AIAA JOURNAL

Vol. 42, No. 6, June 2004

First-Order Saddlepoint Approximation for Reliability Analysis

Xiaoping Du∗
University of Missouri–Rolla, Rolla, Missouri 65409-0050
and
Agus Sudjianto†
Ford Motor Company, Dearborn, Michigan 48121-4091

In the approximation methods of reliability analysis, nonnormal random variables are transformed into equiv-
alent standard normal random variables. This transformation tends to increase the nonlinearity of a limit-state
function and, hence, results in less accurate reliability approximation. The first-order saddlepoint approxima-
tion for reliability analysis is proposed to improve the accuracy of reliability analysis. By the approximation of
a limit-state function at the most likelihood point in the original random space and employment of the accurate
saddlepoint approximation, the proposed method reduces the chance of an increase in the nonlinearity of the limit-
state function. This approach generates more accurate reliability approximation than the first-order reliability
Downloaded by RMIT UNIV BUNDOORA on June 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.3877

method without an increase in the computational effort. The effectiveness of the proposed method is demonstrated
with two examples and is compared with the first- and second-order reliability methods.

Nomenclature I. Introduction
E
F
f
=
=
=
expectation
cumulative distribution function
probability density function
N UMERICAL simulations are routinely used to capture physi-
cal phenomena in detail to predict engineering system behav-
iors and to reduce the amount of physical testing. Because the per-
g = limit-state function formance and reliability of engineering systems are directly affected
H = Hessian matrix by the uncertainties of model parameters and model structures, it is
K = cumulant generating function necessary to consider uncertainties with computational simulations
k = main curvature of the limit-state function at u∗ in the design process to ensure high reliability. Typical applications
nt = number of tractable random variables include reliability-based design1−4 and integrated design for relia-
n ∼t = number of intractable random variables bility and robustness.5−8 Because of higher reliability requirements
P = probability of an engineering system, the accuracy of the calculation of relia-
pf = probability of failure bility or the probability of failure becomes very critical. The tradi-
R = reliability tional Monte Carlo simulation9 is generally accurate if a sufficient
t = saddlepoint number of simulations are used. However, for high reliability, an ex-
U = standard normal random variable cessively large number of simulations are often needed. This high
U = vector of standard normal random variables computational demand is often prohibitive for complex engineering
u = realization of random variable U simulations such as finite element analysis and computational fluid
u∗ = most probable point or most likelihood dynamics. To overcome the shortcoming of the expensive computa-
point in u space tional cost, approximation methods have been developed10−18 such
X = random variable as the first-order reliability method (FORM) and the second-order
X = vector of random variables reliability method (SORM) to reduce the number of function eval-
x = realization of random variable X uations (simulation runs). Compared to Monte Carlo simulation,
x∗ = most probable point or most likelihood both FORM and SORM are much more efficient, especially when
point in x space the reliability is extremely high. Generally, SORM is more accurate
Y = system response than FORM, but needs more computations than FORM. In spite of
β = reliability index its usefulness, FORM is often not accurate enough in many cases.
 = cumulative distribution function of standard This arouses a tradeoff consideration between the efficiency and ac-
normal distribution curacy and leads to the need for a more accurate reliability analysis
−1 = inverse cumulative distribution function of standard method without large computational demand. To meet this need,
normal distribution we propose a new approximation method for reliability analysis:
φ = probability density function of standard first-order saddlepoint approximation (FOSPA). FOSPA is gener-
normal distribution ally more accurate than FORM, and in some cases more accurate
∇ = gradient than SORM, while it maintains the same order of magnitude of
computational effort as FORM.
In the next section, we present the theoretical and mathematical
background of this paper, including FORM, SORM, and the saddle-
Received 15 July 2003; revision received 17 December 2003; accepted for point approximation. Thereafter, we present the proposed FOSPA
publication 2 January 2004. Copyright  c 2004 by the American Institute of in detail, as well as examples to demonstrate its effectiveness. The
Aeronautics and Astronautics, Inc. All rights reserved. Copies of this paper discussion and conclusion are given in more depth at the end of this
may be made for personal or internal use, on condition that the copier pay paper.
the $10.00 per-copy fee to the Copyright Clearance Center, Inc., 222 Rose-
wood Drive, Danvers, MA 01923; include the code 0001-1452/04 $10.00 in II. Methods for Probability Evaluation
correspondence with the CCC.
∗ Assistant Professor, Department of Mechanical and Aerospace Engineer- Essentially, the evaluation of reliability or the probability of fail-
ing, 1870 Miner Circle; [email protected]. Member AIAA. ure by FORM and SORM is to estimate a probability, or the cu-
† Manager, Subsystem Engineering C, V-Engine Engineering, Powertrain mulative distribution function (CDF) of a random variable that is
Operations, 21500 Oakwood Building; [email protected]. a function, that is, a limit-state function, of other random variables
1199
 1200 DU AND SUDJIANTO

(basic variables), provided that the distributions of the later variables where u∗ is the expansion point. Equation (6) is used in FORM and
are given. Saddlepoint approximation19 was originally developed for Eq. (7) is used in SORM.
a related purpose, that is, to approximate the CDF of statistics of To reduce the loss of accuracy to a minimum degree, it is natural to
a random variable, for example, mean of random variable. In the expand the function g(U) at a point that has the highest contribution
following discussion, we will briefly review FORM, SORM, and to the probability integration. Therefore, the MPP is considered as
the saddlepoint approximation. Thereafter, we will discuss the need the expansion point. The MPP is the point on the surface of g(U) = 0
to extend the saddlepoint approximation to reliability analysis. for which the PDF of U is at its maximum. Maximizing the joint PDF
of U on the surface ofg(U) = 0 and noting that the contour f u (u)
A. SORM and FORM is a concentric hypersphere, we have the following formulation for
The reliability is defined as locating the MPP:

R = P{g(X) ≥ 0} (1) min u subject to g(u) = 0 (8)

u

The probability of failure is given by where   stands for the norm (length) of a vector.
Geometrically, the MPP is the shortest distance point from sur-
p f = 1 − R = P{g(X) < 0} (2) face g(u) = 0 to the origin in u space, and the minimum distance
β = u∗  is called reliability index. From Eqs. (5) and (6), the prob-
If the joint probability density function (PDF) of X is f x , the ability of failure is approximated by FORM as
probability of failure is evaluated with the integral
p f = P{g(X) < 0} = (−β) (9)


