Journal of Multivariate Analysis 156 (2017) 103–115

Contents lists available at ScienceDirect

Journal of Multivariate Analysis

journal homepage: www.elsevier.com/locate/jmva

On local linear regression for strongly mixing random fields

Mohamed El Machkouri a,∗ , Khalifa Es-Sebaiy b , Idir Ouassou b
a
Université de Rouen Normandie, LMRS UMR CNRS 6085, avenue de l’Université, 76801 Saint-Étienne-du-Rouvray, France
b
National School of Applied Sciences-Marrakesh, Cadi Ayyad University, Av. Abdelkrim Khattabi, 40000 Guéliz-Marrakech, Morocco

article info abstract

Article history: We investigate the local linear kernel estimator of the regression function g of a stationary
Received 22 February 2016 and strongly mixing real random field observed over a general subset of the lattice Zd .
Available online 20 February 2017 Assuming that g is differentiable with derivative g ′ , we provide a new criterion on the
mixing coefficients for the consistency and the asymptotic normality of the estimators of g
AMS 2000 subject classifications: and g ′ under mild conditions on the bandwidth parameter. Our results improve the work
62G05
of Hallin et al. (2004) in several directions.
60J25
62G07
© 2017 Elsevier Inc. All rights reserved.

Keywords:
Local linear regression estimation
Strong mixing
Random fields
Asymptotic normality

1. Introduction and main results

In a variety of fields such as soil science, geology, oceanography, econometrics, epidemiology, image processing and
many others, the aim of practioners is to handle phenomena observed on spatial sets. In particular, one of the fundamental
questions is how to understand the phenomenon from a set of (dependent) observations based on regression models. In
this work, we investigate the problem in the context of strongly mixing spatial processes (or random fields) and we focus
on local linear regression estimation.
Let d be a positive integer and let {(Yi , Xi ) : i ∈ Zd } be a strictly stationary R2 -valued random field defined on a probability
space (Ω , F , Pr). The estimation of its regression function g defined by g (x) = E(Y0 |X0 = x) for almost all real x is a
natural question and a very important task in statistics. The non-spatial case d = 1, i.e., dependent time series, has been
extensively studied. One can refer, e.g., to Lu and Cheng [17], Masry and Fan [18], Robinson [21], Roussas [23] and many
references therein. For d > 2, contributions to the case of strongly mixing random fields were made by Biau and Cadre [1],
Carbon et al. [2], Dabo-Niang and Rachdi [4], Dabo-Niang and Yao [5], El Machkouri [7], El Machkouri and Stoica [10], Hallin
et al. [12] and Lu and Chen [15,16].
To our knowledge, there are no results on regression estimation for dependent random fields which are not strongly
mixing. However, the density estimation problem has been investigated, e.g., by El Machkouri [9] for a class of stationary
ergodic random fields which contains non mixing linear random fields studied by Cheng et al. [3], Hallin et al. [11], and
Wang and Woodroofe [25].
Given two σ -algebras U and V , the α -mixing coefficient introduced by Rosenblatt [22] is defined by
α(U, V ) = sup{| Pr(A ∩ B) − Pr(A) Pr(B)|, A ∈ U, B ∈ V }.

∗ Corresponding author.
E-mail addresses: mohamed.elmachkouri@univ-rouen.fr (M. El Machkouri), k.essebaiy@uca.ma (K. Es-Sebaiy), i.ouassou@uca.ma (I. Ouassou).

http://dx.doi.org/10.1016/j.jmva.2017.02.002
0047-259X/© 2017 Elsevier Inc. All rights reserved.
104 M. El Machkouri et al. / Journal of Multivariate Analysis 156 (2017) 103–115

Let p be fixed in [1, ∞]. The strong mixing coefficients (α1,p (n))n>0 associated to {(Yi , Xi ) : i ∈ Zd } are defined by

α1,p (n) = sup {α(σ (Yk , Xk ), FΓ ), k ∈ Zd , Γ ⊂ Zd , |Γ | 6 p, ρ(Γ , {k}) > n},

where FΓ = σ (Yi , Xi ; i ∈ Γ ), |Γ | is the number of element in Γ and the distance ρ is defined for any subsets Γ1 and Γ2 of
Zd by ρ(Γ1 , Γ2 ) = min{|i − j|, i ∈ Γ1 , j ∈ Γ2 } with |i − j| = max16s6d |is − js | for any i = (i1 , . . . , id ) and j = (j1 , . . . , jd ) in
Zd . We say that the random field (Yi , Xi )i∈Zd is strongly mixing if limn→∞ α1,p (n) = 0.
Fix x ∈ R. Assuming that g is differentiable at x with derivate g ′ (x), the main idea in local linear regression is to
approximate g (z ) by g (x) + g ′ (x)(z − x) for any z in the neighborhood of x and to estimate (g (x), g ′ (x)) instead of using a
classical nonparametric estimation method (e.g., Nadaraya–Watson kernel estimator) for g itself. Following [12], we define
⊤
the local linear kernel regression estimator of g (x), g ′ (x)

by

gn (x)
   
 Xi − x
= argmin {Yi − s − t (Xi − x)}2 K , (1)
gn′ (x) bn
(s,t )∈R2 i∈Λn

where bn is the bandwidth parameter going to zero as n → ∞, Λn is a finite subset of Zd whose cardinality |Λn | goes to
infinity as n → ∞, and K is a probability kernel, i.e., a function K : R → R such that R K (s)ds = 1. We introduce the
following notations:

1  
Xi − x

1   Xi − x 2 
Xi − x

u00 (n) = K , u11 (n) = K ,
|Λn |bn i∈Λn
bn |Λn |bn i∈Λn
bn bn

1   Xi − x   Xi − x 
u01 (n) = u10 (n) = K ,
|Λn |bn i∈Λn
bn bn
     
1  Xi − x 1  Xi − x Xi − x
v0 (n) = Yi K , v1 (n) = Yi K ,
|Λn |bn i∈Λn
bn |Λn |bn i∈Λn
bn bn
     
1  Xi − x 1  Xi − x Xi − x
w0 (n) = Zi K and w1 (n) = Zi K
|Λn |bn i∈Λn
bn |Λn |bn i∈Λn
bn bn

with Zi = Yi − g (x) − g ′ (x)(Xi − x). A straightforward calculation gives

gn (x) u00 (n) u10 (n) v 0 ( n)

     
= Un−1 Vn , where Un = and Vn = .
gn′ (x)bn u01 (n) u11 (n) v 1 ( n)
⊤
Denoting Wn = Vn − Un g (x), g ′ (x)bn = (w0 (n), w1 (n))⊤ , we obtain


 ′gn (x) − ′g (x)

 
G(n, x) ≡ = Un−1 Wn . (2)
gn (x) − g (x) bn

This paper’s main contribution is to provide sufficient conditions ensuring the consistency (Theorem 1) and the
asymptotic normality (Theorem 2) of the estimator defined by (1) under mild conditions on the bandwidth parameter; see
Assumption (A6). Our approach is based on the so-called Lindeberg method [8–10,14] instead of Bernstein’s blocking method
used in several previous works for proving limit theorems in the random field setting [2,12,24]. Finally, we emphasize that
our work can be extended without any effort to any (RN × R)-valued random field {(Yi , Xi ) : i ∈ Zd } with arbitrary positive
integer N. The present discussion is limited to the case N = 1 only to simplify notation.
Let K : R → R be a probability kernel. For any c = (c0 , c1 ) ∈ R2 and any s in R, we define Kc (s) = (c0 + c1 s)K (s). In the
sequel, we consider the following assumptions:

(A1) For any c in R2 , we have supt ∈R |Kc (t )| < ∞, R |Kc (t )|dt < ∞ and Kc has an integrable second-order radial upper


bound, i.e., the function ψ defined for any real x by r (x) = sup|t |>|x| t 2 Kc (t ) is integrable.
(A2) g is twice differentiable and g ′′ is continuous.
(A3) There exists a positive constant κ such that supk̸=0 |f0,k (x, y) − f (x)f (y)| ≤ κ for any (x, y) in R2 , where f0,k is the
continuous joint density of (X0 , Xk ) and f is the continuous marginal density of X0 .
(A4) E |Y0 |2+δ < ∞ for some δ > 0.
 

(A5) bn → 0 in such a way that |Λn |b3n → ∞.

(A6) bn → 0 in such a way that |Λn |bn → ∞ and |Λn |b5n → 0.

Our first main result ensures the consistency of the estimator.

 M. El Machkouri et al. / Journal of Multivariate Analysis 156 (2017) 103–115 105

Theorem 1. If (A1)–(A5) hold and

∞ δ
 (2d−1)δ+6d−2
m 2+δ α12,∞
+δ
(m) < ∞, (3)
m=1

then for any x ∈ R,

Pr
G(n, x)/bn −−−→ 0, (4)
n→∞

where G(n, x) is defined by (2).

The second main contribution of this paper is a new central limit theorem.

Theorem 2. If (A1)–(A4), (A6) and (3) hold, then for any x ∈ R such that f (x) > 0,
D
|Λn |bn G(n, x) −−−→ N [0, U −1 Σ (U −1 )⊤ ],

n→∞

where
  
K 2 (t )dt tK 2 (t )dt
Σ = var(Y0 |X0 = x)f (x) 
 R 
 R 
tK 2 (t )dt t 2 K 2 (t )dt

R R

and
  
1 tK (t )dt
U = f ( x)  .
 R

  (5)
tK (t )dt t 2 K (t ) dt

R R

Theorem 2 improves Theorem 3.1 in Hallin et al. [12] since the mixing condition (3) is less restrictive than the assumption
α1,∞ (m) = O m−µ where µ > 2(3 + δ)d/δ . Moreover, the regions Λn that we consider in our work are general (Λn


is a finite subset of Zd such that |Λn | → ∞) whereas, in [12], these regions are assumed to be rectangular in order to
enforce Bernstein’s blocking technique. Finally, simple conditions on the bandwidth parameter are assumed compared to
Assumptions (B2), (B2′ ) and (B3) in [12], where conditions on the bandwidth parameter and the strong mixing coefficient are
interlaced in a complicated way. Moreover, the condition |Λn |b5n → 0 in our assumption (A6) is also a necessary condition
in Theorem 3.1 in [12]. We elaborate below on the latter point.
Using the notations in [12], we have

 ′gn (x) − ′g (x)

 
Hn ≡ = Un−1 Wn .
gn (x) − g (x) bn
At the end of the proof of Theorem 3.1 in [12], the authors find

Hn − U −1 E (Wn ) = Un−1 {Wn − E (Wn )} + (Un−1 − U −1 )E (Wn ) (6)

and they complete the proof of their result by applying Lemmas 2.1, 2.3 and 3.1 in their paper. But, using their notations,
from (6), we have
   
n̂bdn Hn − U −1 n̂bdn E (Wn ) = Un−1 n̂bdn {Wn − E (Wn )} + (Un−1 − U −1 ) n̂bdn E (Wn ) .

