Page 1
Electronic copy available at: http://ssrn.com/abstract=1984844
Two estimators of the long-run variance: beyond short
memory
Karim M. Abadir∗
Imperial College Business School, Imperial College London, London SW7 2AZ, UK
Walter Distaso
Imperial College Business School, Imperial College London, London SW7 2AZ, UK
Liudas Giraitis
Department of Economics, Queen Mary, University of London, London E14 NS, UK
January 19, 2009
Abstract
This paper deals with the estimation of the long-run variance of a sta-
tionary sequence. We extend the usual Bartlett-kernel heteroskedasticity and
autocorrelation consistent (HAC) estimator to deal with long memory and an-
tipersistence. We then derive asymptotic expansions for this estimator and the
memory and autocorrelation consistent (MAC) estimator introduced by Robin-
son (2005). We offer a theoretical explanation for the sensitivity of HAC to the
bandwidth choice, a feature which has been observed in the special case of short
memory. Using these analytical results, we determine the MSE-optimal band-
width rates for each estimator. We analyze by simulations the finite-sample
performance of HAC and MAC estimators, and the coverage probabilities for
the studentized sample mean, giving practical recommendations for the choice
of bandwidths.
JEL Classification: C22, C14.
Keywords: long-run variance, long memory, heteroskedasticity and autocorrelation
consistent (HAC) estimator, memory and autocorrelation consistent (MAC) estima-
tor.
∗Corresponding author. E-mail address: k.m.abadir@imperial.ac.uk.
1
Page 2
Electronic copy available at: http://ssrn.com/abstract=1984844
1 Introduction and setup
In empirical studies, it is now standard practice to produce robust estimates of stan-
dard errors (SEs). Popular references in econometrics for such procedures include
White (1980), Newey and West (1987), Andrews and Monahan (1992). In statis-
tics, the literature goes further back to Jowett (1955) and Hannan (1957). These
procedures for estimating covariance matrices account for heteroskedasticity and
autocorrelation of unknown form, for short memory models.
There is now an increasing body of evidence suggesting the existence of long
memory in macroeconomic and financial series; e.g. see Diebold and Rudebusch
(1989), Baillie and Bollerslev (1994), Gil-Alaña and Robinson (1997), Chambers
(1998), Cavaliere (2001), Abadir and Talmain (2002). It is therefore of interest to
adapt the most popular of these procedures, the Bartlett-kernel heteroskedastic-
ity and autocorrelation consistent (HAC) estimator, to account for the possibility
of long memory and antipersistence. In addition to HAC, we study the alterna-
tive memory and autocorrelation consistent (MAC) estimator recently introduced
by Robinson (2005). He established the consistency of his MAC estimator of the
covariance matrix, leaving open the issue of its higher-order expansion.
Our first contribution is to derive second order expansions for HAC and MAC
in the univariate case, reducing the problem to the estimation of a scalar (the long
run variance) instead of estimating the covariance matrix. Our derivations give
an insight into the more difficult multivariate case and provide the first step in
understanding this problem.
The second contribution of this paper is to provide a theoretical explanation for
the sensitivity of HAC estimators to the choice of bandwidth, a feature that has
been widely observed in the special case of short memory. Our results show that
the HAC estimator is sensitive because the minimum-MSE bandwidth depends on
the persistence in the series. The theoretical part of this paper explains where the
problem comes from and gives some practical advice for selecting the bandwidth.
We also show that, on the other hand, the MAC estimator is more robust to the
bandwidth selection, since its asymptotic properties are not affected by long memory
or antipersistence.
The final theoretical contribution of this paper is to obtain the distribution
of the estimated normalized spectrum at the origin, by virtue of its link to the
long-run variance. The distribution is Gaussian for MAC, but the one for HAC is
2
Page 3
Gaussian only if the long memory is below some threshold. In the case of short
memory, HAC is the usual Bartlett-kernel estimator of the spectral density at zero
frequency, and its bias and asymptotic distribution are well investigated in the
literature. The asymptotic results for the HAC estimator provide the background
for the development of kernel estimation of a spectral density under long memory
and antipersistence.
