Computer methods

in applied
mechanics and
engineering
EISEVlER Comput. Methods Appl. Mech. Engrg. 156 (1998) 45-74

Structural instability analyses based on generalised

path-following
Anders Eriksson
Structural Mechanics group, Department of Structural Engineering, Royal Institute of Technology, S-100 44 Stockholm. Sweden

Received 18 September 1996

Abstract

This paper describes how quasi-static, conservative instability problems can he completely described, using generalised path-
following procedures for augmented equilibrium formulations. In particular, methods for treatment of compound critical states are
discussed. The numerical methods are seen as extensions to common equilibrium path methods, allowing the solution of subsets 01
equilibrium states, also fulfilling auxiliary relations, e.g. criticality. These formulations are in general used to descrtbe the parameter
dependence in structural response, in instability analyses and in optimisation. The paper describes the general setting of these
generalised equilibrium problems, and discusses some details in their numerical treatment. Emphasis is given to the evaluation of
path tangent vectors, in the presence of critical eigenvectors for the structural tangential stiffness matrix. Also. the isolation of
special states, i.e. vanishing variables, turning points and exchanges of stability, is discussed.
Numerical examples are used to show the possibilities and properties of the obtained solution paths, together with some aspects
of the numerical procedures. @ 1998 Elsevier Science S.A.

1. Introduction

Previous papers by the author have described generalised path-following methods for nonlinear quasi-
static discretised structural problems [l-3]. These include methods for the tracing of equilibrium paths,
i.e. load-displacement relations, as a special case. The methods, however, also allow the calculation of
more general paths, i.e. the direct evaluation of parameter dependence in different aspects of structural
response.
The primary target in using these general paths has been the detailed description of structural insta-
bilities. Although they can in many cases be seen as over-simplifications with respect to real be:haviour,
elastic instability analyses are useful tools for understanding the behaviour of many structures. On the
other hand, formal instability analyses can be considered as overly sophisticated, but it is possible to
show that many instability aspects are missed by common analysis models. Simplified analysis models
also sometimes give incorrect information concerning the stability or instability of critical situations. At
least from a theoretical viewpoint, it is therefore interesting to study the complex instability phenomena
that occur for particular instances of a structure, when the geometry is parameterised.
In order to accurately evaluate the instability behaviour of structures, three basic ingredients are
needed: First, sufficiently accurate numerical elements are needed for the discretised structural model,
as discussed in [4,5]. Second, suitable numerical methods are needed for treatment of the nonlinear struc-
tural relations that arise from these discrete models (cf. [l-3]). Further, the analyses require methods for
handling and describing the critical states. Some methods for this, related to e.g. catastrophe theoretical
concepts [6], are one of the bases for this paper.

00457825/98/$19.00 @? 1998 Elsevier Science S.A. All rights reserved

PII SOO45-7825(97)00200-4
46 A. Eriksson/Comput. Methods Appl. Mech. Engrg. 156 (1998) 45-74

From the application viewpoint, the generalised formulations allow studies of a structural equilibrium
problem from an unconventional direction. Rather than seeing the response as a function of a variable
load, for a specified structure, the response is seen as a function also of the parameters defining the
structure. By adding new parameters to the model, the dependence of the results can be obtained,
without complete re-calculation of alternative structural models. Although the stated problems will be
slightly more complicated than the basic equilibrium problem, their use can in many cases be time-
saving in the total context of analysis. The study of structures in a multi-parametric setting also allows
the treatment of the genericity of the behaviour (cf. [7]).
One example can be demonstrated by Fig. 1, where a clamped toggle frame under a vertical point
load at the apex is studied with conventional load-deflection analyses; the example is discussed below,
with further details in [4]. The three-dimensional diagram shows as dash-dotted lines the calculated
load-deflection graphs for the structure, for six different structural heights.
A few important conclusions can be drawn from Fig. 1. A mechanical conclusion is that the qualitative
behaviour of the structure changes when the height is changed, i.e. the snap-through behaviour disappears
for low frames. It is obvious that a ‘third-order’ limit point (cf. [8]), is obtained for a certain height.
It is also noted that, although the whole paths are normally evaluated, only the solutions to a few
specific loads are really used; these are, e.g. the maximum load, possible bifurcation loads and loads
where certain stress values are reached. It is further noted that, in order to study variations in structural
geometry or imperfections, a whole new path must normally be evaluated.
In order to avoid re-calculation of a new path, Fig. 1 leads naturally to the evaluation of ‘fold lines’,
i.e. curves connecting corresponding singular states on the different load-displacement paths. The curves
connecting the extremum load states for the six studied cases are in the figure shown to be parts of a
continuous curve in the displacement-load-height space. The projections of the fold line for the specific
example are shown on two of the sides of the three-dimensional box. A turning point exists for the
fold line at the height where the snap-through behaviour first appears. It will in the sequel be described
how these fold lines can be immediately obtained, by using a generalised path-following procedure. This
continuous solution path will then be represented by a discrete set of points, each being an equilibrium
state for an instance of the parameterised structure.
Another useful tool in analyses of critical structural behaviour is the method for treatment of all post-
critical paths in compound, i.e. coinciding or nearly coinciding, singularities, developed by Huitfeldt [9].
The properties of the critical states for the ideal model are then investigated, using a ‘branch connecting’
path. This possibility is the only reliable tool for treating compound critical states, as it can accurately
find all paths going out from a critical state. It can also give information on the post-critical stiffness for
ordinary instabilities. It is, therefore, an important tool for theoretical analyses of complicated structures.

Fig. 1. Set of load-displacement paths for variable structure. The solutions refer to the problem in Fig. 5(a), with variable height
H. Dash-dotted lines are calculated load-deflection relations (cf. Fig. 5(b)). Thick solid lines connect the extremum load states
for these. The thin solid line is the calculated ‘fold line’. Projections are shown on the box sides.
 A. Eriksson/Comput. Methods Appl. Mech. Engrg. IS6 (i998) 45-74 37

It can, however, become computationally expensive, as rather many solution points are needed; the
evaluation of each point is in itself, however, only of marginally higher complexity than one for a basic
state. The method needs further development for practical purposes, but is discussed as a concept below.
Another tool for treatment of post-critical branches is also considered below. This can be seen as
a method for simultaneous evaluation of fundamental and post-critical path. Although of a larger size
and complexity, it can well be used for treatment of, e.g. asymmetric bifurcations in frame and shell
structures.
With the mentioned tools, a full investigation of instability-sensitive structures can be performed. Start-
ing from an idealised structural model with chosen configuration and chosen measures, the fundamental
behaviour under load can be evaluated, including instances of critical behaviour. As an addition to this
ideal situation, the parameter dependence, and thereby the sensitivity to different modelling assump-
tions, can be obtained with the evaluation of suitable equilibrium subsets. A new path, with some extra
conditions included, is then started from a chosen state on the fundamental path, in a separate run. The
procedure of restarting the analysis is very simple manually, but is believed to be difficult to make fully
automatic in a general setting.
The objective of the discussed methods is two-fold. One aim is to give practical methods for the
description of structural behaviour in the presence of instabilities. The second objective is of a more
theoretical character and is related to the mechanical understanding of the properties of different classes
of critical behaviour. One particular aspect of this is the information concerning unfolding of the relevant
classes of catastrophes. As will be shown by the numerical examples below, these aspects can be well
understood by the proposed tools, where the focus is set on five classes of generalised one-dimensional
paths:
l fundamental equilibrium paths,
l non-critical subset paths,
l critical subset paths,

l branch connecting paths,

l simultaneous paths.
For the further analysis of critical behaviour, also higher-dimensional solution spaces can be evaluated
(cf. e.g. [lo]). Such methods have not knowingly been used for structural equilibrium problems, but seem
to have the potential of incorporating catastrophe theoretical concepts into the basic solution. According
to Seydel [7, p. 2831 reduction of ‘a parameter space to lower-dimensional subsets will remain a standard
approach for obtaining quantitative insight for a long time’. A ‘full qualitative interpretation of multi-
parameter models requires instruments . . . provided by singularity and catastrophe theory’, which are
not yet fully developed.
This paper is presented as follows: First, the mentioned classes of generalised paths are discussed,
concerning characteristic features and formulations. Then, the specific numerical procedures needed for
generalised path-following are discussed, particularly with respect to aspects differing from an ordinary
equilibrium path algorithm. Emphasis is given to methods that can be efficiently implemented for typical
structural equilibrium problems; the properties that must be considered are the size of the stated problem,
and symmetricity and handedness of the tangential stiffness matrix. Finally, a number of numerical results
are discussed, indicating the mechanical conclusions, which can be drawn from the results. The developed
program code will be discussed in a separate paper.

2. Paths to follow

The introduction described the needs for path-following procedures, as a tool for description of the
sensitivity and parameter dependence of obtained results. This section will give the detailed description
of the formulations used for the different classes of paths discussed.
The general problem setting, as used in the present algorithm, is that of a discretised model of a quasi-
static elastic structural problem. It is assumed that the internal forces in the structure can be uniquely
defined from the current displacement state. No history is therefore allowed in the formulation. This will
48 A. Eriksson/Comput. Methods Appl. Mech. Engrg. 156 (1998) 45-74

normally exclude all problems involving e.g. plastic behaviour or finite space rotations, unless suitably
formulated (cf. [5,11]).
The basic mathematical methods for the discussed augmented systems are well known, with a number
of aspects discussed in [12-H]. They have, however, not been much used in connection with structural
mechanical problems, where normally more problem-adapted methods are used (cf. [16,17]).

2.1. Problem formulation

The formulated problems seek solution sequences to systems, which can be written in the generic
residual form

G(z) =

with F the equilibrium equations, and g a set of auxiliary equations, specifying the required subset
of equilibrium states. The solution to the problem gives a set of combinations z = [dT, ATIT of state
variables d and control variables A fulfilling both sets of equations. In the general form, d is the
N discretised displacement components, whereas A contains p control variables. In order to get one-
dimensional solution sets-curve segments-the relation r = p - 1 is needed, with r the number of
auxiliary equations (cf. [3]).
The common form of equilibrium problem is a special case, which uses only one control parameter,
the load factor, and no auxiliary equations, i.e. p = 1, r = 0. In the more general cases, extra control
variables are used to specify the structural geometry or auxiliary loads. The auxiliary equations can
be used to specify the values for certain response quantities or to demand criticality of the structure;
examples have also been given in [3], where one equation is used to specify e.g. structural weight in an
optimisation context.
The variables used in the formulation can be of different types. These include load factors and load
directions, but can also be related to the structural model, defining structural geometry and sectional
or material properties. The examples in this paper primarily discuss the dependence of results on a
geometrical measure, but they also treat the effects of a perturbing force in calculations of a branch
connecting path.
The problem formulation starts by defining the nonlinear set of equations to be solved. In all the
considered cases, the main part of the set consists of the equilibrium equations; in the present setting,
these are seen as
F(z) -f(d,A) -P(A) (2)
where internal forcesf are dependent on the current displacements d, but possibly also on parameters
in the model. The external forces p are dependent on-at least one of-the control variables in A; it is
assumed that Ai is the principal load factor. The solution to Eq. (2) is normally p-dimensional.
The auxiliary equations are also written in residual form, i.e. as
g(z) z g(d, A) = 0 (3)
with the equations g often of simple form, cf. the following subsections. Obviously, the simultaneous
fulfilment of Eqs. (2) and (3) reduces the dimension of the solution set.
In the problem formulation, it is important to consider the scaling of the used variables and the stated
equations. Compared to a normal equilibrium formulation, it is seen that the auxiliary equations can
be of another scale than the acting forces in the system. The scaling of equations and variables will be
further discussed in Section 3.1.1.

