and education use, including for instruction at the authors institution
and sharing with colleagues.
Other uses, including reproduction and distribution, or selling or
licensing copies, or posting to personal, institutional or third party
websites are prohibited.
In most cases authors are permitted to post their version of the
article (e.g. in Word or Tex form) to their personal website or
institutional repository. Authors requiring further information
regarding Elsevier’s archiving and manuscript policies are
encouraged to visit:
Asymptotic linearity and limit distributions, approximations
Jo~ao T. Mexia
a, Manuela M. Oliveira
b,aFCT Nova University of Lisbon, Department of Mathematics, Quinta da Torre, 2825 Monte da Caparica, Portugal
bUniversity of Evora, Department of Mathematics and CIMA – Center for Research on Mathematics and its Applications, Cole´gio Luı´s Anto´nio Verney, Rua Rom~ao
Ramalho 59, 7000-671 E´vora, Portugal
a r t i c l e i n f o
Article history:
Received 20 January 2009 Received in revised form 13 September 2009 Accepted 14 September 2009 Available online 19 September 2009 Keywords:
Asymptotic linearity Linear and quadratic forms Polynomials
Normal distributions Central limit theorems
a b s t r a c t
Linear and quadratic forms as well as other low degree polynomials play an important role in statistical inference. Asymptotic results and limit distributions are obtained for a class of statistics depending onmþX, with X any random vector and mnon-random vector with JmJ- þ 1. This class contain the polynomials inmþX. An application to the case of normal X is presented. This application includes a new central limit theorem which is connected with the increase of non-centrality for samples of fixed size. Moreover upper bounds for the suprema of the differences between exact and approximate distributions and their quantiles are obtained.
&2009 Elsevier B.V.. All rights reserved.
Contents
1. Introduction . . . 353
2. Notations and concepts . . . 354
3. Asymptotic results . . . 354
4. Application: the normal case . . . 356
5. Final remarks . . . 357
6. Uncited reference. . . 357
1. Introduction
Linear and quadratic forms as well as other low degree polynomials play an important role in statistics (e.g. Ito, 1980; Lazraq and Cleroux, 2001; Oliveira and Mexia, 2007a, 2007b; Sabatier and Traissac, 1994; Scheffe´, 1959). InAreia et al. (2008)it is pointed out that low degree polynomials in normal independent variables, with sufficiently small variation coefficients, are approximately normal. Moreover, also in Areia et al. (2008), are presented simulations suggesting that, when J
l
J- þ 1, a polynomial Pðl
þXÞ has a leading component linear in X. This linear component will be normal.We will obtain asymptotic approximations and limit distributions for a class of statistics containing Pð
l
þXÞ.In the next section we will present the concept of asymptotic linear functions which will play a central role in our study. We will also introduce the relevant notations and point out that polynomials are asymptotically linear functions. Next we
Contents lists available atScienceDirect
journal homepage: www.elsevier.com/locate/jspi
Journal of Statistical Planning and Inference
0378-3758/$ - see front matter & 2009 Elsevier B.V.. All rights reserved. doi:10.1016/j.jspi.2009.09.014
Corresponding author. Tel.: þ351 266 745300; fax: þ351 266 744546.
use a variant of thedmethod (e.g.Lehmann and Casella, 2001) to obtain asymptotic and limit results. In the last section we will consider the case when
m
þX will be normal with mean vectorm
. This variant will display a good control of the approximation errors both for distribution functions and their quantiles.If Vþ is the Moore–Penrose matrix inverse of the variance–covariance matrix V of
m
þX, the quadratic form ðm
þXÞtVþð
m
þXÞ will have a chi-square distribution with r ¼ rankðVÞ degrees of freedom and non-centrality parameterd¼
m
tVþm
, (e.g.Mexia, 1990). Thus we may usedto measure the non-centrality of the sample. In our study of the normalcase we will obtain a central limit theorem, for fixed size samples whose non-centrality parameter diverges to þ1. Moreoverdmay be used to measure sample non-centrality whenever V, the variance–covariance matrix of X is defined, andd! þ 1will imply J
l
J- þ 1. Since our asymptotic and limit results are derived assuming that Jl
J- þ 1, they are connected with the increase of non-centrality in samples, normal or not, with fixed size. When the components X1; . . . ; Xmof X are i.i.d. with mean value
m
and variances
2 we havel
¼m
1m and V ¼
s
2Im so Vþ¼ ð1=s
2ÞIm andd¼mð
m
2=s
2Þ ¼m=VC2, with VC ¼s
=
m
the variation coefficient. Thus when the X1; . . . ; Xm are observations with lowvariation coefficients we have a sample with larged. The variation coefficient with high precision observations are used.