From Lemma 2.3 in [12], for any c = (c0 , c1⊤ )⊤ ∈ R1+d ,

  
n̂bdn E (cWn ) = n̂bdn b2n c0 B0 (x) + c1⊤ B1 (x) + o( n̂bdn b2n ),
 

where B0 (x) and B1 (x) are defined in Lemma 2.3 in [12]. So, we obtain
 
n̂bdn Hn − U −1 n̂bdn b2n (B0 (x), B1 (x))⊤
  
= Un−1 n̂bdn {Wn − E (Wn )} + (Un−1 − U −1 ) n̂bdn E (Wn ) + o( n̂bdn b2n ).
The asymptotic normality of
 
n̂bdn Hn − U −1 n̂bdn b2n (B0 (x), B1 (x))⊤
106 M. El Machkouri et al. / Journal of Multivariate Analysis 156 (2017) 103–115

n̂bdn {Wn − E (Wn )} (see Lemma 3.1 in [12]) if the term

−1

will follow from that of Un
 
(Un−1 − U −1 ) n̂bdn E (Wn ) + o( n̂bdn b2n ) (7)

is shown to be negligible. However, the convergence in probability of Un to U is not sufficient to ensure the negligibility of
the term (7). In fact, it is necessary to assume n̂bdn+4 → 0, which correspond to |Λn |b5n → 0 in our work. So, assumption
(A6) in Theorem 2 is less restrictive than conditions (B1), (B2), (B2′ ) and (B3) in [12]. More precisely, (B1) corresponds to
the condition bn → 0 and |Λn |bn → ∞ in our paper whereas conditions (B2), (B2′ ) and (B3) are not necessary for the
asymptotic normality of the estimator in our work.

2. Proofs

In the sequel, for any sequences (pn )n>1 and (qn )n>1 of real numbers, we denote pn E qn if and only if there exists κ > 0
(not depending on n) such that pn 6 κ qn . Moreover, proofs of some technical lemmas in this section are postponed to the
Appendix. Consider the sequence (mn )n>1 of positive integers defined by
 
 −3δ   3d1  
δ

d(4+δ)
mn = max τn ,  bn |i| 2+δ α12,∞ (|i|)  + 1 ,
 4+δ +δ

(8)
|i|>τn

−δ
2d(4+δ)
where τn = ⌊bn ⌋ and ⌊·⌋ denotes the integer part function. The proof of the following lemma is left to the reader; see
Lemma 2 in [8].

Lemma 1. If (3) holds, then

δ
 δ
− 42+δ δ
+δ  d(4+δ)
mn → ∞, mdn bn4+δ → 0 and mdn bn4+δ |i| 2+δ α12,∞
+δ
(|i|) → 0.
|i|>mn

2.1. Proof of Theorem 1

Let x and c = (c0 , c1 ) be fixed in R and R2 , respectively. Denote


η = var(Y0 |X0 = x)f (x) Kc2 (t )dt .
R

Lemma 2. E(cWn ) E b2n and |Λn |bn var(cWn ) → η as n → ∞.

Proof of Lemma 2. For any positive integer n, we have
  
1 X0 − x
E (cWn ) = E Z 0 Kc
bn bn
  
1 X0 − x
g (X0 ) − g (x) − g ′ (x)(X0 − x) Kc
 
= E
bn bn
  
1 u−x
g (u) − g (x) − g ′ (x)(u − x) Kc f (u)du
 
=
bn R bn

g (x + v bn ) − g (x) − g ′ (x)v bn Kc (v)f (x + v bn )dv
 
=
R
b2n

= g ′′ (θn (x, v))v 2 Kc (v)f (x + v bn )dv,
2 R

where θn (x, v) is a real number between x and x + v bn . By the Lebesgue density theorem (see Chapter 2 in [6]), we have
 
g ′′ {θn (x, v)}v 2 Kc (v)f (x + v bn )dv −−−→ g ′′ (x)f (x) v 2 Kc (v)dv.
R n→∞ R

So, we obtain E (cWn ) E b2n . Furthermore,

  2 
 
1  
 |Λn |bn var(cWn ) − E(∆2 ) =  ∆i − E(∆0 ) 6 |E(∆0 ∆j )|,
   2 

0  |Λ | E (9)
 n i∈Λn  j∈Zd \{0}
 M. El Machkouri et al. / Journal of Multivariate Analysis 156 (2017) 103–115 107

where
    
Zi Xi − x Zi Xi − x
∆i = √ Kc − E √ Kc . (10)
bn bn bn bn

Next, we call on the following auxiliary result.

Lemma 3. E ∆20 → η as n → ∞. In addition,

 

δ −δ δ
sup E (|∆0 ∆i |) E bn4+δ and |E (∆0 ∆i ) | E bn2+δ α12,+δ
1 (|i|) for any i ̸= 0.
i∈Zd \{0}

Combining Lemmas 3 and 1, we obtain

δ
 δ
− 42+δ δ
 +δ  d(4+δ)
|E(∆0 ∆j )| E mdn bn4+δ + mdn bn4+δ |i| 2+δ α12,∞
+δ
(|i|) −−−→ 0.
n→∞
j∈Zd \{0} i∈Zd
|i|>mn

Using (9) and Lemma 3, we deduce that |Λn |bn var(cWn ) → η as n → ∞. Thus the proof of Lemma 2 is complete. 

L2
Lemma 4. Un −−−→ U, where U is defined by (5).
n→∞

Proof of Lemma 4. Let k be fixed in {0, 1, 2} and fix x ∈ R. Then,

 
1   Xi − x k 
Xi − x

1  1 X0 − x
k 
X0 − x
K = √ ∆i,k + E K ,
|Λn |bn i∈Λn
bn bn |Λn | bn i∈Λn
bn bn bn

where
 k    k  
1 Xi − x Xi − x 1 X0 − x X0 − x
∆i,k = √ K −√ E K .
bn bn bn bn bn bn

First, using again the Lebesgue density theorem, we have

 k    
1 X0 − x X0 − x
E K = v K (v)f (x + v bn )dv −−−→ f (x)
k
v k K (v)dv. (11)
bn bn bn R n→∞ R

2
Next, arguing as in the proof of Lemma 3, we have E(∆0,k ) converges to f (x) t 2k K 2 (t )dt as n

R
→ ∞ and
E ∆0,k ∆j,k  goes to zero as n → ∞. Consequently,
   
j∈Zd \{0}

 
2

E ∆0,k 
  2
 1
6 1
   
∆ E ∆0,k ∆j,k  −−−→ 0.


 |Λ |2 b E i, k − (12)
 n n i∈Λn
| Λ |b
n n 
 |Λ |b
n n n→∞
j∈Zd \{0}

Combining (11) and (12) and keeping in mind that |Λn |bn goes to infinity as n → ∞, we obtain

1   Xi − x k 
Xi − x

L2

K −−−→ f (x) v k K (v)dv.
|Λn |bn i∈Λn
bn bn n→∞ R

Thus the proof of Lemma 4 is complete. 

