Papers About Extreme Value and Crossing Levels by Arvid Naess

Hindawi Publishing Corporation
Journal of Probability and Statistics

Volume 2013, Article ID 797014, 15 pages
http://dx.doi.org/10.1155/2013/797014
Research Article
Estimation of Extreme Values by the Average Conditional
Exceedance Rate Method
A. Naess,1 O. Gaidai,2 and O. Karpa3
1
Department of Mathematical Sciences and CeSOS, Norwegian University of Science and Technology, 7491 Trondheim, Norway
Norwegian Marine Technology Research Institute, 7491 Trondheim, Norway
3
Centre for Ships and Ocean Structures (CeSOS), Norwegian University of Science and Technology, 7491 Trondheim, Norway
2
Correspondence should be addressed to A. Naess; arvidn@math.ntnu.no

Received 18 October 2012; Revised 22 December 2012; Accepted 9 January 2013
Academic Editor: A. Thavaneswaran
Copyright 2013 A. Naess et al. This is an open access article distributed under the Creative Commons Attribution License, which
permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
This paper details a method for extreme value prediction on the basis of a sampled time series. The method is specifically designed
to account for statistical dependence between the sampled data points in a precise manner. In fact, if properly used, the new method
will provide statistical estimates of the exact extreme value distribution provided by the data in most cases of practical interest. It
avoids the problem of having to decluster the data to ensure independence, which is a requisite component in the application of, for
example, the standard peaks-over-threshold method. The proposed method also targets the use of subasymptotic data to improve
prediction accuracy. The method will be demonstrated by application to both synthetic and real data. From a practical point of
view, it seems to perform better than the POT and block extremes methods, and, with an appropriate modification, it is directly
applicable to nonstationary time series.
1. Introduction
Extreme value statistics, even in applications, are generally
based on asymptotic results. This is done either by assuming
that the epochal extremes, for example, yearly extreme wind
speeds at a given location, are distributed according to the
generalized (asymptotic) extreme value distribution with
unknown parameters to be estimated on the basis of the observed data [1, 2]. Or it is assumed that the exceedances above
high thresholds follow a generalized (asymptotic) Pareto
distribution with parameters that are estimated from the data
[14]. The major problem with both of these approaches is
that the asymptotic extreme value theory itself cannot be used
in practice to decide to what extent it is applicable for the
observed data. And since the statistical tests to decide this
issue are rarely precise enough to completely settle this problem, the assumption that a specific asymptotic extreme value
distribution is the appropriate distribution for the observed
data is based more or less on faith or convenience.
On the other hand, one can reasonably assume that in
most cases long time series obtained from practical measurements do contain values that are large enough to provide
useful information about extreme events that are truly

asymptotic. This cannot be strictly proved in general, of
course, but the accumulated experience indicates that asymptotic extreme value distributions do provide reasonable, if not
always very accurate, predictions when based on measured
data. This is amply documented in the vast literature on the
subject, and good references to this literature are [2, 5, 6]. In
an effort to improve on the current situation, we have tried to
develop an approach to the extreme value prediction problem
that is less restrictive and more flexible than the ones based
on asymptotic theory. The approach is based on two separate
components which are designed to improve on two important
aspects of extreme value prediction based on observed data.
The first component has the capability to accurately capture
and display the effect of statistical dependence in the data,
which opens for the opportunity of using all the available data
in the analysis. The second component is then constructed
so as to make it possible to incorporate to a certain extent
also the subasymptotic part of the data into the estimation
of extreme values, which is of some importance for accurate
estimation. We have used the proposed method on a wide
variety of estimation problems, and our experience is that
it represents a very powerful addition to the toolbox of

methods for extreme value estimation. Needless to say, what
is presented in this paper is by no means considered a closed
chapter. It is a novel method, and it is to be expected that
several aspects of the proposed approach will see significant
improvements.
where := means by definition, the following one-step memory approximation will, to a certain extent, account for the
dependence between the s,
Prob { | 1 , . . . , 1 }
Prob { | 1 } ,
2. Cascade of Conditioning Approximations
(3)
for 2 . With this approximation, it is obtained that
In this section, a sequence of nonparametric distribution

functions will be constructed that converges to the exact
extreme value distribution for the time series considered. This
constitutes the core of the proposed approach.
Consider a stochastic process (), which has been
observed over a time interval, (0, ) say. Assume that values
1 , . . . , , which have been derived from the observed
process, are allocated to the discrete times 1 , . . . , in (0, ).
This could be simply the observed values of () at each
, = 1, . . . , , or it could be average values or peak
values over smaller time intervals centered at the s. Our
goal in this paper is to accurately determine the distribution
function of the extreme value = max{ ; = 1, . . . , }.
Specifically, we want to estimate () = Prob( )
accurately for large values of . An underlying premise for the
development in this paper is that a rational approach to the
study of the extreme values of the sampled time series is to
consider exceedances of the individual random variables
above given thresholds, as in classical extreme value theory.
The alternative approach of considering the exceedances by
upcrossing of given thresholds by a continuous stochastic
process has been developed in [7, 8] along lines similar to that
adopted here. The approach taken in the present paper seems
to be the appropriate way to deal with the recorded data time
series of, for example, the hourly or daily largest wind speeds
observed at a given location.
From the definition of () it follows that
() 2 ()
:= Prob { | 1 } Prob (1 ) .
=2
(4)
By conditioning on one more data point, the one-step memory approximation is extended to
Prob { | 1 , . . . , 1 }
Prob { | 1 , 2 } ,
(5)
where 3 , which leads to the approximation
() 3 () := Prob { | 1 , 2 }
=3
Prob {2 | 1 } Prob (1 ) .
(6)
For a general , 2 , it is obtained that
() ()
:= Prob { | 1 , . . . , +1 }
=
() = Prob ( ) = Prob { , . . . , 1 }
= Prob { | 1 , . . . , 1 }
Prob {1 , . . . , 1 }
Prob { | 1 . . . , 1 }
=2
(1)
= Prob { | 1 , . . . , 1 }
=2
Prob (1 ) .
In general, the variables are statistically dependent.
Hence, instead of assuming that all the are statistically
independent, which leads to the classical approximation
() 1 () := Prob ( ) ,
=1
(7)
Prob (1 ) ,
where () = ().
It should be noted that the one-step memory approximation adopted above is not a Markov chain approximation
[911], nor do the -step memory approximations lead to
th-order Markov chains [12, 13]. An effort to relinquish the
Markov chain assumption to obtain an approximate distribution of clusters of extremes is reported in [14].
It is of interest to have a closer look at the values for ()
obtained by using (7) as compared to (2). Now, (2) can be
rewritten in the form
(2)
() 1 () = (1 1 ()) ,
=1
(8)
where 1 () = Prob{ > }, = 1, . . . , . Then the approximation based on assuming independent data can be written
as
() 1 () := exp ( 1 ()) ,
(9)
=1
() () = (1 ()) (1 ()) ,
(10)
=1
where () = Prob{ > | 1 , . . . , +1 }, for

2, denotes the exceedance probability conditional on
1 previous nonexceedances. From (10) it is now obtained
that
=1
() () := exp ( () ()) ,
(11)
,
and () () as with () = () for .
For the cascade of approximations () to have practical
significance, it is implicitly assumed that there is a cut-off
value satisfying such that effectively () =
(). It may be noted that for -dependent stationary data
sequences, that is, for data where and are independent
whenever | | > , then () = +1 () exactly, and, under
rather mild conditions on the joint distributions of the data,
lim 1 () = lim () [15]. In fact, it can be shown
that lim 1 () = lim () is true for weaker conditions than -dependence [16]. However, for finite values of
the picture is much more complex, and purely asymptotic
results should be used with some caution. Cartwright [17]
used the notion of -dependence to investigate the effect on
extremes of correlation in sea wave data time series.
Returning to (11), extreme value prediction by the conditioning approach described above reduces to estimation of
(combinations) of the () functions. In accordance with
the previous assumption about a cut-off value , for all values of interest, , so that 1
=1 () is effectively
negligible compared to
= (). Hence, for simplicity, the
following approximation is adopted, which is applicable to
both stationary and nonstationary data,
() = exp ( ()) ,
The concept of average conditional exceedance rate (ACER)

of order is now introduced as follows:
() =
Alternatively, (7) can be expressed as,
3. Empirical Estimation of the Average

Conditional Exceedance Rates
1.
(12)
Going back to the definition of 1 (), it follows that
=1 1 () is equal to the expected number of exceedances

of the threshold during the time interval (0, ). Equation
(9) therefore expresses the approximation that the stream of
exceedance events constitute a (nonstationary) Poisson process. This opens for an understanding of (12) by interpreting
the expressions
= () as the expected effective number
of independent exceedance events provided by conditioning
on 1 previous observations.
1
() ,
+ 1 =
= 1, 2, . . . .
(13)
In general, this ACER function also depends on the number

of data points .
In practice, there are typically two scenarios for the
underlying process (). Either we may consider to be a
stationary process, or, in fact, even an ergodic process. The
alternative is to view () as a process that depends on certain
parameters whose variation in time may be modelled as an
ergodic process in its own right. For each set of values of the
parameters, the premise is that () can then be modelled as
an ergodic process. This would be the scenario that can be
used to model long-term statistics [18, 19].
For both these scenarios, the empirical estimation of the
ACER function () proceeds in a completely analogous
way by counting the total number of favourable incidents,
that is, exceedances combined with the requisite number of
preceding nonexceedances, for the total data time series and
then finally dividing by + 1 . This can be shown to
apply for the long-term situation.
A few more details on the numerical estimation of ()
for 2 may be appropriate. We start by introducing the
following random functions:
() = 1 { > , 1 , . . . , +1 } ,
= , . . . , , = 2, 3, . . . ,
() = 1 {1 , . . . , +1 } ,
(14)
= , . . . , , = 2, . . . ,
where 1{A} denotes the indicator function of some event A.
Then,
() =
E [ ()]
E [ ()]
= , . . . , , = 2, . . . ,
(15)
where E[] denotes the expectation operator. Assuming an

ergodic process, then obviously () = () = =
(), and by replacing ensemble means with corresponding time averages, it may be assumed that for the time series
at hand
() = lim
= ()
()
(16)
where () and () are the realized values of () and

(), respectively, for the observed time series.
Clearly, lim E[ ()] = 1. Hence, lim ()/
() = 1, where
() =
= E [ ()]
+1
(17)
The advantage of using the modified ACER function () for

2 is that it is easier to use for nonstationary or longterm statistics than (). Since our focus is on the values of
the ACER functions at the extreme levels, we may use any
function that provides correct predictions of the appropriate
ACER function at these extreme levels.
To see why (17) may be applicable for nonstationary time
series, it is recognized that
E [ ()]
E [ ()]
() exp ( ()) = exp (
)
(18)
exp ( E [ ()]) .
If the time series can be segmented into blocks, such that

E[ ()] remains approximately constant within each block
and such that E[ ()] () for a sufficient
range of -values, where denotes the set of indices for block
no. , = 1, . . . , , then
= E[ ()] = (). Hence,
(19)
where
() =
1
() .
+ 1 =
() =
2
1 ()
( () ()) ,
1 =1
(20)
It is of interest to note what events are actually counted

for the estimation of the various (), 2. Let us
start with 2 (). It follows from the definition of 2 () that
2 () ( 1) can be interpreted as the expected number of
exceedances above the level , satisfying the condition that an
exceedance is counted only if it is immediately preceded by a
non-exceedance. A reinterpretation of this is that 2 () (
1) equals the average number of clumps of exceedances
above , for the realizations considered, where a clump of
exceedances is defined as a maximum number of consecutive
exceedances above . In general, () ( + 1) then
equals the average number of clumps of exceedances above
separated by at least 1 nonexceedances. If the time
series analysed is obtained by extracting local peak values
from a narrow band response process, it is interesting to
note the similarity between the ACER approximations and
the envelope approximations for extreme value prediction
[7, 20]. For alternative statistical approaches to account for
the effect of clustering on the extreme value distribution, the
reader may consult [2126]. In these works, the emphasis is
on the notion of an extremal index, which characterizes the
clumping or clustering tendency of the data and its effect on
the extreme value distribution. In the ACER functions, these
effects are automatically accounted for.
Now, let us look at the problem of estimating a confidence
interval for (), assuming a stationary time series. If
realizations of the requisite length of the time series is
available, or, if one long realization can be segmented into
(21)
where () () denotes the ACER function estimate from realization no. , and () = =1 () ()/.
Assuming that realizations are independent, for a suitable
number , for example, 20, (21) leads to a good approximation of the 95% confidence interval CI = ( (), + ())
for the value (), where
() = ()
() exp ( ( + 1) ()) ,
subseries, then the sample standard deviation () can be

estimated by the standard formula
1.96 ()
.
(22)
Alternatively, and which also applies to the non-stationary case, it is consistent with the adopted approach to assume
that the stream of conditional exceedances over a threshold
constitute a Poisson process, possibly non-homogeneous.
Hence, the variance of the estimator () of (), where
() =
= ()
(23)
+1
is Var[ ()] = (). Therefore, for high levels , the approximate limits of a 95% confidence interval of (), and also
(), can be written as
() = () (1
1.96
( + 1) ()
).
(24)
4. Estimation of Extremes for the Asymptotic

Gumbel Case
The second component of the approach to extreme value
estimation presented in this paper was originally derived for
a time series with an asymptotic extreme value distribution
of the Gumbel type, compared with [27]. We have therefore
chosen to highlight this case first, also because the extension
of the asymptotic distribution to a parametric class of extreme
value distribution tails that are capable of capturing to some
extent subasymptotic behaviour is more transparent, and
perhaps more obvious, for the Gumbel case. The reason
behind the efforts to extend the extreme value distributions to
the subasymptotic range is the fact that the ACER functions
allow us to use not only asymptotic data, which is clearly an
advantage since proving that observed extremes are truly
asymptotic is really a nontrivial task.
The implication of the asymptotic distribution being of
the Gumbel type on the possible subasymptotic functional
forms of () cannot easily be decided in any detail. However,
using the asymptotic form as a guide, it is assumed that
the behaviour of the mean exceedance rate in the tail is
dominated by a function of the form exp{( ) } (
1 ), where , , and are suitable constants, and 1 is an
appropriately chosen tail marker. Hence, it will be assumed

that,
() = () exp { ( ) } ,
1 ,
(25)
where the function () is slowly varying, compared with

the exponential function exp{ ( ) } and , , and
are suitable constants, that in general will be dependent on .
Note that the value = () = 1 corresponds to the asymptotic Gumbel distribution, which is then a special case of the
assumed tail behaviour.
From (25) it follows that
()
) = log ( ) log ( ) .
log log (
()
(26)
Therefore, under the assumptions made, a plot of
log | log( ()/ ())| versus log( ) will exhibit a
perfectly linear tail behaviour.
It is realized that if the function () could be replaced
by a constant value, say , one would immediately be in a
position to apply a linear extrapolation strategy for deep tail
prediction problems. In general, () is not constant, but its
variation in the tail region is often sufficiently slow to allow
for its replacement by a constant, possibly by adjusting the tail
marker 1 . The proposed statistical approach to the prediction
of extreme values is therefore based on the assumption that
we can write,
() = exp { ( ) } ,
1 ,
(27)
where , , , and are appropriately chosen constants. In

a certain sense, this is a minimal class of parametric functions
that can be used for this purpose which makes it possible to
achieve three important goals. Firstly, the parametric class
contains the asymptotic form given by = = 1 as a
special case. Secondly, the class is flexible enough to capture,
to a certain extent, subasymptotic behaviour of any extreme
value distribution, that is, asymptotically Gumbel. Thirdly,
the parametric functions agree with a wide range of known
special cases, of which a very important example is the
extreme value distribution for a regular stationary Gaussian
process, which has = 2.
The viability of this approach has been successfully demonstrated by the authors for mean up-crossing rate estimation
for extreme value statistics of the response processes related
to a wide range of different dynamical systems, compared
with [7, 8].
As to the question of finding the parameters , , ,
(the subscript , if it applies, is suppressed), the adopted
approach is to determine these parameters by minimizing the
following mean square error function, with respect to all four
arguments,
2
(, , , ) = log ( ) log + ( ) , (28)
=1
where 1 < < denotes the levels where the ACER function has been estimated, denotes a weight factor that
puts more emphasis on the more reliably estimated ( ).

The choice of weight factor is to some extent arbitrary. We
have previously used = (log + ( ) log ( )) with
= 1 and 2, combined with a Levenberg-Marquardt least
squares optimization method [28]. This has usually worked
well provided reasonable and initial values for the parameters
were chosen. Note that the form of puts some restriction
on the use of the data. Usually, there is a level beyond which
is no longer defined, that is, ( ) becomes negative.
Hence, the summation in (28) has to stop before that happens.
Also, the data should be preconditioned by establishing the
tail marker 1 based on inspection of the empirical ACER
functions.
In general, to improve robustness of results, it is recommended to apply a nonlinearly constrained optimization [29].
The set of constraints is written as
log ( ) 0,
0 < < +,
min < 1 ,
(29)
0 < < +,
0 < < 5.
Here, the first nonlinear inequality constraint is evident, since

under our assumption we have ( ) = exp{( ) }, and
( ) < 1 by definition.
A Note of Caution. When the parameter is equal to 1.0 or
close to it, that is, the distribution is close to the Gumbel
distribution, the optimization problem becomes ill-defined
or close to ill-defined. It is seen that when = 1.0, there is
an infinity of (, ) values that gives exactly the same value
of (, , , ). Hence, there is no well-defined optimum in
parameter space. There are simply too many parameters. This
problem is alleviated by fixing the -value, and the obvious
choice is = 1.
Although the Levenberg-Marquardt method generally
works well with four or, when appropriate, three parameters,
we have also developed a more direct and transparent
optimization method for the problem at hand. It is realized
by scrutinizing (28) that if and are fixed, the optimization
problem reduces to a standard weighted linear regression
problem. That is, with both and fixed, the optimal values
of and log are found using closed form weighted linear
regression formulas in terms of , = log ( ) and =
( ) . In that light, it can also be concluded that the best
2
linear unbiased estimators (BLUE) are obtained for =
,
2
where = Var[ ] (empirical) [30, 31]. Unfortunately, this
is not a very practical weight factor for the kind of problem
we have here because the summation in (28) then typically
would have to stop at undesirably small values of .
It is obtained that the optimal values of and are given

by the relations
(, ) =
=1 ( ) ( )
2
=1 ( )
,
(30)
log (, ) = + (, ) ,
5. Estimation of Extremes for the General Case

For independent data in the general case, the ACER function
1 () can be expressed asymptotically as
1/
1/
() = () [1 + ( ( ) )]
where =
=1 / =1 , with a similar definition of .
To calculate the final optimal set of parameters, one
may use the Levenberg-Marquardt method on the function
) = ( (, ), , , (, )) to find the optimal values
(,
and , and then use (30) to calculate the corresponding
and .
For a simple construction of a confidence interval for
the predicted, deep tail extreme value given by a particular
ACER function as provided by the fitted parametric curve, the
empirical confidence band is reanchored to the fitted curve
by centering the individual confidence intervals CI0.95 for the
point estimates of the ACER function on the fitted curve.
Under the premise that the specified class of parametric
curves fully describes the behaviour of the ACER functions in
the tail, parametric curves are fitted as described above to the
boundaries of the reanchored confidence band. These curves
are used to determine a first estimate of a 95% confidence
interval of the predicted extreme value. To obtain a more
precise estimate of the confidence interval, a bootstrapping
method would be recommended. A comparison of estimated
confidence intervals by both these methods will be presented
in the section on extreme value prediction for synthetic data.
As a final point, it has been observed that the predicted value
is not very sensitive to the choice of 1 , provided it is chosen
with some care. This property is easily recognized by looking
at the way the optimized fitting is done. If the tail marker is
in the appropriate domain of the ACER function, the optimal
fitted curve does not change appreciably by moving the tail
marker.
1 () [1 + ( ( ))]
of approximation, the behaviour of the mean exceedance rate

in the subasymptotic part of the tail is assumed to follow a
1/
( 1
function largely of the form [1 + (( ) )]
), where > 0, , > 0, and > 0 are suitable constants,
and 1 is an appropriately chosen tail level. Hence, it will be
assumed that [32]
(31)
where > 0, , are constants. This follows from the

explicit form of the so-called generalized extreme value
(GEV) distribution Coles [1].
Again, the implication of this assumption on the possible
subasymptotic functional forms of () in the general case
is not a trivial matter. The approach we have chosen is to
assume that the class of parametric functions needed for
the prediction of extreme values for the general case can be
modelled on the relation between the Gumbel distribution
and the general extreme value distribution. While the extension of the asymptotic Gumbel case to the proposed class of
subasymptotic distributions was fairly transparent, this is not
equally so for the general case. However, using a similar kind
1 , (32)
where the function () is weakly varying, compared with

1/
the function [1 + ( ( ) )] and > 0, , > 0
and > 0 are suitable constants, that in general will be
dependent on . Note that the values = 1 and () = 1 corresponds to the asymptotic limit, which is then a special case
of the general expression given in (25). Another method to
account for subasymptotic effects has recently been proposed
by Eastoe and Tawn [33], building on ideas developed by
Tawn [34], Ledford and Tawn [35] and Heffernan and Tawn
[36]. In this approach, the asymptotic form of the marginal
distribution of exceedances is kept, but it is modified by a
multiplicative factor accounting for the dependence structure
of exceedances within a cluster.
An alternative form to (32) would be to assume that
1/
() = [1 + ( ( ) + ())]
1 ,
(33)
where the function () is weakly varying compared with

the function ( ) . However, for estimation purposes, it
turns out that the form given by (25) is preferable as it leads to
simpler estimation procedures. This aspect will be discussed
later in the paper.
For practical identification of the ACER functions given
by (32), it is expedient to assume that the unknown function
() varies sufficiently slowly to be replaced by a constant.
In general, () is not constant, but its variation in the
tail region is assumed to be sufficiently slow to allow for its
replacement by a constant. Hence, as in the Gumbel case, it
is in effect assumed that () can be replaced by a constant
for 1 , for an appropriate choice of tail marker 1 . For
simplicity of notation, in the following we will suppress the
index on the ACER functions, which will then be written as

() = [1 + ( ) ] ,
1 ,
(34)
where = 1/, = .
For the analysis of data, first the tail marker 1 is
provisionally identified from visual inspection of the log
plot (, ln ()). The value chosen for 1 corresponds to the
beginning of regular tail behaviour in a sense to be discussed
below.
The optimization process to estimate the parameters is
done relative to the log plot, as for the Gumbel case. The mean
square error function to be minimized is in the general case
written as
(
, , , , ) = log ( ) log
(35)
=1
2
+ log [1 + ( ) ] ,
where is a weight factor as previously defined.
An option for estimating the five parameters , , ,

, is again to use the Levenberg-Marquardt least squares
optimization method, which can be simplified also in this
case by observing that if , , and are fixed in (28), the
optimization problem reduces to a standard weighted linear
regression problem. That is, with , , and fixed, the optimal
values of and log are found using closed form weighted
linear regression formulas in terms of , = log ( ) and
= 1 + ( ) .
It is obtained that the optimal values of and log
are given by relations similar to (30). To calculate the final
optimal set of parameters, the Levenberg-Marquardt meth , , ) =
od may then be used on the function (
, , ), (
, , )) to find the optimal values ,
(
, , , (
, and , and then the corresponding and can be
calculated. The optimal values of the parameters may, for
example, also be found by a sequential quadratic programming (SQP) method [37].
6. The Gumbel Method

To offer a comparison of the predictions obtained by the
method proposed in this paper with those obtained by other
methods, we will use the predictions given by the two methods that seem to be most favored by practitioners, the Gumbel
method and the peaks-over-threshold (POT) method, provided, of course, that the correct asymptotic extreme value
distribution is of the Gumbel type.
The Gumbel method is based on recording epochal
extreme values and fitting these values to a corresponding
Gumbel distribution [38]. By assuming that the recorded
extreme value data are Gumbel distributed, then representing
the obtained data set of extreme values as a Gumbel probability plot should ideally result in a straight line. In practice, one
cannot expect this to happen, but on the premise that the data
follow a Gumbel distribution, a straight line can be fitted to
the data. Due to its simplicity, a popular method for fitting
this straight line is the method of moments, which is also
reasonably stable for limited sets of data. That is, writing the
Gumbel distribution of the extreme value as
Prob ( ) = exp { exp ( ( ))} ,
(36)
it is known that the parameters and are related to the mean

value and standard deviation of () as follows:
= 0.577221 and = 1.28255/ [39]. The estimates
of and obtained from the available sample therefore
provides estimates of and , which leads to the fitted
Gumbel distribution by the moment method.
Typically, a specified quantile value of the fitted Gumbel
distribution is then extracted and used in a design consideration. To be specific, let us assume that the requested quantile
value is the 100(1 )% fractile, where is usually a small
number, for example, = 0.1. To quantify the uncertainty
associated with the obtained 100(1 )% fractile value based
the 95% confidence interval of this
on a sample of size ,
value is often used. A good estimate of this confidence
interval can be obtained by using a parametric bootstrapping
method [40, 41]. Note that the assumption that the initial
extreme values are actually generated with good approximation from a Gumbel distribution cannot easily be verified with
any accuracy in general, which is a drawback of this method.
Compared with the POT method, the Gumbel method would
also seem to use much less of the information available in
the data. This may explain why the POT method has become
increasingly popular over the past years, but the Gumbel
method is still widely used in practice.
7. The Peaks-over-Threshold Method

7.1. The Generalized Pareto Distribution. The POT method for
independent data is based on what is called the generalized
Pareto (GP) distribution (defined below) in the following
manner: it has been shown in [42] that asymptotically the
excess values above a high level will follow a GP distribution
if and only if the parent distribution belongs to the domain
of attraction of one of the extreme value distributions. The
assumption of a Poisson process model for the exceedance
times combined with GP distributed excesses can be shown
to lead to the generalized extreme value (GEV) distribution
for corresponding extremes, see below. The expression for the
GP distribution is
1/
() = (; , ) = Prob ( ) = 1 (1 + ) .
+
(37)
Here > 0 is a scale parameter and ( < < ) determines the shape of the distribution. ()+ = max(0, ).
The asymptotic result referred to above implies that
(37) can be used to represent the conditional cumulative
distribution function of the excess = of the observed
variates over the threshold , given that > for is
sufficiently large [42]. The cases > 0, = 0, and < 0
correspond to Frechet (Type II), Gumbel (Type I), and
reverse Weibull (Type III) domains of attraction, respectively,
compared with section below.
For = 0, which corresponds to the Gumbel extreme
value distribution, the expression between the parentheses
in (37) is understood in a limiting sense as the exponential
distribution
(38)
() = (; , 0) = exp ( ) .
Since the recorded data in practice are rarely independent, a declustering technique is commonly used to filter the
data to achieve approximate independence [1, 2].
7.2. Return Periods. The return period of a given value
of in terms of a specified length of time , for example,
a year, is defined as the inverse of the probability that the
specified value will be exceeded in any time interval of length
. If denotes the mean exceedance rate of the threshold
per length of time (i.e., the average number of data points
above the threshold per ), then the return period of the
value of corresponding to the level = + is given by
the relation
1
1
=
.
=
(39)
Prob ( > ) Prob ( > )
Hence, it follows that

Prob ( ) = 1
1
.
()
(40)
Invoking (1) for = 0 leads to the result

=
[1 () ]
.
(41)
Similarly, for = 0, it is found that,

= + ln () ,
(42)
where is the threshold used in the estimation of and .
8. Extreme Value Prediction for Synthetic Data