2. Notations and concepts
Given a sufficiently regular function g : Rk-R, let g and g be the gradient and the Hessian matrix of g. Then, with r dðxÞ
the supremum of the spectral radius rðyÞ of g ðyÞ, when Jy xJrd, the function g is asymptotically linear, if whatever d40, we have kdðuÞu --10; with kdðuÞ ¼ sup rdðxÞ Jg ðxÞJ; JxJZu ( ) :
Since the first and second order partial derivatives of polynomials are themselves polynomials with degree decreasing with successive derivations, polynomials will be asymptotically linear.
Given a random vector X and an asymptotically linear function gðÞ, we will show that, when J
l
J-1, the distribution FYof
Y ¼ gð
l
þXÞapproaches the distribution of Z ¼ gð
l
Þ þg ðl
ÞtX:With lpthe p th quantile of the random variable L, given
Z3 ¼ðZ gð
l
ÞÞ Jg ðl
ÞJ and Y3 ¼ðY gðl
ÞÞ Jg ðl
ÞJwe will obtain upper bounds for jy3
pz3pj. Moreover, when 1 Jg ð
l
ÞJg ðl
Þ -JmJ-1 bwe will establish that FZ3converges to FZ3 3, with Z
3
3¼b
tX.
As we shall see these results are easy to apply when X is normal. Then Z3
3will also be normal and a central limit theorem
will apply.
3. Asymptotic results
If g : Rk-R is sufficiently regular we have, seeKhuri (2003)
Y ¼ gð
l
þXÞ ¼ gðl
Þ þg ðl
ÞtX þ Z 1 0 ð1 hÞðXtg ðl
þhXÞXÞ dh: Thus, when JXJrd, jY Zj ¼ Z 1 0 ð1 hÞðXtg ðl
þhXÞXÞ dh rrdðl
Þd2J.T. Mexia, M.M. Oliveira / Journal of Statistical Planning and Inference 140 (2010) 353–357 354
as well as jY3Z3jrk
dðJ
l
JÞd2: We now establishProposition 3.1. If gðÞ is asymptotically linear, then jY3Z3j -a:s JmJ-1
0, where a.s. indicates almost sure convergence.
Proof. If gðÞ is asymptotically linear, whatever
e
40 and d40, there is uðe
Þsuch that, for u4uðe
Þ, kdðuÞoe
. Thus, whenJ
l
JZuðe
Þ, JXJod implies jY3Z3jre
. Whateverd40, we may choose d40 to ensure that PrðJXJodÞZ1 d, thus when J
l
JZuðe
Þ, we will also have PrðjY3Z3jre
ÞZ1 dwhich establishes the thesis. &Since kdðuÞ decreases with u we may define
uðd;
e
Þ ¼Minfu; kdðuÞd2re
gthus, with lqthe p th quantile of the distribution of L ¼ JXJ2, whenever Lrlq, J