2.2. Fundamental paths

The treatment of fundamental paths is extensively discussed in literature (cf. [16,17] and many others).
In the present setting, the basic equilibrium problem can be written as
F(d,A) zf(d) -p(A) = 0 (4)
 A. Eriksson/Comput. Methods Appl. Mech. Engrg. 156 (1998) 45-74 49

with A = At, using a one-parameter description for the loading and assuming that the internal forces
can be obtained from the displacements only. This set of equations defines the whole problem on a
fundamental path and is used as basic equilibrium relation in the more general problem setting.
The tangential stiffness expression is obtained from differentiation of the equilibrium residual, as
&6d -p,,.tih=O (3
where an index following a comma indicates differentiation in a formally simplified manner and it is
assumed that the load description is proportional, i.e. p(h) z A .p.*; this will be assumed in the sequel.

2.3. Non-critical subset paths

A non-critical subset of equilibrium states can be formulated by demanding a specific value for one
displacement component, i.e. from
g(d) s di - d* = 0 (6)
with d* a constant. Similarly, a control variable can be specified by
g(A) !Z Ai - A* = 0 (7)
A reactive or sectional force component (cf. [2]), is a function of the displacements and the geometry.
Demanding a specific value therefore leads to the general form
g(z) G Ri(d, A) - R’ = 0 (8)
The two first forms are easily obtained from the solved variables themselves, whereas the third can
be evaluated in connection with the evaluation of the internal forces. The whole set of residuals is thus
easily obtained.

2.4. Critical subset paths

When following a fold line of critical states, the problem definition includes an auxiliary equation
of criticality, which must be fulfilled for all solution points. The criticality could be described by e.g.
the determinant of the tangential stiffness matrix; a few other formulations are listed in [18]. Further
possibilities have been discussed in [1,2]. The preferred form, necessary in order to allow for multiple
critical states, seeks a number of the lowest (in magnitude) eigenvalues of the tangential stiffness matrix
K. A suitable equation, obtained from [19-221, can, in its general form, be written as

[:]=[;T:“I-‘[-;.I] (9)

with 1 the unit matrix and @’ an orthonormal matrix approximating a basis for the critical eigenspace.
This system is used to seek a certain number of the lowest eigenvalues to K and the corresponding
eigenvectors @, chosen such that Qi OT@ = -1. Lagrange multipliers are used to enforce the constraint
on the eigenvectors. The matrix ,u introduced means that Eq. (9) solves the linear minimisation problem

1
min (rC) = min(m+,) 5C @TK@j+Cpij(G,j+ @PT@j)
(10)
i j ij
1

The formulated problem in Eq. (9) can be seen as one step of inverse subspace iteration, using a
method that is stable in the presence of (almost) vanishing eigenvalues. Thus, the Lagrangian multipliers
in ,u can be seen as

p = (@q-QG-’ (11)
 50 A. ErikssonlComput. Methods Appl. Mech. Engrg. 156 (1998) 45-74

resembling Rayleigh quotients for the eigenvalues. These values correspond to the improved approxi-
mations Qi = -K-l @‘Op. With this interpretation, the p matrix is the interaction matrix, describing the
relations between the previous and current approximations to the sought eigenvectors. By solving the
small eigenvalue problem pP = Pl, an improved approximation to the eigenvectors is obtained from
an orthogonal rotation
@‘=@P (12)
giving improved eigendirections in @‘. This is followed by an orthonormalisation, yielding @“T@” = 1.
The process is repeated for improved representation of the critical eigenvectors, by replacing a0 by
@“, until the or. matrix is diagonal, and @ ” is identical to Go, measured by the norm I(QjrrT@’ - 1 I(.
The lowest eigenvalues of the K matrix are then extracted from the diagonal of the 5 matrix and are
used as residual functions describing criticality.
In Eq. (9), approximate null vectors are initially specified in the columns of <PO.In the present context,
these are normally easily approximated from the previous analyses. Consideration of the formulated
equation (9) and numerical experience verifies that the choice of initial @” is not very sensitive, as
long as the matrix columns contain components from the critical eigenspace. The problem of auxiliary
equations for compound criticalities is discussed in [23].
The described form is only used for critical states existing on a whole path of critical states and not
in the isolation of one particular state on a path; the latter problem is considered in Section 3.4. As
the sought critical eigenvalues are normally several magnitudes lower than the non-critical ones, when
iterating close to a fold line, convergence in the subspace iterations is assured to be rapid; some problems
have occasionally been seen in examples where two fold lines cross.
When following a fold line of simple critical states, the a0 matrix is a vector, approximating the
critical eigenvector at the sought state. The iteration formulation is here used as

(13)

for i = 0,1, . . . . Compared to [l], the iterative improvement of the eigenvector has been shown by
numerical experiments to be necessary in order to follow a fold line of bifurcation states. It has, however,
been concluded that no improvement is normally obtained by saving the approximations between path
steps, which is in agreement with [21]. By doing this, it has, however, in [ll] been possible to find the
correct lowest buckling loads for a rectangular plate, for variable length-to-width ratios.
The expression in Eq. (13) essentially performs inverse power iterations in the 4 vectors, not af-
fected by the near singularity of the K matrix. The g value is similar to the Rayleigh quotient for the
approximations

g(z) = g(K(z)) = 4TK C/I with (diT+) = 1 (14)

which converges towards the lowest absolute eigenvalue of K and thus can be used as a residual equation.
A specific problem occurs when Eq. (9) is used for cases with higher than one-dimensional critical
eigenspace, e.g. the fold line for hilltop branching states in Section 4.1.6. When evaluating the basis
for the critical space, as in Eq. (9), the resulting columns in @ are arbitrary within the space, if the
eigenvalues are equal or close to equal. The normally rather natural directions introduced by the 4p”
matrix are thereby lost in the evaluation, which complicates the specification of the residual. In order to
handle these problems, it has been found useful to orthogonally transform the final @ matrix, in such
a way that @ OT@ is diagonal, with Q” the initially specified matrix. The eigenvalues corresponding to
these directions are then updated, according to p = GTK@.
As all auxiliary equations for criticality use the tangential stiffness matrix-and this thereby has to be
evaluated in connection with the residual evaluation-efficiency aspects are important for relevant size
problems. The tangential stiffness matrix should thus be stored in these cases, in order to be used in
several phases of an iteration cycle.
 A. Eriksson/Comput. Methods Appl. Mech. Engrg. 156 (1998) 45-74 51

2.5. Branch connecting paths

Special methods are needed in an ordinary path-following continuation algorithm. to switch branches
at a bifurcation state. Different methods for these tasks have been developed and are well functioning.
Essentially, these either treat a slightly perturbed version of the ideal problem [24], or start the new path
by special prediction methods [25,26]. Both strategies rely on a knowledge about the type of critical state
at hand and are only reliable for simple bifurcation states.
Another method, which does not need this detailed knowledge, is the method based on a ‘branch
connecting’ path. The method is developed by Huitfeldt [9] and is further commented on by Allgower
and Georg [15]. In the present context, the method is based on an equilibrium equation of the form
P(d,A)~f(d)-hp.,-7.x=0 withhEn,, r--A2 (15)
if a basic formulation with one control parameter is considered. The vector x describes a disturbing load
case, which is essentially arbitrary, although the method functions better with a clever choice of x; it must
also, for easy interpretation of the results, be orthogonal to P,~. It is obvious that the subset of solutions
to Eq. (15) with r = 0 are solutions to the fundamental one-parameter equilibrium problem in Eq. (4).
This motivates a search for such states in the vicinity of a critical state. In order to find a relevant set of
these, Huitfeldt suggests that a path of solutions to Eq. (15) is sought on a hyper-sphere, centred at the
approximation to the critical state. By introduction of the sphere equation, this curve is one-dimensional.
It can be shown that the sign of the auxiliary load factor r changes every time this curve intersects a
solution path for the basic problem, i.e. when crossing the fundamental and all bifurcation paths.
The equation of the sphere is in [9] given as
g(d, A, r) - (d - fz*)T(d - d*) + (A - A*)* + r2 - p2 = 0 (16)
where z* - [d*T , A*lT is the estimated critical state and p is a chosen radius. In the present investigation,
a simplified version
gd(d, A, T) = (d - d*)T(d - d*) - p* = 0 (17)
has been found slightly more suitable, as it reduces the scaling problem. It should be noted that this is
not an arc length constraint but a problem-defining equation; obviously, this auxiliary equation together
with the two-parameter formulation in Eq. (15), leads to a one-dimensional solution path. When solving
new points on this branch connecting path, a seemingly similar equation must be added as a constraint
in the iterations, as discussed below in Section 3.
The path following can best be started from a passed equilibrium state on the fundamental path, close
to a more or less accurately decided critical state. Referring to Fig. 3 (see Section 3.4.4 below), one
could in an automatic procedure use the solution points ‘0’ and ‘1’ to define their midpoint as Z* and
use the distance between them to decide p. Linear inter- or extrapolation gives a good starting point for
the path on or close to the fundamental path with r = 0. The disturbing load vector x can preferably be
taken as a point load vector exciting the critical eigenvector. The vector is given a scaling, such that r
will be of the same order of magnitude as A - A*.
The choice of radius p is important. A too low value can lead to convergence problems from the
nearby singularity. Too large values will lead to too many points needed or that closed paths are missed.
In performed tests, the radii have been manually chosen in the order of p M 0.1. /Id* (1.
By following the path until it returns to the starting point, an even number of distinct states with T = 0
are found (cf. Fig 10). These are the desired crossings with the basic solution paths from which normal
path following can be initiated for the basic problem, without problems related to the criticality of the
structure. Based on symmetry, the branch connecting path needs often only to be followed until it again
crosses the fundamental path of the basic problem-a condition which is often easily identified. This is,
however, not necessarily the case for all problems and all radii, as one-sided closed solution paths can
exist for some classes of instabilities [6].
One particular aspect of the branch connecting path evaluations is that the algorithm must find all
solutions to the augmented set of equations for which T vanishes; this is further discussed below. It
is noted that this isolation procedure need not be very accurate, and that the vanishing value is well
52 A. Eriksson/Comput. Methods Appl. Mech. Engrg. 156 (1998) 45-74

defined. Experience shows that linear interpolation between solved points on the curve can be used with
good accuracy.
Numerical experiments by Huitfeldt [9], and in the present investigation, show that the method is very
reliable and finds all paths in the vicinity of a critical state. The only cases where the algorithm has
missed any basic solutions have been when the vector x was chosen in exactly the same directions as
one of the paths leaving the critical state. This means that the eigenvector itself is not normally a good
x vector; a better choice would be a unit vector for the dominant component of the mode.
Obviously, the evaluation of a whole branch connecting path is computationally expensive, although it
is easily formulated in a general algorithm. The same basic idea is believed to be possible to introduce into
an automatic procedure of limited computational requirements. Some success has been seen in methods
starting from an arbitrary perturbation in the eigenspace initiated from the critical state. From this state,
the vector x can be chosen from the internal force increment related to the critical state and iterations
performed towards 7 = 0. By using symmetry aspects and keeping a record of all obtained solutions,
a good knowledge of the region around the critical state can be obtained with limited computational
effort. The practical implementation of the ideas, however, need further studies.

