Professional Documents
Culture Documents
Topic1.4a Reference Point Estimate Method
Topic1.4a Reference Point Estimate Method
ABSTRACT: Rosenblueth’s point-estimate method for approximating the low-order moments of functions of
random variables is widely used in geotechnical reliability analyses. It is a special case of numerical quadrature
based on orthogonal polynomials. For normal variables, it corresponds to Gauss-Hermite quadrature, but Ro-
senblueth’s procedure automatically generates the weights and abscissas of Gauss-Legendre and Gauss-Laguerre
Downloaded from ascelibrary.org by Hong Kong University of Sci and Tech (HKUST) on 03/01/22. Copyright ASCE. For personal use only; all rights reserved.
quadrature as well. Despite beliefs to the contrary, the method is not a form of Monte Carlo simulation or FOSM
Taylor series expansion. The method is reasonably robust and can be satisfactorily accurate for a wide range of
practical problems, although the computational requirements increase rapidly with the number of uncertain
quantities. Caution should be exercised in applying the method when transformation of the uncertain quantities
severely changes the form of the distributions or when moments of order greater than the two are involved.
Examples with closed-form solutions serve to illustrate the use and accuracy of the method.
practice of numerical analysis, and he knew the appropriate The higher-order central moments of fX (x) of order m are
and inappropriate applications of the point-estimate method to
engineering problems. This paper attempts to explicate these
relations and applications in a way that is more accessible to
the practicing geotechnical profession, and to place the point-
Xm = 冕 (x ⫺ X)m ⭈ fX (x) ⭈ dx (2)
estimate within the broader category of numerical models, We use Xm for the mth central moment to distinguish it from
such as the finite-element method. The method is a powerful the mth power of X . The second central moment, X2 , is the
tool, and more rigorously based than many workers in geo- variance, and its square root is the standard deviation, X . The
technical engineering and reliability analysis would credit. discrete pmf has nonzero probabilities at a limited number of
discrete points only. The corresponding moments of the dis-
MATHEMATICAL BACKGROUND crete pmf pX (x) are
The basic mathematical problem is that of a random variable
or variables, X, with probability distribution function (pdf ),
Xm = 冘 (x ⫺ X)m ⭈ pX (x) (3)
Then, equating the moments of fX (x) and pX (x) yields
冕 冘
1
Consulting Engr., 23 Fredana Rd., Waban, MA 02468.
2
Prof. and Chair., Dept. of Civ. and Envir. Engrg., Univ. of Maryland,
Xm = (x ⫺ X)m ⭈ fX (x) ⭈ dx = (x ⫺ X)m ⭈ pX (x) (4)
College Park, MD 20742.
Note. Discussion open until February 1, 2000. To extend the closing
date one month, a written request must be filed with the ASCE Manager It is well known that any probability density function can
of Journals. The manuscript for this paper was submitted for review and
possible publication on October 19, 1998. This paper is part of the Jour-
be represented to any desired degree of accuracy by taking
nal of Geotechnical and Geoenvironmental Engineering, Vol. 125, No. moments of high enough order. In principle, there are infinitely
9, September, 1999. 䉷ASCE, ISSN 1090-0241/99/0009-0779–0786/ many pmf’s of X satisfying the low-order moments of (4). To
$8.00 ⫹ $.50 per page. Paper No. 19475. limit these to a unique representation, Rosenblueth considered
JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING / SEPTEMBER 1999 / 779
x⫹ = X ⫹
2
冋 冑 冉 冊册
X
⫹ 1⫹
X
2
2
X (10)
x⫺ = X ⫹
2
冋 冑 冉 冊册
X
⫺ 1⫹
X
2
2
X (11)
P⫹ =
1
2
冋
1⫺
X
⭈册 1
2 兹1 ⫹ (X /2)2
(12)
P⫺ = 1 ⫺ P⫹ (13)
A great simplification occurs when the skewness is zero or
negligible. The distribution of X is then symmetric, and
1
P⫹ = P⫺ = ; x⫹ = X ⫹ X ; x⫺ = X ⫺ X (14)
2
Case 2. In the case that X is symmetric and approximately
Gaussian, Rosenblueth posits that x can be estimated at more FIG. 1. Rosenblueth’s Points and Weights for Two Variables,
than two points. For example, a three-point estimate would Correlated or Uncorrelated
1 lated as if the ith variable were the only random variable and
P⫹⫹⫹ = P⫺⫺⫺ = (1 ⫹ 12 ⫹ 23 ⫹ 31) (21a) the others were all held constant at their mean values. The Ȳi
8
and VYi can be calculated using the procedures of (5) and (14).
