Structural Safety 22 (2000) 233±249

[Link]/locate/strusafe

Directional methods for structural reliability analysis

Jinsuo Nie, Bruce R. Ellingwood *
Department of Civil Engineering, The Johns Hopkins University, Baltimore, MD 21218, USA

Abstract
Directional simulation reduces the dimension of the limit state probability integral by identifying a set of
directions for integration, integrating either in closed-form or by approximation in those directions, and
estimating the probability as a weighted average of the directional integrals. Most existing methods identify
these directions by a set of points distributed on the unit hypersphere. The accuracy of the directional
simulation depends on how the points are identi®ed. When the limit state is highly nonlinear, or the
inherent failure probability is small, a very large number of points may be required, and the method can
become inecient. This paper introduces several new approaches for identifying directions for evaluating
the probability integral Ð Spherical t-design, Spiral Points, and Fekete Points Ð and compares the failure
probabilities with those determined in a number of examples in previously published work. Once these
points have been identi®ed for a probability integral of given dimension, they can be used repeatedly for
other probability integrals of the same dimension in a fashion analogous to Gauss Quadrature. # 2000
Elsevier Science Ltd. All rights reserved.
Keywords: Directional sampling; Engineering mechanics; Limit states; Monte Carlo simulation; Probability; Relia-
bility; Statistics; Structural engineering

1. Introduction

Practical structural reliability analyses often require the evaluation of the failure probability for
limit states involving a vector, X, of from 5 to 20 random variables described by a joint prob-
ability density function fX x†. Given the limit state function G x† ˆ 0, de®ned such that safe
domain s ˆ fxjG x† > 0g and failure domain f ˆ fxjG x† < 0g, the failure probability is given by

Pf ˆ fX x†dx 1†
f

In general, this integral is hard to evaluate, particularly in high-dimensional space. A common

approach is ®rst to transform the random vector X ˆ ‰X1 ; X2 ; . . . ; Xd ŠT to an independent

* Corresponding author.

0167-4730/00/$ - see front matter # 2000 Elsevier Science Ltd. All rights reserved.
PII: S0167-4730(00)00014-X
234 J. Nie, B.R. Ellingwood / Structural Safety 22 (2000) 233±249

standard normal random vector U ˆ ‰U1 ; U2 ; . . . ; Ud ŠT by the Rosenblatt transformation

U ˆ T X†, and then apply ®rst-order or second-order reliability methods (FORM, SORM),
Monte Carlo simulation, or other methods. With the Rosenblatt transformation, the limit state
function in the space of independent standard normal variables may become a highly nonlinear
function. For highly nonlinear structural component and system limit states, FORM and SORM
may not be suciently accurate [1].
Directional simulation and importance directional simulation have been studied by Ditlevsen,
et al. [2], and Melchers et al. [3,4], among others. The directional simulation method involves
generating uniformly distributed direction vectors and performing a one-dimensional integration
along each direction. The importance directional simulation method uses the importance sam-
pling technique to concentrate the direction vectors in the regions of interest. Although direc-
tional simulation methods are relatively ecient compared to other Monte Carlo simulation
approaches, these methods may diminish in accuracy when the limit state G u† is highly nonlinear
unless the number of sampling directions is large. Moreover, for a system reliability analysis that
is supported by ®nite element modeling, the number of directions required to limit the error in Pf
must be held to a minimum for the analysis to be performed eciently.
Developing an idea proposed by Kendall [5] and DeaÂk [6], Katsuki and Frangopol [7]
approximated the actual limit state surface in the independent standard normal space by a series
of hyperspherical segments, each having a radius described by a 2 distribution. The failure
probability was then approximated by the sum of the failure probabilities associated with those
segments approximating the limit state hypersurface. This Hyperspace Division Method (HDM)
can achieve fair accuracy in approximating the failure probability for both structural components
and systems, the latter of which may have highly nonlinear limit states with multiple local
extrema. Improvements to enhance the convergence and eciency of the HDM for realistic sys-
tems were discussed in a more recent paper by the same authors [8].
The key idea in directional simulation and the HDM is the same: (a) to seek a set of points (as
directions) uniformly distributed on the unit hypersphere either by simulation or by construction
(in which all points are obtained from some speci®ed formula), and (b) to perform the reliability
analysis as a sequence of one-dimensional integrations in the directions thus identi®ed. This idea
is developed further in this paper, in which a directional method is proposed that yields accurate
failure probabilities of components and systems described by limit states involving nonlinearities
and local extrema. Once such points (directions) are determined, they can be used repeatedly for
directional integration in an analogy with Gaussian Quadrature, rather than having to be regen-
erated for each reliability analysis. Numerical examples illustrate the eciency of the method in
comparison with other methods.

