WORST SHAPES OF TMPERFECTIONS FOR SPACE

TRUSSES WITH MANY SIMULTANEOUSLY

BUCKLING MEMBERS

RALF PEEK and NICOLAS TRIASTAFYLLIDIS

College of Engineering. The University of Michigan. Ann Arbor, Ml 48109. U.S.A.

Abstract-Optimally designed truss type structures whose joints do not transmit moments often
have ;1 large number of coincident buckling modes. Each mode corresponds to the buckling of an
individual member. Due to the interaction between various simultaneous modes. such trusses
can be sensitive to imperfections. This problem is analysed using the Lyapunov-Schmidt-Koiter
decomposition and asymptotic expnnsion technique. The shape of the imperfection that maximizes
the load drop is determined from the postbifurcated equilibrium branch of the perfect structure on
which the loud drops most rapidly. It is shown that this branch is obtained by minimizing a
homogeneous quadratic form subject to linear inequality constraints. The generAI theory is illus-
trated by several examples involving two- and three-dimensional trusses.

I INTRODUCTION

Lattice type structures ;lrc’ not only cornmm for applications on the ground such as lattice
roofs. radio antennae, crane booms, dishes of radio telescopes and lattice domes; but are
also being envisioned for applications in space. including lattice columns. communications
platforms. ractio-astronotny dishus. solar panels, reflectors and other structures. With the
advent and increasing avail;tbility of hi&r strength materials. buckling often becomes a
critical design consideration. This is espccialiy true for applications in S~CC. due to the
combination ot’ large structlllXl dimensions, small loads and the importance of weight
minimization (although the limitation of dcllcclions is ills0 iln CSSCntiill considcrntion in
this case).
Often such structures are optimized so that. as the applied load is increased, a number
of members reach their buckling load at the same time. Buckling ofany member corresponds
to an eigenmode for the system as a whole. Thus, an optimized structure can have a large
number of coincident eigenmodcs.
It has long been established (Koitcr, 1945) that multiple coincident eigenmodcs can
lead to high imperfection sensitivity of the structure’s load carrying capacity. For example,
the ioad carrying capacity of a cylinder in axial compression can drop by a factor of five
or more due to imperfections. Lattice structures can also be sensitive to imperfections
(Wright, 1965. I966a; Castajlo, 1989; Britvec. 1973; Britvec and Davister. 1985): a lattice
dome in Bucharest failed at an estimated load of about 40% of the theoretical buckling
load for a perfect dome (Wright. 1965. 1966a). The high imperfection sensitivity of lattice
domes was further confirmed by experimental results reported in Wright (1966b). Lattice
columns are also very imperfection sensitive when the overall column buckling and member
buckling occur at approximately the same level of applied load (Thompson and Hunt,
1973 ; Crawford and Benton. I980 : Elyada, 1985). In this paper a general methodology for
determining the imperfection sensitivity for space trusses with multiple coincident modes
involving buckling of individual members in the elastic range is presented. The advantages
of the approach presented here over previous studies on this subject are : (I) it provides the
worst shape of imperfection. as well as a simple relationship between the magnitude of the
imperfection (of the worst shape) and the corresponding drop in load carrying capacity of
the structure; (2) it is applicable for any truss type structure; and (3) it is computationally
efficient and avoids severely ill-conditioned calculations.
The approach is based on the decomposition and asymptotic expansion technique that
was pioneered by the mathematicians Lyapunov and Schmidt around the turn of the
century. and also later (apparently independently) by Koitcr (1945) who applied the meth-
odology to structural problems. For the reader’s convenience. the main general results

2385
2386 R. PEEK and \ THMSTAFYLLIDIS

from this Lyapunov-Schmidt-Koiter (LSK) decomposition and asymptotic expansion are

reviewed in Appendix A. Results for the worst shape of imperfection (Koiter. 1976:
Triantafyllidis and Peek, 1992) are also summarized in this appendix.

2. RESIJLTS FOR .A SINGLE MEMBER

In order to determine the postbuckling behavior of the structure. it is first necessary

to establish a single valued and smooth expression for the strain energy of an individual
member as a function of member deformation parameters. One possibility ~vould be to use
the member elongation. L’,,, for member VI. as the deformation parameter. However. a plot
of asial force versus elongation exhibits a sudden change in slope when the member buckles.
This means that second and higher derivatives of this strain energy function for the member
do not exist at this point. and renders the asymptotic expansion of the LSK approach
inapplicable. To overcome this problem, an additional out-of-straight member deformation
parameter n*,,,,, is introduced. which arises naturally in the solution for the postbuckling
behavior of a single member by the LSK approach.
More specifically, for the cast of a member with constant cross-sectional propcrtic?.
the transvcrsc displacements (in a coordinate ~ramc that remains aligned with the mcmhcr
endpoints) can bc cnprcsscd as

whcrc L,,,,, dcnotcs the length of the mcmbcr bcforc doformation. and Cc-(.\‘) must satisfy
ttio orthogonality condition.

Note that upper case symbols U or CI’dcnotc displaccmcnts that vary along the Icngth of
the mcmbcr as a function of .Y, whereas lower cast symbols N’,,,,, or c,,,,, rlcnotc scalar
dcformation paramctcrs. Following the LSK approach, the mcmbcr strain cncrgy for
spccilicd number elongation c,,,,) and out-of-straight deformation II’,,,,, is minimized by an
appropriate choice of U(.t’) and W(S). whcrc U(X) denotes the axial displaccmcnts within
the member and bi’(.Y) must satisfy the orthogonality condition (eqn (2)). This Icads to
a unique and adequately smooth strain energy function. a, ,,,,(C,,,, N’,,,,,, If’,,,,, ). whcrc IF;,,,,
rcprescnts the imperfection of the member. At this point, the pcrtcct structure is con-
sidered so that Cp,,,,,= 0. Using a formulation for slender columns (in the sense that shear
deformations are negligible) but arbitrarily large displacements (Appendix R), yields
the following results for this member strain energy (b(,,,, and its derivatives evaluated at
“.(,,I) = IV,,,,, = 0 :

0
E.4
4 ,a,, (f’, 0. 0) =
I
2 ( ),,,,,e2.
EA
-I!.
(b,“,,,,(f~.O.O)=
() ;I ,nr,
1’. d ,,,,,.,,(f~.O.O)= -7‘
, ,I,,
, (3)

(b,,“)JC,O,O) = 0. (b,“,, (L’.0-O) = 0.

