Sunteți pe pagina 1din 31

INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING

Int. J. Numer. Meth. Engng 2006; 66:604634


Published online 12 December 2005 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/nme.1567

Lower bound limit analysis of cohesive-frictional materials


using second-order cone programming

A. Makrodimopoulos and C. M. Martin,


Department of Engineering Science, University of Oxford, Parks Road, Oxford, OX1 3PJ, U.K.

SUMMARY
The formulation of limit analysis by means of the nite element method leads to an optimization
problem with a large number of variables and constraints. Here we present a method for obtaining
strict lower bound solutions using second-order cone programming (SOCP), for which efcient primal-
dual interior-point algorithms have recently been developed. Following a review of previous work, we
provide a brief introduction to SOCP and describe how lower bound limit analysis can be formulated in
this way. Some methods for exploiting the data structure of the problem are also described, including
an efcient strategy for detecting and removing linearly dependent constraints at the assembly stage.
The benets of employing SOCP are then illustrated with numerical examples. Through the use of an
effective algorithm/software, very large optimization problems with up to 700 000 variables are solved
in minutes on a desktop machine. The numerical examples concern plane strain conditions and the
MohrCoulomb criterion, however we show that SOCP can also be applied to any other problem of
lower bound limit analysis involving a yield function with a conic quadratic form (notable examples
being the DruckerPrager criterion in 2D or 3D, and Nielsens criterion for plates). Copyright 2005
John Wiley & Sons, Ltd.

KEY WORDS: limit analysis; lower bound; cohesive-frictional; nite element; optimization; conic
programming

1. INTRODUCTION

The theory of classical limit analysis for perfectly plastic structures is based on the lower
and upper bound theorems established by Gvozdev [1] and Drucker et al. [2]. However since
limit analysis is a subset of shakedown analysis, it would be fair to consider Melan [3] as

Correspondence to: C. M. Martin, Department of Engineering Science, University of Oxford, Parks Road, Oxford,
OX1 3PJ, U.K.
E-mail: chris.martin@eng.ox.ac.uk

Contract/grant sponsor: Engineering and Physical Sciences Research Council (EPSRC); contract/grant number:
GR/S26897/01
Received 18 April 2005
Revised 24 September 2005
Copyright 2005 John Wiley & Sons, Ltd. Accepted 28 September 2005
LOWER BOUND LIMIT ANALYSIS USING SOCP 605

the originator of the lower bound theorem. More theoretical aspects such as duality and large
strains are covered by Kamenjarzh [4], Christiansen [5] and Gao [6].
Together, the bound theorems allow the exact value of the limit load to be bracketed in a
rigorous manner. It is also possible to obtain a direct estimate of the limit load using various
exact methods (typically based on mixed nite elements), though with this approach there is no
obvious estimate of the error involved. In this paper, we focus on the computationally efcient
determination of strict lower bound solutions. In most engineering applications, solutions derived
from the lower bound theorem are particularly valuable because they provide a safe estimate
of the load that will cause collapse.

1.1. Lower bound analysis using nite elements


When the lower bound theorem is implemented via the nite element method, it is usual
to employ a piecewise continuous stress eld in which adjoining elements are separated by
statically admissible discontinuities. These discontinuities introduce additional degrees of free-
dom, giving results that are generally much better than those obtained from a fully continuous
stress eld. Some of the rst analyses of this type were made by Belytschko and Hodge
[7]. They used quadratic stress elements and obtained strict lower bounds for several exam-
ple problems, though sophisticated procedures were needed to ensure that the yield criterion
was satised throughout each element. In subsequent developments by Lysmer [8], Pastor [9]
and Sloan [10], the use of three-node triangles soon became the favoured approach. With this
element the stress components vary linearly, so it is only necessary to enforce the (convex)
yield restriction at the nodes. In the same way, the use of four-node tetrahedra allows a sim-
ple yet rigorous approach to lower bound limit analysis in 3D (see e.g. References [1113]).
If a strict lower bound solution is sought, there is now a strong preference for adopting
piecewise linear stress elds of this type, despite their low order. This preference has been
reinforced by the development of special linear extension elements that allow the construc-
tion of statically and plastically admissible stress elds for unbounded domains. The original
2D extension elements of Pastor [9] have recently been generalized to 3D by Lyamin and
Sloan [14].

1.2. Limit analysis and optimization


In most formulations of nite element limit analysiswhether they are based on the lower
bound, upper bound or exact approachthe nature of the optimization problem is dened by
the yield function. In early works such as those by Maier [15], Lysmer [8], Anderheggen and
Knpfel [16] and Pastor [9], the yield function was linearized and limit analysis was formulated
as a linear programming (LP) problem. This allowed the use of algorithms and codes that were
mature and robust. However, with the size of the optimization problem depending on both
the neness of the discretization and the number of facets used to approximate the yield
surface, even the best available LP algorithms proved to be relatively weak for large problems.
This was overcome, to some extent, by advances in computing power and in large-scale LP
technology. For example Sloan [17] developed an active set LP algorithm specically for
use with his lower and upper bound methods; Andersen and Christiansen [18] solved large
problems by developing a bespoke interior-point algorithm for LP; and recently Pastor et al.
[19] exploited the capabilities of the commercial LP code XA (also based on the interior-point
method).

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
606 A. MAKRODIMOPOULOS AND C. M. MARTIN

Over the last decade, the emergence of efcient algorithms for large-scale non-linear pro-
gramming (NLP) has seen a gradual move away from LP-based methods for limit analysis.
Following the pioneering work of Belytschko and Hodge [7] and Gao [20] in the use of
yield functions without linearization, Zouain et al. [21] were possibly the rst who anal-
ysed structures with more than 1000 elements, using an NLP algorithm based on the method
of feasible directions. Andersen et al. [22] and Christiansen and Andersen [23] adopted a
different approach, formulating limit analysis as a problem requiring the minimization of a
sum of norms, and applying an appropriate optimization algorithm based on the interior-point
method [24]. Lyamin and Sloan [12, 25] modied the algorithm in Reference [21] and applied
it to various problems in geomechanics, while Krabbenhoft and Damkilde [26] developed a
new interior-point algorithm, also for use with a general (smooth) non-linear yield function.
Tin-Loi and Ngo [27] employed the general-purpose NLP codes CONOPT and MINOS, how-
ever they did not provide many details about the efciency of their analyses. Recent studies
by Yang et al. [13] and Li and Yu [28] have also adopted NLP-based approaches to limit
analysis.
Summarizing, the method of Andersen and Christiansen seems to be the most effective.
The reason is that they use a highly specialized optimization algorithm [24] that is ideally
suited to the arising problem. On the other hand, their method can only be applied to von
Mises-type yield functions. The development of an algorithm able to handle problems with
any non-linear yield function (even restricted to convexity) is complicated by factors such
as possible non-differentiability, and possible singularity of the Hessian. All of the numerical
examples in References [21, 29, 30] were conned to the von Mises criterion, and it seems
that the substantial modications by Pontes et al. [31] were necessary to handle cones with an
apex. Lyamin and Sloan [12, 25] solved several problems involving cohesive-frictional materials,
but in order to face the issue of non-differentiability they smoothed the MohrCoulomb yield
surface. Krabbenhoft and Damkilde [26] also included a cohesive-frictional example problem,
but they did not discuss how the apex of the yield surface was treated (it appears that they
employed a squared form of the MohrCoulomb criterion, and this is discussed in more detail
in Section 3.5).

1.3. Aims and scope of the paper


The basic purpose of the present paper is to show how lower bound limit analysis of cohesive-
frictional materials can be formulated as a second-order cone programming (SOCP) problem.
The advantages of this approach include its efciency, and the fact that it can treat a range of
cone-shaped yield restrictions in their native form (yet avoid the usual difculties encountered
when applying general non-linear optimization methods, i.e. yield function non-differentiability
and Hessian singularity). Our description concentrates on the case of the MohrCoulomb cri-
terion in plane strain, however a directly analogous formulation is feasible for a number of
other important yield criteria and problem types, notably the DruckerPrager criterion for 2D
or 3D continua, as well as Nielsens criterion for plates and Ilyushins criterion for shells.
All of these yield restrictions represent second-order cone sets, and thus SOCP is a class of
non-linear optimization that suits perfectly.
Although SOCP has been used in several previous studies involving both shakedown and
limit analysis [3237], these were all restricted to von Mises materials. In fact such problems
can be solved in many other ways, for example by applying an NLP algorithm specialized

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
LOWER BOUND LIMIT ANALYSIS USING SOCP 607

for quadratic constraints, or indeed by a general-purpose NLP algorithm (there is no problem


of non-differentiability). The use of SOCP in these earlier works was therefore not essential,
but happened to provide a convenient and efcient way of handling the quadratic constraints
generated by the von Mises criterion. In this paper, the constraints are not simply quadratic
because the formulation encompasses cohesive-frictional materials as well. We aim to show
that SOCP remains a natural choice for solving these more difcult problems, where the yield
function is no longer differentiable everywhere.
The use of an algorithm (and associated software) for SOCP that is considered to represent
the current state of the art proves that our formulation is advantageous. Several numerical
examples, all involving very large optimization problems, are solved with remarkable speed. It
is also noteworthy that these results are obtained without making any changes to the default
settings of the software, and without encountering any computational difculties (e.g. instability
of the interior-point iterations). Additionally, we explore the structure of the equality constraints
generated by the inter-element equilibrium conditions and the stress boundary conditions in
lower bound analysis. By examining when linearly dependent constraints arise, we are able to
identify and remove redundant rows during (rather than after) the assembly of the constraint
matrix, further enhancing efciency and stability. We also outline a method for substantially
reducing the number of variables and constraints, though it appears that this only brings marginal
benets in terms of overall run time.
Finally, a key advantage of our approach is that both SOCP and the interior-point method
are subjects of great interest within the mathematical programming community, and thus re-
lated algorithms are under continuous development. This means that in the future, even more
effective algorithms for SOCP will appear, and the proposed formulation will be even more
attractive.

