 Hz. In this example, the eigenfrequencies Aa i
are considered to be identical when the relative difference The OFD method is used as a reference method whose
between the values is < 10 -4. The fourth eigenfrequency is limit with regard to design sensitivity accuracy is known to
simple and equal to 175.03 Hz. The eigenmodes correspond- be set only by the solution procedure, the discretization and
ing to the 4 eigenfrequencies can be seen in Figs. 2-5, and the the usual accuracy capabilities of the applied finite element,
influence of the ribs is very clear. when the design perturbation Aa i is sufficiently small.
AS the second eigenfrequency is multiple, an infinite num- In the usual SA method, the derivatives of the element
ber of linear combinations of the eigenvectors shown in Figs. stiffness and mass matrices [see (7), (8), (9) and (10)] are
3 and 4 corresponding to the multiple eigenfrequency will calculated by simple finite difference approximations. It has
 212

been shown by Barthelemy and Haftka (1988) that the usual will increase with increasing plate thickness.
SA method is prone to severe inaccuracy problems when For this design change, the erroneous assumption of using
used for calculation of design sensitivities of beams, plates (6) to the double eigenfrequency (see last column in Table 1)
and shells, so we apply a new, recently developed version in fact gives the same sensitivities as obtained by (22). This
of the SA method (see Olhoff et al. 1993; Lund and Olhoff shows that the influence of the off-diagonal terms f T e =
1993). In this new method, the element matrix derivatives
f T e in the sensitivity matrix in (22) is very weak for this
are upgraded to "exact" numerical derivatives (exact up to
symmetric design change.
round-off errors). Then the eigenfrequency sensitivities #j
Next, the eigenfrequency sensitivities are found with re-
are determined by using (6) for the simple eigenfrequencies
spect to the thickness of the ribs as shown in Fig. 7. The
f l , f4 and using (22) for the multiple eigenfrequ¢ncy f2 = f3 results are shown in Table 2.
as we consider the case of change of single design variables.
Furthermore, we shall calculate the sensitivities of all four
eigenfrequencies regarding them as simple in order to see the
consequences of this erroneous assumption.
The first design sensitivity analysis is with respect to the
thickness of the plate as shown in Fig. 6.
Fig. 7. Design variable 2: rib thickness

Table 2. Eigenfrequency sensitivities with respect to design vari-

able 2: rib thickness
Frequency OFD method (6) and (22) (6)
Fig. 6. Design variable 1: plate thickness
J Aa 2
The sensitivities are shown in Table 1 and it is seen that
1878.6 1878.6 1878.6
the same results are obtained by the OFD method and the
1714.2 1714.2 1714.2
SA method using (6) and (22).
1714.2 1714.2 1714.2
Table 1. Eigenfrequency sensitivities with respect to design vari- 304.7 305.1 305.1
able 1: plate thickness
It is seen in Table 2 that again the same results are ob-
Frequency OFD method (6) and (22) (6) !
tained by the OFD method and the SA method using (6) and
0f.
J (22). All the sensitivities are positive, and again the multiple
-1551.6 -1551.6 -1551.6 eigenfrequency remains multiple with this design change.
14093.2 14094.1 14094.1 Next we want to determine sensitivities of the eigenfre-
14093.2 14094.1 14094.1 queneies when the position of the horizontal ribs is changed.
31407.9 31406.8 31406.8 The design variable is the distance between the horizontal
ribs, see Fig. 8. The results are shown in Table 3.
Increasing the plate thickness is a symmetric design
change and therefore we may expect that the multiple eigen-
frequency remains multiple. Note that while the sensitivities
of the other eigenfrequencies are positive, the sensitivity of
the lowest eigenfrequency is negative, i.e. the first eigenfre-
quency will decrease if the thickness of the plate is increased. Fig. 8. Design variable 3: position of horizontal ribs
These results require some explanation.
The eigenmode corresponding to the lowest eigenfre-
quency can be characterized as a global mode while all higher T a b l e 3. Eigenfrequency sensitivities with respect to design vari-
order eigenmodes can be regarded as nearly-local modes for able 3: position of horizontal ribs
each of the nine subdomains of the plate, see Figs. 2-5. Thus, Frequency OFD method (6) and (22) (6)
increasing the plate thickness will have little effect on the asj
overall stiffness of the structure because it is mainly governed J Aa 3 PJ 0as
by the ribs, whereas the thickness of the plate has a large in- -40.9 -40.9 -40.9
fluence on the total mass, i.e. the inertia forces. Therefore, -380.6 -380.6 -287.5
increasing the plate thickness will decrease the lowest eigen- 186.9 186.9 93.7
frequency which mainly depends on the overall stiffness and -169.1 -168.6 -168.6
total mass of the structure.
The higher order eigenmodes, on the other hand, mainly It is seen from Table 3 that the multiple eigenfrequency
depend on the local stiffness and mass of each of the nine ] = f2 -- f3 will split when the distance between the hori-
subdomains, and increase of the plate thickness has a larger zontal ribs is increased. Now we can also see differences be-
effect on the local stiffness than on the local mass of each sub- tween the two last columns in Table 3 showing the influence
domain. This is the reason why the higher eigenfrequencies of off-diagonal terms in (22), which implies that calculation
 213

of sensitivities of the double eigenfrequency by means of the as simple leads to completely wrong results for this design
single-modal formula in (6) gives erroneous results. change.
Next the distance between the vertical ribs is used as a de- The last design variable is the width of the vertical ribs
sign variable as shown in Fig. 9, and the results ar e presented as shown in Fig. 11.
in Table 4.

Fig. 11. Design variable 6: width of vertical ribs

Fig. 9. Design variable 4: position of vertical ribs
Table 6. Eigenfrequency sensitivities with respect to design vari-
Table 4. Eigenfrequency sensitivities with respect to design vari- able 6: width of vertical ribs
able 4: position of vertical ribs Frequency OFD method (6) and (22) (6)
Frequency OFD method (6) and (22) (6) ass 0I.
J Aa6
J Aa4 PJ Oa4 -273.4 -273.4 -273.4
-40.9 -40.9 -40.9 -866.2 26.4 -120.1
-380.6 186.9 93.7 26.4 -866.2 -719.7
186.9 -380.6 -287.5 -399.5 -399.8 -399.8
-166.0 -167.0 -167.0
As before, the sensitivities with respect to this design vari-
The sensitivities for this design variable should be the able should be the same as those obtained for design variable
same as those obtained for design variable 3, the distance 5, the width of the horizontal ribs, except that the sensi-
between the horizontal ribs, except that the sensitivities for tivities for the multiple eigenfrequency f2 = f3 should be
the multiple eigenfrequency f2 = f3 should be interchanged interchanged when using the SA method and (6) and (22).
when using (22). It is seen that the OFD method does not Again it is seen that very similar results are obtained by
display this situation because the eigenfrequencies are or- the OFD and SA methods when (6) and (22) are used prop-
dered by magnitude in each analysis when using the OFD erly, whereas the last column again witnesses a shortcoming
method. More importantly, we again note that application of (6) when applied to the double eigenfrequency f2 = f3.
of the single-modal equation (6) yields erroneous sensitivities Up to now the eigenfrequency sensitivities have been
of the multiple eigenfrequency f2 = f3" found with respect to single design dhanges of each of the
The next design variable is the width of the horizontal 6 design variables: thickness of plate, thickness of ribs, po-
ribs, see Fig. 10, and the results are shown in Table 5. sition of horizontal ribs, position of vertical ribs, width of
horizontal ribs, and width of vertical ribs.
Let us finally show that it is possible to determine sensi-
tivities for any direction in the space of the 6 design variables
when some of them are changed simultaneously. The posi-
tion of both the horizontal and vertical ribs, see Figs. 8 and 9,
will be changed simultaneously and again the OFD method
Fig. 10. Design variable 5: width of horizontal ribs is used as a reference. The two design variables will be given
unit increments.
All the generalized gradient vectors fsk in (27) have been
Table 5. Eigenfrequency sensitivities with respect to design vari-
calculated and (28) is used for determining the increments
able 5: width of horizontal ribs
A f = A f2 and AS = A f3 of the multiple eigenfrequency
Frequency OFD method (6) and (22) (6) f2 = f3. Equation (12) is used for determining increments of
Af. 0$. the simple eigenfrequencies fl and f4. The sensiti,~ities for
J
this simultaneous design change are shown in Table 7.
1 -273.4 -273.4 -273.4
2 -866.2 -866.2 -719.7 Table 7. Eigenfrequency sensitivities for unit increments of de-
3 26.4 26.4 -120.1 sign variable 3: position of horizontal ribs and design variable 4:
4 -401.4 -400.9 -400.9 )osition of vertical ribs
Frequency OFD method (12) and (28)
Again the results obtained by the OFD method and the
SA method using (6) and (22) are very similar, and it is j asj
seen that the multiple eigenfrequency splits with this design 1 -81.7 -81.7
change. Furthermore, the sensitivities of the double eigenfre- 2 -193.8 -193.8
quency obtained by using (6) even have a wrong sign, so using 3 -193.8 -193.8
the erroneous assumption of regarding the eigenfrequencies 4 -335.5 -335.5
214

It is seen that very accurate results are obtained by using

(12) and (28) for determining sensitivities of single and bi-
modal eigenfrequencies, respectively, in any direction in the
space of design parameters.

