Professional Documents
Culture Documents
PROBABILITY
AND
MATHEMATICAL STATISTICS
Vol. 0, Fasc. 0 (0000), pp. 000–000
doi:
i
scr
COMPLETE CONSISTENCY FOR RECURSIVE PROBABILITY DENSITY
ESTIMATOR OF WIDELY ORTHANT DEPENDENT SAMPLES∗
BY
anu
CHENLU Z H UA N S U N (H EFEI ) AND XIAOXIN L I (C HIZHOU )
1. INTRODUCTION
we will consider a rather weak and applicable dependence structure, i.e., widely or-
thant dependence structure, the concept of which was first introduced by Wang et
al. [16] as follows.
Definition 1.1. A finite collection of random variables X1 , X2 , · · · , Xn is said
cep
to be widely upper orthant dependent (WUOD) if there exists a finite real number
gU (n) such that for all finite real numbers xi , 1 ¬ i ¬ n,
∏
n
(1.1) P (X1 > x1 , X2 > x2 , · · · , Xn > xn ) ¬ gU (n) P (Xi > xi ).
i=1
∗
Supported by the key project of the outstanding young talent support program of the university
of Anhui Province (gxyqZD2016367).
pt
2 Che nlu Z hu ans un and Xiaoxin Li
∏
n
i
(1.2) P (X1 ¬ x1 , X2 ¬ x2 , · · · , Xn ¬ xn ) ¬ gL (n) P (Xi ¬ xi ).
i=1
scr
If X1 , X2 , · · · , Xn is both WUOD and WLOD, we then say that X1 , X2 , · · · , Xn is
widely orthant dependent random variables (WOD), and gU (n), gL (n) are called
dominating coefficients. A sequence of random variables {Xn , n 1} is said to
be WOD if every finite subcollection is WOD.
With various dominating coefficients, the WOD structure contains many other
dependence structures. Wang et al. [16] presented some examples showing that
anu
WOD random variables contain negatively dependent random variables, positively
dependent random variables, and some other classes of dependent random vari-
ables, and also they presented some examples of WOD random variables which do
not satisfy these other dependence structures.
It can be easily checked that gU (n) 1, gL (n) 1. If both (1.1) and (1.4)
hold with gU (n) = gL (n) = M for all n 1, where M is a positive constant, then
the random variables are called END, which was introduced by [7]. If M = 1 for
all n 1, then the random variables are called NOD, which was introduced by
dm
Lehmann [4] (cf. also Joag-Dev and Proschan [3]). As is well known, NA random
variables are NOD. Furthermore, Hu [2] pointed out that negatively superaddi-
tive dependent (NSD, for short) random variables are NOD. Hence, the class of
WOD random variables includes independent sequences, NA sequences, NSD se-
quences, NOD sequences, and END sequences as special cases. Thus, studying the
limit behavior of WOD random variables is of general interest. There are many re-
sults investigating the WOD random variables. For example, Wang and Cheng [19]
studied the basic renewal theorems for random walks with widely dependent incre-
ments; Chen et al. [1] investigated the uniform asymptotics for the finite-time ruin
te
i
nhn i=1 hn
scr
Wolverton and Wagner [20] introduced the following recursive kernel estimator of
f (x):
( )
1 ∑n
1 x − Xi
(1.3) fˆn (x) = K ,
n i=1 hi hi
anu
where 0 < hn ↓ 0 are bandwidths and K is some kernel function. Note that (1.3)
can be computed recursively by
( )
n−1 ˆ 1 x − Xn
(1.4) fˆn (x) = fn−1 (x) + K .
n nhn hn
This recursive property is particularly useful in large sample sizes since fˆn (x) can
be easily updated with each additional observation. This is especially relevant in a
time series context, where there has been an interest in the use of nonparametric
dm
estimates in very long financial time series. Also, under certain circumstances, the
recursive estimator is more efficient than its nonrecursive estimator fn (x) when ef-
ficiency is measured in terms of the variance of an appropriate asymptotic distribu-
tion. Moreover, the estimator can be applied in estimating the hazard rate function,
which is defined as r(x) = f (x)/(1 − F (x)), where f (x) is the unknown marginal
probability density function and F (x) is the distribution function. A general hazard
rate estimator for r(x) is
fˆn (x)
te
Baek [5] discussed the point asymptotic normality for fˆn (x) under NA random
variables. Masry [10] obtained the quadratic mean convergence and asymptotic
normality of the recursive estimator under various assumptions on the dependence
of Xi ; Li et al. [6] discuss the asymptotic bias, quadratic-mean convergence and
establish the pointwise asymptotic normality of fˆn (x) for a stationary sequence of
NA sequences. Li and Yang [8] studied the strong convergence rate of recursive
probability density estimator fˆn (x) based on NA samples. Li [9] extend the results
of Li and Yang [8] from NA samples to END samples.
