AN INCREMENTAL APPROACH TO THE SOLUTION

OF SNAPPING AND BUCKLING PROBLEMS-t

National Aerospace Laboratory, Anthony Fokkenveg2. Amsterdam-1017,Holland

jorpublication
(Received 16 June 1978; received 28 November 1978)

M-This paper is concerned with the numerical solution of systems of equations of discrete
variables, which represent the nonlinear behaviour of elastic systems under conservativeloading con-
ditions. In particular, an incremental approach to the solution of buckling and snapping probkms is
explored.
The topics that are covered can be summarizedas follows:-The computationof nonlinearequilibrium
paths with continuationthrough limit points and bifurcation points.-The determinationof critical equili-
brium states.
Characteristic to the proceduresemployed is the use of the length of the equilibrium path as control
parameter.This feature, togetherwith the secondorder iteration methodof Newton, offers a reliable basis
for the proceduresdescribed. Actual computations.carried out on a finite element model of a shallow
circular arch, illustrate the effectivenessof the methodsproposed.

I. INTRODUCTION

The elastic stability analysis of plate and shell structures by means of the finite element method
or finite difIerence method inevitably involves the solution of large systems of nonlinear
equations. In principle, two broad classes of solution methods are available today. One consists
of computer adaptations of a perturbation method developed as far back as 1945. by
Koiter[ 1.21. Examples of this approach in terms of suggested, or actual programme im-
plementations are given by[5-91, in particular[7] and [8]. The other approach, in existence since
the evolution of the present day computer, covers methods which try to solve the equations
step by step, i.e. in a pointwise fashion. Examples of this mode of attack can be found
in[1&22]. The present paper is exclusively devoted to this second type of approach.
It is a consequence of any discrete formulation (e.g. the finite element method) that
the deformation of a given structure is described by a set of (N) deformation parameters, also
called generalized coordinates. In this context the loaddeformation history of a structure
presents itself as a curve in a (N + 1) dimensional space spanned by the deformation
parameters and the magnitude of the applied loads. Such a curve is usually referred to as
equilibrium path or deformation path. The problem of elastic stability is intimately connected
with singularities that occur somewhere along the path under consideration (usually the path
which is connected with the undeformed state of the structure, the so-called primary path).
These singular points are better known as critical points. Well known is their classification into
limit points and bifurcation points, geometrical concepts that are connected with the physical
concepts of snapping and buckling respectively[3]. In principle, the stability analysis of a
structure or structural system consists of a computation and evaluation of the critical point of
the path that is considered to be relevant to the problem at hand. To be able to carry out such
analysis, a computational procedure should have the capability: (1) to compute the critical
points, i.e. limit or bifurcation points, (2) to trace parts of the path or path’s (branches)
connected with these point&
The numerical strategies set forward in this paper are designed to meet these requirements.
They are based on an incremental method proposed earlier in[14,15] which features Newton’s
method and a special parameter controlling the progress of the computations along the
equilibrium path(s). In the geometrical terms the control parameter selected corresponds, in
good approximation to the “arc length” of the equilibrium path to be computed. It is introduced
by means of an auxiliary equation which is added to the set of equations governing the
tPaper presented at the 14th ht. Gong. tireticol and Applied Mechanics. Ddft, The Netherlands, 30 August4
September 1976.
SThis is another way of sayingthat the methodshouldhave the capability to computepost-criticalquilibrium states.

529
!3sVd IX No.7-a
530 E.hKS

equilibrium of the structure at hand. An important advantage of this particular choice is that the
modified set of equations become singular only at a bifurcation point. This means that the
domain of application of the method is increased considerably in comparison with that of more
conventional methods.
The present paper is a shortened version of [16lt where the application of this method to the
problems of the type (1) + (2) was studied in detail. Just as in[16], the emphasis will here be put
on an exposition of possible strategies and an evaluation of their feasibility.
The reliability of Newton’s method in procedures of the type considered presently is well
known[23], and it is this property which prompted its choice. A serious drawback is however,
that it requires a large volume of operations when the dimension of the system of equations is
large. This detracts, at first sight, much from the practical significance of this method in
connection with the solution of large scale problems. Nevertheless, the use of Newton’s method
makes it possible to deal with the question of effectiveness first which is the main objective of
this paper. It should be reahsed however, that improvements in the economy are possible by
measures which do not impair the intrinsic advantages of the methods presently proposed.
Some of these modifications are mentioned, but a directed discussion of possible ways to
improve the computational efficiency will be deferred to a future paper.
The references[lO-13, 17-221 constitute a sample of contributions to the present subject,
which were readily available at the time [ 161was completed. The principal diflerence between
the present paper and the references mentioned is the difference in choice of the control
parameter in the computations. In (10, 111a family of control parameters is considered, which
are suitable for the continuation of the computations through limit points. However, the
selection of the one to be applied is left to depend on the particular problem at hand. In the
examples presented, the parameter chosen is called for only in a close neighbourhood of the
limit point. In the other references, the difhculties connected with limit points are either
circumvented by a “jump” technique(221, or solved by replacing the load by one of the
displacement variables as control parameter(l2, 13, 21, 221. Predictorcorrector schemes are
applied with either Newton’s method as corrector[lt&l3, 17, 181 or a modification of this
method: the so-called direct substitution method or modified Newton-Raphson method[1%22,
291. The question of continuation onto the branch of a bifurcation point is discussed in[13,221.
The subject of the present paper will be preceded with a summary of some elementary
concepts of elastic stability theory for as far as this appears to be necessary for the
developments that follow.

BASICCONCEPTS
2.1 Equilibtium equations
The class of problems considered here concerns purely elastic structures under conservative
loading conditions. It is weU known that for these structures a potential energy function can be
defined from which all pertinent information as regards their behaviour can be derived. It is
assumed here that the kinematical configurations of a structure are described by a finite set of
generalized coordinates:
r-= ]t,, rZ.f3.. . ., INI (2.1)

which are also referred to as displacement variables or deformation parameters. As is

customary in the theory of elastic stability, the loading of the structure is supposed to be
dependent on a single intensity parameter, here denoted by p.
Let the potential energy be given by the function

P = P(t; p). (2.2)

The equilibrium of the structure is then determined by a stationary value of this function.
Stationary values of the energy P are governed by the set of equations

f(f*p)=dP(i;p)=O
I
i=l 2 N
(2.3)
9
ati ’ ’ ‘*-. *
YThis report was distriiutcd only on a limited scale.
 An incremental approach to the solution of snapping ad buckling problems 531

In general, these equations are nonlinear in i and pi and this is always assumed to be the case in
the sequel. The coordinates t, are supposed to be defined in such a way that the undeformed
state is the reference state

p=O; i=6. (2.4)

The configuration [p; i] of the structure can be visualized as a point in a (N + 1) dimensional

Euclidean Space RN+,, with an orthonormal basis

[iq = [&, P,, &. . . . ipf]. (2.5)

This point is then given by

Here and in the following repeated indices imply the well known summation convention.
In general, more than one solution i exists at a given value of the load p The solutions vary
when the load p is varied and describe curves in R,,,+I, the equilibrium paths of the structure.
They can be described in parametric form by

P = p(q);i= i(q) (2.7)

where n is a suitably chosen path parameter. In the physical model of the structures considered,
the load is prescribed, which implies that p takes the role of path parameter

p = p; i= i(p). (2.8)

For the mathematical model the choice (2.8) is not always appropriate. Here, it will prove to be
more useful to choose q to satisfy

(2.9)

This defines 11as the arc length of the curve (2.7) and it will be denoted by “s” in the sequel.

