This action might not be possible to undo. Are you sure you want to continue?
WITH TRENDING DATA
BY
DEAN CORBAE, SAM OULIARIS
and PETER C. B. PHILLIPS
COWLES FOUNDATION PAPER NO. 1039
COWLES FOUNDATION FOR RESEARCH IN ECONOMICS
YALE UNIVERSITY
Box 208281
New Haven, Connecticut 065208281
2002
http://cowles.econ.yale.edu/
Econometrica, Vol. 70, No. 3 (May, 2002), 1067–1109
BAND SPECTRAL REGRESSION WITH TRENDING DATA
By Dean Corbae, Sam Ouliaris, and Peter C. B. Phillips
1
Band spectral regression with both deterministic and stochastic trends is considered. It
is shown that trend removal by regression in the time domain prior to band spectral regres
sion can lead to biased and inconsistent estimates in models with frequency dependent
coefﬁcients. Both semiparametric and nonparametric regression formulations are consid
ered, the latter including general systems of twosided distributed lags such as those aris
ing in lead and lag regressions. The bias problem arises through omitted variables and is
avoided by careful speciﬁcation of the regression equation. Trend removal in the frequency
domain is shown to be a convenient option in practice. An asymptotic theory is developed
and the two cases of stationary data and cointegrated nonstationary data are compared.
In the latter case, a levels and differences regression formulation is shown to be useful in
estimating the frequency response function at nonzero as well as zero frequencies.
Keywords: Band spectral regression, deterministic and stochastic trends, discrete
Fourier transform, distributed lag, integrated process, leads and lags regression, nonsta
tionary time series, twosided spectral BN decomposition.
1. introduction
Hannan’s (1963a, b) bandspectrum regression procedure is a useful
regression device that has been adopted in some applied econometric work,
notably Engle (1974), where there are latent variables that may be regarded as
frequency dependent (like permanent and transitory income) and where there is
reason to expect that the relationship between the variables may depend on fre
quency. More recently, band spectral regression has been used to estimate coin
tegrating relations, which describe low frequency or long run relations between
economic variables. In particular, Phillips (1991a) showed how frequency domain
regressions that concentrate on a band around the origin are capable of pro
ducing asymptotically efﬁcient estimates of cointegrating vectors. Interestingly,
in that case the spectral methods are used with nonstationary integrated time
series in precisely the same way as they were originally designed for stationary
series, on which there is now a large literature (see, inter alia, Hannan (1970 Ch.
VII), Hannan and Thomson (1971), Hannan and Robinson (1973), and Robinson
(1991)). Related methods were used in Choi and Phillips (1993) to construct fre
quency domain tests for unit roots. In all of this work, the capacity of frequency
domain techniques to deal nonparametrically with short memory components in
1
Our thanks go to a coeditor and three referees for helpful comments. An early draft was written
in the summer of 1994. The ﬁrst complete version of the paper was written over the period 1 January–
1 July 1997 while Phillips was visiting the University of Auckland. Lemmas A–D are due to Phillips
and are taken from other work to make this paper self contained. Phillips thanks the NSF for research
support under Grant Nos. SBR 9422922, SBR 9730295, and SES 0092509.
1067
1068 d. corbae, s. ouliaris, and p. c. b. phillips
the data is exploited. Robinson (1991), who allowed for band dependent model
formulations, showed how, in such semiparametric models, data based methods
can be used to determine the smoothing parameter that plays a central role in
the treatment of the nonparametric component. Other data based methods have
been used directly in the study of cointegrating relations and testing problems by
Xiao and Phillips (1998, 1999).
To the extent that relations between variables may be band dependent, it is
also reasonable to expect that causal relations between variables may vary accord
ing to frequency. Extending work by Geweke (1982), both Hosoya (1991) and
Granger and Lin (1995) have studied causal links from this perspective and con
structed measures of causality that are frequency dependent. Most recently, fre
quency band methods have been suggested for the estimation of models with long
memory components (e.g., Marinucci (1998), Robinson and Marinucci (1998),
Kim and Phillips (1999)), where salient features of the variables are naturally
evident in frequency domain formulations.
This paper studies some properties of frequency band regression in the pres
ence of both deterministic and stochastic trends, a common feature in economic
time series applications. In such cases, a seemingly minor issue relates to the
manner of deterministic trend removal. In particular, should the deterministic
trends be eliminated by regression in the time domain prior to the use of the
frequency domain regression or not? Since spectral regression procedures were
originally developed for stationary time series and since the removal of deter
ministic trends by least squares regression is well known to be asymptotically
efﬁcient under weak conditions (Grenander and Rosenblatt (1957, p. 244)), it
may seem natural to perform the trend removal in the time domain prior to the
use of spectral methods. Indeed, this is recommended in Hannan (1963a, p. 30;
1970, pp. 443–444), even though the development of spectral regression there
and in Hannan (1970, Ch. VII.4) allows for regressors like deterministic trends
that are susceptible to a generalized harmonic analysis, satisfying the socalled
Grenander conditions (Grenander and Rosenblatt (1957, p. 233), Hannan (1970,
p. 77)). Theorem 11 of Hannan (1970, p. 444) conﬁrms that such time domain
detrending followed by spectral regression is asymptotically efﬁcient in models
where all the variables are trend stationary and where the model coefﬁcients do
not vary with frequency.
In time domain regression, the Frisch–Waugh (1933) theorem assures invari
ance of the regression coefﬁcients to prior trend removal or to the inclusion
of trends in the regression itself. Such invariance is often taken for granted in
empirical work. However, as will be shown here, invariance does not always
apply for bandspectrum regression when one switches from time domain to fre
quency domain regressions. In particular, detrending by removing deterministic
components in the time domain and then applying bandspectrum regression is
not necessarily equivalent to detrending in the frequency domain and applying
bandspectrum regression. The reason is that a band dependent spectral model
is similar in form to a linear regression with a structural change in the coefﬁ
cient vector. So, when there are deterministic trends in the model, the original
band spectral regression 1069
deterministic regressors also need to be augmented by regressors that are rele
vant to the change period (here, the frequency band where the change occurs).
Thus, the seemingly innocuous matter of detrending can have some nontrivial
consequences in practice. In particular, detrending in the time domain yields esti
mates that can be biased in ﬁnite samples, and, in the case of nonstationary data,
inconsistent. The appropriate procedure is to take account of the augmented
regression equation prior to detrending and this can be readily accomplished by
including the deterministic variables explicitly in the frequency domain regres
sion. These issues are relevant whenever band spectral methods are applied,
including those cases that involve cointegrating regressions. In the latter case,
a levels and differences regression formulation is shown to be appropriate and
useful in estimating the frequency response function at nonzero as well as zero
frequencies.
The present paper provides an asymptotic analysis of frequency domain
regression with trending data for the two cases of stationary and cointegrated
nonstationary data. We consider both semiparametric and fully nonparametric
formulations of the regression model, in both cases allowing the model coefﬁ
cients to vary with frequency. In dealing with the nonstationary case, the paper
introduces some new methods for obtaining a limit theory for discrete Fourier
transforms (dft’s) of integrated time series and makes extensive use of two sided
BN decompositions, which lead to levels and differences model formulations.
This asymptotic theory simpliﬁes some earlier results given in Phillips (1991a)
and shows that the dft’s of an 1(1) process are spatially (i.e., frequencywise)
dependent across all the fundamental frequencies, even in the limit as the sam
ple size n →~, due to leakage from the zero frequency. This leakage is strong
enough to ensure that smoothed periodogram estimates of the spectrum away
from the origin are inconsistent at all frequencies w ,= 0. The techniques and
results given here should be useful in other regression contexts where these quan
tities arise. Some extensions of the methods to cases of data with long range
dependence are provided in Phillips (1999).
Most of our notation is standard: [a] signiﬁes the largest integer not exceeding
a, > signiﬁes positive deﬁniteness when applied to matrices, a
∗
is the complex
conjugate transpose of the matrix a, a is a matrix norm of a, a
−
is the Moore
Penrose inverse of a, F
a
=a(a
∗
a)
−
a
∗
is the orthogonal projector onto the range
of a, 1
2
0, 1] is the space of square integrable functions on [0, 1], 1[.] is the indi
cator of .,
J
∼ signiﬁes ‘asymptotically distributed as’ and
J
→ and
µ
→ are used to
denote weak convergence of the associated probability measures and convergence
in probability, respectively, as the sample size, n →~; 1(1) signiﬁes an integrated
process of order one, BM(U) denotes a vector Brownian motion with covariance
matrix U, and we write integrals like
1
0
E(r) Jr as
1
0
E, or simply
E if there
is no ambiguity over limits; MN(0, C) signiﬁes a mixed normal distribution with
matrix mixing variate C, A
c
(0, C) signiﬁes a complex normal distribution with
covariance matrix C, the discrete Fourier transform (dft) of {a
i
; i =1, . . . , n} is
written u
a
(\) = (1¡
√
n)
¸
n
i=1
a
i
e
i\i
, and {\
s
= (2=s¡n) : s = 0, 1, . . . , n−1} are
the fundamental frequencies.
1070 d. corbae, s. ouliaris, and p. c. b. phillips
2. models and estimation preliminaries
Let x
i
(i =1, . . . , n) be an observed time series generated by
x
i
=H
/
2
z
i
÷ ˜ x
i
, (1)
where z
i
is a µ ÷1dimensional deterministic sequence and ˜ x
i
is a zero mean
kdimensional time series. The series x
i
therefore has both a deterministic
component involving the sequence z
i
and a stochastic (latent) component ˜ x
i
. In
developing our theory it will be convenient to allow for both stationary and non
stationary ˜ x
i
. Accordingly, we introduce the following two alternative assump
tions. The mechanism linking the observed variables x
i
, unobserved disturbances
:
i
, and a dependent variable \
i
is made explicit later.
Assumption 1: :
i
=(:
i
, ˜ x
/
i
)
/
is a jointly stationary time series with Wold repre
sentation :
i
=
¸
~
]=0
C
]
£
i−]
, where £
i
=iid (0, l) with ﬁnite fourth moments and the
coefﬁcients C
]
satisfy
¸
~
]=0
]
1
2
C
]
  ~. Partitioned conformably with :
i
, the spec
tral density matrix ]
::
(\) of :
i
is
]
::
(\) =
¸
]
::
(\) 0
0 ]
xx
(\)
¸
,
with ]
::
(\), ]
xx
(\) > 0 ∀\.
Assumption 2: ˜ x
i
is an 1(1) process satisfying ú˜ x
i
= :
i
, initialized at i = 0
by any O
µ
(1) random variable. The shocks :
i
=(:
i
, :
/
i
)
/
satisfy Assumption 1 with
spectral density ]
::
(\) partitioned as
]
::
(\) =
¸
]
::
(\) 0
0 ]
::
(\)
¸
,
with ]
::
(\), ]
::
(\) > 0 ∀\.
Assumptions 1 and 2 sufﬁce for partial sums of :
i
to satisfy the functional law
n
−1¡2
¸
nr]
i=1
:
i
J
→E(r) =EM(U), a vector Brownian motion of dimension (k÷1)
with covariance matrix U = 2=]
::
(0) (e.g., Phillips and Solo (1992, Theorem
3.4)). The vector process E and matrix U can be partitioned conformably with :
as E =(E
:
, E
/
x
)
/
and
U=
¸
U
::
0
0 U
xx
¸
,
where U
xx
>0, so that ˜ x
i
is a full rank vector 1(1) process. In all cases, ˜ x
i
is taken
to be independent of :
i
. This amounts to a strict exogeneity assumption when we
introduce a generating mechanism for a dependent variable \
i
. The assumption
is standard for consistent estimation in the stationary case (as in Hannan (1963a,
b)). In the nonstationary case, it is not required for the consistent estimation of
band spectral regression 1071
(low frequency) cointegrating regression components. However, at bands away
from the origin, some version of incoherency between the regressors and errors
is required for consistency and asserting incoherency enables us to examine the
bias and inconsistency effects of detrending in such regressions, just as in the
case of stationary ˜ x
i
.
We make the following assumptions concerning the deterministic sequence z
i
.
We conﬁne our treatment to time polynomials and set z
i
= (1, i, . . . , i
µ
)
/
, the
most common case in practice, although we anticipate that our conclusions about
bias and inconsistency will apply more generally, for instance to trigonometric
polynomials, after some appropriate extensions of our dft formulae in Lemma A
and the limit results in Lemma C below. Let o
n
=diag(1, n, . . . , n
µ
), and deﬁne
J
i
=o
−1
n
z
i
. Then
J
nr]
=o
−1
n
z
nr]
→u(r) =(1, r, . . . , r
µ
)
/
, (2)
uniformly in r ∈ 0, 1]. The limit functions u(r) are linearly independent in
1
2
0, 1] and n
−1
(
¸
n
i=1
J
i
J
/
i
) →
1
0
uu
/
>0. Now let Z (respectively, D) be the nµ
observation matrix of the nonstochastic regressors z
i
(respectively, standardized
regressors J
i
), F
Z
be the projector onto the range of Z, and O
Z
=1
n
−F
Z
be the
residual projection matrix (respectively, F
D
and 1 −F
D
). Clearly, F
Z
= F
D
and
O
Z
=O
D
.
The Nonparametric Model
The type of model we have in mind for the stochastic component of the data
allows for the regression coefﬁcient to vary across frequency bands. In the time
domain, suppose a (latent) dependent variable ˜ \
i
is related to ˜ x
i
and :
i
according
to a (possibly twosided) distributed lag
˜ \
i
=
~
¸
]=−~
8
/
]
˜ x
i−]
÷:
i
=8(1)
/
˜ x
i
÷:
i
, (3)
where the transfer function of the ﬁlter, b(w) =8(e
iw
) =
¸
~
]=−~
8
]
e
i]w
, is assumed
to converge for all w ∈ −=, =]. Under stationarity (Assumption 1), the cross
spectrum ]
\x
(w) between ˜ \
i
and ˜ x
i
satisﬁes
]
\x
(w) =b(w)
/
]
xx
(w), (4)
whose complex conjugate transpose gives the complex regression coefﬁcient
8
w
=b(−w) =]
xx
(w)
−1
]
x\
(w). (5)
As argued in Hannan (1963a, b), spectral methods provide a natural mechanism
for focusing attention on a particular frequency, or band of frequencies, and for
performing estimation of 8
w
.
1072 d. corbae, s. ouliaris, and p. c. b. phillips
Next, suppose that the coefﬁcients 8
]
in (3) are such that 8(1) has a valid two
sided spectral BN decomposition (see part (b) of Lemma D in the Appendix),
viz.
8(1) =8(e
iw
) ÷
˜
8
w
(e
−iw
1)(e
−iw
1−1), (6)
where
˜
8
w
(1) =
~
¸
]=−~
˜
8
w]
1
]
,
˜
8
w]
=
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
~
¸
k=]÷1
8
k
e
ikw
, ] ≥0,
−
]
¸
k=−~
8
k
e
ikw
, ]  0,
with
¸
~
]=−~

˜
8
]

2
~. Using (6) in (3) we have
˜ \
i
=8(1)
/
˜ x
i
÷:
i
(7)
=
¸
b(w) ÷
˜
8
w
(e
−iw
1)(e
−iw
1−1)
¸
/
˜ x
i
÷:
i
,
=b(w)
/
˜ x
i
÷(e
−iw
1−1) ˜ a
wi
÷:
i
,
in which ˜ a
wi
=
˜
8
w
(e
−iw
1) ˜ x
i
. Then, setting w = \
s
in (7) and taking dft’s at fre
quency \
s
we obtain the following frequency domain form of (3):
u
˜ \
(\
s
) =b(\
s
)
/
u
˜ x
(\
s
) ÷u
:
(\
s
) ÷
1
√
n
¸
e
−i\
s
˜ a
\
s
0
− ˜ a
\
s
n
¸
. (8)
For stationary ˜ x
i
, ˜ a
wi
is also stationary and (8) can be written as
u
˜ \
(\
s
) =b(\
s
)
/
u
˜ x
(\
s
) ÷u
:
(\
s
) ÷O
µ
1
√
n
, (9)
giving an approximate frequency domain version of (3).
The case of integrated ˜ x
i
(Assumption 2) can be handled in a similar way.
Here it is useful to apply to (3) the twosided BN decomposition at unity (part(a)
of Lemma D), viz.
8(1) =8(1) ÷
˜
8(1)(1−1), where (10)
˜
8(1) =
~
¸
]=−~
˜
8
]
1
]
,
˜
8
]
=
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
~
¸
k=]÷1
8
k
, ] ≥0,
−
]
¸
k=−~
8
k
, ]  0,
so that (3) can be interpreted as
˜ \
i
=8(1)
/
˜ x
i
÷:
i
−
˜
8(1)
/
:
i
=8(1)
/
˜ x
i
÷ ˜ :
i
, say, (11)
band spectral regression 1073
which is a cointegrating equation between ˜ \
i
and ˜ x
i
with cointegrating vector
(1, −8(1)
/
) and composite error ˜ :
i
=:
i
−
˜
8(1)
/
:
i
. Setting ¬
i
=
˜
8(1)
/
:
i
and taking
the dft of (11), we obtain
u
˜ \
(\
s
) =8(1)
/
u
˜ x
(\
s
) ÷u
:
(\
s
) −u
¬
(\
s
) (12)
=b(\
s
)
/
u
˜ x
(\
s
) ÷u
:
(\
s
) −u
¬
(\
s
) ÷8(1) −b(\
s
)]
/
u
˜ x
(\
s
). (13)
The ﬁnal two terms of (13) are important, revealing that the approximation (9)
fails in the nonstationary case and has an error of speciﬁcation that arises from
omitted variables.
Assuming it is valid, the spectral BN decomposition of
˜
8(1) at frequency \
s
is
˜
8(1) =
˜
8(e
i\
s
)÷
≈
8
\
s
(e
−i\
s
1)(e
−i\
s
1−1)
where
≈
8
\
s
is constructed as in Lemma D(b). Then
u
¬
(\
s
) =
˜
8(e
i\
s
)
/
u
:
(\
s
) ÷
1
√
n
¸
e
−i\
s
˜ :
\
s
0
− ˜ :
\
s
n
¸
(14)
=
˜
8(e
i\
s
)
/
u
:
(\
s
) ÷O
µ
1
√
n
where ˜ :
\
s
i
=
≈
8
\
s
(e
−i\
s
1):
i
is stationary. It follows that (12) has the form
u
˜ \
(\
s
) =8(1)
/
u
˜ x
(\
s
) −
˜
8(e
i\
s
)
/
u
:
(\
s
) ÷u
:
(\
s
) ÷O
µ
1
√
n
, (15)
giving an approximate frequency domain version of the cointegrated model that
applies for all frequencies. Since :
i
=ú˜ x
i
, it is apparent from (15) that if data on
˜ \
i
and ˜ x
i
were available, the equation could be estimated in the frequency domain
by band spectral regression of u
˜ \
(\
s
) on u
˜ x
(\
s
) and u
ú˜ x
(\
s
) for frequencies \
s
centered around some frequency w. For \
s
centered around w=0, this procedure
was suggested originally in Phillips (1991b) and Phillips and Loretan (1991),
where it is called augmented frequency domain regression, although in that work
deterministic trends were excluded. In the general case where \
s
is centered on
w ,=0, we can recover the spectral coefﬁcient 8
w
=b(−w) from the coefﬁcients
8(1) and
˜
8(e
i\
s
)
/
=
˜
8(e
−i\
s
)
∗
in (15) using the fact that
b(−w) =8(e
−iw
) =8(1) ÷
˜
8(e
−iw
)(e
−iw
−1). (16)
Causal economic relationships are often formulated as in (3) with onesided
relations where 8
]
=0 for ]  0, but twosided relations are not excluded. They
are well known to occur in situations where there is feedback in both directions
between the variables ˜ \ and ˜ x. A notable recent example arises in the simpliﬁ
cation of cointegrating systems to single equation formulations where twosided
1074 d. corbae, s. ouliaris, and p. c. b. phillips
dynamic regressions like (11) arise and where they have been used in regres
sion analysis to obtain efﬁcient estimates of the cointegrating vector (Phillips and
Loretan (1991), Saikonnen (1991), Stock and Watson (1993)).
The mechanism linking the observed dependent variable \
i
to z
i
and ˜ \
i
(and,
hence, ˜ x
i
, and :
i
) is now given by
\
i
=z
/
i
=
1
÷ ˜ \
i
, (17)
where ˜ \
i
follows (3) in the stationary case with ( ˜ x
i
, :
i
) satisfying Assumption 1,
and (11) in the cointegrated case with ( ˜ x
i
, :
i
) satisfying Assumption 2. The
observed series \
i
and x
i
therefore have deterministic trends and stochastic com
ponents that are linked by a system of distributed lags with stationary errors.
The natural statistical approach would appear to be deterministic trend removal
by regression of (\
i
, x
i
) on z
i
, followed by an analysis of the system of lagged
dependencies between the residuals from these regressions. The latter can be
conducted in the frequency domain where the coefﬁcient of interest is the trans
fer function of the ﬁlter b(w)
/
. This coefﬁcient will vary according to frequency
(unless 8
]
=0 for ] ,=0) and can be estimated by nonparametric regression tech
niques. One of the aims of this paper is to examine the ﬁnite sample and asymp
totic performance of this natural procedure against that of a procedure that
seeks to perform the regression analysis in the frequency domain directly on the
observed quantities (\
i
, x
i
, z
i
) rather than on detrended data.
A Fixed Band Regression Model
A simple situation of interest occurs when the response function b(w) is con
stant, let us say on two ﬁxed frequency bands. This case captures the essential
features of the problem that we intend to discuss, is easy to analyze, and will be
considered at some length here and in the asymptotic theory. It has the advan
tage that b(w) is real and parametric and can be estimated at parametric rates
when the bands are of ﬁxed positive length. This case is analogous to a regression
model with a structural break and that analogy helps to explain why time and fre
quency domain detrending have different effects. The nonparametric model can
be interpreted as a limiting case in which one of the bands shrinks to a particular
frequency as the sample size increases. To promote the analogy with a structural
break model we will develop a conventional regression model formulation for
the data.
To this end, we postulate the dual bands
.
= −w
0
, w
0
] and
c
.
= −=,
−w
0
)∪(w
0
, =] for some given frequency w
0
>0 and deﬁne the frequency depen
dent coefﬁcient vector
b(w) =8
.
1w ∈
.
] ÷8
.
c 1w ∈
c
.
], (18)
where 8
.
is a kvector of parameters pertinent to the band
.
, and 8
.
c is the
corresponding kvector of parameters pertinent to the band
c
.
. This formulation
of b(w) enables us to separate low frequency ([w[ ≤ w
0
) from high frequency
band spectral regression 1075
([w[ > w
0
) responses in a regression context. The coefﬁcients 8
]
in the Fourier
representation, b(w) =
¸
~
]=−~
8
]
e
i]w
, of (18) are
8
]
=
1
2=
=
−=
b(w)e
−i]w
Ju =
¸
¸
¸
¸
¸
¸
8
.
w
0
=
÷
8
.
c
=
(= −w
0
), ] =0,
sin]w
0
=]
(8
.
−8
.
c ), ] ,=0,
(19)
so that the ﬁlter in (3) is symmetric and twosided. This is an interesting case
where the coefﬁcients 8
]
decay slowly and do not satisfy the sufﬁcient condition
¸
~
−~
[][
1
2
[8
]
[ ~ given in Lemma D for the validity of the BN decomposition of
8(1) =
¸
~
]=−~
81
]
. Nevertheless, the BN decomposition of 8(1) is still valid in
this case, as shown in Remark (i) following Lemma D in the Appendix.
While data for ˜ \
i
can be generated by (3) and (19) by truncating the ﬁlter, an
alternate frequency domain approach that uses (18) is possible. This mechanism
has the advantage of revealing the impact of the shift in (18) in terms of a stan
dard regression model with a structural change. Like x
i
, the dependent variable
\
i
is assumed to have both deterministic and stochastic components. Its stochas
tic component ˜ \
i
is generated from ˜ x
i
and :
i
by means of a triangular array
formulation which we explain as follows.
Let
¯
A =  ˜ x
1
, . . . , ˜ x
n
]
/
be the nk matrix of observations of the exogenous
regressors ˜ x
i
, let U be the nn unitary matrix with (], k) element e
2=i]k¡n
¡
√
n,
and let W =1
1n
U, where
1
1n
=
¸
0 1
1
n−1
0
¸
is an n n permutation matrix that reorders the rows of U so that the last
row becomes the ﬁrst and succeeding rows are simply advanced by one position.
Then, W
∗
W =1
n
, and the matrix W
¯
A has kth row given by the dft u
˜ x
(\
s
)
/
with
s =k−1.
We introduce the following matrices that are used to provide a frequency
dependent structure to the model for \
i
. Deﬁne:
• .=nn selector matrix that zeroes out frequencies in W
¯
A that are not
relevant to the primary band of interest, say
.
. Then, .
c
W =1 −.]W extracts
the residual frequencies over
c
.
. Note that .
c
.=..
c
=0.
• 4 =W
∗
.W and 4
c
=W
∗
.
c
W =1 −4.
Then, for each n, ˜ \ =( ˜ \
1
, . . . , ˜ \
n
) is generated in the frequency domain in dft
form by the system
.W ˜ \ =.W
¯
A8
.
÷.W:, (20)
.
c
W ˜ \ =.
c
W
¯
A8
.
c ÷.
c
W:, (21)
where : = (:
1
, . . . , :
n
)
/
. Equations (20) and (21) generate observations in the
frequency domain that correspond to the model (8) in the nonparametric case
1076 d. corbae, s. ouliaris, and p. c. b. phillips
with the ﬁnal term of (8) omitted. In the stationary case, this omitted term is
O
µ
(n
−
1
2
) as seen in (9), so the difference between the generating mechanisms
is negligible asymptotically. In the nonstationary case, however, the omitted term
is nonnegligible, as is apparent from (13) and (15). In consequence, the two gen
erating mechanisms differ in the nonstationary case and this leads to a difference
in the asymptotic theory between the two models that will ﬁgure in our discus
sion in Section 5.
Adding (20) and (21) and multiplying by W
∗
gives
˜ \ =4
¯
A8
.
÷4
c
¯
A8
.
c ÷:, (22)
which is the time domain generating mechanism for ˜ \. By construction, the matri
ces 4 and 4
c
have elements that depend on n, and it follows that { ˜ \
i
: i =
1, . . . , n} has a triangular array structure, although we do not emphasize this by
using an additional subscript and we will have no need of triangular array limit
theory in our asymptotic development.
In (22), (:
i
, ˜ x
/
i
) satisﬁes Assumption 1 in the stationary case and (:
i
, ú˜ x
/
i
) sat
isﬁes Assumption 2 in the nonstationary case. Thus, (22) is a variable (band
dependent) coefﬁcient time series regression with exogenous regressors and sta
tionary errors under Assumption 1, and a cointegrating regression with exoge
nous regressors and band dependent coefﬁcients under Assumption 2. Note that
model (22) is semiparametric: parametric in the regression coefﬁcients 8
.
and
8
.
c and nonparametric in the regression errors :.
We now suppose that the observed series \
i
has deterministic components like
x
i
in (1) and is related to the unobserved component ˜ \
i
by means of
\ =Z=
1
÷ ˜ \. (23)
Using (22) and (23), it follows that the observed data satisfy the model
\ =Z(=
1
−H
2
8
.
) ÷A8
.
÷4
c
ZH
2
(8
.
−8
.
c ) −4
c
A(8
.
−8
.
c ) ÷:, (24)
or, equivalently,
\ =Z(=
1
−H
2
8
.
c ) ÷4ZH
2
(8
.
c −8
.
) ÷A8
.
c −4A(8
.
c −8
.
) ÷:, (25)
or
\ =4Z(=
1
−H
2
8
.
) ÷4
c
Z(=
1
−H
2
8
.
c ) ÷4A8
.
