A Near Unit Root Test

Download as pdf or txt
Download as pdf or txt
You are on page 1of 37

ISSN 1440-771X

Department of Econometrics and Business Statistics

http://business.monash.edu/econometrics-and-business-
statistics/research/publications

A Near Unit Root Test for High-


Dimensional Nonstationary Time
Series

Bo Zhang, Jiti Gao and Guangming Pan

May 2019

Working Paper 10/19


A Near Unit Root Test for High–Dimensional
Nonstationary Time Series

Bo Zhang∗ and Jiti Gao† and Guangming Pan‡

May 9, 2019

Abstract

This paper considers a p–dimensional time series model of the form

xt − δt = ϕ(xt−1 − δt−1 ) + Σ1/2 yt , 1 ≤ t ≤ T

where yt = (Yt1 , · · · , Ytp )0 and Σ1/2 is the square root of a symmetric positive definite
matrix. Here ϕ ≤ 1 and T (1 − ϕ) is bounded and the linear processes Ytj is of the
form ∞
P P∞
k=0 bk Zt−k,j where i=0 |bi | < ∞ and {Zij } are independent and identically
distributed (i.i.d.) random variables with EZij = 0, E|Zij |2 = 1 and E|Zij |4 < ∞.
We first investigate the asymptotic behavior of the first k largest eigenvalues of the
sample covariance matrices of the time series model. We then propose an estimator of
ϕ and use it to test for near unit roots. Simulations and empirical applications are also
conducted to demonstrate the performance of the statistic.

Keywords: Asymptotic normality, largest eigenvalue, linear process, near unit root
test.
JEL Classification: C21, C32.

1 Introduction
There is a long literature about testing unit–root time series. Several testing methods have
also been proposed for testing near unit root time series. The literature includes Phillips
and Perron [24] for pioneering a highly popular unit–root test, Phillips [23] for establishing a
novel estimation theory for a regression model associated with a unit root structure, Phillips
and Xiao [25] for a survey of the relevant literature up to 1998, Phillips, Moon and Xiao

Bo Zhang, Department of Econometrics and Business Statistics, Monash University, Caulfield East,
Victoria 3145, Australia. Email: bo.zhang1@monash.edu.

Corresponding Author: Jiti Gao, Department of Econometrics and Business Statistics, Monash University,
VIC 3145, Australia. Email: jiti.gao@monash.edu.

Guangming Pan, School of Physical and Mathematical Sciences, Nanyang Technological University, Sin-
gapore, 637371. Email: gmpan@ntu.edu.sg.

1
[26] for estimating autoregressive roots near unity, and Moon and Phillips [17] for GMM
estimation of panel data models with roots near unity. In the past two decades or so, there
have been studies on testing unit roots and/or near unit roots in panel data settings. One
key contribution is by Choi [9]. Some other contributions include Levin, Lin and Chu [16],
Im, Pesaran, and Shin [13], Chang [7], Pesaran [20], Pesaran, Smith and Yamagata [22] and
Pesaran [21].
As demonstrated in Zhang, Pan and Gao [31], existing unit root tests available for the
panel data setting are not directly applicable for the case where the cross–sectional dimension
is proportional or greater than the time series sample size. To address a corresponding issue
for the near unit case, we propose a new statistic for testing near unit roots in the high–
dimensional setting. One particular feature of the proposed test, which is quite different from
those proposed in Zhang, Pan and Gao [31], is that the proposed test basically mimics the
original version of the unit root test for the low–dimensional case. Since the development of
the theory for the proposed test relies on some asymptotic properties of the largest eigenvalues
of high–dimensional sample covariance matrices, we will briefly discuss the relevant literature
in the rest of this section.
In recent years, large sample properties for high–dimensional sample covariance matrices,
including their eigenvalues and eigenvectors, have been proved to be extremely useful. In fact,
random matrix theory provides many useful methods for estimation and testing procedures
in high-dimensional data analysis. With respect to eigenvalues and eigenvectors, asymptotic
properties for the largest eigenvalues are interesting and useful. There are currently two
main lines of research about asymptotic distributions of the largest eigenvalues of high–
dimensional random matrices. The first line of research is concerned with the Tracy-Widom
law. It is well known that limiting distributions of the largest eigenvalues of high-dimensional
random matrices, such as Wigner matrices, follow the Tracy-Widom law, which was originally
discovered by Tracy and Widom in [28] and [29] for Gaussian Wigner ensembles. The largest
eigenvalue of the Wishart matrix was investigated in Johnstone [14]. Recent progress for
general sample covariance matrices has also been made, and we refer to Bao, Pan and Zhou
[5], Han, Pan and Zhang[12] and El Karoui [11] among others.
Empirical data from finance, speech recognition and wireless communication often implies
that some extreme eigenvalues of sample covariance matrices are well separated from the
rest. This introduces the second line of research about the spiked eigenvalues, which was
first proposed in Johnstone [14]. There are some significant studies in recent years on the
behaviour of these spiked eigenvalues. For instance, the CLTs of the largest eigenvalues of
complex Gaussian sample covariance matrices with a spiked population were investigated in
Baik et al. [3], which also reported an interesting phase transition phenomenon. Baik and
Silverstein [4] further considered almost sure limits of the extreme sample eigenvalues of the
general spiked population. Paul [19] established a CLT for the spiked eigenvalues under the
Gaussian population and the population spikes being simple. The asymptotic distribution of

2
the extreme sample eigenvalues of the general spiked population with arbitrary multiplicity
numbers was further reported in Bai and Yao [2].
Most of the existing studies rely on the assumption that the observations of high dimen-
sional data are independent. However, observations of high–dimensional data in economics
and finance, for example, are often highly dependent. Zhang, Pan and Gao [31] may be the
first paper to deal with the largest eigenvalues of sample covariance matrices generated from
high–dimensional nonstationary time series data. Unfortunately, the result of [31] is also
limited. It only considers the case that there is a unit root in the time series. As we know,
there are many time series with near unit roots and it is hard to use the theory of [31] to
deal with them. So we need some more general results.
This paper establishes the asymptotic behaviour of the first k largest eigenvalues of the
sample covariance matrices of the time series model with near unit root. Another contribution
of this paper is that it develops a new near unit root test. We conclude this section by
giving its organization. Section 2 establishes an asymptotic distributional theory for the
first several largest eigenvalues of the covariance matrix of a high–dimensional dependent
time series. Section 2 also proposes a new near unit root test that is devoted to testing
nonstationarity for high dimensional dependent data. Section 3 evaluates both the size and
power properties of the proposed test. Section 4 concludes the paper with some remarks.
Appendix A establishes some useful lemmas for the largest eigenvalues of sample covariance
matrices and the proof of the main theorems in section 2. Appendix B gives and proves the
truncated versions of lemmas in Appendix A by truncating linear processes. Appendix C
gives the proofs of lemmas in Appendix A.

2 Asymptotic Theory

2.1 Matrix models


The paper is to investigate high–dimensional sample covariance matrices for nonstationary
time series. Let Ytj be MA(q) with q being either fixed, or tending to infinity at a certain
rate, or ∞.
Define the linear processes Ytj by
q
X
Ytj = bk Zt−k,j (2.1)
k=0

q
P
where |bi | < ∞, and {Zij } are independent and identically distributed (i.i.d.) random
i=0
variables with EZij = 0, E|Zij |2 = 1 and E|Zij |4 < ∞. Suppose that yt = (Yt1 , · · · , Ytp )0 is
a p–dimensional time series. Consider a p–dimensional time series model of the form:

xt − δt = ϕ(xt−1 − δt−1 ) + Σ1/2 yt , 1 ≤ t ≤ T, (2.2)

3
where {δt }0≤t≤T is non-random or independent of {xt − δt }0≤t≤T , ϕ ≤ 1 and T (1 − ϕ) is
bounded.
We also define two T × p matrices ∆ = (δ1 , · · · , δT )0 and X́0 = (x0 − δ0 , · · · , x0 − δ0 )0
consisting of the initial vector x0 of the time series.
Throughout the paper, we make the following assumptions about the coefficients bi and
Σ:
Pq
(A1) i=0 i|bi | < ∞.
Pq
(A2) i=0 bi = s 6= 0.
tr(Σ1/2 )
(A3) There exist two positive constants M0 and M1 such that kΣk2 ≤ M0 and p
≥ M1 .

p
(A4) T → ∞ and p → ∞ such that limT →∞ T
= 0.

Assumption A1 implies that the linear process can include MA(q) models and AR(1) models.
Assumption A2 is easily satisfied. Here k · k2 stands for either the spectral norm of a matrix
or the Euclidean norm of a vector. Then Assumption A3 is common. Assumption A4 shows
that we do not require p and T to be of the same order, which is being commonly used in
the random matrix theory literature.

Remark 1. Assumptions A3 and A4 may lead to the main challenge of the model. Since there
is no extra structural assumption on Σ and p may be very large, it’s hard to get a consistent
estimator of Σ (or the cross-section dependence). Then most of the existing methods may
not work.

We can also define ai as follows.



X
ai = bk bk+i . (2.3)
k=0

Then ai = EYtj Yt+i,j .


Now we also need to make some assumptions about Zij and x0 − δ0 .

(A5) {Zi,j } are i.i.d random variables with mean zero, variance one and finite fourth moment.
Let zt = (Zt1 , · · · , Ztp )0 , where t can be either positive or negative integer (for the
purpose of introducing A7 below).

(A6) Ekx0 − δ0 k22 = O(p).



b̃k Σ1 1/2 z−k + b̃−1 Σ2 1/2 z̃ + b̃−2 , where
P
(A7) x0 − δ0 = kΣ1 k2 ≤ M0 , kΣ2 k2 ≤ M0 and z̃ =
k=0
(Z̃1 , · · · , Z̃p )0 is independent of zt for any t, in which
{Z̃j } are i.i.d random variables with

P
mean zero, variance one and finite fourth moments. The coefficients satisfy |b̃k | +
k=0
2
|b̃−1 | < ∞ and kb̃−2 k = O(p).

4
We would like to remark that Assumption A7 implies Assumption A6.
We next specify assumptions about ϕ.

(A8) The unknown ϕ satisfies:


lim T (1 − ϕ) = c ≥ 0. (2.4)
T →∞

(A9)
√ √
min{q, p} p(1 − ϕ) = O(1). (2.5)

The choice of ϕ in (2.4) covers many important cases where both unit–root and near–unit
root scenarios can be accommodated. If Assumptions A4 and A8 hold and q is bounded,
Assumption A9 automatically holds.
We next specify assumptions about {δt }1≤t≤T .

(A10) {δt }0≤t≤T is non-random or independent with {xt − δt }0≤t≤T and trace( √1p ∆(I − H)∆∗ ) =
PT 0
√1 2
p t=1 δt (I − H)δt = op (T ).

min{q, p} PT
(A11) √
p t=2 (δt − δt−1 )0 (I − H)(δt − δt−1 ) = op (T ).
PT 0
PT 0
Remark 2. Note that the order of t=1 δt (I − H)δt may be different from that of t=1 δt δt .
For example, consider δt = gt 1 + δ̃t where gt is large and kδ̃t k2 is small. One may find that
PT 0 PT 0 PT 0
t=1 δt δt is large but t=1 δt (I − H)δt = t=1 δ̃t (I − H)δ̃t is small.

