You are on page 1of 327

A course in Time Series Analysis

Suhasini Subba Rao


Email: suhasini.subbarao@stat.tamu.edu
August 5, 2015

Contents
1 Introduction
1.1

Time Series data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .


1.1.1

R code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

11

Detrending a time series . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

13

1.2.1

Estimation of parametric trend . . . . . . . . . . . . . . . . . . . . .

13

1.2.2

Estimation using nonparametric methods . . . . . . . . . . . . . . . .

14

1.2.3

Estimation of the period . . . . . . . . . . . . . . . . . . . . . . . . .

16

1.3

Some formalism . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

23

1.4

Estimating the mean . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

24

1.5

Stationary processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

27

1.5.1

Types of stationarity (with Ergodicity thrown in) . . . . . . . . . . .

28

1.5.2

Towards statistical inference for time series . . . . . . . . . . . . . . .

33

What makes a covariance a covariance? . . . . . . . . . . . . . . . . . . . . .

34

1.2

1.6

2 Linear time series

38

2.1

Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

39

2.2

Linear time series and moving average models . . . . . . . . . . . . . . . . .

40

2.2.1

Infinite sums of random variables . . . . . . . . . . . . . . . . . . . .

40

The autoregressive model and the solution . . . . . . . . . . . . . . . . . . .

43

2.3.1

Difference equations and back-shift operators . . . . . . . . . . . . . .

43

2.3.2

Solution of two particular AR(1) models . . . . . . . . . . . . . . . .

43

2.3.3

The unique solution of a general AR(1) . . . . . . . . . . . . . . . . .

46

2.3.4

The solution of a general AR(p) . . . . . . . . . . . . . . . . . . . . .

47

2.3

2.3.5

Explicit solution of an AR(2) model . . . . . . . . . . . . . . . . . . .

48

2.3.6

Features of a realisation from an AR(2) . . . . . . . . . . . . . . . . .

51

2.3.7

Solution of the general AR() model . . . . . . . . . . . . . . . . . .

53

An explanation as to why the backshift operator method works . . . . . . .

57

2.4.1

Representing the AR(p) as a vector AR(1) . . . . . . . . . . . . . . .

60

2.5

The ARMA model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

62

2.6

Simulating from an Autoregressive process . . . . . . . . . . . . . . . . . . .

66

2.4

3 The autocovariance function of a linear time series


3.1

3.2

3.3

The autocovariance function . . . . . . . . . . . . . . . . . . . . . . . . . . .

70

3.1.1

The rate of decay of the autocovariance of an ARMA process . . . . .

71

3.1.2

The autocovariance of an autoregressive process . . . . . . . . . . . .

72

3.1.3

The autocovariance of a moving average process . . . . . . . . . . . .

79

3.1.4

The autocovariance of an autoregressive moving average process . . .

80

The partial covariance and correlation of a time series . . . . . . . . . . . . .

81

3.2.1

A review of partial correlation in multivariate analysis . . . . . . . .

82

3.2.2

Partial correlation in time series . . . . . . . . . . . . . . . . . . . . .

85

3.2.3

The variance/covariance matrix and precision matrix of an autoregressive and moving average process . . . . . . . . . . . . . . . . . . . . .

89

Correlation and non-causal time series . . . . . . . . . . . . . . . . . . . . .

91

3.3.1

The Yule-Walker equations of a non-causal process . . . . . . . . . .

94

3.3.2

Filtering non-causal AR models . . . . . . . . . . . . . . . . . . . . .

94

4 Nonlinear Time Series Models


4.1

4.2

70

97

Data Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

99

4.1.1

Yahoo data from 1996-2014 . . . . . . . . . . . . . . . . . . . . . . .

99

4.1.2

FTSE 100 from January - August 2014 . . . . . . . . . . . . . . . . . 102

The ARCH model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103


4.2.1

Features of an ARCH . . . . . . . . . . . . . . . . . . . . . . . . . . . 104

4.2.2

Existence of a strictly stationary solution and second order stationarity


of the ARCH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
2

4.3

4.4

4.5

The GARCH model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107


4.3.1

Existence of a stationary solution of a GARCH(1, 1) . . . . . . . . . . 109

4.3.2

Extensions of the GARCH model . . . . . . . . . . . . . . . . . . . . 111

4.3.3

R code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

Bilinear models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111


4.4.1

Features of the Bilinear model . . . . . . . . . . . . . . . . . . . . . . 112

4.4.2

Solution of the Bilinear model . . . . . . . . . . . . . . . . . . . . . . 113

4.4.3

R code . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115

Nonparametric time series models . . . . . . . . . . . . . . . . . . . . . . . . 115

5 Prediction

117

5.1

Forecasting given the present and infinite past . . . . . . . . . . . . . . . . . 118

5.2

Review of vector spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125


5.2.1

5.3

Spaces spanned by infinite number of elements . . . . . . . . . . . . . 128

Levinson-Durbin algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129


5.3.1

A proof based on projections . . . . . . . . . . . . . . . . . . . . . . . 130

5.3.2

A proof based on symmetric Toeplitz matrices . . . . . . . . . . . . . 132

5.3.3

Using the Durbin-Levinson to obtain the Cholesky decomposition of


the precision matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . 134

5.4

Forecasting for ARMA processes . . . . . . . . . . . . . . . . . . . . . . . . . 135

5.5

Forecasting for nonlinear models . . . . . . . . . . . . . . . . . . . . . . . . . 142


5.5.1

Forecasting volatility using an ARCH(p) model . . . . . . . . . . . . 142

5.5.2

Forecasting volatility using a GARCH(1, 1) model . . . . . . . . . . . 143

5.5.3

Forecasting using a BL(1, 0, 1, 1) model . . . . . . . . . . . . . . . . . 145

5.6

Nonparametric prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146

5.7

The Wold Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148

6 Estimation of the mean and covariance


6.1

An estimator of the mean . . . . . . . . . . . . . . . . . . . . . . . . . . . . 152


6.1.1

6.2

151

The sampling properties of the sample mean . . . . . . . . . . . . . . 152

An estimator of the covariance . . . . . . . . . . . . . . . . . . . . . . . . . . 155


3

6.2.1

Asymptotic properties of the covariance estimator . . . . . . . . . . . 157

6.2.2

Proof of Bartletts formula . . . . . . . . . . . . . . . . . . . . . . . . 158

6.3

Using Bartletts formula for checking for correlation . . . . . . . . . . . . . . 165

6.4

Long range dependence versus changes in the mean . . . . . . . . . . . . . . 168

7 Parameter estimation
7.1

7.2

7.3

171

Estimation for Autoregressive models . . . . . . . . . . . . . . . . . . . . . . 172


7.1.1

The Yule-Walker estimator . . . . . . . . . . . . . . . . . . . . . . . . 172

7.1.2

The Gaussian maximum likelihood . . . . . . . . . . . . . . . . . . . 175

Estimation for ARMA models . . . . . . . . . . . . . . . . . . . . . . . . . . 180


7.2.1

The Gaussian maximum likelihood estimator . . . . . . . . . . . . . . 180

7.2.2

The Hannan-Rissanen AR() expansion method . . . . . . . . . . . 184

The quasi-maximum likelihood for ARCH processes . . . . . . . . . . . . . . 185

8 Spectral Representations

187

8.1

How we have used Fourier transforms so far . . . . . . . . . . . . . . . . . . 188

8.2

The near uncorrelatedness of the Discrete Fourier Transform . . . . . . . . 193


8.2.1

Seeing the decorrelation in practice . . . . . . . . . . . . . . . . . . 194

8.2.2

Proof 1 of Lemma 8.2.1: By approximating Toeplitz with Circulant


matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197

8.3

8.2.3

Proof 2 of Lemma 8.2.1: Using brute force . . . . . . . . . . . . . . . 199

8.2.4

Heuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200

The spectral density and spectral distribution . . . . . . . . . . . . . . . . . 203


8.3.1

The spectral density and some of its properties

. . . . . . . . . . . . 203

8.3.2

The spectral distribution and Bochners theorem . . . . . . . . . . . . 206

8.4

The spectral representation theorem . . . . . . . . . . . . . . . . . . . . . . . 210

8.5

The spectral density functions of MA, AR and ARMA models . . . . . . . . 216

8.6

8.5.1

The spectral representation of linear processes . . . . . . . . . . . . . 216

8.5.2

The spectral density of a linear process . . . . . . . . . . . . . . . . . 217

8.5.3

Approximations of the spectral density to AR and MA spectral densities219

Higher order spectrums . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223


4

8.7

Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
8.7.1

The spectral density of a time series with randomly missing observations225

9 Spectral Analysis

227

9.1

The DFT and the periodogram . . . . . . . . . . . . . . . . . . . . . . . . . 228

9.2

Distribution of the DFT and Periodogram under linearity . . . . . . . . . . . 230

9.3

Estimating the spectral density function . . . . . . . . . . . . . . . . . . . . 236

9.4

The Whittle Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244

9.5

Ratio statistics in Time Series . . . . . . . . . . . . . . . . . . . . . . . . . . 253

9.6

Goodness of fit tests for linear time series models . . . . . . . . . . . . . . . 260

10 Consistency and and asymptotic normality of estimators

262

10.1 Modes of convergence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262


10.2 Sampling properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
10.3 Showing almost sure convergence of an estimator . . . . . . . . . . . . . . . 266
10.3.1 Proof of Theorem 10.3.2 (The stochastic Ascoli theorem) . . . . . . . 268
10.4 Toy Example: Almost sure convergence of the least squares estimator for an
AR(p) process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
10.5 Convergence in probability of an estimator . . . . . . . . . . . . . . . . . . . 273
10.6 Asymptotic normality of an estimator . . . . . . . . . . . . . . . . . . . . . . 274
10.6.1 Martingale central limit theorem . . . . . . . . . . . . . . . . . . . . 276
10.6.2 Example: Asymptotic normality of the least squares estimator . . . . 278
10.6.3 Example: Asymptotic normality of the weighted periodogram . . . . 281
10.7 Asymptotic properties of the Hannan and Rissanen estimation method . . . 283
T bT k2 ) . . . . . . . . . . . 288
10.7.1 Proof of Theorem 10.7.1 (A rate for kb
10.8 Asymptotic properties of the GMLE . . . . . . . . . . . . . . . . . . . . . . 291
A Background

300

A.1 Some definitions and inequalities . . . . . . . . . . . . . . . . . . . . . . . . 300


A.2 Martingales . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
A.3 The Fourier series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305

A.4 Application of Burkholders inequality . . . . . . . . . . . . . . . . . . . . . 309


A.5 The Fast Fourier Transform (FFT) . . . . . . . . . . . . . . . . . . . . . . . 311
B Mixingales

316

B.1 Obtaining almost sure rates of convergence for some sums . . . . . . . . . . 317
B.2 Proof of Theorem 10.7.3 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 318

Preface
The material for these notes come from several different places, in particular:
Brockwell and Davis (1998)
Shumway and Stoffer (2006) (a shortened version is Shumway and Stoffer EZ).
Fuller (1995)
Pourahmadi (2001)
Priestley (1983)
Box and Jenkins (1970)
A whole bunch of articles.
Tata Subba Rao and Piotr Fryzlewicz were very generous in giving advice and sharing
homework problems.
When doing the homework, you are encouraged to use all materials available, including
Wikipedia, Mathematica/Maple (software which allows you to easily derive analytic
expressions, a web-based version which is not sensitive to syntax is Wolfram-alpha).
You are encouraged to use R (see David Stoffers tutorial). I have tried to include
Rcode in the notes so that you can replicate some of the results.
Exercise questions will be in the notes and will be set at regular intervals.
You will be given some projects are the start of semester which you should select and
then present in November.

Chapter 1
Introduction
A time series is a series of observations xt , observed over a period of time. Typically the
observations can be over an entire interval, randomly sampled on an interval or at fixed time
points. Different types of time sampling require different approaches to the data analysis.
In this course we will focus on the case that observations are observed at fixed equidistant
time points, hence we will suppose we observe {xt : t Z} (Z = {. . . , 0, 1, 2 . . .}).
Let us start with a simple example, independent, uncorrelated random variables (the
simplest example of a time series). A plot is given in Figure 1.1. We observe that there
arent any clear patterns in the data. Our best forecast (predictor) of the next observation
is zero (which appears to be the mean). The feature that distinguishes a time series from
classical statistics is that there is dependence in the observations. This allows us to obtain
better forecasts of future observations. Keep Figure 1.1 in mind, and compare this to the
following real examples of time series (observe in all these examples you see patterns).

1.1

Time Series data

Below we discuss four different data sets.


The Southern Oscillation Index from 1876-present
The Southern Oscillation Index (SOI) is an indicator of intensity of the El Nino effect (see
wiki). The SOI measures the fluctuations in air surface pressures between Tahiti and Darwin.
8

2
1
0
2

whitenoise

50

100

150

200

Time

Figure 1.1: Plot of independent uncorrelated random variables


In Figure 1.2 we give a plot of monthly SOI from January 1876 - July 2014 (note that
there is some doubt on the reliability of the data before 1930). The data was obtained
from http://www.bom.gov.au/climate/current/soihtm1.shtml. Using this data set one

40

20

soi1

20

major goal is to look for patterns, in particular periodicities in the data.

1880

1900

1920

1940

1960

1980

2000

2020

Time

Figure 1.2: Plot of monthly Southern Oscillation Index, 1876-2014

Nasdaq Data from 1985-present


The daily closing Nasdaq price from 1st October, 1985- 8th August, 2014 is given in Figure
1.3. The (historical) data was obtained from https://uk.finance.yahoo.com. See also
http://www.federalreserve.gov/releases/h10/Hist/. Of course with this type of data

2000
1000

nasdaq1

3000

4000

the goal is to make money! Therefore the main object is to forecast (predict future volatility).

1985

1990

1995

2000

2005

2010

2015

Time

Figure 1.3: Plot of daily closing price of Nasdaq 1985-2014

Yearly sunspot data from 1700-2013


Sunspot activity is measured by the number of sunspots seen on the sun. In recent years it has
had renewed interest because times in which there are high activity causes huge disruptions
to communication networks (see wiki and NASA).
In Figure 1.4 we give a plot of yearly sunspot numbers from 1700-2013. The data was
obtained from http://www.sidc.be/silso/datafiles. For this type of data the main aim
is to both look for patterns in the data and also to forecast (predict future sunspot activity).
Yearly and monthly temperature data
Given that climate change is a very topical subject we consider global temperature data.
Figure 1.5 gives the yearly temperature anomalies from 1880-2013 and in Figure 1.6 we plot
10

150
100
0

50

sunspot1

1700

1750

1800

1850

1900

1950

2000

Time

Figure 1.4: Plot of Sunspot numbers 1700-2013


the monthly temperatures from January 1996 - July 2014. The data was obtained from
http://data.giss.nasa.gov/gistemp/graphs_v3/Fig.A2.txt and http://data.giss.
nasa.gov/gistemp/graphs_v3/Fig.C.txt respectively. For this type of data one may be
trying to detect for global warming (a long term change/increase in the average temperatures). This would be done by fitting trend functions through the data. However, sophisticated time series analysis is required to determine whether these estimators are statistically
significant.

1.1.1

R code

A large number of the methods and concepts will be illustrated in R. If you are not familar
with this language please learn the very basics.
Here we give the R code for making the plots above.
# assuming the data is stored in your main directory we scan the data into R
soi <soi1 <-

scan("~/soi.txt")
ts(monthlytemp,start=c(1876,1),frequency=12)

# the function ts creates a timeseries object, start = starting year,


# where 1 denotes January. Frequency = number of observations in a
11

0.5
0.5

0.0

temp

1880

1900

1920

1940

1960

1980

2000

Time

0.6
0.2

0.4

monthlytemp1

0.8

Figure 1.5: Plot of global, yearly average, temperature anomalies, 1880 - 2013

2000

2005

2010

2015

Time

Figure 1.6: Plot of global, monthly average, temperatures January, 1996 - July, 2014.
# unit of time (year). As the data is monthly it is 12.
plot.ts(soi1)

12

1.2

Detrending a time series

In time series, the main focus is on modelling the relationship between observations. Time
series analysis is usually performed after the data has been detrended. In other words, if
Yt = t + t , where {t } is zero mean time series, we first estimate t and then conduct the
time series analysis on the residuals. Once the analysis has been performed, we return to
the trend estimators and use the results from the time series analysis to construct confidence
intervals etc. In this course the main focus will be on the data after detrending. However,
we start by reviewing some well known detrending methods. A very good primer is given in
Shumway and Stoffer, Chapter 2, and you are strongly encouraged to read it.

1.2.1

Estimation of parametric trend

Often a parametric trend is assumed. Common examples include a linear trend


Yt = 0 + 1 t + t

(1.1)

Yt = 0 + 1 t + 2 t2 + t .

(1.2)

and the quadratic trend

For example we may fit such models to the yearly average temperature data. Alternatively
we may want to include seasonal terms

Yt = 0 + 1 sin

2t
12


+ 3 cos

2t
12


+ t .

For example, we may believe that the Southern Oscillation Index has a period 12 (since
the observations are taken monthly) and we use sine and cosine functions to model the
seasonality. For these type of models, least squares can be used to estimate the parameters.
Remark 1.2.1 (Taking differences to avoid fitting linear and higher order trends)
A commonly used method to avoid fitting linear trend to a model is to take first differences.

13

For example if Yt = 0 + 1 t + t , then


Zt = Yt+1 Yt = 1 + t+1 t .
Taking higher order differences (ie. taking first differences of {Zt } removes quadratic terms)
removes higher order polynomials.
Exercise 1.1

(i) Import the yearly temperature data (file global mean temp.txt) into R

and fit the linear model in (1.1) to the data (use the R command lsfit).
(ii) Suppose the errors in (1.1) are correlated. Under the correlated assumption, explain
why the standard errors reported in the R output are unreliable.
(iii) Make a plot of the residuals after fitting the linear model in (i). Make a plot of the
first differences. What do you notice about the two plots, similar?
(What I found was quite strange)
The AIC (Akaike Information Criterion) is usually used to select the parameters in the model
(see wiki). You should have studied the AIC/AICc/BIC in several of the prerequists you
have taken. In this course it will be assumed that you are familiar with it.

1.2.2

Estimation using nonparametric methods

In Section 1.2.1 we assumed that the mean had a certain known parametric form. This may
not always be the case. If we have no apriori idea of what features may be in the mean, we
can estimate the mean trend using a nonparametric approach. If we do not have any apriori
knowledge of the mean function we cannot estimate it without placing some assumptions on
its structure. The most common is to assume that the mean t is a sample from a smooth
function, ie. t = ( nt ). Under this assumption the following approaches are valid.
Possibly one of the most simplest methods is to use a rolling window. There are several
windows that one can use. We describe, below, the exponential window, since it can be

14

evaluated in an online way. For t = 1 let


1 = Y1 , then for t > 1 define

t = (1 )
t1 + Yt ,
where 0 < < 1. The choice of depends on how much weight one wants to give the present
observation. It is straightforward to show that

t =

t1
X

(1 )

tj

Yj =

t
X

j=1

[1 exp()] exp [(t j)] Yj

j=1

where = log(1 ). Let b = 1/ and K(u) = exp(u)I(u 0), then


t can be written
as
1/b

t = (1 e )
| {z }
b1

n
X


K

j=1

tj
b


Yj ,

This we observe that the exponential rolling window estimator is very close to a nonparametric kernel estimator of the mean, which has the form

t =

n
X
1
j=1


K

tj
b


Yj .

it is likely you came across such estimators in your nonparametric classes. The main difference between the rolling window estimator and the nonparametric kernel estimator is that
the kernel/window for the rolling window is not symmetric. This is because we are trying
to estimate the mean at time t, given only the observations up to time t. Whereas for nonparametric kernel estimators we can be observations on both sides of the neighbourhood of
t.
Other type of estimators include sieve-estimators. This is where we expand (u) in terms
of an orthogonal basis {k (u); k Z}
(u) =

ak k (u).

k=1

15

Examples of basis functions are the Fourier k (u) = exp(iku), Haar/other wavelet functions
etc. We observe that the unknown coefficients ak are a linear in the regressors k . Thus
we can use least squares to estimate the coefficients, {ak }. To estimate these coefficients, we
truncate the above expansion to order M , and use least squares to estimate the coefficients
n
X
t=1

"

#2
t
Yt
ak k ( ) .
n
k=1
M
X

(1.3)

The orthogonality of the basis means that the least squares estimator a
k is
n

1X
Yt k
a
k
n t=1

 
t
.
n

It is worth pointing out that regardless of the method used, correlations in the errors
{t } will play an role in quality of the estimator and even on the choice of bandwidth, b, or
equivalently the number of basis functions, M (see Hart (1991)). To understand why, suppose
t
the mean function is t = ( 200
) (the sample size n = 200), where (u) = 5(2u2.5u2 )+20.

We corrupt this quadratic function with both iid and dependent noise (the dependent noise is
the AR(2) process defined in equation (1.6)). The plots are given in Figure 1.7. We observe
that the dependent noise looks smooth (dependent can induce smoothness in a realisation).
This means that in the case that the mean has been corrupted by dependent noise it difficult
to see that the underlying trend is a simple quadratic function.

1.2.3

Estimation of the period

Suppose that the observations {Yt ; t = 1, . . . , n} satisfy the following regression model
Yt = A cos(t) + B sin(t) + t
where {t } are iid standard normal random variables and 0 < < . The parameters A, B,
and are real and unknown. Unlike the regression models given in (1.2.1) the model here
is nonlinear, since the unknown parameter, , is inside a trignometric function. Standard
least squares methods cannot be used to estimate the parameters. Assuming Gaussianity of
16

3
2

3 2 1

ar2

1
0
2 1

iid

0.0

0.2

0.4

0.6

0.8

1.0

0.0

0.2

0.6

0.8

1.0

0.8

1.0

temp

22

quadraticar2

18

20

24
22
20
18

16

16

quadraticiid

0.4

24

temp

0.0

0.2

0.4

0.6

0.8

1.0

0.0

0.2

temp

0.4

0.6

temp

Figure 1.7: Top: realisations from iid random noise and dependent noise (left = iid and right
= dependent). Bottom: Quadratic trend plus corresponding noise.
{t }, the maximum likelihood corresponding to the model is
n

1X
Ln (A, B, ) =
(Yt A cos(t) B sin(t))2 .
2 t=1
Nonlinear least squares method (which would require the use of a numerical maximisation
scheme) can be employed to estimate the parameters. However, using some algebraic manipulations, explicit expressions for the estimators can be obtained (see Walker (1971) and
Exercise 1.3). These are

n = arg max In ()

where

2
n

1 X

In () =
Yt exp(it)

n t=1
17

(1.4)

(we look for the maximum over the fundamental frequencies k =


n

2k
n

for 1 k n),

X
2X
n = 2
Yt cos(
n t) and B
Yt sin(
n t).
An =
n t=1
n t=1
The rather remarkable aspect of this result is that the rate of convergence of |
n | =
O(n3/2 ), which is faster than the standard O(n1/2 ) that we usually encounter (we will see
this in Example 1.2.1).
In () is usually called the periodogram. Searching for peaks in the periodogram is a long
established method for detecting periodicities. If we believe that there were two or more
periods in the time series, we can generalize the method to searching for the largest and
second largest peak etc. We consider an example below.
Example 1.2.1 Consider the following model

Yt = 2 sin

2t
8


+ t

t = 1, . . . , n.

(1.5)

where t are iid standard normal random variables. It is clear that {Yt } is made up of
a periodic signal with period eight. We make a plot of one realisation (using sample size
n = 128) together with the periodogram I() (defined in (1.4)). In Figure 1.8 we give a plot of
one realisation together with a plot of the periodogram. We observe that there is a symmetry,
this is because of the ei in the definition of I() we can show that I() = I(2 ). Notice
there is a clear peak at frequency 2/8 0.78 (where we recall that 8 is the period).
This method works extremely well if the error process {t } is uncorrelated. However, problems arise when the errors are correlated. To illustrate this issue, consider again model (1.5)
but this time let us suppose the errors are correlated. More precisely, they satisfy the AR(2)
model,
t = 1.5t1 0.75t2 + t ,

(1.6)

where {t } are iid random variables (do not worry if this does not make sense to you we
define this class of models precisely in Chapter 2). As in the iid case we use a sample size
18

1 2 3
1
3

signal

20

40

60

80

100

120

Time

0.2
0.0

0.4

frequency

Figure 1.8: Left: Realisation of (1.5) with iid noise, Right: Periodogram
n = 128. In Figure 1.9 we give a plot of one realisation and the corresponding periodogram.
We observe that the peak at 2/8 is not the highest. The correlated errors (often called

0
5

signal2

coloured noise) is masking the peak by introducing new peaks. To see what happens for

20

40

60

80

100

120

Time

0.4
0.0

P2

0.8

frequency

Figure 1.9: Left: Realisation of (1.5) with correlated noise and n = 128, Right: Periodogram

larger sample sizes, we consider exactly the same model (1.5) with the noise generated as
19

in (1.6). But this time we use n = 1024 (8 time the previous sample size). A plot of one
realisation, together with the periodogram is given in Figure 1.10. In contrast to the smaller

0
5

signal2

sample size, a large peak is seen at 2/8. These examples illustrates two important points:

200

400

600

800

1000

Time

0.0 0.2 0.4 0.6

P2

frequency

Figure 1.10: Left: Realisation of (1.5) with correlated noise and n = 1024, Right: Periodogram

(i) When the noise is correlated and the sample size is relatively small it is difficult to
disentangle the deterministic period from the noise. Indeed we will show in Chapters
2 and 3 that linear time series can exhibit similar types of behaviour to a periodic
deterministic signal. This is a subject of on going research that dates back at least 60
years (see Quinn and Hannan (2001)).
However, the similarity is only to a point. Given a large enough sample size (which
may in practice not be realistic), the deterministic frequency dominates again.
(ii) The periodogram holds important properties about the correlations in the noise (observe the periodogram in both Figures 1.9 and 1.10), there is some interesting activity
in the lower frequencies, that appear to be due to noise.
This is called spectral analysis and is explored in Chapters 8 and 9. Indeed a lot of
time series analysis can be done within the so called frequency or time domain.
20

Exercise 1.2 (Understanding Fourier transforms)

(i) Let Yt = 1. Plot the Peri-

odogram of {Yt ; t = 1, . . . , 128}.


(ii) Let Yt = 1 + t , where {t } are iid standard normal random variables. Plot the Periodogram of {Yt ; t = 1, . . . , 128}.
t
(iii) Let Yt = ( 128
) where (u) = 5 (2u 2.5u2 ) + 20. Plot the Periodogram of {Yt ; t =

1, . . . , 128}.
(iv) Let Yt = 2 sin( 2t
). Plot the Periodogram of {Yt ; t = 1, . . . , 128}.
8
(v) Let Yt = 2 sin( 2t
) + 4 cos( 2t
). Plot the Periodogram of {Yt ; t = 1, . . . , 128}.
8
12
Exercise 1.3

(i) Let

Sn (A, B, ) =

X
n

Yt2

n
X
t=1

t=1


 1
2
2
Yt A cos(t) + B sin(t) + n(A + B ) .
2

Show that
n

2Ln (A, B, ) + Sn (A, B, ) =

X
(A2 B 2 ) X
sin(2t).
cos(2t) + AB
2
t=1
t=1

and thus |Ln (A, B, ) + 12 Sn (A, B, )| = O(1) (ie. the difference does not grow with
n).
Since Ln (A, B, ) and 12 Sn (A, B, ) are asymptotically equivalent (i) shows that we
can maximise

1
S (A, B, )
2 n

instead of the likelihood Ln (A, B, ).

(ii) By profiling out the parameters A and B, use the the profile likelihood to show that
P

n = arg max | nt=1 Yt exp(it)|2 .


(iii) By using the identity (which is the one-sided Dirichlet kernel)
n
X
t=1

exp(it) =

exp( 12 i(n+1)) sin( 12 n)


sin( 12 )

0 < < 2

= 0 or 2.

21

(1.7)

we can show that for 0 < < 2 we have


n
X

t cos(t) = O(n)

t=1
n
X

t2 cos(t) = O(n2 )

n
X

t sin(t) = O(n)

t=1
n
X

t=1

t2 sin(t) = O(n2 ).

t=1

Using the above identities, show that the Fisher Information of Ln (A, B, ) (denoted
as I(A, B, )) is asymptotically equivalent to

2I(A, B, ) = E

2 Sn 

=
2

n2
2

n
2

n2
B
2

n
2

2
n2 A

n2

B + O(n) 2 A + O(n)

n3
3

+ O(n)
+ O(n)

(A2 + B 2 ) + O(n2 )

(iv) Use the Fisher information to show that |


n | = O(n3/2 ).
Exercise 1.4

(i) Simulate three hundred times from model (1.5) using n = 128. Esti-

mate , A and B for each simulation and obtain the empirical mean squared error
P300
1
2

i=1 (i ) (where denotes the parameter and i the estimate).


300
In your simulations, is the estimate of the period, superior to the estimator of coefficients, A and B?
(ii) Do the same as above but now use coloured noise given in (1.6) as the errors. How do
your estimates compare to (i)?
R Code
Simulation and periodogram for model (1.5) with iid errors:
temp <- rnorm(128)
signal <- 1.5*sin(2*pi*c(1:128)/8) + temp # this simulates the series
# Use the command fft to make the periodogram
P <- abs(fft(signal)/128)**2
frequency <- 2*pi*c(0:127)/128
# To plot the series and periodogram

22

par(mfrow=c(2,1))
plot.ts(signal)
plot(frequency, P,type="o")
Simulation and periodogram for model (1.5) with correlated errors:
set.seed(10)
ar2 <- arima.sim(list(order=c(2,0,0), ar = c(1.5, -0.75)), n=128)
signal2 <- 1.5*sin(2*pi*c(1:128)/8) + ar2
P2 <- abs(fft(signal2)/128)**2
frequency <- 2*pi*c(0:127)/128
par(mfrow=c(2,1))
plot.ts(signal2)
plot(frequency, P2,type="o")

1.3

Some formalism

When we observe the time series {xt }, usually we assume that {xt } is a realisation from a random
process {Xt }. We formalise this notion below. The random process {Xt ; t Z} (where Z denotes
the integers) is defined on the probability space {, F, P }. We explain what these mean below:
(i) is the set of all possible outcomes. Suppose that , then {Xt ()} is one realisation
from the random process. For any given , {Xt ()} is not random. In time series we will
usually assume that what we observe xt = Xt () (for some ) is a typical realisation. That
is, for any other , Xt ( ) will be different, but its general or overall characteristics
will be similar.
(ii) F is known as a sigma algebra. It is a set of subsets of (though not necessarily the set of
all subsets, as this can be too large). But it consists of all sets for which a probability can
be assigned. That is if A F, then a probability is assigned to the set A.
(iii) P is the probability.
Different types of convergence we will be using in class:

23

a.s.

(i) Almost sure convergence: Xn a as n (in this course a will always be a constant).
This means for every Xn a, where P () = 1 (this is classical limit of a sequence,
see Wiki for a definition).
P

(ii) Convergence in probability: Xn a. This means that for every > 0, P (|Xn a| > ) 0
as n (see Wiki)
2

(iii) Convergence in mean square Xn a. This means E|Xn a|2 0 as n (see Wiki).
(iv) Convergence in distribution. This means the distribution of Xn converges to the distribution
of X, ie. for all x where FX is continuous, we have Fn (x) FX (x) as n (where Fn and
FX are the distribution functions of Xn and X respectively). This is the simplest definition
(see Wiki).
Which implies which?
(i), (ii) and (iii) imply (iv).
(i) implies (ii).
(iii) implies (ii).
Central limit theorems require (iv). It is often easiest to show (iii) (since this only requires
mean and variance calculations).

1.4

Estimating the mean

Based on one realisation of a time series we want to make inference about parameters associated
with the process {Xt }, such as the mean etc. Let us consider the simplest case, estimating the
mean. We recall that in classical statistics we usually assume we observe several independent
realisations, {Xt } from a random variable X, and use the multiple realisations to make inference
= 1 Pn Xk . Roughly speaking, by using several independent realisations we
about the mean: X
k=1
n
are sampling over the entire probability space and obtaining a good estimate of the mean. On the
other hand if the samples were highly dependent, then it is likely that {Xt } would be concentrated
over small parts of the probability space. In this case, the variance of the sample mean would not
converge to zero as the sample size grows.

24

A typical time series is a half way house between totally dependent data and independent data.
Unlike, classical statistics, in time series, parameter estimation is based on only one realisation
xt = Xt () (not multiple, independent, replications). Therefore, it would appear impossible to
obtain a good estimator of the mean. However good estimates, of the mean, can be made, based
on just one realisation so long as certain assumptions are satisfied (i) the process has a constant
mean (a type of stationarity) and (ii) despite the fact that each time series is generated from one
realisation there is short memory in the observations. That is, what is observed today, xt has
little influence on observations in the future, xt+k (when k is relatively large). Hence, even though
we observe one tragectory, that trajectory traverses much of the probability space. The amount
of dependency in the time series determines the quality of the estimator. There are several ways
to measure the dependency. We know that the most common measure of linear dependency is the
covariance. The covariance in the stochastic process {Xt } is defined as
cov(Xt , Xt+k ) = E(Xt Xt+k ) E(Xt )E(Xt+k ).
Noting that if {Xt } has zero mean, then the above reduces to cov(Xt , Xt+k ) = E(Xt Xt+k ).
Remark 1.4.1 It is worth bearing in mind that the covariance only measures linear dependence.
For some statistical analysis, such as deriving an expression for the variance of an estimator, the
covariance is often sufficient as a measure. However, given cov(Xt , Xt+k ) we cannot say anything
about cov(g(Xt ), g(Xt+k )), where g is a nonlinear function. There are occassions where we require
a more general measure of dependence (for example, to show asymptotic normality). Examples of
more general measures include mixing (and other related notions, such as Mixingales, Near-Epoch
dependence, approximate m-dependence, physical dependence, weak dependence), first introduced by
Rosenblatt in the 50s (M. and Grenander (1997)). In this course we will not cover mixing.
Returning to the sample mean example suppose that {Xt } is a time series. In order to estimate the
mean we need to be sure that the mean is constant over time (else the estimator will be meaningless).
Therefore we will assume that {Xt } is a time series with constant mean . We observe {Xt }nt=1
= 1 Pn Xt . It is clear that this is an unbiased
and estimate the mean with the sample mean X
t=1
n
= (it is unbiased). Thus to see whether it converges in mean square
estimator of , since E(X)

25

to we consider its variance


=
var(X)

n
n
n
1 X
2 X X
var(X
)
+
cov(Xt , X ).
t
n2
n2
t=1

(1.8)

t=1 =t+1

If the covariance structure decays at such a rate that the sum of all lags is finite (supt

= |cov(Xt , X )|

, often called short memory), then the variance is O( n1 ), just as in the iid case. However, even
in order to test/construct CI for .
with this assumption we need to be able to estimate var(X)
Usually this requires the stronger assumption of stationarity, which we define in Section 1.5
Example 1.4.1 (The variance of a regression model with correlated errors) Let us return
to the parametric models discussed in Section 1.2.1. The general model is

Yt = 0 +

p
X

j ut,j + t = 0 ut + t ,

j=1

where E[t ] = 0 and we will assume that {ut,j } are nonrandom regressors. Note this includes the
parametric trend models discussed in Section 1.2.1. We use least squares to estimate
Ln () =

n
X

(Yt 0 ut )2 ,

t=1

with
n = arg min Ln () = (

n
X
t=1

Thus

)
Ln (
n

ut u0t )1

n
X

Yt ut .

t=1

we will derive an expression for


(this
= 0. To evaluate the variance of
n
n

expression also applies to many nonlinear estimators too). We note that by using straightforward
algebra we can show that
n

i0 X
) Ln () h
Ln (
n

ut u0t .

=
n

t=1

26

(1.9)

<

Moreoover because

)
Ln (
n

= 0 we have

) Ln ()
Ln (
n

Ln ()

n
n
X
X
u t t .
=
[Yt 0 ut ] ut =
| {z }
=

t=1

(1.10)

t=1

Altogether (1.9) and (1.10) give


h

n
i0 X

ut u0t =

t=1

n
X

u0t t .

t=1

and
h

n
X

!1
ut u0t

t=1

n
X

u t t .

t=1

Using this expression we can see that


h

=
var
n

1X
ut u0t
n

!1

var

t=1
1
n

Finally we need only evaluate var


n

var

1X
u t t
n

!
=

t=1

Pn

t=1 ut t

1X
u t t
n
t=1

1X
ut u0t
n

!1
.

t=1

which is

n
1 X
cov[t , ]ut u0
n2
t, =1
n
X

1
n2
|

var[t ]ut u0t +

t=1

{z

expression if independent

n
n
2 X X
cov[t , ]ut u0
n2
t=1 =t+1
|
{z
}

additional term due to correlation in the errors


P
P
Under the assumption that n1 nt=1 ut u0t is non-singular, supt kut k1 < and supt
= |cov(t , )| <
h
i
= O(n1 ), but just as in the case of the sample mean we need to
, we can see that var
n
impose some additional conditions on {t } we want to construct confidence intervals/test for .