÷4
c
A8
.
c ÷:. (26)
These models extend (22) to cases where the observed data contain deterministic
trends. Observe that the trend regressors z
i
now appear in the model with band
dependent coefﬁcients, just like the exogenous regressors x
i
.
The formulations (24)–(26) make it clear that detrending the data in the time
domain is not a simple matter of applying the projection matrix O
Z
, as might
be expected immediately from (1) and (23). In fact, correct trend removal is
accomplished by the use of the operator O
I
= 1 −F
I
, where I = Z, 4
c
Z] or,
band spectral regression 1077
equivalently, in view of (26) I = 4Z, 4
c
Z]. Methods that rely on preﬁltering
by means of O
Z
do not fully remove the trends and this can have important
consequences, like biased and inconsistent estimates of the regression coefﬁcients
8
.
and 8
.
c . Of course, when . = 1
n
the coefﬁcient is invariant across bands
(8
.
=8, say), we have 4 =1
n
and 4
c
is null, so that I =Z and usual detrending
by O
Z
is appropriate. In that case (26) reduces to
\ =Z(=
1
−H
2
8) ÷A8÷:.
Put another way, conventional detrending by O
Z
implicitly assumes that there is
no variation in the coefﬁcient across frequency bands. If there are such variations,
then correct detrending needs to take into account the effects of such variations
on the trends, as in (26). Otherwise, there will be misspeciﬁcation bias from
omitted variables in the deterministic detrending.
Fixed Band Regressions and Misspeciﬁcation Bias
A simple illustration with the Hannan (1963a) inefﬁcient bandspectrum
regression estimator shows the effects of such misspeciﬁcation. This estimator
can be constructed for the band
.
and then has the form
ˆ
8
.
=(A
/
O
Z
4O
Z
A)
−1
(A
/
O
Z
4O
Z
\), (27)
with a corresponding formula for
ˆ
8
.
c , the estimator over the band
c
.
. In form
ing
ˆ
8
.
and
ˆ
8
.
c , the data are ﬁltered by a trend removal regression via the projec
tion O
Z
before performing the bandspectrum regression. This procedure follows
Hannan’s (1963a, b) recommendation for dealing with deterministic trends and,
as we have discussed in the introduction, is the conventional approach in this
context. Using (24) and (27), we ﬁnd
ˆ
8
.
=8
.
÷{A
/
O
Z
4O
Z
A}
−1
{A
/
O
Z
4O
Z
4
c
ZH
2
(8
.
−8
.
c ) (28)
−4
c
A(8
.
−8
.
c ) ÷:]}
=8
.
−{A
/
O
Z
4O
Z
A}
−1
{A
/
O
Z
4O
Z
4
c
¯
A(8
.
−8
.
c ) −:]}
=8
.
−{
¯
A
/
O
Z
4O
Z
¯
A}
−1
{
¯
A
/
O
Z
4O
Z
4
c
¯
A(8
.
−8
.
c ) −:]},
and the corresponding formula for
ˆ
8
.
c is
ˆ
8
.
c =8
.
c −{
¯
A
/
O
Z
4
c
O
Z
¯
A}
−1
{
¯
A
/
O
Z
4
c
O
Z
4
¯
A(8
.
c −8
.
) −:]}. (29)
It follows that
1(
ˆ
8
.
[A) =8
.
−{A
/
O
Z
4O
Z
A}
−1
{A
/
O
Z
4O
Z
4
c
¯
A(8
.
−8
.
c )}, (30)
and
1(
ˆ
8
.
c [A) =8
.
c −{A
/
O
Z
4
c
O
Z
A}
−1
{A
/
O
Z
4
c
O
Z
4
¯
A(8
.
c −8
.
)}.
1078 d. corbae, s. ouliaris, and p. c. b. phillips
Hence, bandspectrum regression yields biased estimates of the coefﬁcients when
8
.
,=8
.
c and trend removal regression is conducted via the projection O
Z
.
To some extent, the problem with
ˆ
8
.
is a ﬁnite sample one. In fact, for sta
tionary ˜ x
i
, the bias disappears as n →~. But, as we shall see in Section 5, when
˜ x
i
is integrated the bias in
ˆ
8
.
c does not always disappear in the limit.
Narrow Band Regressions
In the nonparametric case we estimate 8
w
in (5) at a particular frequency w,
rather than over a ﬁxed discrete band. In that case the band spectral estimator
has the same form as (27) but the selector matrix .=.
w
(respectively, 4 =4
w
)
chooses frequencies in a shrinking band about w. We write the estimator as
ˆ
8
w
=(A
/
O
Z
4
w
O
Z
A)
−1
(A
/
O
Z
4
w
O
Z
\). (31)
From the form of (15) it is apparent that the narrow band regression leading
to (31) omits the term involving the dft, u
:
(\
s
), of the differences :
i
= ú˜ x
i
. So
this narrow band regression also involves omitted variable misspeciﬁcation.
An alternate nonparametric procedure is to use an augmented narrow band
regression in which the differences are included. To ﬁx ideas, we detrend the
data using O
Z
and then perform a band spectral regression of the form
u
\.z
(\
s
) = ˜ a
/
1w
u
x.z
(\
s
) ÷ ˜ a
/
2w
u
úx.z
(\
s
) ÷residual (32)
for frequencies \
s
in a shrinking band around w. In equation (32) the afﬁx ‘.z’ on
the subscript (e.g., in u
x.z
(\
s
)) signiﬁes that the data have been detrended before
taking dft’s. This approach is analogous to the augmented spectral regression
approach of Phillips and Loretan (1991) for estimating a cointegrating vector.
However, the narrow band regression (32) now applies for frequencies w that
are nonzero as well as zero and prior detrending of the data has occurred in the
time domain. Deﬁne A =A, úA] and ˜ a
/
w
=( ˜ a
/
1w
, ˜ a
/
2w
). Then
˜ a
w
=(A
/
O
Z
4
w
O
Z
A)
−1
(A
/
O
Z
4
w
O
Z
\). (33)
In view of the relation (16), an estimate of 8
w
=8(e
−iw
) may be recovered from
˜ a
w
using the linear combination
˜
8
w
=
¸
˜ a
1, −w
−(e
−iw
−1) ˜ a
2, −w
, w ,=0,
˜ a
10
, w =0.
(34)
The asymptotic theory for the narrow band estimates
ˆ
8
w
and
˜
8
w
is developed
in Section 5.
3. dft recursion formulae
Here we develop some useful formulae for dft’s of the deterministic sequence
z
i
. The following lemma gives general recursion formulae for the quantity
W
k
(\
s
) =
¸
n
i=1
i
k
e
i\
s
i
. The case for \
s
=0 is well known, of course, but the result
for \
s
,=0 appears to be new.
band spectral regression 1079
Lemma A: (a) For s =0,
W
k
(\
s
) =
n
¸
i=1
i
k
=
1
k÷1
¸
n
k÷1
−
k÷1
1
E
1
n
k
÷
k÷1
2
E
2
n
k−1
÷
k÷1
3
E
3
n
k−2
÷· · · ÷
k÷1
k
E
k−1
n
¸
,
where E
]
are the Bernoulli numbers.
(b) For s ,=0, W
k
(\
s
) satisﬁes the recursion
W
k
(\
s
) =n
k
e
i\
s
e
i\
s
−1
÷
1
e
i\
s
−1
k
¸
]=1
k
]
(−1)
]
W
k−]
(\
s
) (35)
with initialization
W
0
(\
s
) =
n
¸
i=1
e
i\
s
i
=
¸
n, s =0,
0, s ,=0.
(36)
Using (35), the standardized quantities W
k
(\
s
) =
1
√
n
¸
n
i=1
i
n
k
e
i\
s
i
satisfy the
recursion
W
k
(\
s
) =
1
√
n
e
i\
s
e
i\
s
−1
÷
1
e
i\
s
−1
k
¸
]=1
1
n
]
k
]
(−1)
]
W
k−]
(\
s
), (37)
and then the dft of J
i
is simply
u
J
(\
s
) =
1
√
n
n
¸
i=1
J
i
e
i\
s
i
=
W
0
(\
s
), . . . , W
µ
(\
s
)
/
.
The main cases of interest are low order polynomials, where explicit expres
sions for the discrete Fourier transforms W
k
(\
s
) are easily obtained from
Lemma A. Thus, when k =0, 1, 2 we have
W
0
(\
s
) =
¸
√
n, s =0,
0, s ,=0,
(38)
W
1
(\
s
) =
¸
¸
¸
¸
¸
¸
n÷1
2n
1¡2
, s =0,
1
n
1¡2
e
i\
s
e
i\
s
−1
, s ,=0,
(39)
W
2
(\
s
) =
¸
¸
¸
¸
¸
¸
(n÷1)(2n÷1)
6n
3¡2
, s =0,
1
n
1¡2
e
i\
s
e
i\
s
−1
−
2
n
3¡2
e
i\
s
(e
i\
s
−1)
2
, s ,=0.
(40)
1080 d. corbae, s. ouliaris, and p. c. b. phillips
In case (38), it is apparent that eliminating the zero frequency will demean
the data and leave the model unchanged for \
s
,= 0. Then, 4
c
Z = 0 and so
4O
Z
4
c
=44
c
−4F
z
4
c
=0, and O
Z
=O
I
. It follows that A
/
O
Z
4O
Z
4
c
¯
A=0
in (30) and therefore
ˆ
8
.
is unbiased in this case of simple data demeaning.
On the other hand, when z
/
i
=(1, i) and J
/
i
=(1, i¡n), it follows from (38) and
(39) that
u
J
(\
s
)
/
=
1
√
n
n
¸
i=1
J
/
i
e
i\
s
i
=
¸
¸
¸
¸
¸
¸
√
n,
n÷1
2n
1¡2
, s =0,
0,
1
n
1¡2
e
i\
s
e
i\
s
−1
, s ,=0,
(41)
and the second component of u
J
(\
s
) has nonzero elements for s ,= 0. Hence,
4
c
Z ,= 0 and O
Z
,= O
I
, so that
ˆ
8
.
is biased in this case. The same applies
for higher order trends.
Frequencies in the band
c
.
satisfy [\
s
[ > w
0
> 0. It therefore follows from
Lemma A and (37) that W
k
(\
s
) =O(n
−
1
2
) uniformly for \
s
∈
c
.
. Hence,
u
J
(\
s
) =O(n
−1¡2
) for all \
s
∈
c
.
(42)
and, thus, 4
c
D =W
∗
.
c
WD has elements that are of O(n
−1¡2
).
4. frequency domain detrending
We consider ﬁrst the ﬁxed band regression model (22) and (23). The problem
of the omitted variable bias evident in (28) can be dealt with either in the fre
quency domain or in the time domain. In the time domain, one simply detrends
using the projection operator O
I
=1 −F
I
, where I =4Z, 4
c
Z], as is appar
ent from the form of (26). In the frequency domain, the alternative is to leave
the detrending until the regression is performed in the frequency domain.
To do frequency domain detrending, we simply apply the discrete Fourier
transform operator W to (24) and then perform the band spectrum regression.
The transformed model is
W\ =WZ(=
1
−H
2
8
.
) ÷W4
c
ZH
2
(8
.
−8
.
c ) ÷WA8
.
−W4
c
A(8
.
−8
.
c ) ÷W:
=WZ(=
1
−H
2
8
.
) ÷.
c
WZH
2
(8
.
−8
.
c ) ÷WA8
.
−.
c
WA(8
.
−8
.
c ) ÷W:.
The resulting band spectral estimator for the band
.
is equivalent to a regres
sion on
.W\ =.WZ(=
1
−H
2
8
.
) ÷.WA8
.
÷.W:, (43)
since ..
c
=0, and therefore this estimator has the form
ˆ
8
]
.
=(A
/
W
∗
.O
.WZ
.WA)
−1
(A
/
W
∗
.O
.WZ
.W\) (44)
=8
.
÷(A
/
W
∗
.O
.WZ
.WA)
−1
(A
/
W
∗
.O
.WZ
.W:).
band spectral regression 1081
Clearly, 1(
ˆ
8
]
.
[A) =8
.
, and the estimator is unbiased. A similar result holds for
the corresponding estimator
ˆ
8
]
.
c
of 8
.
c .
In this frequency domain approach to detrending, the so called Frisch–Waugh
(1933) theorem clearly holds, i.e., the regression coefﬁcient
ˆ
8
]
.
on the variable
AWX in (43) is invariant to whether the regressor AWZ is included in the regres
sion or whether all the data have been previously detrended in the frequency
domain by regression on AWZ.
Following (32), the natural narrow band approach is to use an augmented
regression model that includes the dft of the trend in the regression, viz.
u
\
(\
s
) = ˜ c
/
1w
u
x
(\
s
) ÷ ˜ c
/
2w
u
úx
(\
s
) ÷ ˜ c
/
3w
u
z
(\
s
) ÷residual, (45)
for frequencies \
s
centered on w. This narrowband regression leads to the estimate
˜
8
]
w
=
¸
˜ c
1, −w
−(e
−iw
−1) ˜ c
2, −w
, w ,=0,
˜ c
10
, w =0,
(46)
similar to (34).
5. asymptotic theory
We derive a limit distribution theory for the detrended band spectral regres
sion estimates and consider what happens to the bias as n → ~. We start by
introducing some notation and making the framework for the limit theory more
precise.
We start with the parametric case where there are the two discrete bands
.
and
.
c . Let n
a
= #{\
s
∈
.
} and n
c
= #{\
s
∈
.
c } be the number of
fundamental frequencies in the bands
.
and
.
c . It is convenient to subdivide
−=, =] into subbands
]
of equal width (say, =¡J ) that center on frequencies
{w
]
==]¡J : ] =−J ÷1, . . . , J −1}. Let n=#(\
s
∈
]
) and suppose that J
a
of
these bands lie in
.
. It follows that n and n
a
can be approximated by n =2nJ
and n
a
=2nJ
a
, respectively.
In the nonparametric case, we focus on a single frequency w and consider a
shrinking band
w
of width =¡J centered on w. Again, we let n=#(\
s
∈
w
).
The following condition will be useful in the development of the asymptotics
and will be taken to hold throughout the remainder of the paper. Additional
requirements will be stated as needed.
Assumption 3: (a) n
a
¡n → 0 and n
c
¡n → 1 −0 for some ﬁxed number 0 ∈
0, 1] as n →~.
(b) n, J →~, and J ¡
√
n →0 as n →~.
For the bias in
ˆ
8
.
to vanish asymptotically, the deviation that depends on the
term
{A
/
O
Z
4O
Z
A}
−1
{A
/
O
Z
4O
Z
4
c
¯
A(8
.
−8
.
c )}
1082 d. corbae, s. ouliaris, and p. c. b. phillips
in (28) needs to disappear as n →~. We will distinguish the two cases of sta
tionary and nonstationary (integrated) ˜ x
i
corresponding to Assumptions 1 and 2
in the following discussion.
5.1. The Stationary Case
Here, the bias in
ˆ
8
.
will disappear when
A
/
O
Z
4O
Z
A
n
−1
A
/
O
Z
4O
Z
4
c
¯
A
n
µ
→0. (47)
A similar requirement, obtained by interchanging 4 and 4
c
in (47), holds for
the bias in
ˆ
8
.
c . We have the following result.
Theorem 1 (Semiparametric Case): If ˜ x
i
and :
i
are zero mean, stationary,
and ergodic time series satisfying Assumption 1, and ˜ \
i
is generated by (22), band
spectral regression with detrending in the time domain or in the frequency domain
is consistent for both 8
.
and 8
.
c . The common limit distribution of
ˆ
8
.
and
ˆ
8
]
.
is
given by
√
n(
ˆ
8
.
−8
.
),
√
n(
ˆ
8
]
.
−8
.
)
J
→A(0, I
a
),
where
I
a
=
.
]
xx
(w) Jw
−1
2=
.
]
xx
(w)]
::
(w) Jw
.
]
xx
(w) Jw
−1
, (48)
with an analogous result for
ˆ
8
.
c and
ˆ
8
]
.
c
.
In the stationary case, therefore, the bias in band spectral regression from time
domain detrending disappears as n →~ and there is no difference between the
two bands
.
and
c
.
in terms of the limit theory. It is therefore irrelevant
whether the main focus of interest is high or low frequency regression. The form
of the asymptotic covariance matrix I
a
is a band spectral version of the familiar
sandwich formula for the robust covariance matrix in least squares regression.
I
a
can be estimated by replacing the spectra in the above formula with cor
responding consistent estimates and averaging over the band
.
. For the full
band case where
.
=−=, =], the matrix I
a
is the well known formula for the
asymptotic covariance matrix of the least squares regression estimator in a time
series regression (c.f. Hannan (1970, p. 426)). A formula related to I
a
was given
in Hannan (1963b, equation (16)) for band spectral estimates in the context of
models with measurement error.
A similar result holds in the nonparametric case, but with a different conver
gence rate.
band spectral regression 1083
Theorem 1
/
(Nonparametric Case): If ˜ x
i
and :
i
are zero mean, stationary,
and ergodic time series satisfying Assumption 1, and ˜ \
i
is generated by (3), non
parametric band spectral regression with detrending in the time domain or in the
frequency domain is consistent for 8
w
= b(−w). The common limit distribution of
ˆ
8
w
and
ˆ
8
]
w
is given by
√
n(
ˆ
8
w
−8
w
),
√
n(
ˆ
8
]
w
−8
w
)
J
→A
c
(0, I
w
),
where
I
w
=]
::
(w)]
xx
(w)
−1
.
5.2. The Nonstationary Case
Here, the distinction between low frequency regression and regression at other
frequencies becomes important. In the band regression model (22), there are
only two bands
.
and
c
.
. Over
.
, which includes the zero frequency, the
estimator is known to be nconsistent when 8
.
=8
.
c (see Phillips (1991a)) since
in that case, the regression equation is a conventional cointegrating relation.
When 8
.
,=8
.
c , the same result continues to hold over the band
.
, as shown
in Theorem 2 below. In this case, the bias in (28) disappears when
A
/
O
Z
4O
Z
A
n
2
−1
A
/
O
Z
4O
Z
4
c
¯
A
n
2
µ
→0. (49)
Note that in (49) the moment matrices are standardized by n
2
, because the data
nonstationarity is manifest in bands like
.
that include the zero frequency. Over
frequency bands like
c
.
that exclude the zero frequency the rate of convergence
of the moment matrices is slower and the bias in
ˆ
8
.
c will disappear when
A
/
O
Z
4
c
O
Z
A
n
−1
A
/
O
Z
4
c
O
Z
4
¯
A
n
µ
→0. (50)
Our starting point in the 1(1) case is to provide some limit theory for dis
crete Fourier transforms of 1(1) and detrended 1(1) processes. The following
two lemmas do this and provide a limit theory for periodogram averages of such
processes. The remainder of the limit theory then follows in a fairly straightfor
ward way from these lemmas. To make the derivations simpler we conﬁne our
attention here to the case of a linear trend.
Lemma B: Let ˜ x
i
be an 1(1) process satisfying Assumption 2. Then, the discrete
Fourier transform of ˜ x
i
for \
s
,=0 is given by
u
˜ x
(\
s
) =
1
1−e
i\
s
u
:
(\
s
) −
e
i\
s
1−e
i\
s
 ˜ x
n
− ˜ x
0
]
n
1¡2
. (51)
1084 d. corbae, s. ouliaris, and p. c. b. phillips
Equation (51) shows that the discrete Fourier transforms of an 1(1) process are
not asymptotically independent across fundamental frequencies. Indeed, they are
frequencywise dependent by virtue of the component n
−1¡2
˜ x
n
, which produces a
common leakage into all frequencies \
s
,=0, even in the limit as n →~. As the
next lemma shows, this leakage is strong enough to ensure that smoothed peri
odogram estimates of the spectral density ]
xx
(w) =[1−e
iw
[
−2
]
::
(w) are inconsis
tent at frequencies w ∈
c
.
. Lemma C(f) shows that the leakage is still manifest
when the data are ﬁrst detrended in the time domain. On the other hand, it is
apparent from (41) that we can write (51) as
u
˜ x
(\
s
) =
1
1−e
i\
s
u
:
(\
s
) ÷ui
n
(\
s
) ˜ x
n
− ˜ x
0
], (52)
from which it is clear that frequency domain detrending (i.e., using residuals from
regressions of the frequency domain data on u
i¡n
(\
s
) or u
J
(\
s
)) will remove the
second term of (51) and thereby eliminate the common leakage from the low
frequency.
Lemma C: Let ˜ x
i
be an 1(1) process satisfying Assumption 2 and J
i
=(1, i¡n).
Deﬁne ˜ x
/
.J, i
= ˜ x
/
i
−J
/
i
(D
/
D)
−1
(D
/
¯
A) and let u
x.J
(\) be the discrete Fourier trans
form of ˜ x
.J, i
. Then, as n →~:
(a) n
−2
¸
\
s
∈
.
u
˜ x
(\
s
)u
˜ x
(\
s
)
∗
J
→
1
0
E
x
E
/
x
;
(b) n
−3¡2
¸
\
s
∈
.
u
˜ x
(\
s
)u
J
(\
s
)
∗
J
→
1
0
E
x
u
/
;
(c) n
−2
¸
\
s
∈
.
u
x.J
(\
s
)u
x.J
(\
s
)
∗
J
→
1
0
E
x.u
E
/
x.u
;
(d) n
−1¡2
¸
\
s
∈
c
.
u
˜ x
(\
s
)u
J
(\
s
)
∗
J
→−
1
2=
E
x
(1)
c
.
(e
i\
]
1
(\)
∗
¡(1−e
i\
)) J\;
(e) n
−1
¸
\
s
∈
c
.
u
˜ x
(\
s
)u
˜ x
(\
s
)
∗
J
→
c
.
¸
]
xx
(w)÷
1
2=
(1¡([1−e
iw
[
2
))E
x
(1)E
x
(1)
/
¸
Jw;
(f) n
−1
¸
\
s
∈
c
.
u
x.J
(\
s
)u
x.J
(\
s
)
∗
J
→
c
.
¸
]
xx
(w)÷(2=)
−1
g(w, E
x
)g(w, E
x
)
∗
¸
Jw;
(g) n
−1¡2
¸
\
s
∈
c
.
u
x.J
(\
s
)u
J
(\
s
)
∗
J
→−
1
2=
c
.
g(w, E
x
)]
1
(w)
∗
Jw;
(h)
¸
\
s
∈
c
.
u
J
(\
s
)u
J
(\
s
)
∗
→
1
2=
c
.
]
1
(w)]
1
(w)
∗
Jw;
(i) n
−1
¸
\
s
∈
.
u
J
(\
s
)u
J
(\
s
)
∗
→
1
0
uu
/
;
(j) n
−2
¸
\
s
∈
0
u
˜ x
(\
s
)u
˜ x
(\
s
)
∗
J
→
1
0
E
x
E
/
x
;
(k) n
−2
¸
\
s
∈
0
u
x.J
(\
s
)u
x.J
(\
s
)
∗
J
→
1
0
E
x.u
E
/
x.u
;
(l)
1
n
¸
\
s
∈
0
u
J
(\
s
)u
J
(\
s
)
∗
→
1
0
uu
/
;
(m) n
−3¡2
¸
\
s
∈
0
u
˜ x
(\
s
)u
J
(\
s
)
∗
J
→
1
0
E
x
u
/
;
(n) (n¡n)
¸
\
s
∈
w
u
J
(\
s
)u
J
(\
s
)
∗
→]
1
(w)]
1
(w)
∗
, w ,=0;
(o)
√
nn
−1
¸
\
s
∈
w
u
˜ x
(\
s
)u
J
(\
s
)
∗
J
→(e
iw
¡(1−e
iw
))E(1)]
1
(w)
∗
, w ,=0;
band spectral regression 1085
where ]
xx
(w) =[1−e
iw
[
−2
]
::
(w), ]
1
(w) =(0, e
iw
¡(e
iw
−1))
/
for w ,=0, and
E
x.u
(r) =E
x
(r) −
1
0
E
x
u
/
1
0
uu
/
−1
u(r),
and
g(w, E
x
) =
e
iw
1−e
iw
E
x
(1) ÷
1
0
E
x
u
/
1
0
uu
/
−1
]
1
(w).
Joint convergence applies in (a)–(o).
With this result in hand, the 1(1) band regression model can be analyzed. The
limit theory for time domain detrended band spectral regression is as follows.
Theorem 2: Suppose (:
i
, ˜ x
i
) satisﬁes Assumption 2 and ˜ \
i
is generated by (22).
Then
n(
ˆ
8
.
−8
.
)
J
→
1
0
E
x.u
E
/
x.u
−1
1
0
E
x.u
JE
:
(53)
= MN
0,
1
0
E
x.u
E
/
x.u
−1
2=]
::
(0)
,
and
ˆ
8
.
c
J
→8
.
c −E
−1
£(8
.
c −8
.
), (54)
where
E =
¸
c
.
2=]
xx
(w) ÷g(w, E
x
)g(w, E
x
)
∗
] Jw
¸
,
and
£ =
¸
c
.
g(w, E
x
)]
1
(w)
∗
Jw
1
0
uu
/
−1
1
0
uE
/
x
. (55)
So, when the regressors are 1(1), band spectral regression is inconsistent in
frequency bands that exclude the origin when detrending is performed in the time
domain prior to frequency domain regression. The bias is random and is linear
in the differential, 8
.
c −8
.
, between the coefﬁcients in the frequency bands. In
the case of a linear trend z
i
=i, the limit function ]
1
(w) =e
iw
¡(e
iw
−1) and then
the bias depends on
g(w, E
x
) =
e
iw
1−e
iw
¸
E
x
(1) −
1
0
E
x
r
1
0
r
2
−1
.
1086 d. corbae, s. ouliaris, and p. c. b. phillips
When ]
xx
(w) = (1¡2=)[1−e
iw
[
−2
(i.e., when :
i
is iid(0, 1)), a simple calculation
reveals that the probability limit (54) of
ˆ
8
.
c simpliﬁes to
8
.
c ÷(8
.
−8
.
c )
¸
1
0
E
x
r
1
0
r
2
−1
−E
x
(1)
¸
1
0
r
2
−1
1
0
E
x
r
1÷
¸
E
x
(1) −
1
0
E
x
r
1
0
r
2
−1
¸
2
. (56)
So, in this case, the bias is not dependent on the width of the band 8
.
c .
The following theorem gives the corresponding results for the nonparametric
estimators
ˆ
8
w
and
˜
8
w
.
Theorem 2
/
: Suppose (:
i
, ˜ x
i
) satisﬁes Assumption 2 and ˜ \
i
is generated by (11)
where 8(1) and
˜
8(1) both have valid BN decompositions.
(a) For w =0,
n(
ˆ
8
0
−8
0
)
J
→
¸
1
0
E
x.u
E
/
x.u
¸
−1
¸
1
0
E
x.u
JE
:
−
¸
1
0
E
x.u
JE
/
x
÷ú
x
¸
˜
8(1)
¸
(57)
where 8
0
=8(1) and ú
x
=
¸
~
]=−~
1(:
0
:
/
]
). For w ,=0,
ˆ
8
w
→
J
8
w
÷
¸
2=E
−1
w
]
xx
(w) −1
¸
˜
8(e
−iw
)(e
−iw
−1), (58)
where 8
w
=8(e
−iw
) =b(−w),
E
w
=2=]
xx
(w) ÷g(w, E
x
)g(w, E
x
)
∗
,
and ]
xx
(w) =[1−e
iw
[
−2
]
::
(w).
(b) For w =0,
n(
˜
8
0
−8
0
)
J
→
¸
1
0
E
x.u
E
/
x.u
¸
−1
¸
1
0
E
x.u
JE
:
¸
.
For w ,=0,
√
n(
˜
8
w
−8
w
)→
J
A
c
0,
]
::
(w)
]
xx
(w)
.