2.5 Numerical example 2: ribbed plate with a cluster of

eigenfrequencies Fig. 15. Fourth eigenmode, f4 = 72.24689 Hz
Next we shall illustrate how important it is for sensitiv-
ity analysis to decide correctly from the numerical results
whether some eigenvalues coalesce and become multiple or
remain distinct.
We consider the same example as before, i.e. the square
plate with ribs shown in Fig. 1, but now the thickness of
the plate is 2 1/2 times less, i.e. 0.0020 m. This causes the
9 lowest eigenfrequencies to become very close because their Fig. 16. Fifth eigenmode, f5 = 72.27588 Hz
corresponding eigenmodes can be regarded as local modes for
each of the nine subdomains of the plate, see Figs. 12-21.
The first eigenfrequency is 70.404 Hz and the second
through the ninth eigenfrequency are close to 72.2 Hz. These
9 eigenfrequencies are so close that it is difficult to decide
which of them are multiple.
If we use as a criterion for identical eigenfrequencies that
the relative difference between the frequencies must be < Fig. 17. Sixth eigenmode, f8 = 72.31243 Hz
10 - 3 , then the 8 eigenfrequeneies from the second to the
ninth should be considered as multiple, but if we use a tighter
criterion such as 10 - 4 then the second and third should be
considered as a double eigenfrequency and the sixth, seventh,
eighth, and ninth should be considered as a 4-fold multiple
eigenfrequency. If the criterion 10 - 5 is used, the second and
third should be considered as a double eigenfrequency, and
similarly with the seventh and eighth. Fig. 18. Seventh eigenmode, f7 = 72.31890 Hz

method which is used as reference, (ii) using (6) and (22)

and assuming two double eigenfrequencies, (iii) using (6) and
(22) and assuming one double and one 4-fold multiple eigen-
frequency, (iv) using (6) and (22) and assuming one 8-fold
multiple eigenfrequency, and (v) only using (6) which is only
valid in cases of simple eigenvalues. The results are shown in
Fig. 12. First eigenmode, fl = 70.40406 ttz Table 8.
It is seen that we obtain the same results by the SA
method assuming either two double eigenfrequencies f2 = f3
and f7 = f8, or one double eigenfrequency f2 = f3 and one
4-fold multiple eigenfrequency f6 = f7 = f8 - f9- These two
columns are in excellent agreement with the OFD method,

Fig. 13. Second eigenmode, f2 = 72.17485 Hz

Fig. 19. Eighth eigenmode, fs = 72.31890 Hz

Fig. 14. Third eigenmode, f3 -- 72.17485 Hz

We will determine sensitivities of the eigenfrequencies

when the width of the horizontal ribs is changed as shown
in Fig. 10. The sensitivities are calculated (i) by the OFD Fig. 20. Ninth eigenmode, f9 = 72.32089 Hz
 215

K=[l+x, 20)
The characteristic equation for this system is
A2 - 2A + 1 - x 2 - y2 = 0,
and
Fig. 21. Tenth eigenmode, flo = 109.96799 Hz A1,2 = 1 ~ ~ . (30)
Table 8. Eigenfrequency sensitivities with respect to design vari- The level curves A = c, where c is a constant, for this function
able5: width of horizontal ribs, see Fig. 10 are described by the equations
Freq. OFD (6) and (22) (6) and (22) (6) and (22) and (6) c=li~x2+y 2,
method f2 = fa and f2 = f 3 and .¢~ = f 3 = f , =
~=~ and
f 6 = f7 = f5 = f s = f r =
fa = f9 fs=f9 z2 + y2 = ( c - 1) 2 .
j ,a/j oyj This surface is a circular cone, see Fig. 22.
Aa~ ttJ #J #J Oa5
Bimodality occurs at x = 0, y = 0, for which we have
1 -247.7 -247.7 -247.7 -247.7 -247.7 A1 = A2 = 1.
2 -270.6 -270.6 -270.6 -290.2 -270.2
It can be seen immediately from (30) and Fig. 22 that the
3 -1.1 -1.1 -1.1 -1.1 -1.5
eigenvalues A are not differentiable at the bimodal point in
4 -145.5 -145.1 -145.1 -44.6 -145.1
5 -174.7 -175.1 -175.1 -289.0 -175.1 the usual (Frdchet) sense. Indeed,
6 -0.9 -0.9 -0.9 -0.9 -0.9 0~1,2 x 0)tl,2 y
7 -20.6 -20.6 -20.6 -1.1 -11.4 Oz - + ~ ' Oy - + ~ "
8 -0.8 -0.8 -0.8 -0.8 -10.0
9 -14.4 -14.5 -14.5 -0.9 -14.5 As z --* 0 and y --+ 0, the two right-hand side expres-
10 -146.5 -146,5 -146.5 -146.5 -146.5 sions become undefined and have no limit. Furthermore,
L'HSpital's rule cannot help either because the derivatives
but it is not a general situation that different choices of mul- of the denominators also tend to zero.
tiplicity lead to the same results. This is also illustrated by
the next column in Table 8 because if we assume having an
8-fold multiple eigenfrequency f2 = f3 = f4 = f5 = f6 =
f7 = f8 = f9, wrong sensitivities are obtained for several of
the eigenfrequencies, e.g. f5 and fg. Finally, if we consider
all eigenfrequencies as simple, i.e. use (6), then the results
.......t
for the multiple eigenfrequencies f2 = f3 and f7 = f8 are
erroneous while the other sensitivities are correct.
This illustrates that the influence of the off-diagonal terms
in the sensitivity matrix in (22) is quite small for this design
change. This is also the reason why the same sensitivities / ................. X
are obtained for f6, fT, fS, and f9 using (22) independently
of whether the assumption f7 = f8 or the assumption f6 =
f7 = f8 = f9 is used. In both cases, the off-diagonal terms Fig. 22. Circular cone surface Ior eigenvalue ,~
in the sensitivity matrix in (22) are small compared to the
diagonal terms for this design change, and therefore the same Now we proceed to sensitivity analysis of the double eigen-
sensitivities are obtained for both assumptions. value A = 1 at x = 0, y = 0, using the results of Section 2.
These results show the importance for the design sensi- Taking the variation ~e we have
tivity analysis of deciding correctly whether an eigenfrequency
x = ~e 1, y = ce 2, ~ + e 2 = 1.
is multiple or simple. Thus, wrong assumptions concerning
the multiplicity of an eigenfrequency may lead to erroneous For the sake of simplicity, we can introduce the angle a
sensitivity results. and write the directional vector e in the form
e I = cosa, e2 = s i n s . (31)
3 Optimization problems for eigenvalues Let us determine the directional derivatives #1, #2 for the
double eigenvalue ~ = 1. The orthonormalized eigenvectors
Before formulating optimization problems for eigenvalues, we
corresponding to A are
consider three simple examples that illustrate the main ideas.
Only examples involving 2 × 2 matrices depending on two q~l=- ( ; ) , ~b2= (01). (32)
design variables x and y will be considered.
Using the expressions in (29) and (32), we obtain the
3.1 Example 1 generalized gradient vectors fsk according to (26)
The first example is described by the following K and M
matrices: flTl=((X0)(~ 71)(:)'(10)(01 10)(10)):(1'0)'
 216

fT= ((10)(~ 0 1 ) (01),(1 0)(01 ~) (~))=(0,1), A1 = 1 + 2 x 1 when y = 0 (41)

A2 = l + x J
The last equation implies that along the direction y = 0,
fT= ((01)(~ O 1 ) (01),(0 1 ) ( ~ ~) (~))=(-1,0). x > 0 we have A1 > A, A2 > A, where A = 1. Based on
Thus, (27) takes the form (38)-(41) we can plot surfaces of the eigenvalues A1,2, see
Fig. 24.
dee [ c o s a - p sins ]
sin a - cos s - # j = 0, (33)
or

#~,2 = sin2 s + cos 2 s = 1. (34)

So, #1 = i and #2 = -1 for any direction e = (cos a, sins).
Hence, the double eigenvalue A splits into A1,2 = 1 4- ¢ for
any variation ce. This means that the bimodal solution x =
0, y = 0 is the optimum solution to the problem
max min
x,y j=l,2
Aj. (35)
/~ ~ i
::, "/'J/
I;
......... / /
This result, of course, can be seen immediately from Fig. 22. ,/ ii
3.2 Example 2 // ~!:
/ , /
Consider another example with the K and M matrices given
by /
/
/

The characteristic equation for this system is Fig. 24. Surface plot of eigenvalue )~
A2 _A(2 + 3x) + (1 + 3 x + 2x 2 - y2) = 0, (37) Let us proceed to sensitivity analysis near the bimodal
and point z = 0, y = 0 on the basis of the results of Section 2. The
2+3x+ V/~ +4y 2 orthonormalized eigenvectors corresponding to A1 = A2 = 1
A,,2 = 2 (38) are
The bimodal solution )t 1 = A2 = 1 occurs at the point z
=O,y=O.
Let us study the level curves. Inserting A = c, where c is According to (26) and with use of (36) and (42), we can
a constant, into (37) we obtain find the vectors fsk
2 2 y2
(x 3 ( c ~ 1 ) ) = ( ~ _ _ 1 ) + - 2 ' fT= ((10)(~ 01)(~),(10)(
10)
0 1

