J. Math. Anal. Appl.

429 (2015) 497–512

Contents lists available at ScienceDirect

Journal of Mathematical Analysis and Applications

www.elsevier.com/locate/jmaa

The consistency of the nearest neighbor estimator of the density

function based on WOD samples ✩
Xuejun Wang a,b,∗ , Shuhe Hu a
a
School of Mathematical Sciences, Anhui University, Hefei 230601, PR China
b
Division of Mathematical Sciences, School of Physical and Mathematical Sciences,
Nanyang Technological University, Singapore 637371, Singapore

a r t i c l e i n f o a b s t r a c t

Article history: In this paper, the consistency of the nearest neighbor estimator of the density
Received 22 August 2014 function based on widely orthant dependent (WOD, in short) samples is investigated.
Available online 9 April 2015 The convergence rate of strong consistency, the complete consistency, the uniformly
Submitted by V. Pozdnyakov
complete consistency and uniformly strong consistency of the nearest neighbor
Keywords: estimator of the density function based on WOD samples are established. Our
Nearest neighbor estimator results established in the paper generalize or improve the corresponding ones for
WOD random variables independent samples and some negatively dependent samples.
Strong consistency © 2015 Elsevier Inc. All rights reserved.
Complete consistency
Uniform consistency

1. Introduction

1.1. Brief review

Suppose that the population X has the unknown density function f (x) and X1 , X2 , · · · , Xn are samples
of X. Let {kn , n ≥ 1} be a sequence of positive integers such that 1 ≤ kn ≤ n. For fixed x and n, denote

an (x) = min{a : there exist at least kn i such that Xi ∈ [x − a, x + a]}.

Loftsgarden and Quesenberry [14] introduced the nearest neighbor estimator of f (x) as follows:

kn
fn (x) = . (1.1)
2nan (x)

✩
Supported by the National Natural Science Foundation of China (11201001, 11171001, 11426032), the Natural Science
Foundation of Anhui Province (1508085J06) and the Research Teaching Model Curriculum of Anhui University (xjyjkc1407).
* Corresponding author.
E-mail address: wxjahdx2000@126.com (X. Wang).

http://dx.doi.org/10.1016/j.jmaa.2015.04.016
0022-247X/© 2015 Elsevier Inc. All rights reserved.
498 X. Wang, S. Hu / J. Math. Anal. Appl. 429 (2015) 497–512

Since the concept of the nearest neighbor estimator of the density function was introduced by Loftsgarden
and Quesenberry [14], many authors devoted to study the asymptotic properties of the nearest neighbor
estimator of the density function. For independent samples, Loftsgarden and Quesenberry [14] and Wag-
ner [19] established the weak consistency and strong consistency, respectively; Moore and Henrichon [15]
and Devroye and Wagner [6] obtained the uniform consistency and strong uniform consistency, respectively;
Chen [3] established the convergence rate of the consistency of the nearest neighbor density estimator, and
so on. For dependent samples, Boente and Fraiman [1] studied the strong consistency of the nearest neighbor
density estimator based on ϕ-mixing and α-mixing samples; Chai [2] obtained the strong consistency, weak
consistency, uniformly strong consistency and convergence rate of the nearest neighbor density estimator
based on ϕ-mixing stationary processes; Yang [26] investigated the weak consistency, strong consistency,
uniformly strong consistency of the nearest neighbor density estimator based on negatively associated (NA,
in short) samples; Liu and Zhang [13] established the consistency and asymptotic normality of nearest
neighbor density estimator based on ϕ-mixing samples, and so on.
Recently, Yang [26] established the following weak consistency and strong consistency for NA samples.

Theorem A. Let {Xn , n ≥ 1} be a sequence of NA random variables and kn → ∞, kn

n → 0 as n → ∞.

2
kn
(i) If n → ∞ as n → ∞, then

fn (x) −−
P
→ f (x). (1.2)

(ii) If for any γ > 0,

∞
  
γk2
exp − n < ∞, (1.3)
n=1
n

then

fn (x) → f (x) a.s., as n → ∞. (1.4)

The main purpose of this article is to generalize and improve the result of Theorem A for NA random
variables to the case of widely orthant dependent (WOD) random variables, which includes NA as a special
case. In addition, we will present the convergence rate of strong consistency, uniformly complete consistency
and uniformly strong consistency for the nearest neighbor estimator of density function based on WOD
samples. Our results generalize or improve the corresponding ones of Yang [26] for NA random variables to
the case of WOD random variables. The key techniques used in the paper are the Bernstein type inequality
and the truncated method.

1.2. Concepts of widely orthant dependence

In this subsection, we will recall the definition of WOD random variables, which was introduced by Wang
et al. [23] as follows.