In this section, we illustrate the performance of the ACER
method and also the 95% CI estimation. We consider 20 years
of synthetic wind speed data, amounting to 2000 data points,
which is not much for detailed statistics. However, this case
may represent a real situation when nothing but a limited data
sample is available. In this case, it is crucial to provide extreme
value estimates utilizing all data available. As we will see, the
tail extrapolation technique proposed in this paper performs
better than asymptotic methods such as POT or Gumbel.
The extreme value statistics will first be analyzed by
application to synthetic data for which the exact extreme
values can be calculated [43]. In particular, it is assumed
that the underlying (normalized) stochastic process () is
stationary and Gaussian with mean value zero and standard
deviation equal to one. It is also assumed that the mean zero
up-crossing rate + (0) is such that the product + (0) = 103 ,
where = 1 year, which seems to be typical for the wind
speed process. Using the Poisson assumption, the distribution
of the yearly extreme value of () is then calculated by the
formula
1 yr () = exp {+ () } = exp {103 exp (
2
)} ,
2
(43)
where = 1 year and + () is the mean up-crossing rate per

year, is the scaled wind speed. The 100-year return period
value 100 yr is then calculated from the relation 1 yr (100 yr ) =
1 1/100, which gives 100 yr = 4.80.
The Monte Carlo simulated data to be used for the
synthetic example are generated based on the observation
that the peak events extracted from measurements of the
wind speed process, are usually separated by 3-4 days. This is
done to obtain approximately independent data, as required
by the POT method. In accordance with this, peak event data
are generated from the extreme value distribution
3 d () = exp { exp (
2
)} ,
2
(44)
where = + (0) = 10, which corresponds to = 3.65 days,

and 1 yr () = (3 d ())100 .
Since the data points (i.e., = 3.65 days maxima) are

independent, () is independent of . Therefore, we put =
1. Since we have 100 data from one year, the data amounts to
2000 data points. For estimation of a 95% confidence interval
for each estimated value of the ACER function 1 () for the
chosen range of -values, the required standard deviation in
(22) was based on 20 estimates of the ACER function using
the yearly data. This provided a 95% confidence band on
the optimally fitted curve based on 2000 data. From these
data, the predicted 100-year return level is obtained from
1 (100 yr ) = 104 . A nonparametric bootstrapping method
was also used to estimate a 95% confidence interval based on
1000 resamples of size 2000.
The POT prediction of the 100-year return level was
based on using maximum likelihood estimates (MLE) of the
parameters in (37) for a specific choice of threshold. The
95% confidence interval was obtained from the parametrically bootstrapped PDF of the POT estimate for the given
threshold. A sample of 1000 data sets was used. One of the
unfortunate features of the POT method is that the estimated
100 year value may vary significantly with the choice of
threshold. So also for the synthetic data. We have followed the
standard recommended procedures for identifying a suitable
threshold [1].
Note that in spite of the fact that the true asymptotic
distribution of exceedances is the exponential distribution in
(38), the POT method used here is based on adopting (37).
The reason is simply that this is the recommended procedure
[1], which is somewhat unfortunate but understandable.
The reason being that the GP distribution provides greater
flexibility in terms of curve fitting. If the correct asymptotic
distribution of exceedances had been used on this example,
poor results for the estimated return period values would be
obtained. The price to pay for using the GP distribution is that
the estimated parameters may easily lead to an asymptotically
inconsistent extreme value distribution.
The 100-year return level predicted by the Gumbel method was based on using the method of moments for parameter
estimation on the sample of 20 yearly extremes. This choice
of estimation method is due to the small sample of extreme
values. The 95% confidence interval was obtained from the
parametrically bootstrapped PDF of the Gumbel prediction.
This was based on a sample of size 10,000 data sets of 20 yearly
extremes. The results obtained by the method of moments
were compared with the corresponding results obtained by
using the maximum likelihood method. While there were
individual differences, the overall picture was one of very
good agreement.
In order to get an idea about the performance of the
ACER, POT, and Gumbel methods, 100 independent 20year MC simulations as discussed above were done. Table 1
compares predicted values and confidence intervals for a
selection of 10 cases together with average values over the 100
simulated cases. It is seen that the average of the 100 predicted
100-year return levels is slightly better for the ACER method
than for both the POT and the Gumbel methods. But more
significantly, the range of predicted 100-year return levels by
the ACER method is 4.345.36, while the same for the POT
method is 4.195.87 and for the Gumbel method is 4.415.71.
Table 1: 100-year return level estimates and 95% CI (BCI = CI by bootstrap) for A = ACER, G = Gumbel, and P = POT. Exact value = 4.80.
A 100
5.07
4.65
4.86
4.75
4.54
4.80
4.84
5.02
4.59
4.84
4.62
4.82
A CI
(4.67, 5.21)
(4.27, 4.94)
(4.49, 5.06)
(4.22, 5.01)
(4.20, 4.74)
(4.35, 5.05)
(4.36, 5.20)
(4.47, 5.31)
(4.33, 4.81)
(4.49, 5.11)
(4.29, 5.05)
(4.41, 5.09)
A BCI
(4.69, 5.42)
(4.37, 5.03)
(4.44, 5.19)
(4.33, 5.02)
(4.27, 4.88)
(4.42, 5.14)
(4.48, 5.19)
(4.62, 5.36)
(4.38, 4.98)
(4.60, 5.30)
(4.45, 5.09)
(4.48, 5.18)
Hence, in this case the ACER method performs consistently

better than both these methods. It is also observed from the
estimated 95% confidence intervals that the ACER method, as
implemented in this paper, provides slightly higher accuracy
than the other two methods. Lastly, it is pointed out that the
confidence intervals of the 100-year return level estimated
by the ACER method obtained by either the simplified
extrapolated confidence band approach or by nonparametric
bootstrapping are very similar, except for a slight mean shift.
As a final comparison, the 100 bootstrapped confidence intervals obtained for the ACER and Gumbel methods missed
the target value three times, while for the POT method this
number was 18.
An example of the ACER plot and results obtained for
one set of data is presented in Figure 1. The predicted 100-year
value is 4.85 with a predicted 95% confidence interval (4.45,
5.09). Figure 2 presents POT predictions based on MLE for
different thresholds in terms of the number of data points
above the threshold. The predicted value is 4.7 at = 204,
while the 95% confidence interval is (4.25, 5.28). The same
data set as in Figure 1 was used. This was also used for the
Gumbel plot shown in Figure 3. In this case the predicted
100 yr
value based on the method of moments (MM) is MM = 4.75
with a parametric bootstrapped 95% confidence interval of
(4.34, 5.27). Prediction based on the Gumbel-Lieblein BLUE
method (GL), compared with for example, Cook [44], is
100 yr
GL = 4.73 with a parametric bootstrapped 95% confidence
interval equal to (4.35, 5.14).
9. Measured Wind Speed Data

In this section, we analyze real wind speed data, measured
at two weather stations off the coast of Norway: at Nordyan
and at Hekkingen, see Figure 4. Extreme wind speed prediction is an important issue for design of structures exposed to
the weather variations. Significant efforts have been devoted
to the problem of predicting extreme wind speeds on the basis
of measured data by various authors over several decades,
see, for example, [4548] for extensive references to previous
work.
G 100
4.41
4.92
5.04
4.75
4.80
4.91
4.85
4.96
4.76
4.77
4.79
4.84
P 100
4.29
4.88
5.04
4.69
4.73
4.79
4.71
4.97
4.68
4.41
4.53
4.72
G BCI
(4.14, 4.73)
(4.40, 5.58)
(4.54, 5.63)
(4.27, 5.32)
(4.31, 5.39)
(4.41, 5.50)
(4.36, 5.43)
(4.47, 5.53)
(4.31, 5.31)
(4.34, 5.32)
(4.31, 5.41)
(4.37, 5.40)
P BCI
(4.13, 4.52)
(4.42, 5.40)
(4.48, 5.74)
(4.24, 5.26)
(4.19, 5.31)
(4.31, 5.34)
(4.32, 5.23)
(4.47, 5.71)
(4.15, 5.27)
(4.23, 4.64)
(4.05, 4.88)
(4.27, 5.23)
101
102
ACER1 ()
Sim. No.
1
10
20
30
40
50
60
70
80
90
100
Av. 100
103
104
105
4.85
2.5
3.5
4.5
5.5
Figure 1: Synthetic data ACER 1 , Monte Carlo simulation ();

optimized curve fit (); empirical 95% confidence band (- -);
optimized confidence band ( ). Tail marker 1 = 2.3.
Hourly maximum gust wind was recorded during the

13 years 19992012 at Nordyan and the 14 years 1998
2012 at Hekkingen. The objective is to estimate a 100-year
wind speed. Variation in the wind speed caused by seasonal
variations in the wind climate during the year makes the
wind speed a non-stationary process on the scale of months.
Moreover, due to global climate change, yearly statistics may
vary on the scale of years. The latter is, however, a slow
process, and for the purpose of long-term prediction we
assume here that within a time span of 100 years a quasistationary model of the wind speeds applies. This may not be
entirely true, of course.
9.1. Nordyan. Figure 5 highlights the cascade of ACER
estimates 1 , . . . , 96 , for the case of 13 years of hourly data
recorded at the Nordyan weather station. Here, 96 is considered to represent the final converged results. By converged,
we mean that 96 for > 96 in the tail, so that there is no
10

4.76
4.74
Hekkingen Fyr 88690
4.72
4.7
100 yr
4.7
4.68
4.66
4.64
Nordyan Fyr 75410
4.62
4.6
100
130
160
190 204
230
250
Figure 2: The point estimate 100 yr of the 100-year return period

value based on 20 years synthetic data as a function of the number
of data points above threshold. The return level estimate = 4.7 at
= 204.
5
Figure 4: Wind speed measurement stations.

101
3
2
102
ACER ()
ln(ln(( + 1)/))
1
0
103
4.75
1
3.6
3.8
4.2
4.4
4.6
4.8
Figure 3: The point estimate 100 yr of the 100-year return period

value based on 20 years synthetic data. Lines are fitted by the
method of momentssolid line () and the Gumbel-Lieblein
BLUE methoddash-dotted lite (- -). The return level estimate
by the method of moments is 4.75, by the Gumbel-Lieblein BLUE
method is 4.73.
need to consider conditioning of an even higher order than

96. Figure 5 reveals a rather strong statistical dependence
between consecutive data, which is clearly reflected in the
effect of conditioning on previous data values. It is also interesting to observe that this effect is to some extent captured
already by 2 , that is, by conditioning only on the value of the
previous data point. Subsequent conditioning on more than
one previous data point does not lead to substantial changes
in ACER values, especially for tail values. On the other hand,
to bring out fully the dependence structure of these data, it
was necessary to carry the conditioning process to (at least)
the 96th ACER function, as discussed above.
However, from a practical point of view, the most important information provided by the ACER plot of Figure 5 is that
104
3.5
4.5
5.5
6.5
/
=1
=2
=4
= 24
= 48
= 72
= 96
Figure 5: Nordyan wind speed statistics, 13 years hourly data.

Comparison between ACER estimates for different degrees of conditioning. = 6.01 m/s.
for the prediction of a 100-year value, one may use the first
ACER function. The reason for this is that Figure 5 shows that
all the ACER functions coalesce in the far tail. Hence, we may
use any of the ACER functions for the prediction. Then, the
obvious choice is to use the first ACER function, which allows
us to use all the data in its estimation and thereby increase
accuracy.
In Figure 6 is shown the results of parametric estimation
of the return value and its 95% CI for 13 years of hourly
11
5
101
4
3
ln(ln(( + 1)/))
ACER1 ()
102
103
104
105
2
1
0
106
8.62
3
6
/
Figure 6: Nordyan wind speed statistics, 13 years hourly data. 1 ()

(); optimized curve fit (); empirical 95% confidence band (- -);
optimized confidence band ( ). Tail marker 1 = 12.5 m/s = 2.08
( = 6.01 m/s).
6.5
7.5
/
8.5
9.23
9
9.5
Figure 8: Nordyan wind speed statistics, 13 years of hourly

data. Gumbel plot of yearly extremes. Lines are fitted by the
BLUE methoddash-dotted lite (- -). = 6.01 m/s.
Table 2: Predicted 100-year return period levels for Nordyan
Fyr weather station by the ACER method for different degrees of
conditioning, annual maxima, and POT methods, respectively.
8.1
8
7.95
100 yr /
8.56
Method
7.9
ACER, various
7.8
7.7
7.6
Annual maxima
100
120
140
161
180
200
Figure 7: The point estimate 100 yr of the 100-year return level based
on 13 years hourly data as a function of the number of data points
above threshold. = 6.01 m/s.
maxima. The predicted 100-year return speed is 100 yr =

51.85 m/s with 95% confidence interval (48.4, 53.1). = 13
years of data may not be enough to guarantee (22), since we
required 20. Nevertheless, for simplicity, we use it here
even with = 13, accepting that it may not be very accurate.
Figure 7 presents POT predictions for different threshold
numbers based on MLE. The POT prediction is 100 yr =
47.8 m/s at threshold = 161, while the bootstrapped 95%
confidence interval is found to be (44.8, 52.7) m/s based
on 10,000 generated samples. It is interesting to observe the
unstable characteristics of the predictions over a range of
threshold values, while they are quite stable on either side of
this range giving predictions that are more in line with the
results from the other two methods.
Figure 8 presents a Gumbel plot based on the 13 yearly
extremes extracted from the 13 years of hourly data. The
POT
Spec
1
2
4
24
48
72
96
MM
GL
100 yr , m/s
51.85
51.48
52.56
52.90
54.62
53.81
54.97
51.5
55.5
47.8
95% CI (
100 yr ), m/s
(48.4, 53.1)
(46.1, 54.1)
(46.7, 55.7)
(47.0, 56.2)
(47.7, 57.6)
(46.9, 58.3)
(47.5, 60.5)
(45.2, 59.3)
(48.0, 64.9)
(44.8, 52.7)
Gumbel prediction based on the method of moments (MM)

100 yr
is MM = 51.5 m/s, with a parametric bootstrapped 95%
confidence interval equal to (45.2, 59.3) m/s, while prediction
100 yr
based on the Gumbel-Lieblein BLUE method (GL) is GL =
55.5 m/s, with a parametric bootstrapped 95% confidence
interval equal to (48.0, 64.9) m/s.
In Table 2 the 100-year return period values for the
Nordyan station are listed together with the predicted 95%
confidence intervals for all methods.
9.2. Hekkingen. Figure 9 shows the cascade of estimated
ACER functions 1 , . . . , 96 for the case of 14 years of hourly
data. As for Nordyan, 96 is used to represent the final
converged results. Figure 9 also reveals a rather strong statistical dependence between consecutive data at moderate
wind speed levels. This effect is again to some extent captured
already by 2 , so that subsequent conditioning on more than
one previous data point does not lead to substantial changes
in ACER values, especially for tail values.
12

Table 3: Predicted 100-year return period levels for Nordyan
Fyr weather station by the ACER method for different degrees of
conditioning, annual maxima, and POT methods, respectively.
102
ACER ()
Method
10
Spec
100 yr , m/s
95% CI (
100 yr ), m/s
60.47
(53.1, 64.9)
62.23
(53.3, 70.0)
63.03
(53.0, 74.5)
24
60.63
(51.3, 70.7)
48
60.44
(51.3, 77.0)
72
58.06
(51.2, 66.4)
96
59.19
(52.0, 68.3)
MM
58.10
(50.8, 67.3)
GL
60.63
(53.0, 70.1)
53.48
(48.9, 57.0)
ACER, various
104
4.5
5.5
6.5
7.5
8.5
Annual maxima
= 48
= 72
= 96
=1
=2
=4
= 24
POT
Figure 9: Hekkingen wind speed statistics, 14 years hourly data.

Comparison between ACER estimates for different degrees of conditioning. = 5.72 m/s.
9.42
9.38
9.35
100 yr /
102
9.34
ACER1 ()
103
9.3
104
9.26
105
100
140
185
220
260
300
106
10.6
5
10
11
12
Figure 11: The point estimate 100 yr of the 100-year return level
based on 14 years hourly data as a function of the number of data
points above threshold. = 5.72 m/s.
Figure 10: Hekkingen wind speed statistics, 14 years hourly data.

1 () (); optimized curve fit (); empirical 95% confidence band
(- -); optimized confidence band ( ). Tail marker 1 = 23 m/s =
4.02 ( = 5.72 m/s).
Also, for the Hekkingen data, the ACER plot of Figure 9

indicates that the ACER functions coalesce in the far tail.
Hence, for the practical prediction of a 100-year value, one
may use the first ACER function.
In Figure 10 is shown the results of parametric estimation
of the return value and its 95% CI for 14 years of hourly
maxima. The predicted 100-year return speed is 100 yr =
60.47 m/s with 95% confidence interval (53.1, 64.9). Equation
(22) has been used also for this example with = 14.
Figure 11 presents POT predictions for different threshold
numbers based on MLE. The POT prediction is 100 yr =
53.48 m/s at threshold = 183, while the bootstrapped
95% confidence interval is found to be (48.9, 57.0) m/s based

on 10,000 generated samples. It is interesting to observe
the unstable characteristics of the predictions over a range of
threshold values, while they are quite stable on either side of
this range giving predictions that are more in line with the
results from the other two methods.
Figure 12 presents a Gumbel plot based on the 14 yearly
extremes extracted from the 14 years of hourly data. The
Gumbel prediction based on the method of moments (MM)
100 yr
is MM = 58.10 m/s, with a parametric bootstrapped 95%
confidence interval equal to (50.8, 67.3) m/s. Prediction based
100 yr
=
on the Gumbel-Lieblein BLUE method (GL) is GL
60.63 m/s, with a parametric bootstrapped 95% confidence
interval equal to (53.0, 70.1) m/s.
In Table 3, the 100-year return period values for the
Hekkingen station are listed together with the predicted 95%
confidence intervals for all methods.
13
2.5
2
1.5
1
0.5
()
ln(ln(( + 1)/))
0.5
1
1.5
0
10.2 10.6
1
7
7.5
8.5
9.5
10
10.5
Figure 12: Hekkingen wind speed statistics, 14 years of hourly

data. Gumbel plot of yearly extremes. Lines are fitted by the
BLUE methoddash-dotted lite (- -). = 5.72 m/s.
10. Extreme Value Prediction for a Narrow

Band Process
In engineering mechanics, a classical extreme response prediction problem is the case of a lightly damped mechanical oscillator subjected to random forces. To illustrate this
prediction problem, we will investigate the response process of a linear mechanical oscillator driven by a Gaussian
white noise. Let () denote the displacement response; the
+
+ 2 ()
dynamic model can then be expressed as, ()
2
() = (), where = relative damping, = undamped
eigenfrequency, and () = a stationary Gaussian white noise
(of suitable intensity). By choosing a small value for , the
response time series will exhibit narrow band characteristics,
that is, the spectral density of the response process ()
will assume significant values only over a narrow range
of frequencies. This manifests itself by producing a strong
beating of the response time series, which means that the size
of the response peaks will change slowly in time, see Figure 13.
A consequence of this is that neighbouring peaks are strongly
correlated, and there is a conspicuous clumping of the peak
values. Hence the problem with accurate prediction, since the
usual assumption of independent peak values is then violated.
Many approximations have been proposed to deal with
this correlation problem, but no completely satisfactory
solution has been presented. In this section, we will show
that the ACER method solves this problem efficiently and
elegantly in a statistical sense. In Figure 14 are shown some
of the ACER functions for the example time series. It may
be verified from Figure 13 that there are approximately 32
sample points between two neighbouring peaks in the time
series. To illustrate a point, we have chosen to analyze the
time series consisting of all sample points. Usually, in practice,
only the time series obtained by extracting the peak values
would be used for the ACER analysis. In the present case,
the first ACER function is then based on assuming that all
2
2.5
60
70
80
90
100
110
120
Time (s)
Figure 13: Part of the narrow-band response time series of the linear
oscillator with fully sampled and peak values indicated.
the sampled data points are independent, which is obviously

completely wrong. The second ACER function, which is
based on counting each exceedance with an immediately
preceding nonexceedance, is nothing but an upcrossing rate.
Using this ACER function is largely equivalent to assuming
independent peak values. It is now interesting to observe
that the 25th ACER function can hardly be distinguished
from the second ACER function. In fact, the ACER functions
after the second do not change appreciably until one starts to
approach the 32nd, which corresponds to hitting the previous
peak value in the conditioning process. So, the important
information concerning the dependence structure in the
present time series seems to reside in the peak values, which
may not be very surprising. It is seen that the ACER functions
show a significant change in value as a result of accounting
for the correlation effects in the time series. To verify the
full dependence structure in the time series, it is necessary
to continue the conditioning process down to at least the
64th ACER function. In the present case, there is virtually
no difference between the 32nd and the 64th, which shows
that the dependence structure in this particular time series is
captured almost completely by conditioning on the previous
peak value. It is interesting to contrast the method of dealing
with the effect of sampling frequency discussed here with that
of [49].
To illustrate the results obtained by extracting only the
peak values from the time series, which would be the
approach typically chosen in an engineering analysis, the
ACER plots for this case is shown in Figure 15. By comparing
results from Figures 14 and 15, it can be verified that they
are in very close agreement by recognizing that the second
ACER function in Figure 14 corresponds to the first ACER
function in Figure 15, and by noting that there is a factor of
approximately 32 between corresponding ACER functions in
the two figures. This is due to the fact that the time series of
peak values contains about 32 times less data than the original
time series.
14

101
ACER ()
102
103
104
105
1.5
2.5
3.5
/
= 32
= 64
=1
=2
= 25
Figure 14: Comparison between ACER estimates for different

degrees of conditioning for the narrow-band time series.
related to applications in mechanics is also presented. The

validation of the method is done by comparison with exact
results (when available), or other widely used methods for
extreme value statistics, such as the Gumbel and the peaksover-threshold (POT) methods. Comparison of the various
estimates indicate that the proposed method provides more
accurate results than the Gumbel and POT methods.
Subject to certain restrictions, the proposed method also
applies to nonstationary time series, but it cannot directly
predict for example, the effect of climate change in the form
of long-term trends in the average exceedance rates extending
beyond the data. This must be incorporated into the analysis
by explicit modelling techniques.
As a final remark, it may be noted that the ACER method
as described in this paper has a natural extension to higher
dimensional distributions. The implication is that, it is then
possible to provide estimates of for example, the exact bivariate extreme value distribution for a suitable set of data [50].
However, as is easily recognized, the extrapolation problem is
not as simply dealt with as for the univariate case studied in
this paper.
ACER ()
Acknowledgment
101
This work was supported by the Research Council of Norway

(NFR) through the Centre for Ships and Ocean Structures
(CeSOS) at the Norwegian University of Science and Technology.
102
References
103
1.5
2.5
3.5
/
=1
=2
=3
=4
=5
Figure 15: Comparison between ACER estimates for different

degrees of conditioning based on the time series of the peak values,
compared with Figure 13.
11. Concluding Remarks

This paper studies a new method for extreme value prediction
for sampled time series. The method is based on the introduction of a conditional average exceedance rate (ACER), which
allows dependence in the time series to be properly and easily
accounted for. Declustering of the data is therefore avoided,
and all the data are used in the analysis. Significantly, the
proposed method also aims at capturing to some extent the
subasymptotic form of the extreme value distribution.
Results for wind speeds, both synthetic and measured,
are used to illustrate the method. An estimation problem
[1] S. Coles, An Introduction to Statistical Modeling of Extreme

Values, Springer Series in Statistics, Springer, London, UK, 2001.
[2] J. Beirlant, Y. Goegebeur, J. Teugels, and J. Segers, Statistics of
Extremes, Wiley Series in Probability and Statistics, John Wiley
& Sons, Chichester, UK, 2004.
[3] A. C. Davison and R. L. Smith, Models for exceedances over
high thresholds, Journal of the Royal Statistical Society. Series B.
Methodological, vol. 52, no. 3, pp. 393442, 1990.
[4] R.-D. Reiss and M. Thomas, Statistical Analysis of Extreme Values, Birkhauser, Basel, Switzerland, 3rd edition, 1997.
[5] P. Embrechts, C. Kluppelberg, and T. Mikosch, Modelling Extremal Events, vol. 33 of Applications of Mathematics (New York),
Springer, Berlin, Germany, 1997.
[6] M. Falk, J. Husler, and R.-D. Reiss, Laws of Small Numbers:
Extremes and Rare Events, Birkhauser, Basel, Switzerland, 2 nd
edition, 2004.
[7] A. Naess and O. Gaidai, Monte Carlo methods for estimating
the extreme response of dynamical systems, Journal of Engineering Mechanics, vol. 134, no. 8, pp. 628636, 2008.
[8] A. Naess, O. Gaidai, and S. Haver, Efficient estimation of
extreme response of drag-dominated offshore structures by
Monte Carlo simulation, Ocean Engineering, vol. 34, no. 16, pp.
21882197, 2007.
[9] R. L. Smith, The extremal index for a Markov chain, Journal
of Applied Probability, vol. 29, no. 1, pp. 3745, 1992.
[10] S. G. Coles, A temporal study of extreme rainfall, in Statistics
for the Environment 2-Water Related Issues, V. Barnett and K.
F. Turkman, Eds., chapter 4, pp. 6178, John Wiley & Sons,
Chichester, UK, 1994.