If we denote b = - ~ this equation takes the simple form f 5 = ((1 0) ( ~ 01 ) (01)'(10)(01 01 ) ( 0 1 ) ) = ( 0 , 1 ) ,

( x - 3b) 2 = b2 + ~ , (39)

which defines a hyperbola with the asymptotes y = -t-v~(z -

(,01) 2 0 1
,1,0,
3b) , see Fig. 23. (43)
y Taking the directional vector e in the form e =
(cos s, sin s) and using (43), we obtain from (27) the char-
acteristic equation
[2coss- # sins ]
det / sins cosa-#j =0"
From this equation we obtain
#2 _ 3# cos s + 3 cos 2 a - 1 = 0, (44)
or

Fig. 23. Level curves for eigenvalue A 3 cos s + X/1 + 3 sin 2 s

~,,2 = 2 (45)
If e = 1 then b = 0, and we obtain from (39)
Let us consider the term 3cos 2 s-1 in (44). According to
y = (40) the Vieta theorem, it is equal to the product of the roots of
Near the bimodal point x = 0, y = 0, according to (38) (44)
we have #art2 = 3 cos2 s -- 1. (46)
A1,2-- 14-y when x = O , This quantity can be positive as well as negative depend-
and ing on the value of s. If s = 0 then #lit2 = 2 > 0. This
 217

means that Pl and #2 are of the same sign for the directions 4 Formulation of the optimization problem
e : (1, 0) and e -- (-1, 0), see also Fig. 24.
Consider again the eigenvalue problem
For a = 0 we obtain from (45)
Kd~j = A j M ¢ j , j = l,...,n, (53)
~1 "- 1, ~u2 = 2, (47)
where the components of K and M are smooth functions of
which is in good agreement with (41). the design variables ai, i = 1,..., I.
if ~ = =t=~/2 then it follows from 46 that #1#2 = -1(0. The optimization problem is formulated as follows:
This means that for the directions e -- (0, 1) and e -- (0, -1) max min A- (54)
the deiivatives pl and #2 are of opposite signs, see also Fig. al,...,al j~_l,..., u Y
24. under the constraint
This implies that in contrast to Example 1 the bimodal
F ( a l , . . . , a i ) = 0, (55)
solution ~ -- 0, y = 0 does not constitute the optimum solu-
tion $o the problem where F is a smooth scalar function of the design variables
ai, i = l , . . . , I .
max. min Aj , (48) In mechanical problems this constraint usually reflects a con-
x,y j--l,2
stant volume restriction.
because there exists an improving variation e = (1, 0) for
which #1 > 0, #2 > 0. 4.1 Simple optimum fundamental eigenvalue
If the optimum is achieved at the simple lowest eigenvalue
3.3 Example 3 )~1 with A1 < )~2 -~ .~3 ~- ..., then the necessary optimahty
Let us consider another simple example in which the double condition implies linear dependence of the gradient vectors
eigenvalue does not split along some direction. of AI and F
VA 1 - 70f0 = 0, (56)
where
The characteristic equation for this system leads to VA 1 =
A1,2 -~ 1 + x ± y. (50)
1 )01 ....
Bimodality takes place at y = 0, where the eigenvalues
remain double for arbitrary x.
,...,On i , (57)
and 70 is a (Lagrangian) multiplier to be determined from
(55).

4.2 Example
In this optimization example we consider the numerical ex-
ample from Section 2.4 where the position of the ribs is taken
as a design variable, see Fig. 26. We want to maximize the
lowest eigenfrequency with a constant volume constraint, see
(54) and (55).
All data for the problem are given in Section 2.4. The
iteration history for the 4 lowest eigenfrequencies is shown in
Fig. 27 and the final design is shown in Fig. 28.
X
The distance between the ribs has been reduced from 0.50
m to 0.2835 m. It turns out that the optimum solution is
characterized by a distinct lowest eigenvalue. This is natural
to expect because one-parameter symmetric eigenvalue prob-
lems usually possess simple eigenvalues (see Arnold 1989).
Fig. 25. Surface plot of eigenvalue )~
4.3 Double optimum fundamental eigenvalue
The surfaces for )~ are plotted in Fig. 25. They represent Consider now the case when the optimum is achieved at the
two intersecting planes. double lowest eigenvalue )~1 = $2, where )~1 = )~2 < ~3 -~ ...
If we put a constraint as illustrated in Fig. 25 This is the nondifferentiable case and we must use directional
x 2 + y2 :. 1, (51) derivatives.
Taking the vector of varied design variables in the form
then the bimodal solution x = 1, y = 0 will constitute the a + ee, [[ e [[= 1, according to (27), we obtain the directional
optimum one for the problem derivatives Pl and #2 from
max min )~j, subject to x 2 + y 2 = 1. (52) [ flTle-# -flT2e ] = 0 . (58)
x,y j=1,2 det ~T e fTe_ p
218

The necessary optimality condition for a maximum is

min(/q, P2) - 0, (60)
for any direction e satisfying the condition f i e = 0.
From (58) we see that if we take the direction as - e , then
both #1 and ]~2 will change their signs to the opposite ones.
This means that if for some direction e both derivatives Pl,
/l 2 are negative, then the design point is not a maximum,
since a change in sign of the direction e leads t o / J l > 0,
#2 > 0, i.e. a better design. This means that the necessary
optimality condition in the bimodal case is
#1/~2 _< 0, (61)
!= a II a ktl u
for any admissible variation e, i.e. a variatiori that satisfies
2! !: b --I !~ b
the condition
f T e : o. (65)
The optimality condition in (61) is fundamentally different
from that of the differentiable case due to its nonlinear na-
ture. The condition was first formulated by Masur and Mroz
(1979, 1980).
Using (58) and (59) we can express the necessary opti-
mality condition of (61) in the form
Fig. 26. Design variable for square plate with ribs: position of
horizontal and vertical ribs (fTe)(f2T2e) - (fTe)2 _< 0, (63)
Frequency [Hz] for any arbitrary direction e satisfying the condition in (62).
Let us formulate the following lemma.
190 • ..i. ................ i ..... i ..... i ....... i .... i ....
L e m m a 1. If the vectors f l l , f12, f22, f0 are linearly indepen-
170 dent, then there exists an improving variation e
for which #1 > 0, P2 > 0.
150
Proof. Consider the system of linear algebraic equations of
130 I. eigenfrequency the variables e l , . . . , e I

110
~- 2. elgenfrequency flTle = u0 > 0, f T e = 0,
-~- 3. elgenfrequency
90 . . . - . . . . :
:. . . . . :. . . . . !
.. . . . . !
. . . . . . i
. .... :
~ . . . . ! i
: . . . . . . .
f T e = u0 > 0, f T e = 0. (64)
-'- 4. elgenfrequency
: : :

70 ...:. .... :..... : ..... : ..... : ..... : .......... : .... ,. .... If the vectors fll, f12, t"22, f0 are linearly independent,
then a solution e to the system in (64) exists for arbitrary
values of u 0 and u20. The vector ~ = ~ is then an improving
1 2 3 4 5 6 7 8 9 10
Iteration number
variation since from (58) and (64) we have

Fig. 27. Iteration history pl=fT~= ~ > 0, p2 =f2T2~= ~ >0, (65)

/ / / / / / / / / / / . /

/
which proves the lemma.
Now let us formulate the necessary optimality conditions
/ for the bimodal case.
Z T h e o r e m 1. If the vector of design variables a constitutes the
/I
/ solution of the optimization problem, (54) and
/
/
(55), with the double eigenvalue A1 = A2 < A3 -<
I .... then the vectors f11, f12, t"22, fo are linearly
dependent
/
/
711fll + 2712f12 + 722f22 -- 70fo : O, (66)
/////// ////// with the coefficients 7sk satisfying the inequality
Fig. 28. Optimum design 711722 ~ 722 • (67)

This is a quadratic equation in #. Solving it we obtain for Here it is assumed that the rank of the matrix consisting
any direction e of the vectors f l l , f12, t"22, f0 is equal to 3. Note that the
linear independence of the four vectors mentioned above is
f T e -t- f T e -4- ¢ ( f T e -- fTe)2 + 4(fTe) 2 possible only when the dimension I of the vector of design
/11'2 = 2 (59) variables a is greater than 3.
 219