Downloaded by RMIT UNIV BUNDOORA on June 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.3877

p f = P{g(X) < 0} = f x (x) dx (3) From Eqs. (5) and (7), SORM11 gives the following approxima-
g(x) < 0 tion,
The limit-state function g(X) is usually a nonlinear function of X; 
n−1
1
therefore, the integration boundary is nonlinear. Because the num- p f = P{g(X) < 0} = (−β) (1 + β κi ) 2 (10)
ber of random variables in practical applications is usually high, i =1
multidimensional integration is involved. Because of these com-
plexities, there is rarely a closed-form solution to Eq. (3); it is also Generally, because the approximation of the limit-state in SORM
often difficult to evaluate the probability with numerical integra- [Eq. (7)] is better than that in FORM, the accuracy of SORM is
tion methods. When the computation cost of the limit-state function higher than that of FORM [Eq. (6)].
is relatively cheap, Monte Carlo integration is often applied to the
problem. However, when Monte Carlo simulation is not computa- B. Saddlepoint Approximation
tionally affordable, approximation methods such as FORM10 and Daniels19 introduced the saddlepoint approximation technique for
SORM11 have become the methods of choice in practical applica- the approximation distribution of statistics, for example, the mean,
tions. These approximation methods involve the following steps: by integration of its density estimate. Since Daniels’s work, espe-
1) transformation of random variables from their original random cially after 1980, research and applications in this area have vastly
space into a standard normal space, 2) optimization process to find increased.21−30 Instead of direct approximation of the probability
the most probable point (MPP): the design point with the high- integration in Eq. (2), saddlepoint approximation uses a Fourier in-
est contribution to the integral calculation in Eq. (3), 3) linear version formula (in an integral form) to approximate a PDF. Let Y
(in FORM) or quadratic approximation (in SORM) of the limit-state be a random variable distributed according to the density function
function in the standard normal space at the MPP, and 4) calculation f (y). The moment generating function of Y is defined as
of probability by the use of normal distribution tail approximation.
+∞
In the first step, the original random variables X = M(ξ ) = eξ y f (y) dy (11)
{X 1 , X 2 , . . . , X n } (in x space) are transformed into a set of ran- −∞
dom variables U = {U1 , U2 , . . . , Un } (in u space) whose elements
follow a standard normal distribution. The transformation is given and the cumulant generating function (CGF) of Y is defined as
by20
K (ξ ) = log{M(ξ )} (12)
−1
 
ui =  Fxi (xi ) (4) To restore f (y) from K (ξ ), we can apply the inverse Fourier
formula
The probability integration is then rewritten as
+∞

1
f (y) = M(iξ )e−iξ y dξ
P{g(X) < 0} = f u (u) du (5) 2π −∞
g(u) < 0

+i∞
1
Note that after the transformation, the integration in Eq. (5) in u = exp{K (ξ ) − ξ y} dξ (13)
space is identical to the integration in Eq. (3) in x space without any 2π −i∞
loss of accuracy, and the contours of the integrand f u (u) become Using exponential power series expansions to evaluate the integral
concentric hyperspheres. The motivation for the use of the trans- in Eq. (13) and the Hermite polynomials approximation, Daniels19
formation formulation in Eq. (5) instead of Eq. (3) to calculate the arrived to the so-called saddlepoint approximation to f (y) as
probability of failure will become clear in the following discussion.
To make the integration calculation in Eq. (5) easier, in addition 1
f (y) = {1/2π K (t)} 2 exp{K (t) − t y} (14)
to making the integrand more regular (concentric hypercircle con-
tours), the integral boundary g(u) is also approximated linearly with where K (t) is the second derivative of the CGF with respect to
the first-order Taylor expansion as t, where t is the saddlepoint corresponding to the solution to the
following equation:
g(U) ≈ g(u∗ ) + ∇(u∗ )(U − u∗ ) (6)
K (t) = y (15)
or with the second-order Taylor expansion as
The central idea of deriviation of Eq. (14) is to choose the integral
g(U) ≈ g(u∗ )+∇(u∗ )(U−u∗ )+ 12 (U−u∗ )T H(u∗ )(U−u∗ ) (7) path passing through the saddlepoint of the integrand, where the
 DU AND SUDJIANTO 1201

integrand is approximated. Because the saddlepoint is an extreme nonlinearity of a limit-state function because the transformation in
point, the function of integrand falls away rapidly as we move from Eq. (4) is nonlinear. For example, if a limit state is a linear function
this point. Thus, the influence of neighboring points on the integral of nonnormal random variables, after the transformation by the use
in Eq. (13) is diminished.29 Interested readers should consult Goutis of Eq. (4), it will become a nonlinear function of standard normal
and Casella29 for a good explanation of saddlepoint approximation. random variables. If the approximation to the limit-state function at
For the comprehensive methodology, one can refer to Ref. 30. the MPP in u space cannot capture the nonlinearity well, the accu-
Although the theory of saddlepoint approximation is quite com- racy of the probability approximation will become unacceptable. To
plex, its use, especially the CDF approximation version, is fairly reduce the accuracy loss to the minimum extent, we need to avoid
straightforward.21 The approximation of CDF of Y by the saddle- or reduce the chance of an increase in the nonlinearity due to the
point approximation derived by Lugannani and Rice23 is transformation of random variables. In other words, we may con-
sider approximating a limit-state function in the original x space or
FY = P{Y ≤ y} = (w) + φ(w)(1/w − 1/v) (16) avoid unnecessary transformation as much as possible.
To address the aforementioned concerns, we propose the FOSPA
or, alternatively by Barndorff-Nielsen,24 to improve the accuracy of reliability analysis while maintaining
the same efficiency as FORM. In FOSPA, the limit-state function
FY = P{Y ≤ y} = {w + (1/w) log(v/w)} (17)
is linearized in the original random space at the so-called most
where likelihood point (MLP). If all of the random variables are tractable,
then the saddlepoint approximation can be directly applied. If some
1
w = sign(t){2[t y − K (t)]} 2 (18) of the random variables do not have CGF, they are transformed into
other random variables that have CGF before the linearization. In
Downloaded by RMIT UNIV BUNDOORA on June 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.3877