Ac
ones of Li [9] for END samples and Li and Yang [8] for NA samples. We will also
study the complete consistency for the estimator (1.3) under some mild conditions.
The layout of the paper is as follows. Main results are presented in Section
i
2. Some lemmas are provided in Section 3. The proofs are given in Section 4.
Throughout this paper, C, c0 , c1 , · · · represent some positive constants whose value
scr
may be different in different places. a = O(b) implies that a ¬ Cb. C(f ) denotes
all the continuity points of function f and C 2 (f ) stands for a point set in where the
′′
second-order derivative f exists and is bounded and continuous.
2. MAIN RESULTS
∞ ∞
anu
In this section, we will present the strong convergence rate for the recursive
kernel estimator fˆn (x). We adopt the following assumptions which were also used
in Li ∫and Yang [8] and Li∫[9]. ∫∞
(A1 ) −∞ K(u)du = 1, −∞ uK(u)du = 0, −∞ u2 K(u)du < ∞, K(·) ∈ L1 ;
(A2 ) The bandwidths hn satisfy that 0 < hn ↓ 0 and nhn → ∞ as n → ∞.
Now we state our main results as follows.
Theorem 2.1. Suppose that (A1 )-(A2 ) hold. Let {Xn , n 1} be a sequence of
strictly stationary WOD random variables with g(n) = O(nδ ) for some δ 0.
dm
Suppose that the kernel K(·) is a bounded monotone density function and the
bandwidth hn = O(n−1/5 log1/5 n). Then for any x ∈ C 2 (f ),
( )
(2.1) |fˆn (x) − f (x)| = O [log n/(nhn )]1/2 , completely.
Remark 2.1. Li and Yang [8] and Li [9] obtained the similar results under NA
and
( END samples, respectively. ) The convergence rate obtained in their result is
o [log n(log log n)l /(nbn )]1/2 for some l > 0 under the meaning of almost surely
te
(a.s.). Noting that WOD contains END and NA, complete convergence is stronger
than a.s. convergence (by the Borel-Cantelli lemma), and the rate in our result is
slightly faster, thus our result improves and extends the corresponding ones of Li
[9] as well as Li and Yang [8].
cep
Relaxing the restriction on the dominating coefficients g(n), we have the fol-
lowing more general result.
Theorem 2.2. Suppose that (A1 )-(A2 ) hold. Let γn = [log(ng(n))/(nhn )]1/2 →
0. Let {Xn , n 1} be a sequence of strictly stationary WOD random variables.
Suppose that the kernel K(·) is a bounded monotone density function and the band-
width satisfies that hn = O(n−1/5 log1/5 n). Then for any x ∈ C 2 (f ),
(2.2) |fˆn (x) − f (x)| = O(γn ), completely.
Ac
Remark 2.2. In Theorem 2.1, we required that the dominating coefficients g(n)
are polynomial increasing, which is always assumed in many papers. However,
pt
Complete consistency for recursive probability density estimator of WOD samples 5
Theorem 2.2 allows that the dominating coefficients g(n) can be geometric in-
creasing. If g(n) = O(nδ ) for some δ 0, the strong convergence rate is the same
as that in Theorem 2.1. Consequently, Theorem 2.2 is much more general and ap-
i
plicable.
scr
Furthermore, by relaxing the restriction on the bandwidth hn , we have the
following result.
Theorem 2.3. Suppose that (A1 )-(A2 ) hold. Let {Xn , n 1} be a sequence
of strictly stationary WOD random variables. Suppose that the kernel K(·) is a
bounded monotone density function and log(ng(n))/(nhn ) → 0. Then for any
x ∈ C 2 (f ),
(2.3)
anu
fˆn (x) − f (x) → 0, completely.