,,.I I
(4)

At the bifurcation point for the member under axial compression. (I,,,,, = -e,(,,,. where
E rc’(s,,,)‘L~,,, is the member shortening at buckling,
C’C(,“) and .r(,,,, E (I/A),!,,f/L,,,, is the
slenderness ratio, the following expressions are obtained :
 Worst shapes or imperfections 2317

(6)

where E,,,, = IL’&,,,)’ is the magnitude of the axial strain at criticality.

Equations (3)-(6) contain all the information that is needed in the subsequent analysis
of the overall system. They are based on an exact formulation for a compressible column
(Appendix B). However, the factor I +O($) will be dropped in the subsequent analysis
since E: is very small compared to unity in most applications. Using the corresponding
exact expressions given in Appendix B would introduce no additional difficulty into the
analysis. However, approximations introduced by neglecting shear deformations of the
buckled members would remain, and are probably of no less importance than the O(G)
terms in eqns (Sb) and (6).

3. ANALYSIS OF OVERALLL SYSI-EM (PERFECT CASE)

Let L’denote the collection of all joint displacement vectors for the structure, and
\I’ = ( . . . , w,,,,. . . .) be the collection of all member out-of-straight deformations. All these
quantities can be collected into a vector u = (0, w), which fully defines the configuration of
the structure. If the loads are applied at the joints only, and their magnitude and direction
is a function of I, only. the total potential energy can be written as

whcrc P(%) is a linear operator such that P(l)c: is the work done by the applied loads
through joint displacements I’. ti dcscribcs the impcrfcctions for the cntirc structure, and
the summation is carried out over all mcmbcrs !II in the structure. In evaluating the
dorivativcs (or variations) of this potential energy, the following functional dcpcndcncics
must be considcrcd :

where t’(,,,)is the diffcrcncet in the joint displacement vectors at each end of member m.
The Gatcaux dcrivativcs (SW Appendix C for detinition) of these quantities are

L’,,” ,.,‘SU
= SL’,,), 11’,,“).&’ = SW,,,. (9a,b, IO)

O,,” ,,,‘,‘S2l’SI/’ = f fs2P,,,,,*s,U(,,) - !-(K

11 WI1 *~2I’,,,,J.q,,,,
‘S,qn,,). (I 1)

where Sr. S,r. etc. and SW. S, W. etc. denote variations in the displacements ~1and I(
respectively ; Jr,,, is the difference of the variations in joint displacement at each end of
member m ; 6w,,, is the variation in the out-of-straight deformation of member m; x~,,,,is
the ditfercnce in joint coordinates at each end of member 111in the deformed configuration ;
and I = (+,, * x,,~,) Ii2 is the deformed length of member m.
With the aid of eqns (7)-(9). the Gateaux derivatives of the potential energy can be
evaluated :
@.,Su = c N,,,e,,,,,.dr- P(l)bu, 4.Av = 1 (b,m,.hq,“,. (12. 13)
m m

4.cy&u6~u = ~~,m,.n(e,,,,d,v)(e,,,,6,u)+~N,,,e,,,.”,62ud,u. (14)

m
tSuch differences in displacements at each end of the member should be taken by defining a beginning and
end for each member. and always subtracting the displacement at the beginning from that at the end.
where iv,,,,, = b,,,,, is the axial force in member m.
For the perfect structure with no out-of-straight deformations of the members. sub-
stituting eqns (3) and (4) into eqns (13) and (15) yields

&,,r(~*, 0, 2, O)li~ = 0. (b,,. (r.0, i..o)d\,61. = 0. (30a.b)

In view ofeqn (ZOa) the equilibrium conditions. cb,,,du = 0 ($,,.iir+4..3~ = 0). are satisfied
if I(’ = 0 and $,,.Sr = 0. Therefore the principal solution must be of the form :

II = &i.) = (:!(j.).O), C’
,,,,, = S ,,,, ,(i.), :V,,,,, = i’,,,,,(L) (21)

(11Ill
and tlic equations for the buckling modus 1;: E ( IX, )I’) red ucc to

(Jl
C/~,,-,.‘;!
A- = 0 v ii~, fb,,,,, b(’SW = 0 v &,. (7’)

Il‘thc operator (I),,,. bccomcs singular. global buckling modes woultl dcvclop. This C;I\C
is not consitlcrd hcrc. Instead it is assumed that (b,,,, remains positive dclinitc, while ;I set
of rncmlxrs IIIE .\I reach their buckling load at L = j.,. III this GISC the folhnving rcl;ltions
apply for iill mcrnbcrs ttt fz ,CI :

s,,,,,(i.,) = -G,,,,,, ~,,,,,(j.,) = -NC,,,,,. h,j.s,, = 0. (2.1)

whcrc A’, ,,,,, = n’(.s’E.4) ,,,,, is the buckling load for member IH. From the last condition. and
cqn (IO) it is seen that ;I typical buckling mode involves out-ol-straight tlcl~orm;~tions ot
some mcmbcr in .\I only, and c;m be writtsn as

(1)
I, = (0, :;!), Y; = (. . .s ,,),,, .), z,,,,, = OV’Irl # i. (174)

f:or c(>rlvcnicncc. the mode identifier i is taken to bc the number of the mcmbcr that buckle\
for mode i. The mode normulizution condition, eqn (AZ), reduces to

where ;I dot placed itbovc any quantity indicates that this quantity should be evaluated on
the principal branch (as ;I function of 2 only), then dilTcrcntiatcd with rospcct to E. and
evaluated at i. = i.,; 3,, is the Kronccker deltat ; and b0 is an arbitrary positive constant.

Choosing $,, determines the viIlues of :;!,,,. In order to facilitate the interpretation of the
algebraic results. this normalization is defined via the properties of a rcfcrcnce member.
which may. but need not, correspond to any of the members in the structure. Thus

tThcrc arc no imphcd summattons for rcpccatcd indtccs throughout this paper.
where any member properties identified by a subscript zero in parentheses correspond to
properties of the reference member. For linear prebuckling behavior. -&l?,,, = NC,,,. and
therefore c,,, = L,,,. for any member i whose properties coincide with those of the reference
member.
It follows from eqns (24). (IS) and (5a) that c#J:~,‘;:‘ii ‘:i vanishes for all i. j. k. This
indicates a symmetric bifurcation. Higher order postbuckling coefficients ‘#J,,~,defined
in eqn (A6) (Appendix A) are therefore required to determine the directions and curvatures
of the bifurcated equilibrium branches. The postbuckling coefficients in turn de-
pend on second order displacements. w). To determine
%’ = (‘?. ,j’! these quantities. the
procedure outlined in Appendix A is followed : first the space of admissible displacements,
,-t. is decomposed as the linear span of spaces A, and k, where rl, is the space spanned by
the eipenvectors and A is a complementary space, such that the zero vector is the only
element that is common to both a,, and A’. The obvious choice for this complementary
space is