Proof. Linear dependence of vectors fsk, f0 in (66) is Theorem 2. If the vector of design variables a renders a lowest
a consequence of Lemma 1. To prove (67) we express, for N-fold eigenvalue ~1 = A2 = ... = AN a max-
example, f22 from (66) imum, it is necessary that the vectors f0, fsk,
s, k = 1,..., N, k > s are linearly dependent
f22 : -- 711~ 0712~ , 70
722 xl 1 -- ~7-~2t12 -1- -722
- f0 , (68) N
and substitute this expression into (63). Then we obtain E 7skfsk - 70f0 = 0, (73)
s,k=l
711/fT
7-~2~ 11e)2 -4- 2712
~ fT12e f T11 e -{- (flT2e)2 -> 0. (69) with the coefficients 70, 7sk satisfying conditions
of positive semidefiniteness of the symmetric ma-
This quadratic form of f T e and f T e is positive semidef- trix 7sk, s,k = 1 , . . . , N .
inite only if its coefficients satisfy the inequality in (67). Here it is assumed that the rank of the matrix consisting
Lemma 1 and Theorem 1 were formulated and proved for of the vectors f0 and fsk, s,k = 1 , . . . , N , k > s, is equal to
the fi~st time by Bratus and Seyranian (1983). N ( N + 1)/2, a n d t h a t this rank is less than I.
Note that due to the symmetry fsk = fks we delimit our-
4.4 N-fold optimum fundamental eigenvalue selves to symmetric coefficients 7sk = "/ks such that the sum
Consider next the general case when in the optimization in (73) can be written in the form
problem, (54) and (55), the maximum is attained at an N- N N N
fold multiple lowest eigenvalue A1 -- A2 -- . . . . AN ( E 7skfsk= E T s s f s s + 2 E 7skfsk" (74)
AN+I _< .... s,k=l 3=1 s,k=l
In this case, for any admissible direction e, i.e. direction s>k
satisfying the condition in (62), we find directional derivatives Nevertheless, we prefer the form of (73) due to its conve-
#j from (27) nience.
Proof. Linear dependence of the vectors f0, fsk, s,k =
det[f T e - #hsk] = O, s,k = 1 , . . . , N . (70) 1 , . . . , N , k ~ s, is an obvious consequence of Lemma 2.
Note that we have just assumed the rank of the matrix of
If the maximum is attained then there must be no admissible
these vectors to be N ( N + 1)/2, i.e. one less than the total
direction e for which all # = ~tj, j = 1 , . . . , N are of the
same sign. This is an obvious generalization of the necessary number of the vectors. Let us prove the necessity of positive
optimality condition in (61). semidefiniteness of the matrix 7sk. To this end, we choose
The lemma and the theorem for the general case can be a new basis of eigenvectors ¢ 1 , . . . , ~ N for which the matrix
formulated as follows. 7sk is diagonal, and show that if an optimum is attained then
all%s > 0 , s = 1 , . . . , N .
Lemma 2. If the vectors f0, fsk, s,k = 1 , . . . , N , k > s Let us transform the eigenvectors
[the total number of these vectors is equal to
N
( g + 1)N/2 + 1] are linearly independent, then
there exists an improving variation e for which ~s = E gksek , s = I , . . . , N . (75)
I~j>O,j=~,...,N. k=l
Here (~s are transformed eigenvectors satisfy!ng the or-
Note that the linear independence of the vectors is only
thonormality condition in (15), and gsk is the transformation
possible if I > (N + 1)N/2 + 1, where I is the dimension of
matrix.
the vector a of design variables.
Using (75) in (15), we obtain
Proof. Consider the system of linear equations in e l , . . . , eI
f T e = O, ~bTMej = gsi¢ M gkjfk =
f T e = ~ s k V O, s,k=l,...,N, k>s, (71)
N N
where v0 are given positive constants. If the vectors f0, fsk g igk#'TMek = g igk 6 k =
are linearly independent, a solution to (71) exists for any Vs0, s,k=l s,k=l
in particular when Vs0 > 0. Suppose the vector e is a solution
to the system in (71) and let us normalize this vector as
N
= He-
~ . Then we obtain from (71) and (70) EgsigsJ=hiJ' i,j= l,...,N. (76)
s=l
0 In matrix form this equation is equivalent to
gj:fT~:~ >0, j:I,...,N, (72) g T g = I and gT = g - l , (77)
which implies the existence of an improving variation. This where I is the unit matrix. The last equation means that the
proves the lemma (see also Seyranian 1987). transformation matrix g is an orthogonal matrix. Now let us
When I < (N + 1)N/2 + 1, the vectors f0, fsk are always express vectors ~bk from (75) by ¢s. Due to (77) we have
linearly dependent and hence an improving variation may not N
exist. ¢s = ~ gskek, s = l , . . . , Y . (78)
Let us formulate the theorem for the necessary optimality k=l
conditions. Using the notation
220

OK OK of vectors is equal to N ( N + 1)/2, i.e. ~one less than the num-

~al""' OaI ~al,] ' ber of vectors in the set. The vector fjj can be expressed as
and (26), we obtain a linear combination of the remaining vectors tO, fss in (82)
N N since q~j ~ 0. Then, if the vector fjj is removed from the
E "/ksfks = E 7 k s O T ( v K - A V M ) ¢ s = set f0, fsk, s, k = 1,..., N, k > s, the remaining vectors are
k,s=l k,s=l linearly independent.
Now, determining a vector of variation e from (85) and
e
normalizing this vector, 6 = [ - ~ , we obtain
E 7ks gkt¢ (VK - ArM) gsmq)m =
k,s=l t=l =
[jell>0, s=l,...,g, s#j,
E gktVksgsm ~T( V K - AVM)(~m =
t,m=l k,s=l (86)
N Thus we can find . j from (84) taking the variation
N
t,m=l ,j ~T.~ 1
So, in the new basis the matrix 7ks takes the form 7jj ,=1
,#j
N
= gk ksg . (80)
k,s=l 1 ~(Z/ss~uO. (87)
In matrix form we have s#j
,~ = gT,yg = g - l , T g " (81)
Here we have used (86).
This means that there exists a basis in which the matrix If the maximum of the lowest N-fold eigenvalue is
is diagonal. Then the optimality condition in (73) takes the achieved, then for any admissible variation e the sensitivi-
form ties "k, k = 1,... , N must not he of the same sign. Since
N we have chosen 6 such that all . s , s = 1,... ,N, s ¢ j, are
- 0f0 = o . (82) positive, . j must.be less than or equal to zero, i . e . . j _< 0.
s~-i Using (87) we obtain
The rank of the matrix of the vectors f0 and f'sk, s, k =
1,... ,N, k >_ s, is equal to the rank of the matrix of the
vectors f0 and fsk, s,k = 1 , . . . , N , k >_ s, since the two ,=i \vii/
sets of vectors are equivalent and can be expressed linearly sCj
in terms of one another (see Kurosh 1962), for an arbitrary choice of the positive constants us0, s =
N N 1,...,N,s~j.
f,k = E gsigkJfiJ ' fsk = E gisgJkfij" (83) The inequality in (88) can be satisfied only if
i,j=l i,j=l
~,__.~s__>0, s=l,...,N, sej, (89)
To show that the condition 7ss >_ 0, s = 1,..., N is the vii
necessary condition for optimality, let us consider an admissi-
since, otherwise, the constants u0 can be chosen such that
ble variation e, i.e. a variation satisfying the condition (62).
the inequality in (88) is violated. This means that all Z/ss,
Multiplying (82) by e, we obtain
s = 1 , . . . , N must be of the same sign. Without loss of
N
generality, all Z[ss can be regarded as nonnegative quanti-
= o. (84)
ties. So, we have proved that Z/s, >_ O, s = 1,..., N, which
implies positive semidefiniteness of the matrix of coefficients
Suppose that in (82) the j-th coefficient ~jj ~ O. Let us take 7,k, s, k = 1,..., N, and this constitutes the necessary opti-
the admissible variation e such that mality condition.
f i e = 0, Note that Masur (1984, 1985) formulated the condition
of positive semidefiniteness of the matrix ~, as a necessary
fTe=0, s,k=l,...,N~ k>s, optimality condition. Unfortunately, the proof presented by
~T e = .O, t= 1,...,N, t e j, (85) Masur (1985) is true only for the bimodal case N = 2. The
admissible variation, used by Masur, possesses two nonzero
where ,0 are arbitrary positive constants. values "1 ¢ 0, "2 ¢ 0 and N - 2 zero values , s = O,
Such a variation e exists for arbitrary ,t0 because the s = 3 , . . . , N. Masur assumed that at the optimum for this
vectors f0, fsk, s,k = 1.... ,N, k > s, and the vectors variation "1 and , 2 must be of opposite signs. This is not
ftt, t = 1, . . . , N , t 7k j, are linearly independent. This is true when N > 2, since the necessary optimality condition
true because due to the assumption made earlier, and the m i n ( , 1 , , 2 , . . . , , N ) -< 0 is satisfied for arbitrary ,1 and ,2.
equivalence of the set of vectors f0, f,k and the set f0, fsk, Above, we have proved the necessity of this condition in the
s, k = 1,..., N, k > s, the rank of the matrix of the latter set general case.
 221