[11] R. L. Smith, J. A. Tawn, and S. G. Coles, Markov chain models
for threshold exceedances, Biometrika, vol. 84, no. 2, pp. 249
268, 1997.
[12] S. Yun, The extremal index of a higher-order stationary Markov chain, The Annals of Applied Probability, vol. 8, no. 2, pp.
408437, 1998.
[13] S. Yun, The distributions of cluster functionals of extreme
events in a dth-order Markov chain, Journal of Applied Probability, vol. 37, no. 1, pp. 2944, 2000.
[14] J. Segers, Approximate distributions of clusters of extremes,
Statistics & Probability Letters, vol. 74, no. 4, pp. 330336, 2005.
[15] G. S. Watson, Extreme values in samples from -dependent
stationary stochastic processes, Annals of Mathematical Statistics, vol. 25, pp. 798800, 1954.
[16] M. R. Leadbetter, G. Lindgren, and H. Rootzen, Extremes and
Related Properties of Random Sequences and Processes, Springer
Series in Statistics, Springer, New York, NY, USA, 1983.
[17] D. E. Cartwright, On estimating the mean energy of sea waves
from the highest waves in a record, Proceedings of the Royal
Society of London. Series A, vol. 247, pp. 2228, 1958.
[18] A. Naess, On the long-term statistics of extremes, Applied
Ocean Research, vol. 6, no. 4, pp. 227228, 1984.
[19] G. Schall, M. H. Faber, and R. Rackwitz, Ergodicity assumption
for sea states in the reliability estimation of offshore structures,
Journal of Offshore Mechanics and Arctic Engineering, vol. 113,
no. 3, pp. 241246, 1991.
[20] E. H. Vanmarcke, On the distribution of the first-passage time
for normal stationary random processes, Journal of Applied
Mechanics, vol. 42, no. 1, pp. 215220, 1975.
[21] M. R. Leadbetter, Extremes and local dependence in stationary
sequences, Zeitschrift fur Wahrscheinlichkeitstheorie und Verwandte Gebiete, vol. 65, no. 2, pp. 291306, 1983.
[22] T. Hsing, On the characterization of certain point processes,
Stochastic Processes and Their Applications, vol. 26, no. 2, pp.
297316, 1987.
[23] T. Hsing, Estimating the parameters of rare events, Stochastic
Processes and Their Applications, vol. 37, no. 1, pp. 117139, 1991.
[24] M. R. Leadbetter, On high level exceedance modeling and tail
inference, Journal of Statistical Planning and Inference, vol. 45,
no. 1-2, pp. 247260, 1995.
[25] C. A. T. Ferro and J. Segers, Inference for clusters of extreme
values, Journal of the Royal Statistical Society. Series B, vol. 65,
no. 2, pp. 545556, 2003.
[26] C. Y. Robert, Inference for the limiting cluster size distribution
of extreme values, The Annals of Statistics, vol. 37, no. 1, pp. 271
310, 2009.
[27] A. Naess and O. Gaidai, Estimation of extreme values from
sampled time series, Structural Safety, vol. 31, no. 4, pp. 325
334, 2009.
[28] P. E. Gill, W. Murray, and M. H. Wright, Practical Optimization,
Academic Press Inc, London, UK, 1981.
[29] W. Forst and D. Hoffmann, OptimizationTheory and Practice,
Springer Undergraduate Texts in Mathematics and Technology,
Springer, New York, NY, USA, 2010.
[30] N. R. Draper and H. Smith, Applied Regression Analysis, Wiley
Series in Probability and Statistics: Texts and References Section, John Wiley & Sons, New York, NY, USA, 3rd edition, 1998.
[31] D. C. Montgomery, E. A. Peck, and G. G. Vining, Introduction
to Linear Regression Analysis, Wiley Series in Probability and
Statistics: Texts, References, and Pocketbooks Section, WileyInterscience, New York, NY, USA, Third edition, 2001.
15
[32] A. Naess, Estimation of Extreme Values of Time Series with
Heavy Tails, Preprint Statistics No. 14/2010, Department of
Mathematical Sciences, Norwegian University of Science and
Technology, Trondheim, Norway, 2010.
[33] E. F. Eastoe and J. A. Tawn, Modelling the distribution of the
cluster maxima of exceedances of subasymptotic thresholds,
Biometrika, vol. 99, no. 1, pp. 4355, 2012.
[34] J. A. Tawn, Discussion of paper by A. C. Davison and R. L.
Smith, Journal of the Royal Statistical Society. Series B, vol. 52,
no. 3, pp. 393442, 1990.
[35] A. W. Ledford and J. A. Tawn, Statistics for near independence
in multivariate extreme values, Biometrika, vol. 83, no. 1, pp.
169187, 1996.
[36] J. E. Heffernan and J. A. Tawn, A conditional approach for
multivariate extreme values, Journal of the Royal Statistical
Society. Series B, vol. 66, no. 3, pp. 497546, 2004.
[37] Numerical Algorithms Group, NAG Toolbox for Matlab, NAG,
Oxford, UK, 2010.
[38] E. J. Gumbel, Statistics of Extremes, Columbia University Press,
New York, NY, USA, 1958.
[39] K. V. Bury, Statistical Models in Applied Science, Wiley Series
in Probability and Mathematical Statistics, John Wiley & Sons,
New York, NY, USA, 1975.
[40] B. Efron and R. J. Tibshirani, An Introduction to the Bootstrap,
vol. 57 of Monographs on Statistics and Applied Probability,
Chapman and Hall, New York, NY, USA, 1993.
[41] A. C. Davison and D. V. Hinkley, Bootstrap Methods and Their
Application, vol. 1 of Cambridge Series in Statistical and Probabilistic Mathematics, Cambridge University Press, Cambridge,
UK, 1997.
[42] J. Pickands, III, Statistical inference using extreme order
statistics, The Annals of Statistics, vol. 3, pp. 119131, 1975.
[43] A. Naess and P. H. Clausen, Combination of the peaks-overthreshold and bootstrapping methods for extreme value prediction, Structural Safety, vol. 23, no. 4, pp. 315330, 2001.
[44] N. J. Cook, The Designers Guide to Wind Loading of Building
Structures, Butterworths, London, UK, 1985.
[45] N. J. Cook, Towards better estimation of extreme winds,
Journal of Wind Engineering and Industrial Aerodynamics, vol.
9, no. 3, pp. 295323, 1982.
[46] A. Naess, Estimation of long return period design values for
wind speeds, Journal of Engineering Mechanics, vol. 124, no. 3,
pp. 252259, 1998.
[47] J. P. Palutikof, B. B. Brabson, D. H. Lister, and S. T. Adcock, A
review of methods to calculate extreme wind speeds, Meteorological Applications, vol. 6, no. 2, pp. 119132, 1999.
[48] O. Perrin, H. Rootzen, and R. Taesler, A discussion of statistical
methods used to estimate extreme wind speeds, Theoretical and
Applied Climatology, vol. 85, no. 3-4, pp. 203215, 2006.
[49] M. E. Robinson and J. A. Tawn, Extremal analysis of processes
sampled at different frequencies, Journal of the Royal Statistical
Society. Series B, vol. 62, no. 1, pp. 117135, 2000.
[50] A. Naess, A Note on the Bivariate ACER Method, Preprint Statistics No. 01/2011, Department of Mathematical Sciences, Norwegian University of Science and Technology, Trondheim,
Norway, 2011.
Advances in
Operations Research
http://www.hindawi.com
Volume 2014
Advances in
Decision Sciences
Volume 2014
Mathematical Problems
in Engineering
Volume 2014
Journal of
Algebra
Probability and Statistics

Volume 2014
The Scientific
World Journal

Volume 2014
International Journal of
Differential Equations
Volume 2014
Volume 2014
Submit your manuscripts at

Advances in
Combinatorics
Mathematical Physics
Volume 2014
Journal of
Complex Analysis
Volume 2014
International
Journal of
Mathematics and
Mathematical
Sciences
Journal of

Stochastic Analysis
Abstract and
Applied Analysis


Mathematics
Volume 2014
Volume 2014
Discrete Dynamics in
Nature and Society
Volume 2014
Volume 2014
Journal of
Journal of
Discrete Mathematics
Journal of
Volume 2014

Applied Mathematics
Journal of
Function Spaces
Volume 2014

Volume 2014

Volume 2014
Optimization
Volume 2014

Volume 2014
Applied Ocean Research 28 (2006) 18

www.elsevier.com/locate/apor
Numerical methods for calculating the crossing rate of high and extreme
response levels of compliant offshore structures subjected to random waves
A. Naess a,b, , H.C. Karlsen b , P.S. Teigen c
a Centre for Ships and Ocean Structures, Norwegian University of Science and Technology, A. Getz vei 1, NO-7491, Trondheim, Norway
b Department of Mathematical Sciences, Norwegian University of Science and Technology, A. Getz vei 1, NO-7491, Trondheim, Norway
c Statoil Research Centre, Rotvoll, Trondheim, Norway
Received 16 December 2005; received in revised form 3 April 2006; accepted 8 April 2006
Available online 5 June 2006
Abstract
The focus of the paper is on methods for calculating the mean upcrossing rate of stationary stochastic processes that can be represented as
second order stochastic Volterra series. This is the current state-of-the-art representation of the horizontal motion response of e.g. a tension leg
platform in random seas. Until recently, there has been no method available for accurately calculating the mean level upcrossing rate of such
response processes. Since the mean upcrossing rate is a key parameter for estimating the large and extreme responses it is clearly of importance
to develop methods for its calculation. The paper describes in some detail a numerical method for calculating the mean level upcrossing rate
of a stochastic response process of the type considered. Since no approximations are made, the only source of inaccuracy is in the numerical
calculation, which can be controlled. In addition to this exact method, two approximate methods are also discussed.
c 2006 Elsevier Ltd. All rights reserved.
Keywords: Second order stochastic Volterra model; Mean crossing rate; Extreme response; Slow drift response; Method of steepest descent
1. Introduction
The problem of calculating the extreme response of
compliant offshore structures like tension leg platforms or
moored spar buoys in random seas, has been a challenge for
many years, and, in fact, it still represents a challenge. Starting
with the state-of-the-art representation of the horizontal
excursions of moored, floating offshore structures in random
seas as a second order stochastic Volterra series, we shall in
this paper develop a general method for estimating the extreme
response of such structures. Even if the Volterra series model
was formulated more than 30 years ago, it is not until quite
recently that general numerical methods have become available
that allow accurate calculation of the probability distribution
and, perhaps more importantly, the mean upcrossing rate of
the total response process. This last quantity is the crucial
parameter for estimating extreme responses.
During the 1980s significant efforts were directed towards
developing methods for calculating the response statistics of
Corresponding author. Tel.: +47 73 59 70 53; fax: +47 73 59 35 24.
E-mail address: arvidn@math.ntnu.no (A. Naess).

0141-1187/$ - see front matter c 2006 Elsevier Ltd. All rights reserved.
doi:10.1016/j.apor.2006.04.001
compliant offshore structures subjected to random waves. The

list of contributions is long. To mention but a few, which
also contain references to other work focussing on response
statistics, see [16]. However, none of these works succeeded in
developing a general method that made it possible to calculate
the exact statistical distribution of the total response process,
not to mention the much harder problem of calculating the mean
upcrossing rate.
A general method for solving the first problem of calculating
the exact statistical distribution was presented in [7]. Then it
took almost a decade before a general method for solving the
second problem was outlined [8]. This method was developed
further in [9]. While the method is mathematically sound, initial
efforts to carry out the requisite calculations have revealed that
some care is needed in setting up the numerical algorithms.
The work presented in this paper is part of continued efforts
to set up a robust and accurate numerical procedure. It should
be emphasized that while the discussion in this paper is limited
to the long-crested seas case, the method described also covers
the short-crested seas case, cf. [6]. A recent paper presenting
a discussion and comparison of some approximate procedures
for calculating the mean upcrossing rate is notable [10].
A. Naess et al. / Applied Ocean Research 28 (2006) 18
2. The response process

The response process Z (t) that is considered here is assumed
to be represented as a second order stochastic Volterra series.
This would apply to the state of the art representation of
e.g. the surge response of a large volume, compliant offshore
structure in random waves. This response would consist of a
combination of the wave frequency component Z 1 (t) and the
slow-drift component Z 2 (t), that is, Z (t) = Z 1 (t) + Z 2 (t).
Naess [6] describes the standard representation of the two
response components leading to a second order Volterra series
model for the total response. To alleviate the statistical analysis
of the response process, it has been shown [2,6] that the slowdrift response Z 2 (t) can be expressed as
N
Z 2 (t) =
j {W2 j1 (t)2 + W2 j (t)2 }.
(1)
where H 1 () denotes the linear transfer function between the

waves and the surge response. Based on the representations
given by Eqs. (1) and (5), [11] describes how to calculate
the statistical moments of the response process Z (t), while
a general and accurate numerical method for calculating the
PDF of Z (t) is given in [7]. However, for important prediction
purposes, like extreme response estimation, the crucial quantity
is the mean rate of level upcrossings by the response process.
3. The mean crossing rate
Let N Z+ ( ) denote the rate of upcrossings of the level
by Z (t), cf. [12], and let Z+ ( ) = E[N Z+ ( )], that is,
Z+ ( ) denotes the mean rate of upcrossings of the level . As
discussed in [9], under suitable regularity conditions on the
response process, which can be adopted here, the following
formula can be used
j=1
Here W j (t), j = 1, . . . , 2N are real stationary Gaussian

N (0, 1)-processes. The coefficients j are obtained by solving
the eigenvalue problem (assumed nonsingular)
Qu j = j u j
(2)
to find the eigenvalues j and orthonormal eigenvectors u j ,

j = 1, . . . , N , of the N N -matrix Q = (Q i j ), where
1
Q i j = H 2 (i , j ) [S X (i )S X ( j )]1/2 .
2
(3)
Here H 2 (, ) denotes the quadratic transfer function between

the waves and the surge response, cf. [2,6], S X () denotes
the one-sided spectral density of the waves, and 0 < 1 <
< N is a suitable discretization of the frequency axis.
The stochastic
processes W j (t) can be represented as follows
(i = 1)
W2 j1 (t) + i W2 j (t) =
u j (k )Bk eik t
(4)
k=1
where u j (k ) denotes the kth component of u j and {Bk }

is a set of independent, complex Gaussian N (0, 1)-variables
with independent and identically distributed real and imaginary
parts. The representation can be arranged so that W2 j (t)
becomes the Hilbert transform of W2 j1 (t), cf. [6]. For each
fixed t, {W j (t)} becomes a set of independent Gaussian
variables.
Having achieved the desired representation of the quadratic
response Z 2 (t), it can then be shown that the linear response
can be expressed as
2N
Z 1 (t) =
j W j (t).
(5)
j=1
The (real) parameters j are given by the relations

H 1 (k )
k=1
S X (k ) u j (k )
(6)
s f Z Z (, s)ds
(7)
where f Z Z (, ) denotes the joint PDF of Z (0) and Z (0) =

dZ (t)/dt|t=0 . Eq. (7) is often referred to as the Rice
formula [13]. Z+ ( ) is assumed throughout to be finite.
Calculating the mean crossing rate of a stochastic process
represented as a second order stochastic Volterra series directly
from Eq. (7) has turned out to be very difficult due to the
difficulties of calculating the joint PDF f Z Z (, ). However, this
can be circumvented by invoking the concept of characteristic
function.
Denote the characteristic function of the joint variable
(Z , Z ) by M Z Z (, ), or, for simplicity of notation, by M(, ).
Then
M(u, v) = M Z Z (u, v) = E[exp(iu Z + iv Z )].
(8)
Assuming that M(, ) is an integrable function, that is,

M(, ) L 1 (R2 ), it follows that
f Z Z (, s) =
1
(2 )2
M(u, v)

exp (iu ivs) dudv.
(9)
By substituting from Eq. (9) back into Eq. (7), the mean
crossing rate is formally expressed in terms of the characteristic
function, but this is not a very practical expression.
The solution to this is obtained by considering the
characteristic function as a function of two complex variables.
It can then often be shown that this new function becomes
holomorphic in suitable regions of C2 , where C denotes the
complex plane. As shown in detail in [14], under suitable
conditions, the use of complex function theory allows the
derivation of two alternative expressions for the crossing rate.
Here we shall focus on one of these alternatives, viz.
Z+ ( ) =
2 j1 + i 2 j =
Z+ ( ) =
1
(2 )2
ia
ib
ia
ib
1
M(z, w)eiz dzdw (10)
w2
where 0 < a < a1 for some positive constant a1 , and b0 < b <
b1 for some constants b0 < 0 and b1 > 0.
To actually carry out the calculations, the joint characteristic

function needs to be known. It has been shown [8] that for
the second order stochastic Volterra series, it can be given
in closed form. To this end, consider the multidimensional
Gaussian vectors W = (W1 , . . . , Wn ) ( denotes transposition)
and W = (W 1 , . . . , W n ) , where n = 2N . It is obtained that the
covariance matrix of (W , W ) is given by
11
=
21
12
22
(11)
(ik )u i (k )u j (k ) .
(12)
k=1
Then it can be shown that

r2i1,2 j1 = r2i,2 j = R(Ri j )
(13)
while
r2i1,2 j = r2i,2 j1 = (Ri j )
(14)
where R(z) denotes the real part and (z) the imaginary part of
z. Similarly, let
N
k2 u i (k )u j (k ) .
(15)
s2i1,2 j1 = s2i,2 j = R(Si j )
(16)
Si j =
k=1
Then
while
s2i1,2 j = s2i,2 j1 = (Si j ).
(17)
By this, the covariance matrix is completely specified.

It is convenient to introduce a new set of eigenvalues j ,
j = 1, . . . , n defined by 2i1 = 2i = i , i = 1, . . . , N .
Let = diag(1 , . . . , n ) be the diagonal matrix with the
parameters j on the diagonal, and let = (1 , . . . , n ) . It
can now be shown that [8]
M(u, v)
= exp
Previous efforts to carry out numerical calculation of the

mean crossing rate using Eq. (10) have been reported in [9].
These initial investigations indicated that the method had the
potential to provide very accurate numerical results. We shall
rewrite Eq. (10) as follows
Z+ ( ) =
1
(2 )2
ia
ia
1
I (, w)dw
w2
(22)
where
where 11 = I = the n n identity matrix, 12 =

(ri j ) = (E[Wi W j ]), 21 = (E[W i W j ]) and 22 = (si j ) =
(E[W i W j ]); i, j = 1, . . . , n. ri j = r ji and 12 = 21 . It
follows from Eq. (4), that the entries of the covariance matrix
can be expressed in terms of the eigenvectors u j , cf. [2]. Let
Ri j =
4. Numerical calculation
1
1
1
ln(det(A)) v 2 V + t A1 t
2
2
2
(18)
ib
ib
ib
exp{iz + ln M(z, w)}dz.
(23)
ib
A numerical calculation of the mean crossing rate can start by

calculating the function I (, w) for specified values of and
w. However, a direct numerical integration of Eq. (23) is made
difficult by the oscillatory term exp{iR(z) }. This problem
can be avoided by invoking the method of steepest descent, also
called the saddle point method. For this purpose, we write
g(z) = g(z; w) = iz + ln M(z, w)
= (x, y) + i (x, y)
(24)
where z = x + iy. (x, y) and (x, y) become real harmonic
functions when g(z) is holomorphic. The idea is to identify
the saddle point of the surface (x, y) (x, y) closest to
the integration line from ib to ib. By shifting this
integration line to a new integration contour that passes through
the saddle point, and then follows the path of steepest descent
away from the saddle point, it can be shown that the function
(x, y) stays constant, and therefore the oscillatory term in the
integral degenerates to a constant. This is a main advantage of
the method of steepest descent for numerical calculations. It
can be shown that the integral does not change its value as long
as the function g(z) is a holomorphic function in the region
bounded by the two integration contours and if the integrals
vanish along the contour segments required to close the region.
If z s denotes the identified saddle point, where g (z s ) =
0, the steepest descent path away from the saddle point will
follow the direction given by g (z) , for z = z s , cf. [15].
Typically, the singular points of the function g will be around
the imaginary axis, which indicates that the direction of the
paths of steepest descent emanating from the saddle point will
typically not deviate substantially from a direction orthogonal
to the imaginary axis. This provides a guide for setting up a
numerical integration procedure based on the path of steepest
descent. First the saddle point z s is identified. Then the path of
steepest descent starting at z s and going right is approximated
by the sequence of points {z j }
j=0 calculated as follows:
z0 = zs
where
M(z, w) eiz dz
I = I (, w) =
z1 = zs + h
(25)
A = I 2 i u 2 i v ( 21 + 12 ) + 4 v 2 V
(19)
V = 22 21 12
(20)
t = i u I + i v 12 2 v V
2
(21)
zj =
g (z j )
h,
|g (z j )|
z j+1 = z j +
z j,
j = 1, 2, . . .
j = 1, 2, . . .
where h is a small positive constant.
(26)
(27)
Similarly, the path of steepest descent going left is

approximated by the sequence {z j }
j=0 calculated by
(28)
z 1 = z s h
zj =
g (z j )
h,
|g (z j )|
z j1 = z j +
z j,
j = 1, 2, . . .
j = 1, 2, . . . .
(29)
(30)
A numerical estimate I of I can be obtained as follows.

I = I+ + I
(31)
+
+
Z ( ) = Z (ref )
where
h
I+ = exp{g(z s )} +
2
z j exp{g(z j )}
(32)
j=1
and
h
I = exp{g(z s )}
2
z j exp{g(z j )}
(33)
j=1
for a suitably large integer K .

A numerical estimate Z+ ( ) of the mean crossing rate can
now be obtained by the sum
Z+ ( ) =
criterion. By this procedure, the CPU time was reduced by

a factor of about 3. This method will be referred to as the
hybrid method, and the corresponding approximation of Z+ ( )
is denoted by Z+ ( ).
A simple approximation proposed in [16,17] is worth a
closer scrutiny. It is based on the widely adopted simplifying
assumption that the displacement process is independent of the
velocity process. This leads to an alternative approximation of
Z+ ( ), which we shall denote by +
Z ( ). It is given by the
formula
1
R
(2 )2
L
j=L
1
I (, w j ) w j
w 2j
(34)
2
2
s ;w)
g(z
x2
exp{g(z s ; w)}
(35)
which can be substituted directly into Eq. (34), leading to an

approximation of Z+ ( ), which is denoted by Z+ ( ).
This approximation can be exploited in the following way:
(1) The full method is used for an inner interval of w-values,
which contribute significantly to the integral in Eq. (22). (2) A
Laplace approximation is then used in an outer interval of wvalues where the contribution is less than significant. Of course,
the level of significance is chosen according to some suitable
(36)
where f Z denotes the marginal PDF of the surge response,

and ref denotes a suitable reference level, typically the mean
response. Here, ref has been chosen as the point where f Z
assumes its maximum, which corresponds well with the mean
response level. A general approximation for Z+ (ref ) is given
in [17]. If only slow-drift response is considered, a good
approximation is obtained by putting Z+ (ref ) 1/T0 , where
T0 = 2/0 is the slow-drift period. The advantage of Eq. (36)
is that the rhs is much faster to calculate than the exact formula.
An approximation developed by Langley and McWilliam
[18] expresses the joint PDF of Z (t) and Z (t) as a series in
the following way
f Z Z (z, s) = f Z (z) f Z (s)
where the discretization points w j are chosen to follow the

negative real axis from a suitably large negative number up to
a point at , where 0 < a, then follow a semi-circle in
the lower half plane to on the positive real axis, and finally
follow this axis to a suitably large positive number. Since the
numerical estimate does not necessarily have an imaginary part
that is exactly equal to zero, the real part operator has been
applied.
Generally, the CPU time required to carry out the
computations above can be quite long, depending on the size of
the problem, which is related to the number N of eigenvalues.
It is therefore of interest to see if approximating formulas
are accurate enough. The first such approximation we shall
have a look at is the Laplace approximation for the inner
integral over the saddle point [15]. The simplest version of this
approximation, adapted to the situation at hand, leads to the
result
I = I (, w)
f Z ( )
f Z (ref )
(n)
(m)
(1)n+m Anm f Z (z) f Z (s)
(37)
n=1 m=1
(n)
(m)
where f Z (z) is the nth derivative of f Z (z), and f Z (s) is the

mth derivative of f Z (s). f Z (z) and f Z (s) are given by infinite
series expressions, which are truncated in practice, cf. [18,
10]. The coefficients Anm are defined in terms of cumulants
for the displacement and velocity. Expressions for calculating
the mean crossing rate are provided in [19]. Since explicit
expressions are given, the computational burden incurred
by adopting this approximation is practically negligible. For
an example structure studied in [10], it is shown that the
approximation expressed in Eq. (37) gives results very similar
to Eq. (36).
5. Distribution of extreme response
To provide estimates of the extreme values of the response
process, it is necessary to know the probability law of the
extreme value of Z (t) over a specified period of time T , that
is, of M Z (T ) = max{Z (t); 0 t T }. An exact expression
for this probability law is in general unknown, but a good
approximation is usually obtained by assuming that upcrossings
of high response levels are statistically independent events.
Under this assumption, the probability distribution of M Z (T )
can be written as
Prob (M Z (T ) ) = exp{ Z+ ( ) T }
(38)
which clearly displays the crucial role of the mean upcrossing

rate Z+ ( ) in determining the extreme value distribution.
As already pointed out, Eq. (38) is only an approximation

to the true extreme value distribution. If this is so, why is
it of practical significance to be able to perform an accurate
calculation of the upcrossing rate? It is a well-known fact
that the exact extreme value distribution is not completely
determined by the upcrossing rate alone. That would be true
only when the upcrossing events of the high response levels
can be assumed to be statistically independent. Usually that
is a good approximation except when the total damping is
small. For such cases, Naess [20] has developed an effective
method to account for the effect of low damping on the extreme
value distribution, and the key to an accurate estimation of
the extreme value distribution is good estimates of the mean
upcrossing rate at high response levels.
So far the analysis has been limited to a short-term sea
state characterized by a wave spectrum. This means that the
extreme value distribution of Eq. (38) applies to the response
values within this short-term sea state. In the context of a design
situation, this would imply that the sea state has to be chosen as
a design sea state. In practice this would usually mean that the
spectral parameters would be chosen along a sea state contour
having a prescribed return period.
An alternative approach is to use a long-term analysis. Let us
assume that a short-term sea state is characterized by significant
wave height Hs and a spectral period Ts , which have a joint PDF
f Hs Ts (h, t). Then, for each set of values Hs = h and Ts = t,
the mean upcrossing rate can be calculated as we have described
above. Let it be denoted by Z+ ( | h, t). As shown by [21], the
distribution of the long-term extreme value over the time period
T is given as follows
Prob (M Z (T ) )
= exp T
0
Z+ (
| h, t) f Hs Ts (h, t) dh dt .
(39)
Haver and Nyhus [22] have proposed a model for the joint PDF
f Hs Ts (h, t).
Table 1
Main particulars of the MDF
Draught (m)
Column diameter (m)
Natural period surge/sway (s)
Natural period yaw (s)
80.0
10.0
133.5
121
Fig. 1. Computer mesh of the submerged part of the moored deep floater.
The random stationary sea state is specified by a JONSWAP

spectrum, which is given as follows
S X () =
g 2
5
exp
5 p
4
+ ln exp
6. Numerical examples
[2
1
.
+ 2i 0 + 02 ]
1
p
(41)
The numerical results presented in this section are based on

a specific example structure. It is a moored deep floater (MDF)
with main particulars as detailed in Table 1. Fig. 1 shows the
submerged part of the floater in the form of a computer mesh,
which is used for the calculation of the hydrodynamic transfer
functions. The total mass (including added mass) of the MDF is
M = 12.5106 kg. The damping ratio is set equal to = 0.06,
and the natural frequency in sway is 0 = 0.047 rad/s. Note
that the second order theory is based on the assumption that the
i j ) K 2 (i , j ), where K 2 (, )
QTF H 2 (i , j ) = L(
is a QTF characterizing the slowly varying surge forces on the
is a linear transfer function for the surge motion
MDF, and L()
of the MDF, that is
L()
=
1
2 2
(40)
where g = 9.81 ms2 , p denotes the peak frequency in

rad/s and , and are parameters related to the spectral
shape. = 0.07 when p , and = 0.09 when > p .
The parameter is chosen to be equal to 3.0. The parameter
is determined from the following empirical relationship
= 5.06
Hs
T p2
(1 0.287 ln )
(42)
Hs = significant wave height and T p = 2/ p = spectral peak

wave period. For the subsequent calculations, Hs = 10.0 m and
T p = 12 s. The natural frequency in surge is 0.047 rad/s, which
is well below the range where the waves have noticeable energy.
This is why the second order, nonlinear term in the Volterra
expansion is needed to capture the resonant motions in surge of
the MDF.