1.5

Stationary processes

We have established that one of the main features that distinguish time series analysis from classical
methods is that observations taken over time (a time series) can be dependent and this dependency

27

tends to decline the further apart in time these two observations. However, to do any sort of analysis
of this time series we have to assume some sort of invariance in the time series, for example the mean
or variance of the time series does not change over time. If the marginal distributions of the time
series were totally different no sort of inference would be possible (suppose in classical statistics you
were given independent random variables all with different distributions, what parameter would
you be estimating, it is not possible to estimate anything!).
The typical assumption that is made is that a time series is stationary. Stationarity is a rather
intuitive concept, it is an invariant property which means that statistical characteristics of the time
series do not change over time. For example, the yearly rainfall may vary year by year, but the
average rainfall in two equal length time intervals will be roughly the same as would the number of
times the rainfall exceeds a certain threshold. Of course, over long periods of time this assumption
may not be so plausible. For example, the climate change that we are currently experiencing is
causing changes in the overall weather patterns (we will consider nonstationary time series towards
the end of this course). However in many situations, including short time intervals, the assumption
of stationarity is quite a plausible. Indeed often the statistical analysis of a time series is done
under the assumption that a time series is stationary.

1.5.1

Types of stationarity (with Ergodicity thrown in)

There are two definitions of stationarity, weak stationarity which only concerns the covariance of a
process and strict stationarity which is a much stronger condition and supposes the distributions
are invariant over time.
Definition 1.5.1 (Strict stationarity) The time series {Xt } is said to be strictly stationary
if for any finite sequence of integers t1 , . . . , tk and shift h the distribution of (Xt1 , . . . , Xtk ) and
(Xt1 +h , . . . , Xtk +h ) are the same.
The above assumption is often considered to be rather strong (and given a data it is very
hard to check). Often it is possible to work under a weaker assumption called weak/second order
stationarity.
Definition 1.5.2 (Second order stationarity/weak stationarity) The time series {Xt } is said
to be second order stationary if the mean is constant for all t and if for any t and k the covariance
between Xt and Xt+k only depends on the lag difference k. In other words there exists a function

28

c : Z R such that for all t and k we have


c(k) = cov(Xt , Xt+k ).
Remark 1.5.1 (Strict and second order stationarity)

(i) If a process is strictly stationar-

ity and E|Xt2 | < , then it is also second order stationary. But the converse is not necessarily
true. To show that strict stationarity (with E|Xt2 | < ) implies second order stationarity,
suppose that {Xt } is a strictly stationary process, then
cov(Xt , Xt+k ) = E(Xt Xt+k ) E(Xt )E(Xt+k )
Z


=
xy PXt ,Xt+k (dx, dy) PXt (dx)PXt+k (dy)
Z
=
xy [PX0 ,Xk (dx, dy) PX0 (dx)PXk (dy)] = cov(X0 , Xk ),
where PXt ,Xt+k and PXt is the joint distribution and marginal distribution of Xt , Xt+k respectively. The above shows that cov(Xt , Xt+k ) does not depend on t and {Xt } is second order
stationary.
(ii) If a process is strictly stationary but the second moment is not finite, then it is not second
order stationary.
(iii) It should be noted that a weakly stationary Gaussian time series is also strictly stationary too
(this is the only case where weakly stationary implies strictly stationary).
Example 1.5.1 (The sample mean and its variance under stationarity) Returning the variance of the sample mean discussed (1.8), if a time series is second order stationary, then the sample
is estimating the mean and the variance of X
is
mean X
=
var(X)

n
n
1
2 X X
var(X0 ) + 2
cov(Xt , X )
n
T
t=1 =t+1

n
1
2 X n r
var(X0 ) +
cov(X0 , Xr ) .
|
{z
}
n | {z } n
n
r=1

c(0)

c(r)

P
We approximate the above, by using that the covariances
r |c(r)| < . Therefore for all r,
Pn
P
(1 r/n)c(r) c(r) and | r=1 (1 |r|/n)c(r)| r |c(r)|, thus by dominated convergence (see

29

Chapter A)

Pn

r=1 (1

r/n)c(r)

r=1 c(r).

This implies that


var(X)

1
2X
1
c(0) +
c(r) = O( ).
n
n
n
r=1

The above is often called the long term variance. The above implies that
)2 = var(X)
0,
E(X
P

which we recall is convergence in mean square. Thus we have convergence in probability X


.

The example above illustrates how second order stationarity gave a rather elegant expression
for the variance. We now motivate the concept of ergodicity.
Sometimes, it is difficult to evaluate the mean and variance of an estimator, but often we
may only require almost sure or convergence in probability. Therefore, we may want to find an
alternative method to evaluating the mean squared error. To see whether this is possible we recall
that for iid random variables we have the very useful law of large numbers
n

1X
a.s.
Xt
n
t=1

and in general

1
n

Pn

t=1 g(Xt )

a.s.

E[g(X0 )] (if E[g(X0 )] < ). Does such a result exists in time

series? It does, but we require the slightly stronger condition that a time series is ergodic (which
is a slightly stronger condition than the strictly stationary).
Definition 1.5.3 (Ergodicity - very rough) Let (, F, P ) be a probability space. A transformation T : is said to be measure preserving if for every set A F, P (T 1 A) = P (A).
Moreover, it is said to be an ergodic transformation if T 1 A = A implies that P (A) = 0 or 1.
It is not obvious what this has to do with stochastic processes, but we attempt to make a link. Let
us suppose that X = {Xt } is a strictly stationary process defined on the probability space (, F, P ).
By strict stationarity the transformation (shifting a sequence by one)
T (x1 , x2 , . . .) = (x2 , x3 , . . .),
is a measure preserving transformation. Thus a process which is stationarity is measure preserving.

30

To understand ergodicity we define the set A, where


A = { : (X1 (), X0 (), . . .) H}. = { : X1 (), . . . , X2 (), . . .) H}.
The stochastic process is said to be ergodic, if the only sets which satisfies the above are such that
P (A) = 0 or 1. Roughly, this means there cannot be too many outcomes which generate sequences
which repeat itself (are periodic in some sense).
See Billingsley (1994), page 312-314, for examples and a better explanation.
The definition of ergodicity, given above, is quite complex and is rarely used in time series analysis.
However, one consequence of ergodicity is the ergodic theorem, which is extremely useful in time
series. It states that if {Xt } is an ergodic stochastic process then
n

1X
a.s.
g(Xt ) E[g(X0 )]
n
t=1

for any function g(). And in general for any shift 1 , . . . , k and function g : Rk+1 R we have
n

1X
a.s.
g(Xt , Xt+1 , . . . , Xt+k ) E[g(X0 , . . . , Xt+k )]
n

(1.11)

t=1

(often (1.11) is used as the definition of ergodicity, as it is an iff with the ergodic definition). Later
you will see how useful this.
(1.11) gives us an idea of what constitutes an ergodic process. Suppose that {t } is an ergodic
process (a classical example are iid random variables) then any reasonable (meaning measurable)
function of Xt is also ergodic. More precisely, if Xt is defined as
Xt = h(. . . , t , t1 , . . .),

(1.12)

where {t } are iid random variables and h() is a measureable function, then {Xt } is an Ergodic
process. For full details see Stout (1974), Theorem 3.4.5.
Remark 1.5.2 As mentioned above all Ergodic processes are stationary, but a stationary process
is not necessarily ergodic. Here is one simple example. Suppose that {t } are iid random variables
and Z is a Bernoulli random variable with outcomes {1, 2} (where the chance of either outcome is

31

half ). Suppose that Z stays the same for all t. Define

1 + t Z = 1
Xt =
+ Z = 2.
2
t
It is clear that E(Xt |Z = i) = i and E(Xt ) = 12 (1 + 2 ). This sequence is stationary. However,
P
we observe that T1 Tt=1 Xt will only converge to one of the means, hence we do not have almost
sure convergence (or convergence in probability) to 12 (1 + 2 ).
Exercise 1.5 State, with explanation, which of the following time series is second order stationary,
which are strictly stationary and which are both.
(i) {t } are iid random variables with mean zero and variance one.
(ii) {t } are iid random variables from a Cauchy distributon.
(iii) Xt+1 = Xt + t , where {t } are iid random variables with mean zero and variance one.
(iv) Xt = Y where Y is a random variable with mean zero and variance one.
(iv) Xt = Ut +Ut1 +Vt , where {(Ut , Vt )} is a strictly stationary vector time series with E[Ut2 ] <
and E[Vt2 ] < .
Example 1.5.2 In Chapter 6 we consider estimation of the autocovariance function. However for
now use R command acf. In Figure 1.11 we give the sample acf plots of the Southern Oscillation
Index and the Sunspot data. We observe that are very different. The acf of the SOI decays rapidly,
but there does appear to be some sort of pattern in the correlations. On the other hand, there is
more persistence in the acf of the Sunspot data. The correlations of the acf data appear to decay
but over a longer period of time and there is a clear periodicity in the correlation.
Exercise 1.6

(i) Make an ACF plot of the monthly temperature data from 1996-2014.

(ii) Make and ACF plot of the yearly temperature data from 1880-2013.
(iii) Make and ACF plot of the residuals (after fitting a line through the data (using the command
lsfit(..)$res)) of the yearly temperature data from 1880-2013.
Briefly describe what you see.

32

0.4
0.0

ACF

0.8

Series soi

50

100

150

200

250

300

40

50

60

Lag

0.4 0.0 0.4 0.8

ACF

Series sunspot

10

20

30
Lag

Figure 1.11: Top: ACF of Southern Oscillation data. Bottom ACF plot of Sunspot data.
R code
To make the above plots we use the commands
par(mfrow=c(2,1))
acf(soi,lag.max=300)
acf(sunspot,lag.max=60)

1.5.2

Towards statistical inference for time series

Returning to the sample mean Example 1.5.1. Suppose we want to construct CIs or apply statistical
tests on the mean. This requires us to estimate the long run variance (assuming stationarity)


var(X)

2X
1
c(0) +
c(r).
n
n
r=1

There are several ways this can be done, either by fitting a model to the data and from the model
estimate the covariance or doing it nonparametrically. This example motivates the contents of the
course:
(i) Modelling, finding suitable time series models to fit to the data.
(ii) Forecasting, this is essentially predicting the future given current and past observations.

33

(iii) Estimation of the parameters in the time series model.


(iv) The spectral density function and frequency domain approaches, sometimes within the frequency domain time series methods become extremely elegant.
(v) Analysis of nonstationary time series.
(vi) Analysis of nonlinear time series.
(vii) How to derive sampling properties.

1.6

What makes a covariance a covariance?

The covariance of a stationary process has several very interesting properties. The most important
is that it is positive semi-definite, which we define below.
Definition 1.6.1 (Positive semi-definite sequence)

(i) A sequence {c(k); k Z} (Z is the

set of all integers) is said to be positive semi-definite if for any n Z and sequence x =
(x1 , . . . , xn ) Rn the following is satisfied
n
X

c(i j)xi xj 0.

i,j=1

(ii) A function is said to be an even positive semi-definite sequence if (i) is satisfied and c(k) =
c(k) for all k Z.
An extension of this notion is the positive semi-definite function.
Definition 1.6.2 (Positive semi-definite function)

(i) A function {c(u); u R} is said to

be positive semi-definite if for any n Z and sequence x = (x1 , . . . , xn ) Rn the following


is satisfied
n
X

c(ui uj )xi xj 0.

i,j=1

(ii) A function is said to be an even positive semi-definite function if (i) is satisfied and c(u) =
c(u) for all u R.

34

Remark 1.6.1 You have probably encountered this positive definite notion before, when dealing
with positive definite matrices. Recall the n n matrix n is positive semi-definite if for all x Rn
x0 n x 0. To see how this is related to positive semi-definite matrices, suppose that the matrix n
P
has a special form, that is the elements of n are (n )i,j = c(ij). Then x0 n x = ni,j c(ij)xi xj .
We observe that in the case that {Xt } is a stationary process with covariance c(k), the variance
covariance matrix of X n = (X1 , . . . , Xn ) is n , where (n )i,j = c(i j).
We now take the above remark further and show that the covariance of a stationary process is
positive semi-definite.
Theorem 1.6.1 Suppose that {Xt } is a discrete time/continuous stationary time series with covariance function {c(k)}, then {c(k)} is a positive semi-definite sequence/function. Conversely for
any even positive semi-definite sequence/function there exists a stationary time series with this
positive semi-definite sequence/function as its covariance function.
PROOF. We prove the result in the case that {Xt } is a discrete time time series, ie. {Xt ; t Z}.
We first show that {c(k)} is a positive semi-definite sequence. Consider any sequence x =
Pn
(x1 , . . . , xn ) Rn , and the double sum
i,j xi c(i j)xj . Define the random variable Y =
Pn
Pn
0
i=1 xi Xi . It is straightforward to see that var(Y ) = x var(X n )x =
i,j=1 c(ij)xi xj where X n =
P
(X1 , . . . , Xn ). Since for any random variable Y , var(Y ) 0, this means that ni,j=1 xi c(ij)xj 0,
hence {c(k)} is a positive definite sequence.
To show the converse, that is for any positive semi-definite sequence {c(k)} we can find a
corresponding stationary time series with the covariance {c(k)} is relatively straightfoward, but
depends on defining the characteristic function of a process and using Komologorovs extension
theorem. We omit the details but refer an interested reader to Brockwell and Davis (1998), Section
1.5.


In time series analysis usually the data is analysed by fitting a model to the data. The model

(so long as it is correctly specified, we will see what this means in later chapters) guarantees the
covariance function corresponding to the model (again we cover this in later chapters) is positive
definite. This means, in general we do not have to worry about positive definiteness of the covariance
function, as it is implicitly implied.
On the other hand, in spatial statistics, often the object of interest is the covariance function
and specific classes of covariance functions are fitted to the data. In which case it is necessary to

35

ensure that the covariance function is semi-positive definite (noting that once a covariance function
has been found by Theorem 1.6.1 there must exist a spatial process which has this covariance
function). It is impossible to check for positive definiteness using Definitions 1.6.1 or 1.6.1. Instead
an alternative but equivalent criterion is used. The general result, which does not impose any
conditions on {c(k)} is stated in terms of positive measures (this result is often called Bochners
theorem). Instead, we place some conditions on {c(k)}, and state a simpler version of the theorem.
Theorem 1.6.2 Suppose the coefficients {c(k); k Z} are absolutely summable (that is

|c(k)| <

). Then the sequence {c(k)} is positive semi-definite if an only if the function f (), where

f () =

1 X
c(k) exp(ik),
2
k=

is nonnegative for all [0, 2].


We also state a variant of this result for positive semi-definite functions. Suppose the function
R
{c(u); k R} is absolutely summable (that is R |c(u)|du < ). Then the function {c(u)} is positive
semi-definite if and only if the function f (), where
1
f () =
2

c(u) exp(iu)du,

is nonnegative for all R.


PROOF. See Section 8.3.1.

Example 1.6.1 We will show that sequence c(0) = 1, c(1) = 0.5, c(1) = 0.5 and c(k) = 0 for
|k| > 1 a positive definite sequence.
From the definition of spectral density given above we see that the spectral density corresponding
to the above sequence is
f () = 1 + 2 0.5 cos().
Since | cos()| 1, f () 0, thus the sequence is positive definite. An alternative method is to
find a model which has this as the covariance structure. Let Xt = t + t1 , where t are iid random
variables with E[t ] = 0 and var(t ) = 0.5. This model has this covariance structure.

36

We note that Theorem 1.6.2 can easily be generalized to higher dimensions, d, by taking Fourier
transforms over Zd or Rd .
Exercise 1.7 Which of these sequences can used as the autocovariance function of a second order
stationary time series?
(i) c(1) = 1/2, c(0) = 1, c(1) = 1/2 and for all |k| > 1, c(k) = 0.
(ii) c(1) = 1/2, c(0) = 1, c(1) = 1/2 and for all |k| > 1, c(k) = 0.
(iii) c(2) = 0.8, c(1) = 0.5, c(0) = 1, c(1) = 0.5 and c(2) = 0.8 and for all |k| > 2,
c(k) = 0.
Exercise 1.8

(i) Show that the function c(u) = exp(a|u|) where a > 0 is a positive semi-

definite function.
(ii) Show that the commonly used exponential spatial covariance defined on R2 , c(u1 , u2 ) =
p
exp(a u21 + u22 ), where a > 0, is a positive semi-definite function.

37

Chapter 2
Linear time series
Prerequisites
Familarity with linear models.
Solve polynomial equations.
Be familiar with complex numbers.
Understand under what conditions the partial sum Sn =
P
P
(ie. if
j=1 aj .
j=1 |aj | < , then Sn S, where S =

Pn

j=1 aj

has a well defined limits

Objectives
Understand what causal and invertible is.
Know what an AR, MA and ARMA time series model is.
Know how to find a solution of an ARMA time series, and understand why this is important (how the roots determine causality and why this is important to know - in terms of
characteristics in the process and also simulations).
Understand how the roots of the AR can determine features in the time series and covariance
structure (such as pseudo periodicities).

38

2.1

Motivation

The objective of this chapter is to introduce the linear time series model. Linear time series models
are designed to model the covariance structure in the time series. There are two popular subgroups of linear time models (a) the autoregressive and (a) the moving average models, which can
be combined to make the autoregressive moving average models.
We motivate the autoregressive from the perspective of classical linear regression. We recall one
objective in linear regression is to predict the response variable given variables that are observed.
To do this, typically linear dependence between response and variable is assumed and we model Yi
as
Yi =

p
X

aj Xij + i ,

j=1

where i is such that E[i |Xij ] = 0 and more commonly i and Xij are independent. In linear
regression once the model has been defined, we can immediately find estimators of the parameters,
do model selection etc.
Returning to time series, one major objective is to predict/forecast the future given current and
past observations (just as in linear regression our aim is to predict the response given the observed
variables). At least formally, it seems reasonable to represent this as

Xt =

p
X

aj Xtj + t ,

(2.1)

j=1

where we assume that {t } are independent, identically distributed, zero mean random variables.
Model (2.1) is called an autoregressive model of order p (AR(p) for short). It is easy to see that

E(Xt |Xt1 , . . . , Xtp ) =

p
X

aj Xtj

j=1

(this is the exected value of Xt given that Xt1 , . . . , Xtp have already been observed), thus the
past values of Xt have a linear influence on the conditional mean of Xt (compare with the linear
P
P
model Yt = pj=1 aj Xt,j + t , then E(Yt |Xt,j ) = pj=1 aj Xt,j ). Conceptionally, the autoregressive
model appears to be a straightforward extension of the linear regression model. Dont be fooled
by this, it is a more complex object. Unlike the linear regression model, (2.1) is an infinite set of
linear difference equations. This means, for this systems of equations to be well defined, it needs

39

to have a solution which is meaningful. To understand why, recall that equation is defined for all
t Z, so let us start the equation at the beginning of time (t = ) and run it on. Without any
constraint on the parameters {aj }, there is no reason to believe the solution is finite (contrast this
with linear regression where these issues are not relevant). Therefore, the first thing to understand
is under what conditions will the AR model (2.1) have a well defined stationary solution and what
features in a time series is the solution able to capture.
Of course, one could ask why go through to the effort. One could simply use least squares to
estimate the parameters. This is possible, but without a proper analysis it is not clear whether
model has a meaningful solution (for example in Section 3.3 we show that the least squares estimator can lead to misspecified models), its not even possible to make simulations of the process.
Therefore, there is a practical motivation behind our theoretical treatment.
In this chapter we will be deriving conditions for a strictly stationary solution of (2.1). We
will place moment conditions on the innovations {t }, these conditions will be sufficient but not
necessary conditions Under these conditions we obtain a strictly stationary solution but not a
second order stationary process. In Chapter 3 we obtain conditions for (2.1) to have be both a
strictly and second order stationary solution. It is possible to obtain strictly stationary solution
under far weaker conditions (see Theorem 4.0.1), but these wont be considered here.
Example 2.1.1 How would you simulate from the model
Xt = 1 Xt1 + 2 Xt1 + t .

2.2
2.2.1

Linear time series and moving average models


Infinite sums of random variables

Before defining a linear time series, we define the MA(q) model which is a subclass of linear time
series. Let us supppose that {t } are iid random variables with mean zero and finite variance. The
time series {Xt } is said to have a MA(q) representation if it satisfies

Xt =

q
X

j tj ,

j=0

40

where E(t ) = 0 and var(t ) = 1. It is clear that Xt is a rolling finite weighted sum of {t },
therefore {Xt } must be well defined (which for finite sums means it is almost surely finite, this
you can see because it has a finite variance). We extend this notion and consider infinite sums of
random variables. Now, things become more complicated, since care must be always be taken with
anything involving infinite sums. More precisely, for the sum

j tj ,

j=

to be well defined (has a finite limit), the partial sums Sn =

Pn

j=n j tj

should be (almost

surely) finite and the sequence Sn should converge (ie. |Sn1 Sn2 | 0 as n1 , n2 ). Below, we
give conditions under which this is true.
Lemma 2.2.1 Suppose {Xt } is a strictly stationary time series with E|Xt | < , then {Yt } defined
by
Yt =

j Xtj ,

j=

< , is a strictly stationary time series. Furthermore, the partial sum converges
P
almost surely, Yn,t = nj=0 j Xtj Yt . If var(Xt ) < , then {Yt } is second order stationary
where

j=0 |j |

and converges in mean square (that is E(Yn,t Yt )2 0).


PROOF. See Brockwell and Davis (1998), Proposition 3.1.1 or Fuller (1995), Theorem 2.1.1 (page
31) (also Shumway and Stoffer (2006), page 86).

Example 2.2.1 Suppose {Xt } is a strictly stationary time series with var(Xt ) < . Define {Yt }
as the following infinite sum
Yt =

j k j |Xtj |

j=0

where || < 1. Then {Yt } is also a strictly stationary time series with a finite variance.
We will use this example later in the course.
Having derived conditions under which infinite sums are well defined (good), we can now define
the general class of linear and MA() processes.

41

Definition 2.2.1 (The linear process and moving average (MA)()) Suppose that {t }
P
are iid random variables,
j=0 |j | < and E(|t |) < .
A time series is said to be a linear time series if it can be represented as

Xt =

j tj ,

j=

where {t } are iid random variables with finite variance.


(ii)
(i) The time series {Xt } has a MA() representation if it satisfies

Xt =

j tj .

j=0

Note that since that as these sums are well defined by equation (1.12) {Xt } is a strictly stationary
(ergodic) time series.
The difference between an MA() process and a linear process is quite subtle. A linear process
involves both past, present and future innovations {t }, whereas the MA() uses only past and
present innovations.
Definition 2.2.2 (Causal and invertible)
(i) A process is said to be causal if it has the representation

Xt =

aj tj ,

j=0

(ii) A process is said to be invertible if it has the representation

Xt =

bj Xtj + t ,

j=1

(so far we have yet to give conditions under which the above has a well defined solution).
Causal and invertible solutions are useful in both estimation and forecasting (predicting the future
based on the current and past).
A very interesting class of models which have MA() representations are autoregressive and
autoregressive moving average models. In the following sections we prove this.

42

2.3

The autoregressive model and the solution

In this section we will examine under what conditions the AR process has a stationary solution.

2.3.1

Difference equations and back-shift operators

The autoregressive model is defined in terms of inhomogenuous difference equations. Difference


equations can often be represented in terms of backshift operators, so we start by defining them
and see why this representation may be useful (and why it should work).
The time series {Xt } is said to be an autoregressive (AR(p)) if it satisfies the equation
Xt 1 Xt1 . . . p Xtp = t ,

t Z,

where {t } are zero mean, finite variance random variables. As we mentioned previously, the autoregressive model is a difference equation (which can be treated as a infinite number of simultaneous
equations). Therefore for it to make any sense it must have a solution. To obtain a general solution
we write the autoregressive model in terms of backshift operators:
Xt 1 BXt . . . p B p Xt = t ,
where (B) = 1

Pp

j=1 j B

j,

(B)Xt = t

B is the backshift operator and is defined such that B k Xt = Xtk .

Simply rearranging (B)Xt = t , gives the solution of the autoregressive difference equation to
be Xt = (B)1 t , however this is just an algebraic manipulation, below we investigate whether it
really has any meaning. To do this, we start with an example.

2.3.2

Solution of two particular AR(1) models

Below we consider two different AR(1) models and obtain their solutions.
(i) Consider the AR(1) process
Xt = 0.5Xt1 + t ,

t Z.

(2.2)

Notice this is an equation (rather like 3x2 + 2x + 1 = 0, or an infinite number of simultaneous


equations), which may or may not have a solution. To obtain the solution we note that

43

Xt = 0.5Xt1 + t and Xt1 = 0.5Xt2 + t1 . Using this we get Xt = t + 0.5(0.5Xt2 +


t1 ) = t + 0.5t1 + 0.52 Xt2 . Continuing this backward iteration we obtain at the kth
P
iteration, Xt = kj=0 (0.5)j tj + (0.5)k+1 Xtk . Because (0.5)k+1 0 as k by taking
P
j
the limit we can show that Xt =
j=0 (0.5) tj is almost surely finite and a solution of
(2.2). Of course like any other equation one may wonder whether it is the unique solution
(recalling that 3x2 + 2x + 1 = 0 has two solutions). We show in Section 2.3.3 that this is the
unique stationary solution of (2.2).
Let us see whether we can obtain a solution using the difference equation representation. We
recall, that by crudely taking inverses, the solution is Xt = (1 0.5B)1 t . The obvious
P
j
question is whether this has any meaning. Note that (1 0.5B)1 =
j=0 (0.5B) , for
|B| 2, hence substituting this power series expansion into Xt we have

X
X
X
Xt = (1 0.5B)1 t = ( (0.5B)j )t = ( (0.5j B j ))t =
(0.5)j tj ,
j=0

j=0

j=0

which corresponds to the solution above. Hence the backshift operator in this example helps
us to obtain a solution. Moreover, because the solution can be written in terms of past values
of t , it is causal.
(ii) Let us consider the AR model, which we will see has a very different solution:
Xt = 2Xt1 + t .
Doing what we did in (i) we find that after the kth back iteration we have Xt =

(2.3)
Pk

j
j=0 2 tj +

2k+1 Xtk . However, unlike example (i) 2k does not converge as k . This suggest that if
P
j
we continue the iteration Xt =
j=0 2 tj is not a quantity that is finite (when t are iid).
P
j
Therefore Xt =
j=0 2 tj cannot be considered as a solution of (2.3). We need to write
(2.3) in a slightly different way in order to obtain a meaningful solution.
Rewriting (2.3) we have Xt1 = 0.5Xt + 0.5t . Forward iterating this we get Xt1 =
P
(0.5) kj=0 (0.5)j t+j (0.5)t+k+1 Xt+k . Since (0.5)t+k+1 0 we have

Xt1 = (0.5)

X
j=0

as a solution of (2.3).

44

(0.5)j t+j

Let us see whether the difference equation can also offer a solution. Since (1 2B)Xt = t ,
using the crude manipulation we have Xt = (1 2B)1 t . Now we see that (1 2B)1 =
P
P j j
j
j=0 (2B) for |B| < 1/2. Using this expansion gives Xt =
j=0 2 B Xt , but as pointed out
above this sum is not well defined. What we find is that (B)1 t only makes sense (is well
defined) if the series expansion of (B)1 converges in a region that includes the unit circle
|B| = 1.
What we need is another series expansion of (1 2B)1 which converges in a region which
includes the unit circle |B| = 1 (as an aside, we note that a function does not necessarily
have a unique series expansion, it can have difference series expansions which may converge
in different regions). We now show that a convergent series expansion needs to be defined in
terms of negative powers of B not positive powers. Writing (1 2B) = (2B)(1 (2B)1 ),
therefore

(1 2B)1 = (2B)1

(2B)j ,

j=0

which converges for |B| > 1/2. Using this expansion we have

Xt =

X
X
(0.5)j+1 B j1 t =
(0.5)j+1 t+j+1 ,
j=0

j=0

which we have shown above is a well defined solution of (2.3).


In summary (1 2B)1 has two series expansions

X
1
=
(2B)j
(1 2B)
j=0

which converges for |B| < 1/2 and

X
1
= (2B)1
(2B)j ,
(1 2B)
j=0

which converges for |B| > 1/2. The one that is useful for us is the series which converges
when |B| = 1.
It is clear from the above examples how to obtain the solution of a general AR(1). We now
show that this solution is the unique stationary solution.

45

2.3.3

The unique solution of a general AR(1)

Consider the AR(1) process Xt = Xt1 + t , where || < 1. Using the method outlined in (i), it is
P
j
straightforward to show that Xt =
j=0 tj is its stationary solution, we now show that this
solution is unique.
P
j
We first show that Xt =
j=0 tj is well defined (that it is almost surely finite). We note
P
P
j
j
that |Xt |
j=0 | | |tj |. Thus we will show that
j=0 | | |tj | is almost surely finite,
which will imply that Xt is almost surely finite. By montone convergence we can exchange sum
P
P
and expectation and we have E(|Xt |) E(limn nj=0 |j tj |) = limn nj=0 |j |E|tj |) =
P j
P
j
E(|0 |)
j=0 tj is a well defined solution of
j=0 | | < . Therefore since E|Xt | < ,
Xt = Xt1 + t .
To show that it is the unique stationary causal solution, let us suppose there is another (causal)
solution, call it Yt (note that this part of the proof is useful to know as such methods are often used
when obtaining solutions of time series models). Clearly, by recursively applying the difference
equation to Yt , for every s we have
Yt =

s
X

j tj + s Yts1 .

j=0

Evaluating the difference between the two solutions gives Yt Xt = As Bs where As = s Yts1
P
j
and Bs =
j=s+1 tj for all s. To show that Yt and Xt coincide almost surely we will show that
P
for every  > 0,
s=1 P (|As Bs | > ) < (and then apply the Borel-Cantelli lemma). We note
if |As Bs | > ), then either |As | > /2 or |Bs | > /2. Therefore P (|As Bs | > ) P (|As | >
/2)+P (|Bs | > /2). To bound these two terms we use Markovs inequality. It is straightforward to
show that P (|Bs | > /2) Cs /. To bound E|As |, we note that |Ys | || |Ys1 | + |s |, since {Yt }
is a stationary solution then E|Ys |(1 ||) E|s |, thus E|Yt | E|t |/(1 ||) < . Altogether
P
this gives P (|As Bs | > ) Cs / (for some finite constant C). Hence
s=1 P (|As Bs | > ) <
P
s
s=1 C / < . Thus by the Borel-Cantelli lemma, this implies that the event {|As Bs | > }
happens only finitely often (almost surely). Since for every , {|As Bs | > } occurs (almost surely)
P
j
only finite often for all , then Yt = Xt almost surely. Hence Xt =
j=0 tj is (almost surely)
the unique causal solution.

46

2.3.4

The solution of a general AR(p)

Let us now summarise our observation for the general AR(1) process Xt = Xt1 + t . If || < 1,
then the solution is in terms of past values of {t }, if on the other hand || > 1 the solution is in
terms of future values of {t }.
Now we try to understand this in terms of the expansions of the characteristic polynomial
(B) = 1 B (using the AR(1) as a starting point). From what we learnt in the previous
section, we require the characteristic polynomial of the AR process to have a convergent power
series expansion in the region including the ring |B| = 1. In terms of the AR(1) process, if the root
of (B) is greater than one, then the power series of (B)1 is in terms of positive powers, if it is
less than one, then (B)1 is in terms of negative powers.
Generalising this argument to a general polynomial, if the roots of (B) are greater than one,
then the power series of (B)1 (which converges for |B| = 1) is in terms of positive powers (hence
the solution (B)1 t will be in past terms of {t }). On the other hand, if the roots are both less
than and greater than one (but do not lie on the unit circle), then the power series of (B)1 will
be in both negative and positive powers. Thus the solution Xt = (B)1 t will be in terms of both
past and future values of {t }. We summarize this result in a lemma below.
Lemma 2.3.1 Suppose that the AR(p) process satisfies the representation (B)Xt = t , where
none of the roots of the characteristic polynomial lie on the unit circle and E|t | < . Then {Xt }
has a stationary, almost surely unique, solution.
We see that where the roots of the characteristic polynomial (B) lie defines the solution of
the AR process. We will show in Sections 2.3.6 and 3.1.2 that it not only defines the solution but
determines some of the characteristics of the time series.
Exercise 2.1 Suppose {Xt } satisfies the AR(p) representation

Xt =

p
X

j Xtj + t ,

j=1

where

Pp

j=1 |j |

< 1 and E|t | < . Show that {Xt } will always have a causal stationary solution.

47

2.3.5

Explicit solution of an AR(2) model

Specific example
Suppose {Xt } satisfies
Xt = 0.75Xt1 0.125Xt2 + t ,
where {t } are iid random variables. We want to obtain a solution for the above equations.
It is not easy to use the backward (or forward) iterating techique for AR processes beyond
order one. This is where using the backshift operator becomes useful. We start by writing Xt =
0.75Xt1 0.125Xt2 + t as (B)Xt = , where (B) = 1 0.75B + 0.125B 2 , which leads to what
is commonly known as the characteristic polynomial (z) = 1 0.75z + 0.125z 2 . If we can find a
power series expansion of (B)1 , which is valid for |B| = 1, then the solution is Xt = (B)1 t .
We first observe that (z) = 1 0.75z + 0.125z 2 = (1 0.5z)(1 0.25z). Therefore by using
partial fractions we have
1
1
2
1
=
=
+
.
(z)
(1 0.5z)(1 0.25z)
(1 0.5z) (1 0.25z)
We recall from geometric expansions that

X
1
=
(0.5)j z j
(1 0.5z)

|z| 2,

j=0

X
2
=2
(0.25)j z j
(1 0.25z)

|z| 4.

j=0

Putting the above together gives

X
1
=
{(0.5)j + 2(0.25)j }z j
(1 0.5z)(1 0.25z)

|z| < 2.

j=0

The above expansion is valid for |z| = 1, because

j=0 |

(0.5)j + 2(0.25)j | < (see Lemma

2.3.2). Hence
Xt = {(1 0.5B)(1 0.25B)}1 t =

X

{(0.5)j + 2(0.25)j }B j t =
{(0.5)j + 2(0.25)j }tj ,

j=0

j=0

which gives a stationary solution to the AR(2) process (see Lemma 2.2.1).
The discussion above shows how the backshift operator can be applied and how it can be used

48

to obtain solutions to AR(p) processes.

The solution of a general AR(2) model


We now generalise the above to general AR(2) models
Xt = (a + b)Xt1 abXt2 + t ,
the characteristic polynomial of the above is 1 (a + b)z + abz 2 = (1 az)(1 bz). This means
the solution of Xt is
Xt = (1 Ba)1 (1 Bb)1 t ,
thus we need an expansion of (1 Ba)1 (1 Bb)1 . Assuming that a 6= b, as above we have


1
b
a
1
=

(1 za)(1 zb)
b a 1 bz 1 az
Cases:
(i) |a| < 1 and |b| < 1, this means the roots lie outside the unit circle. Thus the expansion is

j=0

j=0

X
X

1
1
=
b
bj z j a
aj z j ,
(1 za)(1 zb)
(b a)
which leads to the causal solution
1
Xt =
ba

X

j+1

j+1


)tj .

(2.4)

j=0

(ii) Case that |a| > 1 and |b| < 1, this means the roots lie inside and outside the unit circle and
we have the expansion
1
(1 za)(1 zb)

1
ba

1
(b a)

49


b
a

1 bz (az)((az)1 1)
 X


X
bj z j + z 1
aj z j ,
b
j=0

j=0

(2.5)

which leads to the non-causal solution


Xt =

X
X

1
bj+1 tj +
aj t+1+j .
ba
j=0

(2.6)

j=0

Later we show that the non-causal solution has the same correlation structure as the causal
solution when a = a1 .
This solution throws up additional interesting results. Let us return to the expansion in (2.5)
and apply it to Xt

Xt =

1
b
1
1

t =
t +

t
1
1
(1 Ba)(1 Bb)
b a |1 {z
bB }
B(1 a B )
{z
}
|
causal AR(1)

noncausal AR(1)

1
(Yt + Zt+1 )
ba

where Yt = bYt1 + t and Zt+1 = a1 Zt+2 + t+1 . In other words, the noncausal AR(2)
process is the sum of a causal and afuture AR(1) process. This is true for all noncausal
time series (except when there is multiplicity in the roots) and is discussed further in Section
2.6.
Several authors including Richard Davis, Jay Breidt and Beth Andrews argue that noncausal
time series can model features in data which causal time series cannot.
(iii) a = b > 1. The characteristic polynomial is (1 az)2 . To obtain the convergent expansion
when |z| = 1 we note that (1 az)2 = (1) d(1az)
d(az)

. Thus

X
(1)
=
(1)
j(az)j1 .
(1 az)2
j=0

This leads to the causal solution


Xt = (1)

jaj1 tj .

j=1

Exercise 2.2 Show for the AR(2) model Xt = 1 Xt1 + 2 Xt2 + t to have a causal stationary

50

solution the parameters 1 , 2 must lie in the region


2 + 1 < 1,
Exercise 2.3

2 1 < 1

|2 | < 1.

