The Annals of Statistics

2015, Vol. 43, No. 1, 276–298

DOI: 10.1214/14-AOS1280
c Institute of Mathematical Statistics, 2015

ON THE BLOCK MAXIMA METHOD IN EXTREME VALUE

THEORY: PWM ESTIMATORS1
arXiv:1310.3222v2 [math.ST] 30 Dec 2014

By Ana Ferreira and Laurens de Haan

University of Lisbon and Erasmus Univ Rotterdam
In extreme value theory, there are two fundamental approaches,
both widely used: the block maxima (BM) method and the peaks-
over-threshold (POT) method. Whereas much theoretical research
has gone into the POT method, the BM method has not been studied
thoroughly. The present paper aims at providing conditions under
which the BM method can be justified. We also provide a theoretical
comparative study of the methods, which is in general consistent with
the vast literature on comparing the methods all based on simulated
data and fully parametric models. The results indicate that the BM
method is a rather efficient method under usual practical conditions.
In this paper, we restrict attention to the i.i.d. case and focus
on the probability weighted moment (PWM) estimators of Hosking,
Wallis and Wood [Technometrics (1985) 27 251–261].

1. Introduction. The block maxima (BM) approach in extreme value

theory (EVT), consists of dividing the observation period into nonover-
lapping periods of equal size and restricts attention to the maximum ob-
servation in each period [see, e.g., Gumbel (1958)]. The new observations
thus created follow—under domain of attraction conditions, cf. (2) below—
approximately an extreme value distribution, Gγ for some real γ. Parametric
statistical methods for the extreme value distributions are then applied to
those observations.
In the peaks-over-threshold (POT) approach in EVT, one selects those of
the initial observations that exceed a certain high threshold. The probability
distribution of those selected observations is approximately a generalized
Pareto distribution Pickands (1975).

Received April 2014; revised August 2014.

1
Supported in part by FCT Project PTDC /MAT /112770 /2009; EXPL/MAT-
STA/0622/2013 and PEst-OE/MAT/UI0006/2014.
AMS 2000 subject classifications. Primary 62G32; secondary 62G20, 62G30.
Key words and phrases. Block maxima, peaks-over-threshold, probability weighted mo-
ment estimators, extreme value index, asymptotic normality, extreme quantile estimation.

This is an electronic reprint of the original article published by the

Institute of Mathematical Statistics in The Annals of Statistics,
2015, Vol. 43, No. 1, 276–298. This reprint differs from the original in pagination
and typographic detail.
1
2 A. FERREIRA AND L. DE HAAN

In the case of the POT method, exact conditions under which the statis-
tical method is justified can be described by a second-order term [see, e.g.,
Drees (1998) and de Haan and Ferreira (2006), Section 2.3]. In the case of
block maxima, usually it is taken for granted that the maxima follow very
well an extreme value distribution. In this paper, we take this misspecifi-
cation into account by quantifying it in terms of a second-order expansion;
cf. Condition 2.1 below. Since Gγ is not the exact distribution for those
observations, a bias may appear.
The POT method picks up all “relevant” high observations. The BM
method on the one hand misses some of these high observations and, on the
other hand, might retain some lower observations. Hence the POT seems to
make better use of the available information.
There are practical reasons for using the BM method:
• The only available information may be block maxima, for example, yearly
maxima with long historical records or long range simulated data sets
Kharin et al. (2007).
• The BM method may be preferable when the observations are not ex-
actly independent and identically distributed (i.i.d.). For example, there
may be a seasonal periodicity in case of yearly maxima or, there may be
short range dependence that plays a role within blocks but not between
blocks; cf. for example, Katz, Parlange and Naveau (2002) and Madsen,
Rasmussen and Rosbjerg (1997) for further discussion.
• The BM method may be easier to apply since the block periods ap-
pear naturally in many situations [Naveau et al. (2009), van den Brink,
Können and Opsteegh (2005), de Valk (1993)]. On the other hand, the
POT method allows for greater flexibility in many cases since it might be
difficult to change the block size in practice.
When working with BM, there are two sets of estimators that are widely
used: the maximum likelihood (ML) estimators [e.g., Prescott and Walden
(1980)] and the probability weighted moment (PWM) estimators Hosking,
Wallis and Wood (1985). Recently, Dombry (2013) has proved consistency
of the former. The present paper concentrates on the latter. Our work has
given rise to the paper Bücher and Segers (2014) on the multivariate case.
The PWM estimators under the Gγ model are very popular, for exam-
ple, in applications to hydrologic and climatologic extremes, because of their
computational simplicity, good performance for small sample sizes and ro-
bustness even for location and scale parameters [Diebolt et al. (2008), Katz,
Parlange and Naveau (2002), Caires (2009), Hosking (1990)].
The relative merits of POT and BM have been discussed in several pa-
pers, all based on simulated data: Cunnane (1973) states that for γ = 0
and ML estimators, the POT estimate for a high quantile is better only if
the number of exceedances is larger than 1.65 times the number of blocks;
 ON THE BLOCK MAXIMA METHOD 3

Wang (1991) writes that POT is as efficient as BM model for high quan-
tiles, based on PWM estimators; Madsen, Pearson and Rosbjerg (1997) and
Madsen, Rasmussen and Rosbjerg (1997) write that POT is preferable for
γ > 0, whereas for γ < 0, BM is more efficient, again with the number of ex-
ceedances larger than the number of blocks; Martins and Stedinger (2001)
state that the gains (when using historical data) with the BM model are in
the range of the gains with the POT model, based on ML estimators; Caires
(2009) in a vast simulation study writes that with POT samples having an
average of two or more observations per block, the estimates are more accu-
rate than the corresponding BM estimates, and with more than 200 years
of data the accuracies of the two approaches are similar and rather good,
based on several estimators including the PWM and ML estimators.
From all these studies, some even with mixed views, the following two
features seem dominant. First, POT is more efficient than BM in many
circumstances, though needing, on average, a number of exceedances larger
than the number of blocks. Secondly, POT and BM often have comparable
performances, for example, for large sample sizes.
Our theoretical comparison shows that BM is rather efficient. The asymp-
totic variances of both extreme value index and quantile estimators are
always lower for BM than for POT. Moreover, the approximate minimal
mean square error is also lower for BM under usual circumstances. The op-
timal number of exceedances is generally higher than the optimal number
of blocks.
The paper is organized as follows. In Section 2, we state exact conditions
to justify the BM method, along with the asymptotic normality result for
the PWM estimators including high quantile estimators. In Section 3, we
provide a theoretical comparison between the two methods, BM and POT.
The analysis is based on a uniform expansion of the relevant quantile pro-
cess given in Section 2.1. This expansion also provides a basis for analysing
alternative estimators besides the PWM estimator. Proofs are postponed to
Section 4.
Throughout the paper, we assume that the observations are i.i.d. In future
work, we shall extend the results to the non-i.i.d. case and to the maximum
likelihood estimator.