Table 2
Comparison of calculated upcrossing rate +
Z () for different step lengths
2
= /1
h = 1.0 103
h = 1.0 102
2.0
5.0
10.0
15.0
20.0
25.0
8.38 103
3.93 103
5.53 104
5.70 105
5.34 106
4.81 107
8.38 103
3.93 103
5.50 104
5.65 105
5.26 106
4.71 107
Table 3
Calculated upcrossing rates for 10 eigenvalues
Fig. 2. The 100 normalized eigenvalues j /1 .
To get an accurate representation of the response process,

there is a specific requirement that must be observed. Since
the damping ratio is only 6%, the resonance peak of the linear
transfer function for the dynamics is quite narrow. Hence,
to capture the dynamics correctly, the frequency resolution
must secure a sufficient number of frequency values over the
resonance peak. This usually leads to an eigenvalue problem
with the Q-matrix of size of the order of magnitude 100 100.
Using the full representation of this size in calculating the
mean crossing rate by the general method described here, would
lead to very heavy calculations. In order to reduce this, we
have investigated the effect of restricting the calculations by
retaining only some of the terms in Eq. (1).
For the specific example considered here, where we have
used exactly 100 (positive) frequencies, the values of the
obtained eigenvalues j have been plotted in Fig. 2. It is seen
that a substantial portion of the response variance, which is
given by Var[Z 2 (t)] = 4 Nj=1 2j , would be lost if only 10 or
20 eigenvalues were retained. This is also a factor to consider
when deciding on the number of terms to retain.
In this paper, we have focussed on the slow-drift response.
Hence, only results for Z 2 (t) will be presented. In the tables,
Z+2 ( ), Z+2 ( ), Z+2 ( ), and +
Z 2 ( ) denote the mean upcrossing
rate of Z 2 (t) calculated by the full numerical method, the
hybrid method, the Laplace approximation, and the simplified
method of Eq. (36), respectively.
To highlight the effect of the increment parameter h, Table 2
compares the results obtained by the full numerical method
for two values of h for 10 eigenvalues, that is, for a response
representation retaining the first 10 terms. The CPU time differs
by a factor of roughly 10 between the two choices of a value for
h. Since the differences between the calculated crossing rates
are fairly small, we have chosen to use the larger value to save
CPU time.
In Tables 36 we have written down the results obtained
for 10, 20, 30 and 50 eigenvalues, respectively. It is apparent
that there is some variability of the calculated upcrossing
rates depending on the number of eigenvalues included in the
analysis. Ideally, it would therefore be desirable to carry out
= /1
+
Z ()
+
Z ()
+
Z ()
2.0
5.0
10.0
15.0
20.0
25.0
8.38 103
3.93 103
5.50 104
5.65 105
5.26 106
4.71 107
8.38 103
3.93 103
5.50 104
5.65 105
5.26 106
4.71 107
7.41 103
3.59 103
5.23 104
5.59 105
5.36 106
4.92 107
Table 4
= /1
+
Z ()
+
Z ()
+
Z ()
2.0
5.0
10.0
15.0
20.0
25.0
8.09 103
4.19 103
6.51 104
6.95 105
6.56 106
5.91 107
8.10 103
4.19 103
6.51 104
6.95 105
6.56 106
5.91 107
7.27 103
3.82 103
6.06 104
6.74 105
6.58 106
6.11 107
Table 5
= /1
+
Z ()
+
Z ()
+
Z ()
2.0
5.0
10.0
15.0
20.0
25.0
7.21 103
3.58 103
5.48 104
5.85 105
5.54 106
5.00 107
7.21 103
3.58 103
5.48 104
5.85 105
5.54 106
5.00 107
6.51 103
3.27 103
5.11 104
5.67 105
5.55 106
5.17 107
the calculations with at least 50 eigenvalues. However, our

present implementation of the exact method is too expensive
computationally to allow for more than about 50 eigenvalues.
Even if this can be improved significantly, like for the
hybrid method, it would still be a method requiring heavy
computations.
To get a more detailed picture of how the crossing rate varies
with the number of eigenvalues retained, the mean upcrossing
rate was calculated for the level = 20 as a function of the
number of eigenvalues. The result is shown in Fig. 3. It was
also decided to investigate the effect of updating the truncated
response representation so that it had the correct variance. This
was achieved by multiplying the retained eigenvalues by a
suitable factor. The effect of this updating on the calculated

Table 6
Table 7
= /1
+
Z ()
+
Z ()
+
Z ()
= /1
+
Z ()
+
Z ()
+
Z ()
2.0
5.0
10.0
15.0
20.0
25.0
6.55 103
3.25 103
5.04 104
5.44 105
5.19 106
4.71 107
6.55 103
3.25 103
5.04 104
5.44 105
5.19 106
4.71 107
5.93 103
2.98 103
4.70 104
5.28 105
5.20 106
4.86 107
2.0
5.0
10.0
15.0
20.0
25.0
6.17 103
3.03 103
4.71 104
5.11 105
4.88 106
4.44 107
5.59 103
2.78 103
4.40 104
4.96 105
4.90 106
4.63 107
6.03 103
2.74 103
3.65 104
3.44 105
2.94 106
2.44 107
Fig. 3. The mean upcrossing rate of the level = 20 as a function of the

number of retained eigenvalues.
upcrossing rate is also shown in Fig. 3. The figure indicates

a couple of interesting conclusions. Updating for variance can
lead to inaccurate results for the crossing rate for small to
moderate number of eigenvalues retained. Comparing Figs. 2
and 3 it is seen that surprisingly accurate results are obtained
for even a small number of retained eigenvalues when the
truncation is done exactly where negative eigenvalues are
followed by positive eigenvalues. This seems to provide the
right balance between the terms in the response representation,
and it indicates a useful criterion for truncating the response
representation for crossing rate calculations.
It is also of great interest to observe that the simple Laplace
approximation in fact provides quite accurate estimates of
the mean upcrossing rates, and for this method the number
of eigenvalues has practically no effect on the computational
burden. Hence, from a practical point of view this is an
extremely appealing method. In Table 7 we have listed
the results obtained by the hybrid method, the Laplace
approximation and also the simple approximation of Eq. (36)
for 100 eigenvalues. It is seen that while there is excellent
agreement between the hybrid method and the Laplace
approximation, the simple approximation leads to significantly
lower values. In terms of extreme value predictions, for
the example structure at hand the Laplace approximation is
within about 1% of the hybrid method, while the simple
approximations would lead to an underestimation of typically
5%10% compared with the two more accurate methods. While
Fig. 4. The exceedance probability PF (; T ) as a function of = /1 for

T = 3, 6, 18 h for 22 and 100 eigenvalues.
Fig. 5. The exceedance probability PF (; T ) as function of = /1 for T =

3 h and 100 eigenvalues for the hybrid, Laplace and simple approximations.
the results obtained by the simple approximation are somewhat

lower than those obtained by the more accurate methods,
for some applications an accuracy within 5%10% would be
considered quite satisfactory. However, the discrepancy needs
to be checked for each specific case at hand. The results for
the extreme value distributions are summarized in Figs. 4 and
5, which show the exceedance probability PF (; T ) = 1
Prob(M Z (T ) 1 ) for the various cases.
7. Concluding remarks
In this paper we have described a general numerical method
for calculating the mean level upcrossing rate of a second order
stochastic Volterra series. The specific model considered was
the surge response of a moored deep floater in random seas.
The method has been successfully implemented, and results
have been calculated. However, it has been pointed out that
a numerical calculation for the full problem by the general
method is in many cases too demanding computationally
for practical use. Therefore, two approximations have been
investigated, and it has been shown that the approximation
obtained by using what is known as a Laplace approximation
for integrals, appears to provide estimates that are sufficiently
accurate for practical purposes. The second approximation,
which was based on a widely used simplification, did not in
general perform that well.
The main accomplishment of the present work, is that there
is now available a general numerical method for calculating the
mean upcrossing rate of any second order stochastic Volterra
system with a stationary Gaussian process as input. Since this is
the state-of-the-art representation of the motion response of any
large volume floating offshore structure subjected to random
seas, this makes it possible to calculate accurately the response
statistics of such structures.
Acknowledgements
The financial support from the Statoil Research Centre
and the Research Council of Norway (NFR) through the
Centre for Ships and Ocean Structures (CeSOS) at the
Norwegian University of Science and Technology is gratefully
acknowledged.
References
[1] Vinje T. On the statistical distribution of second-order forces and motions.
International Shipbuilding Progress 1983;30:5868.
[2] Naess A. Statistical analysis of second-order response of marine
structures. Journal of Ship Research 1985;29(4):27084.
[3] Naess A. The statistical distribution of second-order slowly-varying forces
and motions. Applied Ocean Research 1986;8(2):1108.
[4] Langley RS. A statistical analysis of low frequency second-order forces
[5] Kato S, Ando S, Kinoshita T. On the statistical theory of total second order
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
response of moored floating structures. In: Proc. 19th annual offshore

technology conference. 1987 [Paper No. 5579].
Naess A. Statistical analysis of nonlinear, second-order forces and
motions of offshore structures in short-crested random seas. Probabilistic
Engineering Mechanics 1990;5(4):192203.
Naess A, Johnsen JM. An efficient numerical method for calculating the
statistical distribution of combined first-order and wave-drift response.
Journal of Offshore Mechanics and Arctic Engineering 1992;114(3):
195204.
Naess A. Crossing rate statistics of quadratic transformations of gaussian
processes. Probabilistic Engineering Mechanics 2001;16(3):20917.
Naess A, Karlsen HC. Numerical calculation of the level crossing rate
of second order stochastic Volterra systems. Probabilistic Engineering
Mechanics 2004;19(2):15560.
Grime AJ, Langley RS. On the efficiency of crossing rate prediction
methods used to determine extreme motions of moored offshore
structures. Applied Ocean Research 2003;25(3):12735.
Naess A. The response statistics of non-linear second-order transformations to gaussian loads. Journal of Sound and Vibration 1987;115(1):
10329.
Leadbetter RM, Lindgren G, Rootzen H. Extremes and related properties
of random sequences and processes. New York: Springer-Verlag; 1983.
Rice SO. Mathematical analysis of random noise. In: Wax N, editor.
Selected papers on noise and stochastic processes. New York: Dover
Publications, Inc.; 1954. p. 133294.
Naess A. The mean rate of level crossings of a stochastic process
expressed in terms of a characteristic function. Preprint statistics no.
4/2002. Trondheim: Department of Mathematical Sciences, Norwegian
University of Science and Technology; 2002.
Henrici P. Applied and computational complex analysis, vol. II. New
York: John Wiley and Sons, Inc.; 1977.
Teigen P, Naess A. Stochastic response analysis of deepwater structures
in short-crested random waves. Journal of Offshore Mechanics and Arctic
Engineering, ASME 1999;121:1816.
Teigen P, Naess A. Extreme response of floating structures in combined
wind and waves. Journal of Offshore Mechanics and Arctic Engineering,
ASME 2003;125:8793.
Langley RS, McWilliam S. A statistical analysis of first and second
order vessel motions induced by waves and wind gusts. Applied Ocean
Research 1993;15(1):1323.
McWilliam S, Langley RS. Extreme values of first- and second-order
wave-induced vessel motions. Applied Ocean Research 1993;15(3):
16981.
Naess A. Extreme response of nonlinear structures with low damping
subjected to stochastic loading. Journal of Offshore Mechanics and Arctic
Engineering, ASME 1999;121:25560.
Naess A. On the long-term statistics of extremes. Applied Ocean Research
1984;6(4):2278.
Haver S, Nyhus KA. A wave climate description for long term response
calculations. In: Proceedings of the 9th int conference on offshore
mechanics and arctic engineering. Tokyo (Japan): ASME; 1986.
Mori
CH219.tex
5/6/2007
18: 24
Page 1
Simulation methods
Mori
CH219.tex
5/6/2007
18: 24
Page 2
Mori
CH219.tex
5/6/2007
18: 24
Page 3
Applications of Statistics and Probability in Civil Engineering Kanda, Takada & Furuta (eds)
2007 Taylor & Francis Group, London, ISBN 978-0-415-45134-5
Estimation of extreme response of dynamic structural systems subjected

to stochastic loading by direct Monte Carlo simulation
A. Naess
Centre for Ships and Ocean Structures & Department of Mathematical Sciences
Norwegian University of Science and Technology, Trondheim, Norway
O. Gaidai
Centre for Ships and Ocean Structures
ABSTRACT: The development of simple and efficient methods for estimation of the extreme response of
dynamical systems subjected to random excitations is discussed in the present paper. The key parameter for
calculating the statistical distribution of extreme response is the mean level upcrossing rate function. Exploiting
the regularity of the tail behaviour of this function, an efficient simulation based methodology for estimating the
extreme response distribution function is developed.
Keywords:
Monte Carlo simulation, mean crossing rate, extreme response, nonlinear dynamic systems.
INTRODUCTION
is determined by the mean level upcrossing rate. This

implies that the upcrossing events can be assumed to be
independent, that is, the so-called Poisson assumption
is adopted. For most practical cases this assumption
is perfectly acceptable, and will usually lead to very
accurate extreme value estimates. Exceptions to this
would be extreme value distributions for very narrow
band processes, but even in such cases the Poisson
assumption often leads to fairly good extreme value
estimates. Quite simple procedures for correcting for
band-width effects are discussed in (Naess 1990). Correction procedures very similar in spirit to the methods
described in this paper are currently being developed.
During the last decade it has become increasingly feasible to estimate the response statistics of dynamical
systems driven by stationary stochastic loading processes by standard Monte Carlo methods. However,
the problem of estimating the extreme response is still
largely considered to be too costly computationally if
done by a standard Monte Carlo technique. This is due
to the necessity of providing estimates of events with
a very low probability of occurring, thereby requiring
very long simulation times. Recent years have seen the
appearance of importance sampling techniques also
for dynamical systems (Tanaka 1998; A. Naess 2000;
Au and Beck 2003; Macke and Bucher 2003; Olsen and
Naess 2005), which can reduce substantially the computational cost. But these methods are fairly involved
and are not likely to become widely used in practical
applications.
Recognizing the huge increase in computational
power over the last 510 years, which is not showing
any signs of weakening, we consider it of some importance to develop methods based on simple, standard
Monte Carlo simulation methods that have the capability to offer robust and accurate estimates of extreme
response of quite general nonlinear dynamical systems
subjected to stationary stochastic inputs. To simplify
matters slightly, it is expedient to adopt the simplification that the statistical distribution of extreme values
2 THE STATE SPACE MODEL

The equations of motion for the dynamic system
considered is assumed to be of the form
where M denotes a generalized n n mass matrix,

X = (X1 , . . . , Xn )T = the system response vector, H a
nonlinear vector function, and F(t) denotes a stationary stochastic loading process. Hence the solution X(t)
of Eq. (1) is also a stationary vector process.
T , it is
Introducing the state space vector Z = (X, X)
seen that Eq. (1) can be rewritten in state space form as
Mori
CH219.tex
5/6/2007
18: 24
Page 4
posed: Can it be effectively estimated on the basis of

simulated response time histories? However, this may
not be the relevant question to ask. Since the desired
function is actually not the joint density fX X (, ) itself,
but rather the mean crossing rate + (; t), it is easily
realized that a better strategy could be to estimate this
last function directly from the simulated time series. It
is then expedient to rewrite Eq. (5) as
for suitable choice of h and G(t).

For specific prediction purposes, it is usually the
extreme values of one, or a few, of the component
processes of X(t) that is sought. For simplicity, let us
assume that the component in question is X (t) = X1 (t),
and let Z(t) = (X (t), X (t))T .
3 THE MEAN CROSSING RATE
Let N + (; t1 , t2 ) denote the random number of times
that the process X (t) upcrosses the level during a
time interval (t1 , t2 ). By introducing the mean rate of upcrossings of X (t) at time t, denoted by + (; t), then
under suitable regularity conditions on the response
process, which can be adopted here, the following
formula obtains
where
Here the averaged mean upcrossing rate + () is

conveniently estimated from simulated response time
histories.
4
EMPIRICAL ESTIMATION OF THE MEAN

CROSSING RATE
Based on the discussion in the previous section, the

adopted method for estimating the averaged mean
upcrossing rate + () is based on counting the number of upcrossings of the level for each simulated response time history. Let n+ (; 0, T ) denote the
counted number of upcrossings during time T from a
particular simulated time history, that is, n+ (; 0, T )
is an outcome of the random variable N + (; 0, T ).
Assuming that k time histories of length T have been
simulated, the appropriate sample mean value estimate
of + () is then
where fX (t)X (t) (, ) denotes the joint PDF of X (t) and

X (t) = dX (t)/dt. Equation (3) is often referred to as
the Rice formula (Rice 1954). + (; t) is assumed
throughout to be finite.
Note that Eq. (3) can also be written as
where X + = max (X , 0).

Let M (T ) = max{X (t); 0 t T } denote the
extreme value of the response process X (t) over the
time interval of length T . The distribution of M (T )
under the Poisson assumption is given in terms of the
mean upcrossing rate by the following relation
where n+
j (; 0, T ) denotes the counted number of
upcrossings of the level for time history no. j.
For a suitable number k, e.g. k 20 30, a good
approximation of the 95% confidence interval for the
value + () can be obtained as
where the empirical standard deviation s () is given as
which brings out the crucial role of the mean upcrossing rate X+ (; t) in determining the extreme value
distribution. To calculate the mean upcrossing rate of
the stochastic process X (t), it is necessary to find the
joint PDF fX X (, ). However, for most nonlinear systems, this function cannot be calculated analytically, or
even numerically. The following question can then be
An alternative estimate of the standard deviation can be

obtained by invoking the assumption that N + (; 0, T )
is a Poisson random variable. Its variance is given
+
as Var[N + (; 0, T )] = + ()T . Hence, s ()2 ()T .
The advantage of Eq. (10) is that it is distribution-free.
Mori
CH219.tex
5/6/2007
18: 24
Page 5
RELATION BETWEEN MEAN UPCROSSING

RATE AND MARGINAL PDF IN THE
EXTREME RESPONSE RANGE
From the previous comments concerning the function (x), it may be assumed that [1 (y)]/y will
be negligible for relevant extreme values. This leads
to the conclusion that the transformed process Y (t)
has a mean upcrossing rate that is very close to an
exponential function.
In this section we analyze the relation between the

mean upcrossing rate and the marginal PDF of a stationary response process with an emphasis on the
extreme response range, i.e. in the tail region.As is well
known, for a stationary Gaussian process the factor
E[X + |X = ] in Eq. (4) reduces to a constant. Hence,
if this constant is known, the mean upcrossing rate is
determined entirely by the PDF. Apparently, this is a
very important property when it comes to the estimation of extreme values. It is therefore of interest to
investigate to what extent the tail behaviour of mean
upcrossing rate differs from that of the marginal PDF.
The PDF fX (x) of X (t) is now written as
EXTRAPOLATION OF THE MEAN

UPCROSSING RATE
The previous section leads to two alternative ways of

setting up an extrapolation procedure for the mean
upcrossing rate function. The obvious alternative is
to use the simulated time series of X (t) to identify the
function (x) (x x0 ), and thereby providing the function (x). Transforming the time series of X (t) into
corresponding time series of Y (t) = [X (t)], the mean
upcrossing rate function Y+ (y) can then estimated as
described above. Fitting a straight line to a plot of
log (Y+ (y)/q) versus y would then provide the basis
for extrapolation.
The drawback of this procedure is the necessity to
identify the function (x). The second alternative is
therefore based on the following assumption
where (x) = a well-behaved function that is strictly

increasing for increasing x for x x0 for some x0 .
According to Eq. (4), we can write
where q = E[X + ], exp{(x)} = E[X + |X = x]/E[X + ].

Hence, q exp{(x)} expresses the mean upcrossing
rate for the case with independent X (t) and X (t). The
function (x) therefore accounts for the effect of dependence between the displacement and velocity on the
mean upcrossing rate, which is usually nonzero. However, in general, |(x)| is of much slower increase than
(x) as x . From Eqs. (11) and (12) it is seen that
where a, b and c are suitable constants. Hence, we

assume that
where q is a suitable constant. Therefore, log | log

(X+ (x)/q)| plotted versus log (x b), which we shall
call loglog-log plot, exhibits almost linear tail
behaviour, which suggests an accurate linear extrapolation strategy. In practice, a choice of value for
the parameter q would be based on looking at the
ratio X+ (x)/fX (x) = q exp{(x)}, which is usually sufficiently constant for large x to serve as a guide for a
first choice. The value of the parameter b must then be
based on optimizing the linear fit, while at the same
time keeping x0 as small as possible. The last proviso
comes from the desire to obtain accurate extrapolation, which is related to the accuracy in the slope of
the fitted straight line. In this paper the least squares
method was used to perform the linear fit. The region
over which the fitting is done, should be as long as the
linear fit allows.
As an approximation of the q it is suggested to take
an average of the ratio q 0 = X+ (x)/fX (x) from the data
tail and further optimization is done only with respect
to the single parameter b. In principle however one
might envisage the possibility to optimize q 0 as well,
if needed.
With this approach, it is demonstrated that linear
extrapolation can be achieved with good accuracy over
Plotting ln X+ (x) versus ln fX (x) will then clearly show

to what extent |(x)| is dominated by (x) as x .
This will be investigated for several case studies in the
section on numerical examples below, and it will be
demonstrated that the relation is close to linear with
slope approximately equal to one for large values of the
variable x, meaning that (x) can be largely neglected
in the tail. An alternative way to take advantage of
the property above is to introduce a new stationary stochastic process Y (t) as follows: Y (t) = [X (t)],
where ( ) is a strictly increasing function such that
(x) = (x) for x x0 . The mean upcrossing rate Y+ (y)
of Y (t) is then given by the expression
Hence, for y y0 = (x0 ), it is obtained that
Mori
CH219.tex
5/6/2007
18: 24
Page 6
several decimal orders of magnitude of + (x). Obviously, one should neglect data points from the very tail
where the confidence band width is greater than the
mean value itself (9), that is,
The linearity was observed starting from 0 with

+ (0 ) (103 , 102 ), and it was clearly developed
over 2-3 decimal orders of magnitude. It is therefore
plausible to assume that linear extrapolation can be
carried out over several orders of magnitude beyond
the simulated data.
As a result of the extrapolation technique described
above, the empirical estimation of the x-upcrossing
rate needed for a specific extreme value estimation can
be achieved with sufficient accuracy for most practical
prediction purposes with much less numerical efforts
than a direct estimation of the pertinent extreme values. Note that on the premise of a linear tail behaviour
relative to a specified plotting procedure, the estimation uncertainty is basically related to the estimation
of a slope. Hence, if the confidence band is very narrow over part of the range of linear behaviour, the
uncertainty in the slope will tend to be small, and
consequently, so will the uncertainty of the estimated
upcrossing rate at the extreme response levels.
7
Figure 1. Monte Carlo simulated upcrossing rates versus analytical results for the mean upcrossing rate on the
loglog-log scale, Duffing oscillator: Monte Carlo (); linear
fit (); confidence bands (- -); range of linearity (--).
NUMERICAL EXAMPLES
To exemplify the use of the extreme value estimation

procedure described in this paper, we shall use three
particular nonlinear dynamic systems. In all cases the
system input is white noise, therefore the Newmark
average acceleration integration procedure is applied.
This allows fast and accurate simulation due to the
possibility of using relatively large time step.
Figure 2. Monte Carlo simulated probability density versus

analytical results for PDF versus mean upcrossing rate on the
log scale, Duffing oscillator: Monte Carlo (); linear fit ().
level upcrossing rate is known in closed form, and

thus offers an easy way of verifying the extrapolation
approach. Figure 1 presents Monte Carlo simulation
results for the mean upcrossing rate on the logloglog plot. The agreement between the linear fit and the
analytical results cannot be distinguished on the plots.
Figure 2 presents Monte Carlo simulation results for
the mean upcrossing rate plotted versus marginal PDF
on the logarithmic scale, response is normalized with
its standard deviation = 0.54; b = 3.7, q = 0.28,
see (17); 20 realizations where generated, 5000 s long
each. The slope of the plotted line is 1.0, as expected.
7.1 Duffing oscillator

As the first example system, we shall use the ubiquitous SDOF Duffing oscillator excited by the stationary
Gaussian white noise W (t)
That is, Eq. (1) takes the specific form
7.2 Benchmark problem deterministic system

parameters and random excitation
where for the present example we have chosen 0 = 1,

= 1, = 0.5. Such a choice of makes Duffing
oscillator strongly non-linear. The attraction of this
model for illustration purposes is the fact that the mean
In this and the next sections we refer to a Benchmark

Study in Reliability Estimation in Higher Dimensions
Mori
CH219.tex
5/6/2007
18: 24
Page 7
of Structural Systems, Communication 4 (IfM 1994).

The case chosen for consideration here is Problem 2,
Case 1 - referred to as Duffing type oscillator: Deterministic system parameters and random excitation. In
the next section Problem 2, Case 3 will be studied. One
can find in (IfM 1994) a detailed problem description
along with an efficient MATLAB code based on the
Newmark integration method.
The primary interest in this paper is the estimation
of the failure probability Pf , for a given failure level
Yf during a time period of T seconds. If the failure
level is high enough, the Poisson independent event
assumption is applicable, meaning that
Figure 3. Monte Carlo results for the mean upcrossing

rate, 0-1st floor relative displacement, Benchmark problem:
Monte Carlo (); linear fit (); confidence bands (- -);
range of linearity (--). Circles indicate highly accurate values deduced with Poisson assumption from Benchmark site
results.
Therefore, estimation of the probability Pf is reduced

to the estimation of the corresponding mean upcrossing rate. Monte Carlo simulation provides a certain
range of mean upcrossing rates within a certain confidence limits. The range and quality of such simulated
data is directly dependent on the number of runs, i.e.
system analyses.
The structure under consideration is a 10 floor
building under specified non-stationary random wind
excitation, acting during T = 20 seconds. The dynamic
system is a Duffing type ten-degree-of-freedom oscillator. The governing equation is given by
where u is a 10 1 vector of floors displacements. M,

C, K are mass, damping and linear part of stiffness.
10 10 matrices M, C are independent of u, while
K(u) is of the Duffing type. Matrices M, C, K are
dependent on key parameters mi , ci , ki , i = 1, . . . , 10
given in (IfM 1994). F(t) represents the input excitation, which is modulated filtered Gaussian white
noise.
The quantities of interest are failure probabilities
(1)
(10)
Pf , Pf for the 01 floor and 910 floor, respectively. Wind loading within each period of T = 20 s
starts from zero and fades away to zero. Therefore
one can join series of T = 20 s together in one quasistationary process. Being aware of the genuine process
non-stationarity, as well as of first passage problem
specifics, we still apply the Poisson assumption since
the failure levels are high enough to ensure independency of upcrossing events. The latter approach
enables time averaging, where one deals with the
averaged mean upcrossing rate + and the averaged
marginal PDF f , i.e.
Figure 4. Monte Carlo results for PDF versus mean upcrossing rate, 0-1st floor relative displacement, Benchmark problem: Monte Carlo (); linear fit ( ).
where + (, t) is the mean upcrossing rate, and f (, t)

is the marginal PDF, which are obtained by averaging over different process realizations of duration T
seconds, while measuring response instantaneously t
seconds after the process start. The quantities + and f
correspond to the averaged stochastic process, which
is stationary. Therefore ideas of the present paper are
directly applicable for the estimation of the failure
probabilities. For the Monte Carlo simulation 50000
realizations 20 s duration each where used.
Figure 3 presents Monte Carlo simulation results
for the mean upcrossing rate on a loglog-log plot, 01st floor relative displacement; b = 0.01 m, q = 0.09,
see (17). Figure 4 presents Monte Carlo simulation
results for the PDF versus mean upcrossing rate on a
log scale, 0-1st floor relative displacement.
Mori
CH219.tex
5/6/2007
18: 24
Page 8

rate, 910th floor relative displacement, Benchmark problem: Monte Carlo (); linear fit (); confidence bands (- -);
range of linearity (- -). Circles indicate highly accurate values deduced with Poisson assumption from Benchmark site
results.

rate, 01st floor relative displacement, Benchmark problem:
Monte Carlo (); linear fit (); confidence bands (- -); range
of linearity (--). Random structural parameters.
7.3 Benchmark problem non-linear system with

random structural parameters and random
excitation
As in the previous section, for a complete problem
description, we refer to (IfM 1994). In this section we
study Problem 2, Case 3, corresponding to a non-linear
Duffing type system Eq. (22) with random structural
parameters and random excitation. It is assumed that
the uncertainties appear in the structural properties
(M, C, K) as well as in the input excitation. The input
excitation is the same as in the previous section. The
key elements mi , ci , ki , i = 1, . . . , 10 of matrices ((M,
C, K)) are modelled by independent Gaussian random
variables whose statistical properties are given in (IfM
1994). We just mention that these are Gaussian variables with standard deviations equal to one tenth of
their expected values. This small standard deviations
make the present case quite similar to the previous section, since the dynamic system under consideration is
the same. In order to simulate such a system, the Latin
Hypercube sampling method was applied. As before,
the estimation of the failure probability Pf is of primary interest, for a given failure level Yf during the
time period of T = 20 seconds.
The simulation results are plotted at Figure 710,
one observes strong similarity with Figure 36 due to
smallness of standard deviations in the system parameters. Parameters b and q are b = 0.021 m q = 0.09,
and b = 0.0045 m q = 0.09 for 01st and 910th floor
relative displacements respectively, see (17). Based
on 50000 system realizations, using Eq. (21) the
response level not being exceeded with a given probability P = 105 was estimated. The accuracy is again
expected to be good due to the narrow confidence band
over the part of linear behaviour shown in the figures,
which provides for accurate linear extrapolation.
Figure 6. Monte Carlo results for PDF versus mean

upcrossing rate, 910th floor relative displacement,
Benchmark problem: Monte Carlo (); linear fit ( ).
Figure 5 presents Monte Carlo simulation results for

the mean upcrossing rate on a loglog-log plot, 9-10th
floor relative displacement; b = 0.0045 m, q = 0.08,
see (17). Figure 6 presents Monte Carlo simulation
results for the PDF versus mean upcrossing rate on
the log scale, 910th floor relative displacement. Vertical dash-dotted lines indicate response range where
linearity was monitored and used for extrapolation.
Based on 50000 system realizations, using Eq. (21)
the response level not being exceeded with a given
probability P = 105 was estimated, deviating less
than 10% from the results reported in (IfM 1994).
This agreement is very satisfactory, taking into account
the relatively small Monte Carlo effort (50000 system realizations) as well as non-stationarity and the
Poisson assumption effects.
Mori
CH219.tex
5/6/2007
18: 24
Page 9
Figure 10. Monte Carlo results for PDF versus mean

upcrossing rate, 910th floor relative displacement, Benchmark problem: Monte Carlo (); linear fit ( ). Random
structural parameters.
Figure 8. Monte Carlo results for PDF versus mean upcrossing rate, 01st floor relative displacement, Benchmark problem: Monte Carlo (); linear fit ( ). Random structural
parameters.
reduced by hundreds of times, compared to straightforward Monte Carlo simulations. This should make
the proposed method of considerable practical interest.
As a last comment, it is easily realized that there
are nonlinear systems for which the proposed simulation procedure might not work very well. For instance,
if the system stiffness characteristics suddenly change
significantly when the response levels reach very high
values that are rarely captured for short simulation
times, then the crossing rate function at extreme values could be difficult to estimate from more moderate
response values. Fortunately, such structural systems
are rarely seen in engineering practice.
rate, 910th floor relative displacement, Benchmark problem: Monte Carlo (); linear fit (); confidence bands (- -);
range of linearity (--). Random structural parameters.
ACKNOWLEDGEMENTS
The financial support from the Research Council of
Norway (NFR) through the Centre for Ships and Ocean
Structures (CeSOS) at the Norwegian University of
Science and Technology is gratefully acknowledged.
Judging from Figures 3, 5, 7, 9 one can conclude that

the case of 910th floor relative displacement is more
involved than 01st floor one and therefore requires a
higher number of Monte Carlo realizations.
REFERENCES
A. Naess, C. S. (2000). Importance sampling for dynamical systems. In ICASP Applications of Statistics and
Probability, Volume 8, pp. 749755. A.A. Balkema.
Au, S. K. and J. L. Beck (2003). Importance sampling in high
dimensions. Structural Safety 25, 139163.
IfM (1994). Benchmark study on reliability estimation in
higher dimensions of structural systems. Technical Report
(see www.uibk.ac.at/mechanik/Benchmark Reliability/),
Institute of Engineering Mechanics, Leopold- Franzens
University, Innsbruch, Austria.
Macke, M. and C. Bucher (2003). Importance sampling for
randomly excited dynamical systems. Journal of Sound
and Vibration 268, 269290.
CONCLUDING REMARKS
From the variety of stochastic systems studied in the

present paper one can conclude that the optimized
extrapolation procedure proposed appears to be quite
general and robust, while it is simple for practical
use. Linear extrapolation on the double logarithmic
scale gives accurate predictions of the mean upcrossing rate and thus extreme response statistics, which
is of particular interest for reliability based design.
The CPU time is in all examples tractable, and it is
Mori
CH219.tex
5/6/2007
18: 24
Page 10
Naess, A. (1990). Approximate first-passage and extremes

of narrow-band Gaussian and non-Gaussian random
vibrations. Journal of Sound and Vibration 138(3),
365380.
Olsen, A. I. and A. Naess (2005). Importance sampling
for dynamic systems by approximate calculation of the
optimal control function. In A. Wilson, N. Limnios, S.
Keller-McNulty, and Y. Armijo (Eds.), Modern Statistical
and Mathematical Methods in Reliability, Volume 10 of
Series on Quality, Reliability and Engineering Statistics,
pp. 339352. World Scientific.
Rice, S. O. (1954). Mathematical analysis of random noise.

