This article was downloaded by: [University of Toronto Libraries]

On: 11 September 2011, At: 08:38

Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered
office: Mortimer House, 37-41 Mortimer Street, London W1T 3JH, UK

Journal of Statistical Computation and

Simulation
Publication details, including instructions for authors and
subscription information:
http://www.tandfonline.com/loi/gscs20

Optimal estimation for the Pareto

distribution
a
Mei Ling Huang
a
Department of Mathematics, Brock University, St. Catharines,
ON, Canada, L2S 3A1

Available online: 16 May 2011

To cite this article: Mei Ling Huang (2011): Optimal estimation for the Pareto distribution, Journal
of Statistical Computation and Simulation, DOI:10.1080/00949655.2010.516751

To link to this article: http://dx.doi.org/10.1080/00949655.2010.516751

PLEASE SCROLL DOWN FOR ARTICLE

Full terms and conditions of use: http://www.tandfonline.com/page/terms-and-

conditions

This article may be used for research, teaching and private study purposes. Any
substantial or systematic reproduction, re-distribution, re-selling, loan, sub-licensing,
systematic supply or distribution in any form to anyone is expressly forbidden.
The publisher does not give any warranty express or implied or make any representation
that the contents will be complete or accurate or up to date. The accuracy of any
instructions, formulae and drug doses should be independently verified with primary
sources. The publisher shall not be liable for any loss, actions, claims, proceedings,
demand or costs or damages whatsoever or howsoever caused arising directly or
indirectly in connection with or arising out of the use of this material.
 Journal of Statistical Computation and Simulation
iFirst, 2011, 1–18

Optimal estimation for the Pareto distribution

Mei Ling Huang*
Downloaded by [University of Toronto Libraries] at 08:38 11 September 2011

Department of Mathematics, Brock University, St. Catharines, ON, Canada L2S 3A1

(Received 20 May 2010; final version received 15 August 2010 )

This paper proposes an optimal estimation method for the shape parameter, probability density function
and upper tail probability of the Pareto distribution. The new method is based on a weighted empirical
distribution function. The exact efficiency functions of the estimators relative to the existing estimators
are derived. The paper gives L1 -optimal and L2 -optimal weights for the new weighted estimator. Monte
Carlo simulation results confirm the theoretical conclusions. Both theoretical and simulation results show
that the new estimation method is more efficient relative to several existing methods in many situations.

Keywords: efficiency; order statistics; weighted empirical distribution function; density estimation; Pareto
upper tail

AMS 1991 Subject Classifications: Primary: 62G30; Secondary: 62G05

1. Introduction

The Pareto distribution is a power-tailed distribution. It has many applications in economics,

actuarial science, survival analysis, queueing networks and other stochastic models [1,2]. It is
important to explore the estimation methods for the Pareto distribution.
Several types of Pareto distribution have been defined [3]. In this paper, we only discuss the
Pareto Type II with x0 = 1. Its probability density function (p.d.f.) and cumulative distribution
function (c.d.f.) are given by
α
f (x) = , x ≥ 0, α > 0, (1)
(1 + x)(α+1)
1
F (x) = 1 − , x ≥ 0, α > 0. (2)
(1 + x)α
The mean and variance of Equation (1) are
1
μ= , α > 1, (3)
α−1
α
σ2 = , α > 2. (4)
(α − 1)2 (α − 2)

*Email: mhuang@brocku.ca

ISSN 0094-9655 print/ISSN 1563-5163 online

© 2011 Taylor & Francis
DOI: 10.1080/00949655.2010.516751
http://www.informaworld.com
 2 M.L. Huang

When the shape parameter α is unknown, we want to estimate α from a random sample
X1 , X2 , . . . , Xn , n > 2, from the p.d.f. in Equation (1). There are several parametric methods
as follows.

1.1. Sample moment method for the shape parameter α

The sample mean estimator for the population mean μ in Equation (3) is

1
n
μ̂X̄ = X̄ = Xi . (5)
Downloaded by [University of Toronto Libraries] at 08:38 11 September 2011

n i=1

Note that Equation (5) is based on the empirical distribution function (EDF) Sn (x),


1
n
1, if x ∈ A,
Sn (x) = I(−∞,x] (Xi ), where IA = (6)
n i=1 0, if x ∈
/ A.

And Sn (x) uses equal weight 1/n for each sample point. Then we have the moment estimator for
the shape parameter α

1
α̂X̄ = + 1. (7)
X̄

1.2. Maximum-likelihood estimator (MLE) for the shape parameter α

Based on Equation (1), we obtain an MLE for α as

n
α̂MLE = n , (8)
i=1 log(Xi + 1)

which is a complete and sufficient estimator with the Vinci distribution (or inverse gamma
distribution) having the p.d.f.

(αn)n −(n+1) −(αn)/x

h(x) = x e , x > 0, and
(n)
αn α
E(α̂MLE ) = ; Bias(α̂MLE ) = ; (9)
n−1 n−1
α 2 n2 α 2 (n + 2)
Var(α̂MLE ) = ; MSE(α̂MLE ) = . (10)
(n − 1)2 (n − 2) (n − 1)2 (n − 2)

Therefore, a MLE for the mean of the Pareto distribution in Equation (3) is given by

1
μ̂MLE = , α̂MLE > 1. (11)
α̂MLE − 1
 Journal of Statistical Computation and Simulation 3

1.3. Minimum risk estimator for the shape parameter α

A minimum risk estimator (MRE) [3] is given by

n−2
α̂MRE = n , with
i=1 log(Xi + 1)
α(n − 2) α
E(α̂MRE ) = ; Bias(α̂MRE ) = − ; (12)
n−1 n−1
α 2 (n − 2) α2
Var(α̂MRE ) = ; MSE(α̂MRE ) = . (13)
(n − 1)2 n−1
Downloaded by [University of Toronto Libraries] at 08:38 11 September 2011

1.4. A proposed weighted estimator for the shape parameter α

Note that the above Moment, MLE and MRE methods use equal weight 1/n on each data point.
We have two questions:

(a) Why should we use equal weight 1/n on the data? Is the EDF in Equation (6), a minimum
variance unbiased estimator for F (x), good enough?
(b) How do we overcome the problem of the lack of efficiency of the tail estimation for the Pareto
distribution? (especially in the heavy tailed case: α ≤ 2, or near 2).