2.6. Simultaneous paths

Another method for the analysis of branching points and in particular for the evaluation of post-
bifurcation paths, can be based on common methods for the analyses of symmetry-breaking situations.
The method, following a basic description from Seydel [7], formulates a problem where two solutions
to the same loads are sought simultaneously. The two solutions are assured to be situated on different
branches by the introduction of a certain distance between them. The present case preferably uses the
length of the projection of the difference on an approximation to the critical eigenvector.
The stated problem can, in a one control variable basic situation, be seen as expressed in the variables

(18)
with 2n state variables and two control variables. The formulation demands simultaneous equilibrium at
the two displacement states dl and d2 for the same load factor, i.e.

F(dl, A) =f(d,) - AP,A = 0

09)
i F(dz, A) -f(d2) - AP,A= 0
with the same evaluation model for the internal forces as in the fundamental problem and assuming
linear proportional loading. By also demanding a certain distance between the two solution states, the
full set of residual equations can be written

F(dl, A)
G(z) - F(dz> A) =o (20)

i &T (d2 - dl) - 6 !

which is (2n + 1) equations in (2n + 2) unknowns, and thus fits into the general setting. The formulated
system is of double size, compared to the basic problem, but the possibilities for parallel treatment
are very good. The doubled size of the differential matrix is then of limited importance in a complete
calculation; this is in particular related to the simple structure of the differential matrix (cf. Section 3.2).
Typical usage of this formulation is when a critical (e.g. bifurcation) state has been found for a
model. By introducing one of the bracketing solutions on the fundamental path as (dl, A) and using an
approximation to the critical eigenvector as 4 to produce d2 = dl + 6. 4 with a suitable 6, a path can
be started, e.g. by incrementing 8.
Very good numerical properties have been obtained with this formulation for treatment of common
symmetric bifurcation states, cf. the numerical example in Section 4.2.1. Also, in cases with asymmetric
bifurcations, the formulation is useful. Although not tested with a reasonable shell element type, this is
believed to be a method that can give information for some notoriously complicated shell problems. A
 A. Eriksson/Cotnput. Methods Appl. Mech. Engrg. 156 (1998) 45-74 53

similar idea has also been tested for cases concerned with the stability of post-critical paths in different
parametric regions (cf. Section 4.2.2).

2.7. Lagrange multiplier relations

The previous section has shown how specific types of paths can be defined by certain auxiliary equa-
tions. Other types of equations can, however, also be used to constrain the solution with other purposes,
possibly in combination with the above cases.
Fold lines for, in particular, symmetric bifurcation states are numerically sensitive. The situation can
be schematically described by Fig. 2, which shows fundamental and post-bifurcation paths for a fictitious
structure, where L governs the geometry. The correct fold line obviously follows the extremum points of
the post-critical paths but is very sensitive. As the eigenvector is a tangent direction of the equilibrium
relation, states along this direction almost fulfil the stated equations. As the eigenvalue of the tangential
matrix has a double zero point at the critical state (when following the eigenvector), solutions with low
residual norms are easily found off the fundamental solution path, especially where the post-bifurcation
paths are very ‘flat’ (cf. Section 4.1.4).
The knowledge about the sought solution points can be utilised in order to reduce these problems, for
a specific model. For instance, all solutions on a fundamental path for the structure in Fig. 5, considering
a vertical force only, must give the rotation dz = 0. This must, therefore, also be valid for the fold line
of bifurcation states, when a variable height is introduced. Similar conclusions can normally be drawn
for other models (cf. the shell buckling problem in [ll]).
In order to assure a zero solution component or other displacement relations, Lagrange multipliers
can be used. These methods introduce an extra control variable p = Ai, corresponding to a load case a,
which reflects a prescribed displacement relation
tTd = 0 (21)
The obtained system will thus be corrected in the equilibrium equations to give the symbolic relation
F(d,h,p)~F(d,A)+pt=O (22)
An auxiliary equation, according to Eq. (21) must also be added; it is noted that these additions can
be made to very general subset formulations, i.e. bifurcational fold lines. The formulation complexity is
thereby only marginally increased.
The Lagrange constraint modifies the original problem as it introduces extra force components. As
the constraint is formulated to ‘pull’ the solution towards the fundamental solution, this force will ideally

Fig. 2. Schematic situation when following a fold line of symmetric bifurcation states in the parametric space. The obtained curve
(thin) easily deviates from the correct one (thick). Dash-dotted lines indicate basic equilibrium paths for fixed structures.
54 A. ErikssodComput. Methods Appl. Mech. Engrg. 156 (1998) 45-74

converge towards zero. Numerical experiments verify that this will be the result, even for numerically
very unstable problems (cf. Section 4.1.4 and 1111). Some care must, however, also be taken when
interpreting results concerning critical loads and modes for problems where Lagrange multipliers have
been introduced in order to prescribe certain displacement relations.

3. Path-following algorithm

With the formulations used in the previous discussion, all types of paths can be followed with the
same algorithm. It is noted that an algorithm capable of handling all these types of paths must have
some special features, not needed in the treatment of fundamental equilibrium paths. The more complex
setting also motivates use of more general and complex basic numerical tools. Some important aspects
of the algorithm are discussed below.
The stated problems can be solved by path-following techniques based on the implicit function theo-
rem. Here, the differential matrix

(23)

plays an important role in many parts. The {(N + r)/(N + p)} ’ matrix, which contains the structural
tangential stiffness matrix as its top left part, is used in the iterative procedure when bringing the
residual function G(z) to zero, but also as an important description of structural properties.
The iteration procedure for a new solution point in the solution set of Eq. (1) can be formulated in es-
sentially the same manner as the equilibrium path algorithms for the restricted problem with r = 0, p = 1.
Introducing a path length measure as a parameter in the solution set, the problem to be solved in each
step can be written as

(24)

where N(z) is a path-measuring function, i.e. a step-wise defined arc-length expression. No general
answer to this parameterising problem has been found and it is believed that the optimal method is
highly problem dependent. In the present implementation, the dominant component from the tangent
vector at the last solution point is normally used as parameter.
Based on the form in Eq. (24) and the corresponding differential matrix Gz, different versions of
Newton-Raphson (‘N-R’) iterations can be used. For efficiency, the special structure of the matrix must
be recognised. It must further be noted that, for many interesting cases of critical subset paths, the
tangential stiffness matrix is (numerically) singular at all converged solutions.
For these problem types, the usual terminology for critical states is no longer relevant. In order to
reflect the new types of interesting points on a subset path, the term ‘special’ solution state on the path
is defined in [2]. These are here defined as points, where the structural properties of the obtained states
change in some way or where the path itself changes qualitatively. The term includes turning points, with
respect to all defined control parameters, regardless of whether they are related to a critical state for an
instance of the parameterised structure. It also includes the exchanges of stability, where the numbers
of negative or zero eigenvalues of the structural tangential stiffness matrix change. This occurs for a
parametric combination where the order of instabilities changes or a certain phenomenon disappears.
Further, it has been described above how the vanishing of a control variable might be highly important.
The needed additions to the conventional path-following algorithm demand more elaborate procedures
in the algorithm and will be discussed below. The solution paths for the augmented systems can be

‘The notation {m/n} is used to describe the dimensions of an m x n matrix.

 A. Eriksson/Comput. Methods Appl. Mech. Engrg. 156 (1998) 45-74 55

followed with any continuation method that allows unsymmetric formulations. For efficiency in practical
problems, the algorithm can, however, be considerably improved by utilising the special properties of
these problems.

3.1. Basic design

In the present implementation, it is possible to define problems with p 3 1 control parameters and
with r 3 0 auxiliary equations. It is thus required that r = p - 1 parameters are varied in the analysis,
but further parameters can be included; these are then kept fixed at their initial values throughout the
calculation. This is, for instance, relevant when a basic case is first studied and this is to be followed by
a set of variations.
If existing, the auxiliary equations are, however, of types that are not easily introduced into a general
program but must be found from a problem dependent description. In the present program, this is
introduced by two or three code segments. Necessary are one routine for initial setup of the problem
data and one routine for evaluation of the residual G(t). Several of the mentioned auxiliary equations,
e.g. those including stresses or reactive force values are easily obtained in the procedure evaluating
the residual forces. If available, the program can also utilise a routine for evaluation of the differential
matrix G,z(z); otherwise, the differential is automatically obtained from a numerical differentiation of
G(z).

3.1.1. Scaling of variables and equations

The use of a suitable scaling for the included variables in a calculation is related to the numerical
behaviour and accuracy of the basic algorithms, for instance, in the solution of sets of simultaneous
equations (cf. e.g. [14]). The scaling of the involved variables is also of great importance for the be-
haviour of any stepwise algorithm. It is noted in [24,27], that any step size procedure for general use
must be based on an automatic scaling of considered quantities. The need for proper scaling is even
more pronounced when several control parameters, possibly of very different types, are included in the
problem formulation. The scaling problem is in the present formulation reduced by manual scaling of
the parameters involved in a specific problem, aiming at ) Ai INN/~ )I.
Also the equations need proper scaling, as this relates directly to the accuracy of the obtained solutions.
It is important, when following a critical subset path, that the residual corresponding to criticality is at
least slightly ‘over-emphasised’. After individual scaling of the auxiliary residual components, the same
tolerances can be used for the auxiliary equations as for the residual forces.

3.1.2. Iterative procedure

The numerical iteration procedure aims at obtaining a set of solutions to Eq. (24), with varying values
for the involved variables. It consists of the following parts, of which none is much affected by the
generalised problem setting.
Starting from one solution point to Eq. (l), an Euler prediction is used and a parameterisation based on
one important component of the tangent vector [28] (cf. Section 3.3). A step length adjustment procedure
can be used, based on the number of iterations needed in the previous step [27]. The solution method is
based on variants of the N-R scheme, where the parameterising component is kept at a constant value
and the differential matrix is re-formed intermittently. It is seen that the major parts of the residual
vector and the differential matrix are obtained from the ordinary force and stiffness formulations.
Numerical experiments show that a solution method without parameterisation, described by Allgower
and Georg [15] can be more suitable for some problems.