2.2 Critical points

According to the theory of stability of conservative systems, the stability of the equilibrium
configurations (2.8) is ensured if the quadratic form

- 1 a2P(i; p)
P2W = 2 at,at, A&At/ = f fuAt,~t, (2.10)

is positive definite. They are unstable if PaAl’> is indefinite, while the transition between a
stable and unstable point of a path is marked by a positive semi-definite Pz(Ai). This point of
the path is called critical state and it will usually be denoted by an asterisk.

[P*, PI = MS*). iW1. (2.11)

The coefficient matrix of the quadratic form PdAi)is symmetric. Using the notation for
partial differentiation introduced in (2.10), this important property is expressed by

fu = fu. (2.12)

The precise definition of a critical quilibrium state is given by means of the generalized
eigenvahre probkm
_
(flJ(r; p) - 4WfjIq(k) = 0. (2.13)
532 E. hK.9

Here, a’(k) denote the characteristic vectors and o(k) the characteristic values of rrrj] at a given
configuration [p, ;I with respect to [TLjJ: a positive definite, but otherwise arbitrary matrix?.
This freedom in the choice of [Ttil is only of .minor importance in the present paper. With the
exception of Section 5.1 eqns (5.2), [Y&J]will here always be equated to the identity matrix, thus

6.j = S,( = the Kronecker delta). (2.14)

The characteristic values o(k) are supposed to be arranged according to the sequence

o(1) 5 w(2) S w(3) 5 * * . * 5 o(N). (2.15)

The quadratic form PAAi) is positive definite, and therefore the equilibrium configuration
[p(s); $)I is stable as long as the characteristic values o(k) satisfy

(2.16)

The stability limit is reached when

0 = w( 1) = w(2) = * * * =w(K)<w(K+l)~...~~(N) (2.17)

thus when one or more eigenvalues w(k) in the sequence (2.15) are zero. In this paper the
restriction will be made to discrete critical points only, i.e. to points defined by

0 = o( 1) c o(2) 5 o(3) S * * * * 5 o(N). (2.18)

In the context of the theory of elastic stability of conservative systems, (2.17) and (2.18) are
unambigous definitions of a critical state. There is another criterion which is useful in
applications but which can only be appreciated as a necessary condition. The determinant

D = det(fiJ( 1; p)} (2.19)

as is wellknown, can be computed from

D= [fi o(k)]det{TJ (2.20)

It follows that at a critical state

D = D(r’i: p*) = 0. (2.21)

The criteria (2.18) and (2.21) will prove useful instruments for the determination of critical
points.

2.3 Limit point and bifurcation point

It is well known that in general, equilibrium becomes critical in either a limit point or
bifurcation point (FM. 1). These geometrical concepts follow from the analysis of the behaviour
of the solutions (2.8) in the neighbourhood of the critical points (see, e.g.[3,41). A
complete analysis of this nature is extensive and falls ouiside the scope of this paper. In what
follows it is only necessary to know the criteria which characterize the diflerence and to have at
the disposal a means to compute the direction vector of the branch of a bifurcation point,
whenever such point is encountered. The present and the following section will deal with these
questions.

tit is convenient to consi&r the vectors d(k) normalised accordii to Id(k)1= 1, whcrc( 1 denotes the
Euclidean norm: (a( = (66)‘” = (u&In. Note that the inner product will be denoted by either (ad) or (a&,).
 An incrementalapproachto the solutionof snappingand bucklingproblems 533

The properties of the path are described by the path derivatives

$1
I&i il= [g.
(2.22)
[ii,
t]=[2, $1, etc.
These vector quantities are determined by the following equations

$ fji(ii p)] = f,i, + fi$ =0 (2.23a)

$ [fdi:PII =fi.i +fi.0b +ti.,ii+hd = 0, etc. (2.23b)

Equations (2.23) are not complete without

&i + ihi,, = 1, (2.24a)

fi + ihih= 0, etc. (2.24b)

which side conditions follow from the definition of the path parameter s, see eqn (2.9). It is
noted that in (2.23) the convention fLo= t3fJap is introduced. Equations (2.23). in conjunction
with (2.24) should hold everywhere along the path [p(s), i(s)]. However, at the critical point,
VU] is singular in accordance to definition (2.21). This means that the non-homogeneous parts of
eqns (2.23) should satisfy the orthogonality conditions

=0
~~f(~)fi,obl,-s* (2.25a)
[ai(l~~o~+~~ii+f~~}l.~,~ =0 (2.25b)

where the asterisk indicates that the evaluation is made at a critical statet.
It follows from the first requirement (2.25a), that two cases should be considered separately

&i(P) = 0 (2.26a)
f#)a,( 1) = 0. (2.26b)

In the general theory, these conditions are known as the condition for a limit point and
bifurcation point respectively. Of course, they should be read in conjunction with (2.18) or
(2.21).
An equivalent formulation of (2.26) can be presented which does not contain the eigenvector
6(l). It is first noted that eqns (2.23a) can be rewritten in the form

LIi;= dgi (2.27)

where the components di of the vector d’ are defined by

d, = - $),,o [14-161’ (2.28)

The transformation of (2.23aH2.27) is accomplished by means of the matrix

[A,1= [g] (2.29)

tEquations (2.25) arc sometimes referred to as consistency relations.

 E. ftIKS

DISPLACEMENT

L._._. __ .-
DISPLACEMENT

Fig. I. Problem types.

which is the matrix of cofactors of D. Note the relationship

g * fik = OS,. (2.30)

As before, examination of the transformed eqns (2.27) subject to (2.18), leads to the conclusion
that two different situations should be considered.

b(s*) = 0 (2.3la)

4W) = - aofco =o.

&
(2.3lb)

In the form (2.3lb) the bifurcation condition is less known than (2.26b). However, it will
prove to be a useful relation in what follows. A proof of the equivalence of (2.3&o-(.26) is given in
the Appendix.

2.4 The tangent to the branch

As was mentioned earlier, the character of a critical point can be exposed by studying the
solutions of (2.23) at such a poitt. At present it will suffice to restrict the considerations to the
solution for the tangent [@(s*), t(P)]. Only a sketch of the derivations will be given.
The computation of the_tangent at the critical point is carried out on the basis of the
observation that a solution t of (2.23a) can be written as the sum of the particular solution py’
and an arbitrary multiple of the eigenvector d(1) associated with the critical state, thus

2=liy’+/&ti(l). (2.32)

It is noted in passing, that according to this proposition

hJW)Y, +h,oo*) = 0; hW)qW = 0; YiUi(l) = 0. (2.33)

When expression (2.32) is subjected to conditions (2.26a) and (2.24a), it follows that

ii(S*), ;-(P)] = (0, d(l)]. (2.34)

Consequently, at the critical point defined by b(P) = 0, the tangent vector corresponds to the
eigenvector 6( 1).
In the second case (2.26b). the situation is more complex. The general solution (2.32) is now
not determined by (2.25a) alone, but also needs the side condition (2.25b) for its completion.
 An incrementalapproachto the solutionof snappingand bucklingproblems 335

The result is that two solutions emerge

where, for convenience 0’ is written for d (1). The specification of by, PII. PI, WI, y’ will nqt be
given because the explicit form of the solution (2.35) is not of interest here. This follows
because once the bifurcation point is computed, one of the solutions (2.35) will automatically be
provided by the methods to be discussed in this paper. It is advantageous therefore, to use this
tangent vector in the construction of the other. Note, that it is possible to express one of the
two solutions (2.35) as a linear combination of the other and the eigenmode d. A convenient
choice of such construction is

@I, &I = a[ILfh*

Pi + 4.