(a) Consider the AR(2) process


Xt = 1 Xt1 + 2 Xt2 + t ,

where {t } are iid random variables with mean zero and variance one. Suppose the roots of
the characteristic polynomial 1 1 z 2 z 2 are greater than one. Show that |1 | + |2 | < 4.
(b) Now consider a generalisation of this result. Consider the AR(p) process
Xt = 1 Xt1 + 2 Xt2 + . . . p Xtp + t .
Suppose the roots of the characteristic polynomial 1 1 z . . . p z p are greater than one.
Show that |1 | + . . . + |p | 2p .

2.3.6

Features of a realisation from an AR(2)

We now explain why the AR(2) (and higher orders) can characterise some very interesting behaviour
(over the rather dull AR(1)). For now we assume that Xt is a causal time series which satisfies the
AR(2) representation
Xt = 1 Xt1 + 2 Xt2 + t
where {t } are iid with mean zero and finite variance. The characteristic polynomial is (B) =
1 1 B 2 B 2 . Let us assume the roots of (B) are complex, since 1 and 2 are real, the roots
are complex conjugates. Thus by using case (i) above we have
1
1
=
1 1 B 2 B 2

where 1 and

1 B 1 B

are the roots of the characteristic. Thus


Xt = C

tj

j=0

X
j=0

51

j tj ,

(2.7)

where C = [ ]1 . Since and C are complex we use the representation = r exp(i) and
C = exp(i) (noting that |r| < 1), and substitute these expressions for and C into (2.7) to give

Xt =

rj cos(j + )tj .

j=0

We can see that Xt is effectively the sum of cosines with frequency that have been modulated by
the iid errors and exponentially damped. This is why for realisations of autoregressive processes
you will often see periodicities (depending on the roots of the characteristic). These arguments can
be generalised to higher orders p.
Exercise 2.4

(a) Obtain the stationary solution of the AR(2) process


2
7
Xt = Xt1 Xt2 + t ,
3
3

where {t } are iid random variables with mean zero and variance 2 .
Does the solution have an MA() representation?
(b) Obtain the stationary solution of the AR(2) process

4 3
42
Xt =
Xt1 2 Xt2 + t ,
5
5
where {t } are iid random variables with mean zero and variance 2 .
Does the solution have an MA() representation?
(c) Obtain the stationary solution of the AR(2) process
Xt = Xt1 4Xt2 + t ,
where {t } are iid random variables with mean zero and variance 2 .
Does the solution have an MA() representation?
Exercise 2.5 Construct a causal stationary AR(2) process with pseudo-period 17. Using the R
function arima.sim simulate a realisation from this process (of length 200) and make a plot of the
periodogram. What do you observe about the peak in this plot?
Below we now consider solutions to general AR() processes.

52

2.3.7

Solution of the general AR() model

AR() models are more general than the AR(p) model and are able to model more complex
behaviour, such as slower decay of the covariance structure. It is arguable how useful these models
are in modelling data, however recently it has become quite popular in time series bootstrap
methods.
In order to obtain the stationary solution of an AR(), we need to define an analytic function
and its inverse.
Definition 2.3.1 (Analytic functions in the region ) Suppose that z C. (z) is an analytic complex function in the region , if it has a power series expansion which converges in , that
P
j
is (z) =
j= j z .
P
j

If there exists a function (z)


=
j= j z such that (z)(z) = 1 for all z , then (z)
is the inverse of (z) in the region .
Well known examples of analytic functions include
(i) Finite order polynomials such as (z) =
(ii) The expansion (1 0.5z)1 =

Pp

j=0 j z

j
j=0 (0.5z)

for = C.

for = {z; |z| 2}.

We observe that for AR processes we can represent the equation as (B)Xt = t , which formally
gives the solution Xt = (B)1 t . This raises the question, under what conditions on (B)1 is
(B)1 t a valid solution. For (B)1 t to make sense (B)1 should be represented as a power
series expansion. Below, we give conditions on the power series expansion which give a stationary
solution. It is worth noting this is closely related to Lemma 2.2.1.
P

is finite on a region that includes |z| = 1 (hence


P
it is analytic) and {Xt } is a strictly stationary process with E|Xt | < . Then
j= |j | <
P
and Yt = (B)Xtj = j= j Xtj is almost surely finite and strictly stationary time series.
Lemma 2.3.2 Suppose that (z) =

j= j z

P
j
PROOF. It can be shown that if sup|z|=1 |(z)| < , in other words on the unit circle
j= j z <
P
, then
j= |j | < . Since the coefficients are absolutely summable, then by Lemma 2.2.1
P
we have that Yt = (B)Xtj =

j= j Xtj is almost surely finite and strictly stationary.
Using the above we can obtain the solution of an AR() (which includes an AR(p) as a special
case).

53

Corollary 2.3.1 Suppose that


Xt =

j Xtj + t

j=1

and (z) has an inverse (z) =

j= j z

which is analytic in a region including |z| = 1, then

the solution of Xt is
Xt =

j tj .

j=

Corollary 2.3.2 Let Xt be an AR(p) time series, where

Xt =

p
X

j Xtj + t .

j=1

Suppose the roots of the characteristic polynomial (B) = 1

Pp

j=1 j B

do not lie on the unit

circle |B| = 1, then Xt admits a strictly stationary solution.


In addition suppose the roots of (B) all lie outside the unit circle, then Xt admits a strictly
stationary, causal solution.
This summarises what we observed in Section 2.3.4.
Rules of the back shift operator:
(i) If a(z) is analytic in a region which includes the unit circle |z| = 1 in its interior, then
a(B)Xt is a well defined random variable.
(ii) The operator is commutative and associative, that is [a(B)b(B)]Xt = a(B)[b(B)Xt ] =
[b(B)a(B)]Xt (the square brackets are used to indicate which parts to multiply first). This
may seems obvious, but remember matrices are not commutative!
(iii) Suppose that a(z) and its inverse

1
a(z)

are both have solutions in the region which includes

the unit circle |z| = 1 in its interior. If a(B)Xt = Zt , then Xt =

1
a(B) Zt .

(i) Clearly a(z) = 1 0.5z is analytic for all z C,


P
1
j
and has no zeros for |z| < 2. The inverse is a(z)
=
j=0 (0.5z) is well defined in the region

Example 2.3.1 (Analytic functions)

|z| < 2.

54

(ii) Clearly a(z) = 1 2z is analytic for all z C, and has no zeros for |z| > 1/2. The inverse is
P
1
1
1
j
a(z) = (2z) (1 (1/2z)) = (2z) ( j=0 (1/(2z)) ) well defined in the region |z| > 1/2.
(iii) The function a(z) =

1
(10.5z)(12z)

is analytic in the region 0.5 < z < 2.

(iv) a(z) = 1 z, is analytic for all z C, but is zero for z = 1. Hence its inverse is not well
defined for regions which involve |z| = 1 (see Example 2.3.2).
Example 2.3.2 (Unit root/integrated processes and non-invertible processes)
(i) If the difference equation has root one, then an (almost sure) stationary solution of the AR model
do not exist. The simplest example is the random walk Xt = Xt1 + t ((z) = (1 z)). This is
an example of an Autoregressive Integrated Moving Average ARIMA(0, 1, 0) model (1 B)Xt = t .
To see that it does not have a stationary solution, we iterate the equation n steps backwards
P
P
and we see that Xt = nj=0 tj + Xtn . St,n = nj=0 tj is the partial sum, but it is clear that
the partial sum St,n does not have a limit, since it is not a Cauchy sequence, ie. |St,n St,m | does
not have a limit. However, given some initial value X0 , for t > 0 we can define the unit process
P
Xt = Xt1 +. Notice that the nonstationary solution of this sequence is Xt = X0 + tj=1 tj which
has variance var(Xt ) = var(X0 ) + t (assuming that {t } are iid random variables with variance one
and independent of X0 ).
We observe that we can stationarize the process by taking first differences, ie. defining Yt =
Xt Xt1 = t .
(ii) The unit process described above can be generalised ARIMA(0, d, 0), where (1 B)d Xt = t .
In this case to stationarize the sequence we take d differences, ie. let Yt,0 = Xt and for 1 i d
define the iteration
Yt,i = Yt,i1 Yt1,i1
and Yt = Yt,d will be a stationary sequence. Or define

Yt =

d
X
j=0

d!
(1)j Xtj ,
j!(d j)!

in which case Yt as defined above will be a stationary sequence.

55

(iii) The general ARIMA(p, d, q) is defined as (1 B)d (B)Xt = (B)t , where (B) and (B) are
p and q order polynomials respectively and the roots of (B) lie outside the unit circle.
Another way of describing the above model is that after taking d differences (as detailed in (ii))
the resulting process is an ARMA(p, q) process (see Section 2.5 for the definition of an ARMA
model).
To illustrate the difference between stationary ARMA and ARIMA processes, in Figure 2.1
(iv) In examples (i) and (ii) a stationary solution does not exist. We now consider an example
where the process is stationary but an autoregressive representation does not exist.
Consider the MA(1) model Xt = t t1 . We recall that this can be written as Xt = (B)t
where (B) = 1 B. From Example 2.3.1(iv) we know that (z)1 does not exist, therefore it does

20
60

40

20

ar2

ar2I

40

60

80

not have an AR() representation since (1 B)1 Xt = t is not well defined.

100

200

300

400

Time

100

200

300

400

Time

(a) Xt = 1.5Xt1 0.75Xt2 + t

(b) (1 B)Yt = Xt , where is defined in (a)

Figure 2.1: Realisations from an AR process and its corresponding integrated process, using
N (0, 1) innovations (generated using the same seed).

56

2.4

An explanation as to why the backshift operator


method works

To understand why the magic backshift operator works, we use matrix notation to rewrite the
AR(p) model as an infinite set of difference equations

..
.

... ... ... ...


...
...
...

Xt

... 0
1 1 . . . p . . .

t1
... 0
0
1
1 . . . p

X
t2
.
... ... ... ...
...
...
...
..

..
.



t


=
t1


t2
.
..

The above is an infinite dimensional equation (and the matrix is an infinite upper triangular matrix).
Formally to obtain a simulation we invert the matrix to get a solution of Xt in terms of t . Of course
in reality it is not straightfoward to define this inverse. Instead let us consider a finite (truncated)
version of the above matrix equation. Except for the edge effects this is a circulant matrix (where
the rows are repeated, but each time shifted by one, see wiki for a description). Truncating the
matrix to have dimension n, we approximate the above by the finite set of n-equations

...

0
1
1

...
...
...

1 2 . . .

...

...

...

...

...

Xn

Xn1

..
.

X0


= n1

..

C n X n n .
The approximation of the AR(p) equation only arises in the first p-equations, where

X0

p
X

j Xnj

= 0

j Xn+1j

= 1

j=1

X1 1 X0

p
X
j=2

..
.
Xp

p1
X

..
.

j Xpj p Xn = p .

j=1

57

We now define the n n matrix Un , where

Un =

0 1 0 0 ... 0
0 0 1 0 ... 0
.. .. .. .. .. ..
. . . . . .

1 0 0 0 ... 0
We observe that Un is a deformed diagonal matrix where all the ones along the diagonal have
been shifted once to the right, and the left over one is placed in the bottom left hand corner. Un
is another example of a circulant matrix, moreover Un2 shifts once again all the ones to the right

2
Un =

0 1 0 0 ... 0
0 0 1 0 ... 0
.. .. .. .. .. ..
. . . . . .
1 0 0 0 ... 0

0 1 0 0 ... 0
Un3 shifts the ones to the third off-diagonal and so forth until Unn = I. Thus all circulant matrices
can be written in terms of powers of Un (the matrix Un can be considered as the building blocks
of circulant matrices). In particular

Cn = In

p
X

j Unj ,

j=1

[In

Pp

j
j=1 j Un ]X n

= n and the solution to the equation is

X n = (In

p
X

j Unj )1 n .

j=1

Our aim is to write (In

Pp

j 1
j=1 j Un )

as a power series in Un , with Un playing the role of the

backshift operator.
P
To do this we recall the similarity between the matrix In pj=1 j Unj and the characteristic
P
equation (B) = 1 pj=1 j z j . In particular since we can factorize the characteristic equation as
Q
P
Q
(B) = pj=1 [1 j B], we can factorize the matrix In pj=1 j Unj = pj=1 [In j Un ]. To obtain
the inverse, for simplicity, we assume that the roots of the characteristic function are greater than

58

one (ie. |j | < 1, which we recall corresponds to a causal solution) and are all different. Then there
exists constants cj where

[In

p
X

j Unj ]1

j=1

p
X

cj (In j Un )1

j=1

P
(just as in partial fractions) - to see why multiply the above by [In pj=1 j Unj ]. Finally, we
P
j
recall that if the eigenvalues of A are less than one, then (1 A)1 =
j=0 A . The eigenvalues
of Un are {exp( 2ij
n ); j = 1, . . . , n}, thus the eigenvalues of j Un are less than one. This gives
P
k k
(In j Un )1 =
k=0 j Un and

[In

p
X

j Unj ]1 =

j=1

p
X

cj

j=1

kj Unk .

(2.8)

k=0

Therefore, the solution of Cn X n = n is

X n = Cn1 n =

p
X

cj

j=1

kj Unk n .

k=0

Let us focus on the first element of the vector X n , which is Xn . Since Unk n shifts the elements of
n up by k (note that this shift is with wrapping of the vector) we have
Xn =

p
X
j=1

cj

n
X

kj nk

k=0

p
X

cj

j=1

kj nk
{z

mod (n) .

(2.9)

k=n+1
0

Note that the second term decays geometrically fast to zero. Thus giving the stationary solution
P
P
k
Xn = pj=1 cj
k=0 j nk .
P
To recollect, we have shown that [In pj=1 j Unj ]1 admits the solution in (2.8) (which is
the same as the solution of the inverse of (B)1 ) and that Unj n plays the role of the backshift
operator. Therefore, we can use the backshift operator in obtaining a solution of an AR process
because it plays the role of the matrix Un .

59

Example 2.4.1 The AR(1) model, Xt 1 Xt1 = t is written as

...

1
1

...

1 . . .

...

...

0 0

Xn

Xn1

..
.

X0

... ...
0


= n1

..

C n X n = n .
The approximation of the AR(1) is only for the first equation, where X0 1 Xn = 0 . Using the
matrix Un , the above equation can be written as (In 1 Un )X n = n , which gives the solution
X n = (In 1 Un )1 n .
Let us suppose that |1 | > 1 (ie, the root lies inside the unit circle and the solution is noncausal),
1
then to get a convergent expansion of (1n 1 Un )1 we rewrite (In 1 Un ) = 1 Un (In 1
1 Un ).

Thus we have
"
(In 1 Un )

#
k
k
1 Un

(1 Un )1 .

k=0

Therefore the solution is

Xn =

!
Unk+1
k+1
1

n ,

k=0

which in its limit gives the same solution as Section 2.3.2(ii).


Notice that Unj and B j are playing the same role.

2.4.1

Representing the AR(p) as a vector AR(1)

Let us suppose Xt is an AR(p) process, with the representation

Xt =

p
X

j Xtj + t .

j=1

For the rest of this section we will assume that the roots of the characteristic function, (z), lie
outside the unit circle, thus the solution causal. We can rewrite the above as a Vector Autoregressive

60

(VAR(1)) process
X t = AX t1 + t

(2.10)

where

1 2 . . . p1 p

...

...

...

0
,

(2.11)

X 0t = (Xt , . . . , Xtp+1 ) and 0t = (t , 0, . . . , 0). It is straightforward to show that the eigenvalues of


A are the inverse of the roots of (z) (since
det(A zI) = z p

p
X

i z pi = z p (1

i=1

p
X

i z i )),

i=1

{z

=z p (z 1 )

thus the eigenvalues of A lie inside the unit circle. It can be shown that for any |max (A)| < < 1,
there exists a constant C such that |kAj kspec C j (see Appendix A). Note that result is
extremely obvious if the eigenvalues are distinct (in which case the spectral decomposition can be
used), in which case |kAj kspec C |max (A)|j (note that kAkspec is the spectral norm of A, which
is the largest eigenvalue of the symmetric matrix AA0 ).
We can apply the same back iterating that we did for the AR(1) to the vector AR(1). Iterating
(2.10) backwards k times gives

Xt =

k1
X

Aj tj + Ak X tk .

j=0
P

Since kAk X tk k2 kAk kspec kX tk k 0 we have

Xt =

Aj tj .

j=0

61

2.5

The ARMA model

Up to now, we have defined the moving average and the autoregressive model. The MA(q) average
has the feature that after q lags there isnt any correlation between two random variables. On
the other hand, there are correlations at all lags for an AR(p) model. In addition as we shall see
later on, it is much easier to estimate the parameters of an AR model than an MA. Therefore,
there are several advantages in fitting an AR model to the data (note that when the roots are
of the characteristic polynomial lie inside the unit circle, then the AR can also be written as an
MA(), since it is causal). However, if we do fit an AR model to the data, what order of model
should we use? Usually one uses the AIC (BIC or similar criterion) to determine the order. But
for many data sets, the selected order tends to be relative large, for example order 14. The large
order is usually chosen when correlations tend to decay slowly and/or the autcorrelations structure
is quite complex (not just monotonically decaying). However, a model involving 10-15 unknown
parameters is not particularly parsimonious and more parsimonious models which can model the
same behaviour would be useful. A very useful generalisation which can be more flexible (and
parsimonious) is the ARMA(p, q) model, in this case Xt satisfies

Xt

p
X

i Xti = t +

i=1

q
X

j tj .

j=1

Definition 2.5.1 (Summary of AR, ARMA and MA models)

(i) The autoregressive AR(p)

model: {Xt } satisfies

Xt =

p
X

i Xti + t .

(2.12)

i=1

Observe we can write it as (B)Xt = t


(ii) The moving average M A(q) model: {Xt } satisfies

Xt = t +

q
X
j=1

Observe we can write Xt = (B)t

62

j tj .

(2.13)

(iii) The autoregressive moving average ARM A(p, q) model: {Xt } satisfies

Xt

p
X

q
X

i Xti = t +

i=1

j tj .

(2.14)

j=1

We observe that we can write Xt as (B)Xt = (B)t .


Below we give conditions for the ARMA to have a causal solution and also be invertible. We
also show that the coefficients of the MA() representation of Xt will decay exponentially.
Lemma 2.5.1 Let us suppose Xt is an ARMA(p, q) process with representation given in Definition
2.5.1.
(i) If the roots of the polynomial (z) lie outside the unit circle, and are greater than (1 + ) (for
some > 0), then Xt almost surely has the solution

Xt =

aj tj ,

(2.15)

j=0

where for j > q, aj = [Aj ]1,1 +

Pq

i=1 i [A

A=

1
1
..
.
0

where

2
0
..
.

ji ]

1,1 ,

with

. . . p1 p
...
..
.

... ...

...
..
.

0
..
.

|aj | < (we note that really aj = aj (, ) since its a function of {i } and {i }).

Moreover for all j,


|aj | Kj

(2.16)

for some finite constant K and 1/(1 + ) < < 1.


(ii) If the roots of (z) lie both inside or outside the unit circle and are larger than (1 + ) or less
than (1 + )1 for some > 0, then we have

Xt =

X
j=

63

aj tj ,

(2.17)

(a vector AR(1) is not possible), where


|aj | K|j|

(2.18)

for some finite constant K and 1/(1 + ) < < 1.


(iii) If the absolute value of the roots of (z) = 1 +

Pq

j=1 j z

are greater than (1 + ), then (2.14)

can be written as
Xt =

bj Xtj + t .

(2.19)

j=1

where
|bj | Kj

(2.20)

for some finite constant K and 1/(1 + ) < < 1.


PROOF. We first prove (i) There are several way to prove the result. The proof we consider here,
uses the VAR expansion given in Section 2.4.1. We write the ARMA process as a vector difference
equation
X t = AX t1 + t
where X 0t = (Xt , . . . , Xtp+1 ), 0t = (t +

Pq

j=1 j tj , 0, . . . , 0).

Xt =

(2.21)
Now iterating (2.21), we have

Aj tj ,

(2.22)

j=0

concentrating on the first element of the vector X t we see that


Xt =

[Ai ]1,1 (ti +

i=0

q
X

j tij ).

j=1

Comparing (2.15) with the above it is clear that for j > q, aj = [Aj ]1,1 +

Pq

i=1 i [A

ji ] .
1,1

Observe

that the above representation is very similar to the AR(1). Indeed as we will show below the Aj
behaves in much the same way as the j in AR(1) example. As with j , we will show that Aj
converges to zero as j (because the eigenvalues of A are less than one). We now show that

64

|Xt | K

j
j=1 |tj |

for some 0 < < 1, this will mean that |aj | Kj . To bound |Xt | we use

(2.22)

|Xt | kX t k2

kAj kspec ktj k2 .

j=0

Hence, by using Section 2.4.1 we have |kAj kspec C j (for any |max (A)| < < 1), which gives
the corresponding bound for |aj |.
To prove (ii) we consider a power series expansion of

(z)
(z) .

If the roots of (z) are distinct,

then it is straightforward to write (z)1 in terms of partial fractions and a convergent power
series for |z| = 1. This expansion immediately gives the the linear coefficients aj and show that
|aj | C(1 + )|j| for some finite constant C. On the other hand, if there are multiple roots, say
P
the roots of (z) are 1 , . . . , s with multiplicity m1 , . . . , ms (where sj=1 ms = p) then we need to
adjust the partial fraction expansion. It can be shown that |aj | C|j|maxs |ms | (1 + )|j| . We note
that for every (1 + )1 < < 1, there exists a constant such that |j|maxs |ms | (1 + )|j| C|j| ,
thus we obtain the desired result.
To show (iii) we use a similar proof to (i), and omit the details.

Corollary 2.5.1 An ARMA process is invertible if the roots of (B) (the MA coefficients) lie
outside the unit circle and causal if the roots of (B) (the AR coefficients) lie outside the unit
circle.
The representation of an ARMA process is unique upto AR and MA polynomials (B) and
(B) having common roots. A simplest example is Xt = t , this also satisfies the representation
Xt Xt1 = t t1 etc. Therefore it is not possible to identify common factors in the
polynomials.
One of the main advantages of the invertibility property is in prediction and estimation. We will
consider this in detail below. It is worth noting that even if an ARMA process is not invertible, one
can generate a time series which has identical correlation structure but is invertible (see Section
3.3).

65

2.6

Simulating from an Autoregressive process

Simulating from a Gaussian AR process


It is straightforward to simulate from an AR process with Gaussian innovations, {t }. Given the
autoregressive structure we can deduce the correlation structure (see Chapter 3) (regardless of
the distribution of the innovations). Furthermore, from Lemma 2.5.1(ii) we observe that all AR
processes can be written as the infinite sum of the innovations. Thus if the innovations are Gaussian,
so is the AR process. This allows us to deduce the joint distribution of X1 , . . . , Xp , which in turn
allows us generate the AR(p) process.
We illustrate the details with with an AR(1) process. Suppose Xt = 1 Xt1 + t where {t }
are iid standard normal random variables (note that for Gaussian processes it is impossible to
discriminate between causal and non-causal processes - see Section 3.3, therefore we will assume
|1 | < 1). We will show in Section 3.1, equation (3.1) that the autocovariance of an AR(1) is
c(r) = r1

2j
1 =

j=0

r1
.
1 21

Therefore, the marginal distribution of Xt is Gaussian with variance (1 21 )1 . Therefore, to


simulate an AR(1) Gaussian time series, we draw from a Gaussian time series with mean zero and
variance (1 21 )1 , calling this X1 . We then iterate for 2 t, Xt = 1 Xt1 + t . This will give us
a stationary realization from an AR(1) Gaussian time series.
Note the function arima.sim is a routine in R which does the above. See below for details.

Simulating from a non-Gaussian AR model


Unlike the Gaussian AR process it is difficult to simulate a non-Gaussian model, but we can obtain
a very close approximation. This is because if the innovations are non-Gaussian but known it is not
clear what the distribution of Xt will be. Here we describe how to obtain a close approximation in
the case that the AR process is causal.
Again we describe the method for the AR(1). Let {Xt } be an AR(1) process, Xt = 1 Xt1 + t ,
which has stationary, causal solution
Xt =

j1 tj .

j=0

66

1 = 0. Then obtain the iteration X


t = 1 X
t1 + t for
To simulate from the above model, we set X
t 2. We note that the solution of this equation is
t =
X

t
X

j1 tj .

j=0

t | |1 |t P |j j |, which converges geometrically


We recall from Lemma 2.5.1 that |Xt X
j=0 1
t ; t n} in the simulations
fast to zero. Thus if we choose a large n to allow burn in and use {X
we have a simulation which is close to a stationary solution from an AR(1) process.

Simulating from an Integrated process


To simulate from an integrated process ARIMA(p, 1, q) (1 B)Yt = Xt , where Xt is a causal
ARMA(p, q) process. We first simulate {Xt } using the method above. Then we define the recursion
Y1 = X1 and for t > 1
Yt = Yt1 + Xt .
Thus giving a realisation from an ARIMA(p, 1, q).

Simulating from a non-Gaussian non-causal model


Suppose that Xt satisfies the representation

Xt =

p
X

j Xtj + t ,

j=1

whose characteristic function have roots both inside and outside the unit circle. Thus, the stationary
solution of this equation is not causal. It is not possible to simulate from this equation. To see why,
consider directly simulating from Xt = 2Xt1 + t without rearranging it as Xt1 = 12 Xt 12 t , the
solution would explode. Now if the roots are both inside and outside the unit circle, there would
not be a way to rearrange the equation to iterate a stationary solution. There are two methods to
remedy this problem:

67

(i) From Lemma 2.5.1(ii) we recall that Xt has the solution

Xt =

aj tj ,

(2.23)

j=

where the coefficients aj are determined from the characteristic equation. Thus to simulate
the process we use the above representation, though we do need to truncate the number of
terms in (2.23) and use

t =
X

M
X

aj tj .

j=M

(ii) The above is a brute force method is an approximation which is also difficult to evaluate.
There is a simpler method, if one studies the roots of the characteristic equation.
Let us suppose that {j1 ; j1 = 1, . . . , p1 } are the roots of (z) which lie outside the unit circle
and {j2 ; j2 = 1, . . . , p2 } are the roots which lie inside the unit circle. For ease of calculation
we will assume the roots are distinct. We can rewrite (z)1 as
(z)1 =

hQ

p1
j1 =1 (1

p1
X
j1 =1
p1
X
j1 =1

1
i hQ
i
p2
j1 z)
(1

z)
j
2
j2 =1

p2
X
djd
cj1
+
(1 j1 z)
(1 jd z)

cj1

(1 j1 z)

j2 =1
p2
X
j2

djd
1
z(1 1
jd z )
=1 jd

Thus the solution of Xt is


Xt = (B)1 t =

p2
X
cj1
djd
t
1 1 t
(1 j1 B)

B(1

j
d
jd B )
j =1
=1

p1
X
j1

Let Yj1 ,t = j1 Yj1 ,t1 + t and Zj2 ,t = j2 Zj2 ,t1 + t (thus the stationary solution is gener1
ated with Zj2 ,t1 = 1
j2 Zj2 ,t j2 t ). Generate the time series {Yj1 ,t ; j1 = 1, . . . , p1 } and

{Yj1 ,t ; j1 = 1, . . . , p1 } using the method described above. Then the non-causal time series
can be generated by using

Xt =

p1
X

cj1 Yj1 ,t

j1 =1

p2
X
j2 =1

68

dj2 Zj2 ,t .

Comments:
Remember Yj,t is generated using the past t and Zj,t is generated using future innovations. Therefore to ensure that the generated {Yj,t } and {Zj,t } are close to the
stationary we need to ensure that the initial value of Yj,t is far in the past and the
initial value for Zj,t is far in the future.
If the roots are complex conjugates, then the corresponding {Yj,t } or {Zj,t } should be
written as AR(2) models (to avoid complex processes).

R functions
Shumway and Stoffer (2006) and David Stoffers website gives a comprehensive introduction to time
series R-functions.
The function arima.sim simulates from a Gaussian ARIMA process. For example,
arima.sim(list(order=c(2,0,0), ar = c(1.5, -0.75)), n=150) simulates from the AR(2) model
Xt = 1.5Xt1 0.75Xt2 + t , where the innovations are Gaussian.
Exercise 2.6 In the following simulations, use non-Gaussian innovations.
(i) Simulate an AR(4) process with characteristic function


2
(z) = 1 0.8 exp(i )z
13





2
2
2
1 0.8 exp(i )z 1 1.5 exp(i )z 1 1.5 exp(i )z .
13
5
5

(ii) Simulate an AR(4) process with characteristic function




2
(z) = 1 0.8 exp(i )z
13





2
2
2
2
2
1 0.8 exp(i )z 1 exp(i )z 1 exp(i )z .
13
3
5
3
5

Do you observe any differences between these realisations?

69

Chapter 3
The autocovariance function of a
linear time series
Objectives
Be able to determine the rate of decay of an ARMA time series.
Be able solve the autocovariance structure of an AR process.
Understand what partial correlation is and how this may be useful in determining the order
of an AR model.
Understand why autocovariance is blind to processes which are non-causal. But the higher
order cumulants are not blind to causality.

3.1

The autocovariance function

The autocovariance function (ACF) is defined as the sequence of covariances of a stationary process.
That is suppose that {Xt } is a stationary process with mean zero, then {c(k) : k Z} is the ACF
of {Xt } where c(k) = E(X0 Xk ). Clearly different time series give rise to different features in the
ACF. We will explore some of these features below.
Before investigating the structure of ARMA processes we state a general result connecting linear
time series and the summability of the autocovariance function.

70

Lemma 3.1.1 Suppose the stationary time series Xt satisfies the linear representation
P
The covariance is c(r) =
j= j j+r .
(i) If

(ii) If

(iii) If

j= |j |

< , then

j= |jj |
j= |j |

j= j tj .

|c(k)| < .

< , then

|k c(k)| < .

< , then we cannot say anything about summability of the covariance.

PROOF. It is straightforward to show that


c(k) = var[t ]

j jk .

Using this result, it is easy to see that

|c(k)|

P P
k

|j | |jk |, thus

|c(k)| < , which

proves (i).
P
The proof of (ii) is similar. To prove (iii), we observe that j |j |2 < is a weaker condition
P
then j |j | < (for example the sequence j = |j|1 satisfies the former condition but not the
latter). Thus based on the condition we cannot say anything about summability of the covariances.

First we consider a general result on the covariance of a causal ARMA process (always to obtain
the covariance we use the MA() expansion - you will see why below).

3.1.1

The rate of decay of the autocovariance of an ARMA process

We evaluate the covariance of an ARMA process using its MA() representation. Let us suppose
that {Xt } is a causal ARMA process, then it has the representation in (2.17) (where the roots of
(z) have absolute value greater than 1 + ). Using (2.17) and the independence of {t } we have

cov(Xt , X ) = cov(

aj1 tj1 ,

j1 =0

aj2 j2 )

j2 =0

aj1 aj2 cov(tj , j ) =

j=0

aj aj+|t | var(t )

(3.1)

j=0

(here we see the beauty of the MA() expansion). Using (2.18) we have

|cov(Xt , X )|

var(t )C2

j j+|t |

j=0

C2 |t |

X
j=0

71

2j =

|t |
,
1 2

(3.2)

for any 1/(1 + ) < < 1.


The above bound is useful, it tells us that the ACF of an ARMA process decays exponentially
fast. In other words, there is very little memory in an ARMA process. However, it is not very
enlightening about features within the process. In the following we obtain an explicit expression for
the ACF of an autoregressive process. So far we have used the characteristic polynomial associated
with an AR process to determine whether it was causal. Now we show that the roots of the
characteristic polynomial also give information about the ACF and what a typical realisation of
a autoregressive process could look like.

3.1.2

The autocovariance of an autoregressive process

Let us consider the zero mean AR(p) process {Xt } where

Xt =

p
X

j Xtj + t .

(3.3)

j=1

From now onwards we will assume that {Xt } is causal (the roots of (z) lie outside the unit circle).
Given that {Xt } is causal we can derive a recursion for the covariances. It can be shown that
multipying both sides of the above equation by Xtk (k 0) and taking expectations, gives the
equation

E(Xt Xtk ) =

p
X
j=1

j E(Xtj Xtk ) + E(t Xtk ) =


| {z }
=0

p
X

j E(Xtj Xtk ).

(3.4)

j=1

It is worth mentioning that if the process were not causal this equation would not hold, since t
and Xtk are not necessarily independent. These are the Yule-Walker equations, we will discuss
them in detail when we consider estimation. For now letting c(k) = E(X0 Xk ) and using the above
we see that the autocovariance satisfies the homogenuous difference equation

c(k)

p
X

j c(k j) = 0,

(3.5)

j=1

for k 0. In other words, the autocovariance function of {Xt } is the solution of this difference
equation. The study of difference equations is a entire field of research, however we will now scratch
the surface to obtain a solution for (3.5). Solving (3.5) is very similar to solving homogenuous
differential equations, which some of you may be familar with (do not worry if you are not).

72

Recall the characteristic polynomial of the AR process (z) = 1

Pp

j=1 j z

= 0, which has

the roots 1 , . . . , p . In Section 2.3.4 we used the roots of the characteristic equation to find the
stationary solution of the AR process. In this section we use the roots characteristic to obtain the
solution (3.5). It can be shown if the roots are distinct (the roots are all different) the solution of
(3.5) is

c(k) =

p
X

Cj k
j ,

(3.6)

j=1

where the constants {Cj } are chosen depending on the initial values {c(k) : 1 k p} and are
such that they ensure that c(k) is real (recalling that j ) can be complex.
The simplest way to prove (3.6) is to use a plugin method. Plugging c(k) =

Pp

k
j=1 Cj j

into

(3.5) gives

c(k)

p
X

j c(k j) =

j=1

p
X



p
X
(ki)
k
Cj j
i j

j=1

p
X

i=1



p
X
k
i
Cj j
1
i j = 0.

j=1

i=1

{z

(i )

In the case that the roots of (z) are not distinct, let the roots be 1 , . . . , s with multiplicity
P
m1 , . . . , ms ( sk=1 mk = p). In this case the solution is
c(k) =

s
X

k
j Pmj (k),

(3.7)

j=1

where Pmj (k) is mj th order polynomial and the coefficients {Cj } are now hidden in Pmj (k). We
now study the covariance in greater details and see what it tells us about a realisation. As a
motivation consider the following example.
Example 3.1.1 Consider the AR(2) process
Xt = 1.5Xt1 0.75Xt2 + t ,

(3.8)

where {t } are iid random variables with mean zero and variance one. The corresponding characp
teristic polynomial is 1 1.5z + 0, 75z 2 , which has roots 1 i31/2 = 4/3 exp(i/6). Using the

73

discussion above we see that the autocovariance function of {Xt } is


p
c(k) = ( 4/3)k (C1 exp(ik/6) + C1 exp(ik/6)),
for a particular value of C1 . Now write C1 = a exp(ib), then the above can be written as


p
c(k) = a( 4/3)k cos k + b .
6
We see that the covariance decays at an exponential rate, but there is a periodicity within the decay.
This means that observations separated by a lag k = 12 are more closely correlated than other lags,
this suggests a quasi-periodicity in the time series. The ACF of the process is given in Figure 3.1,
notice that it has decays to zero but also observe that it undulates. A plot of a realisation of the time
series is given in Figure 3.2, notice the quasi-periodicity of about 2/12. Let is now briefly return
P
to the definition of the periodogram given in Section 1.2.3 (In () = n1 | nt=1 Xt exp(it)|2 ). We
used the periodogram to identify the periodogram of a deterministic signal, but showed that when
dependent, correlated noise was added to the dignal that the periodogram exhibited more complex
behaviour. In Figure 6.1 we give a plot of the periodogram corresponding to Figure 3.2. Recall
that this AR(2) gives a quasi-periodicity of 12, which corresponds to the frequency 2/12 0.52,
which matches the main peaks in periodogram. We will learn later that the periodogram is a crude
(meaning inconsistent) estimator of the spectral density function. The spectral density if given in
the lower plot of Figure 6.1.
We now generalise the above example. Let us consider the general AR(p) process defined in (3.3).
Suppose the roots of the corresponding characteristic polynomial are distinct and let us split them
into real and complex roots. Because the characteristic polynomial is comprised of real coefficients,
the complex roots come in complex conjugate pairs. Hence let us suppose the real roots are {j }rj=1
(pr)/2

and the complex roots are {j , j }j=r+1 . The covariance in (3.6) can be written as

c(k) =

r
X
j=1

(p2)/2

Cj k
j

aj |j |k cos(kj + bj )

(3.9)

j=r+1

where for j > r we write j = |j | exp(ij ) and aj and bj are real constants. Notice that as the
example above the covariance decays exponentially with lag, but there is undulation. A typical
realisation from such a process will be quasi-periodic with periods at r+1 , . . . , (pr)/2 , though the

74

1.0
0.8
0.6
0.4
0.4

0.2

0.0

0.2

acf

10

20

30

40

50

lag

0
2
4

ar2

Figure 3.1: The ACF of the time series Xt = 1.5Xt1 0.75Xt2 + t

24

48

72

96

120

144

Time

Figure 3.2: The a simulation of the time series Xt = 1.5Xt1 0.75Xt2 + t

75

0.5

0.3

0.2

Periodogram

0.4

0.1

0.0

frequency

30
0

10

20

spectrum

40

50

60

Autoregressive

0.0

0.1

0.2

0.3

0.4

0.5

frequency

Figure 3.3: Top: Periodogram of Xt = 1.5Xt1 0.75Xt2 + t for sample size n = 144.
Lower: The corresponding spectral density function (note that 0.5 of the x-axis on spectral
density corresponds to on the x-axis of the periodogram).