2. CONIC OPTIMIZATION

2.1. Cone sets


A set C is called a cone if x C and   0, x C. Its dual cone C is dened as

xT y  0 x C y C (1)

If C = C then the cone is self-dual. For example the set + = {x : x  0} is a self-dual cone,
as is the quadratic (or second-order) cone, which has the form

C = {x d : x2:d   x1 , x1  0} (2)

where x2:d = [x2 . . . xd ]T . Note that any constraint of the form Ax + b  cT x + d can be
transformed to a quadratic cone. Another interesting self-dual example is the rotated quadratic
cone

C = {x d : x3:d 2  2x1 x2 , x1 , x2  0} (3)

Although this can readily be transformed into a non-rotated quadratic cone of the form (2), it
can also be treated directly in its original form, see Reference [38].

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
608 A. MAKRODIMOPOULOS AND C. M. MARTIN

2.2. Second-order cone programming (SOCP)


Second-order cone programming (also referred to in the literature as conic quadratic optimiza-
tion) involves an optimization problem of the form

min cT x
s.t. Ax = b (4)
xi Ci i {1, . . . , N}

where xT = [x1T . . . xN
T ] n , x di , A mn , b m , c n and the C are quadratic or
i i
rotated quadratic cones. The dual problem corresponding to (4) is

max bT y
s.t. AT y + s = c (5)
si Ci i {1, . . . , N}

where y m , s n and si di . The optimal point must satisfy the following conditions
[38, 39]:
Ax = b
A y+s=c
T

xi Ci i {1, . . . , N} (6)
si Ci i {1, . . . , N}
Xi Si ei = 0 i {1, . . . , N}

where Xi , Si di di and ei = [1 0 . . . 0]T di . The matrices Xi and Si are given by
mat(Ti xi ) and mat(Ti si ), respectively, where Ti = Idi for a quadratic cone (Ti for a rotated
quadratic cone is given in Reference [38]) and the mat function is dened as
 
T
u1 u2:d
mat(u) = , u d (7)
u2:d u1 Id1
Equations (6) show that SOCP can be considered a generalization of LP. The primal and
dual equality constraints are the same as in LP, the conic constraints are generalized versions
of the non-negativity constraints in LP (xi 0, si 0), and the complementarity condition is a
generalized version of that in LP (xi si = 0, which corresponds to Ti = 1 and ei = 1).
SOCP also encompasses several important classes of non-linear optimization as special cases.
These include minimization of a sum of norms, convex quadratic programming, and convex
quadratically constrained linear programming. In the latter case, recasting the problem in SOCP
form is particularly advantageous both theoretically [40] and in practice [3237].

2.3. Algorithms and software for SOCP


As is the case with LP, large-scale SOCP problems can be solved effectively using primal-dual
algorithms based on the interior-point method. Indeed, it has been shown that this approach
retains its theoretical efciency (polynomial time) when it is generalized from LP to SOCP

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
LOWER BOUND LIMIT ANALYSIS USING SOCP 609

[41, 42]. For this reason, interior-point algorithms for SOCP have received considerable attention
in recent years.
At present one of the leading algorithms is that of Andersen et al. [38], which has been
implemented in MOSEK [43] and has proved to be highly robust and efcient in a suite of
independent benchmark tests [44]. Other algorithms/codes for solving SOCP problems include
SeDuMi [45], SDPT3 [46] and the programs developed by Lobo et al. [47] and Kuo and
Mittelmann [48]. It should also be pointed out that algorithms for semidenite programming
(SDP) can solve SOCP problems as a special case, though as mentioned in Reference [47] this
is not the most efcient approach.
We are inclined to share the view of Christiansen and Andersen [23] that it should not
be necessary for experts in continuum mechanics to develop and implement software for opti-
mization, unless the optimization method is a research topic in itself. For specic engineering
applications of SOCP, there can of course be certain advantages in developing a bespoke al-
gorithm (see e.g. Reference [49]). On the other hand, there are obvious benets in employing
mature SOCP solvers that have been developed and implemented by experts in mathematical
programming. In the case of MOSEK, for example, many aspects of the underlying algorithm
have been studied and tested in detail, including:
use of a simple starting point that need not satisfy the equality constraints [38, 50];
robust detection of infeasible problems [38];
choice of the search direction and step size at each iteration, including a customized
adaptation of Mehrotras predictor-corrector [38, 41];
exploitation of problem structure and sparsity when computing the search direction [38];
techniques for handling dense columns [51] and free variables [52].

3. LOWER BOUND ANALYSIS

3.1. Finite element implementation


Consider a rigidperfectly plastic structure V with boundary surface S. Tractions t are prescribed
on St and displacements u on Su , such that St Su = S and St Su = . For conciseness we
will assume for the time being that u = 0 on the whole of Su ; this will not be the case for
problems involving rigid indenters or footings, but these are easily handled as described in
Section 3.4. According to the lower bound theorem, the structure is safe for any stress eld
that is both statically and plastically admissible, i.e. satises internal equilibrium, the stress
boundary conditions and the yield criterion:

div +  = 0 in V (8a)
=t
n on St (8b)
0
f () in V (8c)
Here  denotes the matrix of the stress tensor (with tensile direct stresses positive),  is the
eld of body forces per unit volume (assumed for simplicity to be prescribed), n denotes the
unit outward normal to St , and f is the yield function. In order for the lower bound theorem
to be valid, f must be convex and the associated ow rule (normality) must apply.

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
610 A. MAKRODIMOPOULOS AND C. M. MARTIN

t2

2 t1
2 2
1
n n
1 t 1 t

3 3

(a) (b)

Figure 1. Edges in nite element mesh: (a) interior; and (b) exterior on traction boundary St .

In this paper, we concentrate on plane strain conditions; some extensions to other problem
types are briey discussed later in Section 5. Following the approach developed by Lysmer [8]
and further by Pastor [9], the structure is discretized into three-node triangular stress elements.
Each element is associated with its own set of internal stress evaluation points, see Figure 1, and
area co-ordinates (linear shape functions) are used to interpolate the nodal stress components
i = [xx,i yy,i xy,i ]T , i = 1, 2, 3 (9)
Equilibrium within an element is governed by Equation (8a), and this leads to the linear equality
constraints

1

[A1 A2 A3 ] 2 =  (10)
3
where
   
x 1 y2 y3 0 x3 x2
= , A1 =
y 2 0 x3 x2 y2 y3

and A2 and A3 are obtained by cyclic permutation of the subscripts.  is the area of the
element.
Equilibrium between two adjacent elements, see Figure 1(a), does not require full continuity
of the stress tensor  across the shared edge; it is sufcient to enforce continuity of the
where n is a unit normal to the edge. The advantage of using internal stress
stress vector n,
evaluation points now becomes apparent: the direct stress acting parallel to the edge can be
discontinuous, allowing greater freedom in the choice of a statically admissible stress eld.
Since the stresses vary linearly within each element, continuity of n at both ends of the
shared edge is sufcient to ensure continuity at all points in between. The requisite equality
constraints are linear, and take the form

1
 
B0 0 B0
0 2
=0 (11)
0 B0 0 B0 1
2

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
LOWER BOUND LIMIT ANALYSIS USING SOCP 611

where
 
cos  0 sin 
B0 =
0 sin  cos 

The stress boundary conditions (8b) are treated similarly. For the exterior edge shown in
Figure 1(b), the equilibrium constraints are
 
 
C0 0 1 t1
= (12)
0 C0 2 t2

where C0 is dened analogously to B0 above.


With regard to the yield criterion, recall that f is convex, and that simplex stress elements
are being employed (at any point within an element, the interpolated stresses are a convex
combination of the nodal stresses). This conveniently means that if (8c) is satised at every
stress evaluation point, it is automatically satised throughout every element. If higher-order
stress elements are used, special optimization procedures are needed to ensure global compliance
with the yield criterion, and thus a rigorous lower bound solution [7].
Once assembled, the various constraints take the form
A = 
B = 0
(13)
C = t
f (i )  0 i {1, . . . , N}

where N is the number of stress evaluation points and  T = [T1 . . . TN ].

3.2. Extension elements


The development above assumes that the structure V has nite dimensions, but in certain prob-
lems (notably in geomechanics) it is necessary to apply the lower bound theorem over innite
or semi-innite domains. This can be achieved with the use of special extension elements that
permit the stress eld to be continued to innity in a statically and plastically admissible man-
ner. Such elements were rst developed by Pastor [9] for plane strain conditions. Subsequently,
the extension element concept has been generalized and applied to a variety of problems in
both 2D and 3D, notably by Sloan and co-workers. Full details are given by Lyamin and
Sloan [14], though it should be noted that the implementation described in their paper does
not necessarily ensure that the yield criterion is satised throughout the extension eld (see
Appendix A).

