asymptotic normality of recursive density estimates under some dependence assumptions
TRANSCRIPT
Abstract. Let fXn; n � 1g be a strictly stationary sequence of negativelyassociated random variables with the marginal probability density functionf ðxÞ, the recursive kernel estimate of f ðxÞ is defined by
fnðxÞ ¼1
n
Xn
j¼1h�1j Kðx� Xj
hjÞ;
where hn is a sequence of positive bandwidths tending to 0, as n!1, Kð�Þ isa univariate kernel function. In this note, we discuss the point asymptoticnormality for fnðxÞ.
Key words: Negatively associated random variables; Recursive kernel esti-mate; Asymptotic normality
AMS 1991 Subject Classification: 62G05
1 Introduction
In many stochastic models, the assumption that random variables are inde-pendent is not plausible. Increases in some random variables are often relatedto decreases in other random variables so an assumption of negative depen-dence is more appropriate than an assumption of independence. Lehmann(1966) investigated various conceptions of positive and negative dependencein the bivariate case. Strong definitions of bivariate positive and negativedependence were introduced by Esary and Proschan (1972). These were laterdeveloped by Alam and Saxena (1981), Block, Savits and Shaked (1982), theirdefinition is:
A finite family of random variables fXi; 1 � i � ng is said to be negativelyassociated (NA) if for every pair of disjoint subsets A and B of f1; 2; � � � ; ng,
Metrika (2004) 60: 155–166DOI 10.1007/s001840300302
Asymptotic normality of recursive density estimatesunder some dependence assumptions
Han-Ying Liang1 and Jong-Il Baek2
1 Department of Applied Mathematics,Tongji University, Shanghai 200092, P. R. China(E-mail: [email protected])
2 School of Mathematics & Informational Statistics and Institute of Basic Natural Science,Wonkwang University, Ik-San 570-749, South Korea (E-mail: [email protected])
Covðf1ðXi; i 2 AÞ; f2ðXj; j 2 BÞÞ � 0
whenever f1 and f2 are coordinatewise increasing and such that the covari-ance exists. An infinite family of random variables is NA if every finitesubfamily is NA.
This definition is carefully studied by Joag-Dev and Proschan (1983). Theycompared it with other concepts of negative dependence, and justified theclaim that NA possesses certain advantages over competing notions of neg-ative dependence. They also derived, as a by-product of their main results,that many well-known multivariate distributions are NA. Because of its wideapplications in multivariate statistical analysis and systems reliability, thenotion of NA has received considerable attention recently. For convergenceresults, we refer to Joag-Dev and Proschan (1983) for fundamental properties,Shao and Su (1999) for law of the iterated logarithm, Liang (2000) forcomplete convergence and Roussas (1994) for the central limit theorem ofrandom fields. Bozorgnia et al. (1993) also derives a wealth of resultsregarding limiting theorems for NA random variables. Asymptotic propertiesof estimates related to NA samples have also been studied extensively. Caiand Roussas (1998) studied uniformly strong consistency, convergence ratesand asymptotic distribution of Kaplan-Meier estimator of a distributedfunction with random censored failure times, Cai and Roussas (1999) gaveBerry-Esseen bounds for smooth estimates of a distribution function, Rous-sas (2001) investigated consistency of the kernel estimate of a probabilitydensity function and Amini and Bozorgnia (2000) dealt with the consistencyand complete convergence of sample quantiles.
Nonparametric estimation of a probability density is an interestingproblem in statistical inference and has an important role in communicationtheory and pattern recognition. The purpose of this paper is to investigaterecursive density estimators when the observations are NA random samples.
Throughout this paper, let fXn; n � 1g be a strictly stationary sequence ofnegatively associated random variables with the marginal probability densityfunction (p.d.f.) f ðxÞ, the recursive kernel estimate of f ðxÞ is defined by
fnðxÞ ¼1
n
Xn
j¼1h�1j Kðx� Xj
hjÞ;
which was introduced by Wolverton and Wagner (1969) and apparentlyindependently by Yamato (1971). Note that fnðxÞ can be computed recur-sively by
fnðxÞ ¼n� 1
nfn�1ðxÞ þ ðnhnÞ�1Kð
x� Xn
hnÞ:
This property is particularly useful in large sample size since fnðxÞ can beeasily updated with each additional observation. Here hn is a sequence ofpositive bandwidths tending to 0, as n!1, Kð�Þ is a univariate kernel.