Combining (2) with Lemmas 2 and 4 and the fact that

√
G(n, x) |Λn |bn {Wn − E (Wn )} E (Wn )
= Un −1
+ Un−1 , (13)
|Λn |bn

bn 3 bn

we obtain (4). Hence the proof of Theorem 1 is complete. 

108 M. El Machkouri et al. / Journal of Multivariate Analysis 156 (2017) 103–115

2.2. Proof of Theorem 2

√
Let c = (c0 , c1 ) be fixed in R2 . By Lemma 2 and Assumption (A6), we derive that |Λn |bn E(cWn ) goes to zero as
√→ ∞. Keeping in mind (13) and using Lemma 4 and Slutsky’s lemma, we only need to prove the asymptotic normality of
n
|Λn |bn {Wn − E (Wn )}. That is what we establish in the following key result, where we recall the notation η = var(Y0 |X0 =
x)f (x) R Kc2 (t )dt.
√ D
Proposition 1. |Λn |bn {cWn − E (cWn )} −→ N (0, η) as n → ∞.
Proof of Proposition 1. Let (ξi )i∈Zd be a field of i.i.d. standard normal random variables independent of (Yi , Xi )i∈Zd and
denote for all i in Zd ,
√
∆i η ξi
Ti = and γi = ,
|Λn |1/2 |Λn |1/2
where ∆i is given by (10). On the lattice Zd we define the lexicographic order as follows: if i = (i1 , . . . , id ) and j = (j1 , . . . , jd )
are distinct elements of Zd , the notation i <lex j means that either i1 < j1 or for some k in {2, . . . , d}, ik < jk and iℓ = jℓ for
1 6 ℓ < k. Recall that |Λn | is the number of element in the region Λn and let ϕ be the unique function from {1, . . . , |Λn |}
to Λn such that ϕ(k) <lex ϕ(ℓ) for 1 6 k < ℓ 6 |Λn |. For all integer 1 6 k 6 |Λn |, we put
k
 |Λn |

Sϕ(k) (T ) = Tϕ(i) c
and Sϕ(k) (γ ) = γϕ(i)
i =1 i=k

with the convention Sϕ(0) (T ) = Sϕ(|

c
Λn |+1) (γ ) = 0. Let h be any measurable function from R to R. For any 1 6 k 6 ℓ 6 |Λn |,
we introduce hk,ℓ = h{Sϕ(k) (T ) + Sϕ(ℓ)
c
(γ )}. We denote by B41 (R) the unit ball of Cb4 (R): h belongs to B41 (R) if and only if it
belongs to C 4 (R) and satisfies max06i64 ∥h(i) ∥∞ 6 1. It suffices to prove that for all h in B41 (R),
 √ 
E h Sϕ(|Λn |) (T ) −−−→ E h ηξ0 .
  
n→∞

We use Lindeberg’s decomposition, viz.

|Λn |
√  
E h Sϕ(|Λn |) (T ) − h ηξ0 = E hk,k+1 − hk−1,k .
    
k=1

Now, we have hk,k+1 − hk−1,k = hk,k+1 − hk−1,k+1 + hk−1,k+1 − hk−1,k and by Taylor’s formula we obtain
1
hk,k+1 − hk−1,k+1 = Tϕ(k) h′k−1,k+1 + k) hk−1,k+1 + Rk ,
2 ′′
Tϕ(
2
1
hk−1,k+1 − hk−1,k = −γϕ(k) h′k−1,k+1 − γϕ(
2
k) hk−1,k+1 + rk ,
′′
2

k) (1 ∧ |Tϕ(k) |) and |rk | 6 γϕ(k) (1 ∧ |γϕ(k) |). Since (T , ξi )i̸=ϕ(k) is independent of ξϕ(k) , it follows that
2 2
where |Rk | 6 Tϕ(

η ′′
 
E γϕ(k) h′k−1,k+1 = 0 E γϕ(
2 ′′
hk−1,k+1 .
   
and k) hk−1,k+1 = E
|Λn |
Hence, we obtain
|Λn | |Λn | |Λn |
η h′′k−1,k+1
  
√    
E h(Sϕ(|Λn |) (T )) − h ηξ0 = E(Tϕ(k) h′k−1,k+1 ) + 2
E (Rk + rk ) .

E Tϕ(k) − +
k=1 k=1
|Λn | 2 k=1

Let L be a positive real number. Then

|Λn |
|∆0 |
   
E|Rk | 6 E ∆0 1 ∧
2

k=1
|Λn |1/2
 2     
Z0 2 X0 − x |Z0 |  X0 − x 
EE Kc 1∧ √ K c

bn bn |Λn |bn  bn 
 2     
Z0 X0 − x |Z0 |  X0 − x 
EE 1|Z0 |6L Kc2 1∧ √ K c

bn bn |Λn |bn  bn 
 2     
Z X0 − x |Z0 |  X0 − x 
+ E 0 1|Z0 |>L Kc2 1∧ √ K c

b n b n |Λ |b  n b n

n
 M. El Machkouri et al. / Journal of Multivariate Analysis 156 (2017) 103–115 109
  3 
L3 |Z0 |2+δ 2 X0 − x
  
 X0 − x  −δ
E √ E Kc  +L E Kc
 
3/2
|Λn |bn bn bn bn

L3
 
|Kc (v)|3 f (x + v bn )dv + L−δ E |Z0 |2+δ |X0 = x + v bn |Kc (v)|3 f (x + v bn )dv.
 