Similar results for minimizing the maximum eigenvalue geometry. The columns are resting on an elastic founda-
were obtained by Overton (1988). Recently, Cox and Over- tion with the stiffness modulus C, are made of a material
ton (1992) derived the necessary optimality conditions for with Young's modulus E, and have a given total volume V
discrete and distributed eigenvalue problems using Clarke's and length L. Similar problems with a linear relationship
generalized gradient (see Clarke 1990), They also considered between I(x) and A ( x ) admit analytical solutions and were
lower and upper bounds on design variables a i < a i < -5i. considered by Plaut et al. (1986).
These results confirm the optimality conditions suggested by In the following, continuum and discrete problem formu-
Olhoff and Rasmussen (1977), and also used by many others lations will be considered in Sections 5.1 and 5.2, respectively,
(see e.g. Gajewski and Zyczkowski 1988). a numerical solution procedure will be developed in Section
5.3, and numerical examples will be presented in Section 5.4.
~.5 Example
Let us consider the case of a diagonal matrix K 5.1 Continuum problem
//11 (a) 0 Consider first a continuum formulation of the problem in
nondimensional form. If we nondimensionalize the coordinate
//= K22(a) (90) x by means of L and define the dimensionless cross-sectional
0 Knn(a) area function a(x), buckling load A and foundation modulus
e by
and suppose that M is the unit matrix. For these matrices
we have a(x) = A ( x ) L / V , ~ = pL4/(E(~V2),

As = K s s ( a l , . . . , a i ) , s= l,...n. (91) c = CL6/(Ec~V2), (96)

This is the case of n separate differentiable functions AAs = then the lateral deflection y(x) at buckling is governed by the
~7T K s s A a , s = 1,. .. ,n. eigenvalue problem consisting of the differential equation
Considering the optimization problem in (54) and (55), [a2(x)ytt (x)] t' :. -Aytt (oQ - cy(x) , (97)
we h a v e
and a set of boundary conditions. We shall assume that the
max
a
m i n [ / / l l ( a ) , . . . ,Knn(a)], F(a) = 0. (92) columns have "classical" boundary conditions, i.e. clamped,
Suppose that the maximum is achieved at the N-fold mul- simply supported or free ends. If e.g. the column is clamped
tiple eigenvalue A1 = ... = AN < AN+ 1 _< ,.. _< An. The and simply supported at the ends x = 0 and x = 1, respec-
necessary optimality conditions of max-min are well-known tively, the boundary conditions are
(see Demyanov and Malozemow 1972) y(0) = y ' ( 0 ) - - 0, y(1) = ( a 2 y ' ) z = l -- 0. (98)
N The continuum problem possesses an infinite number of
E 7 s s V K s s - 70f0 = 0, (93) eigenvalues Aj , 0 < A1 < A2 _< ..., and corresponding eigen-
s----1 functions yj(x). It will be assumed in the following that the
latter are orthonormalized
7ss>_0, s = l , . . . , N . (94)
1
These conditions cannot be directly derived from the
results presented in Section 4.4 since fsk = O, s , k = f = jk. (99)
1 . . . . . N, s ¢ k, whereby the vectors f0, fsk, s , k = 1 , . . . , N , 0
k :> s, will always be linearly dependent [in the linear com- The nondimensional optimization problem consists in de-
bination in (73) we may take 70 = O, 7ii = 0, i = 1,..., N, termining the cross-sectional area function a(x), which for a
and take 71j, i ~k j, arbitrarily]. given value of the foundation modulus e maximizes the small-
However, it is obvious that Lemma 2 and Theorem 2 in est eigenvalue,
Section 4.4 remain valid if we leave out the vectors fsk, s ~ k, max min ~, (100)
and only retain the vectors f0, fss, s = 1 , . . . , N, in the for- a(,~) j=l,...,, ~
mulations. With fss = ~7Kss we then obtain (93). Assum- subject to the condition of given column volume
ing that the rank of the matrix of the vectors f0 and fss, 1
s = 1,... ,N, is equal to N, we may construct an admissible
a(x) dx = 1. (101)
vector of variation e from
0
fTe=90 >0, s= 1,...,N, f0Te=0, (95)
If the maximized smallest eigenvalue ~1 is simple, ~1 <
and show that all the coefficients 7ss, s = 1,..., N, are of ~2 <- A3 -< ..., then the necessary optimality condition takes
the same sign, and this way establish (94). the form of (56) with (see e.g. Tadjbakhsh and Keller 1962)
VA 1 = 2a(x) [Ylt(x)] 2 , f0 = 1. (102)
5 O p t i m i z a t i o n o f c o l u m n s on a n elastic f o u n d a t i o n
The positive real constant 70 in (56) is a Lagrangian multi-
In this section the optimization of single and bimodal eigen- plier to be determined by the dimensionless volume constraint
values will be exemplified by the maximization of the buckling (lO1).
load P of thin elastic columns of variable, but geometrically Consider now the bimodal case where the smallest eigen-
similar cross-sections with the relationship I(x) = c~A2(x) value ~1 = A2 < ~3 - ... is associated with two linearly
between the second area moment I(x) and cross-sectional independent eigenfunctions satisfying (97), the appropriate
area A ( x ) with the constant ~ given by the cross-sectional boundary conditions, and the orthonormality condition (99).
222

If we consider a varied cross-sectional area function in the

form a(x)+ee(x), then the double eigenvalue Ai = A2 = )t as-
1 ]
2ai ys(x)yk~X)dx , s,k= l,2, (106)
sociated with a(x) will generally split into two distinct ones,
Xl-1
Aj .= ~ + ~#j, j = 1,2. By using a perturbation technique
as m Section 2.3, we derive the following quadratic equation it is easily seen that the quadratic equation (103) for the
for determining the two directional derivatives (design sensi- two directional derivatives p = /~1 and # = #2 of the bi-
tivities) # = #1 = AA1/A~ and p = #2 = AA2/A~ of the modal eigenvalue precisely takes the form of (58) with solu-
double eigenvalue, tions given by (59). Similarly, we easily identify the vector
f0 in the volume constraint (101) as

det 2 ( )Ys( ) 9 k ( x ) e ( x ) d x - # 6 s k ] =0,

f0 = ( X l , . . . , x i - z i - 1 , . . . , 1 - X i _ l ) = (t 1 . . . . , g i , ' " , g I ) ,
(107)
for the discrete problem. In the examples in Section 5.4, the
s, k = 1, 2. (103) integrals in (106) are computed by numerical integration with
This equation is analogous to (27) (with N -= 2). the functions y~l(x) and y~l(x) being represented by their dis-
In problems where the boundary conditions and the de- crete values at sets of densely spaced mesh points embedded
sign function a(x) are symmetric, a(x) = a(1 - x), it is in each of the column segments.
useful to apply symmetric and antisymmetric eigenfunctions
Yl (x) = Yl (1 - x) and Y2 (x) = -Y2 (1 - x) in (103). Since the 5.3 Method of optimization
otherwise arbitrary "direction" function e(x) must be sym- In order to exemplify the results in Section 4 and their appli-
metric along with a(x), we see that the mixed term in (103) cability for design problems, we shall now develop an iterative
vanishes: numerical method for the solution of optimization problems
1 for stepped columns on an elastic foundation. For these prob-
P
lems it is unknown a priori whether the optimum buckling
2 / a(x)y~'(x)y~(x)e(x) dx = 0, (104)
load is a simple or a double eigenvalue. Extension of the
0 method to any degree of multiplicity of the optimum eigen-
since its integrand will be antisymmetric for any arbitrary value is straightforward.
admissible function e(x). Hereby (103) gives the following Assume first that at a given iteration stage t h e design is
simple expressions for the directional derivatives of the double associated with a bimodal eigenvalue A1 = A2 = A, or that
eigenvalue: the two lowest eigenvalues A1 < A2 are very close such that
1 A2 - A1 < 5 where 6 is a small tolerance which we assume to
be specified.
Pl = 2 j a ( x ) [yi'(x)] 2 e(x) dx,
Guided by the results stated in Lemma 1 and Theorem
0 1 in Section 4.3, we take the vector of increments z~a of the
1
vector of design variables a as
#2 = 2 f a(x) [yg(x)] 2 e(x) d x . (105) A a : k(711fll + 2712f12 + 722t"22 - 70f0), (108)
where k is a move-fimit type of scaling factor. We may only
0
need to apply k in iteration steps in the beginning of the
These expressions hold when the double eigenvalue can computational procedure where the vector resulting from the
be treated as an intersection of two differentiable functionals,
expression in the parenthesis may become sufficiently large
and are identical to those resulting from purely single modal to hamper convergence. Application of k is not necessary
formulations. However, in the general case of non-symmetric
later where the expression in the parenthesis, and hence Aa,
boundary conditions and designs, the mixed term in (104) will tend towards the null vector, cf. (66), as we approach the
does not vanish. optimum solution.
In (108), the vectors fll, f12, f22 are defined by (106),
5.2 Columns with piecewise constant cross-sections the vector f0 by (107), and 711, 2"/12, */22 and 70 are to be
Let us now cast the optimization problem in discrete form determined. Both (66) and (108) permit us to normalize the
by assuming the column to be composed of segments xi_ 1 < latter quantities as
x ~ x i of prescribed lengths gi = x i - xi-1, i = 1 , . . . , I , */11 ~" */~1/II*/* II, 2"/12 = 2*/h/I1"/*11,
(where x 0 = 0 and x I .= 1), and individual constant values
a i of the cross-sectioaal areas, i.e. a(x) - ai, xi_ 1 < x < xi, */25 = */:hill< II, */o = II, (lO9)
i = 1,..., I. We treat the values of a i as design variables, where
assemble them in the vector a = ( a l , . . . ,al) , and let e =
( e l , . . . , ei) denote the corresponding unit direction vector of II*/*ll=v,l. +,(2./,~2
12' -'1"-*/,2
2 2 " t - */~2 • (110)
arbitrary variations. Equations for obtaining */51' 2712"' */22" and */~ [and hence
Defining the vectors fsk, s, k = 1, 2 as [] 7" [] by (110)] will be established below.
Now, instead of using the unit vector e and the directional