If [b f] is known explicitly together with the eigenmode 0’.(I and p can be computed from the
requirement (2.25b) and the normalizing condition (2.24a). The result is

(2.37)
0 = ][jQj,,c(;; + a]]-’ = {$ + 2@;; + l}?

It is finally noted, that a detailed derivation of the results shown in this section can be found in
the appendix of [ 161. Only the presentation of the result (2.36) is modified with respect to the
one shown in[16]. This is done here to allow for the inclusion of the special case: aif@+& = 0.

3. THElNCREMENTALSOLUTIONOFEQLJILIBRIUMPATHS
3.1 General strategy
In general, eqns (2.3) admit a number of solutions t’ for a given value of the load p but
usually only a few of these solutions are important to the stability analysis. For the time being,
attention will be focused to solutions r which correspond to a continuous deformation of the
structure from the undeformed state. Equilibrium states thus obtained belong to the so-called
basic or fundamental equilibrium path of the structure under consideration.
Iteration methods for the solution of nonlinear equations like (2.3) require, as is well known,
a starting configuration (or initial iterate) which is “close” to the solution to be determined. This
requirement fits well into the scenario of an incremental procedure which is designed to solve
an equilibrium path in terms of a sequence of successive, but distinct points. Each point
obtained offers a means to construct the starting configuration for the next to be computed. The
accuracy of the “initial” iterate can be controlled by keeping the distance between the known
and the still unknown point within certain bounds.
Two well known strategies are depicted in Figs. 2(a) and 2(b). In the first case, the load
parameter p is used as the prescribed variable (denoted by 7). In the second case, one of the
deformation parameters TVis taken to fulfill this role. Each point computed by the first method,
is determined by the intersection of a surface p = constant ( = r)) and the equilibrium path.
governed by eqns (2.3). A point computed by the second method, is determined by the
intersection of a surface 1x = constant ( = 9) with the same solution curve. This explains why
both methods (a) and (b) break down in the neighbourhood of the turning points in Figs. 2(c)
and 2(d) respectively.
The points in question are characterised by

dp
fi=s=O for p=n*

tThere arc more solutionswhen tbc bifurcation point is no longer distinct.

536 E. RIKS

EXAMPLE 1 EXAMPLE 2

(a:

PITFALL5

(C 1 (d)

Fig. 2. Examples incremental solution methods.

and
r+=o for rK=n** (3.1)

where it is noted, that the first definition given corresponds to the condition for a limit point. In
either case, incrementing the value of TIbeyond the limiting values (3.1) leads to specifications
of the surfaces
fo=p-?r=o
fO=fK-t)=O

which do not define intersections with the equilibrium path, at least not in the neighbourhood of
the points defined by (3.1). The breakdown of the procedures described does usually not occur
suddenly, but is announced some time in advance by a marked increase in number of iterations
necessary to obtain converged solutions. There is some evidence [ 14, IS], that this phenomena is
coupled to the decrease in quality of the intersection of the surfaces (3.2) with the equilibrium
patht when the critical points (3.1) are approached.
A measure of the quality of intersection is given by 8, the angle between the tangent of the
equilibrium curve and the normal to the intersecting surface at the point of intersection. The
intersection is considered good if 0 is close or equal to zero and bad if it is close or equal to
7r/2.In this sense, ideal would be a family of surfaces (defined for a range of parameter values 7)

f&XT))=0 (3.3)

which intersects the equilibrium curve everywhere according to the condition 8 = 0 (Fig. 3). Of

Fig. 3. Ideal procedure.

tIn[l4. 151. this connection is not properly formulated. However, it does not invalidate the finding.
 An incremental approach to the solution of snapping and buckling problems 537

course, it is not possible to construct such a set of “ideal” surfaces, because this would require
the advance knowledge of the solution curve defined by (2.3). However, it is not necessary to
insist on 8 = 0 at the points to be computed. Also a small value (0 < ?r/2) would be sufficient to
protect the procedure against failures of the type described. The relaxation of the, condition
8 = 0 to B < 7r/2 opens a whole world of possible formulations of eqns (3.3). The particular
choice made in this payr will be introduced now.
Let [p,, ii] and [b,, r,] denote a point of the path and its corresponding tangent respectively,
and let v1 denote the value of the path or control parameter v at this point. Suppose that these
vectors have been computed at some stage of the computational process and that the next point
is about to follow. The equation

fo(i;P;7))=PI(P-P,)+;;(1-I;)-(11-t7,)=0 (3.4)

defines a surface, i.e. a plane, with [p, i] den@tg the corrdinates of an arbitrary point of the
plane. This plane is normal to the tangent [pi, t,] and its distance to [pi, ii] is (7)- 7,). It will be
intersected by the equilibrium curve with a small angle B if the distance (a - vi) is kept small. As
soon as the point of interseciion is computed for some value of 7, eqn (3.4) can be redefined,
which means that [pi, ii], [ii, tJ in (3.4) are replaced by the values that correspond to the new point
obtained. It becomes clear, that in this way control surfaces are created, which comply very well
with the requirement of a small B along the path.
The control parameter q as defined by (3.4) and the computational procedure based on this
choice, were first introduced in[14]. It is noted that differentiation of (3.4) leads to

(3.5)

where, as before, the dot is used to indicate differentiation with respect to s, the arc length of
the solution curve. It can be deduced from this expression (see also 2.24), that for v1 + 11the
ratio (ds/dq) approaches unity. This means that 11 introduced by (3.4) serves as an ap-
proximation of s for small values of (7 - 7,). Therefore, in the following the symbol 11will be
replaced by s to indicate the geometrical significance of the present choice of the control
parameter. The context in which s appears will usu,ally be sufficient to explain whether this
parameter is used in the exact or in the approximate sense.

3.2 Newton’s method

For what follows it is convenient to introduce a more compact notation. Since the load
parameter p and the deformation parameters th will be treated as unknowns on an equal basis, a
configuration [p, fl can better be written as
_ -
x = xOeo+ xh% = x,&
(3.6)
x,, = p; x,, = r,(h = 1,2,3,. . . . N).

The control parameter s is introduced by adding the equation fo = 0 as a constraint condition to

equation f, = 0. Consequently, one can write

f,,(a:s)=O (a=0,1,2 ,.... N). (3.7)

With this, the convention is introduced that Greek indices cover the range (0.1,. . . . ,N),
while Latin indices remain reserved to cover the range (1.2,. . . , N). It is noted that in
agreement with the foregoing, partial differentiation with respect to x, is denoted by

(3.8)
(a = 0, 1,2,. . . N).
538 E. RIKS

Equation (3.4) in this notation is

imkb) - &(S,)l- 6 - St)= 0

or (3.9)
f(s,)(Z(s) - a@,)} - (s - s,) = 0.