Remarks: (i) In Theorem 2
/
(a) it is apparent that the limit distribution of
ˆ
8
0
has a second order bias term involving
¸
1
0
E
x.u
JE
/
x
÷ú
x
¸
˜
8(1),
which is nonzero except when
˜
8(1) =0, a circumstance that arises when the ﬁlter
8(1) is symmetric (see Remark (ii) following Lemma D in the Appendix) as it
is in the case (19). In that case, the limit (57) reduces to the same mixed normal
distribution as (53). Part (a) shows also that the estimator
ˆ
8
w
is inconsistent at
all frequencies w ,=0 except when
˜
8(e
−iw
) =0. The bias in
ˆ
8
0
and inconsistency
in
ˆ
8
w
are due to omitted variable misspeciﬁcation in the regression.
band spectral regression 1087
(ii) Part (b) of Theorem 2
/
shows that the estimator
˜
8
w
based on the aug
mented regression (32) is consistent for both w =0, and w ,=0. The same mixed
normal distribution as (53) applies when w=0. When w,=0 the limit distribution
is the same as in the stationary case given in Theorem 1
/
.
The following results give the limit theory for the frequency domain detrended
estimators in the ﬁxed band and narrow band cases for nonstationary data.
Theorem 3: Under the conditions of Theorem 2
n
ˆ
8
]
.
−8
.
J
→
1
0
E
x.u
E
/
x.u
−1
1
0
E
x.u
JE
:
(59)
and
√
n
ˆ
8
]
.
c
−8
.
c
J
→A
0,
¸
c
.
]
xx
(w) Jw
¸
−1
¸
2=
c
.
]
xx
(w)]
::
(w) Jw
¸
(60)
¸
c
.
]
xx
(w) Jw
¸
−1
.
In ﬁxed band regression models there is usually some advantage to be gained
by averaging over the band and using weighted regression, as shown in Hannan’s
(1963a, b) original work. Efﬁcient regression is based on a weighted band spec
tral regression that uses a preliminary regression to obtain estimates of the equa
tion errors and a corresponding estimate of the error spectrum, say
ˆ
]
::
(w), that
is uniformly consistent so that sup
w
[
ˆ
]
::
(w) −]
::
(w)[ →
µ
0 (e.g., Hannan (1970,
p. 488)). When such weighted regression is performed with frequency domain
detrending, the resulting estimates have optimal properties for both the nonsta
tionary component and the stationary component. For 8
.
, the limit distribution
is the same as (59) and is optimal in the sense of Phillips (1991b). For 8
.
c ,
the limit distribution of the efﬁcient estimate is normal with variance matrix
2=
c
.
]
xx
(w)]
::
(w)
−1
Jw]
−1
and therefore attains the usual efﬁciency bound in
time series regression (e.g., Hannan (1970, eqn. (3.4), p. 427)) adjusted here for
band limited regression (Hannan (1963, eqn. (16))). Details are omitted and are
available in an earlier version of the present paper.
Theorem 3
/
: Under the conditions of Theorem 2
/
, at w =0
n
˜
8
]
0
−8
0
J
→
1
0
E
x.u
E
/
x.u
−1
1
0
E
x.u
JE
:
and for w ,=0
√
n
˜
8
]
w
−8
w
J
→A
c
0, ]
::
(w)]
xx
(w)
−1
. (61)
1088 d. corbae, s. ouliaris, and p. c. b. phillips
So,
ˆ
8
]
.
and
˜
8
]
0
are consistent and have the same mixed normal limit distribution
as that of
ˆ
8
.
in (53). The limit distribution makes asymptotic inference about 8
.
and 8
0
straightforward, using conventional regression Wald tests adjusted in the
usual fashion so that a consistent estimate of the spectrum of :
i
is used (based
on regression residuals) in place of a variance estimate.
The frequency domain detrended estimators
ˆ
8
]
.
c
and
ˆ
8
]
w
are also consistent,
unlike the time detrended estimator
ˆ
8
w
(w ,= 0) in the nonstationary case. The
limit distribution of
ˆ
8
]
.
c
is the same as it is in the case of trend stationary regres
sors (Theorem 1) and has the same
√
n rate. The nonparametric estimator
˜
8
]
w
is
√
n consistent and is asymptotically equivalent to the augmented regression esti
mator
˜
8
w
. As far as the model (11) is concerned, it is therefore apparent from
Theorems 2
/
and 3
/
that if the correct augmented regression model is used in
estimation, it does not matter asymptotically whether detrending is done in the
time domain or the frequency domain.
6. conclusion
It is natural to eliminate deterministic trends in the time domain by simple least
squares regression because the Grenander–Rosenblatt (1957, p. 244) theorem
shows that such regression is asymptotically efﬁcient when the time series are
trend stationary (although this conclusion does not hold when there are stochas
tic as well as deterministic trends—see Phillips and Lee (1996)). In a similar way,
it seems natural to eliminate deterministic trends in band spectral regressions by
detrending in the time domain prior to the use of spectral methods because these
methods were originally intended for application to stationary time series. How
ever, this paper shows that such time domain detrending will lead to biased coef
ﬁcient estimates in models where the coefﬁcients are frequency dependent. In
models that have both deterministic and stochastic trends, time domain detrend
ing can lead to inconsistent estimates of the coefﬁcients at frequency bands away
from the origin. The inconsistency, which arises from omitted variable effects,
can be substantial and has been conﬁrmed in simulations that are not reported
here (Corbae, Ouliaris, and Phillips (1997)).
The bias and inconsistency arise from omitted variable misspeciﬁcation and
are managed by use of an appropriate augmented regression model. The situa
tion is analogous to a structural break model, but here the coefﬁcients change
across frequency rather than over time. An alternate approach that is suitable in
practice is to model the data and run regressions, including detrending regres
sions, in the frequency domain. In effect, discrete Fourier transforms of all the
variables in the model, including the deterministic trends, are taken and band
regression is performed. When nonparametric estimation is being conducted, the
same principle is employed but one uses a shrinking band that is local to a par
ticular frequency. In the nonstationary case, it turns out to be particularly impor
tant to specify the model in terms of levels and differences as in (15) leading to
the ﬁtted regressions (32) and (45). An estimate of the frequency response coef
ﬁcient at a particular frequency is then recovered from a linear combination of
band spectral regression 1089
the coefﬁcients in the regression as in (34) and (46). This approach, which can
be regarded as a frequency domain version of leads and lags dynamic regression,
provides a convenient single equation method of estimating a long run relation
ship in the presence of deterministic trends and short run dynamics.
Dept. of Economics, University of Texas at Austin, Austin, TX 78712, U.S.A.;
corbae@eco.utexas.edu; http://www.eco.utexas.edu/∼corbae
School of Business, National University of Singapore, Singapore 117591; and
Research Dept., IMF, Washington DC, U.S.A.; souliaris@nus.edu.sg
and
Cowles Foundation for Research in Economics, Yale University, Box 208281,
New Haven, CT 065208281, U.S.A.; and University of Auckland and University of
York; peter.phillips@yale.edu; http://Korora.econ.yale.edu
Manuscript received September, 1997; ﬁnal revision received June, 2001.
APPENDIX
Lemma D (Two Sided BN Decompositions): If C(1) =
¸
~
]=−~
c
]
1
]
and
¸
~
]=−~
[][
1
2
c
]
  ~,
then:
(a) C(1) =C(1) ÷
˜
C(1)(1−1), where
¯
C(1) =
¸
~
]=−~
˜ c
]
1
]
, with
˜ c
]
=
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
~
¸
k=]÷1
c
k
, ] ≥0,
−
]
¸
k=−~
c
k
, ]  0;
(b) C(1) =C(e
iu
) ÷
¯
C
u
(e
−iu
1)(e
−iu
1−1), where
¯
C
u
(1) =
¸
~
]=−~
˜ c
u]
1
]
, with
˜ c
u]
=
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
~
¸
k=]÷1
c
k
e
iku
, ] ≥0,
−
]
¸
k=−~
c
k
e
iku
, ]  0.
Proof of Lemma D: The proof is along the same lines as the proof in Phillips and Solo (1992,
Lemma 2.1) of the one sided BN decomposition.
Remarks: (i) The condition
~
¸
]=−~
[][
1
2
c
]
 ~ (62)
is sufﬁcient for (a) and (b) and ensures that
¸
~
]=−~
˜ c
2
]
~ but it is not necessary for the latter. For
instance, the series 8(1) =
¸
~
]=−~
8
]
1
]
, whose coefﬁcients 8
]
are given by (19), fails (62) but still has
a valid BN decomposition with
¸
~
]=−~
˜
8
2
]
 ~. To see this, let c
]
= e
i]x
¡] and then 8
]
is a constant
times the imaginary part of c
]
. Deﬁne S
n
=
¸
n
]=1
e
i]x
=(e
ix
¡(e
ix
−1))(e
inx
−1). Partial summation gives
F
n
=
n
¸
]=1
e
i]x
]
=
1
n
S
n
÷
e
ix
e
ix
−1
n
¸
]=2
e
i(]−1)x
−1
](] −1)
.
1090 d. corbae, s. ouliaris, and p. c. b. phillips
It follows that
F
n
−F
n
=
S
n
n
−
S
n
n
÷
e
ix
e
ix
−1
n
¸
]=n÷1
e
i(]−1)x
−1
](] −1)
,
and letting n →~ we have
˜ c
n
=
~
¸
]=n÷1
e
i]x
]
=−
S
n
n
÷
e
ix
e
ix
−1
~
¸
]=n÷1
e
i(]−1)x
−1
](] −1)
=O
1
n
,
with a similar result for ˜ c
−n
. We deduce that
¸
~
−~
[ ˜ c
n
[
2
 ~. Hence, 8(1) =
¸
~
]=−~
8
]
1
]
with 8
]
given by (19) has a valid BN decomposition with
¸
~
]=−~
˜
8
2
]
~.
(ii) The case of a symmetric ﬁlter with c
]
= c
−]
for ] ,= 0 is an important specialization, which
includes the series with coefﬁcients 8
]
given by (19). In this case, −˜ c
−n
=
¸
~
]=n
c
−]
=
¸
~
]=n
c
]
=
c
n
÷ ˜ c
n
. Then,
¯
C(1) =
¸
~
]=1
( ˜ c
]
÷ ˜ c
−]
) ÷ ˜ c
0
=
¸
~
]=1
( ˜ c
]
− ˜ c
]
−c
]
) ÷
¸
~
]=1
c
]
= 0. If
¯
C(1) has a valid BN
decomposition, we deduce that
C(1) =C(1) ÷
¯
¯
C(1)(1−1)
2
,
¯
¯
C(1) =
~
¸
]=−~
˜
˜ c
]
1
]
, (63)
where
≈
c
]
=
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
~
¸
k=]÷1
˜ c
k
] ≥0,
−
]
¸
k=−~
˜ c
k
]  0.
A sufﬁcient condition for the validity of (63) is
¸
~
]=−~
[][
1
2
˜ c
]
 ~ or
¸
~
]=−~
[][
3
2
c
]
 ~ in terms
of the original coefﬁcients.
Proof of Lemma A: Part (a) is well known (e.g., Gradshteyn and Ryzhik (1965, formula
0.121)). For part (b) we use partial summation to give
ú
n
¸
i=1
i
k
e
i\
s
i
=
n
¸
i=1
úi
k
e
i\
s
i
÷
n
¸
i=1
(i −1)
k
úe
i\
s
i
=
n
¸
i=1
¸
−
k
¸
]=1
k
]
i
k−]
(−1)
]
¸
e
i\
s
i
÷
n
¸
i=1
(i −1)
k
e
i\
s
(i−1)
e
i\
s
−1
,
or
n
k
=
n
¸
i=1
¸
−
k
¸
]=1
k
]
i
k−]
(−1)
]
¸
e
i\
s
i
÷
n
¸
i=1
(i −1)
k
e
i\
s
(i−1)
e
i\
s
−1
.
Using this formula in the identity
n
¸
i=1
i
k
e
i\
s
i
e
i\
s
−1
=n
k
e
i\
s
−1
÷
n
¸
i=1
(i −1)
k
e
i\
s
(i−1)
e
i\
s
−1
,
we get
n
¸
i=1
i
k
e
i\
s
i
e
i\
s
−1
=n
k
e
i\
s
−1
÷n
k
−
n
¸
i=1
¸
−
k
¸
]=1
k
]
i
k−]
(−1)
]
¸
e
i\
s
i
=n
k
e
i\
s
÷
k
¸
]=1
k
]
(−1)
]
n
¸
i=1
i
k−]
e
i\
s
i
,
band spectral regression 1091
giving the recursion
W
k
(\
s
) =n
k
e
i\
s
e
i\
s
−1
÷
1
e
i\
s
−1
k
¸
]=1
k
]
(−1)
]
W
k−]
(\
s
),
which holds for s =1, . . . , n (i.e., e
i\
s
,=1).
The initial condition
W
0
(\
s
) =
n
¸
i=1
e
i\
s
i
=
¸
n, s =0,
0, s ,=0,
follows by elementary calculation. For higher order trends with k =1, 2, 3 we get
W
1
(\
s
) =
¸
¸
¸
¸
¸
¸
n(n÷1)
2
, s =0,
n
e
i\
s
e
i\
s
−1
, s ,=0,
W
2
(\
s
) =
¸
¸
¸
¸
¸
¸
n(n÷1)(2n÷1)
6
, s =0,
n
e
i\
s
e
i\
s
−1
¸
n−
2
e
i\
s
−1
¸
, s ,=0,
and
W
3
(\
s
) =
¸
¸
¸
¸
¸
¸
¸
¸
n(n÷1)
2
2
, s =0,
n
e
i\
s
e
i\
s
−1
¸
n
2
−3(n−1)
1
e
i\
s
−1
÷6
1
e
i\
s
−1
2
¸
, s ,=0,
which lead to the expressions for W
k
(\
s
), k =0, 1, 2, that are given in Section 2 following Lemma A.
Proof of Theorem 1: This follows standard lines and is omitted (see Corbae, Ouliaris, and
Phillips (1997)).
Proof of Theorem 1
/
: First, observe that
ˆ
8
w
=
¸
n
−1
¯
A
/
O
Z
4
w
O
Z
¯
A
¸
−1
¸
n
−1
¯
A
/
O
Z
4
w
O
Z
˜ \
¸
. (64)
Next, from (8) and (9) we have
u
˜ \
(\
s
) =b(\
s
)
/
u
˜ x
(\
s
) ÷u
:
(\
s
) ÷
1
√
n
¸
e
−i\
s
˜ a
\
s
0
− ˜ a
\
s
n
¸
(65)
=b(w)
/
u
˜ x
(\
s
) ÷u
:
(\
s
) ÷O
µ
1
√
n
(66)
in the stationary case. We now ﬁnd that
n
−1
¯
A
/
O
Z
4
w
O
Z
¯
A =n
−1
¸
\
s
∈
w
u
˜ x
(\
s
)u
˜ x
(\
s
)
∗
÷t
µ
(1) →
µ
2=]
xx
(w), (67)
and using (66) we have
n
−1
¯
A
/
O
Z
4
w
O
Z
˜ \ =n
−1
¯
A
/
4
w
˜ \ ÷t
µ
(1) (68)
=n
−1
¸
\
s
∈
w
u
˜ x
(\
s
)u
˜ x
(\
s
)
∗
b(−w) ÷n
−1
¸
\
s
∈
w
u
˜ x
(\
s
)u
:
(\
s
)
∗
÷t
µ
(1).
1092 d. corbae, s. ouliaris, and p. c. b. phillips
Then, from (64), (67), and (68) we have the expression
√
n
ˆ
8
w
−8
w
=
¸
n
−1
¸
\
s
∈
w
u
˜ x
(\
s
)u
˜ x
(\
s
)
∗
¸
−1
¸
n
−1¡2
¸
\
s
∈
w
u
˜ x
(\
s
)u
:
(\
s
)
∗
¸
÷t
µ
(1). (69)
The family {u
:
(\
s
) : \
s
∈
w
} are known to satisfy a central limit theorem (with limit A
c
(0, 2=]
::
(w)))
for dft’s of stationary processes and are independently distributed
2
as n →~. It follows from this
result and (67) that
n
−1¡2
¸
\
s
∈
w
u
˜ x
(\
s
)u
:
(\
s
)
∗
→
J
A
c
(0, (2=)
2
]
::
(w)]
xx
(w)), (70)
which leads to the stated limit theorem for
ˆ
8
w
. A similar argument gives the result for
ˆ
8
]
w
.
Proof of Lemma B: Take dft’s of the equation ú˜ x
i
=¡
i
, giving
u
˜ x
(\
S
) =n
−1¡2
n
¸
i=1
˜ x
i−1
e
i\
s
i
÷u
:
(\
s
) =e
i\
s
¸
u
˜ x
(\
s
) −n
−1¡2
e
i\
s
n
˜ x
n
− ˜ x
0
¸
÷u
:
(\
s
).
Then, transposing and solving yields the stated formula
u
˜ x
(\
s
) =
1
1−e
i\
s
u
:
(\
s
) −
e
i\
s
1−e
i\
s
 ˜ x
n
− ˜ x
0
]
n
1¡2
. (71)
Proof of Lemma C: Part (a): This result may be proved as in Phillips (1991a). A new and
substantially simpler proof uses part (e) and is as follows.
n
−2
¸
\
s
∈
.
u
˜ x
(\
s
)u
˜ x
(\
s
)
∗
= n
−2
¸
\
s
∈−=, =]
u
˜ x
(\
s
)u
˜ x
(\
s
)
∗
−n
−2
¸
\
s
∈
c
.
u
˜ x
(\
s
)u
˜ x
(\
s
)
∗
= n
−2
n
¸
i=1
˜ x
i
˜ x
/
i
÷O
µ
(n
−1
)
J
→
1
0
E
x
E
/
x
,
where the error magnitude in the second line follows directly from part (e) below.
Part (b): First note from (41) that u
J
(0)
/
=
√
n, (n÷1)¡2n
1¡2
, so that n
−1¡2
u
J
(0)
/
→
1,
1
2
=
]
/
0
=
1
0
u
/
, and
u
J
(\
s
)
/
=
0,
1
n
1¡2
e
i\
s
e
i\
s
−1
=
1
√
n
]
1
(\
s
)
/
, s ,=0.
It follows that as n →~
u
J
(\
s
)
/
=
0,
√
n
2=si
1÷t(1)]
, (72)
2
Hannan (1973, Theorem 3) showed that a ﬁnite collection of u
:
(\
s
) satisfy a central limit theorem
and are independent. Here the collection {\
s
∈
w
} has n members and is asymptotically inﬁnite.
Phillips (2000, Theorem 3.2) showed that an asymptotically inﬁnite collection of u
:
(\
s
) satisfy a
central limit theorem and are asymptotically independent for frequencies \
s
in the neighborhood of
the origin provided the number of frequencies n= t(n
−
1
2
÷
1
µ
) where 1([:
i
[
µ
)  ~ for some µ > 2,
i.e., provided the number of frequencies does not go to inﬁnity too fast. This result can be extended
to frequency bands away from the origin, although a proof was not given in that paper.
band spectral regression 1093
a formula that holds for both s ﬁxed and for s → ~ with (s¡n) → 0. On the other hand, when
\
s
→\ ,=0 as n →~ we have
u
J
(\
s
)
/
=
1
√
n
0,
e
i\
e
i\
−1
÷t
1
√
n
=
1
√
n
]
1
(\)
/
÷t
1
√
n
. (73)
Write the summation over
.
as follows:
¸
\
s
∈
.
=
¸
\
s
∈−=, =]
−
¸
\
s
∈
.
c
. Then
n
−3¡2
¸
\
s
∈
.
u
˜ x
(\
s
)u
J
(\
s
)
∗
=n
−3¡2
¸
\
s
∈−=, =]
u
˜ x
(\
s
)u
J
(\
s
)
∗
−n
−3¡2
¸
\
s
∈
.
c
u
˜ x
(\
s
)u
J
(\
s
)
∗
=n
−3¡2
¸
\
s
∈−=, =]
u
˜ x
(\
s
)u
J
(\
s
)
∗
−n
−2
¸
\
s
∈
.
c
u
˜ x
(\
s
)]
1
(\
s
)
∗
.
First,
n
−2
¸
\
s
∈
.
c
u
˜ x
(\
s
)]
1
(\
s
)
∗
=n
−2
¸
\
s
∈
.
c
¸
1
1−e
i\
s
u
:
(\
s
) −
e
i\
s
1−e
i\
s
 ˜ x
n
− ˜ x
0
]
n
1¡2
¸
]
1
(\
s
)
∗
=O
µ
(n
−1
).
Second,
n
−3¡2
¸
\
s
∈−=, =]
u
˜ x
(\
s
)u
J
(\
s
)
∗
= n
−2
n
¸
i=1
˜ x
i
¸
\
s
∈−=, =]
e
i\
s
i
u
J
(\
s
)
∗
= n
−3¡2
n
¸
i=1
˜ x
i
J
/
i
=n
−1
n
¸
i=1
˜ x
i
√
n
J
/
i
J
→
1
0
E
x
(r)u(r)
/
Jr.
It follows that
n
−3¡2
¸
\
s
∈
.
u
˜ x
(\
s
)u
J
(\
s
)
∗
J
→
1
0
E
x
(r)u(r)
/
Jr.
Part (c): First, observe that
n
−1¡2
˜ x
/
.J, nr]
J
→E
x
(r)
/
−u(r)
/
1
0
uu
/
−1
1
0
uE
/
x
=E
x.u
(r)
/
.
Then
1
n
2
¸
\
s
∈
.
u
x.J
(\
s
)u
x.J
(\
s
)
∗
J
→
1
0
E
x.u
E
/
x.u
,
as in part (a).
Part (d): From Lemma B we have
u
˜ x
(\
s
) =
1
1−e
i\
s
u
:
(\
s
) −
e
i\
s
1−e
i\
s
 ˜ x
n
− ˜ x
0
]
n
1¡2
.
As in the proof of part (b), we have n
−1¡2
˜ x
n
J
→E
x
(1), and, if \
s
→\ ,=0 as n →~,
u
˜ x
(\
s
)
J
→
1
1−e
i\
A
c
(0, 2=]
::
(\)) −
e
i\
1−e
i\
A(0, 2=]
::
(0))
= A
c
0, 2=
]
::
(\) ÷]
::
(0)
[1−e
i\
[
2
.
1094 d. corbae, s. ouliaris, and p. c. b. phillips
It follows that
n
−1¡2
¸
\
s
∈
c
.
u
˜ x
(\
s
)u
J
(\
s
)
∗
=n
−1
¸
\
s
∈
c
.
u
˜ x
(\
s
)]
1
(\
s
)
∗
=n
−1
¸
\
s
∈
c
.
1
1−e
i\
s
u
¡
(\
s
)]
1
(\
s
)
∗
−n
−1
 ˜ x
n
− ˜ x
0
]
n
1¡2
¸
\
s
∈
c
.
e
i\
s
1−e
i\
s
]
1
(\
s
)
∗
=term I ÷term II, say.
Since the u
¡
(\
s
) are asymptotically independent A
c
(0, 2=]
::
(\
s
)), term I
µ
→ 0. For term II, since
˜ x
0
=O
µ
(1), we have
n
−1
 ˜ x
n
− ˜ x
0
]
n
1¡2
¸
\
s
∈
c
.
e
i\
s
1−e
i\
s
]
1
(\
s
)
∗
∼
1
2=
˜ x
n
n
1¡2
2=
n
¸
\
s
∈
c
.
e
i\
s
1−e
i\
s
]
1
(\
s
)
∗
J
→
1
2=
E
x
(1)
c
.
e
i\
]
1
(\)
∗
1−e
i\
J\
so that
term II
J
→−
1
2=
E
x
(1)
c
.
e
i\
]
1
(\)
∗
1−e
i\
J\
, (74)
giving part (d).
Part (e): From Lemma B we get
n
−1
¸
\
s
∈
c
.
u
˜ x
(\
s
)u
˜ x
(\
s
)
∗
J
∼n
−1
¸
\
s
∈
c
.
1
[1−e
i\
s
[
2
¸
u
:
(\
s
)u
:
(\
s
)
∗
÷
˜ x
n
n
1¡2
˜ x
/
n
n
1¡2
¸
=
1
2J
¸
J
a
≤[][≤J
1
n
¸
\
s
∈
]
1
[1−e
i\
s
[
2
¸
u
:
(\
s
)u
:
(\
s
)
∗
÷
˜ x
n
n
1¡2
˜ x
/
n
n
1¡2
¸
÷t
µ
(1)
=
1
2J
¸
J
a
≤[][≤J
1
[1−e
iw
]
[
2
2=
ˆ
]
::
(w
]
) ÷n
−1
¸
\
s
∈
c
.
1
[1−e
i\
s
[
2
˜ x
n
n
1¡2
˜ x
/
n
n
1¡2
÷t
µ
(1),
J
→
c
.
]
::
(w)
[1−e
iw
[
2
Jw÷
1
2=
c
.
1
[1−e
iw
[
2
JwE
x
(1)E
x
(1)
/
,
as required. In the penultimate line above, 2=
ˆ
]
::
(w
]
) = (1¡n)
¸
\
s
∈
]
u
¡
(\
s
)u
¡
(\
s
)
∗
and
ˆ
]
::
(w) −
]
::
(w) →
µ
0.
Part (f): Note that
u
x.J
(\
s
)
∗
= u
˜ x
(\
s
)
∗
−u
J
(\
s
)
∗
(n
−1
D
/
D)
−1
(n
−1
D
/
¯
A) (75)
=
1
1−e
−i\
s
u
:
(\
s
)
∗
−
e
−i\
s
1−e
−i\
s
 ˜ x
n
− ˜ x
0
]
/
n
1¡2
−n
1¡2
u
J
(\
s
)
∗
(n
−1
D
/
D)
−1
(n
−3¡2
D
/
¯
A)
J
→
1
1−e
−i\
A
c
(0, 2=]
¡¡
(\)) −
e
−i\
1−e
−i\
E
x
(1)
/
−]
1
(\)
∗
1
0
uu
/
−1
1
0
uE
/
x
=
1
1−e
−i\
A
c
(0, 2=]
¡¡
(\)) −g(\, E
x
)
∗
, say
when \
s
→\ ,=0 as n →~. Here
g(\, E
x
) =
e
i\
1−e
i\
E
x
(1) ÷
1
0
E
x
u
/
1
0
uu
/
−1
]
1
(\),
band spectral regression 1095
and the complex normal variate A
c
(0, 2=]
¡¡
(\)) in (75) is independent of the Brownian motion E
x
.
If w
]
==]¡J is the midpoint of the band
]
and w
]
→w as n →~, then
2=
ˆ
]
xx.J
(w
]
) =
1
n
¸
\
s
∈
]
u
x.J
(\
s
)u
x.J
(\
s
)
∗
J
→
2=]
¡¡
(w)
[1−e
iw
[
2
÷g(w, E
x
)g(w, E
x
)
∗
=2=]
xx
(w) ÷g(w, E
x
)g(w, E
x
)
∗
.
We deduce that
n
−1
¸
\
s
∈
c
.
u
x.J
(\
s
)u
x.J
(\
s
)
∗
=
1
2J
¸
J
a
≤[][≤J
1
n
¸
\
s
∈
]
u
x.J
(\
s
)u
x.J
(\
s
)
∗
÷t
µ
(1)
=
1
2J
¸
J
a
≤[][≤J
2=
ˆ
]
xx.J
(w
]
) ÷t
µ
(1)
J
→
c
.
¸
]
xx
(w) ÷(2=)
−1
g(w, E
x
)g(w, E
x
)
∗
¸
Jw,
which gives part (f).
Part (g): Observe that
n
−1¡2
¸
\
s
∈
c
.
u
x.J
(\
s
)u
J
(\
s
)
∗
= n
−1
¸
\
s
∈
c
.
u
x.J
(\
s
)]
1
(\
s
)
∗
÷t
µ
(1)
J
∼ −n
−1
¸
\
s
∈
c
.
g(\
s
, E
x
)]
1
(\
s
)
∗
J
∼ −
1
2=
2=
n
¸
\
s
∈
c
.
g(\
s
, E
x
)]
1
(\
s
)
∗
J
→−
1
2=
c
.
g(w, E
x
)]
1
(w)
∗
Jw,
as stated.