2.2 The test statistics and asymptotic theory


We propose our statistics based on a new estimator of ϕ. The basic idea is as follows.
The largest challenge of our model is that Σ is hard to be estimated. From the main
theoretical results of [31], we find that when ϕ = 1, the largest eigenvalues of B depend on
p
ai tr(Σ) and ai tr(Σ2 ), which can be estimated. So we consider the general case of ϕ → 1
p
and then prove that the largest eigenvalues depend on ai tr(Σ), ai tr(Σ2 ) and ϕ. When
p
we obtain the estimators of ai tr(Σ) and ai tr(Σ2 ), we can develop a relation between the
largest eigenvalue and ϕ. Then we can find an estimator for ϕ from the largest eigenvalue.
Since there is some difficult to get an explicit expression based on ϕ in some steps, we
introduce θ as follows. Define π > θϕ,1 > · · · > θϕ,T > 0 are the solutions of the equation

ϕ sin T θ + sin(T + 1)θ = 0. (2.6)

The computational procedure is as follows.


Step 1: Define the sample covariance matrix B by
1
B = X(I − H)X∗ (2.7)
p

where X = (x1 , · · · , xT )0 and H = p1 110 with 1 being a 1 × p column vector.


We calculate the largest eigenvalue ρ1 of B.

5
Step 2: Estimate q by calculating (xt − xt−1 )0 (I − H) (xt+i − xt+i−1 ). Then define m =

min{q̂, p}.
Step 3: Calculate

T m1 XT −j
X x̀t,t X x̀t,t+j
µ̀m = +2 , (2.8)
t=2
p(T − 1) j=1 t=2
p(T − j − 1)

where
x̀t,s = (xt − xt−1 )0 (I − H) (xs − xs−1 ) .

Here µ̀m is the estimator of the term which includes ai tr(Σ).


Step 4: Define
 2
sin(T + 1)θ sin(T + 1)θ 1 − cos 2θ
g2 (θ) = −2 (1 + cos θ) + 1 + = ,
sin T θ sin T θ 1 − cos 2T θ

 3  2T −m
sin(T +1)θ sin(T +1)θ

sin(T + 1)θ
m − sin T θ
− − sin T θ
κ(θ) = − −m ,
sin T θ 2(T − 1)

f (θ) = g2 (θ)κ(θ). (2.9)

and
h(θ) = ρ1 f (θ) − µ̀m . (2.10)

We can obtain the largest solution θ̂ of h(θ) = 0 on (0, π).


Step 5: Define
sin(T + 1)θ
g1 (θ) = − .
sin T θ
From (2.6), we can calculate ϕ̂ = g1 (θ̂).
Step 6: Define
[T /2] T r
x̀2t,s
P P
S̀σ2 ,0
t=2 s=t+[T /2]
|µ̀m | 2 p
S̀σ2 ,0 = , S̀σ2 ,m = . (2.11)
(T − 32 [T /2])([T /2] − 1) T
P x̀t,t
p(T −1)
i=2
p
Here S̀σ2 ,m is the estimator of the term which includes ai tr(Σ2 ).
We first establish the following theorem, which will be used in the proof of Theorem 2
below.

Theorem 1. Suppose that Assumptions A1–A5 and A7–A11 hold. Then,

√ ρ1 g2 (θϕ,1 )κ(θϕ,1 ) − µ̀m


p → N (0, 1). (2.12)
S̀σ2 ,m

p ρ1 f 0 (θ̂)
Define TeN = S̀m2 g10 (θ̂)
(ϕ̂ − ϕ), which generalizes the test statistics TN of [31]. Then we
can establish our main results as follows.

6
Theorem 2. Suppose that Assumptions A1–A5 and A7–A11 hold.

(1) Then the solution θ̂ is unique such that


   
Tπ Tπ
lim P ∃θ̂ ∈ , so that h(θ̂) = 0 = 1. (2.13)
p,T →∞ T + 1 T + 14

(2) As T → ∞, we have
√ ρ1 f 0 (θ̂)
p (θ̂ − θϕ,1 ) → N (0, 1). (2.14)
Sσ2 ,m

(3) Moreover, we have as T → ∞


TeN → N (0, 1). (2.15)

It is also stressed that TeN is applicable in many different choices of ϕ involved in (2.4).
For a general form of ϕ = 1 − T χ√p , we consider both unit–root and near unit root scenarios:

χ = 0, χ = √1p , χ = 1 or χ = p. As shown in the simulation studies in Section 3 below, TeN
works well numerically.

3 Simulation

3.1 Estimation of q
Generally, the estimation of q can be done as follows. Calculate
PT −j
t=2 x̀t,t+j
ζj = (3.1)
p(T − j − 1)

and
√ √
q̂ = min{0 ≤ i < [ p] : |ζj | < p−1/4 T −1/2 , i < j < [ p]}. (3.2)

3.2 Comparison with some existing tests


There are several existing unit root tests for panel data. Some of them consider the case where
there is no cross–sectional dependence and such test don’t work when there is cross–sectional
dependence (see, for example, the IPS test proposed in [13]). To test for nonstationarity in
the panel data case with cross-sectional dependence, [7] showed that the Bootstrap method
with estimation of Σ performs better for the case where p is fixed and T is large. In [31], the
simulation results have showed that Bootstrap-OLS t∗ols and the F–statistic corresponding to

Bootstrap-OLS Fols in [7] don’t produce good size and power properties when p is large. Now
we run some simulations to show that there are similar problems for the near unit root case
when p is large.    
|i−j|
We use the setting yt = zt , δt = 0 and Σ = Σi,j = 0.3 . We compare the size
performance of our test TN with the two tests t∗ols and ∗
Fols under H0 with x0 = 0. Table 1–3

7
report the results of the three tests based on 1000 replications, 500 bootstrap replications
and different values of p and T . The nominal size throughout this section is set to be 0.05.
Tables 1–3 below show that when p becomes large, both t∗ols and Fols

have a poor size
property even though yt is independent over t. They don’t work for the near unit root when
p is large.

1
Table 1: The empirical size of two tests under H0 : ϕ = 1 − Tp

the test T \p 10 20 40 60 80

t∗ols 40 0.011 0.007 0.001 0.000 0.000



Fols 40 0.025 0.014 0.001 0.003 0.001
t∗ols 60 0.024 0.016 0.000 0.000 0.000

Fols 60 0.042 0.023 0.000 0.000 0.002
t∗ols 80 0.042 0.020 0.009 0.001 0.000

Fols 80 0.050 0.025 0.012 0.001 0.000
t∗ols 160 0.043 0.029 0.017 0.010 0.003

Fols 160 0.066 0.037 0.020 0.011 0.003
t∗ols 320 0.045 0.032 0.036 0.024 0.012

Fols 320 0.061 0.048 0.047 0.032 0.015

1
Table 2: The empirical size of two tests under H0 : ϕ = 1 − T

p

the test T \p 10 20 40 60 80

t∗ols 40 0.021 0.005 0.000 0.001 0.000



Fols 40 0.040 0.007 0.000 0.001 0.000
t∗ols 60 0.026 0.011 0.002 0.000 0.000

Fols 60 0.035 0.017 0.002 0.000 0.000
t∗ols 80 0.042 0.020 0.002 0.000 0.000

Fols 80 0.047 0.026 0.003 0.000 0.000
t∗ols 160 0.040 0.036 0.021 0.011 0.001

Fols 160 0.052 0.046 0.026 0.032 0.013
t∗ols 320 0.058 0.047 0.027 0.012 0.001

Fols 320 0.066 0.058 0.029 0.037 0.014

8
1
Table 3: The empirical size of two tests under H0 : ϕ = 1 − T

the test T \p 10 20 40 60 80

t∗ols 40 0.021 0.002 0.000 0.000 0.000



Fols 40 0.027 0.0002 0.000 0.000 0.000
t∗ols 60 0.032 0.016 0.002 0.000 0.000

Fols 60 0.042 0.018 0.002 0.000 0.000
t∗ols 80 0.036 0.019 0.002 0.000 0.000

Fols 80 0.047 0.022 0.002 0.000 0.000
t∗ols 160 0.040 0.037 0.016 0.008 0.005

Fols 160 0.051 0.037 0.029 0.023 0.018
t∗ols 320 0.047 0.041 0.017 0.008 0.005

Fols 320 0.052 0.035 0.032 0.024 0.019

3.3 The simulation for TeN


This subsection is to conduct some simulations to investigate the size
 and power
 of T̃N  . We
|i−j|
now consider the setting where yt = ψzt−1 + zt , ψ = 0.5 and Σ = Σi,j = 0.3 . We
 
also set the elements of δt as δit = cos 2π(i+t)
T
.
Since TeN works for different ϕ, we rewrite ϕ = 1 − χ√ . The bigger χ is, the further ϕ
T p
goes to 1. The size and power results of TeN based on 1000 replications and different values of
p, T , χ for the different hypothesises are reported in Tables 4–7. The sizes and power values
were all computed using asymptotic critical values from the standard normal distribution.

Tables 4–7 below provide us with the following findings.

• Table 4 shows that the test TeN has good sizes and power values even when p and T
are as small as 10 and 20, respectively. Meanwhile, the power values increase when χ
increases, and the sizes become stable even when p increases to 80. Table 4 particularly
reveals some findings for the case where the null hypothesis is a unit root of the form:
H0 : ϕ = 1, and the alternative is a near unit root of the form: H1 : ϕ = 1 − T1 . In
this setting, the power values are quite substantial with 0.239 for the case of (p, T ) =
(10, 20).

• Tables 5–7 have similar features to those observed in Table 4, but there are some
differences. For instances, Table 5 shows that when H0 : χ = √1
holds, the so–called
p

power values under H1 : χ = 0 or H1 : χ = 1 look more like sizes. Under H1 : χ = p,
however, when the departure is substantial, the power values increase significantly, and
they also increase when either p or T increases.

9
Table 4: Sizes and power values under H0 : χ = 0

p T χ = 0 (size) χ= √1 (power) χ = 1 (power) χ= p (power)
p

10 20 0.040 0.041 0.066 0.239


10 30 0.044 0.044 0.079 0.308
10 40 0.051 0.043 0.081 0.347
10 60 0.053 0.057 0.094 0.371
10 80 0.049 0.053 0.096 0.369
20 20 0.039 0.031 0.047 0.408
20 30 0.044 0.045 0.081 0.495
20 40 0.053 0.054 0.090 0.539
20 60 0.048 0.054 0.089 0.593
20 80 0.042 0.054 0.082 0.606
30 20 0.038 0.035 0.058 0.598
30 30 0.049 0.049 0.064 0.651
30 40 0.052 0.048 0.086 0.707
30 60 0.039 0.044 0.076 0.731
30 80 0.055 0.057 0.089 0.748
30 20 0.043 0.054 0.093 0.482
30 30 0.027 0.043 0.068 0.505
30 40 0.035 0.040 0.056 0.523
30 60 0.038 0.037 0.069 0.581
30 80 0.034 0.033 0.059 0.593
40 20 0.049 0.045 0.092 0.630
40 30 0.045 0.036 0.075 0.666
40 40 0.035 0.046 0.068 0.685
40 60 0.026 0.038 0.072 0.734
40 80 0.039 0.041 0.061 0.751
60 20 0.054 0.061 0.089 0.801
60 30 0.038 0.047 0.073 0.833
60 40 0.040 0.045 0.060 0.857
60 60 0.040 0.030 0.065 0.874
60 80 0.040 0.036 0.055 0.913
80 20 0.064 0.053 0.080 0.897
80 30 0.032 0.051 0.066 0.946
80 40 0.038 0.042 0.074 0.943
80 60 0.040 0.044 0.062 0.953
80 80 0.050 0.051 0.070 0.962