In the independent case, fnðxÞ has been thoroughly examined in Wegmanand Davies (1979). In the dependent case, quadratic mean convergence andasymptotic normality of these recursive estimators have been obtained byMasry (1986) under various assumptions on the dependence of Xi. Strongpointwise consistency of fnðxÞ has been proved by Gyorfi (1981). Takahata(1980) and Masry and Gyorfi (1987) obtained sharp almost sure rates of fnðxÞ
156 H.-Y. Liang and J.-Il. Baek
to f ðxÞ for the class of asymptotically uncorrelated processes, the definition ofwhich can be found in Masry and Gyorfi (1987). Masry (1987) establishedsharp rates of almost sure convergence of fnðxÞ to f ðxÞ for vector-valuedstationary strong mixing processes under weak assumptions on the strongmixing condition, these rates were improved by Tran (1989). Tran (1990)studied the uniform convergence and asymptotic normality of fnðxÞ undersome dependent assumption defined in terms of joint densities.
In addition, a closely related estimator is the Rosenblatt-Parzen kernelestimate of f ðxÞ
f nðxÞ ¼1
nhn
Xn
j¼1Kðx� Xj
hnÞ;
which has been extensively studied by many authors, such as Wolverton andWagner (1969), Roussas (2000,2001), Bosq, Merlevede and Peligrad (1999)and Lu (2001).
In this note, we shall discuss the point asymptotic normality for fnðxÞ. Themethods of proof are closely related to those of Masry (1986), Tran (1990)and Roussas (2000). Here, unlike mixing cases, the negatively associatedrandom variable X1;X2; � � � ;Xn are subject to the transformationKðx� Xj=hjÞ; j ¼ 1; 2; � � � ; n, losing in this process the negatively associatedproperty, i.e. the kernel weights Kðx� Xj=hjÞ are not necessarily negativelyassociated.
In the sequel, let C denote a positive constant, CðxÞ and Cðx; yÞ denotepositive constants depending on x and x; y, respectively, whose values areunimportant and may vary at different place. The set cðf Þ, < and N denotethe continuity points of the function f , the real numbers and the naturalnumbers, respectively; suppðf Þ ¼ fx 2 <; f ðxÞ > 0g.
Now, we shall give some assumptions:
(A1) If f ðx; y; kÞ is the joint p.d.f. of the random variable Xj and Xjþk, thensupx;y jf ðx; y; kÞ � f ðxÞf ðyÞj � M0 for k � 1.
(A2) (i) The kernel function K satisfies
K 2 L1;
Z
<KðuÞdu ¼ 1; sup
x2<ð1þ jxjÞjKðxÞj <1:
(ii) The derivative ðd=duÞKðuÞ ¼ K 0ðuÞ exists for all u 2 < and isbounded jK 0ðuÞj � B for u 2 <.
(A3) 0 < hn # 0; hnPn
j¼1 h�1j =n! h ð0 < h <1Þ.(A4) Let 0 < p ¼ pn < n; 0 < q ¼ qn < n be integers tending to1 along with
n, and let hn > 0 be bandwidths and k ¼ kn ¼ ½ npþq� ! 1 so that
kðp þ qÞ=n! 1 and (i) pnknn ! 1, (ii) pnhn ! 0 and p2
n=nhn ! 0, (iii)1h3n
P1j¼qnjCovðX1;Xjþ1Þj ! 0:
Remark 1.1 (a) The assumption (A1) was used by many authors, (A2)(i) and(A3) are similar to that used by Masry (1986); in addition, (A2)(i)implies that K is bounded and K 2 L2 from K 2 L1.
(b) Since qnkn=n ¼ ðpn þ qnÞkn=n� pnkn=n; (A4)(i) implies qnkn=n! 0.Also,qn=pn ¼ ðqnkn=nÞ=ðpnkn=nÞ ! 0; so that qn < pn, eventually.