Theorem 2.5. Suppose that (A1 )-(A2 ) hold. Let γn = [log(ng(n))/(nhn )]1/2 →
0. Let {Xn , n 1} be a sequence of strictly stationary WOD random variables.
te
Suppose that the kernel K(·) is a bounded monotone density function and the band-
width satisfies that hn = O(n−1/5 log1/5 n). If there exists a point x0 such that
F (x0 ) < 1, then for any x ∈ C 2 (f ) and x ¬ x0 ,
cep
Theorem 2.6. Suppose that (A1 )-(A2 ) hold. Let {Xn , n 1} be a sequence
of strictly stationary WOD random variables. Suppose that the kernel K(·) is a
bounded monotone density function and log(ng(n))/(nhn ) → 0. If there exists a
point x0 such that F (x0 ) < 1, then for any x ∈ C 2 (f ) and x ¬ x0 ,
Ac
3. SOME LEMMAS
In this section, we will present some lemmas which will be used in proving
i
our main results.
Lemma 3.1. (Wang et al. [17]). Let {Xn , n 1} be a sequence of WOD random
scr
variables.
(i) If {fn (·), n 1} are all nondecreasing (or nonincreasing), then {fn (Xn ), n
1} are still WOD.
(ii) For each n 1 and any t ∈ R,
{ }
n
∑ ∏
n
anu
E exp t Xi ¬ g(n) E exp{tXi }.
i=1 i=1
Proof. Nothing that |tXi | ¬ 1 a.s. and EXi = 0 for each i 1, we have
∞ ∞
∑ E(tXi )k ∑ 1
E exp{tXi } = 1 + ¬ 1 + t2 EXi2
k=2 k! k=2 k!
(3.1) ¬ 1 + t2 EXi2 ¬ exp{t2 EXi2 }.
te
By Markov’s inequality, Lemma 3.1 (ii) and (3.1), we can see that
( ) { }
n
∑ n
∑ ∏
n
P Xi ε ¬ e−tε E exp t Xi ¬ g(n)e−tε E exp{tXi }
cep
The desired result follows by replacing Xi by −Xi in (3.2). This completes the
proof of the lemma.
Lemma 3.3. (Li and Yang [8]). Suppose that (A1 ) holds, then for all x ∈ C 2 (f ),
∫
Ac
Lemma 3.4. (Li and Yang [8]). Suppose that (A1 ) holds, then for all x ∈ C 2 (f ),
( )−1
n
1 ∑
i
h2
|E fˆn (x) − f (x)| ¬ C < ∞.
n i=1 i
scr
Lemma 3.5. Let {Xn , n 1} be a sequence of WOD random variables with un-
known distribution function F (x) and bounded probability density function f (x).
Let Fn (x) be the empirical distribution function. If µn =: [log(ng(n))/n]1/2 → 0,
then
sup |Fn (x) − F (x)| = O(µn ), completely.
anu
x
Noting that nµn → ∞, then for any positive constant D1 , we have that 2/n <
D1 µn /2 for all n large enough. Then if follows from (3.3) that
( ) ( )
P sup |Fn (x) − F (x)| > D1 µn ¬ P max |Fn (xnj ) − F (xnj )| > D1 µn /2
−∞<x<∞ 1¬j¬n−1
n−1
∑
(3.4) ¬ P (|Fn (xnj ) − F (xnj )| > D1 µn /2).
te
j=1
Let ξi = I(Xi < xnj ) − EI(Xi < xnj ). By Lemma 3.1, {ξi , i 1} is still a se-
quence of WOD random variables with Eξi = 0, |ξi | ¬ 2 and Eξi2 ¬ 1. Thus by
choosing t = D1 µn /4 in Lemma 3.2 we have that for all n large enough,
cep
( )
n
∑
P (|Fn (xnj ) − F (xnj )| > D1 µn /2) = P ξi > D1 nµn /2
i=1
{ }
n
2 ∑
¬ 2g(n) exp −D1 nµn t/2 + t Eξi2
i=1
{ }
¬ 2g(n) exp −D1 nµn t/2 + nt 2
{ }
¬ 2g(n) exp −D12 nµ2n /16
Ac
{ }
¬ 2g(n) exp −D12 c0 log(ng(n))/16
¬ 2g(n)(ng(n))−D1 c0 /16 .
2
(3.5)
pt
8 Che nlu Z hu ans un and Xiaoxin Li
Recall that g(n) 1. Taking D1 sufficiently large such that D12 c0 /16 > 3, by (3.4)
and (3.5) we have that
( )
i
∑∞ ∑∞ n−1
∑
g(n)(ng(n))−D1 c0 /16 < ∞.