3.2. Differential matrix

As noted above, usage of a correct differential matrix is important. The differential matrix is a gen-
eralisation of the tangential stiffness matrix and contains differentials of all equations, with respect to
all variables. It mainly consists of the structural tangential matrix, to which has been added r extra rows
and p extra columns. The accuracy of the matrix is important from the efficiency viewpoint, as it is used
56 A. Eriksson/Comput. Methods Appl. Mech. Engrg. 156 (1998) 45-74

in the iterations for a new solution point. It is, however, even more important as the matrix is used for
the quantitative description of the properties of a solution point.
Analytically derived matrices are therefore preferred. For many of the interesting auxiliary equations,
the differentials are trivial or easily obtained. The most complicated quantity in this respect is the
determinant of the tangential stiffness matrix. The preferred expression for criticality in Eq. (9) can be
rather easily numerically handled. For one critical vector, the differential with respect to the displacement
components can be approximately evaluated as (cf. Eq. (14))

c&d) = bPTK+),d = 4TK,d+ = 4TV+K = "cp+f

=+(f(d+Y4)-2f(d)+f(d - r4))T (25)

with y suitably chosen and 4 the converged approximation to the critical eigenvector. The notation V is
here introduced to denote the directional derivative, in this case of first and second order, with respect
to 4. It is emphasised that only internal forces are needed, which makes the approximation cheaper
than when utilising the tangential stiffness K, as in [l].
The added columns to the differential matrix will normally need numerical differentials if they do not
relate to auxiliary load cases. In particular, all parameters affecting the overall structural geometry will
need such treatment. The auxiliary variables in particular cause problems in complex equations, such as
e.g. the criticality defining equation in Eq. (13). Practical experience shows that the differential of this
quantity with respect to geometrical parameters cannot normally be neglected in the differential matrix.
By using Eq. (14), and assuming that the eigenvector approximation resulting from Eq. (13) does not
change much when altering the geometry, one can evaluate

g,Aj x +TK,/t,+ w & (4’K(z+8Aj)Qj - +TK(z)4) (26)

I
with obvious notation, in a case of one-dimensional critical eigenspace. In Eq. (26), the first quadratic
form can further be approximated as

+TK(~ + aA,)+ = +TV,f (Z + 6Aj)

M + .4T (f(Z+sAj
+r+)-f(~+sAj)) (27)

with a chosen y. It is noted that the second of these internal force vectors is also needed in order to
produce the residual force differential F ,“,. This means that the differential matrix, for a critical case
with one geometric variable, can be evaluated from one tangential stiffness matrix and five internal force
vectors. It is often, however, optimal from the accuracy viewpoint to use a central difference expression
for the numerical differential with respect to the 4 vector in Eq. (27); this leads to a need for one more
evaluation of internal forces. Obviously, several of these can be done in parallel.
With several considered columns, it is assumed that the eigenvectors keep constant and orthonormal
for small changes of the parameter values. This means that the same procedures as in the above case
can be used, but repeated for each obtained 4i vector. This means that Eq. (25) needs two additional
evaluations of the internal forces for each considered mode and Eq. (27) one. For the case in Section 4.1.6,
with two critical directions and two geometrical parameters, that means that the total differential is
evaluated from one tangential stiffness matrix and eleven internal force vectors.
The differential of the set of equations used for the branch connecting paths in Section 2.5 is easily
produced. The two load cases considered give the two columns to be added to the right of the tangential
stiffness matrix. The auxiliary row is given by the equation defining the sphere, i.e. for the formulation
in Eq. (17), as
gd,d = 2(d - d*)T gd,A = o (28)
If the slightly more complex form in Eq. (16) is used, the differential includes
g,A = 2(A - h*) and g,, = 27 (29)
 A. ErikssonICotnput. Methods Appl. Mech. Engrg. 1.56 (1998) 45-74 57

The differential matrix related to the simultaneous path formulation in Section 2.6 is very simple.
Differentiation of Eq. (20) with respect to the used variables in Eq. (18) gives

1
cf@l) 0 -P,n 0

G.Z E 0 F,d(dz) -PA 0 (30)

i -Cpr Cpr 0 -1

where the tangential stiffness matrix F,d(d) in this case is evaluated for two different displacement
states. It is obvious that this matrix can be rather efficiently evaluated. Similar expressions can be easily
formulated for other applications of the simultaneous path idea.

3.3. Tangent evaluation

In all applications of path following for solution manifolds, the concept of tangent space is important.
The space consists of the null space of G,Z, i.e. of all vectors t fulfilling

G,z . t E [G,d G,A] (31)

indicating the local directions of the solution manifold. For a non-critical state in a fundamental path
evaluation (cf. Eq. (4)), the space is a one-dimensional vector of the form
tf = [6dT, &A] T (32)
where Sd is the current linearised displacement response to a load increment Sh. It is obvious that the
tangent vector tf can be arbitrarily scaled.
At a simple bifurcation state in the equilibrium path, the tangent space is of higher dimension. From
the singularity of the tangential stiffness matrix F,d, i.e. F,d 4 = 0 for one eigenvector 4, and the
condition for a bifurcation state c,bTF,h = 0, it can be seen that the tangent space is spanned by the
columns of the two-dimensional matrix

T=
! SA 0 1
with i3d orthogonal to 4, c$T6d = 0. It is seen that any linear combination of the columns fulfills
(33)

Eq. (31); the vectors span a two-dimensional tangent space. Locally, all vectors in this space fulfil the
differential equilibrium relation. The decision on the vectors in this space that are real path tangents
needs consideration of higher-order differentials of the equilibrium relation and leads to solution of the
Algebraic Bifurcation Equation (‘A.B.E.‘) e.g. in the general form from [29,30]. It is particularly noted
that neither of the two columns is a path tangent for an asymmetric bifurcation state.
At a limit state, with an extremum value for the control parameter, c#J~F,~# 0, and only one vector
remains. This vector obviously contains the eigenvector C#Jas the displacement part, corresponding to a
zero load factor increment.
As the singularity of the tangential stiffness matrix exists only at distinct critical states in normal
equilibrium path-following, handling of the higher dimensional tangent spaces at these states is of limited
difficulty. Essentially, problems are avoided by evaluating the tangent direction of the path close to the
singular state and utilising the known critical eigenvector.
In the more general context, the evaluation of the tangent space becomes more complicated. This
is particularly the case when paths of critical equilibrium states are traced. Limiting the extra control
variables to geometrical parameters in the model, a few interesting cases are shown in Table 1. For each
type of solution curve, the situation at a general point is shown, together with the situations occurring
at ‘special states’ along the curve, i.e. points where the qualitative behaviour of the structure changes.
The situations marked with ‘-’ are common on the curve and must therefore be possible to handle. The
two cases marked ‘- -’ are examples of more exotic situations, only occurring for special instances of
the parameterised structure. In the table, t, is used for a non-critical null vector, whereas c,z~~
denotes a
58 A. Eriksson/Comput. Methods Appl. Mech. Engrg. 156 (1998) 45-74

Table 1
Schematic bases for tangent spaces at common situations in generalised path-following. Further explanation is given in the text
Case (Path, point) Tangent space Comments
Fundamental path:
General point
- limit point
- bifurcation point
- - hilltop point
- - double bifurcation point

Non-critical subset path

General point
- turning point turning w.r.t. added parameter
_ crossing bifurcation path bP? 41
Limit state subset path
General point
- turning point turning w.r.t. added parameter
_ crossing bifurcation path

Bifurcation subset path

General point
- turning point turning w.r.t. added parameter
_ crossing limit state path
_ crossing bifurcation path

critical displacement eigenvector for zero control variable increments. Most of the discussed cases are
shown by numerical examples below.
It is obvious from the table that two- or higher-dimensional tangent spaces can exist, even when the
interesting solution curve has a distinct one-dimensional tangent. The higher-dimensional space is then
related to a critical eigenvector existing for the tangential stiffness matrix. Obviously, this existence of
additional null vectors can cause problems in the path-following procedure; in a pictorial language, the
problem can be seen as trying to follow the line of lowest points in a shallow valley (cf. Fig. 2). It is
again noted in the table that the path tangents are a few specific combinations within this space.
The following subsection discusses how the tangent space can be evaluated in a general case. The
primary objective of the development is to find an orthonormal matrix T containing, as its columns, a
set of vectors that span the whole tangent space. As indicated by Table 1, it is important that the basis
vectors in this space can be chosen to reflect inherent properties of the problem at hand. Essentially,
methods for this can be based on projections on known critical eigenvectors or on orthogonal rotations
in the null space.

3.3.1. Singular value decomposition

The most reliable method for evaluation of the null space of an {m/n} matrix A is the singular value
decomposition A = UX VT, with U and V orthonormal matrices and Z an {m/n} matrix, with the
singular values q 2 0 along the main diagonal. A zero or low (T value indicates a column in V that
belongs to the numerical right null space of the matrix A. As the tangent matrix is always evaluated at
a state that is only numerically critical, a tolerance must be chosen for zero singular value. The main
advantage of the singular value decomposition is that these values are clearly shown. Procedures are
available for handling these special situations, but are normally considered too expensive to use in a
path-following algorithm.

3.3.2. QR factorisation
An alternative to the singular value decomposition of a matrix, is the QR decomposition. During the
last years, much interest has been focussed on so-called ‘rank revealing’ QR decompositions 131-341. In
 A. Eriksson/Comput. Methods Appl. Mech. Engrg. 156 (1998) 45-74 59

general, these factorise an {m/n} matrix A(m > n) into a product of an orthogonal matrix Q , an upper
triangular matrix R and a permutation matrix PT, in such a way that the numerical rank of the matrix
is obvious. The numerical rank of a matrix is in this context defined as the number of singular values
greater than a certain tolerance. It is assumed that a clear gap exists in the singular values around this
tolerance, so that a number of singular values are lower, while the others are clearly higher than this
tolerance. Under these assumptions, the rank revealing QR factorisation gives good estimates to the low
singular values and the numerical null vectors.
An interesting algorithm is described in [35]. The paper describes how the sparsity aspects can be
taken into account in a rank revealing QR decomposition, and how fill-in in the studied matrix can be
reduced, although not completely avoided. The algorithm uses restricted column pivoting, together with
an ‘incremental condition estimation’ [36], in order to push possibly linearly dependent columns to the
rightmost part of the matrix. These pushed columns are then treated in order to obtain a matrix Rzz of
low norm, in the lower right part of R, according to

where R,, is upper triangular and has no low diagonal values. Based on this factorisation, the singular
vectors of A can be evaluated. The needed singular vectors are easily obtained in the factorisation
process.
A variation of this factorisation has been tested in the developed algorithm. The {(n + r)/(n +p)}
matrix G,Z is thereby brought to upper triangular form by Householder transformations [37]. A restricted
column pivoting is used in each cycle, with the objective to bring an optimally independent column into
the pivoting position. The pivoting column is chosen only among the s first unused columns (with
5 2 s 2 10 for realistic problems), but with the first remaining column preferred. The chosen column
is the one that is farthest away from the subspace spanned by the already computed columns, i.e. has
the highest norm of its components remaining below the already finished level in its unit length form.
The result from the first column is in the comparison multiplied by a factor higher than unity-typically
2. Before performing the Householder transformation, the lowest singular value of the potentially new
matrix is estimated by the incremental method from [36], where a new singular vector for the matrix of
increased size is guessed from available information. If this estimate gives an unacceptably low value,
the whole pivoting window is cyclically permuted to the right part of the matrix and a new pivoting
column candidate is chosen.
When all the columns have been tried as pivoting columns, according to the two tests above, the
remaining columns are tested. No fill-in aspects are considered at this stage and standard column pivoting
is therefore used, accepting all columns as candidates, but again preferring the first available one. When
a low estimated singular value is obtained in this phase, the process is discontinued, with p columns
processed, thus indicating the numerical rank of the matrix. The result is then a factorisation
RII RI?
G., = Q PT (35)
[ 0 =O 1

where P describes the performed column pivoting, Q is the product of the Householder transformations
and RI1 is a non-singular square matrix of size corresponding to the rank p of A. RI2 is a rectangular
matrix of columns, which are numerically Iinearly dependent on the columns of RI ,, It is noted that Q
is never explicitly evaluated and that P can be represented by an index table.
The null vectors of G,z can now be obtained from the null vectors of the R matrix. If all columns in

tR R;,%
=-1
RI2 are linearly independent, a basis for this null space can be obtained from

