Testing for Directional Symmetry in Spatial Dependence Using the Periodogram Nelson Lu 1 Dale L. Zimmerman 2 1 Nelson Lu is Biostatistician, Wyeth Research, Pearl River, NY 10965 (E-mail: [email protected]) 2 Dale L. Zimmerman is Professor, Department of Statistics and Actuarial Science, University of Iowa, Iowa City, IA 52242 (E-mail: [email protected]; Phone: 319-335-0818; Fax: 319- 335-3017). Please address all correspondence to this author.
24
Embed
Testing for directional symmetry in spatial dependence using the periodogram
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Testing for Directional Symmetry in Spatial Dependence
Using the Periodogram
Nelson Lu1 Dale L. Zimmerman2
1Nelson Lu is Biostatistician, Wyeth Research, Pearl River, NY 10965 (E-mail: [email protected])
2Dale L. Zimmerman is Professor, Department of Statistics and Actuarial Science, University
of Iowa, Iowa City, IA 52242 (E-mail: [email protected]; Phone: 319-335-0818; Fax: 319-
335-3017). Please address all correspondence to this author.
Abstract
The characterization of spatial dependence is an important component of a spatial modeling
exercise. For reasons of convenience, model parsimony, or computational efficiency, the spa-
tial covariance structure is often assumed to be isotropic (direction-invariant), completely
symmetric, or reflection symmetric (the latter two being forms of directional invariance some-
what weaker than isotropy). We propose some diagnostic tests of the latter two properties
that are based on the two-dimensional periodogram. An advantage of basing tests on the
periodogram rather than the sample semivariogram or covariance function is that the pe-
riodogram ordinates corresponding to different frequencies are asymptotically independent,
leading to simpler distribution theory. A simulation study and two examples illustrate the
usefulness and limitations of the proposed tests.
1 Introduction
For spatial and spatio-temporal data, characterization of the covariance structure is an im-
portant component of a modeling exercise. For reasons of convenience, model parsimony,
and computational efficiency, the covariance structure often is assumed to have some type
of directional invariance property. For a two-dimensional, second-order stationary random
process defined on a square lattice, three important directional invariance properties are re-
flection symmetry, complete symmetry, and isotropy, defined as follows. In these definitions
C(j, k) denotes the covariance between two random variables lagged j columns and k rows
apart in the lattice.
Definition. A second-order stationary random process on a square lattice is said to be
reflection symmetric if its covariance function satisfies
C(j, k) = C(−j, k) for all (j, k) ∈ Z2.
Definition. A second-order stationary random process on a square lattice is said to be
completely symmetric if its covariance function satisfies
C(j, k) = C(−j, k) = C(k, j) = C(−k, j) for all (j, k) ∈ Z2.
Definition. A second-order stationary random process on a square lattice is said to be
isotropic if its covariance function satisfies
C(j, k) = C0(√
j2 + k2) for all (j, k) ∈ Z2,
where C0(·) is the covariance function of a one-dimensional random process.
The usage of these terms is quite widespread, though not universal: Gneiting (2002), for
example, refers to a spatio-temporal version of reflection symmetry as “full symmetry.” Note
1
that isotropy implies complete symmetry, which in turn implies reflection symmetry, but the
converses are not true.
In spatial analyses in which the dependence is assumed to satisfy a directional symmetry
property, diagnostic checking for the validity of the assumption has been spotty. At best,
the checking may be done informally through such graphical diagnostics as direction-specific
sample semivariogram plots and the rose diagram (Isaaks and Srivastava, 1989, pp. 149-154)
or via a qualitative comparison of the estimates of C(j, k) to those of C(−j, k), (j, k) ∈ Z2
(Modjeska and Rawlings, 1983). At worst, the assumption is simply not checked at all.
There is a real need for diagnostics that can be accompanied by formal testing. A step in
this direction was taken by Lu (1994), who developed a method for formally testing for any
directional symmetry hypothesis that can be expressed as linear equality constraints on the
covariance function or semivariogram (as can all three types of symmetry defined above).