P⫹⫹⫺ = P⫺⫺⫹ =
1
(1 ⫹ 12 ⫺ 23 ⫺ 31) (21b)
This method requires only 2n ⫹ 1 evaluations of Y—two
8 evaluations for each variable and one for the estimate of y.
1
P⫹⫺⫹ = P⫺⫹⫺ = (1 ⫺ 12 ⫺ 23 ⫹ 31) (21c) RELATION TO ORTHOGONAL POLYNOMIAL
8 QUADRATURE
1 Case 2 is actually a direct application of a well-known tech-
P⫹⫺⫺ = P⫺⫹⫹ = (1 ⫺ 12 ⫹ 23 ⫺ 31) (21d )
8 nique for numerical integration known as Gauss-Hermite
For any two variables, when both of the point coordinates have quadrature, which provides an approximate value for the in-
the same sense of the standard deviation, the sign of the cor- tegral
冕
responding correlation coefficient is positive, and, when the ⫹⬁
above the mean and negative when the value is one standard I⬇ Hi ⭈ g(zi) (27)
deviation below the mean, then the weights are i=1
冋 冘冘 册
n ⫺1 n
1 in which n = number of points used in the approximation; zi
P(s1s2 . . . sn ) = 1⫹ (si)(sj)ij (22) is a point at which g(z) is to be evaluated; and Hi = corre-
2n i=1 j=i⫹1
sponding weight. Salzer et al. (1952) provide a table of the
冘
Again, values of zi and Hi for n ⱕ 20, but a more accessible reference
E[Y m ] ⬇ Pi ( yi)m (23) is Abramozitz and Stegun (1964). The first three columns of
Table 1 give the values for n ⱕ 5. The obvious example of a
where yi = value of Y evaluated at xi , and i is an appropriate function containing exp(⫺z 2) integrated along the real axis
combination of ⫹ and ⫺ signs indicating the location of xi . from minus to plus infinity is the Gaussian, or Normal, prob-
When the variables are uncorrelated, (21) reduces to P = 1/ ability distribution. Hildebrand (1956) points out that Hermite
2n. This still requires evaluating Y at 2n points, which can be polynomials are ‘‘of particular importance in the theory of sta-
onerous. Rosenblueth proposed that, when the variables are tistics.’’ The probability density function for the normal dis-
uncorrelated, the following approximations (in his notation) tribution with zero mean is
冉 冊
are also a valid generalization of (14):
1 ⫺x 2
pdf(x) = exp (28)
兹2 2 2
冘 冘
n n
Hi
⬇ (兹2zi) = xiwi
i=1 兹 i=1 (30)
The values of xi and wi are in the last two columns of Table
1. It can be seen that for n = 3 the values are the same as
those given by Rosenblueth in (18) and (19). In effect, Case
2 is a restatement of Gauss-Hermite quadrature with a change
Downloaded from ascelibrary.org by Hong Kong University of Sci and Tech (HKUST) on 03/01/22. Copyright ASCE. For personal use only; all rights reserved.
of variable.
In Case 1, when the skewness is zero, the values of the
variable x and the weights (⫾X and 0.5) are the same as the
values in Table 1 for n = 2. However, Rosenblueth’s method
in Case 1 is not simply an application of Gauss-Hermite quad-
rature with two points. It is actually based on a more radical
approach, namely, the assumption that the continuously dis-
tributed variable x can be replaced by a discrete variable that
has values only at the discrete points and that the locations of
FIG. 3. Typical Four-Point Gaussian Quadrature and Mapping
the points and their weights can be chosen so that the moments Pattern Employed for Plane Finite Element
of x are recovered. Then the integrals of functions of the dis-
crete variable should approximate the integrals of functions of
the continuous variable. The approach is then extended to the technique for numerical integration of various components,
case of a skewed variable in Case 1 and multiple variables in such as the stiffness matrix, over an irregularly shaped ele-
Case 3. ment. First, the element is mapped onto a square with nor-
To illustrate that the approach is more than simply an ap- malized coordinates s and t. The points for Gaussian quadra-
plication of a well-known numerical technique for integrating ture are identified along each coordinate direction and a grid
functions of a normally distributed variable, assume that x is is formed locating the points throughout the element. The
uniformly distributed between a and b. Eqs. (10)–(13) then weights are the products of the weights for the individual di-
give rections. Then the entire process is mapped back onto the ac-
tual element. Extensions to more than two dimensions or to
1 (b ⫺ a) (b ⫺ a) curved isoparametric elements are obvious.