1
v = t{K (t)} 2 (19) the following, we will discuss FOSPA in three cases: 1) all of the
random variables are tractable, 2) some of the random variables are
in which sign(t) = +1, −1, or 0, depending on whether t is positive, tractable, and 3) none of the random variables is tractable. Strictly
negative, or zero. speaking, by tractable we mean that a random variable has a closed-
Daniels19 discusses the existence and properties of the real roots form of CGF; otherwise, we call the random variable intractable.
to Eq. (15), on which the saddlepoint approximation depends, and At the end of this section, we will present a general procedure and
concludes that the saddlepoint approximation can be used whenever computational aspect of FOSPA implementation.
t lies with the restricted range assumed by K (t) where Eq. (15) has
a unique real root. A. Case 1: All of the Random Variables Are Tractable
From Eqs. (16) and (17), we see that the CDF of Y is approximated The limit-state function g(X) is first linearized at some point x∗ ,
by the use of standard normal distribution, as shown by the use of namely, the integral boundary of Eq. (3) is approximated by a hyper-
CDF and PDF of the standard normal distribution in Eqs. (16) and plane at x∗ . Similar to the concept of the MPP, the expansion point
(17). Wood et al.27 derived a general saddlepoint formula where the x∗ is chosen such that the joint PDF of X is at its maximum value on
normal-base distribution is replaced by a general-base distribution. the boundary of the limit state g(X) = 0; this point is called the MLP.
As indicated by many previous researchers, for example, in In other words, the MLP is the point on the boundary g(X) = 0, that
Ref. 19, the saddlepoint approximation yields extremely good accu- has the highest contribution to the probability of failure:
racy for CDF, especially for the tail area of a distribution, whereas
it requires only the process of finding one saddlepoint without any

integration. In terms of accuracy and efficiency, there is a great po- pf = f x (x) dx
tential to extend this technique to reliability analysis and eventually g(x) < 0
to probabilistic engineering design.
Because the saddlepoint approximation method involves the CGF The following model is used to identify the MLP x∗ :
and its derivatives, the major requirement for applications of the
technique is the tractability, that is, the existence of a CGF of the 
n
distribution of random variable Y . For an engineering application, max f i (xi )
Y is a system performance, that is, limit-state function, that is de- x
i =1
pendent on basic random variables X, that is, Y = g(X). The key subject to g(x) = 0 (20)
to application of the saddlepoint approximation to a general perfor-
mance Y is to find the CGF of Y provided that distributions of X are
given. In this papers, a general FOSPA method is developed with the The linear form of g(X) at x∗ is
capability to evaluate the CDF of a limit-state function accurately
for any continuous distributions of basic variables. g(X) ≈ ∇(x∗ )(X − x∗ ) (21)

III. FOSPA Reliability Method Then the CGF of g(X) is given by

The calculation error of the probability of failure of FORM comes
from the linear approximation [Eq. (6)] to the limit-state state func- 
n
tion in u space. The error of SORM comes from two sources, one is K (t) = K i (t) (22)
the quadratic approximation [Eq. (7)] to the limit-state function in u i =1
space and the other is the approximation of probability integration
for the approximated limit-state function in the quadratic form. For where K i (t) is the CGF of ∇i (x∗ )(Xi − xi∗ ).
a detailed discussion on the error of FORM and SORM, refer to The first and second derivatives of K (t) are
Ref. 31. Even though FORM gives an accurate solution to the prob-
ability integration for the approximated limit-state function (a linear 
n
function), it is generally less accurate than SORM because of the lin- K (t) = K i (t) (23)
ear approximation. The fact that SORM is generally more accurate i =1
than FORM implies that the accuracy of the limit-state function ap-
proximation is very important to ensure a highly accurate reliability 
n

estimation. K (t) = K i (t) (24)

Though the nonnormal to normal transformation makes it pos- i =1
sible and easy to calculate the probability of failure analytically
(without simulations), the transformation generally increases the respectively.
 1202 DU AND SUDJIANTO

Table 1 CGF of some distributions

Distribution PDF CGF
√
Normal f (x) = (1/ 2π σ ) exp [(x − µ)2 /2σ 2 ] K (t) = µt + 2σ t
1 2 2

Exponential f (x) = β exp (−βx) K (t) = −ln(1 − t/β)

Uniform f (x) = 1/(b − a) K (t) = ln[exp (bt) − exp (at)] − ln(b − a) − ln(t)
Gumbel f (x) = (1/σ ) exp [(x − µ)/σ ] exp{−exp [(x − µ)/σ ]} K (t) = µt + log (1 − σ t)
χ2 f (x) = [1/(n/2)2n/2 ]x n/2 − 1 exp (− 12 x) K (t) = − 12 n ln(1 − 2t)
Gamma f (x) = [β α /(α)]x α − 1 exp (−βx) K (t) = α{ln(β) − n(β − t)}

According to Eq. (15), the saddlepoint t is identified by the solu- C. Case 3: None of the Random Variables Is Tractable
tion to the equation When all random variables are intractable, they must be trans-
formed into selected tractable random variables such as standard

n
normal variables. If all of the random variables are transformed into
K (t) − y = K i (t) − y = 0 (25)
standard normal variables, after the transformation, the model of
i =1
searching the MLP becomes
Once the saddlepoint t is identified, the probability P{g(X) ≤ y}

n
can be calculated from Eq. (16) with the equations max φ(u i )
Downloaded by RMIT UNIV BUNDOORA on June 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.3877

  
u
1 i =1
1

n 2
subject to g(u) = 0 (30)
w = sign(t){2[t y − K (t)]} 2 = sign(t) 2 t y − K i (t)
i =1
which is equivalent to the model in Eq. (8) for the MPP search.
(26) Therefore, the solution u∗ to the model in Eq. (30) is exactly the
1
MPP defined in the model (8). At the MLP, the linearization of the
1