Definition 1.1. For the random variables {Xn , n ≥ 1}, if there exists a finite real sequence {gU (n), n ≥ 1}
satisfying for each n ≥ 1 and for all xi ∈ (−∞, ∞), 1 ≤ i ≤ n,


n
P (X1 > x1 , X2 > x2 , · · · , Xn > xn ) ≤ gU (n) P (Xi > xi ),
i=1

then we say that the {Xn , n ≥ 1} are widely upper orthant dependent (WUOD, in short); if there exists a
finite real sequence {gL (n), n ≥ 1} satisfying for each n ≥ 1 and for all xi ∈ (−∞, ∞), 1 ≤ i ≤ n,
 X. Wang, S. Hu / J. Math. Anal. Appl. 429 (2015) 497–512 499


n
P (X1 ≤ x1 , X2 ≤ x2 , · · · , Xn ≤ xn ) ≤ gL (n) P (Xi ≤ xi ),
i=1

then we say that the {Xn , n ≥ 1} are widely lower orthant dependent (WLOD, in short); if they are both
WUOD and WLOD, then we say that the {Xn , n ≥ 1} are widely orthant dependent (WOD, in short), and
gU (n), gL (n), n ≥ 1, are called dominating coefficients.

Recall that when gL (n) = gU (n) = M for some positive constant M , the random variables {Xn , n ≥ 1}
are called extended negatively upper orthant dependent (ENUOD, in short) and extended negatively lower
orthant dependent (ENLOD, in short), respectively. If they are both ENUOD and ENLOD, then we say that
the random variables {Xn , n ≥ 1} are extended negatively orthant dependent (END, in short). The concept
of END random variables was proposed by Liu [11] and further promoted by Chen et al. [4], Shen [16,18],
Wu and Guan [25], Wang and Wang [22], Yang and Wang [27], and so forth. When gL(n) = gU (n) = 1 for
any n ≥ 1, the random variables {Xn , n ≥ 1} are called negatively upper orthant dependent (NUOD, in
short) and negatively lower orthant dependent (NLOD, in short), respectively. If they are both NUOD and
NLOD, then we say that the random variables {Xn , n ≥ 1} are negatively orthant dependent (NOD, in
short). The concept of negative dependence was introduced by Ebrahimi and Ghosh [7] and carefully studied
by Joag-Dev and Proschan [10]. Joag-Dev and Proschan [10] pointed out that negatively associated (NA,
in short) random variables are NOD. Hu [9] introduced the concept of negatively superadditive dependence
(NSD, in short) and gave an example illustrating that NSD does not imply NA. Christofides and Vaggelatou
[5] indicated that NA implies NSD. In addition, Hu [9] pointed out that NSD implies NOD. From the
statements above, we can see that the class of WOD random variables includes END random variables,
NOD random variables, NSD random variables, NA random variables and independent random variables
as special cases. Hence, studying the limiting behavior of WOD random variables and its applications are
of great interest.
The concept of WOD random variables was introduced by Wang et al. [23] and many applications have
been found. See, for example, Wang and Chen [20] presented some basic renewal theorems for random walks
with widely dependent increments. Wang et al. [21] studied the uniform asymptotics for the finite-time
ruin probability of a new dependent risk model with a constant interest rate. Liu et al. [12] gave the
asymptotically equivalent formula for the finite-time ruin probability under a dependent risk model with
constant interest rate. He et al. [8] provided the asymptotic lower bounds of precise large deviations with
nonnegative and dependent random variables. Shen [17] established the Bernstein type inequality for WOD
random variables and gave some applications, Wang et al. [24] studied the complete convergence for WOD
random variables and gave its applications in nonparametric regression models, and so forth.
This work is organized as follows: main results are presented in Section 2. Some important lemmas are
provided in Section 3 and the proofs of the main results are provided in Section 4, respectively.
Throughout this article, let {Xn , n ≥ 1} be a sequence of WOD random variables with the dominating
coefficients gU (n), gL (n), n ≥ 1. Denote g(n) = max{gU (n), gL (n)}. The symbol C denotes a positive
constant which is not necessarily the same one in each appearance. an = o(bn ) and an = O(bn ) stand for
limn→∞ abnn = 0 and an ≤ Cbn , respectively. Let I(A) be the indicator function of the event A. Denote
log x = ln max(x, e). x denotes the integer part of x. Let F (x) be the corresponding distribution function
n
of the density function f (x) and Fn (x) = n1 i=1 I(Xi < x) be the empirical distribution function of
X1 , X2 , · · · , Xn .

2. Main results

In this section, let c(f ) denote all continuity points of the density function f . Our main results are as
follows.
500 X. Wang, S. Hu / J. Math. Anal. Appl. 429 (2015) 497–512

2.1. Weak consistency

2
kn
Theorem 2.1. Let {Xn , n ≥ 1} be a sequence of WOD random variables and kn
n → 0, n → ∞ as n → ∞.
If for any γ > 0,
 
γkn2
lim g(n) exp − = 0, (2.1)
n→∞ n

then for any x ∈ c(f ), (1.2) holds.

2.2. Complete consistency and strong consistency

Theorem 2.2. Let {Xn , n ≥ 1} be a sequence of WOD random variables and kn

n → 0 as n → ∞. If for any
γ > 0,

∞
  
γk2
g(n) exp − n < ∞, (2.2)
n=1
n

then for any x ∈ c(f ),

fn (x) → f (x) completely, as n → ∞, (2.3)

and thus (1.4) holds.