It is expedient to make a distinction between exact or converged solutions of eqns (3.7) and
approximate solutions or iterates. This will be done by the f&owing convention

i=(xo,x,,x2 ,.... xp.J) (exact)

C?= (Co,ur, C72,.. . . ON) (approximate solution or iterate) (3.10)
i = (70,71,72,. . . . T&t) (improved approximation or higher order iterate).

Applied to eqns (3.7) the method of Newton[25] is defined by the system of equations

f~.&?;s)Aue + fa(ti: s) = 0 (3.11)

with
T== CT,+ Au,. (3.12)
It can be defined in inverse form

70 = @cd
- E,& s)f@; s) (3.13)

where F,, correspond to the definition

[Fasl = vq7-‘. (3.14)

A configuration f(s) defined by eqns (3.7) is computed as follows. An approximation 6 is

entered into (3.1I), the system is solved for Ati and an improved solution 7’results, (3.12). The
process is repeated and convergence is obtained if C?approaches x’ sufliciently close. The
literature about the method is extensive so that a detailed description can be omitted (for
instance[25-271). Three gJobaJ points should be mentioned however: (a) The method makes use
of the inverse of I’&,] which means that the Jacobian must satisfy

J = detues} # 0.

(b) If Jf 0, a finite domain ]c?- a(’ = s2(a) around the point f can be defined so that any initial
iterate 6 = Go) which satisfies ]c?- X-1’ < S2(S) leads to a convergent process. (c) The rate of
convergence is quadratic, i.e.
(i-fJ=L(C+-i(2

where L is some positive number which depends on the configuration f(s).

The effectiveness of the procedure to be discussed now is for a great deal based on these
properties.

3.3 Basic procedure

Suppose that of the equilibrium path to be computed, one point is known in advance together
with its path derivative, and denote these vectors by

i, = f(s,) (3.15)
1, = i(s,).

An estimate of a next point of the path is then readily given by

c+“)= Z, + As& = Z, + (s - s,)& (3.16)

 An incremental approach to the solution of snapping and buckling problems 539

This approximate solution C?(I)can be improved by the successive solution of

f~s(~(‘?Au~)+for(~(‘); s1 + As) = 0
(3.17)
#+I) = $0 + A&i’

i= l,2,3,...M

until for i = iU,a sufficiently accurate solution is obtained. (The question how this accuracy can
be controlled effectively is not a part of the discussion at the present time. The way it is usually
done is mentioned in Section 5.1.) A picture of the procedure defined by (3.16) and (3.17) is
given in Fig. 4.
@”

Fig. 4. Basic procedure.

For the computation of the next point of the path (s + As) it is necessary to compute the
tangent i(s). It can be done by means of the set of equations

= ti@‘?b, - S,, = 0
fa,@(x’ (3.18a)

i(s)=;. (3.18b)

Equation (3.18a) follow from eqns (3.7) when the latter are differentiated with respect to
s(b,, = dx=/ds). The normahzing condition (3.18b) is a consequence of the circumstance that the
parameter s in (3.7) can only be appreciated as an approximation to the arc length of the path. It
goes without saying that one can quite well do away with the “exact” value of the tangent
vector i as determined by (3.18) and use an approximation instead. A simple example of such
alternative is
(3.19)

The choice depends somewhat on the way the complete procedure is implemented. However,
the first possibility should always be preferred in cases where the extra volume of operations
connected with (3.18) can be accepted.

3.4 Justificationof the method

The effectiveness of Newton’s method as it was formulated in Section 3.2 depends in the
first place on the existence of a non-singular Jacobian f&l. In the second place, successful
application of the method depends on the “closeness” of the initial iterate c?‘) to the true
solution i.
The first requirement concerns

J = det V..d (3.20)

Expansion of this determinant with respect of the first row gives

J = fo.oD+ fo.iAi
where A, are the co-factors of f0.i in J. It can easily be verified that the co-factors Ai must be
540 E. RIKS

identical to the vector components di introduced in Section 2. Thus,

J = ~o.oD+ f0.d 114, IS]. (3.21)

With the present choice of fO. eqn (3.4). this expression can be written in the form

J(s) = ~~s,)~s) t ij(s,)si(s) = ~=(s,)d=(s) (3.22)

where do(s) is used to denote the determinant D(s) for convenience. The parameter value s
corresponds to the intersection of f. = 0 and fi = 0, while sI corresponds to a previous point of
the solution curve in agreement with the foregoing notation. Although evaluated at the point
n(s), J(s) = det&&)} is also dependent on the point $(si) through the derivative ,$s,).
In the compact notation introduced earlier, eqns (2.27) read

do(sf&(s) = dds)io(s). (3.23)

The square of the “length” of both sides leads to the equality

d:(s)= ~o?sKda(~Ms)l. (3.24)

The scalar product of both sides of (3.23) with &(st) and use of (3.21) resuits in

(3.25)

Equations (3.24) and (3.25) can be combined to give

J’(s)= ~;i,(s,)~~(s)}*I~(s)d,(s)l. (3.26)

It is noted that there are two possibilities for J(s) to become zero. The first occurs if the
angle B between i(s) and &i) is equal to 7r12,because then

cos 8 = .i(s&) = 0. (3.27)

It is clear that this situation can easily be avoided by taking the stepsize (s - s,) small.
The second possibility occurs when

(3.28)

which is the condition that is satisfied at a bifurcation point. Consequently, vols] will be non-
singular along the path if bifurcation points are excluded and if the steps As = s - sI are taken
small enough.
The second requirement for convergence of Newton’s method concerns the distance of the
initial iterate c?(” to the exact solution Z(s). It must be clear that this distance can be made as
small as one wishes because the use of the predictor (3.16) implies

(3.29)

Therefore, the procedure described in the previous section should be effective along any
smooth equilibrium path governed by eqns (2.3), as long as the stepsize As is kept small enough
and as long as no attempt is made to compute a bifurcation point or a point in a close
neighbourhood thereof. A rigorous account of this method has recently been given in[32]: see
also Section 5.4. It is finally noted, that eqn (3.25) can written as

D(s) = b(s)l(sfli(s,)k(s)}-’ (3.30)

 An incrementalapproachto the solutionof snappingand bucklingproblems 541

which illustrates with more clarity the connection between the stability determinant D and the
determinant 1 of the extended system of equations. Relation (3.30) is useful for the detection of
critical points as will be seen later.

4. APPLICATIONS

4.1 Detennination of a critical state

An important part of the stability analysis is the determination of the critical state. In the
case of a snapping problem, the computation of the critical state (the limit point) is usually the
final goal of the analysis. In the case of a buckling problem the determination of the bifurcation
state is the point of departure for an evaluation of the specific properties of this state.
Conditions for the occurrence of critical points were stated in Section 2. The lowest
eigenvalue of [fLi] should satisfy

o,{(i( s)} = 0.

It is also possible however to detect a critical point by the vanishing of the determinant
o(a(s>}, thus
D{x’(s)} = 0.