2. The estimators and their properties. Let X̃1 , X̃2 , . . . be i.i.d. ran-
dom variables with distribution function F . Define for m = 1, 2, . . . and
i = 1, 2, . . . , k the block maxima
(1) Xi = max X̃j .
(i−1)m<j≤im

Hence, the m × k observations are divided into k blocks of size m. Write

n = m × k, the total number of observations. We study the model for large
4 A. FERREIRA AND L. DE HAAN

k and m, hence we shall assume that n → ∞; in order to obtain meaningful

limit results, we have to require that both m = mn → ∞ and k = kn → ∞,
as n → ∞.
The main assumption is that F is in the domain of attraction of some
extreme value distribution
Gγ (x) = exp(−(1 + γx)−1/γ ), γ ∈ R, 1 + γx > 0,
that is, for appropriately chosen am > 0 and bm and all x
 
Xi − bm
lim P ≤ x = lim F m (am x + bm )
m→∞ am m→∞
(2)
= Gγ (x), i = 1, 2, . . . , k.
This can be written as
1 1
lim = (1 + γx)1/γ ,
m→∞ m − log F (am x + bm )
which is equivalent to the convergence of the inverse functions:
V (mx) − bm xγ − 1
lim = , x > 0,
m→∞ am γ
with V = (−1/ log F )← . Hence, bm can be chosen to be V (m). This is the
first-order condition. For our analysis, we also need a second-order expansion
as follows.

Condition 2.1 (Second-order condition). Suppose that for some posi-

tive function a and some positive or negative function A with limt→∞ A(t) =
0,
Z x Z s
(V (tx) − V (t))/a(t) − (xγ − 1)/γ γ−1
lim = s uρ−1 du ds = Hγ,ρ (x),
t→∞ A(t) 1 1

for all x > 0 [see, e.g., de Haan and Ferreira (2006), Theorem B.3.1]. Note
that the function |A| is regularly varying with index ρ ≤ 0.

Let X1,k , . . . , Xk,k be the order statistics of the block maxima X1 , . . . , Xk .

The statistics β0 = k−1 ki=1 Xi,k and
P

k
1 X (i − 1) · · · (i − r)
(3) βr = Xi,k , r = 1, 2, 3, . . . , k > r,
k (k − 1) · · · (k − r)
i=1

are unbiased estimators of EX1 F rm (X1 ) [Landwehr, Matalas and Wallis

(1979)]. The PWM estimators for γ, as well as the location bm and scale
 ON THE BLOCK MAXIMA METHOD 5

am = a([m]), are simple functionals of β0 , β1 and β2 . The estimator γ̂k,m for

γ is defined as the solution of the equation
3γ̂k,m − 1 3β2 − β0
= ,
2γ̂k,m − 1 2β1 − β0
γ̂k,m 2β1 − β0 1 − Γ(1 − γ̂k,m )
(4) âk,m = γ̂
and b̂k,m = β0 + âk,m ,
2 k,m −1 Γ(1 − γ̂k,m ) γ̂k,m
R∞
where Γ(x) = 0 tx−1 e−t dt, x > 0 [Hosking, Wallis and Wood (1985)]. The
rationale behind the estimator of γ becomes clear when checking the state-
ment of Theorem 2.3 below.

2.1. Asymptotic normality. The following theorem is the basis for

analysing estimators in the BM approach. Let ⌈u⌉ represent the smallest
integer larger than or equal to u.

Theorem 2.1. Assume that F is in the domain of attraction of an ex-

treme value distribution Gγ and that Condition 2.1 holds. √ Let m = mn →
∞ and k = kn → ∞ as n → ∞, in such a way that kA(m) → λ ∈ R.
Let 0 < ε < 1/2 and {Xi,k }ki=1 be the order statistics of the block maxima
X1 , X2 , . . . , Xk . Then, with {Ek }k≥1 an appropriate sequence of Brownian
bridges,
√ X⌈ks⌉,k − bm (− log s)−γ − 1
 
k −
a0 (m) γ
√
 
Ek (s) 1
= + kA0 (m)Hγ,ρ
s(− log s)1+γ − log s
+ (s−1/2−ε (1 − s)−1/2−γ−ρ−ε )oP (1),
as n → ∞, where the oP (1) term is uniform for 1/(k + 1) ≤ s ≤ k/(k + 1).
The functions a0 (m) and A0 (m) are chosen as in Lemma 4.2 below.

Theorem 2.2. Assume the conditions of Theorem 2.1 with γ < 1/2.
Then
√ (r + 1)βr − bm
 
k − Dr (γ)
am
Z 1
d
→ (r + 1) sr−1 (− log s)−1−γ E(s) ds + λIr (γ, ρ) =: Qr ,
0
as n → ∞, jointly for r = 0, 1, 2, 3, . . . , where →d means convergence in dis-
tribution, E is Brownian bridge,
(r + 1)ξ Γ(1 − ξ) − 1
Dr (ξ) = , ξ<1
ξ
6 A. FERREIRA AND L. DE HAAN

[Dr (0) = log(r + 1) − Γ′ (1) as defined by continuity], and

1


 (Dr (γ + ρ) − Dr (γ)),
ρ






 ρ 6= 0,
(r + 1)γ


 Dr′ (γ) = (−Γ′ (1 − γ) + log(r + 1)Γ(1 − γ)



Ir (γ, ρ) = γ


 − (r + 1)−γ Dr (γ)),
γ 6= 0, ρ = 0,




1

Dr′ (0) = (log2 (r + 1) + Γ′′ (1) − 2 log(r + 1)Γ′ (1)),



2



γ = 0, ρ = 0.