Asymptotic normality of recursive density estimates 157
(c) The first of assumptions (A4)(ii) and (b) imply that qnhn ! 0, which alsoimplies hn ! 0: The second of the assumptions in (A4)(ii) impliesnhn !1:
(d) Assumptions (A4)(i)(ii) are easily satisfied, if pn and qn are chosen asfollows: With hn ! 0, let pn � h�d1
n ; qn � h�d2n ð0 < d2 < d1 < 1Þ, where
xn � ynmeans that, as n!1; xn=yn tends to a constant. It is easily seenthat kn � nhd1
n , so that (A4)(i)(ii) are satisfied, provided nh1þ2d1n !1:
Also, for Assumption (A4)(iii), let jCovðX1;Xjþ1Þj ¼ Ckj ð0 < k < 1Þ, then
h�3n
X1
j¼qn
jCovðX1;Xjþ1Þj � h�3n kqn � h�3n = exp½ð� log kÞh�d2n � ! 0
as n!1. Thus, (A4) is satisfied under the condition that nh1þ2d1n !1:
Next, let jCovðX1;Xjþ1Þj ¼ Cj�a ða > 1Þ, then
h�3n
X1
j¼qn
jCovðX1;Xjþ1Þj � h�3þða�1Þd2n ! 0;
provided, a > 1þ 3=d2.Thus, (A4) is satisfied for a > 1þ 3=d2 andnh1þ2d1
n !1.(e) The assumptions in Assumptions (A4) can be seen as purely technical for
proving our main result. But (A4)(iii) can also be seen as a ‘‘weakdependence’’ assumption, which is a further restriction for NA variables.
Our main results are as follows:
Theorem 1.1 Assume that (A1)-(A4) hold true.
(1) For x 2suppðf Þ \ cðf Þ,lim
n!1nhnVarðfnðxÞÞ ¼ r2ðxÞ :¼ hf ðxÞ
Z
<K2ðuÞdu: ð1:1Þ
ZnðxÞ ¼ffiffiffiffiffiffiffinhn
p½fnðxÞ � EfnðxÞ�
D�!n!1Nð0; r2ðxÞÞ :¼ ZðxÞ: ð1:2Þ
(2) In addition, if x1; x2; � � � ; xd are distinct points of suppðf Þ \ cðf Þ, thenðZnðx1Þ; Znðx2Þ; � � � ; ZnðxdÞÞ
D�!n!1ðZðx1Þ; Zðx2Þ; � � � ; ZðxdÞÞ; ð1:3Þ
where Zðx1Þ; Zðx2Þ; � � � ; ZðxdÞ are independent.
By applying the Toeplitz lemma (see Hall and Heyde (1980), p. 31 or Masry(1986)) and the Taylor expansion, we can obtain that:
Lemma 1.1 Suppose that (A2)(i) holds.
(1) For x 2 cðf Þ, we have limn!1 EfnðxÞ ¼ f ðxÞ:(2) Assume that the second-order derivative f 00 of f exists and is continuous and
bounded, and that Kð�Þ satisfiesZ
<uKðuÞdu ¼ 0;
Z
<u2KðuÞdu <1;
158 H.-Y. Liang and J.-Il. Baek
fhn; n � 1g satisfyP1
n¼1 h2n ¼ 1: Then
limn!1ð1n
Xn
j¼1h2
j Þ�1½EfnðxÞ � f ðxÞ� ¼ 1
2f 00ðxÞ
Z
<u2KðuÞdu:
Corollary 1.1 Denote by Z�nðxÞ ¼ffiffiffiffiffiffiffinhnp
½fnðxÞ � f ðxÞ�. Suppose that all theassumptions of Theorem 1.1 are fulfilled, and that the conditions of Lemma 1.1hold and ðn�1hnÞ1=2
Pnj¼1 h2
j ! 0. If Z�n replaces Zn, then the results in Theorem1.1 are still true.
2 Proofs of Main Result
Lemma 2.1 (Masry (1986)) Assume that K satisfies (A2)(i). If g 2 L1, then forx 2 cðgÞ,
limh!0
h�1Z
<Kðx� u
hÞgðuÞdu ¼ gðxÞ
and
limh!0
h�1Z
<Kðx� u
hÞKðy � u
hÞgðuÞdu ¼ gðxÞ
R< K2ðuÞdu; x ¼ y;0; x 6¼ y:
�
Lemma 2.2 (Cai and Roussas (1999) Let A and B be disjoint subsets of N,andlet fXj; j 2 A [ Bg be NA. Assume that f : <#A ! < and g : <#B ! < arepartially differentiable with bounded partial derivatives, denote byk@f =@tik1stands for the supnorm. Let q : < ! < be a bounded differentiablefunction with bounded derivative. Then
jCovff ðXi; i 2 AÞ; gðXj; j 2 BÞgj �X
i2A
X
j2B
k @f@tik1 � k
@g@tjk1½�CovðXi;XjÞ�;
jCovfY
i2A
qðXiÞ;Y
j2A
qðXjÞgj � kqk#Aþ#B�21 kq0k21
X
i2A
X
j2B
jCovðXi;XjÞj:
Proof of Theorem 1.1 We only prove (1.3), the proof of (1.1) and (1.2) isanalogous. By the Cramer-Wold method, we only have to verify that for anynon-zero real number a1; a2; � � � ; ad ,
Xd
i¼1aiZnðxiÞ �!