It is feasible to determine the critical state of the basic path using a strategy which can be
described as follows. Compute points along the path a(%) together with the values w(si) or
D(sJ until the quantities last mentioned undergo a change of signt. Construct polynomials
through Z(s,) and W&I) or D(g) and construct with use of these polynomials the approximation
of x’(P) for which the zero of wI or D is attained. It is known that procedures of this nature
have been used successfully[20,21]. However, an unpleasant aspect of this method is the
uncertainty about the stepsize Ahsto be used. The stepsize As is determinate for the number of
points computed before a zero of wl, or D will be encountered. Too many points are wasteful
from an economical point of view, too few points will lead to inaccurate results. Therefore, at
present a strategy is adopted which guides the computation in the direction of the critical point.
The aim is to reduce the number of intermediate results (the non-critical points) as much as
possible.
Suppose that a sequence of pre-critical points of the basic path have been computed and
denote these points by
i(Sr): a(.Yd:. . . . . . .: . . . . . . . -f(Si) (4.1)
O<Sr<Sz<“‘*“<Si.

Suppose further that the stability coefficients, i.e. the eigenvalue or(s) and the determinant D(s)
can be expanded according to the series

WI(S) = WI(Si)+ OI(Si)(S - Si) + f iil(SiMS - sI) + * * .

D(S) = D(q) + B(Si)(S - Si) + i si(Si)(S - Si)’ + . . . . (4.2)

where Si corresponds to the last point of the sequence (4.1). Within a certain neighbourhood of
the critical state, the slopes &r(s), as) must be negative*. Suppose now that f(s,) is within this
neighbourhood. It then follows that a first estimate of the critical state x’*= x’(P) (relative to
the point x’(s,)) can be given by

C* = .f(~,) + As*i(Si) (4.3)

tNotc, that at this point tk notationw(k) is changedto ok.

*See also the 6rst remark of Section 5.3.
542 E. !ttKS

where AS is determined by either

Asf=(s*--S,)=-~ (4.4a)
I

AS* = (s* - Xi)= -$$.

Postpone the examination into the possibilities to compute (4.4a) or (4.4b) for the time being
and assume that one of these estimates has been obtained.
If the approximation (4.3) is not satisfactory, a new estimate should be constructed relative
to a point which is nearer to the critical state. Such a point can conveniently be computed with
the basic procedure in which the predictor (3.16) is modified to

In this expression n is some measure to serve as a means to keep the stepsize AS = uA.s* within
acceptable bounds. After the new approximation (4.3) has been obtained the process can be
repeated and so on. The complete procedure is convergent if As* approaches zero in the
subsequent steps. Figure 5 shows a picture of such a procedure.

MS~ACEMENT Xi

Fii. 5. Compu~tion of critical point.

It does not seem possible to give a recipe for the choice of the factor 7. If one insists on
approaching the critical state from below, i.e. from the stable part of the path (as was tacitly
assumed in the foregoing), it should be given the value O< r) 5 1 in most cases: with q = I
probably as one of the best choices. Much depends on the accuracy of the estimate As* and the
error of the tirst iterate &o)(8= n(s) + nAs*.&) -Z(si+i)) which should be within a certain
bound as was discussed earlier. (The strategy whereby the critical state is reached from below
is a convenient one from a programming point of view but it is by no means necessary. This
particular aspect of the impkmentation of the procedure will not be discussed in this paper).
The time has come to deal with the question how the estimates (4.C) and (4.4b) can be
computed.
(a) ?7re esfj~ufe (4.4~). The ch~acte~stic vahre oi(s) of Ifu] is deter~~ by eqns (2.13)

fi,,{~(S)bj- W(S)T@j= 0 (2.13)

where Yf,,is some suitably chosen, positive detinite reference matrix not considered dependent
on s. Note that G is written for d(l).
These equations are &id along the path for any s. Cons~uently, it is permissible to
 An incremental approach to the solution of snapping and buckling probkms 543

differentiate (2.13) with respect to s, which gives

fd, - 01 T,#, + f,,,a, - 4 Tip, = 0. (4.6)

By taking the scalar product of a’ with the expression on the 1.h.s. of (2.13) and (4.6) one can
show that
W(s)=&
(4.7)
O,(S)=&
I i
where

(4.8)

Thus (4.4a) can be computed if aI, Cs,i and rfU] are known at the point &. The hrst eigenvalue
and corresponding vector can be obtained from (2.13) by standard procedures. The slope J(si)
at the point a(s) is automatically produced when the procedure of Section 3.3 is used.
However, the computation of vti] needs more consideration. The r.h.s. of (4.8) looks straight-
forward but nevertheless necessitates the evaluation of the third order derivatives f*t which
occupy an array of dimension N*(N + 1). Operation with such a large set of functions is
cumbersome and will soon lead to insurmountable storage problems. Therefore, in this paper
the point of view is taken that the use of derivatives of order three and higher should be
avoided if it is possible to do so. A possible way to circumvent this difficulty is to use an
approximation, e.g.

(4.9)

The disadvantage here is that [fU(si_,)] must be stored and to a lesser extent that at least two
points of the path must be determined before it can be used. Yet another possibility to
approximate cj(s,) is

(4.10)

which is the simplest of all, but which needs two evahrations of o before it can be applied.
(b) The estimate (4.46). Application of (4.C) requires the evaluation of the smallest
eigenvalue of &,,I (in absolute sense), which adds considerably to the volume of the com-
putations, at least when standard routines are used. It might well be that the second estimate
(4.4b) is a more economical alternative. The evaluation of the derivative as) in an exact sense
is impractical for large systems of equations because it requires too many numerical operations.
Therefore, also in this case, hnite difference approximations are a possible way out. The
simplest choice is

(4.11)

Clearly, (4.11) can only be used with advantage if the determinant values D(s,) can be obtained
in a relatively easy way. If the solution of eqns (3.17), a decomposition method is used,
the determinant J = detud} will automatically be provided. In that case the computation of
(4.11) is straightforward because, see eqn (3.30)

WhimIorderderivatives of the potential energy J?

544 E. itIKS

It is noted that eqns (4.12) require three points of the path to be known in advance which
follows from the appearance of the three different tangents in these expressions. If this
property is undesirable, one can drop the terms [&(si_z) f(si-1)J and [f(si-1) f(si)]. This
follows because the zero’s of D(s) are essentially contained in the zero’s of b(s) or J(s)?.

4.2 Continuation beyond the critical points

It is often desirable to extend the computation of the basic equilibrium path beyond the
critical states. It must be clear that this task can be carried out with the procedure discussed in
Section 3. As was shown in Section 3.4, the basic incremental method used here is only
sensitive to a bifurcation point because there the Jacobian matrix LfflJ of eqns (3.17) becomes
singular. Passing a limit point or a turning point with respect to any other variable TVdoes
therefore not pose a problem. Passing a bifurcation point need not be a problem too, as long as
the use of eqns (3.17) close to or at this particular point can be avoided. In fact, when the
stepsize As is chosen arbitrarily there is little chance that this will occur. Consequently,
continuation of the computation beyond limit and bifurcation points (Fig. 6), can be carried out
without special or difficult measures in the implementation of the procedure.

EXAMPlE 2

Fig. 6. Continuationalong basic path.

