Modeling Extreme Events from Computer Simulations

Part II

Emmanuel Vazquez

SUPELEC, Gif-sur-Yvette, France

Summer School CEA-EDF-INRIA, 2011

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 1 / 43

Outline of Part II

1 Structural reliability
Limit-state functions
Reliability indexes
Limit-state function defined on the factor space
Affine limit-state functions
First-order reliability method
Reliability of systems
References

2 Elicitation of subjective probability distributions

Definition
Copulas
Linear correlation
Dependence measures
References

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 2 / 43

 Structural reliability Limit-state functions

Reliability approach in mechanics

◮ In mechanics, the term reliability describes the ability of a system to accomplish a

required function
◮ Mechanical materials or structures are considered as systems comprising an input, a
state and an output
◮ The classical point of view

Design Resistance Model R = r (x)

Parameters

x gR,S (R, S)

Load Stress Model S = s(x)

◮ The success of a design (or serviceability) is seen in the verification of the inequality

gR,S (R, S) = R − S ≥ 0
◮ The quantity gR,S (R, S) = R − S is called the (safety) margin
◮ gR,S is called the limit-state function.

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 3 / 43

Structural reliability Limit-state functions

Example: deformation of a rod

11111111
00000000 ◮ The yield point of a material is defined as the stress
00000000
11111111
00000000
11111111
00000000
11111111
at which a material begins to deform plastically
(prior to the yield point the material will deform
elastically and will return to its original shape when
the applied stress is removed)
◮ Stress: σ = F /A (Pa)
2
cross section A (m ) ◮ Yield limit: σy (Pa)
◮ Design rule: F /A ≤ σy
◮ Set
0
1
0
1
x = (F , A, σy ) ∈ R3
11
00 0
1
1111
0000
00
11 0
1 F ·L (m)
00
11 0
1 ∆L = A·E r (x) = Aσy
0
1
00 0000
11 1111
0
1
0
1
0
1 s(x) = F
◮ Define gR,S : (R, S) 7→ R − S. The design must
load F (N) verify
gR,S (r (x), s(x)) = Aσy − F ≥ 0

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 4 / 43

 Structural reliability Limit-state functions

Failure region, Limit-state function

◮ The operating scenario is the availability of a resistance R greater than the stress S,
such that
gR,S (R, S) = R − S ≥ 0
◮ The failure scenario is
gR,S (R, S) = R − S < 0
and the failure region is defined as the region of R2+

Γ = {(r , s) ∈ R2+ ; gR,S (r , s) < 0}

◮ The limit state is defined as the region of R2+

∂Γ = {(r , s) ∈ R2+ ; gR,S (r , s) = 0}

(hence, the name limit-state function for gR,S )

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 5 / 43

Structural reliability Limit-state functions

Probability of failure

◮ In practice, the load applied on a system is unknown, and the design parameters are
subjected to dispersions
→ the parameter vector x is uncertain, and can be modeled by a random vector

X ∼ PX
◮ Then, R = r (X ) and S = s(X ) are also random variables
◮ Let fR,S be the joint pdf of (R, S) (wrt to the Lebesgue measure)
◮ The probability of failure of the system corresponds to
Z
α = P{R − S < 0} = fR,S (r , s)dr ds
r −s<0

◮ Reliability is defined as 1 − α

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 6 / 43

 Structural reliability Limit-state functions

Probability of failure

◮ Usually, R and S assumed to be independent

◮ Then,
Z
α = fS (s)fR (r )dr ds
r −s<0
Z Z !
= fS (s) fR (r )dr ds
s∈[0,∞[ r ∈[0,s]
Z
= fS (s)FR (s)ds
s∈[0,∞[

where fS and fR are the pdf of S and R, and FR is the cdf of R

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 7 / 43

Structural reliability Limit-state functions

Probability of failure: case of a known resistance

◮ When the resistance is deterministic: R = r0 =⇒ FR (r ) = 1r0 ,∞[ (r )