D Xd
i¼1aiZðxiÞ: ð2:4Þ
Without loss of generality, we shall prove (2.4) for d ¼ 2, i.e.
aZnðxÞ þ bZnðyÞ �!D
aZðxÞ þ bZðyÞ; ð2:5Þwhere a ¼ a1; b ¼ a2; x ¼ x1; y ¼ x2:
Denote by Kðx�Xj
hjÞ ¼ Kðx�Xj
hjÞ � EKðx�Xj
hjÞ. Hence, (2.5) becomes into
Asymptotic normality of recursive density estimates 159
Xn
j¼1
ffiffiffiffiffihn
n
rh�1j ½aKðx� Xj
hjÞ þ bKðy � Xj
hjÞ� �!D Nð0; a2r2ðxÞ þ b2r2ðyÞÞ: ð2:6Þ
For m ¼ 1; 2; � � � ; k, split the set f1; 2; � � � ; ng into k(large) p-blocks, Im, andk(small) q-blocks, Jm, as follows:
Im ¼ fi : i ¼ ðm� 1Þðp þ qÞ þ 1; � � � ; ðm� 1Þðp þ qÞ þ pg;Jm ¼ fj : j ¼ ðm� 1Þðp þ qÞ þ p þ 1; � � � ;mðp þ qÞg;
the remaining points form the set fl : kðp þ qÞ þ 1 � l � ng (which may be ;).Put
nnj ¼ffiffiffiffiffihn
n
rh�1j ½aK ðx� Xj
hjÞ þ bKðy � Xj
hj�
and for m ¼ 1; 2; � � � ; k, set
ynm ¼X
i2Im
nni; y0nm ¼X
j2Jm
nnj; y00nk ¼Xn
l¼kðpþqÞþ1nnl:
Also set
Sn ¼Xn
j¼1nnj; Tn ¼
Xk
m¼1ynm; T 0n ¼
Xk
m¼1y0nm; T 00n ¼ y00nk :
So that
Sn ¼ Tn þ T 0n þ T 00n :
and convergence (2.6) will be established by showing that
Tn �!D
Nð0; a2r2ðxÞ þ b2r2ðyÞÞ: ð2:7Þand
EðT 0nÞ2 þ EðT 00n Þ
2 ! 0: ð2:8ÞTo prove (2.7) and (2.8), we shall divide them into the following four lemmas.
Lemma 2.3 Under the assumptions of Theorem 1.1, we have VarðT 0nÞ ! 0:
Proof. Note that
VarðT 0nÞ ¼Xk
m¼1Varðy0nmÞ þ 2
X
1�i<j�k
Covðy0ni; y0njÞ: ð2:9Þ
While
Varðy0nmÞ¼XmðpþqÞ
i¼ðm�1ÞðpþqÞþpþ1VarðnniÞþ2
X
ðm�1ÞðpþqÞþpþ1�i<j�mðpþqÞCovðnni;nnjÞ
:¼I1ðmÞþ I2ðmÞ:Note that h�1j VarðKðx�Xj
hjÞÞ ! f ðxÞ
R< K2ðuÞdu; j!1; which shows that for
all j � 1,
160 H.-Y. Liang and J.-Il. Baek
h�1j VarðKðx� Xj
hjÞÞ � CðxÞ: ð2:10Þ
By Lemma 2.1 , we get
jh�2j CovðKðx� Xj
hjÞ;Kðy � Xj
hjÞÞj
¼ jh�2j
ZKðx� u
hjÞKðy � u
hjÞf ðuÞdu
� h�2j
ZKðx� u
hjÞf ðuÞdu �
ZKðy � u
hjÞf ðuÞduj
� Ch�1j þ Cðx; yÞ: ð2:11Þ
Applying (2.10) and (2.11) and the monotonocity of hj, we get
I1ðmÞ ¼XmðpþqÞ
j¼ðm�1ÞðpþqÞþpþ1
hn
nh�2j ½a2VarðKðx� Xj
hjÞÞ þ b2VarðKðy � Xj
hjÞÞ
þ 2abCovðKðx� Xj