2. Fundamental procedure

All directional methods require identi®cation of directions along which the integration is per-
formed in closed form, by simulation or numerical means. In the independent standard normal
space, the integral along each direction is obtained exactly by utilizing the 2 distribution. To
summarize, given a hyperspace with d independent standard normal variables U ˆ
‰U1 ; U2 ; . . . ; Ud ŠT , the new random variable Z2 , de®ned by
 J. Nie, B.R. Ellingwood / Structural Safety 22 (2000) 233±249 235

Z2 ˆ U21 ‡ U22 ‡


‡ U2d 2†

is a chi-square random variable with d degrees of freedom. If the limit state function is a hyper-
sphere of radius R in the hyperspace , then

Gd u† ˆ ÿu21 ÿ u22 ÿ


ÿ u2d ‡ R2 ˆ ÿz2 ‡ R2 ˆ 0 3†

and the failure probability associated with this hypersphere can be obtained exactly [9]:

Pf ˆ P‰Gd u†40Š ˆ 1 ÿ 2d R2 † 4†

Fig. 1 shows a two-dimensional illustration of with the limit state simpli®ed to a circular
function G2 u† ˆ ÿu21 ÿ u22 ‡ R2 ˆ 0: Subdomains fi and si are radially split from f and s
respectively. Let the arc length on the limit state associated with fi be Afi , and the total length of
the circle (area in higher dimensions) be A, where A is the surface area of a hypersphere in d-
dimension space. A is given by
8
>
> d=2 dÿ1
>
< d=2†! r
d if d is even
Aˆ 5†
>
> 2 d dÿ1†=2
 d ÿ 1†=2†!
>
: d rdÿ1 if d is odd
d!

Then the failure probability associated with fi is

Pfi ˆ P‰U  fi Š ˆ ‰1 ÿ 2d R2 †ŠAfi =A 6†

Although the weight Afi =A allows one to divide the hypersphere unevenly (it leads to an
adaptive division scheme), it is dicult to ®nd Afi when d54. If the hypersphere is radially
divided evenly, the contribution of each subdomain, Pfi , becomes simply [7]

Fig. 1. Failure domain and segments for limit states u21 ‡ u22 ˆ R2 (after [7]).
236 J. Nie, B.R. Ellingwood / Structural Safety 22 (2000) 233±249

Pfi ˆ ‰1 ÿ 2d R2 †Š=m 7a†

where m is the total number of subdomains. Note that m may have various interpretations in
di€erent point-generating methods, such as the number of sampling directions in directional
simulation, the number of subdomains in the HDM, and the number of points in the Spherical t-
design, Spiral points and Fekete points methods to be described subsequently. In the more usual
case where the limit state is a hypersurface rather than a hypersphere (Fig. 2), the hypersurface is
approximated by a series of hyperspherical segments, each with its central point Qi lying on the
actual limit state. The failure probability of any subdomain fi is approximated by Eq. (7b),

Pfi ˆ ‰1 ÿ 2d R2i †Š=m 7b†

where Ri ˆradius of the hypersphere segment of subdomain i, Ri 50.

The total failure probability Pf is the summation over all subdomains, i.e.
X
m X
m
Pf  Pfi ˆ ‰1 ÿ 2d R2i †Š=m 8†
iˆ1 iˆ1

The Ri 's can be obtained through numerical methods (e.g. [10]). This approach eliminates the
limitations of FORM and SORM that occur when nonlinearities in Gd u† or multiple local
extremum points of the probability density function exist on the limit state hypersurface, because
it approximates the failure surface over a wider domain of x than do FORM/SORM. System

 
Fig. 2. Limit state G u ˆ 0 and its approximation with spherical segments (after [7]).

 J. Nie, B.R. Ellingwood / Structural Safety 22 (2000) 233±249 237

reliability problems may be solved in the same way as component problems except that one needs
a special way to calculate the Ri 's. Given k limit state functions describing component (or modal)
failures,
(
min Rik † for series systems
k
Ri ˆ 9†
max Rik † for parallel systems
k

For systems that are not modeled as either series or parallel systems, one might utilize the
response surface determined from a ®nite element analysis as the limit state surface.
Eq. (8) is tantamount to an equally weighted average of probabilities evaluated in m directions.
Its accuracy is contingent on getting evenly distributed points on a unit hypersphere. In 2-
dimensional space, this is trivial; in dimensions of 3 and higher, it is increasingly dicult. Our
interest lies mainly in d53.

