Spatial Statistics 3 (2013) 120

Contents lists available at SciVerse ScienceDirect

Spatial Statistics
journal homepage: www.elsevier.com/locate/spasta

Disaggregation of spatial autoregressive

processes
Nikolai Leonenko a , Emanuele Taufer b,
a

School of Mathematics, University of Cardiff, Cardiff, United Kingdom

Department of Economics and Management, University of Trento, Trento, Italy

article

info

Article history:
Received 3 August 2012
Accepted 24 January 2013
Available online 4 February 2013
Keywords:
Gegenbauer polynomials
Aggregated random fields
Gaussian random field
Density estimation
-convergence

abstract
An aggregated Gaussian random field, possibly strong-dependent,
is obtained from accumulation of i.i.d. short memory fields via an
unknown mixing density which is to be estimated. The so-called
disaggregation problem is considered, i.e. is estimated from a
sample of the limiting aggregated field while samples of the elementary processes remain unobserved. Estimation of the density is
via its expansion in terms of orthogonal Gegenbauer polynomials.
After defining the estimators, their consistency and convergence
rates are discussed. An example of application to -convergence in
EU GDP per capita is discussed.
2013 Elsevier B.V. All rights reserved.

1. Introduction
Consider a random field defined on a regular rectangular lattice, the so-called doubly-geometric
process, defined by the equation
Ys,t = 1 Ys1,t + 2 Ys,t 1 1 2 Ys1,t 1 + s,t ,

( s, t ) Z 2

(1)

where 1 < 1 , 2 < 1 and s,t , (s, t ) Z is a white noise with null mean and variance . This
process was first studied by Martin (1979) with the aim of providing an easy to use spatial process,
which may serve as a good approximation in many applications.
In this paper we consider the so-called disaggregation problem: starting from independent
individual fields of the form (1) with random coefficients 1 and 2 , a new random field is constructed
by aggregating the individual fields. The aggregating mechanism, due to randomness of the s
coefficients, originates from an unknown mixing density which is to be estimated. The task here is
2

Corresponding author. Tel.: +39 0461 282368.

E-mail addresses: LeonenkoN@Cardiff.ac.uk (N. Leonenko), Emanuele.Taufer@unitn.it (E. Taufer).

2211-6753/$ see front matter 2013 Elsevier B.V. All rights reserved.
doi:10.1016/j.spasta.2013.01.001

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

to accomplish the estimation procedure having availability of data from an aggregated field. This is a
typical situation in practical problems where an observed phenomenon may arise as a contribution of
a myriad of micro-phenomena.
(j)
To be more precise, consider a sequence of independent random fields Ys,t , (s, t ) Z2 , j = 1,
2, . . . , defined as
(j)

(j) (j)

(j) (j)

(j) (j) (j)

(j)

Ys,t = 1 Ys1,t + 2 Ys,t 1 1 2 Ys1,t 1 + s,t ,

(s, t ) Z2

(2)

where
(j)

I. s,t , j = 1, 2, . . . is a sequence of independent copies of zero mean and unit variance strong white
(j)

noise, i.e. s,t , (s, t ) Z2 are independent in (s, t ).

(j)

(j)

II. (1 , 2 ), j = 1, 2, . . . are independent copies (in j) of a random vector (1 , 2 ) supported in

[1, 1] [1, 1] and satisfying
1

2
1

< ,
(j)

1
1 22

< .

(j)

III. The sequences {1 , 2 }j1 and { (j) }j1 are independent.

(j)
(j)
IV. It is assumed that the distribution of (1 , 2 ) admits a mixture density (1 , 2 ) = 1 (1 )2 (2 )
(independent case) such that

1 (1 )2 (2 )
d1 d2 < .
(1 12 )(1 22 )
N

(3)

(j)

An aggregated field Xs,t = limN N1

j=1 Ys,t , whose convergence properties will be characterized more precisely below, is constructed. In the next section, we will define an estimator of the mixing
density (1 , 2 ) based on data obtained from the aggregated field Xs,t . The technique of estimation
is based on an expansion in terms of orthogonal Gegenbauer polynomials. This approach has been
used by Leipus et al. (2006) and Celov et al. (2010) in the context of random processes. In this paper
we will use a bivariate setting; however the set-up provided lends itself to a straightforward extension to a multi-dimensional setting. It will turn out that the proposed estimator of the mixing density
is consistent while asymptotic normality will not follow given the presence of edge-effects. To avoid
edge-effects one could use unbiased covariance estimators instead, which however have the drawback to be not always positive-definite. Alternatively, tapered covariance estimators could be used;
we do not pursue this approach here as, we will see, our estimation procedure requires the choice
of weighting functions; the necessity of selecting a data taper as well would unduly complicate the
estimation procedure. For further discussion on these issues one can consult Dahlhaus and Knsch
(1987) or Guyon (1982).
A seminal paper on aggregation of processes is that of Granger (1980) where it is also shown that
accumulation of short-memory random processes can lead to long-memory macro phenomena; we
will allow for the possible presence of long-memory in our development. Several papers have been
successively devoted to the subject, among them we mention Davidson (1991, 2002), Zaffaroni (2004),
Leonenko and Taufer (2005), Beran et al. (2010), Davidson and Monticina (2010) in the context of
random processes. Extensions to random fields on a lattice have been considered by various authors,
see Lavancier (2005) for a review and Lavancier (2011) for more recent developments. In our set-up
note that model (1) can be seen as a special case of spatial unilateral AR models as defined by Whittle
(1954), i.e.
Ys,t =

p1
p2

kl Ysk,t l + s,t ,

00 = 0,

k=0 l=0

with p1 = p2 = 1, 11 = 01 10 ; it has also been discussed in connection with the spatial unilateral
first order ARMA model as defined by Basu and Reinsel (1993). More recently Baran and Pap (2009)
consider the simpler case Ys,t = 1 Ys1,t + 2 Ys,t 1 + s,t , (s, t ) Z2 which has a stationary solution

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

in the case |1 | + |2 | < 1. Boissy et al. (2005) extend model (1) to include long-memory by defining a

two-indexes operator, d , to be used in a spatial setting, where d = (d1 , d2 ) and d = (1 B1 )d1 (1

d2
B2 ) , where B1 Ys,t = Ys1,t , B2 Ys,t = Ys,t 1 are the backshift operators. The operator d is defined

by its corresponding power series representation in (z1 , z2 ) that is d Ys,t =

k,l=0

kl Ysk,t l , where

the coefficients are found from the power series expansion of (1 z1 )d1 (1 z2 )d2 in the unit polydisc
1 (0) 1 (0) with 1 (0) = {z C : |z | < 1}. The fractional autoregressive model of the type (1) is

then defined as (B1 , B2 , 1 , 2 )d Ys,t = s,t , (s, t ) Z2 , (d1 , d2 ) ( 21 , 12 ). Recently Ruiz-Medina

(2011, 2012) has introduced a spatial functional process with values in a separable Hilbert space H
satisfying the equation
Ys,t = R + L1 (Ys1,t ) + L2 (Ys,t 1 ) + L3 (Ys1,t 1 ) + s,t ,

( s, t ) Z 2 ,

where s,t is assumed to be a martingale difference satisfying certain conditions, R H and Li