Lu’s approach was based on a statistic that measured how well the sample covariance function
(or semivariogram) satisfied these linear equality constraints.
In this paper, we propose tests for reflection symmetry and complete symmetry that
are based on certain ratios of periodogram ordinates. An advantage of basing tests on the
periodogram rather than the sample covariance function or semivariogram is that the pe-
riodogram ordinates corresponding to different frequencies are asymptotically independent,
leading to simpler distribution theory and simpler implementation of the tests. There is a
long history of using the periodogram of one-dimensional processes to construct tests for pe-
riodicity, independence, and other properties; see, for example, Fisher (1929), Durbin (1967),
and Priestley and Rao (1969). The approach taken in the present article is very similar in
spirit to that of these earlier authors.
Consideration of directionally symmetric covariance structures plays a role in both spatial
analysis and design. In spatial experimental design, for example, Modjeska and Rawlings
(1983) determined the optimal plot size and shape for field-plot experiments under a two-
2
dimensional extension of Smith’s (1938) empirical model relating soil heterogeneity to plot
size and shape, and showed that this two-dimensional model implied a reflection symmetric
correlation structure. Martin and Eccleston (1993) developed spatial experimental designs
that are optimal under reflection symmetry or complete symmetry.
The remainder of the article is organized as follows. In section 2 we review some necessary
background material on the spectral density and periodogram of a lattice process. Tests for
reflection symmetry and complete symmetry are developed in sections 3 and 4, respectively.
Section 5 presents results from a simulation study of the performance of the tests. Section
6 applies our methodology to data from two spatial experiments, and section 7 gives a brief
concluding discussion.
2 The Two-Dimensional Spectral Density and Peri-
odogram
Suppose that X = {X(u, v) : (u, v) ∈ Z2} is a real-valued random process, where the index
set Z2 is a square lattice with unit spacing in both dimensions. In addition we assume that
X is second-order stationary; that is, that
E(X(u, v)) = µ (a constant) (1)
and
cov(X(u + j, v + k), X(u, v)) = C(j, k) for all (u, v) ∈ Z2, (2)
with (2) holding for all (j, k) ∈ Z2. The covariance function, C(·, ·), necessarily is even and
nonnegative definite, i.e. C(−j,−k) = C(j, k) for all (j, k) ∈ Z2 and
∑nj=1
∑nk=1 ajakC(j, k) ≥
0 for any finite number n and any real numbers {aj : j = 1, . . . , n}.
We assume further that X has associated with it a spectral density function, defined as
f(λ, ω) =1
(2π)2
∞∑
j=−∞
∞∑
k=−∞
C(j, k) exp(−i(λj + ωk))
3
=1
(2π)2
∞∑
j=−∞
∞∑
k=−∞
C(j, k) cos(λj + ωk) (3)
where (λ, ω) ∈ (−π, π]2. The domain of f(·, ·) can be restricted to (−π, π]2 because f(·, ·) is
periodic with a period of 2π in the row and column directions of the lattice. The spectral
density also satisfies the following properties:
f(λ, ω) = f(−λ,−ω) for all (λ, ω) ∈ Z2, (4)
f(λ, ω) ≥ 0 for all (λ, ω) ∈ Z2, (5)
C(j, k) =
∫ π
−π
∫ π
−π
f(λ, ω) cos(λj + ωk) dλ dω for all (j, k) ∈ Z2. (6)
Suppose now that X is observed on a finite subset of Z2 forming an M (columns) by N
(rows) rectangular lattice. From these data, both the covariance function and spectral density
can be estimated. An estimator of the former at lags (j, k) (j = 0,±1,±2, . . . ,±(M − 1),
k = 0, 1, 2, . . . , N − 1) is the sample covariance function,
C(j, k) =
1MN
∑M−ju=1
∑N−kv=1 Y (u, v) Y (u + j, v + k), j ≥ 0
1MN
∑Mu=−j+1
∑N−kv=1 Y (u, v) Y (u + j, v + k), j < 0
(7)
where Y (u, v) = X(u, v)− 1MN
∑Mu=1
∑Nv=1 X(u, v). The spectral density is estimated by the
periodogram, denoted by I(λ, ω). For frequencies belonging to the set
Q = {(λp, ωq) : λp =2π
Mp, p = −[
M − 1
2], . . . , 0, . . . , [
M
2]
and ωq =2π
Nq, q = −[
N − 1
2], . . . , 0, . . . , [
N
2]}
(where [ · ] is the greatest integer function), the periodogram is defined as
I(λp, ωq) =1
(2π)2MN
∣∣∣∣∣
M∑
u=1
N∑
v=1
X(u, v) exp(−i(λpu + ωqv))
∣∣∣∣∣
2
=1
(2π)2MN
(M∑
u=1
N∑
v=1
X(u, v) (cos(λpu + ωqv) − i sin(λpu + ωqv))
)
·
(M∑
u=1
N∑
v=1
X(u, v) (cos(λpu + ωqv) + i sin(λpu + ωqv))
).