In N. Wax (Ed.), Selected Papers on Noise and Stochastic Processes, pp. 133294. New York: Dover Publications, Inc.
Tanaka, H. (1998). Application of an importance sampling method to time-dependent system reliability analyses using the girsanov transformation. In N. Shiraishi,
M. Shinozuka, and Y. K. Wen (Eds.), Proceedings of
ICOSSAR97 - The 7th International Conference on Structural Safety and Reliability, Volume 1, pp. 411418.
A.A. Balkema, Rotterdam.
10
NTNU
A Monte Carlo approach for

efficient estimation of extreme
response statistics
A. Naess1,2) and O. Gaidai1)
1)
Centre for Ships and Ocean Structures

2)
Department of Mathematical Sciences
Norwegian University of Science and Technology
Trondheim, Norway
A Monte Carlo approach for efficient estimation of extreme response statistics p. 1/35
Introduction
NTNU
Even with present day computational power, the problem of

estimating the extreme response of dynamic structural systems
excited by stochastic loading processes is in general not easily
done by standard Monte Carlo simulation.
Introduction
NTNU

This is due to the necessity of providing estimates of events with
a very low probability of occurring, thereby requiring long
simulation times.
Introduction
NTNU

simulation times.
Recent years have seen the appearance of importance sampling
techniques also for dynamical systems, but these are fairly
involved and are not likely to become widely used in practical
applications.
Introduction
NTNU

simulation times.
applications.
Are standard Monte Carlo methods really useless in this context?
Introduction
NTNU

simulation times.
applications.
Are standard Monte Carlo methods really useless in this context?
NOT QUITE!
The Response Process
NTNU
The equations of motion for the dynamic system considered is

assumed to be of the form
+ H X(t), X(t),
M X(t)
t = F(t),
NTNU

+ H X(t), X(t),
M X(t)
t = F(t),
M denotes a generalized n n mass matrix,
X = X(t) = (X1 (t), . . . , Xn (t))T = the system response vector,
H a nonlinear vector function,
F(t) denotes a stochastic loading process.
NTNU

+ H X(t), X(t),
M X(t)
t = F(t),
Hence the solution X(t) is also a stochastic vector process.
NTNU

+ H X(t), X(t),
M X(t)
t = F(t),
Hence the solution X(t) is also a stochastic vector process.
For specific prediction purposes, it is usually the extreme values
of one, or possibly a combination of several, of the component
processes of X(t) that is sought. For simplicity, denote it by
X(t).
The Mean Upcrossing Rate
NTNU
N + (; t1 , t2 ) = the random number of times that the process X(t)

upcrosses the level during a time interval (t1 , t2 ).
NTNU

+ (; t) = the mean rate of -upcrossings of X(t) at time t.
NTNU

Under suitable regularity conditions on the response process the
following formula obtains
+ ((; t t/2, t + t/2)]
E[N
=
+ (; t) = lim
t0
t
s fX(t)X(t)
(, s) ds
where fX(t)X(t)
(, ) denotes the joint PDF of X(t) and
X(t)
= dX(t)/dt.
NTNU

Under suitable regularity conditions on the response process the
following formula obtains
+ ((; t t/2, t + t/2)]
E[N
=
+ (; t) = lim
t0
t
s fX(t)X(t)
(, s) ds
where fX(t)X(t)
(, ) denotes the joint PDF of X(t) and
X(t)
= dX(t)/dt.
+ (; t) =
0
+ |X(t) = ] fX(t) (),

s fX(t)|X(t)
(s|) ds fX(t) () = E[X(t)
0).
where X + = max(X,
NTNU
The extreme value of the response process X(t):

M (T ) = max{X(t); 0 t T }
NTNU

M (T ) = max{X(t); 0 t T }
The distribution of M (T ) under the Poisson assumption is given by the
following relation
T
FM (T ) () = Prob(M (T ) ) = exp
+ (; t) dt
NTNU

M (T ) = max{X(t); 0 t T }
following relation
T
FM (T ) () = Prob(M (T ) ) = exp
+ (; t) dt
+
(; t) in
This brings out the crucial role of the mean upcrossing rate X
determining the extreme value distribution.
NTNU

M (T ) = max{X(t); 0 t T }
following relation
T
FM (T ) () = Prob(M (T ) ) = exp
+ (; t) dt
+
(; t) in
This brings out the crucial role of the mean upcrossing rate X
determining the extreme value distribution.
Note that the parameter of the Poisson distribution is
T
E[N + (; 0, T )] = 0 + (; t) dt.
NTNU
It is expedient to rewrite the extreme value distribution as

FM (T ) () = Prob(M (T ) ) = exp + () T ,
where
1
+
() =
T
+ (; t) dt.
NTNU
It is expedient to rewrite the extreme value distribution as

FM (T ) () = Prob(M (T ) ) = exp + () T ,
where
1
+
() =
T
+ (; t) dt.
The averaged mean upcrossing rate + () is conveniently estimated

from simulated response time histories.
Empirical Estimation of the Mean Upcrossing Rate
NTNU
n+ (; 0, T ) = the counted number of upcrossings during the time

interval (0, T ) from a particular simulated time history.
NTNU

The sample mean value estimate of + ():
1
() =
kT
+
n+
j (; 0, T )
j=1
NTNU

1
() =
kT
+
n+
j (; 0, T )
j=1
For a suitable number k, e.g. k 20 30, a good approximation of the

95 % confidence interval for the value + () is
+
conf. band() = () 1.96 s()/ k
NTNU

1
() =
kT
+
n+
j (; 0, T )
j=1
For a suitable number k, e.g. k 20 30, a good approximation of the

95 % confidence interval for the value + () is
+
conf. band() = () 1.96 s()/ k
The empirical standard deviation s() is given as
1
s() =
k1
j=1
n+
j (; 0, T )
T
()
Mean Upcrossing Rate versus PDF - stationary case
NTNU
The PDF fX (x) of X(t) is written as

fX (x) = exp {(x)} ,
where (x)= a well-behaved function that is strictly increasing
for increasing x for x x0 for some x0 .
NTNU

fX (x) = exp {(x)} ,
Now we can write
+
X
(x) = q exp {(x) + (x)} ,
where q = E[X + ], exp {(x)} = E[X + |X = x]/E[X + ].
NTNU

fX (x) = exp {(x)} ,
Now we can write
+
X
(x) = q exp {(x) + (x)} ,
where q = E[X + ], exp {(x)} = E[X + |X = x]/E[X + ].

q exp {(x)} = q fX (x) expresses the mean upcrossing rate
for the case with independent X(t) and X(t).
NTNU
Assumption: |(x)| is of much slower increase than (x) as

x .
NTNU

x .
It is seen that
+
ln X
(x) = ln fX (x) + ln q + (x)
= (x) + ln q + (x)
NTNU

x .
It is seen that
+
ln X
(x) = ln fX (x) + ln q + (x)
= (x) + ln q + (x)
+
(x) versus ln fX (x) will then clearly show to what
Plotting ln X
extent |(x)| is dominated by (x) as x .
Extrapolation of Mean Upcrossing Rate
NTNU
Assumption:
(x) = a(x b)c d(x) , x x0 ,
where a, b and c are suitable constants, and d(x) is a function of
much slower increase than (x).
NTNU
Assumption:
(x) = a(x b)c d(x) , x x0 ,
Hence, we assume that
+
X
(x) = q(x) exp{a(x b)c } , x x0 ,
where q(x) = q exp{(x) + d(x)}.
NTNU
Assumption:
(x) = a(x b)c d(x) , x x0 ,
Hence, we assume that
+
X
(x) = q(x) exp{a(x b)c } , x x0 ,
where q(x) = q exp{(x) + d(x)}.

The particular choice for the function (x) reflects the basic
assumption of an asymptotic Gumbel distribution of the extremes.
NTNU
It follows that
+
X
(x)
log log
q(x)
= c log(x b) + log a , x x0 (> b).
NTNU
It follows that
+
X
(x)
log log
q(x)
= c log(x b) + log a , x x0 (> b).
+
Hence, log log X
(x)/
q (x)
exhibits linear tail behaviour.
plotted versus log(x b)
NTNU
It follows that
+
X
(x)
log log
q(x)
= c log(x b) + log a , x x0 (> b).
+
Hence, log log X
(x)/
q (x)
Can q(x) be replaced by a constant?
NTNU
It follows that
+
X
(x)
log log
q(x)
= c log(x b) + log a , x x0 (> b).
+
Hence, log log X
(x)/
q (x)
Can q(x) be replaced by a constant? Lets try.
NTNU
It follows that
+
X
(x)
log log
q(x)
= c log(x b) + log a , x x0 (> b).
+
Hence, log log X
(x)/
q (x)

Choice of initial value q0 for q(x) would be based on looking at
+
the ratio X
(x)/fX (x) = q(x) exp{d(x)} for large x.
NTNU
It follows that
+
X
(x)
log log
q(x)
= c log(x b) + log a , x x0 (> b).
+
Hence, log log X
(x)/
q (x)

Choice of initial value q0 for q(x) would be based on looking at
+
the ratio X
(x)/fX (x) = q(x) exp{d(x)} for large x.
+
Practical solution: q0 = X
(x)/fX (x) (tail average), followed
by optimization wrt b.
Numerical Examples - Duffing oscillator
NTNU
The equation of motion is

+ 20 X + 2 X(1 + X 2 ) = W (t)
X
0
NTNU

+ 20 X + 2 X(1 + X 2 ) = W (t)
X
0
W (t) = stationary Gaussian white noise.
NTNU

+ 20 X + 2 X(1 + X 2 ) = W (t)
X
0
0 = 1, = 1, = 0.5.
NTNU

+ 20 X + 2 X(1 + X 2 ) = W (t)
X
0
0 = 1, = 1, = 0.5.
= 1 strongly nonlinear system.
NTNU

+ 20 X + 2 X(1 + X 2 ) = W (t)
X
0
0 = 1, = 1, = 0.5.
= 1 strongly nonlinear system.
fX X = fX fX is known in closed form.
NTNU
Upcrossing rates estimated from Monte Carlo simulations () with

95% confidence band () versus analytical results ( ) for the mean
upcrossing rate.
Duffing oscillator
+
0.5
1010
0.5
102
1.5
103
6
2.5
10
3
0.9 1
1.1 2
1.2
31.3
1.4 4
1.5
/
NTNU
Monte Carlo () and analytical ( ) results the mean upcrossing rate

versus PDF on the log scale. Slope = 1.0
Duffing oscillator
+0
1
101
2
102
3
103
104
4
5
105
6
106
6 6
10
5
105
4 4
10
3
103
2
102
1
101
pdf
NTNU
Numerical Examples - Hysteretic oscillator

+ 2 X + h(t) = W (t)
X
NTNU

+ 2 X + h(t) = W (t)
X
W (t) = stationary Gaussian white noise. The hysteretic restoring
force is
h(t) = 02 X + (1 )Z(t)
where = post-yielding stiffness parameter (0 1).
NTNU

+ 2 X + h(t) = W (t)
X
force is
h(t) = 02 X + (1 )Z(t)
The hysteretic component:
Z |Z|1 X|Z|
Z = |X|
+ AX
NTNU

+ 2 X + h(t) = W (t)
X
force is
h(t) = 02 X + (1 )Z(t)
Z |Z|1 X|Z|
Z = |X|
+ AX
0 = 1, = 0.05, = 0.1, A = = 1, = = 0.5.
NTNU

+ 2 X + h(t) = W (t)
X
force is
h(t) = 02 X + (1 )Z(t)
Z |Z|1 X|Z|
Z = |X|
+ AX
0 = 1, = 0.05, = 0.1, A = = 1, = = 0.5.

fX X unknown.
NTNU

Monte Carlo results for the mean upcrossing rate, 100 realizations (*)
along with 95% confidence bands () versus 50000 realizations ( ).
Hysteretic oscillator
+
0.8
1
1.2
2
1.4
10
1.6
1.8
103
2
4
10
2.2
2.4
6
102.6
2.8
2.8
2.9
3
1
3.1
3.2
2
3.33
3.4 4
3.5
5
3.6
NTNU

Monte Carlo results for the mean upcrossing rate versus PDF, 100
realizations (*) versus 50000 realizations (). Slope = 1.02
Hysteretic oscillator
+2
2.5
3
103
3.5
4
104
4.5
105
5
5.5
106
6
6 105
5 104
4 103
3 102
2 101
pdf
NTNU
Numerical Examples - Jacket structure

The Kvitebjrn jacket platform with the superstructure removed.
NTNU

The equation of motion for the horizontal excursions of the jacket
at main deck level is
+ CX + KX = Q.
MX
NTNU

+ CX + KX = Q.
MX
X = (X1 , . . . , XN )T where Xk = Xk (t), k = 1, . . . , N , denote
displacement of the k-th node xk = (xk , yk , zk ) in the wave
direction, which is the positive x-direction.
NTNU

+ CX + KX = Q.
MX
X = (X1 , . . . , XN )T where Xk = Xk (t), k = 1, . . . , N , denote
displacement of the k-th node xk = (xk , yk , zk ) in the wave
direction, which is the positive x-direction.
Q = (Q(t, x1 ), . . . , Q(t, xN ))T , where
Q(t, xk ) = Fin (t, xk ) + Fd (t, xk ), k = 1, . . . , N and
d = z1 zk zN = L d, where d = 190 m is the water
depth and L = 216 m is the jacket support height.
NTNU

The inertia force components are given as
Fin (t, xk ) = km U (t, xk )
NTNU

The drag force components
Fd (t, xk ) = kd U (t, xk ) + Uc |U (t, xk ) + Uc |
NTNU

The drag force components
Fd (t, xk ) = kd U (t, xk ) + Uc |U (t, xk ) + Uc |
km = Cm D2 /4, kd = Cd D/2
Gumbel plot of 20 simulated 3 hour extremes with fitted Gumbel

distribution. Sea state with Hs = 12 m, Tp = 12 s.
Gumbel plot, H=12m, T=12s
5
3
P0 level
ln(ln((n+1)/k))
NTNU
2
0.25
0.3
0.35
0.4
0.45 L
G
M
0.5
0.55
0.6
0.65
Gumbel plot of 20 simulated 3 hour extremes with fitted Gumbel

distribution. Sea state with Hs = 14.7 m, Tp = 15 s.
Gumbel plot for 20 3 hour maxima Mk, k=1,..,20
4
3
P0 level
2
ln(ln((n+1)/k))
NTNU
2
0.35
0.4
0.45
0.5
0.55
M
0.6 L 0.65
G
0.7
0.75
Empirical PDF of the 90% fractile value based on samples of size 20

for the sea state with Hs = 12 m, Tp = 12 s.
Bootstrapped PDF, Hs=12m, Tp=12s
12
10
PDF
NTNU
0
0.3
0.35
0.4
0.45
0.5
0.55
Level estimate
0.6
0.65
0.7
Empirical PDF of the 90% fractile value based on samples of size 20

for the sea state with Hs = 14.7 m, Tp = 15 s.
Bootstrapped PDF, Hs=14.7m, Tp=15s
12
10
PDF
NTNU
0
0.45
0.5
0.55
0.6
0.65
0.7
Level estimate (m)
0.75
0.8
0.85
Mean upcrossing rate statistics along with 95% confidence bands ()

for the sea state with Hs = 12 m, Tp = 12 s, = 0.047 m. : Monte
Carlo; : linear fit.
Sea state Hs=12m, T=12s
0
2
log10+()
NTNU
6
15
10
0
/
10
15
Mean upcrossing rate statistics along with 95% confidence bands ()

for the sea state with Hs = 14.7 m, Tp = 15 s, = 0.066 m. : Monte
Carlo; : linear fit.
Sea state Hs=14.7m, T=15s
0
2
log10+()
NTNU
6
15
10
0
/
10
15
NTNU

Transformed plot along with 95% confidence bands () for the sea
state with Hs = 12 m, Tp = 12 s, = 0.047 m. : Monte Carlo;
: linear fit, q = 0.04, b = 1.4.
+
Sea state Hs=12m, Tp=12s
2
10
0.4
0.6
0.8
1
1.2
103
1.4
1.6
4
1.8
10
2
105 P level
2.2 0
6
2.4
10
/
0.5
3
14
1.56
8 2
10
2.5
12
NTNU

Transformed plot along with 95% confidence bands () for the sea
state with Hs = 14.7 m, Tp = 15 s, = 0.066 m. : Monte Carlo;
: linear fit, q = 0.06, b = 0.9.
+
Sea state Hs=14.7m, Tp=15s
2
10
0.4
0.6
0.8
1
1.2
103
1.4
1.6
4
1.8
10
2
5
10
P level
2.2 0
6
2.4
10
/
0.5
3
14
1.56
8 2
10
12 2.5
NTNU
Numerical Examples - TLP

Sketch of submerged part of TLP.
0
5
10
15
20
25
30
20
20
10
10
0
0
10
10
20
20
NTNU

The equation of motion for the horizontal excursions of the TLP
is
+ D(t) Z(t)
MZ(t)
+ C(Z(t))
+ K(Z(t)) = F(t)
NTNU

is
+ D(t) Z(t)
MZ(t)
+ C(Z(t))
+ K(Z(t)) = F(t)
F(t) = F1 (t) + F2 (t)
NTNU

is
+ D(t) Z(t)
MZ(t)
+ C(Z(t))
+ K(Z(t)) = F(t)
F(t) = F1 (t) + F2 (t)
A simplified model for the surge response is adopted here:
NTNU

is
+ D(t) Z(t)
MZ(t)
+ C(Z(t))
+ K(Z(t)) = F(t)
F(t) = F1 (t) + F2 (t)
A simplified model for the surge response is adopted here:
1
2
3
Z + 2e (0 + cF2 (t))Z + e (Z + Z ) =
F1 (t) + F2 (t)
M
NTNU

Crossing rates by Monte Carlo simulation (*) with 95% confidence
bands () and by saddle point integration (
) for the case of
linear dynamics (
c = = 0). Sea state with Hs = 10 m, Tp = 11 s.
sea state with H =10m, T =11s
s
+1.8
3
101.9
2
2.1
2.2
104
2.3
2.4
5
102.5
2.6
106
2.7
7
102.8
/
3.7
3 3.8
3.9
4.1
5
4.2
4.3
4.4
7
4.5
94.6
NTNU

bands () and by saddle point integration (
) for the case of
linear dynamics (
c = = 0). Sea state with Hs = 15 m, Tp = 17 s.
sea state with H =15m, T =17s
s
+1.6
1.8
103
2.2
104
2.4
105
2.6
106
7
102.8
3.4
/
23.5
3.6
3 3.7
3.8
3.95
4.1
7
4.2
NTNU

bands () for the case of nonlinear dynamics, q = 0.2, b = 5.8 Z .
Sea state with Hs = 10 m, Tp = 11 s.
nonlinear TLP,sea state with H =10m, T =11s
s
+0.8
1
1.2
3
101.4
1.6
4
101.8
2
5
102.2
6
102.4
7
10
2.6
0.9
1.1
1.2
1.3
1.4
1.5
1.6
5 1.7
1.8
NTNU

bands () for the case of nonlinear dynamics, q = 0.2, b = 2.9 Z .
Sea state with Hs = 15 m, Tp = 17 s.
nonlinear TLP,sea state with H =15m, T =17s
s
1.9
10
2
2.1
4
102.2
2.3
2.4
5
10
2.5
6
102.6
2.7
7
10
0.8
1.2 4
1.4
1.6
1.8
Conclusions
NTNU
From the variety of stochastic systems studied, one can conclude

that the extrapolation procedure proposed appears to be quite
general and robust, while it is simple and practical to use.
Conclusions
NTNU

Optimized linear fit and extrapolation on a double logarithmic
scale gives accurate predictions of the mean upcrossing rate and
thus extreme response statistics.
Conclusions
NTNU

Optimized linear fit and extrapolation on a double logarithmic
scale gives accurate predictions of the mean upcrossing rate and
thus extreme response statistics.
The CPU time is in all examples tractable, and it is reduced by a
factor of 100, compared to straight-forward Monte Carlo
simulations down to the same extreme value levels.
ARTICLE IN PRESS
Ocean Engineering 34 (2007) 21882197

www.elsevier.com/locate/oceaneng
Efcient estimation of extreme response of drag-dominated offshore

structures by Monte Carlo simulation
A. Naessa,b,, O. Gaidaib, S. Haverc
a
Department of Mathematical Sciences, Norwegian University of Science and Technology, NO-7491 Trondheim, Norway
Centre for Ships and Ocean Structures, Norwegian University of Science and Technology, NO-7491 Trondheim, Norway
c
Statoil ASA, NO-4035 Stavanger, Norway
Received 30 January 2007; accepted 22 March 2007

Available online 13 April 2007
Abstract
The paper describes a novel approach to the problem of estimating the extreme response statistics of a drag-dominated offshore
structure exhibiting a pronounced dynamic behaviour when subjected to harsh weather conditions. It is shown that the key quantity for
extreme response prediction is the mean upcrossing rate function, which can be simply extracted from simulated response time histories.
A commonly adopted practice for obtaining adequate extremes for design purposes requires the execution of 20 or more 3-h time domain
analyses for several extreme sea states. For early phase considerations, it would be convenient if extremes of a reasonable accuracy could
be obtained based on shorter and fewer simulations. The aim of the work reported in the present paper has therefore been to develop
specic methods which make it possible to extract the necessary information about the extreme response from relatively short time
histories.
The method proposed in this paper opens up the possibility to predict simply and efciently both short-term and long-term extreme
response statistics. The results presented are based on extensive simulation results for the Kvitebjrn jacket structure, in operation on the
Norwegian Continental Shelf. Specically, deck response time histories for different sea states simulated from an MDOF model were
used as the basis for our analyses.
r 2007 Elsevier Ltd. All rights reserved.
Keywords: Monte Carlo simulation; Mean crossing rate; Extreme response prediction; Jacket platform
1. Introduction
For drag-dominated offshore structures like jacket and
jack-up platforms that exhibit pronounced dynamic
behaviour in severe seas, there has not been available a
simple, accurate and fast method for estimating the
extreme response values. There are in particular two
different aspects of this problem that have caused this lack
of easily accessible methods. Firstly, there is the nonlinear
characteristics of the drag force itself. Secondly, due to the
fact that the structures considered are bottom-xed, there
is the added complexity of the inundation effect. These two
complications make simple approximations very hard to
achieve, if accurate results are required.
Corresponding author.