Remark 2.1. We point out that the condition kn → ∞ as n → ∞ is not needed in Theorem 2.1 and
Theorem 2.2. If g(n) = M , where M is a positive constant, then (2.1) is obvious, and (2.2) is equivalent
to (1.3). In this case, {Xn , n ≥ 1} is a sequence of END random variables, which includes independent
random variables, NA random variables, NOD random variables as special cases. Hence, Theorem 2.1 and
Theorem 2.2 hold for independent random variables, NA random variables, NOD random variables and
END random variables. In addition, Borel–Cantelli lemma yields that complete convergence implies strong
convergence. Hence, the results of Theorem 2.1 and Theorem 2.2 generalize and improve the corresponding
ones of (i) and (ii) in Theorem A for NA random variables, respectively.

By using Theorem 2.2, we can get the following corollary.

Corollary 2.1. Let {Xn , n ≥ 1} be a sequence of WOD random variables with g(n) = O(nδ ) for some δ ≥ 0
and knn → 0 as n → ∞. If

kn
√ → ∞, as n → ∞, (2.4)
n log n

then for any x ∈ c(f ), (2.3) holds, and thus (1.4) holds.

The following result presents the complete consistency and strong convergence rate for the nearest neigh-
bor estimator of the density function.

Theorem 2.3. Let {Xn , n ≥ 1} be a sequence of WOD random variables with g(n) = O(nδ ) for δ ≥ 0. Let
f (x) satisfy the local Lipschitz condition at x and f (x) > 0. If there exists a sequence {qn , n ≥ 1} of positive
numbers such that
 X. Wang, S. Hu / J. Math. Anal. Appl. 429 (2015) 497–512 501

kn k q
qn → 0, → 0, √ n n → ∞, as n → ∞, (2.5)
nqn n log n

then for any ε > 0,

∞

P (|fn (x) − f (x)| > εqn ) < ∞, (2.6)
n=1

and thus,

|fn (x) − f (x)| = o(qn ) a.s., as n → ∞. (2.7)

Remark 2.2. We point out that the condition (2.5) yields that kn → ∞ as n → ∞. Actually,
√
kn qn n log n
kn = √ · → ∞, as n → ∞.
n log n qn

Taking kn = n3/4 log1/4 n and qn = n−1/4 (log n)1/4 log log n in Theorem 2.3, we can get the following
result on strong convergence rate.

Corollary 2.2. Let {Xn , n ≥ 1} be a sequence of WOD random variables with g(n) = O(nδ ) for δ ≥ 0. Let
f (x) satisfy the local Lipschitz condition at x and f (x) > 0. If kn = n3/4 log1/4 n, then
 
|fn (x) − f (x)| = o n−1/4 (log n)1/4 log log n a.s., as n → ∞.

2.3. Uniformly complete consistency and uniformly strong consistency

In this subsection, we will provide some results on uniformly complete consistency and uniformly strong
consistency for the nearest neighbor estimator of the density function.

Theorem 2.4. Let {Xn , n ≥ 1} be a sequence of WOD random variables with g(n) = O(nδ ) for some δ ≥ 0
and f (x) be uniformly continuous. If (2.4) holds and knn → 0 as n → ∞, then for any ε > 0,

∞

P sup |fn (x) − f (x)| > ε < ∞, (2.8)
x
n=1

and thus,

lim sup |fn (x) − f (x)| = 0 a.s. (2.9)

n→∞ x

Theorem 2.5. Let {Xn , n ≥ 1} be a sequence of WOD random variables with g(n) = O(nδ ) for some δ ≥ 0
and f (x) satisfy the Lipschitz condition on R. If there exists a sequence {qn , n ≥ 1} of positive numbers
such that

kn k q
kn → ∞, qn → 0, → 0, √ n n → ∞, as n → ∞, (2.10)
nqn2 n log n

then for any ε > 0,

∞

P sup |fn (x) − f (x)| > εqn < ∞, (2.11)
x
n=1
502 X. Wang, S. Hu / J. Math. Anal. Appl. 429 (2015) 497–512

and thus,

sup |fn (x) − f (x)| = o(qn ) a.s., as n → ∞. (2.12)

x

Taking kn = n2/3 log1/3 n and qn = n−1/6 (log n)1/6 log log n in Theorem 2.5, we can get the following
strong convergence rate for the nearest neighbor estimator of the density function.

Corollary 2.3. Let {Xn , n ≥ 1} be a sequence of WOD random variables with g(n) = O(nδ ) for some δ ≥ 0
and f (x) satisfy the Lipschitz condition on R. If kn = n2/3 log1/3 n, then

 
sup |fn (x) − f (x)| = o n−1/6 (log n)1/6 log log n a.s., as n → ∞.
x

3. Some lemmas

To prove the main results of the paper, we need the following important lemmas.
The first one is the basic property for WOD random variables, which comes from Wang et al. [23].

Lemma 3.1. Let {Xn , n ≥ 1} be WOD.

(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 s ∈ R,


n 
n
E exp s Xi ≤ g(n) E exp{sXi }. (3.1)
i=1 i=1

The next one is the Bernstein type inequality for WOD random variables, which was established by Shen
[17].

Lemma 3.2. Let {Xn , n ≥ 1} be a sequence of WOD random variables with EXi = 0 and |Xi | ≤ b for each
n
i ≥ 1, where b is a positive constant. Denote Bn2 = i=1 EXi2 for each n ≥ 1. Then for any ε > 0,

    
 n  ε2
 
P  Xi  ≥ ε ≤ 2g(n) exp − 2 2 . (3.2)
  2Bn + 3 bε
i=1

The last one is a basic property for empirical distribution function, which has been proved by Yang [26].