n 2 limit-state function is given by
v = t{K (t)} = t 2 K i (t) (27) 
i =1

n
∂g 
g(X) = Ui − u i∗ (31)
∂u i 
i =1 u∗
The CGFs of some common distributions are listed in Table 1.
For further details, refer to Ref. 32.
Appendix A shows that the calculated probability of failure from
B. Case 2: Some of the Random Variables Are Tractable
saddlepoint approximation based on Eq. (31) is the same result as
that of FORM. In other words, FORM is identical to FOSPA when
Some random variables may not have a closed form that is, in-
all random variables are transformed into standard normal variables.
tractable, CGF, for example, Weibull distribution and lognormal
Therefore, FORM is a special case of FOSPA.
distribution. There are two ways to approach intractable CGF: 1) ap-
proximate the CGF using polynomial expansions33 or 2) transform
the random variable into another random variable with tractable D. General Procedure and Computation Implementation of FOSPA
CGF. The latter approach is adopted in this paper for the purpose of The procedure of FOSPA is summarized as follows:
simplicity. One possible transformation is similar to the one used in 1) Determine whether a random variable has tractable or in-
FORM and SORM as shown in Eq. (4), which is the transformation tractable CGF and form two sets of random variables, one set with
from a random variable with intractable CGF to a standard normal tractable CGF, Xt , and the other set without tractable CGF, X∼t .
variable. In general, any distribution with tractable CGF can be used Transform the latter set into standard normal variables U.
∗
for the transformation. 2) Solve the model in Eq. (28) to identify the MLP {xt , u∗ }.
Let the set of variables that have tractable CGF be Xt = {X it ; i = 3) Linearize the limit-state function at the MLP as shown in
1, 2, . . . , n t } and the set of variables without tractable CGF be Eq. (29).
X∼t = {X ∼t j ; j = 1, 2, . . . , n ∼t }. After the nonnormal–normal trans-
4) Formulate the saddlepoint equation and solve it to obtain the
formation, X∼t is transformed into a set of standard normal variables saddlepoint t.
U = {U j ; j = 1, 2, . . . , n ∼t }. Then, the formulation for searching the 5) Use Eqs. (15–19) to find the probability of failure.
MLP {x∗t , u∗ } becomes Note that if all of the random variables are tractable, X∼t will
be an empty set and the problem belongs to case 1, and if none of

nt 
n ∼t the random variables is tractable, Xt will be an empty set and the
max f i xit φ(u j ) problem belongs to case 3 where the same result as that of FORM
x t ,u
i =1 j =1 will be obtained.
subject to g(xt , u) = 0 (28) To make the numerical computation process of FOSPA more sta-
ble, several practical measures may be considered, and some of them
∗
After linearization, the limit-state function at the MLP {xt , u∗ } are briefly discussed here. The variables in Eqs. (20) and (28) for
is given by MLP search are normalized by the means and standard deviations of
 the random variables. This normalization sets the design variables in

nt
∂g  ∗ the same scales. Note that this normalization is a linear transforma-
g(X) ≈ q(X , U) =t
X it − xit
∂xt  i =1 i t∗
x ,u ∗
tion and will not affect the nonlinearity of the limit-state function,
but it will help the convergence of the iterative process of finding
  the MLP. To avoid the objective functions of the MLP search in
∂g 
n ∼t
+ U j − u ∗j (29) Eqs. (20) and (28) becoming too small, one may choose to use the
j =1
∂u j  ∗
x t ,u ∗
natural logarithm of the objective functions. To avoid singularities
in Eqs. (18) and (19), one may use the reverse sign of the limit-sate
Because the limit state in Eq. (29) is a linear combination of function when a square root of a negative value occurs.
tractable random variables, the saddlepoint approximation method Given that there is a strong need to minimize the number of
in Eqs. (22–27) can be applied in conjunction with Eq. (29) to eval- limit-state function evaluations, so that the technique is practical
uate the probability of failure. for computationally expensive engineering simulation models, for
 DU AND SUDJIANTO 1203

example, finite element analysis and computational fluid dynamics, For this specific example, the theoretical solution can be found.
we compare the efficiencies of the methods by counting the number The probability of failure p f = P{g(X) < 0} is listed in Table 2 and
of function evaluations of limit-state function. Because FOSPA uses depicted in Fig. 1 for n = 2.
a similar optimization formulation to find the MLP as FORM does Figure 1 shows the probability of failure for different values of
for the MPP, and because it uses a less nonlinear constraint func- a. The probability of failure changes in the range roughly between
tion, the computational effort (measured by the number of function 0.4 and 0 as a varies. The curves of FOSPA and the exact solution
evaluations) of FOSPA is less than or at most the same as that of almost overlap each other over the whole range of the probability.
FORM. This indicates that FOSPA is consistently good over the range of
probability of failure. SORM is more accurate than FORM, but
IV. Numerical Examples when the probability of failure is high, for example, 0.4, SORM
is not accurate, as shown in Fig. 1. The accuracy of solution from
In this section, two examples are used to demonstrate the effec-
SORM increases as the probability of failure becomes lower. This
tiveness of the proposed method. The first example is associated
phenomenon conforms to the fact that SORM is only accurate at
with a linear limit-state function and the other with a nonlinear
the tail of a distribution due to its asymptotic approximation to
limit-state function. We will compare the accuracy and efficiency
the probability integration.11 In this example, with linear limit-state
among FOSPA, FORM, and SORM. If no theoretical solution exists,
function and tractable CGF random variables, the results show that
we will use the result of Monte Carlo simulation with a relatively
FOSPA is the most accurate method.
large sample size as a reference. In the following examples, the first-
Figure 2 shows that when a = 3.5, the original linear limit-state
and second-order derivatives are evaluated numerically with the fi-
function becomes highly nonlinear after the transformation to stan-
nite difference method. Because of this finite difference calculation,
dard normal distributions required by both FORM and SORM. The
SORM, which requires second-order derivative information, has an
Downloaded by RMIT UNIV BUNDOORA on June 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.3877

linear approximation of FORM is far away from the transformed

inherent inefficiency in terms of the number of function evaluations.