R∞
Note that Γ′ (1 − γ) = 0 u−γ e−u (log u) du and Γ′′ (1) = 1.97811.
√
Remark 2.1. The condition kA(m) → λ ∈ R means that the growth
of kn , the number of blocks, is restricted with respect to the growth of
mn , the size of a block, as n → ∞. In particular this condition implies that
(log k)/m → 0, as n → ∞.

Theorem 2.3. Under the conditions of Theorem 2.2, as n → ∞,

−1
√

d 1 log 3 log 2
k(γ̂k,m − γ) → −
Γ(1 − γ) 1 − 3−γ 1 − 2−γ
 
γ γ
× (Q2 − Q0 ) − γ (Q1 − Q0 )
3γ − 1 2 −1
=: ∆,
√
 
âk,m γ
k − 1 →d γ (Q1 − Q0 )
am (2 − 1)Γ(1 − γ)
Γ′ (1 − γ)
   
log 2 −γ 1
+∆ + +
γ 1 − 2−γ log 2 Γ(1 − γ)
=: Λ,
√ b̂k,m − bm γΓ′ (1 − γ) − 1 + Γ(1 − γ) 1 − Γ(1 − γ)
k →d Q 0 + 2
∆+ Λ
am γ γ
=: Ξ;
where for γ = 0 the formulas should read as (defined by continuity):
√ log 3 log 2 −1
   
d 1 1
kγ̂k,m → − (Q2 − Q0 ) − (Q1 − Q0 ) ,
2 2 log 3 log 2
 ON THE BLOCK MAXIMA METHOD 7

√
   
âk,m d 1 log 2 ′
k −1 → (Q1 − Q0 ) + ∆ + Γ (1) ,
am log 2 2
√ b̂k,m − bm
k →d Q0 − Γ′′ (1)∆ + Γ′ (1)Λ.
am

Remark 2.2. A slight modification of γ̂k,m produces the explicit esti-

mator
 
∗ 1 4β3 − β0
(5) γ̂k,m = log −1 ,
log 2 2β1 − β0
∗ ∗
which is the solution of (4γ̂k,m − 1)(2γ̂k,m − 1)−1 = (4β3 − β0 )(2β1 − β0 )−1 .
The conditions of Theorem 2.2 imply
−1
√

∗ 1 log 4 log 2
k(γ̂k,m − γ) →d −
Γ(1 − γ) 1 − 4−γ 1 − 2−γ
 
γ γ
× (Q3 − Q0 ) − γ (Q1 − Q0 ) .
4γ − 1 2 −1

2.2. High quantile estimation. Our estimator for xn = F ← (1 − pn ) =

V (1/(− log(1 − pn ))), with pn small, is
(mpn )−γ̂k,m − 1
x̂k,m = b̂k,m + âk,m .
γ̂k,m

Theorem 2.4. Assume the conditions of Theorem 2.2 with ρ negative,

or zero with γ negative. Moreover, assume that the probabilities pn satisfy
log(mpn )
lim mpn = 0 and lim√ =0
n→∞ n→∞ k
√
[in case ρ < 0 the latter can be simplified to limn→∞ (log pn )/ k = 0]. Then
√ (x̂k,m − xn ) γ−
k →d ∆ + (γ− )2 Ξ − γ− Λ − λ
am qγ (1/(mpn )) γ− + ρ
R t γ−1
as n → ∞, where γ− = min(0, γ) and qγ (t) = 1 s log s ds.

3. Theoretical comparison between BM and POT methods. In this sec-

tion, we develop a theoretical comparison between the BM and POT meth-
ods, by comparing the two PWM estimators for the two methods [Hosking
and Wallis (1987) and Hosking, Wallis and Wood (1985), resp., for POT and
BM].
8 A. FERREIRA AND L. DE HAAN

First, we introduce the PWM-POT estimators for γ and a(n/k), where

k−1
k is the number of selected order statistics, {X̃n−i,n }i=0 , from the original
sample X̃1 , X̃2 , . . . , X̃n . The statistics
k−1 k−1
1X 1Xi
Pn = X̃n−i,n − X̃n−k,n and Qn = (X̃n−i,n − X̃n−k,n )
k k k
i=0 i=0

are estimators for a(n/k)(1 − γ)−1

and a(n/k)(2(2 − γ)−1 ), respectively.
Consequently, the PWM estimators are
 −1  −1
Pn Pn
γ̂k,n = 1 − −1 and â(n/k) = Pn −1 .
2Qn 2Qn
The quantile estimator is
(k/(npn ))γ̂k,n − 1
x̂k,n = X̃n−k,n + â(n/k) .
γ̂k,n
Asymptotic normality under conditions equivalent to the ones in Theorems
2.3 and 2.4 holds [see, e.g., Cai, de Haan and Zhou (2013)], if ρ ∈ [−1, 0] with
a caveat for ρ = −1 [for certain cases the functions A in the corresponding
second-order conditions may not be the same asymptotically resulting in
different values of λ in the limiting distributions; cf. Drees, de Haan and Li
(2003)].
For BM, k is defined as the number of blocks and, for POT, k is defined as
the number of selected top order statistics. Hence, in both cases k means the
number of selected observations. For the theoretical comparison, we confine
ourselves to the range ρ ∈ [−1, 0] and γ ∈ [−1, 1/2), a usual range in many
applications.

Extreme value index estimators.

• First, we compare asymptotic variance and bias for a common value of k:
The asymptotic variances of the two γ estimators are shown in Figure 1:
the curve from BM is always below the other one, meaning lower values
for the asymptotic variance for all values of γ. The asymptotic biases are
compared in Figure 2, through the ratio “bias BM/bias POT”. Recall
that the bias depends on both first- and second-order parameters γ and
ρ. Contrary to what is observed for the variance, the bias of BM is always
larger but for ρ = 0 they are the same regardless the value of γ,√
equal to 1
[or λ if one takes into account the asymptotic contribution of kA(n/k)
to the biases].
• Next, we compare asymptotic mean square errors for the “optimal choice”
of k (i.e., that value that makes the limiting mean square error of γ̂ − γ
minimal), which is different in the two cases:
 ON THE BLOCK MAXIMA METHOD 9

Fig. 1. Asymptotic variances of γ PWM estimators with dashed line for POT.