Lemma 3.3. Let F (x) be a continuous distribution function. For n ≥ 3, assume that xn,k satisfy

k
F (xn,k ) = , k = 1, 2, · · · , n − 1.
n

Then

2
sup |Fn (x) − F (x)| ≤ max |Fn (xn,k ) − F (xn,k )| + , (3.3)
x 1≤k≤n−1 n
n
where Fn (x) = 1
n i=1 I(Xi < x) is the empirical distribution function of X1 , X2 , · · · , Xn .
 X. Wang, S. Hu / J. Math. Anal. Appl. 429 (2015) 497–512 503

4. Proofs of the main results

In this section, we will present the proofs of the main results.

Proof of Theorem 2.1. Let x ∈ c(f ). For any ε > 0 and n ≥ 1, denote

kn kn
bn (x) = , cn (x) = (f (x) > ε is needed).
2n(f (x) + ε) 2n(f (x) − ε/2)

It follows by the definition of fn (x) that

.
Ax = {|fn (x) − f (x)| > ε}

= {fn (x) > f (x) + ε} {fn (x) < f (x) − ε}

= {fn (x) > f (x) + ε} {fn (x) < f (x) − ε, f (x) > ε}

⊂ {fn (x) > f (x) + ε} {fn (x) < f (x) − ε/2, f (x) > ε}

= {an (x) < bn (x)} {an (x) > cn (x), f (x) > ε}
 
kn
⊂ Fn (x + bn (x)) − Fn (x − bn (x)) ≥
n
  
kn
Fn (x + cn (x)) − Fn (x − cn (x)) ≤ , f (x) > ε
n
. 
= A1x A2x . (4.1)

Noting that bn (x) → 0, cn (x) → 0, and F  (x) = f (x), we have for n → ∞ that

F (x + bn (x)) − F (x − bn (x))
→ f (x)
2bn (x)

and

F (x + cn (x)) − F (x − cn (x))
→ f (x),
2cn (x)

which imply that for all n large enough,

kn f (x) + ε/2
F (x + bn (x)) − F (x − bn (x)) < 2bn (x)(f (x) + ε/2) = (4.2)
n f (x) + ε

and

kn f (x) − ε/4
F (x + cn (x)) − F (x − cn (x)) > 2cn (x)(f (x) − ε/4) = , (4.3)
n f (x) − ε/2

respectively.
504 X. Wang, S. Hu / J. Math. Anal. Appl. 429 (2015) 497–512

Denote e(x) = ε
8(f (x)+ε) ≤ 18 , it follows by (4.1) and (4.2) that
 
kn
A1x = Fn (x + bn (x)) − Fn (x − bn (x)) ≥
n
 
kn kn f (x) + ε/2
⊂ Fn (x + bn (x)) − Fn (x − bn (x)) − F (x + bn (x)) + F (x − bn (x)) ≥ −
n n f (x) + ε
 
4kn
= Fn (x + bn (x)) − Fn (x − bn (x)) − F (x + bn (x)) + F (x − bn (x)) ≥ e(x)
n
 
kn
⊂ |Fn (x + bn (x)) − F (x + bn (x))| ≥ e(x)
n
 kn

|Fn (x − bn (x)) − F (x − bn (x))| ≥ e(x)
n
. 
= E1x E2x . (4.4)

Similarly, it follows by (4.1) and (4.3) that

 
kn
A2x = Fn (x + cn (x)) − Fn (x − cn (x)) ≤ , f (x) > ε
n
⊂ {Fn (x + cn (x)) − Fn (x − cn (x)) − F (x + cn (x)) + F (x − cn (x))
kn kn f (x) − ε/4
≤ − , f (x) > ε}
n n f (x) − ε/2
= {Fn (x + cn (x)) − Fn (x − cn (x)) − F (x + cn (x)) + F (x − cn (x))
kn ε
≤− , f (x) > ε}
4n f (x) − ε/2
 
kn ε
⊂ Fn (x + cn (x)) − Fn (x − cn (x)) − F (x + cn (x)) + F (x − cn (x)) ≤ −
4n f (x) + ε
 
2kn
= Fn (x + cn (x)) − Fn (x − cn (x)) − F (x + cn (x)) + F (x − cn (x)) ≤ − e(x)
n
 
kn
⊂ |Fn (x + cn (x)) − F (x + cn (x))| ≥ e(x)
n
 kn

|Fn (x − cn (x)) − F (x − cn (x))| ≥ e(x)
n
. 
= E3x E4x . (4.5)

By (4.1), (4.4) and (4.5), we have

  
Ax ⊂ E1x E2x E3x E4x . (4.6)

For fixed x and n ≥ 1, denote

(n)
Xi = I(Xi < x + bn (x)) − EI(Xi < x + bn (x)), 1 ≤ i ≤ n.