3.3. MohrCoulomb criterion


For plane strain conditions, the MohrCoulomb yield criterion can be expressed as

(xx yy )2 + 42xy + (xx + yy ) sin  2c cos   0 (14)

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
612 A. MAKRODIMOPOULOS AND C. M. MARTIN

where c is the cohesion and  is the angle of internal friction. Adopting the notation

m = 21 (xx + yy ), sij = ij m ij ( is Kroneckers ) (15)

the yield criterion takes the simple form



2 + s 2 +  sin  c cos   0
sxx (16)
xy m

By introducing an auxiliary variable, say z, this restriction can be transformed into a linear
equality constraint coupled with a quadratic cone constraint:

z + am = k
(17)
2 + s2  z
sxx xy

where a = sin  and k = c cos . Expressed in this way, the yield criterion is now compatible
with the standard SOCP formulation (4). For a purely cohesive material,  = 0 and the Mohr
Coulomb criterion reduces to the Tresca criterion.
Note that for plane strain conditions, the MohrCoulomb (Tresca) and DruckerPrager (von
Mises) criteria are equivalent. The corresponding relationship between the two sets of model
parameters is well known, and is given for example on p. 23 of the book by Chen [53].

3.4. Formulation as SOCP problem


In the simplest case, lower bound limit analysis involves maximizing a load multiplier that
acts on the whole of the prescribed traction eld, while ensuring that the equilibrium and yield
constraints outlined above are satised. When expressing the optimization problem in SOCP
form, it is convenient in light of (17) to use as the nodal stress components

si = [m, i sxx,i sxy,i ]T (18)

rather than those dened in (9). We then obtain the problem

max 
s.t. A s = 
B s = 0
C s t = 0 (19)
zi + am, i = k i {1, . . . , N}

2
sxx,i + sxy,i
2 z
i i {1, . . . , N}

where a = sin , k = c cos , N is the number of stress evaluation points and sT = [s1T . . . sN
T ].

Spatial variation of the material parameters is permissible, but within a given element a must
be constant and k can vary at most linearly (otherwise convexity is lost and the yield criterion
may be violated inside the element). In terms of the MohrCoulomb parameters,  is required
to be constant within a given element but c can vary linearly. The matrices A , B and C are

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
LOWER BOUND LIMIT ANALYSIS USING SOCP 613

induced from their counterparts in (13) considering



1 1 0

i = Qsi where Q = 1 1 0 (20)
0 0 1
When the objective function consists of a load multiplier, as in (19), it is obviously linear.
More complex situations may involve maximizing the resultant of a surface pressure eld (e.g.
indentation or bearing capacity problems), or optimization with respect to body forces (e.g.
slope stability problems). In all these cases the objective function is again linear, as required
for the use of SOCP. Appropriate modications to the basic structure (19) must also be made
if any extension elements are present.
It is clear that the nite element formulation of the lower bound theorem leads to an
optimization problem that is extremely sparse. Considering the equality constraints in (19), the
number of non-zero elements per column does not exceed six, with the possible exception of the
column corresponding to the load multiplier . This sparsity is very important for the efcient
operation of SOCP algorithms based on the interior-point method. Most modern algorithms
incorporate special routines to ensure that occasional dense columns do not degrade the overall
efciency of the solution process.
We also observe that if a > 0 (i.e.  > 0 in terms of MohrCoulomb) then the m variables
and the equality constraints zi + am, i = k can be eliminated from the problem by expressing
each m, i in terms of its respective zi . However, this cannot happen if a = 0 (i.e.  = 0); to
obtain the standard SOCP formulation (4) we must retain both the m and z variables, along
with the equality constraints zi = k. This suggests that when  = 0 it would possibly be more
efcient to use an NLP algorithm specialized for convex quadratic constraints, in which case
the auxiliary z variables would not be required, though the reduced cost of each iteration could
well be outweighed by an increased number of iterations. This was investigated by the rst
author [32] for some cases of plane stress shakedown analysis using the von Mises criterion,
and in fact better efciency was obtained when the analyses were formulated and solved as
SOCP problems.

3.5. Difculties with general NLP approaches


An important advantage of using SOCP is that when  > 0, the singular apex point of the
MohrCoulomb yield function in (14) poses no difculty. By contrast, general NLP algorithms
such as those employed by Lyamin and Sloan [12] and Krabbenhoft and Damkilde [26] require
the yield function to be twice continuously differentiable, so that calculations involving its
gradient and Hessian can be performed. In Reference [12] a hyperbolic approximation was
used to smooth the MohrCoulomb yield surface (internally) in the vicinity of the apex. On
the other hand, in Reference [26] there was no discussion of smoothing since these authors
implemented the squared form of (14), namely

(xx yy )2 + 42xy (2c cos  (xx + yy ) sin )2  0 (21)

Although this solves the problem of non-differentiability, we now have a non-linear inequality
constraint g()0 where g is non-convex, except when  = 0. If the additional (linear) inequal-
ity constraint xx + yy 2c cot  is applied when  > 0, which it should be in any case to

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
614 A. MAKRODIMOPOULOS AND C. M. MARTIN

exclude physically meaningless stress states, then the two restrictions together dene a convex
set. However, the gradient of the squared yield function in (21) vanishes where the inequalities
intersect, so no reasonable constraint qualication (needed to establish the KarushKuhnTucker
conditions, see e.g. Reference [54]) is satised at the apex point xx = yy = c cot , xy = 0.
Moreover, the Hessian of the squared yield function is not positive denite, and this may have
important consequences for the efcient computation of the search direction at each iteration.
This whole issue is complicated, but it would seem that considerable caution is needed if a
derived yield criterion such as (21) is employed.
Another feature of the (original) MohrCoulomb yield function in (14) is that its Hessian
is singular. While this is not an issue for SOCP algorithms, it is inconvenient for general
NLP algorithms since the block diagonal matrix of Hessians (one for each yield constraint)
needs to be inverted at each iteration. To allow this to be done efciently, it is common
practice to regularize each Hessian by adding a small positive multiple of either the identity
matrix [12, 26] or some other perturbation matrix [31]. In choosing the small multiple there is
obviously a tradeoff between the stability of the inversion and the accuracy of the calculated
search direction. Although simple strategies such as those employed in References [12, 26, 31]
appear to perform satisfactorily in practice, it is also possible for the perturbations to be adapted
dynamically (in relation to the current values of the primal and dual variables) to achieve a
balance between stability and accuracy. These issues were examined by Mszros [55] when
investigating the analogous singularity problems that arise in interior-point algorithms when
free variables are present.

4. EXPLOITATION OF PROBLEM STRUCTURE

In this section we consider some characteristic features of lower bound limit analysis using
nite elements, and discuss how these can be exploited to improve the efciency of the
optimization process. These techniques are by no means specic to SOCP, and could also be
used to advantage in LP- and general NLP-based formulations.

4.1. Avoiding linearly dependent constraints


It is usual for optimization software to carry out a presolve procedure before entering the main
solution algorithm. Some objectives of the presolve phase are to reduce the size of the problem
by eliminating redundant variables and/or constraints, to identify inconsistent constraints (in
which case the problem is trivially infeasible), and to improve the scaling of the problem. As
a result, the optimization can potentially become faster and more stable.
When the matrix of equality constraintsdenoted A in (4)is analysed during presolve, it
is relatively easy to detect and remove obvious redundancies such as zero rows, or rows that
are multiples of other rows [56]. The detection and removal of hidden linear dependencies is
more difcult. This is still desirable, however, since in most optimization algorithms (interior-
point and otherwise) the linear system used to determine the search direction becomes singular
if A does not have full row rank. If dependent rows are present and they are not removed, the
effectiveness of the optimization will depend to a large extent on the ability of the software to
nd (or approximate) the solution of a semidenite linear system at each iteration. Even if the
code has been designed to do this, the presolve routine should still endeavour to remove as

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
LOWER BOUND LIMIT ANALYSIS USING SOCP 615

3 3
4 2 4 2
1 n
t 1 n
t
12 12
(a) (b)

Figure 2. Interior node belonging to N = 4 elements: (a) general case; and


(b) case giving linear dependency.

many dependencies as possible. Tests conducted by Andersen and Andersen [56] have indicated
that the removal of dependent rows during presolve can provide substantial net savings in
solution time, as well as better stability during the main optimization phase.
In previous work on lower bound limit analysis using nite elements, the topic of linearly
dependent constraints has received very little attention. Lyamin and Sloan [12] mention that,
as part of their presolve routine, they factor the square matrix AAT . If any diagonal element is
zero (to within some tolerance) they eliminate the respective row, recalculate AAT and repeat
the factorization. In other studies it appears that no particular effort has been made to detect
and remove dependent rows, even though it could be very benecial to do so [56].
We now describe a method for avoiding redundant constraints, without the need to check
the assembled constraint matrix A for hidden dependencies. The rst step is to identify the
precise conditions that lead to linearly dependent constraints in lower bound limit analysis.
This motivates a simple procedure for detecting and removing dependent rows while A is
being assembled, with the result that it has full row rank from the outset (there are no
dependencies for the presolve routine to nd). The computational cost of this procedure is
absolutely negligible.