An asymptotic expression of the “asymptotic minimal mean square

error” (MINMSE in the sequel) is obtained in the following way. Sup-
pose ρ < 0. First we find for each estimator the optimal k in the sense
of minimizing the approximate asymptotic mean square error. Denote by
σi2 = σi2 (γ) and Bi2 = Bi2 (γ, ρ) (i = 1, 2; “1” refers to PWM-BM and “2”
refers to PWM-POT) the asymptotic variance and squared R ∞ bias of the
estimators. Under Condition 2.1, we can write A2 (t) = t s(u) du with
s(·) decreasing and 2ρ − 1 regularly varying. The limiting mean square
error is, approximately,
 2 
σi 2 2
(6) inf + A (n/k)Bi
k k

Fig. 2. Ratio of asymptotic bias of γ PWM estimators.

10 A. FERREIRA AND L. DE HAAN

Fig. 3. Contour plot for the ratio of asymptotic minimal mean square error of γ PWM
estimators.

R∞
or, writing r for n/k, inf r ((r/n)σi2 + Bi2 r s(u) du). Setting the derivative
equal to zero and using properties of regularly varying functions one finds
(i)
for the optimal choice of r, r0 ∼ (1/s)← (n)(Bi2 /σi2 )1/(1−2ρ) and, in terms
of k,
 2 1/(1−2ρ)
(i) n σi
k0 ∼ .
(1/s) (n) Bi2
←

(i)
Note that the optimal k0 is different but of the same order for both
(i)
methods. Next, inserting k0 in (6), after some manipulation we get the
following asymptotic expression for MINMSE,
1 − 2ρ (1/s)← (n) 2 1/(1−2ρ) 2 −2ρ/(1−2ρ)
(Bi ) (σi ) .
−2ρ n
It follows that MINMSE(BM)/MINMSE(POT) is, approximately,

B1 (γ, ρ) 1/(1−2ρ) σ12 (γ) −2ρ/(1−2ρ)

 2   
,
B22 (γ, ρ) σ22 (γ)
which does not depend on n, just on γ and ρ.
The contour plot of “MINMSE(BM)/MINMSE(POT)” is represented
in Figure 3. It can be seen that the BM has lower MINMSE for a large
range of (γ, ρ) combinations. Note that this range includes γ negative and
γ positive close to zero which seem to be common values in many practical
 ON THE BLOCK MAXIMA METHOD 11

Fig. 4. Contour plot for the ratio of the optimal values of k.

situations, for example, in hydrologic and climatologic extremes. Only for

γ > 0.2 approximately, MINMSE for POT can be lower depending on ρ.
Finally, comparing the optimal sample sizes (cf. Figure 4 with con-
tour plot of the ratio of the optimal values of k), one sees that POT
requires systematically larger optimal sample size even when the approx-
imate MINMSE is smaller for POT than BM.

Quantile estimators. We repeat the previous analysis for the quantile

estimators:
• The asymptotic variances of the two estimators are compared in Figure 5:
again the curve from BM is always below the other one meaning lower
values for the asymptotic variance for all values of γ. In Figures 6 and 7,
the asymptotic bias is represented for each case separately. Note that for
γ negative, the bias for BM approaches zero when ρ ↑ 0 whereas in the
POT case it escapes to −∞.
• The contour plot for the ratio “MINMSE(BM)/MINMSE(POT)” is rep-
resented in Figure 8. Again the BM method has lower MINMSE for a
large range of (γ, ρ) combinations. The “irregularity” around γ ≈ −0.2 is
due to a change of sign in the bias in the POT case. Finally, Figure 9 gives
the contour plot for the ratio of the optimal values of k, which is smaller
than one when γ is small and ρ is closer to zero.
In conclusion, for both the extreme value index and quantile PWM esti-
mators, the ones from the BM method have always lower asymptotic vari-
ances. Moreover, at an optimal level the BM gives lower MINMSE, thus
12 A. FERREIRA AND L. DE HAAN

Fig. 5. Asymptotic variances of quantile PWM estimators with dashed line for POT.

being more efficient, under many practical situations. This is in agreement

with some of Sofia Caires’ (2009) conclusions, for example, that for equal
sample sizes or with more than 200 years of data the uncertainty or the error
of the estimates are lower for BM than for POT.

4. Proofs. Throughout this section, Z represents a unit Fréchet random

variable, that is, one with distribution function F (x) = e−1/x , x > 0, and
{Zi,k }ki=1 are the order statistics from the associated i.i.d. sample of size k,
Z1 , . . . , Zk . Similarly, {Xi,k }ki=1 represents the order statistics of the block
maxima X1 , . . . , Xk from (1) and, X⌈u⌉,k = Xr,k for r −1 < u ≤ r, r = 1, . . . , k.
Recall the function V from Section 2. The following representation will be
useful:
(7) X =d V (mZ).

Fig. 6. Asymptotic bias of quantile PWM estimator under BM method.

 ON THE BLOCK MAXIMA METHOD 13

Fig. 7. Asymptotic bias of quantile PWM estimator under POT method.

We start by formulating a number of auxiliary results.

Lemma 4.1. 1. As k → ∞,

(log k)Z1,k →P 1.

Fig. 8. Contour plot for the ratio of asymptotic minimal mean square error of quantile
PWM estimators.
14 A. FERREIRA AND L. DE HAAN

Fig. 9. Contour plot for the ratio of the optimal values of k in quantile case.