,j = (I, = (1.. 11.): II’,,,,) = 0 v nr E M). (27)

With the aid ofcqns (20) and (2-I). the gcncral equation for the second order displacements.
cqn (A5). rcrluccs to

(f~:“~;;!‘~;!+f~:,‘;!‘),i,,+((~
,,,,,c’tG’;I! = 0 v (5~. 5~)
+ ~~~;.*.‘~~!)SIV E ..i. (3)

whcrc ‘ii’ = (I;:‘, ‘ii!) E ri are the dcsirctl second order displnccmcnts. The variation with
rcspoct to II’ togcthsr with cqns (16) and (IX). and the observation that ,/I,,,,,,_,~> 0 for m$ ,C/

Icacls to ‘ii! = 0. The vari;ttion with rcspcct to ~1. and application of cqns (5). (17) and (26)
Icads to the conclusion th;,t I;!’ = 0 for i # j, and ,$’ are determined from

(N ,),, ,e,, ,,,.+ ~/P~,,‘:‘!‘)SL~ = 0. (“l))

N INI, s ,.(:‘i,,J2
cp;,,,<
),,,
4”
= ---. (30)
(-&,)

Thus the second order displaccmcnts ‘FJrepresent the joint displacements due to an initial
tension N,,,,, in the member i. as calculated by linearization about the critical point. (The
initial tension is trcatcd as if it wcrc caused by an infinitesimal thermal contraction.) By the
assumption that Ioc;~l buckling modes develop before the global modes, positive dcfinitcncss
ol’~b:.~is assured. Indeed. thisoperator represents the tangent stiffness matrix of the structure
with out-of-straight mcmbcr deformations restrained. All second order displaccmcnts can
bc obtained with only one factorization of this tangent stiffness matrix.
For the calculation of the fourth order postbuckling coefficients, ,#I,,,,. eqn (A6) reduces
to

(31)

Evaluating eqn (31) with the aid of cqns (5). (6), (I 7) and (19) yields

(32)
‘390 R. PEEK and N. THIA\T.\F>LLIL)IS

where

(33a.b)

represents the change in axial force in member i associated \vith the second order dis-
placements for the buckling ofmemberj. Symmetry of the coeEcients tz,, can be verified bq

substituting Sr = ‘i!’ into eqn (79). Hokvcver. in pencral neither the matrix [tr,,] nor the
matrix [-a,,] is positive definite.
Having calculated the required postbuckling coeflicients. nttcntion is now focused on
the bifurcated equilibrium branches for the pcrfcct structure. By application ofcqns (A?)
and (AJ). these can bc written as :

whcrc < is the path paramctcr. ,Intl i.: ;lnil r, must bc solirtiow lo

(35&b)

Solutions to thcsc equations can readily bc obtained as follows: (I) Partition the set ot
modes ,\I into Iwo mutually csclusivc ~~ricl collcclivcly cdiaustivc scls .\I,, and A/, , and take
1, = 0 V iE M,,; (7) solve the system of liner quations

,F,u,,J’, = I ViE.W,. (30)

for _I;, Jo M, ; and (3) obtain the solutions for this partition as

(37)

The solution is real only when all y, arc of the s;unc sign.
Regarding the stability ofthc bifurcated equilibrium branches, applying cqn (A7) Icads
to the incremental stability matrix

This matrix must bc positive dcfinitc to cnsurc stability of the bifurcated branches. Altcr-
natively positive scmidctinitcncss is ;I necessary condition for stability. This in turn requires

(39)

Violation of this last condition for sonic ie ;CI,, involves loss of stability for the member i
while it remains straight. This mans that the member must be carrying an axial load in
 Worst shapes of imperfections 2391

excess of its buckling load. Therefore eqn (39) will be referred to as the local. or member
stability condition. Solutions which violate this condition will be said to be locally unstable.
Locally stable solutions are not necessarily stable (since B,, still can have negative eigenvalues
corresponding to eigenvectors 1, with x, = 0 V i E M,). Indeed ali bifurcated branches on
which the load drops away from the bifurcation point (AZ< 0) are unstable.
Of particular interest is the postbuckling branch for which i2 is a minimum. since this
also provides the worst imperfection shape (see Appendix B). One way of determining this
branch is to find all postbuckling branches with the aid of eqns (36) and (37). and pick the
real solution for which A2is a minimum. This method quickly becomes impractical, since the
number of such solutions increases exponentially with the number of potentially interacting
modes. The preferable alternative is to solve the minimization problem given in eqn (AIO).
After substituting for the coefficients 4,,k, from eqn (32). this reduces to :

minimize

(JOa)

subject to

(40b.c)

where x, = (a,)‘. Despite the simple appearance of the minimization problem, it may have
many local minima, since the matrix of coefficients N,,is not positive (or negative) definite.
Indeed any solution of eqns (36) and (37) is a potential local minimum. Fortunately,
howover. the numerical cxamplcs of Section 5 suggest that the actual number of local
minima is much smaller.
It can bc shown, using the Kuhn-Tucker conditions for optimality of the solution to
cqns (JO) (SW Appendix D). that the postbuckling branch on which the load drops most
rapidly is locally stable in the scnsc of cqn (39). Marc gcncrally. local minima for cqns (40)
correspond to locally stable bifurcated equilibrium branches, and rice tww.

4. IMPERFECT STRUCTURE AND WORST IMPERf‘ECTION SHAPE

The imprrfcctions considered can be written as E = E& where I: is the scalar magnitude
of the impcrfcction and G is the shape of the impcrfcction. which will be normalized in
some fashion. The following types of imperfections are considered : (1) errors in the joint
coordinates in the rcfcrcnce configuration; (2) member misfit (a member is too long or too
short); (3) curved initial geometry of the members; and (4) eccentricities at the joints. For
the present case. when local buckling modes develop before the global modes, it is found
that the leading order effects are due to the imperfections (3) and (4) only. These
cffccts lead to a drop in load carrying capacity of the structure which is of the order E’ ‘.
The other imperfections [(I) and (2)], on the other hand, have effects which are of the order
t: or higher. and will therefore not be included. The analysis provides results which are
ilSylTlptOtiCillly Cxilct for small imperfections.
In view of thcsc considerations. the imperfection h can be characterized as

Ii = rt = (. . . , CV,,“),. . .) = ECV
= c(. . . , LV
(,,,,, . . .), (41)
where It;,., = Iv,,,,,(X) define a transverse geometric imperfection for member nt. with
rt ,m,= 0 V X representing a straight member with zero eccentricities at the ends (see Appen-
dix B for details). The projections of the imperfection shape [eqn (A9b)] which determine
the postbuckling behavior of the structure, are given by

N C(I)
G,,,, (42a-c)
J-~.~~~,,Nu(o,L,i,L,o,
 (43)

is the amplitude of an equivalent sinusoidal imperfection.

The worst shape of imperfection corresponds to z = r,. and hence to
---__
r?cr, -- li’t-;.,~~,,,)fv,,,,L,,,L,,,,
---__ .,_
hc,,,
..--____ r,,
f-w
It can readily be verified that the imperfection norm

satisfies the conditions outlined in the last paragraph of Appendix A. and leads to a worst
imperfection shape that involves sinusoidal imperfections of amplitude G,,, given by eqn
(44) for the members in hit, and no imperfection for the other members.