4.3 Continuution onto a branch

The discussion is slightly more involved in cases where one wishes to compute the branch
of a bifurcation point. It wiIl be assumed here, that the bifurcation point together with its
tangent to the basic path are obtained beforehand, e.g. as a result of the method discussed in
Section 4.1. They will be denoted by xC(s*),&(s*) respectively.
As should have become clear from the foregoing, continuation along any smooth path is
possible as long as at least one regular point of the path is known in advance. In that case, the
tangent to the path can be obtained from eqn (3.18) and the subsequent operations can be
carried out in a straight-forward manner, i.e. in the way it is described in Section 3.3.
However, of the branch of the basic path, only the configuration x’(s*) is known in advance and
this state corresponds to a singular matrix [fps]. The tangent to the branch at this point cannot
be computed from (3.18), but follows from eqns (2.37). In the shortened notation they, read

tChlcsh0Uldexpect loss of accuracyof (4.1I) in that case.

SNote~tlisim~~d=[O,oc,cr~ ,...., aNI.
 An incrementalapproachto the solutionof snappingand bucklingproblems 54s

Continuation onto the branch can now be started with the following initial configuration

#‘) = i(s;) +AS&,@*). (4.14)

Unfortunately the calculation of &s*) from expressions (4.13) involves the evaluation of the
derivatives fti which is, as was explained before, not attractive from a practical point of view.
It was shown in [ 161,that it is possible to compute ~1in (4.13) without the use of these functions.
However, it was also demonstrated that another start of the procedure onto the branch can be
made which abandons the direction vector (4.13) altogether.
First it is noted that the tangent ff, is not a necessary choice for fO.. in the formulation of
the auxiliary surface f0 = 0. The computation of the set of eqns fa = 0, (3.7), can be carried out
with any other normal vector [fO..] as long as intersection of the plane f0 = 0 with the path
remains reasonably well preserved. With this in mind, an effective choice for the normal l’J,J is
given by

(4.15a)

where the factor CL,is determined by the product

/Ll = - {&(s+)a’}-‘. (4.15b)

It follows by inspection, that the scalar product of 4 and & with lj0.J gives

&&*)fo.. = 0
hlwlfo.17~0.
Thus the surface f. = 0 defined by the normal (4.15) is parallel to the tangent &(s*>. This implies
that in the neighbourhood of the critical state intersections of the basic path with the surface
fo = 0 cannot occur. In other words, the computational process is prevented to return to the
basic state. The predictor in the first step onto the path is now given by

CT-(I)= i(s*) + A&&,. (4.16)

Once the point corresponding to (4.16) is determined, the tangent to the branch at this point can
be computed with (3.18) or (3.19) and continuation of the procedure can further be carried out
in the manner described before. The advantage of the predictor (4.16) over the one discussed
earlier, (4.14), is the simplicity of its construction. Only the explicit knowledge of a’ the lirst
eigenvector of [fii(s*)] is required. Possible failures as a result of the predictor (4.16)
(determined by (4.15)) are anticipated when the angle between the basic path and branch is
small at the critical state, thus when

cos e = [i,(P) - i&*)1 - 1.

In that case one is forced to attempt the alternatives mentioned earlier.

5. EXAMPLES AND CONCLUSION

5.1 Shallow circular arch
The stability of a circular arch, uniformly loaded in radial direction, is lost in either a limit
point or a bifurcation point depending on the dimensions of the arch and the boundary
conditions enforced at the endsl27.281. This property offers the opportunity to test the
numerical techniques discussed in the foregoing on one and the same model. In[ 161, a finite
element method was employed for the discretization of the governing equilibrium equations. In
this section, the results of the computations carried out in[16] will briefly be discussed.
Apart from the boundary conditions, the determining factor in the behaviour of the arch is
546 E. &KS

the parameter

(5.1)

where l?/h = radius to thickness ratio and 2u = arc-length of the arch. In the examples that
follow K was set equal to either K = 4 or K = 10.
The first two examples in Figs. 7 and 8 respectively, represent two different calculations of

wo WO

Fig. 7. Example limit point computation(I). Fig. 8. Example limit point computation(II).

the same problem, i.e. the case “clamped-clamped; K = 4; (I = 22.5”“. Procedure I made use of
the predictions As* of the type (4.4a) which was computed from the set of equations

Procedure II was based on the predictor AS* of the type (4.4b), the approximate value of which
is given by
(5.3)

The direct way in which As* is determined by (5.2) follows from the definition of (4.4a) together
with (4.7) if T,, in (2.13) is taken to be

More details on this particular choice of ‘IVcan be found in [ 161.

The iterative improvements of each solution point were terminated after the improvement
AC+”satisfied
jA&“I< rl = l,As*(‘) (5.5)

where Q, is some small, preassigned, positive number. The value AS*“’ corresponded to the first
estimate
cr,i(& + A&(~)}u, = 0. (5.6)

The approximation of the critical state

p = j(q) + As*%(s,) (5.7)

was considered to be adequate if the prediction AS*(‘)in the step (i) satisfied

IAs+( C r2 = ,2As+t’) (5.8)

again with c2 preassigned.
tin this @icdaf cae Y, = &*; 16,= 1,[16].
 An incrementalapproachto the solutionof snappingand bucklingproblems 541

Table I.
-
r iTEP
T PROCE CURE I

rmua
T PROCEDURE II

IERATIOI
P w. l(T’ P
COUNT COUNT
-

1 2.4354 4.6033 2.43534

2 3.0290 8.0578 2.5540 50246

3 3.0753 9.2763 2.8529 6.4525

4 3.0777 9.5919 2.9800 74538

5 3.0780 9.6413 3.0490 8.4085

-

6 3.0722 9.0871

7 3.0722 9.4694

8 30777 9.6074

9 3.0778
-
ITERATI
BIRATION COUNT TOTAL 7
TOTAL

The numerical results of the two runs are presented in Table 1. Measured in terms of the
total number of iterations carried out with eqn (3.17). Procedure I is the most efficient of the
two examined here.
However, the total computer run time used by a particular method also depends on other
factors such as the time required to compute the estimates (5.2) and (5.3). A discussion about
this important question and others connected with the economy of the methods proposed here
is postponed to the next section.
Further examples of calculations carried out with Procedure I are given by Figs. 9-l 1. In
none of these cases difficulties were encountered. Iterative improvement of the initial iterate
c?(‘)seldom exceeded 2 iterations if the termination condition (5.5) was applied with l1= 0.005.
As expected, the largest number of iterations are needed at points which correspond to the
largest step AS.
An example of numerical continuation through a limit point is given in Fig. 12. An example
of continuation onto a branch is presented in Fig. 13.
All these calculations were carried out on a system of equations of dimension 46, 47 or
50 depending on the boundary conditions used (a M-element representation).

K ~4
a=2250
lj .l
P dh
P ‘E-i_

0 01 02 03 04 05 06
Hb

Fii. 9. Example computationbifurcation point (1). Fii 10. Example computationbifurcationpoint (2).
 E. RIKS

Fig. Il. Example computationlimit point.

K ~4
‘X.225’
lJ=l

I-., c I ,
--L L.--.L_ I-.-L_
0 02 04 06 08 10 12 14 16 0 02 04 06 08 10 12 14 16
WI WC

Fig. 12. Example continuationalong basic path. Fig. 13. Example continuationonto branch.