2. [Csörgő and Horváth (1993), page 381] Let 0 < ν < 1/2. With {Ek }k≥1 ,
an appropriate sequence of Brownian bridges,
s(− log s) √
 
Ek (s) 
sup k((− log s)Z⌈ks⌉,k −1)− = oP (1),
(s(1 − s))ν 
1/(k+1)≤s≤k/(k+1) s(− log s) 
as k → ∞ (⌈u⌉ represents the smallest integer larger or equal to u).
3. Similarly, with 0 < ν < 1/2 for an appropriate sequence {Ek }k≥1 of Brow-
nian bridges and ξ ∈ R,
sup (s(1 − s))−ν
1/(k+1)≤s≤k/(k+1)
 Zξ
√ ⌈ks⌉,k − 1 (− log s)−ξ − 1
  
1+ξ

×  ks(− log s)
 − − Ek (s) = oP (1),
ξ ξ
as k → ∞.

The following is an easily obtained variant of Theorem B.3.10 of de Haan

and Ferreira (2006).

Lemma 4.2. Under Condition 2.1, there are functions A0 (t) ∼ A(t) and
a0 (t) = a(t)(1 + o(A(t))), as t → ∞, such that for all ε, δ > 0 there exists
t0 = t0 (ε, δ) such that for t, tx > t0 ,
 (V (tx) − V (t))/a0 (t) − (xγ − 1)/γ
 

 − Hγ,ρ (x)
 A0 (t)
 ON THE BLOCK MAXIMA METHOD 15

(8)
≤ ε max(xγ+ρ+δ , xγ+ρ−δ ).
Moreover,
 a0 (tx)/a0 (t) − xγ
 
γ ρ

 − x (x − 1)/ρ
 A0 (t)
(9)
≤ ε max(xγ+ρ+δ , xγ+ρ−δ )
and
 
 A0 (tx) ρ
 ρ+δ ρ−δ

 A0 (t) − x  ≤ ε max(x , x ).

Note that
 
1 xγ+ρ − 1 xγ − 1

− , ρ 6= 0 6= γ,





 ρ γ+ρ γ
xγ − 1
  
1 γ


x log x − , ρ = 0 6= γ,



Hγ,ρ (x) = γ  γ
1 xρ − 1


− log x , ρ 6= 0 = γ,


ρ ρ




1


 (log x)2 ,

 ρ = 0 = γ.
2

Proof of Theorem 2.1. By representation (7),

X⌈ks⌉,k − bm (− log s)−γ − 1
−
a0 (m) γ
V (mZ⌈ks⌉,k ) − bm V (m/− log s) − bm
 
d
= −
a0 (m) a0 (m)
V (m/− log s) − bm (− log s)−γ − 1
 
+ −
a0 (m) γ
= I (random part) + II (bias part).
We start with part I,

−γ V ((− log s)Z⌈ks⌉,k m/− log s) − V (m/− log s)

 
I = (− log s)
a0 (m/− log s)
 
a0 (m/− log s)
× (− log s)γ
a0 (m)
= I.1 × I.2.
16 A. FERREIRA AND L. DE HAAN

According to (9) of Lemma 4.2, for each ε, δ > 0 there exists t0 such that
the factor I.2 is bounded (above and below) by
(− log s)−ρ − 1
 
−ρ+δ −ρ−δ
1 + A0 (m) ± ε max((− log s) , (− log s) )
ρ
provided m ≥ t0 and s ≥ e−m/t0 . According to (8) of Lemma 4.2, for factor
I.1 we have the bounds
γ
−γ ((− log s)Z⌈ks⌉,k ) − 1
 
m
(− log s) + A0 (− log s)−γ
γ − log s
× {Hγ,ρ ((− log s)Z⌈ks⌉,k ) ± ε max(((− log s)Z⌈ks⌉,k )γ+ρ+δ ,

((− log s)Z⌈ks⌉,k )γ+ρ−δ )}

= I.1a + I.1b
provided s√≥ e−m/t0 and m/ log k ≥ t0 [the latter inequality eventually holds
true since kA0 (m) is bounded]. Note that m/ log k ≥ t0 implies mZ1,k ≥ 2t0
which implies (Lemma 4.1) mZ⌈ks⌉,k ≥ 2t0 for all s.
For term I.1a, we use Lemma 4.1.3:
(Z⌈ks⌉,k )γ − 1 (− log s)−γ − 1
−
γ γ
is bounded (above and below) by
1 Ek (s) ε (s(1 − s))ν
√ ± √ ,
k s(− log s)1+γ k s(− log s)1+γ
for some ε > 0, 0 < ν < 1/2 and all s ∈ [1/(k + 1), k/(k + 1)].
m
Next, we turn to term I.1b. By Lemma 4.2, (− log s)−γ A0 ( − log s ) is bounded
(above and below) by
A0 (m){(− log s)−γ−ρ ± ε max((− log s)−γ−ρ+δ , (− log s)−γ−ρ−δ )}
provided s > e−m/t0 and m/ log k > t0 . Furthermore for ρ 6= 0 6= γ and s ∈
[1/(k + 1), k/(k + 1)], by Lemma 4.1.3,
Hγ,ρ ((− log s)Z⌈ks⌉,k )
1 ((− log s)Z⌈ks⌉,k )γ+ρ − 1 ((− log s)Z⌈ks⌉,k )γ − 1
 
= −
ρ γ+ρ γ
 Z γ+ρ − 1
(− log s)−γ−ρ − 1
 
1 γ+ρ ⌈ks⌉,k
= (− log s) −
ρ γ +ρ γ +ρ
 Zγ
⌈ks⌉,k − 1 (− log s)−γ − 1

γ
− (− log s) −
γ γ
 ON THE BLOCK MAXIMA METHOD 17

is bounded by
(s(1 − s))ν
  
1 γ+ρ 1 Ek (s) ε
(− log s) √ ±√
ρ k s(− log s)1+γ+ρ k s(− log s)1+γ+ρ
ε (s(1 − s))ν
 
γ 1 Ek (s)
− (− log s) √ ∓√
k s(− log s)1+γ k s(− log s)1+γ
2ε (s(1 − s))ν
=± √ ,
ρ k s(− log s)
and similarly for cases other than ρ 6= 0 6= γ. The remaining part of I.1b,
namely ±ε max(((− log s)Z⌈ks⌉,k )γ+ρ+δ , ((− log s)Z⌈ks⌉,k )γ+ρ−δ ), is similar.
Part II, by the inequalities of Lemma 4.2, is bounded by
   
1 −γ−ρ+δ −γ−ρ−δ
A0 (m) Hγ,ρ ± ε max((− log s) , (− log s) )
− log s
√ 1
hence it contributes kA0 (m)Hγ,ρ ( − log s ) to the result.
Collecting all the terms, one finds the result. 