L(H ), i = 1, 2, 3, with L(H ) denoting the space of bounded linear operators on (H ).
See also Besag (1974) for an interesting discussion and examples of fields on a lattice. Estimation
problems related to model (1) are discussed in Culles and Gleeson (1991), Basu and Reinsel (1994),
Bhattacharyya et al. (1996, 1997), Beran et al. (2009), Guo et al. (2009); recently Roknossadati and
Zarepour (2010) have studied M-estimation for model (1) when the innovations have infinite variance.
Note that the widely debated topic on Modifiable Areal Unit Problem (MAUP), where the analysis
of the effect of spatial aggregation on statistical estimates is analyzed, is close but does not coincide
with the topic of the present paper where the concern is to estimate an aggregating mechanism which
may or may not be due to geographical clustering of units. The technique discussed here could indeed
used as an additional tool of analysis which can shed different light on aggregation problems. We will
confront ourselves with this exercise in an example of application. For MAUP we refer the interested
reader to Cressie (1996) and Arbia and Petrarca (2011).
In the next section the estimator will be defined precisely and its convergence properties will
be analyzed. Section 3 will present a small simulation study and an example of application will be
presented in Section 4. Proofs are in the Appendix.
2. Estimating the mixing density
Under assumption s I.IV., for a fixed (s, t ) Z2 , as N ,
(N )

Xs,t =

N
1 (j)
Ys,t
N j =1

almost surely tends to zero. This situation permits to recover ergodicity and a Gaussian limit
(j)
distribution with the same second order characteristics as those of Ys,t . In fact the limit process is
ergodic because it is Gaussian and has an absolutely continuous spectral measure. More precisely,
(N )
due to the Central Limit Theorem, the finite dimensional distributions of the random field Xs,t tends
2
to those of a Gaussian random field Xs,t , (s, t ) Z ; we denote this by

(N ) D

NXs,t Xs,t ,

(s, t ) Z2 , N .

(4)

The field Xs,t (s, t ) Z2 is called an aggregated. It has zero mean and covariance function

(s, t ) = Cov(Xs,t , X0,0 ) = Cov(Ys(,jt) , Y0(,j)0 ) =

|| ||

1s 2t 1 (1 )2 (2 )
d1 d2
(1 12 )(1 22 )

ei[s1 +t 2 ] f (1 , 2 )d1 d2 ,

(5)

where the spectral density

f (1 , 2 ) =

1
4 2

1 (1 )2 (2 )
d1 d2 ,
|1 1 ei1 |2 |1 2 ei2 |2

(1 , 2 ) ( , )2 .

(6)

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

The long-range dependent case, that is

(s,t )Z2

| (s, t )| = , is obtained if and only if

1 (1 )2 (2 )
d1 d2 = .
(1 12 )2 (1 22 )2
(N )

Indeed, a stronger convergence result holds for Xs,t . To formulate it, we will use some results
of Suquet (1992, 1999) which discusses tightness problems in a large class of Banach spaces. We
formulate the result for the case (s, t ) N2 ; the case (s, t ) Z2 can be obtained in an analogous
fashion. For a proof of Theorem 1 see the Appendix.
(N )

Theorem 1. Under conditions I.IV, for every > 0, as N , the sequence of random fields Xs,t ,
(s, t ) N2 weakly converges to the random field Xs,t in a Hilbert space H2 (N2 ). The limiting process Xs,t
is a zero mean Gaussian random field with spectral density given by (6).
The disaggregation problem deals with finding the individual fields (if they exist) of the form
(2) which produce the aggregated field Xs,t with given spectral density (6) or covariance (5). This is
equivalent to finding (1 , 2 ), the mixture density such that (6) or (5) hold, or to estimate (1 , 2 )
using the aggregate observations Xs,t , 1 s n1 , 1 t n2 .
()

To define our estimation procedure, let GK (x), x [1, 1], K N, > 1 denote the normalized
()

1/2 +1/2

Gegenbauer polynomials. In our context, GK (x) = hK

(1972, formula 22.2.3),
hK =

CK

(x) where, see Abramowitz and Stegun

(K + 2 + 1)

2 ,
! K + + 12 + 12

22 K

and
CK (x) =

[
K /2]

(1)k

k=0

(K k + )
k!(K 2k)! ()

(2x)K 2k .

The polynomials Ck (x) can be also defined in terms of their generating function
1

(1 2xt +

t2

Ck (x)t k ,

> 1/2, x [1, 1], t (1, 1),

k=0

and satisfy the recurrence relation

C0 (x) = 1,
CK (x) =

1
K

C1 (x) = 2 x,

[2x(K + 1)CK1 (x) 2(K + 2 2)CK2 (x)].

Let Au = [u, u] and consider the functions wj (j ) = (1 j2 )j , j > 1, j A1 , j = 1, 2. Defining

( , )

( )

( )

GK11,K22 (1 , 2 ) = GK11 (1 )GK22 (2 ), K1 , K2 {0, 1, . . .}, then the set of bivariate Gegenbauer orthog( , )

onal polynomials GK11,K22

K1 ,K2 N

forms a complete orthonormal system in the Hilbert space L2 (A21 ,

w1 (1 )w2 (2 )d1 d2 ) of functions h(1 , 2 ) such that

h2 (1 , 2 )w1 (1 )w2 (2 )d1 d2 < ,

A21

with scalar product

h1 , h2 =

A21

h1 (1 , 2 )h2 (1 , 2 )w1 (1 )w2 (2 )d1 d2 < .

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

A function h L2 (A21 , w1 (1 )w2 (2 )d1 d2 ) admits an expansion in mean square convergent series
of the following form
h(1 , 2 ) =

( , )

hK1 ,K2 GK11,K22 (1 , 2 ),

(7)

K1 =0 K2 =0

( , )

with hK1 ,K2 = h(1 , 2 ), GK11,K22 (1 , 2 ) . Note that by the orthogonality property of the Gegenbauer
polynomials, it holds
( , )

( , )

GK11,K22 (1 , 2 ), GM11,M22 (1 , 2 ) = M11 M22 .

()

Letting gK ,j denote the coefficients of the Gegenbauer polynomial, then

( )

GK11 (1 ) =

K1

( )

( )

gK1 ,1j1 11 ,

GK22 (2 ) =

j1 =0

K2

( )

gK2 ,2j2 22 .

j 2 =0

Exploiting the ideas developed in Leipus et al. (2006) an estimator of the bivariate density 1 (1 )2 (2 )
will be developed in terms of the first few terms of the orthogonal expansion (7). To this end consider
the re-scaled bivariate density

(1 , 2 ) =

1 (1 )2 (2 )
.
(1 12 )1 (1 22 )2

(8)

Note that (1 , 2 ) L2 (A21 , w1 (1 )w2 (2 )d1 d2 ) if

12 (1 )22 (2 )
d1 d2 < ,
(1 12 )1 (1 22 )2

j > 1, j = 1, 2.

(9)

The coefficients K1 ,K2 in the expansion (7) for (1 , 2 ) have the form

( , )
K1 ,K2 = (1 , 2 ), GK11,K22 (1 , 2 )
=

K2
K1

( )

( )

gK1 ,1j1 gK2 ,2j2 (1 , 2 ), 11 22 .