3. Implementation of directional methods

3.1. Directional simulation

DeaÂk [6] was among the ®rst who studied the directional simulation method as a tool for
evaluating multidimensional normal probability integrals. It is a very ecient method of Monte
Carlo simulation, provided that radius vectors to the limit state surface in any direction can be
obtained eciently [2]. The failure probability formulation is similar as in Eq. (8) except that m
now is the sample size. Directional simulation mixes simulation in (d ÿ 1) dimensions with
numerical integration in one dimension. To get acceptable (with respect to evenness) directional
samples in this method, a reasonable large number of points on the unit hypersphere (direc-
tions) are needed. This entails extra cost for computing radii, which becomes critical when the limit
state function is highly nonlinear or the reliability analysis involves a large number of variables.
In directional simulation, a set of N points P ˆ fP1 ; P2 ; . . . ; PN g uniformly distributed on the
unit hypersphere de®ne the directions. Two approaches to generating these points are common.
The ®rst is to generate N vectors of the form ‰u1 ; u2 ; . . . ; ud ŠT , where d is the dimension of the
space, and ui 's are realizations of a vector of independent standard normal random variables,
each of which has been normalized to unit length. The second approach is to generate N vectors
of the form x ˆ ‰x1 ; x2 ; . . . ; xd ŠT by the rejection method, where xi 's are independent samples
from a one-dimensional uniform distribution. A vector is retained if jxj41:0; otherwise, another
is generated. Finally, all vectors are normalized. The directional simulation examples presented
subsequently for comparisons to other methods studied were prepared using the rejection method.

3.2. Other directional methods

In recent years, a number of alternate methods to generate ``evenly distributed'' points (direc-
tions) on the unit hypersphere have been developed in other ®elds. Some of these may prove to be
useful in structural reliability applications, and are described below.
238 J. Nie, B.R. Ellingwood / Structural Safety 22 (2000) 233±249

3.2.1. Spherical t-design

A spherical t-design is de®ned as follows, using the notation of Hardin and Sloane [11],

De®nition.
A set of N points P ˆ fP1 ; P2 ; . . . ; PN g on the unit hypersphere

d ˆ Sdÿ1 ˆ x ˆ ‰x1 ; x2 ; . . . ; xd Š 2 Rd : x
x ˆ 1 10†

forms a spherical t-design if the identity

1X N
f x†d x† ˆ f Pi † 11†
d
N iˆ1

(where  is a Lebesgue measure on d normalized to have total measure 1) holds for all poly-
nomials f of degree 4t.
In other words, the integral of a polynomial function over the hypersphere d can be
approximated by its average value at the points P. If P forms a spherical t-design, Eq. (11) is
exact for any polynomials of degree 4t. Delsarte et al. [12] have shown that, given d and t, the
smallest number of points, M4N, of a spherical t-design is obtained from,
!
tÿ1
Mˆ2 ‡dÿ1 if t is odd 12a†
2
dÿ1
! !
t t
‡dÿ1 ‡dÿ2
Mˆ 2 ‡ 2 if t is even 12b†
dÿ1 dÿ1

in which † is the binomial coecient. A spherical t-design with M points is said to be tight. Very
few tight t-designs exist; however, Eq. (12) serves as a benchmark later in this paper against which
to compare the number of points required by other methods. Table 1 presents the number of
points, M, vs d and t. Hardin and Sloane [11] have developed a series of spherical t-designs up to
t ˆ 21 in three-dimensional space using a pattern search algorithm, but so far have been unable
to provide t-designs in hyperspace of degree 54. Some examples in three-dimensional space will
be presented later in this paper. As will be seen in these examples, the 240-point spherical 21-
design leads to a very accurate and ecient estimate of Pf .

3.2.2. Construction methods

A construction method is one in which all the ``evenly'' distributed points (directions) are
described explicitly by some formula. These methods, if available, are the most ecient way to
get the points. Two methods are introduced below.