In recent years, there have been studies putting unequal weights on the data points, to obtain
weighted EDFs or processes [4, 5]. But there are some difficulties to determine what weights
should be used for the data points. Huang [6] studied a symmetric weighted empirical distribution
function (SWEDF) and its efficiency function relative to the EDF in Equation (6) for estimating
any population c.d.f. The SWEDF is defined by


n
n∗ (x) =
F I(−∞,x] (X(i) )pn,i , x ∈ R, n > 2, (14)
i=1

where pn,i ’s are symmetric general weights

⎧
⎪
⎨w, i = 2, . . . , n − 1, 0 < w <
1
,
pn,i ≡ n−2
⎪
⎩w1,n ≡ 1 (1 − (n − 2)w), i = 1, n,
2

and X(i) ≤ X(2) ≤ · · · ≤ X(n) are the order statistics of the random sample. Note that


n
0 < pn,i < 1, i = 1, . . . , n; pn,i = 1.
i=1

The parameter w in Equation (14) is the weight for the middle n − 2 data points, and w1,n is
the weight for the extreme data points. Huang [6] indicated that if w > 1/n, then the SWEDF in
Equation (14) has good efficiency of estimating any distribution function in the tails relative to
the EDF in Equation (6). Since the Pareto distribution is power-tailed, it is important to estimate
the upper tail probability. It is interesting to explore how different weights affect the estimation
of the mean, density function and upper tail probability. In this paper, we first obtain a weighted
 4 M.L. Huang

estimator μ̂w based on Fˆn∗ (x) in Equation (14) to estimate the mean μ in Equation (3), namely


n−1
1
μ̂w = wX(i) + (1 − (n − 2)w)[X(1) + X(n) ]. (15)
i=2
2

Then a weighted estimator for α is defined by

1
α̂w = + 1. (16)
μ̂w
This paper shows good efficiencies of the weighted estimator in Equation (14) for estimating the
Downloaded by [University of Toronto Libraries] at 08:38 11 September 2011

Pareto mean, density function and upper tail probability relative to existing estimators. Section 2
gives an exact efficiency function (EFF) of the weighted estimator μ̂w in Equation (15) relative to
the sample mean X̄ in Equation (5) and μ̂MLE in Equation (11) for estimating the population mean
μ in (3). The L1 -optimal and L2 -optimal weights for the new weighted estimators are given in
Section 3. A weighted semiparametric density estimator is introduced in Section 4. In Section 5,
the Monte Carlo simulation results show that the new weighted estimation method for estimating
the Pareto mean, density function and upper tail probability is more efficient relative to the existing
methods in many situations. The simulation results confirm the theoretical conclusions. Finally,
we give all the proofs in Appendix 1.

2. An exact EFF for μ̂w relative to X̄

In this section, we derive the EFF of μ̂w in Equation (15) relative to the sample mean X̄ in
Equation (5) for estimating the population mean μ in Equation (3).

Theorem 2.1 when α > 2, The mean square error (MSE) function of the μ̂w in Equation (15)
for estimating the population mean of the Pareto distribution (3) is given by

MSE(μ̂w ) = aw2 + bw + c, where 0 < w < 1, (17)

1
a = A + (n − 2)2 B − (n − 2)C,
4
1
b = − (n − 2)B + C − 2μD + μ(n − 2)E,
2
1
c = B − μE + μ2 .
4

n−1
(n + 1) (n − i + 1 − 2/α) (n − i + 1 − 1/α)
A= −2 +1
i=2
(n − i + 1) (n + 1 − 2/α) (n + 1 − 1/α)


n−2 
n−1
(n + 1)(n − j + 1 − 1/α)
+2 •
i=2 j =i+1
(n − j + 1)

(n − i + 1 − 2/α) 1
× −
(n + 1 − 2/α)(n − i + 1 − 1/α) (n + 1 − 1/α)
(n + 1) (n − i + 1 − 1/α)
− +1 .
(n + 1 − 1/α) (n − i + 1)
 Journal of Statistical Computation and Simulation 5

n(n − 2/α) + n!(1 − 2/α) n(n − 1/α) + n!(1 − 1/α)

B= −2
(n + 1 − 2/α) (n + 1 − 1/α)
1 1
+ 2n!(1 − 1/α) −
(n − 2/α)(n − 1/α) (n + 1 − 1/α)
n
−2 + 4;
(n − 1/α)

n−1
n!(n − i + 1 − 1/α) 1 1 n
C= − −
i=2
(n − i + 1)(n − 1/α) n − 2/α n − 1/α n − 1/α
(n − i + 1 − 2/α)
Downloaded by [University of Toronto Libraries] at 08:38 11 September 2011

1
+ n!(1 − 1/α) −
(n + 1 − 2/α)(n − i + 1 − 1/α) (n + 1 − 1/α)
(n + 1) (n − i + 1 − 1/α)
− +2 ;
(n + 1 − 1/α) (n − i + 1)

n−1
(n + 1) (n − i + 1 − 1/α)
D= −1 ;
i=2
(n − i + 1) (n + 1 − 1/α)
n(n − 1/α) + n!(1 − 1/α)
E= − 2,
(n + 1 − 1/α)
∞
where (x) = 0 x α−1 e−x dx, α > 0, is a Gamma function.

Corollary 2.1 The EFFs of μ̂w in Equation (15) for estimating μ in Equation (3) relative to
X̄ in Equation (5) and μ̂MLE in Equation (11) are given by

Var(X̄) α/[n(α − 1)2 (α − 2)]

EFF(μ̂w(X̄) ) = = , α > 2. (18)
MSE(μ̂w ) MSE(μ̂w )
MSE(μ̂MLE )
EFF(μ̂w(μ̂MLE ) ) =
MSE(μ̂w )
 
∞ 2
1 1 1
= − h(x) dx , (19)
MSE(μ̂w ) 0 x−1 α−1

where MSE(μ̂w ) is in Equation (17), μ̂MLE = 1/(α̂MLE − 1) is in Equation (11) and h(x) is in
Equation (9).

3. The optimal weights

3.1. The L1 -optimal weights

Huang and Brill [7] proposed a level-crossing weighted empirical distribution function (LCEDF)
to estimate the c.d.f. F (x), x ∈ R. This method leads to an L1 -optimal choice of the weights for
w in Equation (14),
1 1
wL1 -opt = √ > . (20)
n(n − 1) n
Huang [6] indicates that if more weight is given to the mid-data, then the efficiencies of the LCEDF
relative to the classical EDF exceeds 1 on the tails of x values. Using this idea, we have
 6 M.L. Huang

Definition 3.1 L1 -optimal estimators for the mean μ and the shape parameter α of the Pareto
distribution are defined by
μ̂wL1 -opt = μ̂w |w=wL1 -opt , where μ̂w is given in Equation (15); (21)
1
α̂wL1 -opt = + 1. (22)
μ̂wL1 -opt

3.2. The L2 -optimal weights

In this section, we use the MSE criteria to find an L2 -optimal weight w in Equation (14) for
Downloaded by [University of Toronto Libraries] at 08:38 11 September 2011

estimating the Pareto mean.

Corollary 3.1 The L2 -optimal weight w for minimizing MSE(μ̂w ) in Equation (17) when
α > 2, n > 4C/B + 2, is given by
b b2
wL2 -opt = − , and MSEmin (μ̂w ) = − + c, and (23)
2a 4a
μ̂wL2 -opt = μ̂w |w=wL2 -opt , where μ̂w is given in Equation (15); (24)
1
α̂wL2 -opt = + 1, (25)
μ̂wL2 -opt
where a, b, c, B and C are given in Equation (17).
√
In Table A1,Appendix 2, we list the values of wL1 -opt = 1/ n(n − 1), MSE(μ̂wL1 -opt ), MSE(X̄),
MSE(μ̂MLE ) and the exact EFF(μ̂w(X̄) ) relative to X̄, and EFF(μ̂w(MLE) ) relative to μ̂MLE , respec-
tively, for n = 10, 20, 30 and 50 here we choose α = 2.5, 3, 4 (since the variance of Equation (1)
is infinite when α ≤ 2), and use Equations (18)–(20), The efficiencies are greater than 1 in 21 out
of 24 (87.5%) cases.
Similarly, in Table A2, we list the values of wL2 -opt , MSEmin (μ̂wL2 -opt ), MSE(X̄), MSE(μ̂MLE )
and the exact EFFmax (μ̂w ) relative to X̄, and EFF(μ̂w(MLE) ) relative to μ̂MLE , respectively, for
n = 10, 20, 30 and 50; α = 2.5, 3, 4, by using Equation (23). The efficiencies are greater than 1
in 24 out of 24 (100%) cases.

