Seoul National University Instructor: Junho Song

Dept. of Civil and Environmental Engineering [email protected]

457.212 Statistics for Civil & Environmental Engineers

In-Class Material: Class 15
Extreme Value Distributions (A&T 4.2.3)

Given: The distribution model of a random quantity 𝑋, i.e., PDF or CDF

Question: From a sample of size n , the distribution of the minimum or maximum?
→ Deriving an ( ) distribution
e.g., maximum flood (or drought) in the next 100 years, maximum traffic load on bridge in
the next 50 years
(More generally, the distribution of the k-th largest or smallest from a sample → “_______
Statistics”)

1. Deriving “Exact Distributions”

 Maximum: Yn = max ( X1 , X 2 ,..., X n )

𝐹𝑌𝑛 (𝑦) = 𝑃(𝑋1 ≤ 𝑦 𝑋2 ≤ 𝑦 ⋯ 𝑋𝑛 ≤ 𝑦)

Under the assumption that X 1 , X 2 ,..., X n are statistically ( ) and ( )

distributed,

𝐹𝑌 𝑛 (𝑦) = [ ]𝑛

Its corresponding PDF is,

dFYn ( y)
= n 
n −1
fYn ( y) = f X ( y)
dy

 Minimum: Y1 = min ( X1 , X 2 ,..., X n )

1 − 𝐹𝑌 1 (𝑦) = 𝑃(𝑋1 > 𝑦 𝑋2 > 𝑦 ⋯ 𝑋𝑛 > 𝑦) = [ ]𝑛

Therefore, the CDF of Y1 is

FY1 ( y) = 1 −  
n

The corresponding pdf is

dFY1 ( y)
= n 
n −1
fY1 ( y) = f X ( y)
dy

1
Seoul National University Instructor: Junho Song
Dept. of Civil and Environmental Engineering [email protected]

Example 1: Suppose the PDF of a random variable X is given as below.

1
f X ( x) = , x 1
x2

When someone constructs a sample of size n, derive the CDF and PDF of the largest in the
sample, i.e., Yn = max ( X1 , X 2 ,..., X n ) .

2
Seoul National University Instructor: Junho Song
Dept. of Civil and Environmental Engineering [email protected]

Example 2: Suppose a certain

class of drivers will merge into
freely flowing traffic only if the
time between passing cars, 𝑋𝑖 is
at least 𝑦 sec. If the arrivals of
cars are following a Poisson
arrival process with the mean
occurrence rate 𝜈, what is the probability that after 𝑛 cars have passed, a driver will not have
been able to merge?

2. Asymptotic Distributions

An asymptotic distribution can be derived for a large-size sample, i.e., n →  , using Cramer’s
method (1946). For the extreme value distribution in Example 1, the exact (i.e., derived) and
asymptotic distributions by Cramer’s method are compared as follows:

3
Seoul National University Instructor: Junho Song
Dept. of Civil and Environmental Engineering [email protected]

The asymptotic distributions of the extremes tend to converge on certain limiting forms
(Gumbel 1958):

 Type I: The ( ) exponential form, exp −e− A( n ) y 

Gumbel distribution (largest)

𝐹𝑌𝑛 (𝑦) = exp[−𝑒 −𝛼𝑛 (𝑦−𝑢𝑛 ) ]
𝑓𝑌𝑛 (𝑦) = 𝛼𝑛 𝑒 −𝛼𝑛 (𝑦−𝑢𝑛 ) exp[−𝑒 −𝛼𝑛 (𝑦−𝑢𝑛 ) ]
Note: 𝑢𝑛 ~ location parameter, 𝛽𝑛 = 1/𝛼𝑛 ~ scale parameter

 Type II: The exponential form, exp − A(n) / y k 

Fréchet distribution (largest)

𝑣𝑛 𝑘
𝐹𝑌𝑛 (𝑦) = exp [− ( ) ]
𝑦
𝑘 𝑣𝑛 𝑘+1 𝑣𝑛 𝑘
𝑓𝑌𝑛 (𝑦) = ( ) exp [− ( ) ]
𝑣𝑛 𝑦 𝑦

 Type III: The exponential form with upper/lower bound, exp  − A(n) / ( − y) k 

Weibull distribution (smallest)

𝑦−𝜀 𝑘
𝐹𝑌1 (𝑦) = 1 − exp [− ( ) ]
𝑤1 − 𝜀

The type is determined by the ( ) behavior of the original probability density function.