4.1.1. Interior nodes. Figure 2(a) shows an interior node of the mesh belonging to N = 4
elements. From (11), (19) and (20), the conditions for inter-element equilibrium between two
sequential stress evaluation points i, j are:
    
cos ij cos ij sin ij cos ij cos ij sin ij si 0
= (22)
sin ij sin ij cos ij sin ij sin ij cos ij sj 0
This pair of equations can be transformed into various alternative forms, including
    
1 cos 2ij sin 2ij 1 cos 2ij sin 2ij si 0
= (23)
0 sin 2ij cos 2ij 0 sin 2ij cos 2ij sj 0
and
    
cos 2ij 1 0 cos 2ij 1 0 si 0
= (24)
sin 2ij 0 1 sin 2ij 0 1 sj 0
Version (23) has a transparent physical interpretation, namely continuity of the normal and
shear stresses nn and nt across the shared edge, though (24) is more convenient for the
development below (it also provides slightly better sparsity).

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
616 A. MAKRODIMOPOULOS AND C. M. MARTIN

Consider all constraints of the form (24) derived from the N = 4 shared edges in
Figure 2(a). The only other equality constraints involving the N sets of stress variables are
those related to equilibrium within the N elements. It can be demonstratedthough a formal
proof is tediousthat provided the mesh is free of degenerate elements having zero area, there
is never any dependency between the inter-element and intra-element equilibrium constraints.
Since the latter are themselves independent, we can isolate the former and perform a purely
local check for dependency. Dening

aij = [cos 2ij sin 2ij ]T (25)

and grouping the s and m variables separately, we nd for the case N = 4:



I I 0 0 a12 a12 0 0
 node 
0 I 0 a23 a23 0
I 0 s
=0 (26)
0 0 I I 0 0 a34 a34 nodem
I 0 0 I a41 0 0 a41
where

snode = [sxx,1 sxy,1 ... sxx,4 sxy,4 ]T , node


m = [m, 1 . . . m, 4 ]T

and the identity blocks I have dimension 2 2. Applying Gaussian elimination gives

I I 0 0 a12 a12 0 0
 
0 I I 0 a23 snode
0 a23 0
=0 (27)
0 0 I I 0 0 a34 a34 nodem
0 0 0 0 a12 a41 a23 a12 a34 a23 a41 a34
Analogous matrices are obtained for other values of N . By inspection, the rst 2(N 1) rows
of the reduced system are independent, and furthermore these rows are independent from the
nal two. This implies that linear dependence, if any, can only exist between the nal two
rows. It is easy to verify that for the case where a quadrilateral is divided into four triangles
by its diagonals, i.e.

N = 4, |34 12 | = , |41 23 | =  (28)

there is always dependence. This case is illustrated in Figure 2(b). In Appendix B we prove that
the conditions (28) are the only ones under which a redundant equality constraint is generated
at an interior node, assuming no degenerate triangles are present. This turns out to be quite a
common conguration when structured meshes are employed in lower bound limit analysis, not
only for benchmarks such as those in Sections 6.1 and 6.2, but also for engineering applications
(see e.g. References [57, 58]). On the other hand, an unstructured mesh is unlikely to give rise
to any dependencies, at least at interior nodes.
To avoid introducing linearly dependent rows into the constraint matrix, the key is to assemble
the inter-element equilibrium equations by traversing the mesh from node to node, rather than
edge to edge. Whenever an interior node shared by four elements is encountered, it is simply
a matter of checking whether the other conditions in (28) are both satised. If the conditions

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
LOWER BOUND LIMIT ANALYSIS USING SOCP 617

180 180

(a) (b) (c)

Figure 3. Exterior node on traction boundary St : cases giving linear


dependency: (a) N = 1; (b) N = 2; and (c) N = 3.

for dependency are conrmed, only seven of the eight inter-element equilibrium equations
are independent, and one can be omitted from the assembly (though of course the choice
is not arbitrary). The obvious benet of this approach is that we never have to search the
fully assembled constraint matrix for hidden dependencies during presolve. A similarly efcient
strategy can be developed for 3D problems.

4.1.2. Exterior nodes. It is easy to show that a node on the displacement boundary Su can
never be responsible for a dependency. However, for nodes on the traction boundary St there
are three congurations that lead to linearly dependent (or possibly inconsistent) constraints.
These are shown in Figure 3. Some instances of dependency-generating exterior nodes, for both
N = 1 and 2, can be seen on the boundaries of the unstructured meshes in Figures 5(b) and
8, though there are no such cases on the boundaries of the structured meshes in Figures 5(a)
and 6. The detection of dependency (or infeasibility) is as simple and cheap for a node on St
as it is for an interior node; again the key is to assemble the equilibrium constraints node by
node, rather than edge by edge.

4.2. Eliminating free variables


We now touch briey on another issue related to the presolve phase, namely the elimination
of certain variables with a view to reducing the size of the problem. In LP it is only the free
(unbounded) variables that can be eliminated, otherwise the standard form of the non-negativity
constraints is lost. In the same way, when lower bound limit analysis is cast as a standard
SOCP problem in (19), only the m variables can be substituted out of the problem without
disrupting the simple structure of the quadratic cone constraints. (In general NLP formulations,
the desire to preserve the uniformity and sparsity of the non-linear yield constraints would
impose a similar restriction.) As noted in Section 3.4, the m variables are eliminated anyway
when  > 0; they only remain when  = 0.
In general it can be counterproductive to eliminate free variables from a large, sparse op-
timization problem because of excessive ll-in. This, however, is a consequence of the global
Gaussian elimination performed on the full constraint matrix. Here we outline a local reduction
technique that allows a large proportion (though not all) of the m variables to be eliminated,
with only a mild loss of sparsity. Consider once more the equality constraints associated with
a node where N elements meet. We note that with the conditions for inter-element equilibrium
expressed in the form (23), the second equation does not involve m, i or m, j . This makes it

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
618 A. MAKRODIMOPOULOS AND C. M. MARTIN

simple to express m, 2 . . . m, N in terms of just the s variables and m, 1 as follows:

m,2:N = Rsnode (29)

where

snode = [sxx,1 sxy,1 . . . sxx,N sxy,N m, 1 ]T 2N +1

and R (N 1)(2N+1) . At an interior node, this allows us to eliminate (N 1) of the 3N


stress variables, and (N 1) of the 2N equality constraints. At an exterior node on St , we can
again eliminate (N 1) of the 3N stress variables, along with (N 1) of the 2(N + 1) equality
constraints. (In either case there may still be a dependency among the remaining equality
constraints, allowing one to be eliminated as discussed above.) The drawback of this process
is that the constraint matrix becomes denser, with the degree of ll-in depending on the value
of N. Some tests on typical problems have shown that the speed of the optimization can be
improved, but the difference is not very noteworthy (less than 10%). This type of elimination
could be more protable for algorithms that cannot handle a large number of free variables
successfully.
In closing, we re-emphasize that this reduction strategy is only protable when  = 0 and the
m variables are genuinely free, i.e. not participating (even indirectly) in any yield constraint.
If the strategy were applied to a problem where  > 0, it would be necessary to retain the
equality constraints zi + am, i = k in (19), instead of using them to eliminate all of the m, i
trivially. So in practice we would not have removed any rows from the constraint matrix, and
despite the removal of some columns, the remaining ones would have become much denser.

5. EXTENSIONS TO OTHER PROBLEM TYPES

5.1. DruckerPrager criterion (3D solids)


The present formulation of lower bound limit analysis can readily be extended to continuum
problems in 3D, employing four-node tetrahedral stress elements [12, 13] in conjunction with
the cohesive-frictional yield criterion of Drucker and Prager [59]. With notation

m = 13 (xx + yy + zz ), sij = ij m ij (30)

the DruckerPrager criterion has the form



J2 + am k  0 (31)

where J2 = 21 sij sij is the second invariant of the deviatoric stress tensor and a, k are material
parameters (when a = 0 the von Mises criterion is recovered). Following the approach used in
Section 3.3, restriction (31) can be transformed into a linear equality constraint coupled with
a quadratic cone constraint:

z + am = k
(32)
y  z

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
LOWER BOUND LIMIT ANALYSIS USING SOCP 619

where y 5 is given (non-uniquely) by



1 1/2 0 0 0 sxx

0 3/2 0 0 0 syy

y=
0 0 1 0 0
sxy (33)

0 0 0 1 0 sxz
0 0 0 0 1 syz

Specialization to plane strain or plane stress is straightforward. In general we can apply SOCP
to any problem of limit (or shakedown) analysis involving a yield criterion of the form

T M + qT  k  0 (34)

where , q d and M dd is at least positive semidenite. As noted previously, SOCP


was used in References [3237] to solve various problems involving the von Mises criterion,
for which q = 0 and (34) reduces to a quadratic constraint. However no analyses with pressure-
dependent yield criteria such as (14) and (31) were performed in these earlier works.