76

magnitude of each period will vary.


An interesting discussion on covariances of an AR process and realisation of an AR process is
given in Shumway and Stoffer (2006), Chapter 3.3 (it uses the example above). A discussion of
difference equations is also given in Brockwell and Davis (1998), Sections 3.3 and 3.6 and Fuller
(1995), Section 2.4.
Example 3.1.2 (Autocovariance of an AR(2)) Let us suppose that Xt satisfies the model Xt =
(a + b)Xt1 abXt2 + t . We have shown that if |a| < 1 and |b| < 1, then it has the solution

Xt =

X

1
bj+1 aj+1 )tj .
ba
j=0

By writing a timeline it is straightfoward to show that for r > 1


cov(Xt , Xtr ) =

X
(bj+1 aj+1 )(bj+1+r aj+1+r ).
j=0

Example 3.1.3 The autocorrelation of a causal and noncausal time series Let us consider the two
AR(1) processes considered in Section 2.3.2. We recall that the model
Xt = 0.5Xt1 + t
has the stationary causal solution
Xt =

0.5j tj .

j=0

Assuming the innovations has variance one, the ACF of Xt is


cX (0) =

1
1 0.52

cX (k) =

On the other hand the model


Yt = 2Yt1 + t

77

0.5|k|
1 0.52

has the noncausal stationary solution


Yt =

X
(0.5)j+1 t+j+1 .
j=0

Thus process has the ACF


cY (0) =

0.52
1 0.52

cX (k) =

0.52+|k|
.
1 0.52

Thus we observe that except for a factor (0.5)2 both models has an identical autocovariance function.
Indeed their autocorrelation function would be same. Furthermore, by letting the innovation of Xt
have standard deviation 0.5, both time series would have the same autocovariance function.
Therefore, we observe an interesting feature, that the non-causal time series has the same
correlation structure of a causal time series. In Section 3.3 that for every non-causal time series
there exists a causal time series with the same autocovariance function. Therefore autocorrelation
is blind to non-causality.
Exercise 3.1 Recall the AR(2) models considered in Exercise 2.4. Now we want to derive their
ACF functions.
(i)

(a) Obtain the ACF corresponding to


2
7
Xt = Xt1 Xt2 + t ,
3
3
where {t } are iid random variables with mean zero and variance 2 .
(b) Obtain the ACF corresponding to

4 3
42
Xt =
Xt1 2 Xt2 + t ,
5
5
where {t } are iid random variables with mean zero and variance 2 .
(c) Obtain the ACF corresponding to
Xt = Xt1 4Xt2 + t ,
where {t } are iid random variables with mean zero and variance 2 .

78

(ii) For all these models plot the true ACF in R. You will need to use the function ARMAacf.
BEWARE of the ACF it gives for non-causal solutions. Find a method of plotting a causal
solution in the non-causal case.
Exercise 3.2 In Exercise 2.5 you constructed a causal AR(2) process with period 17.
Load Shumway and Stoffers package asta into R (use the command install.packages("astsa")
and then library("astsa").
Use the command arma.spec to make a plot of the corresponding spectral density function. How
does your periodogram compare with the true spectral density function?

R code
We use the code given in Shumway and Stoffer (2006), page 101 to make Figures 3.1 and 3.2.
To make Figure 3.1:
acf = ARMAacf(ar=c(1.5,-0.75),ma=0,50)
plot(acf,type="h",xlab="lag")
abline(h=0)
To make Figures 3.2 and 6.1:
set.seed(5)
ar2 <- arima.sim(list(order=c(2,0,0), ar = c(1.5, -0.75)), n=144)
plot.ts(ar2, axes=F); box(); axis(2)
axis(1,seq(0,144,24))
abline(v=seq(0,144,12),lty="dotted")
plot(frequency, Periodogram,type="o")
library("astsa")
arma.spec( ar = c(1.5, -0.75), log = "no", main = "Autoregressive")

3.1.3

The autocovariance of a moving average process

Suppose that {Xt } satisfies

Xt = t +

q
X
j=1

79

j tj .

The covariance is

cov(Xt , Xtk ) =

P
p

i=0 i ik

k = q, . . . , q

otherwise

where 0 = 1 and i = 0 for i < 0 and i q. Therefore we see that there is no correlation when
the lag between Xt and Xtk is greater than q.

3.1.4

The autocovariance of an autoregressive moving average process

We see from the above that an MA(q) model is only really suitable when we believe that there
is no correlaton between two random variables separated by more than a certain distance. Often
autoregressive models are fitted. However in several applications we find that autoregressive models
of a very high order are needed to fit the data. If a very long autoregressive model is required
a more suitable model may be the autoregressive moving average process. It has several of the
properties of an autoregressive process, but can be more parsimonuous than a long autoregressive
process. In this section we consider the ACF of an ARMA process.
Let us suppose that the causal time series {Xt } satisfies the equations

Xt

p
X

i Xti = t +

i=1

q
X

j tj .

j=1

We now define a recursion for ACF, which is similar to the ACF recursion for AR processes. Let
us suppose that the lag k is such that k > q, then it can be shown that the autocovariance function
of the ARMA process satisfies

E(Xt Xtk )

p
X

i E(Xti Xtk ) = 0

i=1

On the other hand, if k q, then we have

E(Xt Xtk )

p
X
i=1

i E(Xti Xtk ) =

q
X

j E(tj Xtk ) =

j=1

We recall that Xt has the MA() representation Xt =

80

q
X

j E(tj Xtk ).

j=k

j=0 aj tj

(see (2.17)), therefore for

k j q we have E(tj Xtk ) = ajk var(t ) (where a(z) = (z)(z)1 ). Altogether the above
gives the difference equations

c(k)

c(k)

p
X
i=1
p
X

i c(k i) = var(t )

q
X

j ajk

for 1 k q

(3.10)

j=k

i c(k i) = 0, for k > q,

i=1

where c(k) = E(X0 Xk ). (3.10) is homogenuous difference equation, then it can be shown that the
solution is
c(k) =

s
X

k
j Pmj (k),

j=1

P
where 1 , . . . , s with multiplicity m1 , . . . , ms ( k ms = p) are the roots of the characteristic
P
polynomial 1 pj=1 j z j . Observe the similarity to the autocovariance function of the AR process
(see (3.7)). The coefficients in the polynomials Pmj are determined by the initial condition given
in (3.10).
You can also look at Brockwell and Davis (1998), Chapter 3.3 and Shumway and Stoffer (2006),
Chapter 3.4.

3.2

The partial covariance and correlation of a time


series

We see that by using the autocovariance function we are able to identify the order of an MA(q)
process: when the covariance lag is greater than q the covariance is zero. However the same is
not true for AR(p) processes. The autocovariances do not enlighten us on the order p. However
a variant of the autocovariance, called the partial autocovariance is quite informative about order
of AR(p). We start by reviewing the partial autocovariance, and its relationship to the inverse
variance/covariance matrix (often called the precision matrix).

81

3.2.1

A review of partial correlation in multivariate analysis

Partial correlation
Suppose X = (X1 , . . . , Xd ) is a zero mean random vector (we impose the zero mean condition to
simplify notation and its not necessary). The partial correlation is the covariance between Xi and
Xj , conditioned on the other elements in the vector. In other words, the covariance between the
residuals of Xi conditioned on X (ij) (the vector not containing Xi and Xj ) and the residual of Xj
conditioned on X (ij) . That is the partial covariance between Xi and Xj given X (ij) is defined
as
cov Xi var[X (ij) ]1 E[X (ij) Xi ]X (ij) , Xj var[X (ij) ]1 E[X (ij) Xj ]X (ij)

= cov[Xi Xj ] E[X (ij) Xi ]0 var[X (ij) ]1 E[X (ij) Xj ].


Taking the above argument further, the variance/covariance matrix of the residual of X ij =
(Xi , Xj )0 given X (ij) is defined as

var X ij E[X ij X (ij) ]0 var[X (ij) ]1 X (ij) = ij c0ij 1
(ij) cij

(3.11)

where ij = var(X ij ), cij = E(X ij X (ij) ) (=cov(X ij , X (ij) )) and (ij) = var(X (ij) )
( denotes the tensor product). Let sij denote the (i, j)th element of the (2 2) matrix ij
c0ij 1
(ij) cij . The partial correlation between Xi and Xj given X (ij) is
ij =

s12
,
s11 s22

observing that
(i) s12 is the partial covariance between Xi and Xj .
(ii) s11 = E(Xi

k6=i,j

i,k Xk )2 (where i,k are the coefficients of the best linear predictor of

Xi given {Xk ; k 6= i, j}).


(ii) s22 = E(Xj

k6=i,j

j,k Xk )2 (where j,k are the coefficients of the best linear predictor of

Xj given {Xk ; k 6= i, j}).


In the following section we relate partial correlation to the inverse of the variance/covariance matrix
(often called the precision matrix).

82

The precision matrix and its properties


Let us suppose that X = (X1 , . . . , Xd ) is a zero mean random vector with variance . The (i, j)th
element of the covariance cov(Xi , Xj ) = ij . Here we consider the inverse of , and what
information the (i, j)th of the inverse tells us about the correlation between Xi and Xj . Let ij
denote the (i, j)th element of 1 . We will show that with appropriate standardisation, ij is the
negative partial correlation between Xi and Xj . More precisely,

ij
ii jj

= ij .

(3.12)

The proof uses the inverse of block matrices. To simplify the notation, we will focus on the (1, 2)th
element of and 1 (which concerns the correlation between X1 and X2 ). Let X 1,2 = (X1 , X2 )0 ,
X (1,2) = (X3 , . . . , Xd )0 , (1,2) = var(X (1,2) ), c1,2 = cov(X (1,2) , X (1,2) ) and 1,2 = var(X 1,2 ).
Using this notation it is clear that

var(X) = =

1,2

c1,2

c01,2

(1,2)

(3.13)

It is well know that the inverse of the above block matrix is

1 =

P 1 c01,2 1
(1,2)

P 1

1 P 1 + 1
1 c0 1
1
1,2 (1,2)
(1,2) c1,2 P
(1,2) c1,2 P

(3.14)

where P = (1,2 c01,2 1


(1,2) c1,2 ). Comparing P with (3.11), we see that P is the 2 2 variance/covariance matrix of the residuals of X(1,2) conditioned on X (1,2) . Thus the partial correlation
between X1 and X2 is
P1,2
1,2 = p
P1,1 P2,2

(3.15)

where Pij denotes the elements of the matrix P . Inverting P (since it is a two by two matrix), we
see that

P 1 =

P2,2

2
P1,1 P2,2 P1,2
P1,2

83

P1,2
P11

(3.16)

Thus, by comparing (3.14) and (3.16) and by the definition of partial correlation given in (3.15) we
have
1
P1,2
= 1,2 .

Let ij denote the (i, j)th element of 1 . Thus we have shown (3.12):
ij =

ij
ii jj

In other words, the (i, j)th element of 1 divided by the square root of its diagonal gives negative
partial correlation. Therefore, if the partial correlation between Xi and Xj given Xij is zero, then
i,j = 0.
The precision matrix, 1 , contains many other hidden treasures. For example, the coefficients
of 1 convey information about the best linear predictor Xi given X i = (X1 , . . . , Xi1 , Xi+1 , . . . , Xd )
(all elements of X except Xi ). Let
Xi =

i,j Xj + i ,

j6=i

where {i,j } are the coefficients of the best linear predictor. Then it can be shown that
i,j =

ij
ii

and

ii =

E[Xi

1
P

j6=i i,j Xj ]

(3.17)

The proof uses the same arguments as those in (3.13).


Exercise 3.3 By using the decomposition

var(X) = =

c1

c01

(1)

(3.18)

where 1 = var(X1 ), c1 = E[X1 X 01 ] and (1) = var[X 1 ] prove (3.17).

The Cholesky decomposition and the precision matrix


We now represent the precision matrix through its Cholesky decomposition. It should be mentioned
that Mohsen Pourahmadi has done a lot of interesting research in this area and he recently wrote

84

a review paper, which can be found here.


We define the sequence of linear equations

Xt =

t1
X

t,j Xj + t ,

t = 2, . . . , k,

(3.19)

j=1

where {t,j ; 1 j t 1} are the coefficeints of the best linear predictor of Xt given X1 , . . . , Xt1 .
P
2
2
Let t2 = var[t ] = E[Xt t1
j=1 t,j Xj ] and 1 = var[X1 ]. We standardize (3.19) and define
t
X

t,j Xj =

j=1

1
Xt
t

t1
X

t,j Xj ,

(3.20)

j=1

where we note that t,t = t1 and for 1 j < t 1, t,j = t,j /i . By construction it is clear
that var(LX) = Ik , where

1,1

...

0 ...
0
0
2,1 2,2

L = 3,1 3,2 3,3 . . .


0
0

..
..
..
..
..
..
.
.
.
.
.
.

k,1 k,2 k,3 . . . k,k1 k,k

(3.21)

and LL0 = 1 (see ?, equation (18)), where = var(X k ). Let t = var[X t ], then
ij
t

t
X

ik jk .

k=1

We use apply these results to the analysis of the partial correlations of autoregressive processes
and the inverse of its variance/covariance matrix.

3.2.2

Partial correlation in time series

The partial covariance/correlation of a time series is defined in a similar way.


Definition 3.2.1 The partial covariance/correlation between Xt and Xt+k+1 is defined as the partial covariance/correlation between Xt and Xt+k+1 after conditioning out the inbetween time series
Xt+1 , . . . , Xt+k .

85

We now obtain an expression for the partial correlation between Xt and Xt+k+1 in terms of their
autocovariance function (for the final result see equation (3.22)). As the underlying assumption
is that the time series is stationary it is the same as the partial covariance/correlation Xk+1 and
X0 . In Chapter 5 we will introduce the idea of linear predictor of a future time point given the
present and the past (usually called forecasting) this can be neatly described using the idea of
projections onto subspaces. This notation is quite succinct, therefore we derive an expression for
the partial correlation using projection notation. The projection of Xk+1 onto the space spanned
by X k = (X1 , X2 , . . . , Xk ), is the best linear predictor of Xk+1 given X k . We will denote the
projection of Xk onto the space spanned by X1 , X2 , . . . , Xk as PX k (Xk+1 ) (note that this is the
same as the best linear predictor). Thus

PX k (Xk+1 ) =

X 0k (var[X k ]1 E[Xk+1 X k ])1

X 0k 1
k ck

:=

k
X

k,j Xj ,

j=1

where k = var(X k ) and ck = E(Xk+1 X k ). To derive a similar expression for PX k (X0 ) we use the
stationarity property
PX k (Xk+1 ) = X 0k (var[X k ]1 E[X0 X k ])
= X 0k (var[X k ]1 Ek E[Xk+1 X k ])
1
0
= X 0k 1
k Ek ck = X k Ek k ck :=

k
X

k,k+1j Xj ,

j=1

where Ek is a matrix which swops round all the elements in a vector

Ek =

0 0 0 ... 0 1
0 0 0 ... 1 0
.. .. .. .. ..
. . . . .
.
1 0 .. 0 0 0

Thus the partial correlation between Xt and Xt+k (where k > 0) is the correlation X0 PX k (X0 )
and Xk+1 PX k (Xk+1 ), which is
cov(Xk+1 PX k (Xk+1 ), X0 PX k (X0 ))
= cov(Xk+1 X0 ) c0k 1
k Ek ck .

86

(3.22)

We consider an example.
Example 3.2.1 (The PACF of an AR(1) process) Consider the causal AR(1) process Xt =
0.5Xt1 + t where E(t ) = 0 and var(t ) = 1. Using (3.1) it can be shown that cov(Xt , Xt2 ) =
20.52 (compare with the MA(1) process Xt = t +0.5t1 , where the covariance cov(Xt , Xt2 ) = 0).
We evaluate the partial covariance between Xt and Xt2 . Remember we have to condition out the
random variables inbetween, which in this case is Xt1 . It is clear that the projection of Xt onto
Xt1 is 0.5Xt1 (since Xt = 0.5Xt1 + t ). Therefore Xt Psp(X
Xt = Xt 0.5Xt1 = t . The

t1 )
projection of Xt2 onto Xt1 is a little more complicated, it is Psp(X
Xt2 =

t1 )

E(Xt1 Xt2 )
Xt1 .
2 )
E(Xt1

Therefore the partial correlation between Xt and Xt2


cov Xt PXt1 Xt , Xt2 PXt1 ) Xt2

E(Xt1 Xt2 )
Xt1
= cov t , Xt2
2 )
E(Xt1


= 0.

In fact the above is true for the partial covariance between Xt and Xtk , for all k 2. Hence we
see that despite the covariance not being zero for the autocovariance of an AR process greater than
order two, the partial covariance is zero for all lags greater than or equal to two.
Using the same argument as above, it is easy to show that partial covariance of an AR(p) for
lags greater than p is zero. Hence in may respects the partial covariance can be considered as an
analogue of the autocovariance. It should be noted that though the covariance of MA(q) is zero
for lag greater than q, the same is not true for the parial covariance. Whereas partial covariances
removes correlation for autoregressive processes it seems to add correlation for moving average
processes!
Model identification:
If the autocovariances after a certain lag are zero q, it may be appropriate to fit an MA(q)
model to the time series.
On the other hand, the autocovariances of any AR(p) process will only decay to zero as the
lag increases.
If the partial autocovariances after a certain lag are zero p, it may be appropriate to fit an
AR(p) model to the time series.
On the other hand, the partial covariances of any MA(p) process will only decay to zero as
the lag increases.

87

Exercise 3.4 (The partial correlation of an invertible MA(1)) Let t,t denote the partial correlation between Xt+1 and X1 . It is well known (this is the Levinson-Durbin algorithm, which we
cover in Chapter 5) that t,t can be deduced recursively from the autocovariance funciton using the
algorithm:
Step 1 1,1 = c(1)/c(0) and r(2) = E[X2 X2|1 ]2 = E[X2 1,1 X1 ]2 = c(0) 1,1 c(1).
Step 2 For j = t

t,t =
t,j

c(t)

= t1,j

Pt1

j=1 t1,j c(t

r(t)
t,t t1,tj

j)
1 j t 1,

and r(t + 1) = r(t)(1 2t,t ).

(i) Using this algorithm show that the PACF of the MA(1) process Xt = t + t1 , where || < 1
(so it is invertible) is
t,t =

(1)t+1 ()t (1 2 )
.
1 2(t+1)

(ii) Explain how this partial correlation is similar to the ACF of the AR(1) model Xt = Xt1 +
t .
Exercise 3.5 (Comparing the ACF and PACF of an AR process) Compare the below plots:
(i) Compare the ACF and PACF of the AR(2) model Xt = 1.5Xt1 0.75Xt2 + t using
ARIMAacf(ar=c(1.5,-0.75),ma=0,30) and ARIMAacf(ar=c(1.5,-0.75),ma=0,pacf=T,30).
(ii) Compare the ACF and PACF of the MA(1) model Xt = t 0.5t using ARIMAacf(ar=0,ma=c(-1.5),30)
and ARIMAacf(ar=0,ma=c(-1.5),pacf=T,30).
(ii) Compare the ACF and PACF of the ARMA(2, 1) model Xt 1.5Xt1 + 0.75Xt2 = t 0.5t
using ARIMAacf(ar=c(1.5,-0.75),ma=c(-1.5),30) and
ARIMAacf(ar=c(1.5,0.75),ma=c(-1.5),pacf=T,30).
Exercise 3.6 Compare the ACF and PACF plots of the monthly temperature data from 1996-2014.
Would you fit an AR, MA or ARMA model to this data?

88

Rcode
The sample partial autocorrelation of a time series can be obtained using the command pacf.
However, remember just because the sample PACF is not zero, does not mean the true PACF is
non-zero. This is why we require error bars!

3.2.3

The variance/covariance matrix and precision matrix of an


autoregressive and moving average process

Let us suppose that {Xt } is a stationary time series. In this section we consider the variance/covariance matrix var(X k ) = k , where X k = (X1 , . . . , Xk )0 . We will consider two cases (i) when
Xt follows an MA(p) models and (ii) when Xt follows an AR(p) model. The variance and inverse
of the variance matrices for both cases yield quite interesting results. We will use classical results
from multivariate analysis, stated in Section 3.2.1.
We recall that the variance/covariance matrix of a stationary time series has a (symmetric)
Toeplitz structure (see wiki for a definition). Let X k = (X1 , . . . , Xk )0 , then

k = var(X k ) =

c(0)

c(1)

. . . c(k 2) c(k 1)

c(1) . . . c(k 3) c(k 2)


..
..
..
.
.
.
..
c(k 1) c(k 2)
.
...
c(1)
c(0)
c(1)
..
.

c(0)
..
.

k for AR(p) and MA(p) models


(i) If {Xt } satisfies an MA(p) model and k > p, then k will be bandlimited, where p offdiagonals above and below the diagonal will be non-zero and the rest of the off-diagonal will
be zero.
(ii) If {Xt } satisfies an AR(p) model, then k will not be bandlimited.

1
k for an AR(p) model
We now consider the inverse of k . Warning: note that the inverse of a Toeplitz is not necessarily
Toeplitz (unlike the circulant which is). We use the results in Section 3.2.1. Suppose that we have
an AR(p) process and we consider the precision matrix of X k = (X1 , . . . , Xk ), where k > p.

89

Recall the (i, j)th element of 1


k divided by the square roots of the corresponding diagonals is
the negative partial correlation of between Xi and Xj conditioned on all the elements in X k . In
Section 3.2.2 we showed that if |i j| > p, then the partial correlation between Xi and Xj given
Xi+1 , . . . , Xj1 (assuming without loss of generality that i < j) is zero. We now show that the
ij
precision matrix of 1
k will be bandlimited (note that it is not immediate obvious since k is the

negative partial correlation between Xi and Xj given X(ij) not just the elements between Xi and
Xj ). To show this we use the Cholesky decomposition given in (3.19). Since Xt is an autoregressive
process of order p and plugging this information into (3.19), for t > p we have

Xt =

t1
X

t,j Xj + t =

j=1

p
X

j Xtj + t ,

j=1

thus t,tj = j for 1 j p otherwise t,tj = 0. Moreover, for t > p we have t2 = var(t ) = 1.
For t p we use the same notation as that used in (3.19). This gives the lower triangular pbandlimited matrix

1,1

2,1
2,2

..
..
.
.

p p1

.
..
.
Lk =
.
.

0
0

0
0

.
..
.
.
.

0
0

...

...

...
..
.

0
..
.

0
..
.

...
..
.

0
..
.

0
..
.

1
..
.

...
..
.

0
..
.

0
..
.

. . . 1
..
..
.
.

. . . p p1 . . . 1
...
..
.

0
..
.

p
..
.

...

. . . 2 1
..
..
..
.
.
.
...

0 ... 0

0 ... 0

.. .. ..
. . .

0 ... 0

.. .. ..
. . .

0 ... 0

1 ... 0

.. .. ..
. . .

0 ... 1

(3.23)

(the above matrix has not been formated well, but after the first p 1 rows, there are ones along
the diagonal and the p lower off-diagonals are non-zero).
0
We recall that 1
k = Lk Lk , thus we observe that since Lk is a lower triangular bandlimited
0
matrix, 1
k = Lk Lk is a bandlimited matrix with the p off-diagonals either side of the diagonal
(i,j) = 0 if |i j| > p.
non-zero. Let ij denote the (i, j)th element of 1
k . Then we observe that
P
Moreover, if 0 < |i j| p and either i or j is greater than p, then ij = 2 pk=|ij| k k|ij|+1

|ij| .

90

The coefficients (i,j) gives us a fascinating insight into the prediction of Xt given the past
and future observations. We recall from equation (3.17) that ij /ii are the coffficients of the
best linear predictor of Xi given X i . This result tells if the observations came from a stationary
AR(p) process, then the best linear predictor of Xi given Xi1 , . . . , Xia and Xi+1 , . . . , Xi+b (where
a and b > p) is the same as the best linear predictor of Xi given Xi1 , . . . , Xip and Xi+1 , . . . , Xi+p
(knowledge of other values will not improve the prediction).
There is an interesting duality between the AR and MA model which we will explore further
in the course.

3.3

Correlation and non-causal time series

Here we demonstrate that it is not possible to identify whether a process is noninvertible/noncausal


from its covariance structure. The simplest way to show result this uses the spectral density
function, which will now define and then return to and study in depth in Chapter 8.
Definition 3.3.1 (The spectral density) Given the covariances c(k) (with

|c(k)|2 < ) the

spectral density function is defined as


f () =

c(k) exp(ik).

The covariances can be obtained from the spectral density by using the inverse fourier transform
c(k) =

1
2

f () exp(ik).
0

Hence the covariance yields the spectral density and visa-versa.


For reference below, we point out that the spectral density function uniquely identifies the autocovariance function.
Let us suppose that {Xt } satisfies the AR(p) representation

Xt =

p
X

i Xti + t

i=1

where var(t ) = 1 and the roots of (z) = 1

Pp

j=1 j z

can lie inside and outside the unit circle,

but not on the unit circle (thus it has a stationary solution). We will show in Chapter 8 that the

91

spectral density of this AR process is


f () =

|1

1
.
2
j=1 j exp(ij)|

(3.24)

Pp

Factorizing f ().
P
Let us supose the roots of the characteristic polynomial (z) = 1 + qj=1 j z j are {j }pj=1 ,
P
Q
thus we can factorize (x) 1 + pj=1 j z j = pj=1 (1 j z). Using this factorization we have
(3.24) can be written as
f () =

1
.
2
j=1 |1 j exp(i)|

(3.25)

Qp

As we have not assumed {Xt } is causal, the roots of (z) can lie both inside and outside the
unit circle. We separate the roots, into those outside the unit circle {O,j1 ; j1 = 1, . . . , p1 }
and inside the unit circle {I,j2 ; j2 = 1, . . . , p2 } (p1 + p2 = p). Thus

(z) = [

p1
Y

(1 O,j1 z)][

j1 =1

p2
Y

(1 I,j2 z)]

j2 =1
p
1
Y

p2
Y

j1 =1

j2 =1

= (1)p2 I,j2 z p2 [

(1 O,j1 z)][

(1 1
I,j2 z)].

(3.26)

Thus we can rewrite the spectral density in (3.27)


f () =

1
1
.
Qp 1
Qp2
1
2
2
2
j2 =1 |I,j2 |
j1 =1 |1 O,j exp(i)|
j2 =1 |1 I,j2 exp(i)|

Qp 2

(3.27)

Let
fO () =

Then f () =

Qp 1

j1 =1 |1

O,j

1
Qp 2

exp(i)|2

j2 =1 |1

2
1
I,j2 exp(i)|

Qp 2

2
j2 =1 |I,j2 | fO ().

A parallel causal AR(p) process with the same covariance structure always exists.
We now define a process which has the same autocovariance function as {Xt } but is causal.

92

Using (3.26) we define the polynomial

e
(z)
=[

p1
Y

(1 O,j1 z)][

j1 =1

p2
Y

(1 1
I,j2 z)].

(3.28)

j2 =1

By construction, the roots of this polynomial lie outside the unit circle. We then define the
AR(p) process
e
e t = t ,
(B)
X

(3.29)

et } has a stationary, almost sure unique solution. Morefrom Lemma 2.3.1 we know that {X
over, because the roots lie outside the unit circle the solution is causal.
et } is fe(). We know that the spectral density
By using (3.24) the spectral density of {X
et }
function uniquely gives the autocovariance function. Comparing the spectral density of {X
with the spectral density of {Xt } we see that they both are the same up to a multiplicative
constant. Thus they both have the same autocovariance structure up to a multiplicative
constant (which can be made the same, if in the definition (3.29) the innovation process has
Q
variance pj22=1 |I,j2 |2 ).
Therefore, for every non-causal process, there exists a causal process with the same autocovariance function.
By using the same arguments above, we can generalize to result to ARMA processes.
Definition 3.3.2 An ARMA process is said to have minimum phase when the roots of (z) and
(z) both lie outside of the unit circle.
Remark 3.3.1 For Gaussian random processes it is impossible to discriminate between a causal
and non-causal time series, this is because the mean and autocovariance function uniquely identify
the process.
However, if the innovations are non-Gaussian, even though the autocovariance function is blind
to non-causal processes, by looking for other features in the time series we are able to discriminate
between a causal and non-causal process.

93

3.3.1

The Yule-Walker equations of a non-causal process

Once again let us consider the zero mean AR(p) model

Xt =

p
X

j Xtj + t ,

j=1

and var(t ) < . Suppose the roots of the corresponding characteristic polynomial lie outside the
unit circle, then {Xt } is strictly stationary where the solution of Xt is only in terms of past and
present values of {t }. Moreover, it is second order stationary with covariance {c(k)}. We recall
from Section 3.1.2, equation (3.4) that we derived the Yule-Walker equations for causal AR(p)
processes, where

E(Xt Xtk ) =

p
X

j E(Xtj Xtk ) c(k)

j=1

p
X

j c(k j) = 0.

(3.30)

j=1

Let us now consider the case that the roots of the characteristic polynomial lie both outside
and inside the unit circle, thus Xt does not have a causal solution but it is still strictly and second
order stationary (with autocovariance, say {c(k)}). In the previous section we showed that there
P
e
e
et = t (where (B) and (B)
exists a causal AR(p) (B)
X
= 1 pj=1 j z j are the characteristic
polynomials defined in (3.26) and (3.28)). We showed that both have the same autocovariance
structure. Therefore,

c(k)

p
X

j c(k j) = 0

j=1

t }.
This means the Yule-Walker equations for {Xt } would actually give the AR(p) coefficients of {X
Thus if the Yule-Walker equations were used to estimate the AR coefficients of {Xt }, in reality we
t }.
would be estimating the AR coefficients of the corresponding causal {X

3.3.2

Filtering non-causal AR models

Here we discuss the surprising result that filtering a non-causal time series with the corresponding
causal AR parameters leaves a sequence which is uncorrelated but not independent. Let us suppose

94

that
Xt =

p
X

j Xtj + t ,

j=1

where t are iid, E(t ) = 0 and var(t ) < . It is clear that given the input Xt , if we apply the
P
filter Xt pj=1 j Xtj we obtain an iid sequence (which is {t }).
P
Suppose that we filter {Xt } with the causal coefficients {ej }, the output et = Xt pj=1 ej Xtj
is not an independent sequence. However, it is an uncorrelated sequence. We illustrate this with an
example.
Example 3.3.1 Let us return to the AR(1) example, where Xt = Xt1 + t . Let us suppose that
> 1, which corresponds to a non-causal time series, then Xt has the solution
Xt =

X
1
t+j+1 .
j
j=1

et =
The causal time series with the same covariance structure as Xt is X

1 e
Xt1

+ (which has

backshift representation (1 1/(B))Xt = t ). Suppose we pass Xt through the causal filter

et = (1

(1 1 B)
1
1
B)Xt = Xt Xt1 =
1 t

B(1 B
)

1
1 X 1
= t + (1 2 )
t+j .

j1
j=1

Evaluating the covariance of the above (assuming wlog that var() = 1) is

1
1 1
1 X 1
cov(e
t , et+r ) = (1 2 ) r + (1 2 )2
= 0.

2j
j=0

Thus we see that {e


t } is an uncorrelated sequence, but unless it is Gaussian it is clearly not independent. One method to study the higher order dependence of {e
t }, by considering its higher order
cumulant structure etc.
The above above result can be generalised to general AR models, and it is relatively straightforward
to prove using the Cr
amer representation of a stationary process (see Section 8.4, Theorem ??).

95

Exercise 3.7

(i) Consider the causal AR(p) process


Xt = 1.5Xt1 0.75Xt2 + t .

Derive a parallel process with the same autocovariance structure but that is non-causal (it
should be real).
(ii) Simulate both from the causal process above and the corresponding non-causal process with
non-Gaussian innovations (see Section 2.6). Show that they have the same ACF function.
(iii) Find features which allow you to discriminate between the causal and non-causal process.

96

Chapter 4
Nonlinear Time Series Models
Prerequisites
A basic understanding of expectations, conditional expectations and how one can use conditioning to obtain an expectation.
Objectives:
Use relevant results to show that a model has a stationary, solution.
Derive moments of these processes.
Understand the differences between linear and nonlinear time series.
So far we have focused on linear time series, that is time series which have the representation
Xt =

j tj ,

j=

where {t } are iid random variables. Such models are extremely useful, because they are designed
to model the autocovariance structure and are straightforward to use for forecasting. These are
some of the reasons that they are used widely in several applications.
A typical realisation from a linear time series, will be quite regular with no suddent bursts
or jumps. This is due to the linearity of the system. However, if one looks at financial data, for
example, there are sudden bursts in volatility (variation) and extreme values, which calm down
after a while. It is not possible to model such behaviour well with a linear time series. In order to
capture nonlinear behaviour several nonlinear models have been proposed. The models typically

97

consists of products of random variables which make possible the sudden irratic bursts seen in
the data. Over the past 30 years there has been a lot research into nonlinear time series models.
Probably one of the first nonlinear models proposed for time series analysis is the bilinear model,
this model is used extensively in signal processing and engineering. A popular model for modelling
financial data are (G)ARCH-family of models. Other popular models are random autoregressive
coefficient models and threshold models, to name but a few (see, for example, Subba Rao (1977),
Granger and Andersen (1978), Nicholls and Quinn (1982), Engle (1982), Subba Rao and Gabr
(1984), Bollerslev (1986), Terdik (1999), Fan and Yao (2003), Straumann (2005) and Douc et al.
(2014)).
Once a model has been defined, the first difficult task is to show that it actually has a solution
which is almost surely finite (recall these models have dynamics which start at the , so if they
are not well defined they could be infinite), with a stationary solution. Typically, in the nonlinear
world, we look for causal solutions. I suspect this is because the mathematics behind existence of
non-causal solution makes the problem even more complex.
We state a result that gives sufficient conditions for a stationary, causal solution of a certain
class of models. These models include ARCH/GARCH and Bilinear models. We note that the
theorem guarantees a solution, but does not give conditions for its moments. The result is based
on Brandt (1986), but under stronger conditions.
Theorem 4.0.1 (Brandt (1986)) Let us suppose that {X t } is a d-dimensional time series defined by the stochastic recurrence relation
X t = At X t1 + B t ,

(4.1)

where {At } and {Bt } are iid random matrices and vectors respectively. If E log kAt k < 0 and
E log kB t k < (where k k denotes the spectral norm of a matrix), then

X t = Bt +

k1
Y

k=1

i=0

!
Ati

B tk

(4.2)

converges almost surely and is the unique strictly stationary causal solution.
Note: The conditions given above are very strong and Brandt (1986) states the result under
which weaker conditions, we outline the differences here. Firstly, the assumption {At , Bt } are iid
can be relaxed to their being Ergodic sequences. Secondly, the assumption E log kAt k < 0 can be

98

relaxed to E log kAt k < and that {At } has a negative Lyapunov exponent, where the Lyapunov
Q
exponent is defined as limn n1 k nj=1 Aj k = , with < 0 (see Brandt (1986)).
The conditions given in the above theorem may appear a little cryptic. However, the condition
E log |At | < 0 (in the unvariate case) becomes quite clear if you compare the SRE model with
the AR(1) model Xt = Xt1 + t , where || < 1 (which is the special case of the SRE, where
P
j
the coefficients is deterministic). We recall that the solution of the AR(1) is Xt =
k=1 tj .
The important part in this decomposition is that |j | decays geometrically fast to zero. Now let
Q
us compare this to (4.2), we see that j plays a similar role to k1
i=0 Ati . Given that there are
Q
similarities between the AR(1) and SRE, it seems reasonable that for (4.2) to converge, k1
i=0 Ati
should converge geometrically too (at least almost surely). However analysis of a product is not
straight forward, therefore we take logarithms to turn it into a sum
k1
k1
Y
1
1X
a.s.
log
Ati =
log Ati E[log At ] := ,
k
k
i=0

i=0

since it is the sum of iid random variables. Thus taking anti-logs


k1
Y

Ati exp[k],

i=0

which only converges to zero if < 0, in other words E[log At ] < 0. Thus we see that the condition
E log |At | < 0 is quite a logical conditional afterall.