l
JZuqðe
Þ ¼uðlq;e
ÞimpliesjY3Z3jr
e
:
So, if qð
e
Þis the probability of the event Aðe
Þ ¼ fjY3Z3jre
gwe see that qðe
ÞZq, when J
l
J4uqðe
Þ.Let Acð
e
Þbe the complement of Aðe
Þand qcðe
Þ ¼1 qðe
Þ. We now haveProposition 3.2. Whatever z and
e
40, FZ3ðze
Þ qcðe
ÞrFY3ðzÞrFZ3ðz þe
Þ þqcðe
Þ.Proof. The thesis follows from FY3ðzÞ ¼ PrðY
3rzÞ ¼ qð
e
ÞPrðY3rzjAðe
ÞÞ þqcð
e
ÞPrðY3rzjAcðe
ÞÞrqðe
ÞPrðZ3rz þe
jAðe
ÞÞ þqcð
e
Þand from
FZ3ðz
e
Þ qcðe
Þrqðe
ÞPrðZ3rze
jAðe
ÞÞrqðe
ÞPrðY3rzjAðe
ÞÞrPrðY3rzÞ ¼ FY3ðzÞ: &Corollary 3.1. When FZ3ðzÞ has density fZ3ðzÞ bounded by C, we have SupfjFYFZjg ¼SupfjFY3FZ3jgrdð
e
Þ withdð
e
Þ ¼2ðCe
þqðe
ÞÞ. Proof. Since FYðxÞ ¼ FY3 x gðl
Þ Jg ðl
ÞJ ! and FZðxÞ ¼ FZ3 x gðl
Þ Jg ðl
ÞJ !we have SupfjFYFZjg ¼SupfjFY3FZ3g. To complete the proof we have only to apply Proposition 3.1 remembering that,
since fZ3 is bounded by C; FZ3ðz þ
e
Þ FZ3ðze
Þr2Ce
. &Corollary 3.2. When fZ3 is bounded by C and gðÞ is asymptotically linear, SupfjFYFZjg
-JmJ-þ1
0. This last corollary gives conditions for FZ to approach FY uniformly when J
l
J!1.Next we have
Proposition 3.3. If fZ3is bounded by C and if fZ3ðzÞZmðdÞ40, whenever z3
ddðeÞrzrz 3 1dþdðeÞ, we have Supfjy3 pz 3 pj;drpr1 dgr2
e
þ dðe
Þ mðdÞ :Proof. According to Corollary 3.1 of Proposition 3.2, we have FZ3ðy 3 p
e
Þ dðe
ÞrFY3ðy 3 pÞ ¼prFZ3ðy 3 pþe
Þ þdðe
Þso FZ3ðy3p
e
Þrp þ dðe
Þ and p dðe
ÞrFZ3ðy3pþe
Þ. Thus y3pe
rzpþdðeÞ and z3pdðeÞry 3
pþ
e
, so z3pdðeÞe
ry 3prz3pþdðeÞþ
e
. Tocomplete the proof we have only to point out that
z3 pþdðeÞz 3 pdðeÞr 2dð
e
Þ mðdÞ: &Corollary 3.3. If J
l
JZuqðe
Þ, fZ3 is bounded by C and fZ3ðzÞZmðdÞ40, then whenever z3ddðeÞrzrz 3 1dþdðeÞ, we have Supfjy3 pz 3 pj;drpr1 dgr2ð
e
þdqðe
ÞÞmðdÞÞ; with dqðe
Þ ¼2ðCe
þqcÞwhere qc¼1 q.Proof. The thesis follows from Proposition 3.3 since uqð
e
Þ4q, when Jl
J4uqðe
Þ. & We also have Proposition 3.4. If 1 Jg ðl
ÞJg ðl
ÞJm-J-1 b; FZ3 -JmJ-1 FZ3 3this convergence being uniform when fZ3 is bounded by C and gð:Þ is asymptotically linear.