5.2. Nielsen criterion (plates)


Nielsens criterion [60] is widely used in the calculation of limit loads for plate bending
problems (e.g. plastic design of reinforced concrete slabs). In Reference [26] the benchmark
problem of a clamped square plate under uniform pressure was solved using triangular plate
elements and a general NLP algorithm. It is also possible to express and solve this problem in
SOCP form, thus overcoming the issues related to non-convexity analogous to those discussed
in Section 3.5. The Nielsen criterion has the general form
+
Mpx  Mxx Mpx
+
Mpy  Myy Mpy
(35)
+ +
2
Mxy  (Mpx Mxx )(Mpy Myy )

2
Mxy  (Mpx + Mxx )(Mpy + Myy )

where Mpx + , M + are the positive yielding moments and M , M the negative ones. In fact
py px py
this restriction is the intersection of two rotated quadratic cones, which can be seen if we set
+ +
2X = Mpx Mxx
+ +
2Y = Mpy Myy
+
+ (36)
2X = Mpx + Mpx 2X
+
+
2Y = Mpy + Mpy 2Y

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
620 A. MAKRODIMOPOULOS AND C. M. MARTIN

Finally we obtain a system compatible with the standard SOCP formulation (4), containing two
linear equality constraints and two rotated quadratic cone constraints:
+ +
2X + 2X = Mpx + Mpx
+ +
2Y + 2Y = Mpy + Mpy
2
Mxy  2X + Y + (37)
2
Mxy  2X Y
X+ , X , Y + , Y  0

6. NUMERICAL EXAMPLES

The example problems below are all well known plane strain benchmarks in the area of limit
analysis. The MohrCoulomb criterion is employed throughout. For brevity we focus on the
lower bound solutions and the efciency with which they were obtained, without entering into
detailed interpretation or discussion of the respective stress elds, or a detailed interpretation
of the dual variables.
The computations were performed on a Dell Pentium IV machine (2.66 GHz CPU, 2 GB
RAM) in the Windows XP environment, using the conic interior-point optimizer of the MOSEK
package [38, 43] mentioned in Section 2.3. For all analyses the default convergence tolerances
of the software were retained. The input data les were prepared using an extended version
of the standard MPS format. The reported CPU times refer to the time actually spent on the
interior-point iterations, i.e. they exclude the time taken to read the data le and execute the
presolve routine (which spends most of its time reordering the rows of the constraint matrix).
The presolve typically required about 30% of the CPU time needed for the main optimization
phase, except for the nest mesh for the thick cylinder problem (Section 6.2) where the presolve
took slightly longer than the optimization. Note that since all linearly dependent constraints
were detected and removed a priori using the technique described in Section 4.1, it was possible
to disable the corresponding aspect of MOSEKs presolve procedure.

6.1. Block with thin symmetric cuts


Figure 4 shows a slotted plane strain block under uniform tension. This problem has been
studied extensively, for example in References [15, 18, 2224] where mixed elements were
used, and in Reference [26] where strict lower bounds were obtained using three-node triangular
stress elements with discontinuities (as used here). The problem has also been studied using
displacement elements of various orders [27]. In all of these cases the analysis was limited to
purely cohesive material (c = 1,  = 0). Here we also present results for a unit cohesion and
friction angles of 15 and 30 .
We rst considered various N N meshes of squares subdivided into four triangles, modelling
only the upper right quarter of the structure because of the horizontal and vertical symmetry, as
shown in Figure 5(a). Details of the meshes used, and the sizes of the resulting optimization
problems, are given in Table I. For the analyses with  = 0, the reduction procedure described
in Section 4.2 was employed; this resulted in the elimination of around 80% of the m variables

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
LOWER BOUND LIMIT ANALYSIS USING SOCP 621

p p

Figure 4. Block with thin symmetric cuts.

L
xy = yy = 0
xx = xy = 0

L/2
xx = p, xy = 0
xy = 0

L/2

(a) xy = 0 (b)

Figure 5. Meshes for upper right quarter of block with thin symmetric cuts: (a) structured with
N = 4; and (b) unstructured with h/L = 0.075.

Table I. Results for block with thin symmetric cuts (structured meshes).
=0  = 15  = 30
p/c CPU (s) p/c CPU (s) p/c CPU (s)
N NE NVS (error %) (iter) (error %) (iter) (error %) (iter)

4 64 576 1.12703 <0.1 0.94975 <0.1 0.70976 <0.1


(0.40) (13) (1.12) (14) (2.81) (16)
8 256 2304 1.13050 0.2 0.95730 0.2 0.72036 0.2
(0.09) (15) (0.33) (20) (1.36) (20)
10 400 3600 1.13088 0.3 0.95836 0.3 0.72266 0.3
(0.06) (18) (0.22) (19) (1.04) (24)
20 1600 14 400 1.13138 1.9 0.95975 1.6 0.72608 2.6
(0.01) (19) (0.08) (22) (0.58) (30)
30 3600 32 400 1.13147 6.9 0.96008 5.1 0.72735 6.1
(0.01) (23) (0.04) (23) (0.40) (28)
50 10 000 90 000 1.13152 27.3 0.96029 19.7 0.72844 25.4
(0.00) (24) (0.02) (24) (0.25) (31)
100 40 000 360 000 1.13154 200.4 0.96042 125.4 0.72925 165.8
(0.00) (27) (0.01) (28) (0.14) (36)

NE = no. of elements, NVS = no. of stress variables.

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
622 A. MAKRODIMOPOULOS AND C. M. MARTIN

Table II. Results for block with thin symmetric cuts (unstructured meshes).
=0  = 15  = 30
h/L NE p/c error % p/c error % p/c error %

0.075 401 1.13080 0.07 0.95589 0.48 0.71259 2.42


0.025 3624 1.13101 0.05 0.95972 0.08 0.72269 1.04
0.0075 40 751 1.13131 0.02 0.96006 0.05 0.72808 0.30

NE = no. of elements.

(the exact gure depends on N ) or about 25% of the total stress variable count NVS. We also
generated three unstructured meshes using GiD [61], one of which is shown in Figure 5(b).
These meshes were designed to have element counts corresponding approximately to those of
the structured meshes with N = 10, 30 and 100.
Since no exact solution is available, when evaluating the error entries in Tables I and II we
took the correct results to be the values obtained from Richardson extrapolation applied to
the results from the nest three structured meshes (N = 30, 50, 100). This gave the following
estimates:

p/c = 1.13154 for  = 0

p/c = 0.96050 for  = 15

p/c = 0.73028 for  = 30

Note how an increase in  leads to a drop in capacity, since the structure is in tension. The
results for structured meshes (Table I) conrm that, as expected, the limit load is approached
monotonically from below as N is increased. However, the level of underestimation with
respect to the (extrapolated) correct solution becomes greater as  increases. When  = 0 it
is interesting that, in relation to the use of mixed elements [5, 18, 2224], the lower bound
method seems to give results that come much closer to the (extrapolated) correct solution.
This was also noted and discussed by Krabbenhoft and Damkilde [26]. Incidentally, for equal
values of N , our lower bound results when  = 0 are (with one minor exception in the nal
digit) identical to those obtained in Reference [26], though N = 38 was the nest discretization
considered there. The results for unstructured meshes (Table II) are only slightly inferior to
those obtained from the structured meshes with a comparable number of elements. Thus the
good performance observed in this example is not dependent on a special pattern of elements.
Finally, we see in Table I that the number of iterations required for the solution of the SOCP
problems is generally in the range 1530, which is characteristic of algorithms based on the
interior-point method.

6.2. Thick cylinder expansion


Another well known test problem [12, 26] is that of a thick cylinder of cohesive-frictional
soil subjected to uniform internal pressure. The analytical solution for the collapse load is
often attributed to Yu [62], but it was actually derived much earlier by Nadai [63, p. 465].

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
LOWER BOUND LIMIT ANALYSIS USING SOCP 623

2a
2b

Figure 6. Notation and coarse mesh (5 15) for thick cylinder.

Table III. Results for thick cylinder.


Mesh NE NVS p/c Error % CPU (s) Iter

5 15 300 2700 0.52974 1.46 0.2 12


10 30 1200 10 800 0.53547 0.39 0.8 10
20 60 4800 43 200 0.53703 0.10 2.7 8
40 120 19 200 172 800 0.53741 0.03 16.6 9
80 240 76 800 691 200 0.53748 0.02 95.4 10

NE = no. of elements, NVS = no. of stress variables.

When  > 0 and the external pressure is zero, the limit load is
  
b ( 1)/
p = c cot  1 (38)
a

where = tan2 (/4 + /2) and a and b are the inner and outer radii. For the case b/a = 1.5,
c = 1 and  = 30 the same parameters as those considered in Reference [12]the exact
solution is p = 0.53758.
After taking advantage of the horizontal and vertical symmetry, several structured meshes
were employed, the coarsest of which is shown in Figure 6. The mesh was rened uniformly,
so that the innermost quadrilateral cells remained approximately square. Each quadrilateral cell
was divided into four triangular elements by intersecting the two diagonals. The nest mesh
contained 64 times as many elements (and hence variables) as the nest meshes considered in
References [12, 26].
The results are given in Table III. Although this benchmark is not a difcult test, it does serve
to demonstrate how the use of SOCP allows us to optimize huge problems (nearly 700 000
variables for the nest mesh) and thus approach the exact solution very closely in modest
CPU times. What is most noteworthy is the small number of iterations, even in relation to the
previous example. By contrast, the iteration counts in both [12, 26] seem to increase with the
size of the mesh for this problem.

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
624 A. MAKRODIMOPOULOS AND C. M. MARTIN

5.5 5.5

p 2.5 p
2.5 50

100

Figure 7. Block with asymmetric holes (after Zouain et al. [30]).

Figure 8. Mesh 1 for block with asymmetric holes.