0029-8018/$ - see front matter r 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.oceaneng.2007.03.006
Since both effects have long been acknowledged to be of

importance for the extreme response of such structures, a
substantial amount of effort has been invested over the
years in the pursuit of design procedures that would
incorporate these effects in a satisfactory manner.
A number of papers have been published that have
succeeded to some extent to solve part of the problem for
structures with signicant dynamics, to name just a few,
see, e.g. Naess and Pisano (1997), Karunakaran et al.
(2001), Liaw and Zheng (2003). In the paper by Gaidai and
Naess (2006), both aspects were taken into account in the
modelling, and the problem was solved for a simplied
SDOF model of the structure. Recently, a paper applying
FORM methodology to a similar problem succeeded in
including also second-order wave effects (Jensen and
Capul, 2006). This last paper follows to some extent the
idea underlying the Constrained NewWave methodology
ARTICLE IN PRESS
A. Naess et al. / Ocean Engineering 34 (2007) 21882197
2189
discussed at length by Cassidy et al. (2001). The papers

mentioned apply different approximations, enabling nonlinear effects to be successfully incorporated into the
analyses, but a number of assumptions have to be made
before results are obtained. It appears that among these
approximate methods, the FORM-based method is the one
which is the most versatile and accurate.
In this paper the authors aim at developing a simple,
robust and general method, which makes it possible to
tackle all nonlinear effects without any simplication
beyond those implicit in any computer model. This is
achieved by combining Monte Carlo simulations with a
CPU time-saving optimized linear extrapolation technique,
see Naess and Gaidai (2006). The resulting method has the
advantage that it ties up easily with present day design
practice, which to a large extent is based on providing
design characteristic values using time domain simulations
in combination with, e.g. the environmental contour line
principle. We will return to a brief explanation of the
environmental contour line method later in this paper.
A brute force approach often used in design is presented in
Karunakaran et al. (2001).
The work described in the present paper extends previous
work by the authors reported in Naess et al. (2007).
2. Hydrodynamic loads on the structure
To obtain a suitable computational model, the hydrodynamic forces acting on the structure are assumed to be
distributed over discrete nodes, located from the deck
structure to the sea bottom. That is, the lumped parameter
model described in Karunakaran et al. (2001) has been
adopted. Fig. 1 depicts the Kvitebjrn jacket platform with
the superstructure removed together with the corresponding 3D computer model used for the Monte Carlo
simulations. The 16 conductors installed on the platform
have been idealized as one hydrodynamic load element
extending from the bottom to the top at the centre of the
jacket.
The horizontal excursions of the structure in the
direction of propagation of a long-crested wave eld
propagating in the x-direction are studied. Therefore, only
the horizontal x-components of the hydrodynamic forces
are considered. For this conguration, the standard
engineering model of the hydrodynamic forces F t; x; z
per unit length at a given location x; y; z with a nonnegligible drag force component is the Morison formula
(Sarpkaya and Isaacson, 1981; Najaan and Burrows,
1994)
F t; x; z F in t; x; z F d t; x; z,
(1)
where the inertia force component is

_ x; z,
F in t; x; z km Ut;
(2)
and the drag force component is

F d t; x; z kd Ut; x; z U c jUt; x; z U c j.
(3)
Fig. 1. Left: Sketch of the Kvitebjrn platform with the superstructure

removed. Right: The computer model of the Kvitebjrn platform
(Karunakaran et al., 2001).
For generality, a relative velocity formulation could have

been adopted. However, for the case study in this paper the
relative velocities are of minor inuence, since the platform
oscillation amplitudes are far less than the diameter of the
main structural members. The coefcients of the inertia
and drag force terms are expressed, respectively, as
km C m rpD2 =4;
kd C d rD=2,
(4)
where r is the water density, the inertia and drag

coefcients C m and C d , respectively, can be considered to
some extent as empirical parameters depending on the
structural surface roughness. U denotes the zero-mean
water particle velocity in the horizontal direction, U c is the
constant current. In this paper only the zero current case is
studied, meaning that U c 0. The random stationary sea
state is specied by a Torsethaugen wave spectrum
(Torsethaugen, 1993). The Torsethaugen spectrum is a
model for the frequency spectrum which accounts for the
fact that the real sea typically is of a combined nature, i.e. a
wind sea system superimposed on an incoming swell
system. The sea state is assumed to be long crested.
From a given wave spectrum, a realization of the sea
surface process of a duration slightly in the excess of 3 h is
simulated. As the rst-order realization is generated, the
second-order correction is calculated, see, e.g. Marthinsen
ARTICLE IN PRESS
2190
and Winterstein (1992). This means that the surface

elevation used for the load simulations is of second order,
i.e. the crest heights are on the average slightly larger than
the depths of the wave troughs. In principle, the
corresponding kinematics could also be calculated to
second order. However, that was not done for the
simulations used in this paper. Instead, a hybrid solution
was chosen. After a second-order realization of the random
wave surface was generated, the obtained result was then
(somewhat inconsistently) written as a sum of harmonic
components. The kinematics, speed and acceleration of
the uid particles were thereafter calculated using Wheeler
stretching (Wheeler, 1970), see, e.g. Stansberg et al.
(2006). Subsequently, time histories for the nodal loads
could be calculated. This means that the resulting response
to a 3 h realization of a stationary sea state is represented
by a corresponding 3 h realization of the vector process of
nodal responses. Note that we assume throughout that
there were no wave impact loading effects on the deck
structure itself.
4. Extreme value prediction

Naess (1984a, b) showed that the crucial function for
determining a good approximation to the extreme value
distribution for both a stationary and a non-stationary
process is the mean rate of level upcrossings. Let us start
with the stationary case, and denote by N x; T the
random number of times that the process X t upcrosses
the level x during a time interval of length T. By the
assumed stationarity, EN x; T EN x; 1T, where
EN x; 1 is referred to as the mean rate of x-upcrossings
of X t. We shall use the notation n x EN x; 1.
Let MT maxfX t: 0ptpTg denote the extreme
value of the deck response process X t over the time
interval of length T. The cumulative distribution function
of MT under the Poisson assumption (Naess, 1984a) is
given in terms of the mean upcrossing rate by the following
relation for a stationary short-term sea state
ProbMTpx expfn xTg,
(6)
3. Dynamic response
In agreement with the lumped parameter model referred
in the previous section (Karunakaran et al., 2001), the
following dynamic model is adopted
CX
_ KX Q.
MX
(5)
_ and KX are suitable

Here, M denotes a mass matrix, CX
matrix functions describing the models for damping and
stiffness characteristics. X X 1 ; . . . ; X N T , where X k
X k t; k 1; . . . ; N, denotes displacement of the kth node
xk xk ; yk ; zk in the wave direction, which is the positive
x-direction. The total uid loading forces is expressed by
the vector Q Qt; x1 ; . . . ; Qt; xN T , where Qt; xk
F in t; xk F d t; xk , k 1; . . . ; N and d z1 pzk p
zN L d, where d 190 m is the water depth and L
216 m is the jacket support height. Therefore the mean
water level corresponds to z 0, z1 is the sea bottom, while
zN is the jacket deck. Each of the F in t; xk and F d t; xk
represent locally integrated forces appropriately allocated
to the nodal points. Eq. (5) represents the MDOF dynamic
model which has to be simulated in order to obtain the
deck response X N t time history. For the simulations
_
presented in this paper it was assumed that CX
_ and KX KX, that is, linear damping and stiffness.
CX
Even for the extreme displacement responses, they are still
relatively small compared to the dimensions of the
structure, and the linear dynamic model is expected to
be reasonably accurate also for the extreme sea states.
Without going into details, we just mention that the
MDOF model is also quite accurate in the sense that it
faithfully accounts for structural geometry. On the other
hand, the MDOF model becomes CPU time demanding
when the number of degrees-of-freedom N reaches high
values.
where n x denotes the mean rate of upcrossings of the

level x by the deck response process X t. Eq. (6) brings out
the crucial role of the mean upcrossing rate in determining
the extreme value distribution.
For a long-term situation, which implies a nonstationary response process, Eq. (6) is replaced by (Naess,
1984b)
Z T

ProbMTpx exp
n x; t dt ,
(7)
0
where n x; t denotes the mean level upcrossing rate at

time t, and T equals the long-term period considered. For
practical purposes, this is rewritten as (Naess, 1984b; Naess
and Moan, 2005)

Z
ProbMTpx exp T
n x; wf W w dw ,
(8)
w
where f W w denotes the long-term (ergodic) PDF of

relevant sea state parameters. In our case, W H s ; T p ,
which implies that the long-term PDF can be estimated
from smoothed joint distributions of the sea state
characteristics for the specied location.
In recent years, the environmental contour line method
has become increasingly popular as a simplied approach
to capture the long-term statistics (Karunakaran et al.,
2001). Typically, the response for sea states along a design
contour line is mapped, and the most critical sea state is
identied. The procedure is discussed in more detail below
in Section 6.
5. Empirical estimation of the mean crossing rate
To provide estimates of the extreme values of the
response process X t on the basis of simulated response
time histories, we saw in the previous section that a key to
achieve this is the estimation of the mean upcrossing rate
for each short-term sea state. By assuming the requisite
ARTICLE IN PRESS
ergodic properties of the short-term response process, the

mean upcrossing rate is conveniently estimated from the
ergodic mean value. That is, it can be assumed that
n x; T
,
T!1
T
n x lim
(9)
where n x; T denotes a realization of N x; T, that is,

n x; T denotes the counted number of upcrossings during
time T from a particular simulated time history. In
practice, k time histories of a suitable length, T 0 say, are
provided by simulation. The appropriate mean value estimate of n x is then
n^ x
k
1 X
n x; T 0 ,
kT 0 j1 j
(10)
where n
j x; T 0 denotes the counted number of upcrossings of the level x by time history no. j. This will be the
approach to the estimation of the mean upcrossing rate
adopted in this paper.
For a suitable number k, e.g. kX20 30, and provided
that T 0 is sufciently large, a fair approximation of the
95% condence interval (CI0:95 ) for the value n x can be
obtained as

s^x
s^x
CI0:95 x n^ x 1:96 p ; n^ x 1:96 p ,

(11)
k
k
where the empirical standard deviation s^x is given as
!2
k
n
1 X
j x; T 0
2
n^ x .
(12)
s^x
k 1 j1
T0
Note that k and T 0 may not necessarily be the number and
length of the actually simulated response time series.
Rather, they have been chosen to optimize the estimate
of Eq. (12). If initially, k~ time series of length T~ are
~ 0 and T~ k0 T 0 . That is, each initial
simulated, then k kk
time series of length T~ has been divided into k0 time series
of length T 0 , assuming, of course, that T~ is large enough to
allow for this in an acceptable way. The consistency of the
estimates obtained by Eq. (12) can be checked by the
observation that VarN x; t n xt since N x; t is a
Poisson random variable by assumption. This leads to the
equation
"
#
k N x; T
X
1
n x
0
j
2
,
(13)
sx Var
k
T0
T0
j1
where fN
1 x; T 0 ; . . . ; N k x; T 0 g denotes a random sample
with a possible outcome fn

1 x; T 0 ; . . . ; nk x; T 0 g. Hence,
2
s^x =k % n^ x=kT 0 . Since this last relation is consistent
with the adopted assumptions, it could have been used as
the empirical estimate of the sample variance in the rst
place. It is also insensitive to the blocking of data discussed
~ However, the advantage of Eq. (12)
above since kT 0 k~T.
is that it does not rely on any specic assumptions about
the statistical distributions involved.
2191
The idea underlying the development of the new

approach described in this paper is based on the observation that for dynamic models described by Eq. (5), the
mean x-upcrossing rate as a function of the level x is in
general highly regular. As will be shown in the next section,
the mean upcrossing rate tail, say for xXx0 , behaves very
closely like expfax bc g xXx0 where a, b and c are
suitable constants. Hence, as discussed in detail in Naess
and Gaidai (2006), it may be assumed that
n x % qx expfax bc g;
xXx0 ,
(14)
where the function qx is slowly varying compared with

the exponential function expfax bc g.
By plotting log j logn x=qxj versus logx b, it is
expected that an almost perfectly linear tail behaviour will
be obtained. Now, as it turns out, the function qx can be
largely considered as a constant. This suggests a linear
extrapolation strategy obtained by replacing qx by a
suitable constant value, q say. In practice, an initial choice
of the parameter q is extracted from an average of the ratio
n x=f X x for large values of x, because this ratio largely
determines q (Naess and Gaidai, 2006). f X x denotes the
PDF of X t. The value of the parameter b must be based
on optimizing the linear t by, e.g. least squares, while at
the same time keeping x0 as small as possible. The last
proviso comes from the desire to obtain accurate extrapolation, which is related to the accuracy in the slope of the
tted straight line. Hence the region over which the tting
is done should be as long as the linear t allows. It may be
noted that the predicted extreme values are not very
sensitive to the choice of x0 , indicating a fair amount of
robustness of the proposed procedure. However, as
discussed by Naess and Gaidai (2006), since the initial
value of q may not be optimal, it is suggested to optimize
also with respect to the value of q, if found necessary.
However, it should be noted that in all cases considered so
far, practically nothing has been gained in the accuracy of
the nal prediction by optimizing with respect to q,
emphasizing the statement above on the low sensitivity of
the method to the choice of x0 .
As a result of the extrapolation technique described
above, the empirical estimation of the x-upcrossing rate
needed for a specic extreme value estimation can be
achieved with sufcient accuracy for most practical
prediction purposes with much less numerical efforts than
a direct estimation of the pertinent extreme values. Note
that on the premise of a linear tail behaviour relative to a
specied plotting procedure, the estimation uncertainty is
related to the estimation of a slope. Hence, if the
condence band is very narrow over part of the range of
linear behaviour, the uncertainty in the slope will tend to be
small, and consequently, so will the uncertainty of the
estimated upcrossing rate at the extreme response levels.
A closer scrutiny of the data obtained from the Monte
Carlo simulations reveal that for the particular case studied
in this paper, the parameter c is very close to 1. This means
that logn x=qx plotted versus x gives a curve that is
ARTICLE IN PRESS
2192
very close to a straight line, at least in the tail. Adopting

this plotting strategy has the obvious advantage of
avoiding any optimization (if qx is easily replaced by a
constant). The accuracy of both the possible plotting
strategies described above will be tested in the next section
on numerical results.
6. Environmental contour line method
1. Combinations of signicant wave height, H s , and

spectral peak period, T p , corresponding to an annual
exceedance probability of 102 are identied. These sea
states are represented by a contour line which is
indicated in Fig. 3.
2. The worst sea state along this contour line is identied
by doing a few simulations for a number of sea states
along the contour.
3. As the worst sea state is identied, the aim is to establish
the distribution function for the 3 h target extreme
value. This will typically be done by repeating a number
X20 of 3 h simulations for this sea state. An extreme
value distribution, e.g. Gumbel, is thereafter tted to the
set of observed 3 h extremes.
4. An estimate for the response value corresponding to an
annual exceedance probability of 102 is obtained by
adopting a suitable a-percentile of the 3 h extreme value
distribution. Reasonable values for a are in the range
8595%.
The aim of this paper is to show that a reasonable
estimate for the 102 annual probability response extreme
can be obtained using less simulation time than what is
necessary by the approach reviewed above.
7. Numerical results
To demonstrate the performance of the proposed
method for extreme response prediction compared with
that of the Gumbel method of episodical extremes, the
results obtained for three representative long-crested
JONSWAP sea states from the appropriate 100-year
contour diagram, see Fig. 2 (Karunakaran et al., 2001),
as listed in Table 1, will be reported.
Twenty independent response time histories, each of 3 h
duration, were simulated for each sea state, and the
extreme horizontal deck response in the wave direction is
identied for each time series. These extreme value data are
assumed to be Gumbel distributed, and plotting each of
these data sets as a Gumbel probability plot results in
Fig. 2. Return period contour plots for the Kvitebjrn platform.
Table 1
Representative sea states
H s (m)
T p (s)
12.0
14.7
14.7
12.0
15.0
16.5
Gumbel plot, H=12m, T=12s

5
4
3
P0 level
2
ln(ln((n+1)/k))
As mentioned in the Introduction, the environmental

contour line approach is used for obtaining adequate
estimates for response values corresponding to an annual
exceedance probability of 102 ( return period of 100
years). We shall not present the method in detail, reference
is made to Kleiven and Haver (2004), here we will mainly
present the steps of this method:
1
0
1
2
0.25
0.3
0.35
0.4
0.45 LG 0.5
Mk
0.55
0.6
0.65
Fig. 3. Empirical Gumbel plot of the 20 simulated 3 h extremes for the sea
state H s 12 m and T p 12 s together with the tted Gumbel distribution ( ).
Figs. 35. Specically, the observed 3 h extremes M k are

plotted versus lnln21=k, for k 1; . . . ; 20.
The tted straight line in each gure, which represents
the tted Gumbel distribution, is based on the moment
estimation method. That is, writing the Gumbel distribution of the 3 h extreme value M3 h as

xa
ProbM3 hpx exp exp
,
(15)
b
it is known that the parameters a and b are related to the
mean value mM and standard deviation sM of M3 h as
ARTICLE IN PRESS
2193
Table 2
90% Fractile values of tted Gumbel distributions
Gumbel plot for 20 3 hour maxima Mk, k=1,..,20

4
LG (m)
CI0:95
0.47
0.618
0.632
(0.396, 0.540)
(0.548, 0.695)
(0.558, 0.726)
1
Bootstrapped PDF, Hs=12m, Tp=12s
0
12
10
2
0.35
0.4
0.45
0.5
0.55
Mk
0.6 LG 0.65
0.7
0.75
state H s 14:7 m and T p 15 s together with the tted Gumbel
distribution ( ).
PDF
ln(ln((n+1)/k))
P0 level
2
Gumbel plot Hs=14.7m, Tp=16.5s

5
0
0.3
4
3
0.4
0.45
0.5
0.55
Level estimate
0.6
0.65
0.7
Fig. 6. Thee empirical PDF of the predicted 90% fractile value based on
sample of size 20 for the sea state with H s 12 m, T p 12 s. The
indicates the limits of CI0:95 .
P0 level
2
ln(ln((n+1)/k))
0.35
Bootstrapped PDF, Hs=14.7m, Tp=15s

12
0
10
8
0.4
0.5
0.6 LG
Mk
0.7
0.8
state H s 14:7 m and T p 16:5 s together with the tted Gumbel
distribution ( ).
PDF
follows: a mM 0:57722b and b sM =1:28255 (Bury,

1975). The estimates of mM and sM obtained from the
available sample therefore provides estimates of a and b,
which leads to the tted Gumbel distribution by the
moment method.
The 90% fractile value LG of the tted Gumbel
distribution is identied and shown in each gure. The
obtained 90% fractile values are listed in Table 2. To
quantify the uncertainty associated with the obtained 90%
fractile value based on a sample of size 20, the 95%
condence interval CI0:95 of this value will be used.
A good estimate of this condence interval can be obtained
0
0.45
0.5
0.55
0.6
0.65
0.7
Level estimate (m)
0.75
0.8
0.85
Fig. 7. The empirical PDF of the predicted 90% fractile value based on
sample of size 20 for the sea state with H s 14:7 m, T p 15 s. The
by using a parametric bootstrapping method (Efron and

Tibshirani, 1993; Davison and Hinkley, 1997). In our
context, this simply means that the initial sample of twenty
ARTICLE IN PRESS
2194
Bootstrapped PDF, Hs=14.7m, Tp=16.5s
Sea state Hs=14.7m, T=15s

0
11
10
9
8
2
log10+()
PDF
7
6
5
3
4
4
3
2
1
0
0.4
0.5
0.6
0.7
Level estimate
0.8
0.9
Fig. 8. The empirical PDF of the predicted 90% fractile value based on
sample of size 20 for the sea state with H s 14:7 m, T p 16:5 s. The
Sea state Hs=12m, T=12s

0
1
log10+()
2
3
4
5
6
15
10
0
/
10
15
Fig. 9. Mean upcrossing rate statistics along with 95% condence bands
() for the sea state with H s 12 m, T p 12 s, s 0:047 m. : Monte
Carlo; : linear t.
3 h extreme values is assumed to have been generated from

an underlying Gumbel distribution, whose parameters are,
of course, unknown. If this Gumbel distribution had been
known, it could have been used to generate many samples
of size 20. For each sample, a new Gumbel distribution
would be tted and the corresponding 90% fractile value
identied. If the number of samples had been large enough,
then a very accurate estimate of the 95% condence
interval on the 90% fractile value based on a sample of size
20 could be found. Since the true parameter values of the
underlying Gumbel distribution are unknown, they are
replaced by the estimated values obtained from the initial
6
15
10
10
15
/
() for the sea state with H s 14:7 m, T p 15 s, s 0:066 m. : Monte
Carlo; : linear t.
sample. This tted Gumbel distribution is then used as

described above to provide an approximate 95% condence interval. Note that the assumption that the initial
twenty 3 h extreme values are actually generated from a
Gumbel distribution is quite accurate in the present case, as
discussed below.
Invoking the parametric bootstrap, the 95% condence
interval is estimated for each case based on 100 000 bootstrap samples. The obtained results are listed in Table 2.
This way of estimating the 90% fractile value of the 3 h
extreme value distribution will be referred to as the Gumbel
method. The empirical PDFs obtained for the predicted
90% fractile values with the CI0:95 indicated are shown in
Figs. 68.
Let us now compare the results provided by the Gumbel
method discussed above with the results obtained by the
method proposed in the present paper. Figs. 911 present
mean upcrossing rates for the deck response along with
95% condence bands on a logarithmic scale for the three
sea states from Table 1. In order to dimensionalize the
response in Figs. 914 one should multiply the nondimensionalized response x=s (given on the horizonal axis)
with the response standard deviation s. For the three sea
states from Table 1, s is as follows: 0.047, 0.066, 0.068 m.
It is observed that the left and right tails on the gures
are largely symmetric. This is primarily due to the absence
of current in these simulations. The so-called inundation
effects, arising from the wave elevation prole, which
include second-order effects, do not change the symmetry
signicantly, which is believed to be due in most part to the
use of Wheeler stretching in calculating the wave forces on
the structure. However, the dynamics of the structure will
also contribute to some extent to the observed symmetry.
The main message of Figs. 911 is a distinctive tail
linearity, suggesting, in fact, that linear extrapolation can
ARTICLE IN PRESS
Sea state Hs=14.7m, T=16.5s
2195
Sea state Hs=14.7m, Tp=15s
102
log10+()
103
104
105
P0 level
106
6
15
10
10
15
/
() for the sea state with H s 14:7 m, T p 16:5 s, s 0:068 m. :
Monte Carlo; : linear t.
10
12
Sea state Hs=14.7m, Tp=16.5s
102
Fig. 13. Transformed plot along with 95% condence bands () for the
sea state with H s 14:7 m, T p 15 s, s 0:066 m. : Monte Carlo; :
linear t, q 0:06, b 0:9s.
Sea state Hs=12m, Tp=12s

102
103
103
104
105
6
/
104
P0 level
105
106
P0 level
106
3
10
12
/
sea state with H s 12 m, T p 12 s, s 0:047 m. : Monte Carlo; :
be carried out at least two orders of magnitude down on

the logarithmic scale (dashed line on the plots). Alternatively, the general transformation approach sketched in
Section 5 may be applied, see Figs. 1214 where
log j logn x=qj is plotted versus logx b, see (14).
Both extrapolation methods agree well with each other,
and will provide accurate estimates of the extreme
upcrossing rates. The explanation of this phenomenon lies
in the statistics of the drag component UjUj of the
Morison force. This drag force exhibits a mean upcrossing
rate that is linear in the tails on a logarithmic scale. Since
the dynamics of the jacket structure under consideration,
as described by Eq. (5), is linear, the system is not expected
10
12
/
sea state with H s 14:7 m, T p 16:5 s, s 0:068 m. : Monte Carlo; :
to change signicantly the basic statistical features of the

input (force) stochastic process.
It is worth pointing out that when the mean upcrossing
rate is linear in the tails on a logarithmic scale, as it is here,
then the associated extreme value distribution can be
expected to be very close to a Gumbel distribution, which
would correspond to exactly linear tails. While it may not
at all be obvious that the data plotted in, e.g Fig. 3 come
from a Gumbel distribution, the plot in Fig. 9 strongly
supports this assumption.
Aiming at T 3 h extreme response prediction, one
needs upcrossing rates down to about 104 105 . Accurate
estimates based on direct Monte Carlo simulation down to
ARTICLE IN PRESS
2196
Table 3
90% Fractile values by the new method
LCR (m)
PI0:95
0.489
0.618
0.623
(0.464, 0.515)
(0.607, 0.630)
(0.608, 0.645)
this order is expensive in terms of CPU time. It is therefore

convenient when accurately estimated upcrossing rates
down to about 103 can be used as a basis for linear
extrapolation down to the appropriate response level x
(with n x % 105 ), as illustrated in Figs. 914.
For extreme responses, the moderate sea states do not
contribute noticeably. Therefore, severe sea states are of
particular importance. Using the contour line approach we
can limit the considerations to a few short-term 3 h sea
states along the contour line. The quantity of interest is the
response level x being exceeded with a certain probability
(usually between 5% and 15%) during the specied time
period T 3 h.
Returning now to the specic prediction of the 90%
fractile of the 3 h extreme value distribution, LCR say,
Figs. 1214. lead to the estimates listed in Table 3. The
estimated 95% condence intervals are also given in the
table, and are indicated in the gures. It is seen that they
are signicantly smaller than by the Gumbel method.
The prediction accuracy is thus signicantly higher for
the proposed method. However, it is also observed that
there is generally reasonably good agreement between the
LCR -values and the LG -values. Note also that on the
premise of a linear tail behaviour, the 95% condence
interval for LCR is completely controlled by the most
narrow part of the 95% condence band in Figs. 1214.
It is also of interest to investigate the effect of reducing
the amount of data. This is done by dividing the 20
simulated time histories into four groups of ve 3 h time
series in each group. The results for the three cases were:
Case 1: H s 12 m and T 12 s
The four groups gave the following predictions of the
90% fractile: 0.468, 0.448, 0.480, 0.514 m. The 95%
condence intervals for 0.480 m was estimated to be
(0.43,0.53), with similar results for the other three cases,
which is slightly better but largely comparable to the result
for the Gumbel method based on twenty 3 h time series.
Case 2: H s 14:7 m and T 15 s
For this case the following predictions of the 90%
fractile were found: 0.575, 0.635, 0.634, 0.615 m. The 95%
(0.59,0.64), with similar results for the other three cases,
which is signicantly better than for the Gumbel method
based on twenty 3 h time series.
Case 3: H s 14:7 m and T 16:5 s
For this case the following predictions of the 90%
fractile were found: 0.595, 0.621, 0.667, 0.632 m. The 95%
(0.60,0.68), with similar results for the other three cases.

