Proban - TM - Manual de Probabilidade - Veritas

SESAM
USER MANUAL
Proban Theory
General Purpose Probabilistic Analysis Program
DET NORSKE VERITAS

SESAM
User Manual
Proban Theory
General Purpose Probabilistic Analysis
Program
October 1st, 2004
Valid from program version 4.4
Developed and marketed by

DET NORSKE VERITAS
DNV Software Report No.: 96-7017 / Revision 2, October 1st, 2004
Copyright © 2004 Det Norske Veritas
All rights reserved. No part of this book may be reproduced, in any form or by any means, without permission in
writing from the publisher.
Published by:
Det Norske Veritas
Veritasveien 1
N-1322 Høvik
Norway
Telephone: +47 67 57 99 00
Facsimile: +47 67 57 72 72
E-mail, sales: [email protected]
E-mail, support: [email protected]
Website: www.dnv.com
If any person suffers loss or damage which is proved to have been caused by any negligent act or omission of Det Norske Veritas, then Det Norske Veritas shall pay compensation to such person for his proved
direct loss or damage. However, the compensation shall not exceed an amount equal to ten times the fee charged for the service in question, provided that the maximum compensation shall never exceed USD
2 millions. In this provision “Det Norske Veritas” shall mean the Foundation Det Norske Veritas as well as all its subsidiaries, directors, officers, employees, agents and any other acting on behalf of Det Norske
Veritas.
Table of Contents
1 INTRODUCTION ............................................................................................................1-1
1.1 Proban in the SESAM System ......................................................................................................... 1-2

1.2 How to read the Manual................................................................................................................... 1-4
2 PROBABILISTIC MODEL - VARIABLES..................................................................2-1
2.1 Definition of Variables .................................................................................................................... 2-1

2.2 Types of Variables ........................................................................................................................... 2-2
2.2.1 General .............................................................................................................................. 2-2
2.2.2 Distribution........................................................................................................................ 2-2
2.2.3 Event Probability............................................................................................................... 2-3
2.2.4 Fitted Distribution ............................................................................................................. 2-3
2.2.5 Fixed Value ....................................................................................................................... 2-4
2.2.6 Function............................................................................................................................. 2-4
2.2.7 Generated Distribution ...................................................................................................... 2-4
2.2.8 Process Time ..................................................................................................................... 2-5
2.3 Correlation of Random Variables .................................................................................................... 2-5
2.4 Extreme (Max/Min) of a Number of Random variables.................................................................. 2-7
2.5 Time Derivative Process.................................................................................................................. 2-7
2.6 Measured Value ............................................................................................................................... 2-8
3 PROBABILISTIC MODEL - EVENTS .........................................................................3-1
3.1 Definition of Event .......................................................................................................................... 3-1

3.2 Single event...................................................................................................................................... 3-2
3.3 Union of Events ............................................................................................................................... 3-3
3.4 Intersection of Events ...................................................................................................................... 3-4
3.5 Conditional Event ............................................................................................................................ 3-5
4 MODEL INFORMATION OBTAINED BY USE OF PROBAN ................................ 4-1
4.1 General............................................................................................................................................. 4-1

4.2 Event Probability ............................................................................................................................. 4-2
4.3 Reliability Index............................................................................................................................... 4-3
4.4 Distribution of Variable, Simulation................................................................................................ 4-4
4.5 Distribution of a Variable, FORM/SORM Methods........................................................................ 4-5
4.6 Crossing Rate ................................................................................................................................... 4-5
4.7 First Passage Probability.................................................................................................................. 4-6
4.8 Derivative of Probability.................................................................................................................. 4-6
4.9 Conditional Probability.................................................................................................................... 4-7
4.10 Conditional Distribution .................................................................................................................. 4-8
4.11 Parametric Sensitivity Factor for Probability, Crossing Rate and Derivative of Probability .......... 4-8
4.12 Parametric Sensitivity Factor for Distribution................................................................................. 4-9
4.13 Uncertainty Importance Factors....................................................................................................... 4-9
5 GUIDE TO METHODS................................................................................................... 5-1
5.1 Introduction...................................................................................................................................... 5-1

5.2 Event Probability ............................................................................................................................. 5-1
5.2.1 FORM................................................................................................................................ 5-1
5.2.2 SORM - SINGLE EVENT ................................................................................................ 5-2
5.2.3 SORM - SYSTEM............................................................................................................. 5-3
5.2.4 Monte-Carlo Simulation.................................................................................................... 5-3
5.2.5 Directional Simulation....................................................................................................... 5-4
5.2.6 Axis-Orthogonal Simulation ............................................................................................. 5-4
5.2.7 Designpoint Simulation ..................................................................................................... 5-4
5.3 Conditional Probability.................................................................................................................... 5-5
5.4 Crossing Rate ................................................................................................................................... 5-5
5.5 First Passage Probability.................................................................................................................. 5-6
5.6 Distribution of Variable ................................................................................................................... 5-6
5.6.1 Monte-Carlo Simulation.................................................................................................... 5-6
5.6.2 Latin-Hypercube Simulation ............................................................................................. 5-6
5.6.3 Mean-Value FORM........................................................................................................... 5-7
5.6.4 Distribution of Variable from Parameter Study ................................................................ 5-7
5.7 Parametric Sensitivity for Probability, Crossing Rate and Derivative of Probability ..................... 5-7
5.8 Parametric Sensitivity Factor for Distribution Moments................................................................. 5-8
5.9 Uncertainty Importance Factors....................................................................................................... 5-8
6 TRANSFORMATION OF VARIABLES.......................................................................6-1
6.1 Motivation........................................................................................................................................ 6-1

6.2 Rosenblatt Transformation .............................................................................................................. 6-2
7 FIRST ORDER RELIABILITY METHOD, FORM ....................................................7-1
7.1 Motivation........................................................................................................................................ 7-1

7.2 Single Event..................................................................................................................................... 7-2
7.3 Intersection of Single Events ........................................................................................................... 7-3
7.3.1 Linearization of Small Intersection ................................................................................... 7-3
7.3.2 Linearization of Large Intersection ................................................................................... 7-4
7.3.3 Proban Options .................................................................................................................. 7-5
7.4 Union of Single Events .................................................................................................................... 7-5
7.4.1 The Complementary Problem ........................................................................................... 7-5
7.4.2 Bounds............................................................................................................................... 7-5
7.4.3 Proban Options .................................................................................................................. 7-6
8 SECOND ORDER RELIABILITY METHOD, SORM ...............................................8-1
8.1 Motivation........................................................................................................................................ 8-1

8.2 Single Event..................................................................................................................................... 8-1
8.3 Intersection of Single Events ........................................................................................................... 8-3
8.3.1 Approximation of Small Intersection ................................................................................ 8-3
8.3.2 Approximation of Large Intersection ................................................................................ 8-4
8.3.3 Proban Options .................................................................................................................. 8-5
8.4 Union of Single Events .................................................................................................................... 8-5
8.4.1 The Complementary Problem ........................................................................................... 8-5
8.4.2 Bounds............................................................................................................................... 8-5
8.4.3 Proban Options .................................................................................................................. 8-5
8.5 Union of Intersections Single Events............................................................................................... 8-5
8.5.1 Bounds............................................................................................................................... 8-5
9 NESTED RELIABILITY ANALYSIS ...........................................................................9-1
9.1 Method ............................................................................................................................................. 9-1

9.1.1 Restrictions........................................................................................................................ 9-1
10 CROSSING RATE ......................................................................................................... 10-1
11 FIRST PASSAGE PROBABILITY.............................................................................. 11-1
12 APPLICATIONS INVOLVING EQUALITY EVENTS ............................................ 12-1
13 APPLICATIONS INVOLVING MEASURED VALUES .......................................... 13-1
14 MEAN VALUE FORM FOR DISTRIBUTIONS ....................................................... 14-1
15 GENERATED DISTRIBUTION .................................................................................. 15-1
16 DESIGN POINT SIMULATION .................................................................................. 16-1
16.1 Motivation...................................................................................................................................... 16-1
17 MONTE-CARLO SIMULATION ................................................................................ 17-1
17.1 Simulation of Probability............................................................................................................... 17-1

17.2 Adjusted Simulation of Probability ............................................................................................... 17-2
17.3 Simulation of Distribution ............................................................................................................. 17-3
17.4 Options for Simulation of Probability............................................................................................ 17-3
17.5 Options for Simulation of Distribution .......................................................................................... 17-3
18 DIRECTIONAL SIMULATION .................................................................................. 18-1
18.1 Motivation...................................................................................................................................... 18-1

18.2 Sampling Formula.......................................................................................................................... 18-2
18.3 Options........................................................................................................................................... 18-3
19 AXIS-ORTHOGONAL SIMULATION ...................................................................... 19-1
19.1 Motivation...................................................................................................................................... 19-1

19.2 Selection of Axis............................................................................................................................ 19-2
19.3 Sampling of a Multiplicative Correction ....................................................................................... 19-2
19.4 Sampling of an Additive Correction .............................................................................................. 19-3
19.5 Computation of PA ......................................................................................................................... 19-3
20 LATIN-HYPERCUBE SIMULATION ........................................................................20-1
20.1 Method ........................................................................................................................................... 20-1

20.2 Options for Simulation of Probability ........................................................................................... 20-2
21 PARAMETRIC SENSITIVITY FOR PROBABILITY AND CROSSING RATE..21-1
21.1 Introduction.................................................................................................................................... 21-1

21.2 Single Event................................................................................................................................... 21-2
21.3 Small Intersection .......................................................................................................................... 21-3
21.4 Large Intersection .......................................................................................................................... 21-3
21.5 Union ............................................................................................................................................. 21-3
22 PARAMETRIC SENSITIVITY FOR DISTRIBUTION ...........................................22-1
APPENDIX A OPTIMIZATION.......................................................................................... A-1
A1 NLPQL............................................................................................................................................ A-1
A2 The SQP optimizer.......................................................................................................................... A-2
A3 The RSM optimizer ........................................................................................................................ A-3
A4 The RFCRC optimizer .................................................................................................................... A-4
A5 The UNMIN optimizer ................................................................................................................... A-5
APPENDIX B CODING OF A FUNCTION ....................................................................... B-1
B1 General.............................................................................................................................................B-1
B2 Differentiability ...............................................................................................................................B-1
B3 Consistent Model .............................................................................................................................B-4
APPENDIX C QUADRATIC FORMS IN NORMAL SPACE.......................................... C-1
C1 Parabola ...........................................................................................................................................C-1
C2 Second Order Taylor Expansion......................................................................................................C-1
C3 Asymptotic Results ..........................................................................................................................C-2
APPENDIX D MULTINORMAL INTEGRAL .................................................................. D-1
APPENDIX E BOUNDS ON PROBABILITY OF A UNION........................................... E-1
E1 Union................................................................................................................................................E-1
E2 Simple Bounds .................................................................................................................................E-1
E3 Ditlevsen Bounds .............................................................................................................................E-2
E4 Reliability Index for Bounds............................................................................................................E-2
APPENDIX F TERMINOLOGY..........................................................................................F-1
F1 Introduction......................................................................................................................................F-1
F2 Events...............................................................................................................................................F-1
F3 Variables ..........................................................................................................................................F-3
F4 Probability Measures .......................................................................................................................F-3
F5 Parametric Sensitivity and Uncertainty Importance ........................................................................F-3
F6 Approximation of Event ..................................................................................................................F-4
F7 Simulation Methods .........................................................................................................................F-5
F8 Statistical Quantities ........................................................................................................................F-5
REFERENCES...................................................................................................ΡΕΦΕΡΕΝΧΕΣ−1
SESAM Proban Theory
Program version 4.4 01-OCT-2004 1-1
1 INTRODUCTION
Proban provides the industry with an advanced probabilistic analysis tool, i.e. a program for the computa-
tion of probabilities, crossing rates, distributions and sensitivity factors associated with structural safety.
Proban also includes methods which covers demands in applications such as economic risk analysis, tradi-
tional risk and reliability analysis, stochastic reservoir simulation, operations analysis etc.
Hence, Proban includes the First Order Reliability Method, FORM, and the Second Order reliability
Method, SORM, that has been introduced to solve the high reliability problems often encountered in studies
of structural safety. These methods are theoretically justified by asymptotic theory. In many applications the
probability is typically 10-3 or less, making the use of asymptotic methods attractive.
Through Proban a number of sampling methods are also made available. Those are the (adjusted) Monte
Carlo simulation method, the Directional simulation method, the Axis-orthogonal simulation method, the
Designpoint simulation method, and the Latin-Hypercube simulation method. The Monte Carlo simulation
method and the Latin-Hypercube simulation method are efficient for sampling the central part of a distribu-
tion. The Designpoint simulation method, is also efficient for sampling probabilities in the tail of a distribu-
tion, provided the number of variables in the problem is not too large. The Axis-orthogonal simulation
method samples a correction to the FORM-reliability for an intersection. The Designpoint simulation
method samples the probability around the design point using an adjusted Monte Carlo simulation method.
The events that Proban can handle are method dependent. If the method relies on FORM or SORM, then the
event can at most be a union of intersections of single events. If the method is the Monte Carlo simulation
method or the Directional simulation method then an arbitrary composition of intersections and unions can
be treated.
A variable can be modelled as a numeric constant, a function, a distribution, a time dependent stochastic
process and the probability of an event. A parameter of a variable can be assigned a coordinate of an other
variable so that a network structure for dependencies between variables can be defined. Additional statisti-
cal dependence between variables can be modelled through correlations.
A library of functions and a library of distributions are made available with Proban. Simple functions and
integrals of functions can be input interactively. Thus quite complex statistical models can be modelled
through the input.
Proban Theory SESAM
1-2 01-OCT-2004 Program version 4.4
The results of a probabilistic analysis are a) the probability of a stochastic process being in or crossing into
an event domain, b) the crossing rate, c) the probability of an event conditioned on a measured variable, d)
the partial derivative of the probability with respect to a measured value or a threshold value of the event
function, together with sensitivity factors and uncertainty importance factors.
Alternatively the distribution of a variable or an event function as well as its first four moments, the mean,
the standard deviation, the skewness and the kurtosis are estimated. The distribution may be restricted to an
event domain. Sensitivity factors are provided for the statistical moments.
The purpose of this manual is to provide the user a detailed description of the methods implemented in
Proban. Hopefully the manual will guide the user to an appropriate model for the problem at hand.
1.1 Proban in the SESAM System

SESAM is comprised of preprocessors, environmental analysis programs, structural analysis programs and
post processors. An overview of SESAM is shown in Figure 1.1.
SESAM Proban Theory
1.1
Figure 1.1 SESAM overview

Proban Theory SESAM
1.2 How to read the Manual

• Chapter 2 describes the models for variables that Proban can handle.
• Chapter 3 describes the models for events that Proban can handle.
• Chapter 4 describes the results that Proban delivers.
• Chapter 5 is a guide to the methods.
• Chapter 6 describes the mapping of the random model variables to a set of independent standard normal
random variables.
• Chapter 7 to Chapter 22 describes the methods implemented in Proban.
• Appendix A to Appendix E gives more specialized implementation notes.
• Appendix F tabulates the terminology used in this manual.
• Appendix REFERENCES lists the literature referenced in this manual.

SESAM Proban Theory
2 PROBABILISTIC MODEL - VARIABLES
2.1 Definition of Variables

A stochastic model is defined through variables. A variable describes a functional relationship in the physi-
cal model or the randomness of a parameter in the model.
A variable has the following attributes, a name, a description, a type and a parameter list. The name of a
variable is a user-defined attribute which is unique among the variables and is used to reference the variable.
The description of a variable is a user-defined attribute which is output when the variable is printed. The
type of variable identifies how its value is calculated. The parameter list is a set of user-defined values
which are input to the specific calculation method.
A variable X is a vector (X1, X2,..., Xn) of dimension n and coordinates Xi, i=1,2,...,n
A number of assignments can be made to change the calculation method for variables. Thus assignment of
extreme values (max/min of a number of independent realizations) and assignment of correlation coeffi-
cients are available.
Assignments can also change the interpretation of a variable. Thus a variable can be assigned as a time
derivative process or as a measured value. Its integration level in an analysis of a time dependent stochastic
process can be changed. It can be specified as a parameter for a generated distribution variable and for an
event probability variable.
A parameter of a variable is assigned a numeric constant or a coordinate of another variable. This is permis-
sible for as long as a variable does not (in)directly reference itself. Thus, Proban handles a network structure
of dependencies between variables.
A model can thus split into several variables that are modelled and studied separately and also assembled to
study the joint behaviour.
Example: The variable X has the Lognormal distribution with mean µ = y (value of variable Y), standard
deviation σ = z3 (value of coordinate 3 of variable Z) and lower bound γ = 0.5. The type attribute of X is the
Lognormal distribution and the parameter list to be used when results are required for X is µ, σ and γ.
Proban Theory SESAM
Variables are created by using the command CREATE VARIABLE.
Variables are changed by using the command CHANGE VARIABLE.
Variables are deleted by using the command DELETE VARIABLE.
2.2 Types of Variables
2.2.1 General
A variable has one of the following types: Fixed value, distribution, fitted distribution, event probability,
function, generated distribution, and time.
2.2.2 Distribution
The variable is assigned a distribution from the distribution library, e.g., the exponential distribution, the
normal distribution etc. A distribution describes the randomness of a parameter in the model. The distribu-
tions available in Proban are described in the Proban Distribution Manual.
A distribution variable may be assigned to parameters of any other variable.
2.1
Figure 2.1 Lognormal Distribution

SESAM Proban Theory
2.2.3 Event Probability

A probability variable is assigned the probability of an event E (see Chapter 3). The probability is delivered
in terms of the probability, the log10 of the probability or the reliability index, see Section 4.3.
The event probability may be conditioned on the value of some of the variables in the event model. Those
variables then work as parameters for the probability variable.
If the variables conditioned on are represented by the vector Y, then the event probability becomes PE(W;y)
where y is a realisation of Y and W is the vector of variables defining E minus the variables Y conditioned
on.
Fixed valued variables are always conditioned on and are not required to be specified as conditioning varia-
bles.
The variables Y cannot be dependent on the variables W. Thus in(direct) reference to W for the variables Y
is not permitted. Also, correlation of variables in Y and variables in W is not permitted.
A probability variable invokes a two level nested analysis with the variables Y being random at the outer
level and the variables W being random at the inner level. The analysis method is restricted to FORM. This
restricts the models for the event E to be models that can be treated by FORM, see Chapter 7. The outer
level analysis permits one single event only. At most one probability variable can be active in a model calcu-
lated on.
The variables W must permit a FORM analysis to take place. For example, the result of the conditioning
assignment must not make a subevent E non-random or without a FORM solution.
The conditioning variables Y are assigned by using the command ASSIGN CONDITIONING.
2.2.4 Fitted Distribution