(10)

j1 =0 j2 =0

Using the equality

j +2 j 2
2

11 22 11

j +2

j +2 j 2 +2
2

11 22 + 11
(1 12 )(1 22 )

1 (1 )2 (2 )d1 d2

= (j1 , j2 ) (j1 , j2 + 2) (j1 + 2, j2 ) + (j1 + 2, j2 + 2)

(11)

which follows from (5), we obtain that the function (8) has the Gegenbauer expansion (7) with
coefficients

K1 ,K2 =

K1
K2

( )

( )

gK1 ,1j1 gK2 ,2j2

j1 =0 j2 =0

K1
K2

( )

11 22 (1 )(2 )d1 d2

( )

gK1 ,1j1 gK2 ,2j2 [ (j1 , j2 ) (j1 , j2 + 2) (j1 + 2, j2 ) + (j1 + 2, j2 + 2)].

j1 =0 j2 =0

Having a random sample Xi1 ,i2 , 1 i1 n1 , 1 i2 n2 , the covariance (j1 , j2 ) can simply be
estimated by the sample covariance

n1 ,n2 (j1 , j2 ) =

1
n1 n2

n1 j1 n2 j2

i 1 =1 i 2 =1

Xi1 ,i2 Xi1 +j1 ,i2 +j2

(12)

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

and the natural estimate of the mixture density 1 (1 )2 (2 ) is of the form

n1 ,n2 (1 , 2 ) = n1 ,n2 (1 , 2 )(1 12 )1 (1 22 )2

= (1 12 )1 (1 22 )2

K1 (n1 ) K2 (n2 )

(n ,n ) ( , )
K1 ,1K2 2 GK11,K22 (1 , 2 )

(13)

K1 =0 K2 =0

where
(n ,n )
K1 ,1K2 2 =

K2
K1

( )

( )

gK1 ,1j1 gK2 ,2j2 [ n1 ,n2 (j1 , j2 )

j 1 =0 j 2 =0

n1 ,n2 (j1 , j2 + 2) n1 ,n2 (j1 + 2, j2 ) + n1 ,n2 (j1 + 2, j2 + 2)]

(14)

and K1 (n1 ), K2 (n2 ) are nondecreasing sequences which tend to infinity with special rate.
Some asymptotic results on the estimator are now provided. The data Xj1 ,j2 are observed over a
rectangular grid, i.e. 1 j1 n1 , 1 j2 n2 . Asymptotically we require that each nj , j = 1, 2
be increasing with the overall sample size of the field n1 n2 . Since the devices used in the proofs will
remain the same, to avoid heavier notation we will simply define n1 = n2 = n. In this section, C will
indicate a generic constant, not depending on n, which may change through different formulas. A first
theorem on the asymptotic mean integrated square error (MISE) is provided.
Theorem 2. Let {Xs,t } be the aggregated field defined by (4) with a mixture density satisfying (3) and
j > 1, j = 1, 2 are such that (9) holds. For K1 (n) and K2 (n) satisfying
Kj (n) = j log n,

j > 0, j = 1, 2,

1 + 2 <

1
2 log(1 +

2)

(15)

then
1

lim

E[ n (1 , 2 ) (1 , 2 )]2

(1 12 )1 (1 22 )2

d1 d2 = 0

(16)

with n defined in (13) and (14).

We now state a semi-parametric hypothesis for the mixing density (1 , 2 ) under which the
theorem gives a rate of L2 -convergence of the estimators.
A semi-parametric form for the mixture density (1 , 2 ) = 1 (1 )2 (2 ) is

1 (1 ) = (1 1 )d1,1 (1 + 1 )d1,2 1 (1 ),

2 (2 ) = (1 2 )d2,1 (1 + 2 )d2,2 2 (2 ),

(17)

where di,j > 0, i, j = 1, 2, and 1 , 2 are continuous functions on (1, 1) and do not vanish at 1.
The long-range dependent case is obtained when d1,1 or d1,2 and d2,1 or d2,2 are less than 1.
Theorem 3. Let the mixture density have the form (17) with i having continuous derivative on [1, 1]
and does not vanish at 1. Then, for any i > 1 such that min(di,1 , di,2 ) > i /2 + 3/4, i = 1, 2 and
K1 (n) and K2 (n) satisfying (15),

E[ n (1 , 2 ) (1 , 2 )]2

(1 12 )1 (1 22 )2

d1 d2 = O

(log n)6

(18)

with n defined in (13) and (14).

The next theorem provides a uniform convergence rate for the estimators in the semi-parametric
hypothesis case.

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

Fig. 1. Sample path (left) and theoretical spatial auto-covariance (right) of the aggregated field with mixing density (21).

Theorem 4. Let (1 , 2 ), be defined as in (17) where 1 and 2 are analytic on the unit disc, continuous
in A1 and not vanishing at 1. Let dj = min(dj,1 , dj,2 ), j = 1, 2. Then for any [0, 1[, if 0 j <
dj + 1/2, j = 1, 2,

sup ( n (1 , 2 ) (1 , 2 ))2

C ()

1 ,2 A

2(d1 +d2 1 2 +1)

(log n)

(19)

If 5/2 j < dj + 1/2, j = 1, 2, the result holds also with = 1.

3. Monte Carlo validation
A small simulation study is proposed here in order to evaluate the performance of the estimators
defined in Section 2.
The data are obtained by generating the limit Gaussian field (4). In order to generate a Gaussian
field with given covariance structure the procedure described in Wood and Chan (1994) has been
implemented. The second order properties of the generated data have been checked at all instances
and we always found a good agreement between theoretical and sample properties of the simulated
fields.
With the data we proceed to calculate the estimator n (1 , 2 ) for different choices of the
parameters 1 and 2 , K1 (n) and K2 (n). Note that since the estimator n (1 , 2 ) maybe negative in
some sub-regions, in the simulations its re-normalized positive part computed as

n+ (1 , 2 ) = 1 1
1 1

max[0, n (1 , 2 )]
max[0, n (1 , 2 )] d1 d2

(20)

is used. As an example of application, consider the mixing density

(1 , 2 )

(1 i )(1 + i )2.25 .

(21)

i=1

The choice of the parameters d1 = 1, d2 = 2.25 for the mixing density has been made in order to
have a constantly decreasing auto-covariance function of the resulting aggregated field. Fig. 1 reports
the sample path of a randomly generated field and a portion of the theoretical spatial auto-covariance
of the underlying Gaussian field.
To obtain the simulation results, for each choice of 1 and 2 , K1 (n) and K2 (n), m = 100 Gaussian
fields of size n2 with n = 128 have been generated. The number m of replications has been set at 100 to

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

(a) K1 (n) = 1, K2 (n) = 1.

(b) K1 (n) = 2, K2 (n) = 1.

(c) K1 (n) = 2, K2 (n) = 2.

(d) K1 (n) = 3, K2 (n) = 3.