Proof of Theorem 2.2. Let, for r = 0, 1, 2, 3, . . . ,

(r) (⌈ks⌉ − 1) · · · (⌈ks⌉ − r)
Jk (s) = , s ∈ [0, 1].
(k − 1) · · · (k − r)
(r)
Note that Jk (s) → sr , as k → ∞, uniformly in s ∈ [0, 1], and
k 1
1 X (i − 1) · · · (i − r) 1
Z
(r)
= Jk (s) ds =
k (k − 1) · · · (k − r) 0 r+1
i=1
Z 1
= sr ds.
0

Then
√ (r + 1)βr − bm (r + 1)γ Γ(1 − γ) − 1
 
k −
am γ
√ (r + 1) 0 X⌈ks⌉,k Jk(r) (s) ds − bm
 R1
= k
am
Z 1
(− log s)−γ − 1 r

− (r + 1) s ds
0 γ
√ Z 1
X⌈ks⌉,k − bm (− log s)−γ − 1 (r)

= k(r + 1) − Jk (s) ds
0 am γ
18 A. FERREIRA AND L. DE HAAN

√ Z 1
(− log s)−γ − 1 r (r)
− k(r + 1) (s − Jk (s)) ds
0 γ
√ 1/(k+1) X
⌈ks⌉,k − bm (− log s)−γ − 1 (r)
Z  
= k(r + 1) − Jk (s) ds
0 am γ
√ Z k/(k+1) 
X⌈ks⌉,k − bm (− log s)−γ − 1

(r)
+ k(r + 1) − Jk (s) ds
1/(k+1) a m γ
√ 1 X⌈ks⌉,k − bm (− log s)−γ − 1 (r)
Z  
+ k(r + 1) − Jk (s) ds
k/(k+1) am γ
√ Z 1
(− log s)−γ − 1 r (r)
− k(r + 1) (s − Jk (s)) ds
0 γ
= I.1 + I.2 + I.3 + I.4.
(r) √
For I.4: since (sr − Jk (s)) = O(1/k) uniformly in s, I.4 = O(1/ k).
For I.1, note that
Z 1/(k+1) √
X⌈ks⌉,k − bm r
(10) k s ds = oP (1).
0 am
This follows since, the left-hand side of (10) equals, in distribution,
√
k V (mZ1,k ) − V (m)
(k + 1) r+1 am
which, by Lemmas 4.1.1, 4.2 and the fact that m/ log k → ∞, is bounded
(below and above) by
√  γ
Z1,k − 1

k γ+ρ+δ γ+ρ−δ
+ A(m)Hγ,ρ (Z1,k ) ± εA(m) max(Z1,k , Z1,k ) .
(k + 1)r+1 γ
ξ
√
This is easily seen to converge to zero in probability, since Z1,k / k=
√ √
{(log k)Z1,k }ξ log−ξ k/ k →P 0 for all real ξ and kA(m) → λ. Hence, I.1 =
oP (1).
Next, we show that
Z 1 √ X⌈ks⌉,k − bm (r)
(11) k Jk (s) ds = oP (1).
k/(k+1) am
(r)
The left-hand side equals, in distribution, since Jk (s) ≡ 1 for s ∈ (k(k +
1)−1 , 1),
√ V (mZk,k ) − V (m)
 
k
1− k .
k+1 am
 ON THE BLOCK MAXIMA METHOD 19

Lemma 4.1 yields

γ
V (mZk,k ) − V (m) Zk,k − 1 γ+ρ+δ
= + A(m){Hγ,ρ (Zk,k ) ± εZk,k },
am γ
γ
which is (since Zk,k /kγ converges to a positive random variable) of the order
√
OP (kγ ). Hence, the integral is of order (k + 1)−1 kkγ which tends to zero
since γ < 1/2.
Finally, I.2 has the same asymptotic behaviour as
Z k/(k+1) √ 
X⌈ks⌉,k − bm (− log s)−γ − 1 r

(r + 1) k − s ds,
1/(k+1) am γ
which, by Theorem 2.1 tends to
Z 1 1  
1
Z
r−1 −1−γ
(r + 1) s (− log s) E(s) ds + λ(r + 1) Hγ,ρ sr ds.
0 0 − log s
For the evaluation of the latter integral note that for ξ < 1,
Z 1 Z ∞
(r + 1) sr (− log s)−ξ ds = (r + 1)ξ−1 v −ξ e−v dv = (r + 1)ξ−1 Γ(1 − ξ).
0 0
Moreover, note that
Z 1
(− log s)−ξ − 1 (r + 1)ξ Γ(1 − ξ) − 1
(r + 1) sr ds = , ξ<1
0 ξ ξ
[D (0) = log(r + 1) − Γ′ (1) as defined by continuity], and (r + 1) ×
R 1r 1 r
0 Hγ,ρ ( − log s )s ds = Ir (γ, ρ). 