Remark It is interesting to see that 100% efficiencies are greater than 1 in Tables A2 and A3
by using wL2 -opt . But the wL1 -opt in Equation (21) is totally nonparametric, it is more robust and
easy to use. The wL2 -opt in Equation (23) depends on α. In practice, we may estimate α first, then
obtain a wL2 -opt while still keeping the optimal advantage. Of course, we use the given values of
α in simulations.

4. Optimal weighted density estimators

A semiparametric weighted density estimator is given in this section. We estimate α first and then
substitute it into Equation (1).

Definition 4.1 A semiparametric weighted density estimator is defined by

 
αw
fw (x) = , x ≥ 0, n ≥ 2, (26)
(1 + x)(αw +1)
where 
αw is defined in Equation (16).
 Journal of Statistical Computation and Simulation 7

We substitute 
αwL1 -opt , 
αwL2 -opt in Equations (22) and (25) replacing 
αw in Equation (26) to
obtain the semiparametric L1 - and L2 -optimal weighted density estimators,


αwL1 -opt

fw L1 -opt (x) = , x ≥ 0, (27)
(
α +1)
(1 + x) wL1 -opt

αwL2 -opt

fw L2 -opt (x) = , x ≥ 0. (28)
(
αwL +1)
(1 + x) -
2 opt

αX̄ and 
We also define two density estimators by using  αMLE in Equations (7) and (8) as
Downloaded by [University of Toronto Libraries] at 08:38 11 September 2011

 
αX̄ 
αMLE
fX̄ (x) = α +1)
, x ≥ 0, and f
MLE (x) = , x ≥ 0. (29)
(1 + x) (
X̄ (1 + x)(αMLE +1)

5. Simulations

5.1. Simulation efficiencies of the optimal weighted mean

√
The simulation results are given in Tables A3 and A4 use the weight wL1 -opt = 1/ n(n − 1) in
Equation (20) to compute μ̂wL1 -opt in Equation (21), and use the weight wL2 -opt in Equation (23)
to compute μ̂wL2 -opt in Equation (24). We compute the simulation efficiencies of these estimators
relative to the estimators μ̂X̄ in Equation (5) and μ̂MLE in Equation (11) at selected values α =
2.5, 3 and 4, and generate m = 1, 000, 000 random samples for each case with sample sizes of
10, 20, 30 and 50, respectively.
The simulation MSEs are defined by

1 
m
SMSE(μ̂w ) = (μ̂w(i) − μ)2 ,
m i=1

1 
m
SMSE(μ̂X̄ ) = (μ̂ − μ)2 ,
m i=1 X̄(i)

1 
m
SMSE(μ̂MLE ) = (μ̂MLE(i) − μ)2 ,
m i=1

where μ̂w(i) , μ̂X̄(i) and μ̂MLE(i) are estimates from the ith sample, i = 1, 2, . . . , m.
The simulation efficiencies of μ̂w relative to μ̂X̄ and μ̂MLE are given by

SMSE(μ̂X̄ ) SMSE(μ̂MLE )
SEFF(μ̂w(X̄) (x)) = , SEFF(μ̂w(MLE) (x)) = . (30)
SMSE(μ̂w ) SMSE(μ̂w )

The one-million times simulation efficiencies in Tables A3 and A4 are almost perfectly consistent
with the exact efficiencies in Tables A1 and A2. The simulation results confirm the theoretical
results in Theorem 2.1, Corollary 2.1. In each table, simulation efficiencies show that the optimal
estimator μ̂w is more efficient relative to the classical estimators μ̂X̄ and μ̂MLE in 22 out of 24
(91.7%) cases, when using the L1 -optimal weight, and in 24 out of 24 (100%) cases when using
the L2 -optimal weight.
 8 M.L. Huang

5.2. Simulation efficiencies of the optimal weighted density estimator

√
Table A5 uses the weight wL1 -opt = 1/ n(n − 1) in Equation (20) to compute f  w L1 -opt (x) in

Equation (27). Table A6 uses the weight wL2 -opt in Equation (25) to compute fw L2 -opt (x) in
Equation (28). The efficiencies of these estimators relative to f 
X̄ (x) and fMLE (x) in Equation (29)
are at selected values x = 1, 2, 3, 4, 5, 6, 8, 10, 12 for α = 2.5, 3, and m = 1, 000, 000 generated
random samples with sample sizes 10, 20, 30 and 50. There is a total of 216 cases for each table.
The simulation MSEs are defined by

1 
m

SMSE(fw (x)) = (fw (i) (x) − f (x))2 ;
Downloaded by [University of Toronto Libraries] at 08:38 11 September 2011

m i=1

1  
m

SMSE(fX̄ (x)) = (f (x) − f (x))2 ;
m i=1 X̄ (i)

1  
m
SMSE(f
MLE (x)) = (fMLE (i) (x) − f (x))2 ,
m i=1

where f (x) is given in Equation (1); f   

w (i) , fX̄ (i) and fMLE (i) are estimates from the ith sample,
i = 1, 2, . . . , m.
The simulation efficiencies in Tables A5 and A6 show that the weighted estimator f w (x) is more

efficient relative to the classical estimators fX̄ (x), f MLE (x) in 120 out of 144 (83.3%) cases, when
using the L1 -optimal weights, and in 113 out of 144 (78.5%) cases when using the L2 -optimal
weights. Note that the higher efficiencies occurred on the right tail of the distribution.

5.3. Simulation efficiencies for the Pareto upper tail

In many situations, we are interested in estimating the Pareto upper tail probability

1
1 − F (x) = P {X > x} = , x ≥ 0, α > 0. (31)
(1 + x)α
We consider the following estimation methods:

1
1
− Fw (x) = 
, where αw is given in Equation (16),
(1 + x)αw
1
1− FX̄ (x) = 
, where α X̄ is given in Equation (7),
(1 + x)αX̄
 1
1− FMLE (x) = , where α MLE is given in Equation (8),
(1 + x)αMLE

 1
1− FMRE (x) = , where α MRE is given in Equation (12).
(1 + x)αMRE

We use the wL1 -opt in Equation (20) to compute 1 − Fw (x) in Table A7, and use the wL2 -opt
in Equation (23) to compute 1 − Fw (x) in Table A8. The efficiencies of 1 − Fw (x) relative to
1 
− FX̄ (x), 1 − FMLE (x) and 1 −  FMRE (x) are at selected values x = 1, 2, 3, 4, 5, 6, 8, 10, 12 for
α = 2.5, 3, and m = 1000 generated random samples with sample sizes 10, 20, 30 and 50. There
are 216 cases for each table. The simulation efficiencies in TablesA7 andA8 show that the weighted
 Journal of Statistical Computation and Simulation 9

estimator 1− Fw (x) is more efficient relative to the classical estimators 1 

− FX̄ (x), 1 − FMLE (x)

and 1 − FMRE (x) in 213 out of 216 (98.6%) cases, when using wL1 -opt , and in 202 out of 216
(93.5%) cases when using wL2 -opt . Note that the higher efficiencies occurred on the right tail of
the distribution.
The simulations were run by using MAPLE 13 and C++ with double precision.