◮ Then α = P{S > r0 } = P{s(X ) > r0 }

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 8 / 43

 Structural reliability Reliability indexes

Reliability indexes in structural reliability

◮ In the domain of structural reliability, the serviceability of a design is often quantified

using the notion of a reliability index
◮ When the resistance and the stress are deterministic, the notion of reliability index
may be defined arbitrarily as the numbers
R
or R − S
S
◮ When R and S are viewed as independent random variables, a notion reliability index
may be defined as the number
E(gR,S (R, S)) mR − mS
βC = = p
var(gR,S (R, S))1/2 σR2 + σS2
◮ However, the interpretation of βC is not simple and, above all, βC is generally not
related to the probability of failure of the system

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 9 / 43

Structural reliability Limit-state function defined on the factor space

Probability of failure expressed in terms of the uncertain factors

◮ Note that the point of view of structural reliability is not different from that
presented in Part I
◮ In practice, fR and fS cannot be determined directly by the user → only the
distribution PX of the vector of uncertain factors can
◮ Thus, it is generally easier to express a limit-state function gX in the factor space:

gX : x 7→ gR,S (r (x), s(x))

so that the probability of failure can be expressed as

α = P{gX (X ) < 0} , X ∼ PX

or
α = PX {x ∈ X; gX (x) < 0} = PX {gX < 0}
(The probability of failure is the volume of excursion of g above zero.)

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 10 / 43

 Structural reliability Affine limit-state functions

Probability of failure for Gaussian random factors and affine limit-state

function

◮ Assume X ∼ N(0, Id ) ∈ Rd
◮ Assume moreover that gX is affine: ∀x ∈ Rd

gX (x) = a0 + a1 x[1] + · · · + ad x[d] = a0 + (a, x)

◮ Then, the limit-state ∂Γ is the hyperplane defined by a0 + (a, x) = 0

fX = cst

x⋆

gX = 0

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 11 / 43

Structural reliability Affine limit-state functions

Probability of failure for Gaussian random factors and affine limit-state

function

fX = cst

x⋆

gX = 0

◮ Let x ⋆ be the nearest point of the hyperplane ∂Γ = {x ∈ Rd ; a0 + (a, x) = 0} to

the origin
◮ x ∈ ∂Γ ⇐⇒ (x − x ⋆ , x ⋆ ) = (x ⋆ , x) − kx ⋆ k2 = 0
◮ Thus, for a0 6= 0,
a x⋆ |a0 |
= − ⋆ 2 =⇒ β := kx ⋆ k =
a0 kx k kak
◮ Note that U = (X , η), with X ∼ N(0, Id ) and η = x ⋆ /β, is a random variable with
distribution N(0, 1)
E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 12 / 43
 Structural reliability Affine limit-state functions

Probability of failure for Gaussian random factors and affine limit-state

function

◮ A failure corresponds to the event {U > β} = {(X , x ⋆ ) > β 2 }, which has probability

α = 1 − P{U ≤ β} = 1 − Φ(β) = Φ(−β)

◮ Therefore, to compute a probability of failure in the case of standard normal factors

and affine limit-state function gX :
1. Solve
x ⋆ = arg min kxk
x∈Rd
subject to gX (x) = 0
2. The probability of failure is α = Φ(−β), with β = kx ⋆ k

◮ In the literature of structural reliability, x ⋆ is called a design point, or most central

failure point

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 13 / 43

Structural reliability Affine limit-state functions

Probability of failure for Gaussian random factors and affine limit-state

function

◮ Consider again an affine limit-state function gX : x 7→ a0 + (a, x)

◮ Define the margin as the random variable Z = gX (X )
◮ Note that
E (Z ) = a0
and
var(Z ) = a12 + · · · + ad2 = kak2
◮ Thus,
|a0 | |E (Z )|
β= =
kak var(Z )1/2
E (Z )
◮ In the literature of structural reliability, the ratio is interpreted as a reliability
var(Z )1/2
index
◮ β is called the Hasofer-Lind reliability index

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 14 / 43

 Structural reliability Affine limit-state functions

Measure of the importance of the factors with respect to a failure

◮ The sensitivity of the probability of failure with respect to changes in the factors is
an important information for the design of a system (makes it possible to understand
which factor are most important to control)