f k=
[! Y
xi-1
.... ,
derivatives Pl = AA1/z~¢ and #5 ----AA2/A¢ as in (58) and
(59), we wish to work with the increments AA 1 and AA2
of the double eigenvalue that directly correspond to a given
 223

vector A a of increments of the design variables. To this end, substitute these into (108) to obtain the new vector A a of
we multiply each component in (58) by e, and obtain' the increments of design variables.
following quadratic equation for the increments AA = AA 1 We note that the determinant of the coefficient matrix
and AA = AA2, in (116) is nonnegative and that it only vanishes if the vec-
tors fll, f12, f22 and f0 become linearly dependent, which
[ fTAa-A)~ fTAa ] is the necessary condition for an optimum bimodal solution,
det flT2Aa f T A a - AA = 0 , (111) cf. Theorem 1 and (66) in Section 4.3. The latter implies,
see (66), that the vector A a in (108) vanishes at the bimodal
just as we earlier transformed (27) into (28). optimum. In the above iterative computational procedure,
Obviously, the increments A a for the iterative computa- the numerical values of 7~1, 27~2, 722. and 7~ obtained from
tional procedure can be chosen in many ways. Equation (111) (116), and hence the value of [I 7" [I from (110), rapidly in-
leads us to choose the following four simultaneous conditions creases as we approach the bimodal optimum point with lin-
as a basis for determining the four coefficients 711, 2712, 722 ear dependence of the vectors f l l , f12, f22 and f0. This im-
and 70 in the expression for A a in (108) (where we disregard plies that in addition to A a ~ 0 in (108), due to [17" []~ c~,
the scaling factor k): in (112) and (113)we have A)~1 ~ 0 and A)~2 ---* 0 as we
AA 1 = fTAa---- 1/[17" [I, (112) approach the bimodal optimum solution.
Let us finally consider the case where the column design
ziA 2 = f T A a = (1 + )~1 --"~2)/II7" II, (113) obtained at a given iteration s~tage is associated with a simple
buckling eigenvalue )~1 with A2 - A 1 > 5. In this case we wish
flT2Aa = 0, (114) to increase A1 by single modal steps of redesign. The eigen-
value increment and the volume constraint are then expressed
f [ A a = O. (115) by
Here, (115) expresses the volume constraint (which is lin- = (117)
ear in the design variables in the present case; the satisfaction
of a nonlinear constraint requires a trivial modification of the f [ A a = 0, (118)
approach). Equation (114) imposes a diagonalization of the where f l l and f0 are defined by (106) and (107), and the
matrix in (111), which reduces the solutions AA = AA 1 and vector of increments of the design variables is taken in the
A)~ = AA 2 of (111) to those expressed by the first equality single modal form
signs in (112) and (113). These expressions imply that AA 1
and AA 2 are precisely the increments of A1 and A2, respec- A a = k(fll - 70f0), (119)
tively, whose indices correspond to the directly identifiable where k is again a positive scaling factor. The a priori un-
modes yl(x) and y2(x) obtained from the solution of the known constant 70 is determined by substituting (119) into
eigenvalue problem for the current design. (118), which gives
It is seen from (112) and (113) that if A1 = A2 at a given
f[f11 (120)
iteration step, we specify AA 1 and AA 2 to be equal and pos-
70- f[f0'
itive with a view to obtaining an increase of the bimodal
eigenvalue. If the two lowest eigenvalues are different (within and substitution of (119) and (120) into (117) yields the
the small difference 5 defined earlier), we assign the index 1 Cauchy-Bunyakowski inequality
/
to the smallest eigenvalue and the corresponding mode, pre-
scribe the value 1/ l] 7* [1 for its increment A~I, see (112), AA 1 k (f1T1fxl - ( q f l l ) ( f t ~ f11)~ > 0 (121)
and assign a slightly smaller value to the increment AA 2 of \ (f0 ]-'
the next eigenvalue A2, see (113), in order to increase A1 for the increment A)q of the eigenvalue. Thus, each step
while diminishing the difference between )~1 and A2. of redesign increases the eigenvalue A1 while satisfying the
Now, by substituting (108) (where we disregard k) into volume constraint, (118), and this continues until f l l and f0
(112)-(115) and making use of (109), we obtain the following become linearly dependent, i.e.
system of equations for determining the unknown coefficients f l l - 70f0 = 0, (122)
7~1, 27~2, 722. and 7~: which is the necessary condition for single modal optimality
~f12 flTlf0 of the fundamental eigenvalue AI"
f1T1f11 f1T1f22

symm
fTf12
r f12
752
27h / =
If, during this single modal iterative procedure, the dis-
tance between A2 and ~1 decreases and we obtain )~2-A1 _< 5
at a certain stage, then we perform subsequent iterations
using the bimodal eigenvalue optimization procedure de-
scribed earlier. This may give rise to a decrease of the dis-
tance between the eigenvalues A3 and the bimodal eigenvalue
1 A1 = A2, and if coalescence takes place, it is necessary to
1 + A1 - A2 (116) adopt a trimodal optimization scheme for subsequent iter-
0 ations, and so on. Note that in addition to the (single or
0 multimodal) fundamental eigenvalue subject to treatment at
Having solved (116), we compute 117* II from (110), the nor- a given stage of the iterative procedure, it is also necessary to
malized coefficients 711, 2712, 722 and 70 from (109), and know the value of the next (higher order) eigenvalue in order
224

to be able to capture its possible coalescence with the funda- All the designs in Fig. 29 are found to be symmetric and
mental eigenvalue and update the subsequent computations. associated with a bimodal optimum buckling load. The buck-
As mentioned earlier, it is straightforward to construct a mul- ling load of the optimum column without elastic foundation
timodal scheme for optimization of eigenvalues of any multi- (c = 0) in Fig. 29a is A = A1 = A2 = 52.105 and we have
plicity N by generalization of the method described above. A/A u = 1.320, i.e. the optimum (bimodal) buckling load is
32.0% higher than the (simple) buckling load of a correspond-
5.4 Numerical examples ing uniform column with both ends clamped and the same
volume and length.
Subject to some selected values of the elastic foundation mod- It is interesting to note that the value A = 52.105 for the
ulus c, see (96), we now solve the dimensionless buckling stepped optimum column design in Fig. 29a with 50 design
load optimization problem for a stepped column with I = 50 variables, is only marginally less than the bimodal buckling
segments of equal lengths li = 1/50 but individual cross- load A = 52.356 for the classical clamped-clamped eonlinuum
sectional areas ai, i -- 1 , . . . , I, which play the role as design column (with, in principle, infinitely many design variables)
variables of the discretized problem. No minimum or maxi- determined by Olhoff and Rasmussen (1977).
mum constraints are specified for the design variables. The Figure 30 displays the iteration history behind the opti-
optimization problems are solved by the method developed in mum solution in Fig. 29a, when using the method of Section
Section 5.3, and we have chosen to determine the eigenvalues 5.3 and starting out from uniform design. We see that the
and eigenmodes in the steps of redesign by a finite difference initially distinct eigenvalues A1 and A2 coalesce after around
procedure based on successive iterations (which involves fur- 10 iterations and that subsequent iterations are carried out
ther discretization of each of the segments of the column). by the bimodal eigenvalue optimization procedure. Figure 30
Examples will be presented for (symmetric) clamped- also shows our monitoring of A3 which turns out to remain
clamped boundary conditions and for (non-symmetric) distinct from A1 and A2. (This is the case for all the examples
clamped-simply supported conditions. In each of the exam- considered in this section.)
ples, the initial column design (iteration 0) has been chosen

iiii
Eigenvalue
to be uniform with a i = 1, i = 1,..., I = 50, which meets
1 ,
the nondimensional volume constraint ~ aig i = 1. For each
value of c and set of boundary conditions, the uniform design l l O . . . . . . i . . . . . . . j . . . . . . . T . . . . . . . T. . . . . . . [. . . . . . . . ~. . . . . . . i . . . . . . .
will also be used as a reference for the gain of the optimiza-
tion, and we shall denote the fundamental buckling eigen-
value of the uniform column by Au.
Figure 29 illustrates optimum designs of columns with
both ends champed. The designs are shown to suitable scale,
100t ..... i ....... ~ - ~ + ~ - , ~ . ~ - ~ . ~ , ~ - ~ = ~

8# ......
i
T. . . . . . . I ....... T .......
i
T. . . . . . .
i
I ........
!
F. . . . . . .
i 7./!
l .......
- -