hjÞ;Kðy � Xj
hjÞÞ�
� a2qCðxÞn
þ b2qCðyÞn
þ Cqn� 2abþ 2abqhn
n� Cðx; yÞ: ð2:12Þ
Note that, for i < j, by using (A1), we have
supx;yjCovðKðx�Xi
hiÞ;Kðy�Xj
hjÞÞj
¼ hihj supx;yjZ
<2
KðsÞKðtÞ½f ðx� shi;y� thj;j� iÞÞ� f ðx� shiÞf ðy� thjÞdsdt�j
�M0hihj;
and hence
I2ðmÞ ¼2hn
n
X
ðm�1ÞðpþqÞþpþ1�i<j�mðpþqÞh�1i h�1j fa2CovðKðx� Xi
hiÞ;Kðx� Xj
hjÞÞ
þ b2CovðKðy � Xi
hiÞ;Kðy � Xj
hjÞÞ þ ab½CovðKðx� Xi
hiÞ;Kðy � Xj
hjÞÞ
þ CovðKðy � Xi
hiÞ;Kðx� Xj
hjÞÞ�g � 2M0ða2 þ b2 þ 2abÞ q
2hn
n:
ð2:13ÞTherefore, by (2.12), (2.13) and applying (A4) we get
Xk
m¼1Varðy0nmÞ �ða2CðxÞ þ b2CðyÞ þ 2abCÞ kq
nþ 2abCðx; yÞ � kq
nhn
þ 2M0ða2 þ b2 þ 2abÞ kq2hn
n! 0: ð2:14Þ
By Lemma 2.2 , (A2)(ii) and noticing i 6¼ j, we get
Asymptotic normality of recursive density estimates 161
jCovðy0ni;y0njÞj�
hn
n
XiðpþqÞ
s¼ði�1ÞðpþqÞþpþ1
XjðpþqÞ
t¼ðj�1ÞðpþqÞþpþ1
�h�1s h�1t fa2jCovðKðx�Xs
hsÞ;Kðx�Xt
htÞÞj
þ b2jCovðKðy�Xs
hsÞ;Kðy�Xt
htÞÞjþab½jCovðKðx�Xs
hsÞ;Kðy�Xt
htÞÞj
þjCovðKðy�Xs
hsÞ;Kðx�Xt
htÞÞj�g
� B2ða2þb2þ2abÞ 1
nh3n½�CovðX 0ni;X
0njÞ�; ð2:15Þ
where X 0ni ¼PiðpþqÞ
s¼ði�1ÞðpþqÞþpþ1 Xs; X 0nj ¼PjðpþqÞ
t¼ðj�1ÞðpþqÞþpþ1 Xt: While, by sta-tionarity
jCovðX 0n1;X 0nðlþ1ÞÞj
¼ jXq
r¼1ðq� r þ 1ÞCovðX1;XlðpþqÞþrÞ þ
Xq�1
r¼1ðq� rÞCovðXrþ1;XlðpþqÞþ1Þj
¼ jXq
r¼1ðq� r þ 1ÞCovðX1;XlðpþqÞþrÞ þ
Xq�1
r¼1ðq� rÞCovðX1;XlðpþqÞ�rþ1Þj
�XlðpþqÞþq
r¼lðpþqÞ�ðq�2ÞjCovðX1;XrÞj ¼ q
XlðpþqÞþðq�1Þ
r¼lðpþqÞ�ðq�1ÞjCovðX1;Xrþ1Þj: ð2:16Þ
Hence, by (2.15), (2.16), (A4) and using stationarity again, we get
2X
1�i<j�k
jCovðy0ni; y0njÞj
� �2B2ða2 þ b2 þ 2abÞ 1
nh3n
Xk�1
l¼1ðk � lÞCovðX 0n1;X 0nðlþ1ÞÞ
� Ckqnh3
n
Xk�1
l¼1
XlðpþqÞþðq�1Þ
r¼lðpþqÞ�ðq�1ÞjCovðX1;Xrþ1Þj
� Ckqn� 1h3
n
X1
r¼p
jCovðX1;Xrþ1Þj ! 0;
which, together with (2.14) and (2.9), implies that VarðT 0nÞ ! 0:
Lemma 2.4 Under the assumptions of Theorem 1.1, we have VarðT 00n Þ ! 0:
Proof. Note that
VarðT 00n Þ ¼Xn
l¼kðpþqÞþ1VarðnnlÞ þ 2
X
kðpþqÞþ1�i<j�n
Covðnni; nnjÞ :¼ J1 þ J2:
162 H.-Y. Liang and J.-Il. Baek