4.1
4.1.1

Data Motivation
Yahoo data from 1996-2014

We consider here the closing share price of the Yahoo daily data downloaded from https://uk.
finance.yahoo.com/q/hp?s=YHOO. The data starts from from 10th April 1996 to 8th August 2014
(over 4000 observations). A plot is given in Figure 4.1. Typically the logarithm of such data taken,
and in order to remove linear and/or stochastic trend the first difference of the logarithm is taken
(ie. Xt = log St log St1 ). The hope is that after taking differences the data has been stationarized
(see Example 2.3.2). However, the data set spans almost 20 years and this assumption is rather
precarious and will be investigated later. A plot of the data after taking first differences together

99

From the QQplot in Figure 4.2, we observe that log

100

200

yahoo

300

400

with the QQplot is given in Figure 4.2.

1000

2000

3000

4000

Time

Figure 4.1: Plot of daily closing Yahoo share price 1996-2014

0.4
0.2
0.4
0.6

0.2

Sample Quantiles

0.0

0.4
0.2
0.4
0.6

yahoo.log.diff

0.0

0.2

Normal QQ Plot

0.8

0.8

1000

3000

Time

Theoretical Quantiles

Figure 4.2: Plot of log differences of daily Yahoo share price 1996-2014 and the corresponding
QQplot
differences {Xt } appears to have very thick tails, which may mean that higher order moments of
the log returns do not exist (not finite).
In Figure 4.3 we give the autocorrelation (ACF) plots of the log differences, absolute log differences and squares of the log differences. Note that the sample autocorrelation is defined as

b(k) =

b
c(k)
,
b
c(0)

where

b
c(k) =

T |k|
1 X

(Xt X)(X
t+k X).
T

(4.3)

t=1

The dotted lines are the errors bars (the 95% confidence of the sample correlations constructed
under the assumption the observations are independent, see Section 6.2.1). From Figure 4.3a
we see that there appears to be no correlation in the data. More precisely, most of the sample

100

correlations are within the errors bars, the few that are outside it could be by chance, as the error
bars are constructed pointwise. However, Figure 4.3b the ACF plot of the absolutes gives significant
large correlations. In contrast, in Figure 4.3c we give the ACF plot of the squares, where there
does not appear to be any significant correlations.

10

15

20

25

30

35

Lag

0.8
0.6
ACF
0.0

0.2

0.4

ACF
0.0

0.0

0.2

0.2

0.4

0.4

ACF

0.6

0.6

0.8

0.8

1.0

Series (yahoo.log.diff)^2

1.0

Series abs(yahoo.log.diff)

1.0

Series yahoo.log.diff

10

15

20

25

30

35

10

Lag

15

20

25

30

35

Lag

(a) ACF plot of the log differ- (b) ACF plot of the absolute (c) ACF plot of the square of
ences
of the log differences
the log differences

Figure 4.3: ACF plots of the transformed Yahoo data


To summarise, {Xt } appears to be uncorrelated (white noise). However, once absolutes have
been taken there does appear to be dependence. This type of behaviour cannot be modelled with
a linear model. What is quite interesting is that there does not appear to be any significant
correlation in the squares. However, on explanation for this can be found in the QQplot. The
data has extremely thick tails which suggest that the forth moment of the process may not exist
(the empirical variance of Xt will be extremely large). Since correlation is defined as (4.3) involves
division by b
c(0), which could be extremely large, this would hide the sample covariance.

R code for Yahoo data


Here we give the R code for making the plots above.
yahoo <- scan("~/yahoo304.96.8.14.txt")
yahoo <- yahoo[c(length(yahoo):1)] # switches the entries to ascending order 1996-2014
yahoo.log.diff <- log(yahoo[-1]) - log(yahoo[-length(yahoo)])
# Takelog differences
par(mfrow=c(1,1))
plot.ts(yahoo)
par(mfrow=c(1,2))

101

plot.ts(yahoo.log.diff)
qqnorm(yahoo.log.diff)
qqline(yahoo.log.diff)
par(mfrow=c(1,3))
acf(yahoo.log.diff) # ACF plot of log differences
acf(abs(yahoo.log.diff)) # ACF plot of absolute log differences
acf((yahoo.log.diff)**2) # ACF plot of square of log differences

4.1.2

FTSE 100 from January - August 2014

For completeness we discuss a much shorter data set, the daily closing price of the FTSE 100
from 20th January - 8th August, 2014 (141 observations). This data was downloaded from http:
//markets.ft.com/research//Tearsheets/PriceHistoryPopup?symbol=FTSE:FSI.
Exactly the same analysis that was applied to the Yahoo data is applied to the FTSE data and

6500

6600

ftse

6700

6800

the plots are given in Figure 4.4-4.6.

20

40

60

80

100

120

140

Time

Figure 4.4: Plot of daily closing FTSE price Jan-August, 2014


We observe that for this (much shorter) data set, the observations do not appear to deviate
much from normality. Furthermore, the ACF plot of the log differences, absolutes and squares
do not suggest any evidence of correlation. Could it be, that after taking log differences, there is
no dependence in the data (the data is a realisation from iid random variables). Or that there is
dependence but it lies in a higher order structure or over more sophisticated transformations.
Comparing this to the Yahoo data, may be we see dependence in the Yahoo data because it
is actually nonstationary. The mystery continues. It would be worth while conducting a similar
analysis on a similar portion of the Yahoo data.

102

Normal QQ Plot

20

60

100

140

0.015 0.010 0.005 0.000 0.005 0.010 0.015

Sample Quantiles

0.015 0.010 0.005 0.000 0.005 0.010 0.015

ftse.log.diff

Time

Theoretical Quantiles

Figure 4.5: Plot of log differences of daily FTSE price Jan-August, 2014 and the corresponding QQplot

10

15

1.0
0.8
0.2

0.0

0.2

0.4

ACF

0.6

0.8
0.2

0.0

0.2

0.4

ACF

0.6

0.8
0.6
0.4
0.2

0.0

0.2

ACF

Series (ftse.log.diff)^2

1.0

Series abs(ftse.log.diff)

1.0

Series ftse.log.diff

Lag

10

15

Lag

10

15

Lag

(a) ACF plot of the log differ- (b) ACF plot of the absolute (c) ACF plot of the square of
ences
of the log differences
the log differences

Figure 4.6: ACF plots of the transformed FTSE data

4.2

The ARCH model

During the early 80s Econometricians were trying to find a suitable model for forecasting stock
prices. They were faced with data similar to the log differences of the Yahoo data in Figure 4.2. As
Figure 4.3a demonstrates, there does not appear to be any linear dependence in the data, which
makes the best linear predictor quite useless for forecasting. Instead, they tried to predict the
variance of future prices given the past, var[Xt+1 |Xt , Xt1 , . . .]. This called for a model that has a
zero autocorrelation function, but models the conditional variance.
To address this need, Engle (1982) proposed the autoregressive conditionally heteroskadastic
(ARCH) model (note that Rob Engle, together with Clive Granger, in 2004, received the Noble prize

103

for Economics for Cointegration). He proposed the ARCH(p) which satisfies the representation
t2

Xt = t Zt

= a0 +

p
X

2
aj Xtj
,

j=1

where Zt are iid random variables where E(Zt ) = 0 and var(Zt ) = 1, a0 > 0 and for 1 j p
aj 0.
Before, worrying about whether a solution of such a model exists, let us consider the reasons
behind why this model was first proposed.

4.2.1

Features of an ARCH

Let us suppose that a causal, stationary solution of the ARCH model exists (Xt is a function of
Zt , Zt1 , Zt1 , . . .) and all the necessary moments exist. Then we obtain the following.
(i) The first moment:
E[Xt ] = E[Zt t ] = E[E(Zt t |Xt1 , Xt2 , . . .)] = E[t E(Zt |Xt1 , Xt2 , . . .)]
|
{z
}
t function of Xt1 ,...,Xtp

= E[t

E(Zt ) ] = E[0 t ] = 0.
| {z }

by causality

Thus the ARCH process has a zero mean.


(ii) The conditional variance:
var(Xt |Xt1 , Xt2 , . . . , Xtp ) = E(Xt2 |Xt1 , Xt2 , . . . , Xtp )
= E(Zt2 t2 |Xt1 , Xt2 , . . . , Xtp ) = t2 E[Zt2 ] = t2 .
Thus the conditional variance is t2 = a0 +

Pp

2
j=1 aj Xtj

(a weighted sum of the squared

past).
(iii) The autocovariance function:
Without loss of generality assume k > 0
cov[Xt , Xt+k ] = E[Xt Xt+k ] = E[Xt E(Xt+k |Xt+k1 , . . . , Xt )]
= E[Xt t+k E(Zt+k |Xt+k1 , . . . , Xt )] = E[Xt t+k E(Zt+k )] = E[Xt t+k 0] = 0.

104

The autocorrelation function is zero (it is a white noise process).


P
(iv) We will show in Section 4.2.2 that E[X 2d ] < iff [ pj=1 aj ]E[Zt2d ]1/d < 1. It is well known
that even for Gaussian innovations E[Zt2d ]1/d grows with d, therefore if any of the aj are
non-zero (recall all need to be positive), there will exist a d0 such that for all d d0 E[Xtd ]
will not be finite. Thus the we see that the ARCH process has thick tails.
Usually we measure the thickness of tails in data using the Kurtosis measure (see wiki).
Points (i-iv) demonstrate that the ARCH model is able to model many of the features seen in the
stock price data.
In some sense the ARCH model can be considered as a generalisation of the AR model. That
is the squares of ARCH model satisfy
Xt2

Zt2

= a0 +

p
X

2
aj Xtj
+ (Zt2 1)t2 ,

(4.4)

j=1

with characteristic polynomial (z) = 1

Pp

j=1 aj z

j.

It can be shown that if

Pp

j=1 aj

< 1, then the

roots of the characteristic polynomial (z) lie outside the unit circle (see Exercise 2.1). Moreover,
the innovations t = (Zt2 1)t2 are martingale differences (see wiki). This can be shown by noting
that
E[(Zt2 1)t2 |Xt1 , Xt2 , . . .] = t2 E(Zt2 1|Xt1 , Xt2 , . . .) = t2 E(Zt2 1) = 0.
| {z }
=0

Thus cov(t , s ) = 0 for s 6= t. Martingales are a useful asymptotic tool in time series, we demonstrate how they can be used in Chapter 10.
To summarise, in many respects the ARCH(p) model resembles the AR(p) except that the
innovations {t } are martingale differences and not iid random variables. This means that despite
the resemblence, it is not a linear time series.
We show that a unique, stationary causal solution of the ARCH model exists and derive conditions under which the moments exist.

105

4.2.2

Existence of a strictly stationary solution and second order


stationarity of the ARCH

To simplify notation we will consider the ARCH(1) model


Xt = t Zt

2
t2 = a0 + a1 Xt1
.

(4.5)

It is difficult to directly obtain a solution of Xt , instead we obtain a solution for t2 (since Xt can
2
2 Z2
immediately be obtained from this). Using that Xt1
= t1
t1 and substituting this into (4.5)

we obtain
2
2
2
t2 = a0 + a1 Xt1
= (a1 Zt1
)t1
+ a0 .

(4.6)

We observe that (4.6) can be written in the stochastic recurrence relation form given in (4.1)
2
2 ] = log a +
with At = a1 Zt1
and Bt = a0 . Therefore, by using Theorem 4.0.1, if E[log a1 Zt1
1
2 ] < 0, then 2 has the strictly stationary causal solution
E[log Zt1
t

t2 = a0 + a0

ak1

k
Y

2
Ztj
.

j=1

k=1

The condition for existence using Theorem 4.0.1 and (4.6) is


E[log(a1 Zt2 )] = log a1 + E[log Zt2 ] < 0,

(4.7)

which is immediately implied if a1 < 1 (since E[log Zt2 ] log E[Zt2 ] = 0), but it is also satisfied
under weaker conditions on a1 .
To obtain the moments of Xt2 we use that it has the solution is

Xt2 = Zt2 a0 + a0

ak1

k=1

k
Y

2
Ztj
,

j=1

therefore taking expectations we have

E[Xt2 ] = E[Zt2 ]E a0 + a0

ak1

k=1

106

k
Y
j=1

2
Ztj
= a0 + a0

X
k=1

ak1 .

Thus E[Xt2 ] < if and only if a1 < 1 (heuristically we can see this from E[Xt2 ] = E[Z22 ](a0 +
2 ])).
a1 E[Xt1

By placing stricter conditions on a1 , namely a1 E(Zt2d )1/d < 1, we can show that E[Xt2d ] < .

The ARCH(p) model


We can generalize the above results to ARCH(p) processes (but to show existence of a solution we
need to write the ARCH(p) process as a vector process similar to the Vector AR(1) representation of
an AR(p) given in Section 2.4.1). It can be shown that under sufficient conditions on the coefficients
{aj } that the stationary, causal solution of the ARCH(p) model is
Xt2 = a0 Zt2 +

mt (k)

(4.8)

k1

where mt (k) =

k
Y

a0

j1 ,...,jk 1

r=1

ajr

k
Y

2 P
r
Zt

s=0 js

(j0 = 0).

r=0

The above solution belongs to a general class of functions called a Volterra expansion. We note
P
that E[Xt2 ] < iff pj=1 aj < 1.

4.3

The GARCH model

A possible drawback of the ARCH(p) model is that the conditional variance only depends on finite
number of the past squared observations/log returns (in finance, the share price is often called
the return). However, when fitting the model to the data, analogous to order selection of an
autoregressive model (using, say, the AIC), often a large order p is selected. This suggests that
the conditional variance should involve a large (infinite?) number of past terms. This observation
motivated the GARCH model (first proposed in Bollerslev (1986) and Taylor (1986)), which in
many respects is analogous to the ARMA. The conditional variance of the GARCH model is a
weighted average of the squared returns, the weights decline with the lag, but never go completely
to zero. The GARCH class of models is a rather parsimonous class of models and is extremely
popular in finance. The GARCH(p, q) model is defined as

Xt = t Zt

t2

= a0 +

p
X
j=1

107

2
aj Xtj

q
X
i=1

2
bi ti

(4.9)

where Zt are iid random variables where E(Zt ) = 0 and var(Zt ) = 1, a0 > 0 and for 1 j p
aj 0 and 1 i q bi 0.
Under the assumption that a causal solution with sufficient moments exist, the same properties
defined for the ARCH(p) in Section 4.2.1 also apply to the GARCH(p, q) model.
It can be shown that under suitable conditions on {bj } that Xt satisfies an ARCH() representation. Formally, we can write the conditional variance t2 (assuming that a stationarity solution
exists) as

(1

q
X

bi B i )t2 = (a0 +

i=1

p
X

2
aj Xtj
),

j=1

where B denotes the backshift notation defined in Chapter 2. Therefore if the roots of b(z) =
P
P
(1 qi=1 bi z i ) lie outside the unit circle (which is satisfied if i bi < 1) then
t2 =

(1

1
Pq

j
j=1 bj B )

(a0 +

p
X

2
aj Xtj
) = 0 +

j=1

2
j Xtj
,

(4.10)

j=1

where a recursive equation for the derivation of j can be found in Berkes et al. (2003). In other
words the GARCH(p, q) process can be written as a ARCH() process. This is analogous to the
invertibility representation given in Definition 2.2.2. This representation is useful when estimating
the parameters of a GARCH process (see Berkes et al. (2003)) and also prediction. The expansion
in (4.10) helps explain why the GARCH(p, q) process is so popular. As we stated at the start of this
section, the conditional variance of the GARCH is a weighted average of the squared returns, the
weights decline with the lag, but never go completely to zero, a property that is highly desirable.
Example 4.3.1 (Inverting the GARCH(1, 1)) If b1 < 1, then we can write t2 as

t2

X
X


a0
2
2
+ a1
bj Xt1j
.
=
bj B j a0 + a1 Xt1
=
1b
j=0

j=0

This expression is useful in both prediction and estimation.


In the following section we derive conditions for existence of the GARCH model and also its
moments.

108

4.3.1

Existence of a stationary solution of a GARCH(1, 1)

We will focus on the GARCH(1, 1) model as this substantially simplifies the conditions. We recall
the conditional variance of the GARCH(1, 1) can be written as
 2
2
2
2
t2 = a0 + a1 Xt1
+ b1 t1
= a1 Zt1
+ b1 t1
+ a0 .

(4.11)

We observe that (4.11) can be written in the stochastic recurrence relation form given in (4.1) with
2
2
At = (a1 Zt1
+ b1 ) and Bt = a0 . Therefore, by using Theorem 4.0.1, if E[log(a1 Zt1
+ b1 )] < 0,

then t2 has the strictly stationary causal solution


t2

= a0 + a0

Y
k
X

2
(a1 Ztj
+ b1 ).

(4.12)

k=1 j=1

These conditions are relatively weak and depend on the distribution of Zt . They are definitely
2
2
satisfied if a1 + b1 < 1, since E[log(a1 Zt1
+ b1 )] log E[a1 Zt1
+ b1 ] = log(a1 + b1 ). However

existence of a stationary solution does not require such a strong condition on the coefficients (and
there can still exist a stationary solution if a1 + b1 > 1, so long as the distribution of Zt2 is such
that E[log(a1 Zt2 + b1 )] < 0).
By taking expectations of (4.12) we can see that

E[Xt2 ] = E[t2 ] = a0 + a0

Y
k
X

(a1 + b1 ) = a0 + a0

k=1 j=1

(a1 + b1 )k .

k=1

Thus E[Xt2 ] < iff a1 + b1 < 1 (noting that a1 and b1 are both positive). Expanding on this
argument, if d > 1 we can use Minkowski inequality to show

(E[t2d ])1/d

a0 +

k
Y

2
a0
(E[ (a1 Ztj
j=1
k=1

d 1/d

+ b1 )] )

Y
k
X
2
a0 + a0
(
E[(a1 Ztj
+ b1 )d ])1/d .
k=1 j=1

2
Therefore, if E[(a1 Ztj
+ b1 )d ] < 1, then E[Xt2d ] < . This is an iff condition, since from the

definition in (4.11) we have


2
2
2
2d
2
2
]d E[(a1 Zt1
+ b1 )t1
]d = E[(a1 Zt1
+ b1 )d ]E[t1
],
E[t2d ] = E[a0 + (a1 Zt1
+ b1 )t1
{z
}
|
every term is positive

2
2 . We observe that by stationarity and if
since t1
has a causal solution, it is independent of Zt1

109

2d ]. Thus the above inequality only holds if E[(a Z 2 + b )d ] < 1.


E[t2d ] < , then E[t2d ] = E[t1
1 t1
1
2 + b )d ] < 1.
Therefore, E[Xt2d ] < iff E[(a1 Zt1
1

Indeed in order for E[Xt2d ] < a huge constraint needs to be placed on the parameter space
of a1 and b1 .
Exercise 4.1 Suppose {Zt } are standard normal random variables. Find conditions on a1 and b1
such that E[Xt4 ] < .
The above results can be generalised to GARCH(p, q) model. Conditions for existence of a
stationary solution hinge on the random matrix corresponding to the SRE representation of the
GARCH model (see Bougerol and Picard (1992a) and Bougerol and Picard (1992b)), which are
nearly impossible to verify. Sufficient and necessary conditions for both a stationary (causal)
P
P
solution and second order stationarity (E[Xt2 ] < ) is pj=1 aj + qi=1 bi < 1. However, many
econometricians believe this condition places an unreasonable constraint on the parameter space of
{aj } and {bj }. A large amount of research has been done on finding consistent parameter estimators
under weaker conditions. Indeed, in the very interesting paper by Berkes et al. (2003) (see also
Straumann (2005)) they derive consistent estimates of GARCH parameters on far milder set of
conditions on {aj } and {bi } (which dont require E(Xt2 ) < ).
Definition 4.3.1 The IGARCH model is a GARCH model where
Xt = t Zt

t2 = a0 +

p
X

2
aj Xtj
+

j=1

where the coefficients are such that

Pp

j=1 aj

Pq

i=1 bi

q
X

2
bi ti

(4.13)

i=1

= 1. This is an example of a time series

model which has a strictly stationary solution but it is not second order stationary.
Exercise 4.2 Simulate realisations of ARCH(1) and GARCH(1, 1) models. Simulate with both iid
Gaussian and t-distribution errors ({Zt } where E[Zt2 ] = 1). Remember to burn-in each realisation.
In all cases fix a0 > 0. Then
(i) Simulate an ARCH(1) with a1 = 0.3 and a1 = 0.9.
(ii) Simulate a GARCH(1, 1) with a1 = 0.1 and b1 = 0.85, and a GARCH(1, 1) with a1 = 0.85
and b1 = 0.1. Compare the two behaviours.

110

4.3.2

Extensions of the GARCH model

One criticism of the GARCH model is that it is blind to negative the sign of the return Xt . In
other words, the conditional variance of Xt only takes into account the magnitude of Xt and does
not depend on increases or a decreases in St (which corresponds to Xt being positive or negative).
In contrast it is largely believed that the financial markets react differently to negative or positive
Xt . The general view is that there is greater volatility/uncertainity/variation in the market when
the price goes down.
This observation has motivated extensions to the GARCH, such as the EGARCH which take
into account the sign of Xt . Deriving conditions for such a stationary solution to exist can be
difficult task, and the reader is refered to Straumann (2005) and more the details.
Other extensions to the GARCH include an Autoregressive type model with GARCH innovations.

4.3.3

R code

install.packages("tseries"), library("tseries") recently there have been a new package


developed library("fGARCH").

4.4

Bilinear models

The Bilinear model was first proposed in Subba Rao (1977) and Granger and Andersen (1978) (see
also Subba Rao (1981)). The general Bilinear (BL(p, q, r, s)) model is defined as

Xt

p
X
j=1

j Xtj = t +

q
X

i ti +

r X
s
X

bk,k0 Xtk tk0 ,

k=1 k0 =1

i=1

where {t } are iid random variables with mean zero and variance 2 .
To motivate the Bilinear model let us consider the simplest version of the model BL(1, 0, 1, 1)
Xt = 1 Xt1 + b1,1 Xt1 t1 + t = [1 + b1,1 t1 ]Xt1 + t .

(4.14)

Comparing (4.16) with the conditional variane of the GARCH(1, 1) in (4.11) we see that they are
very similar, the main differences are that (a) the bilinear model does not constrain the coefficients
to be positive (whereas the conditional variance requires the coefficients to be positive) (b) the

111

2
2
t1 depends on Xt1 , whereas in the GARCH(1, 1) Zt1
and t1
are independent coefficients and

(c) the innovation in the GARCH(1, 1) model is deterministic, whereas in the innovation in the
Bilinear model is random. (b) and (c) makes the analysis of the Bilinear more complicated than
the GARCH model.

4.4.1

Features of the Bilinear model

In this section we assume a causal, stationary solution of the bilinear model exists, the appropriate
number of moments and that it is invertible in the sense that there exists a function g such that
t = g(Xt1 , Xt2 , . . .).
Under the assumption that the Bilinear process is invertible we can show that
E[Xt |Xt1 , Xt2 , . . .] = E[(1 + b1,1 t1 )Xt1 |Xt1 , Xt2 , . . .] + E[t |Xt1 , Xt2 , . . .]
= (1 + b1,1 t1 )Xt1 ,

(4.15)

thus unlike the autoregressive model the conditional expectation of the Xt given the past is a
nonlinear function of the past. It is this nonlinearity that gives rise to the spontaneous peaks that
we see a typical realisation.
To see how the bilinear model was motivated in Figure 4.7 we give a plot of
Xt = 1 Xt1 + b1,1 Xt1 t1 + t ,

(4.16)

where 1 = 0.5 and b1,1 = 0, 0.35, 0.65 and 0.65. and {t } are iid standard normal random
variables. We observe that Figure 4.7a is a realisation from an AR(1) process and the subsequent
plots are for different values of b1,1 . Figure 4.7a is quite regular, whereas the sudden bursts in
activity become more pronounced as b1,1 grows (see Figures 4.7b and 4.7c). In Figure 4.7d we give
a plot realisation from a model where b1,1 is negative and we see that the fluctation has changed
direction.
Remark 4.4.1 (Markov Bilinear model) Some authors define the BL(1, 0, 1, 1) as
Yt = 1 Yt1 + b1,1 Yt1 t + t = [1 + b11 t ]Yt1 + t .
The fundamental difference between this model and (4.16) is that the multiplicative innovation

112

bilinear(400, 0.5, 0.3)

3
2
1
0

bilinear(400, 0.5, 0)

2
3
0

100

200

300

400

100

Time

200

300

400

Time

(b) 1 = 0.5 and b1,1 = 0.35

5
10

bilinear(400, 0.5, 0.6)

20

15

bilinear(400, 0.5, 0.6)

10

(a) 1 = 0.5 and b1,1 = 0

100

200

300

400

Time

100

200

300

400

Time

(d) 1 = 0.5 and b1,1 = 0.65

(c) 1 = 0.5 and b1,1 = 0.65

Figure 4.7: Realisations from different BL(1, 0, 1, 1) models


(using t rather than t1 ) does not depend on Yt1 . This means that E[Yt |Yt1 , Yt2 , . . .] = 1 Yt1
and the autocovariance function is the same as the autocovariance function of an AR(1) model
with the same AR parameter. Therefore, it is unclear the advantage of using this version of the
model if the aim is to forecast, since the forecast of this model is the same as a forecast using the
corresponding AR(1) process Xt = 1 Xt1 + t . Forecasting with this model does not take into
account its nonlinear behaviour.

4.4.2

Solution of the Bilinear model

We observe that (4.16) can be written in the stochastic recurrence relation form given in (4.1) with
At = (1 + b11 t1 ) and Bt = a0 . Therefore, by using Theorem 4.0.1, if E[log(1 + b11 t1 )] < 0

113

and E[t ] < , then Xt has the strictly stationary, causal solution

k1
X
Y
(1 + b1,1 tj ) [(1 + b1,1 tk )tk ] + t .
Xt =
k=1

(4.17)

j=1

To show that it is second order stationary we require that E[Xt2 ] < , which imposes additional
conditions on the parameters. To derive conditions for E[Xt2 ] we use (4.18) and the Minkowski
inequality to give

(E[Xt2 ])1/2

k1
Y

(1 + b1,1 tj )


1/2
E [(1 + b11 tk )tk ]2

j=1

k=1

2 1/2

k1
X
Y


1/2 
1/2
E [(1 + b1,1 tj )]2
E [(1 + b1,1 tk )tk ]2
.

(4.18)

k=1 j=1

Therefore if E[4t ] < and


E [(1 + b1,1 t )]2 = 2 + b211 var(t ) < 1,
then E[Xt2 ] < (note that the above equality is due to E[t ] = 0).
Remark 4.4.2 (Inverting the Bilinear model) We note that
t = (bXt1 )t1 + [Xt Xt1 ],
thus by iterating backwards with respect to tj we have

t =

X
j=0

j1

(b)

j
Y

!
Xt1j

[Xtj Xtj1 ].

i=0

This invertible representation is useful both in forecasting and estimation (see Section 5.5.3).
Exercise 4.3 Simulate the BL(2, 0, 1, 1) model (using the AR(2) parameters 1 = 1.5 and 2 =
0.75). Experiment with different parameters to give different types of behaviour.
Exercise 4.4 The random coefficient AR model is a nonlinear time series proposed by Barry Quinn
(see Nicholls and Quinn (1982) and Aue et al. (2006)). The random coefficient AR(1) model is

114

defined as
Xt = ( + t )Xt1 + t
where {t } and {t } are iid random variables which are independent of each other.
(i) State sufficient conditions which ensure that {Xt } has a strictly stationary solution.
(ii) State conditions which ensure that {Xt } is second order stationary.
(iii) Simulate from this model for different and var[t ].

4.4.3

R code

Code to simulate a BL(1, 0, 1, 1) model:


# Bilinear Simulation
# Bilinear(1,0,1,1) model, we use the first n0 observations are burn-in
# in order to get close to the stationary solution.
bilinear <- function(n,phi,b,n0=400) {
y <- rnorm(n+n0)
w <- rnorm(n + n0)
for (t in 2:(n+n0)) {
y[t] <- phi * y[t-1] + b * w[t-1] * y[t-1] + w[t]
}
return(y[(n0+1):(n0+n)])
}

4.5

Nonparametric time series models

Many researchers argue that fitting parametric models can lead to misspecification and argue that
it may be more realistic to fit nonparametric or semi-parametric time series models instead. There
exists several nonstationary and semi-parametric time series (see Fan and Yao (2003) and Douc
et al. (2014) for a comprehensive summary), we give a few examples below. The most general

115

nonparametric model is
Xt = m(Xt1 , . . . , Xtp , t ),
but this is so general it looses all meaning, especially if the need is to predict. A slight restriction
is make the innovation term additive (see Jones (1978))
Xt = m(Xt1 , . . . , Xtp ) + t ,
it is clear that for this model E[Xt |Xt1 , . . . , Xtp ] = m(Xt1 , . . . , Xtp ). However this model has
the distinct disadvantage that without placing any structure on m(), for p > 2 nonparametric
estimators of m() are lousy (as the suffer from the curse of dimensionality).
Thus such a generalisation renders the model useless. Instead semi-parametric approaches have
been developed. Examples include the functional AR(p) model defined as

Xt =

p
X

j (Xtp )Xtj + t

j=1

the semi-parametric AR(1) model


Xt = Xt1 + (Xt1 ) + t ,
the nonparametric ARCH(p)

Xt = t Zt

t2 = a0 +

p
X

2
aj (Xtj
).

j=1

In the case of all these models it is not easy to establish conditions in which a stationary solution
exists. More often then not, if conditions are established they are similar in spirit to those that
are used in the parametric setting. For some details on the proof see Vogt (2013) (also here), who
considers nonparametric and nonstationary models (note the nonstationarity he considers is when
the covariance structure changes over time, not the unit root type). For example in the case of the
the semi-parametric AR(1) model, a stationary causal solution exists if | + 0 (0)| < 1.

116

Chapter 5
Prediction
Prerequisites
The best linear predictor.
Some idea of what a basis of a vector space is.

Objectives
Understand that prediction using a long past can be difficult because a large matrix has to
be inverted, thus alternative, recursive method are often used to avoid direct inversion.
Understand the derivation of the Levinson-Durbin algorithm, and why the coefficient, t,t ,
corresponds to the partial correlation between X1 and Xt+1 .
Understand how these predictive schemes can be used write space of sp(Xt , Xt1 , . . . , X1 ) in
terms of an orthogonal basis sp(Xt PXt1 ,Xt2 ,...,X1 (Xt ), . . . , X1 ).
Understand how the above leads to the Wold decomposition of a second order stationary
time series.
To understand how to approximate the prediction for an ARMA time series into a scheme
which explicitly uses the ARMA structure. And this approximation improves geometrically,
when the past is large.
One motivation behind fitting models to a time series is to forecast future unobserved observations - which would not be possible without a model. In this chapter we consider forecasting, based
on the assumption that the model and/or autocovariance structure is known.

117

5.1

Forecasting given the present and infinite past

In this section we will assume that the linear time series {Xt } is both causal and invertible, that is

Xt =

aj tj =

j=0

bi Xti + t ,

(5.1)

i=1

where {t } are iid random variables (recall Definition 2.2.2). Both these representations play an
important role in prediction. Furthermore, in order to predict Xt+k given Xt , Xt1 , . . . we will
assume that the infinite past is observed. In later sections we consider the more realistic situation
that only the finite past is observed. We note that since Xt , Xt1 , Xt2 , . . . is observed that we can
obtain (for t) by using the invertibility condition

= X

bi X i .

i=1

Now we consider the prediction of Xt+k given {X ; t}. Using the MA() presentation
(since the time series is causal) of Xt+k we have

Xt+k =

k1
X

aj+k tj

aj t+kj

j=0

j=0

{z

innovations are observed

{z

future innovations impossible to predict

P
Pk1
since E[ k1
j=0 aj t+kj |Xt , Xt1 , . . .] = E[ j=0 aj t+kj ] = 0. Therefore, the best linear predictor
of Xt+k given Xt , Xt1 , . . ., which we denote as Xt (k) is

Xt (k) =

aj+k tj =

j=0

X
j=0

aj+k (Xtj

bi Xtij ).

(5.2)

i=1

Xt (k) is called the k-step ahead predictor and it is straightforward to see that its mean squared
error is

2
k1
k
X
X
E [Xt+k Xt (k)]2 = E
aj t+kj = var[t ]
a2j ,
j=0

(5.3)

j=0

where the last line is due to the uncorrelatedness and zero mean of the innovations.
Often we would like to obtain the k-step ahead predictor for k = 1, . . . , n where n is some

118

time in the future. We now explain how Xt (k) can be evaluated recursively using the invertibility
assumption.
Step 1 Use invertibility in (5.1) to give

Xt (1) =

bi Xt+1i ,

i=1

and E [Xt+1 Xt (1)]2 = var[t ]


Step 2 To obtain the 2-step ahead predictor we note that
Xt+2 =
=

X
i=2

bi Xt+2i + b1 Xt+1 + t+2


bi Xt+2i + b1 [Xt (1) + t+1 ] + t+2 ,

i=2

thus it is clear that


Xt (2) =

bi Xt+2i + b1 Xt (1)

i=2



and E [Xt+2 Xt (2)]2 = var[t ] b21 + 1 = var[t ] a22 + a21 .
Step 3 To obtain the 3-step ahead predictor we note that
Xt+3 =
=

X
i=3

bi Xt+2i + b2 Xt+1 + b1 Xt+2 + t+3


bi Xt+2i + b2 (Xt (1) + t+1 ) + b1 (Xt (2) + b1 t+1 + t+2 ) + t+3 .

i=3

Thus
Xt (3) =

bi Xt+2i + b2 Xt (1) + b1 Xt (2)

i=3




and E [Xt+3 Xt (3)]2 = var[t ] (b2 + b21 )2 + b21 + 1 = var[t ] a23 + a22 + a21 .

119

Step k Using the arguments above it is easily seen that

Xt (k) =

bi Xt+ki +

k1
X

bi Xt (k i).

i=1

i=k

Thus the k-step ahead predictor can be recursively estimated.


We note that the predictor given above it based on the assumption that the infinite past is
observed. In practice this is not a realistic assumption. However, in the special case that time
series is an autoregressive process of order p (with AR parameters {j }pj=1 ) and Xt , . . . , Xtm is
observed where m p 1, then the above scheme can be used for forecasting. More precisely,

Xt (1) =

p
X

j Xt+1j

j=1

Xt (k) =

Xt (k) =

p
X
j=k
p
X

j Xt+kj +

k1
X

j Xt (k j) for 2 k p

j=1

j Xt (k j) for k > p.

(5.4)

j=1

However, in the general case more sophisticated algorithms are required when only the finite
past is known.

Example: Forecasting yearly temperatures


We now fit an autoregressive model to the yearly temperatures from 1880-2008 and use this model
to forecast the temperatures from 2009-2013. In Figure 5.1 we give a plot of the temperature time
series together with its ACF. It is clear there is some trend in the temperature data, therefore we
have taken second differences, a plot of the second difference and its ACF is given in Figure 5.2.
We now use the command ar.yule(res1,order.max=10) (we will discuss in Chapter 7 how this
function estimates the AR parameters) to estimate the the AR parameters. The function ar.yule
uses the AIC to select the order of the AR model. When fitting the second differences from (from
1880-2008 - a data set of length of 127) the AIC chooses the AR(7) model
Xt = 1.1472Xt1 1.1565Xt2 1.0784Xt3 0.7745Xt4 0.6132Xt5 0.3515Xt6 0.1575Xt7 + t ,

120

ACF

0.2

0.5

0.0

0.2

0.0

0.4

temp

0.6

0.5

0.8

1.0

Series global.mean

1880

1900

1920

1940

1960

1980

2000

10

Time

15

20

25

30

25

30

Lag

Figure 5.1: Yearly temperature from 1880-2013 and the ACF.

1.0
0.5
ACF

0.2

0.0

0.0
0.6

0.5

0.4

0.2

second.differences

0.4

0.6

Series diff2

1880

1900

1920

1940

1960

1980

2000

Time

10

15

20

Lag

Figure 5.2: Second differences of yearly temperature from 1880-2013 and its ACF.
with var[t ] = 2 = 0.02294. An ACF plot after fitting this model and then estimating the residuals
{t } is given in Figure 5.3. We observe that the ACF of the residuals appears to be uncorrelated,
which suggests that the AR(7) model fitted the data well (there is a formal test for this called the
Ljung-Box test which we cover later). By using the sequence of equations

121

0.4
0.2

0.0

0.2

ACF

0.6

0.8

1.0

Series residuals

10

15

20

Lag

Figure 5.3: An ACF plot of the estimated residuals {bt }.