%- l,
L eilenvlllue
eitenvalue

and the linear dimensions perpendicular to the column axes

are proportional to the square root of the cross-sectional ar- 60 7J77 77777;;
eas.
2040 ~ i i i i i l l ii~iiiiii iiiiiiil]iiiiiiiiiiiiiiii~iiiil]

3*
0
0 5 10 15 20 25 30 35 40
Iteration n u m b e r
(A) c=0
Fig. 30. Iteration history, optimization of clamped-clamped col-
umn with c = 0. The optimum design is shown in Fig. 29a

For the solution in Fig. 29a, the vectors fll, f12, f22 and
f0 are linearly dependent, i.e. satisfy (66). The (normalized)
values of the coefficients 711, 2712 and 722 are determined
(B) c = 876.7 to be 711 : 3.459.10 - 4 , 2712 = 0 and 722 = 9.250.10 - 3 ,
which implies that the inequality for positive semidefiniteness
in (67) is also satisfied. Thus, both necessary conditions (66)
and (67) for a bimodal optimum are satisfied by the solution
X > ~ • X, in Fig. 29a. The same holds true for the solutions in Figs.
29b and 29c. Here, we again find 2712 = 0, which is not
surprising since the optimum designs obtained and shown in
(C) c = 1500 Fig. 29 are symmetric.
Fig. 29. Clamped-clamped optimum columns subject to differ- Figure 29b shows the optimum design of the stepped col-
ent values c of modulus of elastic foundation. Columns have 50 umn with two clamped ends when the modulus of the elastic
segments of equal lengths. (a) e = 0, optimum A = 52.105 is foundation is taken to be c = 876.7. The design is associated
bimodal, )~/)~ = 1.320. (b) c = 876.7, optimum A = 102.55 is with a bimodal optimum buckling load A = A1 = A2 = 102.55
bimodal, )~/£~ -----1.039. (e) c = 1500, optimum A = 123.91 is which is only 3.9% larger than the buckling load of the cor-
bimodal, ) , / ~ = 1.114 responding uniform column with the same volume, length,
 225

foundation modulus, and boundary conditions. The inter-

esting point associated with c = 876.7 is that at this value k
of the foundation modulus, the buckling load of the uniform
comparison column is bimodal, Au = A~ = A~ = 98.67. This
implies, as can be seen in Fig. 31, that when starting the op- (A) c = 50
timization procedure from uniform design, the final optimum
bimodal design in Fig. 29b will be obtained through an iter-
ation history that exclusively encompasses steps of redesign
involving himodal buckling loads. > +777 @ J',, k
Elgenvalue
105 ~ I I I I l (B) c 500
=
i i i i i i
lo<l ........ f ....... i ........ r ....... l ...... i ....... T ....... { ........
i i I i i i I

lo2 ........ +......... ~ ....... T....... i ........ F....... T........ k

,. ================================.
i i i i i i i
lOl , I
,
I
,
I I I
,
l
, + ,1,. -etgen'value
~
-
(C) c = I000
Fig. 32. Clamped-simply supported optimum columns subject to
different values c of modulus of elastic foundation. Columns have
...... i ....... ........ i ....... ! ...... 4........ i ....... f ........
9 ,"/L ....... i ....... 4 ....... i- ....... ~ ....... 4 ........ i- ....... ~-.......
50 segments of equal lengths. (a) c = 50, optimum A = 31.042
/ i i i '~ i i i is simple, A/),~ = 1.278. (b) c = 500, optimum A = 66.296 is
96 p....... +....... -t ........ I- ....... ~-....... -i ....... b ...... ÷...... bimodal, A/A~ = 1.188. (c) c = 1000, optimum A = 90.420 is
9 5| ' i i i i i i i bimodal, A/Aa = 1.214
0 10 20 30 40 50 60 70 80
Eigenvalue
Iteration number
12~. i ] ] i i
Fig. 31. Iteration history, optimization of clamped-clamped col- 110 ......... ,~.......... ,~..........;.......... •.........."...........
umn with c = 876.7. The optimum design is shown is Fig. 29b

Figure 29c illustrates the stepped optimum column de-

sign for c = 1500. For this solution, the optimum bimodal so ........... .~.......... .i.......... .÷.......... T.......... i . . . . . . . . . .

I I I ' '
buckling load is A = A1 = A2 = 123.91, which is 11.4% larger 70 . . . . . . . . . . . [ . . . . . . . . . . [ . . . . . . . . . . i . . . . . . . . . . i . . . . . . . . . . ] . . . . . . . . . . . . 1. eigenvalue
than the (simple) buckling load of the corresponding uniform
column on the same foundation. The iteration history is qual- 50 ............. -~- 3. eigenvalue

itatively similar to the one shown in Fig. 30. The solution in 40

' [ ! . . . . ! _ ! _
Fig. 29c illustrates the fact that the waviness of the optimum 30 ~ ~ - ~ . . . . . . . . . . r . . . . . . . . . . C. . . . . . . . . . ~. . . . . . . . . . .
designs increases with increasing values of the modulus c of zo! .......... L.......... i .......... i .......... i ......... ~ . . . . . . . . . . .

the elastic foundation. ........... L .......... "i .......... i.......... "i ......... ,"...........
l~! i i i I !
Next, we consider optimum designs of stepped columns
with one end clamped and the other simply supported, see 5 10 15 20 25 30
Fig. 32. It is characteristic that both for these boundary con- Iteration number
ditions and for doubly clamped boundary conditions, corre-
Fig. 33. Iteration history, optimization of clamped-simply sup-
sponding optimum solutions for continuum columns are as- ported column with c = 50. The optimum design is shown in Fig.
sociated with bimodal buckling loads for any value c >_ 0 32a
of the modulus of the elastic foundation. However, for the
stepped clamped-simply supported columns under considera- is 18.8% larger than the (simple) buckling load of the cor-
tion, with the finite dimension I = 50 of the design space, the responding uniform reference column. The iteration history
optimum designs are associated with simple buckling loads is qualitatively similar to that shown in Fig. 30, and the fi-
for values of c up to a threshhold value of approximately 250, nal solution satisfies the necessary conditions in (66) for a
and only associated with bimodal optimum buckling loads bimodal optimum solution. The (normalized) values of 711,
for larger values of c. 2712 and 722 associated with the solution are found to be
Figure 32a shows the stepped optimum column design 711 = 2.488'10-3, 2712 = 1.882'10 - 3 and T22 = 1.319"10-2,
for c = 50. Its buckling load A -- A1 = 31.042 is simple, and so the necessary condition in (67) is also satisfied. We note
27.8% larger than that of a corresponding uniform column on that the value of 2712 is now non-zero as the optimum design
the same foundation. Figure 33 displays the iteration history, in Fig. 32b is nonsymmetric.
which entirely consists of single modal steps of redesign, and For c = 1000, we obtain the optimum design in Fig. 32c
we have checked that the final solution satisfies the necessary with the bimodal buckling load A = A1 = ~2 = 90.420, which
condition in (56) for a single modal optimum solution. is 21.4% higher than the (simple) buckling load of the corre-
For c = 500, the stepped optimum column, see Fig. 32b, sponding uniform comparison column. The iteration history
has a bimodal buckling load A = A1 = A2 = 66.296 which behind this optimum solution is also qualitatively similar to
226