6. Conclusions

In this paper, the theoretical and simulation results consistently show that the proposed weighted
Downloaded by [University of Toronto Libraries] at 08:38 11 September 2011

method has better efficiencies for estimating the Pareto mean, density function and upper tail
relative to the existing methods. The crucial point is that wL1 -opt and wL2 -opt are all greater than
1/n (intuitively, we can see this in Tables A1–A4). This means that the proposed weighted method
gives less weight on extreme data values. It is interesting to see how further studies on a variety
of weights affect the estimation of the Pareto distribution. Such studies open an alternative area
for inference of heavy tailed distributions.

Acknowledgements
Research of the author is supported by an NSERC Canada grant.

References

[1] L. Brown, N. Gans, A. Mandelbaum, A. Sakov, H. Shen, S. Zeltyn, and L. Zhao, Statistical analysis of a telephone
call center, J. Amer. Statist. Assoc. 100(469) (2005), pp. 36–50.
[2] D. Cooley, D. Nychka, and P. Naveau, Bayesian spatial modeling of extreme precipitation return levels, J. Amer.
Statist. Assoc. 102(479) (2007), pp. 824–840.
[3] C.K. Kleiber and S. Kotz, Statistical Size Distribution in Economics and Actuarial Sciences, John Wiley & Sons,
New York, 2003.
[4] P. Barbe and P. Bertail, The Weighted Bootstrap, Springer, New York, 1995.
[5] G.R. Shorack and J.A. Wellner, Empirical Processes with Applications to Statistics, John Wiley & Sons, New York,
1986.
[6] M.L. Huang, The Efficiencies of a Weighted Distribution Function Estimator, The Proceeding of American Statistical
Association, Nonparametric Statistics Section, 2003, pp. 1502–1506.
[7] M.L. Huang and P.H. Brill, A distribution estimation method based on level crossings, J. Statist. Plann. Inference
124(1) (2004), pp. 45–62.

Appendix 1

Lemma A.1 For a Pareto random variable X with density given in Equation (1), for i ∈ {1, 2, . . . , n}, 1 ≤ i ≤ n, α > 2,
(n + 1) (n − i + 1 − 1/α)
E[X(i) ] = • − 1;
(n − i + 1) (n + 1 − 1/α)
(n + 1) (n − i + 1 − 2/α) (n − i + 1 − 1/α)
2
E[X(i) ]= −2 + 1.
(n − i + 1) (n + 1 − 2/α) (n + 1 − 1/α)
For i, j ∈ {1, 2, . . . , n}, 1 ≤ i < j ≤ n, when α > max(1/(n − j + 1), 2/(n − i + 1)),
(n + 1)(n − j + 1 − 1/α)
E[X(i) X(j ) ] =
(n − j + 1)
(n − i + 1 − 2/α) 1
• −
(n + 1 − 2/α)(n − i + 1 − 1/α) (n + 1 − 1/α)
(n + 1) (n − i + 1 − 1/α)
− + 1.
(n + 1 − 1/α) (n − i + 1)
 10 M.L. Huang

Proof Based on the theory of order statistics, the uth quantile of Equation (2) is

1
F −1 (u) = − 1, 0 < u < 1,
(1 − u)1/α

for i ∈ I + , 1 ≤ i ≤ n, α > 2/(n − i + 1); we have

 ∞
n−1
E[X(i) ] = n x[F (x)]i−1 f (x)[1 − F (x)]n−i dx
i−1 0
 1
n−1
Downloaded by [University of Toronto Libraries] at 08:38 11 September 2011

1
=n − 1 ui−1 (1 − u)n−i du
i−1 0 (1 − u)1/α
 1  1
n−1
=n ui−1 (1 − u)n−i−1/α du − ui−1 (1 − u)n−i du
i−1 0 0

(n + 1) 1
= B i, n − i + 1 − − B(i, n − i + 1)
(i)(n − i + 1) α
(n + 1) (n − i + 1 − 1/α)
= − 1.
(n − i + 1) (n + 1 − 1/α)
 ∞
n−1
2
E[X(i) ]=n x 2 [F (x)]i−1 f (x)[1 − F (x)]n−i dx
i−1 0
 1 2
n−1 1
=n − 1 ui−1 (1 − u)n−i du
i − 1 0 (1 − u)1/α

n−1 2 1
=n B i, n − i + 1 − − 2B i, n − i + 1 − + B(i, n − i + 1)
i−1 α α
(n + 1) (n − i + 1 − 2/α) (n − i + 1 − 1/α)
= −2 + 1.
(n − i + 1) (n + 1 − 2/α) (n + 1 − 1/α)

For i, j ∈ I + , 1 ≤ i < j ≤ n, α > max(1/(n − i + 1), 2/(n − j + 1)),

 ∞  ∞
n!
E[X(i) X(j ) ] = • xy[F (x)]i−1 f (x)
(i − 1)!(j − i − 1)!(n − j )! 0 0

× [F (y) − F (x)]j −i−1 f (y)[1 − F (y)]n−j dx dy

where the integral part becomes, for 0 ≤ u < v ≤ 1,

 1  v
[F −1 (u)][F −1 (v)]ui−1 (v − u)j −i−1 (1 − v)n−j du dv
0 0
 1  v 1 1
= −1 − 1 ui−1 (v − u)j −i−1 (1 − v)n−j du dv
0 0 (1 − u)1/α (1 − v)1/α
 1  v 1 1
= ui−1 (v − u)j −i−1 (1 − v)n−j du dv
0 0 (1 − u)1/α (1 − v)1/α
 1  v 1
− ui−1 (v − u)j −i−1 (1 − v)n−j du dv
0 0 (1 − u)1/α
 1  v 1
− ui−1 (v − u)j −i−1 (1 − v)n−j du dv
0 0 (1 − v)1/α
 1  v
+ ui−1 (v − u)j −i−1 (1 − v)n−j du dv.
0 0
 Journal of Statistical Computation and Simulation 11

Let 1 − u = x, 1 − v = y. Since 0 ≤ u < v ≤ 1, then 0 ≤ y < x ≤ 1, such that 0 < y/x < 1. Then the last integral
becomes
 1  x
x −1/α y −1/α (1 − x)i−1 (x − y)j −i−1 y n−j dy dx
0 0
 1  x
− x −1/α (1 − x)i−1 (x − y)j −i−1 y n−j dy dx
0 0
 1  x
− y −1/α (1 − x)i−1 (x − y)j −i−1 y n−j dy dx
0 0
 1  x
+ (1 − x)i−1 (x − y)j −i−1 y n−j dy dx.
Downloaded by [University of Toronto Libraries] at 08:38 11 September 2011

0 0

In this expression, the first integral is

  
1 1 y j −i−1  y n−j −1/α  y 
x −1/α (1 − x)i−1 x (j −i−1)+(n−j −1/α)+1 1 − d dx
0 0 x x x
2 1
=B n−i+1− ,i B n − j + 1 − ,j − i .
α α

The second integral is

  
1 1 y j −i−1  y n−j  y 
x −1/α (1 − x)i−1 x (j −i−1)+(n−j )+1 1 − d dx
0 0 x x x
1
=B n−i+1− , i B(n − j + 1, j − i).
α

The third integral is

  
1 1 y j −i−1  y n−j −1/α  y 
(1 − x)i−1 x (j −i−1)+(n−j −1/α)+1 1 − d dx
0 0 x x x
1 1
=B n−i+1− ,i B n − j + 1 − ,j − i .
α α

The fourth integral is

  
1 1 y j −i−1  y n−j  y 
(1 − x)i−1 x (j −i−1)+(n−j )+1 1 − d dx
0 0 x x x
= B(n − i + 1, i)B(n − j + 1, j − i).