A. Example 1: Linear Limit-State Function Table 2 Probability pf = P{g(X) < 0} for n = 2

A linear limit-state function is given by a sum of independent a FORM SORM FOSPA Exact
random variables,11 as follows:
0.0001 0.3166 0.3612 0.4068 0.4060
√ 
n 0.5 0.1795 0.2301 0.2482 0.2474
g(X) = n + a n − Xi (32) 1.0 0.0990 0.1393 0.1459 0.1452
i =1
1.5 0.0536 0.0816 0.0835 0.0831
2.0 0.0286 0.0466 0.0469 0.0466
2.5 0.0152 0.0262 0.0260 0.0258
where a is a constant and X i are n independent random variables. 3.0 0.0079 0.0145 0.0142 0.0141
3.5 0.0041 0.0079 0.0077 0.0076
1. Case 1: All Random Variables Are Tractable 4.0 0.0021 0.0043 0.0041 0.0041
Let each of the random variables follows a standard exponential 4.5 0.0011 0.0023 0.0022 0.0022
distribution with CDF 5.0 0.0006 0.0012 0.0012 0.0012
5.5 0.0003 0.0007 0.0006 0.0006
6.0 0.0001 0.0003 0.0003 0.0003
F(xi ) = 1 − exp(−xi ) (33)

Fig. 1 Probability of failure when n = 2.

 1204 DU AND SUDJIANTO

a) b)
Fig. 2 Limit-state function in a) x and b) u spaces.
Downloaded by RMIT UNIV BUNDOORA on June 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.3877

Fig. 3 Probability of failure when n = 10.

nonlinear limit-state function in u space, and even the quadratic Monte Carlo simulation (MCS) is employed, and its result is used
approximation in SORM cannot capture the nonlinearity of the as a reference for comparison of the accuracy of other methods. The
transformed limit-state function very well. Therefore, both FORM number of simulations in the MCS is 106 . The calculated proba-
and SORM are not as accurate as FOSPA in this example. Be- bility of failure is shown in Table 5. Note that FOSPA is the most
cause FOSPA uses the original linear limit-state function without accurate method and that SORM is more accurate than FORM.
the increase of nonlinearity and saddlepoint approximation, a high- With FORM and SORM, the transformation of {X 1 , X 2 , X 3 } into
accuracy approximation results. That is, the overall accuracy of a standard normal variable {U1 , U2 , U3 } makes the original linear
FOSPA is superior to FORM and SORM. limit-state function become nonlinear in terms of {U1 , U2 , U3 }. On
The result for higher dimension with n = 10 is listed in Table 3 and the other hand, FOSPA only involves the transformation of X 3 into
shown in Fig. 3. The result still shows that the FOSPA is much more a standard normal variable U3 . That is, the original limit state is only
accurate than FORM and SORM. The related detailed equations nonlinear in terms of U3 , and the remaining terms of X 1 and X 2 are
used in this example are given in Appendix B. kept linear. As a result of the minimum increase of nonlinearity of
the limit state, FOSPA is more accurate than FORM and SORM.
2. Case 2: Some Random Variables Are Not Tractable The numbers of function evaluations used by FOSPA, FORM, and
In the following case, we choose X 3 to follow a Weibull dis- SORM (including finite difference calculation and iterations to find
tribution that does not have a closed-form CGF. The distribution MLP/MPP) are 25, 37, and 57, respectively. In this case, the min-
information is shown in Table 4. imum increase of nonlinearity also helps FOSPA to be the most
Because X 3 is not tractable, it is transformed into a standard efficient method for finding the MLP, whereas SORM is the least
normal variable before the saddlepoint approximation is applied. efficient method for this specific case.
 DU AND SUDJIANTO 1205

Table 3 Probability pf = P{g(X) < 0} for n = 10 Table 8 Distributions of random variables

a FORM SORM FOSPA Exact Variable Parameter 1 Parameter 2 Distribution

0.0001 0.1429 0.4683 0.4580 0.4579 Diameter D 39 mm 0.1 mm Normala

0.5000 0.0628 0.3392 0.2810 0.2809 Span L 400 mm 0.1 mm Normal
1.0000 0.0253 0.2131 0.1554 0.1554 External force F 1500 N 350 N Gumbelb
1.5000 0.0094 0.1195 0.0786 0.0786 Torque T 250 Nm 35 Nm Normal
2.0000 0.0033 0.0610 0.0369 0.0369 Strength S 70 MPa 80 MPa Uniformc
2.5000 0.0011 0.0288 0.0162 0.0162 a
For normal distribution, parameters 1 and 2 are mean and standard deviation,
3.0000 0.0004 0.0127 0.0067 0.0067 respectively.
3.5000 0.0001 0.0053 0.0027 0.0027 b
For Gumbel distribution, parameters 1 and 2 are mean and standard deviation,
4.0000 0.0000 0.0021 0.0010 0.0010 respectively.
c
4.5000 0.0000 0.0008 0.0004 0.0004 For a uniform distribution, parameters 1 and 2 are lower and upper bounds,
5.0000 2.85e–6 2.96e–4 1.29e–4 1.29e–4 respectively.
5.5000 8.03e–7 1.05e–4 4.43e–5 4.44e–5
6.0000 2.22e–7 3.63e–5 1.47e–5 1.47e–5 Table 9 Probability P{g(X) < 0}

Failure
probability FORM SORM FOSPA MCS
Table 4 Information of random variables
Variable Parameter 1 Parameter 2 Distribution P{g(X) < 0} 7.007×10−7 4.3581 × 10−7 6.1754 × 10−4 7.850 × 10−4
N 1472 1514 102 106
Downloaded by RMIT UNIV BUNDOORA on June 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.3877

X1 1.2 —— Exponentiala
X2 1.2 —— Exponential
X3 2 1.5 Weibullb Table 10 Probability at the tails of distribution
a
For an exponential distribution, parameter 1 is the mean. Failure
b
For a Weibull distribution, parameters 1 and 2 are parameters a and b, respec- probability FORM SORM FOSPA MCS
b
tively, in the PDF of a Weibull distribution f (x) = abx b − 1 e−ax .
P{g(X) < 4.5 × 107 } 0.96798 0.97406 0.99927 0.99938
N 212 254 55 106
Table 5 Probability of failure for case 2