.L+pp
1
[
The produced matrix can be orthonormalised by a Gram-Schmidt process [37]. The null space basis for
(36)

the matrix C,z is then obtained by a permutation of the rows described by the matrix P.
60 A. Eriksson/Cornput. Methods Appl. Mech. Engrg. 156 (1998) 45-74

The restricted pivoting leads to a limited fill-in in the matrix if a suitable storage is used. This means
that the procedure can be reasonably efficient also for banded matrices. For the augmented equilibrium
problems, special measures must be taken for the rows and columns added to the tangential stiffness
matrix. A suitable storage scheme for large-scale problems is suggested in Section 3.5.

3.3.3. Eigenspace separated tangents

The above described methods can be seen as general procedures for evaluation of the null space of the
differential matrix and has been tested in the algorithm. As they are, however, rather computationally
expensive, the method described in [3] has been further developed to handle the generalised situations.
The main idea in this procedure is to separate the behaviour of the path into parts occurring in the critical
eigenspace of the tangential stiffness matrix and parts transversal to it. Although less time-consuming
than the QR-based method, this procedure could not-in the form given in [3]-reliably handle more
complicated situations. It has, therefore, been further developed.
The method starts from the null space of only the equilibrium equations and then studies how this
null space is reduced by the auxiliary equations. That means that the procedure first seeks a matrix T F
with columns giving the basis for solutions to the restricted problem

IF& F/I] CF = 0 (37)

If the tangential stiffness matrix F,d is non-singular, this space obviously contains the p columns, obtained
from

(38)

The matrix T> can easily be reduced to an orthonormal matrix TF of dimension 6 p, using e.g. the
Gram-Schmidt process 1371,
T; = TFR (39)
The non-singularity of the matrix is here judged from an evaluated number of vanishing eigenvalues,
further discussed in Section 3.4.3. This number, Z, decides the number of eigenvalues that are numerically
vanishing, i.e. are within a certain tolerance from zero.
If the tangential stiffness matrix is found to be numerically singular, eigenvectors corresponding to
almost zero eigenvalues are evaluated, using a subspace iteration technique, which can utilise the storage
properties of the matrix. As vectors corresponding to low eigenvalues are sought, the convergence is
rapid. The critical eigenvectors are collected in a matrix Qi and are obviously contained in the numerical
null space described by TF, when a zero control variable path is appended.
The vectors in @ will give F ,d@ M 0, but there may also exist null vectors to the matrix in Eq. (37)
corresponding to the control parameter columns in F ,A. The parts of these that are orthogonal to the
critical eigenvectors in @ can be solved from an augmented system, which in [3] is given as

[m:
u”]
(:J =-(F(g (40)

This system can be solved in a numerically stable way, as the inclusion of the @ matrices give a non-
singular matrix. The solution to this system contains the orthogonal parts (‘minimal length solutions’)
of the responses to the control parameter columns in td, (aTtd = 0) and the lengths of the projections
of the columns in F ,A on the critical directions in Hz = @ TF ,n. The projection matrix is here of size
{Z/P)-
If the values in nz are not numerically zero, the vectors in td are not tangent vectors to the system.
The result can, however, be used to establish a set of columns

T:, =
If” 1 (41)
 A. Eriksson/Comput. Methods Appl. Mech. Engrg. 156 (1998) 45-74 hl

which is correct in the sense that the product F,zTI, reflects the projection of F.A on the range of F ,d.
The matrix Tk may, however, contain incorrect null vectors. According to a discussion in 131,the system
can only carry ‘force-like’ vectors that are orthogonal to @. By combining the studied columns, such
vectors can be obtained from the null space of the projection matrix n,y. It is thus obvious that the null
space of F ,t contains all vectors in the space spanned by

TK. = Tk . null(Z7.7) 5
1 (42)

where null(n:) seeks out independent linear combinations of control parameter columns that are or-
thogonal to the vectors in 0. This might be empty, if the number of critical eigenvalues is equal to or
higher than the number of control variables. In general, one notes that the number of columns in T$ is
(2 p, 3 z). The columns in T: can give an orthonormal basis TF, similarly as in Eq. (39).
The result from the first phase of tangent evaluation is the matrix TF obtained by orthonormalisation
of the results from Eqs. (38) or (42), depending on the singularity of the tangential stiffness matrix.
It is obvious that the results to some extent depend on the choice of numerical tolerance for a zero
eigenvalue. Due to the expressions used, this influence has been considered minimal in performed tests.
As a second phase, the auxiliary equations are introduced. The key question is then which subspace
of the space spanned by TF also fulfills

g,,t = 0 (43)
This subspace is spanned by

T’ = TF . null(g,Z . TF) (44)

which, after orthonormalisation e.g. by a QR procedure T’ = TR, is the result from the tangent evalu-
ation. It is obvious from the dimensions of the matrices that at least one column must remain.
A special note is needed with respect to the evaluated critical eigenvectors in matrix @ . As discussed
above, it is of interest to differentiate between different inherent vectors in a two- or higher-dimensional
critical space. In order to obtain this, the evaluated columns of @ are orthogonally transformed in order
to reflect the directions for which g,d @ = 0, before introducing them into Eq. (42). This transformation
can be based on an evaluation of null(g,d a).
The obtained basis vectors in T are arbitrary, when the space is more complex than one-dimensional.
Using the information schematically shown by Table 1, it is often possible to orthogonally transform
this in such a way that a first vector, which is orthogonal to the critical vectors, is obtained. In order to
consider this, the obtained columns are orthogonally transformed, aiming at a lower triangular region
of zeroes in the matrix T. From this matrix, the critical directions and the primary response are found,
if existing.
The described procedures have been tested for different relevant situations. Both the method based
on the RRQR factorisation and the eigenspace separated method have been shown to work well. As
the RRQR-based method has been found slightly sensitive to the choices of pivoting window and to the
chosen threshold value, and is also considerably more computationally expensive, the latter method has
been used for the numerical examples below.

3.3.4. Null space of small matrix

The null space of a matrix is needed in several phases of the tangent evaluation. It is worth noting
that all these null space evaluations refer to matrices of small sizes, rather than to the full differential
matrix. Any method can be used, as long as an orthonormal matrix is obtained. Due to the need to allow
for numerical inexactness, the used method is based on a singular value decomposition of the matrix.
Right singular vectors corresponding to singular values below a certain tolerance are thereby extracted.

3.4. Special solution states

Three types of special points on the path are considered in the present algorithm. The cases and
62 A. Eriksson/Comput. Methods Appl. Mech. Engrg. 156 (1998) 45-74

the methods for isolation of these states are briefly discussed below; further information can be found
from [3].

3.4.1. Turning points

Turning points are essentially of the same type as in normal path-following algorithms. They can,
however, be of more general nature, giving for instance the limits for some geometrical value, when a
certain phenomenon exists. An example is given in Fig. 9, where it is shown that no bifurcation exists
for the structure in Fig. 5, if it is not high enough; this is a turning point in the height parameter for this
fold line.
Turning points can be found, using the signs of the components of the path tangent vector, which is
evaluated at all converged points, The tangent vectors at the previous and the current solution points are
re-scaled to give +l in the current parameterising component, and vanishing components can be found
by a secant method or by bisection. In the present implementation, all control variables are monitored
for turning points.

3.4.2. Vanishing variables

It is, e.g. on branch connecting paths, interesting to note points on the curve, where certain variables
vanish. These can be found from a secant method, starting from a situation where a variable changes
its sign between the previous and the current converged solutions. In the present implementation all
control variables are monitored for vanishing values.

3.4.3. Degree of instability

As a generalisation of methods used in equilibrium path tracing, the present algorithm checks the
degree of stability, noting the exchanges of stability, occurring along the path. This is indicated by the
number of negative pivots in the tangential matrix. In a general parametric study, this for instance marks
the limits for stable structural behaviour.
On a fold line these exchanges indicate border lines for parametric regions, where a certain instability
sequence is valid, thus distinguishing between instances of the parameterised structure with different
qualitative behaviours. As an example, Fig. 9 shows the regions where a bifurcation state is passed before
reaching the maximum load state of the structure. The special states at these borders will always show
a higher degree of criticality than the one valid for the followed path-in the figure a hilltop branching
state [8], where a limit state and a bifurcation state coincide. These situations are characterised by two
fold lines-each with single criticality-that intersect at a double criticality.
In the search for a critical state, in the sense of a singular tangential stiffness matrix, a pivoting strategy
by Chan [38] can be used. The strategy seeks a permutation of the rows and columns in connection with
the LU decomposition of a matrix. The permutation ensures that the absolutely lowest diagonal element
will occur in the last diagonal position. As the permutation strategy is related to the numerical null
vector, the search can be used throughout the search for a critical state. This strategy is an improvement
over strategies that only seek the number of negative diagonal elements in the factors.
In the general context, allowing i.e. critical subset paths, it is not sufficient to count the number
of negative eigenvalues, i.e. the degree of instability, in the tangential stiffness matrix K. The present
formulation keeps counts on both the negative and the zero eigenvalues of the matrix for the converged
points. When either of these numbers changes, the specific point on the path can be isolated by bisection.
In order to avoid unnecessary eigensolution extractions, a simplified method is used in the present
formulation (cf. [3]). After finding a converged solution point to the augmented system, the tangential
stiffness matrix, K, is evaluated. For this matrix, both (K - ~1)and (K + cl)are factorised, counting
the numbers of negative pivots 2. These numbers give the numbers of eigenvalues within and below the
range [-E , E], where E is a constant chosen from the normal magnitudes of the diagonal components
of R (cf. [3]). Only when further information is needed concerning the eigenvectors inverse subspace
iterations are utilised.