E √
|Λn |bn R R

By the Lebesgue density theorem, we have

 
E |Z0 |2+δ |X0 = x + v bn |Kc (v)|3 f (x + v bn )dv −−−→ f (x)E |Z0 |2+δ |X0 = x |Kc (v)|3 dv
   
R n→∞ R

and
 
|Kc (v)| f (x + v bn )dv −−−→ f (x)
3
|Kc (v)|3 dv.
R n→∞ R

Consequently, we obtain
|Λn |
L3
  
E (|Rk |) E √ + L−δ .
k =1
|Λn |bn
1
Choosing L = (|Λn |bn ) 2(3+δ) , we deduce that
|Λn |
 −δ
E (|Rk |) E (|Λn |bn ) 2(3+δ) −−−→ 0.
n→∞
k=1

Moreover,
|Λn |
η3/2 E |ξ0 |3
 

E (|rk |) 6 √ −−−→ 0.
k=1
|Λn | n→∞

So, it is sufficient to show

|Λn |
η h′′k−1,k+1
   
lim E(Tϕ(k) h′k−1,k+1 ) + E 2
Tϕ(k) − = 0. (14)
n→∞
k=1
|Λn | 2

For any i ∈ Zd and any integer k > 1, we define Vik = {j ∈ Zd | j <lex i and |i − j| > k}. For every integers n > 1 and
1 6 k 6 |Λn |, we denote
(n) (m )

Ek = ϕ({1, . . . , k}) ∩ Vϕ(
mn
k) and Sϕ(kn) (T ) = Ti ,
(n)
i∈Ek

(m ) (m )
where mn is defined by (8). In the sequel, for all function ψ from R to R, we adopt the notation ψk−1n ,ℓ = ψ{Sϕ(kn) (T ) +
c
Sϕ(ℓ) (γ )}. More precisely, we are going to use this notation with ψ either equal to h′ or h′′ . Our aim is to show that
|Λn |
(m )

lim E[Tϕ(k) h′k−1,k+1 − Tϕ(k) {Sϕ(k−1) (T ) − Sϕ(kn) (T )}h′′k−1,k+1 ] = 0. (15)
n→∞
k=1

First, we use the decomposition

′(m ) ′(m )
Tϕ(k) h′k−1,k+1 = Tϕ(k) hk−1n,k+1 + Tϕ(k) (h′k−1,k+1 − hk−1n,k+1 ).

Since γ is independent of T , we have E[Tϕ(k) h′ {Sϕ(

c
k+1) (γ )}] = 0 and consequently,
      
′(m ) ′(m ) (mn )
E Tϕ(k) hk−1n,k+1  = E[Tϕ(k) [hk−1n,k+1 − h′ {Sϕ(k+1) (γ )}]] 6 E Tϕ(k) Sϕ(k) (T ) .
   c   

Moreover,
1

(m )
 
E Tϕ(k) Sϕ(kn) (T ) 6 E |∆ϕ(k) ∆i | 6 sup E |∆0 ∆j | .
     
|Λn | (n) j∈Zd \{0}
i∈Ek

So, by Lemma 3, we obtain

  δ
′(m )

E Tϕ(k) hk−1n,k+1  E bn4+δ −−−→ 0.
 
n→∞
110 M. El Machkouri et al. / Journal of Multivariate Analysis 156 (2017) 103–115

Applying again Taylor’s formula,

′(m ) (m )
Tϕ(k) (h′k−1,k+1 − hk−1n,k+1 ) = Tϕ(k) {Sϕ(k−1) (T ) − Sϕ(kn) (T )}h′′k−1,k+1 + R′k ,

where
 
(mn ) (mn )
|R′k | 6 2 Tϕ(k) {Sϕ(k−1) (T ) − Sϕ( (T )}(1 ∧ | S ( T ) − S (T )|).
 
k) ϕ(k − 1) ϕ(k)

Using Lemmas 1 and 3, it follows that

   
|Λn |  
1
  
 
E(|Rk |) 6 2E |∆0 | 
′
|∆i | 1 ∧ √ |∆i |
 
k=1

 |i|6mn |Λn | |i|6mn 

i̸=0 i̸=0

62 E (|∆0 ∆i |) 6 2mdn sup E (|∆0 ∆i |)
|i|6mn i∈Zd \{0}
i̸=0

δ
E mdn bn
4+δ
−−−→ 0.
n→∞

So, we obtain (15). In order to derive (14) it remains to control

|Λn |
  2

 Tϕ(k) (m ) η
F1 := E h′′k−1,k+1 + Tϕ(k) {Sϕ(k−1) (T ) − Sϕ(kn) (T )} − .
k=1
2 2|Λn |

We have
|Λn |
  
1 
     
′′
∆ − E(∆0 )  + η − E ∆20  + 2 E |∆0 ∆j | .
 2 2
  
F1 6 E h

 |Λn | k=1 k−1,k+1 ϕ(k)  |j|6mn
j̸=0

By Lemma 3, we know that

 δ
E ∆20 −−−→ η E |∆0 ∆j | E mdn bn4+δ −−−→ 0.
   
and
n→∞ n→∞
|j|6mn
j̸=0

So, it suffices to prove

|Λn |
  
1 
 
h′′k−1,k+1 ∆2ϕ(k) − E(∆20 )  −−−→ 0.
 