◮ One possibility to define a notion of sensitivity is to look at the variation of the

margin z = gX (x) as a function of x;

◮ We have
a0
∇gX (x) = a = − η
β
1 ⋆
with η = β
x

◮ Thus, if the i th component η[i] of the unit vector η is large, the margin will vary
rapidly as we move along the i th direction
➟ η[i] accounts for the importance of the ith factor

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 15 / 43

Structural reliability First-order reliability method

Extensions

At this point, several extensions can be considered:

◮ Non-affine limit-state functions

◮ Gaussian non-standard random factors
◮ Non-Gaussian independent random factors
◮ Non-Gaussian non-independent factors

→ first-order reliability method (and related methods)

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 16 / 43

 Structural reliability First-order reliability method

Non-affine limit-state functions

◮ In some applications, it may be reasonable to approximate the limit state ∂Γ by a
geometric shape such as a hyperplane

fX = cst

x⋆

gX = 0

◮ Therefore, to approximate a probability of failure in the case of standard normal

factors and a non-affine limit-state function gX :
1. Solve
x ⋆ = arg min kxk
x∈Rd
subject to gX (x) = 0
b = Φ(−β), with β = kx ⋆ k
2. An approximation of the probability of failure is α

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 17 / 43

Structural reliability First-order reliability method

Non-affine limit-state functions: FORM

fX = cst

x⋆

gX = 0

◮ If gX is smooth, note that ∇gX |x=x ⋆ ∝ η, with η = x ⋆ /β

◮ Thus the equation of the approximating hyperplane is
(∇gX |x=x ⋆ , x − x ⋆ ) = 0
◮ Note that x 7→ (∇gX |x=x ⋆ , x − x ⋆ ) is the first-order approximation of gX in the
neighborhood of x ⋆
◮ In structural reliability, using a first-order approximation of gX at x ⋆ in a standard Gaussian
factor space is called First-Order Reliability Method (FORM)

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 18 / 43

 Structural reliability First-order reliability method

Finding the most central failure point

◮ Solving

x ⋆ = arg min kxk

x∈Rd

subject to gX (x) = 0

is a constrained optimization problem, for which many algorithms has been proposed
in the literature:
◮ Penalty methods
◮ Augmented Lagrangian
◮ Projected gradient
◮ BFGS
◮ SQP
◮ ...
(they all are local optimization algorithms)
◮ In general, finding a good approximation of x ⋆ can be done with a moderate number
of evaluations of gX
◮ However: gX generally need to be convex and differentiable
➟ If there exist (even small) numerical instabilities when computing gX , as may be
the case in models based on the numerical solution of some partial differential
equations, then the optimization of gX using standard techniques can fail direly

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 19 / 43

Structural reliability First-order reliability method

How much can go wrong with FORM?

◮ For some applications, finding x ⋆ can be done with only a few evaluations of gX →
interesting when gX is expensive to evaluate
◮ However, having found x ⋆ does not tell if the approximation α ≈ Φ(−β) is good or
not
◮ The probability of failure can be overestimated or underestimated
◮ Consider, for instance the following domain of failure:

Γ = {x ∈ Rd ; kxk > β0 }
2 2
◮ Let V = X[1] + · · · X[d] , so that V ∼ χ2 (d). The failure event is {V > β02 }.
◮ Hence, we have α = 1 − FV (β02 ) and α
b = 1 − Φ(β0 )
◮ b ≈ 2.9 · 10−7 but α ≈ 0.2!
Example: suppose d = 20 and β0 = 5, we obtain α
◮ The FORM approximation should be used only when prior knowledge about the
shape of ∂Γ is available

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 20 / 43

 Structural reliability First-order reliability method

Quadratic approximation: SORM

◮ Instead of using a first-order approximation, one could think of using a second-order
approximation

fX = cst

x⋆
β → Second-Order Reliability Method (SORM)

gX = 0

◮ Such an approximation requires the approximation of x ⋆ and the estimation of the curvature
of the limit-state at x ⋆
◮ A quadratic approximation is more expensive to obtain, but there is, in general, no
guarantee that the approximation of the probability of failure will be better than the
first-order approximation
◮ In fact, we can have
bFORM > α
α>α bSORM