Similarly to the proof in (2.12) and noticing n� kðp þ qÞ � p þ q � 2p, wehave
J1 � ½a2CðxÞ þ b2CðyÞ þ 2abC þ 2abhnCðx; yÞ� � n� kðp þ qÞn
� ½a2CðxÞ þ b2CðyÞ þ 2abC þ 2abhnCðx; yÞ� � 2pn! 0:
Similarly to the proof in (2.13), we have J2 � 8M0ða2 þ b2 þ 2abÞ � pn � phn ! 0:
Lemma 2.5 Under the assumptions of Theorem 1.1, we haveVarðSnÞ ! a2r2ðxÞ þ b2r2ðyÞ; further VarðTnÞ ! a2r2ðxÞ þ b2r2ðyÞ:
Proof. As in the proof of Lemma 2.3, we get
VarðTnÞ ¼Xk
m¼1VarðynmÞ þ 2
X
1�i<j�k
Covðyni; ynjÞ
� ½a2CðxÞ þ b2CðyÞ þ 2abC� � kpnþ 2abCðx; yÞ � kp
n� hn
þ 2M0ða2 þ b2 þ 2abÞ � kp2hn
nþ C � kp
n� 1h3
n
X1
r¼q
jCovðX1;Xrþ1Þj
<1: ð2:17ÞObviously,
VarðSnÞ ¼VarðTnÞ þ VarðT 0nÞ þ VarðT 00n Þþ 2CovðTn; T 0nÞ þ 2CovðTn; T 00n Þ þ 2CovðT 0n; T 00n Þ: ð2:18Þ
Since ETn ¼ ET 0n ¼ ET 00n ¼ 0, by Lemmas 2.3 and 2.4, (2.17) and using theCauchy-Schwarz inequality, we get
jCovðTn; T 0nÞj ! 0; jCovðTn; T 00n Þj ! 0; jCovðT 0n; T 00n Þj ! 0: ð2:19ÞNote that
VarðTnÞ þ VarðT 0nÞ þ VarðT 00n Þ
¼Xk
m¼1VarðynmÞ þ 2
X
1�i<j�k
Covðyni; ynjÞ
þXk
m¼1Varðy0nmÞ þ 2
X
1�i<j�k
Covðy0ni; y0njÞ þ
Xn
l¼kðpþqÞþ1VarðnnlÞ
þ 2X
kðpþqÞþ1�i<j�n
Covðnni; nnjÞ ¼Xn
i¼1VarðnniÞ þ 2An; ð2:20Þ
where
An¼P
1�i<j�k Covðyni; ynjÞþP
1�i<j�k Covðy0ni; y0njÞ þ
PkðpþqÞþ1�i<j�n Covðnni; nnjÞ
þ 12
Pkm¼1 I2ðmÞ þ
Pkm¼1
Pðm�1ÞðpþqÞþ1�i<j�ðm�1ÞðpþqÞþp
Covðnni; nnjÞ and by the
proof in Lemmas 2.3 and 2.4, we have
Asymptotic normality of recursive density estimates 163
An ! 0: ð2:21ÞWhile
Xn
i¼1VarðnniÞ ¼
hn
n
Xn
i¼1h�2i ½a2VarðKðx� Xi
hiÞÞ þ b2VarðKðy � Xi
hiÞÞ
þ 2abCovðKðx� Xi
hiÞ;Kðy � Xi
hiÞÞ�
:¼ L1ðnÞ þ L2ðnÞ þ L3ðnÞ: ð2:22ÞBy Lemma 2.1 and using the Toeplitz lemma and (A3), we get
L1ðnÞ ! a2r2ðxÞ; L2ðnÞ ! b2r2ðyÞ; L3ðnÞ ! 0: ð2:23ÞTherefore, from (2.18)-(2.23) it follows that VarðSnÞ ! a2r2ðxÞ þ b2r2ðyÞ:Further, by Lemmas 2.3 and 2.4 and (2.18)-(2.19), we have
VarðTnÞ ! a2r2ðxÞ þ b2r2ðyÞ:
Lemma 2.6 Under the assumptions of Theorem 1.1, we have
jEeitPk
m¼1 ynm �Yk
m¼1Eeitynm j ! 0; 8t 2 <: ð2:24Þ