Fig. 2. Contour plots for the estimated MISE. Estimated are based on 100 sampled fields of size n2 , n = 128.

control the computing time given the high number of calculations to be performed at each estimation
run. The size n = 128 is due to the use of the Fast Fourier Transform in the generating procedure for
the Gaussian field which requires n = 2i for some integer i. The chosen n provides accuracy in the
generated field, consistent with the theoretical one, and avoids a field of too large total dimension,
of size n2 , for computations. The estimated MISE has been computed as average over the repeated
samples. Fig. 2 reports the estimated values of the MISE (16) for different choices of parameter values.
Each sub figure represent a different choice of the parameters K1 (n) and K2 (n) and the level curves
present in each subgraph represent the estimated MISE for specific choices of 1 and 2 . Values of 1
and 2 range from 0 to 2 by 0.2 intervals; note however that Fig. 2(d) allows the values 0.2.
As far as the bounds on K1 (n) and K2 (n) determined by Theorem 1, considering an equal growth
in both directions, then 1 = 2 = ; from (15) we have that approximately < 1.13; then since
n = 128 it turns out that K1 (n) and K2 (n) must not be greater than 5. As we see, the behavior of the
level curves is quite regular. The optimal results are obtained for K1 (n) = 1 and K2 (n) = 1 when very
low values of the estimated MISE are obtained in correspondence of values of 1 and 2 close to 1. An
analogous behavior, for one dimensional processes, was already noted in Leipus et al. (2006). Using
more terms in the approximation results in a loss of precision, the level curves seem to indicate the

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

Fig. 3. True (left) and estimated (normalized) mixing density (21).

need to lower the values of the parameters 1 and 2 . Recall that 1 , 2 > 1; however we run into
numerical problems in computations by choosing negative values for the s, the only exception being
the case K1 (n) = 3, K2 (n) = 3.
The simulations seem to indicate that values of 1 , 2 in (0, 1) give the best results; with indication
to choose 1 , 2 closer to zero if higher order polynomials are used. In the case of a unimodal mixing
density it seems appropriate to use K1 (n) = K2 (n) = 1 together with 1 = 2 = 1.
Fig. 3(a) and (b) report, respectively, the true and estimated (with K1 (n) = K2 (n) = 1 and
1 = 2 = 1) mixing density n+ (1 , 2 ); the estimation procedure provided an estimated (unnormalized) mixing density as

n (1 , 2 ) = (1 12 )(1 22 )(0.455507 + 0.5526511 + 0.5397682 + 0.5272441 2 ).

The visual effect in Fig. 3 is quite satisfactory as the two densities match quite closely. Note that the
estimated density, given the presence of some negative values on the edges is flat beyond the white
line. The average estimated MISE for this case was equal to 0.01098.
4. An application to EU-15 per capita GDP at the regional level
The -convergence approach, especially developed by Mankiw et al. (1992) and Barro and Salai-Martin (1995), is often used in studying economic growth and inequality in income distribution at
regional or national level. It is said that a -type convergence occurs when a poor economy grows
faster than a richer economy, reaching it in terms of level of income, value added or GDP per capita. A
basic model for the study of -convergence refers to the equation (see, e.g. Arbia, 2005, p. 8 for more
details)
ln(yT ) = (1 e T ) ln(y ) + e T ln y0 ,

(22)

where y0 and yT represent, respectively, the product (or income or value added) at time 0 (the initial
period) and at time T ; y denotes the equilibrium value of the system, given the technology level. As
one can see, the value ln(yt ) is a linear combination of the initial value and the equilibrium point, the
speed of convergence toward the equilibrium is determined by the parameter .
In order to actually estimate the model, given a sequence of data (YT ,i , Y0,i ), i = 1, . . . , n, it is
common practice to cast it in a regression set-up. After rearranging terms, conditionally on Y0,i = y0,i
the model to be estimated is defined through the equation

ln

YT ,i
y0,i

= 0 + 1 ln(y0,i ) + i

(23)

10

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

where 0 is a constant related to the technological level, 1 = (1 e T ) and i is an error term

with null mean and variance 2 . Typically, model (23), postulating independence among equations,
is estimated by means of ordinary least squares (OLS); a significant negative value of 1 in the model
is indicator of presence of -convergence since a negative correlation between growth and initial
income implies a tendency for poor countries to catch up (Baumol, 1986).
There are a variety of studies modeling empirical data on per capita GDP at regional level using
model (23) as a basic starting point. Abreu et al. (2005) report 1650 scientific papers with empirical
studies on economic growth using a regression approach (other approaches are excluded from the
figure above!). A random sample of 610 papers from the selected 1650 provides estimates of the
convergence rate ranging from 2 to 54 with an average of 4.3%. For details and discussion as well
as a thorough meta-analysis of the literature and references one can consult Abreu et al. (2005).
We present here an application of the estimation procedure defined in Section 2 to the per capita
Gross Domestic Product (GDP) of the EU-15 NUTS2 regions. Nomenclature of Territorial Units for
Statistics (NUTS) is a classification system at different levels of the EU area; in our case the NUTS2
level for the EU-15 includes 190 European regions for which the data on head GDP in 1980 (the initial
period) and 2003 have been obtained; the data are freely available at the Eurostat website. Before
getting into the detail of the application it is appropriate to discuss a bit further the characteristics of
the data at hand and of the phenomenon studied.
It is now well documented that the presence of spatial effects in these phenomena cannot be
overlooked, we cite here, for further references, Arbia (2005) and Monfort (2008) which points out
that in presence of spatial effects the -convergence rate maybe overestimated. Another relevant
feature, which is attracting increasing interest and which seems to be confirmed by some studies (e.g.
Huber and Pfaffermayr, 2010, Bollino and Polinori, 2007) is that convergence occurs at micro-area
level, summing up at regional or national level. Indeed, the very nature of GDP implies that it can be
regarded as an aggregated field as it is the sum of several micro-effects of economic agents dislocated
in the territory. However it is quite difficult, if not senseless, to obtain data at micro-level: the smaller
the geographical unit considered, the more likely a substantial fraction of local GDP is attributable to
commuters and the more difficult becomes to interpret the concept of GDP per head (Monfort, 2008).
Hence, extension of the analysis to data obtained at the level of a NUTS3 classification (a sub-regional
level) or finer may not be appropriate.
Notwithstanding a statistical analysis providing information of what is going on at the micro-level
can be important to fully understand the phenomenon. In this sense we will cast (23) in a spatial
context, i.e. information about the geographical location of the i-th unit will be considered and, in
particular, we will assume that in a spatially defined model (23), the error term, now indicated as s,t ,
is an aggregated of the form (4) satisfying assumption IIII. This approach is not new in many spatial
analyses where a spatial error model is introduced in order to analyze spatial inter-dependences. In
our case the accent would be on the mixture density underlying the aggregation of data, i.e. the aim is
to individuate the presence of a well defined structure. The hypothesis here is that the individual field
values aggregating to s,t represent the contribution of micro-economic units (MEUs) on the territory.
The implicit assumption, given our theoretical framework, of independence among MEUs on the same
area is rather strong. We have two comments in this respect: first, this represent a first approach to
the problem which, if proves itself useful, could be further developed. Second, we may reasonably
assume that conditioning on GDP, i.e. the state of the economy (which is given in our setup), will
greatly reduce dependence among MEUs. We hence consider the model

ln

YT ,(s,t )

y0,(s,t )

= 0 + 1 ln(y0,(s,t ) ) + (s,t )

(24)

where (s, t ) indicates the coordinates (latitude and longitude) of the centroids of the spatial unit; a
preliminary estimation of the parameters 0 and 1 will be based on a standard regression model, the
aim here is to analyze a spatial structure in the error terms leaving it as untouched as possible. The
underlying density will then be estimated on the basis of the observed errors

(s,t ) = ln

YT ,(s,t )
y0,(s,t )

0 + 1 ln(y0,(s,t ) )

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

11

Table 1
OLS estimates of model (24) for the per capita GDP of the EU-15 NUTS2
regions.