Proof. The first part of the thesis follows from
1 Jg ð
l
ÞJg ðl
Þ b !t X -a:s: JmJ-1 0 whenever 1 Jg ðl
ÞJg ðl
ÞJm-J-1 b:The second part of the thesis follows from Proposition 3.3 since uqð
e
Þ4q, when Jl
J4uqðe
Þ. &4. Application: the normal case
Let X be normal with null mean vector and regular variance–covariance matrix M. Then
fZ3ðzÞ ¼ eJg ðlÞJ2ðzgðlÞÞ2=g ðlÞtMg ðlÞ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2
p
g ðl
ÞtMg ðl
Þ q Jg ðl
ÞJ is bounded by Cðl
Þ ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiJg ðl
ÞJ 2p
g ðl
ÞtMg ðl
Þ q :Ifykis the smallest eigenvalue of M we have
Cð
l
ÞrC ¼ ffiffiffiffiffiffiffiffiffiffiffi1 2p
ykp :
With xpthe quantile for probability p of the standardized normal density, the quantiles for fZ3 will be
z3 pð
l
Þ ¼g ðl
Þ þxp ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi g ðl
ÞtMg ðl
Þ q Jg ðl
ÞJ ; while the minimum of fZ3in ½z3ddðeÞð
l
Þ; z 3 1dþdðeÞðl
Þwill be mðdjl
Þ ¼ e x2 ddðeÞ=2 ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2p
g ðl
ÞtMg ðl
Þ q Jg ðl
ÞJ so that Cðl
Þ mðdjl
Þ¼e x2 ddðeÞ=2 and dðe
jl
Þ mðdjl
Þ¼ 2ðCðl
Þe
þqcðe
ÞÞ mðdjl
Þ ¼2e x2 ddðeÞ=2e
þ qcðe
ÞJgðl
ÞJ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2p
g ðl
ÞtMg ðl
Þ q 0 B @ 1 C Ar2ex2ddðeÞ=2e
þ q cðe
Þ ffiffiffiffiffiffiffiffiffiffiffi 2p
yk p ! :We thus obtain bounds that do not depend on
l
, both for fZ3ðzÞ and for dðe
jl
Þ=mðdjl
Þ. It is now easy to apply the previousresults to this case.
Moreover if X is normal with null mean vector and variance–covariance matrix M and if y1Z; . . . ; Zyh are the
eigenvalues of M with multiplicities g1; . . . ; gh,
L ¼ JXJ2
may be written (seeImhof, 1961) as a linear combinationPhj¼1yj
w
2gj, of independent chi-square variates. When M is knownit is easy to compute the quantiles lp (Fonseca et al., 2007).
J.T. Mexia, M.M. Oliveira / Journal of Statistical Planning and Inference 140 (2010) 353–357 356
5. Final remarks
In this work we derived asymptotic and limit results for samples, with fixed size, whose non-centrality diverges to þ1. In the case of normal samples a central limit theorem of a new type is established. Our approach is a variant of the d method. This variant will apply to asymptotically linear statistics gð
l
þXÞ such as the polynomials Pðl
þXÞ and leads to a good control of the approximation involved (see Corollary 3.1, Proposition 3.2 and Corollary 3.3).References
Areia, A., Oliveira, M.M., Mexia, J.T., 2008. Models for series of studies based on geometrical representation. Statistical Methodology 1, 277–288. Fonseca, M., Mexia, J.T., Zmys´lony, R., 2007. Jordan Algebras. Generating pivot variables and orthogonal normal models. Journal of Interdisciplinary
Mathematics 1, 305–326.
Imhof, J.P., 1961. Computing the distribution of quadratic forms in normal variables. Biometrics 48 (3–4), 419–426.
Ito, K., 1980. Robustness of anova and manocova test procedures. In: Krishnaiah, P.R. (Ed.), Handbook of Statistics, vol. I. North Holland, Amsterdam. Khuri, A.I., 2003. Advanced Calculus with Applications in Statistics, second ed. Wiley, New York.
Lazraq, A., Cleroux, R., 2001. The PLS multivariate regression model: testing the significance of successive PLS components. Journal of Chemometrics 15, 523–536.
Lehmann, E.L., Casella, G., 2001. Theory of Point Estimation. Springer, Berlin.
Mexia, J.T., 1990. Best linear unbiased estimates, duality of F tests and the Scheffe´ multiple comparison method in presence of controlled heteroscedasticity. Computational Statistics & Data Analysis 10, 3.
Oliveira, M.M., Mexia, T.J., 2007a. ANOVA like analysis of matched series of studies with a common structure. Journal of Statistical Planning and Inference 137, 1862–1870.
Oliveira, M.M., Mexia, J.T., 2007b. Modelling series of studies with a common structure. Computational Statistics and Data Analysis 51, 5876–5885. Sabatier, R., Traissac, P., 1994. The ACT (STATIS method). Computational Statistics & Data Analysis 18, 97–119.