In summary. the procedure for findin, 17 the worst imperfection shape. :tnd the cor-
responding load drop is as follows: (I 1 obtain the solution for the principal brnnch using
cqn (I?) -and locate the bifltr~~lti~~n point i., at which mcmbcrs WE hf run& their buckling
load. (2) Calculate the second ordcrjoint displaccmcnts from cqns (29) and (30). (3) Obtain
ttic coctlicicnts (I,, from cqn (33). (4) Sotvc ttic quadratic: programming prohlcm, cqn (40).
(5) Catculatc rr = F: ,/s,. Finnlly. (6) the worst ir~~p~rf~~ti~~n shape is given by cqn (44).
and ttic corresponding load drop for any given amplitude of inipcrfcction by cqn (A IO).

Typical cxamplcs considered are shown in Fig. I for ptarur trusses and Fig. 2 for qx~e
trusses. Therein members in compression are shown as continuous lines, whereas those in
tension arc dashed. The mcmbcr flcxurat rigidities, Ef, are chosen such that all members
in compression reach their buckling load simultaneously. Consequently the number of
potentially interacting modes in every truss ecluats the number of its members under
compression in the principal solution.
The dimensions of the structures analysed are as follows: All members for the two-
dimensional trusses of Fig. 1 have length L, except the diagonal members for the rectangular
truss in. Fig: I b; which are of length ,,J~L. For the three-dimensional trusses of Fig. 2, the
chord members (i.e. those parallel to the :-axis) are of length L, and the cross-section is
such that the chords fall on a circle of radius L. Thus the length of the ties (i.e. the members
parallel to the X-Y plane) art‘ $3L for the triangular cross-section of Fig. 2a, and J?!L for
the tics parallel to the x or ~9axes in Fig. 2b.
The trusses are constructed from identical unit cells. However, the number of unit cells
and their arrangement varies: for example, the hexagonal truss of Fig. la has a ritdius of
2L. Hexagonal trusses of radii L and 3L (as shown in Table I) are also analysed. The joint
numbering system for these follows the same pattern established in Fig. la. Tab& I and 2
also illustrate the various arrangements of unit ~~‘11sconsidered for the rectangular truss of
Fig. I b, as well as the “antcnnac” of Fig. 2.
Tht: displacement boundary conditions for the two-dimension~ll structures are ibe
trated in Fig. 1. For the three-dimensional structures with tt bays. ail joints in the plane
through points A,,, B, and C, arc restrained in the :-direction. Sufficient restraint in the X-
and y-directions is also provided to prevent rigid body motions. Since these .Yand J restraints
do not produce any reactions, their exact configuration does not inlluence the results
reported here. The applied toads and reactions are shown in Figs 1 and 2. They do not
 Worst shapes of imprzrfections 2393

Aol A12 AZ! A41

2)c-

A00 AIO AZO AM

Fig. I, Two-dimensional example ~WSSES.

change direction. However the magnitude of the reactions shown in Figs I and 2 applies
only for the principal solution.
All members have the same axial rigidity, EA. However, to simplify the prebuckling
solutions, member sites that are shared by two unit cells (e.g. member sites on lines Aol-
A, ,, A,,,-A,?, Azo-Az2 and A3,,-AJz in Fig. I b. and member sites on planes through points
A,, B, and C, in Fig. 2) are occupied by two identical members. This results in identical
stress states in each unit cell for the principal solution, regardless of the overall size of the
structure. Although the hexagonal trusses can also be constructed from unit cells, hcrc only
one member occupies each member site; identical prebuckling forces, N(,,,,= -1, in all
members are achieved by appropriate choice of the external loads.
For the purpose of mode normalization, the length of the reference member is taken
to be L,“, = L. fiL. L, fiL for th e st ructures of Figs lab and 2a,b respectively, and the
rcfcrcnce member buckling load is taken as N,,,, = 4 = EA.c,, where .sE,the buckling strain
of the reference member, controls the slenderness of the members. For the hexagonal truss,
cEalso represents the buckling strain in all the members. For other structures, the member
buckling strains differ, but are still approximately proportional to G.
Following the discussion in Section 4 (see also Appendix A), the worst imperfection
shape for a given truss is determined by the bifurcated equilibrium branch of the perfect
 R. PEEK and S. TRIANTAFILLIDIS

structure for which il is ;I minimum. This will be referred to as the critical bifurcated
branch. It follows from dimensional analysis considerations alone, that the dimensionless
curvature of the critical bifurcated branch, i.?/E.,, depends only on C, (and not on EA or L).

Since the structures considered exhibit essentially linear prebuckling behavior (as long
as E, < 0. I), a standard Newton Iteration procedure provides rapidly converging solutions
on the principal branch of the perfect structure. Such solutions can also readily be extended
to the point where global buckling modes develop (i.e. ~b,~~becomes singular). The value of
the rcfcrcncc strain i,‘E/I at the point where the first global buckling mode develops is
shown in Tables I and 2 as e_. As expected, E,~ is seen to depend on the slenderness of the
structure IS ;1 whole. (Recall that. in contrast, E, determines the slenderness of individual
members.) Only trusses for which E, < cCgarc considcrcd, so that local buckling occurs
before the global bifurcation load is reached. and (6.,.t.remains positive definite.
To find the bifurcated equilibrium branch of the perfect structure with the minimum
curvature L2(j_z < 0). two different strategies are used: a direct approach and an opti-
mization approach. Although it cannot be asserted with 100% certainty that these methods
provide the global minimum. there is every indication that for the examples considered, the
solutions obtained are indeed globally minimal.
 H’orsl shapes of Im~rfecttons 2395

Table I. Results for two-dimensional examples

11 mode Aok. 8384 t- sym modes m N. y)

mode Aoh. 0384 c- sym modes m X. y)

mode AoBl, B384 c- Sym modes m 1. y)

EC- 1 mode AoBs, 0283 (e sym modes cn X, y)

EC- 2 $I&= -27 06 mode A082 (- sym modes MI IL Y)

F_.v “““I &I
,_ hc-
_ -2 88 mode 81C2. 82c2 /82c4. 83c4 (- SYm )

‘2-” = -‘99 mode 8Ic2,82c2/82c4.~3c4~- sym)

1 EC= 01 j+/+= -505 mode Ehct. 02c2m2c4, B3C4t-sym)

26 mode Blct. 82c2/82c4,63c4c- syml

EC- ’ I+/+= -896 mode 6106, cIc2 (a sym modes ln X. y)

I EC- 2 wt= mode gloaal

000 1 %I\= -2 88 mode 002, CtO2 (- sym modes tn X. y)
Ep 001 hfk = -2 89 mode 002, C2D2 (a sym modes tn X. VI
EC- 01 2::: = -3 05 mode 002. C202 t- sym modes tn 1. Y)
EC- 05 * - 3 95 mode C102, CtD2 t. sym modes in X. y)
EC- 1 hf \= -1626 mode 0102 (+ sym modes tn 1. y)

z-’ 2 wxc= mode global

Ec= 0001 &/ +=-227 62 mode 5 members tootlmlzatlon~
% .*
~~~~~~~~~~ 1 EC= 001 [hl’ht=‘Jo; --7 79 mode 6 memoers tootlmlzatlonl
._ _
p21”c= mode glooal
IQ&cc’ mode glooal
mode global
Imode alooal
$I&=-52 20 mode AmAm.A7tAso (* Sym modes m x