4
By taking the Fourier transform of (7), we can also express the periodogram at these fre-
quencies as
I(λp, ωq) =1
(2π)2
M−1∑
j=−M+1
N−1∑
k=−N+1
C(j, k) cos(λj + ωk).
At other frequencies in (−π, π]2, the periodogram is given by the following:
I(λ, ω) =
I(λp, ωq) if λp −πM
< λ ≤ λp + πM
, ωq −πN
< ω ≤ ωq + πN
, λ ≥ 0, ω ≥ 0
I(λp,−ω) if λp −πM
< λ ≤ λp + πM
, λ ≥ 0, ω < 0
I(−λ, ωq) if ωq −πN
< ω ≤ ωq + πN
, λ < 0, ω ≥ 0
I(−λ,−ω) if λ < 0, ω < 0.
This defines I(·, ·) as a piecewise constant function. Clearly, I(λp, ωq) = I(−λp,−ωq) and
thus I(λ, ω) = I(−λ,−ω), i.e. the periodogram is also an even function.
By a simple extension of a well-known one-dimensional result (e.g. Fuller, 1976, Theorem
7.1.1), it can be shown that if the covariance function of X is absolutely summable, then as
M, N → ∞,
E[I(0, 0)] −MN
(2π)2µ2 → f(0, 0)
where µ is given by (1), and
E[I(λ, ω)] → f(λ, ω) if (λ, ω) 6= (0, 0).
Thus, the periodogram is asymptotically unbiased for the spectral density at nonzero fre-
quencies. Furthermore, there is an important result on the asymptotic distribution of I(·, ·),
due to Pagano (1971), which we now give. In the theorem and subsequently, let V denote
the four-point set {(0, 0), (0, π), (π, 0), (π, π)}.
Theorem 1 (a) Let P = {(λj , ωj) ∈ (−π, π]2, j = 1, 2, . . . , J} be a set of distinct points
such that (λv, ωv) 6= (−λu,−ωu) for all (λu, ωu), (λv, ωv) ∈ P . If:
(i) the spectral density function f(·, ·) is continuous, and
5
(ii) for every (u, v) ∈ Z2 the representation
X(u, v) =∞∑
j=−∞
∞∑
k=−∞
a(j, k)Y (u − j, v − k)
holds, where the a(·, ·) are constants and the random variables Y (·, ·) are mutually
independent and satisfy
(MN)−1/2
M∑
m=1
N∑
n=1
Y (m, n)d→ N(0, 1)
as M, N → ∞;
then the periodogram ordinates {I(λ, ω) : (λ, ω) ∈ P} are asymptotically independent, and
I(λ, ω)
f(λ, ω)
d→ χ2
1 if (λ, ω) ∈ V \(0, 0)
2I(λ, ω)
f(λ, ω)
d→ χ2
2 if (λ, ω) ∈ (−π, π]2\V
as M, N → ∞.