10
Table 5: Sizes and power values under H0 : χ = √1
p


p T χ = 0 (power) χ = √1 (size) χ = 1 (power) χ = p (power)
p

10 20 0.050 0.047 0.050 0.189


10 30 0.056 0.057 0.068 0.280
10 40 0.052 0.049 0.075 0.312
10 60 0.039 0.035 0.071 0.312
10 80 0.044 0.051 0.065 0.312
20 20 0.041 0.033 0.040 0.400
20 30 0.044 0.039 0.061 0.476
20 40 0.045 0.041 0.064 0.490
20 60 0.044 0.044 0.083 0.531
20 80 0.043 0.044 0.069 0.562
30 20 0.041 0.048 0.082 0.450
30 30 0.027 0.040 0.060 0.487
30 40 0.034 0.037 0.050 0.490
30 60 0.040 0.036 0.062 0.550
30 80 0.034 0.035 0.051 0.548
40 20 0.046 0.041 0.079 0.611
40 30 0.046 0.037 0.065 0.648
40 40 0.034 0.040 0.058 0.669
40 60 0.027 0.038 0.064 0.711
40 80 0.035 0.039 0.054 0.721
60 20 0.047 0.059 0.079 0.785
60 30 0.032 0.044 0.066 0.824
60 40 0.040 0.046 0.056 0.840
60 60 0.037 0.027 0.061 0.864
60 80 0.043 0.038 0.051 0.902
80 20 0.061 0.046 0.071 0.894
80 30 0.033 0.049 0.057 0.941
80 40 0.036 0.038 0.069 0.931
80 60 0.042 0.045 0.060 0.949
80 80 0.049 0.049 0.060 0.958

11
Table 6: Sizes and power values under H0 : χ = 1

p T χ = 0 (power) χ = √1 (power) χ = 1 (size) χ= p (power)
p

10 20 0.058 0.050 0.044 0.136


10 30 0.062 0.059 0.052 0.177
10 40 0.059 0.052 0.045 0.204
10 60 0.061 0.057 0.052 0.200
10 80 0.047 0.043 0.042 0.227
20 20 0.064 0.056 0.035 0.301
20 30 0.077 0.065 0.042 0.332
20 40 0.069 0.066 0.050 0.381
20 60 0.077 0.062 0.054 0.405
20 80 0.071 0.064 0.048 0.412
30 20 0.046 0.038 0.052 0.343
30 30 0.045 0.031 0.040 0.350
30 40 0.052 0.044 0.029 0.339
30 60 0.067 0.050 0.042 0.396
30 80 0.072 0.064 0.044 0.399
40 20 0.042 0.039 0.051 0.459
40 30 0.048 0.043 0.038 0.508
40 40 0.045 0.044 0.033 0.515
40 60 0.062 0.048 0.051 0.540
40 80 0.067 0.064 0.044 0.544
60 20 0.040 0.046 0.049 0.684
60 30 0.050 0.052 0.046 0.711
60 40 0.051 0.063 0.036 0.735
60 60 0.058 0.056 0.040 0.781
60 80 0.065 0.047 0.032 0.812
80 20 0.057 0.040 0.047 0.830
80 30 0.052 0.042 0.039 0.869
80 40 0.050 0.049 0.046 0.879
80 60 0.060 0.057 0.040 0.893
80 80 0.066 0.068 0.034 0.921

12

Table 7: Sizes and power values under H0 : χ = p

p T χ = 0 (power) χ = √1 (power) χ = 1 (power) χ = p (size)
p

10 20 0.185 0.159 0.109 0.054


10 30 0.189 0.139 0.094 0.056
10 40 0.202 0.163 0.114 0.056
10 60 0.231 0.186 0.113 0.063
10 80 0.218 0.171 0.108 0.053
20 20 0.270 0.259 0.185 0.051
20 30 0.343 0.306 0.204 0.057
20 40 0.389 0.339 0.229 0.046
20 60 0.435 0.391 0.266 0.057
20 80 0.449 0.411 0.267 0.064
30 20 0.502 0.486 0.374 0.050
30 30 0.592 0.561 0.459 0.051
30 40 0.652 0.656 0.505 0.041
30 60 0.704 0.681 0.532 0.061
30 80 0.721 0.730 0.533 0.042
40 20 0.624 0.626 0.498 0.060
40 30 0.749 0.742 0.599 0.046
40 40 0.762 0.747 0.616 0.041
40 60 0.825 0.803 0.666 0.055
40 80 0.846 0.826 0.693 0.061
60 20 0.796 0.795 0.718 0.046
60 30 0.889 0.882 0.797 0.060
60 40 0.908 0.905 0.820 0.049
60 60 0.939 0.927 0.852 0.056
60 80 0.936 0.948 0.870 0.043
80 20 0.889 0.893 0.847 0.055
80 30 0.957 0.941 0.917 0.042
80 40 0.959 0.959 0.922 0.049
80 60 0.980 0.973 0.949 0.061
80 80 0.975 0.977 0.945 0.049

13
4 Conclusions and discussion
This paper has proposed a new near unit test for the case where the dimensionality of a vector
of time series diverges along with the sample size. The proposed test has been constructed
in a similar fashion to that of the original Dickey–Fuller test, and the main difference is that
the asymptotic distribution of the proposed test is standard normal, while the DF test has a
nonstandard limiting distribution.
As shown in the simulation study, some of the existing tests proposed for the low–
dimensional setting don’t work well numerically. The proposed test however has good size
and power properties in the finite–sample cases.

5 Acknowledgements
The authors would like to thank the Australian Research Council Discovery Grants Program
for its support under Grant numbers: DP150101012 & DP170104421.

References
[1] Bai, Z. D. and Silverstein, J. W. (2006). Spectral Analysis of Large Dimensional Random Matrices.
2nd Edition, Springer, New York.

[2] Bai, Z.D. and Yao, J.F (2008).Central limit theorems for eigenvalues in a spiked population model.
Annales de l’Institut Henri Poincaré 44, 447-474.

[3] Baik, J., Ben Arous, G. and Péché, S. (2005). Phase transition of the largest eigenvalue for non–null
complex sample covariance matrices. Annals of Probability 33, 1643–1697.

[4] Baik, J. and Silverstein, J. W. (2006). Eigenvalues of large sample covariance matrices of spiked
population models. Journal of Multivariate Analysis 97, 1382–1408.

[5] Bao, Z. G., Pan, G. M. and Zhou, W. (2015). Universality for the largest eigenvalue of sample
covariance matrices with general population. Annals of Statistics 43, 382–421.

[6] Chan, N.H. and Wei, C.Z. (1988). Limiting distributions of least squares estimates of unstable
autoregressive processes. Annals of Statistics 16, 367–401.

[7] Chang, Y. (2004). Bootstrap unit root tests in panels with cross sectional dependency. Journal of
Econometrics 120, 263–293.

[8] Chen, B. B. and PAN, G. M. (2012). Convergence of the largest eigenvalue of normalized sample
covariance matrices when p and n both tend to infinity with their ratio converging to zero. Bernoulli
18, 1405–1420.

[9] Choi, In. (2001). Unit root tests for panel data. Journal of International Money and Finance 20,
249–272.

[10] Dickey, D.A. and Fuller, W.A. (1979). Distribution of the estimators for autoregressive time series
with an unit root. Journal of the American Statistical Association 74, 423–431.

14
[11] El Karoui, N. (2007). Tracy-Widom limit for the largest eigenvalue of a large class of complex sample
covariance matrices, Annals of Probability 35, 663–714.

[12] Han, X., Pan, G. M. and Zhang, B. (2016). The Tracy–Widom law for the largest eigenvalue of F
type matrix, Annals of Statistics 44, 1564-1592.

[13] Im, K., Pesaran, M.H. and Shin, Y. (2003). Testing for unit roots in heterogeneous panels. Journal
of Econometrics 115, 53–74.

[14] Johnstone, I.M. (2001). On the distribution of the largest eigenvalue in principal component analysis.
Annals of Statististics 29, 295–327.

[15] Johnstone, I.M. (2007). High dimensional statistical inference and random matrices. In International
Congress of Mathematicians I, 307–333.

[16] Levin, A., Lin, C.F. and Chu, C.S.J. (2002). Unit root tests in panel data: asymptotic and finite-
sample properties. Journal of Econometrics 108, 1–24.

[17] Moon H. R. and Phillips, P, C. B. (2004). GMM Estimation of Autoregressive Roots Near Unity
with Panel Data. Econometrica 72, 467–522.

[18] Pan, G. M., Gao, J. and Yang, Y. R. (2014). Testing independence among a large number of high
dimensional random vectors. Journal of the American Statistical Association 109, 600–612.

[19] Paul, D. (2007). Asymptotics of sample eigen–structure for a large dimensional spiked covariance
model. Statistica Sinica 17, 1617–1642.

[20] Pesaran, M.H. (2007). A simple panel unit root test in the presence of cross–sectional dependence.
Journal of Applied Econometrics 22, 265–312.

[21] Pesaran, M.H. (2015). Time Series and Panel Data Econometrics. Oxford University Press, Oxford.

[22] Pesaran, M.H., Smith, L. V. and Yamagata, T. (2013). Panel unit root tests in the presence of a
multifactor error structure. Journal of Econometrics 175, 94–115.

[23] Phillips, P, C. B. (1988). Regression theory for near integrated time series. Econometrica 56, 1021–
1044.

[24] Phillips, P, C. B. and Perron, P. (1988). Testing for a unit root in time series regression. Biometrika
75, 335–346.

[25] Phillips, P, C. B. and Xiao, Z. (1998). A Primer on unit root testing. Journal of Economic Surveys
12, 423–470.

[26] Phillips, P, C. B., Moon H. R. and Xiao, Z. (2001). How to estimate autoregressive roots near
unity ? Econometric Theory 17, 26–69.

[27] Soshnikov, A. (2002). A note on universality of the distribution of the largest eigenvalues in certain
sample covariance matrices. Journal of Statistical Physics 108, 1033–1056.

[28] Tracy, C. A. and Widom, H. (1994). Level-spacing distributions and the Airy kernel. Communications
in Mathematical Physics 159, 151–174.

[29] Tracy, C. A. and Widom, H. (1996). On orthogonal and symplectic matrix ensembles. Communica-
tions in Mathematical Physics 177, 727–754.

15
[30] Yao, J.F., Zheng, S. R. and Bai, Z.D. (2015). Large Sample Covariance Matrices and High-
Dimensional Data Analysis. Cambridge University Press.

[31] Zhang, B , Pan, G. M, and Gao, J. (2018). CLT for Largest eigenvalues and unit root tests for
high–dimensional nonstationary time series. Annals of Statistics 46, 2186–2215.

A Proofs of the main results


We first will give some lemmas which are proved in the supplementary file.
Define for k = 1, · · · , T ,
1
λϕ,k = (A.1)
2ϕ(1 + cos θϕ,k ) + (1 − ϕ)2

where θϕ,k is defined in (2.6).


We also define  

X
γϕ,k = λϕ,k a0 + 2 aj (−1)j cos(jθϕ,k ) , (A.2)
j=1

Lemma 1. Suppose that Assumptions A1–A5 and A7–A9 hold. Let ρk be the kth largest eigenvalue
of B. When k is fixed, the random vector
0


tr(Σ1/2 (I−H)Σ1/2 ) tr(Σ1/2 (I−H)Σ1/2 )
p ρ1 − γϕ,1 p ρk − γϕ,k p
q  ,··· ,  (A.3)
2tr(Σ1/2 (I−H)Σ(I−H)Σ1/2 ) γϕ,1 γϕ,k
p

converges weakly to a zero-mean Gaussian vector w = (w1 , · · · , wk )0 with the covariance function
cov(wi , wj ) = 0 for any i 6= j and var(wi ) = 1.