2
P sup |Fn (x) − F (x)| > D1 µn ¬ C
scr
n=1 −∞<x<∞ n=1 j=1
have that
∑n
i=1
Eηi ¬
2 ∑n
i=1
−2
hi EK
(
anu
i ¬ n. Since K(·) is bounded and monotone, then {ηi , i 1} is still a sequence
of WOD random variables. Moreover, it follows from 0 < hn ↓ 0 that there exists
some positive constant c1 such that max1¬i¬n |ηi | ¬ c1 /hn . By Lemma 3.2 we
2 x − Xi
hi
)
( )
∑ −2 ∫
n
2 x−u
dm
= hi K f (u)du
i=1 R hi
n
∑ ∫
= h−1
i K 2 (u)f (x − hi u)du ¬ c2 nh−1
n .
i=1 R
Set λn = [log n/(nhn )]1/2 . Applying Lemma 3.2 with t = D2 hn λn /(2c2 ), where
D2 is some positive constant which will be specified later. It is easy to check that
t · c1 /hn ¬ 1 for all n large enough, then we get that
( )
te
∑n
P (|fˆn (x) − E fˆn (x)| > D2 λn ) = P ηi > D2 nλn
i=1
{ }
n
2 ∑
¬ 2g(n) exp −D2 nλn t + t Eηi2
cep
i=1
{ }
¬ 2g(n) exp −D2 nλn t + c2 nh−1
n t
2
{ }
¬ 2g(n) exp −D22 nhn λ2n /(4c2 )
{ }
¬ 2g(n) exp − log n · D22 /(4c2 )
2
¬ Cnδ−D2 /(4c2 ) .
Taking D2 large enough such that δ − D22 /(4c2 ) < −2, then we have that
Ac
∞
∑
P (|fˆn (x) − E fˆn (x)| > D2 λn ) < ∞,
n=1
pt
Complete consistency for recursive probability density estimator of WOD samples 9
that is,
( )
i
(4.1) |fˆn (x) − E fˆn (x)| = O [log n/(nhn )]1/2 , completely.
scr
On the other hand, noting that hn = O(n−1/5 log1/5 n), we have by Lemma
3.4 that
n
1 ∑
[log n/(nhn )]−1/2 |E fˆn (x) − f (x)| ¬ C[log n/(nhn )]−1/2 h2
n i=1 i
anu
n
∑
¬ C(hn /(n log n))1/2 h2i
i=1
n
∑
¬ Cn−3/5 log−2/5 n −2/5
i log2/5 i
i=1
¬ C,
Note that
(4.3) |fˆn (x) − f (x)| ¬ |fˆn (x) − E fˆn (x)| + |E fˆn (x) − f (x)|.
te
Therefore, the desired result (2.1) follows immediately by (4.1)-(4.3). The proof is
completed.
Proof of Theorem 2.2. In view of the proof of Theorem 2.1, we only need to
show that
cep
( )
(4.4) |fˆn (x) − E fˆn (x)| = O [log(ng(n))/(nhn )]1/2 , completely.
and
( )
(4.5) |E fˆn (x) − f (x)| = O [log(ng(n))/(nhn )]1/2 .
Ac
Noting that g(n) 1, then (4.5) follows from (4.2) immediately and thus we only
need to prove (4.4). Similar to the proof of (4.1), set γn = [log(ng(n))/(nhn )]1/2 .
pt
10 Che nlu Z hu ans un and Xiaoxin Li
Applying Lemma 3.2 with t = D3 hn γn /(2c2 ) to obtain that for all n large enough,
( )
∑n
P (|fˆn (x) − E fˆn (x)| > D3 γn ) = P
i
ηi > D3 nγn
i=1
{ }
scr
n
2 ∑
¬ 2g(n) exp −D3 nγn t + t Eηi2
i=1
{ }
¬ 2g(n) exp −D3 nγn t + c2 nh−1
n t
2
{ }
¬ 2g(n) exp −D32 nhn γn2 /(4c2 )
{ }
¬ 2g(n) exp −D32 /(4c2 ) · log(ng(n))
anu
¬ 2g(n)(ng(n))−D3 /(4c2 ) .
2
Taking D3 sufficiently large such that D32 /(4c2 ) > 2, then we have that
∞
∑
P (|fˆn (x) − E fˆn (x)| > D3 γn ) < ∞,
n=1
n
1 ∑
lim |E fˆn (x) − f (x)| ¬ C lim h2i = C lim h2n = 0.
n→∞ n→∞ n n→∞
i=1
Consequently, (4.6) is proved and thus the proof of the theorem is completed.
cep
Since the proofs of Theorems 2.4-2.6 are similar, we only present the proof of
Theorem 2.4 as follows.