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 21 / 43

Structural reliability First-order reliability method

Multiple local first-order approximations: multi-FORM

◮ Another possibility proposed in the literature: compute several design points, that is,
several local minimizers of kxk subject to the constraint gX (x) = 0

gX = 0 ◮ Finding several local minimizers can be

fX = cst a difficult task
◮ Consider again the domain of failure:
x1⋆

β1
Γ = {x ∈ Rd ; kxk > β0 }

and the approximation

β2
b
Γ = {x ∈ Rd ; max|x[i] | > β0 }
i
x2⋆
(obtained with 2d design points)

◮ Then, Q
b = P{X ∈ b
α Γ} = 1 − i (Φ(β0 ) − Φ(−β0 )) = 1 − (1 − 2Φ(−β0 ))d ≈ 2dΦ(−β0 )
b
=⇒ for d >> 1, α >> α

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 22 / 43

 Structural reliability First-order reliability method

Non-standard Gaussian random factors

◮ Assume X ∼ N(m, K ) ∈ Rd , with m ∈ Rd and K a d × d symmetric definite positive (SDP)
matrix
◮ To apply the framework above, the idea is to search for a one-to-one whitening
transformation T : Rd → Rd such that U = T (X ) is a standard Gaussian random vector
◮ Consider the eigendecomposition of K such that K = QΛQ T , where Λ is the diagonal
matrix of the eigenvalues of K , and Q is orthogonal
◮ Then, T : X 7→ Λ−1/2 Q T (X − m) is such that U = T (X ) ∼ N(0, Id )
◮ Note that if gX is affine, then gU = gX ◦ T −1 is also affine in the standardized space

4 4

3 3

2 2

1 1

0 0
T
−1 −1

−2 −2

−3 −3

−4 −4
−4 −2 0 2 4 −4 −2 0 2 4

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 23 / 43

Structural reliability First-order reliability method

Non-Gaussian independent random factors

◮ When the components of X are independent but non Gaussian, the idea is to search
for one-to-one transformations Ti : Rd → Rd such that for each i = 1, . . . , d,
U[i] = Ti (X[i] ) is a standard Gaussian random vector
◮ Assume that the cdf Fi of X[i] is continuous and strictly increasing, then

U[i] = Φ−1 (Fi (X[i] )) ∼ N(0, 1),

where Φ−1 stands for the inverse (reciprocal) function of Φ

Indeed,

P{U[i] ≤ u} = P{Φ−1 (Fi (X[i] )) ≤ u}

= P{X[i] ≤ Fi−1 (Φ(u))}
= Φ(u)
◮ Thus, if the components of X are independent but non Gaussian, the random vector

U = T (X ) = (Φ−1 ◦ Fi (X[1] ), . . . , Φ−1 ◦ Fd (X[d] )) ∼ N(0, Id )

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 24 / 43

 Structural reliability First-order reliability method

Non-Gaussian non-independent random factors

◮ As above, the idea is to transform the random vector of factors

◮ This case will be examined separately: in fact it raises two difficult issues:
◮ how to specify the distribution of a random vector in the case of non-independent
components?
◮ what is the influence of the input distribution on the probability of failure?

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 25 / 43

Structural reliability First-order reliability method

FORM and related methods: summing up

◮ Very popular methods in the domain of structural reliability

◮ FORM is easy to understand and to implement

◮ When a first-order approximation is relevant, a good approximation of the

probability of failure can be found with a small number of function evaluations

◮ A major pitfall: the approximation of the probability of failure can be significantly

wrong

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 26 / 43

 Structural reliability First-order reliability method

A personal perspective on the geometrical approximation approaches

◮ If gX is not expensive to evaluate:

◮ Monte Carlo should be preferred over any geometrical approximation
◮ If gX is expensive:
◮ a simple MC approach cannot be used
◮ what can be the use of a geometrical approximation?
◮ It can be very wrong to approximate the limit state with a geometrical shape
◮ can only be justified when it is known in advance that a given geometrical
approximation is correct
◮ SORM can be though as a correction of FORM, but from a mathematical
perspective, it is not → using SORM over FORM can only be justified when it is
known in advance that gX is almost quadratic
◮ A multi-FORM approach seems preferable, but using it is to admit that the shape of
the limit state is unknown, which is dangerous for a geometrical approximation
approach
◮ Fortunately, in a large number of applications, the limit-state function is almost
affine, which explains why FORM remains a very popular method

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 27 / 43

Structural reliability Reliability of systems

Reliability of systems

◮ Until now, we have implicitly considered the case of the failure of a unique
component
◮ In a real system, a failure can happen due to the failure of just one of its (possibly
many) components
◮ The designer can also choose to have redundancy on critical components; in this
case the failure of the system happens when all redundant components fail
◮ To deal with these issues, the domain of structural reliability generally introduces the
notions of parallel and series systems

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 28 / 43

 Structural reliability Reliability of systems

Reliability of systems: series systems

◮ Let x denotes the state of a system, and let Γ be a domain of failure

◮ In structural reliability, a system is called a series system if the occurrence of one
single failure event brings a failure on the whole system
◮ The classical example is that of a chain whose failure is related to any of its links
◮ In other words, it means that we can write
I
[
Γ= Γ(i)
i=1

◮ Assume that each domain of failure Γ(i) is characterized by a limit-state function

(i) (i)
gX , so that the failure of the i th component corresponds to the event {gX (X ) < 0}
◮ Then the failure event for the whole system can be characterized by the limit-state
function
(1) (I )
gX : x 7→ gX (x) ∧ · · · ∧ gX (x)
◮ Conclusion: the case of series systems can be dealt with using the framework we
have exposed previously

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 29 / 43

Structural reliability Reliability of systems

Reliability of systems: parallel systems

◮ In structural reliability, a system is called a parallel system if the failure of all events
is necessary for the failure of the whole system
◮ A parallel system is a principle a redundancy
◮ Using the notations above, the domain of failure for a parallel system corresponds to
\
Γ= Γi
i

◮ Again, the failure event for the whole system can be characterized by a single
limit-state function
(1) (I )
gX : x 7→ gX (x) ∨ · · · ∨ gX (x)
◮ Conclusion: parallel systems can also be dealt with using the framework we have
exposed previously

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 30 / 43

 Structural reliability References

Some references

◮ Ditlevsen O. and Madsen H.O. (1996), Structural reliability methods, Wiley

◮ Lemaire M., Chateauneuf A., and Mitteau J.C., (2009) Structural Reliability, Wiley

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 31 / 43

Elicitation of subjective probability distributions Definition

Elicitation of subjective probability distributions

◮ A very difficult issue: to choose a probability distribution for modeling the uncertain
factors of a system
◮ The process of choosing a distribution for the factors is called elicitation1 in the
literature of decision analysis
◮ Elicitation is particularly difficult when doing risk analysis about new and untried
technologies, for which little data are available
◮ Very often, risk analysis relies on expert judgment
◮ Elicitation of subjective probability distributions is often subject to a number of
serious biases, such as overconfidence in the ability to quantify uncertainty

1
from Latin elicere: draw forth, bring out
E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 32 / 43
 Elicitation of subjective probability distributions Definition

Elicitation of subjective probability distributions

◮ Assume that we are given an approximation of the mean and the standard deviation of a
random variable
◮ An experiment:

◮ X ∼ N(µ, σ 2 )
0.5
2
0.45 ◮ Y ∼ logN(m, s 2 ), with s 2 = log( σ + 1)
µ2
and m = log µ − s 2 /2.
0.4