The variable is created by fitting the parameters of a library distribution to data. The data may be input by
user or be generated by Proban in an analysis. The variable is then employed in further analyses just like a
distribution variable.
A fitted distribution variable may be assigned to parameters of any other variable.
The fitting methods are:
• The (weighted) least squares fit which finds the distribution parameters α which minimises the function
2
. Here wi is the weight of the argument xi, Pi is the corresponding probability and F().is
Σw i ( F ( x i ;α ) – P i )
the fitted distribution function).
• The (weighted) maximum likelihood fit which finds the number of distribution parameters α which min-
imises the function Σw i f ( xi ;α ) . Here wi is the weight of the observation xi and f() is the fitted density dis-
tribution.
• The (weighted) moment fit which calculates the first few moments of the (weighted) observations and
calculates the distribution parameters from the moment formulas.
Proban Theory SESAM
2.2
Figure 2.2 Fitted Distribution
2.2.5 Fixed Value

The variable is a constant value, for example 0.5.
A fixed variable may be assigned to parameters of every other variable. It permits sensitivity factor calcula-
tions and parameter studies for parameters which enters the physical model at more than one place.
2.2.6 Function
A function variable is assigned a function from the functions library.
Proban has a built-in library of simple mathematical functions. Simple one line formula functions as well as
one-dimensional integrals of a function can be created on input. Response fits to user defined or library
functions can also be created on input. Created functions can reference other functions, also functions cre-
ated on input. Thus quite complex functions can be modelled. More elaborate functions can be coded and
linked into the program in a user defined library of functions.
A variable which has the function type attribute may be assigned to parameters of any other variable.
2.2.7 Generated Distribution

A generated distribution variable Z is assigned the distribution of another variable X. This means that ran-
domness and functional relationships of the variables in the model for X defines the functional form of Z.
SESAM Proban Theory
The variable Z may be conditioned on the values of some of the variables in the model for X, say Y. The dis-
tribution then becomes X(W;y) where y is a realization of Y and W is the vector of variables in the model
for X excluding the variables Y.
Fixed valued variables are always conditioned on and are not required to be specified as conditioning varia-
bles.
The distribution of Z(y) has the functional form of the distribution X(W;y) and the vector y is the parameters
for the distribution. Two variables Z1(y1) and Z2(y2) are thus independent unless they are explicitly corre-
lated through specification of a correlation coefficient or through dependency between Y1 and Y2.
Variables of type generated distribution may be mutually correlated and also correlated with variables of
type distribution. A generated distribution variable may be assigned a (max/min of n) extreme value
The generated distribution invokes a FORM analysis for calculation of cumulative probabilities and for dis-
tribution fractiles. The model for X(W;y) must therefore comply with the requirements for FORM.
If a variable of W is assigned a time derivative process, then the time derivative process is neglected within
the generated distribution.
The conditioning variables Y are assigned by using the command ASSIGN CONDITIONING.
Variables of type probability and type generated distribution are not permitted in W.
2.2.8 Process Time

A time variable has two parameters, a) the starting time of a stochastic process and b) its duration.
If a time variable is assigned as an argument of another variable, then the value of the time parameter is
installed, for example during integration over time.
The time variable is required only if there is explicit time dependency in the model. If explicit time depend-
ency is not present, then the process is assumed stationary and the duration, if required, is defined outside
the time variable.
Only one time variable can be defined for a model.
A probability variable cannot be used with models that involves a time dependent stochastic process.
2.3 Correlation of Random Variables

In many cases of probabilistic analysis, the available information about the variables in the model are the
marginal distributions and their mutual correlations. The joint distributions are unknown.
In general there is an infinite number of joint distributions that leads to the same correlation of the variables.
Therefore, it may not be possible to point out one distribution and state that this is the correct one. The prob-
abilistic model information is generally incomplete.
To be able to treat distribution models with incomplete information, one has to decide the type of joint dis-
tribution. Der Kiureghian and Liu (1986) suggest the Nataf model, Nataf (1962), and shows for a variety of
Proban Theory SESAM
cases that the model yields good results. The basic idea of the Nataf correlation model is that if each random
variable is mapped onto a standard normal random variable, the joint distribution becomes the Multivariate
Normal distribution.
The Nataf model may also be a complete distribution model.
Univariate variables of type distribution and type generated may be correlated.
If each variable is a random normal variable, then the joint distribution becomes the Multinormal distribu-
tion which is also available in the Proban distributions library. The joint distribution is then
1 dv 1 dv n
f ( x ) = ( 2π ) – n / 2 det ( R ) – 1 / 2 exp  – --- v T R – 1 v … (2.1)
 2  d x1 d xn
in which R is the correlation matrix for the standard normal variables V,
–1
vi = Φ ( FX ( xi ) ) (2.2)
i
and
dv i fX ( xi )
= --------------
i
(2.3)
d xi ϕ ( vi )
Here f X ( x i ) and F X ( x i ) are the density function for Xi and the cumulative distribution function for Xi
i i
respectively. The functions φ(x) and Φ(x) are the standard normal density distribution function and the
standard normal cumulative distribution function with x as argument, respectively.
If the variables are both of type distribution, then the correlation of the variables may be given directly
(option BASIC). In all other cases, the correlation coefficient must be given for the corresponding standard
normal random variables (option NORMALIZED).
A method for computing ρV from the correlation ρX of X1 and X2 and vice versa has been developed. Win-
terstein et al. (1989) use an expansion in Hermite polynomials. This permits the generality of the correlation
model.
A coefficient of correlation may be a coordinate of a variable if the variable does not (in)directly reference a
time dependent stochastic process.
If a variable is assigned an extreme value (max/min of a number of independent realizations) then the corre-
lation is between the extreme value distributions.
The Nataf correlation model is input using the command ASSIGN CORRELATION.
SESAM Proban Theory
2.4 Extreme (Max/Min) of a Number of Random variables

If a variable X is univariate and of type distribution, fitted distribution or generated distribution, then it can
be redefined to represent the distribution of the maximum or minimum of n independent realizations of X. If
X is distributed FX(x) then the distribution of X is
Y ∼ FX ( y )n (2.4)
while the distribution of Z representing the minimum of n independent realizations of X is
Z ∼ 1 – ( 1 –FX ( z ) ) n (2.5)
The number n is a positive integer.
The max- and min- extreme value distributions are defined by using the command ASSIGN EXTREME.
2.5 Time Derivative Process

In order to define a time dependent stochastic process X(t), t ∈ [ T S ,T S + D ] given by the variable X, a vari-
able Y must be assigned as the time derivative process X· ( t ) , t ∈ [ T ,T + D ] of X.
S S
The starting time TS and the duration D are modelled as attributes assigned to the time variable if present.
A time derivative process must have zero mean. This is required for the methods to give correct results.
A model in which no explicit time dependency is defined, yields a stationary stochastic process with a time
independent event boundary.
All other models yields either a non-stationary stochastic process or a time dependent event boundary.
Since any distribution from the distributions library may be assigned, the process X(t) needs not be Gaus-
sian.
Neither X nor Y can be variables of type probability, time or fixed.
The time derivative process has no meaning within the definition of a generated distribution.
A time derivative process Y is assigned to the process X by use of the command ASSIGN TIME-DERIVA-
TIVE.
Proban includes a time derivative process in the analysis only if one runs a First passage Probability analysis
or a Crossing Rate analysis.
Proban Theory SESAM
2.6 Measured Value

A variable which represents the measured value of a stochastic process must be identified as the measured
value for the equality event (see Chapter 3) which models that the underlying stochastic process has reached
a measured value. For example that a crack has grown to the value am.
Restrictions: A measured value cannot be a variable of type time or type probability.
A measured value is assigned by use of the command ASSIGN MEASURED-VALUE.

SESAM Proban Theory
3 PROBABILISTIC MODEL - EVENTS
3.1 Definition of Event

An event, E(X), is a subset of the sample space for the stochastic process involved, i.e., a subset over all the
possible outcomes of the stochastic process.
Calculation of a probability, a crossing rate, a derivative of a probability and a conditional distribution

requires definition of an event.
An event is in Proban defined through a functional relationship
E ( X ) = { x ;G ( x ) ≤ 0 } (3.1)
The event identifies the outcomes of interest while the random variables X define the nature of the stochas-
tic process.
An event has the following attributes, a name, a description, a type, and a type dependent argument list. The
name of an event is a user defined attribute which is unique among the events and is used to reference the
event. The description of an event is a user defined attribute which is output when the event is printed.
Four event types are implemented in Proban, the single event, the intersection of events, the union of events
and the conditional event.
The intersection of events and the union of events are collections of single events, unions of events and
intersection of events. Thus an arbitrary network structure of events can be modelled. However, the models
are restricted by the calculation method to be used. If the method is the Monte-Carlo simulation method or
the Directional simulation method, then there is no restriction. If the method includes the FORM, then the
model is restricted to a union of intersections of single events.
Events are created by use of the command CREATE EVENT.
Events are changed by use of the command CHANGE EVENT.
Events are deleted by use of the command DELETE EVENT.

Proban Theory SESAM
3.2 Single event

The single event defines a functional relationship which identifies a subdomain of the state space. The sin-
gle event E SE ( X ) is defined by
E SE ( X ) = { x ;G ( x ) ≤ θ } (3.2)
where G ( x ) is a function and θ is a threshold value. The event is an inequality event which defines a vol-
ume in the n-dimensional x-space, the event domain.
Alternatively, the event is defined as
E SE ( X ) = { x ;G ( x ) = θ } (3.3)
The event is an equality event which defines a volume in the n−1-dimensional x-space, the event boundary.
To be meaningful the definition of an equality event requires the definition of a limit process. If a measure-
ment variable is assigned to the event, the limit process is on the measured value. If not, the limit process is
on −θ, the threshold value.
The event function is
G SE ( X ) = G ( x ) – θ (3.4)
In structural reliability the function G SE ( x ,θ ) is denoted as a limit state function, the domain of Equation
(3.2), excluding the boundary is denoted as the failure domain or the failure set, the boundary is denoted as
C
the failure boundary and the domain of E SE ( X ) , the complement of E SE ( X ) is denoted as the safe domain
or the safe set.
3.1
Figure 3.1 Load-Resistance Limit State function

SESAM Proban Theory
A simple single event function is the load-resistance function G(r,l) shown in Figure 3.1,
G ( r ,l ) = r – l (3.5)
The event of interest is failure. Thus the event is
E SE ( R ,S ) = { ( r ,l ) ;r – l ≤ 0 } (3.6)
3.3 Union of Events

The union of m events E i ( X ) is the collection of outcomes which are in at least one subevent E i ( X ) :
E USE ( X ) = ∪ Ei ( X ) (3.7)
i=1
The subevent E i ( X ) may be a single event, a union of events or an intersection of events.
A special case is the union of single events
E USE ( X ) = ∪ ESE ,i ( X ) (3.8)

i=1
The importance of this event is that its probability can be calculated by use of bounding techniques or
through a formulation in terms of complementary events.
Another special case is the union of intersections of single events.
E IUSE ( X ) = ∪ EISE ,i ( X ) (3.9)

i=1
The importance of this event is with the FORM and the SORM, since it is the most general system that can
be treated with FORM and with SORM.
In structural reliability the event E USE ( X ) is used to model series systems, Figure 3.2.
Proban Theory SESAM
3.2
Figure 3.2 Union of Events
3.4 Intersection of Events

The intersection of m events E i ( X ) is the collection of outcomes which are common to all the subevents
Ei ( X ) :
EI ( X ) = ∩ Ei ( X ) (3.10)
i=1
The subevent E i ( X ) may be a single event, a union of events or an intersection of events.
A special case is the intersection of single events
E ISE ( X ) = ∩ EISE ,i ( X ) (3.11)

i=1
The importance of this event is that it is the only intersection system for which the probability can be calcu-
lated by use of FORM or SORM.
In structural reliability the event E ISE ( X ) is used to model parallel systems, Figure 3.3.
SESAM Proban Theory
3.3
Figure 3.3 Intersection of Events
3.5 Conditional Event

The conditional event E C ( X ) is the collection of outcomes in E 1 ( X ) conditioned on the occurrence of the
event E 2 ( X ) :
EC ( X ) = ( E1 E2 ) ( X ) (3.12)
The subevent E 1 ( X ) or E 2 ( X ) is either a single event, a union or single events or an intersection of single
events. The complexity of a subevent depends on the calculation method to be used. If FORM or SORM is
implied, the intersection of E 1 ( X ) and E 2 ( X ) must be an intersection of single events.
In structural reliability a conditional event is used to model inspection no-find situations (no equality events)
and together with the measured value variable (involves equality events) to model inspection find situations.
A conditional event is shown in Figure 3.4.

Proban Theory SESAM
3.4
Figure 3.4 Conditional Event

SESAM Proban Theory
4 MODEL INFORMATION OBTAINED BY USE OF

PROBAN
4.1 General
This section describes the type of information about the model that Proban delivers:
• The event probability.
• The corresponding reliability index.
• The event in-crossing rate.
• The first passage probability into the event.
• The cumulative probability distribution of a random variable.
• The density distribution of a random variable.
• The derivative of the probability with respect to one or more measured values and/or event threshold val-
ues.
• The conditional probability for an event conditioned on another event involving (a mixture of) inequality
events, equality events and measured values.
• The conditional distribution of a variable conditioned on an event.
• A parameter study of a probability or crossing rate as function of a numeric parameter in the model.
• The parametric sensitivity factor of the probability or the crossing rate with respect to a numeric param-
eter in the model.
• The parametric sensitivity factors of the four first statistical moments of the distribution of a random var-
iable.
Proban Theory SESAM
• The uncertainty importance factors which measures the importance to the probability or crossing rate of
modelling a random variable as random rather than as a numerical value, the distribution median being
the numerical value.
A number of calculation methods are available
• The FORM-approximation, which approximates the boundary of an event by linear forms in the u-space.
• The SORM-approximation, which approximates the boundary of an event by linear and quadratic forms
in the u-space.
• The Monte Carlo simulation method which samples points according to the distribution of the model
variables.
• The Directional simulation method which samples directions in the u-space, uniformly distributed on the
surface of the n-dimensional unit hypersphere located at the origin, n being the dimension of the u-space.
• The Axis-Orthogonal simulation method, which first finds the design point in the u-space and subse-
quently samples directions parallel to the u-space axis of the intersection event or single event calculated
on.
• The Designpoint simulation method, which first finds the design point in the u-space and subsequently
employs the Monte Carlo simulation method in the u-space but centred at the design point.
• The Latin-Hypercube simulation method which employs a strategy for stratified sampling of the distri-
bution of a random variable.
• The Adjusted simulation method, which permits a Monte Carlo simulation to be centred at a user defined
point and for variables to sampled according to a stratified method
• The Mean-Value FORM method which estimates the distribution of a random variable assuming that the
u-space gradient changes little in the probability interval of interest.
4.2 Event Probability

The event probability, PE, is the probability that an outcome of the stochastic process X yields the event E,
PE = P ( E ( X ) ) (4.1)
The probability calculations in Proban are formulated in terms of a set of independent standard normal ran-
dom variables U. To facilitate the computations, the model variables X, as defined by the user, are mapped
onto an equivalent model in terms of U. (The mapping X=T(U) is a computational trick. The user does not
have to deal with this.) The mapping is outlined in Chapter 6. The variable space defined by Ui, i=1,2,...,n is
denoted as the u-space.
In Proban an event probability is estimated through the use of
• Exact methods: Exact results for the single event probability is obtained if the event boundary is linear or
quadratic in u-space.
SESAM Proban Theory
• Approximation methods: An approximation to the event is obtained by approximating the event bound-
ary in u-space around a number of approximation points by linear, piecewise linear and quadratic sur-
faces. The choice of approximation points depends on the event in question and on the
computational method used. The purpose is that the probability of the approximation can be calculated
using exact methods, methods known to give good approximations and methods which bounds the prob-
ability. A First Order Reliability Method, FORM, makes use of linear and piecewise linear approxima-
tions to the event boundary. A Second Order Reliability Method, SORM, makes use also of quadratic
approximations to the event boundary.
• Simulation methods: Estimates of the event probability, P̂ E ( N ) , are obtained by use of simulation meth-
ods. The number of samples in the estimate is N. An unbiased sampling strategy has the property that
P̂ E ( N ) → P E ( N ) as N → ∞ . In Proban, two unbiased sampling strategies are implemented, the Monte
carlo simulation method and the Directional simulation method. In addition, two importance sampling
methods, the Axis-Orthogonal simulation method and the Designpoint simulation method are imple-
mented.
4.3 Reliability Index

The reliability index, βR, is defined as the argument of the standard normal distribution which yields 1−the
event probability,
–1 –1
βR = Φ ( 1 – PE ) = –Φ ( PE ) (4.2)
It is customary to distinguish between
• The FORM reliability index, βFORM, where PE is obtained using a FORM approximation method.
• The SORM reliability index, βSORM, where PE is obtained using a SORM approximation method.
• The generalized reliability index, βR, where PE is the exact event probability.
While the event probability is most often a rapidly changing non-linear function of the distribution parame-
ters, θ, the reliability index βR, is often a more ‘linear’ function of θ, Figure 4.1. This implies that the para-
metric sensitivity factors for βR are predictive in a wider range of θ then those for PE.
Proban Theory SESAM
4.1
Figure 4.1 Derivative of Reliability vs Derivative of Probability
The other reason for using the reliability index is historically motivated. In early structural safety applica-
tions the safety was measured in terms of an index, Cornell (1969), Hasofer and Lind (1974), Ditlevsen
(1979a), corresponding to βR for special cases.
Proban permits the probability to be calculated in terms of the probability, the log10 of the probability and
the reliability index.
4.4 Distribution of Variable, Simulation