3 Testing for Reflection Symmetry
In this section we propose using several functions of the periodogram ordinates as test statis-
tics for reflection symmetry. First, however, we establish the following essential theorem.
Theorem 2 Let X , C(·, ·), and f(·, ·) be defined as in section 2. Then X is reflection
symmetric if and only if f(λ, ω) = f(−λ, ω) for all (λ, ω) ∈ (−π, π]2.
Proof:
If X is reflection symmetric, i.e. C(j, k) = C(−j, k) for all (j, k) ∈ Z2, then for every
(λ, ω) ∈ (−π, π]2,
f(−λ, ω) =1
(2π)2
∞∑
j=−∞
∞∑
k=−∞
C(j, k) cos(−λj + ωk)
6
=1
(2π)2
∞∑
j=−∞
∞∑
k=−∞
C(−j, k) cos(λ(−j) + ωk)
=1
(2π)2
∞∑
j=−∞
∞∑
k=−∞
C(j, k) cos(λj + ωk)
= f(λ, ω)
where we have used (3). Conversely, if f(λ, ω) = f(−λ, ω) for all (λ, ω) ∈ (−π, π]2, then for
all (j, k) ∈ Z2,
C(−j, k) =
∫ π
−π
∫ π
−π
f(λ, ω) cos(−jλ + kω) dλ dω
=
∫ π
−π
∫ π
−π
f(−λ, ω) cos(−jλ + kω) dλ dω
=
∫ π
−π
∫ π
−π
f(λ, ω) cos(jλ + kω) dλ dω
= C(j, k)
where we have used (6). Q.E.D.
Theorem 2 indicates that the hypothesis of reflection symmetry can be expressed as either
H0 : C(j, k) = C(−j, k) for all (j, k) ∈ Z2
or
H0 : f(λ, ω) = f(−λ, ω) for all (λ, ω) ∈ (−π, π]2.
We use the periodogram to test this second representation of H0 against the alternative
hypothesis HA: not H0.
Only a subset of periodogram ordinates will be useful for this purpose. More specifically,
because (a) I(·, ·) is even, (b) I(λ, π) = I(−λ, π) for all λ and (c) I(π, ω) = I(π,−ω) for all
ω, we may restrict our attention to ordinates belonging to the set
S = {(λp, ωq) : λp =2π
Mp, p = −[
M − 1
2], . . . , [
M
2]; ωq =
2π
Nq, q = 0, . . . , [
N
2]}
\{(λ, ω) : λ < 0, ω = 0 or π}.
7
Figure 1 displays S when M = N = 7 and M = N = 8.
Now, when M and N are large, Theorem 1 indicates that the periodogram ordinates
{I(λ, ω) : (λ, ω) ∈ S} are approximately independent, and for (λ, ω) ∈ S\V ,
2I(λ, ω)
f(λ, ω)
·∼ χ2
2.
Consequently, for (λ, ω) ∈ S+, where S+ = S ∩ (0, π)2,
2I(λ, ω)/f(λ, ω)
2I(−λ, ω)/f(−λ, ω)=
f(−λ, ω)I(λ, ω)
f(λ, ω)I(−λ, ω)
·∼ F (2, 2). (8)
Denoting I(λ, ω)/I(−λ, ω) by R(λ, ω) and f(λ, ω)/f(−λ, ω) by cλω, we can express (8)
as
R(λ, ω)·∼ cλωF (2, 2).
Under H0, cλω = 1 (by Theorem 2) and thus R(λ, ω) is distributed approximately as a F (2, 2)
random variable. Moreover, under H0, {R(λ, ω): (λ, ω) ∈ S+} is a set of statistics that are
approximately independent and identically distributed as F (2, 2). Note that the cardinality
of {R(λ, ω) : (λ, ω) ∈ S+} is given by
nR =
[M − 1
2
]×
[N − 1
2
].