Lemma 2. If the Assumptions A1-A7 and A9 hold, then

κ(θϕ,1 ) = 1 + Op (m(1 − ϕ)) (A.4)

and

tr(Σ1/2 (I − H)Σ1/2 ) X 1
µ̀m = κ(θϕ,1 )(a0 + 2 aj ) + op (m(1 − ϕ)) + Op (1 − ϕ) + op ( √ ). (A.5)
p p
j=1

Lemma 3. Under the conditions specified in lemma 2,

S̀σ2 ,0 tr(Σ1/2 (I − H)Σ(I − H)Σ1/2 )


= a20 + op (1), (A.6)
p p

r
S̀σ2 ,0 s
|µ̀m2 | 2 p

X 2tr(Σ1/2 (I − H)Σ(I − H)Σ1/2 )
S̀σ2 ,m = T
= (a0 + 2 ai ) + op (1) (A.7)
P x̆i,i p
i=1
p(T −1)
i=2

and
q
2tr(Σ1/2 (I−H)Σ(I−H)Σ1/2 )
γϕ,1 p
= 1 + op (1). (A.8)
λϕ,1 S̀σ2 ,m

16
Proof of Theorem 1: We conclude from Lemmas 1-3 that under the conditions in Theorem 1,
ϕ,1 λ
√ ρ1 − κ(θϕ,1 ) µ̀m
p → N (0, 1). (A.9)
λϕ,1 S̀σ2 ,m
1
Note that λϕ,1 = g2 (θϕ,1 ) and (A.4), (2.12) then follows.
ρ1
Proof of Theorems 2: At first we prove (2.13). From Lemmas 1-2 we have µ̀m = λϕ,1 (1 + op (1)).
1
We can define ϕ̌ = ϕ − T and θ̌ ∈ ( TT+1
π
, π) such that g1 (θ̌) = ϕ̌. Then

lim P (h(θ̌) > 0) = 1.


p,T →∞

Similarly we obtain

lim P (h( ) < 0) = 1.
p,T →∞ T + 41
Thus (2.13) follows from the continuity of h(θ).
We next prove that the solution is unique. It suffices to prove that f (θ) is monotone on the
interval ( TT+1
π
, TT+π1 ). From the chain rule f 0 (θ) = g20 (θ)κ(θ) + g2 (θ)κ0 (θ). It is straightforward to
4
verify that κ(θ) = 1 + op (1), κ0 (θ) = Op (m1 ) and g2 (θ) = Op ( T12 ). As for g20 (θ) we have

sin 2θ sin T θ − 2T cos T θ sin2 θ


g20 (θ) = . (A.10)
sin3 T θ

We rewrite θ = T +ı where 0 < ı < 1. Then

sin(T + 1)θ sin T (1−ı)π


T +ı
g1 (θ) = − = . (A.11)
sin T θ sin TT +ı
ıπ

sin 2T π
T +ı sin(T π −
T ıπ
T +ı ) − 2T cos(T π − T ıπ 2 Tπ
T +ı ) sin T +ı
g20 (θ) = 3 T ıπ
sin (T π − T +ı )
Tπ T ıπ T ıπ Tπ
T π cos T +ı sin(T π − T +ı ) −
T cos(T π − T +ı ) sin T +ı
= 2 sin
T +ı sin (T π − TT +ı
3 ıπ
)
Tπ T ıπ T ıπ Tπ
T π cos T +ı (−1)T +1 sin T +ı + T (−1)T +1 cos T +ı sin T +ı
= 2 sin
T +ı (−1)T +1 sin3 TT +ı
ıπ

ıπ T ıπ T ıπ ıπ
ıπ − cos T +ı sin T +ı + T cos T +ı sin T +ı
= 2 sin . (A.12)
T +ı sin3 TT +ı
ıπ

1
When 4 < ı < 1, − cos Tıπ+ı sin TT +ı
ıπ
+ T cos TT +ı
ıπ
sin Tıπ+ı < 0 has a constant order. It follows that
there is a constant  > 0 such that

g20 (θ)T < − (A.13)

when 1
4 < ı < 1. This, together with κ(θ) = 1 + op (1), κ0 (θ) = Op (m1 ) and g2 (θ) = Op ( T12 ), implies
that there is a constant 1 > 0 such that

T f 0 (θ) < −1 . (A.14)

Then the solution is unique.

17
Now we prove (2.14). Since θ̂ ∈ ( TT+1
π
, π) is the solution of h(θ) = 0, we can rewrite (2.12) as

√ ρ1 f (θϕ,1 ) − ρ1 f (θ̂)
p → N (0, 1). (A.15)
S̀σ2 ,m2

This leads to f (θϕ,1 ) − f (θ̂) = Op ( T 21√p ). This, together with (A.14), implies that

1
θ̂ − θϕ,1 = Op ( √ ). (A.16)
T p

Then we use Taylor expansions. We only need to prove f 00 (θ̃) = Op (1) for any θ̃ between θϕ,1
and θ̂. Since (A.16) ensures that θϕ,1 and θ̂ are very close, it’s not hard. The proof of (2.15) is
similar. We ignore the details.

B Results for truncated matrices


This section is to consider the truncated version of the sample covariance matrix. The overall
strategy of the proof, which needs more delicate analysis, is similar to that of [31]. However for the
completeness of the proof and for facilitating the readers we repeat some necessary steps in [31] and
still use the same notation as in [31].
To this end, define
l
X
Yij,l = bk Zi−k,j
k=0
with l = max{p, T }, a truncated version of Ytj in (2.1). However, to simplify notation, we let bi = 0
for all i > l in this section, so that we still use Yij instead of Yij,l . In this way ai defined in (2.3)
and Ytj in (2.1) respectively become
l−i
X l
X
ai = bk bk+i , Ytj = bk Zt−k,j .
k=0 k=0

Furthermore let F = (Fij ) be a T × (T + l) matrix with



bl+i−j i ≤ j ≤ i + l,
Fij = (B.1)
0 otherwise.

It follows that Y = FZp , where Zp is a (T +l)×p random matrix with (Zp )i,j = Zi−l,j . For the sake
of notation simplicity, we below denote Zp by Z and (Zp )i,j by Zij . Let A = (Aij )T ×T = (a|i−j| )T ×T .
We then have A = FF0 . We would remind readers that l depends on T so that a|i−j| depends on T.
We also assume that x0 = 0 and δt = 0 in this section.
Note that we have studied the case ϕ = 1 in [31]. Below we deal with with the case of ϕ < 1.

B.1 Eigenvalues of Cϕ ACϕ ∗


Write
∗ ∗
B = (1/p)X(I − H)X∗ = (1/p)Cϕ YΣ1/2 (I − H)Σ1/2 Y∗ Cϕ = (1/p)Cϕ FZp Σ1/2 (I − H)Σ1/2 Z∗p F∗ Cϕ .
(B.2)

18
We below investigate the eigenvalues and eigenvectors of Cϕ FF∗ Cϕ ∗ = Cϕ ACϕ ∗ at first. These
are crucial steps.
Since it’s very hard to find the eigenvalues of Cϕ ACϕ ∗ directly we use the following strategy. At
first, we note that the eigenvalues of Cϕ ACϕ ∗ and AC∗ϕ Cϕ are the same. We obtain the eigenvalues
and eigenvectors of C∗ϕ Cϕ by first studying (C∗ϕ Cϕ )−1 . The next key step is to approximate the
eigenvalues of CAC∗ . Our results are summarized in the following series of Lemmas and Theorems.
The first two lemmas describe the eigenvalues of C∗ϕ Cϕ and decide their limits.

Lemma 4. Let λϕ,1 ≥ λϕ,2 ≥ · · · ≥ λϕ,T ≥ 0 be the eigenvalues of C∗ϕ Cϕ . We then have

1
λϕ,k = with π > θϕ,1 > · · · > θϕ,T > 0, (B.3)
2ϕ(1 + cos θϕ,k ) + (1 − ϕ)2
where π > θϕ,1 > · · · > θϕ,T > 0 are the solutions of the equation

ϕ sin T θ + sin(T + 1)θ = 0. (B.4)

Lemma 5. (Lemma C.2 in [31]) Using the notation in Lemma 4,


λ1,k 4
lim 2
= 2 (B.5)
T →∞ T π (2k − 1)2
for any fixed k.

Lemma 6. Using the notation in Lemma 4, under the assumption (A9),


λϕ,k
lim >0 (B.6)
T →∞ T 2

for any fixed k. Futhermore, when T is big enough there exists a independent constant Υ > 0 such
that
λ1,k − λϕ,k
lim ≤ ΥT (1 − ϕ) (B.7)
T →∞ λ1,k
for any k.

Lemma 7 below specifies the eigenvectors of C∗ϕ Cϕ .

Lemma 7. Let x̃k = (xk,1 , · · · , xk,T )0 be a T × 1 vector with

xk,i = (−1)T −i sin(T − i + 1)θϕ,k , −l ≤ i ≤ T + l. (B.8)

Then {x̃k , 1 ≤ k ≤ T } are orthogonal and satisfy for any k

C∗ϕ Cϕ x̃k = λϕ,k x̃k . (B.9)

Lemma 8.
T
ΣTj=1 (xk,j )2 = + o(T ).
2
Let
x̃k
ỹk = . (B.10)
kx̃k k
Then {ỹk }1≤k≤T are orthogonal and the jth element of ỹk , yk,j , satisfies

|xk,j | 1
|yk,j | = = O( √ ). (B.11)
kx̃k k T

19
Lemma 9 gives the approximation to the eigenvalues of AC∗ϕ Cϕ .

Lemma 9. Define γϕ,k by


X
γϕ,k = λϕ,k (a0 + 2 aj (−1)j cos(jθϕ,k )). (B.12)
1≤j≤T −1

When T (1 − ϕ) is bounded, for any fixed constant k ≥ 1, there is a constant cϕ,k such that
γϕ,k
lim = cϕ,k > 0. (B.13)
T →∞ T 2

Let βϕ,1 ≥ βϕ,2 ≥ · · · ≥ βϕ,T be the eigenvalues of AC∗ϕ Cϕ . If A satisfies the assumptions (A1)
and (A2), then for any fixed integers i ≥ 1 and j ≥ 1 the following holds
βϕ,i − γϕ,i
| | = O(T −1 ). (B.14)
γϕ,j
For any  > 0 there exists T0 and k0 where k0 is a fixed number independent of T such that when
T ≥ T0 and k ≥ k0 ,
βϕ,k
| | ≤ . (B.15)
γϕ,1
Lemma 10. (Lemma C.5 in [31]) Suppose that A satisfies the assumptions (A1) and (A2). Then
(T + 1)T X
tr(AC∗1 C1 ) = a0 + aj (T − j + 1)(T − j)
2
1≤j≤T −1

and
β1,k γ1,k 8
lim = lim = 2 .
T →∞ tr(AC∗1 C1 ) T →∞ tr(AC∗1 C1 ) π (2k − 1)2
Lemma 11. (Lemma C.6 in [31]) Suppose that A satisfies the assumptions (A1) and (A2). For
any  > 0, we can find T0 and k0 , where k0 is a finite number independent of T , such that when
T ≥ T0 , P
k>k0 β1,k
| | < . (B.16)
γ1,1
Lemma 12. Suppose that A satisfies the assumptions (A1) and (A2). For any  > 0, we can find
T0 and k0 , where k0 is a finite number independent of T , such that when T ≥ T0 ,
P
βϕ,k
| k>k0 | < . (B.17)
γϕ,1
Proof of Lemma 4:
Let MT = (C∗ϕ Cϕ )−1 . Define the characteristic function of MT by gT (λ) = det(λIT − MT ).
We can verify that the entries of the inverse matrix Cϕ −1 , a T × T lower triangular matrix, are of
the form 


1 i = j,

−1
Cϕ,ij = −ϕ i = j + 1,



0 otherwise.