6.3. Block with asymmetric holes


The plane strain test problem in Figure 7 has been studied by Zouain et al. [30], who used
mixed six-node triangular elements with continuous quadratic displacements and discontinuous
stresses. Within each of their elements the deviatoric stresses varied linearly, but the mean
stress was constant. The limit loads obtained (using the von Mises criterion) were 1.052y for
an initial mesh of 716 elements, and 1.035y for an improved mesh after renement. In terms
of the Tresca criterion, these results correspond to 1.822c and 1.793c, respectively (cf. 2c for a
block with no holes). Obviously it is not possible to make a direct comparison between Zouain
et al.s discretization and ours; even with the same mesh, the number of variables would be
different (seven per element in their case, nine in ours). However a common point is that in
both cases there is a cost of three yield constraints per element.
For our analyses we employed unstructured meshes created by GiD [61], using settings hgen
and hc to control the element sizes in general and on the circumference of the two holes. The
coarsest mesh (see Figure 8 and Table IV ) consisted of 604 elements and gave a limit load
of 1.783c, which is 0.56% lower than the rened value of Zouain et al. [30]. For the second
analysis, the value of hgen was reduced substantially (Mesh 2), but the result was almost the

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
LOWER BOUND LIMIT ANALYSIS USING SOCP 625

Table IV. Results for block with asymmetric holes.


=0  = 30
Mesh hgen hc NE p/c CPU (s) iter p/c CPU (s) iter

1 12.0 0.6 604 1.7832 0.9 27 1.0450 0.6 17


2 2.0 0.6 2996 1.7840 5.6 19 1.0464 4.0 20
3 2.0 0.1 4406 1.8087 5.7 17 1.0555 5.2 21
4 1.0 0.1 12 738 1.8089 28.2 22 1.0562 21.8 23

NE = no. of elements.

same. However, when the value of hc was also reduced to give more elements in the vicinity
of the holes (Mesh 3), the limit load increased to 1.809c. Although this is a strict lower
bound solution, it is interesting to note that it is 0.89% higher than Zouain et al.s rened
value, possibly because in their analyses the mean stress was constrained to be constant in
each element. The nal analysis with Mesh 4 again conrmed that reducing hgen was not very
productive; to improve the solution further it would have been necessary to place even more
elements near the holes. Results for the same four meshes, but with material properties c = 1
and  = 30 , are also given in Table IV. These exhibit similar trends to the purely cohesive
case in terms of convergence, iteration count and CPU time.

6.4. Strip footing bearing capacity


On weightless MohrCoulomb soil (c0, 0,  = 0) in the absence of surcharge, the bearing
capacity of a rigid, symmetrically loaded strip footing of width B is given by
Q
= cNc (39)
B
where Q is the limit load (force per unit length) and Nc is a dimensionless bearing capacity
factor that depends on . Exact values of Nc can be determined using the well-known equation
of Prandtl [64]:


 
Nc = e tan  tan2 + 1 cot  (40)
4 2
Note that if  = 0 then Nc = 2+. Another fundamental case is that of weighty cohesionless soil
(c = 0,  > 0,  > 0) with no surcharge; the bearing capacity is then traditionally expressed as
Q 1
= BN (41)
B 2
where N is another dimensionless factor that depends on . At present there is no analyti-
cal solution for N , but it can be evaluated using a variety of numerical methods. To assist
with benchmarking exercises such as this, the second author has recently published a selec-
tion of high-precision N values for both smooth and rough footings [65]. These numbers (for
 = 5 , 10 , . . . , 45 ) were obtained using the method of stress characteristics, though they have
now been checked using an alternative technique (RungeKutta integration of the governing
ordinary differential equations), and also formally conrmed as exact plasticity solutions [66].

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
626 A. MAKRODIMOPOULOS AND C. M. MARTIN

B/2 9.5B

Q/2 totally 80 divisions


extension boundary
in the finest mesh

dummy node
unidirectional extension element

bidirectional extension element

unidirectional extension zone

bidirectional extension zone

Figure 9. General scheme of semi-structured meshes for strip footing.

Generally speaking, bearing capacity problems pose a difcult test for nite element methods
because of the singularities at the edges of the footing. In order to obtain good lower bound
solutions, particularly with a piecewise linear stress eld, it is desirable to have a very ne
fan of elements at the singular point (see e.g. Reference [12]). For our analyses, GiD [61] was
used to generate three semi-structured meshes based on the scheme shown in Figure 9, with
the lines of the fan continuing to the borders of the mesh. In the nest mesh, the fan was
divided into 80 elements at the singular point. As in Reference [12] we employed extension
elements (though incorporating the correction given in Appendix A) to model the semi-innite
soil domain. To maximize the compressive bearing capacity, the integral of the (tensile) vertical
stress yy over the width of the footing was minimized.
In the initial set of analyses, Meshes 13 were used to determine lower bound solutions
for the bearing capacity factors Nc ( = 0, 35 ) and N ( = 35 ). The footing was assumed
to be rough, i.e. no constraint was placed on xy on the relevant portion of the boundary.
Table V shows how the results compare with the respective exact solutions: Nc = 5.142, 46.12
and N = 34.48. Clearly the N problem is the most difcult of the three, needing the nest
mesh to obtain a lower bound that comes within 1% of the exact value.
Further analyses were devoted to the evaluation of lower bounds on N for various friction
angles, considering footings both smooth (xy = 0) and rough (xy unconstrained). The results
obtained using the nest discretization (Mesh 3) are reported in Table VI. In all but four
cases (all involving rough footings) the underestimation with respect to the exact solution is
less than 1%. Results from a similar series of analyses have recently been reported by Hjiaj

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
LOWER BOUND LIMIT ANALYSIS USING SOCP 627

Table V. Results for rough strip footing.


=0  = 35  = 35
Nc CPU (s) Nc CPU (s) N CPU (s)
Mesh NE (error %) (iter) (error %) (iter) (error %) (iter)

1 1110 5.114 1.4 45.02 1.4 29.37 1.6


(0.54) (21) (2.40) (25) (14.83) (28)
2 6111 5.135 12.8 45.68 22.5 33.18 13.5
(0.13) (22) (0.96) (48) (3.77) (30)
3 24 163 5.141 59.2 46.07 97.1 34.23 98.0
(0.01) (22) (0.12) (41) (0.71) (42)

NE = no. of elements.

Table VI. Results for bearing capacity factor N using Mesh 3.

Smooth Rough
N CPU (s) N CPU (s) Exact values
 ( ) (error %) (iter) (error %) (iter) smooth rough

5 0.08415 55.5 0.1119 56.8 0.08446 0.1134


(0.37) (24) (1.30) (22)
10 0.2800 67.5 0.4273 66.7 0.2809 0.4332
(0.31) (29) (1.36) (27)
15 0.6973 75.1 1.166 67.4 0.6991 1.181
(0.26) (32) (1.26) (29)
20 1.575 76.4 2.811 74.7 1.579 2.839
(0.26) (32) (0.98) (32)
25 3.452 91.6 6.447 76.5 3.461 6.491
(0.27) (37) (0.69) (33)
30 7.630 83.1 14.67 83.6 7.653 14.75
(0.30) (36) (0.55) (36)
35 17.51 91.9 34.23 98.0 17.58 34.48
(0.37) (39) (0.71) (42)
40 42.97 91.8 84.73 103.9 43.19 85.57
(0.51) (40) (0.98) (45)
45 116.8 96.2 231.4 115.6 117.6 234.2
(0.68) (42) (1.22) (50)

et al. [67]. If the two sets of lower bound N values are compared, ours are lower for small
, but become higher as  increases. However several of the lower bound solutions of Hjiaj et
al. lie above the exact values, and even (in two cases) above upper bound solutions obtained
by the authors [68]. These discrepancies are apparently attributable to the procedure used by
Hjiaj et al. to smooth the apex of the yield surface. Although no details were given in the
original paper [67], this procedure involved the introduction of a non-zero cohesion (Lyamin,
personal communication, 2005). As has been noted already, an advantage of using SOCP is
that no such ad hoc modication of the true yield function is required.
It is noteworthy in both Tables V and VI that as the friction angle increases, the itera-
tion counts gradually become larger than those in the previous examples. Problems involving

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
628 A. MAKRODIMOPOULOS AND C. M. MARTIN

singular
point

Figure 10. Mesh near edge of strip footing, showing principal stresses (all compressive) at
element centroids (Mesh 2, rough footing, c = 1,  = 35 ,  = 0).

Figure 11. Variation of yield function (Mesh 2, rough footing, c = 1,  = 35 ,  = 0).

bearing capacity, especially the evaluation of N for high friction angles, appear to be the most
challenging of the test cases considered in this paper. The use of quite slender elements in the
vicinity of the singular point, see Figure 10, would be expected to give entries in the constraint
matrix whose magnitudes vary considerably. However, post-analysis checks have shown that
the satisfaction of the equality constraints is equally good when compared with analyses using

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
LOWER BOUND LIMIT ANALYSIS USING SOCP 629

unstructured meshes where no elements are slender (as in Figure 5(b) for example). It is also of
interest to perform a post-analysis check on the satisfaction of the yield criterion when slender
elements are present. The contour plot of the value of the yield function (16) in Figure 11
conrms the behaviour expected from an interior-point algorithm: during the nal iterations the
solution comes very close to the boundary dened by the active yield constraints, but always
remains plastically admissible (f  0).