This is again signicantly better than for the Gumbel
method based on twenty 3 h time series.
In this paper we have discussed two methods for
predicting extreme response statistics by Monte Carlo
simulation. One was the commonly adopted Gumbel
method, the other was a new method based on combining
the Monte Carlo simulations with an optimized linear
extrapolation technique. The performance of the two
methods were studied for a specic jacket structure
subjected to three different sea states chosen along a 100year environmental contour line. The 90% fractile value of
the 3 h extreme value distribution was estimated for each
sea state together with the corresponding 95% condence
interval. By comparing the various estimates obtained, it
can be concluded that the new method is signicantly more
accurate when based on the same amount of simulated
data.
Acknowledgements
The nancial support from the Research Council of
Norway (NFR) through the Centre for Ships and Ocean
Structures (CeSOS) at the Norwegian University of Science
and Technology is gratefully acknowledged. Finally, the
authors wish to thank Dr. Daniel Karunakaran, Subsea 7,
for carrying out the time domain simulations used in this
paper.
References
Bury, K.V., 1975. Statistical Models in Applied Sciences. Wiley, New
York.
Cassidy, M.J., Eatock Taylor, R., Houlsby, G.T., 2001. Analysis of jackup units using constrained NewWave methodology. Applied Ocean
Research 23 (4), 221234.
Davison, A.C., Hinkley, D.V., 1997. Bootstrap Methods and Their
Applications. Cambridge University Press, London.
Efron, B., Tibshirani, R.J., 1993. An Introduction to the Bootstrap.
Chapman & Hall, New York.
Gaidai, O., Naess, A., 2006. Extreme response statistics for drag
dominated offshore structures. In: Proceedings 5th International
Conference on Computational Stochastic Mechanics. Rhodos, Greece.
Jensen, J.J., Capul, J., 2006. Extreme response predictions for jack-up
units in second-order stochastic waves by FORM. Probabilistic
Engineering Mechanics 21 (4), 330337.
Karunakaran, D., Haver, S., Baerheim, M., Spidsoe, N., 2001. Dynamic
behaviour of Kvitebjrn jacket in the North Sea. In: Proceedings
OMAE. Rio de Janeiro, Brasil.
Kleiven, G., Haver, S., 2004. Environmental contour lines for design
purposes. In: Proceedings OMAE. Vancouver, Canada.
Liaw, C.Y., Zheng, X.Y., 2003. Polynomial approximations of wave
loading and superharmonic responses of xed structures. Journal of
Offshore Mechanics and Arctic Engineering 125 (3), 161167.
Marthinsen, T., Winterstein, S., 1992. On the skewness of random surface
waves. In: Proceedings ISOPE, San Francisco, CA.
Naess, A., 1984a. On a rational approach to extreme value analysis.
Applied Ocean Research 6 (3), 173174.
ARTICLE IN PRESS
Naess, A., 1984b. On the long-term statistics of extremes. Applied Ocean
Research 6 (4), 227228.
Naess, A., Gaidai, O., 2006. Monte Carlo methods for estimating the
extreme response of dynamical systems. Technical Report, Norwegian
University of Science and Technology, Trondheim, submitted for
publication.
Naess, A., Gaidai, O., Haver, S., 2007. Estimating extreme response of
drag dominated offshore structures from simulated time series of
structural response. In: Proceedings 26th International Conference on
Offshore Mechanics and Arctic Engineering. ASME, New York, Paper
no. OMAE200729119.
Naess, A., Moan, T., 2005. Probabilistic design of offshore structures. In:
Chakrabarti, S.K. (Ed.), Handbook of Offshore Engineering, vol. 1.
Elsevier, Amsterdam, pp. 197277 (Chapter 5).
2197
Naess, A., Pisano, A., 1997. Frequency domain analysis of dynamic

response of drag dominated offshore structures. Applied Ocean
Research 19 (56), 251262.
Najaan, G., Burrows, R., 1994. Critical assessment of the least square
error method used in derivation of Morisons force coefcients.
Journal of Offshore Mechanics and Arctic Engineering 116 (1), 16.
Sarpkaya, T., Isaacson, M., 1981. Mechanics of Wave Forces on Offshore
Structures. Van Nostrand Reinhold, New York.
Stansberg, C.T., Gudmestad, O.T., Haver, S., 2006. Kinematics under extreme
waves. In: Proceedings OMAE 2006. ASME, Hamburg, Germany.
Torsethaugen, K., 1993. A two-peak wave spectral model. In: Proceedings
OMAE. Glasgow, Scotland.
Wheeler, J.D.E., 1970. Method for calculating forces produced by
irregular waves. Journal of Petroleum Technology 22 (3), 359367.
8th ASCE Specialty Conference on Probabilistic Mechanics and Structural Reliability
PMC2000-151
THE PEAKS OVER THRESHOLD METHOD AND BOOTSTRAPPING FOR

ESTIMATING LONG RETURN PERIOD DESIGN VALUES
A. Naess, M. ASCE and P. H. Clausen
arvid.naess@bygg.ntnu.no, pal.clausen@bygg.ntnu.no
Abstract
The paper investigates the application of the peaks over threshold (POT) method (Smith 1989; Davison and
Smith 1990) in combination with the bootstrapping method (Davison and Hinkley 1997) for estimating long
return period characteristic values of environmental loads, e.g. wind loads, on the basis of observed data.
In particular, attention is focused on the effect that different statistical estimators of the parameters inherent
in the POT method have on the predicted characteristic values. The accuracy of the predicted long return
period characteristic values provided by the different methods will be evaluated by application to synthetic
data where the relevant characteristic values can be calculated. The bootstrapping method is advocated as
a practical tool for the estimation of confidence intervals on the provided point estimates of the long return
period values.
Introduction
The focus of this paper is the practical implementation of the peaks over threshold (POT)
method for the estimation of extreme values characterized by return periods that are much
longer than the periods of data acquisition. Typically, the goal is to estimate extremes of an
environmental process like for instance wind speed with a return period of say 50 or 100
years on the basis of 25 years of recorded data.
The traditional method adopted for estimating 50 or 100 year return period values is usually
based on the assumption that the distribution of yearly extreme values can be approximated
by a Gumbel or Type I extreme value distribution (Gumbel 1958; Castillo 1988). The
general procedure is then to plot the data on Gumbel probability paper and carry out some
form of linear regression on the data to fit a straight line. By restricting the database to
comprise only yearly extreme values, a great amount of potentially useful information is
discarded. This procedure can be improved as discussed by Naess (1998a). In recent years
the POT has also received considerable attention.
The POT method is based on utilizing all peak events of a given time series exceeding
a specified threshold. The word event is used deliberately, because in many practical
applications, like wind speed time series, one would consider peak values with a time
separation less than a certain value to belong to the same peak event. For example, the
passage of a low pressure system with a duration of a few days is seen as an event in this
context. By considering peak events instead of yearly maxima, the number of datapoints
for statistical processing may be increased considerably. However, in this connection it is
Naess, Clausen
appropriate to recall the basic assumption of the POT method that the threshold is high
relative to typical values. Excessive lowering of the threshold to obtain more data may lead
to substantial bias. Gross et al. (1994) have performed numerical simulations that indicate
that in samples taken from extreme value populations, optimal results are obtained by a
choice of threshold corresponding to a number of exceedances of the order of ten per year
on the average.
Naess (1998b) used a POT analysis based on estimators proposed by de Haan (1994) to
analyse wind speed data recorded at 44 US weather stations. It was shown that the predicted
50 year values of wind speed obtained by the POT method were in general significantly
below the corresponding predictions obtained by using the traditional Gumbel method.
Naess and Clausen (1999) investigated the performance of the POT method using several
different estimators. The data used for the analyses were synthetic data sampled from
a stationary Gaussian process. For such processes the 50 year values can be accurately
given. It was established that the POT method in combination with the chosen estimators
significantly underestimated the correct values.
In the present paper a new estimator for use with the POT method will be introduced. It
will be demonstrated that this estimator provides good agreement with the target values.
The problem of estimating confidence intervals for the obtained long return period values
is also discussed, and it is pointed out that the bootstrapping technique offers a practical
solution.
The Peaks Over Threshold Method
The POT method is based on what is called the generalized Pareto (GP) distribution (defined below) in the following manner: It has been shown (Pickands 1975) that asymptotically, the excess values above a high level will follow a GP distribution if and only if the
parent distribution belongs to the domain of attraction of one of the extreme value distributions. The assumption of a Poisson process model for the exceedance times combined
with GP distributed excesses can be shown to lead to the generalized extreme value (GEV)
distribution for corresponding extremes, see below. The expression for the GP distribution
is

is a scale parameter and

Here
distribution. .
(1)
determines the shape of the
The asymptotic result referred to above implies that equation (1) can be used to represent
the conditional cumulative distribution function of the excess
of the observed
variate over the threshold u, given that
for sufficiently large (Pickands 1975).
,
and
correspond to Frechet (Type II), Gumbel (Type I), and
The cases
reverse Weibull (Type III) domains of attraction, respectively. For
, the expression
between the parentheses in equation (1) is understood in a limiting sense as the exponential
.
Naess, Clausen
The return period of a given wind speed, in years, is defined as the inverse of the probability that the specified wind speed will be exceeded in any one year. In this section we give
expressions that allow the estimation from the GP distribution of the excess value above
a level corresponding to any percentage point , where is the mean crossing
rate of the threshold per year (i.e., the average number of data points above the threshold
per year). That is,
. Invoking equation (1) for
leads to
the result
. Similarly, for
, it is found that
. The
, where is the threshold
required value with the return period is then
used in the estimation of and .
Statistical Estimators
The de Haan Estimators
Let denote the total number of data points, while the number of observations above the
threshold value is denoted by . The threshold then represents the -th highest
data point(s). An estimate for is
, where denotes the length of the record
in years. The highest, second highest, ... , -th highest, -th highest variates are
denoted by , , ..., ,
, respectively.
The parameter estimators proposed by de Haan (1994) are based on the following two
quantities
Estimators for
and
and are then given by the relations
and
(2)
(3)
where
if
, while
if
. de Haan (1994) showed that under general
conditions on the underlying probability law,
and
as
(in prob.).
The Moment Estimators
In terms of the the mean value and the standard deviation of the exceedance
variate , it can be shown that (Hosking and Wallis 1987)

and

(4)
which provide the moment estimators for and .

The c0 Estimators
Naess (1998a) showed that if the underlying long term probability density of the stochastic
process giving rise to the observations can be assumed of the form
Naess, Clausen
(5)
3
where = a positive constant; and = a well-behaved function that is strictly increasing

for increasing for
for some , then, under reasonable assumptions, the data
obtained by transforming the initial data using the function are Gumbel distributed
with god approximation. That is, the transformed data will have
. The estimators
then apply to the transformed data, for which only the -parameter has to be estimated.
This leads to two estimators: One based on the de Haan estimator for . This estimator
will be referred to as the de Haan estimator. The other is besed on the moment estimator
for , and is referred to as the moment estimator.
Numerical Results
Synthetic Data
The accuracy and behaviour of the various estimation procedures discussed above, will
be studied by application to synthetic data for which the values to be estimated can be
calculated accurately. In particular, it is assumed that the underlying stochastic process
is stationary and Gaussian with mean value zero and standard deviation equal to
1.0. It is also assumed that the mean zero upcrossing rate is such that the product
, where = 1 year, which would be typical for the wind speed process. The
distribution of the yearly extreme value of is then calculated by the formula
. The 50-year return period value is then calculated from
the relation , which gives
.
The corresponding data to be used for the POT analyses, are generated by observing that the
peak events extracted from measurements of the wind speed process are usually peak events
that are separated by three to four days. This is done to obtain approximately independent
data, cf. Naess (1998b). In accordance with this, the peak event data are generated from
the extreme value distribution
, where
,
which implies that corresponds to approximately 3.65 days.
For the POT analyses, 5000 independent time series of 2000 peak events in each time series
were generated from . Each time series then corresponds to 20 years of data.
The POT analyses were carried out as follows: First, a threshold level was chosen so that
about 40 peak events exceeded this threshold. The subsequent threshold levels were chosen
so that approximately 10 new peak events were added for each new threshold. A typical
feature of the resulting estimates is a significant variation as function of threshold. To
reduce the variability and facilitate the practical choice of a suitable threshold to be used
for the prediction of , the results are smoothed by a running
filter. Specifically, in
and denotes
Figure 1 are plotted the results for , where =
the estimate for data points above the threshold. For comparison, the result obtained by
the Gumbel method for yearly extremes combined with transformation of data was 4.64,
cf. Naess (1998a). The target value is 4.65.
Figure 2 shows the mean values of the calculated estimates from the 5000 independent samples as a function of the number above the threshold. The target value of 4.65 is
Naess, Clausen
4.9
Dehaan
c0
c0 mom
mom
4.9
Dehaan
c0
c0 mom
mom
4.8
4.8
50yr
x50yr
4.7
4.7
4.6
4.6
4.5
4.5
50
100
150
200
250
Figure 1. The running mean value

as function of the number of peak events
above threshold.
50
100
150
200
250
Figure 2. The mean value of the calculated estimates from 5000 independent
samples.
accurately predicted by the Gumbel method combined with transformation of data giving
4.66. Judged from the mean values, Figure 2 indicates that the POT analysis in combination with the de Haan and the de Haan estimators lead to varying degrees of significant
underestimation of the target value. The best performance is clearly offered by the moment
and the moment estimators. On the average they are also more stable as a function of
threshold level expressed through the parameter . While the moment estimator leads on
the average to a slight underestimation, the moment estimator leads to a slight overestimation of the target value.
As mentioned in the Introduction, another topic that we want to discuss is the problem of
estimating confidence intervals associated with the predicted long return period values. In
particular, our concern is the performance of the bootstrapping method for this purpose.
To elucidate on this, for the specific choice of = 145, two empirical PDFs have been
calculated for each one of the estimators for discussed above. One is obtained from
the 5000 independent samples, providing an accurate approximation to the target PDF,
from which the confidence interval can be derived. The other is obtained on the basis of
one arbitrary sample for which 5000 nonparametric bootstrap samples have been generated.
The conclusion reached from a comparison of these PDFs is that the bootstrap method will
provide reasonably accurate estimates of the confidence intervals associated with the various estimators. In particular, there is excellent agreement between the target and bootstrap
PDFs for the two estimators. A feature of significant practical interest is the observation
that the estimators have much better efficiency than the other two estimators. This is
clearly connected to the fact that effectively only one estimated parameter enters the
estimator, viz. , while estimates of both and enter the other two estimators.
Conclusions
The main focus of this paper has been the performance of the POT method in combination
with various estimators. The goal has been the prediction of characteristic values associated
with long return periods, typically 50 or 100 years, on the basis of recorded data for shorter
Naess, Clausen
periods of time, e.g. 20 years. In particular, the attention has been directed toward the de
Haan and moment estimators applied to the data as recorded, that is, to untransformed data.
In addition, two new estimators have been proposed, which are related to transformed data.
The transformation is derived from the underlying distribution of the instantaneous value
of the process under study.
An extensive set of synthetic data have been used to test the performance of the various estimators. The overall conclusion is that the de Haan estimator for untransformed data and the
de Haan estimator for transformed data on the average lead to significant underestimation of the target values. On the other hand, the moment estimator for untransformed data
and the proposed moment estimator for transformed data, provide on the average much
better agreement with the target values. It has also been pointed out that the estimators
are superior in terms of efficiency.
References
Castillo, E. (1988). Extreme Value Theory in Engineering. San Diego: Academic Press.
Davison, A. C. and D. V. Hinkley (1997). Bootstrap Methods and their Applications. London: Cambridge
University Press.
Davison, A. C. and R. L. Smith (1990). Models of Exceedances Over High Thresholds. Journal of the
Royal Statistical Society 52(Ser. B), 393442.
de Haan, L. (1994). Extreme Value Statistics. In J. Galambos, J. A. Lechner, and E. Simiu (eds.), Extreme
Value Theory and Applications. Dordrecht: Kluwer Academic Publishers.
Gross, J. L. et al. (1994). Novel Extreme Value Procedures: Application to Extreme Wind Data. In J. Galambos, J. A. Lechner, and E. Simiu (eds.), Extreme Value Theory and Applications. Dordrecht: Kluwer Academic Publishers.
Gumbel, E. (1958). Statistics of Extremes. New York: Columbia University Press.
Hosking, J. R. M. and J. R. Wallis (1987). Parameter and Quantile Estimation for the Generalized Pareto
Distribution. Technometrics 29, 339349.
Naess, A. (1998a). Estimation of Long Return Period Design Values for Wind Speeds. Journal of Engineering Mechanics, ASCE 124(3), 252259.
Naess, A. (1998b). Statistical Extrapolation of Extreme Value Data Based on the Peaks Over Threshold
Method. Journal of Offshore Mechanics and Arctic Engineering, ASME 120, 9196.
Naess, A. and P. H. Clausen (1999). Statistical Extrapolation and the Peaks Over Threshold Method.
OMAE996422. In Proceedings 18th International Conference on Offshore Mechanics and Arctic Engineering. New York: ASME.
Pickands, J. (1975). Statistical Interference Using Order Statistics. Annals of Statistics 3, 119131.
Smith, R. L. (1989). Extreme Value Theory. In W. Ledermann, E. Lloyd, S. Vajda, and C. Alexander (eds.),
Handbook of Applicable Mathematics, pp. 437472. New York: John Wiley and Sons.
Naess, Clausen
This article was published in an Elsevier journal. The attached copy

is furnished to the author for non-commercial research and
education use, including for instruction at the authors institution,
sharing with colleagues and providing to institution administration.
Other uses, including reproduction and distribution, or selling or
licensing copies, or posting to personal, institutional or third party
websites are prohibited.
In most cases authors are permitted to post their version of the
article (e.g. in Word or Tex form) to their personal website or
institutional repository. Authors requiring further information
regarding Elseviers archiving and manuscript policies are
encouraged to visit:
http://www.elsevier.com/copyright
Author's personal copy
Probabilistic Engineering Mechanics 22 (2007) 343352

www.elsevier.com/locate/probengmech
The asymptotic behaviour of second-order stochastic Volterra series models

of slow drift response
A. Naess a,b, , O. Gaidai a
a Centre for Ships and Ocean Structures, Norwegian University of Science and Technology, Trondheim, Norway
b Department of Mathematical Sciences, Norwegian University of Science and Technology, Trondheim, Norway
Received 4 October 2006; received in revised form 27 August 2007; accepted 30 August 2007
Available online 12 September 2007
Abstract
The paper presents a detailed study of the structure and asymptotic behaviour of a second-order stochastic Volterra series model of the slow
drift response of large volume compliant offshore structures subjected to random seas. A long standing challenge has been to develop efficient
and accurate methods for calculating the response statistics of compliant offshore structures to random seas. Recent work has revealed that
the statistical properties of the response process, which consist of a linear, first-order component and a nonlinear, second-order component, is
surprisingly complex. The goal of the research work presented here is to complement efforts to develop numerical procedures to calculate the
statistics of the response process.
c 2007 Elsevier Ltd. All rights reserved.
Keywords: Stochastic Volterra series; Slow drift response; Offshore structures; Response statistics
1. Introduction
The response of compliant large volume offshore structures
to random seas is typically characterized by two timescales.
In the case of the motion response of for example a moored
drilling rig or a tension leg platform, a common feature
is the presence of slowly varying response at frequencies
well below those of the wave energy spectrum. This motion
behaviour occurs in the lateral motion modes (surge, sway,
yaw) due to the existence of the mooring system, which often
leads to resonance periods in these motion modes of the
order of minutes. The common model for the hydrodynamic
excitation forces at these periods is a nonlinear, second-order
transformation from the wave elevation process. Since the wave
elevation in this context is usually modelled as a Gaussian
process, this leads to a non-Gaussian excitation and therefore
in general to a non-Gaussian slow drift response. In such cases
the total response is consequently also non-Gaussian. As will be
shown, when the equations of motion for the floating structure
Corresponding address: CeSOS/NTNU, Otto Nielsens vei 10, NO-7491
Trondheim, Norway. Tel.: +47 73 59 46 31; fax: +47 73 59 35 24.
0266-8920/$ - see front matter c 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.probengmech.2007.08.003
can be assumed linear or linearized and time invariant, then

the response process can be represented by a second-order
stochastic Volterra series with Gaussian input.
Over the years, considerable efforts have been invested
in trying to develop methods for calculating the statistics of
a response process that can be expressed as a second-order
stochastic Volterra series. For the type of analysis considered
in this paper, the pioneering contribution is that of [1]. Their
work paved the way to a representation of the response
process that greatly facilitated further analysis. Among the
early contributions to the statistical analysis of the response
of compliant offshore structures, we mention just a few:
[26]. Later contributions include [711]. These papers contain
references to the extensive literature on this topic.
In this paper we shall expound further on a stochastic
analysis of the slow drift forces and motions given by the
first author in some previous papers [4,12]. The analysis
is particularly directed towards the prediction of large and
extreme responses. In the light of results from numerical
simulations of the slow drift response of a moored vessel
in random waves, as well as the possibility to use recently
developed numerical methods for accurate calculations [13,14],
there appears to be a need for a modification of the previous
344
A. Naess, O. Gaidai / Probabilistic Engineering Mechanics 22 (2007) 343352
analysis as given in the previous papers. At the same time this

allows us to correct an error and also to clarify the structure of
the slow drift approximation introduced in [4,12].
Wind induced excitation on the structure may also cause
slow drift response, but this can in general be considered as
a linear force transfer mechanism leading to a Gaussian force
process under standard assumptions. Modelling of stochastic
response of structures due to wind forces by means of Volterra
series is discussed in [15]. For brevity, only hydrodynamic
excitation forces are considered in this paper.
where l( ) denotes the impulse response function of the linear

dynamic model. From Eqs. (1)(3) it is seen that the total surge
response Z (t) = Z 1 (t) + Z 2 (t), where the linear part Z 1 (t) is
given as
Z 1 (t) =
0
F1 (t) =
0
k1 ( ; ) d X (t ) d,
(1)
k2 (, ; , )
d X (t ) d X (t ) d d ,
(2)
where k2 (, ; , ) is called the bidirectional quadratic

impulse response function for surge. The total surge excitation
forces F(t) on the structure is then F(t) = F1 (t) + F2 (t),
which is a general second-order stochastic Volterra series
model. This representation gives a fully adequate description
of the hydrodynamic forces accurate to second order on a large
volume structure, that is, a case where inertia forces dominate
and potential theory is applicable.
The dynamic response of a structure subjected to a
hydrodynamic loading model as detailed above can be
completely described within the framework of a second-order
stochastic Volterra series model if the equations of motion are
linear or have been linearized. It may be noted that a dynamic
analysis based on the linear equations of motion is common
practice in the evaluation of many floating offshore structures.
In most cases this invariably leads to conservative estimates of
the large and extreme responses provided the excitation process
is modelled with sufficient accuracy.
Let Z (t) denote the horizontal excursion in surge of the
moored structure. For simplicity of exposition, coupling with
other motion modes is neglected. Adopting a linear dynamic
model for the surge response, it is obtained that
Z (t) =
l( ) F(t ) d,
0
h 2 (, ; , )
(5)
The impulse response functions h 1 and h 2 are given as follows
l(t) k1 ( t; ) dt,
h 1 ( ; ) =
(6)
d X (t ) d X (t ) d d .
where d X (t) denotes the incremental directional components

adding up to give the short-crested random seas X (t), which
represents the wave elevation at some reference point. Formally,
X (t) = d X (t), cf. [8]. X (t) is assumed to be a

stationary zero-mean Gaussian process describing a short-term
sea state. k1 ( ; ) denotes the impulse response function of the
linear transfer mechanism between a long-crested wave field
propagating in the direction and the surge excitation forces.
Similarly, the nonlinear, second-order wave excitation forces in
e.g. surge can be expressed as
F2 (t) =
Z 2 (t) =
2. The dynamic model

The linear, first-order wave excitation in e.g. surge on an
offshore structure in short-crested random seas in deep water
can be expressed as a stochastic process F1 (t) as follows
(4)
and the second-order response Z 2 (t) is determined by

0
h 1 ( ; ) d X (t ) d,
(3)
and
h 2 (, ; , ) =
l(t) k1 ( t, t; , ) dt. (7)

0
3. Response process representation

The representation of Z (t) as given by Eqs. (4) and (5) is not
expedient for the statistical analysis of the total surge response.
To this end we shall adapt the results obtained by [8]. Since any
numerical analysis of the response process by necessity requires
that the system functions be provided, we need to relate to the
way this is done in practice. System functions here means either
the impulse response functions or their equivalents, which
in ocean and offshore engineering almost invariably means
their Fourier transforms. The
following Fourier transforms are
therefore introduced (i = 1)
K 1 (; ) =
k1 ( ; ) ei d,
(8)
and
K 2 (, ; , ) =
k2 (, ; , )
0
ei( + ) d d .
(9)
These two functions, generally referred to as the linear and

quadratic (surge force) transfer functions, respectively, are
calculated on a discrete grid in parameter space. Specifically,
let 0 < 0 < < L+1 be the chosen discretization of
the (positive) part of the frequency axis, and < 0 <
< M+1 < the corresponding discretization of (, ).
The corresponding spectral representation of the short-crested
random wave elevation X (t) at the chosen reference location is
then expressed as follows
L
X (t) =
j=L k=1
1
S X (| j |; k ) j k B jk ei j t .
2
(10)
Here S X ( j ; k ) denotes the directional one-sided spectral

density of the short-crested random seas. j = ( j+1
j1 )/2 and k = (k+1 k1 )/2. j = j . {B jk },
345
j = 1, . . . , L, k = 1, . . . , M, is a set of independent, complex

Gaussian N (0, 1)-variables with independent and identically
distributed real and imaginary parts. These variables also
satisfy the relations B j,k = B jk , where * signifies complex
conjugation. Also note that throughout this paper, when the
summation index runs from negative to positive values, it
invariably omits zero.
Analogously to Eqs. (8) and (9), let H 1 (; ) and
H2 (, ; , ) denote the corresponding Fourier transforms

of h 1 and h 2 , respectively. Invoking Eqs. (6)(9), it follows that
H 1 (; ) = L()
K 1 (; ) and H 2 (, ; , ) = L(
+
) K 2 (, ; , ), where L() denotes the Fourier transform
of l(t), that is, L()

is the linear transfer function for the surge
motion, which may be written as
L()
= [2 M() + i C() + K ]1 ,
(11)
where M(), C() and K are appropriate mass, damping and

stiffness coefficients, respectively. The mass and damping are
in general frequency dependent.
Substituting the expression for the wave elevation process
given by Eq. (10) into the equations for the surge response, it
can now be shown that
L
Z 1 (t) =
q jk B jk ei j t ,
(12)
j=L k=1
where
q jk = H 1 ( j ; k )
1
S X (| j |; k ) j k ,
2
(13)
and
L
Z 2 (t) =
Q i jkl Bik B jl ei(i j ) t ,
(14)
j=L j=L k=1 k=1
where
Q i jkl = H 2 (i , j ; k , l )
1
2
S X (|i |; k ) S X (| j |; l ) i j k l . (15)
As the focus of the present paper is on the first order and

second order, slowly varying response, it is appropriate to
introduce the slow drift approximation. This amounts to putting
K 2 (, ; , ) = 0
for < 0.
(16)
This condition clearly implies that all sum-frequency components of the second-order excitation forces are neglected, retaining only the difference frequency components, which are
the only terms that can give rise to the slowly varying response.
Introducing the L M L M-matrix G = (G ), where
G = Q i jkl ,
= (k 1)L + i
= (l 1)L + j
i, j = 1, . . . , L
(17)
k, l = 1, . . . , M,
now be shown that by solving the eigenvalue problem (assumed

nonsingular)
G = ,
(18)
to find the eigenvalues and the orthonormal eigenvectors (of

dimension N = L M) , = 1, . . . , N , the slowly varying
difference frequency response can be represented as
N
Z 2 (t) =
(W21 (t)2 + W2 (t)2 ).
(19)
=1
Here W (t), = 1, . . . , 2N , are real stationary Gaussian

N (0, 1)-processes, which can be determined as follows, [4]
L
W (t) =
w ( j , k ) B jk ei j t ,
(20)
j=L k=1
where the functions w ( j , k ) are defined as
( j , k )
2
w21 ( j , k ) =
1
( j , k )
j > 0
j < 0,
(21)
and
( j , k )
2
w2 ( j , k ) =
i
( j , k )
2
j > 0
j < 0,
(22)
where ( j , k ) denotes the (k 1)L + jth component of

the eigenvector . For each fixed t, the W = W (t), =
1, . . . , 2N , becomes a set of independent Gaussian variables.
Note that w2 ( j , k ) = ( j ) w21 ( j , k ), where () =
i for > 0, ( (0) = 0) and () = i for < 0.
() is recognized as the transfer function of the Hilbert
transform. Hence it follows that W2 (t) is the Hilbert transform
of W21 (t). That is, W21 (t)2 + W2 (t)2 is the square of the
Hilbert envelope of W21 (t), pointing to the slowly varying
characteristics of Z 2 (t).
Having achieved the representation of the slowly varying
response Z 2 (t) given by Eq. (19), we may proceed to establish
a similar representation of the first-order response Z 1 (t). The
following relation obtains,
N
(d21 W21 (t) + d2 W2 (t)),
Z 1 (t) =
(23)
=1
where the real expansion coefficients d are given by the

equation
L
H 1 ( j ; k )
d21 i d2 =
j=1 k=1
it can be demonstrated that G becomes a Hermitian matrix. By

adapting the analysis in [8] to the present formulation, it can
S X (| j |; k ) j k ( j , k ) .
(24)
346
The total surge response process may now be written succinctly

as
f W1 ...Wn Z (w1 , . . . , wn , z ) = f Z |W1 ...Wn (z |w1 , . . . , wn )

f W1 ...Wn (w1 , . . . , wn ).
Z (t) =
( W (t)2 + d W (t)),
(25)
From the equation
=1
where n = 2N , and the eigenvalue parameters 21 =

2 = . This response process representation facilitates
significantly the statistical analysis we seek to carry out. It may
be noted that the PDF of the slow drift response Z 2 (t) can be
calculated in closed form [16]. Unfortunately, the same does not
apply to the total response Z (t). However, accurate numerical
methods are now available for calculating the PDF of Z (t),
cf. [17].
4. The asymptotic behaviour of the mean crossing rate
The statistics of high level excursions and extreme values
of the response process Z (t) are largely determined by the
mean upcrossing rate Z+ ( ) for large values of . The mean
upcrossing rate is given by the Rice formula
Z+ ( ) =
(31)
s f Z Z (, s) ds.
(26)
In this section asymptotic bounds on the mean upcrossing

rates of the response process Z (t) will be established. For
convenience it will be assumed that the eigenvalues have been
arranged so that 1 = 2 > | j | for j = 3, . . . , n. To obtain
the desired asymptotic results, we shall proceed as follows. By
the law of marginal probability (Z = Z (t), W j = W j (t))
f Z Z (z, z ) =
f W2 ...Wn Z Z
(w2 , . . . , wn , z, z ) dw2 . . . dwn .
f W2 ...Wn Z Z (w2 , . . . , wn , z, z )
w1+
z
+ f W1 W2 ...Wn Z (w1 , w2 , . . . , wn , z )
w1
z
11
21
(33)
it is seen that 11 = I , where I denotes the identity matrix of

order n. The joint PDF of W given W = w = (w1 , . . . , wn )
is multivariate normal with mean value 21 w and covariance
matrix 22 21 12 . Hence the expected value of { Z |W = w}
is given as
= (w)
= E[ Z |W = w] = p 21 w,
(34)
and the corresponding variance as

2 = (w)2 = Var[ Z |W = w] = p (22 21 12 ) p,
(35)
where p = (21 w1 + c1 , . . . , 2n wn + cn ) . The PDF of the
conditional random variable { Z |W1 = w1 , . . . , Wn = wn } is
given as
f Z |W (z |w) =
1
2
exp
(z )
2
.
2 2
(36)
Each of the random variables W j is an N (0, 1)-variable.