(n) (n) (n) (n)

It is easily seen that X1 , X2 , . . . , Xn are still WOD random variables by Lemma 3.1 (i), and EXi = 0,
(n) (n) . n (n)
|Xi | ≤ 1. Applying Lemma 3.2 with Xi = Xi , b = 1 and Bn2 = i=1 E(Xi )2 ≤ n, we have by (3.2)
that,
 X. Wang, S. Hu / J. Math. Anal. Appl. 429 (2015) 497–512 505

kn
P (E1x ) = P |Fn (x + bn (x)) − F (x + bn (x))| ≥ e(x)
n
  
n 
 (n) 
=P  Xi  ≥ kn e(x)
 
i=1
 
kn2 e2 (x)
≤ 2g(n) exp −
2Bn2 + 23 kn e(x)
 
k2 e2 (x) 1
≤ 2g(n) exp − n 1 (since kn ≤ n and e(x) ≤ )
2n + 12 n 8
 
12e2 (x)kn2
= 2g(n) exp − . (4.7)
25n

Similarly, we have

 
12e2 (x)kn2
P (Ejx ) ≤ 2g(n) exp − , j = 2, 3, 4. (4.8)
25n

Therefore, we have by (4.6)–(4.8) and (2.1) that


4
P (|fn (x) − f (x)| > ε) = P (Ax ) ≤ P (Ejx )
j=1
 
12e2 (x)kn2
≤ 8g(n) exp −
25n
→ 0, as n → ∞. (4.9)

This completes the proof of the theorem. 2

Proof of Theorem 2.2. We use the same notations as those in Theorem 2.1. The desired result (2.3) follows by
(4.9) and (2.2), and (1.4) follows by (2.3) and Borel–Cantelli lemma, respectively. The proof is completed. 2

Proof of Corollary 2.1. It suffices to prove (2.2). In fact, for any γ > 0, we have by g(n) = O(nδ ) and (2.4)
that

∞
    ∞  
γkn2 γkn2
g(n) exp − = g(n) exp − log n
n=1
n n=1
n log n
∞

≤C n−2 < ∞.
n=1

The proof is completed. 2

Proof of Theorem 2.3. For any ε > 0 and n ≥ 1, denote

kn kn
un (x) = , vn (x) = (f (x) > εqn is needed).
2n(f (x) + εqn ) 2n(f (x) − εqn /2)
506 X. Wang, S. Hu / J. Math. Anal. Appl. 429 (2015) 497–512

Similar to the proof of (4.1), we have

.
Bx = {|fn (x) − f (x)| > εqn }

⊂ {fn (x) > f (x) + εqn } {fn (x) < f (x) − εqn /2, f (x) > εqn }
 
kn
⊂ Fn (x + un (x)) − Fn (x − un (x)) ≥
n
 kn

Fn (x + vn (x)) − Fn (x − vn (x)) ≤ , f (x) > εqn
n
. 
= B1x B2x . (4.10)

By the Differential Mean Value Theorem, we can see that there exist θ1 ∈ (x − un (x), x + un (x)) and
θ2 ∈ (x − vn (x), x + vn (x)) such that

F (x + un (x)) − F (x − un (x)) = 2un (x)f (θ1 ) (4.11)

and

F (x + vn (x)) − F (x − vn (x)) = 2vn (x)f (θ2 ). (4.12)

If ω ∈ B1x , then we have by (4.11) that

Fn (x + un (x)) − Fn (x − un (x)) − F (x + un (x)) + F (x − un (x))

kn kn f (x) − f (θ1 ) + εqn
≥ − 2un (x)f (θ1 ) = · . (4.13)
n n f (x) + εqn

Similarly, If ω ∈ B2x , then we have by (4.12) that

Fn (x + vn (x)) − Fn (x − vn (x)) − F (x + vn (x)) + F (x − vn (x))

kn kn f (x) − f (θ2 ) − εqn /2
≤ − 2vn (x)f (θ2 ) = · . (4.14)
n n f (x) − εqn /2

Since f (x) satisfies the local Lipschitz condition at x and f (x) > 0, qn → 0, nq
kn
n
→ 0 as n → ∞, we can see
that there exists a positive constant L(x) depending only on x such that for all n large enough,

L(x)kn L(x)kn εqn

|f (x) − f (θ1 )| ≤ L(x)|x − θ1 | ≤ L(x)un (x) = ≤ ≤ , (4.15)
2n(f (x) + εqn ) 2nf (x) 2
L(x)kn L(x)kn εqn
|f (x) − f (θ2 )| ≤ L(x)|x − θ2 | ≤ L(x)vn (x) = ≤ ≤ . (4.16)
2n(f (x) − εqn /2) nf (x) 4

Note that the density function f (x) is bounded. Denote M = sup f (x) < ∞. By (4.15) and (4.16), we can
x
see that for all n large enough,

kn f (x) − f (θ1 ) + εqn kn −εqn /2 + εqn kn qn ε

· ≥ · ≥ · (4.17)
n f (x) + εqn n f (x) + εqn n 4M

and

kn f (x) − f (θ2 ) − εqn /2 kn εqn /4 − εqn /2 kn qn ε

· ≤ · ≤− · . (4.18)
n f (x) − εqn /2 n f (x) − εqn /2 n 4M
 X. Wang, S. Hu / J. Math. Anal. Appl. 429 (2015) 497–512 507

ε
Denote d = 8M , we have by (4.13) and (4.17) that for all n large enough,

 
kn
B1x ⊂ − 2un (x)f (θ1 )
Fn (x + un (x)) − Fn (x − un (x)) − F (x + un (x)) + F (x − un (x)) ≥
n
 