. Spiral points method

Rakhmanov et al. [13] introduced the spiral point procedure to construct a large number,
 J. Nie, B.R. Ellingwood / Structural Safety 22 (2000) 233±249 239

N, of points on a sphere S2. This method can only be used in 3-dimensional problems. For
spherical coordinates ; †, 044, 0442,

2 k ÿ 1†
k ˆ arcos hk †; hk ˆ ÿ1 ‡ ; 14k4N
Nÿ1
0 1
B 3:6 1 C
k ˆ @kÿ1 ‡ p qA mod 2†; 24k4N ÿ 1; 1 ˆ N ˆ 0 13†
N 1 ÿ h2
k

where the parameter 3.6 is based on Habicht and van der Waerden's [14] best packing suggestion
and numerical experiments [13].
. Hyperspace division method (HDM)
The HDM procedure [7] will be explained for a 3-d space for simplicity. It can be extended
to hyperspaces.

The unit sphere x2 ‡ y2 ‡ z2 ˆ 1 can be expressed in polar coordinates, by a system of equa-

tions with parameters  and ';
8
< x ˆ cos ' cos 
y ˆ sin ' cos  14†
:
z ˆ sin 

Table 1
The number of points M in a tight t-design vs dimension d [Eq. (12)]
Degree t
Dimension d t=8 t=21 t=32
2 9 22 33
3 25 132 289
4 55 572 1785
5 105 2002 8721
6 182 6006 35,853
7 294 16,016 128,877
8 450 38,896 415,701
9 660 87,516 1,225,785
10 935 184,756 3,350,479
11 1287 369,512 8,580,495
12 1729 705,432 20,764,055
13 2275 1,293,292 47,805,615
14 2940 2,288,132 105,306,075
15 3740 3,922,512 222,981,435
16 4692 6,537,520 455,657,715
17 5814 10,623,470 300,329,863
18 7125 16,872,570 566,092,360
19 8645 26,246,220 1036,974,872
20 10,395 40,060,020 5,915,896,470
240 J. Nie, B.R. Ellingwood / Structural Safety 22 (2000) 233±249

where 04'42, ÿ =244=2. To determine the points, ®rst ‰ÿ=2; =2Š is equally divided
into m ÿ 1 intervals, in which m is speci®ed in advance. Each of these intervals is = m ÿ 1†.
This gives a series of latitude circles, each of which is de®ned by a function

x2 ‡ y2 ˆ 1 ÿ z2 ˆ cos2 i 15†

where i is given by

i ˆ ÿ=2 ‡  i ÿ 1†= m ÿ 1†; i ˆ 1; 2; . . .; m 16†

Second, the circle i is divided equally into m'i arcs, in which m'i is given by an integer function
INT( ), i.e.

m'i ˆ INT 2 m ÿ 1†cosi † 17†

Consequently, the m'i points are almost evenly distributed on the latitude circle i. The angles of
these points are

'ij ˆ 2=m'i  j; j ˆ 1; 2; . . .; m'i 18†

and the total number of points are

X
m
mˆ m'i 19†
iˆ1

Finally, the rectangular Cartesian coordinates of each point are calculated by Eq. (14). Gen-
erating the points is slow because of the large number of cosine operations. Recent work [8], has
enhanced the eciency of the HDM.

3.2.3. Pattern search for Fekete Points

A numerical method can be derived to generate points evenly distributed on the unit sphere
from minimizing the potential energy (PE) in a set of points with forces of mutual repulsion. Such
points are denoted as Fekete Points in 3-d space [15]; they are generated conceptually for higher
dimensional spaces in this paper.

De®nition.
Fekete Points P ˆ fP1 ; P2 ; . . . ; PN g on the unit sphere are points that minimize
X ÿ1
E 1; P† ˆ Pj ÿ Pk 20†
14j<k4N

in which Pi ˆ ‰xi1 ; xi2 ; xi3 ŠT [15]. Eq. (20) describes physically the potential energy of N particles
on the unit sphere with unit charges that repel each other according
P to Coulomb's law. Although
ÿs
there are more general forms for the energy, e.g.E s; P† ˆ 14j<k4N Pj ÿ Pk , experience has
 J. Nie, B.R. Ellingwood / Structural Safety 22 (2000) 233±249 241