127 (1) = 1.1472X127 1.1565X126 1.0784X125 0.7745X124 0.6132X123
X
0.3515X122 0.1575X121
127 (2) = 1.1472X
127 (1) 1.1565X127 1.0784X126 0.7745X125 0.6132X124
X
0.3515X123 0.1575X122
127 (3) = 1.1472X
127 (2) 1.1565X
127 (1) 1.0784X127 0.7745X126 0.6132X125
X
0.3515X124 0.1575X123
127 (4) = 1.1472X
127 (3) 1.1565X
127 (2) 1.0784X
127 (1) 0.7745X127 0.6132X126
X
0.3515X125 0.1575X124
127 (5) = 1.1472X
127 (4) 1.1565X
127 (3) 1.0784X
127 (2) 0.7745X
127 (1) 0.6132X127
X
0.3515X126 0.1575X125 .
127 (1), . . . , X
127 (5) as forecasts of X128 , . . . , X132 (we recall are the second differences),
We can use X
which we then use to construct forecasts of the temperatures. A plot of the second difference
forecasts together with the true values are given in Figure 5.4. We note that (5.3) can be used to

122

give the mse error. For example


127 (1)]2 = 2
E[X128 X
t
127 (1)]2 = (1 + 21 )t2
E[X128 X

If we believe the residuals are Gaussian we can use the mean squared error to construct confidence
intervals for the predictions.
On the other hand, if the residuals are not Gaussian, we can construct 95% confidence intervals
for the forecast using bootstrap. Specifically, we rewrite the AR(7) process as an MA() process

Xt =

tj .
j ()

j=0

Hence the best linear predictor can be rewritten as


Xt (k) =

t+kj
j ()

j=k

thus giving the prediction error

Xt+k Xt (k) =

k1
X

t+kj .
j ()

j=0

We have the prediction estimates, therefore all we need is to obtain the distribution of

Pk1
j=0

t+kj .
j ()

This can be done by estimating the residuals and then using bootstrap to estimate the distribuP
Pk1


tion of k1
j=0 j ()t+kj , using the empirical distribution of
j=0 j ()t+kj . From this we can
construct the 95% CI for the forecasts.
A small criticism of our approach is that we have fitted a rather large AR(7) model to time
series of length of 127. It may be more appropriate to fit an ARMA model to this time series.
Exercise 5.1 In this exercise we analyze the Sunspot data found on the course website. In the data
analysis below only use the data from 1700 - 2003 (the remaining data we will use for prediction).
In this section you will need to use the function ar.yw in R.
(i) Fit the following models to the data and study the residuals (using the ACF). Using this

123

0.3

= forecast

0.2

= true value

0.1

0.0

0.1

second difference

0.2

0.3

2000

2002

2004

2006

2008

2010

2012

year

Figure 5.4: Forecasts of second differences.


decide which model
Xt = + A cos(t) + B sin(t) + t
|{z}

or

AR

Xt = + t
|{z}
AR

is more appropriate (take into account the number of parameters estimated overall).
(ii) Use these models to forecast the sunspot numbers from 2004-2013.

diff1 = global.mean[c(2:134)] - global.mean[c(1:133)]


diff2 = diff1[c(2:133)] - diff1[c(1:132)]
res1 = diff2[c(1:127)]
residualsar7 <- ar.yw(res1,order.max=10)$resid
residuals <- residualsar7[-c(1:7)]
Forecast using the above model
res = c(res1,rep(0,5))
res[128] = -1.1472*res[127]

-1.1565*res[126]

124

-1.0784*res[125]

-0.7745*res[124]

-0.6132*res[

res[129] = -1.1472*res[128]

-1.1565*res[127]

-1.0784*res[126]

-0.7745*res[125]

-0.6132*res[

res[130] = -1.1472*res[129]

-1.1565*res[128]

-1.0784*res[127]

-0.7745*res[126]

-0.6132*res[

res[131] = -1.1472*res[130]

-1.1565*res[129]

-1.0784*res[128]

-0.7745*res[127]

-0.6132*res[

res[132] = -1.1472*res[131]

-1.1565*res[130]

-1.0784*res[129]

-0.7745*res[128]

-0.6132*res[

5.2

Review of vector spaces

In next few sections we will consider prediction/forecasting for stationary time series. In particular
to find the best linear predictor of Xt+1 given the finite past Xt , . . . , X1 . Setting up notation our
aim is to find
Xt+1|t = PX1 ,...,Xt (Xt+1 ) = Xt+1|t,...,1 =

t
X

t,j Xt+1j ,

j=1

where {t,j } are chosen to minimise the mean squared error min E(Xt+1
t

Pt

2
j=1 t,j Xt+1j ) .

Basic results from multiple regression show that

t,1

..
. = 1
t rt ,

t,t
where (t )i,j = E(Xi Xj ) and (rt )i = E(Xti Xt+1 ). Given the covariances this can easily be done.
However, if t is large a brute force method would require O(t3 ) computing operations to calculate
(5.7). Our aim is to exploit stationarity to reduce the number of operations. To do this, we will
briefly discuss the notion of projections on a space, which help in our derivation of computationally
more efficient methods.
Before we continue we first discuss briefly the idea of a a vector space, inner product spaces,
Hilbert spaces, spans and basis. A more complete review is given in Brockwell and Davis (1998),
Chapter 2.
First a brief definition of a vector space. X is called an vector space if for every x, y X and
a, b R (this can be generalised to C), then ax + by X . An inner product space is a vector
space which comes with an inner product, in other words for every element x, y X we can defined
an innerproduct hx, yi, where h, i satisfies all the conditions of an inner product. Thus for every
element x X we can define its norm as kxk = hx, xi. If the inner product space is complete

125

(meaning the limit of every sequence in the space is also in the space) then the innerproduct space
is a Hilbert space (see wiki).
(i) The classical example of a Hilbert space is the Euclidean space Rn where
P
the innerproduct between two elements is simply the scalar product, hx, yi = ni=1 xi yi .

Example 5.2.1

(ii) The subset of the probability space (, F, P ), where all the random variables defined on
R
have a finite second moment, ie. E(X 2 ) = X()2 dP () < . This space is denoted as
L2 (, F, P ). In this case, the inner product is hX, Y i = E(XY ).
(iii) The function space L2 [R, ], where f L2 [R, ] if f is mu-measureable and
Z

|f (x)|2 d(x) < ,

is a Hilbert space. For this space, the inner product is defined as


Z
hf, gi =

f (x)g(x)d(x).
R

In this chapter we will not use this function space, but it will be used in Chapter ?? (when
we prove the Spectral representation theorem).
It is straightforward to generalize the above to complex random variables and functions defined
on C. We simply need to remember to take conjugates when defining the innerproduct, ie. hX, Y i =
R
E(XY ) and hf, gi = C f (z)g(z)d(z).
In this chapter our focus will be on certain spaces of random variables which have a finite variance.

Basis
The random variables {Xt , Xt1 , . . . , X1 } span the space Xt1 (denoted as sp(Xt , Xt1 , . . . , X1 )), if
for every Y Xt1 , there exists coefficients {aj R} such that

Y =

t
X

aj Xt+1j .

(5.5)

j=1

Moreover, sp(Xt , Xt1 , . . . , X1 ) = Xt1 if for every {aj R},

Pt

j=1 aj Xt+1j

Xt1 . We now

define the basis of a vector space, which is closely related to the span. The random variables
{Xt , . . . , X1 } form a basis of the space Xt1 , if for every Y Xt1 we have a representation (5.5) and

126

this representation is unique. More precisely, there does not exist another set of coefficients {bj }
P
such that Y = tj=1 bj Xt+1j . For this reason, one can consider a basis as the minimal span, that
is the smallest set of elements which can span a space.
Definition 5.2.1 (Projections) The projection of the random variable Y onto the space spanned
P
by sp(Xt , Xt1 , . . . , X1 ) (often denoted as PXt ,Xt1 ,...,X1 (Y)) is defined as PXt ,Xt1 ,...,X1 (Y) = tj=1 cj Xt+1j ,
where {cj } is chosen such that the difference Y P( Xt ,Xt1 ,...,X1 ) (Yt ) is uncorrelated (orthogonal/perpendicular) to any element in sp(Xt , Xt1 , . . . , X1 ). In other words, PXt ,Xt1 ,...,X1 (Yt ) is the best
linear predictor of Y given Xt , . . . , X1 .

Orthogonal basis
An orthogonal basis is a basis, where every element in the basis is orthogonal to every other element
in the basis. It is straightforward to orthogonalize any given basis using the method of projections.
To simplify notation let Xt|t1 = PXt1 ,...,X1 (Xt ). By definition, Xt Xt|t1 is orthogonal to
the space sp(Xt1 , Xt1 , . . . , X1 ). In other words Xt Xt|t1 and Xs (1 s t) are orthogonal
(cov(Xs , (Xt Xt|t1 )), and by a similar argument Xt Xt|t1 and Xs Xs|s1 are orthogonal.
Thus by using projections we have created an orthogonal basis X1 , (X2 X2|1 ), . . . , (Xt Xt|t1 )
of the space sp(X1 , (X2 X2|1 ), . . . , (Xt Xt|t1 )). By construction it clear that sp(X1 , (X2
X2|1 ), . . . , (Xt Xt|t1 )) is a subspace of sp(Xt , . . . , X1 ). We now show that
sp(X1 , (X2 X2|1 ), . . . , (Xt Xt|t1 )) = sp(Xt , . . . , X1 ).
To do this we define the sum of spaces. If U and V are two orthogonal vector spaces (which
share the same innerproduct), then y U V , if there exists a u U and v V such that
1 .
y = u + v. By the definition of Xt1 , it is clear that (Xt Xt|t1 ) Xt1 , but (Xt Xt|t1 )
/ Xt1
1 . Continuing this argument we see that X 1 = sp(X
Hence Xt1 = sp(X
t Xt|t1 ) Xt1
t Xt|t1 )
t

sp(X
t1 Xt1|t2 ), . . . , sp(X
1 ). Hence sp(X
t , . . . , X1 ) = sp(X
t Xt|t1 , . . . , X2 X2|1 , X1 ).
Pt
Therefore for every PXt ,...,X1 (Y ) = j=1 aj Xt+1j , there exists coefficients {bj } such that

PXt ,...,X1 (Y ) = PXt Xt|t1 ,...,X2 X2|1 ,X1 (Y ) =

t
X

PXt+1j Xt+1j|tj (Y ) =

j=1

t1
X

bj (Xt+1j Xt+1j|tj ) + bt X1 ,

j=1

where bj = E(Y (Xj Xj|j1 ))/E(Xj Xj|j1 ))2 . A useful application of orthogonal basis is the
ease of obtaining the coefficients bj , which avoids the inversion of a matrix. This is the underlying
idea behind the innovations algorithm proposed in Brockwell and Davis (1998), Chapter 5.

127

5.2.1

Spaces spanned by infinite number of elements

The notions above can be generalised to spaces which have an infinite number of elements in their
basis (and are useful to prove Wolds decomposition theorem). Let now construct the space spanned
by infinite number random variables {Xt , Xt1 , . . .}. As with anything that involves we need to
define precisely what we mean by an infinite basis. To do this we construct a sequence of subspaces,
each defined with a finite number of elements in the basis. We increase the number of elements in
the subspace and consider the limit of this space. Let Xtn = sp(Xt , . . . , Xn ), clearly if m > n,
n

then Xtn Xtm . We define Xt , as Xt =


, then there
n=1 Xt , in other words if Y Xt

exists an n such that Y Xtn . However, we also need to ensure that the limits of all the sequences
lie in this infinite dimensional space, therefore we close the space by defining defining a new space
which includes the old space and also includes all the limits. To make this precise suppose the
sequence of random variables is such that Ys Xts , and E(Ys1 Ys2 )2 0 as s1 , s2 . Since
the sequence {Ys } is a Cauchy sequence there exists a limit. More precisely, there exists a random
n

variable Y , such that E(Ys Y )2 0 as s . Since the closure of the space, X t , contains the
set Xtn and all the limits of the Cauchy sequences in this set, then Y Xt . We let
Xt = sp(Xt , Xt1 , . . .),

(5.6)

The orthogonal basis of sp(Xt , Xt1 , . . .)


An orthogonal basis of sp(Xt , Xt1 , . . .) can be constructed using the same method used to orthogonalize sp(Xt , Xt1 , . . . , X1 ). The main difference is how to deal with the initial value, which in the
case of sp(Xt , Xt1 , . . . , X1 ) is X1 . The analogous version of the initial value in infinite dimension
space sp(Xt , Xt1 , . . .) is X , but this it not a well defined quantity (again we have to be careful
with these pesky infinities).
Let Xt1 (1) denote the best linear predictor of Xt given Xt1 , Xt2 , . . .. As in Section 5.2 it is

clear that (Xt Xt1 (1)) and Xs for s t1 are uncorrelated and Xt = sp(Xt Xt1 (1))Xt1
,

where Xt = sp(Xt , Xt1 , . . .). Thus we can construct the orthogonal basis (Xt Xt1 (1)), (Xt1
Xt2 (1)), . . . and the corresponding space sp((Xt Xt1 (1)), (Xt1 Xt2 (1)), . . .). It is clear that
sp((Xt Xt1 (1)), (Xt1 Xt2 (1)), . . .) sp(Xt , Xt1 , . . .). However, unlike the finite dimensional
case it is not clear that they are equal, roughly speaking this is because sp((Xt Xt1 (1)), (Xt1
Xt2 (1)), . . .) lacks the inital value X . Of course the time in the past is not really a well

128

defined quantity. Instead, the way we overcome this issue is that we define the initial starting

random variable as the intersection of the subspaces, more precisely let X =


.
n= Xt

Furthermore, we note that since Xn Xn1 (1) and Xs (for any s n) are orthogonal, then
sp((Xt Xt1 (1)), (Xt1 Xt2 (1)), . . .) and X are orthogonal spaces. Using X , we have
tj=0 sp((Xtj Xtj1 (1)) X = sp(Xt , Xt1 , . . .).
We will use this result when we prove the Wold decomposition theorem (in Section 5.7).

5.3

Levinson-Durbin algorithm

We recall that in prediction the aim is to predict Xt+1 given Xt , Xt1 , . . . , X1 . The best linear
predictor is

Xt+1|t = PX1 ,...,Xt (Xt+1 ) = Xt+1|t,...,1 =

t
X

t,j Xt+1j ,

(5.7)

j=1

where {t,j } are chosen to minimise the mean squared error, and are the solution of the equation

t,1

..
. = 1
t rt ,

t,t

(5.8)

where (t )i,j = E(Xi Xj ) and (rt )i = E(Xti Xt+1 ). Using standard methods, such as Gauss-Jordan
elimination, to solve this system of equations requires O(t3 ) operations. However, we recall that
{Xt } is a stationary time series, thus t is a Toeplitz matrix, by using this information in the 1940s
Norman Levinson proposed an algorithm which reduced the number of operations to O(t2 ). In the
1960s, Jim Durbin adapted the algorithm to time series and improved it.
We first outline the algorithm. We recall that the best linear predictor of Xt+1 given Xt , . . . , X1
is
Xt+1|t =

t
X

t,j Xt+1j .

(5.9)

j=1

The mean squared error is r(t + 1) = E[Xt+1 Xt+1|t ]2 . Given that the second order stationary
covariance structure, the idea of the Levinson-Durbin algorithm is to recursively estimate {t,j ; j =
1, . . . , t} given {t1,j ; j = 1, . . . , t 1} (which are the coefficients of the best linear predictor of Xt

129

given Xt1 , . . . , X1 ). Let us suppose that the autocovariance function c(k) = cov[X0 , Xk ] is known.
The Levinson-Durbin algorithm is calculated using the following recursion.
Step 1 1,1 = c(1)/c(0) and r(2) = E[X2 X2|1 ]2 = E[X2 1,1 X1 ]2 = 2c(0) 21,1 c(1).
Step 2 For j = t

t,t =
t,j

c(t)

= t1,j

Pt1

j=1 t1,j c(t

r(t)
t,t t1,tj

j)
1 j t 1,

and r(t + 1) = r(t)(1 2t,t ).

We give two proofs of the above recursion.


Exercise 5.2

(i) Suppose Xt = Xt1 +t (where || < 1). Use the Levinson-Durbin algorithm,

to deduce an expression for t,j for (1 j t).


(ii) Suppose Xt = t1 + t (where || < 1). Use the Levinson-Durbin algorithm (and possibly
Maple/Matlab), deduce an expression for t,j for (1 j t). (recall from Exercise 3.4 that
you already have an analytic expression for t,t ).

5.3.1

A proof based on projections

Let us suppose {Xt } is a zero mean stationary time series and c(k) = E(Xk X0 ). Let PXt ,...,X2 (X1 )
denote the best linear predictor of X1 given Xt , . . . , X2 and PXt ,...,X2 (Xt+1 ) denote the best linear
predictor of Xt+1 given Xt , . . . , X2 . Stationarity means that the following predictors share the same
coefficients
Xt|t1 =

t1
X

t1,j Xtj

PXt ,...,X2 (Xt+1 ) =

j=1

PXt ,...,X2 (X1 ) =

t1
X

t1
X

t1,j Xt+1j

(5.10)

j=1

t1,j Xj+1 .

j=1

The last line is because stationarity means that flipping a time series round has the same correlation
structure. These three relations are an important component of the proof.
Recall our objective is to derive the coefficients of the best linear predictor of PXt ,...,X1 (Xt+1 )
based on the coefficients of the best linear predictor PXt1 ,...,X1 (Xt ). To do this we partition the

130

space sp(Xt , . . . , X2 , X1 ) into two orthogonal spaces sp(Xt , . . . , X2 , X1 ) = sp(Xt , . . . , X2 , X1 )


sp(X1 PXt ,...,X2 (X1 )). Therefore by uncorrelatedness we have the partition
Xt+1|t = PXt ,...,X2 (Xt+1 ) + PX1 PXt ,...,X2 (X1 ) (Xt+1 )
=

t1
X
j=1

t1,j Xt+1j + tt (X1 PXt ,...,X2 (X1 ))


|
{z
}
by
projection
onto
one
variable
{z
}
by (5.10)

t1
X

t1,j Xt+1j

j=1

t1

+ t,t X1
t1,j Xj+1 .

j=1

{z
}
|

(5.11)

by (5.10)

We start by evaluating an expression for t,t (which in turn will give the expression for the other
coefficients). It is straightforward to see that
t,t =
=
=

E(Xt+1 (X1 PXt ,...,X2 (X1 )))


E(X1 PXt ,...,X2 (X1 ))2
E[(Xt+1 PXt ,...,X2 (Xt+1 ) + PXt ,...,X2 (Xt+1 ))(X1 PXt ,...,X2 (X1 ))]
E(X1 PXt ,...,X2 (X1 ))2
E[(Xt+1 PXt ,...,X2 (Xt+1 ))(X1 PXt ,...,X2 (X1 ))]
E(X1 PXt ,...,X2 (X1 ))2

(5.12)

Therefore we see that the numerator of t,t is the partial covariance between Xt+1 and X1 (see
Section 3.2.2), furthermore the denominator of t,t is the mean squared prediction error, since by
stationarity
E(X1 PXt ,...,X2 (X1 ))2 = E(Xt PXt1 ,...,X1 (Xt ))2 = r(t)

(5.13)

Returning to (5.12), expanding out the expectation in the numerator and using (5.13) we have

t,t =

c(0)
c(0) E[Xt+1 PXt ,...,X2 (X1 ))]
E(Xt+1 (X1 PXt ,...,X2 (X1 )))
=
=
r(t)
r(t)

Pt1

j=1 t1,j c(t

j)

r(t)

(5.14)
which immediately gives us the first equation in Step 2 of the Levinson-Durbin algorithm. To

131

obtain the recursion for t,j we use (5.11) to give

Xt+1|t =

t
X

t,j Xt+1j

j=1

t1
X

t1,j Xt+1j + t,t X1

j=1

t1
X

t1,j Xj+1 .

j=1

To obtain the recursion we simply compare coefficients to give


t,j = t1,j t,t t1,tj

1 j t 1.

This gives the middle equation in Step 2. To obtain the recursion for the mean squared prediction
error we note that by orthogonality of {Xt , . . . , X2 } and X1 PXt ,...,X2 (X1 ) we use (5.11) to give
r(t + 1) = E(Xt+1 Xt+1|t )2 = E[Xt+1 PXt ,...,X2 (Xt+1 ) t,t (X1 PXt ,...,X2 (X1 )]2
= E[Xt+1 PX2 ,...,Xt (Xt+1 )]2 + 2t,t E[X1 PXt ,...,X2 (X1 )]2
2t,t E[(Xt+1 PXt ,...,X2 (Xt+1 ))(X1 PXt ,...,X2 (X1 ))]
= r(t) + 2t,t r(t) 2t,t E[Xt+1 (X1 PXt ,...,X2 (X1 ))]
{z
}
|
=r(t)t,t by (5.14)

= r(t)[1

2tt ].

This gives the final part of the equation in Step 2 of the Levinson-Durbin algorithm.
Further references: Brockwell and Davis (1998), Chapter 5 and Fuller (1995), pages 82.

5.3.2

A proof based on symmetric Toeplitz matrices

We now give an alternative proof which is based on properties of the (symmetric) Toeplitz matrix.
We use (5.8), which is a matrix equation where

t,1

.
t .. = rt ,

t,t

132

(5.15)

with

t =

c(2) . . . c(t 1)

c(1) . . . c(t 2)
..
..
..
.
.
.
..
..
c(t 1) c(t 2)
.
.
c(0)

c(0)
c(1)
..
.

c(1)
c(0)
..
.

c(1)

c(2)

rt =
.. .
.

c(t)

and

The proof is based on embedding rt1 and t1 into t1 and using that t1 t1 = rt1 .
To do this, we define the (t 1) (t 1) matrix Et1 which basically swops round all the
elements in a vector

Et1

0 0 0 ... 0 1

0 0 0 ... 1 0
,
.. .. .. .. ..

. . . . .

..
1 0 . 0 0 0

(recall we came across this swopping matrix in Section 3.2.2). Using the above notation, we have
the interesting block matrix structure

t =
and rt =

t1

Et1 rt1

r0t1 Et1
(r0t1 , c(t))0 .

c(0)

Returning to the matrix equations in (5.15) and substituting the above into (5.15) we have

t t = r t ,

t1

Et1 rt1

r0t1 Et1

c(0)

t1,t
t,t

rt1
c(t)

where 0t1,t = (1,t , . . . , t1,t ). This leads to the two equations


t1 t1,t + Et1 rt1 t,t = rt1

(5.16)

r0t1 Et1 t1,t + c(0)t,t = c(t).

(5.17)

We first show that equation (5.16) corresponds to the second equation in the Levinson-Durbin

133

algorithm. Multiplying (5.16) by 1


t1 , and rearranging the equation we have
rt1 1 Et1 rt1 t,t .
t1,t = 1
| t1
{z } | t1 {z
}
=t1

=Et1 t1

Thus we have
t1,t = t1 t,t Et1 t1 .

(5.18)

This proves the second equation in Step 2 of the Levinson-Durbin algorithm.


We now use (5.17) to obtain an expression for t,t , which is the first equation in Step 1.
Substituting (5.18) into t1,t of (5.17) gives


r0t1 Et1 t1 t,t Et1 t1 + c(0)t,t = c(t).

(5.19)

Thus solving for t,t we have

t,t =

c(t) c0t1 Et1 t1


c(0) c0t1 0t1

(5.20)

Noting that r(t) = c(0) c0t1 0t1 . (5.20) is the first equation of Step 2 in the Levinson-Durbin
equation.
Note from this proof it does not appear that we need that the (symmetric) Toeplitz matrix is
positive semi-definite.

5.3.3

Using the Durbin-Levinson to obtain the Cholesky decomposition of the precision matrix

We recall from Section 3.2.1 that by sequentially projecting the elements of random vector on the
past elements in the vector gives rise to Cholesky decomposition of the inverse of the variance/covariance (precision) matrix. This is exactly what was done in when we make the Durbin-Levinson

134

algorithm. In other words,

var

X1

r(1)
X1 1,1 X2

r(2)

= In

..
.
P
Xn n1
j=1 n1,j Xnj

r(n)

0
Therefore, if n = var[X n ], where X n = (X1 , . . . , Xn ), then 1
n = Ln Dn Ln , where

Ln =

...

1,1

2,1
..
.

2,2
..
.

1
..

... ... 0

... ... 0

0 ... 0

. . . . ..
. .
.

(5.21)

n1,n1 n1,n2 n1,n3 . . . . . . 1


and Dn = diag(r11 , r21 , . . . , rn1 ).

5.4

Forecasting for ARMA processes

Given the autocovariance of any stationary process the Levinson-Durbin algorithm allows us to
systematically obtain one-step predictors of second order stationary time series without directly
inverting a matrix.
In this section we consider forecasting for a special case of stationary processes, the ARMA
process. We will assume throughout this section that the parameters of the model are known.
We showed in Section 5.1 that if {Xt } has an AR(p) representation and t > p, then the best
linear predictor can easily be obtained using (5.4). Therefore, when t > p, there is no real gain
in using the Levinson-Durbin for prediction of AR(p) processes. However, we do show in Chapter
?? we can apply the Levinson-Durbin algorithm for obtaining estimators of the autoregressive
parameters.
Similarly if {Xt } satisfies an ARMA(p, q) representation, then the prediction scheme can be
greatly simplified. Unlike the AR(p) process, which is p-Markovian, PXt ,Xt1 ,...,X1 (Xt+1 ) does
involve all regressors Xt , . . . , X1 . However, some simplifications can be made in the scheme. To

135

explain how, let us suppose that Xt satisfies the representation


Xt

p
X

i Xtj = t +

j=1

q
X

i ti ,

i=1

where {t } are iid zero mean random variables and the roots of (z) and (z) lie outside the
unit circle. For the analysis below, let Wt = Xt for 1 t p and for t > max(p, q) let Wt =
P
P
t + qi=1 i ti (which is the MA(q) part of the process). Since Xp+1 = pj=1 j Xt+1j + Wp+1
and so forth it is clear that sp(X1 , . . . , Xt ) = sp(W1 , . . . , Wt ) (ie, they are linear combinations of
each other). We will show for t > max(p, q) that

Xt+1|t = PXt ,...,X1 (Xt+1 ) =

p
X

j Xt+1j +

j=1

q
X

t,i (Xt+1i Xt+1i|ti ).

(5.22)

i=1

for some t,i which can be evaluated from the autocovariance structure. To prove the result we use
the following steps:

PXt ,...,X1 (Xt+1 ) =

p
X
j=1

p
X

j PXt ,...,X1 (Xt+1j ) +


{z
}
|
Xt+1j

j Xt+1j +

j=1

i=1

p
X

q
X

j Xt+1j +

j=1

i=1

p
X

q
X

j Xt+1j +

j=1

q
X

p
X

q
X

i PXt ,...,X1 (t+1i )

i=1

i PXt Xt|t1 ,...,X2 X2|1 ,X1 (t+1i )


{z
}
|
=PWt Wt|t1 ,...,W2 W2|1 ,W1 (t+1i )

i PWt+1i Wt+1i|ti ,...,Wt Wt|t1 (t+1i )

i=1

i1
X

PWt+1i+s Wt+1i+s|ti+s (t+1i )


{z
}
s=0 |
since t+11 is independent of the past

q
X

j=1

t,i (Wt+1i Wt+1i|ti )


|
{z
}
i=1

p
X

q
X

j=1

j Xt+1j +

=Xt+1i Xt+1i|ti

j Xt+1j +

t,i (Xt+1i Xt+1i|ti ),

(5.23)

i=1

this gives the desired result. Thus given the parameters {t,i } is straightforward to construct the
predictor Xt+1|t . It can be shown that t,i i as t (see Brockwell and Davis (1998)),
Chapter 5.
Remark 5.4.1 In terms of notation we can understand the above result for the MA(q) case. In

136

this case,the above result reduces to

bt+1|t =
X

q
X



bt+1i|ti .
t,i Xt+1i X

i=1

We now state a few results which will be useful later.


Lemma 5.4.1 Suppose {Xt } is a stationary time series with spectral density f (). Let X t =
(X1 , . . . , Xt ) and t = var(X t ).
(i) If the spectral density function is bounded away from zero (there is some > 0 such that
inf f () > 0), then for all t, min (t ) (where min and max denote the smallest and
largest absolute eigenvalues of the matrix).
1
(ii) Further, max (1
t ) .

(Since for symmetric matrices the spectral norm and the largest eigenvalue are the same, then
1
k1
t kspec ).

(iii) Analogously, sup f () M < , then max (t ) M (hence kt kspec M ).


PROOF. See Chapter ??.

Remark 5.4.2 Suppose {Xt } is an ARMA process, where the roots (z) and and (z) have absolute value greater than 1 + 1 and less than 2 , then the spectral density f () is bounded by
(1 1 )2p

var(t ) (1(

2
1
)2p
1+1

f () var(t )

1
)2p
(1( 1+
1

(1 1 )2p

. Therefore, from Lemma 5.4.1 we have that max (t )

and max (1
t ) is bounded uniformly over t.
The prediction can be simplified if we make a simple approximation (which works well if t is
bt+1|t = Xt and for t > max(p, q) we define the
relatively large). For 1 t max(p, q), set X
recursion
bt+1|t =
X

p
X
j=1

j Xt+1j +

q
X

bt+1i|ti ).
i (Xt+1i X

(5.24)

i=1

This approximation seems reasonable, since in the exact predictor (5.23), t,i i .
In the following proposition we show that the best linear predictor of Xt+1 given X1 , . . . , Xt ,
bt+1|t and the best linear predictor given the infinite past,
Xt+1|t , the approximating predictor X

137

bt+1|t
Xt (1) are asymptotically equivalent. To do this we obtain expressions for Xt (1) and X

Xt (1) =

bj Xt+1j ( since Xt+1 =

j=1

bj Xt+1j + t+1 ).

j=1

Furthermore, by iterating (5.24) backwards we can show that


tmax(p,q)

bt+1|t =
X

max(p,q)

bj Xt+1j +

j=1

gammaj Xj

(5.25)

j=1

where |j |Ct , with 1/(1 + ) < < 1 and the roots of (z) are outside (1 + ). We give a proof in
the remark below.
Remark 5.4.3 We prove (5.25) for the ARMA(1, 2). We first recall that the AR(1) part in the
ARMA(1, 1) model does not play any role since sp(X1 , Xt , . . . , Xt ) = sp(W1 , W2 , . . . , Wt ), where
W1 = X1 and for t 2 we define the corresponding MA(2) process Wt = 1 t1 + 2 t2 + t . The
c2|1 = W1 , W
c3|2 = W2 and for t > 3
corresponding approximating predictor is defined as W
ct|t1 = 1 [Wt1 W
ct1|t2 ] + 2 [Wt2 W
ct2|t3 ].
W
Using this and rearranging (5.24) gives
bt+1|t 1 Xt = 1 [Xt X
bt1|t2 ],
bt|t1 ] +2 [Xt1 X
X
|
|
{z
}
|
{z
}
{z
}
ct|t1 )
=(Wt W

ct+1|t
W

ct1|t2 )
=(Wt1 W

By subtracting the above from Wt+1 we have


ct|t1 ) 2 (Wt1 W
ct1|t2 ) + Wt+1 .
ct+1|t = 1 (Wt W
Wt+1 W

(5.26)

ct+1|t as the matrix difference equation


It is straightforward to rewrite Wt+1 W

ct+1|t
Wt+1 W
ct|t1
Wt W
{z
=b
t+1

=
}

1 0
|
{z
=Q

ct|t1
Wt W

+
ct1|t2
Wt1 W
}|
{z
} |

=b
t

Wt+1
0
{z

W t+1

ct+1|t lead to the same difference equation except for some


We now show that t+1 and Wt+1 W

138

initial conditions, it is this that will give us the result. To do this we


write t as function of {Wt } (the irreducible condition). We first note that t can be written as
the matrix difference equation

t+1

t
{z

1 0
|
{z

=t+1

t1
} | {z }
t

Wt+1
0
{z

W t+1

(5.27)

Thus iterating backwards we can write


t+1 =

X
X
bj Wt+1j ,
(1)j [Qj ](1,1) Wt+1j =
j=0

j=0

where bj = (1)j [Qj ](1,1) (noting that b0 = 1) denotes the (1, 1)th element of the matrix Qj (note
we did something similar in Section 2.4.1). Furthermore the same iteration shows that

t+1 =

t3
X

(1)j [Qj ](1,1) Wt+1j + (1)t2 [Qt2 ](1,1) 3

j=0

t3
X

bj Wt+1j + (1)t2 [Qt2 ](1,1) 3 .

(5.28)

j=0

Therefore, by comparison we see that

t+1

t3
X

bj Wt+1j = (1)t2 [Qt2 ]1 =


3

bj Wt+1j .

j=t2

j=0

We now return to the approximation prediction in (5.26). Comparing (5.27) and (5.27) we see
that they are almost the same difference equations. The only difference is the point at which the
algorithm starts. t goes all the way back to the start of time. Whereas we have set initial values
c2|1 = W1 , W
c3|2 = W2 , thus b
for W
03 = (W3 W2 , W2 W1 ).Therefore, by iterating both (5.27) and
(5.27) backwards, focusing on the first element of the vector and using (5.28) we have
3 ]1
t+1 bt+1 = (1)t2 [Qt2 3 ]1 +(1)t2 [Qt2b
|
{z
}
P

=
j=t2 bj Wt+1j

We recall that t+1 = Wt+1 +

P
ct+1|t . Substituting this into
bt+1 = Wt+1 W
j=1 bj Wt+1j and that

139

the above gives


ct+1|t
W

bj Wt+1j =

j=1

bj Wt+1j + (1)t2 [Qt2b


3 ]1 .

j=t2

Replacing Wt with Xt 1 Xt1 gives (5.25), where the bj can be easily deduced from bj and 1 .
Proposition 5.4.1 Suppose {Xt } is an ARMA process where the roots of (z) and (z) have roots
bt+1|t and Xt (1) be defined as in (5.23),
which are greater in absolute value than 1 + . Let Xt+1|t , X
(5.24) and (5.2) respectively. Then

for any

1
1+

bt+1|t ]2 Kt ,
E[Xt+1|t X

(5.29)

bt+1|t Xt (1)]2 Kt
E[X

(5.30)



E[Xt+1 Xt+1|t ]2 2 Kt

(5.31)

< < 1 and var(t ) = 2 .

PROOF. The proof of (5.29) becomes clear when we use the expansion Xt+1 =

j=1 bj Xt+1j

t+1 , noting that by Lemma 2.5.1(iii), |bj | Cj .


Evaluating the best linear predictor of Xt+1 given Xt , . . . , X1 , using the autoregressive expansion
gives
Xt+1|t =

X
j=1

bj PXt ,...,X1 (Xt+1j ) + PXt ,...,X1 (t+1 )


|
{z
}
=0

tmax(p,q)

bj Xt+1j +

j=1

bj PXt ,...,X1 (Xtj+1 ).

j=tmax(p,q)

{z

bt+1|t Pmax(p,q) j Xj
X
j=1

bt+1|t is
Therefore by using (5.25) we see that the difference between the best linear predictor and X

bt+1|t =
Xt+1|t X

max(p,q)

bt+j PXt ,...,X1 (Xj+1 ) +

X
j=1

j= max(p,q)

140

j Xj = I + II.

By using (5.25), it is straightforward to show that the second term E[II 2 ] = E[

Pmax(p,q)
j=1

j Xtj ]2

Ct , therefore what remains is to show that E[II 2 ] attains a similar bound. Heuristically, this seems
reasonable, since bt+j Kt+j , the main obstacle is to show that E[PXt ,...,X1 (Xj+1 )2 ] and does
not grow with t. To obtain a bound, we first obtain a bound for E[PXt ,...,X1 (Xj+1 )]2 . Basic results
in linear regression shows that
PXt ,...,X1 (Xj+1 ) = 0j,t X t ,

(5.32)

0
0
0
where j,t = 1
t r t,j , with j,t = (1,j,t , . . . , t,j,t ), X t = (X1 , . . . , Xt ), t = E(X t X t ) and r t,j =

E(X t Xj ). Substituting (5.32) into I gives

bt+j PXt ,...,X1 (Xj+1 ) =

j= max(p,q)

bt+j 0j,t X t =

j= max(p,q)


bj r 0t,j 1
t X t . (5.33)

j=tmax(p,q)

Therefore the mean squared error of I is

E[I 2 ] =

bt+j r 0t,j 1
t

j= max(p,q)

bt+j r t,j .

j= max(p,q)

To bound the above we use the Cauchy schwarz inequality (kaBbk1 kak2 kBbk2 ), the specP
tral norm inequality (kak2 kBbk2 kak2 kBkspec kbk2 ) and Minkowiskis inequality (k nj=1 aj k2
Pn
j=1 kaj k2 ) we have

X
2
  X
2
2
2
E I2
bt+j r 0t,j 2 k1
k

|bt+j | kr t,j k2 k1
spec
t
t kspec .
j=1

(5.34)

j=1

We now bound each of the terms above. We note that for all t, using Remark 5.4.2 that k1
t kspec
K (for some constant K). We now consider r 0t,j = (E(X1 Xj ), . . . , E(Xt Xj )) = (c(1j), . . . , c(t
j)). By using (3.2) we have |c(k)| Ck , therefore

kr t,j k2 K(

t
X

2(j+r) )1/2 K

r=1

j
.
(1 2 )2

 
Substituting these bounds into (5.34) gives E I 2 Kt . Altogether the bounds for I and II give
bt+1|t )2 K
E(Xt+1|t X

141

j
.
(1 2 )2

Thus proving (5.29).