2kn qn d
⊂ Fn (x + un (x)) − Fn (x − un (x)) − F (x + un (x)) + F (x − un (x)) ≥
n
 
kn qn d
⊂ |Fn (x + un (x)) − F (x + un (x))| ≥
n
  
kn qn d
|Fn (x − un (x)) − F (x − un (x))| ≥
n
. 
= Q1x Q2x . (4.19)

Similarly, we have by (4.14) and (4.18) that for all n large enough,

 
kn qn d
B2x ⊂ |Fn (x + vn (x)) − F (x + vn (x))| ≥
n
 kn qn d

|Fn (x − vn (x)) − F (x − vn (x))| ≥
n
. 
= Q3x Q4x . (4.20)

Therefore, we have by (4.10), (4.19) and (4.20) that for all n large enough,

  
Bx ⊂ Q1x Q2x Q3x Q4x . (4.21)

(n)
For fixed x and n, denote Xi = I(Xi < x + un (x)) − EI(Xi < x + un (x)), 1 ≤ i ≤ n. It is easily seen
(n) (n) (n) (n) (n)
that X1 , X2 , . . . , Xn are still WOD random variables by Lemma 3.1 (i) and EXi = 0, |Xi | ≤ 1.
(n) .  n (n)
Applying Lemma 3.2 with Xi = Xi , b = 1 and Bn2 = i=1 E(Xi )2 ≤ n, we have by g(n) = O(nδ ) and
√ kn q n
n log n
→ ∞ that for all n large enough,

  
 n 
 (n) 
P (Q1x ) = P  Xi  ≥ kn qn d
 
i=1
 
k 2 q 2 d2
≤ 2g(n) exp − 2 n n2
2Bn + 3 kn qn d
 2 2 2
k q d
≤ 2g(n) exp − n n
4n
 2 
d k2 q2
= 2g(n) exp − · n n · log n
4 n log n
≤ 2n−2 . (4.22)

Similar to the proof of (4.22), we have that for all n large enough,

P (Qjx ) ≤ 2n−2 , j = 2, 3, 4. (4.23)

508 X. Wang, S. Hu / J. Math. Anal. Appl. 429 (2015) 497–512

Therefore, it follows by (4.21)–(4.23) that

∞
 ∞

P (|fn (x) − f (x)| > εqn ) ≤ C n−2 < ∞, (4.24)
n=1 n=1

which implies (2.6). The desired result (2.7) follows by (4.24) and the Borel–Cantelli lemma immediately.
This completes the proof of the theorem. 2

Proof of Theorem 2.4. We use the same notations as those in the proof of Theorem 2.1.
Since f (x) is uniformly continuous, we have that for any ε > 0, there exists a δ > 0 such that when
|x − y| < δ,

ε
|f (x) − f (y)| < . (4.25)
4

It follows by kn
n → 0 that kn
εn < δ for all n large enough. Hence,

kn kn
bn (x) = ≤ < δ, for all x (4.26)
2n(f (x) + ε) 2εn

and

kn kn kn
cn (x) = ≤ = < δ, for all x. (4.27)
2n(f (x) − ε/2) 2n(ε − ε/2) εn

By the Differential Mean Value Theorem, we can see that there exist θ1 ∈ (x − bn (x), x + bn (x)) and
θ2 ∈ (x − cn (x), x + cn (x)) such that

F (x + bn (x)) − F (x − bn (x)) = 2bn (x)f (θ1 ) (4.28)

and

F (x + cn (x)) − F (x − cn (x)) = 2cn (x)f (θ2 ). (4.29)

It follows by (4.26) and (4.27) that |x − θ1 | < δ and |x − θ2 | < δ, respectively, which together with (4.25)
yield that
ε ε
|f (x) − f (θ1 )| < , |f (x) − f (θ2 )| < . (4.30)
4 4
Denote
 
ε dkn
M = sup f (x) < ∞, d = and B = sup |Fn (x) − F (x)| ≥ .
x 8(M + ε) x n

By (4.28) and (4.30), we can see that

 
kn
A1x = Fn (x + bn (x)) − Fn (x − bn (x)) − F (x + bn (x)) + F (x − bn (x)) ≥ − 2bn (x)f (θ1 )
n
 
kn f (x) − f (θ1 ) + ε
= ·
Fn (x + bn (x)) − Fn (x − bn (x)) − F (x + bn (x)) + F (x − bn (x)) ≥
n f (x) + ε
 
kn −ε/4 + ε
⊂ Fn (x + bn (x)) − Fn (x − bn (x)) − F (x + bn (x)) + F (x − bn (x)) ≥ ·
n f (x) + ε
 X. Wang, S. Hu / J. Math. Anal. Appl. 429 (2015) 497–512 509

 
kn ε
⊂ Fn (x + bn (x)) − Fn (x − bn (x)) − F (x + bn (x)) + F (x − bn (x)) ≥ ·
n 4(M + ε)
 
2dkn
⊂ Fn (x + bn (x)) − Fn (x − bn (x)) − F (x + bn (x)) + F (x − bn (x)) ≥
n
  
dkn dkn
⊂ |Fn (x + bn (x)) − F (x + bn (x))| ≥ |Fn (x − bn (x)) − F (x − bn (x))| ≥
n n
⊂ B. (4.31)