shown that their use slows down the convergence to minimum PE or may lead to convergence to
a local minimum. Therefore, this approach will be based on Coulomb's law. It is assumed to
remain valid when Eq. (20) is generalized to dimensions higher than 3.
We begin by generating N points on the unit hypersphere by any appropriate method, as
described previously, assigning a unit charge to each point, and calculating the initial potential
energy of the system. Second, using a generalized point repulsion method (e.g. Leech's algorithm,
19961), in which all the points are considered to repel each other according to a 1=r2 force law
(Coulomb's law), the forces between the points are computed. The tangent component Fti of the
total force Fi acting on each point i is used to determine a pattern, which de®nes the moving
direction and relative magnitude of point i; the maximum tangent force Ftmax among these tan-
gent components of the forces acting on all points is recorded. Third, set an initial maximum
moving step u ˆ A=N†1= dÿ1† =2, by which the point with the largest tangent force will move, and
Fti
search for a lower PE by letting each point i move a step of u. Test the PE to see if it is
jFtmax j
reduced by this step. If not, reduce u by half, and repeat this step untill a lower PE is obtained.
Finally, when the di€erence of PEnÿ1 and PEn between two steps (n ÿ 1) and n falls below some
speci®ed tolerance (say, 10ÿ8PEn), then the procedure stops and yields an approximate set of
Fekete points. Although generating Fekete points (44d420) is time-consuming, once they are
generated and stored in the computer, they can be used repeatedly for reliability analyses.
Examples in spaces with dimensions from 3 to 7 are presented in the sequel.

4. Numerical examples

4.1. Linear limit state functions

The following linear limit state functions with three to ®ve variables:

X
d p
gd x† ˆ ÿ xi ‡ 3 d ; 34 d45† 21†
iˆ1

where x ˆ ‰x1 ; x2 ; . . . ; xd ŠT , is a vector of independent Gaussian random variables, were evaluated

previously by Katsuki and Frangopol [7], and are analyzed in order to compare the proposed
methods with previous results. It is clear that the FORM will ®nd the exact solutions since the
limit state functions are linear. Indeed, the probability of failure is Pf ˆ 1:34997  10ÿ3 for each
of these limit state functions.
Table 2(a) shows that for the three-variable linear limit state, the use of t-design points or
Fekete points yields the most ecient (with respect to the cost to compute radii and Pfi ) and
accurate solutions. The Spiral Points Method is more accurate than the HDM for the same
number of points, from which one can infer that the Spiral Points Method distributes points more

1
Source code is available at [Link] The code deals with three-dimen-
sional space only.
242 J. Nie, B.R. Ellingwood / Structural Safety 22 (2000) 233±249

Table 2
Failure probabilities for linear limit states [Eq. (21)] computed by di€erent methods
Method to generate Number of Computed probability Error with respect to
the points points N of failure (10ÿ3) the exact Pf (%)

a. g3 X† for three variables

HDM 320 1.34275 ÿ0.535
1292 1.35154 0.116
Spiral Points 320 1.34560 ÿ0.324
1292 1.34886 ÿ0.101
Spherical t-design t=8 36 1.36486 +1.103
t=10 60 1.34603 ÿ0.292
t=11 72 1.35208 +0.156
t=13 96 1.34643 ÿ0.262
t=14 108 1.34757 ÿ0.178
t=16 144 1.34940 ÿ0.042
t=21 240 1.34999 +0.0015
Fekete Points 36 1.35795 +0.591
60 1.34698 ÿ0.221
72 1.35476 +0.355
96 1.35030 +0.024
108 1.34905 ÿ0.068
144 1.35034 +0.027
240 1.34951 ÿ0.034
300 1.34979 ÿ0.013
b. g4 X† for four variables
HDM 2170 1.35269 +0.201
1,303,124 1.35180 +0.136
Fekete Points 100 1.34753 ÿ0.181
240 1.35165 +0.124
300 1.35035 +0.028
400 1.35000 +0.002
640 1.35000 +0.002
c. g5 X† for ®ve variables
HDM 604 1.26767 ÿ6.096
229,794 1.36102 +0.819
Fekete Points 320 1.34709 ÿ0.213
480 1.35359 +0.268
604 1.34801 ÿ0.145
640 1.35064 +0.050
800 1.35097 +0.074
960 1.35368 +0.275
2080 1.34866 ÿ0.097

evenly than does HDM. The same conclusions can be drawn from the results presented for the 4-
and 5-variables problems in Tables 2(b) and (c); note that the Spherical t-design and Spiral Points
methods are available only in three-dimensional space. For 6- and 7-variables problems, the e-
ciency and accuracy of the Fekete Points method is shown in Table 3. Note that the maximum
 J. Nie, B.R. Ellingwood / Structural Safety 22 (2000) 233±249 243

Table 3
Failure probabilities for G6 X† and G7 X† computed by Fekete points method

Dimensions of Number of Computed probability Error with respect

the space points N of failure ( 10ÿ3) to the exact Pf (%)

d=6 640 1.36259 +0.935

1280 1.35345 +0.258
5120 1.35191 +0.144
6080 1.35016 +0.014
d=7 1280 1.34392 ÿ0.448
2560 1.33627 ÿ1.015
5120 1.35137 +0.104
20,000 1.34887 ÿ0.081

number of points used when d ˆ 6 or 7 (6080 or 20,000) is only slightly larger than the minimum
number required by Eq. (12) for a tight 21-design (6006 or 16016 in Table 1). This number is the
minimum necessary for the approximation of Eq. (1) by Eq. (8) to be exact when the integrand in
Eq. (1) is described by a polynomial of degree 21 or less.