Proof of Theorem 2.4. Set F̄n (x) = 1 − Fn (x) and F̄ (x) = 1 − F (x). It
follows from (1.5) that
F̄ (x)|fˆn (x) − f (x)| + |Fn (x) − F (x)|f (x)
(4.7) |r̂n (x) − r(x)| ¬ .
F̄n (x)F̄ (x)
From 0 ¬ F (x) ¬ F (x0 ) < 1 for all x ¬ x0 , supx f (x) ¬ C < ∞, applying The-
Ac
orem 2.1 and taking µn = (log n/n)1/2 in Lemma 3.5, we can see that
( )
(4.8) |fˆn (x) − f (x)| = O [log n/(nhn )]1/2 , completely.
pt
Complete consistency for recursive probability density estimator of WOD samples 11
and
( )
(4.9) sup |Fn (x) − F (x)| = O (log n/n)1/2 , completely.
i
x¬x0
On the other hand, from (4.9) one has that for x ¬ x0 and all n large enough,
scr
(4.10) F̄n (x) F̄ (x)/2 F̄ (x0 )/2 > 0.
Consequently, the desired result (2.4) follows from (4.7)-(4.10). The proof is com-
pleted.
REFERENCES
anu
[1] Y. C hen, L . Wang and Y. Wang, Uniform asymptotics for the finite-time ruin probabilities
of two kinds of nonstandard bidimensional risk models, Journal of Mathematical Analysis and
Applications, 401 (2013), pp. 114–129.
[2] T. H u, Negatively superadditive dependence of random variables with applications, Chinese
J. Appl. Probab. Statist. 16 (2000), pp. 133–144.
[3] K. Joag-D ev and F. Proschan, Negative association of random variables with applica-
tions, Ann. Statist. 11 (1983), pp. 286–295.
[4] E. L ehm ann, Some concepts of dependence, Ann. Math. Statist. 37 (1966), pp. 1137–1153.
[5] H. L i ang and J . Baek, Asymptotic normality of recursive density estimates under some
dm
dependence assumptions, Metrika, 60 (2004), pp. 155–166.
[6] Y. L i , C . We i and S. Yang, The recursive kernel distribution function estimator based
on negatively and positively associated sequences, Communications in Statistics-Theory and
Methods, 39:20 (2010), pp. 3585–3595.
[7] L. L iu, Precise large deviations for dependent random variables with heavy tails, Statist.
Probab. Lett. 79 (2009), pp. 1290–1298.
[8] Y. L i and S. Yang, Strong convergence rate of recursive probability density estimator for
NA sequences, Chinese J. Engineering Mathematics, 22:4 (2005), pp. 659–665.
[9] Y. L i, On the rate of strong convergence for a recursive probability density estimator of END
samples and its applications, Journal of Mathematical Inequalities, 11:2 (2017), pp. 335–343.
[10] E. Masr y, Recursive probability density estimation for weakly dependent stationary pro-
te
[18] X. Wa ng, Y. Wu, A. Rosalsky, Complete convergence for arrays of rowwise widely or-
thant dependent random variables and its applications, Stochastics: An International Journal
of Probability and Stochastic Processes, (2017, to appear).
i
[19] Y. Wa ng, D . Cheng, Basic renewal theorems for random walks with widely dependent in-
crements, Journal of Mathematical Analysis and Applications, 384 (2011), pp. 597–606.
scr
[20] C. Wol ve rt on and T. Wag ne r, Asymptotically optimal discriminant functions for pattern
classification, IEEE Trans Inform Theory, 15 (1969), pp. 258–265.
[21] S. Yang, Consistency of nearest neighbor estimator of density function for negative associated
samples, Acta Mathematice applicatae Sinica, 26:3 (2003), pp. 385–395.
[22] W. Ya ng, T. Liu, X. Wang and S . Hu, On the Bahadur representation of sample quan-
tiles for widely orthant dependent sequences, Filomat, 28 (2014), pp. 1333–1343.
anu
School of Mathematical Sciences Department of Mathematics and Computer
Anhui University, P.R. China Sciences
E-mail: zhuansuncl@163.com Chizhou University, P.R. China
E-mail: lxx@czu.edu.cn
Received on 5.5.2017;
revised version on 4.10.2017
te dm
cep
Ac