2The latter is not necessary, if already the first gives zero negative pivots, which simplifies the treatment of common fundamental
paths.
 A. ErikssonIComput. Methods Appl. Mech. Engrg. 156 (1998) 45-74 63

3.4.4. Isolation procedure

The special points indicated by the above-mentioned quantities can be isolated to high accuracy by
a general procedure. It is obvious that all the considered quantities are related to the properties of
the studied converged solution point z and can be seen as a function of a path length s along the
step, measured with the current parameterising component, and normalised such that s = 0 and s = 1
correspond to solutions zo and tr, before and after the step.
The difference between the three cases must be noted. For the vanishing variable or tangent component
cases, exact values are easily obtained. For the exchange of stability, only the signs of the eigenvalues are
obtained, without an elaborate evaluation procedure. As a general procedure is considered necessary.
all the considered quantities are therefore mapped to { -1, 0,1) indicators, when compared to relevant E
borders-dependent on the quantity considered and the specific problem. As a procedure for handling
such variables cannot rely on secant searches, bisection is used.
Fig. 3 shows the possible situations for a step, where this isolation should be initiated. For the basic case
to the left, one variable changes its sign, in this case from positive to negative. This also reflects a situation
where the number of negative eigenvalues increases by one, while the number of zero eigenvalues keeps
the same. A special point, with a vanishing quantity, occurs within this interval.
The procedure aims at finding the points a and b-corresponding to solution points za and zh--in the
sub-figure, i.e. the points where the studied quantity leaves the ‘almost zero’ region. By a systematic
comparison of the properties of points within the interval to those of the starting point zo, the rightmost
point za with properties equal to those at zo and the leftmost point zb with different properties, can be
found.
The bracketing points around the special point are found by a version of bisection, starting from the
initial solutions s, = 0, sb = 1. New points are then solved by adding normalised steps to the current
left bracketing point, s* = sa + 2-i (j = 1,2, . . . , j+). A converged solution z* with this parameter value
is sought and the considered quantities evaluated. Based on the signs of these, compared to the ones at
point zo, the current bracketing parameter sa or sb can be updated. If a point, where the studied quantity
is within the E zone, is found, this can be noted as a lowest limit for the ‘b’ bordering point. This reduces
the interval in which new ‘b’ points can be sought, when in the second phase, similar steps are tested
from the bordering point sb.
In both phases, the procedure is discontinued when a predefined step size 2-j’ has been reached,
indicating the desired accuracy. It is then possible to solve for a point ‘c’ at the midpoint of the interval,
but the special point is normally sufficiently accurately approximated by zC = l/2 (za + tb). This avoids
iterations close to a singular state.
Other possible situations after completion of a step are shown in the centre and right cases. Here,
one of the studied quantities changes from a vanishing to a non-vanishing value, or vice versa, in the
completed step. This isolation then becomes one-sided, aiming to find a point ‘b’ as close as possible
after the passed e-border. The second phase of the above isolation method can be used for this.

Fig. 3. Basic situations for isolation of special solution states. The path length s is measured in the current parameter; c{(s) is any
of the studied quantities.
64 A. ErikssonlComput. Methods Appl. Mech. Engrg. 156 (1998) 45-74

Fig. 4. General structure for storage of general differential matrix, allowing full column pivoting, but disregarding symmetry.
Augmenting bottom block is stored separately.

In the present algorithm, all special states are normally isolated to rather high accuracy. The computa-
tional work therefore easily becomes rather large. This is in particular the case for the branch connecting
paths, where many points of these types are present. In an efficient algorithm for practical problems,
the number of studied quantities must be reduced. This is also very easily done, as only a few criteria
are relevant for a specific problem. For a practical algorithm, it is also easy to synchronise the threshold
values, in order to avoid double isolation procedures.

3.5. Treatment of augmented system

In the systems formulated, the extra rows and columns destroy the symmetry and bandedness, inherent
in the tangential stiffness matrix itself. Special methods must therefore be used, in order to efficiently
allow the solution of the different systems in the algorithm.
In order to treat the ‘arrow-shaped’ systems [39], a procedure, discussed in [l], is a further development
of the postponed factorisation method described in [28]. The improvement in the method is a pivoting
possibility in the tangential stiffness matrix. This is needed, e.g. in cases where a critical subset path is
followed. With this method, the upper left part of the system remains banded and symmetric. The lower
and right parts have no specific structure and must be stored as matrices with a low number of rows
and columns, respectively. The pivoting strategy, where one row and column are moved away from the
tangential stiffness matrix, is successful in most cases.
As the general path-following procedure is often used for systems where the tangential stiffness matrix
is singular, the pivoting possibilities are important for reliably accurate solutions. This demands another
storage form for the matrices. In order to also make the column pivoting in the RRQR procedure
possible, a new matrix storage form is being developed. The storage is based on a column-wise storage
scheme for the active parts of the columns, not making use of the symmetry properties. Allowing for
auxiliary equations and several control variable columns, a suitable storage could be according to Fig. 4.
Although not utilising the symmetry in order to save computational time and storage, this scheme allows
full column pivoting.

4. Numerical examples

A number of examples have been analysed with the present algorithm; some of these are discussed in
[4,5]. In the present paper, only two illustrative structures will be described. For these, extensive analyses
 A. Eriksson/Comput. Methods Appl. Mech. Engrg. 156 (1998) 45-74 65

have been performed, in order to completely reveal their instability behaviour. Some comparisons will
here be made of the different numerical procedures being used, including some reliability and efficiency
aspects. Application of the method to shell problems is studied by Lidstrom [ll].

4.1. Clamped toggle frame

A clamped toggle frame, previously studied analytically by Williams [40] and numerically by several
authors, can show all the discussed possibilities. The basic model is shown in Fig. 5; the basic case
used L = 12.943, E = 1.03. 107, and a beam section of height 0.243, width 0.753. The structure was
modelled with eight elements of the type ‘ST’ (cf. [4]), not utilising the symmetry. The model shows
a snap-through behaviour under a vertical downwards point load -Pi, if sufficiently high; it can also
give several bifurcation states. The curves in Fig. 5b show the behaviour of the structure, when various
heights 0.2 < H 6 0.7 were introduced (cf. the three-dimensional plot of the curves in Fig. 1).
All the described problems below have been studied with the same set of tolerances and switches in
the algorithm. In all cases, modified N-R iterations have been used and a residual tolerance of 1.0 10-7.
The thresholds for low singular and eigenvalues have been set to 0.1 and 0.3, respectively; these can be
compared to the lowest eigenvalue and the lowest singular value of the unloaded structure, which were
228 and 231, respectively. All isolations of special states were performed down to steps of 2--s of the
initial step.

4.1.1. Fundamental path

The behaviour of the structure was described by state variables, corresponding to the displacements
in the 21 degrees-of-freedom, and a control variable PI = 100 Ai. No auxiliary equations were used for
the fundamental path. The solution paths were parameterised by the vertical deflection d,, for which
equal steps of Adi = -0.05 were introduced in each case.
The interesting points in the fundamental equilibrium path for the case H = 0.7 are marked in the
load-displacement path in Fig. 6. These represent the limit load states (marked with ‘0’) at PI = -115.19
and PI = -11.21, respectively, together with the bifurcation states (‘x’), PI = -102.18 and PI = -18.35.
Also, as an example, the states on the path where the load is PI = -100 (‘+‘) are marked. All these
states were easily isolated with the present algorithm. Only the bifurcation states gave two-dimensional
tangent spaces: the presented algorithm succeeds in separating the path tangent direction from the critical
eigenvector.

1.5

1
8
T
a-
l
0.5

n
0.6
-dl

Fig. 5. Load-deflection of clamped toggle frame, according to Williams [40], for different structural heights. A three-dimensional
view is given in Fig. 1. Components of approximate critical modes are indicated.
66 A. Eriksson/Comput. Methods Appl. Mech. Engrg. 156 (1998) 45-74

4.1.2. Non-critical subset path

The response under a specified load was studied, when varying the height H of the structure. In the
formulation, a second control variable was introduced as H = AZ, affecting the internal forcesf(d, A).
An auxiliary equation g(A) = lOO(A, + 1) = 0 was introduced to fix the load value to PI = -100, with a
suitable scaling. Paths were started from the first and third ‘+‘-marked points in Fig. 6 and followed in
both directions towards lower and higher H. The results from these four analyses are collected in Fig. 7,
where the two ‘+’ mark the starting points for the curve segments.
Three special points were detected on the lower curve. Two structural bifurcation points (‘x’) were
found at heights H = 0.68969 and H = 0.76755, respectively. These indicate heights for which buckling
will occur at the specified load. Also, one turning point (‘o’), with respect to the height was found at
H = 0.65889. This is the lowest height, where the specified load can be carried without snapping or
large deflections. At these points, the degree of instability changes for the parameterised structure. The
numbers on the curves indicate this degree in different regions. This can be used to decide the sequence
of criticality at the obtained equilibrium state.
One special point is found when the upper path crosses H = A2 = 0, ‘0’. All points on this curve
segment are stable.
Only the two bifurcation states give tangent spaces of dimension two. The tangential stiffness matrix
has a vanishing eigenvalue and the total differential matrix has two zero singular values. Using the
proposed method, the two-dimensional tangent space is spanned by the path tangent vector and the
buckling mode.
Also at the turning point, an eigenvalue of the tangential stiffness matrix vanishes. Only one zero sin-
gular value exists in the complete differential matrix, however, and the tangent space is one-dimensional.

4.1.3. Limit state subset path

The maximum load was studied, when varying the structural height as in the previous example. An
auxiliary equation was introduced, according to Eq. (13), with 4(O) resembling the snap-through mode,
and containing only the values (1,2,1) as vertical displacements (cf. Fig. 5). Paths were started from the
maximum point ‘0’ in Fig. 6 and followed in both directions. The results from these two calculations are
collected in Fig. S(a), with the path starting point marked with ‘+‘.
Five special points were detected during the analyses. These are the intersections of the limit state
subset with the bifurcation state subsets, giving hilltop branching points [8], with two-dimensional critical

2-

degree of instability
1.5-

7
a-
l
I-
q

0I.5 . /L
/

O-
0 0.2 0.4 0.6 0.8 1 1.2
H
-_ 1 r

0.5 -d, ’ 1.3

Fig. 6. Fundamental solution, i.e. equilibrium path for clamped toggle frame in Fig. 5, with H = 0.7. Interesting points on the
curve are marked for coming examples.