Proof of Theorem 2.3. From Theorem 2.2, we obtain

√ 2β1 − β0 2γ − 1
 
k − Γ(1 − γ) →d Q1 − Q0 ,
am γ
√ 3β2 − β0 3γ − 1
 
k − Γ(1 − γ) →d Q2 − Q0
am γ
hence, by Cramér’s delta method,
√ 3γ̂k,m − 1 3γ − 1
 
k γ̂ −
2 k,m − 1 2γ − 1
√ 3β2 − β0 3γ − 1
 
= k −
2β1 − β0 2γ − 1
3γ − 1
 
d 1 γ γ
→ (Q2 − Q0 ) − γ (Q1 − Q0 ) .
Γ(1 − γ) 2γ − 1 3γ − 1 2 −1
20 A. FERREIRA AND L. DE HAAN

It follows that γ̂k,m →P γ, and hence

√ r γ̂k,m − 1 √ r γ̂k,m −γ − 1
 
k − 1 = k
rγ − 1 1 − r −γ
has the same limit distribution as
√ log r
k(γ̂k,m − γ) , r = 2, 3.
1 − r −γ
It follows that
√ 3γ̂k,m − 1 3γ − 1
 
k γ̂ −
2 k,m − 1 2γ − 1
3γ − 1 √ 3γ̂k,m − 1 √ 2γ̂k,m − 1
    
= γ k − 1 − k − 1
2 −1 3γ − 1 2γ − 1
has the same limit distribution as
3γ − 1 √
 
log 3 log 2
k(γ̂k,m − γ) −
2γ − 1 1 − 3−γ 1 − 2−γ
and, consequently,
√
k(γ̂k,m − γ)
 −1
d 1 log 3 log 2
→ −
Γ(1 − γ) 1 − 3−γ 1 − 2−γ
 
γ γ
× (Q2 − Q0 ) − γ (Q1 − Q0 ) .
3γ − 1 2 −1
For the asymptotic distribution of âk,m we write
√ âk,m
 
γ̂k,m
k −1 γ̂
am (2 k,m − 1)Γ(1 − γ̂k,m)
√ 2β1 − β0 2γ − 1
  
× k − Γ(1 − γ)
am γ
√ 2γ − 1 2γ̂k,m − 1
 
+ k Γ(1 − γ) − Γ(1 − γ̂k,m) ,
γ γ̂k,m
and the statement follows, for example, by Cramér’s delta method.
For the asymptotic distribution of b̂k,m , we write
√ b̂k,m − bm
 
k
am
√ β0 − bm Γ(1 − γ) − 1
 
= k −
am γ
 ON THE BLOCK MAXIMA METHOD 21

âk,m √ Γ(1 − γ̂k,m) − 1 Γ(1 − γ) − 1

 
− k −
am γ̂k,m γ
Γ(1 − γ) − 1 √ âk,m
 
+ k −1
γ am
and the statement follows, for example, by Cramér’s delta method. 

Proof of Theorem 2.4. The proof follows the line of the comparable
result for the POT method [see, e.g., de Haan and Ferreira (2006), Chap-
ter 4.3]. Let cn = 1/(mpn ). Then
√
k(x̂k,m − xn )
am qγ (cn )
√ γ̂
cnk,m − 1
  
k 1
= b̂k,m + âk,m −V
am qγ (cn ) γ̂k,m − log(1 − pn )
√ √  γ̂k,m
cn − 1 cγn − 1

k b̂k,m − bm âk,m k
= + −
qγ (cn ) am am qγ (cn ) γ̂ γ
√ 
V (m/(−m log(1 − pn ))) − V (m) cγn − 1

k
− −
qγ (cn ) am γ
γ
cn − 1 √ âk,m
 
+ k −1 .
γqγ (cn ) am
Similarly, as on pages 135–137 of de Haan and Ferreira (2006), this converges
in distribution to
γ−
∆ + (γ− )2 Ξ − γ− Λ − λ . 
γ− + ρ

APPENDIX: ASYMPTOTIC VARIANCES AND BIASES OF THE

PWM ESTIMATORS
The following provides a basis for an algorithm to calculate the asymptotic
variances/covariances and biases of the PWM estimators.
R1
Let Qr = (r + 1) 0 sr−1 (− log s)−1−γ E(s) ds + λIr (γ, ρ), r = 0, 1, 2, as de-
fined in Theorem 2.2. For r, m = 0, 1, 2,
Cov(Qr , Qm )
= (r + 1)(m + 1)
Z 1Z 1
× sr−1 um−1 (− log s)−1−γ (− log u)−1−γ EB(s)B(u) ds du
0 0
(12)
22 A. FERREIRA AND L. DE HAAN
Z 1 Z u
m−1 −1−γ
= (r + 1)(m + 1) u (1 − u)(− log u) sr (− log s)−1−γ ds du
0 0
+ (r + 1)(m + 1)
Z 1 Z s
r−1 −1−γ
× s (1 − s)(− log s) um (− log u)−1−γ du ds
0 0
using the fact that EB(s)B(u) = min(s, u) − su = s(1 − u) for 0 < s < u.
These integrals can be evaluated numerically (we have used Mathematica
software).
From Theorems 2.3 and 2.4, after some calculations,
√
(13) k(γ̂k,m − γ) →d Cγ (kγ,0 Q0 + kγ,1 Q1 + kγ,2 Q2 ),
√ âk,m
 
(14) k − 1 →d ka,0 Q0 + ka,1 Q1 + ka,2 Q2 ,
am
√ b̂k,m − bm
(15) k →d kb,0 Q0 + kb,1 Q1 + kb,2 Q2 ,
am
√ (x̂k,m − xn )
(16) k →d kx,0 Q0 + kx,1 Q1 + kx,2 Q2 ,
am qγ (cn )
where, for γ 6= 0,
 −1
1 log 3 log 2
Cγ = − ,
Γ(1 − γ) 1 − 3−γ 1 − 2−γ
γ(3γ − 2γ ) −γ γ
kγ,0 = γ γ
, kγ,1 = γ , kγ,2 = ;
(3 − 1)(2 − 1) 2 −1 3γ −1
Γ′ (1 − γ)
 
log 2 1 γ
Ca = − + ,
γ log 2 1 − 2−γ Γ(1 − γ)
γ
ka,0 = Cγ kγ,0 Ca − γ ,
(2 − 1)Γ(1 − γ)
γ
ka,1 = Cγ kγ,1 Ca + γ ,
(2 − 1)Γ(1 − γ)
γ
ka,2 = Cγ kγ,2 Ca − γ ;
(2 − 1)Γ(1 − γ)
γΓ′ (1 − γ) − 1 + Γ(1 − γ)
Cb = ,
γ2
1 − Γ(1 − γ)
kb,0 = 1 + Cγ kγ,0 Cb + kγ,0 ,
γ
1 − Γ(1 − γ)
kb,1 = Cγ kγ,1 Cb + kγ,1 ,
γ
 ON THE BLOCK MAXIMA METHOD 23

1 − Γ(1 − γ)
kb,2 = Cγ kγ,2 Cb + kγ,2 ;
γ
kx,0 = Cγ kγ,0 + (γ− )2 kb,0 − γ− ka,0 ,
kx,1 = Cγ kγ,1 + (γ− )2 kb,1 − γ− ka,1 , kx,2 = Cγ kγ,2 + (γ− )2 kb,2 − γ− ka,2
and, for γ = 0,
Cγ = 2(log(3/2)) −1 , kγ,0 = (log 2)−1 − (log 3)−1 ,
kγ,1 = −(log 2)−1 , kγ,2 = (log 3)−1 ;
Ca = 2−1 log 2 + Γ′ (1), Cb = −Γ′′ (1)
and the rest follow similarly by continuity. Then the asymptotic variances,
covariances and biases follow by combining (12) with (13)–(16) in the obvi-
ous way.