Part (h): From (73) we have
¸
\
s
∈
c
.
u
J
(\
s
)u
J
(\
s
)
∗
∼
1
n
¸
\
s
∈
c
.
]
1
(\
s
)]
1
(\
s
)
∗
→
1
2=
c
.
]
1
(w)]
1
(w)
∗
Jw,
as required.
Part (i): Using part (h), we get
n
−1
¸
\
s
∈
.
u
J
(\
s
)u
J
(\
s
)
∗
= n
−1
¸
\
s
∈−=, =]
u
J
(\
s
)u
J
(\
s
)
∗
÷O
µ
n
−1
= n
−1
n
¸
i=1
J
i
¸
\
s
∈−=, =]
e
i\
s
i
u
J
(\
s
)
∗
÷O
µ
n
−1
= n
−1
n
¸
i=1
J
i
J
/
i
÷O
µ
n
−1
→
1
0
u(r)u(r)
/
Jr,
giving the stated result.
1096 d. corbae, s. ouliaris, and p. c. b. phillips
Part (j): Using the representation of u
˜ x
(\
s
) in Lemma B, the fact that [s[ > J →~ when \
s
∈
−=, =] −
0
, and proceeding as in part (a), we ﬁnd that
n
−2
¸
\
s
∈
0
u
˜ x
(\
s
)u
˜ x
(\
s
)
∗
=n
−2
¸
¸
\
s
∈−=, =]
−
¸
\
s
∈−=, =]−
0
u
˜ x
(\
s
)u
˜ x
(\
s
)
∗
÷t
µ
(1)
=n
−2
¸
\
s
∈−=, =]
u
˜ x
(\
s
)u
˜ x
(\
s
)
∗
÷O
µ
¸
n≥[s[≥J
1
s
2
¸
u
:
(\
s
)u
:
(\
s
)
∗
÷
˜ x
n
n
1¡2
˜ x
/
n
n
1¡2
¸
=
¸
n
−2
n
¸
i=1
˜ x
i
˜ x
/
i
÷t
µ
(1)
J
→
1
0
E
x
E
/
x
.
Part (k): In the same way as parts (j) and (c) we ﬁnd that
n
−2
¸
\
s
∈
0
u
x.J
(\
s
)u
x.J
(\
s
)
∗
=n
−2
n
¸
i=1
˜ x
.J, i
˜
x
/
.J, i
÷t
µ
(1)
J
→
1
0
E
x.u
E
/
x.u
.
Part (l): From part (b) we have n
−1¡2
u
J
(0)
/
→]
/
0
=
1,
1
2
=
1
0
u
/
and, for \
s
,=0 with (s¡n) →0,
we have u
J
(\
s
)
/
=
0, (
√
n¡2=si)1÷t(1)]
from (72). Thus, for w =0, we ﬁnd that
1
n
¸
\
s
∈
0
u
J
(\
s
)u
J
(\
s
)
∗
=
1
n
u
J
(0)u
J
(0)
/
÷
1
n
¸
\
s
∈
0
−{0}
u
J
(\
s
)u
J
(\
s
)
∗
=
1
n
u
J
(0)u
J
(0)
/
÷
¸
¸
0 0
0 2
1
2=
2
¸
n
s=1
1
s
2
→]
0
]
/
0
÷
¸
0 0
0
1
12
¸
=
1
0
uu
/
.
Part (m): As in part (b) we ﬁnd that
n
−3¡2
¸
\
s
∈
0
u
˜ x
(\
s
)u
J
(\
s
)
∗
= n
−3¡2
¸
\
s
∈−=, =]
u
˜ x
(\
s
)u
J
(\
s
)
∗
−n
−3¡2
¸
\
s
,∈
0
u
˜ x
(\
s
)u
J
(\
s
)
∗
= n
−3¡2
n
¸
i=1
˜ x
i
J
/
i
−t
µ
(1)
J
→
1
0
E
x
(r)u(r)
/
Jr,
since
n
−3¡2
¸
\
s
¢
0
u
˜ x
(\
s
)u
J
(\
s
)
∗
=O
µ
1
nn
¸
\
s
¢
0
u
˜ x
(\
s
)
=O
µ
1
nn
¸
\
s
¢
0
¸
u
:
(\
s
) −e
i\
s
˜ x
n
−˜ x
0
√
n
1−e
i\
s
=t
µ
(1).
Part (n): For w ,=0, using (73) we obtain
n
n
¸
\
s
∈
w
u
J
(\
s
)u
J
(\
s
)
∗
∼
1
n
¸
\
s
∈
w
]
1
(\
s
)]
1
(\
s
)
∗
→]
1
(w)]
1
(w)
∗
.
band spectral regression 1097
Part (o): For w ,=0, using (73) we obtain
√
n
n
¸
\
s
∈
w
u
˜ x
(\
s
)u
J
(\
s
)
∗
=
1
n
¸
\
s
∈
w
u
˜ x
(\
s
)]
1
(\
s
)
∗
÷t(1)]
=
1
n
¸
\
s
∈
w
¸
u
:
(\
s
) −e
i\
s
˜ x
n
−˜ x
0
√
n
1−e
i\
s
]
1
(\
s
)
∗
÷t
µ
(1)
J
→
e
iw
1−e
iw
E(1)]
1
(w)
∗
.
Finally, joint weak convergence in (a)–(o) applies because the component elements jointly converge
and one may apply the continuous mapping theorem in a routine fashion. In particular, Assumptions 1
and 2 ensure that
n
−1¡2
nr]
¸
i=1
:
/
i
, u
:
(\)
/
J
→(E(r)
/
, £
/
\
), with £
\
=A
c
(0, 2=]
::
(\)),
and where £
\
is independent of E(r) for all \ ,=0. The required quantities in (a)–(k) are functionals
of these elements and smoothed periodogram ordinates (like (1¡n)
¸
\
s
∈
]
u
:
(\
s
)u
:
(\
s
)
∗
for w
]
→w
as in the case of part (f)) that converge in probability to constants.
Proof of Theorem 2: Note that
ˆ
8
.
−8
.
=−{
¯
A
/
O
Z
4O
Z
¯
A}
−1
{
¯
A
/
O
Z
4O
Z
4
c
¯
A(8
.
−8
.
c ) −:]}.
We consider the limiting behavior of n
−2
¯
A
/
O
Z
4O
Z
¯
A and n
−1
¯
A
/
O
Z
4O
Z
4
c
¯
A. Deﬁne ˜ x
/
.J,i
= ˜ x
/
i
−
J
/
i
(D
/
D)
−1
(D
/
¯
A). Since ˜ x
i
is an 1(1) process and satisﬁes an invariance principle when standardized
by n
−1¡2
, we have
n
−1¡2
˜ x
/
.J, nr]
J
→E
x
(r)
/
−u(r)
/
1
0
uu
/
−1
1
0
uE
/
x
=E
x.u
(r)
/
, say. (76)
Write the discrete Fourier transform of ˜ x
/
.J, i
as u
x.J
(\
s
) and then from Lemma C(c) we have
n
−2
¯
A
/
O
Z
4O
Z
¯
A =
1
n
2
¸
\
s
∈
.
u
x.J
(\
s
)u
x.J
(\
s
)
∗
J
→
1
0
E
x.u
E
/
x.u
. (77)
Since
1
0
E
x.u
E
/
x.u
> 0 (see Phillips and Hansen (1990)), n
−2
¯
A
/
O
Z
4O
Z
¯
A has a positive deﬁnite limit
as n →~.
Next, decompose n
−1
¯
A
/
O
Z
4O
Z
4
c
¯
A as follows:
¯
A
/
O
Z
4O
Z
4
c
¯
A
n
=−
¯
A
/
O
Z
W
∗
.WF
Z
W
∗
.
c
W
¯
A
n
−
¯
A
/
O
D
W
∗
.WF
D
W
∗
.
c
W
¯
A
n
(78)
=
¯
A
/
F
D
W
∗
.WF
D
W
∗
.
c
W
¯
A
n
−
¯
A
/
W
∗
.WF
D
W
∗
.
c
W
¯
A
n
=term .−term E.
Take each of these terms in turn. Factor term . as follows and consider each factor separately. Write
¯
A
/
F
D
W
∗
.WF
D
W
∗
.
c
W
¯
A
n
=
¯
A
/
F
D
W
∗
.WD
n
3¡2
D
/
D
n
−1
D
/
W
∗
.
c
W
¯
A
n
1¡2
. (79)
1098 d. corbae, s. ouliaris, and p. c. b. phillips
The ﬁrst factor is
n
−1¡2
¯
A
/
F
D
W
∗
.WD
n
=
n
−1¡2
¯
A
/
D
n
D
/
D
n
−1
D
/
W
∗
.WD
n
(80)
J
→
1
0
E
x
u
/
1
0
uu
/
−1
1
0
uu
/
=
1
0
E
x
u
/
,
in view of (76) and Lemma C(i). The third factor is the conjugate transpose of
n
−1¡2
¯
A
/
W
∗
.
c
WD = n
−1¡2
¸
\
s
∈
c
.
u
˜ x
(\
s
)u
J
(\
s
)
∗
(81)
J
→−
1
2=
E
x
(1)
c
.
e
i\
]
1
(\)
∗
1−e
i\
J\
,
from Lemma C(d). The limit of term . now follows by combining (81) and (80) and using joint weak
convergence:
¯
A
/
F
D
W
∗
.WF
D
W
∗
.
c
W
¯
A
n
=
n
−1¡2
¯
A
/
F
D
W
∗
.WD
n
D
/
D
n
−1
(D
/
W
∗
.
c
W(n
−1¡2
¯
A)) (82)
J
→
¸
1
0
E
x
u
/
¸
1
0
uu
/
−1
¸
−
1
2=
c
.
e
−i\
]
1
(\)
1−e
−i\
J\
E
x
(1)
/
¸
.
Next consider term E of (78). Using Lemma C, we obtain
¯
A
/
W
∗
.WF
D
W
∗
.
c
W
¯
A
n
(83)
=
1
n
n
−1¡2
¸
\
s
∈
.
u
˜ x
(\
s
)u
J
(\
s
)
∗
D
/
D
n
−1
n
−1¡2
¸
\
s
∈
c
.
u
J
(\
s
)u
˜ x
(\
s
)
∗
J
→
1
0
E
x
u
/
1
0
uu
/
−1
−
1
2=
c
.
e
−i\
]
1
(\)
1−e
−i\
J\E
x
(1)
/
.
Combining (82) and (83) in (78) we ﬁnd
3
n
−1
¯
A
/
O
Z
4O
Z
4
c
¯
A =t
µ
(1). (84)
As for the limit distribution of
ˆ
8
.
, we have
n(
ˆ
8
.
−8
.
) =
¸
¯
A
/
O
Z
4O
Z
¯
A
n
2
¸
−1
¸
¯
A
/
O
Z
4O
Z
:
n
¸
(85)
−
¸
¯
A
/
O
Z
4O
Z
¯
A
n
2
¸
−1
¸
¯
A
/
O
Z
4O
Z
4
c
¯
A
n
(8
.
−8
.
c )
¸
.
From Lemma C(c) and (77) above, we have
n
−2
¯
A
/
O
Z
4O
Z
¯
A
J
→
1
0
E
x.u
E
/
x.u
=C, (86)
3
By changing the probability space (on which the random sequence ˜ x
i
is deﬁned), we can ensure
that both terms tend almost surely to the same random variable (see, for example, Theorem 4,
page 47 of Shorack and Wellner (1986)). Then, in the original space the difference of the terms tends
in distribution to zero giving the stated result.
band spectral regression 1099
and from (84)
¯
A
/
O
Z
4O
Z
¯
A
n
2
−1
¯
A
/
O
Z
4O
Z
4
c
¯
A
n
µ
→0. (87)
Next,
¯
A
/
O
Z
4O
Z
:
n
=n
−1
¸
\
s
∈
.
u
x.J
(\
s
)u
:.J
(\
s
)
∗
=
1
n
¸
\
s
∈
.
u
x.J
(\
s
)u
:
(\
s
)
∗
(88)
−
1
n
3
2
¸
\
s
∈
.
u
x.J
(\
s
)u
J
(\
s
)
∗
D
/
D
n
−1
D
/
:
√
n
.
Then, using (76) and proceeding as in the proof of Lemma C(a) above, we ﬁnd
n
−1
¸
\
s
∈
.
u
x.J
(\
s
)u
:
(\
s
)
∗
=n
−1
¸
\
s
∈−=, =]
u
x.J
(\
s
)u
:
(\
s
)
∗
−n
−1
¸
\
s
∈
.
c
u
x.J
(\
s
)u
:
(\
s
)
∗
(89)
=n
−1
n
¸
i=1
˜ x
.J, i
:
i
−t
µ
(1)
J
→
1
0
E
x.u
JE
:
.
Further, from Lemma C(b), (i) and Assumption 2, we obtain
1
n
3
2
¸
\
s
∈
.
u
x.J
(\
s
)u
J
(\
s
)
∗
D
/
D
n
−1
D
/
:
√
n
J
→
1
0
E
x.u
u
/
1
0
uu
/
1
0
uJE
:
=0, (90)
since
1
0
E
x.u
u
/
=0. Combining (88), (89), and (90), we deduce that
¯
A
/
O
Z
4O
Z
:
n
J
→
1
0
E
x.u
JE
:
. (91)
The limit distribution (91) is a mixture normal distribution with mixing matrix variate
1
0
E
x.u
E
/
x.u
.
It now follows from (86) and (91) that
¸
¯
A
/
O
Z
4O
Z
¯
A
n
2
¸
−1
¸
¯
A
/
O
Z
4O
Z
:
n
¸
J
→
1
0
E
x.u
E
/
x.u
−1
1
0
E
x.u
JE
:
(92)
≡ MN
0, 2=]
::
(0)
1
0
E
x.u
E
/
x.u
−1
.
Using (85), (87), and (92), we deduce that
n(
ˆ
8
.
−8
.
)
J
→
¸
1
0
E
x.u
E
/
x.u
¸
−1
¸
1
0
E
x.u
JE
:
¸
,
which gives the stated result.
1100 d. corbae, s. ouliaris, and p. c. b. phillips
For the limit of
ˆ
8
.
c , we need to examine the asymptotic behavior of the bias term in (29), which
depends on the matrix quotient (n
−1
¯
A
/
O
Z
4
c
O
Z
¯
A)
−1
(n
−1
¯
A
/
O
Z
4
c
O
Z
4
¯
A). Take each of these fac
tors in turn. First,
¯
A
/
O
Z
4
c
O
Z
¯
A
n
=
¯
A
/
O
D
4
c
O
D
¯
A
n
=n
−1
¸
\
s
∈
c
.
u
x.J
(\
s
)u
x.J
(\
s
)
∗
where u
x.J
(\
s
)
∗
=u
˜ x
(\
s
)
∗
−u
J
(\
s
)
∗
(n
−1
D
/
D)
−1
(n
−1
D
/
¯
A). From Lemma C(f) we have
n
−1
¸
\
s
∈
c
.
u
x.J
(\
s
)u
x.J
(\
s
)
∗
J
→
c
.
¸
]
xx
(w) ÷(2=)
−1
g(w, E
x
)g(w, E
x
)
∗
¸
Jw, (93)
which is a positive deﬁnite limit. Next,
n
−1
¯
A
/
O
Z
4
c
O
Z
4
¯
A
=−n
−1
¯
A
/
O
Z
4
c
F
Z
4
¯
A =−n
−1
¯
A
/
O
D
4
c
F
D
4
¯
A
=−
n
−1¡2
¸
\
s
∈
c
.
u
x.J
(\
s
)u
J
(\
s
)
∗
(n
−1
D
/
D)
−1
n
−3¡2
¸
\
s
∈
.
u
J
(\
s
)u
˜ x
(\
s
)
∗
.
From Lemma C(g) and (b), we obtain
n
−1
¯
A
/
O
Z
4
c
O
Z
4
¯
A
J
→−
−
1
2=
c
.
g(w, E
x
)]
1
(w)
∗
Jw
1
0
uu
/
−1
1
0
uE
/
x
. (94)
It follows from (93) and (94) and joint weak convergence that the asymptotic bias term for
ˆ
8
.
c
involves
(n
−1
¯
A
/
O
Z
4
c
O
Z
¯
A)
−1
(n
−1
¯
A
/
O
Z
4
c
O
Z
4
¯
A)
J
→
¸
c
.
¸
2=]
xx
(w) ÷g(w, E
x
)g(w, E
x
)
∗
¸
Jw
¸
−1
¸
c
.
g(w, E
x
)]
1
(w)
∗
Jw
1
0
uu
/
−1
1
0
uE
/
x
¸
,
establishing the stated result.
Proof of Theorem 2
/
: Part (a): From (64) and (11) we have
ˆ
8
w
=8(1) ÷
¸
¯
A
/
O
Z
4
w
O
Z
¯
A
¸
−1
¸
¯
A
/
O
Z
4
w
O
Z
˜ :
¸
. (95)
Since ˜ :
i
is a strictly stationary and ergodic sequence with mean zero and satisﬁes a central limit
theorem, (n
−1
D
/
D)
−1
(n
−1
D
/
˜ :) =O
µ
(n
−1¡2
), and so
˜ :
/
i
−z
/
i
(n
−1
Z
/
Z)
−1
(n
−1
Z
/
˜ :) = ˜ :
/
i
−z
/
i
o
−1
n
(n
−1
D
/
D)
−1
(n
−1
D
/
˜ :)
µ
→ ˜ :
/
i
. (96)
Then, since
˜
8(1) has a valid BN decomposition,
n
−1
¯
A
/
O
Z
4
w
O
Z
˜ : =n
−1
¸
\
s
∈
w
u
˜ x.J
(\
s
)u
˜ :
(\
s
)
∗
÷t
µ
(1) (97)
=n
−1
¸
\
s
∈
w
u
˜ x.J
(\
s
)u
:
(\
s
)
∗
−n
−1
¸
\
s
∈
w
u
˜ x.J
(\
s
)u
:
(\
s
)
∗
˜
8(e
−i\
s
) ÷t
µ
(1).
band spectral regression 1101
When w =0, we ﬁnd as in (89) and Theorem 3.1 of Phillips (1991a) that
n
−1
¸
\
s
∈
0
u
˜ x.J
(\
s
)u
:
(\
s
)
∗
J
→
1
0
E
x.u
JE
:
, (98)
and
n
−1
¸
\
s
∈
0
u
˜ x.J
(\
s
)u
:
(\
s
)
∗
J
→
1
0
E
x.u
JE
/
x
÷ú
x
, (99)
where ú
x
=
¸
~
]=−~
1(:
0
:
/
]
). We deduce that
n
−1
¯
A
/
O
Z
4
w
O
Z
˜ :
J
→
1
0
E
x.u
JE
:
−
¸
1
0
E
x.u
JE
/
x
÷ú
x
¸
˜
8(1). (100)
Next, as in Lemma C(c) we ﬁnd that
n
−2
¯
A
/
O
Z
4
w
O
Z
¯
A
J
→
1
0
E
x.u
E
/
x.u
. (101)
Combining (100) and (101), we obtain
n(
ˆ
8
0
−8(1))
J
→
¸
1
0
E
x.u
E
/
x.u
¸
−1
¸
1
0
E
x.u
JE
:
−
¸
1
0
E
x.u
JE
/
x
÷ú
x
¸
˜
8(1)
¸
.
Now consider the case where w ,=0. First we have
n
−1
¯
A
/
O
Z
4
w
O
Z
¯
A =n
−1
¸
\
s
∈
w
u
x.J
(\
s
)u
x.J
(\
s
)
∗
,
and in a similar fashion to (93) we ﬁnd that
n
−1
¸
\
s
∈
w
u
x.J
(\
s
)u
x.J
(\
s
)
∗
J
→2=]
xx
(w) ÷g(w, E
x
)g(w, E
x
)
∗
=E
w
, say. (102)
Next, from (15) we have
u
˜ \
(\
s
) =8(1)
/
u
˜ x
(\
s
) −
˜
8(e
i\
s
)
/
u
:
(\
s
) ÷u
:
(\
s
) ÷O
µ
1
√
n
. (103)
Now
u
\.J
(\
s
)
∗
=u
˜ \
(\
s
)
∗
−u
J
(\
s
)
∗
(n
−1
D
/
D)
−1
(n
−1
D
/
˜ \), (104)
and, in view of the cointegrating relation (11), we have
n
−
1
2
˜ \
nr]
→
J
8(1)
/
E
x
(r) =E
\
(r), (105)
and it follows as in part (f) of Lemma C that
u
\.J
(\
s
)
∗
∼u
˜ \
(\
s
)
∗
−]
1
(\
s
)
∗
1
0
uu
/
−1
1
0
uE
/
x
8(1) ÷t
µ
(1). (106)
However, from the proof of Lemma C(f) we also have
u
x.J
(\
s
)
∗
=u
x
(\
s
)
∗
−]
1
(\
s
)
∗
1
0
uu
/
−1
1
0
uE
/
x
÷t
µ
(1). (107)
1102 d. corbae, s. ouliaris, and p. c. b. phillips
We deduce from (103), (106), and (107) that
u
\.J
(\
s
)
∗
∼u
˜ x
(\
s
)
∗
8(1) −]
1
(\
s
)
∗
1
0
uu
/
−1
1
0
uE
/
x
8(1) ÷u
:
(\
s
)
∗
(108)
−u
:
(\
s
)
∗
˜
8(e
−i\
s
) ÷t
µ
(1)
=u
x.J
(\
s
)
∗
8(1) −u
:
(\
s
)
∗
˜
8(e
−i\
s
) ÷u
:
(\
s
)
∗
÷t
µ
(1).
Then
n
−1
¯
A
/
O
Z
4
w
O
Z
˜ \ =n
−1
¸
\
s
∈
w
u
x.J
(\
s
)u
\.J
(\
s
)
∗
(109)
∼n
−1
¸
\
s
∈
w
u
x.J
(\
s
)u
x.J
(\
s
)
∗
8(1)
−n
−1
¸
\
s
∈
w
u
x.J
(\
s
)u
:
(\
s
)
∗
˜
8(e
−iw
) ÷t
µ
(1).
Using (51) we ﬁnd that
n
−1
¸
\
s
∈
w
u
x.J
(\
s
)u
:
(\
s
)
∗
= n
−1
¸
\
s
∈
w
u
x
(\
s
)u
:
(\
s
)
∗
÷t
µ
(1) (110)
= n
−1
¸
\
s
∈
w
1
1−e
i\
s
u
:
(\
s
)u
:
(\
s
)
∗
÷t
µ
(1)
→
µ
2=
1−e
iw
]
::
(w).
It follows from (95), (102), (109), and (110) that
ˆ
8
w
→
J
8(1) −
2=
1−e
iw
E
−1
w
]
::
(w)
˜
8(e
−iw
) =8(1) ÷
2=
[1−e
iw
[
2
E
−1
w
]
::
(w)
˜
8(e
−iw
)(e
−iw
−1).
The true coefﬁcient is
8
w
=b(−w) =8(e
−iw
) =8(1) ÷
˜
8(e
−iw
)(e
−iw
−1),
so we have
ˆ
8
w
→
J
8
w
÷
¸
2=
[1−e
iw
[
2
E
−1
w
]
::
(w) −1
¸
˜
8(e
−iw
)(e
−iw
−1)
which gives the stated result since ]
xx
(w) =[1−e
iw
[
−2
]
::
(w).
Part (b): The estimator
˜
8
w
is derived from the augmented spectral regression (32) and the relation
(34). In (32), \
i
and x
i
are ﬁrst detrended in the time domain and then dft’s are taken of the detrended
data and differences of the detrended data. Since u
úx.J
(\
s
) ∼u
:
(\
s
) ÷t
µ
(1), (108) can be written as
u
\.J
(\
s
)
∗
=u
x.J
(\
s
)
∗
8(1) −u
úx.J
(\
s
)
∗
˜
8(e
−i\
s
) ÷u
:
(\
s
)
∗
÷t
µ
(1). (111)
First, take w = 0. In this case, the regressors u
x.J
(\
s
) and u
úx.J
(\
s
) in (111) are asymptotically
orthogonal after appropriate scaling because
n
−1¡2
n
−1
¸
\
s
∈
0
u
˜ x.J
(\
s
)u
:
(\
s
)
∗
=O
µ
(n
−1¡2
), (112)
in view of (99). It then follows from (98), Lemma C(k), (112), and (111) that
n(
˜
8
0
−8(1))
J
→
¸
1
0
E
x.u
E
/
x.u
¸
−1
¸
1
0
E
x.u
JE
:
¸
,
as required.
band spectral regression 1103
Now take the case w ,=0. From part (f) of Lemma C and (52) we have
u
x.z
(\
s
) =u
x.J
(\
s
) =
u
:
(\
s
)
1−e
i\
s
−
e
i\
s
1−e
i\
s
 ˜ x
n
− ˜ x
0
]
n
1¡2
−(n
−3¡2
¯
A
/
D)(n
−1
D
/
D)
−1
(n
1¡2
u
J
(\
s
))
=
u
:
(\
s
)
1−e
i\
s
÷ui
n
(\
s
) ˜ x
n
− ˜ x
0
] −(n
−3¡2
¯
A
/
D)(n
−1
D
/
D)
−1
(n
1¡2
u
J
(\
s
))
=
u
:
(\
s
)
1−e
i\
s
÷.
n
(n
1¡2
u
J
(\
s
)).
Here, n
1¡2
u
J
(\
s
) =O
µ
(1) from (42) and
.
/
n
=
¸
0,
˜ x
n
− ˜ x
0
√
n
¸
−(n
−3¡2
¯
A
/
D)(n
−1
D
/
D)
−1
=O
µ
(1).
Then (111) is
u
\.J
(\
s
)
∗
=
¸
u
:
(\
s
)
1−e
i\
s
÷.
n
(n
1¡2
u
J
(\
s
))
¸
∗
8(1) −u
:
(\
s
)
∗
˜
8(e
−i\
s
) ÷u
:
(\
s
)
∗
÷t
µ
(1) (113)
=u
:
(\
s
)
∗
¸
8(1)
1−e
−i\
s
−
˜
8(e
−i\
s
)
¸
÷u
J
(\
s
)
∗
(n
1¡2
.
n
) ÷u
:
(\
s
)
∗
÷t
µ
(1).
The augmented narrow band regression (32) around frequency w can therefore be written as
u
\.z
(\
s
) = ˜ a
/
1w
¸
u
:
(\
s
)
1−e
i\
s
÷.
n
(n
1¡2
u
J
(\
s
))
¸
÷ ˜ a
/
2w
u
:
(\
s
) ÷t
µ
(1)] ÷residual, (114)
which is asymptotically equivalent to the regression
u
\.z
(\
s
)
∗
=u
:
(\
s
)
∗
˜
b
1, −w
÷(n
1¡2
u
J
(\
s
))
∗
˜
b
2, −w
÷residual (115)
with
˜
b
1, −w
=
˜ a
1, −w
1−e
−iw
÷ ˜ a
2, −w
,
˜
b
2w
= ˜ a
1, −w
.
n
.
In view of (113), (115), and the asymptotic orthogonality of the regressors in (115) (i.e., n
−1
¸
\
s
∈
u
:
(\
s
)(n
1¡2
u
J
(\
s
))
∗
→
µ
0), we ﬁnd that
˜
b
1, −w
→
µ
8(1)
1−e
−iw
−
˜
8(e
−iw
) =
8
w
1−e
−iw
and
√
n
¸
˜
b
1, −w
−
8
w
1−e
−iw
¸
→
J
A
c
0,
2=]
::
(w)
2=]
::
(w)
.