gnð�Þ ¼Xk
m¼1E½X 2
nmIðjXnmj � �Þ� ! 0; 8� > 0; ð2:25Þ
where Xnm ¼ ynm=sn; sn ¼Pk
m¼1 VarðynmÞ:
Proof. By Lemma 2.2 , using the monotonicity of hj and stationarity of Xj,we have
jEeitPk
m¼1 ynm �Yk
m¼1Eeitynm j
� jCovðeitPk�1
m¼1 ynm ; eitynk Þj þ jEeitPk�1
m¼1 ynm �Yk�1
m¼1Eeitynm j � � � �
� jCovðeitPk�1
m¼1 ynm ; eitynk Þj þ jCovðeitPk�2
m¼1 ynm ; eityn;k�1Þjþ � � � þ jCovðeityn2 ; eityn1Þj
� ½Bðjaj þ jbjÞ�2t2 � hn
n½X
j2I1
X
l2I2
h�2j h�2l jCovðXj;XlÞj
þX
j2ðI1þI2Þ
X
l2I3
h�2j h�2l jCovðXj;XlÞj
þ � � � þX
j2ðI1þI2þ���þIk�1Þ
X
l2Ik
h�2j h�2l jCovðXj;XlÞj�
� ½Btðjaj þ jbjÞ�2 � 1
nh3n½X
j2I1
X
l2I2
jCovðXj;XlÞj þX
j2ðI1þI2Þ
X
l2I3
jCovðXj;XlÞj
164 H.-Y. Liang and J.-Il. Baek
þ � � � þX
j2ðI1þI2þ���þIk�1Þ
X
l2Ik
jCovðXj;XlÞj�
� ½Btðjaj þ jbjÞ�2 � 1
nh3n½ðk � 1Þ
X
j2I1
X
l2I2
jCovðXj;XlÞj
þ ðk � 2ÞX
j2I1
X
l2I3
jCovðXj;XlÞj þ � � � þX
j2I1
X
l2Ik
jCovðXj;XlÞj� ð2:26Þ
Once again, by stationarity,
X
j2I1
X
l2Ik
jCovðXj;XlÞj ¼ pjCovðX1;Xðk�1ÞðpþqÞþ1Þj
þ ðp � 1ÞjCovðX1;Xðk�1ÞðpþqÞþ2Þj þ � � � þ jCovðX1;Xðk�1ÞðpþqÞþpÞj ð2:27Þ
Hence, from (2.26), (2.27) and (A4) it follows that
jEeitPk
m¼1 ynm �Yk
m¼1Eeitynm j � C
pkn� 1h3
n
X1
j¼p
jCovðX1;Xjþ1Þj ! 0 as n!1:
Thus, (2.24) is verified. Now, we prove (2.25)From (A2)(i) we know that jKðxÞj � C; x 2 <, and hence
jXnmj �1
snpffiffiffiffiffiffiffiffiffiffihn=n
p� h�1n ð2aC þ 2bCÞ � Cp
snffiffiffiffiffiffiffinhnp :
Therefore, by (A4)(ii),
gnð�Þ �Xk
m¼1
C2p2
s2n � nhnP ðjXnmj � �Þ �
C2
�2s2n� p2
nhn! 0:
Proof of Corollary 1.1. Note that Z�nðxÞ ¼ffiffiffiffiffiffiffinhnp
½fnðxÞ� EfnðxÞ�þffiffiffiffiffiffiffinhnp
½EfnðxÞ � f ðxÞ� and
ffiffiffiffiffiffiffinhn
p½EfnðxÞ � f ðxÞ� ¼
ffiffiffiffiffiffiffinhn
pð1n
Xn
j¼1h2
j Þð1
n
Xn
j¼1h2
j Þ�1½EfnðxÞ � f ðxÞ�
¼ffiffiffiffiffiffiffiffiffiffiffiffin�1hn
p Xn
j¼1h2
j � ð1
n
Xn
j¼1h2
j Þ�1½EfnðxÞ � f ðxÞ�:
Theorefore, Theorem 1.1, Lemma 1.1 andffiffiffiffiffiffiffiffiffiffiffiffin�1hn
p Pnj¼1 h2
j ! 0 imply theconclusion.