P⫺ = P⫹ = ; x⫾ = X ⫾ X = ⫾ (31a,b) Rosenblueth’s third case is a similar extension of the one-
2 2 2兹3
dimensional Gaussian quadrature points to a grid of higher
These are identically equal to the weights and integration dimensions. Whereas the finite-element method works almost
points called for by two-point Gauss-Legendre quadrature, exclusively with polynomial definitions of the functions to be
which is the form of Gaussian quadrature that applies in the integrated, Rosenblueth’s method incorporates the effects of
case of a uniform weighting function. different probabilistic distributions for the underlying varia-
Now consider a function that is exponentially distributed. bles. In fact, the method implicitly allows each of these vari-
The skewness coefficient is 2 and, if the parameter (hazard ables to have a different distribution, so long as the mean and
rate) is , the mean and standard deviation are both 1/ (Evans variance of the variable is known.
et al. 1993). Eqs. (10)–(13) give
2 ⫿ 兹2 1 LIMITATIONS OF ORTHOGONAL POLYNOMIAL
P⫾ = ; x⫾ = (2 ⫾ 兹2) (32a,b) QUADRATURE
4
Again these are identical to the weights and coordinates called Although Rosenblueth’s method is a simple way of obtain-
for by Gauss-Laguerre quadrature, which is the technique ap- ing the coordinates and weights for evaluating the mean and
propriate for an exponential weighting function. variance of a function of an arbitrarily distributed variable x,
In these three cases the Rosenblueth procedure yields values it does have limitations. First, two points may not be adequate
of the coordinates and weights that are identical to the opti- to obtain accurate estimates of the moments for the function
mum values for numerical integration using two points. In fact, Y in a particular case. Rosenblueth gives some examples where
for any distribution of x, the Rosenblueth procedure in Case the procedure has to be modified for some difficult functions,
1 gives the values of the two coordinates and two weights that but in most practical cases the functions are well behaved.
recover exactly the first three moments of x and the condition A more serious deficiency appears when the method is used
that the sum of all probabilities must be 1.0. This is the optimal to evaluate moments of Y higher than the second moment. By
choice for a two-point integration scheme. Rosenblueth’s evaluating the skewness coefficient of an arbitrary function Y
method is thus analogous to the generalized method of nu- from the Case 1 procedure and by working through the alge-
merical integration with orthogonal polynomials (Press et al. bra, one discovers that, regardless of the function Y, Rosen-
1992), and it applies for any probabilistic distribution of X. blueth’s two-point procedure will always return the skewness
Miller and Rice (1983), apparently independently of Rosen- of X instead of the skewness of Y. The procedure should not
blueth, developed a similar point-estimate approach based on be used for moments higher than the second, but most practical
generalized numerical quadrature and described how to obtain reliability calculations do not use the higher moments anyway.
the coordinates and weights for multiple concentration points. Harr (1987) observes that the point-estimate method cannot be
used for moments higher than those used to define the points
RELATION WITH ‘‘GAUSS POINTS’’ IN and weights. The limitation is actually stronger: Because the
FINITE-ELEMENT METHOD two-point method returns the skewness of X and not Y, it
should not be used for moments higher than the second order
Figures similar to Fig. 3 are to be found in most textbooks even when the third moments are used to find the points and
on the finite-element method. They illustrate the most common weights.
782 / JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING / SEPTEMBER 1999
冕
points, gives exactly the same answers. When the correlation
coefficient is set to zero, M is unchanged, and M in both the I= g(z) ⭈ f (z) ⭈ dz (39)
exact and Rosenblueth calculations becomes 30.41 kPa.
A more complicated example is the margin of safety of a
slope against sliding on a single plane, otherwise known as in which the limits of integration have been left out because
the Cullman analysis (Taylor 1948). The margin of safety is they can vary depending on the form or the functions to be
冋 册
integrated. In (39), g(z) = function to be evaluated at the in-
1 H tegration points, and f (z) = so-called weighting function. The
M=c⫹ sin( ⫺ )cos [tan ⫺ tan ]␥ (36)
2 sin approximate integral is
where = inclination of the failure plane; = inclination of
the front of the slope; H = vertical distance from the crest to
I⬇ 冘 Hi ⭈ g(zi) (40)
the toe of the slope; and the other variables have the conven- The values of the Hi’s are determined by the function f (z).
tional soil mechanics meetings. If the only uncertain variables Eqs. (26)–(30) deal with the condition that f (z) is the pdf for
are the unit weight (␥), the cohesion (c), and the tangent of the Normal distribution.
the friction angle (tan ), and if these variables are mutually Rosenblueth’s method establishes the locations of the inte-
independent, the mean and variance of M can be expressed gration points and the values of the Hi’s for whatever pdf ap-
exactly. For the values of the parameters— plies to the variable X. Gaussian quadrature with n points then
guarantees that the integration will be exact for any polyno-
H = 10 m, = 20⬚, = 26⬚, ␥ = 22 kN/m3, ␥ = 2.2 kN/m3, mial function g(x) of an order up to and including (2n ⫺ 1).