Failure
Table 11 Probability near the median
probability a FORM SORM FOSPA MCS
P{g(X) < 0} 4.3 1.224 3.025 2.770 2.289 Failure
× 10−4 × 10−4 × 10−4 × 10−4 probability FORM SORM FOSPA MCS
Na —— 25 37 25 106 P{g(X) < 2.48 × 107 } 0.1538 0.1406 0.4825 0.5038
a
Number of function evaluations. N 93 135 43 106

Table 6 Information of random variables where S is the material strength, D is the diameter of the shaft, F
is the external force, T is the external torque, and L is the length of
Variable Parameter 1 Parameter 2 Distribution the shaft. The limit-state function represents the difference between
X1 2 1.5 Weibull the strength and the maximum stress.
X2 2 1.5 Weibull The variable information is given in Table 8.
X3 2 1.5 Weibull This problem belongs to case 1 where all of the random vari-
ables are tractable. The results of probability of failure as compared
with MCS (106 simulations) are shown in Table 9. Referenced to
Table 7 Probability P{g(X) < 0} for case 3 MCS, FOSPA generates the most accurate solution with the least
computational demand.
Failure The preceding result indicates that FOSPA provides an accurate
probability a FORM SORM FOSPA MCS CDF estimate at the right tail of the distribution of the limit-state
P{g(X) < 0} 0.5 0.0026 0.0038 0.0026 0.0040 function. To illustrate the accuracy of FOSPA over the whole dis-
N 21 27 21 106 tribution range, the CDF of the limit-state function at the left tail
and near the median are also calculated and given in Tables 10 and
11, respectively. From Tables 9 and 10, note that FOSPA is also
3. Case 3: All of the Random Variables Are Not Tractable superior to FORM and SORM at both tails in terms of accuracy
In the following case, all random variables follow Weibull dis- and efficiency. Table 11 shows that FOSPA also produces a reason-
tributions as shown in Table 6. Because a Weibull distribution does ably accurate CDF estimate around the median of the distribution
not have tractable CGF, the transformation from {X 1 , X 2 , X 3 } to a whereas both FORM and SORM have very large errors. This ex-
standard normal variable {U1 , U2 , U3 } is required. ample demonstrates that FOSPA is evenly accurate over the whole
As expected, FOSPA has the same result as FORM as distribution and, therefore, beneficial for the generation of a com-
shown in Table 7. The MLP from FOSPA and the MPP plete distribution of a performance (limit-state function).
from the FORM are identical, that is, {x1MLP , x2MLP , x3MLP } =
{x1MPP , x2MPP , x3MPP } = {1.2887, 1.2887, 1.2887}. In this case, SORM V. Discussion
is the most accurate method because the second-order approxima- In this section, we summarize the proposed FOSPA method with
tion in SORM provides a better approximation to the limit-state a detailed discussion on its accuracy and efficiency in comparison
function in u space. to FORM and SORM. Based on the discussion, recommendations
for the selection of the reliability analysis methods under various
B. Example 2: Nonlinear Limit-State Function circumstances will be provided in the next section.
Consider the limit-state function of a shaft in a speed reducer Saddlepoint approximation is an accurate method for the estima-
defined as tion of CDF of a random variable if its CGF is known. The central
 idea of the proposed FOSPA is to approximate the CGF of a general
g(X) = S − (32/π D 3 ) F 2 L 2 /16 + T 2 (34) limit-state function through linearization of limit-sate function. The
 1206 DU AND SUDJIANTO

linearization is conducted at the MLP, the point where the joint PDF VI. Conclusions
of the random variables is at its maximum value for a given limit- In summary, the proposed FOSPA method for reliability analysis
state value. If a random variable does not have a closed-form CGF is an attractive alternative to the existing reliability analysis methods
(intractable), it is transformed to another random variable with a FORM and SORM. One may consider the following facts when
tractable CGF before the linearization. In this paper, an intractable selecting the reliability methods: FORM is a special case of FOSPA,
random variable is transformed to a standard normal variable. Note and the latter is more accurate than the former with less or at most the
that other types of random variables with tractable CGF can also be same computational effort. If the limit-state function in the original
used for the transformation. Once the limit-state function is in the space is less nonlinear than that of standard normal transformed
form of a linear combination of tractable variables, the CGF of the space, FOSPA may be more accurate than SORM. SORM is less
limit-state function is easily obtained. The saddlepoint is the solu- efficient, that is, it requires more function evaluations, than FOSPA
tion to the equation of the first derivative of the CGF equal to the and FORM.
limit-state value. Thereafter, the saddlepoint approximation solution
is used to approximate the probability of failure or the reliability.
In contrast to FORM, which conducts linearization in the trans- Appendix A: FORM Is a Special Case of FOSPA
formed standard normal space (which imposes nonlinear transfor- If none of the random variables is tractable, FORM produces
mations), FOSPA linearizes the limit-state function in the original the same result as FOSPA when standard normal transformation is
space of tractable random variables. As a consequence of the min- employed. After a limit-state function g(X) is approximated by a
imization of random variable transformation, FOSPA reduces the linear function in Eq. (30), the CGF of q(U) is given by
chance of an increase in the nonlinearity of the limit-state function.
  2
n 
Therefore, the linearization of the limit-state function in FOSPA 
 n
∂g  ∗ 1  ∂g 
Downloaded by RMIT UNIV BUNDOORA on June 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.3877

gives a more accurate approximation than that of FORM. Generally,

K (t) = −  u t+ t2 (A1)
FOSPA is more accurate than FORM, except in the following cases
i =1
∂u i  i 2 i = 1 ∂u i u ∗
∗
where they are equivalent: 1) all random variables have intractable u