5.2 Questions about eficiency

Applied to the arch problem, the methods discussed in this paper prove to be effective in the
sense that all equilibrium points relevant to the analysis of the arch can be obtained by them.
However, as tools in the framework of a general purpose program for the stability analysis of
structures, one desires next to effectiveness also an acceptable degree of efficiency of the
methods employed.
The method of Newton on which the present procedures are based requires the repeated
solution of the set of equations

f=~(~“‘)Aus”‘+f~(~.“‘; s) = 0 (5.9)
$i+l) = $0 + A$). . j = 1
, 2 (.... M

in order to obtain one particular solution C+(S).The iteration count (M) varies from point to
point in principle, and is strongly dependent on the accuracy of the predictor #! The
operations connected with the solution of (5.9) increase cubically with the dimension (N) of the
problem at hand. It is expedient therefore to minimize the number of iterations with (5.9) as
much as possible.
A strongly recommended way to deal with this problem is to improve the predictor (3.16) for
instance by taking into account the curvature i(s) of the path

,_+‘) = f(s,) + As&) + i(s,). (5.10)

 to thesolutionof snapping
An increment approach andbucklingproblems 549

Another improvement can be induced by a relaxation of the termination condition for the
Newton iteration cycle at points of the path which are not of particular interest (see[ 161Section
6).
A different means to reduce the volume of the compu~tions is based on a m~ification of
Newton’s method.
The idea is to reduce the ~ornpu~tion~ effort connected with the fa~to~tion of I.&,.&. It
is well known that the iterative process defined by (5.9) will also be effective with a suitable
approximation to the matrix [fJ6)]. For instance, in some neigh~urh~ of the point
1, = x’(s,), l&&r)] can be used in

fD,a(a,)Ao,“‘+l=(cs”‘: sI + As) = 0
(5.11)
a(s) = &(‘)f 8, A$’

to compute points Z(s) for s = sl + As.

In this scheme, &(&)I is factorized once. Each iteration involves only the compu~tion of
&] and the cons~ction of the improvement AC?(by back substitution). The effort spent is
considerably less than the effort spent with a true Newton iteration. However, the m~ification
does not converge quad~tic~ly[24] but linearly. Consequently the number of iterations
required is larger than that required by Newton’s method. The linear character of the method
implies.
fg@,_ nl = L’i’l$M _ 21 (5.12)

where the factor L”’ depends on ir and a(s).

It can be shown that the leading term of L”’ is linearly dependent on the distance
fZ(si) -i( - (Q - ~8). F or small (Si - st). Lo’ < 1 and convergence is assured. However, when
(s, - sI) increases, the rate of convergence will decrease until at some distance convergence
breaks down altogether. Consequently, the method requires a redefinition of l&J from time to
time but this will be required anyway if one wants to make use of the stability coefficients wi or
D.
The modification just described is used in the computer code STAGS (29,301 which is a code
to analyse the nonlinear behaviour of shells and plate structures. The authors report consider-
able improvement of the efficiency with respect to the original procedure based on Newton’s
method. In STAGS, the control parameter is either the load factor @) or a multiplier connected
with prescribed displacements. It will be interesting to see how this modification works when
the length of the path is prescribed, because this choice of the control parameter seems to have
a definite edge over other possible choices.
In the procedures for the computation of critical points the predictors based on (4.4a) or
(4.4b) were introduced in order to reduce the number of intermediate results, that is the
non-critical points x”(si).However, the use of these estimates can only be justified if it does not
add considerably to the volume of the computations. When the solution of the linear system
(5.9) or (5.11) is based on a decomposition, method, the value of the determinant J(ci) will be
produced as a consequence of this method. In that Fase, D, and therefore the estimate (4.4b),
can easily be computed as is described in Section 4.1.
The situation is less simple for the estimate (4.4a). The calculation of WIand the corresponding
characteristic vector a’ of the matrix &J] is expensive if standard computer codes are used.*
However, improvement in economy is possible if more specialized methods are introduced. For
instance, it is possible to obtain W(S) and a’(s) from the decomposed &a], using an iterative cycle
similar to that defined by (5.11). An evaluation of this possibility is documented in [33] and will be
published in the near future.

5.3 Concluding remarks

(a) From the present description of the computation of critical points should not be deduced

tBy factorizationis meantthe computation

of [l.&+)l and decomposition,
i.e. the determination
of its inverse of
semi-inverse.
550 E. R~KS

that the predictors (4.4a) or (4.4b) can be used along the whole length of the path in all practical
cases. Successful application of these estimates is based on monotonically decreasing values of
D(Si) or wI(Si) and in general this condition is not fulfilled in the complete interval 0 5 s 5 s*.
Examples of cases where D first increases for 0 6 s 6 sI are given in[28,31]. Consequently,
implementation of the estimates (4.4a) and (4.4b) must be accompanied with appropriate
safeguards.
(b) It sometimes happens that the path contains a stationary point i = 0 or a point where
6 3 0, after which p starts to increase again [3I]. The advantage of the parameter s as compared
to p is clearly demonstrated in such cases.
(c) Just after the completion of this manuscript, the author received a study by Menzel and
Schwetlick(321, which concerns a method quite similar to the one denoted here by basic
procedure. It is shown in a rigorous fashion that the method converges if the solution curve
R(s) exists, the matrix [fi,a(x)] has the rank N and the stepsize As is taken small enough. These
conditions correspond to the conditions stated in Section 3.4.