ILlh,=-s426 mode AmAw.AnAm (- sym modes m x

15 mode AmAuo.AnAuo t- svm moo** m x

Ec- 0001 [%/~,=-I0

v 1 1 mode 7 memocrs (ootlmlzatlon)
Ecq-o 150

The direct approach is based on finding solutions to eqn (36) for various choices of
the set 1LI,. Since the number of possible sets M, grows exponentially with the number of
potentially interacting modes, it quickly becomes impractical to consider all possible sets
LCI,. Furthermore, it was observed that the critical postbuckling branch typically involves
buckling of only a few mcmbcrs. Thcrcfore. only sets hf, that involve buckling of up to
five members are considered. Even then, the number of possible sets bIcI, (a sum of binomial
coefficients) can bccomc staggeringly high. For expediency, the number of possible sets M,
is therefore further limited to 500,000 starting with those that involve buckling of the fewest
members.
As an altcrnativc to the above direct approach. the quadratic programming algorithm
by Schittkowski (1986) is used to solve the problem defined by eqns (40). This proved very
efiicicnt even for rather large numbers ofinteracting modes (up to 100). For all the examples
considered, several initial gucsscs lead to the same final optimal solution. Moreover, for
problems with a rclativcly low number of interacting modes (up to about 40) where the
direct approach is fcasiblc, the results from the optimization method arc always in agreement
with those from the direct approach. In other cases (idcntificd in Tables I and 2 by the
word optimization in parentheses). none of the 500,000 solutions calculated by the direct
procedure were real and locally stable, and it was necessary to rely on the results from the
optimization approach alone.
 Table 2. Results for three-dimensional sumpies

- 00126
IO bays

Ecg = a

V
EC- 01 ~l~c=-sasl mode AoDI,CODI.DODI
c- I ~l~,=-a.9l4 mode AoDI.CODI
Ecs - 0 0237 EC- 000001 %I kc=-21 9 73 mod e A~D*.CID~D,D~
hl+-2t9a1 mode ANXCID$DID~
EC- 0001 $1 kc=-?20 eymode AJD~,CID~.DYDI
EC- 001 1 kc=-229 06 mode A~D~.C~D~,DID~
EC- 01 41 +J74 63 mode AIDI.C~D~.DIDI

EC- 00001 I%I _

I,=-443 55 mode A?Da.C?Ds,D?Ds
._
$1 At=-448 I16mode A7Da,C~Da,07Ds
J ;

Ecq - 000818
EC- 0001
EC- 001
EC. o,
%I ~~=-s00~~rn~~d~~~A7Da,~7Da.D7D1
,&I kc” mode qloaal
8 bavs . . L-I 1.X

The results for all I2 trusses considered are shown in Tables I and 2. for various values
of the parameter E,. For small values of E, (slender members), the curvature of the critical
bifurcated branch, -Lz/i.,, is seen to approach an asymptotic value (independent of E,).
This corresponds to the case investigated by Britvec (1973) and Britvec and Davistcr (1985)
in which compressibility of the members is neglected. Indeed, for a selection of the examples
considered by Britvec (1973). the current analysis gives results that are in agreement with
his. Even for essentially incompressible members, the slcnderncss of the structure as a whole
has ;I strong clyect on the postbuckling behavior and imperfection sensitivity. with Az:‘i.,
being larger when the structure as a whole is more slender.
For larger I:, (stubbier members). A2/L, increases in magnitude, and it becomes very
large as the local buckling load approaches the global (i.e. as E, -+ E,~). This is not surprising,
since b,,, bccomcs singular in this case, resulting [via cqn (29)] in large second order
displacements, which in turn [via cqn (33)] product postbuckling coellicitnts N,, of large
magnitude.
The dcpcndence of i.J& on E, is further illustmtcd in Fig. 3 for a I x 4 bay rectangular
truss made of four identical square unit cells each of side L(,,, and a I x I bar truss of
dimensions L,,, x 4&,,. Global bifurcation for these structures occurs at cLp= 0.0375 for
the I x 4 bay truss, and at c,.~= 0.0309 for the somewhat more flexible I x I bay truss of
 Worst shapes of imperfections 2397

0 2 bay antenna
0 lbayantenrla

0.00 0.01

CRITICAL REFERENCE STRAIN e,

Fig. 3. Results for rectangular trusses of overall dimensions f.,,, x 4L,,,.

the wnt’ overall dimensions. Clearly A,/& is seen to approach infinity as the local buckling
load approuchcs the global enc. This implies that the imperfection sensitivity bccomcs
intinitc under such conditions. It must bc borne in mind. however, thut this result applies
only for intinitcsimally small impcrfcctions. Actual imperfections arc finite. and so is the
drop in load carrying capacity they produce. Hcncc the limits of applicability of the
asymptotic results must bccomc vanishingly small as the local modes approach the glob;ll.
lindcr such circumstances. the interaction bctwccn local and global modes should bc
explicitly includctl in the analysis. bcforc any meaningful conclusions in regard to impcr-
fation sensitivity GUI bc drawn.
Similar results xc obscrvcd in Fig. 4 for a 2 bay space truss of the type shown in Fig.
3 with ~111equilateral triangular section of sidelength L,,,, and an overall length IO&,,,. and

0 lx4baytnJaa
0 lxlbaytruss

-I
0.00 0.01 0.02 0.03 0.04

CRlllCAL REFERENCE STRAIN EC

Fig. 4. Results for triangular section antennae of overall dimensions L,., x IO&,.