CGF and they are transformed into standard normal variables, 2) all
tractable random variables are normally distributed and all of the and its derivative is
intractable random variables are transformed into standard normal   2
 n 
∂g  ∗  ∂g 
variables, and 3) all random variables are normally distributed. In n
K (t) = − u + t (A2)
∂u i  ∗ i ∂u i  ∗
the aforementioned three cases, the MLP from FOSPA is identical
to the MPP from FORM and, therefore, both methods have the same i =1 u i =1 u
accuracy. In this sense, FORM is a special case of FOSPA.
It is generally recognized that SORM is more accurate than The saddlepoint is obtained from K (t) = 0:
FORM, although there are a few exceptions; however, there is no
such direct conclusion about the comparison between FOSPA and
   n   2
∂g  ∗  ∂g 
n
SORM in terms of their accuracy. One method is more accurate than t= u (A3)
the other depending on the problem under consideration. Generally ∂u i  ∗ i
i =1
∂u i  ∗ u i =1 u
speaking, when the limit-state function is less nonlinear in terms of
original random variables or the nonnormal to normal transforma-
tion increases nonlinearity of the limit-state function significantly, The CGF at the saddlepoint becomes
FOSPA may have a higher accuracy than SORM.   2   2 
 n 
The search of the MLP needs an iterative process where the limit- ∂g  ∗ ∂g 
n
1
state function is evaluated repeatedly. Because a search for an MLP is K (t) = − u (A4)
a task similar to a search for an MPP, it is expected that FOSPA has at 2 ∂u i  ∗ i
i =1 u
∂u i  ∗ i =1 u
most the same order of magnitude of computational demand as that
of FORM. In many cases, the search for the MLP is more efficient and its second-order derivative with respect to the saddlepoint is
than the search for the MPP because the constraint function in the given by
optimization model of the MLP is less nonlinear than that of the
MPP. Note that the search of the saddlepoint does not consume any n   2
limit-state function evaluations. Because SORM needs the second-
 ∂g 
K (t) = (A5)
order derivative of a limit-sate function, it is generally much less ∂u i  ∗ i =1 u
efficient when the derivative is evaluated numerically.
Considering the same computational effort and higher accuracy
of FOSPA compared to FORM, one may choose FOSPA for a reli- Substitution of Eqs. (A4) and (A5) into Eqs. (18) and (19) yields
    12
n 
ability analysis. When higher accuracy is needed, one should also
consider that, depending on the linearity of the limit-state and ran- 
n
∂g  ∗  ∂g 
w= u = −β (A6)
dom distribution, SORM is not always better than FOSPA in terms ∂u i  ∗ i
i =1 u
∂u i  ∗ i =1 u
of accuracy. The computational efficiency, accuracy, and implemen-
tation simplicity of the proposed method make it attractive for real-     2    2
n  n 
1
world reliability analysis. The authors have extensively applied the 
n
∂g  ∗  ∂g  ∂g 
2

v= u 2
∂u i  ∗ i ∂u i  ∗ ∂u i  ∗
proposed method to various computationally intensive simulation
models used in automotive engine design (see Hoffman et al.34 ). i =1 u i =1 u i =1 u

    12
n 
As indicated in example 2, FOSPA can also be used to generate
accurate CDF associated with a range of limit-state values. This is 
n
∂g  ∗  ∂g 
= u = −β (A7)
accomplished by enumeration of the limit-state values, performance ∂u i  ∗ i
i =1
∂u i  ∗
u i =1 u
of linearization at all MLP associated with the limit state values, and
calculation of the probability by the use of Eq. (16) or (17). By the respectively.
use of this approach, FOSPA can accurately calculate the CDF at The combination of Eqs. (A6), (A7), and Eq. (16) results in
both tails, as well as around the median (or mean) of a distribution.
To further improve accuracy, the second-order saddlepoint ap-
proximation can be considered, and the key to the new development P{g < 0} = (−β) (A8)
is how to identify the CGF of a second-order approximation of a
limit-state function. which is the same result as that of FORM.
 DU AND SUDJIANTO 1207

Appendix B: Case 1 of Example 1 5 Du, X., and Chen, W., “Towards a Better Understanding of Modeling

For FOSPA, the CGF of the limit-state function of example 1 is Feasibility Robustness in Engineering,” Journal of Mechanical Design, Vol.
122, No. 4, 2000, pp. 357–583.
given by 6 Du, X., and Chen, W., “Efficient Uncertainty Analysis Methods for Mul-
√ tidisciplinary Robust Design,” AIAA Journal, Vol. 4, No. 3, 2002, pp. 545–
K (t) = −n ln(1 − t) − n + a n t (B9) 552.
7 Du, X., and Chen, W., “Integrated Methodology for Uncertainty Propaga-
and its derivatives are tion and Management in Simulation-Based Systems Design,” AIAA Journal,
√ Vol. 38, No. 8, 2000, pp. 1471–1478.
K (t) = n/(1 − t) − n + a n (B10) 8 Du, X., Sudjianto, A., and Chen, W., “An Integrated framework for Op-
timization Under Uncertainty Using Inverse Reliability Strategy,” Proceed-
√ 2 ings of 2003 ASME International Design Engineering Technical Conferences
K (t) = n + a n n (B11) and the Computers and Information in Engineering Conference [CD-ROM],
American Society of Mechanical Engineers, Fairfield, NJ, 2003.
respectively. 9 Harbitz, A., “An Efficient Sampling Method for Probability of Failure
The solution of K (t) = 0 produces the saddlepoint Calculation,” Structural Safety, Vol. 3, No. 2, 1986, pp. 109–115.
√ 
10 Hasofer, A. M., and Lind, N. C., “Exact and Invariant Second-Moment
√
t =a n n+a n >0 (B12) Code Format,” Journal of the Engineering Mechanics Division, ASCE,
Vol. 100(EM1), 1974, pp. 111–121.
11 Breitung, K., “Asymptotic Approximations for Multinomial Integrals,”
Combining Eqs. (B9–B12), we obtain
Journal of Engineering Mechanics, Vol. 110, No. 3, 1984, pp. 357–367.
1    √  √ 1 12 Ditlevsen, O., “Generalized Second Moment Reliability Index,” Journal
w = sign(t){2(t y − K (t)} 2 = 2 n ln n n+a n +a n 2 of Structural Mechanics, Vol. 7, No. 4, 1979, pp. 435–451.
Downloaded by RMIT UNIV BUNDOORA on June 14, 2013 | http://arc.aiaa.org | DOI: 10.2514/1.3877