Accordingly, their independence implies that the joint PDF of
W is
(w j ),
(37)
j=1
(28)
(29)
j=2
which is w1 = c1 /(21 ) [(z + c12 /(41 ) q(w))/1 ]1/2 ,

where q(w) = nj=2 ( j w 2j + c j w j ). Hence
where (x) = (21)1/2 exp{ x2 } denotes the PDF of an N (0, 1)variable.

Introducing the parameters = (w
1 , . . . , wn ) and
= (w1 , . . . , wn ), it follows from the equations above that
Z+ ( ) =
(30)
The joint PDF f W1 W2 ...Wn Z is determined by applying the

decomposition
1
(2)n/2
w1
1
= [1 (z + c12 /(41 ) q(w))]1/2 .
z
2
12
22 ,
( j w 2j + c j w j ),
it is seen that the conditional random variable { Z |W1 =

w1 , . . . , Wn = wn } is a sum of jointly Gaussian variables
and therefore itself a Gaussian variable. To derive its mean
and variance, consider the multi-dimensional random vector
W = (W1 , . . . , Wn ) and W = (W 1 , . . . , W n ) , where the
prime indicates transposition. Writing the covariance matrix
of (W , W ) as
f W (w) =
Here w1 denote the two possible solutions of the equation

1 w12 + c1 w1 = z
(32)
j=1
(27)
By invoking the law of transformation of variables, the

following result is obtained
= f W1 W2 ...Wn Z (w1+ , w2 , . . . , wn , z )
2 j W j W j + c j W j ,
Z =
...
exp{ 12 (w22 + + wn2 )}

2[1 ( + c12 /(41 ) q(w))]1/2
s +
+
1
1
exp
+
1
1
exp

2
s ds dw2 . . . dwn .
1
(w1+ )2
2
1
(w1 )2
2
(38)
347
The domain of integration is determined by the relation

+ (c1 /21 )2 q(w) > 0 (1 > 0). It can be shown that
1
2
exp
1
2
It is also the case that s j ri j = si r ji , that is, 12 = 21 =

12 . This is seen as follows,
s j ri j = E[Wi W j ]
x dx
= (/) + (/) = G(/),
(39)
where and are constants and (x) =

The
function G(x) is defined by the last equality in Eq. (39).
Recalling that 1 = 2 = 1 , and making the assumption that
1 > | j |, j = 3, 4, . . . , it can now be shown that
1
exp
21
(2 )n/2
n
exp 21
...
2[1 (
=+,
c1
21
ik w j (k , l ) Bk l eik t
k =L l =1
L
ik wi (k , l ) w j (k , l )
and, since this equation holds for any i, j = 1, . . . , n, it also

follows that
(l j w 2j c j w j /1 )
j=3
ik wi (k , l ) w j (k , l ) .
c1
c2
w2
[ + c12 /(41 ) q(w)]1/2 +
21
21 1
dw2 . . . dwn ,
(40)
where l j = 1 j /1 and = / . Eq. (40) will
now be used to study the asymptotic behaviour of Z+ ( ) as
. To this end it is necessary to first establish the
asymptotic behaviour of and . The following parameters
are introduced
r jk
E[W j W k ]
= qk j .
s j sk
(41)
The last equation follows from the fact that si r ji is a real

parameter, hence complex conjugation does not change its
value. Comparison of the last term of the two equations reveals
that
s j ri j = si r ji ,
i, j = 1, . . . , n.
j = 1, . . . , N .
(42)
ik w2i (k , l ) w2 j (k , l )
s2 j r2i,2 j =
k=L l=1
L
and
=
n
= Sk j ,
j = k.
(43)
l=1
k=L l=1
L
ik w2i1 (k , l ) w2 j1 (k , l )
k=L l=1
= s2 j1 r2i1,2 j1 ,
ik w2i1 (k , l ) w2 j (k , l )
k=L l=1
k2 | w2 j1 (k , l ) |2 = s22 j1 .
k=L l=1
(50)
s2 j r2i1,2 j =
k=L l=1
k2 |(k ) w2 j1 (k , l )|2
L
ik (k ) w2i1 (k , l )
k=L l=1
(k ) w2 j1 (k , l )
since |(k )|2 = 1. From the relation s2 j = s2 j1 , it follows

that r2i,2 j = r2i1,2 j1 . Further, it follows that,
k2 |w2 j (k , l )|2
Some relations between these parameters will be useful in

the sequel. First we prove that s2 j = s2 j1 , j = 1, . . . , N . This
holds since
s22 j =
(49)
From Eq. (45) it is obtained that
k=1
(48)
and
rl j rlk
and r2i1,2 j = r2i,2 j1 ,
r2i,2 j = r2i1,2 j1
S jk = s j sk q jk
(47)
The following results will now be proved
q2 j1,2 j = 0 ,
rk2j ,
(46)
k=L l=1
i, j = 1, . . . , N ,
Writing S = (S jk ) = 22 21 12 , it is found that

S j j = s 2j = s 2j 1
ik w j (k , l ) wi (k , l )
k=L l=1
G( )
E[W j W k ]
=
,
sk
si r ji = E[W j W i ] =
+ c12 /(41 ) q(w))]1/2
s 2j = Var[W j ] = E[W j2 ],
(45)
k=L l=1
exp
q jk =
wi (k , l ) Bkl eik t
k=L l=1
x
(t) dt.
Z+ ( ) =
=E
(44)
ik w2i1 (k , l ) (k ) w2 j1 (k , l )
=
k=L l=1
348

L
ik (k ) w2i1 (k , l ) w2 j1 (k , l )
S j j (2 j w j + c j )2
+
j=2
k=L l=1
L
Si j (2i wi + ci )(2 j w j + c j ).
+2
ik w2i (k , l ) w2 j1 (k , l )
(58)
2i< jn
k=L l=1
= s2 j1 r2i,2 j1 ,
(51)
since (k ) = (k ), giving the asserted equation

r2i1,2 j = r2i,2 j1 . From Eq. (48) the following two relations
can now be derived
r2i1,2 j1 r2i1,2 j + r2i,2 j1 r2i,2 j = 0,
i, j = 1, . . . , N ,
(52)
Exploiting the fact that for the slow drift approximation

typically |r2 j1,2 j | = |r2 j,2 j1 | 1, while all the other
correlation coefficients are approximately zero, the following
approximate relation obtains ( = /1 )
s1 r12 c2 w1 c1 w2
s2 j1r2 j1,2 j c2 j w2 j1 c2 j1 w2 j
and
j=2
2
2
2
2
r2i1,2
j1 + r2i,2 j1 = r2i1,2 j + r2i,2 j ,
i, j = 1, . . . , N .
(53)
21 s1r12 c2 + c12 q(w)/1 c1 w2 c1 c2
s2 j1r2 j1,2 j c2 j w2 j1 c2 j1 w2 j
+ 21
This leads to the two results
j=2
j = 1, . . . , N ,
ri,2 j1 ri,2 j = 0,
(54)
21 s1r12 c2 + c12 + c12 (w2 + c2 )2
i=1
and
c1 (w2 + c2 ) ,
n
2
(ri,2
j1
2
ri,2
j)
= 0,
j = 1, . . . , N ,
(55)
i=1
which implies that s22 j1 = s22 j .

Finally,
L
k2 w2 j1 (k , l ) w2 j (k , l )
s22 j1 q2 j1,2 j =
k=L l=1
L
k2 (k ) |w2 j1 (k , l )|2 = 0,
(56)
where the notation c j = c j /(21 ) has been introduced. It

has also been assumed that c1 c2 = 0. The asymptotic
relation above was based on the observation that l j > 0
for j = 3, . . . , n, which implies that the main contribution
to the multiple integral over the variables w3 , . . . , wn in Eq.
(40) comes from a bounded region in Rn2 . The following
asymptotic relation is also established.
( )2
421 [( + c12 + c22 (w2 + c2 )2 )s12
k=L l=1
since (k ) = (k ) and |w2 j1 (k , l )| =

|w2 j1 (k , l )|, giving q2 j1,2 j = 0, j = 1, . . . , N .
We now proceed to calculate the expressions for and
. Let p = (21 w1 + c1 , . . . , 2n wn + cn ) and w =

(w1 , . . . , wn ) . It is found that
+ (w2 + c2 )2 s22 ] = 4 21 s12 [ + c12 + c22 ],
+ s1 r12 c2 w1 c1 w2 + c2 s1
r j2 w j
(2k wk + ck ) sk r jk w j ,
j=3 k=3
where l j = 1 j /1 .
2
( ) = ( p ) S p =
+ 2(21 w1
(21 w1
n
+ c1 )
+ c1 ) S11
2
S1 j (2 j w j + c j )
j=2
+ c22
c2 + c12 + c12 (w2 + c2 )2
(61)
where the parameter jk = r jk / 1
n
2
l=1 rlk .
Asymptotically the domain of integration can be replaced
+ c12
c1 (w2 + c2 ) ,
j=3
n
12
(w1 r j1 + w2r j2 ) l j w j c j /(21 )
j=3
(60)
since s22 = s12 . This leads to the asymptotic relation
= ( p ) 21 w
= 21 s1
(59)
(57)
by ( + c12 + c22 c2 , + c12 + c22 c2 ) nj=3 (t j , t j ),

where t j = for j = 3, . . . , n. The following asymptotic
relation is obtained.
Z+ ( )
s1 + c12 + c22
(2)n/2
+c12 +c22
+c12 +c22
exp c12 c22

2
dw3 . . .
du
dwn
349
exp
1
2
exp c1 [
[
(l j w 2j 2c j w j )
j=3
+ c12 + c22 u 2 ]1/2

+ c12 + c22 u 2 ]1/2
=+,
+ c2 u
upcrossings of high response levels are statistically independent

events. Under this assumption, the probability distribution of
M Z (T ) can be written as
G( )
Prob (M Z (T ) ) = exp{ Z+ ( ) T },
(62)
where the new integration variable u = w2 + c2 has been

introduced. The integration with respect to w3 , . . . , wn in Eq.
(62) can be carried through in closed form. This leads to the
+ c12 + c22 x)
expression (u =
Z+ ( )

n
s1 + c12 + c22
2
exp{c2j /(2l j )}
j=3
= exp + eb
exp c12 c22

2
1 =+,
+ =
dx.
(63)
To finalize the behaviour of the last integral as it is seen

that the term with c1 > 0 is the one providing the dominating
contribution to the integral. The asymptotic value of the integral
is found by using Laplaces method [18]. It gives (N > 1,
= /1 )
s1
2
Z+ ( )
exp{(c22 j1 + c22 j )/(2k j )}
+ c12 + c22
c12 + c22
+ c12 + c22
c12 + c22
, (64)
where the parameters k j = 1 j /1 , j = 1, . . . , N , have

been introduced.
To find the asymptotic behaviour of the upcrossing rate
n+
Z 2 ( ) in the case of a pure second-order response, that is,
Z 2 (t), it is not legitimate to put all c j = 0 in Eq. (64). Instead,
we have to go back to Eq. (62). Carrying through an asymptotic
analysis similar to the one leading to Eq. (64) for the present
case gives the expression (N > 1, = /1 )
N
j=2
1
kj
exp
y
1
+ b y + ln y
2
4
exp{(c22 j1 + c22 j ) /(2k j )}

kj
j=2
(65)
5. Extreme value prediction

To provide estimates of the extreme values of the response
process, it is necessary to know the probability law of the
extreme value of Z (t) over a specified period of time T , that
is, M Z (T ) = max{Z (t); 0 t T }. An exact expression for
this probability law is in general unknown, but a suitable, and
usually accurate, approximation is obtained by assuming that
(68)
Following the analysis in [4], it can be shown that the mean

value of Y is given by the expression
y dPY (y) y0 +
E[Y ] =
0
,
(y0 )
(69)
where (y) = 2y b y 41 ln y ln( + eb /2 T ), (y) =

d(y)/dy, and y0 is the solution of the equation (y) = 0. It
then follows that [4]
E[Y ] = Q + 2b Q +
1/4
exp c12 c22 +

2
2
s1
n+
Z 2 ( )
2
s1
2 b
kj
j=2
T exp
where
G( )
+ c12 + c22 c1 1 x 2 + c2 x
1 x2
2 /2
(67)
exp
where the expression for Z+ ( ) given by the rhs of Eq. (64) will
be used.
Let Y = 1 M Z (T ) + b2 , where b2 = c12 + c22 . Then the
probability distribution of Y is given as
PY (y) = Prob (Y y)
lj
(66)
1
ln Q
2
b
1+
Q
b
+ 2 + 2b2 + (1 + b2 2 ) + O
Q
1
Q
, (70)
where Q = 2 ln( + eb /2 T ). The corresponding expression

for the mean value of M Z (T ) is given as
2
E[M Z (T )] = 21 ln( + T ) + 1 2b Q +
b
1+
Q
1
ln Q
2
b
+ 2 + (1 + b2 2 )
+O
Q
1
Q
.
(71)
An entirely similar analysis for the pure slow drift response

Z 2 (t) gives, with M Z 2 (T ) = max{Z 2 (t); 0 t T },
Prob M Z 2 (T ) = exp{n +
Z 2 ( ) T }
= exp n + T exp
+ ln
21
2
1
(72)
where
n+ =
2s1
1
.
k
j=2 j
This leads to the formula
(73)
350
Table 1
Particulars of the TLP
Column diameter D (m)
Eigenperiod surge Te (s)
Relative damping
Total mass (incl. added mass) M (kg)
E[M Z 2 (T )] = 1 Q 2 + ln Q 2 + 2
+O
1
3/2
Q2
10.0
128.8
0.05
1.5 107
1
1
Q2
(74)
where Q 2 = 2 ln(n + T ). It may be noted that [4]

Var[M Z (T )] Var[M Z 2 (T )] 2 2 21 /3.
6. Numerical example
In this section asymptotic estimates of the mean upcrossing
rate will be compared with exact values, computed using the
saddle point method [13]. A particular model of an offshore
tension leg platform (TLP) is considered and the corresponding
linear transfer function (LTF) and quadratic transfer function
(QTF) are computed using a second-order diffraction program
(WAMIT). For simplicity, unidirectional seas are considered,
meaning that the directional argument is skipped. This
simplification should have no effect on the conclusion based on
the comparison of accuracy. The slow drift surge deck motion
is studied applying a single degree of freedom model. The TLP
particulars are detailed in Table 1, these are used to obtain
second-order response. As one can see, for the second-order
response we have used a simplified version of (11), where
mass M and damping coefficient = C Te /4 are frequency
independent, which is a good approximation for the slow drift
motion. However, for the linear response a more accurate
frequency dependent transfer function (11) was applied directly
in WAMIT.
The discrete frequency range is the following: 1 =
2/30.0, . . . , n = 2/4.0 (rad/s), n = 30. To get an
accurate representation of the response process, there is a
specific requirement that must be observed. Since the damping
ratio is only 5%, the frequency resolution must secure a
sufficient number of frequency values over the resonance peak.
This will ensure that the second-order difference frequency
response component captures the TLP surge dynamics with
sufficient accuracy. It is commonly required that there are at
least 5 discrete frequencies over the frequency range where
| L()|
is equal to or higher than half of the resonance peak
height max[| L()|],

1 < < n , see Eq. (11). For the surge
force QTF K 2 (, ) (9) a suitable initial frequency grid must
be chosen for which the values of the force QTF are calculated.
The calculation of the force QTF is generally by far the most
time consuming part of the numerical analysis. Therefore the
initial grid is usually rather coarse to avoid excessive computer
time. This necessitates the use of an interpolation procedure
to be able to provide values of the QTF on a much finer
grid than the initial one to comply with the requirement of
sufficient frequency resolution to capture the dynamics of slow
drift motion. In this paper cubic spline interpolation has been

used. In the particular case considered here, the resolution
requirement led to the choice = 0.0018 rad/s and
L = 760 interpolated discrete frequencies. Consequently the
corresponding G-matrix (17) and its eigenvalue problem are of
size L L. The numerical software package MATLAB tackles
efficiently the two-dimensional (matrix) interpolation problem,
as well as the eigenvalue problem. It is apparent that there is
some variability in the calculated upcrossing rates depending
on the number of eigenvalues j (19) included in the analysis.
In the present example 150 eigenvalues j , j = 1, . . . , 150
proved to be amply sufficient.
The random stationary sea state is specified by a JONSWAP
spectrum, which is given as follows
S () =
5 p
4
g 2
exp
5
exp
1
2 2
+ ln
1
p
(75)
where g = 9.81 m s2 , p denotes the peak frequency in rad/s

and where and are parameters affecting the spectral shape.
= 0.07 when p , and = 0.09 when > p . The
parameter is chosen to be equal to 3.3. The parameter is
determined from the following empirical relationship
= 5.06
Hs
T p2
(1 0.287 ln ) ,
(76)
where Hs denotes the significant wave height and T p = 2/ p

is the spectral peak wave period. For the subsequent test
cases, two different JONSWAP sea states are chosen with the
following parameters: Hs = 7 m and T p = 10 s (moderate
sea), and Hs = 10 m and T p = 16 s (severe sea).
For the asymptotic analysis we suggest to neglect all
correlation coefficients rk1 except for r21 used in the calculation
of s1 , cf. (42). Since |r21 | 1 and rk1 0, k = 2, the latter
replacement is equivalent to a minor tuning of w (k ) in Eq.
(20), which should have only minor influence on the statistics,
but it increases the numerical stability of the results.
Fig. 1 presents dimensional, spline interpolated one-sided
power spectral density (PSD) of the linear response part Z 1
(12) expressed in m2 s/rad. Fig. 2 presents dimensional, spline
interpolated difference frequency surge response function
Q(, ) (14) expressed in m s/rad. The severe sea state was
chosen. Since the QTF is complex valued, only its absolute
value is plotted.
Table 2 presents a comparison between the exact and
asymptotically approximated upcrossing rates. Since for
engineering applications it is not the upcrossing rates
themselves, but rather the response levels not being exceeded
with a given probability P that are important, the comparison
(A)
is given in terms of the ratio P / P . The level P is derived
from (66), where T = 3 h
P = Prob(M Z (T ) P ),
(77)
351
in focus, which is beneficial for the asymptotic approximation.

Increasing the storm duration T from 3 h to e.g. 6 or 18 h will
further increase accuracy of the asymptotic estimation.
Fig. 1. PSD of the linear response part Z 1 (m2 s/rad).
The asymptotic behaviour of the mean upcrossing rate at

high response levels of the combined first-order and secondorder surge motions of moored offshore structures in random
short-crested seas has been investigated. The main motivation
behind this study was a desire to correct and amend previous
asymptotic results [4,12], which were numerically unstable
and tended to give very unreliable results. The present study
has succeeded in deriving an asymptotic result for the mean
upcrossing rate that does not suffer from the weaknesses of the
previous results cited above, and which appears to give fairly
accurate estimates of extreme responses.
Acknowledgements
The authors are grateful for the financial support of this work
by the Research Council of Norway (NFR) through the Centre
for Ships and Ocean Structures (CeSOS) at the Norwegian
University of Science and Technology. The authors also want
to express their gratitude to Dr. P.S. Teigen for the numerical
calculations of the transfer functions for the example structure
used in this paper.
References
Fig. 2. Absolute value of the dimensional (ms/rad) surge response QTF.

Table 2
(A)
Ratio P / P of levels not being exceeded with a given probability P

P
Z 2 , mod.
sea
Z 1 + Z 2 , mod.
sea
Z 2 , sev.
sea
Z 1 + Z 2 , sev. sea
0.70
0.75
0.80
0.85
0.90
0.95
0.98
0.98
0.98
0.98
0.98
0.99
0.89
0.89
0.90
0.90
0.90
0.91
0.99
0.99
0.99
0.99
0.99
0.99
0.97
0.97
0.98
0.98
0.99
0.99
(A)
while P denotes the level obtained by the asymptotic

approximation (64), or (65). For comparison, both the
combined response Z = Z 1 + Z 2 and the second-order response
Z 2 are considered for two different sea statesmoderate and
severe. It is seen that the asymptotic approximation is very
accurate for pure second-order response. For the combined
response, the asymptotic approximation is very accurate for
the severe sea state, while the agreement for moderate sea is
less pronounced, but still acceptable for practical purposes,
giving 10% relative error in the response level estimation. It is
important to note that for extreme response estimation, usually
it is not the moderate but rather the severe sea states that are
[1] Kac M, Siegert AJF. On the theory of noise in radio receivers with square
law detectors. Journal of Applied Physics 1947;18:38397.
[2] Neal E. Second order hydrodynamic forces due to stochastic excitation.
in: Proc. 10th ONR symposium. 1974.
[3] Vinje T. On the statistical distribution of second-order forces and motions.
International Shipbuilding Progress 1983;30:5868.
[4] Naess A. Statistical analysis of second-order response of marine
structures. Journal of Ship Research 1985;29(4):27084.
[5] Langley RS. A statistical analysis of low frequency second-order forces
and motions. Applied Ocean Research 1987;9(3).
[6] Kato S, Ando S, Kinoshita T. On the statistical theory of total second order
response of moored floating structures. in: Proc. 19th annual offshore
technology conference. Paper no. 5579. 1987.
[7] Donley MG, Spanos PD. Dynamic analysis of non-linear structures by the
method of statistical quadratization. Lecture notes in engineering, vol. 57.
Berlin: Springer-Verlag; 1990.
[8] Naess A. Statistical analysis of nonlinear, second-order forces and
motions of offshore structures in short-crested random seas. Probabilistic
Engineering Mechanics 1990;5(4):192203.
[9] Langley RS, McWilliam S. A statistical analysis of first and second
order vessel motions induced by waves and wind gusts. Applied Ocean
Research 1993;15(1):1323.
[10] McWilliam S, Langley RS. Extreme values of first and second order wave
induced vessel motions. Applied Ocean Research 1993;15(3):16981.
[11] Kareem A, Zhao J, Tognarelli MA, Gurley KR. Dynamics of nonlinear
stochastic systems: A frequency domain approach. In: Haldar A,
Ayyub BM, Guran A, editors. Uncertainty modelling in vibration, control
and fuzzy analysis of structural systems. Singapore: World Scientific
Publishing Co.; 1997. p. 10146.
[12] Naess A. The response statistics of non-linear second-order transformations to Gaussian loads. Journal of Sound and Vibration 1987;115(1):
10329.
352
[13] Naess A, Karlsen HC. Numerical calculation of the level crossing rate
of second order stochastic Volterra systems. Probabilistic Engineering
Mechanics 2004;19(2):15560.
[14] Naess A, Karlsen HC, Teigen PS. Numerical methods for calculating the
crossing rate of high and extreme response levels of compliant offshore
structures subjected to random waves. Applied Ocean Research 2006;
28(1):18.
[15] Benfratello S, Di Paola M, Spanos PD. Stochastic response of MDOF
wind excited structures by means of Volterra series approach. Journal of
Wind Engineering and Industrial Aerodynamics 1998;7476:113545.

[16] Naess A. The statistical distribution of second-order slowly-varying forces
[17] Naess A, Johnsen JM. An efficient numerical method for calculating the
statistical distribution of combined first-order and wave-drift response.
Journal of Offshore Mechanics and Arctic Engineering 1992;114(3):
195204.
[18] Bender CM, Orszag SA. Advanced mathematical methods for scientists
and engineers. Singapore: McGraw-Hill, Inc.; 1978.

Papers About Extreme Value and Crossing Levels by Arvid Naess

Transféré par

Informations du document

Copyright

Formats disponibles

Partager ce document

Partager ou intégrer le document

Options de partage

Avez-vous trouvé ce document utile ?

Ce contenu est-il inapproprié ?

Droits d'auteur :

Formats disponibles

Papers About Extreme Value and Crossing Levels by Arvid Naess

Transféré par

Droits d'auteur :

Formats disponibles

Hindawi Publishing Corporation

Journal of Probability and Statistics

Correspondence should be addressed to A. Naess; arvidn@math.ntnu.no

useful information about extreme events that are truly

Journal of Probability and Statistics

it represents a very powerful addition to the toolbox of

2. Cascade of Conditioning Approximations

for 2 . With this approximation, it is obtained that

In this section, a sequence of nonparametric distribution

where 3 , which leads to the approximation

Journal of Probability and Statistics

where () = Prob{ > | 1 , . . . , +1 }, for

The concept of average conditional exceedance rate (ACER)

Alternatively, (7) can be expressed as,

3. Empirical Estimation of the Average

Going back to the definition of 1 (), it follows that

=1 1 () is equal to the expected number of exceedances

In general, this ACER function also depends on the number

where E[] denotes the expectation operator. Assuming an

where () and () are the realized values of () and

Journal of Probability and Statistics

The advantage of using the modified ACER function () for

() exp ( ()) = exp (

If the time series can be segmented into blocks, such that

It is of interest to note what events are actually counted

subseries, then the sample standard deviation () can be

4. Estimation of Extremes for the Asymptotic

Journal of Probability and Statistics

appropriately chosen tail marker. Hence, it will be assumed

where the function () is slowly varying, compared with

where , , , and are appropriately chosen constants. In

puts more emphasis on the more reliably estimated ( ).

Here, the first nonlinear inequality constraint is evident, since

Journal of Probability and Statistics

It is obtained that the optimal values of and are given

5. Estimation of Extremes for the General Case

of approximation, the behaviour of the mean exceedance rate

where > 0, , are constants. This follows from the

where the function () is weakly varying, compared with

where the function () is weakly varying compared with

where is a weight factor as previously defined.

Journal of Probability and Statistics

An option for estimating the five parameters , , ,

6. The Gumbel Method

it is known that the parameters and are related to the mean

7. The Peaks-over-Threshold Method

Journal of Probability and Statistics

Hence, it follows that

Invoking (1) for = 0 leads to the result

Similarly, for = 0, it is found that,

where is the threshold used in the estimation of and .

8. Extreme Value Prediction for Synthetic Data

where = 1 year and + () is the mean up-crossing rate per

where = + (0) = 10, which corresponds to = 3.65 days,

Since the data points (i.e., = 3.65 days maxima) are

Journal of Probability and Statistics

Hence, in this case the ACER method performs consistently

9. Measured Wind Speed Data

Figure 1: Synthetic data ACER 1 , Monte Carlo simulation ();

Hourly maximum gust wind was recorded during the

Journal of Probability and Statistics

Hekkingen Fyr 88690

Nordyan Fyr 75410

Figure 2: The point estimate 100 yr of the 100-year return period