0.35

0.3 ◮ Then, E[X ] = E[Y ] = µ and

0.25
var[X ] = var[Y ] = σ 2 .
0.2

0.15

◮
0.1

0.05
Assume µ = 3 and σ 2 = 1. Then,
0 P(X > 7) = 3.1 · 10−5 and
P(Y > 7) = 2.8 · 10−3
0 1 2 3 4 5 6 7 8

◮ For risk analysis, “knowing” only the mean and the standard deviation of a random variable
is a very poor information
◮ In fact, in many applications, the occurrence of a failure is likely to be related to an extreme
event in the factor space
→ it is probably more important to characterize the tail behavior of the factors than the
central behavior (EVT can help)

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 33 / 43

Elicitation of subjective probability distributions Definition

Elicitation of subjective probability distributions

◮ Elicitation gets even more complicated if we want to introduce dependence

information between random variables. In fact, the question is: “how to measure
dependence?”
◮ In probability theory and statistics, the dependence between random variables is
described by the concept of copula (from Latin co-apere “join together”)

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 34 / 43

 Elicitation of subjective probability distributions Copulas

Copulas
◮ Consider a random vector X = (X1 , . . . , Xd ) ∈ Rd . The dependence between the
component random variables X1 , . . . , Xd is completely described by the joint cdf
F (x1 , . . . , xd ) = P{X1 ≤ x1 , . . . , Xd ≤ xd }
◮ For simplicity, assume that the components Xi , i = 1, . . . , d, have continuous,
strictly increasing, marginal cdfs Fi
◮ The concept of copula: separate F into a part that describes the dependence
structure and parts which describe the marginal behavior only
◮ Transform X component-wise to obtain standard-uniform marginal
distributions U([0, 1])
T: dom F → [0, 1]d
(x1 , . . . , xd ) 7 → (F1 (x1 ), . . . , Fd (xd ))
◮ The joint cdf of U = T (X ) is called the copula of the random vector X
◮ It follows that for x ∈ dom F
F (x1 , . . . , xd ) = C (F1 (x1 ), . . . , Fd (xd ))
and for u ∈ [0, 1]d
C (u1 , . . . , ud ) = F (F1−1 (u1 ), . . . , Fd−1 (ud ))
E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 35 / 43

Elicitation of subjective probability distributions Copulas

Copulas
◮ Some copulas/dependence structures:
◮ The independent copula: Cind (u) = u1 u2 · · · ud
◮ The Gaussian copula
Cg,R (u1 , . . . , ud ) =
Z Φ−1 (u1 ) Z Φ−1 (ud )  1 
···(2π)−d/2 (det R)−1/2 exp − u T R −1 u du1 · · · dud
−∞ −∞ 2
h  β i
◮ The bivariate Gumbel copula CGu,β (u, v ) = exp − (− log u)1/β + (− log v )1/β

Realizations from two

distributions with identical
Gamma(3, 1) marginal
distributions and identical
correlation ρ = 0.7, but
different dependence
structures.

from Embrechts et al. (1999)

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 36 / 43

 Elicitation of subjective probability distributions Linear correlation

Linear correlation (or Pearson’s correlation)

◮ In some applications, choosing independent random factors is irrelevant

◮ However, choosing a dependence structure can be a difficult task because
information may be scarce (e.g., no data available)
◮ Very often, the dependence structure of a random vector is summarized through
(linear) correlations between components
Recall that for (X , Y ) ∈ R2 , a second-order random vector, the linear correlation
coefficient is
cov(X , Y )
ρ(X , Y ) = p
var(X ) var(Y )
If |ρ(X , Y )| = 1, then there is a perfect linear dependence between X and Y , i.e.,
Y = aX + b, for some a, b ∈ R

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 37 / 43

Elicitation of subjective probability distributions Linear correlation

Nataf transformation

◮ The Nataf transformation TNataf is a one-to-one function which maps a random

vector X with a Gaussian copula to a random vector U with standard Gaussian
distribution
◮ Makes it possible to apply FORM for non-Gaussian non-independent random vectors
◮ Conversely, the Nataf transformation makes it possible to define implicitly the
distribution F of a random vector X = (X1 , . . . , Xd ) with
1. Gaussian copula
2. prescribed continuous, strictly increasing, marginals F1 , . . . , Fd
3. prescribed correlation matrix RX = (ρ(Xi , Xj ))i,j