0
1

Estimate

Std. error

t-statistic

P-value

0.00019
0.01204

0.0003687
0.0009532

0.5121
12.6291

0.6092
0.0000

Fig. 4. Regression residuals for the NUTS2 data.

where 0 and 1 are the OLS estimates. Table 1 report the regression values for estimating the
coefficients. As we see, there definitely is a convergence effect in the area under study.
Fig. 4 depicts the regression residuals for the 190 NUTS2 regions. It can be seen from the map that
regions do show a pattern of spatial dependence: the smooth change of colors indicates a smooth
variation from large negative residuals to large positive ones showing that the regression (24) does
not capture the full structure present in the data: a regional dependence appears evident.
We then analyze whether we can find a meaningful structure in the residuals by using model (1).
Given the data are not on a regular grid the application of model (1) can be criticized; however, since
we are dealing with residuals we may reasonably assume that, and indeed the smooth change of colors
in the map supports this idea, a stationary autoregressive structure holds; model (1) here is used as
a discrete approximation to a continuous phenomenon to describe the basic structure of relations on
the territory. The idea here is anyway to stretch to the maximum extent the theoretical model to see
if consistent and interpretable results are obtained with the estimation technique proposed.
In order to get a closer match between the theoretical model and the data, we define the distance
between region i and region j, for i, j = 1, . . . , 190, to be the distance in km between the centroids of
the regions with a reference unit distance set at 321 km. This is the minimum distance that assures
that every region in the data-set has at least a neighbor, i.e., if dij denotes the distance in km between
region i and j, then max1i190 min1j190 (di,j ) = 321. In a regular grid, the estimator (12), n1 ,n2 (j1 , j2 )
considers all couple of points which are

j21 + j22 apart. In our example, with an irregular grid and

12

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

Fig. 5. Estimated (normalized) mixing density for the EU-15 NUTS2 GDP data.

limited number of observations one would have too few observations for any given exact distance d.
Our proposed solution, for a given distance d, considers all data couples whose
distance is within a
given interval containing d. More precisely, in computing n1 ,n2 (j1 , j2 ), for d =

j21 + j22 we use all data

couples whose centroids are within a km distance less than 321 if d = 1 and between 321d
321/2
and 321d + 321/2 otherwise. For example, if n1 ,n2 (1, 1) has to be estimated, then d = 2 and all
regions between a distance of 293.46 and 614.46 km are included in the computation of n1 ,n2 (1, 1).
We are aware that this creates further bias in the estimation however it seems a reasonable procedure
since we are essentially exploiting an isotropy condition on the residuals and not on the original GDP
data.
We proceed by estimating the mixing density as described in Section 2; given the limited number
of data we restrict to the case K1n = K2n = 1 with 1 = 2 = 1; we controlled for other values of the
s in the range [0.5 1.5] obtaining similar results in all cases. The estimated mixing density has the
form

1 , 2 ) = (1 12 )(6.77726 106 0.000032761 0.000032762

(
+ 0.0005077211 2 )(1 22 )
and it is represented in Fig. 5. As we see the estimated mixing density shows that there is a clear cut
behavior of the coefficients 1 and 2 where they tend to be both positive or both negative implying
MEUs are either positively or negatively associated with neighbors. The residuals should be essentially
free from -type effects as they should have been removed by the regression; what is left is the
structure of relationships within the territory which seems quite rich. Schematically we can note the
following: (a) given the fact that there are either positive and negative relationship this could be due to
links between economic sectors rather than a generalized micro-structure territorial level; one sector
could grow at lower speed (or decrease) given its state in a larger economy; (b) the varied structure
suggests that there is much less regional pattern than at the aggregated level (ecological correlation
effects are present) and consequently there seems to be no further evidence of -type effects at microlevel. Overall the results go along with the intuition that nowadays dependences between economies
are less and less local and that analyzing GDP at too fine level maybe inappropriate.

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

13

The clear and consistent results show that this technique could find relevant applications for the
analysis of specific territorial areas or economic sectors in order to analyze micro-effects as interaction
of MEUs cannot avoid considering the spatial side of the phenomenon.
5. Conclusions
We have presented an estimation procedure for the mixing density for autoregressive random
fields with random coefficients. The relatively simple set-up proposed, based on independence
assumptions, is to be seen as a starting point and as an approximation to practical situations. Applications of the technique proposed can be developed in all those situation where aggregation of data at
micro-level is done for administrative, privacy or constraints of other nature. The proofs of the theorems show that the methodology can be easily extended to higher order spatial autoregressive models
by considering multi-dimensional Gegenbauer polynomials constructed from higher order tensorial
products of one-dimensional Gegenbauer polynomials. This fact would allow the formulation of the
aggregation procedure studied in general classes of spatial autoregressive processes, including the
ones considered in Whittle (1954).
Acknowledgments
The authors would like to thank three anonymous referees for their valuable comments which led
to an improved version of the paper.
Nikolai Leonenko was partially supported by the Marie-Curie grant PIRES-GA-2008-230804.
Emanuele Taufer gratefully acknowledges financial contribution of PRIN 2007 JRTXFL (Analysis and
modeling of efficiency, productivity and public policies at the micro-area level).
The authors are indebted to prof. Charles Suquet for providing references related to compactness
of probabilistic measures in Hilbert Spaces and to Dr. Diego Giuliani for providing Fig. 4.
Appendix. Proofs
For tightness in Hilbert spaces, there is a necessary and sufficient condition, see Theorem 1 in
Suquet (1999). It can be easily adapted to the case of Hilbert spaces having a basis indexed by some
countable set I. Indeed an Hilbertian basis in a separable Hilbert space is always unconditional, so the
convergence of the expansion into its basis is preserved under any permutation of the indexes or any
grouping of terms.
In our setting, one can use Theorem 3 in Suquet (1999) with a Schauder decomposition with the
space Gi defined as the space of sequences (at1 (i),t2 )t2 N . This leads to the following statement.
Theorem 5. Let F be a set of random elements X = (Xt1 ,t2 )(t1 ,t2 )N2 in the Hilbert space H2 (N2 ) of real
sequences a = (at1 ,t2 )(t1 ,t2 )N2 such that
a2t1 ,t2

a2 :=

t =(t1 ,t2 )N2

(1 + |t |)2+

< .

Then F is tight if and only if

1. For every j N,

lim sup P

c X F

Xi2,t2

i=0 t2 N

(1 + |(i, t2 )|)2+

>c

= 0.

2. For each positive ,

lim sup P

j X F

Xi2,t2

i>j t2 N

(1 + |(i, t2 )|)2+

>

= 0.

14

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

Another possibility is to use an increasing sequence of finite index sets, say Cn , such that nN Cn =
N2 . Indeed with Bn := Cn \ Cn1 , we have a Schauder decomposition of H2 (N2 ) into the spaces of finite
sequences indexed by Bn . This leads to the following.
Theorem 6. F is tight if and only if
1. For every n N,

lim sup P

c X F

Xt2

t Cn

(1 + |t |)2+

>c

= 0.

2. For each positive ,

lim sup P

j X F

Xt2

t N2 \C

> = 0.