The cumulative distribution of (the value of) the variable Y(X) is
FY ( y ) = P ( Y ( X ) ≤ y ) (4.3)
and the density distribution is the corresponding derivative fY(y).
Proban calculates the distribution of a variable by use of simulation methods. The simulation methods avail-
able are a) the Monte Carlo simulation method which samples points according to the distribution of the
random variables in the model, and b) the Latin-Hypercube simulation method which is a stratified sam-
pling strategy where each coordinate axis is divided into a prescribed number, m, of intervals of equal prob-
ability. The m points of each latin hypercube are sampled using one point from each axis interval only once.
During sampling, the four first statistical moments of the distribution, i.e. the mean, µ, the standard devia-
tion, σ, the skewness, δ, and the kurtosis, κ, are estimated. (The kurtosis proper is used here. Kurtosis excess
is obtained from kurtosis proper by subtracting 3.) Using N points in the sample, the unbiased estimator for
the mean is
N
ˆ 1
µ = ---- ∑ g ( u i ) (4.4)
N
i=1
The unbiased estimator for the standard deviation is
SESAM Proban Theory
N 1/2
 
ˆ 1 ˆ 2
σ = ------------- ∑ ( g ( u i ) – µ ) 
 (4.5)
N – 1 
 i=1 
The unbiased estimator for the skewness is
N
N ˆ 3
------------------------------------ ∑ ( g ( u i ) – µ )
(N – 1)(N – 2)
ˆ i=1 (4.6)
δ = ---------------------------------------------------------------------------
ˆ3
σ
The unbiased estimator for the kurtosis (proper) is

N
N(N + 1) - ∑ ( g ( ui ) – µ )
---------------------------------------------------
ˆ 4
( N – 1 )N – 2 ) ( N – 3 )
ˆ i=1 3 ( 3N – 5 ) -
κ = -------------------------------------------------------------------------------------------
ˆ4
- – ----------------------------------- (4.7)
σ ( N – 2)(N – 3)
4.5 Distribution of a Variable, FORM/SORM Methods

The probability distribution of the variable Y(X) is
FG ( θ ) = P ( Y ( X ) ≤ θ ) (4.8)
The parameter study on FG(θ) with θ as parameter is executed by user defining the variable G(X)=Y(X)−θ,
user creating the event E=G(X) ≤ 0 and user executing a parameter study for θ. Both FORM, SORM and
probability simulation methods may be used. Subsequently one may fit a distribution to the result of the
parameter study to obtain the density distribution fG(θ) as well.
The Mean Value FORM is a parameter study on FG(θ) with θ as parameter and is executed by Proban defin-
ing the variable G(X)=Y(X)−θ, Proban creating the event E=G(X) ≤ 0 and Proban executing a parameter
study for θ. The assumption behind the Mean Value FORM method is that the direction to the approxima-
tion point of the θ-shifted function in terms of u-space variables, g(u)−θ is closely approximated by ∇ g(0).
Here g(u) =G(T(U)). It is thus sufficient to compute the value of the event function at a number of points
within prescribed probability limits in order to obtain an approximation to the distribution of the variable.
Optionally the gradient ∇ g(ul) where ul corresponds to the lower probability bound, and the gradient
∇ g(uu), where uu corresponds to the upper probability bound, are used to modify the directions to the
approximation points of the θ-shifted event.
4.6 Crossing Rate

The calculation of the crossing rate uses the parallel system formulation developed from Madsens formula,
Hagen and Tvedt (1990). If there are m single event functions, Gi, i=1,2,...,m corresponding to the events Ei,
then the crossing rate for the intersection event is
Proban Theory SESAM
m
 m 
∂ 
ν(t) = ∑ P ∩ G j ≤ 0 ∩ G i + θ i G· i ≤ 0 ∩ – G· i ≤ 0 (4.9)
∂ θi  
i=1 j ≠ i  θi = 0
Here G· i is the time derivative process of Gi.
Notice that the parallel system formulation works with any distribution model that can be mapped onto the
u-space. However, the time derivative process must have zero mean since nonstationarity is handled only
through explicit time dependency.
If the model includes variables not assigned time derivatives, then these are averaged over.
If the process is non-stationary, and the duration of the process is given, then the crossing rate is averaged
over the duration. A trapezoidal quadrature is used to integrate over time.
A variable which is assigned as a time derivative, may itself be assigned a time derivative.
The starting time and the duration of the process are assigned to the process time variable.
The number of points in the quadrature is defaulted to 6. The value may be changed by use of the command
DEFINE ANALYSIS OPTION TIME-DEPENDENT.
4.7 First Passage Probability

The first passage probability is calculated by use of the parallel system method for calculation of the cross-
ing rate conditioned on the time-independent variables in the model and using the Poisson assumption for
the crossing probability, Hagen and Tvedt (1990). The time-independent variables are integrated by using a
formulation by Wen and Chen (1989). The Poisson assumption yields
T0 + D
 
 – mn ∫ ν ( t ) dt
 
P FPP = 1 – ( 1 – P ( T 0 ) ) exp
m  T0
-------------------------------------- (4.10)
 ( 1 – P ( T0 ) ) 
 
 
 
in which the crossing rate ν(t) calculated by the parallel system method while P(T0) is the probability of
being in the failure set at the starting time T0. The duration of the process, D, is measured from the starting
time. A trapezoidal quadrature is used for the integration over time. The constant m permits an extreme
value formulation in terms of the minimum of m independent realizations of the process. The constant n
allows for the duration of the process to be split into n equal intervals (if periodical).
The default value of both m and n is 1. The default number of points in the quadrature is 6. The value may
be changed by use of the command DEFINE ANALYSIS OPTION TIME-DEPENDENT.
4.8 Derivative of Probability

The partial derivative of the event probability with respect to one or more equality event threshold values or
measurement values permits both the calculation of the density distribution of a variable and a separate cal-
culation of the nominator and the denominator of a conditional probability calculation.
SESAM Proban Theory
If the event E involves the single equality event, SE, as E = Ẽ ∩ { x ;G SE ( x ) + θ = 0 } then
∂P E
= ∂ P ( Ẽ ∩ G SE ( x ) + θ ≤ 0 ) (4.11)
∂θ ∂θ θ=0
where θ is an auxiliary variable added to the equality event function.
If a measured value is defined for the event SE, then
∂P E
= ∂ P ( Ẽ ∩ G SE ( x ,θ ) ≤ 0 ) (4.12)
∂θ ∂θ θ=0
where θ is an auxiliary variable added to the measured value.
Proban checks the definition of the system at hand for single equality events and for measured values and
sets up the appropriate formulation.
Note that the partial derivative of Equation (4.11) or of Equation (4.12) can be simulated if only one single
equality event is included. This is done by simulating the parametric sensitivity factor with respect to θ.
4.9 Conditional Probability

In many cases the interest is on computing the probability that an event occur conditioned on the occurrence
of another event, e.g., the event that a structural element fails within a specified number of years condi-
tioned on the non-detection of cracks or the detection of a crack of a measured size, Madsen (1987).
4.2
Figure 4.2 Law of Conditional Probability
If the problem at hand includes no single equality events, then the law of conditional probability gives:
PE
EA
∩E B
= --------------------- (4.13)
EB
A PE
B
Proban Theory SESAM
If the event EB includes the single equality event, SE, as E B = E B̃ ∩ { x ;G SE ( x ) = 0 } then
∂
P ( E A ∩ E B̃ ∩ G SE ( x ) + θ ≤ 0 )
∂θ θ=0
PE EB = ------------------------------------------------------------------------------------------- (4.14)
∂
P ( E B̃ ∩ G SE ( x ) + θ ≤ 0 )
A
∂θ θ=0
where θ is an auxiliary variable added to the single equality event function.
If a measured value is defined for the event SE, then

∂ P(E
A ∩ E B̃ ∩ G SE ( x ;θ ) ≤ 0 )
∂θ θ=0
P EA EB = --------------------------------------------------------------------------------------- (4.15)
∂ P(E
B̃ ∩ G SE ( x ;θ ) ≤ 0 )
∂θ θ=0
where θ is an auxiliary variable added to the measured value.
Proban checks the definition of the system at hand. If it is a conditional system, then Proban checks for sin-
gle equality events and for measured values and sets up the appropriate formulation.
4.10 Conditional Distribution

The distribution of a variable Y E ( X ) conditioned on the event E(X) is simulated by excluding samples out-
side the event.
Y E ( X ) = { Y ( X ) ;X ∈ E } (4.16)
4.11 Parametric Sensitivity Factor for Probability, Crossing Rate and

Derivative of Probability
The parametric sensitivity factor, f θ , is defined as the derivative of the reliability measure with respect to
the parameter θ. The parameter θ is either a variable of type Fixed or a numerical parameter of a function or
of a distribution. Thus, if the main result of the calculation is the reliability index, then
∂β R
βθ = (4.17)
∂θ
and similarly for the probability P or log10(P). The parametric sensitivity factor can be used to estimate the
change of the reliability index due to a change ∆θ of the parameter:
β R ( θ + ∆θ ) = β R ( θ ) + β θ ∆θ (4.18)
The method used for the computation of the parametric sensitivity factors depends on the calculation
method for the main result. If FORM or SORM is used, then an asymptotic result and a result consistent
SESAM Proban Theory
with the FORM approximation are available. If the Directional simulation methods is used, then the para-
metric sensitivity factors are simulated simultaneously, see Chapter 21.
A subset of the parameter is selected for sensitivity calculation by use of the command ASSIGN SENSI-
TIVITY. The computation of parametric sensitivity factors is optional. The command DEFINE ANALY-
SIS-OPTION SENSTIVITY allows the user the options of none, all or selected parameters for parametric
sensitivity calculation.
4.12 Parametric Sensitivity Factor for Distribution

The parametric sensitivity factor for the four first statistical moments of the distribution of a variable with
respect to the parameter θ are defined as the derivatives of the statistical moments with respect to the param-
eter θ. The parameter θ is either a variable of type Fixed or a numerical parameter of a function or of a dis-
tribution. Thus, the definition of the parametric sensitivity factor of the mean of the distribution is:
∂µ
µθ = (4.19)
∂θ
For the standard deviation of the distribution the definition is
∂σ
σθ = (4.20)
∂θ
For the skewness of the distribution the definition is
∂δ
δθ = (4.21)
∂θ
For the kurtosis of the distribution the definition is
∂κ
κθ = (4.22)
∂θ
The parametric sensitivity factors are computed during the sampling. The sensitivity factors are the deriva-
tives of the sample estimates µ̂ , σ̂ , δ̂ and κ̂ .
A subset of the parameter is selected for sensitivity calculation by use of the command ASSIGN SENSI-
TIVITY. The computation of parametric sensitivity factors is optional. The command DEFINE ANALY-
SIS-OPTION SENSTIVITY allows the user the options of none, all or selected parameters for parametric
sensitivity calculation.
4.13 Uncertainty Importance Factors

The uncertainty importance factor, a µ , is defined as the derivative of the reliability index with respect to the
mean µ of the corresponding u-space variable, Hohenbichler and Rackwitz (1986). Thus
Proban Theory SESAM
∂β R
αµ = (4.23)
∂µ
The uncertainty importance factor indicates the importance of modelling the random variable X as a distrib-
uted variable rather than as a fixed valued variable, the median of the distribution being the fixed value.
The uncertainty importance factor are computed only if the FORM or SORM is involved or if the Direc-
tional simulation method is employed.
In the sequel, let a µ be a u-space unit vector. In the case of a single event, the vector a µ is simply
a µ = sgn ( g ( 0 ) )a (4.24)
where a is the direction vector to the design point, see Chapter 6. If the random variables X are independ-
ent, an estimate of the reliability index obtained by representing Xi by its median, is given by the omission
factor, Madsen (1988a),
ω = ( 1 – a µ2 ,i ) 1 / 2 (4.25)
The new reliability index is
β R ,new = ωβ R ,old (4.26)
If the random variable Xi is one out of a number m mutually independent random variables, the omission
factor resulting from representing the random variables Xi, i=1,2,...,m by their successive median, is
m 1/2
 
ω = 1 – ∑ a µ2 ,i
 (4.27)
 
 i=1 
In Proban, for an independent variable, 100 a i2 is printed. For a group of mutually dependent variables,
m
100 ∑ a µ2 ,i is printed. These numbers can be interpreted as the percentage of the total uncertainty, reflected
i=1
in the reliability index βR, due to the corresponding random variable or group of random variables.
In the case of compound events the uncertainty importance factors printed by Proban are the same percent-
ages of the reliability index as in the case of a single event. However, the interpretation of the factors as
omission factors, Equation (4.27), may not apply.
SESAM Proban Theory
5 GUIDE TO METHODS
5.1 Introduction
A variety of methods are made available with Proban. The application of a particular method depends, as in
all mathematical modelling and analysis, on the problem at hand and on the objective of the analysis. In the
following some important characteristics of the methods are tabulated to assist the user to choose a method
appropriate to the problem at hand. Important to the selection of a method is:
• The objective of the analysis.
• The number of random variables involved.
• The computational cost of evaluating the event function.
• The properties of the event unction, e.g. existence, continuity and differentiability.
• The reliability level of interest.
• The probability range of interest for a variable or an event function.
The guide is divided into
5.2 Event Probability
5.2.1 FORM
The First Order Reliability Method linearizes the single event boundaries at a selected set of u-space points.
The applicability of the method is as follows:
• Systems: The method is available to SE, USE, ISE and UISE systems
• Event function: The event function must be continuously differentiable.

Proban Theory SESAM
• Results: Probability, parametric sensitivity factors, and uncertainty importance factors.
• Applicability: The method is particularly useful for the high reliability problems often encountered in
engineering. The method assumes that each single event boundary is linear or nearly linear in the neigh-
bourhood of each approximation point. The method is also useful if qualitative statements about the
model are sufficient, e.g., if the results are used to compare models.
• Check: If the applicability of FORM to a specific model is uncertain, then the FORM result should be
compared for selected cases with the result obtained by using simulation. Often comparison with the
SORM result will suffice.
• Computation cost: The cost of a FORM calculation is essential the cost of solving a number of optimiza-
tion problems, one for each subsystem in the model. Assume that the system is composed of m subsys-
tems, that subsystem no. i is composed of ki single events and the single event no. j of subsystem no. i
have nij random variables. If solving the optimization problem for subsystem no. i requires ri iterations,
then the computational effort in terms of function evaluations is of the order
m ki
∑ ri ∑ ( nij + 1 ) (5.1)
i=1 j=1
5.2.2 SORM - SINGLE EVENT

In case of a single event, the Second Order Reliability Method, SORM, fits a quadratic surface to the bound-
ary of the event. Three types of quadratic fits are available, the parabolic fit, the diagonal fit and a fit based
on the Taylor expansion to the second order of the event function. The two first are invariant with respect to
the formulation of the event function, for as long as the event boundary is preserved, while this is not true
for the fit based on the second order Taylor expansion. The applicability of the method is as follows:
• Event function: The event function must be twice continuously differentiable.
• Applicability: The method assumes that the event boundary is closely approximated by a quadratic sur-
face in the neighbourhood of the approximation point. The invariance property of the parabolic form
makes it generally applicable. The diagonal fit is generally applicable if the event function can be written
n
as G ( x ) = ∑ hi ( xi ) and the random variables Xi are mutually independent. It is also applicable when
i=1
the curvatures of a full parabolic approximation have small values. The lack of invariance implies that
care has to be exercised whenever the Taylor expansion to the second order is used.
• Check: If the applicability of SORM to a specific model is uncertain, then the SORM result should, for
selected cases be compared to the result obtained by using simulation.
• Computation cost: The cost of a SORM calculation is the cost of a FORM calculation plus the cost of
deriving a quadratic approximation to the event boundary. If there are n random variables in the model,
then the additional number of function evaluations is: a) Parabolic approximation: n(n-1)/2, b) Diagonal
approximation: 2n, c) Taylor expansion to the second order: n(n+1)/2. If n is large, say 100, then simula-
tion methods become competitive also for small probabilities, say 10-3. In the cases a) and c) the
additional cost is 5000 function evaluations.
SESAM Proban Theory
5.2.3 SORM - SYSTEM

In case the system comprises more than one single event, the parabolic SORM is used. The SORM is used
to compute a correction factor to the FORM.
• Systems: The method is applicable to USE, ISE and UISE systems.
• Event function: The event function must be twice continuously differentiable.
• Applicability: Small intersection (the median of the distribution, u=0 in u-space, is not in the event): The
method assumes that the joint intersection of the boundaries of active single events is closely approxi-
mated by a parabolic surface in the neighbourhood of the approximation point.
Large intersection (u=0 is in the event): The method assumes that each single event boundary is closely
approximated by a parabolic surface in the neighbourhood of its approximation point.
These properties are typical for the high reliability problems often encountered in engineering.
• Check: If the applicability of SORM to a specific model is uncertain, then the SORM result should, for
selected cases be compared to the result obtained by using simulation.
• Computation cost: Small intersection: If there are n random variables and k active single events in the
model, then the additional number of function evaluations is: k(n-k)(n-k-1)/2. Large intersection:
Number of single events × cost of SORM for single events.
5.2.4 Monte-Carlo Simulation

The Monte Carlo simulation method samples from the joint distribution of the random basic variables.
• Systems: All systems except systems involving equality events.
• Event function: The event function should be possible for all realizations of the basic variables, i.e., all
the points where the joint density distribution has a positive value. If the event function cannot be calcu-
lated at some point x, then the point contributes zero to the probability.
• Results: Probability.
• Applicability: The Monte Carlo simulation method gives an unbiased estimate of the probability. The
method is efficient for the simulation of the central part of the distribution, i.e., probabilities in the range,
say [0.1,0.9].
• Computation cost: The number of function evaluations required to obtain a reasonable reliable estimate
of the probability PE for the un-adjusted Monte Carlo method is approximately 100/PE. The purpose of
the adjusted method is to improve on the efficiency of the method by exploiting user knowledge about
the domain of the event which contribute to the probability.
Proban Theory SESAM
5.2.5 Directional Simulation

The Directional simulation method samples directions uniformly distributed on the n-dimensional unit
sphere in u-space and accumulates a weighted sum of the event probabilities conditioned on the sampled
directions.
• Systems: All systems except systems involving equality events.
• Event function: The calculation of the event function should be possible for all valid realizations of the
basic random variables.
• Results: Probability, parametric sensitivity factors and uncertainty factors.
• Applicability: The Directional simulation method gives an unbiased estimate of the probability. The
method is efficient for simulation of small probabilities, i.e., in the ranges (0,0.1] and [0.9,1].
• Computation cost: The computational cost is proportional to the area of the surface of the n-dimensional
unit sphere that generates directions intersecting the event. As this proportion tends to be reduced with
increase in the number of random variables, the Monte Carlo method becomes competitive if the number
of random variables is large. The Monte Carlo method should be used for n large, say n≥100. The area
mentioned is also reduced for small intersection systems. For such systems the axis-orthogonal simula-
tion method or the design point simulation method may be better suited.
5.2.6 Axis-Orthogonal Simulation

The Axis-Orthogonal simulation method is an importance sampling method that simulates a correction to
the FORM approximation.
• Systems: SE and ISE (small intersection).
• Event function: A FORM approximation must be possible.
• Results: Probability, parametric sensitivity factors and uncertainty importance factors.
• Applicability: The Axis-Orthogonal simulation method gives an unbiased estimate of the probability in
the case there is one local design point only. The method is recommended for The method is recom-
mended for simulation of small probabilities, i.e., in the ranges (0,0.1] and [0.9,1].
• Computation cost: The computational cost is the cost of FORM for the same problem + cost of the addi-
tional sampling. This additional cost depends on how well the event boundary of the SE or ISE system is
approximated by linear surfaces. If well approximated, the number of additional samples may be down
to 50.
5.2.7 Designpoint Simulation

The Designpoint simulation method is an importance sampling method that simulates the probability around
the design point (closest to u-space origin approximation point) by using a u-space Monte Carlo simulation
method with the mean of the u-space samples adjusted to the design point.
• Systems: SE and ISE (small intersection).