To prove (5.30) we note that

X
bt+1|t ]2 = E
E[Xt (1) X
bt+j Xj +
j=0

t
X

2
bj Ytj .

j=tmax(p,q)

Using the above and that bt+j Kt+j , it is straightforward to prove the result.
Finally to prove (5.31), we note that by Minkowskis inequality we have
h
2 i1/2
E Xt+1 Xt+1|t

h
i1/2  
2 1/2  
2 1/2
2
b
b
+ E Xt (1) Xt+1|t
+ E Xt+1|t Xt+1|t
.
E (Xt Xt (1))
{z
} |
|
{z
} |
{z
}
=

Kt/2 by (5.30)

Kt/2 by (5.29)

Thus giving the desired result.

5.5

Forecasting for nonlinear models

In this section we consider forecasting for nonlinear models. The forecasts we construct, may not
necessarily/formally be the best linear predictor, because the best linear predictor is based on
minimising the mean squared error, which we recall from Chapter 4 requires the existence of the
higher order moments. Instead our forecast will be the conditional expection of Xt+1 given the past
(note that we can think of it as the best linear predictor). Furthermore, with the exception of the
ARCH model we will derive approximation of the conditional expectation/best linear predictor,
bt+1|t (given in (5.24)).
analogous to the forecasting approximation for the ARMA model, X

5.5.1

Forecasting volatility using an ARCH(p) model

We recall the ARCH(p) model defined in Section 4.2

Xt = t Zt

t2 = a0 +

p
X
j=1

142

2
aj Xtj
.

Using a similar calculation to those given in Section 4.2.1, we see that


E[Xt+1 |Xt , Xt1 , . . . , Xtp+1 ] = E(Zt+1 t+1 |Xt , Xt1 , . . . , Xtp+1 ) = t+1 E(Zt+1 |Xt , Xt1 , . . . , Xtp+1 )
{z
}
|
t+1 function of Xt ,...,Xtp+1

= t+1 E(Zt+1 ) = 0 t+1 = 0.


| {z }
by causality

In other words, past values of Xt have no influence on the expected value of Xt+1 . On the other
hand, in Section 4.2.1 we showed that
2
2
2
2
2
2
E(Xt+1
|Xt , Xt1 , . . . , Xtp+1 ) = E(Zt+1
t+1
|Xt , Xt2 , . . . , Xtp+1 ) = t+1
E[Zt+1
] = t+1
=

p
X

2
aj Xt+1j
,

j=1

thus Xt has an influence on the conditional mean squared/variance. Therefore, if we let Xt+k|t
denote the conditional variance of Xt+k given Xt , . . . , Xtp+1 , it can be derived using the following
recursion
2
Xt+1|t
=

p
X

2
aj Xt+1j

j=1
2
Xt+k|t
=

2
Xt+k|t
=

p
X
j=k
p
X

2
+
aj Xt+kj

k1
X

2
aj Xt+kj|k

2kp

j=1
2
aj Xt+kj|t

k > p.

j=1

5.5.2

Forecasting volatility using a GARCH(1, 1) model

We recall the GARCH(1, 1) model defined in Section 4.3


 2
2
2
2
+ a0 .
t2 = a0 + a1 Xt1
+ b1 t1
= a1 Zt1
+ b1 t1
Similar to the ARCH model it is straightforward to show that E[Xt+1 |Xt , Xt1 , . . .] = 0 (where we
use the notation Xt , Xt1 , . . . to denote the infinite past or more precisely conditioned on the sigma
algebra Ft = (Xt , Xt1 , . . .)). Therefore, like the ARCH process, our aim is to predict Xt2 .
We recall from Example 4.3.1 that if the GARCH the process is invertible (satisfied if b < 1),

143

then

2
2
2
2
E[Xt+1
|Xt , Xt1 , . . .] = t+1
= a0 + a1 Xt1
+ b1 t1
=

X
a0
2
+ a1
bj Xtj
.
1b

(5.35)

j=0

Of course, in reality we only observe the finite past Xt , Xt1 , . . . , X1 . We can approximate
2 |X , X
2 = 0, then for t 1 let
E[Xt+1
b1|0
t
t1 , . . . , X1 ] using the following recursion, set

2
2
= a0 + a1 Xt2 + b1
bt|t1

bt+1|t

(noting that this is similar in spirit to the recursive approximate one-step ahead predictor defined
in (5.25)). It is straightforward to show that
t1

bt+1|t
=

X
a0 (1 bt+1 )
2
+ a1
bj Xtj
,
1b
j=0

2 |X , . . . , X ] (if the mean square error existed


taking note that this is not the same as E[Xt+1
t
1
2 |X , . . . , X ] would give a smaller mean square error), but just like the ARMA process it will
E[Xt+1
t
1
2
closely approximate it. Furthermore, from (5.35) it can be seen that
bt+1|t
closely approximates
2
t+1

Exercise 5.3 To answer this question you need R install.package("tseries") then remember
library("garch").
(i) You will find the Nasdaq data from 4th January 2010 - 15th October 2014 on my website.
(ii) By taking log differences fit a GARCH(1,1) model to the daily closing data (ignore the adjusted
closing value) from 4th January 2010 - 30th September 2014 (use the function garch(x,
order = c(1, 1)) fit the GARCH(1, 1) model).
(iii) Using the fitted GARCH(1, 1) model, forecast the volatility t2 from October 1st-15th (not2 . Evaluate
ing that no trading is done during the weekends). Denote these forecasts as t|0
P11 2
t=1 t|0

(iv) Compare this to the actual volatility

P11

2
t=1 Xt

144

(where Xt are the log differences).

5.5.3

Forecasting using a BL(1, 0, 1, 1) model

We recall the Bilinear(1, 0, 1, 1) model defined in Section 4.4


Xt = 1 Xt1 + b1,1 Xt1 t1 + t .
Assuming invertibility, so that t can be written in terms of Xt (see Remark 4.4.2):

t =

X
j=0

(b)

j1
Y

!
Xt1j

[Xtj Xtj1 ],

i=0

it can be shown that


Xt (1) = E[Xt+1 |Xt , Xt1 , . . .] = 1 Xt + b1,1 Xt t .
However, just as in the ARMA and GARCH case we can obtain an approximation, by setting
b1|0 = 0 and for t 1 defining the recursion
X


bt+1|t = 1 Xt + b1,1 Xt Xt X
bt|t1 .
X
See ? and ? for further details.
bt+1|t approximate Xt (1)?) We now derive conditions for X
bt+1|t
Remark 5.5.1 (How well does X
to be a close approximation of Xt (1) when t is large. We use a similar technique to that used in
Remark 5.4.3.
We note that Xt+1 Xt (1) = t+1 (since a future innovation, t+1 , cannot be predicted). We
bt+1|t is close to t+1 . Subtracting X
bt+1|t from Xt+1 gives the recursion
will show that Xt+1 X
bt+1|t = b1,1 (Xt X
bt|t1 )Xt + (bt Xt + t+1 ) .
Xt+1 X

(5.36)

We will compare the above recursion to the recursion based on t+1 . Rearranging the bilinear
equation gives
t+1 = bt Xt + (Xt+1 1 Xt ) .
|
{z
}
=bt Xt +t+1

145

(5.37)

We observe that (5.36) and (5.37) are almost the same difference equation, the only difference is
b1|0 . This gives the difference between the two equations as
that an initial value is set for X

bt+1|t ] = (1)t bt X1
t+1 [Xt+1 X

t
Y

b1|0 ]
j + (1)t bt [X1 X

j=1

t
Y

j .

j=1

P
bt+1|t
0 as t , then X
Xt (1) as t . We now show that if
Q
Q
a.s.
E[log |t | < log |b|, then bt tj=1 j 0. Since bt tj=1 j is a product, it seems appropriate to

Thus if bt

Qt

j=1 j

a.s.

take logarithms to transform it into a sum. To ensure that it is positive, we take absolutes and
t-roots
t
Y

log |bt

1X
log |j |
t
j=1
{z
}
|

j |1/t = log |b| +

j=1

average of iid random variables

Therefore by using the law of large numbers we have


t

log |b

t
Y
j=1

Thus we see that |bt

1/t

j |

1X
P
= log |b| +
log |j | log |b| + E log |0 | = .
t

1/t a.s.

j=1 j |

Qt

j=1

exp(). In other words, |bt

Qt

j=1 j |

exp(t), which will only

converge to zero if E[log |t | < log |b|.

5.6

Nonparametric prediction

In this section we briefly consider how prediction can be achieved in the nonparametric world. Let
us assume that {Xt } is a stationary time series. Our objective is to predict Xt+1 given the past.
However, we dont want to make any assumptions about the nature of {Xt }. Instead we want to
obtain a predictor of Xt+1 given Xt which minimises the means squared error, E[Xt+1 g(Xt )]2 . It
is well known that this is conditional expectation E[Xt+1 |Xt ]. (since E[Xt+1 g(Xt )]2 = E[Xt+1
E(Xt+1 |Xt )]2 + E[g(Xt ) E(Xt+1 |Xt )]2 ). Therefore, one can estimate
E[Xt+1 |Xt = x] = m(x)

146

nonparametrically. A classical estimator of m(x) is the Nadaraya-Watson estimator


Pn1

xXt
t=1 Xt+1 K( b )
,
Pn1
xXt
t=1 K( b )

m
b n (x) =

where K : R R is a kernel function (see Fan and Yao (2003), Chapter 5 and 6). Under some
regularity conditions it can be shown that m
b n (x) is a consistent estimator of m(x) and converges
to m(x) in mean square (with the typical mean squared rate O(b4 + (bn)1 )). The advantage of
going the non-parametric route is that we have not imposed any form of structure on the process
(such as linear/(G)ARCH/Bilinear). Therefore, we do not run the risk of misspecifying the model
A disadvantage is that nonparametric estimators tend to be a lot worse than parametric estimators
(in Chapter ?? we show that parametric estimators have O(n1/2 ) convergence which is faster than
the nonparametric rate O(b2 + (bn)1/2 )). Another possible disavantage is that if we wanted to
include more past values in the predictor, ie. m(x1 , . . . , xd ) = E[Xt+1 |Xt = x1 , . . . , Xtp = xd ] then
the estimator will have an extremely poor rate of convergence (due to the curse of dimensionality).
A possible solution to the problem is to assume some structure on the nonparametric model,
and define a semi-parametric time series model. We state some examples below:
(i) An additive structure of the type

Xt =

p
X

gj (Xtj ) + t

j=1

where {t } are iid random variables.


(ii) A functional autoregressive type structure

Xt =

p
X

gj (Xtd )Xtj + t .

j=1

(iii) The semi-parametric GARCH(1, 1)


Xt = t Zt ,

2
t2 = bt1
+ m(Xt1 ).

However, once a structure has been imposed, conditions need to be derived in order that the model
has a stationary solution (just as we did with the fully-parametric models).
See ?, ?, ?, ?, ? etc.

147

5.7

The Wold Decomposition

Section 5.2.1 nicely leads to the Wold decomposition, which we now state and prove. The Wold
decomposition theorem, states that any stationary process, has something that appears close to
an MA() representation (though it is not). We state the theorem below and use some of the
notation introduced in Section 5.2.1.
Theorem 5.7.1 Suppose that {Xt } is a second order stationary time series with a finite variance
(we shall assume that it has mean zero, though this is not necessary). Then Xt can be uniquely
expressed as
Xt =

j Ztj + Vt ,

(5.38)

j=0

where {Zt } are uncorrelated random variables, with var(Zt ) = E(Xt Xt1 (1))2 (noting that Xt1 (1)
, where X
is the best linear predictor of Xt given Xt1 , Xt2 , . . .) and Vt X =
n
n= Xn

is defined in (5.6).
PROOF. First let is consider the one-step ahead prediction of Xt given the infinite past, denoted
P
Xt1 (1). Since {Xt } is a second order stationary process it is clear that Xt1 (1) =
j=1 bj Xtj ,
where the coefficients {bj } do not vary with t. For this reason {Xt1 (1)} and {Xt Xt1 (1)} are
second order stationary random variables. Furthermore, since {Xt Xt1 (1)} is uncorrelated with
Xs for any s t, then {Xs Xs1 (1); s R} are uncorrelated random variables. Define Zs = Xs
Xs1 (1), and observe that Zs is the one-step ahead prediction error. We recall from Section 5.2.1
that Xt sp((Xt Xt1 (1)), (Xt1 Xt2 (1)), . . .) sp(X
) =
). Since
j=0 sp(Ztj ) sp(X

the spaces
j=0 sp(Ztj ) and sp(X ) are orthogonal, we shall first project Xt onto j=0 sp(Ztj ),

due to orthogonality the difference between Xt and its projection will be in sp(X ). This will
lead to the Wold decomposition.
First we consider the projection of Xt onto the space
j=0 sp(Ztj ), which is

PZt ,Zt1 ,... (Xt ) =

j Ztj ,

j=0

where due to orthogonality j = cov(Xt , (Xtj Xtj1 (1)))/var(Xtj Xtj1 (1)). Since Xt

), the difference Xt PZt ,Zt1 ,... Xt is orthogonal to {Zt } and belongs in


j=0 sp(Ztj ) sp(X

148

sp(X
). Hence we have

Xt =

j Ztj + Vt ,

j=0

where Vt = Xt

j=0 j Ztj

and is uncorrelated to {Zt }. Hence we have shown (5.38). To show

that the representation is unique we note that Zt , Zt1 , . . . are an orthogonal basis of sp(Zt , Zt1 , . . .),
which pretty much leads to uniqueness.

Exercise 5.4 Consider the process Xt = A cos(Bt + U ) where A, B and U are random variables
such that A, B and U are independent and U is uniformly distributed on (0, 2).
(i) Show that Xt is second order stationary (actually its stationary) and obtain its means and
covariance function.
(ii) Show that the distribution of A and B can be chosen is such a way that {Xt } has the same
covariance function as the MA(1) process Yt = t + t (where || < 1) (quite amazing).
(iii) Suppose A and B have the same distribution found in (ii).
(a) What is the best predictor of Xt+1 given Xt , Xt1 , . . .?
(b) What is the best linear predictor of Xt+1 given Xt , Xt1 , . . .?
It is worth noting that variants on the proof can be found in Brockwell and Davis (1998),
Section 5.7 and Fuller (1995), page 94.
Remark 5.7.1 Notice that the representation in (5.38) looks like an MA() process. There is,
however, a significant difference. The random variables {Zt } of an MA() process are iid random
variables and not just uncorrelated.
We recall that we have already come across the Wold decomposition of some time series. In
Section 3.3 we showed that a non-causal linear time series could be represented as a causal linear
time series with uncorrelated but dependent innovations. Another example is in Chapter 4, where
we explored ARCH/GARCH process which have an AR and ARMA type representation. Using this
representation we can represent ARCH and GARCH processes as the weighted sum of {(Zt2 1)t2 }
which are uncorrelated random variables.

149

Remark 5.7.2 (Variation on the Wold decomposition) In many technical proofs involving
time series, we often use a results related to the Wold decomposition. More precisely, we often
decompose the time series in terms of an infinite sum of martingale differences. In particular,
we define the sigma-algebra Ft = (Xt , Xt1 , . . .), and suppose that E(Xt |F ) = . Then by
telescoping we can formally write Xt as
Xt =

Zt,j

j=0

where Zt,j = E(Xt |Ftj ) E(Xt |Ftj1 ). It is straightforward to see that Zt,j are martingale
differences, and under certain conditions (mixing, physical dependence, your favourite dependence
P
flavour etc) it can be shown that
j=0 kZt,j kp < (where k kp is the pth moment). This means
the above representation holds almost surely. Thus in several proofs we can replace Xt by
P
j=0 Zt,j . This decomposition allows us to use martingale theorems to prove results.

150

Chapter 6
Estimation of the mean and
covariance
Prerequisites
Some idea of what a cumulant is.

Objectives
To derive the sample autocovariance of a time series, and show that this is a positive definite
sequence.
To show that the variance of the sample covariance involves fourth order cumulants, which
can be unwielding to estimate in practice. But under linearity the expression for the variance
greatly simplifies.
To show that under linearity the correlation does not involve the fourth order cumulant. This
is the Bartlett formula.
To use the above results to construct a test for uncorrelatedness of a time series (the Portmanteau test). And understand how this test may be useful for testing for independence in
various different setting. Also understand situations where the test may fail.

151

6.1

An estimator of the mean

Suppose we observe {Yt }nt=1 , where


Yt = + Xt ,
where is the finite mean, {Xt } is a zero mean stationary time series with absolutely summable
P
covariances ( k |cov(X0 , Xk )| < ). Our aim is to estimate the mean . The most obvious
P
estimator is the sample mean, that is Yn = n1 nt=1 Yt as an estimator of .

6.1.1

The sampling properties of the sample mean

We recall from Example 1.5.1 that we obtained an expression for the sample mean. We showed
that
n

var(Yn ) =

2 X n k
1
var(X0 ) +
c(k).
n
n
n
k=1

Furthermore, if

|c(k)| < , then in Example 1.5.1 we showed that

var(Yn ) =

1
2X
1
var(X0 ) +
c(k) + o( ).
n
n
n
k=1

Thus if the time series has sufficient decay in its correlation structure a mean squared consistent
estimator of the sample mean can be achieved. However, one drawback is that the dependency
means that one observation will influence the next, and if the influence is positive (seen by a positive
covariance), the resulting estimator may have a (much) larger variance than the iid case.
The above result does not require any more conditions on the process, besides second order
stationarity and summability of its covariance. However, to obtain confidence intervals we require
a stronger result, namely a central limit theorem for the sample mean. The above conditions are
P
not enough to give a central limit theorem. To obtain a CLT for sums of the form nt=1 Xt we
need the following main ingredients:
(i) The variance needs to be finite.
(ii) The dependence between Xt decreases the further apart in time the observations. However,
this is more than just the correlation, it really means the dependence.

152

The above conditions are satisfied by linear time series, if the cofficients j decay sufficient fast.
However, these conditions can also be verified for nonlinear time series (for example the (G)ARCH
and Bilinear model described in Chapter 4).
We now state the asymptotic normality result for linear models.
P
Theorem 6.1.1 Suppose that Xt is a linear time series, of the form Xt =
j= j tj , where t
P
P
are iid random variables with mean zero and variance one, j= |j | < and
j= j 6= 0.
Let Yt = + Xt , then we have

where 2 = var(X0 ) + 2


n Yn = N (0, 2 )

k=1 c(k).

PROOF. Later in this course we will give precise details on how to prove asymptotic normality of
several different type of estimators in time series. However, we give a small flavour here by showing
asymptotic normality of Yn in the special case that {Xt }nt=1 satisfy an MA(q) model, then explain
how it can be extended to MA() processes.
The main idea of the proof is to transform/approximate the average into a quantity that we
know is asymptotic normal. We know if {t }nt=1 are iid random variables with mean and variance
one then

n(
n ) N (0, 1).

(6.1)

We aim to use this result to prove the theorem. Returning to Yn by a change of variables (s = t j)
we can show that
n

1X
Yt
n

t=1

q
n
n
1X
1 XX
Xt = +
j tj
n
n
t=1
t=1 j=0

nq
q
q
0
n
ns
X
X
X
X
X
X
1
+
s
j +
s
j +
s
j
n
s=1
s=q+1
s=nq+1
j=0
j=qs
j=0

q
nq
q
0
n
ns
X
X
X
X
X
X
1
1
1
nq
j
s +
s
j +
s
j
+
n
nq
n
n

j=0

:= +

s=1

s=q+1

(n q)
nq + E1 + E2 ,
n

j=q+s

s=nq+1

j=0

(6.2)

153

Pq

It is straightforward to show that E|E1 | Cn1 and E|E2 | Cn1 .


P
nq . We note that if the assumptions are not satisfied and qj=0 j =
Finally we examine (nq)
n

where =

j=0 j .

0 (for example the process Xt = t t1 ), then


n
1
1X
Yt = +
n
n
t=1

0
X

s=q+1

q
X

j +

j=qs

1
n

n
X

s=nq+1

ns
X

j .

j=0

This is a degenerate case, since E1 and E2 only consist of a finite number of terms and thus if t are
non-Gaussian these terms will never be asymptotically normal. Therefore, in this case we simply
P
have that n1 nt=1 Yt = + O( n1 ) (this is why in the assumptions it was stated that 6= 0).
On the other hand, if 6= 0, then the dominating term in Yn is nq . From (6.1) it is
p

P
P
clear that n q nq N (0, 1) as n . However, for finite q, (n q)/n 1, therefore

P
n
nq N (0, 1). Altogether, substituting E|E1 | Cn1 and E|E2 | Cn1 into (6.2) gives




1 P
n Yn = n
nq + Op ( ) N 0, 2 .
n

With a little work, it can be shown that 2 = 2 .


Observe that the proof simply approximated the sum by a sum of iid random variables. In the
case that the process is a MA() or linear time series, a similar method is used. More precisely,
we have

n Yn

1 XX

j tj
n t=1

n
1 X X

j
t + Rn
n
t=1

j=0

j=0

where

Rn

nj

n
X
1 X X

j
s
s
n
s=1
j=0
s=1j

nj
n
0
n

n
X
X
X
X
X
X
1
1

j
s
s +
j
s
s
n
n
s=1
j=0

s=1j

s=nj

:= Rn1 + Rn2 + Rn3 + Rn4 .

154

j=n+1

s=1j

2 ] = o(1) for 1 j 4. We start with R


We will show that E[Rn,j
n,1

2
E[Rn,1
] =

0
0
n
X
X
1 X
s
s
j1 j2 cov
n
s=1j1

j1 ,j2 =0

1
n

n
X

s=1j2

j1 j2 min[j1 1, j2 1]

j1 ,j2 =0

jX
n
n
1 1
1X 2
2 X
j2 min[j2 1]
j (j 1) +
j1 ,
n
n

1
n

j=0
n
X

j2 (j 1) +

j=0

j1 =0
n
X

2
n

j2 =0

|j1 j1 |.

j1 =0

Pn
P
2
< and, thus,
j=0 [1j/n]j
j=0 |j | < , then by dominated convegence
P
P
Pn
n

2
2
j=0 (j/n)j 0 and
j=0 j as n . This implies that
j=0 [1 j/n]j
j=0 j and
Pn
2
2
j=0 (j/n)j 0. Substituting this into the above bounds for E[Rn,1 ] we immediately obtain

Since
P

j=0 |j |

2 ] = o(1). Using the same argument we obtain the same bound for R
E[Rn,1
n,2 , Rn,3 and Rn,4 . Thus

n

1 X
n Yn =
t + op (1)
n
j=1

and the result then immediately follows.

Estimation of the so called long run variance (given in Theorem 6.1.1) can be difficult. There
are various methods that can be used, such as estimating the spectral density function (which we
define in Chapter 8) at zero. An interesting approach advocated by Xiaofeng Shao is to use the
method of so called self-normalization which circumvents the need to estimate the long run mean,
see Shao (2010).

6.2

An estimator of the covariance

Suppose we observe {Yt }nt=1 , to estimate the covariance we can estimate the covariance c(k) =
cov(Y0 , Yk ) from the the observations. A plausible estimator is
n|k|
1 X
cn (k) =
(Yt Yn )(Yt+|k| Yn ),
n
t=1

155

(6.3)

since E[(Yt Yn )(Yt+|k| Yn )] c(k). Of course if the mean of Yt is known to be zero (Yt = Xt ),
then the covariance estimator is
n|k|
1 X
Xt Xt+|k| .
cn (k) =
n

(6.4)

t=1

1 Pn|k|
The eagle-eyed amongst you may wonder why we dont use n|k|
n (k) is a
t=1 Xt Xt+|k| , when c
P
n|k|
1
biased estimator, whereas n|k|
n (k) has some very nice properties
t=1 Xt Xt+|k| is not. However c

which we discuss in the lemma below.


Lemma 6.2.1 Suppose we define the empirical covariances

cn (k) =

1
n

Pn|k|
t=1

Xt Xt+|k| |k| n 1

otherwise

then {
cn (k)} is a positive definite sequence. Therefore, using Lemma 1.6.1 there exists a stationary
time series {Zt } which has the covariance cn (k).
PROOF. There are various ways to show that {
cn (k)} is a positive definite sequence. One method
uses that the spectral density corresponding to this sequence is non-negative, we give this proof in
Section 8.3.1.
Here we give an alternative proof. We recall a sequence is positive definite if for any vector
a = (a1 , . . . , ar )0 we have
r
X

ak1 ak2 cn (k1 k2 ) =

k1 ,k2 =1

n
X

b na 0
ak1 ak2 cn (k1 k2 ) = a0

k1 ,k2 =1

where

b
n =

noting that cn (k) =

1
n

cn (0)

cn (1)

cn (2) . . . cn (n 1)

cn (1)
cn (0)
cn (1) . . . cn (n 2)
,
..
..
..
..

..

.
.
.
.
.

..
..
cn (n 1) cn (n 2)
.
.
cn (0)

Pn|k|
t=1

Xt Xt+|k| . However, cn (k) =

1
n

Pn|k|
t=1

Xt Xt+|k| has a very interesting

b n = Xn X0 , where Xn is a
construction, it can be shown that the above convariance matrix is
n

156

n 2n matrix with

Xn =

...

X1

X2

0
..
.

0
..
.

...
..
.

X1
..
.

X2
..
.

. . . Xn1
..
..
.
.

X1 X2 . . . Xn1 Xn

...

...

Xn1 Xn
Xn
..
.

0
..
.

...

Using the above we have


b n a = a0 Xn X0 a = kX0 ak2 0.
a0
n
2
This this proves that {
cn (k)} is a positive definite sequence.
Finally, by using Theorem 1.6.1, there exists a stochastic process with {
cn (k)} as its autocovariance function.

6.2.1

Asymptotic properties of the covariance estimator

The main reason we construct an estimator is either for testing or constructing a confidence interval
for the parameter of interest. To do this we need the variance and distribution of the estimator. It
is impossible to derive the finite sample distribution, thus we look at their asymptotic distribution.
Besides showing asymptotic normality, it is important to derive an expression for the variance.
In an ideal world the variance will be simple and will not involve unknown parameters. Usually
in time series this will not be the case, and the variance will involve several (often an infinite)
number of parameters which are not straightforward to estimate. Later in this section we show
that the variance of the sample covariance can be extremely complicated. However, a substantial
simplification can arise if we consider only the sample correlation (not variance) and assume linearity
of the time series. This result is known as Bartletts formula (you may have come across Maurice
Bartlett before, besides his fundamental contributions in time series he is well known for proposing
the famous Bartlett correction). This example demonstrates, how the assumption of linearity can
really simplify problems in time series analysis and also how we can circumvent certain problems
in which arise by making slight modifications of the estimator (such as going from covariance to
correlation).
The following theorem gives the asymptotic sampling properties of the covariance estimator
(6.3). One proof of the result can be found in Brockwell and Davis (1998), Chapter 8, Fuller

157

(1995), but it goes back to Bartlett (indeed its called Bartletts formula). We prove the result in
Section 6.2.2.
Theorem 6.2.1 Suppose {Xt } is a linear stationary time series where

Xt = +

j tj ,

j=

where

|j | < , {t } are iid random variables with E(4t ) < . Suppose we observe {Xt :

t = 1, . . . , n} and use (6.3) as an estimator of the covariance c(k) = cov(X0 , Xk ). Define n (r) =
cn (r)/
cn (0) as the sample correlation. Then for each h {1, . . . , n}

n(
n (h) (h)) N (0, Wh )

(6.5)

where n (h) = (
n (1), . . . , n (h)), (h) = ((1), . . . , (h)) and

(Wh )ij =

{(k + i) + (k i) 2(i)(k)}{(k + j) + (k j) 2(j)(k)}.

(6.6)

k=

Equation (6.6) is known as Bartletts formula.


In Section 6.3 we apply the method for checking for correlation in a time series. We first show
how the expression for the asymptotic variance is obtained.

6.2.2

Proof of Bartletts formula

The variance of the sample covariance in the case of strict stationarity


We first derive an expression for cn (r) under the assumption that {Xt } is a strictly stationary time
P
P
series with finite fourth order moment, k |c(k)| < and for all r1 , r2 Z, k |4 (r1 , k, k + r2 )| <
where 4 (k1 , k2 , k3 ) = cum(X0 , Xk1 , Xk2 , Xk3 ).
Remark 6.2.1 (Strict Stationarity and cumulants) We note if the time series is strictly stationary then the cumulants are invariant of shift (just as the covariance is):
cum(Xt , Xt+k1 , Xt+k2 , Xt+k3 ) = cum(X0 , Xk1 , Xk2 , Xk3 ) = 4 (k1 , k2 , k3 ).

158

A simply expansion shows that

var[
cn (r)] =

n|r|
1 X
cov(Xt Xt+r , X X +r ).
n2
t, =1

One approach for the analysis of cov(Xt Xt+r , X X +r ) is to expand it in terms of expectations
cov(Xt Xt+r , X X +r ) = E(Xt Xt+r , X X +r )E(Xt Xt+r )E(X X +r ), however it not clear how this
will give var[Xt Xt+r ] = O(n1 ). Instead we observe that cov(Xt Xt+r , X X +r ) is the covariance
of the product of random variables. This belong to the general class of cumulants of products of
random variables. We now use standard results on cumulants, which show that cov[XY, U V ] =
cov[X, U ]cov[Y, V ] + cov[X, V ]cov[Y, U ] + cum(X, Y, U, V ) (note this result can be generalized to
higher order cumulants, see ?). Using this result we have
var[
cn (r)]
=

n|r|
1 X
n2

t, =1

cov(Xt , X )
|
{z
}


cov(Xt+r , X +r ) + cov(Xt , X +r )cov(Xt+r , X ) + cum(Xt , Xt+r , X , X +r )

=c(t ) by stationarity

n|r|


1 X 
c(t )2 + c(t r)c(t + r ) + k4 (r, t, + r t)
2
n
t, =1

:= I + II + III,
where the above is due to strict stationarity of the time series. We analyse the above term by
term. Either (i) by changing variables and letting k = t and thus changing the limits of the
Pn|r|
summand in an appropriate way or (ii) observing that t, =1 c(t )2 is the sum of the elements
in the Toeplitz matrix

c(0)2

c(1)2

. . . c(n |r| 1)2

c(1)2

c(0)2

2)2

..
.

..
.

. . . c(n |r|
..
..
.
.

c((n |r| 1))2 c((n |r| 2))2 . . .

c(0)2

(noting that c(k) = c(k)) the sum I can be written as

I =

n|r|
1 X
1
c(t )2 = 2
n2
n
t, =1

(n|r|1)

n|r||k|
2

c(k)

X
t=1

k=(n|r|1)

159

1
1=
n

n|r|

X
k=(n|r|)

n |r| |k|
n

c(k)2 .

Pn|r|
P
For all k, (1|k|/n)c(k)2 c(k)2 and | k=(n|r|) (1|k|/n)c(k)2 | k c(k)2 , thus by dominated
P
P
2
convergence (see Chapter A) nk=(n|r|) (1 |k|/n)c(k)2
k= c(k) . This gives

I=

1 X
1
c(k)2 + o( ).
n
n
k=

Using a similar argument we can show that

1 X
1
II =
c(k + r)c(k r) + o( ).
n
n
k=

To derive the limit of III, again we use a change of variables to give


1
III =
n

n|r|

X
k=(n|r|)

n |r| |k|
n


k4 (r, k, k + r).

Pn|r|
To bound we note that for all k, (1 |k|/n)k4 (r, k, k + r) k4 (r, k, k + r) and | k=(n|r|) (1
P
P
|k|/n)k4 (r, k, k + r)| k |k4 (r, k, k + r)|, thus by dominated convergence we have nk=(n|r|) (1
P
|k|/n)k4 (r, k, k + r)
k= k4 (r, k, k + r). This gives
III =

1X
1
4 (r, k, k + r) + o( ).
n
n
k

Therefore altogether we have

nvar[
cn (r)] =

c(k) +

k=

c(k + r)c(k r) +

k=

4 (r, k, k + r) + o(1).

k=

Using similar arguments we obtain


ncov[
cn (r1 ), cn (r2 )] =

X
k=

c(k)c(k + r1 r2 ) +

X
k=

c(k r1 )c(k + r2 ) +

4 (r1 , k, k + r2 ) + o(1).

k=

We observe that the covariance of the covariance estimator contains both covariance and cumulants
terms. Thus if we need to estimate them, for example to construct confidence intervals, this can be
extremely difficult. However, we show below that under linearity the above fourth order cumulant
term has a simpler form.

160

The covariance of the sample covariance under linearity


We recall that

c(k + r1 r2 )c(k) +

k=

c(k r1 )c(k + r2 ) +

k=

4 (r1 , k, k + r2 ) + o(1) = T1 + T2 + T3 + o(1).

k=

We now show that under linearity, T3 (the fourth order cumulant) has a much simpler form. Let
us suppose that the time series is linear
Xt =

j tj

j=

where

|j | < , {t } are iid, E(t ) = 0, var(t ) = 1 and 4 = cum4 (t ). Then T3 is

T3 =

cum

j1 j1 ,

j1 =

k=

j2 r1 j2 ,

j2 =

j3 kj3 ,

j3 =

j4 k+r2 j1

j4 =

j1 j2 j3 j4 cum (j1 , r1 j2 , kj3 , k+r2 j1 ) .

k= j1 ,...,j4 =

Standard results in cumulants (which can be proved using the characteristic function), show that
cum[Y1 , Y2 , . . . , Yn ] = 0, if any of these variables is independent of all the others. Applying this
result to cum (j1 , r1 j2 , kj3 , k+r2 j1 ) reduces T3 to

T3 = 4

j jr1 jk jr2 k .

k= j=

Using a change of variables j1 = j and j2 = j k we have

j1 =

j jr1


j2 j2 r2 = 4 c(r1 )c(r2 ),

j2 =

recalling the covariance of a linear process in Lemma 3.1.1.


Altogether this gives
ncov[
cn (r1 ), cn (r2 )] =

X
k=

c(k)c(k + r1 r2 ) +

c(k r1 )c(k + r2 ) + 4 c(r1 )c(r2 ) + o(1). (6.7)

k=

Thus in the case of linearity our expression for the variance is simpler, and the only difficult

161

parameter to estimate of 4 , which can be done using various methods.

The variance of the sample correlation under linearity


A suprisingly twist in the story is that (6.7) can be reduced further, if we are interested in estimating
the correlation rather than the covariance. We recall the sample correlation is
n (r) =

cn (r)
,
cn (0)

which is an estimator of (r) = c(r)/c(0).


Lemma 6.2.2 (Bartletts formula) Suppose {Xt } is a linear time series, where

|(j)| < .