We deduce that
˜
8
w
= ˜ a
1, −w
÷(1−e
−iw
) ˜ a
2, −w
=
˜
b
1, −w
(1−e
−iw
) →
µ
8
w
and
√
n
¸
˜
8
w
−8
w
¸
→
J
A
c
0,
]
::
(w)
]
xx
(w)
,
giving the stated result.
1104 d. corbae, s. ouliaris, and p. c. b. phillips
Proof of Theorem 3: Note that
n(
ˆ
8
]
.
−8
.
) =(n
−2
A
/
W
∗
.O
.WZ
.WA)
−1
(n
−1
A
/
W
∗
.O
.WZ
.W:)
=(n
−2
A
/
O
I
4O
I
A)
−1
(n
−1
A
/
O
I
4O
I
:)
=(n
−2
¯
A
/
O
I
4O
I
¯
A)
−1
(n
−1
¯
A
/
O
I
4O
I
:)
=(n
−2
¯
A
/
O
4Z
4O
4Z
¯
A)
−1
(n
−1
¯
A
/
O
4Z
4O
4Z
:).
Next,
¯
A
/
O
4Z
4 =
¯
A
/
4 −
¯
A
/
F
4Z
4 =
¯
A
/
4 −
¯
A
/
4Z(Z
/
4Z)
−1
Z
/
4
=
¯
A
/
4 −
¯
A
/
4D(D
/
4D)
−1
D
/
4,
and so
¯
A
/
O
4Z
4O
4Z
¯
A =(
¯
A
/
O
4Z
4)(4O
4Z
¯
A) =
¯
A
/
4
¯
A
/
−
¯
A
/
4D(D
/
4D)
−1
D
/
4
¯
A
/
,
which is
¸
\
s
∈
.
u
˜ x
(\
s
)u
˜ x
(\
s
)
∗
−
¸
\
s
∈
.
u
˜ x
(\
s
)u
J
(\
s
)
∗
¸
\
s
∈
.
u
J
(\
s
)u
J
(\
s
)
∗
−1
¸
\
s
∈
.
u
J
(\
s
)u
˜ x
(\
s
)
∗
.
Now, as in Lemma C(a), (b) and (i), we have
n
−2
¸
\
s
∈
.
u
˜ x
(\
s
)u
˜ x
(\
s
)
∗
J
→
1
0
E
x
E
/
x
,
n
−3¡2
¸
\
s
∈
.
u
˜ x
(\
s
)u
J
(\
s
)
∗
J
→
J
1
0
E
x
u
/
,
n
−1
¸
\
s
∈
.
u
J
(\
s
)u
J
(\
s
)
∗
→
1
0
uu
/
.
Thus,
n
−2
¯
A
/
O
4Z
4O
4Z
¯
A
J
→
1
0
E
x
E
/
x
−
1
0
E
x
u
/
1
0
uu
/
−1
1
0
]
0
E
/
x
=
1
0
E
x.u
E
/
x.u
. (116)
Next, consider the limit behavior of
n
−1
¯
A
/
O
4Z
4O
4Z
: =n
−1

¯
A
/
4:−
¯
A
/
4D(D
/
4D)
−1
D
/
4:]
=n
−1
¸
\
s
∈
.
u
˜ x
(\
s
)u
:
(\
s
)
∗
−
n
−3¡2
¸
\
s
∈
.
u
˜ x
(\
s
)u
J
(\
s
)
∗
n
−1
¸
\
s
∈
.
u
J
(\
s
)u
J
(\
s
)
∗
−1
n
−1¡2
¸
\
s
∈
.
u
J
(\
s
)u
:
(\
s
)
∗
.
As in the proof of (91),
n
−1
¸
\
s
∈
.
u
˜ x
(\
s
)u
:
(\
s
)
∗
J
→
1
0
E
x
JE
:
,
band spectral regression 1105
and
n
−1¡2
¸
\
s
∈
.
u
J
(\
s
)u
:
(\
s
)
∗
J
→
1
0
uJE
:
.
Thus,
n
−1
¯
A
/
O
4Z
4O
4Z
:
J
→
1
0
E
x
JE
:
−
1
0
E
x
u
/
1
0
uu
/
−1
1
0
uJE
:
=
1
0
E
x.u
JE
:
.
It follows that
n(
ˆ
8
]
.
−8
.
)
J
→
1
0
E
x.u
E
/
x.u
−1
1
0
E
x.u
JE
:
≡MN
0,
1
0
E
x.u
E
/
x.u
−1
2=]
::
(0)
,
giving the stated result for the band
.
.
For the band
c
.
, we have
√
n(
ˆ
8
]
.
c
−8
.
c ) =(n
−1
A
/
W
∗
.
c
O
.
c
WZ
.
c
WA)
−1
(n
−1¡2
A
/
W
∗
.
c
O
.
c
WZ
.
c
W:) (117)
=(n
−1
¯
A
/
O
4
c
Z
4
c
O
4
c
Z
¯
A)
−1
(n
−1¡2
¯
A
/
O
4
c
Z
4
c
O
4
c
Z
:).
As above,
¯
A
/
O
4
c
Z
4
c
O
4
c
Z
¯
A =(
¯
A
/
O
4
c
Z
4
c
)(4
c
O
4
c
Z
¯
A)
=
¯
A
/
4
c
¯
A
/
−
¯
A
/
4
c
D(D
/
4
c
D)
−
D
/
4
c
¯
A
/
=
¸
\
s
∈
c
.
u
˜ x
(\
s
)u
˜ x
(\
s
)
∗
−
¸
\
s
∈
c
.
u
˜ x
(\
s
)u
J
(\
s
)
∗
¸
\
s
∈
c
.
u
J
(\
s
)u
J
(\
s
)
∗
−1
¸
\
s
∈
c
.
u
J
(\
s
)u
˜ x
(\
s
)
∗
.
From the above expression and Lemma C(d), (e), and (h) we deduce that
n
−1
¯
A
/
O
4
c
Z
4
c
O
4
c
Z
¯
A (118)
J
→
c
.
¸
]
xx
(w) ÷
1
2=
1
[1−e
iw
[
2
E
x
(1)E
x
(1)
/
¸
Jw−
¸
1
2=
E
x
(1)
c
.
e
iw
]
1
(w)
∗
1−e
iw
Jw
¸
¸
1
2=
c
.
]
1
(w)]
1
(w)
∗
Jw
¸
−
¸
1
2=
c
.
]
1
(w)
e
−iw
1−e
−iw
JwE
x
(1)
/
¸
=
c
.
¸
]
xx
(w) ÷
1
2=
1
[1−e
iw
[
2
E
x
(1)E
x
(1)
/
¸
Jw−
1
2=
E
x
(1)E
x
(1)
/
¸
c
.
e
iw
]
1
(w)
∗
1−e
iw
Jw
¸
¸
c
.
]
1
(w)]
1
(w)
∗
Jw
¸
−
¸
c
.
]
1
(w)
e
−iw
1−e
−iw
Jw
¸
.
Next, observe that
¸
c
.
e
iw
]
1
(w)
∗
1−e
iw
Jw
¸¸
c
.
]
1
(w)]
1
(w)
∗
Jw
¸
−
]
1
(w)
is the 1
2
(
c
.
) projection of the function e
iw
(1 −e
iw
)
−1
onto the space spanned by ]
1
(w). When
the deterministic variable z
i
includes a linear time trend, we know from (39) that the vector ]
1
(w)
includes the function e
iw
(1−e
iw
)
−1
as one of its components. Hence, in this case we have
¸
c
.
e
iw
(1−e
iw
)
−1
]
1
(w)
∗
Jw
¸¸
c
.
]
1
(w)]
1
(w)
∗
Jw
¸
−
]
1
(w) =e
iw
(1−e
iw
)
−1
(119)
for w ∈
c
.
. It follows that (118) is simply
c
.
]
xx
(w) Jw.
1106 d. corbae, s. ouliaris, and p. c. b. phillips
Proceeding with the proof, the second factor of (117) decomposes as
n
−1¡2
¯
A
/
O
4
c
Z
4
c
O
4
c
Z
:
=n
−1¡2
¸
\
s
∈
c
.
u
˜ x
(\
s
)u
:
(\
s
)
∗
−
n
−1¡2
¸
\
s
∈
c
.
u
˜ x
(\
s
)u
J
(\
s
)
∗
¸
\
s
∈
c
.
u
J
(\
s
)u
J
(\
s
)
∗
−
¸
\
s
∈
c
.
u
J
(\
s
)u
:
(\
s
)
∗
.
Using (119) and the independence of ˜ x
i
and :
i
we ﬁnd
n
−1¡2
¯
A
/
O
4
c
Z
4
c
O
4
c
Z
:
J
∼n
−1¡2
¸
\
s
∈
c
.
u
˜ x
(\
s
)u
:
(\
s
)
∗
−
1
2=
E
x
(1)
c
.
e
iw
]
1
(w)
∗
1−e
iw
Jw
1
2=
c
.
]
1
(w)]
1
(w)
∗
Jw
−
n
−1¡2
¸
\
s
∈
c
.
]
1
(\
s
)u
:
(\
s
)
∗
J
∼n
−1¡2
¸
\
s
∈
c
.
¸
u
˜ x
(\
s
) −
1
2=
E
x
(1)
c
.
e
iw
]
1
(w)
∗
1−e
iw
Jw
1
2=
c
.
]
1
(w)]
1
(w)
∗
Jw
−
]
1
(\
s
)
¸
u
:
(\
s
)
∗
=n
−1¡2
¸
\
s
∈
c
.
¸
u
˜ x
(\
s
) −E
x
(1)
e
i\
s
1−e
i\
s
¸
u
:
(\
s
)
∗
=n
−1¡2
¸
\
s
∈
c
.
¸
1
1−e
i\
s
u
¡
(\
s
) −
e
i\
s
1−e
i\
s
 ˜ x
n
− ˜ x
0
]
n
1¡2
−E
x
(1)
e
i\
s
1−e
i\
s
¸
u
:
(\
s
)
∗
=n
−1¡2
¸
\
s
∈
c
.
¸
1
1−e
i\
s
u
¡
(\
s
)
¸
u
:
(\
s
)
∗
÷t
µ
(1)
J
∼A
0,
2=
n
¸
\
s
∈
c
.
¸
1
[1−e
i\
s
[
2
u
¡
(\
s
)u
¡
(\
s
)
∗
¸
]
::
(\
s
)
J
∼A
0, 2=
c
.
]
xx
(w)]
::
(w) Jw
.
We deduce that
√
n(
ˆ
8
]
.
c
−8
.
c )
J
→A
0,
¸
c
.
]
xx
(w) Jw
¸
−1
¸
2=
c
.
]
xx
(w)]
::
(w) Jw
¸¸
c
.
]
xx
(w) Jw
¸
−1
,
giving the stated result.
Proof of Theorem 3
/
: Part (a): First note that
u
x
(\
s
) =H
/
2
u
z
(\
s
) ÷u
˜ x
(\
s
) =H
/
2
o
n
u
J
(\
s
) ÷u
˜ x
(\
s
),
with a similar expression for u
\
(\
s
). The regression (45) is equivalent to the following regression
with frequency domain detrended data:
u
]
˜ \.J
(\
s
) = ˜ c
/
1w
u
]
˜ x.J
(\
s
) ÷ ˜ c
/
2w
u
]
ú˜ x.J
(\
s
) ÷residual, (120)
band spectral regression 1107
where
u
]
x.J
(\
s
)
∗
=u
x
(\
s
)
∗
−u
J
(\
s
)
∗
¸
\
s
∈
w
u
J
(\
s
)u
J
(\
s
)
∗
−1
¸
\
s
∈
w
u
J
(\
s
)u
x
(\
s
)
∗
=u
˜ x
(\
s
)
∗
−u
J
(\
s
)
∗
¸
\
s
∈
w
u
J
(\
s
)u
J
(\
s
)
∗
−1
¸
\
s
∈
w
u
J
(\
s
)u
˜ x
(\
s
)
∗
=u
]
˜ x.J
(\
s
)
∗
, say,
with a similar expression for u
]
\.J
(\
s
) =u
]
˜ \.J
(\
s
), and
u
]
ú˜ x.J
(\
s
)
∗
=u
ú˜ x
(\
s
)
∗
−u
J
(\
s
)
∗
¸
\
s
∈
w
u
J
(\
s
)u
J
(\
s
)
∗
−1
¸
\
s
∈
w
u
J
(\
s
)u
ú˜ x
(\
s
)
∗
=u
:
(\
s
)
∗
÷t
µ
(1).
For w = 0, as in the proof of Theorem 2
/
, the regressors in (120) are asymptotically orthogonal
because
n
−3¡2
¸
\
s
∈
0
u
]
˜ x.J
(\
s
)u
:
(\
s
)
∗
=O
µ
(n
−1¡2
).
Next, using Lemma C(j)–(m), we ﬁnd as in (116) above that
n
−2
¸
\
s
∈
0
u
]
˜ x.J
(\
s
)u
]
˜ x.J
(\
s
)
∗
J
→
1
0
E
x.u
E
/
x.u
,
and as in (89) we get
n
−1
¸
\
s
∈
0
u
]
˜ x.J
(\
s
)u
:
(\
s
)
∗
J
→
1
0
E
x.u
JE
:
.
It follows using (15) and these two results that
n
˜
8
]
w
−8
0
J
→
¸
1
0
E
x.u
E
/
x.u
¸
−1
¸
1
0
E
x.u
JE
:
¸
.
Next consider the w ,=0 case. First we simplify the regression equation (120):
u
]
˜ \.J
(\
s
) = ˜ c
/
1w
u
]
˜ x.J
(\
s
) ÷ ˜ c
/
2w
u
]
ú˜ x.J
(\
s
) ÷residual (121)
= ˜ c
/
1w
u
]
˜ x.J
(\
s
) ÷ ˜ c
/
2w
u
:
(\
s
) ÷t
µ
(1)] ÷residual.
Using (73) and Lemma C(n)–(o) we obtain
u
]
˜ x.J
(\
s
)
∗
=u
˜ x
(\
s
)
∗
−u
J
(\
s
)
∗
¸
\
s
∈
w
u
J
(\
s
)u
J
(\
s
)
∗
−1
¸
\
s
∈
w
u
J
(\
s
)u
˜ x
(\
s
)
∗
(122)
=
u
:
(\
s
)
∗
1−e
−i\
s
−
e
−i\
s
1−e
−i\
s
 ˜ x
n
− ˜ x
0
]
n
1¡2
−
1
√
n
]
1
(w)
∗
¸
n
n
]
1
(w)]
1
(w)
∗
¸
−
¸
n
√
n
]
1
(w)
e
−iw
1−e
−iw
E(1)
¸
1108 d. corbae, s. ouliaris, and p. c. b. phillips
=
u
:
(\
s
)
∗
1−e
−i\
s
−
e
−i\
s
1−e
−i\
s
 ˜ x
n
− ˜ x
0
]
n
1¡2
−
e
−iw
1−e
−iw
E(1)
=
u
:
(\
s
)
∗
1−e
−i\
s
÷t
µ
(1).
So the regression (121) is asymptotically equivalent to
u
]
˜ \.J
(\
s
)
∗
=
¸
u
:
(\
s
)
1−e
i\
s
÷t
µ
(1)
¸
∗
˜ c
1w
÷u
:
(\
s
) ÷t
µ
(1)]
∗
˜ c
2w
÷residual. (123)
From (122) and (15) we deduce, as in Theorem 2
/
, that
˜
b
]
1, −w
:=
˜ c
1, −w
1−e
−iw
÷ ˜ c
2, −wµ
→
µ
8(1)
1−e
−iw
−
˜
8(e
−iw
) =
8
w
1−e
−iw
and
√
n
¸
˜
b
1, −w
−
8
w
1−e
−iw
¸
→
J
A
c
0,
]
::
(w)
]
::
(w)
.
It follows that
˜
8
]
w
= ˜ c
1, −w
÷(1−e
−iw
) ˜ c
2, −w
=
˜
b
1, −w
(1−e
−iw
) →
µ
8
w
and
√
n
¸
˜
8
]
w
−8
w
¸
→
J
A
c
0,
]
::
(w)
]
xx
(w)
.
REFERENCES
Choi, I., and P. C. B. Phillips (1993): “Testing for a Unit Root by Frequency Domain Regression,”
Journal of Econometrics, 59, 263–286.
Corbae, D., S. Ouliaris, and P. C. B. Phillips (1997): “Band Spectral Regression with Trending
Data,” Cowles Foundation Discussion Paper, Yale University, 1163.
Engle, R. (1974): “Band Spectrum Regression,” International Economic Review, 15, 1–11.
Frisch, R., and F. V. Waugh (1993): “Partial Time Series Regression as Compared with Individual
Trends,” Econometrica, 1, 211–223.
Geweke, J. (1982): “Measurement of Linear Dependence and Feedback Between Multiple Time
Series,” Journal of the American Statistical Association, 77, 304–313.
Gradshteyn, I. S., and I. M. Ryzhik (1965): Tables of Integrals, Series and Products. New York:
Academic Press.
Granger, C. W. J., and J.L Lin (1995): “Causality in the Long Run,” Econometric Theory, 11,
530–536.
Grenander, U., and M. Rosenblatt (1957): Statistical Analysis of Stationary Time Series.
New York: Wiley.
Hannan, E. J. (1963a): “Regression for Time Series,” in Time Series Analysis, ed. by M. Rosenblatt.
New York: Wiley.
(1963b): “Regression for Time Series with Errors of Measurement,” Biometrika, 50, 293–302.
(1970): Multiple Time Series. New York: Wiley.
(1973): “Central Limit Theorems for Time Series Regression,” Zeitschrift für Wahrschein
lichkeitsheorie und verwandte Gebiete, 26, 157–170.
Hannan, E. J., and P. J. Thomson (1971): “Spectral Inference over Narrow Bands,” Journal of
Applied Probability, 8, 157–169.
Hannan, E. J., and P. M. Robinson (1973): “Lagged Regression with Unknown Lags,” Journal of
the Royal Statistical Society, Series B, 35, 252–267.
band spectral regression 1109
Hosoya, Y. (1991): “The Decomposition and Measurement of the Interdependence between
Secondorder Stationary Processes,” Probability Theory and Related Fields, 88, 429–444.
Kim, C., and P. C. B. Phillips (1999): “Fully Modiﬁed Estimation of Fractional Cointegration
Models,” Yale University, Mimeo.
Marinucci, D. (1998): “Band Spectrum Regression for Cointegrated Time Series with Long Mem
ory Innovations,” LSE Econometrics Discussion Paper #353.
Phillips, P. C. B. (1991a): “Spectral Regression for Cointegrated Time Series,” in Nonparamet
ric and Semiparametric Methods in Economics and Statistics, ed. by W. Barnett, J. Powell, and
G. Tauchen. Cambridge: Cambridge University Press.
(1991b): “Optimal Inference in Cointegrated Systems,” Econometrica, 59, 283–306.
(1999): “Discrete Fourier Transforms of Fractional Processes,” Cowles Foundation Discussion
Paper #1243, Yale University.
(2000): “Unit Root Log Periodogram Regression,” Cowles Foundation Discussion Paper
#1244, Yale University.
Phillips, P. C. B., and B. Hansen (1990): “Statistical Inference in Instrumental Variable Regres
sion with I(1) processes,” Review of Economic Studies, 57, 99–125.
Phillips, P. C. B., and C. C. Lee (1996): “Efﬁciency Gains from Quasidifferencing under Nonsta
tionarity,” in Athens Conference on Applied Probability and Time Series: Essays in Memory of E. J.
Hannan, ed. by P. M. Robinson and M. Rosenblatt. New York: SpringerVerlag.
Phillips, P. C. B., and M. Loretan (1991): “Estimating Longrun Economic Equilibria,” Review
of Economic Studies, 59, 407–436.
Phillips, P. C. B., and V. Solo (1992): “Asymptotics for Linear Processes,” Annals of Statistics,
20, 971–1001.
Robinson, P. M. (1991): “Automatic Frequency Domain Inference,” Econometrica, 59, 1329–1363.
Robinson P. M., and D. Marinucci (1998): “Semiparametric Frequency Domain Analysis of
Fractional Cointegration,” LSE Econometrics Discussion Paper #348.
Saikkonen, P. (1991): “Asymptotically Efﬁcient Estimation of Cointegration Regressions,” Econo
metric Theory, 7, 1–21.
Shorack, G. R., and J. A. Wellner (1986): Empirical Processes with Applications to Statistics.
New York: Wiley.
Stock, J. H., and M. W. Watson (1993): “A Simple Estimator of Cointegrating Vectors in Higher
Order Integrated Systems,” Econometrica, 61, 783–820.
Xiao, Zhijie, and Peter C. B. Phillips (1998): “Higher Order Approximations for Frequency
Domain Time Series Regression,” Journal of Econometrics, 86, 297–336.
(1999): “Higher Order Approximations for Wald Statistics in Time Series Regressions with
Integrated Processes,” Yale University, Mimeo.
Econometrica, Vol. 70, No. 3 (May, 2002), 1067–1109
BAND SPECTRAL REGRESSION WITH TRENDING DATA By Dean Corbae, Sam Ouliaris, and Peter C. B. Phillips1
Band spectral regression with both deterministic and stochastic trends is considered. It is shown that trend removal by regression in the time domain prior to band spectral regression can lead to biased and inconsistent estimates in models with frequency dependent coefﬁcients. Both semiparametric and nonparametric regression formulations are considered, the latter including general systems of twosided distributed lags such as those arising in lead and lag regressions. The bias problem arises through omitted variables and is avoided by careful speciﬁcation of the regression equation. Trend removal in the frequency domain is shown to be a convenient option in practice. An asymptotic theory is developed and the two cases of stationary data and cointegrated nonstationary data are compared. In the latter case, a levels and differences regression formulation is shown to be useful in estimating the frequency response function at nonzero as well as zero frequencies. Keywords: Band spectral regression, deterministic and stochastic trends, discrete Fourier transform, distributed lag, integrated process, leads and lags regression, nonstationary time series, twosided spectral BN decomposition.
1 introduction Hannan’s (1963a, b) bandspectrum regression procedure is a useful regression device that has been adopted in some applied econometric work, notably Engle (1974), where there are latent variables that may be regarded as frequency dependent (like permanent and transitory income) and where there is reason to expect that the relationship between the variables may depend on frequency. More recently, band spectral regression has been used to estimate cointegrating relations, which describe low frequency or long run relations between economic variables. In particular, Phillips (1991a) showed how frequency domain regressions that concentrate on a band around the origin are capable of producing asymptotically efﬁcient estimates of cointegrating vectors. Interestingly, in that case the spectral methods are used with nonstationary integrated time series in precisely the same way as they were originally designed for stationary series, on which there is now a large literature (see, inter alia, Hannan (1970 Ch. VII), Hannan and Thomson (1971), Hannan and Robinson (1973), and Robinson (1991)). Related methods were used in Choi and Phillips (1993) to construct frequency domain tests for unit roots. In all of this work, the capacity of frequency domain techniques to deal nonparametrically with short memory components in
Our thanks go to a coeditor and three referees for helpful comments. An early draft was written in the summer of 1994. The ﬁrst complete version of the paper was written over the period 1 January– 1 July 1997 while Phillips was visiting the University of Auckland. Lemmas A–D are due to Phillips and are taken from other work to make this paper self contained. Phillips thanks the NSF for research support under Grant Nos. SBR 9422922, SBR 9730295, and SES 0092509. 1067
1
1068
d. corbae, s. ouliaris, and p. c. b. phillips
the data is exploited. Robinson (1991), who allowed for band dependent model formulations, showed how, in such semiparametric models, data based methods can be used to determine the smoothing parameter that plays a central role in the treatment of the nonparametric component. Other data based methods have been used directly in the study of cointegrating relations and testing problems by Xiao and Phillips (1998, 1999). To the extent that relations between variables may be band dependent, it is also reasonable to expect that causal relations between variables may vary according to frequency. Extending work by Geweke (1982), both Hosoya (1991) and Granger and Lin (1995) have studied causal links from this perspective and constructed measures of causality that are frequency dependent. Most recently, frequency band methods have been suggested for the estimation of models with long memory components (e.g., Marinucci (1998), Robinson and Marinucci (1998), Kim and Phillips (1999)), where salient features of the variables are naturally evident in frequency domain formulations. This paper studies some properties of frequency band regression in the presence of both deterministic and stochastic trends, a common feature in economic time series applications. In such cases, a seemingly minor issue relates to the manner of deterministic trend removal. In particular, should the deterministic trends be eliminated by regression in the time domain prior to the use of the frequency domain regression or not? Since spectral regression procedures were originally developed for stationary time series and since the removal of deterministic trends by least squares regression is well known to be asymptotically efﬁcient under weak conditions (Grenander and Rosenblatt (1957, p. 244)), it may seem natural to perform the trend removal in the time domain prior to the use of spectral methods. Indeed, this is recommended in Hannan (1963a, p. 30; 1970, pp. 443–444), even though the development of spectral regression there and in Hannan (1970, Ch. VII.4) allows for regressors like deterministic trends that are susceptible to a generalized harmonic analysis, satisfying the socalled Grenander conditions (Grenander and Rosenblatt (1957, p. 233), Hannan (1970, p. 77)). Theorem 11 of Hannan (1970, p. 444) conﬁrms that such time domain detrending followed by spectral regression is asymptotically efﬁcient in models where all the variables are trend stationary and where the model coefﬁcients do not vary with frequency. In time domain regression, the Frisch–Waugh (1933) theorem assures invariance of the regression coefﬁcients to prior trend removal or to the inclusion of trends in the regression itself. Such invariance is often taken for granted in empirical work. However, as will be shown here, invariance does not always apply for bandspectrum regression when one switches from time domain to frequency domain regressions. In particular, detrending by removing deterministic components in the time domain and then applying bandspectrum regression is not necessarily equivalent to detrending in the frequency domain and applying bandspectrum regression. The reason is that a band dependent spectral model is similar in form to a linear regression with a structural change in the coefﬁcient vector. So, when there are deterministic trends in the model, the original
band spectral regression
1069
deterministic regressors also need to be augmented by regressors that are relevant to the change period (here, the frequency band where the change occurs). Thus, the seemingly innocuous matter of detrending can have some nontrivial consequences in practice. In particular, detrending in the time domain yields estimates that can be biased in ﬁnite samples, and, in the case of nonstationary data, inconsistent. The appropriate procedure is to take account of the augmented regression equation prior to detrending and this can be readily accomplished by including the deterministic variables explicitly in the frequency domain regression. These issues are relevant whenever band spectral methods are applied, including those cases that involve cointegrating regressions. In the latter case, a levels and differences regression formulation is shown to be appropriate and useful in estimating the frequency response function at nonzero as well as zero frequencies. The present paper provides an asymptotic analysis of frequency domain regression with trending data for the two cases of stationary and cointegrated nonstationary data. We consider both semiparametric and fully nonparametric formulations of the regression model, in both cases allowing the model coefﬁcients to vary with frequency. In dealing with the nonstationary case, the paper introduces some new methods for obtaining a limit theory for discrete Fourier transforms (dft’s) of integrated time series and makes extensive use of two sided BN decompositions, which lead to levels and differences model formulations. This asymptotic theory simpliﬁes some earlier results given in Phillips (1991a) and shows that the dft’s of an I(1) process are spatially (i.e., frequencywise) dependent across all the fundamental frequencies, even in the limit as the sample size n → , due to leakage from the zero frequency. This leakage is strong enough to ensure that smoothed periodogram estimates of the spectrum away from the origin are inconsistent at all frequencies = 0. The techniques and results given here should be useful in other regression contexts where these quantities arise. Some extensions of the methods to cases of data with long range dependence are provided in Phillips (1999). Most of our notation is standard: [a] signiﬁes the largest integer not exceeding a > signiﬁes positive deﬁniteness when applied to matrices, a∗ is the complex conjugate transpose of the matrix a, a is a matrix norm of a, a− is the Moore Penrose inverse of a, Pa = a a∗ a − a∗ is the orthogonal projector onto the range of a, L2 0 1 is the space of square integrable functions on [0, 1], 1[A] is the indip d d cator of A, ∼ signiﬁes ‘asymptotically distributed as’ and → and → are used to denote weak convergence of the associated probability measures and convergence in probability, respectively, as the sample size, n → ; I(1) signiﬁes an integrated process of order one, BM denotes a vector Brownian motion with covariance 1 1 matrix , and we write integrals like 0 B r dr as 0 B, or simply B if there is no ambiguity over limits; MN 0 G signiﬁes a mixed normal distribution with matrix mixing variate G, Nc 0 G signiﬁes a complex normal distribution with covariance matrix G, the discrete Fourier transform (dft) of at t = 1 n is √ written wa = 1/ n n at ei t , and s = 2 s/n s = 0 1 n − 1 are t=1 the fundamental frequencies.