It follows that Mi,j , the elements of MT = (C∗ϕ Cϕ )−1 , satisfy



1 i = j = 1,





1 + ϕ2 i = j > 1,

Mij =



−ϕ |i − j| = 1,


0

otherwise.

20
By the cofactor expansion we obtain a recurrence relation as following

gT (λ) = (λ − 1 − ϕ2 )gT −1 (λ) − ϕ2 gT −2 (λ). (B.18)

Consider λ ∈ (0, 4) at first. Hence we may write λ = λ(θ) = 1 + ϕ2 + 2ϕ cos θ. We can further
solve (B.18) to get
ϕT +1 sin T θ + ϕT sin(T + 1)θ
gT (λ) = .
sin θ
6 0, gT (λ) = 0 is equivalent to
When sin θ =

ϕ sin T θ + sin(T + 1)θ = 0.

Let hT (θ) = ϕ sin T θ + sin(T + 1)θ. For any 1 ≤ k ≤ T ,


kπ kT π k(T + 1)π 2kπ
hT ( ) = ϕ sin + sin = (−1)k (1 − ϕ) sin (B.19)
T + 1/2 T + 1/2 T + 1/2 2T + 1
and
kπ kT π (T + 1 − k)π
hT ( ) = ϕ sin = ϕ(−1)k−1 sin . (B.20)
T +1 T +1 T +1
Then ∃θϕ,T +1−k ∈ [ Tkπ kπ
+1 , T +1/2 ] such that hT (θϕ,k ) = 0. So (B.3) gives T different solutions which
satisfy hT (θ) = 0 and sin θ 6= 0. On the other hand, observe that there are at most T solutions for
gT (λ) = 0. The proof of (B.3) is complete.

Proof of Lemma 6: Recalling the proof of Lemma 4, θϕ,k ∈ [ (T +1−k)π


T +1 , (TT+1−k)π
+1/2 ). Then we can find
that both T 2 (1 + cos θϕ,k ) and T 2 (1 − ϕ)2 are bounded for any fixed k. So we can proof (B.6).
We next prove (B.7), Write

1 1 (1 − ϕ)2 + 2(ϕ − 1)(1 + cos θϕ,k ) + 2(cos θϕ,k − cos θ1,k )


λ1,k −λϕ,k = − = .
2 + 2 cos θ1,k 1 + ϕ2 + 2ϕ cos θϕ,k (2 + 2 cos θ1,k )(1 + ϕ2 + 2ϕ cos θϕ,k )
(B.21)
For any k, we define
(T + 1 − k)π
$k = θ1,k − θϕ,k = − θϕ,k . (B.22)
T + 1/2

Then 0 < $T −k+1 ≤ T (2T +1) ,

$k $k
λ1,k − λϕ,k (1 − ϕ)2 + 2(ϕ − 1)(1 + cos θϕ,k ) + 4 sin(θ1,k − 2 ) sin 2
= . (B.23)
λ1,k 1 + ϕ2 + 2ϕ cos θϕ,k
Note that
(1 − ϕ)2 λ1,1
0< ≤ 2 T 2 (1 − ϕ)2 , (B.24)
1+ ϕ2
+ 2ϕ cos θϕ,k T
and
(ϕ − 1)(1 + cos θϕ,k ) (1 + cos θϕ,k ) 1−ϕ
| 2
| = (1 − ϕ)| 2
|≤ . (B.25)
1 + ϕ + 2ϕ cos θϕ,k (1 − ϕ) + 2ϕ(1 + cos θϕ,k ) 2ϕ
From (B.4), (B.19) and (B.20), we can find that
θϕ,k
(1 − ϕ) sin T θϕ,k = sin T θϕ,k + sin(T + 1)θϕ,k = 2 sin(T + 1/2)θϕ,k cos .
2
It follows that
(T + 1 − k)π (T + 1 − k)π θϕ,k
(1 − ϕ) sin T ( − $k ) = 2 sin(T + 1/2)( − $k ) cos .
T + 1/2 T + 1/2 2

21
Then
T +1−k θϕ,k
(1 − ϕ) sin((T + 1 − k)π − π − T $k ) = 2 sin((T + 1 − k)π − (T + 1/2)$k ) cos .
2T + 1 2

T +1−k θϕ,k
(1 − ϕ) sin( π + T $k ) = 2 sin(T + 1/2)$k cos .
2T + 1 2
Then we can find that
T +1−k T +1−k θϕ,k
(1 − ϕ) sin( π) cos T $k + (1 − ϕ) sin T $k cos( π) = 2 sin(T + 1/2)$k cos .
2T + 1 2T + 1 2
θϕ,k
Recall that 0 < $T −k+1 ≤ kπ
T (2T +1) and θϕ,k ∈ [ (T +1−k)π
T +1 , (TT+1−k)π
+1/2 ]. Then sin(T + 1/2)$k cos 2 6=
0. It follows that

sin T2T
+1−k
+1 π cos T $k sin T $k cos T2T
+1−k
+1 π
(1 − ϕ) θϕ,k
+ (1 − ϕ) θϕ,k
= 1.
2 sin(T + 1/2)$k cos 2 2 sin(T + 1/2)$k cos 2
T +1−kπ
Note that 0 < T $k < (T + 1/2)$k ≤ 2T ≤ π2 , and

sin T $k
0< < 1. (B.26)
sin(T + 1/2)$k

Similarly, since θϕ,k ∈ [ (T +1−k)π


T +1 , (TT+1−k)π
+1/2 ), we have

cos T2T
+1−k
+1 π cos T2T
+1−k
+1 π cos T2T
+1−k
+1 π
0< < θϕ,k
< = 1. (B.27)
cos (T 2T
+1−k)π
+2 cos 2 cos (T 2T
+1−k)π
+1

(B.26)-(B.27) imply that

sin T $k cos T2T


+1−k
+1 π
(1 − ϕ) θϕ,k
= O(1 − ϕ).
2 sin(T + 1/2)$k cos 2

This, together with (B.1), implies that

sin T2T
+1−k
+1 π cos T $k
(1 − ϕ) θϕ,k
= 1 + O(1 − ϕ).
2 sin(T + 1/2)$k cos 2

Then
T +1−k
$k $k sin(T + 1/2)$k sin 2T +1 π
tan T $k cos + sin = = (1 − ϕ) θ
(1 + O(1 − ϕ)). (B.28)
2 2 cos T $k 2 cos ϕ,k 2

(T −k+1)π π
Since 0 < $k ≤ T (2T +1) < 2T ,

$k (T − k + 1)2
cos = 1 + O( ) (B.29)
2 T4
and
$k tan T $k
sin = O( ). (B.30)
2 T
(B.28)-(B.30) imply that
T +1−k
1 sin 2T +1 π
tan T $k (1 + O( )) = (1 − ϕ) θ
(1 + O(1 − ϕ)). (B.31)
T 2 cos ϕ,k 2

22
Since T (1 − ϕ) is finite, we have

sin T2T
+1−k
+1 π 1
tan T $k = (1 − ϕ) θϕ,k
(1 + O( )). (B.32)
2 cos T
2

Similarly, since θϕ,k ∈ [ (T +1−k)π


T +1 , (TT+1−k)π
+1/2 ), we conclude that

sin T2T
+1−k
+1 π 1
0< θϕ,k
< (T +1−k)π
≤ ∞. (B.33)
2 Tk cos 2 2 Tk cos 2T +1

This, together with (B.32), implies that


T
tan T $k = O( (1 − ϕ)). (B.34)
k
It follows that
1−ϕ
$k = O( ). (B.35)
k
(B.23)-(B.25), together with the assumption (A8), (B.35) and (B.6), imply that when T is big
enough there exists a constant Υ > 0 such that

λ1,k − λϕ,k
≤ ΥT (1 − ϕ) (B.36)
λ1,k
for any k.

Lemmas 7 can be verified with some straightforward computations and the simple fact that

sin(k + j)θ + sin(k − j)θ = 2 sin kθ cos jθ. (B.37)

We ignore the details here.

Proof of Lemma 8: From (B.8) we obtain

|xk,j | ≤ 1.

Lemma 8 can be then proved with some straightforward computations. We ignore details here.

Proof of Lemma 9:
Let’s prove (B.13) at first. Note that
X X X X
|(a0 +2 aj (−1)j cos(jθϕ,k ))−(a0 +2 aj )| ≤ 2 |aj ||cos(j(π−θϕ,k ))−1|+2 |aj |.
1≤j≤T −1 1≤j≤∞ 1≤j≤T −1 T ≤j
(B.38)
π π
For a fixed k, we can find a jk to satisfy 3 ≤ j(π − θϕ,k ) ≤ 2. It follows that
X 1 X
2 |aj ||cos(j(π − θϕ,k )) − 1| ≤ |aj |j 2 (π − θϕ,k )2
2
1≤j≤jk 1≤j≤jk
jk (π − θϕ,k )2 X
≤ j|aj |
2
1≤j≤jk
kπ 2 X
≤ j|aj | (B.39)
4(T + 1)
1≤j≤∞

23
and that
X X X
2 |aj ||cos(j(π − θϕ,k )) − 1| + 2 |aj | ≤ 4 |aj |
jk <j≤T −1 T ≤j j≥jk
X 3(2k − 1) X
≤ jk−1 4 j|aj | ≤ j|aj |. (B.40)
2T + 1
j≥jk 1≤j≤∞

The assumption (A1) implies that



X
i|ai | < ∞. (B.41)
i=0
From the assumption (A2), (B.41) and truncation conditions we can find
X X X∞
lim (a0 + 2 aj (−1)j cos(jθϕ,k )) = lim (a0 + 2 aj ) = ( bi )2 = s2 > 0. (B.42)
T →∞ T →∞
1≤j≤T −1 1≤j≤∞ i=0

In view of (B.5), (B.12) and (B.42), we have (B.13).


Now we consider the eigenvalues of AC∗ϕ Cϕ . From Lemmas 4-8, we can write C∗ϕ Cϕ as
1/2 1/2
Vϕ∗ Λϕ Λϕ Vϕ , where

Λϕ = diag{λϕ,1 , · · · , λϕ,T }, Vϕ∗ = (ỹ1 , · · · , ỹT ). (B.43)


1/2 1/2
Then we just need to consider the eigenvalues of Ma,ϕ = Λϕ Vϕ AVϕ∗ Λϕ . Define the (i,j) entry
of Ma,ϕ as M̀i,j . Then
1/2 1/2
1/2 1/2 λϕ,i λϕ,j
M̀i,j = λϕ,i λϕ,j ỹi∗ Aỹj = x̃∗i Ax̃j ,
kx̃i kkx̃j k
and
T
X T
X −1 T
X −h T
X
x̃∗i Ax̃j = a0 xih xjh + ah ( xif xj,f +h + xif xj,f −h ).
h=1 h=1 f =1 f =h+1

Note that xj,f −h + xj,f +h = 2 ∗ (−1)h xjf cos(hθϕ,j ). Then


T
X −1 XT T
X −1 T
X h
X
x̃∗i Ax̃j = (a0 + 2 h
ah (−1) cos(hθϕ,j ))( xih xjh ) − ah ( xif xj,f +h + xif xj,f −h ).
h=1 h=1 h=1 f =T −h+1 f =1
(B.44)
PT
Note that h=1 xih xjh = 1{i = j}kx̃i kkx̃j k and
T
X −1 T
X h
X T
X −1
| ah ( xif xj,f +h + xif xj,f −h )| ≤ 2 h|ah |
h=1 f =T −h+1 f =1 h=1

. Then we can write

Ma,ϕ = Ma,ϕ,R + diag{γϕ,1 , · · · , γϕ,T } = Ma,ϕ,R + Λa,ϕ,M . (B.45)

T X
T T −1
X λϕ,i λϕ,j X λ2ϕ,1
kMa,ϕ,R k22 ≤ ∗ (2 h|ah |)2
= O( ). (B.46)
kx̃i k2 kx̃j k2 T2
i=1 j=1 h=1
This, together with (B.13), leads to (B.14).