Similar to the proof of (4.31), we have by (4.29) and (4.30) that

 
kn f (x) − f (θ2 ) − ε/2
A2x = ·
Fn (x + cn (x)) − Fn (x − cn (x)) − F (x + cn (x)) + F (x − cn (x)) ≤
n f (x) − ε/2
 
kn ε/4 − ε/2
⊂ Fn (x + cn (x)) − Fn (x − cn (x)) − F (x + cn (x)) + F (x − cn (x)) ≤ ·
n f (x) + ε
 
2dkn
⊂ Fn (x + cn (x)) − Fn (x − cn (x)) − F (x + cn (x)) + F (x − cn (x)) ≤ −
n
  
dkn dkn
⊂ |Fn (x + cn (x)) − F (x + cn (x))| ≥ |Fn (x − cn (x)) − F (x − cn (x))| ≥
n n
⊂ B. (4.32)

Hence, we have by (4.1), (4.31) and (4.32) that Ax ⊂ B for all x, which together with Lemma 3.3 yields
that for all n large enough,


P sup |fn (x) − f (x)| > ε =P Ax
x
x

dkn
≤P sup |Fn (x) − F (x)| ≥
x n
2 dkn
≤P max |Fn (xn,k ) − F (xn,k )| + ≥
1≤k≤n−1 n n
dkn
≤P max |Fn (xn,k ) − F (xn,k )| ≥
1≤k≤n−1 2n

n−1
dkn
≤ P |Fn (xn,k ) − F (xn,k )| ≥ . (4.33)
2n
k=1

(nk)
For fixed k and n, denote Xi = I(Xi < xn,k ) − EI(Xi < xn,k ), 1 ≤ i ≤ n. It is easily seen that
(nk) (nk) (nk) (nk) (nk)
X1 , X2 , . . . , Xn are still WOD random variables by Lemma 3.1 (i) and EXi = 0, |Xi | ≤ 1.
(nk) 2 .
n (nk) 2
Applying Lemma 3.2 with Xi = Xi , b = 1 and Bn = i=1 E(Xi ) ≤ n, we have by (4.33), g(n) =
O(nδ ) and √nklog
n
n
→ ∞ that for all n large enough,

  

n−1  n  dk
 (nk)  n
P sup |fn (x) − f (x)| > ε ≤ P  Xi  ≥
x   2
k=1 i=1


n−1  
d2 kn2 /4
≤ 2g(n) exp −
k=1
2Bn2 + 13 dkn
510 X. Wang, S. Hu / J. Math. Anal. Appl. 429 (2015) 497–512


n−1  
Ck2
≤ 2g(n) exp − n
n
k=1


n−1  
Ckn2
= 2g(n) exp − · log n
n log n
k=1
−2
≤ 2n , (4.34)

which implies (2.8). Combining (2.8) with Borel–Cantelli lemma, we can get the desired result (2.9) imme-
diately. This completes the proof of the theorem. 2

Proof of Theorem 2.5. We used the same notations as those in the proof of Theorem 2.3.
2 → 0 that there exists a constant L > 0
kn
Since f (x) satisfies Lipschitz condition on R, we have by nq
n
such that for all n large enough,

Lkn Lkn εqn

|f (x) − f (θ1 )| ≤ L|x − θ1 | ≤ Lun (x) = ≤ ≤ (4.35)
2n(f (x) + εqn ) 2nεqn 2

and
Lkn Lkn εqn
|f (x) − f (θ2 )| ≤ L|x − θ2 | ≤ Lvn (x) ≤ = ≤ (4.36)
2n(εqn − εqn /2) nεqn 4

hold for all x.

Denote
 
ε kn qn d
M = sup f (x) < ∞, d = and D = sup |Fn (x) − F (x)| ≥ .
x 8M x n

Similar to the proof of (4.17)–(4.20), we have by (4.35) and (4.36) that

 
kn qn d
B1x ⊂ |Fn (x + un (x)) − F (x + un (x))| ≥
n
  
kn qn d
|Fn (x − un (x)) − F (x − un (x))| ≥
n
⊂D (4.37)

and
 
kn qn d
B2x ⊂ |Fn (x + vn (x)) − F (x + vn (x))| ≥
n
 kn qn d

|Fn (x − vn (x)) − F (x − vn (x))| ≥
n
⊂ D. (4.38)

Hence, by (4.10), (4.37) and (4.38), we obtain that Bx ⊂ D for all x, which together with (4.33) implies
that for all n large enough,

kn qn d
P sup |fn (x) − f (x)| > εqn ≤P sup |Fn (x) − F (x)| ≥
x x n

n−1
kn qn d
≤ P |Fn (xn,k ) − F (xn,k )| ≥ . (4.39)
2n
k=1
 X. Wang, S. Hu / J. Math. Anal. Appl. 429 (2015) 497–512 511

Similar to the proof of (4.34), we have by (4.39) and (2.10) that for all n large enough,


n−1
kn qn d
P sup |fn (x) − f (x)| > εqn ≤ P |Fn (xn,k ) − F (xn,k )| ≥
x 2n
k=1


n−1  
Ckn2 qn2
≤ 2g(n) exp − · log n
n log n
k=1

≤ 2n−2 ,

which implies (2.11). Combining (2.11) with Borel–Cantelli lemma, we can get the desired result (2.12)
immediately. This completes the proof of the theorem. 2

Acknowledgments

The authors are most grateful to the Editor-in-Chief, Associate Editor and anonymous referee for careful
reading of the manuscript and valuable suggestions which helped in improving an earlier version of this
paper.