In the case of the Rosenblueth two-point method, this means
c = 5 kPa, c = 0.5 kPa, tan = 0.267949 ( = 15⬚),
the results will be exact for polynomials of cubic order or less.
tan = 0.018756 (COV = 0.07) (37) Thus, in the trivial case that Y is a linear function of X, its
—the exact analytical results are TABLE 2. Estimates of Moments for Lognormal Distribution
M = 2.633 kPa, M = 0.722 kPa (38) X = exp(Y ); Y = N [0, 1]
冕
⫹X
1 Harr (1989) deals with the case in which the variables may
I= g(x) ⭈ dx (41) be correlated but their skewness coefficients are zero. The pro-
2兹3X X ⫺X
cedure starts by computing the eigenvalues and eigenvectors
Eq. (41) can be evaluated with g(x) and with [g(x)]2 under the of the correlation matrix. Then the hypercube is rotated into a
integral sign. This will give exact values of the first two mo- coordinate system whose axes coincide with the eigenvectors.
ments of g(X ) about the origin, and from these the exact values A hypersphere is imagined to pass through the corners of the
of the mean and variance can be calculated. In many practical hypercube. The calculation points are then the intersections
cases such integrals are tabulated. The point-estimate method between the rotated coordinate axes and the surface of the
is then used to estimate the first two moments of g(X ) about hypersphere. This gives 2n points. Harr further recommends
the origin, and from them the mean and variance. that the contribution of each calculation point be weighted in
Fig. 7 shows the results of performing this calculation when proportion to its eigenvalue.
Y = 1/X. The mean value of X is taken to be 20 kN/m3, which Hong (1996; 1998) deals with the other problem of uncor-
is close to the value for the total unit weight of typical geo- related variables with significant skewness. He develops, for
technical materials. Unless the COV is small, the estimates of each variable, a two-point and a three-point estimation pro-
the variance are poor, but this is satisfactory because in most cedure. The two-point equations are the same as (10)–(13)
practical cases the uncertainty in the unit weight is relatively except that the term 兹1 ⫹ (X /2)2 is replaced by
small. Fig. 8 shows the results when Y = ln X and the mean 兹n ⫹ (X /2)2 wherever it appears. The three-point equations
of X is 2. The pattern for the error in the mean of Y depends include both skewness and kurtosis and are chosen so that the
on the mean of X, but the errors in the variance of Y are not central point is at the mean value of the variables. Since this
affected by the value of the mean of X. The errors are probably point is common to all variables, the composite system has 2n
not acceptable for large values of the COV of X. Figs. 9–11 ⫹ 1 points. Skewness could have been included in the original
show the results for Y = tan X, with three values of the mean Rosenblueth procedure by selecting the points and weights ac-
of X. When the mean of X is 15 or 30⬚, the results range from cording to (10)–(13), but this would still have led to 2 n points.
excellent to satisfactory. When X is 45⬚, the errors in the var- Both Haar’s and Hong’s methods give results superior to
iance of Y become large when the COV exceeds approximately those from FOSM methods, with little or no increase in com-
0.2. However, it should be noted that a case with a mean of putational effort. Because they use fewer points than the orig-
45⬚ and a COV of 0.5 corresponds to an angle varying uni- inal point-estimate methods, they may be somewhat less ac-
formly between 22.5 and 67.5⬚, which seems an unlikely sit- curate for some functions, but limited testing indicates that
uation. Again, it must be emphasized that in all the analyses they are usually satisfactory. As long as the number of varia-
represented by Figs. 7–11, X was assumed to vary uniformly. bles is less than six, the simplicity of the original point-esti-
JOURNAL OF GEOTECHNICAL AND GEOENVIRONMENTAL ENGINEERING / SEPTEMBER 1999 / 785
means that they are outside the range of meaningful definition Press, W. H., Teukolsky, S. A., Vetterling, W. T., and Flannery, B. P.
of the variables. Finally, when faced with a problem that has (1992). Numerical recipes in FORTRAN: The art of scientific compu-
many uncertain variables, the analyst should give serious tation. Cambridge University Press, Cambridge, U.K.
thought to reducing the number of uncertain variables by fix- Rosenblueth, E. (1975). ‘‘Point estimates for probability moments.’’