Acknowledgements. We thank Holger Drees for a useful suggestion.

We would like to thank three unknown referees for their genuine interest
and their insightful comments.

REFERENCES
Bücher, A. and Segers, J. (2014). Extreme value copula estimation based on block
maxima of a multivariate stationary time series. Extremes 17 495–528. MR3252823
Cai, J.-J., de Haan, L. and Zhou, C. (2013). Bias correction in extreme value statistics
with index around zero. Extremes 16 173–201. MR3057195
Caires, S. (2009). A comparative simulation study of the annual maxima and the peaks-
over-threshold methods. SBW-Belastingen: Subproject “Statistics”. Deltares Report
1200264-002.
Csörgő, M. and Horváth, L. (1993). Weighted Approximations in Probability and Statis-
tics. Wiley, Chichester. MR1215046
Cunnane, C. (1973). A particular comparison of annual maxima and partial duration
series methods of flood frequency prediction. J. Hydrol. 18 257–271.
de Haan, L. and Ferreira, A. (2006). Extreme Value Theory: An Introduction. Springer,
New York. MR2234156
de Valk, C. (1993). Estimation of marginals from measurements and hindcast data.
WL—Delft Hydraulics Report H1700.
Diebolt, J., Guillou, A., Naveau, P. and Ribereau, P. (2008). Improving probability-
weighted moment methods for the generalized extreme value distribution. REVSTAT
6 35–50. MR2386298
Dombry, C. (2013). Maximum likelihood estimators for the extreme value index based
on the block maxima method. Available at arXiv:1301.5611.
Drees, H. (1998). On smooth statistical tail functionals. Scand. J. Stat. 25 187–210.
MR1614276
Drees, H., de Haan, L. and Li, D. (2003). On large deviation for extremes. Statist.
Probab. Lett. 64 51–62. MR1995809
Gumbel, E. J. (1958). Statistics of Extremes. Columbia Univ. Press, New York.
MR0096342
24 A. FERREIRA AND L. DE HAAN

Hosking, J. R. M. (1990). L-moments: Analysis and estimation of distributions using

linear combinations of order statistics. J. R. Stat. Soc. Ser. B Stat. Methodol. 52 105–
124. MR1049304
Hosking, J. R. M. and Wallis, J. R. (1987). Parameter and quantile estimation for the
generalized Pareto distribution. Technometrics 29 339–349. MR0906643
Hosking, J. R. M., Wallis, J. R. and Wood, E. F. (1985). Estimation of the gen-
eralized extreme-value distribution by the method of probability-weighted moments.
Technometrics 27 251–261. MR0797563
Katz, R. W. Parlange, M. B. and Naveau, P. (2002). Statistics of extremes in hy-
drology. Advances in Water Resources 25 1287–1304.
Kharin, V. V., Zwiers, F. W., Zhang, X. and Hegerl, G. C. (2007). Changes in
temperature and precipitation extremes in the IPCC ensemble of global coupled model
simulations. Journal of Climate 20 1419–1444.
Landwehr, J., Matalas, N. and Wallis, J. (1979). Probability weighted moments com-
pared with some traditional techniques in estimating Gumbel parameters and quantiles.
Water Resources Research 15 1055–1064.
Madsen, H., Pearson, C. P. and Rosbjerg, D. (1997). Comparison of annual maximum
series and partial duration series methods for modeling extreme hydrologic events 2.
Regional modeling. Water Resources Research 33 759–769.
Madsen, H., Rasmussen, P. F. and Rosbjerg, D. (1997). Comparison of annual maxi-
mum series and partial duration series methods for modeling extreme hydrologic events
1. At-site modeling. Water Resources Research 33 747–757.
Martins, E. S. and Stedinger, J. R. (2001). Historical information in a generalized max-
imum likelihood framework with partial duration and annual maximum series. Water
Resources Research 37 2559–2567.
Naveau, P., Guillou, A., Cooley, D. and Diebolt, J. (2009). Modelling pairwise
dependence of maxima in space. Biometrika 96 1–17. MR2482131
Pickands, J. III (1975). Statistical inference using extreme order statistics. Ann. Statist.
3 119–131. MR0423667
Prescott, P. and Walden, A. T. (1980). Maximum likelihood estimation of the param-
eters of the generalized extreme-value distribution. Biometrika 67 723–724. MR0601119
van den Brink, H. W., Können, G. P. and Opsteegh, J. D. (2005). Uncertainties in
extreme surge level estimates from observational records. Philos. Trans. R. Soc. Lond.
Ser. A Math. Phys. Eng. Sci. 363 1377–1386.
Wang, Q. J. (1991). The POT model described by the generalized Pareto distribution
with Poisson arrival rate. Journal of Hidrology 129 263–280.

ISA Department of Mathematics

Department of Economics Erasmus University Rotterdam
University of Lisbon P.O. Box 1738
Tapada da Ajuda 1349-017 Lisboa 3000 DR Rotterdam
Portugal The Netherlands
and and
CEAUL CEAUL
FCUL FCUL
Bloco C6 - Piso 4 Campo Grande Bloco C6 - Piso 4 Campo Grande
749-016 Lisboa 749-016 Lisboa
Portugal Portugal
E-mail: [email protected] E-mail: [email protected]