(1 + |t |)2+

Of course Theorems 5 and 6 give more practical sufficient conditions if we assume that the Xt s are
square integrable.
Concerning sequence spaces, these results can be somewhat extended beyond the Hilbertian case,
see Section 6 in Suquet (1999).
(N )

Proof of Theorem 1. Almost surely, Xs,t H2 (N2 ), since

t =(t1 ,t2 )N2

[Xs(,Nt ) ]2
= E [X0(,N0) ]2
(1 + |t |)2+

t =(t1 ,t2 )N2

(1 + |t |)2+

< .

(N )

The tightness of the sequence {Xs,t } follows from Theorem 5 by using the Markov inequality
(P (|X | > a) < E |X |/a), and the fact that
(N )
E [Xs,t ]2

N
1

N j=1

(j)
VarYs,t

=
1

neither depends on N nor (s, t ).

1 (1 )2 (2 )
d1 d2
(1 12 )(1 22 )

Proof of Theorem 2. Defining

Rn =

( , )

K1 >K1 (n) K2 >K2 (n)

K1 ,K2 GK11,K22 (1 , 2 )

and using the definition of n in (13) then the MISE (16) can be estimated as

A21

E[ n (1 , 2 ) (1 , 2 )]2 w1 (1 )w2 (2 ) d1 d2

=
A21

(K(1n,)K2
K1 K1 (n) K2 K2 (n)

K1 K1 (n) K2 K2 (n)

( , )
K1 ,K2 )GK11,K22 (1 , 2 )

(n)

E(K1 ,K2 K1 ,K2 )2 +

2
+ Rn

w1 (1 )w2 (2 ) d1 d2

(K1 ,K2 )2 ,

K1 >K1 (n) K2 >K2 (n)

where the last equality is obtained by exploiting the orthogonality property of the Gegenbauer
polynomials. Next, the expected value in the above expression can be estimated as

(n)
E(K1 ,K2 K1 ,K2 )2 = E

K1
K2

j 1 =0 j 2 =0

( )

( )

gK1 ,1j1 gK2 ,2j2 [ n (j1 , j2 ) n (j1 , j2 + 2)

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

15

n (j1 + 2, j2 ) + n (j1 + 2, j2 + 2)]

2

[ (j1 , j2 ) (j1 , j2 + 2) (j1 + 2, j2 ) + (j1 + 2, j2 + 2)]

Using the CauchySchwarz inequality and the relation E(X 2 ) = Var(X ) + E(X )2 the above term can
be bounded by

(K1 + 1)(K2 + 1)

( )

( )

max gK1 ,1j1 max gK2 ,2j2

1 j K
1j K
1

K1
K2

E [ n (j1 , j2 ) n (j1 , j2 + 2) n (j1 + 2, j2 ) + n (j1 + 2, j2 + 2)]

j1 =0 j2 =0

2
[ (j1 , j2 ) (j1 , j2 + 2) (j1 + 2, j2 ) + (j1 + 2, j2 + 2)]

+ Var n (j1 , j2 ) n (j1 , j2 + 2) n (j1 + 2, j2 ) + n (j1 + 2, j2 + 2) .

(25)

We first evaluate the expectation in the above expression. It is well known that

j1
j2
E n (j1 , j2 ) = 1
1
(j1 , j2 ).
n

(26)

The term 1 n1 1 n2 is a source of the so-called edge-effect typical of spatial analysis, see Dahlhaus
and Knsch (1987) or Guyon (1982) for further details. In our case, since the total dimension is n2 , the
bias is simply evaluated of order n1 ; however, in order to determine the rate of convergence of the
MISE (16) a more precise evaluation of the bias of the estimators of the covariance terms (j1 , j2 ) is
needed. To do this we evaluate directly the terms of type (26) appearing in (25). Some lengthy but
straightforward algebra leads to conclude that
K1
K2

E [ n (j1 , j2 ) n (j1 , j2 + 2) n (j1 + 2, j2 ) + n (j1 + 2, j2 + 2)]
j1 =0 j2 =0

[ (j1 , j2 ) (j1 , j2 + 2) (j1 + 2, j2 ) + (j1 + 2, j2 + 2)]

K13 K23
n2

(27)

where we note that, to carefully evaluate the bias, we need to include the terms K1 and K2 in the
numerator. This will be relevant in determining the exact rate of increase of K1 (n) and K2 (n). We next
turn to estimate the variance term in (25), to this end, Lemma 1 below allows to estimate
Var n (j1 , j2 ) n (j1 , j2 + 2) n (j1 + 2, j2 ) + n (j1 + 2, j2 + 2)

C
n2

(28)

Recall from Lemma 5 in Leipus et al. (2006) that

()

max |gk,j | Ck11/2 ek ,

0jk

= log(1 +

2).

(29)

Putting together (27), (28), (29) we can make the estimate

(n)

E(K1 ,K2 K1 ,K2 )2 C

K115 K215 e(K1 +K2 )2

n2

from which, applying (15) we obtain that, as n ,

K1 K1 (n) K2 K2 (n)

(n)

E(K1 ,K2 K1 ,K2 )2

K1 (n)16 K2 (n)16 e(K1 (n)+K2 (n))2

n2

0. 

16

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

Lemma 1. Under the condition that |1 e2i |2 f () is bounded, for n (j1 , j2 ) as defined in (12) then
1

Var n (s, t ) n (s, t + 2) n (s + 2, t ) + n (s + 2, t + 2)

n4

n2

(30)

Proof. We need to compute

1
n4

Var

n s
nt

t +2)
ns n(

Xj1 ,j2 Xj1 +s,j2 +t

j 1 =0

j 1 =0 j 2 =0
n(s+2) n(t +2)

j 1 =0

j 2 =0

Xj1 ,j2 Xj1 +s,j2 +t +2

j 2 =0
n(s+2) nt

Xj1 ,j2 Xj1 +s+2,j2 +t +2

j1 =0

Xj1 ,j2 Xj1 +s+2,j2 +t .

(31)

j 2 =0

Note that there are four terms inside the brackets of the Variance in the above formula, say Var(T1
T2 + T3 T4 ) which, by the CauchySchwarz inequality is bounded by C [Var(T1 T2 ) + Var(T3 T4 )].
Consider first the difference T1 T2 and write it as
t +2 )
ns n(

j1 =0

Xj1 ,j2 Xj1 +s,j2 +t Xj1 +s,j2 +t +2

j 2 =0
n s

Xj1 ,nt 1,j2 +s,n1 +

j 1 =0

ns

Xj1 ,nt ,j2 +s,n =: T11 + T12 + T13 .

j 1 =0

Again, by the CauchySchwarz inequality we have Var(T1 T2 ) C [Var(T11 ) + Var(T12 ) + Var(T13 )].
Consider first the term T11 ; in order to compute its variance we can apply the diagram formula for
cumulants which, since Xt is Gaussian, is given by the expression
Var(T11 ) =

ns n
t 2
ns n
t 2

(Xj1 ,j2 Xj3 ,j4 ) (Dj1 ,j2 Dj3 ,j4 )

j 1 =1 j 2 =1 j 3 =1 j 4 =1

+ (Xj1 ,j2 Dj3 ,j4 ) (Dj1 ,j2 Xj3 ,j4 )

(32)

where Dj1 ,j2 = Xj1 +s,j2 +t Xj1 +s,j2 +t +2 and (X ) indicates the first cumulant of a r.v. X . Note that

(Xj1 ,j2 Xj3 ,j4 ) = (j3 j1 , j4 j2 )