Several functions of {R(λ, ω) : (λ, ω) ∈ S+} could serve as test statistics for testing H0.
We consider three such functions here.
The first is simply the number of periodogram ratios that are “too large” or “too small,”
i.e.,
B = #(R(λ, ω) < Fξ/2(2, 2)
)+ #
(R(λ, ω) > F1−ξ/2(2, 2)
)
where 0 < ξ < 1 and Fξ(2, 2) is the 100ξth percentile of the F (2, 2) distribution. By the
asymptotic independence of the periodogram ratios and the fact that
Pr(R(λ, ω) < Fξ/2(2, 2)
)+ Pr
(R(λ, ω) > F1−ξ/2(2, 2)
) .= ξ
8
under H0 for all (λ, ω) ∈ S+, the asymptotic null distribution of B is binomial with sample
size parameter nR and success probability ξ, i.e., B·∼ bin(nR, ξ) under H0. Accordingly, an
approximately size-α test of H0 versus HA is to reject H0 if B > b, where b can be attained
by solving α = Pr(bin(nR, ξ) > b) for b.
The next two lemmas establish that this test, based on B, is asymptotically unbiased.
Lemma 1 Let
pλω = Pr(cλωF (2, 2) < Fξ/2(2, 2)
)+ Pr
(cλωF (2, 2) > F1−ξ/2(2, 2)
)
where 0 < ξ < 1. Then pλω ≥ ξ, with equality if and only if cλω = 1.
Proof: If X ∼ F (2, 2) then likewise 1/X ∼ F (2, 2), so
pλω = Pr(F (2, 2) > cλωF1−ξ/2(2, 2)
)+ Pr
(F (2, 2) >
1
cλω
F1−ξ/2(2, 2)
).
Now write ν = F1−ξ/2(2, 2). Because the distribution function of an F (2, 2) random variable
is given by
F (x) =x
1 + x, x > 0, (9)
we obtain the relation
ξ
2= Pr(X > ν) =
1
1 + ν.
This relation implies that
ν =2
ǫ− 1 > 1. (10)
Also we can write
pλω =1
1 + cλων+
1
1 + ν/cλω. (11)
Treating (11) as a function of cλω, we can write
g(cλω) =1
1 + cλων+
1
1 + ν/cλω
,
9
which is continuous and differentiable in cλω. Observe that g(1) = ξ, limcλω→0+ g(cλω) = 1
and limcλω→∞ g(cλω) = 1. Furthermore, it is easily shown that
g′(cλω) =ν
(1 + cλων)2(cλω + ν)2(ν2 − 1)(c2
λω − 1).
Thus, since ν > 1 by (10), we see that g(cλω) is strictly decreasing when cλω < 1 and strictly
increasing when cλω > 1, which implies further that g(cλω) attains its minimum value, namely
ξ, only at cλω = 1. Q.E.D.
Note that under HA, B is asymptotically distributed as the sum of nR independent
Bernoulli random variables with different success probabilities pλω, each of which is larger
than ξ by Lemma 1. This, and recursive use of the following lemma, implies that
Pr(B ≥ x|HA) > Pr(B ≥ x|H0), x = 0, . . . , nR.
That is, the test is asymptotically unbiased.
Lemma 2 Suppose that Yi, i = 1, . . . , n are independent Bernoulli(pi) random variables
with pi > 0 for all i and Wi, i = 1, . . . , n are independent Bernoulli(qi) random variables
with qi > 0. In addition, suppose that the Yi’s and Wi’s are independent. Let Si =∑i
t=1 Yt
and Ti =∑i
t=1 Wt. If pi = qi for i = 1, . . . , n − 1 and pn > qn, then
Pr(Sn ≥ x) > Pr(Tn ≥ x), x = 1, . . . , n.
Proof: Note that Pr(Sn−1 = s) = Pr(Tn−1 = s), s = 0, . . . , n − 1. Thus,