it is not required for the consistent estimation of . The mechanism linking the observed variables xt . unobserved disturbances t . we introduce the following two alternative assumptions. The series xt therefore has both a deterministic ˜ component involving the sequence zt and a stochastic (latent) component xt . a vector Brownian motion of dimension k + 1 with covariance matrix = 2 f 0 (e. In the nonstationary case. The shocks t = t vt satisfy Assumption 1 with partitioned as spectral density f f with f fvv = f 0 0 fvv >0 ∀ .4)). and a dependent variable yt is made explicit later. xt is taken to be independent of t . b)). The assumption is standard for consistent estimation in the stationary case (as in Hannan (1963a.1070 d. Theorem 3. The vector process B and matrix can be partitioned conformably with as B = B Bx and = 0 0 xx ˜ ˜ where xx > 0.. the specof t is tral density matrix f f with f fxx = f 0 >0 ∀ . Accordingly. Assumptions 1 and 2 sufﬁce for partial sums of t to satisfy the functional law d nr n−1/2 t=1 t → B r = BM . so that xt is a full rank vector I(1) process. ˜ Assumption 1: t = t xt is a jointly stationary time series with Wold representation t = j=0 Cj t−j . In developing our theory it will be convenient to allow for both stationary and non˜ stationary xt . initialized at t = 0 by any Op (1) random variable. b. s. Phillips and Solo (1992. phillips 2 models and estimation preliminaries Let xt t = 1 (1) xt = 2 zt n be an observed time series generated by ˜ + xt ˜ where zt is a p + 1dimensional deterministic sequence and xt is a zero mean kdimensional time series. ouliaris. 0 fxx ˜ ˜ Assumption 2: xt is an I(1) process satisfying xt = vt . This amounts to a strict exogeneity assumption when we introduce a generating mechanism for a dependent variable yt . corbae. Partitioned conformably with t . and p. c. where t = iid 0 with ﬁnite fourth moments and the 1 coefﬁcients Cj satisfy j=0 j 2 Cj < .g. In all cases.
at bands away from the origin. The limit functions u r are linearly independent in 1 L2 0 1 and n−1 n dt dt → 0 uu > 0. Now let Z (respectively. PD and I − PD ). PZ be the projector onto the range of Z. and deﬁne and the limit results in Lemma C below. b = ei = j=− j eij . We make the following assumptions concerning the deterministic sequence zt . Under stationarity (Assumption 1). . the cross ˜ ˜ between yt and xt satisﬁes spectrum fyx (4) fyx =b fxx whose complex conjugate transpose gives the complex regression coefﬁcient (5) =b − = fxx −1 fxy As argued in Hannan (1963a. spectral methods provide a natural mechanism for focusing attention on a particular frequency. for instance to trigonometric polynomials. D) be the n×p t=1 observation matrix of the nonstochastic regressors zt (respectively. PZ = PD and QZ = QD . and for performing estimation of . The Nonparametric Model The type of model we have in mind for the stochastic component of the data allows for the regression coefﬁcient to vary across frequency bands. However. some version of incoherency between the regressors and errors is required for consistency and asserting incoherency enables us to examine the bias and inconsistency effects of detrending in such regressions. just as in the ˜ case of stationary xt . Then n (2) d nr = −1 n z nr →u r = 1 r rp uniformly in r ∈ 0 1 . or band of frequencies.band spectral regression 1071 (low frequency) cointegrating regression components. b). after some appropriate extensions of our dft formulae in Lemma A np . suppose a (latent) dependent variable yt is related to xt and t according to a (possibly twosided) distributed lag (3) ˜ yt = ˜ j xt−j + t = j=− ˜ L xt + t where the transfer function of the ﬁlter. t p . although we anticipate that our conclusions about bias and inconsistency will apply more generally. Clearly. Let n = diag 1 n dt = −1 zt . and QZ = In − PZ be the residual projection matrix (respectively. standardized regressors dt ). the We conﬁne our treatment to time polynomials and set zt = 1 t most common case in practice. is assumed to converge for all ∈ − . In the time ˜ ˜ domain.
corbae. viz. and p. suppose that the coefﬁcients j in (3) are such that L has a valid twosided spectral BN decomposition (see part (b) of Lemma D in the Appendix). viz. ˜ The case of integrated xt (Assumption 2) can be handled in a similar way. Then.1072 d. setting = s in (7) and taking dft’s at frequency s we obtain the following frequency domain form of (3): (8) wy ˜ s =b t s wx ˜ s +w s 1 ˜ + √ e−i s a n s0 ˜ −a sn ˜ ˜ For stationary xt a (9) wy ˜ s is also stationary and (8) can be written as s =b wx ˜ s +w s 1 + Op √ n giving an approximate frequency domain version of (3). Here it is useful to apply to (3) the twosided BN decomposition at unity (part(a) of Lemma D). (6) where L = ei + ˜ e−i L e−i L − 1 k=j+1 − j k=− ke ik j ≥0 ik ˜ L = j=− ˜ j Lj ˜ j = ke j <0 with (7) j=− ˜j 2 < . phillips Next. Using (6) in (3) we have t ˜ yt = = b =b ˜ L xt + + ˜ e−i L e−i L − 1 ˜ xt + e −i ˜ xt + t ˜ L−1 a t + t ˜ ˜ in which a t = ˜ e−i L xt . s. ouliaris. b. (10) 1 + ˜ L L−1 where k ˜ j = k=j+1 ˜ j Lj ˜ L = j j=− − k L = k=− j ≥0 j <0 so that (3) can be interpreted as (11) ˜ yt = ˜ 1 xt + t − ˜ L vt = ˜ 1 xt + ˜ t say . c.
A notable recent example arises in the simpliﬁcation of cointegrating systems to single equation formulations where twosided . we can recover the spectral coefﬁcient 1 and ˜ ei s = ˜ e−i s ∗ in (15) using the fact that (16) b − = e−i = 1 + ˜ e−i e−i − 1 Causal economic relationships are often formulated as in (3) with onesided relations where j = 0 for j < 0. but twosided relations are not excluded. where it is called augmented frequency domain regression. the equation could be estimated in the frequency domain by band spectral regression of wy s on wx s and w x s for frequencies s ˜ ˜ ˜ centered around some frequency . Assuming it is valid. the spectral BN decomposition of ˜ L at frequency s is ˜ L = ˜ ei where (14) ≈ s s + ≈ s e−i s L e−i s L − 1 is constructed as in Lemma D(b). It follows that (12) has the form = 1 wx ˜ s wy ˜ s − ˜ ei s wv s +w s 1 + Op √ n giving an approximate frequency domain version of the cointegrated model that ˜ applies for all frequencies. For s centered around = 0. They are well known to occur in situations where there is feedback in both directions ˜ ˜ between the variables y and x. Since vt = xt .band spectral regression 1073 ˜ ˜ which is a cointegrating equation between yt and xt with cointegrating vector 1 − 1 and composite error ˜ t = t − ˜ L vt . we obtain (12) (13) wy ˜ s = =b 1 wx ˜ s s s +w +w s s −w −w s s wx ˜ + 1 −b s wx ˜ s The ﬁnal two terms of (13) are important. this procedure was suggested originally in Phillips (1991b) and Phillips and Loretan (1991). Setting t = ˜ L vt and taking the dft of (11). revealing that the approximation (9) fails in the nonstationary case and has an error of speciﬁcation that arises from omitted variables. In the general case where s is centered on = b − from the coefﬁcients = 0. Then w s = ˜ ei = ˜ ei s wv wv s 1 ˜ + √ e−i s v n 1 + Op √ n s0 ˜ −v sn s s ˜ where v (15) st = ≈ s e−i s L vt is stationary. it is apparent from (15) that if data on ˜ ˜ yt and xt were available. although in that work deterministic trends were excluded.
˜ and (11) in the cointegrated case with xt t satisfying Assumption 2. xt . This coefﬁcient will vary according to frequency (unless j = 0 for j = 0) and can be estimated by nonparametric regression techniques. A − 0 ∪ 0 for some given frequency 0 > 0 and deﬁne the frequency dependent coefﬁcient vector (18) b = A1 ∈ A + Ac 1 ∈ c A where A is a kvector of parameters pertinent to the band A . The natural statistical approach would appear to be deterministic trend removal by regression of yt xt on zt . This formulation A of b enables us to separate low frequency ≤ 0 from high frequency . we postulate the dual bands A = − 0 0 and c = − . Stock and Watson (1993)). ˜ The mechanism linking the observed dependent variable yt to zt and yt (and. This case is analogous to a regression model with a structural break and that analogy helps to explain why time and frequency domain detrending have different effects. corbae.1074 d. This case captures the essential features of the problem that we intend to discuss. let us say on two ﬁxed frequency bands. b. The nonparametric model can be interpreted as a limiting case in which one of the bands shrinks to a particular frequency as the sample size increases. The observed series yt and xt therefore have deterministic trends and stochastic components that are linked by a system of distributed lags with stationary errors. c. A Fixed Band Regression Model A simple situation of interest occurs when the response function b is constant. It has the advanis real and parametric and can be estimated at parametric rates tage that b when the bands are of ﬁxed positive length. The latter can be conducted in the frequency domain where the coefﬁcient of interest is the transfer function of the ﬁlter b . ˜ hence. and t ) is now given by (17) yt = zt ˜ 1 + yt ˜ ˜ where yt follows (3) in the stationary case with xt t satisfying Assumption 1. and will be considered at some length here and in the asymptotic theory. One of the aims of this paper is to examine the ﬁnite sample and asymptotic performance of this natural procedure against that of a procedure that seeks to perform the regression analysis in the frequency domain directly on the observed quantities yt xt zt rather than on detrended data. To this end. To promote the analogy with a structural break model we will develop a conventional regression model formulation for the data. followed by an analysis of the system of lagged dependencies between the residuals from these regressions. is easy to analyze. and Ac is the corresponding kvector of parameters pertinent to the band c . s. phillips dynamic regressions like (11) arise and where they have been used in regression analysis to obtain efﬁcient estimates of the cointegrating vector (Phillips and Loretan (1991). Saikonnen (1991). ouliaris. and p.
b = j=− j eij . say A . This mechanism has the advantage of revealing the impact of the shift in (18) in terms of a standard regression model with a structural change. Deﬁne: • A = n × n selector matrix that zeroes out frequencies in W X that are not relevant to the primary band of interest. Note that Ac A = AAc = 0. as shown in Remark (i) following Lemma D in the Appendix. Ac W = I − A W extracts the residual frequencies over c . and the matrix W X has kth row given by the dft wx s with ˜ s = k − 1. y = y1 yn is generated in the frequency domain in dft form by the system (20) (21) ˜ AW y = AW X c c A + AW Ac ˜ A Wy = A WX + Ac W where = 1 n . A = W ∗ AW and c = W ∗ Ac W = I − . We introduce the following matrices that are used to provide a frequency dependent structure to the model for yt . Equations (20) and (21) generate observations in the frequency domain that correspond to the model (8) in the nonparametric case . W ∗ W = In . where E1n = 0 1 In−1 0 is an n × n permutation matrix that reorders the rows of U so that the last row becomes the ﬁrst and succeeding rows are simply advanced by one position. ˜ ˜ Let X = x1 xn be the n × k matrix of observations of the exogenous √ ˜ regressors xt .band spectral regression 1075 > 0 responses in a regression context. • ˜ ˜ ˜ Then. let U be the n × n unitary matrix with j k element e2 ijk/n / n. Then. of (18) are A 0 + Ac − 0 j =0 1 = b e−ij dw = (19) j sin j 0 2 − j =0 A − Ac j so that the ﬁlter in (3) is symmetric and twosided. Then. and let W = E1n U . The coefﬁcients j in the Fourier representation. the BN decomposition of L is still valid in this case. This is an interesting case where the coefﬁcients j decay slowly and do not satisfy the sufﬁcient condition 1 j 2 j < given in Lemma D for the validity of the BN decomposition of − L = j=− Lj . Its stochas˜ ˜ tic component yt is generated from xt and t by means of a triangular array formulation which we explain as follows. an alternate frequency domain approach that uses (18) is possible. Nevertheless. ˜ While data for yt can be generated by (3) and (19) by truncating the ﬁlter. the dependent variable yt is assumed to have both deterministic and stochastic components. Like xt . for each n.
In fact. b. Observe that the trend regressors zt now appear in the model with band dependent coefﬁcients. (22) is a variable (band dependent) coefﬁcient time series regression with exogenous regressors and stationary errors under Assumption 1. Thus. as is apparent from (13) and (15). ouliaris. By construction. as might be expected immediately from (1) and (23). correct trend removal is accomplished by the use of the operator QV = I − PV . The formulations (24)–(26) make it clear that detrending the data in the time domain is not a simple matter of applying the projection matrix QZ . ˜ ˜ In (22). just like the exogenous regressors xt . In the stationary case. it follows that the observed data satisfy the model (24) y=Z 1− 2 A +X A+ c Z 2 A− Ac − c X A− Ac + or. this omitted term is 1 Op n− 2 as seen in (9). In consequence. and a cointegrating regression with exogenous regressors and band dependent coefﬁcients under Assumption 2. Adding (20) and (21) and multiplying by W ∗ gives (22) ˜ y= X A+ c X Ac + ˜ which is the time domain generating mechanism for y. In the nonstationary case. Note that model (22) is semiparametric: parametric in the regression coefﬁcients A and Ac and nonparametric in the regression errors . where V = Z c Z or. the omitted term is nonnegligible. (25) or (26) y= Z 1− 2 A y=Z 1− 2 Ac + Z 2 Ac − A +X Ac − X Ac − A + + c Z 1− 2 Ac + X A+ c X Ac + These models extend (22) to cases where the observed data contain deterministic trends. the matri˜ ces and c have elements that depend on n. the two generating mechanisms differ in the nonstationary case and this leads to a difference in the asymptotic theory between the two models that will ﬁgure in our discussion in Section 5. corbae. t xt satisﬁes Assumption 1 in the stationary case and t xt satisﬁes Assumption 2 in the nonstationary case. and p. phillips with the ﬁnal term of (8) omitted. however. c. s. equivalently. although we do not emphasize this by using an additional subscript and we will have no need of triangular array limit theory in our asymptotic development.1076 d. We now suppose that the observed series yt has deterministic components like ˜ xt in (1) and is related to the unobserved component yt by means of (23) y=Z ˜ 1 +y Using (22) and (23). . and it follows that yt t = 1 n has a triangular array structure. so the difference between the generating mechanisms is negligible asymptotically.
This estimator can be constructed for the band A and then has the form (27) ˆ A = X QZ Q Z X −1 X QZ Q Z y with a corresponding formula for ˆ Ac . we have = In and c is null. so that V = Z and usual detrending A by QZ is appropriate. as we have discussed in the introduction. This procedure follows Hannan’s (1963a. conventional detrending by QZ implicitly assumes that there is no variation in the coefﬁcient across frequency bands. If there are such variations. In that case (26) reduces to y=Z 1− 2 +X + Put another way. the estimator over the band c . we ﬁnd (28) ˆA = = = A+ X QZ Q Z X X QZ Q Z X X QZ Q Z X −1 X QZ Q Z − c c A− c c Z Ac 2 A− Ac X + A− A− Ac Ac A− A− −1 −1 X QZ Q Z X QZ Q Z X X − − and the corresponding formula for ˆ Ac is (29) ˆ Ac = Ac − X QZ c QZ X −1 X QZ c QZ X Ac − A − It follows that (30) and E ˆ Ac X = Ac E ˆA X = A− X QZ Q Z X −1 X QZ Q Z c X A− Ac − X QZ c QZ X −1 X QZ c QZ X Ac − A . Of course. when A = In the coefﬁcient is invariant across bands = say .band spectral regression 1077 equivalently. b) recommendation for dealing with deterministic trends and. In formA ing ˆ A and ˆ Ac . like biased and inconsistent estimates of the regression coefﬁcients A and Ac . Fixed Band Regressions and Misspeciﬁcation Bias A simple illustration with the Hannan (1963a) inefﬁcient bandspectrum regression estimator shows the effects of such misspeciﬁcation. Using (24) and (27). there will be misspeciﬁcation bias from omitted variables in the deterministic detrending. is the conventional approach in this context. in view of (26) V = Z c Z . the data are ﬁltered by a trend removal regression via the projection QZ before performing the bandspectrum regression. then correct detrending needs to take into account the effects of such variations on the trends. as in (26). Methods that rely on preﬁltering by means of QZ do not fully remove the trends and this can have important consequences. Otherwise.
Deﬁne X = X X and a = a1 a2 . We write the estimator as (31) ˆ = X QZ QZ X −1 X QZ QZ y From the form of (15) it is apparent that the narrow band regression leading ˜ to (31) omits the term involving the dft. rather than over a ﬁxed discrete band. In fact. b. The following lemma gives general recursion formulae for the quantity Wk s = n t k ei s t . An alternate nonparametric procedure is to use an augmented narrow band regression in which the differences are included. as we shall see in Section 5. So this narrow band regression also involves omitted variable misspeciﬁcation.. To ﬁx ideas. when ˜ xt is integrated the bias in ˆ Ac does not always disappear in the limit. In that case the band spectral estimator has the same form as (27) but the selector matrix A = A (respectively. phillips Hence. bandspectrum regression yields biased estimates of the coefﬁcients when A = Ac and trend removal regression is conducted via the projection QZ . for sta˜ tionary xt . we detrend the data using QZ and then perform a band spectral regression of the form (32) wy z s ˜ = a 1 wx z s ˜ + a2 w xz s + residual for frequencies s in a shrinking band around . = ) chooses frequencies in a shrinking band about . But. c. the narrow band regression (32) now applies for frequencies that are nonzero as well as zero and prior detrending of the data has occurred in the ˜ ˜ ˜ time domain. Then (33) ˜ a = X QZ QZ X −1 X QZ QZ y = =0 =0 e−i may be recovered from In view of the relation (16). wv s . In equation (32) the afﬁx ‘ z’ on the subscript (e.g. s. 3 dft recursion formulae Here we develop some useful formulae for dft’s of the deterministic sequence zt . . an estimate of ˜ a using the linear combination (34) ˜ = ˜ a1 ˜ a10 − ˜ − e−i − 1 a2 − The asymptotic theory for the narrow band estimates ˆ and ˜ is developed in Section 5. the problem with ˆ A is a ﬁnite sample one. of course. The case for s = 0 is well known. Narrow Band Regressions In the nonparametric case we estimate in (5) at a particular frequency . and p. However. but the result t=1 for s = 0 appears to be new. To some extent. of the differences vt = xt . the bias disappears as n → . corbae. This approach is analogous to the augmented spectral regression approach of Phillips and Loretan (1991) for estimating a cointegrating vector. in wx z s ) signiﬁes that the data have been detrended before taking dft’s. ouliaris.1078 d.
band spectral regression Lemma A: (a) For s = 0. (b) For s = 0. where explicit expressions for the discrete Fourier transforms W k s are easily obtained from Lemma A. Wk s 1079 = n t=1 tk = 1 k+1 k+1 B1 nk + B2 nk−1 nk+1 − 1 2 k+1 + k+1 k+1 B3 nk−2 + · · · + Bk−1 n 3 k where Bj are the Bernoulli numbers. Thus. the standardized quantities W k recursion (37) Wk s = 1 √ n n t=1 t k i e n st satisfy the k 1 k 1 ei s 1 =√ i + i n e s − 1 e s − 1 j=1 nj j −1 j W k−j s and then the dft of dt is simply wd s 1 n =√ dt e i n t=1 st = W0 s Wp s The main cases of interest are low order polynomials. when k = 0 1 2 we have √ (38) W0 s (39) W1 s (40) W2 s 0 n+1 s=0 1/2 2n = ei s 1 s=0 n1/2 ei s − 1 n + 1 2n + 1 6n3/2 = ei s ei s 2 1 − 3/2 i 1/2 ei s − 1 n n e s −1 = n s=0 s=0 s=0 2 s=0 . Wk s satisﬁes the recursion (35) Wk s = nk ei s ei s − 1 + k k 1 ei s − 1 j=1 j −1 j Wk−j s with initialization (36) W0 s = n t=1 ei st = n 0 s=0 s=0 s Using (35).
corbae. It therefore follows from A 1 Lemma A and (37) that W k s = O n− 2 uniformly for s ∈ c . The transformed model is Wy = WZ −W = WZ c 1− c 2 A +W Ac c c Z 2 A− Ac + WX + WX A X A− 2 A +W 2 A− Ac A 1− + A WZ Ac − A WX A− +W A The resulting band spectral estimator for the band sion on (43) (44) AWy = AWZ 1− 2 A is equivalent to a regres + AWX −1 A + AW since AAc = 0. c. thus. Hence. Then. Frequencies in the band c satisfy s > 0 > 0. Hence. phillips In case (38). c Z = 0 and so c QZ c = − Pz c = 0. It follows that X QZ QZ c X = 0 in (30) and therefore ˆ A is unbiased in this case of simple data demeaning. where V = Z c Z . The same applies for higher order trends. A (42) wd c s = O n−1/2 for all s ∈ c A and. and p. we simply apply the discrete Fourier transform operator W to (24) and then perform the band spectrum regression. In the frequency domain. one simply detrends using the projection operator QV = I − PV . b. the alternative is to leave the detrending until the regression is performed in the frequency domain. when zt = 1 t and dt = 1 t/n . it is apparent that eliminating the zero frequency will demean the data and leave the model unchanged for s = 0. and therefore this estimator has the form ˆ f = X W ∗ AQAWZ AWX A = A+ X W ∗ AQAWZ AWy −1 X W ∗ AQAWZ AWX X W ∗ AQAWZ AW .1080 d. On the other hand. so that ˆ A is biased in this case. as is apparent from the form of (26). c Z = 0 and QZ = QV . In the time domain. 4 frequency domain detrending We consider ﬁrst the ﬁxed band regression model (22) and (23). and QZ = QV . ouliaris. The problem of the omitted variable bias evident in (28) can be dealt with either in the frequency domain or in the time domain. s. D = W ∗ Ac WD has elements that are of O n−1/2 . it follows from (38) and (39) that √n n + 1 s=0 1 n 2n1/2 i st wd s = √ dt e = (41) n t=1 ei s 0 1 s=0 n1/2 ei s − 1 and the second component of wd s has nonzero elements for s = 0. To do frequency domain detrending.
the natural narrow band approach is to use an augmented regression model that includes the dft of the trend in the regression. and J / n → 0 as n → . Assumption 3: (a) na /n → and nc /n → 1 − 0 1 as n → . respectively. A similar result holds for f the corresponding estimator ˆ Ac of Ac . /J ) that center on frequencies = j/J j = −J + 1 J − 1 . In the nonparametric case. In this frequency domain approach to detrending. Let na = # s ∈ A fundamental frequencies in the bands A and Ac . for some ﬁxed number ∈ For the bias in ˆ A to vanish asymptotically. the regression coefﬁcient ˆ A on the variable AWX in (43) is invariant to whether the regressor AWZ is included in the regression or whether all the data have been previously detrended in the frequency domain by regression on AWZ. The following condition will be useful in the development of the asymptotics and will be taken to hold throughout the remainder of the paper. We start with the parametric case where there are the two discrete bands and nc = # s ∈ Ac be the number of A and Ac . Again. and the estimator is unbiased. This narrow band regression leads to the estimate − ˜ c1 ˜ c10 ˜ − e−i − 1 c2 − =0 =0 similar to (34). It is convenient to subdivide − into subbands j of equal width (say. the deviation that depends on the term X QZ Q Z X −1 X QZ Q Z c X A− Ac .band spectral regression f 1081 Clearly. viz. It follows that n and na can be approximated by n = 2mJ and na = 2mJa . (45) wy s ˜ = c 1 wx s s ˜ + c2 w x s ˜ + c3 wz s + residual for frequencies (46) ˜f = centered on . Additional requirements will be stated as needed. We start by introducing some notation and making the framework for the limit theory more precise.e. Let m = # s ∈ j and suppose that Ja of j these bands lie in A . √ (b) m J → . i. we let m = # s ∈ . E ˆ A X = A .. 5 asymptotic theory We derive a limit distribution theory for the detrended band spectral regression estimates and consider what happens to the bias as n → . we focus on a single frequency and consider a shrinking band of width /J centered on . the so called Frisch–Waugh f (1933) theorem clearly holds. Following (32).
The common limit distribution of ˆ A and ˆ A is given by √ n ˆA − where −1 −1 √ A f n ˆA − A → N 0 Va d (48) Va = A fxx d 2 A fxx f d A fxx d f with an analogous result for ˆ Ac and ˆ Ac . obtained by interchanging the bias in ˆ Ac . For the full . Va can be estimated by replacing the spectra in the above formula with corresponding consistent estimates and averaging over the band A . the bias in ˆ A will disappear when (47) X QZ Q Z X n −1 X QZ Q Z n c X →0 and c p A similar requirement. ouliaris. 5 1 The Stationary Case Here. We have the following result. p. corbae. . therefore. and yt is generated by (22).f. It is therefore irrelevant A whether the main focus of interest is high or low frequency regression. Hannan (1970. stationary. but with a different convergence rate. in (47). holds for ˜ Theorem 1 (Semiparametric Case): If xt and t are zero mean. The form of the asymptotic covariance matrix Va is a band spectral version of the familiar sandwich formula for the robust covariance matrix in least squares regression. the bias in band spectral regression from time domain detrending disappears as n → and there is no difference between the two bands A and c in terms of the limit theory. equation (16)) for band spectral estimates in the context of models with measurement error. A formula related to Va was given in Hannan (1963b. the matrix Va is the well known formula for the band case where A = − asymptotic covariance matrix of the least squares regression estimator in a time series regression (c. In the stationary case. phillips in (28) needs to disappear as n → . band spectral regression with detrending in the time domain or in the frequency domain f is consistent for both A and Ac . and p. We will distinguish the two cases of sta˜ tionary and nonstationary (integrated) xt corresponding to Assumptions 1 and 2 in the following discussion. c. s. b. ˜ and ergodic time series satisfying Assumption 1. A similar result holds in the nonparametric case.1082 d. 426)).