The plan of the paper is as follows. In Sections 2 and 3, we derive the bias
and asymptotic expansions for both types of estimators, allowing us to describe the
limiting distributions as well as the asymptotic MSEs. This enables us to determine
the rate of the MSE-optimal bandwidth for each estimator. Section 4 investigates
by simulations the finite-sample performance of HAC and MAC estimators, and
coverage probabilities for the studentized sample mean, giving practical recommen-
dations for the choice of bandwidths. Section 5 concludes. The derivations are given
in the Appendix.
We now detail the setting for our paper. Let {}∈Zbe a stationary sequence
with unknown mean := E(). Let the spectral density of {} be denoted by
() and defined over || ≤ . Suppose that it has the property
() = 0||−2+ (||−2) as → 0(1.1)
where || 12 and 0 0.
ARIMA(): when and are finite; but see Abadir and Taylor (1999) for
Special cases include stationary and invertible
identification issues when or are allowed to be infinite. We shall call the
memory parameter of {}; with = 0 indicating short memory, 0 12 long
memory, and −12 0 antipersistence.
To conduct inference on , define the sample mean¯ := −1P
satisfies
var(12−¯ ) = −1−2
−
As → ∞, we can use assumption (1.1) and a change of variable of integration to
get the convergence
Z∞
where we have the continuous function
⎧
⎩
3
=1 which
Z
µsin(2)
sin(2)
¶2
()d
var(12−¯ ) → 2
:= 0
−∞
µsin(2)
2
¶2
||−2d = 0()(1.2)
() :=
⎨
2Γ(1−2)sin()
(1+2)
if 6= 0
if = 02
(1.3)
Page 4
We notice from (1.2) that 2
is just a scaling of 0by the function (), so in the
usual short memory case of = 0 we get
2
= 2(0)and0= (0)
In general, the problem of the estimation of the long-run variance 2
related to the estimation of and 0≡ lim→0||2() appearing in (1.1). The
HAC and MAC procedures mentioned at the start of this section hinge on the
is closely
estimation of the long run variance 2
.
We will consider the behaviour of the estimators under two alternative sets of
assumptions. The first one is stronger than the second one. It allows the derivation
of asymptotic expansions and the resulting investigation of MSE-optimal bandwidth
rates. The second one is sufficient to establish the consistency of the estimators
for a wide class of stationary sequences. It allows the use of estimates of for
robust SEs for¯ . The second type of conditions are very weak, so they yield only
consistency and are not sufficient to obtain other asymptotic results. 2 The first set
of assumptions is common for HAC and MAC:
Assumption L. {} is a linear sequence
= +
∞
X
=0
− ∈ Z
whereP∞
zero mean and unit variance. Moreover, the spectral density () of {} has the
property
() = ||−2()
where ∈ (−1212) and (·) is a continuous bounded function such that () =
0(1 + (||2)) as → 0 and 0= (0) 0.
Let b 2
t := 12−(¯ − )
b
For HAC, the second type of assumptions (to establish consistency) is:
=02
∞, is a real number and {} are i.i.d. random variables with
(1.4)
be a consistent estimator of 2
the sample mean¯ satisfies
. Under condition (1.4), the t-ratio for
→ N(01) → ∞(1.5)
so that a consistent HAC or MAC estimator of 2
allows inference on .
Assumption M. {} is a fourth order stationary process such that, for some
∈ (−1212) and 6= 0,
∼ 2−1if 6= 0
∞
X
=−∞
|| ∞ if = 0
4
Page 5
where := cov(+); and
∞
X
=−∞
|()| ≤ if 0sup
X
=−
|()| ≤ 2if ≥ 0
where denotes a generic constant and () is a fourth-order cumulant defined
by () := E(+++)−(−+−+−). In addition, if
0, then () ≤ ||−2 ∈ [−].