Fig. 7. Non-critical subset path for the problem in Fig. 5, giving the response to a load P , z -100. Markers are used for interesting
points on the curves. Numbers on curve segments denote the degree of instability for points on the segment.
 A. Eriksson/Comput. Methods Appl. Mech. Engrg. 156 (1998) 45-74

6 _ degree of instability

0.2 0.4 0.6 H 0.8 1 1.2

Fig 8. Limit state subset curve for the structure in Fig. 5 under variation of height H. Phenomenological behaviour for this structure
under fixed load Pr = -100.

spaces for the parameterised structures. The hilltop states were accurately isolated, within intervals,
smaller than 0.0012 height units, typically as 0.5671 6 I& 6 0.5677. The degree of instability of the
states on the curve is shown in the figure and is related to the number of bifurcations existing.
The curve is continuous and the tangent unique at the turning point at H = 0.3461, although the
shown projection does not seem to indicate this; cf. the three-dimensional plot in Fig. 1. The passed
state is a ‘third-order’ limit point, existing for this specific height [8].
Figs. 7 and 8(a) can be used to describe the qualitative response of the parameterised structure to a
load PI = -100 (cf. Fig. 8(b)).

4.1.4. Bifurcation state subset path

The buckling load was studied, when varying the height H. An auxiliary equation of criticality was
introduced, but with a vector 4’ of vertical deflections fl (cf. Fig. 5). Paths were started from the first
‘ x ‘-marked point in Fig. 6 and followed in both directions. The results from the analyses are collected
in Fig. 9, where the ‘+’ marks the starting point for the paths.
Three special points were detected during the analyses. One is the turning point at the height where
bifurcations start to exist, ‘0’ at H = 0.50564. The other two are the intersections with the limit state
subset paths studied in Section 4.1.3, ‘*‘. These points led to rather severe convergence problems, which
could, however, be manually overcome.

2.5. degree of instability

0.5.

O- \.
_\ 0
-0.5. limit,loadsX ,
‘\
\\
-1 0 0.5 H i ’ 1:5 2

Fig. 9. Bifurcation state subset curve for the structure in Fig. 5 under variation of height H. Limit state subset curve reproduced
from Fig. 8a.
68 A. ErikssonIComput. Methods Appl. Mech. Engrg. 156 (1998) 45-74

The basic formulation also leads to the solution drifting off from the path, according to the sketched
Fig. 2. This problem was solved by introducing a Lagrange multiplier constraint As, with A3 .lOOOd2 as
the Lagrange force. The A3 became lower than 1.0 lo-l2 for all solutions.
All the points on the curve gave two-dimensional tangent spaces. As the path tangent is always
orthogonal to this critical vector, the choice of the correct tangent vector is simple.

4.1.5. Branch connecting path

Five hilltop branching points can be found in Fig. 8. The one occurring at H = 0.5674 was studied.
Branch connecting paths were evaluated around the point, with PI = -71.25, d’ = -0.2417. The hyper-
sphere centre z* was chosen as the midpoint between the two found bracketing points on a fundamental
path. The radii p, according to Eq. (17) were 0.1 (1d* (1 and 0.05 I)d* 11,
respectively. The initial displace-
ment approximation was set to 0.9d*and 0.95d*, respectively. The auxiliary load vector was chosen as
a point load of size A2 . 100 in the component d2 in Fig. 5. Two paths were thus evaluated, with solution
sequences shown in Fig. 10. It is emphasised that only two, rather arbitrary points, known to bracket
the critical state on the fundamental path, are needed to start the process.
No problems occurred in equilibrium or tangent evaluation, but a large number of special points
needed isolation. The ‘0’ in Fig. 10 are the unnecessarily isolated points of extremum values for the
perturbation load factor A,. The isolated points with vanishing A2 are connected by dash-dotted lines; the
lines are drawn through the hilltop solution state-obtained in another evaluation. All the paths passing
through the hilltop state are clearly indicated, without performing equilibrium iterations close to the
critical state itself. Additional larger or smaller ‘spherical’ paths can improve the available information.
The numbers given in the figure at the found fundamental equilibrium states are not part of the
calculation; they show the differences in primary load -PI, compared to the correct hilltop state. It is
obvious that both the fundamental path and the bifurcating path give second-order curves leaving the
hilltop state, and that the bifurcation is symmetric and unstable.

4.1.6. Hilltop branching state path

Hilltop branching points occur as distinct points on the fold lines for limit or bifurcation states. Refer-
ring to Figs. 8 and 9, these are the ‘*‘-marked points, occurring at specific heights. If one more relevant
parameter is introduced, the hilltop states give continuous curves. Starting from the states H = 0.5814,
PI = -25.25, d, = -0.6770 and H = 0.5674, PI = -71.25, dI = -0.2417, respectively, a model was for-
mulated using the load factor as A,, the structural height H as AZ, the beam sectional height h as A3 and

’ -d,

Fig. 10. Two branch connecting paths around the hilltop state at height H = 0.5674 for the structure in Fig. 5. Solid lines are the
evaluated paths, dash-dotted line (for comparison) connects the paths for the primary problems, drawn through the the correct
critical solution. Load difference compared to correct hilltop state are indicated.
 A. Eriksson/Comput. Methods Appl. Me& Engrg. 156 (1998) 45-74 69

0 I
0 0.2 0.4 0.6 H 0.8 I
“0 0.2 0.4 0.6 H 0.8 1

Fig 11. Hilltop branching states for the problem in Fig. 5, under variation of structural height H and beam sectional height h.

a Lagrange multplier as &. The auxiliary residual equations demanded pll = p22 = d2 = 0, cf. Eq. (11)
and Fig. 5.
The obtained results are shown in Fig. 11. The left sub-figure shows the combinations (H, h) for which
a hilltop state exists. It is noted that these occur for both the ‘maximum’ and the ‘minimum’ load states,
at slightly different combinations. The right sub-figure shows the loads at these hilltop branching states.
With the chosen formulation, no numerical problems were seen in these calculations, even though the
whole paths gave solution states where the tangential stiffness matrix has double zero eigenvalues.

4.1.7. Uptimisation
Based on the structure in Fig. 5, a fictitious optimisation was formulated. The design variables were
chosen as the structural height H, and the sectional height h; the design load was chosen as PI = -100.
The problem was subjected to the constraints that the safety against the limit points should be 2 1.5, the
safety against buckling 2 1.3, and that the maximum stress in the fundamental solution to the design load
should be less than 10000. Due to the overall symmetry of the problem, the maximum stress can only
occur at a few specific spots. In the region of limited deflections, the maximum (compressive) stresses
will be at either the ends or the quarter points of the beams-with the same maximum value occurring
at several places. These two situations were considered individually in the present analysis, Altogether,
four different calculations were thus performed. Each of these used a specification of the studied load
factor as one auxiliary equation, according to Eq. (7). Also, an auxiliary equation of the type in Eq. (8)
for a specified stress position or a criticality equation, according to Eq. (14), was introduced. The modes

0’ I
0 0.5
’ H 1.5 2

Fig 12. Limits for the feasibility region of the structure in Fig. 5, under variations of sectional and structural heights and with
specified criteria.
70 A. ErikssodComput. Methods Appl. Mech. Engrg. 156 (1998) 45-74

according to Fig. 5 were used to describe the bifurcation or the limit state. In the bifurcation case, a
Lagrange constraint for the apex rotation was also included.
The results from the calculations are summarised in Fig. 12, where the combinations of the two
parameters that fulfil the limiting criteria are shown as solid and dotted lines. Obviously these curves
need interpretation to give any conclusions. The dotted parts of the u related curves, are thus excluded,
as these solutions give higher tensile than compressive stresses in the studied sections; they correspond
to structures that must be inverted before being able to carry the specified load. The top half of the
bifurcation curve is also excluded as it refers to the return of the bifurcation path. The turning point on
this curve is the parametric combination for which a degenerated bifurcation point is obtained at the
studied load. The top half of the limit load curve corresponds to the ‘minimum’ load being the specified
value. The turning point on this curve is the parametric combination for which a third order limit point
is obtained at the studied load.
From the remaining parts of the curves, as marked by solid lines in the figure, the feasibility region is
clearly shown. A suitable design, with respect to stated criteria, is obtained for any combination above
these lines.

4.2. Hinged toggle frame

The same structure as in Fig. 5 was analysed, but with the left and right ends hinged, rather than
clamped. This structure is discussed in [4], representing a structure that can show a butterfly buckling

1: L 1
1
L
1
1
0 1 2 3,,-20%, 5 6

Fig. 13. Hinged toggle frame for analysis of butterfly instability. Model, simplified mode and fundamental solution for H = 3.86.
Displacements magnified in dash-dotted curve.

0.04 _ degree. of instability ,.* ’

0 1 2 H3 4 5

Fig. 14. Fold lines for limit and bifurcation states under variation of the height H for the structure in Fig. 13.
 A. Eriksson/Comput. Methods Appl. Mech. Engrg. 156 (1998) 45-74 71

state, for a certain height. The fundamental path for this structure is shown in Fig. 13(b). The basic
structural shape and measures, together with the analysis model, were identical to the problem in the
previous section, except that the standard height was H = 3.86 (cf. Fig. 13(a)). The basic model used 24
unknowns in z = [dT, AlIT.
For H = 3.86, the model will show a bifurcation state (at P1 = -286.3) and a limit state (at
PI = -550.0). An investigation of the bifurcation state, under variation of the height H, shows that
this bifurcation state only exists for H > 0.3818 (cf. Fig. 14); this fold line calculation introduced the
same Lagrange multiplier constraint for the rotation d2 as in Section 4.1.4. Fold lines for the limit states,
starting from the maximum load state in the fundamental case, show that maximum and minimum points
in the response exists for all heights H > 0.1731 (cf. Fig. 14). The fold lines indicate hilltop branching
points at H = 0.449 and H = 0.655.

4.2.1. Post-bifurcation path

The post-bifurcation path for the hinged toggle frame was studied, using the simultaneous path for-
mulation, according to Section 2.6. The model used two displacement states dl and dt, the basic load
factor A,, a control variable A, measuring the distance between the solutions, and a Lagrange multiplier
As, giving the set of residuals as

rf(dl) - &P,A + -43t 1

f(dz) - 0,~
G(z) = =o (45)
eT (d2 - dl) - A2

with c$’ containing two vertical deflections fl (cf. Fig. 13).

The path was initiated from the last solution point before the bifurcation state on the fundamental
path, with PI = -261.6. The solution from this point was introduced as [d;, AlIT. A value A2 = 0.02
was used to create d2 = dl + A,+; obviously the initial approximate solution was very inaccurate. Paths
were initiated in both directions for A,.
Results from calculations with this problem are shown in Fig. 15. The figure shows two projections of
the obtained solution. The dl part stayed on the fundamental path in all points, while the post-critical
displacement vector d2 followed the complete closed loop buckling path.
It is noted that the situation when A2 passes through zero is complicated for the general formulation. In
this passage, one control variable passes through zero, while one changes sign in its tangent component.
Further, an eigenvalue changes from being negative, to being numerically zero, and back to negative.
The number of isolation procedures could be reduced with a more dedicated, ‘intelligent’, algorithm.