Acknowledgements. The authors thank the referee for carefully reading the manuscript and for
valuable suggestions which improved the presentation of this paper. This research was partially
supported by the National Natural Science Foundation of China (No. 10171079), No.R01-2000-
000-00010 and No. 2001-42-D0008 from Korea Research Foundation as well as Wonkwang
University Grant in 2003.
Asymptotic normality of recursive density estimates 165
References
[1] Alam K, Saxena KML (1981) Positive dependence in multivariate distributions. CommunStatist Theor Meth A10:1183–1196
[2] Amini M, Bozorgnia A (2000) Negatively dependenct bounded random variable probabilityinequalities and the strong law of large numbers. J Appl Math & Stochastic Anal 13(3):261–267
[3] Block HW, Savits TH, Sharked M (1982) Some concepts of negative dependence. AnnProbab 10:765–772
[4] Bosq D, Merlevede F and Peligrad M (1999) Asymptotic normality for density kernelestimators in discrete and continuous time. J Multivariate Anal 68:78–95
[5] Bozorgnia A, Patterson RF and Taylor RL (1993) Limit theorems for negatively dependentrandom variables. Unpublished manuscript
[6] Cai ZW, Roussas GG (1998) Kaplan-Meier estimator under association. J MultivariateAnal 76:318–348
[7] Cai ZW, Roussas GG (1999) Berry-Esseen bounds for smooth estimator of a distributionfunction under association. Nonparametric Statist 11:79–106
[8] Esary JD, Proschan F (1972) Relationships among some concepts of bivariate dependence.Ann Math Statist 43: 651–655
[9] Gyorfi L (1981) Strong consistent density estimate from ergodic sample. J Multivariate Anal11:81–84
[10] Hall P and Heyde CC (1980) Martingale Limit Theory and Its Applications. New York:Academic Press.
[11] Joag-Dev K, Proschan F (1983) Negative association of random variables with applications.Ann Statist 11:286–295
[12] Lehmann E (1966) Some concepts of dependence. Ann Math Statist 37:1137–1153[13] Liang HY (2000) Complete convergence for weighted sums of negatively associated random
variables. Statist & Probab Lett 48:317–325[14] Lu Z (2001) Asymptotic normality of kernel density estimators under dependence. Ann Inst
Statist Math 53(3):447–468[15] Masry E (1986) Recursive probability density estimation for weakly dependent processes.
IEEE Trans Inform Theory 32:254–267[16] Masry E (1987) Almost sure convergence of recursive density estimators for stationary
mixing processes. Statist Probab Lett 5:249–254[17] Masry E and Gyorfi L (1987) Strong consistency and rates for recursive probability density
estimators of stationary processes. J Multivariate Anal 22:79–93[18] Parzen E (1962) On estimation of probability density function and mode. Ann Math Statist
33:1065–1076[19] Roussas GG (1994) Asymptotic normality of random fields of positively or negatively
associated processes. J Multivariate Anal 50:152–173[20] Roussas GG (2000) Asymptotic normality of the kernal estimate of a probability density
function under association. Statist & Probab Lett 50:1–12[21] Roussas GG (2001) An Esseen-type inequality for probability density functions, with an
application. Statist & Probab Lett 51:379–408[22] Shao QM, Su C (1999) The law of the iterated logarithm for negatively associated random
variables. Stochastic Process Appl 83:139–148[23] Takahata H (1980) Almost sure convergence of density estimators for weakly dependent
stationary processes. Bull Tokyo Gakugei Univ 32(4):11–32[24] Tran LT (1989) Recursive density estimation under dependence. IEEE Trans Inform Theory
35:1103–1108[25] Tran LT (1990) Recursive density estimation under a weak dependence condition. Ann Inst
Statist Math 42(2):305–329[26] Wegman EJ and Davies HI (1979) Remarks on some recursive estimators of a probability
density function. Ann Statist 7(2):316–327[27] Wolverton CT and Wagner TJ (1969) Asymptotically optimal discriminant functions for
pattern classification. IEEE Trans Inform Theory 15:258–265[28] Yamato H (1971) Sequential estimation of a continuous probability density function and
mode. Bull Math Statist 14:1–12
166 H.-Y. Liang and J.-Il. Baek