◮ Considering a Gaussian copula for the dependence structure of X should not be

considered as a canonical choice → a comprehensive risk analysis procedure should
assess the consequences of this particular choice

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 38 / 43

 Elicitation of subjective probability distributions Linear correlation

Inverse Nataf transformation

◮ Let U ∼ N(0, Id )
◮ Given continuous strictly increasing marginals cdfs F1 , . . . , Fd , and a d × d correlation
matrix RX , the inverse Nataf transformation is defined as
−1
X = TNataf (U) = T3 ◦ T2 ◦ T1 (U), with
T1 : Rd → Rd
u = (u1 , . . . , ud ) 7 → Cu

T2 : Rd → [0, 1]d
v = (v1 , . . . , vd ) 7 → (Φ(v1 ), . . . , Φ(vd ))

T3 : [0, 1]d → Rd
w = (w1 , . . . , wd ) 7 → (F1−1 (w1 ), . . . , Fd−1 (wd ))

where C is a d × d matrix which is computed in such a way that X has correlation

matrix RX
◮ Note that V = T1 (U) is a Gaussian vector. Thus, it has a Gaussian copula. Since T2 and
T3 are component-wise monotonic transformations, T2 and T3 are copula-invariant
◮ Note that the correlation matrix of V = T1 (U) is RV = CC T . In general, RV 6= RX .
◮ Note also that it is not always possible to prescribe any correlation coefficient (depending on
the choice of the marginals)

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 39 / 43

Elicitation of subjective probability distributions Linear correlation

Dependence measures

◮ As mentioned above, considering a Gaussian copula for the dependence structure of

X should not be considered as a canonical choice
◮ Moreover, measuring dependence based on correlation coefficients can be misleading
→ linear dependence should not be taken as a canonical dependence measure

Realizations from seven bi-variate distributions with zero correlation

◮ Other measures of the dependence?

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 40 / 43

 Elicitation of subjective probability distributions Dependence measures

Dependence measures

◮ Non-linear dependence measures have been proposed in the literature

◮ In particular, rank correlation coefficients, such as Spearman’s rank correlation
coefficient ρ, and Kendall’s rank correlation coefficient τ measure the extent to
which two variables increase or decrease simultaneously
◮ For instance, Spearman’s ρ is defined as the Pearson’s correlation coefficient
between the ranked variables

from Wikipedia: article on Spearman’s rank correlation coefficient

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 41 / 43

Elicitation of subjective probability distributions Dependence measures

Tail dependence

◮ In the context of risk analysis, it might be relevant to study the dependence between
extreme values
◮ Let X = (X1 , X2 ) be a vector of continuous random variables with marginals F1
and F2 . The coefficient of upper tail dependence of (X1 , X2 ) is

λU = lim P{X2 > F2−1 (u) | X1 > F1−1 (u)}

u→1

provided that the limit exists. If λU is well-defined, λU ∈ [0, 1]

◮ The value of λU is a property of the copula of X only
◮ Examples:
◮ Consider the Gaussian bi-variate copula, with correlation ρ < 1. Then, λU = 0; that is,
a Gaussian copula with ρ < 1 does not have tail dependence
◮ The Gumbel bi-variate copula, with parameter β > 1, has tail dependence
λU = 2 − 21/β
◮ Choosing a copula with tail-dependence over the Gaussian copula can modify the
probability of failure by several orders of magnitude—see, for instance, Dutfoy and
Lebrun, Congrès Français de Mécanique (2007)

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 42 / 43

 Elicitation of subjective probability distributions References

Some references

◮ Embrechts P., McNeil A., Strautmann D. (1999), Correlation and dependence in risk
management: properties and pitfalls
◮ Embrechts P., Lindskog F., McNeil A. (2001), Modelling Dependence with Copulas
and Applications to Risk Management
◮ Lebrun R. and Dutfoy A. (2009), A generalization of the Nataf transformation to
distributions with elliptical copula

E. Vazquez Extreme events modeling II Summer School CEA-EDF-INRIA, 2011 43 / 43