Proof of Lemma 12: From (B.7) we can find that


X X
| βϕ,k | < (1 + ΥT (1 − ϕ))| β1,k |. (B.47)
k>k0 k>k0

This, together with (B.6) and (B.16), implies (B.17).

24
B.2 Eigenvectors of Cϕ AC∗ϕ
This section is to investigate the eigenvectors of Cϕ AC∗ϕ . Recalling (B.10), we normalize {x̃k }1≤k≤T
to get {ỹk }1≤k≤T . Then we study the eigenvectors of of AC∗ϕ Cϕ by representing them with
{ỹk }1≤k≤T . At last we give some result about the eigenvectors of Cϕ AC∗ϕ which is necessary
for the future proof. Our results are the following.

Lemma 13. Let {uk }1≤k≤T be orthogonal and real vectors such that kuk k = 1 and

Cϕ AC∗ϕ uk = βϕ,k uk . (B.48)

C−1
ϕ uk
Define fk = kC−1
such that
ϕ uk k

fk = ΣTj=1 αkj yj (B.49)

with
ΣTj=1 αkj
2
= 1. (B.50)

Then when k ≥ 1 is fixed,


2 λ
αkk ϕ,k
= 1 + O(T −1 ), (B.51)
ΣTj=1 αkj
2 λ
ϕ,j

where {λϕ,j } are given in Lemma 4.


F∗ C∗ϕ uk
Lemma 14. Let (Sk,1 , · · · , Sk,T +l )0 = sk = √
γϕ,1 . Then {sk }1≤k≤T are orthogonal and

T +l
X
4
Sk,j = O(T −1 ). (B.52)
j=1

C−1
ϕ uk
Proof of Lemma 13: From fk = kC−1 uk k
and (B.48), we have kfk k = 1 and

AC∗ϕ Cϕ fk = βϕ,k fk . (B.53)

From (B.45) and (B.53), we have

1/2 1/2
fk∗ C∗ϕ Cϕ AC∗ϕ Cϕ fk fk∗ Vϕ∗ Λϕ (Λa,ϕ,M + Ma,ϕ,R )Λϕ Vϕ fk
βϕ,k = = .
kCϕ fk k2 kCϕ fk k2
It follows that
1/2 1/2 1/2 1/2
|fk∗ Vϕ∗ Λϕ Λa,ϕ,M Λϕ Vϕ fk | − |fk∗ Vϕ∗ Λϕ Ma,ϕ,R Λϕ Vϕ fk |
≤ βϕ,k (B.54)
kCϕ fk k2

and
1/2 1/2 1/2 1/2
|f ∗ Vϕ∗ Λϕ Λa,ϕ,M Λϕ Vϕ fk | + |fk∗ Vϕ∗ Λϕ Ma,ϕ,R Λϕ Vϕ fk |
βϕ,k ≤ k . (B.55)
kCϕ fk k2
By (B.9), (B.10), (B.43) and (B.49), we have
q
kCϕ fk k = ΣTj=1 αkj
2 λ
ϕ,j . (B.56)

(B.9), (B.10), (B.45) and (B.49) imply that

fk∗ Vϕ∗ Λ1/2 1/2 T 2


ϕ Λa,ϕ,M Λϕ Vϕ fk = Σj=1 αkj γϕ,j λϕ,j . (B.57)

25
From (B.46), we have
1/2 1/2
|fk∗ Vϕ∗ Λϕ Ma,ϕ,R Λϕ Vϕ fk |
≤ kMa,ϕ,R k2 = O(T ).
kCϕ fk k2

This, together with (B.54)-(B.57), implies that

ΣTj=1 αkj
2 γ
ϕ,j λϕ,j ΣTj=1 αkj
2 γ
ϕ,j λϕ,j
− O(T ) ≤ βϕ,k ≤ + O(T ).
ΣTj=1 αkj
2 λ
ϕ,j ΣTj=1 αkj
2 λ
ϕ,j

By Lemma 9, for any fixed k we have


2 λ
αkj 2 λ
αkj
ϕ,j γϕ,j ϕ,j γϕ,j
ΣTj=1 − O(T −1 ) ≤ 1 ≤ ΣTj=1 T + O(T −1 ). (B.58)
ΣTj=1 αkj
2 λ
ϕ,j βϕ,k 2
Σj=1 αkj λϕ,j βϕ,k

Note that {uk }1≤k≤T are orthogonal and {ỹk }1≤k≤T are orthogonal. When k 6= m, from (B.9),
(B.10) and (B.49) we have

fk∗ C∗ϕ Cϕ fm ΣTj=1 αkj αmj λϕ,j


0 = u∗k um = = .
kCϕ fk kkCϕ fm k kCϕ fk kkCϕ fm k

This implies that


ΣTj=1 αkj αmj λϕ,j = 0. (B.59)

αkj λϕ,j
Moreover let υkj = q . We have
ΣT 2
j=1 αkj λϕ,j

ΣTj=1 υkj
2
= 1. (B.60)

Note that (B.58) is equivalent to


γϕ,j 2 γϕ,j
ΣTj=1 υkj
2
− O(T −1 ) ≤ 1 ≤ ΣTj=1 υkj + O(T −1 ). (B.61)
βϕ,k βϕ,k

Also (B.59) implies that


ΣTj=1 υkj υmj = 0. (B.62)

We consider υkj for fixed k below. When k = 1 and T is big enough, Lemma 9, (B.60) and
(B.61) imply

γϕ,j 2 βϕ,1 − γϕ,2 2 |βϕ,1 − γϕ,1 |


O(T −1 ) = |1 − ΣTj=1 υ1j
2
| ≥ (1 − υ11 ) − υ11 . (B.63)
βϕ,1 βϕ,1 βϕ,1
βϕ,1 −γϕ,1
In view of (B.13)-(B.14), we have βϕ,1 = O(T −1 ). Recalling that θϕ,k ∈ [ (T +1−k)π
T +1 , (TT+1−k)π
+1/2 ],
β 2 −1 ) and
we can find that | βϕ,1 ϕ,1
−γϕ,2 | is bounded. It follows that (B.63) implies that υ11 = 1 + O(T
2 = O(T −1 ). From (B.62), for any k 6= 1 we have
ΣTj=2 υ1j
q q
|υk1 υ11 | = |ΣTj=2 υkj υ1j | ≤ ΣTj=2 υkj
2 2 = O(T −1/2 ).
ΣTj=2 υ1j (B.64)

2 = O(T −1 ). It’s similar to obtain that υ 2 = 1 + O(T −1 ) and υ 2 = O(T −1 ) for any
This implies υk1 22 k2
k 6= 2.
2 = 1 + O(T −1 ) for any fixed k. This implies
By repeating these steps we conclude that υkk
(B.51).

26
Proof of Lemma 14: Note that {sk }1≤k≤T are orthogonal and real due to orthogonality of {uk }1≤k≤T .
We conclude from (B.9) and (B.49) that

F∗ C∗ϕ Cϕ fk 1
sk = √ =√ ΣT αkj λϕ,j F∗ ỹj = sk,M + sk,R , (B.65)
γϕ,1 kCϕ fk k γϕ,1 kCϕ fk k j=1

where
1 1
sk,M = √ αkk λϕ,k F∗ ỹk , sk,R = √ Σj6=k αkj λj F∗ ỹj . (B.66)
γϕ,1 kCϕ fk k γϕ,1 kCϕ fk k

By Hölder’s inequality, we have


1 1 q
ksk,R k = k √ Σj6=k αkj λϕ,j F∗ ỹj k ≤ √ 2 λ2 .
kFk2 Σj6=k αkj ϕ,j
γϕ,1 kCϕ fk k γϕ,1 kCϕ fk k

Recalling A = FF∗ , we have


p
kFk2 = kAk2 .

Since A is a Hermitian Toeplitz matrix, from [18],


X
kAk2 ≤ 2 |ak |.
0≤k≤l

By (B.41) we can get


p
kFk2 = kAk2 < ∞.

From Lemma 5, (B.51), and (B.56) we can obtain that for any fixed k,
q v
2 λ2
Σj6=k αkj u 2 λ
ϕ,j u Σj6=k αkj ϕ,j p
≤ t
T 2
λϕ,1 = O(T 1/2 ).
kCϕ fk k Σj=1 αkj λϕ,j

This, together with (B.13), implies that for any fixed k,

ksk,R k = O(T −1/2 ). (B.67)

1
Similarly, we can also obtain that √
γ1 kCfk k αkk λk is bounded for any fixed k.
Let Sk,M,j be the jth element of sk,M and Sk,R,j be the jth element of sk,R . From (B.11), (B.51)
and (B.66) and the assumption (A1) we can obtain that for any fixed k,
l
1 2 X
|Sk,M,j | ≤ √ |αkk |λk √ |bh | = O(T −1/2 ). (B.68)
γϕ,1 kCϕ fk k 2T + 1 h=0

It follows from (B.65)-(B.68) that for any fixed k,


T +l
X T +l
X
4 4 4
Sk,j ≤8 (Sk,R,j + Sk,M,j )
j=1 j=1
 2
T +l
X XT +l
≤8 4
Sk,M,j + 8 2
Sk,R,j  = O(T −1 ). (B.69)
j=1 j=1

27
B.3 The proof of Lemma 1 for truncated matrices
The proof of the remaining part of the Lemma 1 for truncated matrices is the same as that of the
part C.2.3-C.2.4 in [31].

C The proof of Lemmas 1-3


C.1 The proof of Lemma 1

We can deal with p1 Cϕ YΣ1/2 (I − H)Σ1/2 Y∗ C∗ϕ + p1 Cϕ YΣ1/2 (I − H)X́0 Cϕ,0 + p1 Cϕ,0 X́0 (I − H)Σ1/2 Y∗ C∗ϕ +
1
p Cϕ,0 X́0 (I − H)X́∗0 C∗ϕ,0 by the same methods of Appendix B in [31]. For the part about ∆, as-
| √1p ∆(I−H)∆∗ |2 √1 trace(∆(I−H)∆∗ )
p
sumption A8 ensures that T2
≤ T2
= op (1). The other part is similar.

C.2 The proof of Lemmas 2-3


At first we consider the impact of δt . Recalling the definition of µ̀m , we write

T m T −j
X x̀i,i XX x̀i,i+j
µ̀m1 = +2 (C.1)
p(T − 1) p(T − j − 1)
i=2 j=1 i=2
T
X (xi − δi − xi−1 + δi−1 )(I − H)0 (xi − δi − xi−1 + δi−1 )
=
p(T − 1)
i=2
m1 T −j
X X (xi − δi − xi−1 + δi−1 )(I − H)0 (xi+j − δi+j − xi+j−1 + δi+j−1 )
+2
p(T − j − 1)
j=1 i=2
T
X 2(xi − δi − xi−1 + δi−1 )(I − H)0 (δi − δi−1 )
+
p(T − 1)
i=2
m1 T −j
X X (δi − δi−1 )(I − H)0 (xi+j − δi+j − xi+j−1 + δi+j−1 )
+2
p(T − j − 1)
j=1 i=2
m1 T −j
X X (xi − δi − xi−1 + δi−1 )(I − H)0 (δi+j − δi+j−1 )
+2
p(T − j − 1)
j=1 i=2
T
X (δi − δi−1 )(I − H)0 (δi − δi−1 )
+
p(T − 1)
i=2
m1 T −j
X X (δi − δi−1 )(I − H)0 (δi+j − δi+j−1 )
+2 .
p(T − j − 1)
j=1 i=2

Note that

lim P (m ≤ min{q, p}) = 1. (C.2)
p,T →∞

28
Moreover
T m1 T −j
X (δi − δi−1 )(I − H)0 (δi − δi−1 ) X X (δi − δi−1 )(I − H)0 (δi+j − δi+j−1 )
| +2 | (C.3)
p(T − 1) p(T − j − 1)
i=2 j=1 i=2
T 0
(δi − δi−1 )(I − H) (δi − δi−1 )
X
≤ (2m1 + 1)
p(T − m1 − 1)
i=2
T √ !
1 X min{q, p}(δi − δi−1 )(I − H)0 (δi − δi−1 ) 1
=√ O √ = op ( √ ),
p pT p
i=2

which is ensured by Assumption A11. The other parts involving (δi − δi−1 ) are similar. Note that
assumptions A8 and A10-A11 ensure that δt does not change the asymptotic properties of the
estimators. In the following part of this subsection we only consider the case when δt = 0.