7. CONCLUSIONS

A formulation of lower bound limit analysis using second-order cone programming (SOCP)
has been presented. Although the approach is restricted to yield criteria that can be expressed
in a conic quadratic form, this includes a variety of the most popular yield functions such as
MohrCoulomb in plane strain, DruckerPrager (consequently von Mises) in 2D or 3D, and
the Nielsen criterion for plates. For this initial study, attention has been focused on cohesive
and cohesive-frictional problems in plane strain.
By applying an efcient algorithm based on the interior-point method, we have demonstrated
that the use of SOCP for limit analysis is highly advantageous from the aspects of both speed
and stability. The ability to solve problems of unprecedented size means that we can adopt ex-
tremely ne discretizations, and thus obtain strict lower bounds that approximate the true limit
load very closely. The simplicity of the SOCP-based formulation is another signicant advan-
tage. Once a nite element mesh has been prepared, it is quite straightforward to assemble the
optimization data (objective function, linear equality constraints, quadratic cone constraints) and
export them in a standard format; the problem can then be solved using any suitable algorithm
for SOCP optimization. For this study we have used the commercial program MOSEK, but a
number of other effective SOCP codes are available, including several in the public domain.
To obtain even better efciency, some strategies for exploiting the data structure associated
with the inter-element and boundary equilibrium constraints have been suggested. In particular,
the conditions that give rise to linearly dependent constraints have been carefully investigated,
leading to a novel method for detecting and eliminating any redundant rows of the constraint
matrix at the assembly stage. For problems involving purely cohesive (Tresca or von Mises) ma-
terial, a method for eliminating a large proportion of the free variables has also been developed,
though this does not appear to be very protable as far as overall efciency is concerned.
As a general conclusion, SOCP is a powerful (and still developing) technique for non-
linear optimization that allows us to obtain accurate solutions to difcult problems in classical
plasticity involving the respective class of yield functions. Upper bound limit analysis can also
be formulated as an SOCP problem, and this is a topic of ongoing work [68] that includes the
development and application of higher-order elements for the displacement eld.

APPENDIX A: CORRECT FORMULATION OF EXTENSION ELEMENTS

Consider the bidirectional extension element and the generic cone-shaped yield surface f () = 0
shown in Figure A1 (the horizontal and vertical axes of the stress space notionally refer to
the spherical and deviatoric components of ). The usual linear variation of the stresses within


the triangular element is continued into the extension sector bounded by edges 21 projected

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
630 A. MAKRODIMOPOULOS AND C. M. MARTIN

1 0
f 2
2 1 1
1
3 extension 2
zone F( 2) 0
2
2
3

f( ) 0

Figure A1. Bidirectional extension element and various yield constraints.



and 23 projected, i.e. the shape functions perform both interpolation and extrapolation of the
nodal stresses i . According to Lyamin and Sloan [14], the following yield constraints should
be enforced:

f (2 )  0 (A1a)

f (1 2 )  0 (A1b)

f (3 2 )  0 (A1c)

It is clear from Figure A1, however, that (A1b) and (A1c) are not sufciently restrictive. For
example, the stress point 1 lies inside the dotted surface f ( 2 ) = 0, and hence satises
(A1b), yet it lies outside the actual yield surface f () = 0. Even if the additional constraint
f (1 )0 were imposed, this would still not be sufcient; for example the stress point 1


is itself acceptable, but evidently along the projection of edge 21 there would be linearly
extrapolated stress states that violate yield.
The correct procedure for constraining the slave stress points 1 and 3 is to translate the
apex of the yield surfacenot the origin of the stress spaceto the master stress point 2 .
A convenient way of doing this is to dene a new function F that excludes any constant term
from f , i.e.
f () = F () k (A2)
where k0. For a valid extension, it can be seen from Figure A1 that the slave stress points 1
and 3 must lie on or inside the dashed surface F ( 2 ) = 0, so the correct yield constraints
are:

f (2 )  0 (A3a)

F (1 2 )  0 (A3b)

F (3 2 )  0 (A3c)

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
LOWER BOUND LIMIT ANALYSIS USING SOCP 631

This technique was originally suggested by Pastor [9]. Similar arguments can be used to formu-
late the correct yield constraints for unidirectional extension elements (referring to Figure 19(a)
in Reference [14], the condition f (1 2 ) = 0 should be F (1 2 ) = 0). Without the ap-
propriate corrections, the use of extension elements as described in Reference [14] may lead
to stress elds that are not plastically admissible. The only exception is the special case of a
cohesionless material, where the apex of the yield surface and the origin of the stress space
coincide, such that k = 0 and f () = F ().

APPENDIX B: PROOF OF CONDITIONS (28)

Assuming no degenerate elements are present, none of the terms a12 a41 , a23 a12 , etc. in
(27) is equal to 0. Hence the nal two rows of the matrix in (27) are linearly dependent if
and only if all sequential 2 2 submatrices have zero determinant, i.e.
 
 cos 2 cos 2 
 jk ij cos 2kl cos 2j k 
  = 0 j {1, . . . , N} (B1)
 sin 2j k sin 2ij sin 2kl sin 2j k 

where

 
j 1 for 1<j N j +1 for 1j <N
i = m(j ) = , k = n(j ) =
N for j = 1 1 for j = N

and l = n(k) = n(n(j ))

The sequence of stress evaluation points is shown in Figure B1. The determinant leads to the
relation
sin(2j k 2ij ) + sin(2kl 2j k ) + sin(2ij 2kl ) = 0 (B2)
which simplies to
4 sin(j k ij ) sin(kl j k ) sin(kl ij ) = 0 (B3)
Considering that  [0, 2), at least one of the arguments in (B3) must be equal to either zero
(impossible since it implies degeneracy) or . If N = 3, all three arguments refer to the angle

N
1 l
k
j n
i t
jk

Figure B1. Interior node belonging to N elements: sequence of stress evaluation points.

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
632 A. MAKRODIMOPOULOS AND C. M. MARTIN

change between consecutive edges, and if any of these angle changes is  then degeneracy
is implied. We conclude that an interior node with N = 3 can never generate a dependency. If
N 4 then the rst two arguments j k ij and kl j k refer to the angle change between
consecutive edges, whereas the nal argument kl ij refers to the angle change between
edges that are two apart. Only in the latter case is it possible to have an angle change of
 without degeneracy. So the only valid solution to (B3) is
|kl ij | =  (B4)
and thus in view of (B1) we must have
|n(j )n(n(j )) m(j )j | =  j {1, . . . , N} (B5)
It is clear that this equation can only hold for all j under special circumstances, namely
when N = 4 and the junction has been created by the intersection of two straight lines. We
have therefore established that conditions (28) are the only ones under which an interior node
generates a linearly dependent equality constraint.

ACKNOWLEDGEMENTS
The second author worked on this paper while visiting the Centre for Offshore Foundation Systems at
the University of Western Australia, established and supported under the Australian Research Councils
Research Centres Program. The nancial support arranged by Professor Mark Randolph is gratefully
acknowledged.

REFERENCES
1. Gvozdev AA. The determination of the value of the collapse load for statically indeterminate systems
undergoing plastic deformation (in Russian). In Conference on Plastic Deformations 1936, Galerkin BG
(ed.). Akademia Nauk SSSR: Moscow and Leningrad, 1938; 19 38.
2. Drucker DC, Prager W, Greenberg HJ. Extended limit design theorems for continuous media. Quarterly of
Applied Mathematics 1952; 9:381389.
3. Melan E. Theorie statisch unbestimmter Systeme aus ideal-plastischem Baustoff. Sitzungsbericht der
sterreichischen Akademie der Wissenschaften der Mathematisch-Naturwissenschaftlichen Klasse IIa 1936;
145:195 218.
4. Kamenjarzh JA. Limit Analysis of Solids and Structures. CRC Press: Boca Raton, FL, 1996.
5. Christiansen E. Limit analysis of collapse states. In Handbook of Numerical Analysis, vol. 4, Ciarlet PG,
Lions JL (eds). North-Holland: Amsterdam, 1996; 193 312.
6. Gao Y. Extended bounding theorems for nonlinear limit analysis. International Journal of Solids and Structures
1991; 27:523 531.
7. Belytschko T, Hodge PG. Plane stress limit analysis by nite elements. Journal of the Engineering Mechanics
Division (ASCE) 1970; 96:931 944.
8. Lysmer J. Limit analysis of plane problems in soil mechanics. Journal of the Soil Mechanics and Foundations
Division (ASCE) 1970; 96:13111334.
9. Pastor J. Analyse limite: dtermination numrique de solutions statiques compltes. Application au talus
vertical. Journal de Mcanique applique 1978; 2:167196.
10. Sloan SW. Lower bound limit analysis using nite elements and linear programming. International Journal
for Numerical and Analytical Methods in Geomechanics 1988; 12:6177.
11. Pastor J, Turgeman S, Boehler JP. Solution of anisotropic plasticity problems by using associated isotropic
problems. International Journal of Plasticity 1990; 6:143 168.
12. Lyamin AV, Sloan SW. Lower bound limit analysis using non-linear programming. International Journal for
Numerical Methods in Engineering 2002; 55:573 611.