SESAM Proban Theory
• Event function: Normally the same as for FORM. However, by using the RFCRC optimization method,
continuous differentiability at the design point is not required.
• Applicability: The Designpoint simulation method gives an unbiased estimate of the probability in the
case there is one local design point only. The method is useful for simulation of small probabilities, i.e.,
in the ranges (0,0.1] and [0.9,1] and for event boundaries that are not well approximated by linear or
quadratic surfaces in the vicinity of the design point.
• Computation cost: The computational cost is the cost of FORM for the same problem + cost of the addi-
tional sampling.
5.3 Conditional Probability

The conditional event E C ( X ) = ( E 1 E 2 ) ( X ) makes direct calculation of conditional probabilities possi-
ble. It is the method to be used for calculating the probability of an event conditioned on a measured value.
• Systems: Restricted by the choice of calculation method.
• Methods: FORM, SORM and probability simulation methods are available for calculation of conditional
probabilities.
• Applicability: As for the probability calculation method actually used.
• Computation cost: As for the probability calculation method actually used. Notice that two probability
calculations are carried out, one for ( E 1 ∩ E 2 ) ( X ) and one for E 2 ( X ) .
5.4 Crossing Rate

The rate of a stochastic process crossing into an event is calculated by using FORM or SORM in conjunc-
tion with the parallel system method. However, the present Proban implementation permits SORM only for
a model that has no time-independent variables.
• Systems: The crossing rate can be calculated for SE, USE, ISE and UISE systems. Equality events are not
permitted.
• Event function: The event functions must be continuously differentiable (FORM) or twice continuously
differentiable (SORM).
• Results: Crossing rate, parametric sensitivity factors and uncertainty importance factors.
• Applicability: Requirements as for FORM and SORM, dependent on which method is used.
• Computation cost: The cost of a crossing rate calculation depends on the problem at hand and the method
used. If there are no time-independent variables, the cost is as for a single execution of a FORM or a
SORM. If time-independent variables are present, then those variables are averaged over by a nested
Proban Theory SESAM
reliability analysis. The outer loop is a FORM for a single event while the inner loop is a FORM for each
execution of the inner loop. If the process is non-stationary, an integration over time takes place, execut-
ing a FORM or a SORM or a nested analysis for each time integration point.
5.5 First Passage Probability

The probability of a stochastic process being in or crossing into an event is calculated by using FORM or
SORM in conjunction with the parallel system method. However, the present Proban implementation per-
mits SORM only for a model that has no time-independent variables.
• Systems: The crossing rate can be calculated for SE, USE, ISE and UISE systems. Equality events are not
permitted.
• Event function: The event functions must be continuously differentiable (FORM) or twice continuously
differentiable (SORM).
• Applicability: Requirements as for FORM and SORM, dependent on which method is used.
• Computation cost: The cost of a first passage probability calculation depends on the problem at hand and
the method used. If there are no time-independent variables, the cost is as for a single execution of a
FORM or a SORM. If time-independent variables are present, then those variables are integrated over by
a nested reliability analysis. The outer loop is a FORM for a single event while the inner loop is a FORM
for each execution of the inner loop. If the process is non-stationary, an integration over time takes place,
executing a FORM or a SORM or a nested analysis for each time integration point.
5.6 Distribution of Variable
5.6.1 Monte-Carlo Simulation

The Monte Carlo simulation method samples from the joint distribution of the random basic variables.
• Variables: All types of (vector)variables except probability and time variables.
• Results: Distribution, fitted distribution, parametric sensitivity for the four first central moments of the
distribution. (Result can subsequently be fitted to a library distribution.)
• Applicability: The Monte carlo simulation method gives an unbiased estimate of the distribution. The
method is efficient for the simulation of the central part of a distribution, i.e., probabilities in the range,
say [0.01,0.99].
5.6.2 Latin-Hypercube Simulation

The Monte Carlo simulation method uses stratified sampling.
• Variables: All types of (vector)variables except probability and time variables.

SESAM Proban Theory
• Results: Distribution, fitted distribution, parametric sensitivity for the four first central moments of the
distribution. (Result can subsequently be fitted to a library distribution.)
• Applicability: The Latin-Hypercube simulation method gives an unbiased estimate of the distribution.
The method is efficient for the simulation of the central part of a distribution, i.e., probabilities in the
range, say [0.01,0.99]. As compared to the Monte Carlo simulation method, the Latin Hypercube simula-
tion method yields typically a more precise estimate of the mean and the standard deviation of the distri-
bution. The method is particular efficient if the value of the random variable Y(X) is monotone in each of
the basic variables X.
5.6.3 Mean-Value FORM

If the calculation of a variable is costly, then simulation methods may be prohibitively costly for obtaining
the distribution of a random variable. In such cases the Mean-Value FORM may be used if the gradient of
the variable, ∇y ( u ) , is predictive for the function values in the probability range of interest. (This is often
true for event functions defined for structural reliability applications.)
• Variables: All types of variables except probability and time variables.
• Results: Distribution. (Result can subsequently be fitted to a library distribution.)
• Applicability: The central part of the distribution, dependent on the linearity of the function.
• Computation cost: Up to three gradient evaluations and 16 additional function evaluations.
5.6.4 Distribution of Variable from Parameter Study

By subtracting an auxiliary parameter, θ, from the variable,
y(x) – θ ≤ 0 (5.2)
and performing a parameter study to establish P(θ), then the cumulative distribution for θ is established. The
result may subsequently be fitted to a distribution in order to obtain also the density distribution for the var-
iable.
• Variables: All types of variables except time variables.
• Results: Distribution. (Result can subsequently be fitted to a library distribution.)
• Applicability: The whole or the tail part of the distribution, dependent on the linearity of the function and
the calculation method used.
• Computation cost: Depends on the probability calculation method used.
5.7 Parametric Sensitivity for Probability, Crossing Rate and Deriva-

tive of Probability
The parametric sensitivity factor is the derivative of the main result with respect to a parameter in the model.
Proban Theory SESAM
The parametric sensitivity factors are computed using Directional simulation or a method involving FORM.
If FORM is involved, then the sensitivity factors are assembled as a weighted sum of the subsystem sensi-
tivity factors.
The methods available for an intersection of single events are the analytic method and the asymptotic
method. If FORM is used, the analytic method yields a result that is derivative of the FORM approximation.
The asymptotic method yields a result in accordance with an asymptotic formula, see Chapter 21. If SORM
or a FORM based simulation method is used, then the results for FORM are scaled with a relation of the cal-
culated result to the FORM result.
The analytic sensitivity factors are particularly useful for nested analysis, i.e., with variables of type proba-
bility, since this permits the use of gradient based optimization methods for finding the approximation
points.
The analytic sensitivity factors are default.
5.8 Parametric Sensitivity Factor for Distribution Moments

The parametric sensitivity factor is calculated for the mean, the standard deviation, the skewness and the
kurtosis of the variable.
The parametric sensitivity factors are assembled during simulation, and are thus consistent with the simu-
lated results for the distribution.
The calculation of the sensitivity factors is about as costly per parameter as the computation of the general
result. If sensitivity factors are calculated for n parameters, then the extra computational effort is about n
times the calculation of the general result. It is therefore recommended to select a subset of the parameters
for parametric sensitivity calculations.
5.9 Uncertainty Importance Factors

The method used for calculation of uncertainty importance factors is consistent selected for computation of
the parametric sensitivity factors.
However, further evaluations of the event function are not required and the uncertainty importance factors
are therefore inexpensive.
SESAM Proban Theory
6 TRANSFORMATION OF VARIABLES
6.1 Motivation
Basic to the computational methods in use in Proban is the mapping of the probabilistic model in terms of n
random basic variables onto an equivalent model in terms of n independent standard normal random varia-
bles. The mapping preserves all the statistical properties of the model. Denoting the random variables X =
(X1, X2,..., Xn), the independent standard normal random variables U=(U1, U2,..., Un) and the mapping func-
tion T, the mapping of the random variables is
U = T(X) (6.1)
6.1
Figure 6.1 Mapping of the Event Function

Proban Theory SESAM
The mapping of the event function is

–1
g(u) = G(T (u)) (6.2)
An example of the mapping of a single event function is shown in Figure 6.1. It is customary to use the term
x-space for the space of basic random variables and the term u-space for the space of independent standard
normal random variables.
The usefulness of this mapping is due to
• Rotational symmetry: The probability content of domains in the u-space is invariant with the rotation of
the domains about the origin u=0.
• Exact results: The probability can in u-space be obtained exactly for the linear domain, the parabolic
domain, Tvedt (1988), Ref. /35/, and Tvedt (1989), Ref. /36/, the elliptic and hyperbolic domains, Rice
(1980), Ref. /31/, and Helstrom (1983), Ref. /17/.
• Approximations: The probability can in u-space be obtained approximately for a convex domain
bounded by linear surfaces, i.e. a domain defined by a parallel system of single events with linear event
functions. The method used is the approximation method for the cumulative multinormal distribution,
Hohenbichler (1984), Ref. /20/, and Gollwitzer and Rackwitz (1988), Ref. /12/.
If the intersection of a set of single events yields a parabolic surface in direction orthogonal to the subspace
spanned by the gradients of the active constraints, then the probability content of the corresponding domain
is approximated by a split of the linear parallel system and the parabolic domain, Hohenbichler (1984), Ref.
/20/.
• Asymptotic justification: Breitung's theorem, Breitung (1984), Ref. /4/, states that the parabolic approxi-
mation to the single event obtained by fitting curvatures at the u-space design point, u*, (the approxima-
tion point closest to the origin), asymptotically, i.e. as β=|u*|→∞, yields the true probability of the event.
A similar result is valid for a small intersection system, Hohenbichler (1984), Ref. /20/. The rapid decay
of the standard normal distribution away from the origin implies that the asymptotic approximations are
useful also for moderate β, say β≥2.
It follows from this that if an event can be approximated by one of the above mentioned domains, or a com-
position of such domains, then an approximation to the probability content of the domain can be derived.
• Simulation methods: The unbiased directional simulation method, Bjerager (1988), Ref. /2/, and the axis
orthogonal importance sampling method, Schall et al. (1988), Ref. /33/, both require a formulation in u-
space variables.
6.2 Rosenblatt Transformation

Rosenblatt (1952), Ref. /32/, demonstrated a transformation of a set of multivariate random variables onto a
set of independent random variables, each random variable being uniformly distributed on the interval [0,1].
In Hohenbichler and Rackwitz (1981), Ref. /18/, this result is used to define a mapping from the set of basic
random variables onto a set of independent standard normal random variables. The mapping assumes that
the conditional marginal distributions can be obtained in the following sequence:
Φ ( U i ) = F X ( X i X 1 ,X 2 ,… ,X n ) i = 1 ,2, …, n (6.3)
i
SESAM Proban Theory
It is customary to denote this mapping function the Rosenblatt transformation. An example for a single var-
iable is shown in Figure 6.2.
6.2
Figure 6.2 Rosenblatt Transformation for a Single Variable
For multivariate random variables the Rosenblatt transformation in general depends on the conditioning
sequence, i.e. the numbering of the basic random variables (X1, X2,..., Xn). For simulation methods, the
numbering is of little importance, since the mapping preserves all the statistical properties of the model.
However, for a FORM or a SORM calculation this implies that the approximation of the event boundary in
general depends on the numbering of the random variables. However, the effect of changing the numbering
is generally small. (The asymptotic theory yields asymptotically the same result independent of the number-
ing.) In the following cases the transformation yields design points and results independent of the way the
model is input to Proban:
• Independent random variables: In this case there is no conditioning on the right hand side of Equation
(6.3).
• Multinormal random variables: In this case, the point of maximum likelihood on the event boundary in
x-space corresponds to the point of maximum likelihood on the event boundary in u-space. Thus, the
design point is independent of the numbering of the random basic variables.
• Random variable correlated through the Nataf model: In this case, the argument for the multinormal ran-
dom variables applies to the v-space formulation of the function, see Appendix D. Because the v-space to
x-space mapping is as for independent random variables, the statement is valid.
• Random input parameters for distributions: In this case, the natural modelling of the joint distribution
yields a model in the form of the sequence assumed in Equation (6.3). Assuming the conditioning system
of the model, the Rosenblatt transformation yields the same result independent of the numbering of the
random basic variables.
Proban Theory SESAM
SESAM Proban Theory
7 FIRST ORDER RELIABILITY METHOD, FORM
7.1 Motivation
The First Order reliability method, FORM, applies to a probabilistic model formulated in terms of u-space
variables. The basic idea is to approximate the event boundary by linear surfaces at a selected set of points.
The computational methods available for linear and piecewise linear event boundaries are then applied to
obtain an estimate (possibly in terms of bounds) of the event probability. The selection of linearization
points depends on the system configuration and is outlined in the following subsections.
7.1
Figure 7.1 Mapping of Narrow Distributions
The motivation for using the FORM are:
• Almost linear event boundary: It is often so, e.g. in engineering applications, that the distributions used
to model random variables are narrow, i.e., the coefficient of variation have a small value. In such cases
the event boundary in the u-space is often significantly less curved than the event boundary in the x-
space, see Figure 7.1. Because of this, results can be obtained by the FORM with a quality sufficient for
the intended application.
Proban Theory SESAM
• Comparison of probabilistic models: An interesting application of probabilistic methods is to compare

probabilistic models, varying the model assumptions. In such cases the accuracy required is one that per-
mits comparison rather than exact results.
• Computational efficiency: In many applications in engineering, the calculation of a variable may require
computationally costly subroutines, e.g, the value of a variable may require use of the finite element
method. In such cases, the use of other methods, e.g. simulation methods, may be computationally pro-
hibitive for small failure probabilities.
• Theoretical support: Theoretical support for the FORM is offered by Breitung (1984), Ref. /4/, and
Hohenbichler (1984), Ref. /20/. Their theorems state that the system reliability index obtained using the
FORM, asymptotically, i.e. as the norm of the design point |u*| tends to infinity, under certain restrictions
on the curvatures, yields the true reliability index:
β FORM ∼ β R , u* → ∞ (7.1)
7.2 Single Event

The boundary of the single event is linearized at the point of maximum likelihood in u-space, the design
point u*, Figure 7.2:
7.2
Figure 7.2 Linearization of Single Event Boundary
The FORM reliability index is:
β FORM = u* sgn ( g ( 0 ) ) (7.2)

SESAM Proban Theory
The corresponding probability of failure is:

–1
P FORM = Φ ( β FORM ) (7.3)
7.3 Intersection of Single Events
7.3.1 Linearization of Small Intersection

7.3
Figure 7.3 Linearization of Small Intersection
The term small intersection denotes an intersection of single events where the event function is non-nega-
tive at u=0, Figure 7.3. The steps in the linearization procedure are:
• Active constraints: The first linearization point is the point of maximum likelihood, the design point u*,
on the event boundary, see Figure 7.3. The single events which have zero valued functions at this point is
denoted as active.
• Inactive constraints: The single events which are not active, are denoted as inactive. The inactive single
events are linearized in succession, if possible. The linearization point of an inactive single event is the
point of maximum likelihood on the part of its event boundary that intersects the event defined by the
previous linearizations, see Figure 7.3. If this intersection is empty, then the event is deleted from further
calculations.
Proban Theory SESAM
Denoting the linearization of single event no. i:

T
gi ( u ) = βi + ai u (7.4)
the linear surfaces defines a point, −b=(−βi), on the cumulative multivariate normal distribution, with mean
T
0, standard deviation 1 and correlation R=(ρij)= ( a i a j ) . Thus
P FORM = Φ ( – b ;R ) (7.5)
The corresponding reliability index is:

–1
β FORM = – Φ ( P FORM ) (7.6)
7.3.2 Linearization of Large Intersection

7.4
Figure 7.4 Linearization of Large Intersection
The term large intersection denotes an intersection of single events where the event function is strictly neg-
ative at u=0, Figure 7.4. The steps in the linearization procedure are:
SESAM Proban Theory
• Active constraints: The single events are first linearized at their separate design points. A single event
that has its linearization point on the event boundary of the linearized intersection is denoted as active,
see Figure 7.4.
previous linearizations, see Figure 7.4. If this intersection is empty, then the event is deleted from further
calculations.
Denoting the linearization of single event no. i:

T
gi ( u ) = βi + ai u (7.7)
T
P FORM = 1 – Φ ( – b ;R ) (7.8)

–1
β FORM = – Φ ( P FORM ) (7.9)
7.3.3 Proban Options

Proban detects whether the intersection is a small intersection or a large intersection. The user optionally
chooses to exclude inactive constraints from the linearization procedure. The pertinent command is
DEFINE FORM-SORM.
7.4 Union of Single Events
7.4.1 The Complementary Problem