Proof of Lemma 2: At first by (2.2) we write (xt − xt−1 ) as follows


t−2
X
1/2
xt − xt−1 = Σ (yt + (ϕ − 1) ϕk yt−k−1 + (ϕ − 1)ϕt−1 x0 ). (C.4)
k=0

Recalling the assumptions (A7) and (A8) one can find that Ek(ϕ − 1)ϕt−1 x0 k2 → 0. Below for
facilitating statements we ignore the term and write
t−2
X
xt − xt−1 = Σ1/2 (yt + (ϕ − 1) ϕk yt−k−1 ). (C.5)
k=0

Recalling the notation x̆f,g below (2.11) we can write Ex̆t,t−j when j ≥ 1 as

Ex̀t,t−j = E (xt − xt−1 )0 (I − H) (xt−j − xt−j−1 )


t−2
X t−j−2
X
k 0 1/2 1/2
= E[(yt + (ϕ − 1) ϕ yt−k−1 ) Σ (I − H)Σ (yt−j + (ϕ − 1) ϕk yt−j−k−1 )]
k=0 k=0
 
, tr(Σ1/2 (I − H)Σ1/2 ) aj + A1tj + A2tj + A3tj + A4tj + A5tj + A6tj ,

where
j−1
X t−2
X t−j−2
X
A1tj = (ϕ − 1) ϕk aj−k−1 , A2tj = (ϕ − 1) ϕk ak+1−j , A3tj = (ϕ − 1) ϕk ak+j+1 ,
k=0 k=j k=0

j−1 t−j−2
X X t−2
X t−j−2
X
A4tj = (ϕ − 1)2 ϕk1 +k2 ak2 +j−k1 , A5tj = (ϕ − 1)2 ϕk1 +k2 ak2 +j−k1 ,
k1 =0 k2 =0 k1 =j k2 =k1 −j+1

and
1 −j
t−2 kX
X
A6tj = (ϕ − 1)2 ϕk1 +k2 ak1 −k2 −j .
k1 =j k2 =0

Hence
T
1 X
Ex̀t,t−j
p(T − j − 1)
t=j+2
T
tr(Σ1/2 (I − H)Σ1/2 ) 1 X
= (aj + A1tj + A2tj + A3tj + A4tj + A5tj + A6tj(C.6)
).
p (T − j − 1)
t=j+2

29
We next analyze these term by term.
The terms Aktj , k = 1, · · · , 6 can be rewritten as follows:
T
X T
X t−1−j
X
A2tj = (ϕ − 1) ϕk+j−1 ak (C.7)
t=j+2 t=j+2 k=1
−1−j
TX T
X −1−j
TX
k+j−1
= (ϕ − 1) ϕ ak = (ϕ − 1) ϕk+j−1 (T − j − k)ak ;
k=1 t=j+1+k k=1

T
X T
X t−1
X T
X −1
A3tj = (ϕ − 1) ϕk−j−1 ak = (ϕ − 1) ϕk−j−1 (T − k)ak ; (C.8)
t=j+2 t=j+2 k=j+1 k=j+1

T T −2 T −2
X 1−ϕ X |j−k2 | 1 X
ak2 ϕ2+|k2 −j| − ϕmin{2+k2 +j,2T −k2 −j}

A4tj = ak2 ϕ (T − 1 − max{k2 , j}) − 2
1+ϕ (1 + ϕ)
t=j+2 k2 =1 k2 =1
T −2
1−ϕ X
− ak2 ϕk2 +j max{T − k2 − j − 1, 0}; (C.9)
1+ϕ
k2 =1

T T −j−2 −j−2
TX
X 1−ϕ X 1 2T −k2 −j
ak2 ϕj+k2 (T − 1 − j − k2 ) − 2+k2 +j

A5tj = ak2 ϕ − ϕ (C.10)
;
1+ϕ (1 + ϕ)2
t=j+2 k2 =1 k2 =1

and
T T −j−2 −j−2
TX
X 1−ϕ X 1
ak2 ϕj+k2 (T − 1 − j − k2 ) − ak2 ϕ2+k2 +j − ϕ2T −k2 −j (C.11)

A6tj = 2
.
1+ϕ (1 + ϕ)
t=j+2 k2 =0 k2 =0

From (C.6)-(C.11) we conclude that


T
1 X Ex̀t,t−j
(T − j − 1)
t=j+2
tr(Σ (I − H)Σ1/2 )
1/2

−j−2
TX
1−ϕ 1 ϕ2 − ϕ2T −2k 
= ak (1{j = k} 1 + −
1 + ϕ (1 + ϕ)2 T − j − 1
k=0
1 − ϕ j−k 1 ϕ2+j−k − ϕ2T −k−j 
+1{k < j} (ϕ − 1)ϕj−1−k + ϕ −
1+ϕ (1 + ϕ)2 T −j−1
T −k 1 − ϕ k−j T − k − 1 1 ϕ2+k−j − ϕ2T −k−j
+1{k > j} (ϕ − 1)ϕk−1−j + ϕ −
T −1−j 1+ϕ T −1−j (1 + ϕ)2 T −j−1
T − k − j 1 − ϕ k+j T − k − 1 − j 1 ϕ 2+k+j − ϕ2T −k−j 
+(ϕ − 1)ϕk−1+j + ϕ −
T −1−j 1+ϕ T −1−j (1 + ϕ)2 T −j−1
T
X −1

+O (ϕ − 1) |ak | . (C.12)
k=T −j−1

Similarly Ex̀t,t can be written as

Ex̀t,t = E (xt − xt−1 )0 (I − H) (xt − xt−1 )


t−2
X t−2
X
k 0 1/2 1/2
ϕk yt−k−1

= E yt + (ϕ − 1) ϕ yt−k−1 ) Σ (I − H)Σ (yt−j + (ϕ − 1)
k=0 k=0
t−2 t−2
X 1−ϕ X
= tr(Σ1/2 (I − H)Σ1/2 ) a0 + 2(ϕ − 1) ϕk ak+1 + 2 ak (ϕk − ϕ2t−k−2 )
1+ϕ
k=0 k=1
1−ϕ
a0 (1 − ϕ2t−2 ) .

+ (C.13)
1+ϕ

30
It follows that
T
1 X
Ex̀t,t
p(T − 1)
t=2
T −1 T −2
tr(Σ1/2 (I − H)Σ1/2 ) X 2(ϕ − 1)ϕk−1 (T − k) 1−ϕ X k T −k−1
= a0 + ak + 2 ϕ ak
p T −1 1+ϕ T −1
k=1 k=1
T −2
1 X ϕ2+k − ϕ2T −k 1−ϕ ϕ2 − ϕ2T 
−2 ak + a0 − a0 . (C.14)
(1 + ϕ)2 T −1 1+ϕ 2
(1 + ϕ) (T − 1)
k=1

Write E µ̀m as

tr(Σ1/2 (I − H)Σ1/2 ) X
E µ̀m = κj aj . (C.15)
p
j=0

From (C.12) and (C.14) we conclude that


m
1−ϕ ϕ2 − ϕ2T X
j−1 1−ϕ j 1 ϕ2+j − ϕ2T −j 
κ0 = 1 + − + 2 (ϕ − 1)ϕ + ϕ −
1 + ϕ (1 + ϕ)2 (T − 1) 1+ϕ (1 + ϕ)2 T − j − 1
j=1
m m
2ϕm ϕ2 − ϕ2T X 1 ϕ2+j − ϕ2T −j X 1 j(ϕ2+j − ϕ2T −j )
= − −2 + 2
1 + ϕ (1 + ϕ)2 (T − 1) (1 + ϕ)2 T −1 (1 + ϕ)2 (T − 1)(T − j − 1)
j=1 j=1
m
2ϕm ϕ2
− ϕ2T X 1 ϕ2+j− ϕ2T −j m2 (1 − ϕ) 
= − −2 +O , (C.16)
1+ϕ (1 + ϕ)2 (T − 1) (1 + ϕ)2 T −1 T
j=1

and
m m
X 2ϕ X 1−ϕ
1+2 (ϕ − 1)ϕj−1 < κ0 ≤ 1 + (ϕ − 1)ϕj−1 + . (C.17)
1+ϕ 1+ϕ
j=1 j=1

It follows that

|1 − κ0 | = O m(1 − ϕ) . (C.18)

31
Similarly, for 1 ≤ f ≤ m,
T −f 1−ϕ fT −f −1 ϕ2+f − ϕ2T −f
κf /2 = (ϕ − 1)ϕf −1 + ϕ −
T −1 1+ϕ T −1 (1 + ϕ)2 (T − 1)
1−ϕ ϕ2 − ϕ2T −2f
+1 + −
1 + ϕ (1 + ϕ)2 (T − f − 1)
m
X  ϕ2+j−f − ϕ2T −j−f

j−1−f 1 − ϕ j−f 1
+ (ϕ − 1)ϕ + ϕ −
1+ϕ (1 + ϕ)2 T −j−1
j=f +1
f −1 
ϕ2+f −j − ϕ2T −j−f

X
f −1−j T −f 1 − ϕ f −j T − f − 1 1
+ (ϕ − 1)ϕ + ϕ −
T −1−j 1+ϕ T −1−j (1 + ϕ)2 T −j−1
j=1
m 
ϕ2+f +j − ϕ2T −j−f

X T − f − j 1 − ϕ f +j T − f − j − 1 1
+ (ϕ − 1)ϕf −1+j + ϕ − 2
T −1−j 1+ϕ T −1−j (1 + ϕ) T −j−1
j=1
m
X 1 ϕ2+|f −j| − ϕ2T −j−f m2 (1 − ϕ) 
− + O
(1 + ϕ)2 T −1 T
j=1

ϕm−f + ϕm+f − 2ϕm ϕ2+f − ϕ2 − ϕ2T −f + ϕ2T


= κ0 + −
1+ϕ (1 + ϕ)2 (T − 1)
m
X 1 ϕ2+j+f − ϕ2+j − ϕ2T −j−f + ϕ2T −j

(1 + ϕ)2 T −1
j=1
m
X 1 ϕ2+|f −j| − ϕ2+j − ϕ2T −j−f + ϕ2T −j m2 (1 − ϕ) 
− + O
(1 + ϕ)2 T −1 T
j=1
mf (1 − ϕ)  m2 (1 − ϕ) 
= κ0 + O +O
T T
= κ0 + o(m(1 − ϕ)). (C.19)

As for f ≥ m1 , we also have

|κj /2| = O(m(1 − ϕ)) (C.20)

and
∞ ∞
X maxj>m {|κj |} X
|κj ||aj | ≤ j|aj | = O(1 − ϕ). (C.21)
m+1
j=m+1 j=m+1

Then

tr(Σ1/2 (I − H)Σ1/2 ) X
E µ̀m = κ0 (a0 + 2 aj ) + o(m(1 − ϕ)) + O(1 − ϕ). (C.22)
p
j=1

Similarly, one can verify that


 
T 0 1/2 1/2 m T −j 0 1/2 1/2
X yi Σ (I − H)Σ yi X X yi Σ (I − H)Σ yi+j 
V ar(µm ) = O  +2 . (C.23)
p(T − 1) p(T − j − 1)
i=2 j=1 i=2

This, together with (2.66) in [31], implies

V ar(µm ) = o(p−1 ). (C.24)

Write

κ(θϕ,1 ) − κ0 (C.25)
m
ϕ3 − ϕ2T −m 2ϕm ϕ2 − ϕ2T X 1 ϕ2+j − ϕ2T −j m2 (1 − ϕ) 
= ϕm − m − + +2 + O .
2(T − 1) 1 + ϕ (1 + ϕ)2 (T − 1) (1 + ϕ)2 T −1 T
j=1

32
Note that
2ϕm 1−ϕ
ϕm − = ϕm = O(1 − ϕ), (C.26)
1+ϕ 1+ϕ

ϕ2 − ϕ2T
= O(1 − ϕ), (C.27)
(1 + ϕ)2 (T − 1)

m
ϕ3 − ϕ2T −m X 1 ϕ2+j − ϕ2T −j m2 (1 − ϕ) 
m −2 = O (C.28)
2(T − 1) (1 + ϕ)2 T −1 T
j=1

and
m2 (1 − ϕ) 
= o m(1 − ϕ) . (C.29)
T
(C.25)-(C.29) imply (A.4). (C.22)-(C.24) and (A.4) imply (A.5).