4.2. Series system of linear limit state functions

Consider a series system in which the failure region is bounded by the following two limit state
functions:
p
gs1 ˆ ÿx1 ÿ x2 ÿ x3 ‡ 3 3 22a†

gs2 ˆ ÿx3 ‡ 3:0 22b†

S
The failure region is speci®ed by gs1 < 0† gs2 < 0†, as shown in Fig. 3. The second-order
bounds [16] on the failure probability of this series system are 2.5373410ÿ34Pf,series4
2.6186410ÿ3 [7]. The ``exact'' solution 2.5664110ÿ3 was obtained by directional simulation
using 10,000 directions. The sampling error on this estimate is approximately 7.5810ÿ5. Table 4
shows that most of the failure probabilities calculated from the di€erent methods lie within the
bounds, except the Spherical t-design when N ˆ 36 or 60. They all are consistent with the result
from the Monte Carlo simulation.

4.3. Parallel system of linear limit state functions

T
Eqs. (22) now are assumed to bound the failure region gs1 < 0† gs2 < 0† (see Fig. 3) of a
parallel system. The second-order bounds are 8.1297710ÿ54Pf41.6259510ÿ4 [6]; the failure
probability calculated by Hohenbichler's approximation [17] for multinormal integrals, reported
in Katsuki and Frangopol [7], is 1.2421110ÿ4. The ``exact'' solution, 1.5415610ÿ4, was
obtained from directional simulation using 10,000 directions; the sampling error on this estimate
244 J. Nie, B.R. Ellingwood / Structural Safety 22 (2000) 233±249

Fig. 3. Failure domains for series and parallel systems.

Table 4
Failure probabilities for series system [Eq. (22)]

Method to generate the points Number of points N Computed probability of failure (10ÿ3)

HDM 1292 2.56613

5182 2.56625
20,800 2.57087
Spiral Points 1292 2.58502
5182 2.57792
20,800 2.57620
Spherical t-design t=8 36 2.65306
t=10 60 2.61988
t=11 72 2.55116
t=13 96 2.57021
t=14 108 2.56743
t=16 144 2.56544
t=21 240 2.57398
Fekete Points 36 2.57177
60 2.61092
72 2.57039
96 2.58422
108 2.55461
144 2.58010
240 2.57530
300 2.57215
 J. Nie, B.R. Ellingwood / Structural Safety 22 (2000) 233±249 245

is 6.2010ÿ6. Using the same point sets as in the series system yields the results shown in Table 5.
It should be noted that when the Fekete points or the Spherical t-design methods are applied to
this parallel system problem (with a convex failure region, see Fig. 3) with the same point sets as
in the series system (Table 4), too few points are obtained to describe the nonlinear limit state
adequately. However, when N is increased to 1200 (still relatively small, cf. Table 5), the result is
quite accurate. The Spiral Points method appears to be highly accurate, and was found to gen-
erate the points eciently for both series and parallel systems de®ned by piecewise linear func-
tions.

4.4. Nonlinear limit state function

The limit state function

p
gconcave ˆ ÿ0:5 x21 ‡ x22 ‡ x23 ÿ 2x1 x2 ÿ 2x2 x3 ÿ 2x3 x1 † ÿ x1 ‡ x2 ‡ x3 †= 3 ‡ 3:0 23†

has a failure domain that is concave with respect to the origin. The ``exact'' solution 0.1979769
was obtained by directional simulation using 10,000 directions. The sampling error on this

Table 5
Failure probabilities for parallel system [Eq. (22)]

Method to generate the points Number of points N Computed probability of failure(10ÿ4)