In the band regression model (22). the bias in (28) disappears when (49) X QZ Q Z X n2 −1 X QZ Q Z n2 c X →0 p Note that in (49) the moment matrices are standardized by n2 .band spectral regression 1083 ˜ Theorem 1 (Nonparametric Case): If xt and t are zero mean. ˜ and ergodic time series satisfying Assumption 1. the same result continues to hold over the band A . the distinction between low frequency regression and regression at other frequencies becomes important. the regression equation is a conventional cointegrating relation. To make the derivations simpler we conﬁne our attention here to the case of a linear trend. Over A . there are only two bands A and c . the A estimator is known to be nconsistent when A = Ac (see Phillips (1991a)) since in that case. The remainder of the limit theory then follows in a fairly straightforward way from these lemmas. In this case. which includes the zero frequency. Over frequency bands like c that exclude the zero frequency the rate of convergence A of the moment matrices is slower and the bias in ˆ Ac will disappear when (50) X QZ n c QZ X −1 X QZ c QZ X n →0 p Our starting point in the I 1 case is to provide some limit theory for discrete Fourier transforms of I 1 and detrended I 1 processes. ˜ Lemma B: Let xt be an I 1 process satisfying Assumption 2. nonparametric band spectral regression with detrending in the time domain or in the = b − . as shown in Theorem 2 below. The following two lemmas do this and provide a limit theory for periodogram averages of such processes. The common limit distribution of frequency domain is consistent for ˆ and ˆ f is given by √ m ˆ − where V =f fxx −1 √ m ˆf − → Nc 0 V d 5 2 The Nonstationary Case Here. and yt is generated by (3). stationary. Then. the discrete ˜ Fourier transform of xt for s = 0 is given by (51) wx ˜ s = 1 w 1 − ei s v s − ei s 1 − ei s ˜ ˜ x n − x0 1/2 n . When A = Ac . because the data nonstationarity is manifest in bands like A that include the zero frequency.
s. Indeed. corbae.e. even in the limit as n → . which produces a common leakage into all frequencies s = 0.1084 d.. it is apparent from (41) that we can write (51) as (52) wx ˜ = 1 w 1 − ei s v t +wn s s s ˜ ˜ x n − x0 from which it is clear that frequency domain detrending (i. they are ˜ frequencywise dependent by virtue of the component n−1/2 xn . b. ˜ ˜ Deﬁne x d t = xt − dt D D −1 D X and let wx d be the discrete Fourier trans˜ form of x d t . c. this leakage is strong enough to ensure that smoothed peri= 1 − ei −2 fvv are inconsisodogram estimates of the spectral density fxx c tent at frequencies ∈ A . using residuals from regressions of the frequency domain data on wt/n s or wd s ) will remove the second term of (51) and thereby eliminate the common leakage from the low frequency. phillips Equation (51) shows that the discrete Fourier transforms of an I 1 process are not asymptotically independent across fundamental frequencies. Lemma C(f) shows that the leakage is still manifest when the data are ﬁrst detrended in the time domain. ouliaris. As the next lemma shows. as n → : (a) n−2 (b) (c) (d) (e) (f) (g) (h) (i) (j) (k) (l) (m) (n) (o) 1 Bx Bx 0 1 −3/2 n w x s w d s → 0 Bx u ˜ s∈ A d 1 n−2 w x d s w x d s ∗ → 0 Bx u B x u s∈ A ∗ d i ∗ c wx n−1/2 → − 21 Bx 1 / 1 − ei d c e f1 ˜ s wd s s∈ A A ∗ d c wx n−1 fxx + 21 1/ 1−ei 2 Bx 1 Bx 1 d c ˜ ˜ s wx s → s∈ A A ∗ d c wx d n−1 → c fxx + 2 −1 g Bx g Bx ∗ d s wx d s s∈ A A ∗ d ∗ c wx d n−1/2 → − 21 B x f1 d c g s wd s s∈ A A 1 ∗ c wd wd s ∗ → 2 f1 f1 d c s s∈ A A 1 −1 ∗ → 0 uu n s wd s ∈ A wd s d 1 n−2 w x s w x s ∗ → 0 Bx B x ˜ ˜ s∈ 0 d 1 n−2 w x d s w x d s ∗ → 0 Bx u B x u s∈ 0 1 1 wd s wd s ∗ → 0 uu n s∈ 0 d 1 n−3/2 wx s w d s ∗ → 0 Bx u ˜ s∈ 0 ∗ wd s w d s ∗ → f 1 f1 =0 n/m s∈ √ −1 d ∗ i i ∗ nm wx s w d s → e / 1 − e B 1 f1 =0 ˜ s∈ s∈ A wx ˜ s wx ˜ s ∗ d → ∗ d . On the other hand. and p. Then. ˜ Lemma C: Let xt be an I 1 process satisfying Assumption 2 and dt = 1 t/n .
The bias is random and is linear in the differential. the I 1 band regression model can be analyzed. the limit function f1 = ei / ei − 1 and then the bias depends on g Bx = ei 1 − ei Bx 1 − 1 0 Bx r 1 0 r2 −1 . Theorem 2: Suppose Then (53) n ˆA − A t ˜ ˜ xt satisﬁes Assumption 2 and yt is generated by (22). With this result in hand. when the regressors are I 1 . −1 1 0 → d 1 0 Bx u Bx u 1 0 Bx u dB −1 = MN 0 and (54) where = and (55) = g B x f1 c A Bx u Bx u 2 f 0 d ˆ Ac → Ac − −1 Ac − A 2 fxx +g Bx g Bx ∗ d ∗ c A d 1 0 −1 uu 1 0 uBx So. and fvv . f1 1 0 = 0 ei / ei − 1 1 0 −1 Bx u r = Bx r − and g Bx = Bx u uu u r ei B 1 + 1 − ei x 1 0 Bx u 1 0 −1 uu f1 Joint convergence applies in (a)–(o). band spectral regression is inconsistent in frequency bands that exclude the origin when detrending is performed in the time domain prior to frequency domain regression. The limit theory for time domain detrended band spectral regression is as follows. Ac − A .band spectral regression where fxx = 1 − ei −2 1085 for = 0. between the coefﬁcients in the frequency bands. In the case of a linear trend zt = t.
The following theorem gives the corresponding results for the nonparametric estimators ˆ and ˜ . a circumstance that arises when the ﬁlter L is symmetric (see Remark (ii) following Lemma D in the Appendix) as it is in the case (19). s. b. corbae.e. ouliaris. √ m ˜ − 0 −2 +g fvv Bx ∗ → d Bx u Bx u −1 1 0 Bx u dB →d Nc 0 f fxx Remarks: (i) In Theorem 2 (a) it is apparent that the limit distribution of ˆ 0 has a second order bias term involving 1 0 Bx u dBx + x ˜ 1 which is nonzero except when ˜ 1 = 0. (57) where (58) where 0 n ˆ0 − = 0 → x d 1 0 −1 Bx u Bx u j=− −1 1 0 Bx u dB − = 0. the bias is not dependent on the width of the band Ac . n ˜0 − For = 0. (a) For = 0. Part (a) shows also that the estimator ˆ is inconsistent at all frequencies = 0 except when ˜ e−i = 0. 1 0 Bx u dBx + x ˜ 1 1 and = E v0 vj . ˜ ˜ Theorem 2 : Suppose t xt satisﬁes Assumption 2 and yt is generated by (11) where L and ˜ L both have valid BN decompositions. and p. In that case. Bx g . the limit (57) reduces to the same mixed normal distribution as (53).1086 d. 1 0 = 2 fxx = 1 − ei and fxx (b) For = 0.. a simple calculation reveals that the probability limit (54) of ˆ Ac simpliﬁes to (56) + A− 1 Bx r 0 Ac 1 2 r 0 −1 − Bx 1 1 Bx r 0 1 2 r 0 1 2 r 0 −1 1 Bx r 0 Ac 1 + Bx 1 − −1 2 So. c. phillips When fxx = 1/2 1 − ei −2 (i. in this case. 1)). The bias in ˆ 0 and inconsistency in ˆ are due to omitted variable misspeciﬁcation in the regression. when vt is iid(0. For − 1 ˜ e−i ˆ →d = e−i + 2 =b − fxx e−i − 1 . .
g. say fˆ . p. as shown in Hannan’s (1963a. Hannan (1970. the limit distribution of the efﬁcient estimate is normal with variance matrix −1 2 f d −1 and therefore attains the usual efﬁciency bound in c fxx A time series regression (e. eqn. p. When = 0 the limit distribution is the same as in the stationary case given in Theorem 1 . and = 0.g. When such weighted regression is performed with frequency domain detrending. the limit distribution is the same as (59) and is optimal in the sense of Phillips (1991b). 427)) adjusted here for band limited regression (Hannan (1963. Hannan (1970. The same mixed normal distribution as (53) applies when = 0. (3. For A .4). 488)). Theorem 3: Under the conditions of Theorem 2 (59) and (60) √ f n ˆ Ac − Ac f n ˆA − A → d 1 0 −1 Bx u Bx u 1 0 Bx u dB →N 0 × d −1 c A fxx fxx d −1 2 c A fxx f d c A d In ﬁxed band regression models there is usually some advantage to be gained by averaging over the band and using weighted regression.band spectral regression 1087 (ii) Part (b) of Theorem 2 shows that the estimator ˜ based on the augmented regression (32) is consistent for both = 0.. Theorem 3 : Under the conditions of Theorem 2 . Efﬁcient regression is based on a weighted band spectral regression that uses a preliminary regression to obtain estimates of the equation errors and a corresponding estimate of the error spectrum. Details are omitted and are available in an earlier version of the present paper. the resulting estimates have optimal properties for both the nonstationary component and the stationary component. For Ac . (16))). at f n ˜0 − 0 =0 → d 1 0 −1 Bx u B x u 1 0 Bx u dB and for (61) =0 √ m ˜f − → Nc 0 f d fxx −1 . eqn. that is uniformly consistent so that sup fˆ −f →p 0 (e. b) original work.. The following results give the limit theory for the frequency domain detrended estimators in the ﬁxed band and narrow band cases for nonstationary data.
it seems natural to eliminate deterministic trends in band spectral regressions by detrending in the time domain prior to the use of spectral methods because these methods were originally intended for application to stationary time series. including detrending regressions. An alternate approach that is suitable in practice is to model the data and run regressions. As far as the model (11) is concerned. b. phillips f f So. 244) theorem shows that such regression is asymptotically efﬁcient when the time series are trend stationary (although this conclusion does not hold when there are stochastic as well as deterministic trends—see Phillips and Lee (1996)). it is therefore apparent from Theorems 2 and 3 that if the correct augmented regression model is used in estimation. f f The frequency domain detrended estimators ˆ Ac and ˆ are also consistent. In effect. p. can be substantial and has been conﬁrmed in simulations that are not reported here (Corbae. ouliaris. s. The bias and inconsistency arise from omitted variable misspeciﬁcation and are managed by use of an appropriate augmented regression model. The situation is analogous to a structural break model. ˆ A and ˜ 0 are consistent and have the same mixed normal limit distribution as that of ˆ A in (53). The inconsistency. ˆ unlike the time detrended estimator = 0 in the nonstationary case. and Phillips (1997)). it does not matter asymptotically whether detrending is done in the time domain or the frequency domain. which arises from omitted variable effects. time domain detrending can lead to inconsistent estimates of the coefﬁcients at frequency bands away from the origin. An estimate of the frequency response coefﬁcient at a particular frequency is then recovered from a linear combination of . are taken and band regression is performed. this paper shows that such time domain detrending will lead to biased coefﬁcient estimates in models where the coefﬁcients are frequency dependent. The nonparametric estimator ˜ is √ m consistent and is asymptotically equivalent to the augmented regression estimator ˜ . However. In a similar way.1088 d. it turns out to be particularly important to specify the model in terms of levels and differences as in (15) leading to the ﬁtted regressions (32) and (45). and p. but here the coefﬁcients change across frequency rather than over time. in the frequency domain. the same principle is employed but one uses a shrinking band that is local to a particular frequency. In models that have both deterministic and stochastic trends. c. 6 conclusion It is natural to eliminate deterministic trends in the time domain by simple least squares regression because the Grenander–Rosenblatt (1957. In the nonstationary case. When nonparametric estimation is being conducted. The limit distribution makes asymptotic inference about A and 0 straightforward. including the deterministic trends. using conventional regression Wald tests adjusted in the usual fashion so that a consistent estimate of the spectrum of t is used (based on regression residuals) in place of a variance estimate. Ouliaris. corbae. The f limit distribution of ˆ Ac is the same as it is in the case of trend stationary regres√ f sors (Theorem 1) and has the same n rate. discrete Fourier transforms of all the variables in the model.
To see this.phillips@yale. Box 208281.yale. National University of Singapore.. souliaris@nus. and Research Dept.. U. University of Texas at Austin. with ck j k=− j=− j 1 2 cj < j ≥0 ck j <0 where Cw L = ˜ cwj Lj with ˜ cj = k=j+1 − (b) C L = C eiw + Cw e−iw L e−iw L − 1 c eikw k=j+1 k − j k=− j=− j ≥0 j <0 ˜ cwj = ck eikw Proof of Lemma D: The proof is along the same lines as the proof in Phillips and Solo (1992.S. ﬁnal revision received June. the series L = j=− j Lj . Singapore 117591.A.econ. 2001.S.edu. Partial summation gives j=1 Pn = n j=1 eix 1 eijx = Sn + ix j n e −1 n j=2 ei j−1 x − 1 j j −1 . This approach. Washington DC. 1997. Lemma 2. Remarks: (i) The condition (62) j=− j 1 2 cj < ˜ is sufﬁcient for (a) and (b) and ensures that j=− cj2 < but it is not necessary for the latter.edu. http://www. provides a convenient single equation method of estimating a long run relationship in the presence of deterministic trends and short run dynamics..S. fails (62) but still has a valid BN decomposition with j=− ˜ 2 < .sg and Cowles Foundation for Research in Economics.A. peter.edu Manuscript received September. of Economics.edu. For instance. Deﬁne Sn = n eijx = eix / eix − 1 einx − 1 . which can be regarded as a frequency domain version of leads and lags dynamic regression. let cj = eijx /j and then j is a constant j times the imaginary part of cj . TX 78712.band spectral regression 1089 the coefﬁcients in the regression as in (34) and (46). IMF. Dept.. http://Korora.edu/∼corbae School of Business.A. U.utexas. corbae@eco. whose coefﬁcients j are given by (19). and University of Auckland and University of York. where C L = j=− cj Lj .1) of the one sided BN decomposition.eco.utexas. Yale University. CT 065208281. Austin. U. APPENDIX Lemma D (Two Sided BN Decompositions): If C L = j=− cj Lj and then: ˜ ˜ (a) C L = C 1 + C L 1 − L . New Haven.
Hence. c. corbae. We deduce that − cm 2 < .g. we deduce that (63) where C L = C 1 +C L 1 − L 2 C L = j=− ˜ ˜ c j Lj cj = ≈ ˜ ck j k=− j ≥0 ˜ ck j <0 j 1 2 k=j+1 − A sufﬁcient condition for the validity of (63) is of the original coefﬁcients. −c−m = j=m c−j = j=m cj = ˜ ˜ ˜ ˜ ˜ ˜ cm + cm . If C L has a valid BN decomposition.121)). Gradshteyn and Ryzhik (1965. ouliaris. formula 0. C 1 = j=1 cj + c−j + c0 = j=1 cj − cj − cj + j=1 cj = 0. j=− ˜ cj < or j=− j 3 2 cj < in terms Proof of Lemma A: Part (a) is well known (e. b.. which ˜ includes the series with coefﬁcients j given by (19).1090 It follows that d. s. For part (b) we use partial summation to give n t=1 t k ei st = = n t=1 n t=1 t k ei − k j=1 st + n t=1 t −1 j k ei ei st st k k−j −1 t j + n t=1 t − 1 k ei s t−1 ei s − 1 or nk = n t=1 − k j=1 k k−j −1 t j j ei st + n t=1 t − 1 k ei s t−1 ei s − 1 Using this formula in the identity n t=1 t k ei st e i s − 1 = nk e i s − 1 + n t=1 t − 1 k ei s t−1 ei s − 1 we get n t=1 t k ei st e i s − 1 = nk e i s − 1 + n k − = nk e i s + k j=1 n t=1 − j n t=1 k j=1 k k−j −1 t j st j ei st k j −1 t k−j ei . Then. and p. phillips Pn − P m = and letting n → ˜ cm = n eix ei j−1 x − 1 Sn S m − + ix n m e − 1 j=m+1 j j − 1 we have S eix 1 eijx ei j−1 x − 1 = − m + ix =O j m e − 1 j=m+1 j j − 1 m j=m+1 ˜ ˜ with a similar result for c−m . In this case. L = j=− j Lj with j given by (19) has a valid BN decomposition with j=− ˜ 2 < j (ii) The case of a symmetric ﬁlter with cj = c−j for j = 0 is an important specialization.
Ouliaris. Proof of Theorem 1 : First. For higher order trends with k = 1 2 3 we get n n+1 s=0 2 W1 s = i s n e s=0 ei s − 1 n n + 1 2n + 1 s=0 6 W2 s = i 2 n e s n− i s=0 ei s − 1 e s −1 and 2 n n+1 2 = i s 1 1 n e 2 +6 ei s − 1 n − 3 n − 1 i s e −1 ei s − 1 s s=0 2 W3 s s=0 which lead to the expressions for W k k = 0 1 2. We now ﬁnd that (67) m−1 X QZ QZ X = m−1 s∈ wx ˜ s wx ˜ s ∗ + op 1 →p 2 fxx and using (66) we have (68) m−1 X QZ ˜ QZ y = m−1 X =m −1 s∈ ˜ y + op 1 wx ˜ s wx ˜ s ∗ b − + m−1 s∈ wx ˜ s w s ∗ + op 1 ..band spectral regression giving the recursion Wk s 1091 = nk ei e i s s −1 + 1 e −1 i s s k j=1 k j −1 j Wk−j s which holds for s = 1 The initial condition W0 s n (i. and Phillips (1997)). ei n 0 = 1). that are given in Section 2 following Lemma A. s=0 s=0 = n t=1 ei st = follows by elementary calculation. from (8) and (9) we have (65) (66) wy ˜ s =b =b s wx ˜ wx ˜ s +w +w s 1 ˜ + √ e−i s a n 1 + Op √ n s0 ˜ −a sn s s in the stationary case. Proof of Theorem 1: This follows standard lines and is omitted (see Corbae.e. observe that (64) ˆ = m−1 X QZ QZ X −1 m−1 X QZ ˜ QZ y Next.
A similar argument gives the result for ˆ f .2) showed that an asymptotically inﬁnite collection of w s satisfy a central limit theorem and are asymptotically independent for frequencies s in the neighborhood of 1 1 the origin provided the number of frequencies m = o n− 2 + p where E t p < for some p > 2. and wd s 1 2 = = 0 1 ei s n1/2 ei s −1 1 = √ f1 n s s=0 It follows that as n → (72) wd s = 0 √ n 1+o 1 2 si 2 Hannan (1973. (67). from (64). giving sn = n−1/2 n t=1 ˜ xt−1 ei st + wv s = ei s − n−1/2 ei ˜ ˜ xn − x 0 + w v s Then. Proof of Lemma B: Take dft’s of the equation wx ˜ S ˜ xt = wx ˜ s t . and (68) we have the expression (69) √ −1 m ˆ − = m−1 s∈ wx ˜ s wx ˜ s ∗ m−1/2 s∈ wx ˜ s w s ∗ + op 1 The family w s are known to satisfy a central limit theorem (with limit Nc 0 2 f ) s ∈ for dft’s of stationary processes and are independently distributed2 as n → . It follows from this result and (67) that (70) m−1/2 s∈ wx ˜ s w s ∗ →d Nc 0 2 2 f fxx which leads to the stated limit theorem for ˆ . √ n n + 1 /2n1/2 . phillips Then. c. s. n−2 s∈ A wx ˜ s wx ˜ s ∗ = n−2 s∈ − wx ˜ n t=1 s wx ˜ s ∗ − n−2 c s∈ A wx ˜ s wx ˜ s ∗ = n−2 → d 1 0 ˜ ˜ xt xt + Op n−1 Bx Bx where the error magnitude in the second line follows directly from part (e) below.e. Here the collection s ∈ Phillips (2000. This result can be extended to frequency bands away from the origin. b. although a proof was not given in that paper. . Theorem 3. corbae.1092 d. i. Theorem 3) showed that a ﬁnite collection of w s satisfy a central limit theorem has m members and is asymptotically inﬁnite. so that n−1/2 wd 0 → 1 Part (b) First note from (41) that wd 0 = 1 f0 = 0 u . transposing and solving yields the stated formula (71) wx ˜ s = 1 w 1 − ei s v s − ei s 1 − ei s ˜ ˜ xn − x0 n1/2 Proof of Lemma C: Part (a) This result may be proved as in Phillips (1991a). ouliaris. A new and substantially simpler proof uses part (e) and is as follows. provided the number of frequencies does not go to inﬁnity too fast.. and are independent. and p.
we have n−1/2 xn → Bx 1 . On the other hand. Part (d) From Lemma B we have wx ˜ s = 1 w 1 − ei s v s − ei s 1 − ei s d ˜ ˜ xn − x 0 n1/2 s ˜ As in the proof of part (b). observe that ˜ n−1/2 x d Then 1 n2 wx d s nr → Bx r − u r d 1 0 −1 uu 1 0 uBx = Bx u r s∈ A wx d s ∗ → d 1 0 Bx u Bx u as in part (a). n−3/2 s∈ − wx ˜ s wd s ∗ = n−2 n t=1 ˜ xt s∈ − ei st wd s ∗ = n−3/2 → It follows that n−3/2 s∈ A d 1 0 n t=1 ˜ xt dt = n−1 n t=1 ˜ x √t dt n Bx r u r dr wx ˜ s wd s ∗ → d 1 0 Bx r u r dr Part (c) First. when 1 +o √ n − s s ∈ Ac ∗ 1 1 ei =√ 0 i +o √ e −1 n n A 1 = √ f1 n = s∈ − Write the summation over n−3/2 s∈ A as follows: wd s ∗ s∈ A . Then wx ˜ s s wx ˜ s = n−3/2 s∈ − wx ˜ wx ˜ s∈ − wd wd s − n−3/2 s ∈ Ac wd s s ∗ = n−3/2 First.band spectral regression a formula that holds for both s ﬁxed and for s → = 0 as n → we have s → (73) wd s 1093 with s/n → 0. n−2 s ∈ Ac s s ∗ − n−2 s ∈ Ac wx ˜ f1 ∗ wx ˜ s f1 s ∗ = n−2 s ∈ Ac 1 w 1 − ei s v s − ei s 1 − ei s ˜ ˜ xn − x0 n1/2 f1 s ∗ = Op n−1 Second. and. → 1 e N 0 2 fvv − N 0 2 fvv 0 1 − ei c 1 − ei f + fvv 0 = Nc 0 2 vv 1 − ei 2 d i . if wx ˜ s → = 0 as n → .
ouliaris. and p. corbae. 2 fˆvv →p 0. For term II. fvv Part (f) Note that (75) wx d s ∗ w s w s ∗ and fˆvv − = wx ˜ = → = d s ∗ − wd s s ∗ ∗ n−1 D D e 1 − e−i − −i s s −1 n−1 D X s ∗ 1 1 − e−i wv s − ˜ ˜ xn − x0 − n1/2 wd n1/2 n−1 D D 1 0 −1 −1 n−3/2 D X 1 0 1 Nc 0 2 f 1 − e−i 1 Nc 0 2 f 1 − e−i . Since the w s are asymptotically independent Nc 0 2 fvv ˜ x0 = Op 1 . s. In the penultimate line above. c.1094 It follows that n−1/2 d. b. term I → 0. Part (e) From Lemma B we get n−1 wx ˜ s c s∈ A wx ˜ 1 m s ∗ d ∼ n−1 c s∈ A 1 1 − ei wv s s 2 wv ∗ s wv s ∗ + ˜ ˜ xn xn n1/2 n1/2 = 1 2J Ja ≤ j ≤J s∈ j 1 1 − ei s 2 wv s + ˜ ˜ xn xn + op 1 n1/2 n1/2 s 2 1 = 2J → d 1 Ja ≤ j ≤J 1 − ei 2 j 2 2 fˆvv 1 2 j + n−1 c s∈ A 1 1 − ei ˜ ˜ xn xn + op 1 n1/2 n1/2 c A fvv 1 − ei d + c A 1 1 − ei j 2 d Bx 1 Bx 1 = 1/m s∈ j as required. we have n−1 ˜ ˜ xn − x 0 n1/2 ei s 1 − ei s s . Here 1 0 e−i B 1 − f1 1 − e−i x Bx ∗ ∗ uu uBx −g say when s → g = 0 as n → Bx = ei B 1 + 1 − ei x Bx u 1 0 −1 uu f1 . since ei s 1 − ei ∗ p c s∈ A f1 s ∗ ∼ → d ˜ 1 xn 2 2 n1/2 n 1 2 Bx 1 c s∈ A s f1 s ∗ c A e i f1 1 − ei d so that (74) term II → − d 1 2 Bx 1 c A e i f1 1 − ei ∗ d giving part (d). phillips c s∈ A wx ˜ s wd s ∗ = n−1 w c s∈ A wx ˜ ∗ s f1 s ∗ = n−1 c s∈ A 1 1 − ei s s f1 s − n−1 ˜ ˜ xn − x 0 n1/2 c s∈ A ei s 1 − ei s f1 s ∗ = term I + term II say.
as n → . we get n−1 s∈ A wd s wd s ∗ = n−1 s∈ − wd n t=1 s wd st s ∗ + Op n−1 ∗ = n−1 = n−1 → giving the stated result. then j → ∗ = 1 m s∈ j wx d +g s wx d Bx g → Bx d 2 f 1 − ei ∗ 2 +g Bx g Bx ∗ = 2 fxx We deduce that n−1 wx d s c s∈ A wx d s ∗ = = → d 1 2J 1 2J c A Ja ≤ j ≤J 1 m s∈ j wx d s wx d s ∗ + op 1 2 fˆxx d Ja ≤ j ≤J j + op 1 g Bx g Bx ∗ fxx + 2 −1 d which gives part (f).band spectral regression and the complex normal variate Nc 0 2 f If j = j/J is the midpoint of the band 2 fˆxx d j 1095 j and s in (75) is independent of the Brownian motion Bx . Part (g) Observe that n−1/2 wx d s c s∈ A wd s ∗ = n−1 d c s∈ A wx d g s f1 s ∗ + op 1 ∗ ∼ −n−1 ∼− →− as stated. 1 0 dt s∈ − ei wd s + Op n−1 n t=1 dt dt + Op n−1 u r u r dr . Part (i) Using part (h). Part (h) From (73) we have wd s d d c s∈ A s Bx f1 g s 1 2 1 2 2 n g c s∈ A s Bx f1 ∗ s ∗ c A B x f1 d c s∈ A wd s ∗ ∼ 1 n c s∈ A f1 s f1 s ∗ → 1 2 c A f1 f1 ∗ d as required.
for = 0. corbae.1096 − d. s. √ n/2 si 1 + o 1 from (72). Thus. ouliaris. c. for s = 0 with s/n → 0. we ﬁnd that n−2 s∈ 0 ∈ wx ˜ s wx ˜ s ∗ = n−2 s∈ − − s∈ − − 0 s ∗ wx ˜ s s wx ˜ s ∗ + op 1 = n−2 s∈ − wx ˜ wx ˜ + Op n≥ s ≥J 1 wv s2 s wv d 1 0 s ∗ + ˜ ˜ xn xn n1/2 n1/2 = n−2 n t=1 ˜ ˜ xt xt + o p 1 → Bx Bx Part (k) In the same way as parts (j) and (c) we ﬁnd that n−2 s∈ 0 wx d s wx d s ∗ = n−2 n t=1 ˜ ˜ x d tx d t + op 1 → 1 d 1 0 Bx u Bx u 1 Part (l) From part (b) we have n−1/2 wd 0 → f0 = 1 2 = 0 u and. b. we ﬁnd that we have wd s = 0 1 n s∈ 0 wd s wd s ∗ = 1 1 w 0 wd 0 + n d n wd s ∈ 0− 0 s wd s ∗ 0 1 = wd 0 w d 0 + n 0 → f0 f0 + 0 0 1 12 0 2 1 m s=1 1 2 2 1 s2 0 = 0 uu Part (m) As in part (b) we ﬁnd that n−3/2 s∈ 0 wx ˜ s wd s ∗ = n−3/2 s∈ − wx ˜ n t=1 s wd s ∗ − n−3/2 s∈ 0 wx ˜ s wd s ∗ = n−3/2 → since n−3/2 s 0 d 1 0 ˜ xt dt − o p 1 Bx r u r dr wx ˜ s wd s ∗ = Op 1 nm 1 nm wx ˜ s 0 s ˜ ˜ xn −x0 √ n = Op = op 1 Part (n) For n m wv s 0 s − ei s s 1 − ei = 0. and proceeding as in part (a). and p. using (73) we obtain wd s s∈ wd s ∗ ∼ 1 m s∈ f1 s f1 s ∗ → f1 f1 ∗ . phillips when s Part (j) Using the representation of wx s in Lemma B. the fact that s > J → ˜ − 0 .