US 29:19-F
2398 R. PEEK and N. TRIA.I‘TAFYLLIDIS

a 1 bay space truss of the same overall dimensions. Global buckling for these structures
occurs at E,~ = 0.0 I65 and E,~ = 0.0148. Again. for the same overall dimensions, the structure
with a larger number of bays has the higher global buckling load. For essentially incom-
pressible member behavior on the other hand E.,/E., seems not to depend on the number of
bays that a given structure of fixed overall dimensions is divided into.
The members that buckle on the critical bifurcated branch of the structure are listed
in the last column of Tables I and 2 as the “mode”. In most cases a given mode is seen to
persist for a range of values of E,. However. changes in the mode do occur: there seems to
be a tendency towards buckling of a smaller number of members as E, increases towards
tzCp,but changes in mode that are not accompanied by changes in the number of members
that buckle also occur. for example for the hexagonal truss of radius L.

6. CONCLUSIONS

A methodology has been established based on the Lyapunov-Schmidt-Koiter

decomposition and asymptotic expansion, by which the worst imperfection shape and the
corresponding load drop for any truss-type structure with multiple eigenmodes that involve
buckling of individual members can be determined. This problem reduces to a quadratic
programming problem. in which the number of unknowns is equal to the number of
interacting modes. Although the computational effort for finding the global minimum with
lOO%, ccrtnintygrowscxponcntially with the number ofmodes, it appears from theexamplcs
considered that in most cases the global minimum can be found much more eficicntly. The
results arc asymptotically exact for small magnitudes of the imperfection.
The method is ctfectivc for csscntially incomprcssiblc mcmbcr behavior (the cast
considorcd by Rritvcc (1973) and Rritvcc and Davistcr (lOS5)] , as well as for comprcssiblc
mcmbcrs. The impcrfcction sensitivity incrcnscs with increasing mernbcr compressibility
(stubbier mcmbcrs). and becomes infinite as the local mcmbcr buckling load approaches
the bifurcation point for global buckling. Under such conditions. the range of validity 01
the current analysis is cxpcctcd to bccomc vanishingly small ; an analysis in which interaction
of local and global modss is explicitly considcrctl should be used.

.~~~X-rto~r/~~~~~~,,,r~~n,.r
-..Tllc au~l~orsMC indchtcd IO Dr Jinos Ltigi) of ~hc CIVIL Engineering Dcpartrncnt of the
Technical University of Budapest for idsntifying and making avail;tblc Ihc optimiration software used. ‘This work
was supported by the National Scicncc Foundation as Grant No. MSS-9103227 and CES-XX57002. IniIi.11 support
was ;~lso provided by UNISTRUT Space Frams Systems.

REFERENCES

Britvcc. S. J. (1973). 7711,S/trhiliry oj‘E/lsric S~..YICI~. Pcrg;lmon Press, New York.

Britvec. S. J. and Davistcr. M. D. (1985). Post-buckling equihhrium of hypcrsfatlc Iatticcs. J. .G!yrty !\/cL./I. ..l.SCE,
111.2X7-310.
Budiansky. 8. (1974). Theory of buckling and postbuckhng behavior of elastic structures. ..l[/r .,fpp/. .Ift,c/r. 14,
I 65.
Cas~afio. F. (1989). Buckling of viercndccl reticular shells. Steel Structures. ASCE Structures Congress. San
Francisco. May I-5.
Crawford. R. F. and Bcnton. M. D. (19X0). Strength of initially wavy lattice columns. Af.,f.d JI IX, 5X1 SX1.
Elyada. D. (1985). Structural analysis of impcrfcct Ihrcc-lcggcd truss columns for space structures applications.
Ph.D. Thssis. California Institute of Technology. Pasadena. CA 91125.
Koitcr. W. T. (1945). The stability of elastic equilibrium. Thesis. Tcchnischc Hooge School a~ Dclft (In Dutch),
English Translation : Technical Report No. AFFDL-TR-70-25. sponsored by Lockheed Missiles and Space
Co.. February 1970.
Koitcr. W. T. (1976). Current trends in Ihc theory of buckling. Sprposiunr on fhc-klittq q/‘S/rw~urc.~, sponsored
by the International Union ofThcorcIical and Applied Mechanics (IUTAM). llarvard University, Cambridge.
MA. June 17-24. 1974 (EditLul by B. Budiansky). Springer. Berlin.
Kondoh. K. and Atluri, S. N. (1985). lnflucncc of local buckling on global mstability : simplilicd. Iargc dcfor-
mation. post-buckling analysts of plane trusses. Conrpur. Srruc/. 21. 613 627.
Thompson. J. M. T. and tlunt. G. W. (1973). A Gcncral Theory of Elastic Stability. Wiley. London.
Triantafyllidis. N. and Peck. R. (1992). On stability and the worst imperfection shape in solids with nearly
simultaneous cigcnmodcs. Inr. J. .%/ids Sfrucfures 29, 228 l-2299.
Wright, D. T. (1965). Membrane forces and buckling in reticulated shells. Proc. ASCE Structural Division. Vol.
91. pp. 173-201. February 1965.
Wright. D. T. (l966a). A continuum analysis of double layer space frame shells. Publications oi the lnternatlonal
Association for Bridge and Structural Engineering. Ziirich. Vol. 26. pp. 593-610.
 Worst shapes of imperfections 2399

Wrtght. D. T. (I966b). lnstability in reticulated spheroids. experimental results and the effects of nodal imper-
fcctions. In Srrucrural Technology for Large Radio and Rudi~r Telescope S_sstems (Edited by J. W. Mar and H.
Liebowitz). pp. 369-379. MIT Press. Cambridge.