Now, we put them together,

n! 2 1
E[X(i) X(j ) ] = • B n − i + 1 − ,i B n − j + 1 − ,j − i
(i − 1)!(j − i − 1)!(n − j )! α α
1 1 1
−B n−i+1− , i B(n − j + 1, j − i) − B n − i + 1 − , i B n − j + 1 − , j − i
α α α

+ B(n − i + 1, i)B(n − j + 1, j − i)

(n + 1)(n − j + 1 − 1/α) (n − i + 1 − 2/α) 1

= −
(n − j + 1) (n + 1 − 2/α)(n − i + 1 − 1/α) (n + 1 − 1/α)
(n + 1) (n − i + 1 − 1/α)
− + 1.
(n + 1 − 1/α) (n − i + 1)


 12 M.L. Huang

Proof of Theorem 2.1

MSE(μ̂w ) = E[(μ̂w − μ)2 ] = E[μ̂2w ] − 2μE[μ̂w ] + μ2 .
Based on Equation (9), we have

n−1
1
E[μ̂w ] = wE[X(i) ] + (1 − (n − 2)w)(E[X(1) ] + E[X(n) ]);
2
i=2


n−1 
n−2 
n−1
E[μ̂2w ] = w 2 E[X(i)
2
]+2 w 2 E[X(i) X(j ) ]
i=2 i=2 j =i+1

1
+ (1 − (n − 2)w)2 (E[X(1)
2
] + E[X(n)
2
] + 2E[X(1) X(n) ])
Downloaded by [University of Toronto Libraries] at 08:38 11 September 2011

4

n−1
+ w(1 − (n − 2)w)(E[X(1) X(i) ] + E[X(i) X(n) ]).
i=2
Let

n−1 
n−2 
n−1
A= 2
E[X(i) ]+2 E[X(i) X(j ) ];
i=2 i=2 j =i+1

B = E[X(1)
2
] + E[X(n)
2
] + 2E[X(1) X(n) ];


n−1
C= (E[X(1) X(i) ] + E[X(i) X(n) ]);
i=2


n−1
D= E[X(i) ];
i=2

E = E[X(1) ] + E[X(n) ],
then
1
MSE(μ̂w ) = Aw 2 + (1 − (n − 2)w)2 B + w(1 − (n − 2)w)C
4
1
− 2μ Dw + (1 − (n − 2)w)E + μ2
2
1
= A+ (n − 2)2 B − (n − 2)C w2
4
1 1
+ − (n − 2)B + C − 2μD + μ(n − 2)E w + B − μE + μ2 .
2 4
By Lemma A.1, for α > 2,

n−1 
n−2 
n−1
A= 2
E[X(i) ]+2 E[X(i) X(j ) ]
i=2 i=2 j =i+1


n−1
(n + 1) (n − i + 1 − 2/α) (n − i + 1 − 1/α) 
n−2 
n−1
= −2 +1 +2
(n − i + 1) (n + 1 − 2/α) (n + 1 − 1/α)
i=2 i=2 j =i+1

(n + 1)(n − j + 1 − 1/α) (n − i + 1 − 2/α) 1

• −
(n − j + 1) (n + 1 − 2/α)(n − i + 1 − 1/α) (n + 1 − 1/α)
(n + 1) (n − i + 1 − 1/α)
− +1 .
(n + 1 − 1/α) (n − i + 1)
B = E[X(1)
2
] + E[X(n)
2
] + 2E[X(1) X(n) ]
n(n − 2/α) + n!(1 − 2/α) n(n − 1/α) + n!(1 − 1/α)
= −2
(n + 1 − 2/α) (n + 1 − 1/α)
(n + 1)(1 − 1/α) (n − 2/α) 1
+2 −
(1) (n + 1 − 2/α)(n − 1/α) (n + 1 − 1/α)
 Journal of Statistical Computation and Simulation 13

(n + 1) (n − 1/α)

−2 +4
(n + 1 − 1/α) (n)
n(n − 2/α) + n!(1 − 2/α) n(n − 1/α) + n!(1 − 1/α)
= −2
(n + 1 − 2/α) (n + 1 − 1/α)
1 1 n
+ 2n!(1 − 1/α) − −2 + 4.
(n − 2/α)(n − 1/α) (n + 1 − 1/α) (n − 1/α)

n−1
C= (E[X(1) X(i) ] + E[X(i) X(n) ])
i=2


n−1
(n + 1)(n − i + 1 − 1/α) (n − 2/α) 1
= −
(n − i + 1) (n + 1 − 2/α)(n − 1/α) (n + 1 − 1/α)
Downloaded by [University of Toronto Libraries] at 08:38 11 September 2011

i=2

(n + 1) (n − 1/α)

− +1
(n + 1 − 1/α) (n)
(n + 1)(1 − 1/α) (n − i + 1 − 2/α) 1
+ −
(1) (n + 1 − 2/α)(n − i + 1 − 1/α) (n + 1 − 1/α)
(n + 1) (n − i + 1 − 1/α)
− +1
(n + 1 − 1/α) (n − i + 1)

n−1
n!(n − i + 1 − 1/α) 1 1 n
= − −
(n − i + 1)(n − 1/α) n − 2/α n − 1/α n − 1/α
i=2

(n − i + 1 − 2/α) 1
+ n!(1 − 1/α) −
(n + 1 − 2/α)(n − i + 1 − 1/α) (n + 1 − 1/α)
(n + 1) (n − i + 1 − 1/α)
− +2
(n + 1 − 1/α) (n − i + 1)

n−1 
n−1
(n + 1) (n − i + 1 − 1/α)
D= E[X(i) ] = −1 .
(n − i + 1) (n + 1 − 1/α)
i=2 i=2

n(n − 1/α) + n!(1 − 1/α)

E = E[X(1) ] + E[X(n) ] = − 2. 
(n + 1 − 1/α)

Lemma A.2 MSE(μ̂w ) in Equation (17) is a convex function with a > 0, when n > 4C/B + 2, C and B are given in
Equation (17).

Proof Let f (w) = MSE(μ̂w ) = aw 2 + bw + c, we have

1
a =A+ (n − 2)2 B − (n − 2)C, where
4
⎡ 2 ⎤

n−1
A=E⎣ X(i) ⎦ ≥ 0,
i=2
 2 
B = E X(1) + X(n) ≥ 0,
n−1 
  
C=E X(i) X(1) + X(n) ≥ 0.
i=2

Since n > 2, hence

1 4C
(n − 2)B − C > 0, such that a > 0, when n> + 2. 
4 B

Proof of Corollary 3.1 By Theorem 2.1 and Lemma A.2, let

f (w) = MSE(μ̂w ) = aw 2 + bw + c,
 14 M.L. Huang

which is a quadratic function with first and second derivatives of f (w) w.r.t. w

f (w) = 2aw + b, f (w) = 2a > 0.