6. REFERENCES

1. W. T. K&r, On the stability of elastic equilibrium(Translationfrom Dutch). Tech. Rep. AFFDL-TR-70-25, Airforce
Flight Dynamics Lab. (Feb. 1970).
2. W. T. Koiter, Elastic stability and post-bucklingbehaviour.In Proc. Symp. on Nonl. P&l. (Edited by R. E. Langer), p.
257. Univ. of WisconsinPress (1%3).
3. J. M. T. Thompson,A general theory for tbe equilibriumand stability of discrete conservativesystems..Z&schr. fir
Math. Phys. 2#.797-U6 (1%9).
4. J. M. T. Thompsonand G. W. Hunt, A General theory of elastic Stability. Wiley, New York (1973).
5. T. E. Lang and B. J. Hartz, Finite element matrix formulationof postbucklingstability and imperfectionsensitivity.In
PIVC. Symp. Int. Union of I)lcorrtical and Applied Mechanics,L&e. Vd. 61, pp. 727-757 (1970).
6. J. M. T. Thompsonand A. C. Walker, A generaltheory for the branchinganalysisof discretestructuralsystems.Int. 1.
solids SlNclrns 5.281-288 (1%9).
7. R. T. Hafka. R. H. Mallet and W. Nachbar. Adaptation of Koiter’s method to finite element analysisof snap-through
bucklingbehaviour.Inl. 1. Solids SIrvctuns 7, 1427-1447(1971).
8. A. Ecer, Finite element analysisof the post-buddingbehaviourof structures.AIAA I I Ih Aerospace Sciences Meeting,
Washington,DC. AIAA Paper No. 73-255 (IO-12 January 1973).
9. G. A. Thurston. Continuationof Newton’s method throughbifurcation points. 1. Appl. Mech. 9. 427-429 (September
1969).
10. P. M. Anseloneand R. H. Moore, An extensionof the Newton Kantorovic methodfor solvingnonlinearequationswith
an applicationto elasticity. /. Math. Anal. Appl. 13,475-501 (1%6).
II. H. F. Buechner. M. W. JohnsonJr. and R. H. Moore, The calculation of equilibrium states of elastic bodies by
Newton’s method. Ptuc. 91h Mid Western Mech. Con/., Madison (August 1965).
12. A. C. Walker, A method of solutionfor nonlinear simultaneousalgebraic equations.Jnt. 1. Num. Methods Engng 1,
197-180 (1969).
13. A. C. Walker, A nonlinearfinite ekment analysisof shallow circular arches. InL 1. Solids St~clurcs $97-107 (1969).
14. E. Riks, On the numericalsolutionof snappingproblemsin the theory of elastic stability, SUDAAR No. 401, Stanford
University, Stanford. Calif. (1970).
15. E. Riks, The application of Newton’s method to the problem of elastic stability. 1. Appl. Mech. 39. HXO-1066
(December 1972).
16. E. Riks, The incremental solution of some basic problems in elastic stability. Report NLR TR 74005 U. National
AerospaceLaboratory NLR. The Netherlands(1974).
17. 1. Connor and R. Morin, Perturbation techniquesin the analysis of geometrically nonlinear shells. Proc. Symp. Int.
Union of Theonlical and Applied Mechanics, Lii, Vol. 61, pp. 683-705 (1970).
lg. D. Benedetti, C. Bnbbia and L. Ceddin. Geometrical nonlinearanalysis of structuresby finite ekments. Meccanica
7(l) (March 1972).
19. G. A. Dcpuis, D. D. Pfaffinger and P. V. Mar@. Effective use of the incremental stiflness matrices in nonlinear
geometricanalysis. Prvc. Symp. Inl. Union of Theo&Cal and Applied Mechanics, LPge, Vol. 61, pp. 707-725 (1970).
20. S. T. Mau and L. H. Calleober, A finite ekment procedurefor nonlinearorebucklingand initial post bucklinganalysis.
NASA Contractor ReportrNASA CR-1936, (Jan.& 1972).
21. R. H. Gallagher and S. T. Mau. A method of limit wint calculation in finite element structural analvsis. NASA
Contractor &port, NASA CR-2115 (September 1972). .
22. S. Nemat-Nasser and H. D. Shatoff. Numerical analysisof pre- and post-criticalresponseof elastic continuaat finite
strains. Comput. and Structures 3.983-999 (1973).
23. J. A. Stricklin, W. E. Haisler and W. Van R&man, Evaluation of solution procedures for material and/or
geometrically nonlinear structural analysis by the direct stiffnessmethod. AIAAIASME 13fh Strucfures. StNctural
Dynamics and Mareriols Con/., San Antonio, Texas, AIAA Paper No. 72-353 (1972).
24. 1. F. Traub, Iterative Methods for rhe Solution of Equations.Prentice Hall. EnglewoodCliffs, NJ (l%4).
25. J. M. Ortega and W. C. Reinboldt. Itemlive Solution of Nonlinear Eauations in Several Variables. Academic Press.
.
New York-(1970).
26. L. B. Rail, ComputationalS&&n of Nonlinear Operator Equations.Gordon & Breach, London (1970).
27. H. L. Schreyer and E. F. Masur, Buckling of shallow arches. 1. Engng Mech. Lhu., ASCE 92(EM4) 1 (1966).
28. C. L. Dym. Slabiliry Theory and its Application to Structural Mechanics, Chap. 8. Noordhoti. Leyden.
 An incrementalapproachto the solutionof snappingand bucklingproblems 551

29. B 0. Almroth, F. A. Brogan and hf. B. Marlowe, Collapse Analysis for Shells of General Shape. Vol. I, Tech. Rep.
AFFDL TR-71-8 (August 1972).
30. B. 0. Almroth, F. A. Brogan, E. Melkr. F. Zelc and H. T. Petersen, Collapse Analysis for Sbclls of General Shape.
Vol. II, Tech. Rep. AFF’DL TR-71-8 (March 1973).
31. 8. 0. Almroth. E. Melkr and F. A. Brogan, Computer solutionsfor static and dynamic buckling of shells. Proc.
IU’FAM Symp. Buckling of Structum, Cambridge,U.S.AJ974. SpringerVerlag, Berlin (1976).
32. R. Menxel and H. Schwctlick, Zur Behandhmg von Singularitilten bei Einbettungsalgorithmn (Teil I und 2).
Informationen,Tech&& Universit8t. Dresden (1975).
33. E. Riks, A Unified Method for the Computationof Critical EquilibriumStatesof Nonlinear Elastic Systems. NLR Rw~
MP 7704111,National Aerospace Laboratory NLR. The Netherlands(1977).

APPENDIX A
The equivalenceof conditions(2.26a) and (2.31b) folows from the connectionbetween the propertiesof <*me matrices
fjiJ) and [Au]. The characteristicvaluesand vectors{w(i), e(i)) of the symmetricmatrix i&J are definedby

FiJq - a&u, = 0. (Al)

At the critical state

W)

accordingto the definitiongiven in (2.18). The symmetricmatrix [A,J is definedby (2.29)

AU-$. (A3)
I

It satisfiesthe property

A& = f&,, = &,D (A4)

The characteristicvaluesand vectors{A(i): 6(i)) arc definedby

A,&, - A&#, = 0. (M)

They can be expressedin terms of o(i) and d(i)

A(i) = -$j: b(i) = d(i) (A6)

(i = I, 2,3,. . . . . N).

At the critical state the cigenvalucsA(i) become

A(I)= to(k): A(0 = 0 (A7)

a
(i = 2.3,. . . . N).

The characteristicvectors{B(i): i = I, 2,. . ., NJ form a basisof a Ndimensional linear space RN. The setof vectors j
obtainedby the operation

Yi = fi.$/ (A8)

for all i that belongto RN, defines a subspace_T(fu), called the range of &I. In the particular case consideredhere this
spaceis (N - 1) dimensional.A basisof it is: (d(L); ik= 23,. . . N}. If f belongsto the subspaceNV,,) with basisC(I). the
operation (A8) yields j = 0. The subspaceN(f,) is called the null spaceof &,I. It is noted that RN is the union of Tfj,)
and N&j).
The operation
YI = ALA (A9)

for all f in RN definesthe range of [AU]: T(Au). This subspacehas the basis(d(l)) and is equal to UK null spaceof vu].
thus: T(A,J = N&). Moreover, the set of all li leading to f = 0 in (A9) defines UK null space of [AU]. A basis of this
space is {d(h): h = 2.3.4,. . . N) and it is concludedthat N(AU) = T&). Apparently, the geometricalpropertiesof vu]
and [Ay] are complementaryat the critical points.
Consider now the bifurcationcondition(2.26b).

@(l)~&a= 0. (AlO)

This equationconveysthat the vector [fig1Z 8 does not belongto N(fi~) but to TVu) - N(AJ). Consequently,if (AIO) is
satisfied

Aufrp = $0 =0 (All)

must follow. The argumentcan also be reversed.Consequently,(AIO) and (All) are equivalentstatements.