F2 := E (16)
 
 |Λn | k=1  n→∞

Let M be a positive constant and denote EM (∆2ϕ(k) ) = E(∆2ϕ(k) |FV M ), where FV M is the σ -algebra generated by (Xs , Ys ) for
ϕ(k) ϕ(k)
M ′ ′′
s in Vϕ(k) . We have F2 6 F2 + F2 where

|Λn |
  
1
  
F2′ ≡ E h′′k−1,k+1 ∆2ϕ(k) − EM ∆2ϕ(k)
    
 |Λn | k=1



and
|Λn |
  
1 
  
′′ ′′
E ∆ − E(∆0 )  .
  2  2
F2 ≡ E h

 |Λn | k=1 k−1,k+1 M ϕ(k) 

The following technical lemma is proved in the Appendix.

−δ
Lemma 5. ∥∆0 ∥22+δ E bn2+δ .

The next result can be found in [19].

Lemma 6. Let U and V be two σ -algebras and let X be a random variable which is measurable with respect to U. If 1 6 p 6
r 6 ∞, then
1 1
∥E(X |V ) − E(X )∥p 6 2(21/p + 1) {α(U, V )} p − r ∥X ∥r .
 M. El Machkouri et al. / Journal of Multivariate Analysis 156 (2017) 103–115 111

Using Lemmas 5 and 6 with p = 1 and r = (2 + δ)/2, we deduce that

δ −δ δ
F2′′ 6 ∥EM ∆20 − E(∆20 )∥1 6 6∥∆0 ∥22+δ α12,∞ (M ) 6 6 bn2+δ α12,∞ (M ).
  +δ +δ
(17)

Furthermore,
|Λn |
1 
F2′ 6 {Jk1 (M ) + Jk2 (M )},
|Λn | k=1

where
 
′′(M ) 
Jk1 (M ) = E hk−1,k+1 ∆2ϕ(k) − EM ∆2ϕ(k)
  


and
 
′′(M )
 
Jk2 (M ) = E h′′k−1,k+1 − hk−1,k+1 ∆2ϕ(k) − EM ∆2ϕ(k)  .
 

′′(M )
Since hk−1,k+1 is σ (γi ; i ∈ Zd ) ∨ FV M -measurable and (γi )i∈Zd is independent of (Yi , Xi )i∈Zd then Jk1 (M ) = 0. Moreover, if L
ϕ(k)
is a positive real number then
  
|∆i |  2 
 
L
  
Jk2 (M ) 6 E 2 ∧  ∆0 6 √ E (|∆0 ∆i |) + 2E ∆20 1|∆0 |>L
 
√


 |i|<M |Λn | 
 |Λn | |i|<M
i̸=0 i̸=0

d
M L
sup E (|∆0 ∆i |) + 2L−δ E |∆0 |2+δ .
 
6 √
|Λn | i∈Zd \{0}
Applying again Lemmas 3 and 5, we derive
δ
LM d bn4+δ
Jk2 (M ) 6 √ + 2L−δ b−δ/
n
2
.
|Λn |
In particular, for
1
|Λn | 2(1+δ)
L=
δ 2 +6δ
d 2(4+δ)(1+δ)
M 1+δ bn
we get
dδ
M 1+δ
Jk2 (M ) E .
δ −δ 2 +4δ
|Λ | 2(1+δ) b 2(4+δ)(1+δ)
n n

(2d−1)δ+6d−2 −δ
Now, choosing M such that M 2+δ = bn2+δ , we find
bθn
Jk2 (M ) E δ
,
(|Λn |bn ) 2(1+δ)
where
dδ 3 (4 + δ) + δτ 2
θ= and τ = (2d − 1)δ + 6d − 2.
(1 + δ)(4 + δ)τ 2
We can then conclude that F1′ → 0 as n → ∞. Using (3) and (17), we also deduce that

(2d−1)δ+6d−2 δ
F2′′ E M 2+δ α12,∞
+δ
(M ) −−−→ 0.
n→∞

Consequently, we obtain (16), and the proof of Proposition 1 is complete. 

Combining (2) with Lemmas 2 and 4 and Proposition 1, we obtain Theorem 2. 

112 M. El Machkouri et al. / Journal of Multivariate Analysis 156 (2017) 103–115

Table 1
m̂(s) decreases to zero when s increases.
s |Λs | = s2 bs = |Λs |−1/3 m̂(s)

10 100 0.215 −0.408

20 400 0.136 −0.309
30 600 0.104 −0.271
40 1600 0.085 −0.114

3. Numerical results

In this section, we consider the autoregressive random field (Xi,j )(i,j)∈Z2 defined by

Xi,j = 0.75Xi−1,j + 0.2Xi,j−1 + εi,j , (18)

where (εi,j )(i,j)∈Z2 are i.i.d. random variables with standard normal law. From [13], we know that (18) has a stationary
solution Xi,j given by
   k1 + k2 
Xi,j = (0.75)k1 (0.2)k2 εi−k1 ,j−k2 . (19)
k1 >0 k2 >0
k1

We can also check that Xi,j has a normal law with zero mean and variance σ 2 = 3.8346. If we denote by f the density of Xi,j
then f (0) = 0.2037. Let s be a positive integer. We simulate the εi,j ’s over the grid [0, 2s]2 ∩ Z2 and we obtain the data Xi,j
for (i, j) in Λs = [s + 1, 2s]2 ∩ Z2 following (19). Thus, we construct
1  
Xi,j

1   Xi,j 2 
Xi,j

u00 (s) = K , u11 (s) = K
s2 bs (i,j)∈Λ bs s2 bs (i,j)∈Λ bs bs
s s

and
1   Xi,j   Xi,j 
u01 (s) = u10 (s) = K ,
s2 bs (i,j)∈Λ bs bs
s

where K is the Gaussian kernel defined for any real u by K (u) = √1 exp(−u2 /2). From the data set
2π