Hence, f (w) is a convex function with the minimum value

2
b b b2
fmin = MSEmin (μ̂w ) = a − +b − +c =− + c,
2a 2a 4a
at
b
w=− . 
2a
Downloaded by [University of Toronto Libraries] at 08:38 11 September 2011

Appendix 2
√
Table A1. 
Exact efficiencies of μ w relative to X̄, μ̂MLE , wL1 -opt = 1/ n(n − 1)
(sample size n = 10, 20, 30, 50; α = 2.5, 3, 4).

Sample size
n α wL1 -opt w =1/n 
MSE(μ w) MSE(X̄) MSE(μ̂MLE ) EFF(μ̂w(X̄) ) EFF(μ̂w(MLE) )

10 2.5 0.105409 0.1 0.1627 0.2222 2.8511 1.3662 17.5288

3 0.0593 0.0750 0.2121 1.2640 1.8506
4 0.0189 0.0222 0.0262 1.1772 1.3881
20 2.5 0.051299 0.05 0.0819 0.1111 0.1069 1.3561 1.3047
3 0.0302 0.0375 0.0361 1.2401 1.1915
4 0.0097 0.0111 0.0112 1.1505 1.1598
30 2.5 0.033903 0.0333 0.0554 0.0741 0.0534 1.3376 0.9648
3 0.0205 0.0250 0.0219 1.2185 1.0673
4 0.0066 0.0074 0.0072 1.1310 1.0908
50 2.5 0.020203 0.02 0.0339 0.0444 0.0285 1.3094 0.8408
3 0.0126 0.0150 0.0123 1.1896 0.9758
4 0.0040 0.0044 0.0042 1.1072 1.0325


Note: The L1 -optimal μ w has 21 out of 24 cases (87.5%, in bold) with exact efficiencies greater than one.

Table A2. 
Exact efficiencies of μ w relative to X̄, μ̂MLE , wL2 -opt = −b/2a
(sample size n = 10, 20, 30, 50; α = 2.5, 3, 4).

Sample size
n α wL2 -opt w =1/n 
MSE(μ w) MSE(X̄) MSE(μ̂MLE ) EFF(μ̂w(X̄) ) EFF(μ̂w(MLE) )

10 2.5 0.12234 0.1 0.0823 0.2222 2.8511 2.7012 34.6428

3 0.12070 0.0405 0.0750 0.2121 1.8506 5.2246
4 0.11866 0.0155 0.0222 0.0262 1.4361 1.6934
20 2.5 0.05475 0.05 0.0493 0.1111 0.1069 2.2542 2.1687
3 0.05430 0.0233 0.0375 0.0361 1.6065 1.5434
4 0.05378 0.0086 0.0111 0.0112 1.2985 1.3090
30 2.5 0.03532 0.0333 0.0361 0.0741 0.0534 2.0548 1.4821
3 0.03510 0.0167 0.0250 0.0219 1.49∗ 67 1.3110
4 0.03486 0.0060 0.0074 0.0072 1.2365 1.1926
50 2.5 0.02067 0.02 0.0239 0.0444 0.0285 1.8573 1.1927
3 0.02059 0.0108 0.0150 0.0123 1.3884 1.1388
4 0.02050 0.0038 0.0044 0.0042 1.1760 1.0967


Note: The L2 -optimal μ w has 24 out of 24 cases (100%, in bold) with exact efficiencies greater than one.
 Journal of Statistical Computation and Simulation 15

√
Table A3. 
Simulation efficiencies of μ w relative to X̄, μ̂MLE , wL1 -opt = 1/ n(n − 1)
Generated m = 1,000,000 times, sample size n = 10, 20, 50, 100; α = 2.5, 3, 4
SEFF(μ̂w(X̄) ) = SMSE(X̄)/SMSE(μ  w ); SEFF(μ̂w(MLE) ) = SMSE(μ̂MLE )/SMSE(μ
 w ).

Sample size
n α wL1 -opt w =1/n 
MSE(μ w) MSE(X̄) MSE(μ̂MLE ) EFF(μ̂w(X̄) ) EFF(μ̂w(MLE) )

10 2.5 0.105409 0.1 0.1570 0.2131 10.3762 1.3574 66.0935

3 0.0593 0.0750 0.1946 1.2644 3.2806
4 0.0189 0.0223 0.0261 1.1786 1.3798
20 2.5 0.051299 0.05 0.0798 0.1074 0.1063 1.3469 1.3330
3 0.0303 0.0377 0.0361 1.2419 1.1887
Downloaded by [University of Toronto Libraries] at 08:38 11 September 2011

4 0.0097 0.0112 0.0112 1.1525 1.1573

30 2.5 0.033903 0.0333 0.0534 0.0706 0.0535 1.3233 1.0021
3 0.0205 0.0250 0.0219 1.2180 1.0710
4 0.0066 0.0074 0.0072 1.1323 1.0904
50 2.5 0.020203 0.02 0.0333 0.0432 0.0286 1.2996 0.8589
3 0.0126 0.0151 0.0123 1.1908 0.9738
4 0.0040 0.0044 0.0041 1.1088 1.0306


Note: The L1 -optimal μ w has 22 out of 24 cases (91.7%, in bold) with simulation efficiencies greater than one.

Table A4. 
Simulation efficiencies of μ w relative to X̄, 
μMLE , wL2 -opt = −b/2a
Generated m = 1,000,000 times, sample size n = 10, 20, 50, 100; α = 2.5, 3, 4
SEFF(μ̂w(X̄) ) = SMSE(X̄)/SMSE(μ  w ); SEFF(μ̂w(MLE) ) = SMSE(μ̂MLE )/SMSE(μ
 w ).

Sample size
n α wL2 -opt w =1/n 
MSE(μ w) MSE(X̄) MSE(μ̂MLE ) EFF(μ̂w(X̄) ) EFF(μ̂w(MLE) )

10 2.5 0.12234 0.1 0.0823 0.2134 10.3762 2.5475 126.307

3 0.12070 0.0405 0.0751 0.1946 1.8539 4.8027
4 0.11866 0.0155 0.0223 0.0261 1.4425 1.68687
20 2.5 0.05575 0.05 0.0506 0.1075 0.1063 2.1251 2.0445
3 0.05430 0.0233 0.0377 0.0361 1.6152 1.5446
4 0.05378 0.0086 0.0112 0.0112 1.3055 1.3090
30 2.5 0.03532 0.0333 0.0372 0.0707 0.0535 1.9027 1.4885
3 0.03510 0.0167 0.0250 0.0219 1.4970 1.3146
4 0.03486 0.0060 0.0074 0.0072 1.2471 1.1943
50 2.5 0.02067 0.02 0.0248 0.0432 0.0286 1.7435 1.1960
3 0.02059 0.0108 0.0150 0.0123 1.3921 1.1394
4 0.02050 0.0038 0.0045 0.0041 1.1800 1.0970


Note: The L2 -optimal μ w has 24 out of 24 cases (100%, in bold) with simulation efficiencies greater than one.
 16 M.L. Huang