References

[1] G. Boente, R. Fraiman, Consistency of a nonparametric estimate of a density function for dependent variables, J. Multi-
variate Anal. 25 (1) (1988) 90–99.
[2] G. Chai, Consistency of nearest neighbor density estimator of stationary processes, Acta Math. Sin. 32 (3) (1989) 423–432.
[3] X.R. Chen, The rate of consistency of nearest neighbor density estimator, Sci. China Ser. A 12 (1981) 1419–1428.
[4] Y. Chen, A. Chen, K.W. Ng, The strong law of large numbers for extend negatively dependent random variables, J. Appl.
Probab. 47 (2010) 908–922.
[5] T.C. Christofides, E. Vaggelatou, A connection between supermodular ordering and positive/negative association, J. Mul-
tivariate Anal. 88 (2004) 138–151.
[6] L.P. Devroye, T.J. Wagner, The strong uniform consistency of nearest neighbor density estimates, Ann. Statist. 5 (1977)
536–540.
[7] N. Ebrahimi, M. Ghosh, Multivariate negative dependence, Comm. Statist. Theory Methods 10 (1981) 307–337.
[8] W. He, D.Y. Cheng, Y.B. Wang, Asymptotic lower bounds of precise large deviations with nonnegative and dependent
random variables, Statist. Probab. Lett. 83 (2013) 331–338.
[9] T.Z. Hu, Negatively superadditive dependence of random variables with applications, Chinese J. Appl. Probab. Statist.
16 (2000) 133–144.
[10] K. Joag-Dev, F. Proschan, Negative association of random variables with applications, Ann. Statist. 11 (1983) 286–295.
[11] L. Liu, Precise large deviations for dependent random variables with heavy tails, Statist. Probab. Lett. 79 (2009) 1290–1298.
[12] X.J. Liu, Q.W. Gao, Y.B. Wang, A note on a dependent risk model with constant interest rate, Statist. Probab. Lett. 82
(2012) 707–712.
[13] Y. Liu, Y. Zhang, The consistency and asymptotic normality of nearest neighbor density estimator under ϕ-mixing
condition, Acta Math. Sci. 30 (3) (2010) 733–738.
[14] D.O. Loftsgarden, C.P. Quesenberry, A nonparametric estimate of a multivariate density function, Ann. Math. Statist.
36 (3) (1965) 1049–1051.
[15] D.S. Moore, E.G. Henrichon, Uniform consistency of some estimates of a density function, Ann. Math. Statist. 40 (4)
(1969) 1499–1502.
[16] A.T. Shen, Probability inequalities for END sequence and their applications, J. Inequal. Appl. 2011 (2011), Article ID 98,
12 pp.
[17] A.T. Shen, Bernstein-type inequality for widely dependent sequence and its application to nonparametric regression models,
Abstr. Appl. Anal. 2013 (2013), Article ID 862602, 9 pp.
[18] A.T. Shen, On asymptotic approximation of inverse moments for a class of nonnegative random variables, Statistics 48 (6)
(2014) 1371–1379.
[19] T.J. Wagner, Strong consistency of a nonparametric estimate of a density function, IEEE Trans. Syst. Man Cybern. 3
(1973) 289–290.
[20] Y.B. Wang, D.Y. Chen, Basic renewal theorems for random walks with widely dependent increments, J. Math. Anal. Appl.
384 (2011) 597–606.
[21] Y.B. Wang, Z. Cui, K. Wang, X. Ma, Uniform asymptotics of the finite-time ruin probability for all times, J. Math. Anal.
Appl. 390 (2012) 208–223.
[22] S.J. Wang, X.J. Wang, Precise large deviations for random sums of END real-valued random variables with consistent
variation, J. Math. Anal. Appl. 402 (2013) 660–667.
512 X. Wang, S. Hu / J. Math. Anal. Appl. 429 (2015) 497–512

[23] K. Wang, Y. Wang, Q. Gao, Uniform asymptotics for the finite-time ruin probability of a new dependent risk model with
a constant interest rate, Methodol. Comput. Appl. Probab. 15 (2013) 109–124.
[24] X.J. Wang, C. Xu, T.C. Hu, A. Volodin, S.H. Hu, On complete convergence for widely orthant-dependent random variables
and its applications in nonparametric regression models, TEST 23 (2014) 607–629.
[25] Y.F. Wu, M. Guan, Convergence properties of the partial sums for sequences of END random variables, J. Korean Math.
Soc. 49 (6) (2012) 1097–1110.
[26] S.C. Yang, Consistency of nearest neighbor estimator of density function for negative associated samples, Acta Math.
Appl. Sin. 26 (3) (2003) 385–395.
[27] Y. Yang, Y.B. Wang, Tail behavior of the product of two dependent random variables with applications to risk theory,
Extremes 16 (2013) 55–74.