13 Wu, Y.-T., Millwater, H. R., and Cruse, T. A., “Advance Probabilistic

(B13)
Analysis Method for Implicit Performance Function,” AIAA Journal, Vol.
1 28, No. 9, 1990, pp. 1663–1669.
v = t{K (t)} 2 = a (B14) 14 Wu, Y. T., and Wirsching, P. H., “New Algorithm for Structural Re-
liability Estimation,” Journal of Engineering Mechanics, Vol. 113, No. 9,
and the probability of failure 1987, pp. 1319–1336.
15 Chen, X., and Lind, N. C., “Fast Probability Integration by Three-
  √ 1
p f = {w + (1/w) ln(v/w)} =  2 c + a n 2 Parameter Normal Tail Approximation,” Structural Safety, Vol. 1, No. 1,
1983, pp. 169–176.
  √ 1   √ 1 
16 Du, X., and Chen, W., “A Most Probable Point Based Method for Uncer-
+ 1 2 c+a n 2 ln a 2 c+a n 2 (B15) tainty Analysis,” Journal of Design and Manufacturing Automation, Vol. 1,
Nos. 1 and 2, 2001, pp. 47–66.
where 17 Haldar, A., and Mahadevan, S., Reliability Assessment Using Stochastic
 √  Finite Element Analysis, Wiley, New York, 2000, pp. 55–104.
c = n ln n n + a n (B16) 18 Melchers, R. E., Structural Reliability Analysis and Prediction, Wiley,
Chichester, England, U.K. 1999, pp. 32–132.
For FORM the MPP u ∗
= {u i∗ }, i
= 1, 2, . . . , n is given by 19 Daniels, H. E., “Saddlepoint Approximations in Statistics,” Annals of

  √   Mathematical Statistics, Vol. 25, 1954, pp. 631–650.

u i∗ = − −1
exp − n + a n n (B17) 20 Rosenblatt, M., “Remarks on a Multivariate Transformation,” Annals
of Mathematical Statistics, Vol. 23, 1952, pp. 470–472.
21 Huzurbazar, S., “Practical Saddlepoint Approximations,” American
and the reliability index is calculated by
Statistician, Vol. 53, No. 3, 1999, pp. 225–232.
√ 22 Marsh, P. W. N., “Saddlepoint Approximations or Noncentral Quadratic
β = nu i∗ (B18)
Forms,” Econometric Theory, Vol. 14, No. 5, 1998, pp. 539–559.
23 Lugannani, R., and Rice, S. O., “Saddlepoint Approximation for the
Then, the probability of failure is Distribution of the Sum of Independent Random Variables,” Advances in
√ Applied Probability, Vol. 12, 1980, pp. 475–490.
p f = (−β) =  − nu i∗ (B19) 24 Barndorff-Nielsen, O. E., “Inference on Full or Partial Parameters Based
on the Standardized Signed Log Likelihood Ratio,” Biometrika, Vol. 73,
For SORM, the probability of failure is given by 1986, pp. 307–322.
25 Barndorff-Nielsen, O. E., “Approximate Interval Probabilities,” Journal
  −(n − 1)/2
u i∗  −u i∗ − φ −u i∗ of the Royal Statistical Society, Series B, Vol. 52, 1990, pp. 485–496.
−1
√ 26 Gatto, R., and Ronchetti, E., “General Saddlepoint Approximations of
pf = − nu i∗ 1 + u i∗
 −u i∗ Marginal Densities and Tail Probabilities,” Journal of the American Statis-
tical Association, Vol. 91, No. 433, 1996, pp. 666–673.
27 Wood, A. T. A., Booth, J. G. B., and Butler, R. W., “Saddlepoint Ap-
(B20)
proximations to the CDF of Some Statistics with Nonnormal Limit Distribu-
Acknowledgment tions,” Journal of the American Statistical Association, Vol. 422, No. 8, 1993,
pp. 480–486.
Support for the first author from University of Missouri System 28 Kuonen, D., “Computer-Intensive Statistical Methods: Saddlepoint Ap-
Research Board, 943, is gratefully acknowledged. proximations with Applications in Bootstrap and Robust Inference,” Ph.D.
Dissertation, Dept. of Mathematics, Swiss Federal Inst. of Technology, 2001.
29 Goutis, C., and Casella, G., “Explaining the Saddlepoint Approxima-
References
1 Frangopol, D. M., and Klisinski, M., “Reliability-Based Structural Opti- tion,” American Statistician, Vol. 53, No. 3, 1999, pp. 216–224.
30 Jensen, J. L., Saddlepoint Approximations, Clarendon, Oxford, 1995.
mization,” Advances in Design Optimization, edited by H. Adeli, Chapman 31 Mitteau, J. C., “Error Evaluations for the Computation of Failure Prob-
and Hall , London, 1994, pp. 492–570. ability in Static Structural problems,” Probabilistic Engineering Mechanics,
2 Wang, L., Grandhi, R. V., and Hopkins, D. A., “Structural Reliability
Vol. 14, No. 1, 1999, pp. 119–135.
Optimization Using an Efficient Safety Index Calculation Procedure,” Inter- 32 Johnson, N. L., and Kotz, S., Continuous Univariate Distributions-2,
national Journal for Numerical Methods in Engineering, Vol. 38, No. 10, Wiley, New York, 1970.
1995, pp. 1721–1738. 33 Read, K. L. Q., “A Lognormal Approximation for the Collector’s Prob-
3 Tu, J., Choi, K. K., and Young, H. P., “A New Study on Reliability-Based
lem,” American Statistician, Vol. 52, No. 2, 1998, pp. 179–180.
Design Optimization,” Journal of Mechanical Engineering, Vol. 121, No. 4, 34 Hoffman, R. M., Sudjianto, A., Du, X., and Stout, J., “Robust Piston
1999, pp. 557–564. Design and Optimization Using Piston Secondary Motion Analysis,” Society
4 Wu, Y.-T., and Wang, W., “A New Method for Efficient Reliability-
of Automotive Engineers, SAE Paper 2003-01-0148, Warrendale, PA, 2003.
Based Design Optimization,” Probabilistic Mechanics & Structural Relia-
bility: Proceedings of the 7th Special Conference, American Society of Civil A. Messac
Engineers, New York, 1996, pp. 274–277. Associate Editor