Table A5. Simulation efficiencies of L1 -optimal weighted density estimator fˆw relative to the moment
√ and
MLE methods (generated m = 1,000,000 times, sample size n = 10, 20, 30, 50; wL1 -opt = 1/ n(n − 1);
SEFF(fˆw(X̄) (x)) = SMSE(fˆX̄ (x))/SMSE(fˆw (x)); SEFF(fˆw(MLE) (x)) = SMSE(fˆMLE (x))/SMSE(fˆw (x))).

x
Estimator n 1 2 3 4 5 6 8 10 12

(A) Pareto distribution, α = 2.5a

fˆX̄ (x) 10 0.9808 0.9808 1.0472 1.1027 1.1481 1.1854 1.2427 1.2845 1.3166
20 0.8954 0.9877 1.0557 1.1114 1.1580 1.1976 1.2615 1.3111 1.3512
30 0.9037 0.9975 1.0628 1.1159 1.1606 1.1991 1.2624 1.3127 1.3544
50 0.9173 1.0071 1.0662 1.1139 1.1543 1.1894 1.2485 1.2966 1.3375
fˆMLE (x)
Downloaded by [University of Toronto Libraries] at 08:38 11 September 2011

10 0.8130 1.0484 1.2804 1.4872 1.6661 1.8203 2.0720 2.2701 2.4334

20 0.7974 0.9875 1.1317 1.2480 1.3426 1.4203 1.5392 1.6247 1.6881
30 0.7917 0.9433 1.0417 1.1144 1.1695 1.2117 1.2700 1.3041 1.3241
50 0.7807 0.9006 0.9669 1.0124 1.0452 1.0692 1.0999 1.1159 1.1234
(B) Pareto distribution, α = 3b
fˆX̄ (x) 10 0.9317 1.0446 1.1340 1.2008 1.2513 1.2901 1.3497 1.3934 1.4283
20 0.9330 1.0503 1.1379 1.2067 1.2620 1.3078 1.3808 1.4377 1.4846
30 0.9441 1.0562 1.1380 1.2035 1.2579 1.3041 1.3800 1.4409 1.4927
50 0.9568 1.0553 1.1252 1.1814 1.2287 1.2693 1.3368 1.3918 1.4384
fˆMLE (x) 10 0.8959 1.1868 1.4363 1.6321 1.7864 1.9912 2.1066 2.2577 2.3832
20 0.8861 1.0916 1.2410 1.3520 1.4355 1.4999 1.5915 1.6527 1.6962
30 0.8787 1.0322 1.1314 1.1997 1.2475 1.2809 1.3203 1.3379 1.3435
50 0.8685 0.9828 1.0493 1.0928 1.1221 1.1417 1.1630 1.1701 1.1695

Note: There are 9 × 4 × 2 = 72 cases. Overall, the weighted method has 120 out of 144 (83.3%, in bold) cases with efficiencies > 1.
a
The L1 -optimal weighted method has 57 out of 72 (79.9%, in bold) cases with efficiencies >1.
b
The L1 -optimal weighted method has 63 out of 72 (87.5%, in bold) cases with efficiencies > 1.

Table A6. Simulation efficiencies of L2 -optimal weighted density estimator fˆw relative to the moment
and MLE methods (generated m = 1,000,000 times, sample size n = 10, 20, 30, 50; wL2 -opt = −b/2a;
SEFF(fˆw(X̄) (x)) = SMSE(fˆX̄ (x))/SMSE(fˆw (x)); SEFF(fˆw(MLE) (x)) = SMSE(fˆMLE (x))/SMSE(fˆw (x))).

x
Estimator n 1 2 3 4 5 6 8 10 12

(A) Pareto distribution, α = 2.5a

fˆX̄ (x) 10 0.5742 0.7804 1.0091 1.2597 1.5250 1.7989 2.3541 2.9024 3.4331
20 0.4843 0.6948 0.8978 1.1111 1.3354 1.5696 2.0619 2.5789 3.1136
30 0.6165 0.8261 1.0112 1.1943 1.3776 1.5615 1.9301 2.2978 2.6638
50 0.6213 0.8615 1.0222 1.1736 1.3206 1.4641 1.7438 2.0147 2.2782
fˆMLE (x) 10 0.5163 0.8342 1.2337 1.6991 2.2132 2.7621 3.9351 5.1300 6.3451
20 0.4313 0.6946 0.9625 1.2476 1.5482 1.8614 2.5159 3.1950 3.8992
30 0.5402 0.7812 0.9911 1.1926 1.3880 1.5779 1.9411 2.2827 2.6041
50 0.5635 0.7703 0.9270 1.0669 1.1957 1.3161 1.5362 1.7339 1.9135
(B) Pareto distribution, α = 3b
fˆX̄ (x) 10 0.6482 0.9593 1.3089 1.6775 2.0445 2.3978 3.0460 3.6176 4.1260
20 0.7084 1.0102 1.3183 1.6319 1.9338 2.2496 2.8359 3.3865 3.9049
30 0.7487 1.0255 1.2889 1.5501 1.8076 2.0605 2.5511 3.0214 3.4739
50 0.7932 1.0272 1.2307 1.4225 1.6055 1.7809 2.1114 2.4191 2.7077
fˆMLE (x) 10 0.6124 1.0786 1.6453 2.2641 2.8944 3.5119 4.6681 5.7092 6.6491
20 0.6728 1.0499 1.4379 1.8284 2.2110 2.5801 3.2688 3.8931 4.4618
30 0.6968 1.0021 1.2814 1.5452 1.7927 2.0239 2.4411 2.8055 3.1268
50 0.7200 0.9567 1.1477 1.3157 1.4662 1.6019 1.8369 2.0339 2.2017

Note: There are 9 × 4 × 2 = 72 cases. Overall, the weighted method has 113 out of 144 (78.5%, in bold) cases with efficiencies > 1.
a
The L2 -optimal weighted method has 51 out of 72 (70.8%, in bold) cases with efficiencies > 1.
b
The L2 -optimal weighted method has 62 out of 72 (86.1%, in bold) cases with efficiencies > 1.
 Journal of Statistical Computation and Simulation 17

Table A7. Simulation efficiencies of L1 -optimal weighted tail estimator 1

− Fw relative to the classical moment, MLE
and MRE methods √
generated m = 1, 000 times, sample size n = 10, 20, 30, 50; wL1 -opt = 1/ n(n − 1)
SEFF(1 − Fw (X̄) (x)) = SMSE(1 − FX̄ (x))/SMSE(1 − Fw (x));

SEFF(1 − Fw (MLE) (x)) = SMSE(1 − FMLE (x))/SMSE(1
 − Fw (x));
SEFF(1 − Fw (MRE) (x)) = SMSE(1 −  FMRE (x))/SMSE(1 − Fw (x)).

x
Estimator n 1 2 3 4 5 6 8 10 12

(A) Pareto distribution, α = 2.5a

1− FX̄ (x) 10 1.0204 1.1169 1.1880 1.2404 1.2805 1.3121 1.3600 1.3952 1.4229
Downloaded by [University of Toronto Libraries] at 08:38 11 September 2011