For MAC, the second type of assumptions differs from Assumption M and is
straightforward to discuss at the end of Section 3.
2 Asymptotic properties of HAC-type estimators
In this section, we first adapt the HAC estimator to allow for long memory and an-
tipersistence, introducing two HAC-type estimators. Then, we analyze their prop-
erties under Assumption L that {} is a linear process, presenting limiting distrib-
utions and asymptotic expansions for the estimators. To the best of our knowledge,
the asymptotic normality of the HAC estimator was investigated in the literature
only in the short memory case of = 0 and under the assumption that E(4
) ∞.
Our Theorem 2.1(a) will require for {} the existence of only a moment of order
2+ (for some 0), which is a new result in the field. It also shows that, under the
strong persistence 14 12, the asymptotic distribution will be non-Gaussian.
Finally, we show that Assumption M guarantees consistency (but not necessarily
the other properties) of the estimators.
Let
e := −1
−
X
=1
(− E())(+− E()) 0 ≤
be the sample autocovariances of {} centered around E(), and
:= −1
−
X
=1
(−¯ )(+−¯ ) 0 ≤
the sample autocovariances of {} centered around the sample mean¯ .
Define
e 2
() := −1−2
X
=1
e |−|= −2(e 0+ 2
X
=1
(1 − )e )(2.1)
5
Page 6
which uses a known (or correctly hypothesized) E(), and
¯ 2
() := −1−2
X
=1
|−|= −2(0+ 2
X
=1
(1 − ))(2.2)
where the mean is estimated unrestrictedly, and assume that the bandwidth para-
meter satisfies
→ ∞ = (1−) (2.3)
for some 0. The difference between the stochastic expansions of the two
estimators will reveal just how much is the impact of estimating E().The
asymptotically-optimal choice of will arise from the first theorem below. To make
() operational, we can employ any estimatorb that is consistent at the
and two such estimators ofb will be discussed later in Section 3.
of this section, we need to assume that the coefficients decay as
e 2
() and ¯ 2
rate of log or faster, calculating e 2
We start by making Assumption L. In addition, to establish the main theorem
(b) and ¯ 2
(b). This is a very weak condition,
= −1+(1 + (−1)) 6= 0 if 6= 0;
∞
X
=0
= 0 if 0; and(2.4)
∞
X
=
|| = (−2) if = 0(2.5)
Such additional requirements are satisfied, for example, by ∼ ARIMA()
where ∈ (−1212). We now derive asymptotic expansions for the estimators
e 2
:=1
2
−∞
(b) and ¯ 2
(b), where the bias will be expressed in terms of
()
sin2(2)1||≤−0||−2
Z∞
µ
(2)2
¶
d(2.6)
In the case of −12 14, these HAC estimators have Gaussian limit distri-
butions. However, if 14 12, then the limit can be written in terms of a
random variable given by the double Itô-Wiener integral
Z00
where (d) is a standard Gaussian complex measure ((−d) is the conjugate of
(d)) with mean zero and variance E(|(d)|2) = d. The limit variable () has
a (non-Gaussian) Rosenblatt distribution and is well-defined when 14 12.
The symbolR00
6
() :=
R2
ei(1+2)− 1
i(1+ 2)
|1|−|2|−(d1)(d2)(2.7)
R2indicates that one does not integrate on the diagonals 1= ±2.
Page 7
Theorem 2.1. Supppose that {} satisfies Assumption L and (2.4)—(2.5), and that
b is an estimator of such that
(b − )log = (1)
(a) If −12 14 and E(||2+) ∞, for some 0, then, as → ∞,
(2.8)
e 2
(b) − 2
= ()12+ −1−2 + (()12) + (−1−2) (2.9)
and
¯ 2
(b) − 2
= ()12+ −1−2 + (()12) + (−1−2) (2.10)
where
→ N(02
Z∞
),
2
:= 82
0
0
µsin(2)
2
¶4
−4d =
⎧
⎩
⎨
162
0
2(21+4−1)
Γ(4+4)cos(2)if 6= 0
if = 0
1622
03 = 44
3
(2.11)
and it is understood that lim→−14(21+4− 1)cos(2) = log4.