APPENDIX A: GENERAL RESULTS FOR A MULTIPLE SYMMETRIC BIFURCATION

This appendix summarizes general results from the Lyapunov-Schmidt-Koiter asymptotic postbuckling
theory that are used in this paper. These results apply when : (I) The structure is elastic with a sufficiently smooth
potential energy function d = d(u. i. ti). where UE A is the displacement field. 1 is the load parameter, and ric A
represents an imperfection. (2) The sets of admissible displacement and imperfection fields, A and A respectively.

are vector spaces. (3) A smooth principal solution i(1) exists which vanishes at I = 0 and satisfies the equilibrium

condition. &(:(A). i.,O)Su = 0. for all 3uc .4. and ia [O. vi’). where &’ > 4. (4) Equilibrium states on the principal

branch must be stable for rio [O. A,). in the sense that the bilinear operator. 4.,(1!(i). 1.0). is positive definite. (5)
At criticality (i = 4) this stability operator has a finite dimensional nullspace. A,. (6) The basis vectors for this

nullspace are buckling modes denoted by ‘:‘. and satisfy

d:,‘i~& = OVSuc A. &.i’d = 0. (Al)

where the superscript con the potential energy d and its Gateaux derivatives denotes evaluation at criticality, (7)
The buckling modes are orthonormulized so that

d
d> ,~.,,“&~), 1.0,:;’ 2’ =-cb.a,,.
1I-i<
where d,, is any positive constant. and can bc chosen to achieve a convcnicnt normalization.
The bifurcation is said to bc symmetric if

for all combinations of modus. i. j.li. In this cast the solution for the bifurcated equilibrium branches of the
pcrfcct structure is of the form

Z(f) = A,+ !I:C2+O(i’). 140 = i(A(C))+<~a,Y+

pCa,a,‘%‘+o(ly). (A3)
I 1.1
where : IS :I palh par:unctcr which is approximalcly equal to the distance ofthc solution to the princip;ll br;in&.
,‘,I
II :trc second or&r displacements to bc dclincd in cqn (AS). and the paramctcrs e, and Jr arc solutions to

-f12a,+~~(b,,,,a,aka,=0,
~(a,)‘=I,
AA.1 ,
(Ad)

whcrr the postbuckling cocllicicnts &I,,&,arc dctcrmined as follows: (I) Decompose the space of admissible

displacements II into th? space A, which is spanned by the buckling modes ‘II. and a complementary space ri,
dclincd so that A,, and .-I span the rpacr A. but the zero vector is the only element that is common to both A, and

.i. (2) Fond second order displacements, ‘8’~ A such that

(s:““‘;’ ‘0 + ~:;.‘ti’,sti = 0. v 3;; E /i. (AS)

Finally. (3) ths postbuckling coellicicnts can be calculated from

The bifurcated equilibrium branch defined by cqn (A3) is stable in the vicinity of the bifurcation point. if the
incrcmcntal stability matrix

&, = -&&, +Z&,e,a, (A7)

LJ
is positive dctinitc.
For most problems the potcntiul energy d can bc written as a sum of two terms: the first depends on the
displaccmcnts u only. and represents the strain energy of the structure; the other term reprcscnts the potential
cncrgy of the loads, and can be written as -IA(u). where A(.) is a linear function. If the structure is loaded by a
single point load of magnitude 1 and constant direction. then A = A(u) represents the deflection under the load.
Budiansky (1974) rcfcrs to A as the gcncralizcd load shortening. Using his approach to evaluate this quantity on
any of the bifurcated branches for the perfect structure yields :

A = Q(C)) = A,+ !do<’ + OK’). G-w

where A,, is the gcneralizcd load shortening evaluated on the principal equilibrium branch at 1 = A(<).
 2400 R. PEEK and N. TRIANT.\FYLLIUIS

Next consider an imperfection ri = ES, where E IS the scalar imperfectIon amplitude. and u’ IS the shape of
Imperfection. normalized so that

x(f,) = 1. where ;‘, z - ;.T”“.“‘LL (A9a.b)

L I
For small ma$tudes of the imperfection E. the behavtor tn the vximty of the bifurcation pomt depends only on
the projections ;,. The largest load drop (defined as i., - i,. where i., is the value of the load parameter at the first
limit point for the imperfect structure) for a given (small) amplitude of imperfection occurs for i”, = z,. where 3,
is the solution toeqns (A4) for which 1: is smallest. Assuming that this E., is negative, the corresponding normalized
load drop is

(AIO)

So far only the worst shape of the projections ;“, have been determined. In general there is an infinite set of
Imperftytion shapes c. all of which have the same projections $7. In order to obtain a unique solution for the
imperfection shape. it is necessary to define an imperfection norm. such that for any f, satisfying eqn (A9a) : (I)
all i satisfying eqns (A9b) also satisfy the condition 11; I/ 2 I. and (2) there exists a unigue ; which satisfies both
eqns (AYb) and the condition II& = I. This last lj is the worst impcrfcction shape if c, arc worst values for the
projections of the imperfection shape.

APPESDIX B: POSTBUCKLING ANALYSIS OF A COMPRESSIBLE COLUMN

Surprisingly. an exact initial postbucklinp analysis for an CI;ISIICcwnpross~hlccolumn appears not to he

;~v:~~l:~hle. The clowst that could be found is the analysis by Brilvcc (1073). in which an unnecessary assumption
rceardine the Jistrihution of slope along the column is ma&. The reason for this may be that a substantial amount
of algebra is Involved. Furthcrmorc, the results mcrcly conlirm that for slcndcr columns compressibility elTccts
xrc not important m the postbuckling ranpc. The assumption that the column is comprcssihlc in the prcbuckling
r:mgc. hut frcclcs :txially once the critical load is rcxhcd (Kondoh and Atluri. IYKS) is well justllicd. In&cd. for
sm;lllcr WIUCS of the slenderness ratio, when comprcsrihility c!Tccts might play some role. shear deformation wtl[
;11so bccomc important, and perhaps more so. Ncverthclcss. for the sake of completeness. a brief summary of
lhc fcrrmulation Jnd linal rcsulte from the postbuckling ;Inalysiu of a compressible column with specified end
d~~pl~emcnls (but no restraint against end ro(ations) by the I.SK tcchniquc is given hcrc.
I’lan.lr &formation in the .Y % plant IS consitlcrcd. whcrs the S-axis alwavs passes through the endpoints ot
rhc mcmhcr where the loads arc applied. The initial and tlcti>rmcd gcomctry oi the ccntroidal ;lxis of the column
;trc written xi