 Exponentially decaying tail (e.g., Normal) → Type I

 Polynomial tail (e.g., Example 1) → Type II
 Polynomial tail with the limited extreme value → Type III

Let us check whether the distribution of the maximum value among 𝑛 = 10, 100 and 200
standard normal random variables converge to the Gumbel distribution:

install.packages('evd')
library('evd')
NormalMaxGen = function(n,N){
# n: number of sample points in each set
# N: number of sets
Sample = rnorm(n*N,0,1)
Sample_matrix = matrix(Sample,n,N)
SampleMax = apply(Sample_matrix,2,max) # Get maximum of each column
return(SampleMax)
}

4
Seoul National University Instructor: Junho Song
Dept. of Civil and Environmental Engineering [email protected]

u = function(n){
# location parameter u_n of Gumbel (Example 4.18 in A&T)
sqrt(2*log(n))-(log(log(n))+log(4*pi))/2/sqrt(2*log(n))
}

b = function(n){
# scale parameter b_n = 1/alpha_n (See Example 4.18 in A&T)
1/sqrt(2*log(n))
}

X10 = NormalMaxGen(10,10000) # Generate 10 normal random var. 10,000 times

X100 = NormalMaxGen(100,10000)
X200 = NormalMaxGen(200,10000)

xv = seq(-1,5,length=100)

plot(ecdf(X10),main="N=10,000",xlim=c(-1,5),col='red')
lines(xv,pgumbel(xv,u(10),b(10)),col='red',lty='dotted')

lines(ecdf(X100),col='black')
lines(xv,pgumbel(xv,u(100),b(100)),col='black',lty='dotted')

lines(ecdf(X200),col='blue')
lines(xv,pgumbel(xv,u(200),b(200)),col='blue',lty='dotted')

5
Seoul National University Instructor: Junho Song
Dept. of Civil and Environmental Engineering [email protected]

3. Generalized Extreme Value (GEV) Distribution

A general extreme value distribution model that can represent Type I (Gumbel), Type II
(Fréchet) and Type III (Weibull) distributions.

The CDF of GEV distribution is given as

−1/𝜉
𝑦−𝑢
𝐹𝑌 (𝑦) = exp [− (1 + 𝜉 ( )) ]
𝛽

where 𝑢, β(> 0), and 𝜉 are the location, scale and shape parameters. As ξ → 0, GEV becomes
Gumbel. For 𝜉 > 0, GEV is Fréchet distribution. When 𝜉 < 0, GEV is Weibull distribution. The
“evd” package of R provides dgev(), pgev(), qgev() and rgev() functions. One can estimate the
parameters from a data set using fgev() in the package as well.

* Image source and more details:

https://en.wikipedia.org/wiki/Generalized_extreme_value_distribution

6
Seoul National University Instructor: Junho Song
Dept. of Civil and Environmental Engineering [email protected]

Example 3: Download the dataset

‘GEVdata_1.txt’ from the eTL website.
The dataset includes yearly maximum
tension forces of cable #52 of Incheon
Grand Bridge simulated by finite
element analysis using artificially
generated traffic loads.

Source: Kim, J., and J. Song (2019). A

comprehensive probabilistic model of traffic loads based on Weigh-In-Motion data for
applications to bridge structures. KSCE Journal of Civil Engineering. Vol. 23(8), 3628-3643

Assuming the yearly maximum tension force follows a GEV distribution, estimate the
parameters of GEV and plot the corresponding PDF.

GEV=read.table("GEVdata_1.txt") # Read the dataset in numeric form

GEV_v = GEV[,1]
gev_fit = fgev(GEV_v) # Parameter estimation of GEV
# Parameter: shape=ξ, loc=u, scale=β
xv = seq(150,250,length=100)
param = gev_fit$estimate
gev_pdf = dgev(xv,loc=param[1],scale=param[2],shape=param[3])
plot(xv,gev_pdf,type="l",lty=1,col='red')
0.04
0.03
gev_pdf

0.02
0.01
0.00

160 180 200 220 240

xv