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
LOWER BOUND LIMIT ANALYSIS USING SOCP 633

13. Yang H, Shen Z, Wang J. 3D lower bound bearing capacity of smooth rectangular surface footings. Mechanics
Research Communications 2003; 30:481 492.
14. Lyamin AV, Sloan SW. Mesh generation for lower bound limit analysis. Advances in Engineering Software
2003; 34:321338.
15. Maier G. Shakedown theory in perfect elastoplasticity with associated and non-associated ow laws: a nite
element linear programming approach. Meccanica 1969; 4:250 260.
16. Anderheggen E, Knpfel H. Finite element limit analysis using linear programming. International Journal
of Solids and Structures 1972; 8:1413 1431.
17. Sloan SW. A steepest edge active set algorithm for solving sparse linear programming problems. International
Journal for Numerical Methods in Engineering 1988; 26:26712685.
18. Andersen KD, Christiansen E. Limit analysis with the dual afne scaling algorithm. Journal of Computational
and Applied Mathematics 1995; 59:233 243.
19. Pastor J, Thai TH, Francescato P. Interior point optimization and limit analysis: an application. Communications
in Numerical Methods in Engineering 2003; 19:779 785.
20. Gao Y. Panpenalty nite element programming for plastic limit analysis. Computers and Structures 1988;
28:749 755.
21. Zouain N, Herskovits J, Borges LA, Feijo RA. An iterative algorithm for limit analysis with nonlinear
yield functions. International Journal of Solids and Structures 1993; 30:13971417.
22. Andersen KD, Christiansen E, Overton M. Computing limit loads by minimizing a sum of norms. SIAM
Journal on Scientic Computing 1998; 19:1046 1062.
23. Christiansen E, Andersen KD. Computation of collapse states with von Mises type yield condition.
International Journal for Numerical Methods in Engineering 1999; 46:1185 1202.
24. Andersen KD, Christiansen E, Conn A, Overton ML. An efcient primal-dual interior-point method for
minimizing a sum of Euclidean norms. SIAM Journal on Scientic Computing 2000; 22:243 262.
25. Lyamin AV, Sloan SW. Upper bound analysis using linear nite elements and non-linear programming.
International Journal for Numerical and Analytical Methods in Geomechanics 2002; 26:181216.
26. Krabbenhoft K, Damkilde L. A general non-linear optimization algorithm for lower bound limit analysis.
International Journal for Numerical Methods in Engineering 2003; 56:165 184.
27. Tin-Loi F, Ngo NS. Performance of the p-version nite element method for limit analysis. International
Journal of Mechanical Sciences 2003; 45:1149 1166.
28. Li HX, Yu HS. Kinematic limit analysis of frictional materials using nonlinear programming. International
Journal of Solids and Structures 2005; 42:4058 4076.
29. Borges LA, Zouain N, Huespe AE. A nonlinear optimization procedure for limit analysis. European Journal
of Mechanics A/Solids 1996; 15:487512.
30. Zouain N, Borges LA, Silveira JL. An algorithm for shakedown analysis with nonlinear yield functions.
Computer Methods in Applied Mechanics and Engineering 2002; 191:2463 2481.
31. Pontes IDS, Borges LA, Zouain N, Lopes FR. An approach to limit analysis with cone shaped yield-surfaces.
International Journal for Numerical Methods in Engineering 1997; 40:4011 4032.
32. Makrodimopoulos A. Computational approaches to shakedown phenomena of metal structures under plane
and axisymmetric stress states (in Greek). Ph.D. Thesis, Aristotle University of Thessaloniki, Greece, 2001.
33. Makrodimopoulos A, Bisbos C. Shakedown analysis of plane stress problems via SOCP. In Numerical
Methods for Limit and Shakedown Analysis, Staat M, Heitzer M (eds). John von Neumann Institute for
Computing (NIC): Jlich, 2003; 185 216.
34. Ciria H, Peraire J. Computation of upper and lower bounds in limit analysis using second-order cone
programming and mesh adaptivity. 9th ASCE Specialty Conference on Probabilistic Mechanics and Structural
Reliability, Albuquerque, 2004.
35. Bisbos C, Makrodimopoulos A, Pardalos PM. Second-order cone programming approaches to static shakedown
analysis in steel plasticity. Optimization Methods and Software 2005; 20:25 52.
36. Makrodimopoulos A. Computational formulation of shakedown analysis as a conic quadratic optimization
problem. Mechanics Research Communications 2005, in press.
37. Trillat M, Pastor J, Francescato P. Yield criterion for porous media with spherical voids. Mechanics Research
Communications 2005, in press.
38. Andersen ED, Roos C, Terlaky T. On implementing a primal-dual interior-point method for conic quadratic
optimization. Mathematical Programming, Series B 2003; 95:249 277.
39. Tsuchiya T. A polynomial primal-dual path-following algorithm for second-order cone programming. Research
Memorandum No. 649, The Institute of Statistical Mathematics, Tokyo, 1997.

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634
634 A. MAKRODIMOPOULOS AND C. M. MARTIN

40. Andersen ED, Roos C, Terlaky T. Notes on duality in second order and p-order cone optimization.
Optimization 2002; 51:627 643.
41. Nesterov YE, Todd MJ. Primal-dual interior-point methods for self-scaled cones. SIAM Journal on Optimization
1998; 8:324 364.
42. Monteiro RDC, Tsuchiya T. Polynomial convergence of primal-dual algorithms for the second-order cone
program based on the MZ-family of directions. Mathematical Programming 2000; 88(1):61 83.
43. MOSEK ApS. The MOSEK optimization tools version 3.2 (Revision 8). Users Manual and Reference.
Available from http://www.mosek.com [January 2005].
44. Mittelmann HD. An independent benchmarking of SDP and SOCP solvers. Mathematical Programming 2003;
95:407 430.
45. Sturm JF. Using SeDuMi 1.02, a MATLAB toolbox for optimization over symmetric cones. Optimization
Methods and Software 1999; 1112:625 653.
46. Ttnc RH, Toh KC, Todd MJ. Solving semidenite-quadratic-linear programs using SDPT3. Mathematical
Programming 2003; 95:189 217.
47. Lobo MS, Vandenberghe L, Boyd S, Lebret H. Applications of second-order cone programming. Linear
Algebra and its Applications 1998; 284:193 228.
48. Kuo YJ, Mittelmann HD. Interior point methods for second-order cone programming and OR applications.
Computational Optimization and Applications 2004; 28:255 285.
49. Sasakawa T, Tsuchiya T. Optimal magnetic shield design with second-order cone programming. SIAM Journal
on Scientic Computing 2003; 24:1930 1950.
50. Andersen ED, Andersen KD. The MOSEK interior point optimizer for linear programming: an implementation
of the homogeneous algorithm. In High Performance Optimization, Frenk H, Roos C, Terlaky T, Zhang S
(eds). Kluwer Academic Publishers: Dordrecht, 1999; 197232.
51. Andersen KD. A modied Schur-complement method for handling dense columns in interior-point methods
for linear programming. ACM Transactions on Mathematical Software 1996; 22:348 356.
52. Andersen ED. Handling free variables in primal-dual interior-point methods using a quadratic cone. Technical
Report, 2002. Available from http://www.mosek.com [January 2005].
53. Chen WF. Limit Analysis and Soil Plasticity. Elsevier: Amsterdam, 1975.
54. Nocedal J, Wright SJ. Numerical Optimization. Springer: New York, 1999.
55. Mszros C. On free variables in interior point methods. Optimization Methods and Software 1998; 9:121139.
56. Andersen ED, Andersen KD. Presolving in linear programming. Mathematical Programming 1995; 71:221245.
57. Yu HS, Salgado R, Sloan SW, Kim JM. Limit analysis versus limit equilibrium for slope stability. Journal
of Geotechnical and Geoenvironmental Engineering (ASCE) 1998; 124:111.
58. Merield RS, Sloan SW, Yu HS. Rigorous solutions for the bearing capacity of two-layered clay soils.
Gotechnique 1999; 49:471 490.
59. Drucker DC, Prager W. Soil mechanics and plastic analysis or limit design. Quarterly of Applied Mathematics
1952; 10:157165.
60. Nielsen MP. Limit Analysis and Concrete Plasticity. Prentice-Hall: Englewood Cliffs, NJ, 1984.
61. International Center for Numerical Methods in Engineering (CIMNE). GiD version 7.0. Reference Manual.
Available from http://gid.cimne.upc.es [January 2005].
62. Yu HS. Expansion of a thick cylinder of soils. Computers and Geotechnics 1992; 14:21 41.
63. Nadai A. Theory of Flow and Fracture of Solids. McGraw-Hill: New York, 1963.
64. Prandtl L. ber die Eindringungsfestigkeit (Hrte) plastischer Baustoffe und die Festigkeit von Schneiden.
Zeitschrift fr angewandte Mathematik und Mechanik 1921; 1:15 20.
65. Martin CM. Discussion of Calculations of bearing capacity factor N using numerical limit analyses by
Ukritchon et al. Journal of Geotechnical and Geoenvironmental Engineering (ASCE) 2004; 130:1106 1107.
66. Martin CM. Exact bearing capacity calculations using the method of characteristics. Proceedings of the 11th
International Conference of IACMAG, Turin, vol. 4, 2005; 441 450.
67. Hjiaj M, Lyamin AV, Sloan SW. Numerical limit analysis solutions for the bearing capacity factor N .
International Journal of Solids and Structures 2005; 42:16811704.
68. Makrodimopoulos A, Martin CM. Upper bound limit analysis using simplex strain elements and second-order
cone programming. Technical Report No. OUEL 2288/05, University of Oxford, 2005.

Copyright 2005 John Wiley & Sons, Ltd. Int. J. Numer. Meth. Engng 2006; 66:604634

S-ar putea să vă placă și