(Dj1 ,j2 Dj3 ,j4 ) = 2 (j3 j1 , j4 j2 ) (j3 j1 , j4 j2 2) (j3 j1 , j4 j2 + 2)
(Xj1 ,j2 Dj3 ,j4 ) = (j3 j1 + s, j4 j2 + t ) (j3 j1 + s, j4 j2 + t + 2).
Using the spectral representation we have

(Xj1 ,j2 Xj3 ,j4 ) (Dj1 ,j2 Dj3 ,j4 ) =

ei(j3 j1 )1 +i(j4 j2 )2 f (1 )f (2 )d1 d2

[2 e

ei(j3 j1 )3 +i(j4 j2 )4

2i4

e2i4 ]f (3 )f (4 )d3 d4

ei(j3 j1 )(1 +3 )+i(j4 j2 )(2 +4 ) |1 e2i4 |2 f ()d

where, when needed, f ()d = f (1 ) f (4 )d1 d4 . Define the quantity Fn (x) = |

which is proportional to the Feir kernel and satisfies n
ns nt 2 ns nt 2
1

n4

j1 =1 j2 =1 j3 =1 j4 =1

Fn (x)dx = 2 . Then we have

ei(j3 j1 )(1 +3 )+i(j4 j2 )(2 +4 ) |1 e2i4 |2 f ()d

j =1

eijx |2

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

n4

n4

Fns (1 + 3 )Fnt 2 (2 + 4 )|1 e2i4 |2 f ()d

Fns (1 )

f (3 )f (1 3 )

f (2 )

Fnt 2 (2 + 4 )|1 e2i4 |2 f (4 )d.

17

(33)

Since f 2 ()d < and |1 e2i |2 f () is bounded, n1

(33) can be bounded by
1
n2

Fn ()|1 e2i |2 f ()d 2 C and then

4 2 C f 1 f 22 .

(34)

Consider now the second term in (32)

ns nt 2 ns nt 2
1

n4

(Xj1 ,j2 Dj3 ,j4 ) (Dj1 ,j2 Xj3 ,j4 )

j1 =1 j2 =1 j3 =1 j4 =1
ns nt 2 ns nt 2
1

n4

j1 =1 j2 =1 j3 =1 j4 =1

n4

A2

ei(j3 j1 +s)(1 +3 )+i(j4 j2 +t )(2 +4 ) (1 e2i2 )(1 e2i4 )f ()d

Fns (1 + 3 )eis(1 +3 ) f (1 )f (3 )

A2

A4

Fnt 2 (2 + 4 )eit (2 +4 ) (1 e2i2 )(1 e2i4 )f (2 )f (4 )d.

(35)

To see how the above term can be bounded, note that, by the CauchySchwarz inequality we have

(1 e2i2 )(1 e2i4 )f (2 )f (4 )d2 d4

2

|1 e2i2 |2 f (2 )f (4 )d2 d4
|1 e2i4 |2 f (2 )f (4 )d2 d4
A

and f () = |1 e2i |f () is in L2 given f is integrable and |1 e2i |2 f () is bounded. Hence, the

absolute value of (35), can be bounded by

1
n4

Fns (1 )
A

1
n2

f (1 )f (1 3 )
A

Fnt 2 (2 )
A

f (4 )f (2 4 )d
A

4 2 C f 22 f 22 .

(36)

Let us consider now the computation of Var(T12 ). Again, applying the diagram formula for cumulants
of Gaussian r.v. this accounts at computing
ns ns
1

n4

(Xj1 ,nt 1 , Xj2 ,nt 1 ) (Xj1 +s,n1 , Xj2 +s,n1 )

j 1 =1 j 2 =1

+ (Xj1 ,nt 1 , Xj2 +s,n1 ) (Xj1 +s,n1 , Xj2 ,nt 1 )

n s
1
ei(j2 j1 )(1 +3 )+i(2 +4 ) f ()d
= 4
n

1
n4

j 1 =1

A2

Fns (1 )
A

f (3 )f (1 3 )
A

A2

f (2 )f (4 )d

18

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

1
n3

2 C f 22 (f 1 )2 .

Notice that Var(T13 ) can be treated in the same way to see that it is O(n3 ). Overall, the same technique
of analysis can be applied to the other terms resulting from (31) to show that (30) holds. 
( , )

Proof of Theorem 3. Note that from (10), K1 ,K2 = EW (1 , 2 )GK11,K22 (1 , 2 ) which can be rewritten as

A1

with

2 (2 )
( )
G 2 (2 )w2 (2 )
(1 22 )2 K2

j (j )

(1j2 ) j

A1

1 (1 )
( )
G 1 (1 )w1 (1 ) d1 d2 ,
(1 12 )1 K1

, j = 1, 2 satisfying the conditions of Proposition 6 in Leipus et al. (2006); from this it

follows that K1 ,K2 = O(K12 K22 ). Then for K1 (n) and K2 (n) satisfying the conditions of Theorem 1 it
holds

(K(1n,)K2 )2 = O

K1 >K1 (n) K2 >K2 (n)

K1 (n)3 K2 (n)3

=O

13 23 (log n)6

Hence the result of the theorem follows since

K1 K1 (n) K2 K2 (n)

(log n)30

(n)

E(K1 ,K2 K1 ,K2 )2 C

n2

0 < < 1. 

Proof of Theorem 4. Note that we can estimate,

sup (1 12 )21 (1 22 )22

sup ( n (1 , 2 ) (1 , 2 ))2 =

1 ,2 A

1 ,2 A

( , )

(K(1n,)K2 K1 ,K2 )GK11,K22 (1 , 2 )

K1 K1 (n) K2 K2 (n)

K1 >K1 (n) K2 >K2 (n)

( , )
K1 ,K2 GK11,K22 (1 , 2 )

(K(1n,)K2
K1 K1 (n) K2 K2 (n)

K1 ,K2 )sK1 ,K2

+ 2 sup (1 12 )21 (1 22 )22

1 ,2 A

K1 >K1 (n) K2 >K2 (n)

( , )
K1 ,K2 GK11,K22 (1 , 2 )

=: T1 + T2 ,
( , )

where sK1 ,K2 = sup1 ,2 A (1 12 )21 (1 22 )22 GK11,K22 (1 , 2 ). As far as the term T1 is concerned, by
inequality 7.33.6 in Szeg (1967) we have
C

( , )

GK11,K22 (1 , 2 )

(1 12 )1 /2+1/4 (1 22 )2 /2+1/4

1 < 1 , 2 < 1.

Hence sK1 ,K2 C sup1 ,2 A (1 12 )1 /21/4 (1 22 )2 /21/4

CauchySchwarz inequality it follows that

E[T1 ] (K1 (n) + 1)(K2 (n) + 1)

K1 K1 (n) K2 K2 (n)

(n)

C () < . Applying the

s2K1 ,K2 E(K1 ,K2 K1 ,K2 )2

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

19

C ()K1 (n)K2 (n)e(K1 (n)+K2 (n))2 (K1 (n)K2 (n))15

(log n)32
C () 2( + )2 .
n

As far as T2 is concerned, note that since (1 , 2 ) can be factored as 1 (1 )2 (2 ) we can find

coefficients such that

K1 >K1 (n) K2 >K2 (n)

( , )

K1 ,K2 GK11,K22 (1 , 2 ) =

K1 >K1 (n)

( )

K1 GK11 (1 )

K2 >K2 (n)

( )

K2 GK22 (2 ),

by Leipus et al. (2006), Proposition 7 we can bound the above term by

C

(K1 (n))21 2d1 1 (K2 (n))22 2d2 1

(1 |1 |)1 +5/2 (1 |2 |)2 +5/2

and hence
T2

sup (1 |1 |)1 5/2 (1 |2 |)2 5/2 (K1 (n))21 2d1 1 (K2 (n))22 2d2 1 .