Proof of Lemma 3: Now we define

x̆t,s = (xt − xt−1 )0 (xs − xs−1 ) .

We also replace x̀t,s by x̆t,s in S̀σ2 ,0 and S̀σ2 ,m and get Sσ2 ,0 and Sσ2 ,m . In fact the differences
between S̀σ2 ,0 and Sσ2 ,0 are the ones between Σ1/2 (I − H)Σ1/2 and Σ. So we just need to study
Sσ2 ,0 and Sσ2 ,m then replace Σ by Σ1/2 (I − H)Σ1/2 .
. We rewrite (xt − xt−1 ) as follows
t−2
!
X
1/2 k
xt − xt−1 = Σ yt + (ϕ − 1) ϕ yt−k−1 (C.30)
k=0
∞ t−2 ∞
!
X X X
= Σ1/2 bs Zt−s + (ϕ − 1) ϕk bs Zt−1−k−s
s=0 k=0 s=0
t t−2 t−1−k
!
X X X
1/2 k
=Σ bt−s Zs + (ϕ − 1) ϕ bt−1−k−s Zs
s=−∞ k=0 s=−∞
t−1 t−1−s 0 t−2
!
X X X X
1/2 k k
 
=Σ b0 Zt + bt−s + (ϕ − 1) ϕ bt−1−k−s Zs + bt−s + (ϕ − 1) ϕ bt−1−k−s Zs .
s=1 k=0 s=−∞ k=0

Define a new variable b̆t,s as






 0 s > t,


b0

s = t,
b̆t,s = Pt−1−s k (C.31)



 bt−s + (ϕ − 1) k=0 ϕ bt−1−k−s 1 ≤ s ≤ t − 1,

bt−s + (ϕ − 1) t−2 ϕk bt−1−k−s
 P
s ≤ 0,

k=0

so that
t
X ∞
X
xt − xt−1 = Σ1/2 b̆t,s Zs = Σ1/2 b̆t,s Zs . (C.32)
s=−∞ s=−∞

33
Then
 2
p X
X p ∞
X ∞
X
x̆2f,g =  Σij b̆f,h1 b̆g,h2 Zh1 ,i Zh2 ,j  (C.33)
i=1 j=1 h1 =−∞ h2 =−∞
p
X p
X p
X p
X ∞
X ∞
X ∞
X ∞
X
= Σi1 j1 Σi2 j2 b̆f,h1 b̆g,h2 b̆f,h3 b̆g,h4 Zh1 ,i1 Zh2 ,j1 Zh3 ,i2 Zh4 ,j2 .
i1 =1 j1 =1 i2 =1 j2 =1 h1 =−∞ h2 =−∞ h3 =−∞ h4 =−∞

Write Sσ2 ,0 = Sσ2 ,0,h + Sσ2 ,0,l where

[T /2] T p X
p ∞ ∞
1 X X X X X
Sσ2 ,0,h = 3 2 Σii Σij
(T − 2 [T /2])([T /2] − 1) f =2 i=1 j=1 h1 =−∞ h2 =−∞
g=f +[T /2]
Xp ∞
X
Zh31 ,i Zh2 ,j (b̆2f,h1 b̆g,h1 b̆g,h2 + b̆f,h1 b̆f,h2 b̆2g,h1 ) − 3 Σii 2 4 2
b̆f,h b̆2g,h .

Zh,i (C.34)
i=1 h=−∞

Note that when t > 2


[T /2] T p X
p ∞ ∞
X X X X X
|Σii Σij | b̆2f,h1 |b̆g,h1 ||b̆g,h2 |
f =2 g=f +[T /2] i=1 j=1 h1 =−∞ h2 =−∞
p X
p [T /2] f T g
X X X X X
= |Σii Σij | b̆2f,h1 |b̆g,h1 | |b̆g,h2 |. (C.35)
i=1 j=1 f =2 h1 =−∞ g=f +[T /2] h2 =−∞

When h2 ≤ g
g
X ∞
X
|b̆g,h2 | ≤ 2 |bs | < ∞. (C.36)
h2 =−∞ s=0

When h1 ≤ f ≤ g − [T /2]
T
X ∞
X ∞
X
|b̆g,h1 | ≤ |bs | + |bs |T (1 − ϕ) = o(T −1 ) + O (T (1 − ϕ)) < ∞ (C.37)
g=f +[T /2] s=[T /2] s=0

and  2
f
X f
X
b̆2f,h1 ≤  |b̆f,h1 | < ∞. (C.38)
h1 =−∞ h1 =−∞

Moreover,
v
p X
p
u p X
p
X uX
p
|Σii Σij | ≤ max {|Σii |}t Σij 2 p2 = max {|Σii |}p tr(Σ2 ) = O(p3/2 ). (C.39)
1≤i≤p 1≤i≤p
i=1 j=1 i=1 j=1

From (C.36)-(C.39) we conclude that


[T /2] T p X
p ∞ ∞
X X X X X
|Σii Σij | b̆2f,h1 |b̆g,h1 ||b̆g,h2 | = O(p3/2 T ).
f =2 g=f +[T /2] i=1 j=1 h1 =−∞ h2 =−∞

Similarly, we can obtain the order


[T /2] T p
p X ∞ ∞
X X X X X
|Σii Σij | b̆2g,h1 |b̆f,h1 ||b̆f,h2 | = O(p3/2 T )
f =2 g=f +[T /2] i=1 j=1 h1 =−∞ h2 =−∞

34
and
p
X ∞
X
Σii 2 b̆2f,h b̆2g,h = O(p).
i=1 h=−∞

Summarizing the above we have

E|Sσ2 ,0,h | = O(p3/2 T −1 ) = o(p). (C.40)

Similarly write
[T /2] T
1 X X
ESσ2 ,0,l = (C.41)
(T − 32 [T /2])([T /2] − 1) f =2
g=f +[T /2]
X p
p X ∞
X ∞
X X p X p ∞
X ∞
X
Σ2ij b̆2f,h1 b̆2g,h2 + Σ2ij
b̆f,h1 b̆g,h2 b̆f,h2 b̆g,h1
i=1 j=1 h1 =−∞ h2 =−∞ i=1 j=1
h1 =−∞ h2 =−∞
Xp X p X∞ ∞
X Xp X∞
Σ2ii b̆2f,h b̆2g,h .

+ Σii Σjj b̆f,h1 b̆g,h1 b̆f,h2 b̆g,h2 − 3
i=1 j=1 h1 =−∞ h2 =−∞ i=1 h=−∞

Note that

X ∞
X ∞
X
| b̆f,h b̆g,h − ag−f | = | b̆f,h b̆g,h − bs bg−f +s | = O(1 − ϕ) (C.42)
h=−∞ h=−∞ s=0

and

X ∞
X ∞
X
| b̆2f,h − a0 | = | b̆2f,h − b2s | = O(1 − ϕ). (C.43)
h=−∞ h=−∞ s=0

We then conclude that


[TP
/2] T p P
p ∞ ∞
Σ2ij b̆2f,h1 b̆2g,h2
P P P P
f =2 g=f +[T /2] i=1 j=1 h1 =−∞ h2 =−∞
3 = a20 tr(Σ2 ) + O((1 − ϕ)p), (C.44)
(T − 2 [T /2])([T /2] − 1)

[TP
/2] T p P
p ∞ ∞
Σ2ij
P P P P
b̆f,h1 b̆g,h2 b̆f,h2 b̆g,h1
f =2 g=f +[T /2] i=1 j=1 h1 =−∞ h2 =−∞

(T − 32 [T /2])([T /2] − 1)
[T /2] T
tr(Σ2 ) X X
a2g−f + O(ag−f (1 − ϕ)) + O((1 − ϕ)2 )

=
(T − 32 [T /2])([T /2] − 1) f =2
g=f +[T /2]
−3 2
= O(pT + p(1 − ϕ) )
= o(T −1 + T 2 (1 − ϕ)2 ) = o(1), (C.45)

[TP
/2] T
P p P
P p ∞
P ∞
P
Σii Σjj b̆f,h1 b̆g,h1 b̆f,h2 b̆g,h2
f =2 g=f +[T /2] i=1 j=1 h1 =−∞ h2 =−∞

(T − 32 [T /2])([T /2] − 1)
[T /2] T
tr(Σ)2 X X
a2g−f + O(ag−f (1 − ϕ)) + O((1 − ϕ)2 )

=
(T − 32 [T /2])([T /2] − 1) f =2
g=f +[T /2]
2 −3
= O(p T + p (1 − ϕ) ) = o(p1/2 + pT 2 (1 − ϕ)2 ) = o(p),
2 2
(C.46)

35
and
[TP
/2] T p ∞
Σii 2 b̆2f,h b̆2g,h
P P P
f =2 g=f +[T /2] i=1 h=−∞

(T − 32 [T /2])([T /2] − 1)
p
Σii 2
P
[T /2] T f f
! !
i=1
X X X X
≤ b̆2f,h b̆2g,h
(T − 32 [T /2])([T /2] − 1) f =2 h=−∞ h=−∞
g=f +[T /2]

= o(pT −1 ) = o(p1/2 ). (C.47)

These ensure that

ESσ2 ,0,l = a20 tr(Σ2 ) + o(p). (C.48)

One can similarly obtain the order of V ar(Sσ2 ,0,l ) as follows

V ar(Sσ2 ,0,l ) = o(p3 T −2 + p) = o(p2 ). (C.49)

From (C.40), (C.48) and (C.49), we conclude that

Sσ2 ,0 tr(Σ2 )
= a20 + op (1) (C.50)
p p
and

S̀σ2 ,0 tr(Σ1/2 (I − H)Σ(I − H)Σ1/2 )


= a20 + op (1) (C.51)
p p
This, together with Lemma 2, implies that
r
S̀σ2 ,0 s
|µ̀m | 2 p

X 2tr(Σ1/2 (I − H)Σ(I − H)Σ1/2 )
S̀σ2 ,m = T
= (a0 + 2 ai ) + op (1). (C.52)
P x̀i,i p
i=1
p(T −1)
i=2

36

You might also like