(b) If 14 12, E¡4
then
¢ ∞ and () in (1.4) has bounded derivative,
e 2
(b) − 2
= ()1−2e+ −1−2 + (()1−2) + (−1−2)
X
(b) has the property
¯ 2
(2.12)
where
e:= −2
=1
(2
− E(2
))
→ 20();
whereas ¯ 2
(b) − e 2
(b) =
³
()1−2´
) ∞, the MSEs of HAC-type esti-
(2.13)
Under the additional assumption that E(4
mators exist and are minimized asymptotically by
∝
(
1(3+4)−12 14
14 12
12−
(2.14)
where ∝ denotes proportionality. We now list other comments and implications
arising from Theorem 2.1:
Remark 2.1. Since E() = E(e) = 0, the asymptotic bias of the estimators is
given by −1−2. It tends to zero as (hence ) tends to infinity.
7
Page 8
Remark 2.2. When −12 14, the convergence ¯ 2
implies that 2
can be consistently estimated by replacing 0by ¯ 2
(b)
−→ 2
(b)(b) in (2.11).
= ()0
Remark 2.3. If 14, then estimates with known and estimated mean have the
same asymptotic properties. However, if 14, then the rate of convergence of
the sample mean to is rather slow, and replacing by¯ leads to an additional
term in the limiting distribution of the HAC estimator whose consistency is nev-
ertheless unaffected. In the context of hypothesis testing about the mean , one
can estimate the long run variance by treating as unknown and estimating it by
the sample mean. Alternatively, one can compute the long-run variance under the
null hypothesis, treating as known. This will improve the size but may have an
adverse effect on the finite-sample power of tests based on HAC estimators.
Remark 2.4. As a general rule, convergence in distribution does not necessarily
imply a corresponding convergence for moments such as the MSE. However, our
proofs are based on 2expansions for which this implication holds if we make the
additional assumption that E(4
) ∞, hence our stated results for the asymptotic
bias and variance. Note that for the validity of asymptotic expansions (2.9)—(2.10),
only 2 + moments of {} are needed.
Remark 2.5. If {} is a nonlinear process, then Theorem 2.1 might not hold.
For example, the nonlinear transformation = eof a linear process {} will, in
general, increase the bias of estimators. Therefore, the optimal minimizing the
MSE might also change in this case.
Relaxing Assumption L, we obtain the following concistency result.
Theorem 2.2. Suppose that → ∞, = (12), that Assumption M holds, and
thatb − = (1log). Then,
¯ 2
→ 2
(b)
e 2
(b)
−→ 2
as → ∞ (2.15)
3 Robinson’s MAC estimator
In this section, we derive the asymptotic properties of Robinson’s MAC estimator
of 2
= ()0, where () is given by (1.3). We shall show that the asymptotic
properties of the MAC estimator do not depend on the memory parameter , and its
asymptotic distribution is always Gaussian. Hence, it is more robust than HAC to
8
Page 9
the bandwidth selection in practice, something that will be illustrated numerically
in the subsequent section. Define
b 2
() := ()b()
X
where
b() := −1
=1
2
()
is a consistent estimator of 0,
() := (2)−1
¯¯¯¯¯
X
=1
ei
¯¯¯¯¯
2
is the periodogram, = 2 are the Fourier frequencies, and the bandwidth
parameter satisfies → ∞ and = ((log)2).
This estimator has a number of features. First, it does not require estimation
of the unknown mean E() since the periodogram is self-centring at the Fourier
frequencies . Contrast this with HAC estimators; see also Remark 2.3. Second,
as the following theorem will show, the bias and asymptotic distribution of the
estimator do not depend on ∈ (−1212), and the asymptotic distribution is
always Gaussian.