1 ,2 A

The result of the theorem now follows from applying (15).

References
Abramowitz, M., Stegun, I.A. (Eds.), 1972. Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical
Tables, ninth printing ed. Dover, New York.
Abreu, M., de Groot, H.L.F., Florax, R.J.G.M., 2005. A meta-analysis of beta-convergence. Tinbergen Institute Discussion Paper TI
2005-001/3.
Arbia, G., 2005. Spatial Econometrics. Springer, Berlin.
Arbia, G., Petrarca, F., 2011. Effects of MAUP on spatial econometrics models. Lett. Spatial Res. Sci. 4 (3), 173185.
Baran, S., Pap, G., 2009. On the least squares estimator in a nearly unstable sequence of stationary spatial AR models.
J. Multivariate Anal. 100, 686698.
Barro, R.J., Sala-i-Martin, X., 1995. Economic Growth. McGraw-Hill.
Basu, S., Reinsel, G.C., 1993. Properties of the spatial first-order ARMA model. Adv. Appl. Probab. 25, 631648.
Basu, S., Reinsel, G.C., 1994. Regression models with spatially correlated errors. J. Amer. Statist. Assoc. 89, 8899.
Baumol, W.J., 1986. Productivity growth, convergence, and welfare: what the long-run data show. Amer. Econ. Rev. 76,
10721085.
Beran, J., Ghosh, S., Schell, D., 2009. On least squares estimation for long-memory lattice processes. J. Multivariate Anal. 100
(10), 21782194.
Beran, J., Schtzner, M., Ghosh, S., 2010. From short to long memory: aggregation and estimation. Comput. Statist. Data Anal.
54 (11), 24322442.
Besag, J., 1974. Spatial interaction and the statistical analysis of lattice systems. J. Roy. Statist. Soc. Ser. B 36 (2), 192236.
Bhattacharyya, B.B., Khalil, T.M., Richardson, G.D., 1996. GaussNewton estimation of parameters for a spatial autoregression
model. Statist. Probab. Lett. 28, 173179.
Bhattacharyya, B.B., Richardson, G.D., Franklin, L.A., 1997. Asymptotic inference for near unit roots in spatial autoregression.
Ann. Statist. 25, 17091724.
Boissy, B.B., Bhattacharyya, B., Li, X., Richardson, G.D., 2005. Parameter estimates for fractional autoregressive spatial processes.
Ann. Statist. 33, 25532567.
Bollino, C.A., Polinori, P., 2007. Ricostruzione del valore aggiunto su scala comunale e percorsi di crescita a livello microterritoriale: il caso dellUmbria. Scienze Regionali 6 (2), 3573.
Celov, D., Leipus, R., Philippe, A., 2010. Asymptotic normality of the mixture density estimator in a disaggregation scheme.
J. Nonparametr. Stat. 22 (4), 425442.
Cressie, N., 1996. Change of support and the modifiable areal unit problem. Geogr. Syst. 3, 159180.
Culles, B.R., Gleeson, A.C., 1991. Spatial analysis of field experimentsan extension to two dimensions. Biometrics 47, 14491460.
Dahlhaus, R., Knsch, H., 1987. Edge effects and efficient parameter estimation for stationary random fields. Biometrika 74 (4),
877882.
Davidson, J., 1991. The cointegration properties of vector auto-regression models. J. Time Ser. Anal. 12 (1), 2730.
Davidson, J., 2002. A model of fractional cointegration, and tests cointegration using the bootstrap. J. Econometrics 110 (2),
187212.
Davidson, J., Monticina, A., 2010. Tests for cointegration with structural breaks based on sub-samples. Comput. Statist. Data
Anal. 54 (11), 24982511.
Granger, C.W.J., 1980. Long memory relationships and the aggregation of dynamic models. J. Econometrics 14 (2), 227238.
Guo, H., Lim, C.Y., Meerschaert, M.M., 2009. Local Whittle estimator for anisotropic random fields. J. Multivariate Anal. 100 (5),
9931028.
Guyon, X., 1982. Parameter estimation for a stationary process on a d-dimensional lattice. Biometrika 69 (1), 95105.
Huber, P., Pfaffermayr, M., 2010. Testing for conditional convergence in variance and skewness: the firm size distribution
revisited. Oxford Bull. Econ. Stat. 72, 648688.

20

N. Leonenko, E. Taufer / Spatial Statistics 3 (2013) 120

Lavancier, F., 2005. Long memory random fields. In: Bertail, P., Doukhan, P., Soulier, Ph (Eds.), Dependence in Probability and
Statistics. In: Lecture Notes in Statistics, vol. 187. Springer.
Lavancier, F., 2011. Aggregation of isotropic autoregressive fields. J. Statist. Plann. Inference 141 (12), 38623866.
Leipus, R., Oppenheim, G.E, Philippe, A., Viano, M., 2006. Orthogonal series density estimation in a disaggregation scheme.
J. Statist. Plann. Inference 136 (8), 25472571.
Leonenko, N., Taufer, E., 2005. Convergence of integrated superpositions of OrnsteinUhlenbeck processes to fractional
Brownian motion. Stochastics 77 (6), 477499.
Mankiw, N.G., Romer, D., Weil, D., 1992. A contribution to the empirics of economic growth. Quart. J. Econ. 107, 407437.
Martin, R.J., 1979. A subclass of lattice processes applied to a problem in planar sampling. Biometrika 66 (2), 209217.
Monfort, P., 2008. Convergence of EU regionsmeasures and evolution. Working Paper 1/2008Directorate-General for
Regional Policy, EU.
Roknossadati, S.M., Zarepour, M., 2010. M-estimation for a spatial unilateral auto-regressive model with infinite variance
innovations. Econometric Theory 26, 16631682.
Ruiz-Medina, M.D., 2011. Spatial autoregressive and moving average Hilbertian processes. J. Multivariate Anal. 102 (2), 292305.
Ruiz-Medina, M.D., 2012. Spatial functional prediction from spatial autoregressive Hilbertian processes. Environmetrics 23 (1),
119128.
Suquet, Ch., 1992. Relecture des critres de relative compacit dune famille de probabilits sur un espace de Hilbert. Pub. IRMA
Lille 28-III.
Suquet, Ch., 1999. Tightness in Schauder decomposable Banach spaces. Amer. Math. Soc. Transl. Ser. 2 193, 201224.
Szeg, G., 1967. Orthogonal Polynomials, third ed. In: American Mathematical Society Colloquium Publications, vol. 23.
American Mathematical Society, Providence, RI.
Whittle, P., 1954. On stationary processes in the plane. Biometrika 41, 434449.
Wood, A., Chan, G., 1994. Simulation of stationary Gaussian processes in [0, 1]d. J. Comput. Graph. Statist. 3 (4), 409432.
Zaffaroni, P., 2004. Contemporaneous aggregation of linear dynamic models in large economies. J. Econometrics 120 (1), 75102.