Rather than computing a union of single events, one may choose to solve the complementary problem which
is an intersection of single events. The relation of the probabilities is
 m   m 
P ∪ { u ;g i ( u ) ≤ 0 } = 1 – P ∩ { u ;g i ( u ) ≥ 0 }
   (7.10)
   
i = 1  i = 1 
The intersection of single events is computed using the methods of Section 7.3.
7.4.2 Bounds
Using this method, each single event is linearized separately, Figure 7.5. An upper bound and a lower bound
to the event probability of the linearized system is computed by making use of simple bounds, (first order
bounds), see Appendix E 2 or Ditlevsen bounds, (second order bounds), see Appendix E 3. Denoting the
Proban Theory SESAM
upper bound on the probability PU and the lower bound PL, the lower bound on the reliability index
becomes:
–1
β FORM,L = – Φ ( P U ) (7.11)
The upper bound becomes:

–1
β SORM,L = – Φ ( P L ) (7.12)
7.5
Figure 7.5 Linearization of Union of Single Events

In Proban solving the complementary problem is the default method for union of single events. The user
optionally chooses a bounds method. The pertinent command is DEFINE FORM-SORM.
SESAM Proban Theory
8 SECOND ORDER RELIABILITY METHOD, SORM
8.1 Motivation
The Second Order reliability method, SORM, applies to a probabilistic model formulated in terms of u-
space variables. The basic idea is to approximate a single event boundary with a quadratic surface and an
intersection of single events by a linear/parabolic surface. In each case the quadratic approximation is made
at the design point. The computational methods available for quadratic forms and for linear and piecewise
linear forms are then applied to obtain an estimate (possibly in terms of bounds) of the event probability.
The motivation for using the SORM are similar to those for using the FORM. It should, however, be noted
that the SORM extends significantly the class of problems that can be treated by approximation methods
because even highly curved event boundaries can be well approximated.
Theoretical support for the SORM is offered by Breitung (1984), Ref. /4/, and Hohenbichler (1984), Ref. /
20/. Their theorems state that the event probability obtained using the SORM, asymptotically, i.e. as the
norm of the design point |u*| tends to infinity, under certain restrictions on the curvatures, yields the true
event probability.
8.2 Single Event

Three methods are available for approximating the u-space boundary of a single event:
• The parabolic approximation, Figure 8.1, is a parabola fitted to the main curvatures of the boundary at
the design point u*. The exact result for the probability of the parabolic domain is given in Tvedt (1988),
Ref. /35/, and in Tvedt (1989), Ref. /36/. This result and the asymptotic result by Breitung (1984), Ref. /
4/, and by Hohenbichler (unpublished) are implemented in Proban. See Appendix C for more detail.
Proban Theory SESAM
8.1
Figure 8.1 Parabolic Approximation
• The parabolic approximation is derived from the diagonal of G(u*). G(u*) is the matrix of second deriv-
atives of the event function g(u) evaluated at the design point u*. The SORM probability, PSORM, is
computed using the exact result for the parabolic quadratic form. The approximation is a modified ver-
sion of a suggestion by Der Kiureghian et al. (1987), Ref. /8/.
• The approximation derived from the second order Taylor expansion of the event function at u*, Figure
8.2. The exact result for a general quadratic form is given in Rice (1980), Ref. /31/, and Hellstrom
(1989), Ref. /17/. See also Tvedt (1989), Ref. /36/.
8.2
Figure 8.2 Second Order Taylor Approximation

SESAM Proban Theory
The SORM reliability index is:

–1
β SORM = – Φ ( P SORM ) (8.1)
8.3 Intersection of Single Events
8.3.1 Approximation of Small Intersection

8.3
Figure 8.3 SORM Approximation of Small Intersection
The SORM approximation derives a correction factor to the FORM probability, Hohenbichler (1984), Ref. /
20/. The intersection of the single event boundaries yields a non-linear surface in the space orthogonal to the
space spanned by the gradients at the design point, u*, of the active constraints. This surface is approxi-
mated by a parabolic surface. The SORM probability of failure is in this case an asymptotic result for the
linear/parabolic domain.
Q SORM
P SORM = P FORM ---------------------- (8.2)
Φ ( – u* )
where QSORM is the separate probability of the parabolic form and |u*| is the norm of the design point vec-
tor.
The method is illustrated in Figure 8.3.

Proban Theory SESAM
8.3.2 Approximation of Large Intersection

8.4
Figure 8.4 Approximation of Large Intersection
The steps in the linearization procedure are:
• Active constraints: The single events are first linearized at their separate design points. The linearized
version of a single event is as for FORM, however, the distance parameter β is adjusted to account for
the second order probability. A single event that has its adjusted linearization point on the event bound-
ary of the resulting linearized intersection is denoted as active, see Figure 8.4.
previous linearizations. If this intersection is empty, then the event is deleted from further calculations.
In Figure 8.4 the process is demonstrated for an intersection with two single events. Denoting the lineariza-
tion of single event no. i:
T
gi ( u ) = βi + ai u (8.3)
T
P SORM = 1 – Φ ( – b ;R ) (8.4)

SESAM Proban Theory
–1
β SORM = – Φ ( P SORM ) (8.5)

Proban detects whether the intersection is a small intersection or a large intersection. The user optionally
chooses ton exclude inactive constraints from the linearization procedure. The pertinent command is
DEFINE FORM-SORM.
8.4 Union of Single Events
8.4.1 The Complementary Problem

As for FORM, the union of single events is replaced by the complementary intersection of single events.
The SORM for an intersection of single events now applies.
8.4.2 Bounds
Using this method, each single event is replaced by its SORM equivalent linear surface. The bounding tech-
nique for FORM is applied to the linearized single events.

In Proban solving the complementary problem is the default method for union of single events. The user
optionally chooses a bounds method. The pertinent command is DEFINE FORM-SORM.
8.5 Union of Intersections Single Events
8.5.1 Bounds
The event probability is computed using bounds. Each intersection subsystem is treated by the methods of
Section 8.3. The Ditlevsen bounds, see Appendix E, require the probability of the intersection of intersec-
tions. This intersection probability is computed from the linearized intersections.
Proban Theory SESAM
SESAM Proban Theory
9 NESTED RELIABILITY ANALYSIS
9.1 Method
A nested reliability analysis has the form
P E = P ( G ( ( X 1 ,P SUB ( X 2 ) ) ≤ 0 ) ) (9.1)
in which the vectors of random basic variables, X1 and X2, may have elements in common.
The probability function PSUB is modelled as a variable, Xp, of type probability. This attribute assigns an
event and a probability result to Xp. The probability result is one of PSUB, log10(PSUB) and βSUB-the reliabil-
ity index corresponding to PSUB.
The variables X2 on which PSUB depends are all included in the definition of the event. The variables
become stochastic in the outer analysis by assigning them as conditioning variables for Xp.
9.1.1 Restrictions
• The top event must be a single event which is not an equality event.
• The method used for calculation of the event probability of Xp is restricted to FORM and SORM.
• The method used for calculation of the outer event is restricted to FORM.
• If the SORM method is used for calculation of PSUB, then a response surface method should be used for
the design point search for the outer event. This is necessary since the SORM does not give analytic
derivatives of PSUB with respect to PSUB.
Proban Theory SESAM
SESAM Proban Theory
10 CROSSING RATE
The rate of a stochastic process, X, crossing into an event is calculated as

TS + D
 
1
ν = E  ---- ∫ ν ( t, X 1 ) dt (10.1)
D 
 TS 
in which X=(X1,X2(t)), i.e., X1 is the vector of time independent variables and X2(t) is the vector of time
dependent variables. The time dependent variables are integrated over to yield ν ( t, x 1 ) . The expectation E
is taken over X1. TS is the starting time of the process, and D is its duration.
The crossing rate ν ( t, x 1 ) is calculated using FORM or SORM in connection with formulas derived from
the parallel system method, Hagen and Tvedt (1990), Ref. /13/.
If the stochastic process is stationary, i.e., the time t is not a variable in the model, then the formula for the
crossing rate simplifies to:
ν = E ( ν ( X1 ) ) (10.2)
If a subvector Y of X1 is to be integrated together with X2(t), then with X1=(Z,Y) there is:
TS + D
 
1
ν = E  ---- ∫ ν ( t, Z ) dt (10.3)
D 
 TS 
The subvector is assigned to the event by using the command ASSIGN SUBLEVEL INTERGRATION.
Proban Theory SESAM
SESAM Proban Theory
11 FIRST PASSAGE PROBABILITY
The probability of a stochastic process, X, crossing into an event for the first time in the time interval [TS,
TS+D] is calculated as
PE = E ( Pp ( X1 ) ) (11.1)
in which X=(X1,X2(t)), i.e., X1 is the vector of time independent variables and X2(t) is the vector of time
dependent variables. The expectation E is taken over X.
The first passage probability conditioned on the values of X1, is calculated according to the Poisson assump-
tion
TS + D
 
 mn ∫ ν ( t, x 1 ) dt
 
P p ( x 1 ) = 1 – ( 1 – P 0 ( x 1 ) ) m exp  – -------------------------------------------
-
TS
(11.2)
 1 – P0 ( x1 ) 
 
 
 
TS is the starting time of the process, and D is its duration, m yields the extreme value of m independent real-
izations of the process and n is the number of equal periods of length D. The time dependent variables X2(t)
are integrated over to yield ν ( t, x 1 ) .
The crossing rate ν ( t, x 1 ) is calculated using FORM or SORM in connection with formulas derived from
the parallel system method, Hagen and Tvedt (1990), Ref. /13/.
If the stochastic process is stationary, i.e., the time t is not a variable in the model, then the formula for the
first passage probability conditioned on the values of X1 simplifies to:
mnDν ( x 1 )
P p ( x 1 ) = 1 – ( 1 – P 0 ( x 1 ) ) m exp  – -------------------------- (11.3)
 1 – P0 ( x1 ) 
Proban Theory SESAM
If a subvector Y of X1 is to be integrated together with X2(t), then with X1=(Z,Y) the formulas become:
PE = E ( Pp ( Z ) ) (11.4)
and
TS + D
 
 mn ∫ ν ( t, z ) dt
 
P p ( z ) = 1 – ( 1 – P 0 ( z ) ) m exp  – ----------------------------------------
-
TS
(11.5)
 1 – P0 ( z ) 
 
 
 
The subvector is assigned to the event by using the command ASSIGN SUBLEVEL INTERGRATION.
SESAM Proban Theory
12 APPLICATIONS INVOLVING EQUALITY EVENTS
The single equality event defines a surface in the x-space. This surface is defined by the equation
G SE ( x ) = 0 (12.1)
corresponding to the boundary of the event function. If no measured value is assigned to the event, then
Proban computes the partial derivative of the top event probability with respect to θ, θ being the threshold
value of the equality event. Let E be the top event and E= Ẽ ∩ { x;G SE ( x ) = 0 } then
∂P ∂
= P ( Ẽ ∩ { x;G SE ( x ) – θ ≤ 0 } ) (12.2)
∂θ ∂θ θ=0
In case there are k such surfaces
G SE, i ( x ) = 0, i = 1, 2, …, k (12.3)
the formula becomes:

k k
∂P ∂
= P ( Ẽ ∩ G SE, 1 ( x ) – θ 1 ≤ 0 ∩ … ∩ G SE, k ( x ) – θ k ≤ 0 ) θ = … = θ = 0 (12.4)
∂ θ 1 … ∂θ k ∂ θ 1 … ∂θ k 1 k
The events of Equation (12.2) and Equation (12.4) are approximated using FORM or SORM. The k surfaces
G SE, i ( X ) – θ i = 0, i = 1, 2, … , k (12.5)
are treated as equality constraints during the design point search. The derivative is then obtained through a
suitable formula.
Proban Theory SESAM
SESAM Proban Theory
13 APPLICATIONS INVOLVING MEASURED VALUES
If a measurement is carried out, then the following limit process defines the meaning of measuring a varia-
ble
find = lim ( a m – ∆ ≤ a m ≤ a m + ∆ ) (13.1)

∆→0
The process is between a lower bound an upper bound as the distance between the bounds narrows to zero.
This defines the process of measuring a variable uniquely.
In Proban the measurement is defined through a single equality event which represents the surface
a ( X ) = am (13.2)
and the assignment of the variable representing am as the measured value. Notice that the functional form of
Equation (13.2) is not required. The method implemented in Proban ensures that the results of the calcula-
tions are invariant with reformulations of Equation (13.2), for as long as the surface is preserved.
If there are k measured values, each assigned a single equality event with event function
G SE, i ( x, a m, i ) = 0, i = 1, 2, …, k (13.3)
the formula becomes:

k k
∂P ∂
= P ( Ẽ ∩ G SE, 1 ( x,a m, 1 + θ 1 ) ≤ 0 ∩ … ∩ G SE, k ( x,a m, k + θ k ) ≤ 0 ) θ = … = θ = 0(13.4)
∂ θ 1 … ∂θ k ∂ θ 1 … ∂θ k 1 k
The events of Equation (13.2) and Equation (13.4) are approximated using FORM or SORM. The k surfaces
G SE, i ( x, a m, i + θ i ) = 0, i = 1, 2, … , k (13.5)
are treated as equality constraints during the design point search. The derivative is then obtained through a
suitable formula, Hagen et al (1995), Ref. /14/.
Proban Theory SESAM
SESAM Proban Theory
14 MEAN VALUE FORM FOR DISTRIBUTIONS
Since Paloheimo and Hannus (1974), Ref. /30/, a number of suggestions to linearize at points other than the
design point has been put forward. A suggestion by Wu and Wirsching (1987), Ref. /39/, is to base the
choice of linearization point on the gradient of the event function g(u), taken at the origin, ∇g ( 0 ) . The
function is calculated at points δi along the gradient direction. The probability corresponding to a point δi is
P ( g ( u ) ≤ ci ) ≈ Φ ( –δi ) (14.1)
14.1
Figure 14.1 Mean-Value FORM for Distributions

Proban Theory SESAM
If the computation of g(u) is costly, then it may be prohibitively costly to generate the distribution by other
methods than the Mean-Value FORM. (This is the only justification for the method.) However, for the
results to be useful, the level surfaces, c, of the single event function must be close to linear and nearly par-
allel in the range of interest. The event function must also be a monotonous function of δ. This is often true
for single event functions used in engineering.
A variation of this method, using also the gradients at the end-points of a prescribed probability interval, is
also available in Proban.
SESAM Proban Theory
15 GENERATED DISTRIBUTION
The distribution of the value of a variable for which properties of the distribution are calculated by the use
of FORM, is denoted as a generated distribution. Thus, by creating a variable X as generated distribution of
the variable Z, the variable X assumes the distribution of Z as calculated by FORM. The variable X can be
used like any other variable assigned a distribution from the distributions library.
The variable X can be conditioned on a set of variables, say Y, in the definition of Z. The distribution of X(y)
then becomes the distribution of Z with y as arguments to Z.
The optimization algorithm NLPQL is used with the calculation of the probability P(X(y)≤x) and the optimi-
zation algorithm UNMIN is used with the calculation of the fractile x from P(X(y)≤x)=p. These optimization
methods are described in Appendix A.
Proban Theory SESAM
SESAM Proban Theory
16 DESIGN POINT SIMULATION
16.1 Motivation
Monte Carlo sampling around the design point is an importance sampling method suggested by Shinozuka
(1984), Ref. /34/.
The idea is to find the u-space point of maximum likelihood on the event boundary and to perform a Monte-
Carlo simulation by sampling u-space points around this point. The sampling density is for each variable a
normal distribution centred at the design point. The process is demonstrated in Figure 16.1.
16.1
Figure 16.1 Design Point Simulation

Proban Theory SESAM
The probability is simulated by adding a weighted indicator function for each point.
The Monte-Carlo simulation method as implemented here samples from the joint distribution of n independ-
ent normal random variables, V, the sampled point being vi. The sampled u-space point is ui= d+vi where d
is the design point u* or optionally a point shifted from the design point. The basic variables are calculated
taking the inverse xi=T-1(ui), Section 6.2. The indicator function I(u),
1 if g ( u ) ≤ 0
I(u) =  (16.1)
0 if g ( u ) > 0
is evaluated at each sampled point and the probability is estimated as
 
exp  – ∑ u i, j
2
N  
 n   j=1 
1
P̂ E = ---- ∑ I ( u i )  ∏ σ j --------------------------------------------------------- (16.2)
N   n
j = 1   u i, j – d i, j 2 
exp  – ∑  --------------------- 
i=1
  σj  
 j=1 
Here di is the design point coordinate no. i and σi is the standard deviation of the sampling density. Option-
ally di is a point shifted from the design point.
The advantage of the method is that writing v = u-d, then there is
N
 n   n

( 2π ) n / 2 1
P̂ E = φ ( d ) ------------------ ∑ I ( u i ) ∏ σ j exp – d v i – ∑ v i2, j  1 – -----2- 
   T
(16.3)
N     σj  
i=1 j = 1   j=1 
so that one effectively simulates a multiplicative factor to the term φ(d). If each standard deviation is equal
to 1 and d is the design point, then
N
( 2π ) n / 2 T
P̂ E = φ ( u* ) ------------------ ∑ I ( u i ) exp ( – ( u* ) v i ) (16.4)
N
i=1
SESAM Proban Theory
17 MONTE-CARLO SIMULATION
17.1 Simulation of Probability

17.1
Figure 17.1 The Monte-Carlo Simulation Method
The Monte-Carlo simulation method as implemented in Proban samples from the joint distribution of the n
random basic variables as shown in Figure 17.1. This is done by sampling from n independent standard nor-
mal distributions and then taking the inverse xi=T-1(u), Section 6.2. Here ui is the sampled u-space point.
The indicator function I(x),
Proban Theory SESAM
1 if G ( x ) ≤ 0
I(x) =  (17.1)
0 if G ( x ) > 0
is evaluated at each sampled point and the probability is estimated as
N
1
P̂ E = ---- ∑ I ( x i ) (17.2)
N
i=1
An advantage of the method is that it makes use of point values of the event function only. Thus, the event
function is not required to be a smooth function of its variables. Also, the estimate of the event probability is
unbiased.
17.2 Adjusted Simulation of Probability