In addition to Assumption L, we will need the condition that () :=
P∞
d
d() = (|()|)
=0eisatisfies
as → 0+(3.1)
in order to derive the CLT in the following theorem.
Theorem 3.1. Suppose that {} satisfies Assumption L with E(4
Assume thatb is an estimator of such thatb − = (1log). Then
b 2
) ∞ and (3.1).
(b) − 2
= −122
+ 2(b − )(log)2
(1 + (1))(3.2)
+(()2) + (−12)
where
→ N(01)(3.3)
The parameter can be estimated, for example, by the local Whittle estimator
b := argmin∈ [−1212]()
9
Page 10
which minimizes the objective function
() := log
⎛
⎝1
X
=1
2()
⎞
⎠−2
X
=1
log
with bandwidth parameter such that → ∞ and = ((log)2). We
use the notation for the bandwidth of the local Whittle estimator, stressing
that it can be set to values that can differ from the bandwidth used in b 2
Theorem 3.1
√(b − )
For the estimation of , the log-periodogram estimator can be used as an alternative
to the local Whittle estimator; see Robinson (1995a).
(b), whenb is the local Whittle estimator. Let
a CLT to analyze the decline of the MSE as increases. Under Assumption L,
E((b − )2) = (−1
b 2
by (3.2). Since (3.5) is derived using an 2approximation, a more detailed analysis
shows that the MSE is ¡−1+ (log)2−1
optimal bandwidth is therefore the one taking that grows at the maximal rate of
45.
.
If = (45), then Robinson (1995b) showed that under the assumptions of
→ N(014)(3.4)
We now turn to the MSE of b 2
= (45) and = (45), since we only need a consistency rate rather than
). Therefore,
³
(b) − 2
´2
=
µ1
+(log)2
¶
(3.5)
¢, hence decreasing in . The MSE-
In general, without recourse to Assumption L, the consistency of Robinson’s
MAC estimator follows immediately fromb
assume Gaussianity or linearity of {}; see Dalla et al. (2006) and Abadir et al.
(2007). For example, if = 0 and (1.4) holds, then such consistency follows under
the assumptionP∞
4Simulation results
→ andb(b)
→ 0. The estimators
b andb(b) are consistent under very weak general assumptions, which do not
=−∞|()| ≤ ; see Corollary 1 of Dalla et al. (2006).
The objective of this section is to illustrate the asymptotic results for the HAC and
(b) and b 2
MAC estimators ¯ 2
(b), to examine their finite-sample performance, and
to give advice on how to choose the bandwidth parameters in practical applications.
10
Page 11
We focus on the MSE because the primary use of these estimators is the consis-
tent estimation of the long-run variance 2
used in various statistics; e.g. in the
denominator b of HAC and MAC robust t-ratios
t := 12−(¯ − )
b
→ N(01) → ∞(4.1)
For this reason, we also consider the closeness of HAC and MAC robust t-ratios to
their limiting normal distributions; see Velasco and Robinson (2001) for expansions
relating to t-ratios using smoothed autocovariance estimates for (0). We study
the coverage probabilities (CPs) of 95% asymptotic confidence intervals (CIs) for ,
considering how the choice of bandwidths affects the closeness of CPs to the nominal
95% level based on the limiting normal distribution of the t-ratio.
We let {} be a linear Gaussian ARIMA(10) process with unit standard
deviation, for different values of (AR parameter) and . We link to 2
, the object
of our analysis, by means of (1.2)—(1.3). Throughout the simulation exercise, the
number of replications is 5,000. We consider three sample sizes = 2505001000
and we estimate the parameter using the local Whittle estimatorˆ with bandwidth
=¥065¦. We do not report the results for =¥05¦¥08¦because they
Table 1 contains the MSE of the HAC estimator ¯ 2
are dominated by =¥065¦.
values of the bandwidth . The minimum-MSE value for each and is highlighted
(b) calculated for different
by shaded gray boxes. The results for these optima are so scattered across the table,
that in practice it will be difficult to achieve them.