20 1.0306 1.1250 1.1977 1.2547 1.3009 1.3393 1.4001 1.4472 1.4856

30 1.0409 1.1296 1.1990 1.2549 1.3012 1.3405 1.4045 1.4554 1.4975
50 1.0480 1.1263 1.1885 1.2396 1.2830 1.3205 1.3832 1.4343 1.4775
1− FMLE (x) 10 1.2054 1.5687 1.8623 2.0964 2.2878 2.4487 2.7107 2.9214 3.1000
20 1.0750 1.2700 1.4080 1.5122 1.5911 1.6521 1.7392 1.7971 1.8377
30 0.9983 1.1154 1.1916 1.2413 1.2736 1.2940 1.3132 1.3155 1.3088
50 0.9355 1.0009 1.0536 1.0804 1.0965 1.1054 1.1103 1.1056 1.0960
1− FMRE (x) 10 2.3149 4.2025 5.9132 7.4275 8.7869 10.029 12.268 14.285 16.158
20 1.5342 2.2355 2.7988 3.2528 3.6256 3.9387 4.4390 4.8281 5.1442
30 1.2875 1.6822 1.9730 2.1904 2.3561 2.4848 2.6670 2.7854 2.8643
50 1.0946 1.3081 1.4559 1.5618 1.6398 1.6983 1.7769 1.8234 1.8510
(B) Pareto distribution, α = 3b
1− FX̄ (x) 10 1.0701 1.1945 1.2768 1.3346 1.3782 1.4133 1.4688 1.5129 1.5498
20 1.0745 1.1950 1.2832 1.3509 1.4052 1.4507 1.5251 1.5854 1.6372
30 1.0795 1.1912 1.2767 1.3449 1.4020 1.4509 1.5328 1.6014 1.6611
50 1.0751 1.1688 1.2416 1.3007 1.3503 1.3932 1.4648 1.5240 1.5749
1− FMLE (x) 10 1.2677 1.6258 1.8792 2.0668 2.2150 2.3386 2.5411 2.7070 2.8499
20 1.1277 1.3231 1.4526 1.5409 1.6038 1.6502 1.7135 1.7541 1.7818
30 1.0512 1.1728 1.2446 1.2864 1.3102 1.3225 1.3276 1.3192 1.3045
50 0.9936 1.0728 1.1171 1.1415 1.1541 1.1592 1.1571 1.1465 1.1323
1− FMRE (x) 10 2.7718 4.9757 6.8237 8.4138 9.8417 11.163 13.615 15.904 18.091
20 1.7419 2.5622 3.1803 3.6585 4.0432 4.3640 4.8809 5.2915 5.6355
30 1.4353 1.9019 2.2282 2.4616 2.6341 2.7646 2.9459 3.0627 3.1412
50 1.2057 1.4611 1.6315 1.7493 1.8331 1.8943 1.9734 2.0186 2.0443

Note: There are 9 × 4 × 3 = 108 cases. Overall, the weighted method has 213 out of 216 (98.6%, in bold) cases with efficiencies > 1.
a
The L1 -optimal weighted method has 106 out of 108 (98.2%, in Bold) cases with efficiencies > 1.
b
The L1 -optimal weighted method has 107 out of 108 (99.1%, in bold) cases with efficiencies > 1.
 18 M.L. Huang

Table A8. Simulation Efficiencies of L2 -optimal weighted density estimator 1

− Fw relative to the classical moment,
MLE and MRE methods
Generated m = 1000 times, sample size n = 10, 20, 30, 50; wL2 −opt = −b/2a
SEFF(1 − Fw (X̄) (x)) = SMSE(1 − FX̄ (x))/SMSE(1 − Fw (x));

SEFF(1 − Fw (MLE) (x)) = SMSE(1 − FMLE (x))/SMSE(1
 − Fw (x));
SEFF(1 − Fw (MRE) (x)) = SMSE(1 −  FMRE (x))/SMSE(1 − Fw (x)).

x
Estimator n 1 2 3 4 5 6 8 10 12

(A) Pareto distribution, α = 2.5a

1
Downloaded by [University of Toronto Libraries] at 08:38 11 September 2011

− FX̄ (x) 10 0.8857 1.2987 1.7582 2.2373 2.7192 3.1941 4.1067 4.9587 5.7530
20 0.9019 1.2329 1.5825 1.9413 2.3037 2.6665 3.3845 4.0868 4.7711
30 0.9320 1.2279 1.5315 1.8397 2.1505 2.4619 3.0844 3.7027 4.3157
50 0.9594 1.2027 1.4398 1.6717 1.8989 2.1218 2.5551 2.9739 3.3796

1− FMLE (x) 10 1.0463 1.8241 2.7562 3.7812 4.8582 5.9608 8.1855 10.384 12.533
20 0.9399 1.3886 1.8603 2.3395 2.8175 3.2892 4.2039 5.0746 5.9019
30 0.8938 1.2123 1.5220 1.8197 2.1047 2.3765 2.8839 3.3468 3.7719
50 0.8565 1.0778 1.2764 1.4571 1.6229 1.7761 2.0514 2.2925 2.5070

1− FMRE (x) 10 2.0095 4.8864 9.7514 13.397 18.659 24.413 37.043 50.776 65.330
20 1.3414 2.4500 3.6980 5.0325 6.4205 7.8414 10.730 13.634 16.520
30 1.1527 1.8285 2.5200 3.2112 3.8928 4.5634 5.8570 7.0865 8.2552
50 1.0021 1.3969 1.7638 2.1062 2.4270 2.7287 3.2827 3.7809 4.2339
(B) Pareto distribution, α = 3b
1
− FX̄ (x) 10 0.9186 1.3166 1.7466 2.1827 2.6102 3.0228 3.7954 4.4979 4.1410
20 0.9392 1.2549 1.5785 1.9014 2.2201 2.5319 3.1323 3.7013 4.2413
30 0.9728 1.2510 1.5281 1.8018 2.0708 2.3349 2.8476 3.3403 3.8158
50 0.9977 1.2242 1.4393 1.6447 1.8419 2.0316 2.3917 2.7299 3.0496

1− FMLE (x) 10 1.0851 1.8492 2.7381 3.6889 4.6636 5.6411 7.5649 9.4184 11.199
20 0.9797 1.4135 1.8557 2.2916 2.7153 3.1233 3.8906 4.5960 5.2465
30 0.9330 1.2352 1.5186 1.7823 2.0268 2.2540 2.6625 3.0195 3.8158
50 0.8906 1.0971 1.2760 1.4336 1.5742 1.7006 1.9198 2.1043 2.2622

1− FMRE (x) 10 2.0839 4.9538 8.6941 13.069 17.912 23.101 34.235 46.056 58.378
20 1.3982 2.4939 3.6887 4.9294 6.1877 7.4460 9.9307 12.347 14.687
30 1.2032 1.8630 2.5144 3.1450 3.7497 4.3282 5.4071 6.3931 7.2990
50 1.0421 1.4219 1.7632 2.0722 2.3541 2.6126 3.072 3.4706 3.8206

Note: There are 9 × 4 × 3 = 108 cases. Overall, the weighted method has 202 out of 216 (93.5%, in bold) cases with efficiencies >1.
a
The L2 -optimal weighted method has 101 out of 108 (93.5%, in bold) cases with efficiencies >1.
b
The L2 -optimal weighted method has 101 out of 108 (93.5%, in bold) cases with efficiencies >1.