K(X) = .Ye, + W(X),,, r(x) = (X(1 +e/‘I.) + C’(.Y)le, + {G’(X) + W(X);+ (BI)

rcspcc~~vcly. whcrc er. e,. and e, arc the unit vectors in the .Y, t’ and % dlrcctions. Member quantities dclined in
the m;lin body of the paper are not redefined here. and the subscripts in parentheses identifying the member to
which they pertain are omitted. Thus, forcxample. ~deno~ the spcciftcd mcmbcr elongation. The set ofadmissible
dlspl;tccmcnts /I includes all displacement functions (U. IV) that vanibh at the endpoints. X = 0, f.. For the
momcnl. the imperfection w is also assumed to v;mish at the endpoints. The measures of axial and bending
dcformatlon. t and X. are taken to be

where ;I prime denotes ditierentiation with respect to .Y; and /7 = fl(.Y) and 0 = O(.V) &note the rotation of the
longitudinal ilxis of the column bcforc and after deformation rcspcctivcly. (The rotation produced by the loads
is I) IT)
The potcntr;ll energy of the column is given by

whcrc El = .szL’E.4 is the llcxural rigidity of the column. The aryumrnls of this potenlial energy functional are
enclo.scd in square brackets, to distinguish the functional I/J[.. . . ., .] from the rcduccd potential energy function
I)( ,. ,. ) dciincd later. for which the arguments are enclosed in parenrheses instead of brackets. Stationarity of the
potcnlial cncryy with respect to U and W Icads to

EAct = constanl. t= ii:;ilr’, n=txe,.

which coincides with the equation that would be obtatncd from equilibrium considerations, if the axial force and
bcndrng moment at any point in the column are taken to hc MC and EIx. respectively. This illustrates the
advantage of using the deformation measures. e and Y. defined in eqn (BZ). Other deformation measures would
product much the same final results as long as strains in the column are small. but they would also lead to different
(and potentially more complicated) equilibrium and/or constitutivc equations.
 Worst shapes of tmperfcztions 240 I

The principal solution is ti = B’ = 0.

The first bifurcation occurs at c = -e, = -&L. where the buckling strain. e_ is given by

$ = !(I-Jizz). (B5)

This exact expression is within a factor of (I +O(E,)) of the approximation E<= n’s’ used in Sections 2 and 3. The
buckling mode is

,0 , I,
C(X) = 0, W(X) = sin(IrX/L). (86)

In following the LSK technique the set of admissible displacements A is decomposed as the linear span of the
eigenspace rl, and a complementary space A as follows:

.!.
A,=;(G’.W):~=O.W(X)=~~sin(n.Y,‘L)i. i= (u,.: W(X) sin (nX/L) dX = 0 , (B7a.b)
I I 0 I

where w is the scalar out-of-straight deformation of the member. Using this decomposition. the reduced potential
energy can be defined as

Although the authors are not aware of any closed form solution to this minimization problem. it is possible to
resort again to an asymptotic expansion to obtain the first few terms in a series representation of the reduced
potential energy. or equivalently. to obtain the derivatives of the reduced potential energy evaluated at the
bifurcation point. After a substantial amount of algebra (some of which was done with the aid of a computer and
the MACSYMA program) the results obtained for the derivatives of the rcduccd potential energy evaluated at
u* = rf’ = 0 are those given in cqns (3) and (4). These results are exact. Additional dcrivativcs of the reduced
potential energy arc given in eqns (5) and (6). Where thcsc results contain a factor I +O(r$. the corresponding
cxnct results are :

(Boa c)

The approximations in eqns (5) and (6) are within ;I factor of (I +O(ci)) of these exact values. Further approxi-
mation in czn (6) would be possible while maintaining accuracy to within a factor of (I +0(k)). tlowcvcr, in this
case. the results for the postbuckling behavior of the column would reduce to those for the moderate deformation
theory. which predicts that axial load remains constant with increasing displacements after buckling occurs.
Finally. some discussion with regard to the assumption that the imperfection #‘vanishes at the endpoints is
in order: since the formulation is based on arbitrarily large imperfc%tions, it remains valid for an imperfection
that contains rapid changes in #‘in the vicinity of the endpoints. Thus end eccentricities can also be represented.
and eqn (43) is still valid for calculating the equivalent sinusoidal imperfection. Note however that this analysis
applies for link elements (i.e. the elements with length equal to the eccentricities that provide the connection
betwcxn the column end points and the point of load application) consisting of
a short segment of the column.
If these links are to be replaced by rigid links, an alternative formulation is required. since this introduces a
nonlinear essential boundary condition, as a result of which the set of admissible displacements of the centrotd
of the cross-section no longer form a vector held. Such an alternative formulation leads to the same result as eqns
(BY) and (43) to within a factor of (I +O(E,)).

APPENDIX C: DEFINITION AND NOTATION FOR GATEAUX DERIVATIVES

Let F: (Ux V) * W be a mapping, where ti. V and B’ arc vector spxes. Assuming suitable smo0thness.t
the limit

F *u a lim F(u + E SK cl - OK 1’)

_I ,_.,) ----= E
1g F(UfE Su,c)
1‘-0
(Cl)

exists. and is linear in Su. Thus the Gateaux derivative F,. is a linear operator. For fixed 6~. another mapping
(F,,Su) : (Ux V) + Wean then be considcrcd and the above definition can be reapplied to define higher Guteaux
dcrivativcs. For example,

Note that the nth Gatcaux derivative is an n-linear operator. All n arguments always follow the Gatcaux dcrivativc,
except that any scalar arguments (corresponding to the case when f./ and/or V is the set of real numbers) arc

tNorms are required on the spaces LI, V, B’to define the notions of smoothness and existence of the limit.
?‘a? R. PEEK and N. TRIASTAFYLLIDLS

omitted from the list. These may be Included as scalar multtpliers at any other location. since m thus case the
Gateaux derivative reduces to the usual partial derivative. The order of the subscripts is always opposite to that
of the corresponding arguments of the n-linear operator.
Although no norms appear explicitly in this definitron. addressmg extstence of the Gatraux drrtvattvr dots
require norms. Frechet derivatives could also have been used throughout this paper. However their detinition
does require the explicit use of norms, and is less suggestive of their meaning or method of evaluation. That is
the only reason why the Gateaux derivative is preferred here.

APPENDIX D: PROOF OF LOCAL STABILIT\r

Suppose the inequality constraint in eqn (4th~) is replaced by v’, 2 h,. and that these inequality constraints are
active for all in M,,. If b, is increased for some i+s .M,, then the inequality constraint becomes more restrictive and
the minimizing 2.: should increase. Thus a necessary condition for optimality is Si2, I(%, > 0. Thts dertvative IS
given by ?,ir,‘?b, = -p,, where JI, (i~,\f,) is a Lagrange multiplier. which is determined by requirmg that the
Lagrangian.

be stationary with respect to .Y, (IE ,)I). and the Lagrange muitiplters /i, (io ,M,,). and /I. Stattonartty wtth respect
to s, (ic ,M,) recovers eqn (35a). and leads to it = -2E.,. With this result. stationarity of the Lagrangian wi!h
respect to I, (ic M,,) produces.

Thus the lou;tl stability condition, cqn (39). holds for the postbucklmg branch which mintmires /:!. ;ts requtrcd.