Table 2 reports the MSEs of the HAC estimator ¯ 2
(b) when is chosen according
to the asymptotically-optimal rule (2.14). It gives MSEs comparable to the optimal
MSEs of Table 1, except when and are simultaneously large. In this case, the
cost in terms of the MSE can be substantial.
Table 3 contains the MSE of the MAC estimator b 2
that resulted from (3.5): almost all the optima are for = (45) and, in the four
(b) calculated for different
values of the bandwidth . It reveals the accuracy of the simple bandwidth rule
exceptions (shaded boxes), there is little loss in nevertheless sticking to = (45).
Both Tables 2 and 3 show that the MSEs of HAC and MAC estimators usually
increase when || or || increase.
Tables 4 and 5 report CPs for using, respectively, the HAC estimator ¯ 2
(b)
with chosen by the rule (2.14) and the MAC estimator with various bandwidths
. HAC and MAC estimators gives comparable CPs, which are slightly better
11
Page 12
for MAC. CPs approach the nominal 95% level as sample size increases. They are
close to the 95% level except when → 05 or when becomes negative. The
bandwidth =¥08¦tends to give better CPs for MAC, and this is in line with
the recommendations of Table 3.
Because of the specificity of MC studies to the generating process that is used, it
is recommended in practice that the user tries also bandwidths that are smaller than
the maximum allowable =¥08¦which we recommended. This could be used to
check the stability of the estimator as varies near its (unknown) optimal value.
For example, data that are not generated by a linear process (such as ARIMA)
require smaller bandwidths like¥07¦; see Dalla et al. (2006).
5Concluding Remarks
In this paper, the properties of two alternative types of estimators of the long-run
variance have been derived. The first one is an extension of the widely used Bartlett-
kernel HAC estimator, while the second one is the frequency-based MAC estimator
suggested by Robinson (2005). We give guidance on how to choose the bandwidths
in practice, for each estimator. The calculation of both estimators is numerically
straightforward, and allows for the possibility of long-memory or antipersistence in
the data.
Our theoretical results explain that the HAC estimator is sensitive to the se-
lection of the bandwidth , since the order of minimizing the MSE depends on
the extent of the memory in the series. This problem often complicates bandwidth
selection in applied work. The MAC estimator is more robust to the choice of the
bandwidth, which does not depend on the memory. The simulation study confirms
this analytical finding.
On the other hand, the paper does not provide a theory of deriving optimal
estimators, e.g. under MSE-optimality or closeness to normality of the Studentized
t-ratio for . We have studied two types of estimators without establishing whether
or not they are dominated by others, but the asymptotic normality of the MAC
estimator for ∈ (−1212) is an encouraging sign, and so is the good simulation
performance of the two estimators.
12
Page 13
Appendix
A Proofs of the theorems, auxiliary lemmas and propo-
sitions
There are four subsections. The first proves the results relating to the theorems of
Section 2, while the second proves the theorem of Section 3. For the first theorem,
we need lemmas that are derived in the third subsection, and propositions that are
obtained in the fourth one. We require these auxiliary results here, but they can
also be of use beyond our paper.
Throughout this section, we take ∼ to mean that → 1 as → ∞.
A.1Proof of Theorems 2.1 and 2.2
proof of Theorem 2.1. By definitions (2.1), (2.2), and (2.8)
e 2
(b) = e 2
(b) and ¯ 2
()(1 + (1))and¯ 2
(b) = ¯ 2
()(1 + (1))
Condition (2.8) and asymptotic results derived for ¯ 2
(b) by e 2
Also, observe that
Z
where
() and ¯ 2
() below then allows
us to replace e 2
() and ¯ 2
() in the statement of the theorem
without altering the expansions, so we will prove the theorem for e 2
e =
() and ¯ 2
().