we have (76) ˜ n−1/2 x d → Bx r − u r d 1 0 −1 nr uu 1 0 uBx = Bx u r s say ˜ Write the discrete Fourier transform of x d t as wx d (77) 1 and then from Lemma C(c) we have ∗ n−2 X QZ QZ X = 1 n2 s∈ A wx d s wx d s → d 1 0 Bx u Bx u Since 0 Bx u Bx u > 0 (see Phillips and Hansen (1990)). Assumptions 1 and 2 ensure that n−1/2 nr t=1 t wv → B r d with = Nc 0 2 fvv and where is independent of B r for all = 0. Since xt is an I 1 process and satisﬁes an invariance principle when standardized by n−1/2 . The required quantities in (a)–(k) are functionals wv s wv s ∗ for j → of these elements and smoothed periodogram ordinates (like 1/m s∈ j as in the case of part (f)) that converge in probability to constants. Write (79) X PD W ∗ AWPD W ∗ Ac W X = n X PD W ∗ AWD n3/2 DD n −1 D W ∗ Ac W X n1/2 . joint weak convergence in (a)–(o) applies because the component elements jointly converge and one may apply the continuous mapping theorem in a routine fashion. Proof of Theorem 2: Note that ˆA − A = − X QZ Q Z X −1 X QZ Q Z c X A − Ac − ˜ ˜ We consider the limiting behavior of n−2 X QZ QZ X and n−1 X QZ QZ c X.band spectral regression Part (o) For √ n m = 0. Next. Factor term A as follows and consider each factor separately. decompose n−1 X QZ QZ c X as follows: (78) X QZ Q Z n c X =− = X QZ W ∗ AWPZ W ∗ Ac W X X QD W ∗ AWPD W ∗ Ac W X − n n X PD W ∗ AWPD W ∗ Ac W X X W ∗ AWPD W ∗ Ac W X − n n = term A − term B Take each of these terms in turn. Deﬁne x d t = xt − ˜ dt D D −1 D X . n−2 X QZ QZ X has a positive deﬁnite limit as n → . In particular. using (73) we obtain wx ˜ s 1097 s∈ wd s ∗ = 1 m 1 m s∈ wx ˜ wv s f1 s s ∗ +o 1 ˜ ˜ xn −x0 √ n = → d − ei s s s∈ 1 − ei ∗ f1 s ∗ + op 1 ei B 1 f1 1 − ei Finally.
b. for example. we have (85) n ˆA − A = X QZ Q Z X n2 − −1 X QZ Q Z n −1 X QZ Q Z X n2 X QZ Q Z n c X A − Ac From Lemma C(c) and (77) above. Using Lemma C. corbae.1098 The ﬁrst factor is (80) d. . we have (86) 3 n−2 X QZ QZ X → d 1 0 Bx u Bx u = G ˜ By changing the probability space (on which the random sequence xt is deﬁned). The limit of term A now follows by combining (81) and (80) and using joint weak convergence: (82) X PD W ∗ AWPD W ∗ Ac W X n−1/2 X PD W ∗ AWD = n n → d 1 0 DD n − 1 2 −1 D W ∗ Ac W n−1/2 X e−i f1 d 1 − e−i Bx 1 Bx u 1 0 −1 uu c A Next consider term B of (78). we obtain (83) X W ∗ AWPD W ∗ Ac W X n = d 1 −1/2 n n 1 0 s∈ A wx ˜ 1 0 s wd −1 s ∗ DD n −1 n−1/2 c s∈ A wd s wx ˜ s ∗ → Bx u uu − 1 2 c A e−i f1 d Bx 1 1 − e−i Combining (82) and (83) in (78) we ﬁnd3 (84) n−1 X QZ QZ c X = op 1 As for the limit distribution of ˆ A . c. The third factor is the conjugate transpose of (81) n−1/2 X W ∗ Ac WD = n−1/2 →− d c s∈ A wx ˜ s wd s ∗ 1 2 Bx 1 c A e i f1 1 − ei ∗ d from Lemma C(d). Then. in the original space the difference of the terms tends in distribution to zero giving the stated result. phillips n−1/2 X D n−1/2 X PD W ∗ AWD = n n → d 1 0 DD n 1 0 −1 D W ∗ AWD n 1 0 −1 Bx u uu uu = 1 0 Bx u in view of (76) and Lemma C(i). and p. we can ensure that both terms tend almost surely to the same random variable (see. s. page 47 of Shorack and Wellner (1986)). ouliaris. Theorem 4.
and (92). from Lemma C(b).band spectral regression and from (84) X QZ Q Z X n2 −1 1099 (87) Next. Combining (88). (92) X QZ Q Z n → d 1 0 −1 Bx u Bx u 0 1 0 Bx u dB −1 ≡ MN 0 2 f Using (85). we deduce that X QZ Q Z d → n 1 0 (91) Bx u dB 1 0 The limit distribution (91) is a mixture normal distribution with mixing matrix variate It now follows from (86) and (91) that X QZ Q Z X n2 −1 Bx u Bx u . we obtain 1 n 1 0 3 2 s∈ A (90) wx d s wd s ∗ DD n −1 D √ n → d 1 0 Bx u u 1 0 uu 1 0 u dB =0 since Bx u u = 0. (i) and Assumption 2. using (76) and proceeding as in the proof of Lemma C(a) above. X QZ Q Z n c X →0 p (88) X QZ Q Z = n−1 n − s∈ A wx d s w d s ∗ = ∗ 1 n s∈ A wx d −1 s w s ∗ 1 n 3 2 s∈ A wx d s wd s DD n D √ n Then. and (90). we ﬁnd (89) n−1 s∈ A wx d s w s ∗ = n−1 s∈ − wx d n t=1 s w d s ∗ − n−1 s ∈ Ac wx d s w s ∗ = n−1 ˜ xd t t − op 1 → 1 0 Bx u dB Further. (87). we deduce that n ˆA − d 1 0 −1 1 0 1 0 Bx u Bx u A → Bx u Bx u Bx u dB which gives the stated result. . (89).
phillips For the limit of ˆ Ac . c. From Lemma C(f) we have + 2 −1 n−1 c s∈ A wx d wx d s → d c A fxx g Bx g Bx ∗ d which is a positive deﬁnite limit. n−1 X QZ c QZ X c = −n−1 X QZ = − n−1/2 PZ X = −n−1 X QD wx d s c PD X −1 c s∈ A wd s ∗ n−1 D D n−3/2 s∈ A wd s wx ˜ s ∗ From Lemma C(g) and (b). X QZ n where wx d (93) s ∗ c QZ X = X QD n s ∗ c QD X = n−1 −1 c s∈ A wx d s wx d s ∗ = wx ˜ s ∗ − wd s n−1 D D ∗ n−1 D X . b.1100 d. Proof of Theorem 2 : Part (a) From (64) and (11) we have (95) ˆ = 1 + X QZ QZ X −1 X QZ QZ ˜ Since ˜ t is a strictly stationary and ergodic sequence with mean zero and satisﬁes a central limit theorem. we need to examine the asymptotic behavior of the bias term in (29). s. we obtain (94) n−1 X QZ c QZ X → − − d 1 2 c A g B x f1 ∗ d 1 0 −1 uu 1 0 uBx It follows from (93) and (94) and joint weak convergence that the asymptotic bias term for ˆ Ac involves n−1 X QZ → d c QZ X 2 fxx −1 n−1 X QZ +g B x f1 c QZ X Bx 1 0 ∗ −1 c A Bx g ∗ d −1 1 0 × c A g d uu uBx establishing the stated result. corbae. which depends on the matrix quotient n−1 X QZ c QZ X −1 n−1 X QZ c QZ X . Next. and so (96) ˜ t − zt n−1 Z Z −1 n−1 Z ˜ = ˜ t − zt −1 n n−1 D D −1 n−1 D ˜ → ˜ t p Then. ouliaris. (97) n−1 X QZ QZ ˜ = n−1 = n−1 s∈ s∈ wx d ˜ wx d ˜ s w˜ w s ∗ + op 1 − n−1 s∈ s s ∗ wx d ˜ s wv s ∗ ˜ e−i s + op 1 . First. since ˜ L has a valid BN decomposition. and p. Take each of these factors in turn. n−1 D D −1 n−1 D ˜ = Op n−1/2 .
We deduce that QZ ˜ → d 1 0 n−1 X QZ Bx u dB − 1 0 Bx u dBx + x ˜ 1 Next. from (15) we have (103) Now (104) wy d s ∗ wy ˜ s = 1 wx ˜ s − ˜ ei s wv s +w s 1 + Op √ n = wy ˜ s ∗ − wd s ∗ n−1 D D −1 ˜ n−1 D y and. First we have wx d s QZ X = m−1 s∈ wx d s ∗ and in a similar fashion to (93) we ﬁnd that (102) m−1 s∈ wx d s wx d s ∗ → 2 fxx d +g Bx g Bx ∗ = say Next. we have (105) ˜ n− 2 y nr →d 1 1 B x r = By r and it follows as in part (f) of Lemma C that (106) wy d s ∗ ∼ wy ˜ s ∗ − f1 s ∗ 0 1 −1 uu 1 0 uBx 1 + op 1 However.band spectral regression When (98) and (99) where (100) = n−1 s∈ 0 x j=− 1101 = 0. in view of the cointegrating relation (11). we obtain n ˆ0 − 1 → d 1 0 −1 Bx u Bx u 1 0 Bx u dB − 1 0 Bx u dBx + x ˜ 1 Now consider the case where m−1 X QZ = 0. as in Lemma C(c) we ﬁnd that (101) n−2 X QZ QZ X → d 1 0 Bx u Bx u Combining (100) and (101).1 of Phillips (1991a) that n−1 s∈ 0 wx d ˜ s w s ∗ → d 1 0 Bx u dB wx d ˜ s wv s ∗ → d 1 0 Bx u dBx + x E v0 vj . from the proof of Lemma C(f) we also have (107) wx d s ∗ = wx s ∗ − f1 s ∗ 0 1 −1 uu 1 0 uBx + op 1 . we ﬁnd as in (89) and Theorem 3.
b. yt and xt are ﬁrst detrended in the time domain and then dft’s are taken of the detrended data and differences of the detrended data. It then follows from (98). (102). the regressors wx d orthogonal after appropriate scaling because (112) m−1/2 n−1 s∈ 0 xd in (111) are asymptotically wx d ˜ s wv s ∗ = Op m−1/2 in view of (99). (109). (108) can be written as (111) wy d s ∗ = wx d s ∗ 1 −w xd s ∗ ˜ e−i s s +w and w s ∗ + op 1 s First. corbae. ouliaris. s. phillips We deduce from (103). and (111) that n ˜0 − as required. 1 → d 1 0 −1 Bx u Bx u 1 0 Bx u dB . and p. Lemma C(k). and (107) that (108) wy d s ∗ ∼ wx ˜ − wv = wx d s ∗ 1 − f1 ∗ ∗ s ∗ 0 1 −1 uu 1 0 uBx 1 +w s ∗ s s ˜ e−i s + op 1 s ∗ 1 − wv ˜ e−i s +w s ∗ + op 1 Then (109) m−1 X QZ ˜ QZ y = m−1 ∼ m−1 s∈ s∈ wx d wx d wx d s wy d wx d s s ∗ s s ∗ 1 ∗ − m−1 s∈ wv s ˜ e−i + op 1 Using (51) we ﬁnd that (110) m−1 s∈ wx d s wv s ∗ = m−1 s∈ wx s wv wv s ∗ + op 1 wv ∗ = m−1 s∈ 1 1 − ei s s s + op 1 →p 2 f 1 − ei vv 2 1 − ei It follows from (95). c. (106). Part (b) The estimator ˜ is derived from the augmented spectral regression (32) and the relation (34). In (32). take = 0. Since w x d s ∼ wv s + op 1 . In this case. and (110) that ˆ →d 1 − 2 1 − ei −1 fvv ˜ e−i = 1 + −1 2 fvv ˜ e−i e−i − 1 The true coefﬁcient is =b − so we have ˆ →d + 2 1 − ei −1 2 = e−i = 1 + ˜ e−i e−i − 1 fvv − 1 ˜ e−i e−i − 1 which gives the stated result since fxx = 1 − ei −2 fvv .1102 d. (112).
and the asymptotic orthogonality of the regressors in (115) (i. m−1 wv s n1/2 wd s ∗ →p 0).band spectral regression Now take the case wx z s 1103 = 0.e.. From part (f) of Lemma C and (52) we have s = wx d = = wv s ei s − i s 1−e 1 − ei s ˜ ˜ xn − x0 − n−3/2 X D n−1 D D n1/2 −1 −1 n1/2 wd s wv s ˜ ˜ t + w n s xn − x0 − n−3/2 X D n−1 D D 1 − ei s w = v is + An n1/2 wd s 1−e s n1/2 wd s Here. we ﬁnd that s∈ ˜ b1 and √ ˜ m b1 − − →p 1 − ˜ e−i 1 − e−i = 1 − e−i − 1 − e−i → d Nc 0 2 f 2 fvv We deduce that ˜ = a1 ˜ and √ m ˜ − → d Nc 0 f fxx − ˜ + 1 − e−i a2 − ˜ = b1 − 1 − e−i →p giving the stated result. (115). n1/2 wd s = Op 1 from (42) and ˜ ˜ xn − x 0 − n−3/2 X D n−1 D D √ n −1 An = 0 Then (111) is (113) wy d s ∗ = Op 1 = wv s + An n1/2 wd 1 − ei s s ∗ ∗ s 1 − wv s s ∗ ˜ e−i s +w s ∗ s ∗ + op 1 = wv 1 1 − e−i s − ˜ e−i + wd s ∗ n1/2 An + w + op 1 The augmented narrow band regression (32) around frequency (114) wy z s can therefore be written as s ˜ = a1 wv s + An n1/2 wd 1 − ei s s ˜ + a 2 wv + op 1 + residual which is asymptotically equivalent to the regression (115) with ˜ b1 − wy z s ∗ = wv s ∗ ˜ b1 − + n1/2 wd s ∗ ˜ b2 − + residual = ˜ a1 − ˜ + a2 1 − e−i − ˜ ˜ b 2 = a1 − An In view of (113). .
s. ouliaris. as in Lemma C(a).1104 d. phillips Proof of Theorem 3: Note that f n ˆA − A = n−2 X W ∗ AQAWZ AWX = n−2 X QV = n X QV = n XQ −2 Z −2 −1 n−1 X W ∗ AQAWZ AW QV QV Z QV X QV X Q Z −1 −1 n−1 X QV n X QV −1 −1 X n XQ −1 Q Z Next. consider the limit behavior of n−1 X Q Z Q Z = n−1 X = n−1 s∈ A −X wx ˜ s D D w s ∗ D −1 D wx ˜ s − n−3/2 s∈ A ∗ −1 wd s ∗ × n−1 s∈ A wd s wd s n−1/2 s∈ A wd s w s ∗ As in the proof of (91). b. (116) n−2 X Q Z Q Z X→ d 1 0 Bx Bx − 1 0 Bx u 1 0 −1 uu 1 0 f0 B x = 1 0 Bx u Bx u Next. we have n−2 s∈ A wx ˜ s wx ˜ s ∗ → ∗ d 1 0 Bx Bx 1 0 n−3/2 s∈ A wx ˜ s wd ∗ s →d 1 0 d Bx u n−1 s∈ A wd s wd s → uu Thus. (b) and (i). XQ Z =X =X −X P −X Z =X D −X −1 Z Z Z −1 Z D D D and so XQ which is s∈ A Z Q Z X= XQ Z Q Z X =X X −X D D D −1 D X wx ˜ − s wx ˜ wx ˜ s ∗ s∈ A s wd s ∗ s∈ A wd s wd s ∗ −1 s∈ A wd s wx ˜ s ∗ Now. corbae. and p. n−1 s∈ A wx ˜ s w s ∗ → d 1 0 Bx dB . c.
we know from (39) that the vector f1 includes the function ei 1 − ei −1 as one of its components. When A the deterministic variable zt includes a linear time trend. It follows that (118) is simply . XQ cZ c cZ c −1 n−1/2 X W ∗ Ac QAc WZ Ac W cZ c Q cZ X −1 n−1/2 X Q Q cZ Q cZ X= XQ =X = c cZ c c c Q cZ X c X −X s D D s ∗ D −D wx ˜ c X wd s ∗ c s∈ A wx ˜ wx ˜ − c s∈ A −1 s × c s∈ A wd s wd s ∗ c s∈ A wd s wx ˜ s ∗ From the above expression and Lemma C(d). (e). n−1 X Q It follows that f n ˆA − A Z Q Z → d 1 0 Bx dB − 1 0 Bx u 1 0 −1 uu 1 0 udB = 1 0 Bx u dB → d 1 0 −1 Bx u Bx u 1 0 Bx u dB ≡ MN 0 1 0 −1 Bx u Bx u 2 f 0 giving the stated result for the band A . observe that c A d cZ c Q cZ X 1 1 2 1 − ei f1 ∗ 2 c A fxx + f1 + Bx 1 B x 1 − d − f1 1 Bx 1 2 c A e i f1 1 − ei ∗ d 1 2 c A c A d 1 2 c A e−i d Bx 1 1 − e−i 1 Bx 1 B x 1 2 c A fxx f1 1 1 2 1 − ei ∗ − 2 Bx 1 B x 1 f1 d − e i f1 1 − ei ∗ d c A f1 d c A e−i d 1 − e−i − e i f1 1 − ei ∗ d c A f1 f1 ∗ d f1 is the L2 c projection of the function ei 1 − ei −1 onto the space spanned by f1 . Hence. and (h) we deduce that (118) n−1 X Q → × = × Next. in this case we have (119) for ∈ c A c A ei 1 − ei −1 f1 ∗ d c A c A f1 fxx f1 d . ∗ − d f1 = ei 1 − ei −1 . we have A √ f (117) n ˆ Ac − Ac = n−1 X W ∗ Ac QAc WZ Ac WX = n−1 X Q As above. For the band c .band spectral regression and n−1/2 s∈ A 1105 wd s w s ∗ → d 1 0 u dB Thus.
ouliaris. the second factor of (117) decomposes as n−1/2 X Q cZ c Q cZ s = n−1/2 c s∈ A wx ˜ w s ∗ − − n−1/2 c s∈ A wx ˜ s wd s ∗ c s∈ A t wd s wd s ∗ c s∈ A wd s w s ∗ ˜ Using (119) and the independence of xt and n−1/2 X Q cZ c we ﬁnd wx ˜ s Q cZ ∼ n−1/2 − d c s∈ A w s ∗ 1 Bx 1 2 ×n−1/2 c A e i f1 1 − ei f1 s ∗ d s ∗ 1 2 c A f1 f1 ∗ − d c s∈ A w ∼ n−1/2 d c s∈ A wx ˜ × s − 1 Bx 1 2 f1 f1 ei s 1 − ei − c A ∗ e i f1 1 − ei − ∗ d w ∗ 1 2 s c A d w f1 ∗ s s = n−1/2 = n−1/2 ×w = n−1/2 ∼N 0 d d c s∈ A wx ˜ − Bx 1 w s s c s∈ A s ∗ 1 1 − ei s s ei s 1 − ei s ˜ ˜ xn − x0 ei s − Bx 1 n1/2 1 − ei s c s∈ A 1 1 − ei s w s w w d s ∗ + op 1 ∗ 2 n c s∈ A 1 1 − ei f s 2 s w s f s ∼N 0 2 We deduce that √ f n ˆ Ac − giving the stated result. The regression (45) is equivalent to the following regression with frequency domain detrended data: (120) wy d ˜ f s ˜ = c1 wx d ˜ f s ˜ + c2 w f ˜ xd s + residual . c. b. corbae. phillips Proceeding with the proof. and p. s. Ac c A fxx →N 0 d −1 c A −1 fxx d 2 c A fxx f d c A fxx d Proof of Theorem 3 : Part (a) First note that wx s = 2 wz s + wx ˜ s = 2 n wd s + wx ˜ s with a similar expression for wy s .1106 d.
as in the proof of Theorem 2 . using Lemma C(j)–(m). we ﬁnd as in (116) above that n−2 s∈ 0 wx d ˜ f s wx d ˜ f s ∗ → d 1 0 Bx u Bx u and as in (89) we get n−1 s∈ 0 wx d ˜ f s w s ∗ → d 1 0 Bx u dB It follows using (15) and these two results that n ˜f − Next consider the (121) wy d ˜ f s 0 → d 1 0 −1 Bx u B x u 1 0 Bx u dB = 0 case. and −1 ˜ xd s =w = wv ˜ x s ∗ ∗ − wd s s∈ wd s wd s ∗ s∈ wd s w ˜ x s ∗ s + op 1 For = 0. the regressors in (120) are asymptotically orthogonal because n−3/2 s∈ 0 wx d ˜ f s wv s ∗ = Op n−1/2 Next. First we simplify the regression equation (120): ˜ = c 1 wx d ˜ ˜ = c1 w f ˜ xd f s s ˜ + c2 w f ˜ xd s s + residual ˜ + c2 wv + op 1 + residual Using (73) and Lemma C(n)–(o) we obtain −1 (122) wx d ˜ f s ∗ = wx ˜ = s ∗ − wd s ∗ s∈ wd s wd s ∗ s∈ wd s wx ˜ s ∗ wv s ∗ e−i s − 1 − e−i s 1 − e−i 1 − √ f1 n ∗ s ˜ ˜ xn − x 0 n1/2 f1 ∗ − m f n 1 m √ f1 n e−i B 1 1 − e−i .band spectral regression where −1 1107 wx d f s ∗ = wx s ∗ − wd s ∗ s∈ wd s wd s ∗ s∈ −1 wd s wx s ∗ = wx ˜ = wx d ˜ f s ∗ − wd ∗ s ∗ s∈ wd s wd s ∗ s∈ wd s wx ˜ s ∗ s f say s with a similar expression for wy d w f ∗ = wy d ˜ ∗ f s .
Geweke.1108 d. Corbae. M.. and I. 1–11. (1963a): “Regression for Time Series. 8. and P. (1973): “Central Limit Theorems for Time Series Regression. 211–223. Waugh (1993): “Partial Time Series Regression as Compared with Individual Trends.” in Time Series Analysis. 157–169.” Cowles Foundation Discussion Paper. 35. and M. B. E. Phillips (1997): “Band Spectral Regression with Trending Data.” Journal of the American Statistical Association.” Econometrica. B. and J. J. 77. Engle. Ouliaris. C. 252–267.” Journal of Econometrics.” Econometric Theory. Rosenblatt. J. 59. and F. and P. M. C. Granger. I. C. Phillips (1993): “Testing for a Unit Root by Frequency Domain Regression. − − = ˜ c1 − ˜ + c2 1 − e−i − p →p 1 − ˜ e−i 1 − e−i f fvv = 1 − e−i ˜ m b1 − − 1 − e−i → d Nc 0 ˜ + 1 − e−i c2 − ˜ = b1 − 1 − e−i →p . 1. S. Thomson (1971): “Spectral Inference over Narrow Bands. s. by M. I. Rosenblatt (1957): Statistical Analysis of Stationary Time Series. E. 15. 1163.. New York: Academic Press. Grenander. and P. and p. 50. S.. J.” Zeitschrift für Wahrscheinlichkeitsheorie und verwandte Gebiete. R. (1963b): “Regression for Time Series with Errors of Measurement. J. (1982): “Measurement of Linear Dependence and Feedback Between Multiple Time Series. Robinson (1973): “Lagged Regression with Unknown Lags. Frisch.” Journal of the Royal Statistical Society. corbae.. ouliaris.. (1970): Multiple Time Series. that ˜f b1 and √ It follows that ˜ f = c1 ˜ and √ m ˜f − → d Nc 0 f fxx REFERENCES Choi. Hannan. 530–536. W. R.. J. 263–286. as in Theorem 2 .. Series B. ed. 26. D. New York: Wiley. Series and Products.” Biometrika.L Lin (1995): “Causality in the Long Run. J. Yale University. Hannan.. 157–170. 11. 293–302. 304–313. New York: Wiley. New York: Wiley.” International Economic Review. Gradshteyn.” Journal of Applied Probability. c. U. (1974): “Band Spectrum Regression. and P. V. Ryzhik (1965): Tables of Integrals. phillips wv s ∗ e−i s − −i s 1−e 1 − e−i ∗ w s = v −i + op 1 1−e s = s ˜ ˜ xn − x 0 e−i − B 1 1/2 n 1 − e−i So the regression (121) is asymptotically equivalent to (123) wy d ˜ f s ∗ = wv s + op 1 1 − ei s ∗ ˜ c1 + w v s + op 1 ∗ ˜ c2 + residual From (122) and (15) we deduce. b. Hannan. E.
M. Phillips. B. Phillips (1998): “Higher Order Approximations for Frequency Domain Time Series Regression. 783–820. 99–125. Marinucci (1998): “Semiparametric Frequency Domain Analysis of Fractional Cointegration. ed. C. 88. and M. Barnett. Watson (1993): “A Simple Estimator of Cointegrating Vectors in Higher Order Integrated Systems. P. .band spectral regression 1109 Hosoya.. Marinucci. B. and G. and M. 61. J.. 297–336. B. 407–436. C. (1991b): “Optimal Inference in Cointegrated Systems.. C. Wellner (1986): Empirical Processes with Applications to Statistics.” Probability Theory and Related Fields. D.” Review of Economic Studies. 57. Powell. Kim. and B. Xiao.. Lee (1996): “Efﬁciency Gains from Quasidifferencing under Nonstationarity. and J. by P. A. Saikkonen.” in Athens Conference on Applied Probability and Time Series: Essays in Memory of E. 7. 971–1001. Phillips (1999): “Fully Modiﬁed Estimation of Fractional Cointegration Models. Zhijie. 283–306. P. 59. (1991): “The Decomposition and Measurement of the Interdependence between Secondorder Stationary Processes. Robinson. G. C.” LSE Econometrics Discussion Paper #353. (1999): “Discrete Fourier Transforms of Fractional Processes. 59. (1999): “Higher Order Approximations for Wald Statistics in Time Series Regressions with Integrated Processes. J. C.” Yale University. C.. B.. Hannan. M. by W. ed. B.. Yale University. Loretan (1991): “Estimating Longrun Economic Equilibria. (1991a): “Spectral Regression for Cointegrated Time Series. (2000): “Unit Root Log Periodogram Regression. J.” Yale University. (1991): “Asymptotically Efﬁcient Estimation of Cointegration Regressions. Hansen (1990): “Statistical Inference in Instrumental Variable Regression with I(1) processes. P.” Annals of Statistics.” Journal of Econometrics. R. 429–444. P. M.” Cowles Foundation Discussion Paper #1244. Mimeo. (1991): “Automatic Frequency Domain Inference. Phillips.” Econometrica. Y. Shorack. B. 1329–1363.” Review of Economic Studies. Phillips. B. (1998): “Band Spectrum Regression for Cointegrated Time Series with Long Memory Innovations.” in Nonparametric and Semiparametric Methods in Economics and Statistics.” Cowles Foundation Discussion Paper #1243. C. Robinson and M. 59. Rosenblatt. and C. W. 86. P. and V.. Solo (1992): “Asymptotics for Linear Processes.” Econometric Theory.” LSE Econometrics Discussion Paper #348. H.” Econometrica. New York: SpringerVerlag. P. C. 20. Yale University. Robinson P. and D. 1–21. Stock. and P.” Econometrica. and Peter C. Tauchen. Phillips. Cambridge: Cambridge University Press. New York: Wiley. Phillips. P. Mimeo.
This action might not be possible to undo. Are you sure you want to continue?
We've moved you to where you read on your other device.
Get the full title to continue listening from where you left off, or restart the preview.