that in Fig. 30, and we again find that the final solution sat- Dynamics and Materials Conf., Part 1, pp. 562-581. Also: Mech.
isfies the necessary conditions (66) and (67) for bimodal op- Struct. Much. 18, 407-432 (1990)
timality, and that 2712 ik 0. Figure 32b illustrates that also Bathe, K.-J. 1982: Finite element procedures in engineering anal-
the clamped-simply supported optimum columns become in- ysis. New Jersey: Prentice-Hall
creasingly wavy with increasing values of c.
Bendsce, M.P.; Olhoff, N.; Taylor, J.E. 1983: A variational for-
mulation for multicriteria structural optimization. Mech. Struct.
6 Conclusions Much. 11,523-544
Multiple eigenvalues often occur in symmetric structures and Bratus, A.S.; Seyranian, A.P. 1983: Bimodal solutions in eigen-
in complex structures that depend on many design param- value optimization problems. Prikl. Matem. Mekhan. 47, 546-
eters. Thus, multiparameter optimization'problems should 554. Also: Appl. Math. Mech. 47, 451-457
always be considered in this perspective. Simple examples Choi, K.K.; Haug, E.J. 1981: Optimization of structures with re-
show that multiple eigenvalues are not differentiable in the peated eigenvaiues. In: Hang, E.J.; Cea, J. (eds.)Optimization of
common sense. This creates serious problems for derivation distributed parameter structures, Vol. 1, pp. 219-277. Amsterdam:
of optimality conditions and numerical analysis in solving op- Sijthoff and Nordhoff
timization problems. Choi, K.K.; Haug, E.J.; Lam, H.L. 1982: A numerical method
The design sensitivity analysis based on a perturbation for distributed parameter structural optimization problems with
technique presented in this paper shows that directional repeated eigenvalues. J. Struct. Mech. 10, 191-207
derivatives of multiple eigenvalues are eigenvalues of a spe-
Clarke, F. 1990: Optimization and nonsmooth analysis, classics in
cially constructed symmetric matrix with the components applied mathematics 5. Philadelphia: Society for Industrial and
f T e . The efficiency and accuracy of sensitivity analysis of Applied Mathematics
simple and multiple eigenvalues with respect to six design
parameters has been demonstrated by a thorough numerical Cox, S.J. 1992: The shape of the ideal column. The Mathematical
Intelligencer 14, 6-24
study of a multiparameter vibration problem of an elastic
square plate reinforced by stiffeners. Cox, S.J.; Overton, M.L. 1992: On the optimal design of columns
We show that linear dependence of the vectors fsk and the against buckling. SIAM J. Math. Anal. 23, 287-325
gradient vector f0 of the constraint with the coefficients 7ij Courant, R.; Hilbert, D. 1953: Methods of mathematical physics,
and 70 respectively, constitutes the first necessary optimality Vol. 1. New York: Interscience Publishers
condition, and that positive semidefiniteness of the matrix of
Demyanov, V.F.; Malozemov, V.N. 1972: Introduction to mini-
the coefficients 7ij is the second necessary optimality condi-
max. Moscow: Nauka
tion for an optimum multiple eigenvalue. It is demonstrated
and illustrated by examples of optimization of columns on Gajewski, A. 1990: Multimodal optimal design of structural ele-
elastic foundations that these optimality conditions can be ments. Mechanika Teoretyczna i Stosowana 28, 75-92
directly used in the development of a very efficient method Gajewski, A.; Zyczkowski, M. 1988: Optimal structural design
for optimum design. under stability constraints. Dordrecht: Kluwer
It is shown in the paper that there always exists a ba-
Haug, E.J.; Choi, K.K.; Komkov, V. 1986: Design sensitivity anal-
sis of eigenvectors corresponding to the multiple eigenvalue ysis of structural systems. New York: Academic Press
in which the matrix of the coefficients 7ij is diagonal. For
this case, the necessary optimality conditions attain the same Hang, E.J.; Rousselet, B. 1980: Design sensitivity analysis in
form as those obtained by a Lagrange multiplier method ap- structural mechanics. II: eigenvalue variations. J. Struct. Mech.
plied to different and differentiable functionals. This cir- 8, 161-186
cumstance, fortunately, confirms many papers on multimodal Kurosh, A.G. 1962: Course of higher algebra. Moscow: Fizmatgis
optimization which used a formal variational approach or a
Lancaster, P. 1964: On eigenvalues of matrices dependent on a
formal approach based on Pontryagin's maximum principle parameter. Numerisehe Mathematik 6, 377-387
without taking care of the nondifferentiabifity of multiple
eigenvalues. Lund, E.; Olhoff, N. 1993a: Reliable and efficient finite element
based design sensitivity analysis of eigenvalues. In: Herskovits, J.
(ed.) Structural optimization '93, Vol. 2, pp. 197-204. Rio de
Acknowledgement Janeiro: COPPE
The work presented in this paper received support from the Danish Lund, E.; Olhoff, N. 1993b: Shape design sensitivity analysis of
Technical Research Council (Programme of Research on Computer eigenvalues using "exact" munerical differentiation of finite ele-
Aided Design). ment matrices. Report No. 54, Institute of Mechanical engineer-
ing, Aalborg University. Also: Struct. Optim. 8 (1994)
Masur, E.F. 1984: Optimal structural design under multiple eigen-
References value constraints. Int. J. Solids Struct. 20, 211-231
Arnold, V.I. 1989: Mathematical methods of classic mechanics. Masur, E.F. 1985: Some additional comments on optimal struc-
Moscow: Nauka tural design under multiple eigenvalue constraints. Int. J. Solids
Struct. 21, 117-120
Barthelemy, B.; Haffka, R.T. 1988: Accuracy analysis of the semi-
analytical method for shape sensitivity analysis. AIAA Paper 88- Masur, E.F.; Mrdz, Z. 1979: Non-stationary optimality conditions
2284, Proc. AIAA/ASME/ASCE/ASC 29th Structures, Structural in structural design. Int. J. Solids Struct. 15, 503-512
 227

Masur, E.F.; Mr6z, Z. 1980: Singular solutions in structural opti- tion of compressed columns on elastic foundations. J. Appl. Mech.
mization problems. In: Nemat-Nesser, S. (ed.) Variational meth- 53, 130-134
ods in the mechanics of solids, pp. 337-343. New York: Pergamon
Prager~ S.; Prager, W. 1979: A note on optimal design of columns.
Myslinsld, A.; Sokolowski, J. 1985: Nondifferentiable optimization Int. J. Mech. Sci. 21, 249-251
problems for elliptic systems. SIAM J. Control and Optimization
Seyranian, A.P. 1983: A solution of a problem of Lagrange. Dokl.
23, 632-648
Akad. Nauk SSSR 271,337-340. Also: Soy. Phys. Dokl. 28, 550-
Olhoff, N. 1980: Optimal design with respect to structural eigen- 551
values. In: Rimrott, F.P.J.; Tabaxrot, B. (eds.) Prvc. XVth Int.
Seyranian, A.P. 1984: On a problem of Lagrange. Mekhanika
IUTAM Cong. Theoretical and Applied Mechanics, pp. 133-149.
Tverdogo Tela 2, 101-111. Also: Mechanics o] Solids 19, 100-111
Amsterdam: North-Holland
Seyranian, A.P. 1987: Multiple eigenvalues in optimization prob-
Olhoff, N.; Plaut, R.H. 1983: Bimodal optimization of vibrating
lems. Prikl. Mat. Mekh. 51, 349-352. Also: AppL Math. Mech.
shallow arches. Int. J. Solids Struct. 19, 553-570
51, 272-275
Olhoff, N.; Rasmussen, S.H. 1977: On single and bimodal opti-
Tadjbakhsh, I.; Keller, J. 1962: Strongest columns and isoperi-
mum buckling loads of clamped columns. Int. J. Solids Struct.
metric inequalities for eigenvalues. J. Appl. Mech. 29, 159-164
13, 605-614
Taylor, J.E.; Bendsee, M.P. 1984: An interpret'ation for min-max
Olhoff, N.; Rasmussen, J.; Land, E. 1993: A method of "exact"
structural design problems including a method for relaxing con-
numerical differentiation fox error elimination in finite element
straints. Int. J. Solids Struct. 20, 301-314
based semi-analytical shape sensitivity analysis. Mech. Struct.
Mach. 21, 1-66 Wittrick, W.H. 1962: On eigenvalues of matrices dependent of a
parameter. Numerische Mathematik 6, 377-387
Olhoff, N.; Taylor, J.E. 1983: On structural optimization. J. Appl.
Mech. 50, 1139-1151 Zhong, W.; Cheng, G. 1986: Second-order sensitivity analysis
of multimodal eigenvalues and related optimization techniques.
Overton, M.L. 1988: On minimizing the maximum eigenvalue of
Mech. Struct. Mach. 14, 421-436
a symmetric matrix. SIAM J. Matrix Anal. Appl. 9, 256-268
Zyczkowski, M. 1989 (ed.): Structural optimization under stability
Plaut, R.H.; Johnson, L.W.; Olhoff, N. 1986: Bimodal optimiza-
and vibration constraints. Berlin, Heidelberg, New York: Springer

Received Feb. 15, 1994

Revised manuscript received Sept. 30, 1994

Announcement

CISM Control of Flow Instabilities and Unsteady Flows

I n t e r n a t i o n a l C e n t r e for M e c h a n i c a l Sciences Coordinators: G.E.A. Meier (G6ttingen), G.H. Schnerr
P r e l i m i n a r y P r o g r a m m e 1995 (Karlsruhe)
Courses Sept. 18-22, 1995
New Design Concepts for High Speed Air Transport Advanced Methods of Tunnelling
Coordinator: H. Sobieczky (Ghttingen) Coordinator: K. Kovari (Ziirich)
June 5-9, 1995 Sept. 25-29, 1995
Bone Cell and Tissue Mechanics Earthquake Resistant Design
Coordinator: S.C. Cowin (New York) Coordinator: It.T. Duarte (Lisbon)
July 10-14, 1995 Oct. 2-6, 1995
Liquid Bridge Theory and Applications The Flow of Particles in Suspension
Coordinator: J. Meseguer (Madrid) Coordinator: G. Cognet (Grenoble)
July 17-21, 1995 Oct. 9-13, 1995
Mathematical Modelling for Arch Dam Design and Safety
Control Meetings Hosted by GISM
Coordinators: E.it. de Arantes e Oliveira, J.O. Pedro Workshop on "Multimedia GIS Data"
Sept. 4-8, 1995 International Society for Photogrammetry and Remote
Mechanics of Solids with Phase Changes Sensing- Commission 1
Coordinators: M. Berveiller (Metz), F.D. Fischer (Leoben) Coordinators: It. Galetto (Pavia), F. Crosilla (Udine)
Sept. 11-15, 1995 June 12-16, 1995
Modelling and Simulation of Hypersonic Flows for Spatial 10-th International Conference on Artificial Intelligence in
Flights Engineering
Coordinator: It. Brun (Marseille) Chairmen: C.A. Brebbia (Southampton), C. Tasso (Udine)
Sept. 18-22, 1995 July 4-6, 1995