0.03 l.. 0.07@ ! I

0.025. 0.06-
0.02. 0.05;:
fund.
0.01.5*
T
Ly 0.01.

0.005.
O-

-0.005.
-0.01
0 0.1 _dZ 0.2 0.3 0 1 2 3-d14 5 6

Fig. 15. Post-bifurcation path for the structure in Fig. 13, with H = 3.86, evaluated with a simultaneous path formulation.
12 A. Eriksson/Comput. Methods A&. Mech. Engrg. 156 (1998) 4.5-74

0.
.N

Y
-0.01.

-0.02 -

-0.03.

-0.04.

0
0 5 10 15 H 20 25 30 -0.05, 5 10 I5 H 20 25

Fig. 16. Simultaneous path evaluation for analysis of bifurcation states. Left sub-figure gives bifurcation load, right sub-figure gives
a measure for post-critical stiffness.

4.2.2. Butterjiy state

As shown in [4,5], this structure shows a butterfly instability for a certain height. This situation is
the borderline between the unstable symmetric bifurcation occurring for lower structures and the stable
symmetric bifurcation occurring for higher structures. This is in [4] by an analytical model shown to be
H = 18.305, whereas the used beam model indicates a height of H = 18.325; the latter result is found
by manually testing the post-bifurcation paths for different heights, starting from a fold line for the
bifurcation behaviour.
The same model was studied with a simultaneous path calculation. The distance between two solutions
was lixed, measured as in the previous subsection. The difference in load factor between the two solutions
was introduced as AZ, the structural height as H = A3, and a Lagrange load factor as Ad, related to the
component d2 in solution vector d 1. The auxiliary equations demanded criticality for the state d 1, giving
the system

‘f(dl,Ad - @,A +A4e

f(d2, A31 - (4 + MP,A

G(z) = gW(dl, 4)) =o (46)

t$T (d2 - dl) - 0.02
tTdl

where g(K) is the criticality quantity, according to Eqs. (13) and (14).
Fig. 16 shows in the left sub-figure how the bifurcation load depends on the height of the structure,
A;“(H). The other part shows the additional load increment for a certain increment along the critical
direction, i.e. a measure for the post-critical stiffness. Noting the basic quadratic post-critical behaviour,
the load level is approximately A, = Aiif + C2(S)2, where 6 reflects the displacement in the buckling
direction. In the right sub-figure of Fig. 16 the post-critical stiffness is measured as C;(H) = -A,/(&)*,
remembering that this is slightly dependent on the chosen distance along the critical mode. From this
curve can be concluded that the post-critical stiffness changes its sign at a height of H = 18.33, which
agrees excellently with results in [4].

5. Concluding remarks

The paper has described a general setting for the evaluation of one-dimensional generalised solution
paths to augmented structural equilibrium problems. The developed methods have been seen in the light
 A. Eriksson/Comput. Methods Appl. Mech. Engrg. 156 (1998) 45-74 73

of theoretical as well as practical instability analyses of optimised structures. In particular, the parameter
dependence, and thereby the sensitivity, in different response aspects has been considered.
The basic method is the formulation of a set of residual equations, where the equilibrium equations
are augmented by equations, stating the specific path wanted.
Some conclusions from the performed studies are:
l The generalised one-dimensional paths can reveal almost all interesting aspects of parameter de-
pendence and sensitivity in simple and complex instability problems.
l By these methods, seemingly uncommon instability situations can be revealed for rather simple
structures.
l The branch connecting path formulation by Huitfeldt [9] is capable of finding all branching paths
in a compound instability situation.
l The special types of instabilities described by catastrophe theory can be found from the general for-
mulation by introducing a suitable number of parameters and a corresponding number of relations.
l A derived method for the evaluation of the tangent space of the general augmented system is shown
to work reliably and rather efficiently.
l The developed methods, which are aimed at being completely general for one-dimensional solution
paths, must be reduced in generality, in order to be efficiently adopted to specific subsets of problems.
Further studies are needed in order to make the described methods useful tools for the daily engi-
neering work. In narticular, the following topics must be considered in this area:
The storage scheme used for the differential matrix should be further studied, in order that the
algorithm should be able to handle efficiently all types of general situations. A storage scheme is
suggested in Section 3.5, but the needed numerical methods must be developed for this storage.
The development of the used methods, in order that they should be suited for treatment in parallel
or vectorising computers, should be further studied.
The idea behind the branch connecting path method should be developed in a less computationally
demanding direction.
The question of automatic scaling of variables and residual components must be developed.
The possibilities for evaluation of multi-dimensional solution subsets can be utilised in some prob-
lems and need be further developed.

References

IllA. Eriksson, Fold lines for sensitivity analyses in structural instability, Comput. Methods Appl. Mech. Engrg. 114 (1994)
77-101.
PI A. Eriksson, Cutsets and augmented equilibrium equations for multi-parametric structural analysis models, Dept. Struct.
Engrg., Royal Inst. Techn., 1993.
[31A. Eriksson, Equilibrium subsets for multi-parametric structural analysis, Comput. Methods Appl. Mech. Engrg. 140 (1997)
305-327.
I41C. Pacoste and A. Eriksson, Element behaviour in post-critical plane frame analysis, Comput. Methods Appl. Mech. Engrg.
125 (1995) 319-343.
PI C:Pacoste and A. Eriksson, Beam elements in instability problems, Comput. Methods Appl. Mech. Engrg. 144 (1997) 163-197.
161C. Pacoste, On the application of catastrophe theory to stability analyses of elastic structures, Dr. Thesis, Dept. Struct. Engrg.,
Royal Inst. Techn., 1993.
171R. Seydel, From Equilibrium to Chaos. Practical Bifurcation and Stability Analysis (Elsevier, New York, 1988).
PI J.M.T. Thompson and G.W. Hunt, Elastic Instability Phenomena (Wiley, Chichester, 1984).
[91J. Huitfeldt, Nonlinear eigenvalue problems-prediction of bifurcation points and branch switching, Numerical Analysis
Group. Report 17, Chalmers University, Goteborg, 1991.
IlO1W.C. Rheinboldt, On the computation of multi-dimensional solution manifolds of parameterized equations, Numer. Math. 53
(1988) 16%181.
[IllT. Lidstrom, Computational methods for finite element instability analyses, Dr. Thesis, Dept. Struct. Engrg.. Royal Inst. Techn..
Stockholm, 1996.
[12] W.C. Rheinboldt and E. Riks, Solution techniques for nonlinear finite element equations, in: A.K. Noor and W.D. Pilkey, eds.,
State-of-the-Art Surveys on Finite Element Technology (ASME, 1983).
1131 W.C. Rheinboldt, Numerical Analysis of Parameterized Nonlinear Equations (Wiley, New York, 1986).
[14] E.L. Allgower and K. Georg, Numerical Continuation Methods. An Introduction (Springer, Berlin, 1990).
[15] E.L. Allgower and K. Georg, Continuation and path following, in: A. Iserles, ed., Acta Numerica 1993 (1993) l-64.
14 A. Eriksson/Comput. Methods Appl. Mech. Engrg. 156 (1998) 45-74

1161 M.A. Crisfield, New solution procedures for linear and non-linear finite element analysis, in: J. Whitman, ed., The Mathematics
of Finite Elements and Applications V (Academic Press, London, 1986).
[17] M.A. Cristield, Non-linear Finite Element Analysis of Solids and Structures (Wiley, Chichester, 1991).
118)R. Kouhia and M. Mikkola, Strategies for structural stability analysis, in: N.-E. Wiberg, ed., Advances in Finite Element
Technology (CIMNE, Barcelona, 1995) 254-278.
[19] A. Griewank and G.W. Reddien, Characterization and computation of generalized turning points, SIAM J. Numer. Anal. 21
(1984) 176-185.
1201P.J. Rabier and G.W. Reddien, Characterization and computation of singular points with maximum rank deficiency, SIAM J.
Numer. Anal. 23 (1986) 1040-1051.
[21] R.-X. Dai, Characterization and Computation of Foldsets for Parameter-dependent Equations, Ph.D. Thesis, University of
Pittsburgh, 1988.
[22] R.-X. Dai and W.C. Rheinboldt, On the computation of manifolds of foldpoints for parameter-dependent problems, SIAM J.
Numer. Anal. 27 (1990) 437446.
[23] B. Wu, Numerical non-linear analysis of secondary buckling in stability problems, Comput. Methods Appl. Mech. Engrg. 120
(1995) 183-193.
1241R. Kouhia and M. Mikkola, Tracing the equilibrium path beyond simple critical points, Int. J. Numer. Methods Engrg. 28
(1989) 2923-2941.
[25] W. Wagner, A path-following algorithm with quadratic predictor, Comput. Struct. 39 (1991) 339-348.
[26] A. Eriksson, On improved predictions for structural equilibrium path evaluations, Int. J. Numer. Methods Engrg. 36 (1993)
201-220.
[27] A. Eriksson and R. Kouhia, On step size adjustments in structural continuation problems, Comput. Struct. 55 (1995) 495-506.
1281A. Eriksson, On linear constraints for Newton-Raphson corrections and critical point searches in structural FE. problems,
Int. J. Numer. Methods Engrg. 28 (1989) 1317-1334.
[29] H.B. Keller, Numerical Solution of Bifurcation and Nonlinear Eigenvalue Problems, in: P Rabinowitz, ed., Applications of
Bifurcation Theory (Academic Press, New York, 1977).
[30] D.W. Decker and H.B. Keller, Path following near bifurcation, Commun. Pure Appl. Mech. 34 (1981) 149-175.
[31] T. Chan, Rank revealing QR factorizations, Lin. Alg. Appl. 88/89 (1987) 67-82.
[32] T. Chan, Some applications of the rank revealing QR factorization, SIAM J. Sci. Stat. Comput. 13 (1992) 727-741.
[33] Y.P. Hong and C.-T. Pan, Rank-revealing QR factorizations and the singular value decomposition, Math. Comput. 58 (1992)
213-232.
[34] C.H. Bischof and PC. Hansen, A block algorithm for computing rank-revealing QR factorizations, Numerical Algorithms 2
(1992) 371-392.
[35] C.H. Bischof and PC. Hansen, Structure preserving and rank-revealing QR-factorizations, SIAM J. Sci. Stat. Comput. 12
(1991) 1332-1350.
[36] C.H. Bischof, Incremental condition estimation, SIAM J. Matrix Anal. Appl. 11 (1990) 312-322.
137) G. Strang, Linear Algebra and Its Applications, 2nd edition (Academic Press, New York, 1980).
[38] T. Chan, On the existence and computation of LU-factorizations with small pivots, Math. Comput. 42 (1984) 535-547.
[39] T.F. Chan and D.C. Resasco, Generalized deflated block-elimination, SIAM J. Numer. Anal. 23 (1986) 913-924.
[40] F.W. Williams, An approach to the nonlinear behavior of members of a rigid jointed plane framework with finite deflections,
Q. J. Mech. Appl. Math. 17 (1964) 451-469.