HDM 1292 1.24552

5182 1.24428
20,800 1.24308
Spiral Points 1292 1.47737
5182 1.51206
20,800 1.55154
Spherical t-design t=8 36 0.82169
t=10 60 1.26687
t=11 72 1.60022
t=13 96 1.45491
t=14 108 1.60233
t=16 144 1.66062
t=21 240 1.65486
Fekete Points 36 1.55390
60 1.23457
72 1.49641
96 1.32422
108 1.47509
144 1.37152
240 1.55929
300 1.52621
1200 1.55563
246 J. Nie, B.R. Ellingwood / Structural Safety 22 (2000) 233±249

estimate is 1.5610ÿ3. Table 6(a) shows that the Spherical t-design and Fekete points methods
require fewer points than the Spiral Points method for comparable accuracy. However all three
methods yield satisfactory results for the concave failure domain. On the other hand, if the limit
state function has a failure region that is convex,

Table 6
Failure probabilities
Method to generate the points Number of points N Computed probability of failure

a. Concave limit state function [Eq. (23)]

Spiral Points 108 0.19730
240 0.19776
1292 0.19798
5182 0.19801
Spherical t-design t=8 36 0.19939
t=10 60 0.19777
t=11 72 0.20055
t=13 96 0.19800
t=14 108 0.19727
t=16 144 0.19821
t=21 240 0.19801
Fekete Points 36 0.19834
60 0.19703
72 0.19802
96 0.19801
108 0.19788
144 0.19793
240 0.19801
300 0.19792
b. For convex limit state function [Eq. (24)]
Spiral Points 108 1.97195
240 1.93528
1292 1.93867
5182 1.94007
Spherical t-design t=8 36 2.25897
t=10 60 1.86594
t=11 72 1.81391
t=13 96 1.95593
t=14 108 1.93144
t=16 144 2.04789
t=21 240 1.92947
Fekete Points 36 2.19594
60 1.94228
72 2.07587
96 1.89240
108 1.94916
144 1.90733
240 1.92510
300 1.95412
1200 1.94081
 J. Nie, B.R. Ellingwood / Structural Safety 22 (2000) 233±249 247
p
gconvex ˆ 0:5 x21 ‡ x22 ‡ x23 ÿ 2x1 x2 ÿ 2x2 x3 ÿ 2x3 x1 † ÿ x1 ‡ x2 ‡ x3 †= 3 ‡ 3:0 24†

the Spherical t-design and Fekete Points methods need more points to describe the limit state, as
in the previous parallel system example. The ``exact'' solution 1.9304310ÿ2 was obtained by
directional simulation using 10,000 directions. The sampling error on this estimate is 8.0410ÿ4.
Table 6(b) shows that the Fekete Points method (with 1200 points) yields approximately the same
accuracy as the Spiral points method with 5182 points. In both cases, the results are close to the
``exact'' solution.

4.5. Rigid-plastic frame

Ditlevsen et al. [2] examined the rigid-plastic frame structure (illustrated in Fig. 4) by the
directional importance simulation method. This structure can be analyzed as a series system of
three linear limit state functions (collapse mechanisms), which, according to the principle of vir-
tual work, are de®ned as follows:

Beam : gbeam ˆ X2 ‡ 2X3 ‡ X4 ÿ Gb 25†

Sway : gsway ˆ X1 ‡ X2 ‡ X4 ‡ X5 ÿ Fa 26†

Combined : gcombined ˆ X1 ‡ 2X3 ‡ 2X4 ‡ X5 ÿ Fa ÿ Gb 27†

The yield moments Xj ; j ˆ 1; . . . ; 5, at the hinge points in Fig. 4, are independent and identi-
cally distributed lognormal random variables, with mean  ˆ 1 and coecient of variation
ˆ 0:25. The lateral force F, vertical force G and the distances a and b are assumed constant,
with Gb ˆ 1:15 and Fa ˆ 2:40.

Fig. 4. Portal frame modeled as rigid±plastic system.

248 J. Nie, B.R. Ellingwood / Structural Safety 22 (2000) 233±249

Table 7
Failure probabilities for the rigid±plastic frame structure [Eqs. 28)±(30)]

Method to generate the points Number of points N Computed probability of failure(10ÿ5)

Fekete Points 480 5.42718

604 5.28253
640 5.38203
800 5.56352
960 5.49456
2080 5.45102

2
Let logarithmic standard deviation  ˆ ln 1 ‡ † ˆ 0:2462 and logarithmic mean l ˆ ln
ln Xj ÿ l
 ÿ 12 2 ˆ ÿ0:03031. With the transformation Uj ˆ , Uj are independent standard nor-

mal variables. After the transformation, the limit state functions become highly nonlinear forms,