Proc., Nat. Acad. of Sci., 72(10), 3812–3814.
ing those whose variations have little effect on the results and Rosenblueth, E. (1981). ‘‘Two-point estimates in probabilities.’’ Appl.
concentrating on those that contribute most to the global un- Math. Modelling, 5(2), 329–335.
certainty. Salzer, H. E., Zucker, R., and Capuano, R. (1952). ‘‘Table of the zeros
and weight factors of the first twenty Hermite polynomials.’’ J. of Res.,
CONCLUSIONS Nat. Bureau of Standards, 48, RP2294, 111–116.
Taylor, D. W. (1948). Fundamentals of soil mechanics. Wiley, New York.
Rosenblueth’s point-estimate method, which is widely used Wolff, T. F. (1996). ‘‘Probabilistic slope stability in theory and practice.’’
in geotechnical engineering reliability analysis but which has Uncertainty in the Geologic Environment, ASCE, New York, 419–433.
Wong, F. S. (1985). ‘‘First-order, second-moment methods.’’ Comp. and
also been criticized for its simplicity, is a special case of the
Struct., 20(4), 779–791.
orthogonal polynomial approximation procedures leading to
the Gaussian quadrature formulas that are well known in nu- APPENDIX II. NOTATION
merical analysis and finite-element methods. Rosenblueth pre-
scribed the method to be used when there are two or three The following symbols are used in this paper:
variables, and he showed how their correlation structure could
c = cohesion;
be incorporated in the analysis. A simple extension of Rosen- COV = coefficient of variation;
blueth’s approach applies for more than three variables—cor- E[X ] = expected value of X;
related or uncorrelated. fX (x) = probability distribution function for continuous varia-
The method is reasonably robust and is satisfactorily accu- ble X;
rate for a range of practical problems, though computational g(X ) = function of variable X;
requirements increase rapidly with the number of uncertain H = height of cut or slope;
quantities of interest. The method is not Monte Carlo simu- Hi = weight corresponding to point zi in Gauss-Hermite
lation and should not be confused with Taylor series methods. quadrature;
Caution should be used in applying the method to cases in I = integral;
which the transformation of uncertain quantities severely M = margin of safety;
changes the distributional form, or to moments higher than P⫺ , P⫹ = weights associated with x⫺ and x⫹ , respectively;
order two. In particular, when the functions of the variable X pX (x) = probability mass function for discrete variable X;
to be integrated are not well represented by a third-order pol- pmf = probability mass function;
ynomial and when the COV of X is large, the results can be pdf = probability density function;
seriously in error. si = sign(⫾1) depending on whether xi is above or below
Despite these limitations, the method remains a simple, di- mean;
rect, and effective method of computing the low-order mo- VY = COV of Y, (25);
ments of functions of random variables. Its continued use in YYi = COV of Y computed when Xi is only uncertain varia-
ble, (25);
geotechnical reliability analysis is justified by experience and
wi = weight corresponding to point xi in Rosenblueth quad-
theory, but users familiar with the mathematical principles un-
rature;
derlying it will be less likely to apply it outside the regions X, Y = variables;
where it can be expected to yield good results. x⫺ , x⫹ = values of X points above and below X , respectively;
Ȳ = E[Y ] in (24);
ACKNOWLEDGMENTS Ȳi = mean value of Y calculated with all but ith variable
The reviewers’ comments were very helpful in improving the quality held constant, (24);
and scope of the paper, and this contribution is hereby acknowledged. y = value of Y evaluated at mean values of its variables,
(24);
APPENDIX I. REFERENCES y⫺ , y⫹ = values of Y evaluated at x⫺ and x⫹ , respectively;
z = variable;
Abramowitz, M., and Stegun, I. A. (1964). Handbook of mathematical ␥ = unit weight;
functions. Dover Publications, New York. = inclination of failure plane;
Duncan, J. M. (1999). ‘‘The use of back analysis to reduce slope failure X = mean of variable X;
risk: The seventh annual Arthur Casagrande memorial lecture.’’ Civ. X m = mth central moment of variable X;
Engrg. Pract., J. Boston Soc. of Civ. Engrs. Sect., ASCE, Vol. 14(1).
X = skewness coefficient of X;
Evans, M., Hastings, N., and Peacock, B. (1993). Statistical distributions.
Wiley, New York.
= correlation coefficients for two variables;
Harr, M. E. (1987). Reliability-Based Design in Civil Engineering, Mc- ij = correlation coefficient for variables Xi and K j ;
Graw-Hill, New York. X = standard deviation of variable X;
Harr, M. E. (1989). ‘‘Probabilistic estimates for multivariate analyses.’’ = friction angle; and
Appl. Math. Modelling, 13(5), 313–318. = inclination of front of slope.