There are implemented two ways of adjusting the sampling density. The first method samples a point in u-
space as follows:
uj = µj + σj yj (17.3)
where yj is sampled as a standard normal distributed variable. The probability is estimated as the weighted
sum:
 

exp – ∑ u i, j
2 
N  
 n   j=1 
1
P̂ E = ---- ∑ I ( x i )  ∏ σ j ---------------------------------------------------------- (17.4)
N   n
j = 1   u i, j – µ i, j 2 
exp  – ∑  ---------------------- 
i=1
  σj  
 j=1 
The second method introduces strata for the sampled variable. The strata are entered as probability points
for the sampled variable. Two consecutive points define an interval that may optionally be weighted. The
weights are scaled so that their sum is one. The distribution of weights is sampled from as described in the
following. First a probability pu is sampled from the uniform distribution. Then the scaled weights are
summed until the sum exceeds pu. Let the lower and upper probability points for this interval of the ith sam-
ple be pi,l and pi,u. Let the summed scaled weight at the lower probability point be wi,l and the scaled weight
of the interval be ∆ wi. Then the corresponding point ps on the stratified distribution is:
( p u – w i, l )
p s = ------------------------- ( p i, u – p i, l ) + p i, l (17.5)
∆w i
The fractile xi of the distribution corresponding to ps is then calculated and the probability is estimated as
N
1 ∆w i
P̂ E = ---- ∑ I ( x i ) --------------------------- (17.6)
N ( p i, u – p i, l )
i=1
This scales the probability so that a the region outside the probability intervals contributes zero to the prob-
ability. In Proban the probability intervals are entered as an increasing sequence of optionally weighted
probabilities. By letting the sequence start with p>0, then the interval from zero to p contributes zero to the
SESAM Proban Theory
probability. By letting the sequence end with p<1, then the interval from p to 1 contributes zero to the prob-
ability.
Similarly one may enter the intervals as an increasing sequence of fractiles. The region from –∞ to the first
fractile contributes zero to the probability, and so does the region from the last region to ∞ . In Proban a
probability interval corresponding to a fractile interval is calculated by using the fractiles at u=0 to calculate
the distribution parameters that are themselves random variables. These probability intervals are used in the
simulation as above.
17.3 Simulation of Distribution

The distribution of the value of a variable is generated by ordering the sample
y i = y ( x i ), i = 1, 2, …, N (17.7)
where y is the value of the variable, xi is the ith sample of the random variables and N is the number of sam-
ples.
17.4 Options for Simulation of Probability

Maximum number of points to be sampled.
Coefficient of variation for the probability.
Maximum CPU-time.
17.5 Options for Simulation of Distribution

Maximum number of points to be sampled.
Proban Theory SESAM
SESAM Proban Theory
18 DIRECTIONAL SIMULATION
18.1 Motivation
The idea of directional simulation was put forward by Deak (1981), Ref. /6/, who used it to sample a multi-
normal probability. The method samples directions uniformly distributed on the surface of the n-dimen-
sional unit hyper-sphere centred at the origin of the u-space, Figure 18.1. The method assembles the
contributions to the probability integral conditioned on the sampled directions. A number of variance reduc-
ing methods, i.e. sampling of orthogonal sets of directions, were proposed.
18.1
Figure 18.1 Directional Simulation

Proban Theory SESAM
Application of the method to general systems with non-linear event boundaries was suggested by Bjerager
(1988), Ref. /2/. The implementation of the method in Proban follows this paper.
The motivation for the method is that it allows unbiased and efficient sampling of small probabilities, pro-
vided that n is not too large.
18.2 Sampling Formula

The probability of the event may be formulated as
PE = ∫ f 2 ( ν ) dν dΩ
χn
(18.1)
g ( ua ) ≤ 0
2
in which ν is a χ n -distributed random variable, a is a unit vector and dΩ is the surface element of the n-
2
dimensional unit sphere. The unit directions ai are sampled and the function χ n conditioned on a is inte-
grated. This is done by finding the upper bound νu,i and the lower bound νl,i of the intervals where g(ua)≤0
and by adding these contributions to the integral.
m
2 2
P ( ai ) = ∑ ( χn ( νu, i ) – χ n ( νl, i ) ) (18.2)
i=1
assuming there are m such intervals. The estimator for the probability is
N
1
P̂ E = ---- ∑ P ( a i ) (18.3)
N
i=1
where there is N sampled directions.
The simplest method that reduces the variance of the simulated values is to replace P(a) by
1
P 1 ( a ) = --- ( P ( a ) + P ( – a ) ) (18.4)
2
A further reduction of the variance of the sampled set is obtained by sampling of a randomly oriented
orthonormal system of vectors b1, b2,..., bn. A new system of vectors is formed as follows: To each set
n!
{ b i , b i , …, b i } of k out of n vectors, there are m= ----------------------- sets, the following 2k vectors ai(b), are
1 2 n k! ( n – k )!
formed
1
a i ( b ) = ------ ( s 1 b i + s 2 b i + … + s n b i ) (18.5)
k 1 2 n
with sj = ± 1 , j=1,2,...,k. One point is sampled as

m
1
2 m∑ ∑
O k = ---------
k
P ( ai ( b ) ) (18.6)
i = 1 { s ; s j = ± 1 , j = 1 , 2 , …, k }
The estimator for the probability based on sampling N orthonormal systems of vectors, is
N
1
P̂ E, k = ---- ∑ O k, j (18.7)
N
i=1
SESAM Proban Theory
Taking advantage of symmetry, the number of directions calculated, r, is half the number indicated above.
Thus
n!
r = ----------------------- 2 k – 1 (18.8)
k! ( n – k )!
The number of function calls is thus approximately rna, where na is the average number of function calls
required to perform the summation of Equation (18.2).
18.3 Options
The following options for choice of estimator are implemented:
DEFAULT The selection of estimator is based on the u-space dimensions as follows
Use P̂ E, 1 if n≤2 or 16≤n≤50
Use P̂ E, 2 if n=3 or 11≤n≤15
Use P̂ E, 3 if 4≤n≤10
RANDOM-DIRECTION The method uses P1(a) of Equation (18.4).
ORTHOGONAL-1 The method uses P̂ 1, k of Equation (18.7).
The following options for search of the intervals of Equation (18.2) are available:
RISKY-AND-FAST The sign of g(u) at u=0 is compared to the sign of g(ua) at a point, umax, determined
by the demand of accuracy in the solution. If the signs are opposite, it is assumed
that g(ua) has exactly one zero point in the direction of a in the interval [0,umax].
If the signs are the same, then g(ua) has no zero points in this interval.
MEDIUM-SAFE Calculates the function g(ua) at a sequence of equidistant user defined steps for in-
creasing u until a sign shift of the function is detected or the maximum, umax, is
reached. If a sign shift is detected, then the point g(ua)=0 is found. It is assumed
that this is the only zero point for the direction.
SAFE-AND-SLOW Steps out with equidistant user defined steps and checks for sign shifts of the func-
tion g(ua). All contributory intervals will be picked up, unless the step-length is so
large that two solutions are bypassed in one step.
STEP Step length in u-space used in the search for zero points.
SEARCH LIMIT The maximum u-space distance searched in the direction of a. The user may define
this as the corresponding probability of the excluded interval or directly as the max-
Proban Theory SESAM
imum u-space distance searched. When a zero point is found, the maximum u-
space distance searched is further reduced to P(u>umax) = P(u>uzero point)*1.0-7
SESAM Proban Theory
19 AXIS-ORTHOGONAL SIMULATION
19.1 Motivation
Axis-orthogonal sampling around the design point is an importance sampling method suggested by Hohen-
bichler and Rackwitz (1984), Ref. /22/, and Schall et al. (1986), Ref. /33/.
19.1
Figure 19.1 Axis-Orthogonal Simulation
The idea is to define an axis for a small intersection domain, see Section 7.3.1, and to define a sampling den-
sity in a plane orthogonal to this axis. The method assumes that the true boundary of the event is suitably
Proban Theory SESAM
approximated by a set of linear surfaces obtained by using the FORM linearization of small intersection
domains. The probability of the linearized domain is obtained by methods available for the multinormal dis-
tribution and to correct this probability with a simulated estimate of the probability difference of the true
event and the linearized event.
Since the cumulative multinormal distribution is computed only approximately for dimensions greater than
one, using the methods of Appendix D, the axis-orthogonal simulation method estimates an approximation
to the true probability. The quality of this approximation is governed by the quality with which the multinor-
mal integral is calculated.
In addition to the simulation of the additive correction above, there is implemented also a multiplicative cor-
rection, Bjerager (1988), Ref. /2/.
19.2 Selection of Axis

Let ai, i=1,2,...,n be the set of normalized gradients for the single events active at the design point of the
intersection. The axis, aa, is the averaged gradient, pointing into the interior of at the design point, Figure
19.1:
n
∑ ai
i=1 -
a a = – ---------------
n
(19.1)
∑ ai
i=1
The coordinate system is rotated so that the new coordinate un is in the direction of the axis, while
ũ =(u1,u2,...,un-1) are the coordinates normal to un. The axis is now defined as
a u = ũ 0 + u n e n (19.2)
where ũ 0 is the design point in the new coordinates and en is the unit vector in direction of un.
19.3 Sampling of a Multiplicative Correction

The sampling density, HM, is as follows: a) In the space spanned by the gradients of the m single events
active at the design point, the m-dimensional standard normal density, conditioned on the linearized event,
is used, Bjerager (1988), Ref. /2/. b) In the directions orthogonal to this space, and normal to the axis, au, n−
m−1 independent standard normal random variables centred on the axis are sampled from. Thus
P A ( ũ )
PE = PL ∫ --------------------------
Φ ( – u n ( ũ ) ) M
H ( ũ ) dũ = P L C (19.3)
g(u) ≤ 0
in which PL is the probability of the linearized domain, C is the multiplicative correction, u n ( ũ ) is the inter-
section of the linearized domain and the line
l ( u n ) = ũ + u n e n (19.4)
SESAM Proban Theory
P A ( ũ ) is the true probability of the event conditioned on the line defined by l(un). The estimator for C is
N
1 P A ( ũ i )
Ĉ = ---- ∑ ---------------------------
- (19.5)
N Φ ( – u n ( ũ i ) )
i=1
19.4 Sampling of an Additive Correction

The sampling density is in the n-1 dimensional standard normal density function φ ( ṽ ) with ũ = ũ 0 + ṽ ,
i.e. with centre at the design point and normal to the axis aa. The linearized domain is bounded by the linear
approximations of the single events active at the design point and the linear approximations of the con-
T
straints having a positive a i a a . The difference of the probability of the true event and the probability of its
linear approximation is estimated. Thus
φ ( ũ )
PE = PL ∫ ( P A ( ũ ) – Φ ( – u n ( ũ ) ) ) ----------- φ ( ṽ ) dṽ = P L + A
φ ( ṽ )
(19.6)
g(u) ≤ 0
The estimator for A is

N
1 φ ( ũ i )
Â = ---- ∑ ( P A ( ũ i ) – Φ ( – u n ( ũ i ) ) ) ------------- (19.7)
N φ ( ṽ i )
i=1
19.5 Computation of PA
The integral estimator are implemented:
P A ( ũ ) = ∫ φ ( u n ) du n (19.8)
g ( ũ,u n ) ≤ 0
The integral is evaluated through a search for the points un(li) and un(ui) denoting the lower and upper
bounds of the intervals where g ( ũ, u n ) is less than zero, and then to sum up the contributions using the for-
mula
m
P A ( ũ ) = ∑ ( Φ ( un ( ui ) ) – Φ ( un ( li ) ) ) (19.9)
i=1
assuming there are m such intervals. The following options for search of the intervals of Equation (19.9) are
available:
RISKY-AND-FAST The sign of g ( ũ, u n ) at u=0 is compared to the sign at un=un,max determined by the
demand of accuracy in the solution. If the signs are opposite, it is assumed that
g ( ũ, u n ) has exactly one zero point in the direction of un in the interval [0,un,max].
If the signs are the same, then g ( ũ, u n ) has no zero points in this interval.
Proban Theory SESAM
MEDIUM-SAFE Calculates the function g ( ũ, u n ) at a sequence of equidistant user defined steps for
increasing un until a sign shift of the function is detected or the maximum, un,max,
is reached. If a sign shift is detected, then the point g ( ũ, u n ) =0 is found. It is as-
sumed that this is the only zero point for the direction.
RISKY-AND-FAST Steps out with equidistant user defined steps and checks for sign shifts of the func-
tion g ( ũ, u n ) . All contributory intervals will be picked up, unless the step-length
is so large that two solutions are bypassed in one step.
SESAM Proban Theory
20 LATIN-HYPERCUBE SIMULATION
20.1 Method
20.1
Figure 20.1 The Latin-Hypercube Simulation Method
The Latin-hypercube simulation method divides each u-space coordinate axis ui into m segments, Ii(j),
j=1,2,...,m, of equal probability, p=1/m. One point is sampled from each of the segments according to the
standard normal distribution. The ith axis is thus represented by the points uij, j=1,2,...,m. The first point in
the Latin-Hypercube sample is generated by sampling one value uij,1from each axis i. The second point,
uij,2, of the sample is generated the same way, except that the segments containing uij,1 are deleted from the
Proban Theory SESAM
sample. And so on. This way the m points of one Latin-Hyperube are sampled, Mc Kay et al. (1979), Ref. /
28/. An example with n=2 and m=7 is shown in Figure 20.1. The expected number of Latin-Hypercube
points to hit an event is m/PE, the same as for the Monte Carlo method, however, the variance is in general
less.
The Latin-hypercube method is used to sample the central part of the distribution of a variable.
20.2 Options for Simulation of Probability

The number of hypercubes to be sampled.
The number of points per hypercube.
SESAM Proban Theory
21 PARAMETRIC SENSITIVITY FOR PROBABILITY

AND CROSSING RATE
21.1 Introduction
The parametric sensitivity is here discussed for the probability of an event without a time dependent sto-
chastic process in the model. The result for the first passage probability, the crossing rate and the derivative
of a probability is, however, similar.
The definition of the parametric sensitivity factor is
∂P E P E ( θ + ∆θ ) – P E
Pθ = = lim ----------------------------------------
- (21.1)
∂θ ∆θ → 0 ∆θ
where PE is the probability of the event E.
Parametric sensitivity factors can be used to estimate the change in probability resulting from an increment
∆θ of θ. Thus
P E ( θ + ∆θ ) ≈ P E + ∆θP θ (21.2)
The sensitivity factors for the probability are given also for the reliability index, βE, and the logarithm of the
probability, log10(PE). The sensitivity factor for the reliability index is generally more predictive than the
sensitivity factor for the probability.
Two methods are available for the calculation of the sensitivity factor, the asymptotic method and the ana-
lytic method. The method differs in the way intersections of single events are treated. Everything else is
treated the same way.
The asymptotic method was developed by Hohenbichler (1984), Ref. /20/, and Hohenbichler and Rackwitz
(1986), Ref. /21/. The formulas implemented in Proban are outlined below.
Proban Theory SESAM
The analytic method yields analytic derivatives for the FORM probability of intersections of single events.
The method was developed by Bjerager and Krenk (1987), Ref. /1/, Bjerager and Krenk (1989), Ref. /3/ and
Madsen (1988), Ref. /27/.
The analytic method yields analytic correspondence between PE and the partial derivatives P θ which yields
analytic derivatives in a nested reliability analysis when FORM is used for the inner loop analysis.
21.2 Single Event

The u-space formulation of the event function is g(u,θ). The FORM sensitivity factor for the reliability
index is
∂
g ( u*, θ )
∂θ
β θ, FORM = ----------------------------, β FORM → ∞ (21.3)
∇g ( u*, θ )
The expression ∂ g ( u*, θ ) is computed using numerical differentiation.

∂θ
When the SORM method is used, then the parametric sensitivity becomes:
φ ( β SORM ) Φ ( – β SORM )
β θ, SORM = ------------------------ ---------------------------- β θ, FORM, β FORM → ∞ (21.4)
φ ( β SORM ) Φ ( – β SORM )
21.1
Figure 21.1 Design Point for θ and θ + ∆θ

SESAM Proban Theory
21.3 Small Intersection

The asymptotic formula treats the linearized event as a multinormal integral where the correlation matrix is
kept fixed during differentiation. The integration limits may, however, depend on the parameter. The para-
metric sensitivity for the reliability index is
∂ Φ ( – b ;R )
Φ ( –βE ) ∂ βi ∂β i
β θ = ------------------- ∑ ------------------------------- , u* → ∞ (21.5)
φ ( βE ) Φ ( – b ;R ) ∂ θ
i∈A
The elements of b are the reliability indices of the single event linearizations at u* and βE is either of βFORM
and βSORM.
The method of Equation (21.5) neglects the influence of inactive constraints, the derivatives of the coeffi-
cient of correlation. The analytic result accounts for this, but require the additional calculation of G(u*,θ)
∂
and ∇g ( u, θ ) .
∂θ
21.4 Large Intersection