2
Yi,j = + εi,j
1 + Xi2,j

i.e., g (u) = 2/(1 + u2 ) for any u ∈ R, we define also

     
1  X i ,j 1  Xi,j Xi,j
w0 (s) = Zi,j K and w1 (s) = Zi,j K
s2 bs (i,j)∈Λ bs s2 bs (i,j)∈Λ bs bs
s s

with Zi,j = g (Xi,j ) + εi,j − 2. The local linear estimator Ĝ(s, 0) of the regression function g at the point x = 0 is given by
−1 
u00 (s) u10 (s) w0 (s) τ̂ (s)
   
Ĝ(s, 0) = ≡ 0 .
u01 (s) u11 (s) w1 (s) τ̂1 (s)
For s ∈ {10, 20, 30, 40} and bs = |Λs |−1/3 , we take the arithmetic mean value m̂(s) of 300 replications of
τ̂0 (s) + τ̂1 (s)
bs
and we collect the results in Table 1. In order to illustrate the asymptotic normality of the estimator, we consider 300
replications of
√
2 × π 1/4 |Λs |bs {τ̂0 (s) + τ̂1 (s)} 
√ = 1.703 × s2 × bs {τ̂0 (s) + τ̂1 (s)}
3 × f (0) × var(Y0 |X0 = 0)
for (s, bs ) ∈ {(10, 0.215); (20, 0.136); (30, 0.104); (40, 0.085)} and we obtain histograms (see Fig. 1) which fit well to the
standard normal law N (0, 1).

Acknowledgments

The authors would like to thank the Editor-in-Chief and an anonymous referee for valuable comments which helped to
improve the manuscript. The first author was supported by Erasmus Mundus (Battuta project BT14AC0224) funded from
the European commission. He would like to thank the hospitality of Cadi Ayyad University of Marrakesh where a large part
of this work was done.
 M. El Machkouri et al. / Journal of Multivariate Analysis 156 (2017) 103–115 113

Fig. 1. Asymptotic normality of the local linear estimator.

Appendix

Proof of Lemma 3. Let i in Zd \ {0} be fixed. Applying Rio’s covariance inequality [20, Theorem 1.1], we obtain
 α1,1 (|i|)
|E (∆0 ∆i ) | = |cov(∆0 , ∆i )| 6 4 Q∆2 0 (u)du,
0
−1
where Q∆0 is defined by Q∆0 (u) = inf{t > 0 : Pr(|∆0 | > t ) 6 u} for any u ∈ [0, 1]. Since Q∆0 (u) 6 u 2+δ ∥∆0 ∥2+δ , using
Lemma 5, we get
−δ δ
|E (∆0 ∆i )| E bn2+δ α12,+δ
1 (|i|).

(1)
In addition, let L > 1 be a fixed real number. Arguing as in the proof of Lemma 5.2 in [12], we denote ∆j = ∆j 1|∆j |6L and
(2)
∆j = ∆j 1|∆j |>L for any j in Zd . So, we have
(1) (1) (1) (2) (2) (1) (2) (2)
E (|∆0 ∆i |) 6 E(|∆0 ∆i |) + E(|∆0 ∆i |) + E(|∆0 ∆i |) + E(|∆0 ∆i |).

Moreover,
   
Z02 Z02
   
(1) (2) X0 − x X0 − x
sup E(|∆0 ∆i |) E E 1|Z0 |6L Kc2 × E 1|Z0 |>L Kc2
i∈Zd \{0} bn bn bn bn
   
Z02 |Z0 |2+δ
   
−δ/2 X0 − x X0 − x
EL E Kc2 × E Kc2
bn bn bn bn

E L−δ/2 E |Z0 |2 |X0 = x + v bn Kc2 (v)f (x + v bn )dv
 
R

E |Z0 |2+δ |X0 = x + v bn Kc2 (v)f (x + v bn )dv.
 
×
R
114 M. El Machkouri et al. / Journal of Multivariate Analysis 156 (2017) 103–115

Applying again the Lebesgue density theorem, we get

(1) (2)
sup E(|∆0 ∆i |) E L−δ/2 .
i∈Zd \{0}

(2) (1) (2) (2)

Similarly, supi∈Zd \{0} E(|∆0 ∆i |) E L−δ/2 and supi∈Zd \{0} E(|∆0 ∆i |) E L−δ . Finally,
       
(1) (1) |Z0 |
Kc X0 − x  1|Z |6L × √ |Zi |  Xi − x 
sup E(|∆0 ∆i |) 6 E
 
√ K c
 1|Z |6L
bn  bn  0 bn  bn  i
i∈Zd \{0}
     2
|Z0 |  X0 − x 
 1|Z |6L
+ 3 E √ Kc  0
bn bn
2
L2
     
X0 − x Xi − x 
(v)| ( v ) v

6 E Kc Kc  + 3L2 bn |K c f x + b n d
bn bn bn 
R
2  
v
    
L u−x − x 
E  Kc Kc  |f0,i (u, v) − f (u)f (v)|dudv
bn R  bn bn 

Kc u − x Kc v − x  f (u)f (v)dudv + L2 bn

2
L2
      
(v)| ( v ) v .
 
+ | K c f x + bn d
bn 
R bn  bn R

Using Assumption (A3) and the Lebesgue density theorem, we get

2 2
L2 L2
 
(1) (1)
sup E(|∆0 ∆i |) E |Kc (v)|bn dv + |Kc (v)|f (x + v bn )bn dv + L2 bn E L2 bn .
i∈Zd \{0} bn R bn R

−2
Consequently, we obtain supi∈Zd \{0} E (|∆0 ∆i |) E L−δ/2 + L2 bn . Choosing L = bn4+δ , it follows that
 