Beam : gbeam ˆ eu2 ‡l ‡ 2eu3 ‡l ‡ eu4 ‡l ÿ 1:15 28†

Sway : gsway ˆ eu1 ‡l ‡ eu2 ‡l ‡ eu4 ‡l ‡ eu5 ‡l ÿ 2:40 29†

Combined : gcombined ˆ eu1 ‡l ‡ 2eu3 ‡l ‡ 2eu4 ‡l ‡ eu5 ‡l ÿ 3:55 30†

Sÿ
S
The failure domain is then de®ned by gbeam < 0† gsway < 0 gcombined < 0†. The ``exact''
solution 5.4519110ÿ5 was obtained by directional simulation using10ÿ6. The ®rst-order
bounds for the system failure probability are [3.12, 5.68]10ÿ5; the second-order bounds report-
edly were coincident at 5.2010ÿ5 [16]. Table 7 shows that the Fekete Points method yields a very
accurate solution compared to the ``exact'' solution when the number of points 5960.

5. Conclusion

Methods that approximate the limit state surface by a series of spherical segments can provide
accurate estimates of failure probabilities of components or systems. Such methods can deal with
the problems involving high nonlinearities, multiple extrema of the probability density along the
limit state function, and multiple limit states. However, their accuracy depends on the ability to
generate eciently a set of directions along which the probability increments in Eq. (8) are esti-
mated. This paper has presented some procedures to generate these points ``evenly distributed''
on the unit hypersphere. Once the points have been determined, they can be used repeatedly in
the numerical integrations in reliability analysis in a manner somewhat analogous to Gauss
Quadrature.
 J. Nie, B.R. Ellingwood / Structural Safety 22 (2000) 233±249 249

Three factors make the Fekete Points method attractive for this particular method of reliability
analysis. First, advances in computation have made the computations necessary to identify the
points possible. Second, storage of points, once identi®ed, is inexpensive. Third, many practical
structural system reliability problems require only ®ve to 20 random variables, making the e€ort
to identify the points feasible and practical.

References

[1] Bjerager P. On computation methods for structural reliability analysis. In: Frangopol DM, editor. New directions
in structural system reliability. Boulder (CO): University of Colorado, 1988. p. 52±67.
[2] Ditlevsen O, Melchers RE, Gluver H. General multi-dimensional probability integration by directional simulation.
Computers And Structures 1990;36(2):355±68.
[3] Melchers RE. Structural system reliability assessment using directional simulation. Structural Safety 1994;16:23±
37.
[4] Moarefzadeh MR, Melchers RE. Directional importance sampling for ill-proportioned spaces. Structural Safety,
in press.
[5] Kendall MG. A course in the geometry of n dimensions. New York: Hafner Publishing Company, 1961.
[6] DeaÂk I. Three digit accurate multiple normal probabilities. Numer Math 1980;35:369±80.
[7] Katsuki S, Frangopol DM. Hyperspace division method for structural reliability. J Engineering Mechanics, ASCE
1994;120(11):2405±27.
[8] Katsuki S, Frangopol DM. Advanced hyperspace division method for structural reliability. In: Shiraishi, Shino-
zuka, Wen, editors. Structural safety and reliability, 1998. p. 631±8.
[9] Ang AH-S, Tang W. Probability concepts in engineering planning and design, vol. II. New York: John Wiley &
Sons, 1984.
[10] Press WH, Teukolsky SA, Vetterling WT, Flannery BP. Numerical recipes in Fortran 90. 2nd ed., Cambridge
University Press, 1996.
[11] Hardin RH, Sloane NJA. Mclaren's improved snub cube and other new spherical designs in three dimensions.
Discrete Comput Geom 1996;15:429±41.
[12] Delsarte P, Goethals JM, Seidel JJ. Spherical codes and designs. Geom Dedic 1977;6:363±88.
[13] Rakhmanov EA, Sa€ EB, Zhou YM. Minimal discrete energy on the sphere. Math Res Lett 1994;1:647±62.
[14] Habicht W, Van Der Waerden BL. Lagerung Von Punkten Auf Der Kugel. Math Ann 1951;123:223±34.
[15] Sa€ EB, Kuijlaars ABJ. Distributing many points on a sphere. The Mathematical Intelligencer 1997;19(1):5±11.
[16] Ditlevsen O. Narrow reliability bounds for structural systems. J Struct Mech 1979;7(4):453±72.
[17] Hohenbichler M, Rackwitz R. Reliability of parallel systems under imposed strain. J Engineering Mech, ASCE
1983;109(3):896±907.