The asymptotic formula treats the linearized event as the complementary probability of a multinormal inte-
gral where the correlation matrix is kept fixed during differentiation.¨The integration limits may, however,
depend on the parameter. The parametric sensitivity for the reliability index is
– ∂ Φ ( – b ;R )
Φ ( –βE ) ∂ βi ∂β
β θ = ------------------- ∑ - i,
--------------------------------- u i → ∞, i ∈ A (21.6)
φ ( βE ) 1 – Φ ( – b ;R ) ∂ θ
i∈A
with definitions as for formula Equation (21.5).
21.5 Union
The union of m single events and the union of m intersections of single events are treated the same way:
m
β θ ≈ k ∑ φ ( β i )β θ, i (21.7)
i=1
The asymptotic result assumes that the norm, |u*|, of each subsystem design point tends to ∞ .
Using
m
Φ ( –βS ) = ∑ Φ ( –βi ) (21.8)

i=1
then k is:
Proban Theory SESAM
Φ ( –βE )
k = ------------------- φ ( β S ) (21.9)
Φ ( –βS )
where βE is the reliability index which corresponds to the upper bound of the probability.
If there is a union of single events, then by solving the complementary problem, the formulation for large
intersection is used, as described above.
SESAM Proban Theory
22 PARAMETRIC SENSITIVITY FOR DISTRIBUTION
The parametric sensitivity factor is for the mean, µ, the standard deviation, σ, the skewness, δ, and the kur-
tosis, κ, of the distribution:
The definition of the parametric sensitivity factor for the mean is
∂µ µ ( θ + ∆θ ) – µ
= lim ---------------------------------- (22.1)
∂θ ∆θ → 0 ∆θ
and similar for the three other statistical quantities.
The distribution of a variable, the statistical moments of the distribution ant the parametric sensitivity fac-
tors are computed simultaneously. The computation of the sensitivity factors require ∂ g ( u, θ ) to be com-
∂θ
puted at each sampled point.
Proban Theory SESAM
SESAM Proban Theory
Program version 4.4 01-OCT-2004 A-1
APPENDIX A OPTIMIZATION
A1 NLPQL
NLPQL is a Non-Linear Programming method which solves a sequence of Quadratic, Linearly constrained
sub problems. The problem solved is:
minimize f ( u )
subject to g i ( u ) = 0 i = 1, 2, …, m e
(A.1)
gi ( u ) ≤ 0 i = m e + 1, m e + 2, …, m
u L, j ≤ u j ≤ u U, j j = 1, 2, …, n
Here f(u) is the objective function, in Proban essentially equal to |u|, however modified so that the deriva-
tives are continuous at u=0. The number of (usually non-linear) equality constraints is m. The vector uL
stores the lower bounds and the vector uU stores the upper bounds of the n variables u.
Each iteration comprises an estimation of a search direction (requiring one evaluation of the constraint gra-
dients) and one line iteration to find the minimum of the search function.
The Kuhn-Tucker optimality criterion is
∇f ( u* ) + ∑ λi ∇gi ( u* ) = 0
i∈A
(A.2)
∑ λ i g i ( u* ) = 0
i∈A
in which A is the set of constraints active at the optimality point and λi,i∈A is the corresponding set of
Lagrange parameters. (The method implemented excludes inactive constraints, i.e., those inequality con-
straints that have g(u*)>0. Also, the contribution from variables reaching a bound is omitted from the pres-
entation.) The quantities checked for are
Proban Theory SESAM
A-2 01-OCT-2004 Program version 4.4
• Sum of constraint violations, SCV, of the active constraints:
SCV = ∑ g i ( u* ) (A.3)
i∈A
• Norm of the gradient of the Lagrangian, NLG:
 ∂ ∂ 
NLG = max  f ( u ) + ∑ λ j g i ( u* ) , j = 1, 2, …, n (A.4)
 ∂ uj i∈A
∂ uj 
• Kuhn-Tucker optimality criterion, KTO:
KTO = ∑ λ i g i ( u* ) (A.5)
i∈A
The convergence criterion, ε, is checked as follows:
KTO ≤ ε
1⁄2
SCV ≤ ε (A.6)
1⁄4
NLG ≤ ε
The search for an optimal point is performed on an augmented Lagrangian function. The default search
method is BFGS, which is a quasi Newton method that builds a sequence of quadratic approximations to the
Lagrangian from gradients. Optionally, the steepest descent method may be selected.
The search method determines a direction of improvement for the Lagrangian function. The step length in
this direction is by default unrestricted (option free). Optionally the user may specify an upper limit for the
step length in order to avoid overshooting (option value).
If constraints are poorly scaled, then the KTO and the NLG criteria may be fulfilled, while the constraint
violation criterion, SCV, is not satisfied. Therefore this criterion may optionally be excluded from the con-
vergency test or be given a user defined value.
In some analyses, for example in nested reliability analysis, it is important that the analysis continues to run
even if the inner loop optimization fails to converge. Optionally one may therefore specify that the opti-
mizer delivers the best point it found rather than stopping the calculations. When a design point search fails
it is often true that the search converges, but the stop criterion is too strict. The best point option may be use-
ful also in this case. However, the results of such an analysis should be carefully examined, since there are
situations in which the best point option yields an erroneous result.
In addition to this, one can change the maximum number of general iterations on the augmented Lagrangian
and the maximum number of function evaluations during line search for determination of the step length.
A2 The SQP optimizer

The Sequential Quadratic Programming optimizer, SQP, is an old version of the NLPQL method. The con-
vergence criteria are as for NLPQL, but the options are limited to ε, Appendix A.6, the maximum number of
general iterations and the maximum number of function evaluations during line search.
SESAM Proban Theory
A3 The RSM optimizer

The Response Surface Method optimizer works for single events only. The method approximates the u-
space event function by using information from the last m iterations, ui. Then the u-space design point is
found for the approximation surface. Convergency is checked for. If not converged, then new information
about the event function is obtained around the current design point, a new approximation surface is built
and the process is repeated.
The values calculated around design point ui are, if two-way incrementation is used:
 g ( u i ), j=0
i 
hj =  g ( u i + ∆ i e k ), j = k, k = 1, …, n (A.7)

 g ( u i – ∆ i e k ), j = n + k, k = 1, …, n
If one-way incrementation is used, then only the n+1 first values are calculated. The new function is built
from the points where
ui – uk ≤ ri (A.8)
where ri is the radius (range) of contribution, i.e., all the function values generated around design points
closer to ui than ri contribute to the approximation. The current design point is ui.
If a reduction factor different from 1 is employed, then ∆i is divided once per iteration from its initial value
by the reduction factor until its minimum value is reached.
The function built takes the form:

T 1 T
g̃ ( u ) = b + c u + --- u Gu (A.9)
2
If a linear surface is generated, then G is assumed zero. If a linear + diagonal of quadratic is generated, then
the off-diagonal elements of G are assumed zero. If a linear + quadratic function is generated, then the off
diagonal elements are also estimated.
If the unit direction vector at the initial iteration point is a, then the squares of the lowest values are summed
up until the zero constant limit is reached, but not passed, the index set being I. The u-space coordinates in
the sum are set to zero for the rest of the analysis and the final reliability index is adjusted with the omission
factor as follows:
  –1 ⁄ 2
β FORM = β FORM,Surface  1 – ∑ a i2 (A.10)
 i∈I

This is useful if an analysis is costly and it is necessary to reduce the problem size as much as possible, for
example if a FEM code is employed.
Proban Theory SESAM
A4 The RFCRC optimizer

This optimizer is restricted to a single event. It employs a Racwitz-Fiessler method, (the projected gradient
method suggested for use with structural reliability calculations). If the next step is less than the previous
step divided by a reduction factor, then the projected gradient method is used for the next step. If not, a cir-
cle step is employed. The projected gradient step is:
∇g ( u i )
s i = – ------------------------2 g ( u i ) (A.11)
( ∇g ( u i ) )
The next iterate is:
ui + 1 = ui + si (A.12)
The reduction factor is
si – 1
r i = ------------- (A.13)
si
A.1
Figure A.1 Circle Step
If this factor is greater than a user defined limit, then the method converges too slowly, or not at all. Then
the step is replaced by a circle step, Figure A.1 which finds the minimum value point, um, for the event func-
tion along a circle with radius ui, and finds the zero point for the event function along the line tum. This pro-
cedure may converge also when there is a notch at the design point.
SESAM Proban Theory
A5 The UNMIN optimizer

The UN-constrained MINimization algorithm is implemented for generated distributions only. It minimises
the value of a function on a sphere. It is used to calculate a fractile of a generated distribution from the prob-
ability.
The options are maximum number of iterations, maximum number of function evaluations during line
search and the stop criterion.
Proban Theory SESAM
SESAM Proban Theory
Program version 4.4 01-OCT-2004 B-1
APPENDIX B CODING OF A FUNCTION
B1 General
The coding of a function should be carried out with due consideration of the optimization algorithm to be
used. The following is important:
• Differentiability of the function
• Computation cost
• Consistent model
B2 Differentiability
The NLPQL optimizer uses function values and gradients to build a quadratic approximation to an (aug-
mented) Lagrange function. Each function should therefore be twice differentiable in the region of interest
for the optimization.
The RFCRC optimizer uses function values and gradients to define a direction plane for the next step. The
function should therefore differentiable in the region of interest for the optimizer. The method tolerates a
notch at the optimal point if only the direction plane is well approximated during the iterations.
The RSM optimizer is based on function values only. Since it fits a linear or quadratic surface to the u-space
function, the function should be well approximated by such surfaces.
If the gradient is poorly estimated, the NLPQL optimizer may not converge. The RFCRC optimizer will
converge, but slowly.
Some pitfalls and their remedies are given in the following.

Proban Theory SESAM
B-2 01-OCT-2004 Program version 4.4
B.1
Figure B.1 Non-differentiable at the Design Point
• Function non-differentiable at design point: A typical error is to model a system of event functions as
one single event function, for example, the intersection event in Figure B.1. This event function is not
differentiable at the design point. The NLPQL optimizer therefore fails. The RFCRC optimizer will find
the design point, but the FORM or SORM will give poor results for the reliability estimate. The RFCRC
optimizer must therefore be used in conjunction with a simulation method, for example the design point
simulation method.
Such problems should be reformulated into a system of event functions. The intersection of Figure B.1
can then be solved using NLPQL and good approximations to the probability can be calculated using
FORM or SORM.
SESAM Proban Theory
B.2
Figure B.2 Rugged Function
• Rugged function: This situation is shown in Figure B.2. A small differentiation increment ∆u1 yields a
derivative dominated by numeric noise. By using a larger differentiation increment, ∆u2, the numeric
noise is of less significance. However, if the function is strongly non-linear, increase in differentiation
increments may yield poor gradients. The best way to overcome this problem, is to code the analytic
derivative of the function as well, and use the option ANALYTIC for the calculation of derivatives.
B.3
Figure B.3 Strongly Non-linear Function
• Strongly non-linear function: This situation is shown in Figure B.3. A large differentiation increment ∆u2
yields a poor derivative dominated by curvature. By using a smaller differentiation increment, ∆u1, the
Proban Theory SESAM
curvature may be of less significance. However, the derivative may then be dominated by numeric noise.
In this situation one may use the option TWOWAY-INCREMENTATION to estimate the derivative. This
yields an error in the derivative of O((∆u)2) and therefore allows larger increments than the option ONE-
WAY-INCREMENTATION, which yields an error in the derivative of order O(∆u).
A better method is to smooth the function, if possible. If the event function can be written as
G ( x ) = r ( x1 ) – l ( x2 ) (B.1)
with x=(x1,x2), then the following methods are available:
• If r>0 and l>0 always, then
G̃ ( x ) = log ( r ( x 1 ) ) – log ( l ( x 2 ) ) (B.2)
preserves the zero surface for the function. Therefore the probability estimates and sensitivity factors are
the same as for G(x).
• If r arbitrary and/or l arbitrary always, then
G̃ ( x ) = sgn ( r ( x 1 ) ) r ( x 1 ) ξ – sgn ( l ( x 2 ) ) l ( x 2 ) ξ (B.3)
the same as for G(x). The parameter ξ can be modelled as a fixed valued variable, allowing the user to
experiment to find a suitable value of ξ.
• If r>0 and l>0 always, then
 – 1 – log ( – y ) y < – 1

logt ( y ) =  y –1 ≤ y ≤ 1
 (B.4)
 1 + log ( y ) 1<y
G̃ ( x ) = logt ( r ( x 1 ) ) – logt ( l ( x 2 ) )
the same as for G(x).
An often occurring error is to use sloppy tolerances on the results computed by sub-programs called by the
function routine. The results from such programs should have a high precision, for example 6 significant
digits. (The Proban routines that calculates the Rosenblatt transformation and its inverse produce 10 or more
significant digits.) An often occurring error is the use of sloppy tolerances on the precision in the results
from iterative procedures, for example numerical integration.
B3 Consistent Model
A consistent model has the property that the joint probability distribution of the random basic variables has
positive values only in the x-domain where the event function is defined. Thus, a consistent model is defined
for all points in u-space.
SESAM Proban Theory
If the model is not defined consistently, there is a positive probability for an undefined part of the model. If
a sequence of iterates hits a point where the event function is not defined, (ISTAT=1 is returned from a
function subroutine), the optimizer stops. The remedy is to make the model consistent.
Proban Theory SESAM
SESAM Proban Theory
Program version 4.4 01-OCT-2004 C-1
APPENDIX C QUADRATIC FORMS IN NORMAL

SPACE
C1 Parabola
The distribution of the parabolic form
n–1
2
MP = – Yn + ∑ λj Yj (C.1)
j=1
is considered in Tvedt (1988), Ref. /35/, and in Tvedt (1989), Ref. /36/. In Tvedt (1989), Ref. /36/, it is
shown how to obtain the parabolic form MP from the parabolic approximation to the u-space g-function in
the vicinity of the design point. The Laplace transform of MP is
n–1
1 2
= exp  --- u  ∏ ( 1 – 2λ j u )
–1 ⁄ 2
G MP (C.2)
2 
j=1
Using β=βFORM, the probability of the parabolic domain is
PP = P ( MP ≤ –β ) (C.3)
The density integral and the probability integral of MP are computed using the saddle point integration
method to evaluate the inversion integrals. Note that this procedure yields the true result for both the density
integral for MP and the probability integral for MP.
C2 Second Order Taylor Expansion

The distribution of the quadratic form
Proban Theory SESAM
C-2 01-OCT-2004 Program version 4.4
n–1
2
MT = ν + ∑ ( γj Yj + λj Yj ) (C.4)
j=1
is considered in Tvedt (1989), Ref. /36/. In Tvedt (1989), Ref. /36/, it is shown how to obtain the quadratic
form MT from the second order Taylor expansion of the u-space g-function in the vicinity of the design
point. The Laplace transform of MT is
2
 1 γj u 2 
n – 1 exp  --- --------------------
 2 1 – 2λ j u
GM = exp ( νu ) ∏ --------------------------------------- (C.5)
T 1⁄2
( 1 – 2λ j u )
j=1
Using β=βFORM, the probability of the quadratic domain is
PT = P ( M T ≤ –β ) (C.6)
The density integral and the probability integral of MT are computed using the saddle point integration
method to evaluate the inversion integrals, Rice (1980), Ref. /31/, and Helstrom (1983), Ref. /17/. Note that
this procedure yields the true result for both the density integral for MT and the probability integral for MT.
C3 Asymptotic Results
Two asymptotic results, i.e. as β→∞, Breitung (1984), Ref. /4/, and Hohenbichler (unpublished), are imple-
mented. Breitung's result reads
n–1
P P = Φ ( – β ) ∏ ( 1 + 2λ i β ) –1 ⁄ 2 (C.7)
i=1
Hohenbichler's result reads
n–1
–1 ⁄ 2
ϕ(β)
P P = Φ ( – β ) ∏  1 + 2λ i ---------------- (C.8)
 Φ ( – β )
i=1
SESAM Proban Theory
Program version 4.4 01-OCT-2004 D-1
APPENDIX D MULTINORMAL INTEGRAL
Whenever the probability of an intersection of single events is calculated using approximate methods, the
multinormal integral is involved, i.e. the cumulative probability of the multinormal integral must be calcu-
lated. Denoting the n integration limits as x, and the correlation matrix as C, the integral is
x1 x2 xn
1 1
- exp  – --- s C s ds 1 ds 2 … ds n
T –1
Φ ( x ;C ) = ∫ ∫ … ∫ ---------------------------------
( 2π )
n⁄2
C
1⁄2  2 
(D.1)
–∞ –∞ –∞
where |C| is the determinant of C. For dimensions n≥4, there is no tractable method available to compute
this integral with arbitrary precision, except for spacial cases, Madsen et. al. (1986), Ref. /24/, Chapter
4.5.1. However, an approximation method that gives good results also for higher dimensions is available,
Hohenbichler (1982), Ref. /19/, and Gollwitzer and Rackwitz (1988), Ref. /12/. The multinormal probability
can be expressed as the probability of the intersection of the n single linear events
 i 
 
 u ;g i ( u ) =

∑ Lij uj ≤ xi ;i = 1, 2, …, n (D.2)
 j=1 
The matrix L is the lower triangular Cholesky factor of C, i.e. LLT=C and U is a set of n independent stand-
ard normal random variables. The method is based on the following reduction formula, Hohenbichler
(1983), Ref. /19/:
 n   n 
P ∩ g i ( u ) ≤ 0 = Φ ( x 1 )P  ∩ g i ( ũ ) ≤ 0
  (D.3)
   
i = 1  i = 2 
where ũ i = u i ;i = 2, 3, …, n and
–1
u 1 = Φ ( Φ ( x 1 )Φ ( ũ 1 ) ) (D.4)
Proban Theory SESAM
D-2 01-OCT-2004 Program version 4.4
The idea is to approximate the probability of the reduced parallel system by a new multinormal integral of
dimension n−1, Φ(x1;C1), and a correction factor K:
 n 
P ∩ g i ( ũ ) ≤ 0 ≈ KΦ ( x ;C )
 1 1
(D.5)
 
i = 2 
using the SORM method for intersections, Gollwitzer and Rackwitz (1988), Ref. /12/. The elements of x1
are the new integration limits and the elements of C1 are the new correlations. Then the integral Φ(x1;C1) is
reduced the same way. The procedure is repeated until n=1.
SESAM Proban Theory
Program version 4.4 01-OCT-2004 E-1
APPENDIX E BOUNDS ON PROBABILITY OF A

UNION
E1 Union
Consider a union of n events, Ei, i=1,2,...,m, each occurring with probability Pi. The union of the events is
m
E = ∪ Ei (E.1)
i=1
and the probability is
 m 
P E = P  ∪ E i (E.2)
 
i = 1 
The probability PE is calculated from the probabilities Pi and the intersection probabilities Pij=P(Ei∩Ej) in
terms of bounds. A first order bound makes use of Pi only. A second order bound makes use also of Pij.
E2 Simple Bounds
The simple lower bound on PE, PE,L is the maximum probability of a single event
P E, L = max { P i ,i = 1, 2, …, m } (E.3)
The simple upper bound on PE, PE,U is the sum of the single event probabilities:
 m 
 
P E, U = min  ∑ P i ,1  (E.4)
 
i = 1 
Proban Theory SESAM
E-2 01-OCT-2004 Program version 4.4
E3 Ditlevsen Bounds
The Ditlevsen bounds, Ditlevsen (1979), Ref. /10/, Kounias (1968), Ref. /23/, are second order bounds, the
Ditlevsen lower bound on PE, PE,L is:
m  i–1 
 
P E, L = P 1 + ∑ max  0, P i – ∑ P ij  (E.5)
 
i=2  j=1 
The Ditlevsen upper bound on PE, PE,U is:
m
P E, U = P 1 + ∑ ( Pi – max { Pij ,j = 1, 2, …, i – 1 } ) (E.6)

i=2
E4 Reliability Index for Bounds

The lower bound on the reliability index is:
–1
β E, L = – Φ ( P E, U ) (E.7)
The lower bound on the reliability index is:

–1
β E, U = – Φ ( P E, L ) (E.8)
SESAM Proban Theory
Program version 4.4 01-OCT-2004 F-1
APPENDIX F TERMINOLOGY
F1 Introduction
The terminology used in Proban is given in the following sections. Appendix F 2 gives the terminology for
events. Appendix F 3 gives the terminology for variables. Appendix F 4 gives the terminology for measures
of probability. Appendix F 5 gives the terminology for sensitivity and importance. Appendix F 6 gives the
terminology for event approximations. Appendix F 7 gives the terminology for simulation methods. Appen-
dix F 8 gives the definition of statistical quantities used in this manual.
F2 Events
Event function The event function, G(x), defines the event G<0, the event
boundary G=0, and the complementary event G>0. The event
type associated with a given event function is often specified,
e.g. single event function.
Single event, SE The event is assigned a random variable, a threshold and a re-
lation of the variable to the threshold. The event function is
G(x) = z(x) – θ
The corresponding event is
E SE ( x ) = { x ;G ( x ) ≤ 0 }
Intersection of Single Events, ISE The event function value is the maximum of m single event
function values,
G ISE ( x ) = max { G SE, i ( x ) ;i = 1, 2, …, m }
The corresponding event is the intersection of single events

Proban Theory SESAM
F-2 01-OCT-2004 Program version 4.4
E ISE ( x ) = { x ;G ISE ( x ) ≤ 0 } = ∩ ESE, i ( x )

i=1
Union of Single Events, USE The event function value is the minimum of m single event
function values,
G USE ( x ) = min { G SE, i ( x ) ;i = 1, 2, …, m }
The corresponding event is the union of single events

m
E USE ( x ) = { x ;G USE ( x ) ≤ 0 } = ∪ ESE, i ( x )

i=1
Union of Intersections of Single Events, UISEThe event function value is the minimum of m intersection
event function values,
G UISE ( x ) = min { G ISE, i ( x ) ;i = 1, 2, …, m }
The corresponding event is the union of intersections of single

events
m
E UISE ( x ) = { x ;G UISE ( x ) ≤ 0 } = ∪ EISE, i ( x )

i=1
Union of Events, UE The event function value is the minimum of m subevent func-
tion values,
G UE ( x ) = min { G E, i ( x ) ;i = 1, 2, …, m }
The corresponding event is the union of subevents

m
E UE ( x ) = { x ;G UE ( x ) ≤ 0 } = ∪ E E, i ( x )
i=1
Intersection of Events, IE The event function value is the maximum of m subevent func-
tion values,
G IE ( x ) = max { G E, i ( x ) ;i = 1, 2, …, m }
The corresponding event is the intersection of subevents

m
E IE ( x ) = { x ;G IE ( x ) ≤ 0 } = ∩ E E, i ( x )
i=1
SESAM Proban Theory
F3 Variables
Random variable A variable which is uncertain.
Basic variable A variable in the model space.
Random basic variable A basic variable which is also a random variable.
Fixed variable A basic variable which is non-random.
Normalized random variable The standard normal random variable, u-space variable, which
the random basic variable is mapped onto, e.g. using the Rosen-
blatt transformation.
x-space The space of basic variables.
u-space The space of independent standard normal variables the basic

random variables are mapped onto.
v-space The space of dependent standard normal variables the basic

random variables are mapped onto by use of the Nataf model.
F4 Probability Measures
Event probability The probability that the event occur, i.e.
PE = ∫ f X ( x ) dx
G(x) ≤ 0
where G(x) is the event function and fX(x) is the joint distribu-
tion of the random basic variables.
Reliability The probability of the complementary event, i.e. QE=1−PE.
Reliability index The argument of the standard normal distribution correspond-

ing to the reliability, i.e.
–1 –1
βR = Φ ( QE ) = –Φ ( PE )
The reliability index is often indexed with the method used to

estimate βR, for example βFORM and βSORM.
F5 Parametric Sensitivity and Uncertainty Importance

Parametric sensitivity - probability The derivative of the probability with respect to a parameter:
Proban Theory SESAM
dP E
Pθ =
dθ
and similar for the reliability index.
Parametric sensitivity - moments The derivative of the four first moments of the distribution of a
random variable with respect to a parameter, i.e.
dγ
γθ =
dθ
where γ is one of µ, σ, δ and κ, i.e., the mean, standard devia-

tion, skewness and kurtosis.
Uncertainty importance The derivative of the reliability index with respect to the mean
of a u-space variable:
dβ R
aµ =
dµ
In Proban print the uncertainty importance is scaled to give an

impression of the percentage of the reliability index which
stems from the variable.
ã µ = 100a µ2
F6 Approximation of Event
Linearization point The point at which an event function is approximated by a lin-
ear/quadratic surface.
Design point The point on the event boundary which has the highest proba-
bility density in u-space. This is also the u-space event bound-
ary point with the minimum distance to the u-space origin.
Active single events Single events that are active at the design point.
Inactive single events Single events that are not active at the design point.
FORM First Order Reliability Method. The event boundaries of active

single events are linearized at the design point. Each inactive
single event is linearized (in succession) at the point of maxi-
mum likelihood on its event boundary, restricted to the event
defined by the previously linearized events.
SORM Second Order Reliability Method. As FORM, however, the

curvature of the event boundary in the vicinity of a design point
is accounted for through a quadratic form. (Parabolic form for
SESAM Proban Theory
intersections. Parabolic, diagonal or second order Taylor for

single events.)
F7 Simulation Methods
Monte Carlo Samples points according to the distribution of the basic ran-
dom variables. Probability is calculated from evaluation of an
indicator function at each sampled point. The density distribu-
tion is generated from saving the simulated values and ordering
them in an ascending values.
Design point Finds the design point. Then conducts a Monte Carlo sampling
in u-space around the u-space design point. Probability is cal-
culated from evaluation of a scaled indicator function at each
sampled point.
Axis-orthogonal Finds the design point. Defines an axis for a single event or a
small intersection event. Then conducts a Monte Carlo sam-
pling of points in the space orthogonal to the axis. Calculates a
correction to the FORM probability by adding up weighted cor-
rection probabilities for lines parallel to the axis through the
sampled points.
Directional Samples directions in u-space uniformly distributed on the sur-

face of the unit hyper-sphere centred at the u-space origin. A
weighted sum of the event probabilities conditioned on the di-
rections sampled is accumulated.
Latin-hypercube A stratified sampling scheme where each variable is sampled m

times and these outcomes are randomly combined to form m
samples.
F8 Statistical Quantities
Table explains some basic statistical quantities.
The density distribution of the random vari-

Density f( x)
able X.
x
The probability distribution of the random
Cumulative F( x)= ∫ f ( s ) ds variable X.
–∞
∞
n n
Moment E(( X – a) )= ∫ ( s – a ) f ( s ) ds The nth moment of X about a.
–∞
Proban Theory SESAM
Mean µ = E( X) The expectation or average of X.
Median m The value x so that F ( x ) = 0.5
2 0.5 Measures the spread of the distribution of X

Standard Deviation σ = E(( X – µ) ) about µ.
σ
Coefficient of Variation cov = ------ ,µ ≠ 0
µ
3 Measures the symmetry of the distribution

E ( ( X – µ ) -) of X about µ. The Normal distribution has
Skewness δ = ----------------------------
3
σ skewness equal to 0.
4 Measures the peakedness of the distribution
E ( ( X – µ ) -) of X about µ. The Normal distribution has
Kurtosis κ = ----------------------------
4
σ kurtosis equal to 3.
SESAM Proban Theory
Program version 4.4 01-OCT-2004 REFERENCES-1
REFERENCES
1 Bjerager, P. and Krenk, S.

‘Sensitivity measures in structural reliability analysis.’
Proceedings of the 1st IFIP Working Conference on Reliability and Optimization on Structural Systems.
P. Thoft-Christensen, ed., Springer Verlag, Berlin, pp 459-470, 1987
2 Bjerager, P.
‘Probability integration by directional simulation.’
Journal of Engineering Mechanics.
ASCE, Vol. 114(8), pp. 1285-1302, 1988
3 Bjerager, P. and Krenk, S.

‘Parametric sensitivity in first order reliability analysis.’
ASCE, Vol. 115(7), 1989
4 Breitung, K.
‘Asymptotic approximation for multinormal integrals.’
ASCE, Vol. 110(3), pp. 357-366, 1984
5 Cornell, C. A.
‘A probability based structural code.’
Journal of the American Concrete Institute.
Vol. 66(12), pp. 974-985, 1969
6 Deak, I.
‘Three digit accurate multiple normal probabilities.’
Numerische Mathematik.
Vol. 35, pp. 369-380, 1980
7 Der Kiureghian, A. and Liu, P.-L.

‘Structural reliability under incomplete probability information.’
ASCE, Vol. 112(1), pp. 85-104, 1987
8 Der Kiureghian, A., Lin, H.-Z. and Hwang, S.J.

‘Second order reliability approximations.’
Proban Theory SESAM
REFERENCES-2 01-OCT-2004 Program version 4.4

ASCE, Vol. 113(8), pp. 1208-1225, 1987
9 Ditlevsen, O.
‘Generalized second moment reliability index.’
Journal of Structural Mechanics.
ASCE, Vol. 7, pp. 435-451, 1979
10 Ditlevsen, O.
‘Narrow reliability bounds for structural systems.’
Journal of Structural Mechanics.
ASCE, Vol. 7, pp. 453-472, 1979
11 Fiessler, B., Neumann, H.-J. and Rackwitz, R.

‘Quadratic limit states in structural reliability.’
Journal of the Engineering Mechanics Division.
ASCE, Vol. 105, pp. 661-676, 1979.
12 Gollwitzer, S. and Rackwitz, R.

‘An efficient solution to the multinormal integral.’
Probabilistic Engineering Mechanics.
Vol. 3(2), pp. 98-101, 1988.
13 Hagen, Ø. and Tvedt, L.

‘Vector process out-crossing as parallel system sensitivity measure.’
ASCE, Vol. 117(10), pp. 2201-2220, 1991.
14 Hagen, Ø., Sigurdson, G., Tvedt, L., and Cramer, E.H.

‘Reliability Updating Based on Equality Constraints.’
Proceedings of the Fifth (1995) Offshore and Polar Engineering Conference.
The Hague, The netherlands, pp. 361-364, 1995.
15 Harbitz, A.
‘Efficient and accurate probability of failure calculation by use of the importance sampling technique.’
Proceedings of ICASP4.
Firenze, Italy, pp. 825-836, 1983.
16 Hasofer, A. M. and Lind, N. C.

‘Exact and invariant second moment code format.’
SESAM Proban Theory
ASCE, Vol. 100, pp. 111-121, 1974.
17 Helstrom, C. W.
‘Comment: Distribution of quadratic forms in normal random variables- evaluation by numerical integra-
tion.’
SIAM, J. Sci. Stat. Comput.
Vol. 4(2), pp. 353-356, 1983.
18 Hohenbichler, M. and Rackwitz, R.

‘Non-normal dependent vectors in structural reliability.’
ASCE, Vol. 107, pp. 1227-1238, 1981.
19 Hohenbichler, M.
‘An approximation to the multivariate normal distribution.’
Proceedings, EUROMECH 155, Reliability Theory of Structural Engineering Systems.
DIALOG 6-86, Danish Engineering Academy, Denmark, pp. 79-100, 1982.
20 Hohenbichler, M.
‘An asymptotic formula for the probability of intersections.’
Berichte zu Zuverlassigkeitstheorie der Bauwerke, SFB 96.
Technical University of Munich, Heft 69, pp. 21-48.

‘Sensitivity and importance measures in structural reliability.’
Civil Engineering Systems.
Vol. 3(4), pp. 203-209, 1986.

‘Improvements of second order reliability estimates by importance sampling.’
ASCE, Vol. 114(12), pp. 2195-2199, 1988.
23 Kounias, E.
‘Bounds for the probability of a union with applications.’
The Annals of Mathematical Statistics.
Vol. 39, p. 2154, 1968.
24 Madsen, H. O., Krenk, S. and Lind, N. C.

Methods of Structural Safety.
Prentice-Hall, Inc., Englewood Cliffs, N. J.
Proban Theory SESAM
25 Madsen, H. O.
‘Model updating in reliability theory.’
Reliability and Risk in Civil Engineering.
Vol. 1, ICASP5, N. C. Lind, ed., University of Waterloo, Ontario, Canada, pp. 564-577, 1987
26 Madsen, H. O.
‘Omission sensitivity factors.’
Structural Safety.
Vol. 5, pp. 35-45, 1988
27 Madsen, H. O.
‘Sensitivity factors for parallel systems.’
Technical note, Danish Engineering Academy, Denmark, 1988
28 McKay, M. D., Beckman, R. J. and Conover, W. J.

‘A comparison of three methods for selecting values of input variables in the analysis of output from a
computer code.’
Technometrics.
Vol. 21(2), 1979
29 Nataf, A.
‘Determination des distribution dont les marges sont donnees.’
Rendus l’Academie des Sciences.
Vol. 225, Paris, France, pp. 42-43, 1962
30 Paloheimo, E. and Hannus, M.

‘Structural design based on weighted fractiles.’
Journal of the Structural Division.
ASCE, Vol. 100(7), pp. 1367-1378, 1974
31 Rice, S. O.
‘Distribution of quadratic forms in normal random variables- evaluation by numerical integration.’
SIAM, J. Sci. Stat. Comput.
Vol. 1(4), pp. 438-448, 1980.
32 Rosenblatt, M.
‘Remarks on a multivariate transformation.’
The Annals of Mathematical Statistics.
Vol. 23, pp. 470-472, 1952.
33 Schall, G., Gollwitzer, S. and Rackwitz, R.

‘Integration of multinormal densities on surfaces.’
SESAM Proban Theory
Proceedings of the 2nd IFIP Working Conference on Reliability and Optimization on Structural Systems.
P. Thoft-Christensen, ed., Springer Verlag, 1988.
34 Shinozuka, M.
‘Basic analysis of structural safety.’
Journal of Structural Engineering.
ASCE, Vol 109(3), pp. 721-740, 1983
35 Tvedt, L.
‘Second order reliability by an exact integral.’
Proceedings of the 2nd IFIP Working Conference on Reliability and Optimization on Structural Systems.
P. Thoft-Christensen, ed., Springer Verlag, 1988.
36 Tvedt, L.
‘Distribution of quadratic forms in normal space - application to structural reliability.’
ASCE, Vol. 116(6), pp. 1183-1197, 1988.
37 Wen, Y. K. and Chen, H.-C.

‘On fast integration for time variant structural reliability.’
Probabilistic Engineering Mechanics.
Vol. 2(3), pp. 156-162, 1987.
38 Winterstein, S. R., De, R. S. and Bjerager, P.

‘Correlated non-Gaussian models in offshore structural reliability.’
Proceedings, 5th International Conference on Structural Safety and Reliability.
San Francisco, California, 1989.
39 Wu, Y.-T., and Wirsching, P. H.

‘A new algorithm for structural reliability estimation.’
ASCE, Vol. 113, pp. 1319-1336, 1987.
Proban Theory SESAM

Proban - TM - Manual de Probabilidade - Veritas

Uploaded by

Copyright:

Available Formats

Proban - TM - Manual de Probabilidade - Veritas

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Proban - TM - Manual de Probabilidade - Veritas

Uploaded by

Copyright:

Available Formats

SESAM

General Purpose Probabilistic Analysis Program

DET NORSKE VERITAS

October 1st, 2004

Valid from program version 4.4

Developed and marketed by

Copyright © 2004 Det Norske Veritas

1.1 Proban in the SESAM System ......................................................................................................... 1-2

2 PROBABILISTIC MODEL - VARIABLES..................................................................2-1

2.1 Definition of Variables .................................................................................................................... 2-1

3 PROBABILISTIC MODEL - EVENTS .........................................................................3-1

3.1 Definition of Event .......................................................................................................................... 3-1

4.1 General............................................................................................................................................. 4-1

5 GUIDE TO METHODS................................................................................................... 5-1

5.1 Introduction...................................................................................................................................... 5-1

6.1 Motivation........................................................................................................................................ 6-1

7 FIRST ORDER RELIABILITY METHOD, FORM ....................................................7-1

7.1 Motivation........................................................................................................................................ 7-1

8 SECOND ORDER RELIABILITY METHOD, SORM ...............................................8-1

8.1 Motivation........................................................................................................................................ 8-1

9 NESTED RELIABILITY ANALYSIS ...........................................................................9-1

9.1 Method ............................................................................................................................................. 9-1

11 FIRST PASSAGE PROBABILITY.............................................................................. 11-1

12 APPLICATIONS INVOLVING EQUALITY EVENTS ............................................ 12-1

13 APPLICATIONS INVOLVING MEASURED VALUES .......................................... 13-1

14 MEAN VALUE FORM FOR DISTRIBUTIONS ....................................................... 14-1

15 GENERATED DISTRIBUTION .................................................................................. 15-1

16 DESIGN POINT SIMULATION .................................................................................. 16-1

16.1 Motivation...................................................................................................................................... 16-1

17 MONTE-CARLO SIMULATION ................................................................................ 17-1

17.1 Simulation of Probability............................................................................................................... 17-1

18 DIRECTIONAL SIMULATION .................................................................................. 18-1

18.1 Motivation...................................................................................................................................... 18-1

19 AXIS-ORTHOGONAL SIMULATION ...................................................................... 19-1

19.1 Motivation...................................................................................................................................... 19-1

20.1 Method ........................................................................................................................................... 20-1

21 PARAMETRIC SENSITIVITY FOR PROBABILITY AND CROSSING RATE..21-1

21.1 Introduction.................................................................................................................................... 21-1

22 PARAMETRIC SENSITIVITY FOR DISTRIBUTION ...........................................22-1

APPENDIX A OPTIMIZATION.......................................................................................... A-1

APPENDIX B CODING OF A FUNCTION ....................................................................... B-1

APPENDIX C QUADRATIC FORMS IN NORMAL SPACE.......................................... C-1

APPENDIX E BOUNDS ON PROBABILITY OF A UNION........................................... E-1

1.1 Proban in the SESAM System

Figure 1.1 SESAM overview

1.2 How to read the Manual

• Chapter 4 describes the results that Proban delivers.

• Chapter 5 is a guide to the methods.

• Chapter 7 to Chapter 22 describes the methods implemented in Proban.

• Appendix A to Appendix E gives more specialized implementation notes.

• Appendix F tabulates the terminology used in this manual.

• Appendix REFERENCES lists the literature referenced in this manual.

2 PROBABILISTIC MODEL - VARIABLES

2.1 Definition of Variables

Variables are created by using the command CREATE VARIABLE.

Variables are changed by using the command CHANGE VARIABLE.

Variables are deleted by using the command DELETE VARIABLE.

2.2 Types of Variables

A distribution variable may be assigned to parameters of any other variable.