−
ei()d=
Z
−
ei¯()d 0 ≤
() := (2)−1¯¯¯
are the corresponding periodograms. Therefore,
X
=1
ei(− E())
¯¯¯
2
¯() := (2)−1¯¯¯
X
=1
ei(−¯ )
¯¯¯
2
e 2
¯ 2
() =
Z
−
()()d (A.1)
and
() =
Z
−
()¯()d(A.2)
where
() := −1−2¯¯¯
X
=1
ei¯¯¯
2= −1−2
µsin(2)
sin(2)
¶2
(A.3)
13
Page 14
is the renormalized Fejér kernel.
By (A.1) and (A.2), we can write ¯ 2
() = e 2
()(¯() − ())d
() + , where
:=
Z
−
(A.4)
In Lemma A.4, we will show that E(||) ≤ (()1−2+ ()). Hence,
¯ 2
() = e 2
() + (()1−2) (A.5)
If −12 14, then ()1−2= (()12), and we can write (A.5) as
³Z
where
:= ()12³
By Proposition A.1,
−→ N(02
sition A.2,
Z
which proves (2.9) and (2.10).
¯ 2
() − 2
= ()12+
−
()()d − 2
´
+ (()12)
e 2
() −
Z
−
()()d
´
), where 2
is given by (2.11), whereas by Propo-
−
()()d = 2
+ −1−2 + (−1−2)
In the case 14 12, write
e 2
() − 2
=¡e 2
() − E¡e 2
()¢¢+ E¡e 2
()¢− 2
= −1−2+(−1−2)+
Proposition A.3 derives the asymptotic bias E¡e 2
totic behavior
()¢−2
(()1−2) and shows that the stochastic term exhibits the nonstandard asymp-
()1−2¡e 2
Thus, the term on the left-hand side above can be approximated by the normalized
sum −2P
Gaussian limit distribution. These relations imply (2.12) and (2.13).
Proof of Theorem 2.2. The conditionb− = (1log) allows us to prove
(2.15) was shown in Giraitis et al. (2003, Theorem 3.1). For 0,
() − E¡e 2
()¢¢= −2
X
=1
(2
− E(2
)) + (1)
−→ 20()
=1(2
− E(2
)) of strongly dependent variables 2
which has a non-
the theorem for e 2
() and ¯ 2
() instead of e 2
1 −||
(b) and ¯ 2
³
(b). For ≥ 0, convergence
2¯ 2
=
X
||
³
´
e +
X
||
1 −||
´
=: 1+ 2
14
Page 15
where
e
= −1
−||
X
=1
(− )(+||− )
´
=(− ). It suffices to show that
=
³
1 −||
(¯ − )2− −1(¯ − )(1−||+ ||+1)
= E(), and :=P
−21
→ 2
and−22
−→ 0 (A.6)
The verification of the relations −22
→ 0 and −2E(1) → 2
is the same as
in Giraitis et al. (2003).
To prove the convergence (A.6), it remains to check that E((1− E(1))2) =
(4) We have E((1− E(1))2) ≤ || + 0, where
:=
X
|||0|
³
1 −||
´³
1 −|0|
´
−2
−||
X
=1
−|0|
X
0=1
(00)
(00) := −0−0+||−|0|+ −0−|0|−0+||
and
0
:= −2
X
X
|||0|≤
−||
X
∞
X
=1
−|0|
X
0=1
|(||0− 0− + |0|)|
≤ −2X
||≤
=−∞|()| ∞ for 0, −12 0 and = ().
To work out , write = + 0where
=1
00=−∞
|(||00)| ≤ −1= (4)
by the assumptionP∞
:=
X
|||0|
³
1 −||
´³
1 −|0|
´
−2
∞
X
=−∞
−|0|
X
0=1
(00)
whereas 0can be bounded by
|0
| ≤ −2
X
|||0|
X
−|| ≤0
X
0=1
|(00)|
We split summation over into three regions: − || ≤ ≤ , , and ≤ 0. In
the case of − || ≤ ≤ , the order of this part of the sum is straightforward.
15
Download full-text