δ
sup E (|∆0 ∆i |) E bn4+δ .
i∈Zd \{0}

Furthermore, we have
2
Z02
     
X0 − x Z0 X0 − x
E(∆20 ) = E Kc2 − E √ Kc
bn bn bn bn
and
     
X0 − x X0 − x
Y0 − g (x) − g ′ (x)(X0 − x) Kc
 
E Z0 Kc =E
bn bn
  
X0 − x
= E g (X0 ) − g (x) − g (x)(X0 − x) Kc ′
 
bn
  
u−x
g (u) − g (x) − g (x)(u − x) Kc
′
f (u) du
 
=
R bn

g (x + v bn ) − g (x) − g ′ (x)v bn Kc (v)f (x + v bn ) dv
 
= bn
R
b3n

= g ′′ {θn (x, v)}v 2 Kc (v)f (x + v bn )dv,
2 R

where θn (x, v) is a real number between x and x + v bn . Using the Lebesgue density theorem, we obtain
  
X0 − x
E Z0 Kc E b3n .
bn
Furthermore,
   
1 X0 − x
E Z02 Kc2 = E(Z02 |X0 = x + v bn )Kc2 (v)f (x + v bn )dv.
bn bn R

Noting that

E(Z02 |X0 = x + v bn ) = var(Y0 |X0 = x + v bn ) −−−→ var(Y0 |X0 = x)

n→∞
 M. El Machkouri et al. / Journal of Multivariate Analysis 156 (2017) 103–115 115

and applying again the Lebesgue density theorem, we conclude that E ∆20 → η as n → ∞. The proof of Lemma 3 is
 
complete.
Proof of Lemma 5. We have
  2
 2+δ  2+δ
 Z0 X0 − x 
∥∆0 ∥22+δ E E  √ Kc 
b b n

n

  2
  2+δ  2+δ
−1
 2+δ
 X0 − x 
E bn E E |Z0 | |X0 Kc
 
b 
n

 2
 2+δ
−δ
E |Z0 |2+δ |X0 = x + v bn |Kc (v)|2+δ f (x + v bn )dv .
2+δ
 
E bn
R

By the Lebesgue density theorem, we obtain

−δ
∥∆0 ∥22+δ E bn2+δ .
The proof of Lemma 5 is complete. 

References

[1] G. Biau, B. Cadre, Nonparametric spatial prediction, Stat. Inference Stoch. Process. 7 (2004) 327–349.
[2] M. Carbon, C. Francq, L.T. Tran, Kernel regression estimation for random fields, J. Statist. Plann. Inference 137 (2007) 778–798.
[3] T.-L. Cheng, H.-C. Ho, X. Lu, A note on asymptotic normality of kernel estimation for linear random fields on Z2 , J. Theoret. Probab. 21 (2008) 267–286.
[4] S. Dabo-Niang, M. Rachdi, A. Yao, Kernel regression estimation for spatial functional random variables, Far East J. Theor. Stat. 37 (2011) 77–113.
[5] S. Dabo-Niang, A.-F. Yao, Kernel regression estimation for continuous spatial processes, Math. Methods Statist. 16 (2007) 298–317.
[6] L. Devroye, L. Györfi, Nonparametric Density Estimation: The L1 View, Wiley, New York, 1985.
[7] M. El Machkouri, Nonparametric regression estimation for random fields in a fixed-design, Stat. Inference Stoch. Process. 10 (2007) 29–47.
[8] M. El Machkouri, Asymptotic normality of the Parzen–Rosenblatt density estimator for strongly mixing random fields, Stat. Inference Stoch. Process.
14 (2011) 73–84.
[9] M. El Machkouri, Kernel density estimation for stationary random fields, ALEA Lat. Am. J. Probab. Math. Stat. 11 (2014) 259–279.
[10] M. El Machkouri, R. Stoica, Asymptotic normality of kernel estimates in a regression model for random fields, J. Nonparametr. Stat. 22 (2010) 955–971.
[11] M. Hallin, Z. Lu, L.T. Tran, Density estimation for spatial linear processes, Bernoulli 7 (2001) 657–668.
[12] M. Hallin, Z. Lu, L.T. Tran, Local linear spatial regression, Ann. Statist. 32 (2004) 2469–2500.
[13] P.M. Kulkarni, Estimation of parameters of a two-dimensional spatial autoregressive model with regression, Statist. Probab. Lett. 15 (1992) 157–162.
[14] J.W. Lindeberg, Eine neue herleitung des exponentialgezetzes in der wahrscheinlichkeitsrechnung, Math. Z. 15 (1922) 211–225.
[15] Z. Lu, X. Chen, Spatial nonparametric regression estimation: Non-isotropic case, Acta Math. Appl. Sin. Engl. Ser. 18 (2002) 641–656.
[16] Z. Lu, X. Chen, Spatial kernel regression estimation: weak consistency, Statist. Probab. Lett. 68 (2004) 125–136.
[17] Z. Lu, P. Cheng, Distribution-free strong consistency for nonparametric kernel regression involving nonlinear time series, J. Statist. Plann. Inference 65
(1997) 67–86.
[18] E. Masry, J. Fan, Local polynomial estimation of regression functions for mixing processes, Scand. J. Stat. 24 (1997) 165–179.
[19] D.L. McLeish, A maximal inequality and dependent strong laws, Ann. Probab. 3 (1975) 829–839.
[20] E. Rio, Covariance inequalities for strongly mixing processes, Ann. Inst. H. Poincaré Probab. Statist. 29 (1993) 587–597.
[21] P.M. Robinson, Nonparametric estimators for time series, J. Time Ser. Anal. 4 (1983) 185–207.
[22] M. Rosenblatt, A central limit theorem and a strong mixing condition, Proc. Natl. Acad. Sci. USA 42 (1956) 43–47.
[23] G. Roussas, Nonparametric estimation in mixing sequences of random variables, J. Statist. Plann. Inference 18 (1988) 135–149.
[24] L.T. Tran, Kernel density estimation on random fields, J. Multivariate Anal. 34 (1990) 37–53.
[25] Y. Wang, M. Woodroofe, On the asymptotic normality of kernel density estimators for causal linear random fields, J. Multivariate Anal. 123 (2014)
201–213.