Then the variance of the distribution of n (r) is

{(k + r) + (k r) 2(r)(k)}{(k + r) + (k r) 2(r)(k)}

k=

PROOF. By making a Taylor expansion of cn (0)1 about c(0)1 we have


n (r) (r)

=
=

cn (r) c(r)

cn (0) c(0)
[
cn (r) c(r)]
[
cn (0) c(0)]
c(0)

cn (r)
c(0)2
| {z }

replace with c(r)

cn (r)
+ [
cn (0) c(0)]2
cn (0)3
{z
}
|
=O(n1 )

c(r)
cn (r)
[
cn (r) c(r)]
[
cn (r) c(r)]
[
cn (0) c(0)]
+ 2 [
cn (0) c(0)]2
[
cn (0) c(0)]
c(0)
c(0)2
cn (0)3
c(0)2
|
{z
}
O(n1 )

1
:= An + Op ( ),
n
where cn (0) lies between cn (0) and c(0). We observe that the last two terms of the above are
of order O(n1 ) (by (6.7) and that c(0) is bounded away from zero) and the dominating term is
An which is of order O(n1/2 ) (again by (6.7)). Thus the limiting distribution of n (r) (r) is
determined by An and the variance of the limiting distribution is also determined by An . It is
straightforward to show that
nvar[An ] = n

var[
cn (r)]
c(r)2
c(r)2

2ncov[
c
(r),
c

(0)]
+
nvar[
c
(0)]
.
n
n
n
c(0)2
c(0)3
c(0)4

162

(6.8)

By using (6.7) we have

nvar

cn (r)

cn (0)

P
P
P

2
2
2 k= c(k)c(k r) + 4 c(r)c(0)
k= c(k)c(k r) + 4 c(r)
k= c(k) +

=
P
P
P
2
2
2 k= c(k)c(k r) + 4 c(r)c(0)
k= c(k)c(k r) + 4 c(0)
k= c(k) +
+o(1).
Substituting the above into (6.8) gives us

nvar[An ] =

c(k) +

k=

2 2

k=

!
2

c(k)c(k r) + 4 c(r)

k=

!
c(k)c(k r) + 4 c(r)c(0)
2

c(k) +

k=

c(r)2
+
c(0)3
!

c(k)c(k r) + 4 c(0)2

k=

c(0)2

c(r)2
+ o(1).
c(0)4

Focusing on the fourth order cumulant terms, we see that these cancel, which gives the result. 
To prove Theorem 6.2.1, we simply use the Lemma 6.2.2 to obtain an asymptotic expression
for the variance, then we use An to show asymptotic normality of cn (r) (under linearity).
Exercise 6.1 Under the assumption that {Xt } are iid random variables show that cn (1) is asymptotically normal.
Hint: Let m = n/(b + 1) and partition the sum
n1
X

Xt Xt+1 =

t=1

b
X

Pn1
k=1

Xt Xt+1 as follows

(b+1)+b

Xt Xt+1 + Xb+1 Xb+2 +

t=1

Xt Xt+1 + X(b+1)+b+1 X(b+1)+b+2 + . . . +

t=(b+1)+1
(m1)(b+1)+b+1

+X(m1)(b+1) X(m1)(b+1)+1 +

Xt Xt+1

t=(m1)(b+1)+1

m1
X
j=0

where Ub,j =

Pj(b+1)+b
t=j(b+1)

Ub,j +

m1
X

Xj(b+1)+b X(j(b+1)+1

j=0

Xt Xt+1 . Show that the second term in the above summand is asymptotically

negligible and show that the classical CLT for iid random variables can be applied to the first term.

163

Exercise 6.2 Under the assumption that {Xt } is a MA(1) process, show that cn (1) is asymptotically normal.
Exercise 6.3 The block bootstrap scheme is a commonly used method for estimating the finite
sample distribution of a statistic (which includes its variance). The aim in this exercise is to see
how well the bootstrap variance approximates the finite sample variance of a statistic.
(i) In R write a function to calculate the autocovariance b
cn (1) =

1
n

Pn1
t=1

Xt Xt+1 .

Remember the function is defined as cov1 = function(x){...}


(ii) Load the library boot library("boot") into R. We will use the block bootstrap, which partitions the data into blocks of lengths l and then samples from the blocks n/l times to construct
a new bootstrap time series of length n. For each bootstrap time series the covariance is
evaluated and this is done R times. The variance is calculated based on these R bootstrap
estimates.
You will need to use the function tsboot(tseries,statistic,R=100,l=20,sim="fixed").
tseries refers to the original data, statistic to the function you wrote in part (i) (which should
only be a function of the data), R=is the number of bootstrap replications and l is the length
of the block.
Note that tsboot(tseries,statistic,R=100,l=20,sim="fixed")$t will be vector of length
R = 100 which will contain the bootstrap statistics, you can calculate the variance of this
vector.
(iii) Simulate the AR(2) time series arima.sim(list(order = c(2, 0, 0), ar = c(1.5, 0.75)), n =
128) 500 times. For each realisation calculate the sample autocovariance at lag one and also
the bootstrap variance.
(iv) Calculate the mean of the bootstrap variances and also the mean squared error (compared
with the empirical variance), how does the bootstrap perform?
(iv) Play around with the bootstrap block length l. Observe how the block length can influence the
result.
Remark 6.2.2 The above would appear to be a nice trick, but there are two major factors that
lead to the cancellation of the fourth order cumulant term

164

Linearity of the time series


Ratio between cn (r) and cn (0).
Indeed this is not a chance result, in fact there is a logical reason why this result is true (and is
true for many statistics, which have a similar form - commonly called ratio statistics). It is easiest
explained in the Fourier domain. If the estimator can be written as
1
n

Pn

k=1 (k )In (k )
,
1 Pn
k=1 In (k )
n

where In () is the periodogram, and {Xt } is a linear time series, then we will show later that the
asymptotic distribution of the above has a variance which is only in terms of the covariances not
higher order cumulants. We prove this result in Section 9.5.

6.3

Using Bartletts formula for checking for correlation

Bartletts formula if commonly used to check by eye; whether a time series is uncorrelated (there
are more sensitive tests, but this one is often used to construct CI in for the sample autocovariances
in several statistical packages). This is an important problem, for many reasons:
Given a data set, we need to check whether there is dependence, if there is we need to analyse
it in a different way.
Suppose we fit a linear regression to time series data. We may to check whether the residuals
are actually uncorrelated, else the standard errors based on the assumption of uncorrelatedness would be unreliable.
We need to check whether a time series model is the appropriate model. To do this we fit
the model and estimate the residuals. If the residuals appear to be uncorrelated it would
seem likely that the model is correct. If they are correlated, then the model is inappropriate.
For example, we may fit an AR(1) to the data, estimate the residuals t , if there is still
t1 is
correlation in the residuals, then the AR(1) was not the correct model, since Xt X
still correlated (which it would not be, if it were the correct model).

165

ACF

0.0

0.2

0.4

0.6

0.8

1.0

Series iid

10

15

20

Lag

Figure 6.1: The sample ACF of an iid sample with error bars (sample size n = 200).
We now apply Theorem 6.2.1 to the case that the time series are iid random variables. Suppose {Xt }
are iid random variables, then it is clear that it is trivial example of a (not necessarily Gaussian)
linear process. We use (6.3) as an estimator of the autocovariances.
To derive the asymptotic variance of {
cn (r)}, we recall that if {Xt } are iid then (k) = 0 for
k 6= 0. Substituting this into (6.6) we see that

n(
n (h) (h)) N (0, Wh ),

where

1 i=j
(Wh )ij =
0 i=
6 j
In other words,

n(
n (h) (h)) N (0, Ih ). Hence the sample autocovariances at different lags

are asymptotically uncorrelated and have variance one. This allows us to easily construct error
bars for the sample autocovariances under the assumption of independence. If the vast majority of
the sample autocovariance lie inside the error bars there is not enough evidence to suggest that the
data is a realisation of a iid random variables (often called a white noise process). An example of
the empirical ACF and error bars is given in Figure 6.1. We see that the empirical autocorrelations
of the realisation from iid random variables all lie within the error bars.

In contrast in Figure

6.2 we give a plot of the sample ACF of an AR(2). We observe that a large number of the sample
autocorrelations lie outside the error bars.

166

0.4 0.0 0.4 0.8

ACF

Series ar2

10

15

20

0.4 0.0 0.4 0.8

acf

Lag

10

15

20

lag

Figure 6.2: Top: The sample ACF of the AR(2) process Xt = 1.5Xt1 + 0.75Xt2 + t with
error bars n = 200. Bottom: The true ACF.
Of course, simply checking by eye means that we risk misconstruing a sample coefficient that
lies outside the error bars as meaning that the time series is correlated, whereas this could simply
be a false positive (due to multiple testing). To counter this problem, we construct a test statistic

D
n (h) (h)) N (0, I), one method of
for testing uncorrelatedness. Since under the null n(
testing is to use the square correlations

Sh = n

h
X

|
n (r)|2 ,

(6.9)

r=1

under the null it will asymptotically have a 2 -distribution with h degrees of freedom, under the
alternative it will be a non-central (generalised) chi-squared. The non-centrality is what makes us
reject the null if the alternative of correlatedness is true. This is known as the Box-Pierce test (a
test which gives better finite sample results is the Ljung-Box test). Of course, a big question is
how to select h. In general, we do not have to use large h since most correlations will arise when
r is small, However the choice of h will have an influence on power. If h is too large the test will
loose power (since the mean of the chi-squared grows as h ), on the other hand choosing h too
small may mean that certain correlations at higher lags are missed. How to selection h is discussed
in several papers, see for example Escanciano and Lobato (2009).

167

6.4

Long range dependence versus changes in the mean

A process is said to have long range dependence if the autocovariances are not absolutely summable,
P
ie.
k |c(k)| = .
From a practical point of view data is said to exhibit long range dependence if the autocovariances do not decay very fast to zero as the lag increases. Returning to the Yahoo data considered
in Section 4.1.1 we recall that the ACF plot of the absolute log differences, given again in Figure
6.3 appears to exhibit this type of behaviour. However, it has been argued by several authors that

0.0

0.2

0.4

ACF

0.6

0.8

1.0

Series abs(yahoo.log.diff)

10

15

20

25

30

35

Lag

Figure 6.3: ACF plot of the absolute of the log differences.

the appearance of long memory is really because of a time-dependent mean has not been corrected
for. Could this be the reason we see the memory in the log differences?
We now demonstrate that one must be careful when diagnosing long range dependence, because
a slow/none decay of the autocovariance could also imply a time-dependent mean that has not been
corrected for. This was shown in Bhattacharya et al. (1983), and applied to econometric data in
Mikosch and St
aric
a (2000) and Mikosch and Starica (2003). A test for distinguishing between long
range dependence and change points is proposed in Berkes et al. (2006).
Suppose that Yt satisfies
Yt = t + t ,
where {t } are iid random variables and the mean t depends on t. We observe {Yt } but do not
know the mean is changing. We want to evaluate the autocovariance function, hence estimate the

168

autocovariance at lag k using

cn (k) =

n|k|
1 X
(Yt Yn )(Yt+|k| Yn ).
n
t=1

Observe that Yn is not really estimating the mean but the average mean! If we plotted the empirical
ACF {
cn (k)} we would see that the covariances do not decay with time. However the true ACF
would be zero and at all lags but zero. The reason the empirical ACF does not decay to zero is
because we have not corrected for the time dependent mean. Indeed it can be shown that

cn (k) =

n|k|
1 X
(Yt t + t Yn )(Yt+|k| t+k + t+k Yn )
n
t=1

n|k|

n|k|
1 X
1 X
(Yt t )(Yt+|k| t+k ) +
(t Yn )(t+k Yn )
n
n
t=1

t=1

n|k|

c(k)
|{z}

true autocovariance=0

1 X
(t Yn )(t+k Yn )
n
| t=1
{z
}

additional term due to time-dependent mean

Expanding the second term and assuming that k << n and t (t/n) (and is thus smooth) we
have

n|k|
1 X
(t Yn )(t+k Yn )
n
t=1
!2
n
n
1X 2
1X
t
t + op (1)
n
n
t=1
t=1
!2
n
n
n
1 XX 2
1X
t
t + op (1)
n2
n
s=1 t=1

t=1

n
n
n
n
n
n
1 XX
1 XX
1 XX
2
t (t s ) = 2
(t s ) + 2
s (t s )
n2
n
n
s=1 t=1
s=1 t=1
s=1 t=1
|
{z
}
=

1
n2

Pn

s=1

Pn

t=1

t (t s )

Therefore
n|k|
n
n
1 X
1 XX
(t Yn )(t+k Yn ) 2
(t s )2 .
n
2n
t=1

s=1 t=1

Thus we observe that the sample covariances are positive and dont tend to zero for large lags.

169

This gives the false impression of long memory.


It should be noted if you study a realisation of a time series with a large amount of dependence,
it is unclear whether what you see is actually a stochastic time series or an underlying trend. This
makes disentangling a trend from data with a large amount of correlation extremely difficult.

170

Chapter 7
Parameter estimation
Prerequisites
The Gaussian likelihood.

Objectives
To be able to derive the Yule-Walker and least squares estimator of the AR parameters.
To understand what the quasi-Gaussian likelihood for the estimation of ARMA models is,
and how the Durbin-Levinson algorithm is useful in obtaining this likelihood (in practice).
Also how we can approximate it by using approximations of the predictions.
Understand that there exists alternative methods for estimating the ARMA parameters,
which exploit the fact that the ARMA can be written as an AR().
We will consider various methods for estimating the parameters in a stationary time series.
We first consider estimation parameters of an AR and ARMA process. It is worth noting that we
will look at maximum likelihood estimators for the AR and ARMA parameters. The maximum
likelihood will be constructed as if the observations were Gaussian. However, these estimators
work both when the process is Gaussian is also non-Gaussian. In the non-Gaussian case, the
likelihood simply acts as a contrast function (and is commonly called the quasi-likelihood). In time
series, often the distribution of the random variables is unknown and the notion of likelihood has
little meaning. Instead we seek methods that give good estimators of the parameters, meaning that
they are consistent and as close to efficiency as possible without placing too many assumption on

171

the distribution. We need to free ourselves from the notion of likelihood acting as a likelihood
(and attaining the Cr
amer-Rao lower bound).

7.1

Estimation for Autoregressive models

Let us suppose that {Xt } is a zero mean stationary time series which satisfies the AR(p) representation
Xt =

p
X

j Xtj + t ,

j=1

where E(t ) = 0 and var(t ) = 2 and the roots of the characteristic polynomial 1

Pp

j=1 j z

lie

outside the unit circle. We will assume that the AR(p) is causal (the techniques discussed here
will not consistently estimate the parameters in the case that the process is non-causal, they will
only consistently estimate the corresponding causal model). Our aim in this section is to construct
estimator of the AR parameters {j }. We will show that in the case that {Xt } has an AR(p)
representation the estimation is relatively straightforward, and the estimation methods all have
properties which are asymptotically equivalent to the Gaussian maximum estimator.
The Yule-Walker estimator is based on the Yule-Walker equations derived in (3.4) (Section
3.1.4).

7.1.1

The Yule-Walker estimator

We recall that the Yule-Walker equation state that if an AR process is causal, then for i > 0 we
have
E(Xt Xti ) =

p
X

j E(Xtj Xti ), c(i) =

j=1

p
X

j c(i j).

(7.1)

j=1

Putting the cases 1 i p together we can write the above as


rp = p p ,

(7.2)

where (p )i,j = c(i j), (rp )i = c(i) and 0p = (1 , . . . , p ). Thus the autoregressive parameters
solve these equations.

172

The Yule-Walker equations inspire the method of moments estimator called the Yule-Walker
p are estimators of
estimator. We use (7.2) as the basis of the estimator. It is clear that rp and
p )i,j = cn (i j) and (rp )i = cn (i). Therefore we can use
rp and p where (
=
1

rp ,
p
p

(7.3)

as an estimator of the AR parameters 0p = (1 , . . . , p ). We observe that if p is large this involves


by
inverting a large matrix. However, we can use the Durbin-Levinson algorithm to calculate
p
recursively fitting lower order AR processes to the observations and increasing the order. This way
an explicit inversion can be avoided. We detail how the Durbin-Levinson algorithm can be used to
estimate the AR parameters below.
Step 1 Set 1,1 = cn (1)/
cn (0) and rn (2) = 2
cn (0) 21,1 cn (1).
Step 2 For 2 t p, we define the recursion
t,t =
t,j

cn (t)

= t1,j

Pt1
n (t j)
j=1 t1,j c

rn (t)
t,t t1,tj

1 j t 1,

and rn (t + 1) = rn (t)(1 2t,t ).


Step 3 We recall from (5.12) that t,t is the partial correlation between Xt+1 and X1 , therefore tt
are estimators of the partial correlation between Xt+1 and X1 .
As mentioned in Step 3, the Yule-Walker estimators have the useful property that the partial
correlations can easily be evaluated within the procedure. This is useful when trying to determine
the order of the model to fit to the data. In Figure 7.1 we give the partial correlation plot corresponding to Figure 6.1. Notice that only the first two terms are outside the error bars. This
rightly suggests the time series comes from an autoregressive process of order two. Assuming that
t,t is asymptotically normal the error bars (confidence interval) is determined by the variance of
t,t . This can be determined by deriving the variance of t .
The Yule-Walker estimator has the useful property that the parameter estimates {j ; j =

1, . . . , p} correspond to a causal AR(p), in other words, the roots corresponding to (z)


= 1
Pp j
cn (r)} form a positive definite
j=1 j z lie outside the unit circle. This is because the covariances {
p+1 )i,j =
sequence, thus there exists a random vector Z p+1 = (Z1 , . . . , Zp+1 ) where var[Z]p+1 = (

173

0.0
0.5

Partial ACF

0.5

Series ar2

10

15

20

Lag

Figure 7.1: Top: The sample partial autocorrelation plot of the AR(2) process Xt =
1.5Xt1 + 0.75Xt2 + t with error bars n = 200.
cn (ij), using this and the following result it follows that {j ; j = 1, . . . , p} corresponds to a causal
AR process.
Lemma 7.1.1 Let us suppose Z p+1 = (Z1 , . . . , Zp+1 ) is a random vector, where var[Z]p+1 =
(p+1 )i,j = cn (i j) (which is Toeplitz). Let Zp+1|p be the best linear predictor of Zp+1 given
Zp , . . . , Z1 , where p = (1 , . . . , p ) = 1
p r p are the coefficients corresponding to the best linear
P
predictor. Then the roots of the corresponding characteristic polynomial (z) = 1 pj=1 j z j lie
outside the unit circle.
PROOF. We first note that by definition of the best linear predictor, for any coefficients {aj } we
have the inequality

E Zp+1

p
X

j Zp+1j = E ((B)Zp+1 )2 E Yp+1

j=1

p
X

2
aj Zp+1j = E (a(B)Zp+1 )2 .(7.4)

j=1

We use the above inequality to prove the result by contradiction. Let us suppose that there exists
at least one root of (z), which lies inside the unit circle. We denote this root as 1 (|| > 1) and
factorize (z) as (z) = (1 z)R(z), where R(z) contains the other remaining roots and can be
either inside or outside the unit circle.
Define the two new random variables, Yp+1 = R(B)Zp+1 and Yp = R(B)Zp (where B acts
as the backshift operator), which we note is a linear combination of Zp+1 , . . . , Z2 and Zp , . . . , Z1
P
Pp1
respectively ie. Yp+1 = p1
0=1 Ri Zp+1i and Yp =
i=0 Ri Zpi . The most important observation

174

in this construction is that the matrix p+1 is Toeplitz (ie. {Zt } is a stationary vector), therefore
Yp+1 and Yp have the same covariance structure, in particular they have the same variance. Let
=

cov[Yp+1 , Yp ]
cov[Yp+1 , Yp ]
= cor(Yp+1 , Yp ).
=q
var[Yp ]
var[Yp ]var[Yp+1 ]
{z
}
|
by stationarity

We recall that Yp is the best linear predictor of Yp+1 given Yp and || 1.


We now return to the proof. We start by defining a new polynomial 1 (z) = (1 z)R(z). We
will show that the mean squared error of E[1 (B)Zp+1 ]2 E[(B)Zp+1 ]2 , which by (7.4) leads to
a contradiction. Evaluating E[1 (B)Zp+1 ]2 we have
E[1 (B)Zp+1 ]2 = E[(1 B)R(B)Zp+1 ]2 = E[R(B)Zp+1 BR(B)Zp+1 ]2
= E[Yp+1 Yp ]2 E[Yp+1 Yp ]2 = E[(B)Zp+1 ]2 .
From (7.4), the above can only be true if = , and thus the root lies outside the unit circle.
Further, we see that all the roots of (z) can only lie outside the unit circle. Thus giving the
required result.

The above result can immediately be used to show that the Yule-Walker estimators of the AR(p)
coefficients yield a causal solution. Since the autocovariance estimators {
cn (r)} form a positive
p+1 with (
p+1 ) = cn (i j),
semi-definite sequence, there exists a vector Y p where var[Y p+1 ] =
1
thus by the above lemma we have that
p r p are the coefficients of a Causal AR process. We note
that below we define the intuitively obvious least squares estimator, which does not necessarily
have this property.
The least squares estimator is based can either be defined in its own right or be considered as
the conditional Gaussian likelihood. We start by defining the Gaussian likelihood.

7.1.2

The Gaussian maximum likelihood

Our object here is to obtain the maximum likelihood estimator of the AR(p) parameters. We recall
that the maximum likelihood estimator is the parameter which maximises the joint density of the
observations. Since the log-likelihood often has a simpler form, we will focus on the log-likelihood.
We note that the Gaussian MLE is constructed as if the observations {Xt } were Gaussian, though it

175

is not necessary that {Xt } is Gaussian when doing the estimation. In the case that the innovations
are not Gaussian, estimator will be less efficient (will not obtain the Cramer-Rao lower bound)
then the likelihood constructed as if the distribution were known.
Suppose we observe {Xt ; t = 1, . . . , n} where Xt are observations from an AR(p) process. Let
us suppose for the moment that the innovations of the AR process are Gaussian, this implies that
X n = (X1 , . . . , Xn ) is a n-dimension Gaussian random vector, with the corresponding log-likelihood
Ln (a) = log |n (a)| X0n n (a)1 Xn ,

(7.5)

where t (a) the variance covariance matrix of Xn constructed as if Xn came from an AR process
with parameters a. Of course, in practice in the likelihood in the form given above is impossible to
maximise. Therefore we need to rewrite the likelihood in a more tractable form.
We now derive a tractable form of the likelihood under the assumption that the innovations
come from an arbitrary distribution. To construct the likelihood, we use the method of conditioning,
to write the likelihood as the product of conditional likelihoods. In order to do this, we derive the
conditional distribution of Xt+1 given Xt1 , . . . , X1 . We first note that the AR(p) process is pMarkovian, therefore if t p all the information about Xt+1 is contained in the past p observations,
therefore
P(Xt+1 x|Xt , Xt1 , . . . , X1 ) = P(Xt+1 x|Xt , Xt1 , . . . , Xtp+1 ).

(7.6)

Since the Markov property applies to the distribution function it also applied to the density
f (Xt+1 |Xt , . . . , X1 ) = f (Xt+1 |Xt , . . . , Xtp+1 ).
By using the (7.6) we have

P(Xt+1 x|Xt , . . . , X1 ) = P(Xt+1 x|Xt , . . . , X1 ) = P ( x

p
X

aj Xt+1j ),

(7.7)

j=1

where P denotes the distribution of the innovation. Differentiating P with respect to Xt+1 gives

f (Xt+1 |Xt , . . . , Xtp+1 ) =

P ( Xt+1

Pp

j=1 aj Xt+1j )

Xt+1

176

= f Xt+1

p
X
j=1

aj Xt+1j . (7.8)

Example 7.1.1 (AR(1)) To understand why (7.6) is true consider the simple case that p = 1
(AR(1)). Studying the conditional probability gives
P(Xt+1 xt+1 |Xt = xt , . . . , X1 = x1 ) = P(

aX + t xt+1
}
| t {z

|Xt = xt , . . . , X1 = x1 )

all information contained in Xt

= P (t xt+1 axt ) = P(Xt+1 xt+1 |Xt = xt ),


where P denotes the distribution function of the innovation .
Using (7.8) we can derive the joint density of {Xt }nt=1 . By using conditioning we obtain

f(X1 , X2 , . . . , Xn ) = f (X1 , . . . , Xp )

n1
Y

f(Xt+1 |Xt , . . . , X1 )

(by repeated conditioning)

t=p

= f(X1 , . . . , Xp )

n1
Y

f(Xt+1 |Xt , . . . , Xtp+1 )

(by the Markov property)

t=p

= f(X1 , . . . , Xp )

n1
Y

f (Xt+1

t=p

p
X

aj Xt+1j )

(by (7.8)).

j=1

Therefore the log likelihood is


p
n1
X
X
log f (Xt+1
aj Xt+1j ) .
log f(X1 , X2 , . . . , Xn ) = log f(X1 , . . . , Xp ) +
{z
}
|
|
{z
} t=p
j=1
Full log-likelihood Ln (a;X n )
initial observations
|
{z
}
conditional log-likelihood=Ln (a;X n )

In the case that the sample sizes are large n >> p, the contribution of initial observations
log f(X1 , . . . , Xp ) is minimal and the conditional log-likelihood and full log-likelihood are asymptotically equivalent.
So far we have not specified the distribution of . From now on we shall assume that it is
Gaussian. In the case that is Gaussian, log f(X1 , . . . , Xp ) is multivariate normal with mean zero
(since we are assuming, for convenience, that the time series has zero mean) and variance p . We
recall that p (a) is a Toeplitz matrix whose covariance is determined by the AR parameters a, see
(3.7). As can be seen from (3.7), the coefficients are buried within the covariance (which is in
terms of the roots of the characteristic), this makes it quite an unpleasant part of the likelihood to

177

maximise. On the other hand the conditional log-likelihood has a far simpler form

Ln (a; X) = (n p) log 2

1
2

n1
X

Xt+1

t=p

p
X

2
aj Xt+1j .

j=1

The maximum likelihood estimator is




b = arg max log |p (a)| X 0 p (a)1 X + Ln (a; X) .

p
p
n
a

(7.9)

By constraining the parameter space, we an ensure the estimator correspond to a causal AR process.
However, it is clear that despite having the advantage that it attains the Cramer-Rao lower bound
in the case that the innovations are Gaussian, it not simple to evaluate. A far simpler estimator can
be obtained, by simply focusing on the conditiona log likelihood Ln (a; X). An explicit expression
for its maximum can easily be obtained (as long as we do not constrain the parameter space). It
= arg max Ln (a; X) and that
is simply the least squares estimator, in other words,
p
=
1 rp ,

p
p
p )i,j =
where (

1
np

Pn

t=p+1 Xti Xtj

and (rn )i =

1
np

Pn

t=p+1 Xt Xti .

Remark 7.1.1 (A comparison of the Yule-Walker and least squares estimators) Comparing
=
=
1 rp with the Yule-Walker estimator
1 rp we see that
the least squares estimator
p
p
p
p
p and
p (and the corresponding rp and rp ). We see
they are very similar. The difference lies in
p is a Toeplitz matrix, defined entirely by the positive definite sequence cn (r). On the other
that
p is not a Toeplitz matrix, the estimator of c(r) changes subtly at each row. This means
hand,
that the proof given in Lemma 7.1.1 cannot be applied to the least squares estimator as it relies
on the matrix p+1 (which is a combination of p and rp ) being Toeplitz (thus stationary). Thus
the characteristic polynomial corresponding to the least squares estimator will not necessarily have
roots which lie outside the unit circle.
Example 7.1.2 (Toy Example) To illustrate the difference between the Yule-Walker and least
squares estimator (at least for example samples) consider the rather artifical example that the time
series consists of two observations X1 and X2 (we will assume the mean is zero). We fit an AR(1)

178

model to the data, the least squares estimator of the AR(1) parameter is
X1 X2
bLS =
X22
whereas the Yule-Walker estimator of the AR(1) parameter is
bY W =

X1 X2
.
X12 + X22

It is clear that bLS < 1 only if X2 < X1 . On the other hand bY W < 1. Indeed since (X1 X2 )2 > 0,
we see that bY W 1/2.
Exercise 7.1 In R you can estimate the AR parameters using ordinary least squares (ar.ols),
yule-walker (ar.yw) and (Gaussian) maximum likelihood (ar.mle).
Simulate the causal AR(2) model Xt = 1.5Xt1 0.75Xt2 + t using the routine arima.sim
(which gives Gaussian realizations) and also innovations which from a t-distribution with 4df. Use
the sample sizes n = 100 and n = 500 and compare the three methods through a simulation study.
Exercise 7.2 None of these methods are able to consistently estimator the parameters of a noncausal AR(p) time series. This is because all these methods are estimating the autocovariance
function (regardless of whether the Yule-Walker of least squares method is used). It is possible that
other criterions may give a consistent estimator. For example the `1 -norm defined as




p
t
X
X


Xt
Ln () =
j Xtj ,


t=p+1
j=1
with n = arg min Ln ().
(i) Simulate a stationary solution of the non-causal AR(1) process Xt = 2Xt1 + t , where the
innovations come from a double exponential and estimate using Ln (). Do this 100 times,
does this estimator appear to consistently estimate 2?
(ii) Simulate a stationary solution of the non-causal AR(1) process Xt = 2Xt1 + t , where the
innovations come from a t-distribution with 4 df and estimate using Ln (). Do this 100
times, does this estimator appear to consistently estimate 2?
You will need to use a Quantile Regression package to minimise the `1 norm. I suggest using the
package quantreg and the function rq where we set = 0.5 (the median).

179

7.2

Estimation for ARMA models

Let us suppose that {Xt } satisfies the ARMA representation

Xt

p
X

i Xti = t +

i=1

q
X

j tj ,

j=1

and = (1 , . . . , q ), = (1 , . . . , p ) and 2 = var(t ). We will suppose for now that p and


q are known. The objective in this section is to consider various methods for estimating these
parameters.

7.2.1

The Gaussian maximum likelihood estimator

We now derive the Gaussian maximum likelihood estimator (GMLE) to estimate the parameters
and . Let X 0n = (X1 , . . . , Xn ). The criterion (the GMLE) is constructed as if {Xt } were Gaussian,
but this need not be the case. The likelihood is similar to the likelihood given in (7.5), but just as
in the autoregressive case it can be not directly maximised, ie.
Ln (, , ) = log |n (, , )| X 0n n (, , )1 X n ,

(7.10)

where n (, , ) the variance covariance matrix of X n . However, in Section 5.3.3, equation (5.21)
the Cholesky decomposition of n is given and using this we can show that
X 0n n (, , )1 X n

P
n1
X (Xt+1 tj=1 t,j ()Xt+1j )2
X12
=
+
,
r(1; )
r(t + 1; )
t=1

0
1
2
where = (, , 2 ). Furthermore, since 1
n = Ln Dn Ln , then det(n ) = det(Ln ) det(Dn ) =
Qn
Pn
1
t=1 r(t) , this implies log |n (, , )| =
t=1 log r(t; ). Thus the log-likelihood is

Ln (, , ) =

n
X
t=1

P
n1
X (Xt+1 tj=1 t,j ()Xt+1j )2
X12
log r(t; )

.
r(1; )
r(t + 1; )
t=1

180

We recall from (5.24) that best linear predictor,

Pt

j=1 t,j ()Xt+1j ,

can be simplified by taking

into account the ARMA structure


X 0n n (, , )1 X n

X12
+
r(1; )
n1
X

max(p,q)

(Xt+1

Pt

j=1 t,j ()Xt+1j )

+
r(t + 1; )
P
P
(Xt+1 pj=1 j Xt+1j qi=1 t,i (Xt+1i Xt+1i|ti ())2
t=1

r(t + 1; )

max(p,q)

Substituting this into Ln () gives


P
max(p,q)
X (Xt+1 tj=1 t,j ()Xt+1j )2
X12
Ln (, , ) =
log r(t; )

r(1; )
r(t + 1; )
t=1
t=1
P
P
n1
X (Xt+1 pj=1 j Xt+1j qi=1 t,i (Xt+1i Xt+1i|ti ())2
n
X

r(t + 1; )

max(p,q)

, 2 which maximises Ln ().


The maximum likelihood estimator are the parameters n ,
n
We can also use an approximation to the log-likelihood which can simplify the estimation
t+1|t . This motivates the approxischeme. We recall in Section 5.4 we approximated Xt+1|t with X
t+1|t , where X
t+1|t is defined in (5.25) and r(t, )
mation where we replace in Ln () Xt+1|t with X
with 2 to give the approximate Gaussian log-likelihood

b n () =
L

n
X

log 2

t=1

n
X
t=1

log 2

n1
X
t=2
n1
X
t=2

bt+1|t ()]2
[Xt+1 X
2
[((B)1 (B))[t] Xt+1 ]2
2

where ((B)1 (B))[t] denotes the approximation of the polynomial in B to the tth order. This
approximate likelihood greatly simplifies the estimation scheme because the derivatives (which is
the main tool used in the maximising it) can be easily obtained. To do this we note that
d (B)
B i (B)
(B)
Xt =
Xt =
Xti
2
di (B)
(B)
(B)2
d (B)
Bj
1
Xt =
Xt =
Xtj
2
dj (B)
(B)
(B)2

181

(7.11)

therefore
d
di

(B)
Xt
(B)

2


= 2

(B)
Xt
(B)



(B)
Xti
(B)2

d
and
dj

(B)
Xt
(B)

2


= 2

(B)
Xt
(B)




1
Xtj .
(B)2
(7.12)

Substituting this into the approximate likelihood gives the derivatives


b
L
i
b
L
j
b
L
2

#
"
n
i  (B) 

2 Xh
= 2
(B)1 (B) [t] Xt
Xti

(B)2 [ti]
t=1
#
"
n
i  1 

2 Xh
= 2
Xtj
(B)1 (B) [t] Xt

(B) [tj]
t=1

n
i2

1
1 Xh
1

(B)
(B)
X
.
[t] t
2 n 4

(7.13)

t=1

We then use the Newton-Raphson scheme to solve maximise the approximate likelihood. It can be
shown that the approximate likelihood is close the actual true likelihood and asymptotically both
methods are equivalent.
Theorem 7.2.1 Let us suppose that Xt has a causal and invertible ARMA representation
Xt

p
X

j Xtj = t +

j=1

q
X

i ti

i=1

where {t } are iid random variables with mean zero and var[t ] = 2 . Then the the (quasi)-Gaussian

n
n

N (0, 1 ),

with

E(Ut U0t ) E(Vt U0t )

E(Ut Vt0 ) E(Vt Vt0 )

and Ut = (Ut , . . . , Utp+1 ) and Vt = (Vt , . . . , Vtq+1 ), where {Ut } and {Vt } are autoregressive
processes which satisfy (B)Ut = t and (B)Vt = t .
We do not give the proof in this section, however it is possible to understand where this result
comes from. We recall that that the maximum likelihood and the approximate likelihood are

182

asymptotically equivalent. They are both approximations of the unobserved likelihood

e n () =
L

n
X

log

t=1

n1
X
t=2

n1

t=1

t=2

X
X [(B)1 (B)Xt+1 ]2
[Xt+1 Xt (1; )]2
2
=

log

.
2
2

This likelihood is infeasible in the sense that it cannot be maximised since the finite past X0 , X1 , . . .
is unobserved, however is a very convenient tool for doing the asymptotic analysis. Using Lemma
b n and L
e n are all asymptotically equivalent.
5.4.1 we can show that all three likelihoods Ln , L
b n we can simply consider the
Therefore, to obtain the asymptotic sampling properties of Ln or L
en .
unobserved likelihood L
To show asymptotic normality (we assume here that the estimators are consistent) we need to
e n (since the asymptotic properties are determined by
consider the first and second derivative of L
Taylor expansions). In particular we need to consider the distribution of
and the expectation of

en
2L
2

e
L
i
e
L
j

en
L

at its true parameters

at its true parameters. We note that by using (7.12) we have




n
 
2 X
(B)
1
(B)
(B)
X
X
ti
t
2
(B)2
t=1



n
 
1
2 X
(B)1 (B) Xt
Xtj
= 2

(B)
=

(7.14)

t=1


Since we are considering the derivatives at the true parameters we observe that (B)1 (B) Xt =
t ,
(B)
(B) (B)
1
Xti =
ti =
ti = Vti
2
2
(B)
(B) (B)
(B)
and
1 (B)
1
1
Xtj =
tj =
tj = Utj .
(B)
(B) (B)
(B)
Thus (B)Ut = t and (B)Vt = t are autoregressive processes (compare with theorem). This
means that the derivative of the unobserved likelihood can be written as
e
L
i

n
n
e
2 X
L
2 X
= 2
t Uti and
= 2
t Vtj

t=1

(7.15)

t=1

Note that by causality t , Uti and Vtj are independent. Again like many of the other estimators we

183

have encountered this sum is mean-like so can show normality of it by using a central limit theorem
L
; i = 1, . . . , q},
designed for dependent data. Indeed we can show asymptotically normality of {
i
e

L
; j = 1, . . . , p} and their linear combinations using the Martingale central limit theorem, see
{
j
e

Theorem 3.2 (and Corollary 3.1), Hall and Heyde (1980) - note that one can also use m-dependence.
L L
Moreover, it is relatively straightforward to show that n1/2 (
,
) has the limit variance matrix
i j
e

2b

. Finally, by taking second derivative of the likelihood we can show that E[n1 L2 ] = . Thus
giving us the desired result.

7.2.2

The Hannan-Rissanen AR() expansion method

The methods detailed above require good initial values in order to begin the maximisation (in order
to prevent convergence to a local maximum).
We now describe a simple method first propose in Hannan and Rissanen (1982) and An et al.
(1982). It is worth bearing in mind that currently the large p small n problem is a hot topic.
These are generally regression problems where the sample size n is quite small but the number of
regressors p is quite large (usually model selection is of importance in this context). The methods
proposed by Hannan involves expanding the ARMA process (assuming invertibility) as an AR()
process and estimating the parameters of the AR() process. In some sense this can be considered
as a regression problem with an infinite number of regressors. Hence there are some parallels
between the estimation described below and the large p, small n problem.
As we mentioned in Lemma 2.5.1, if an ARMA process is invertible it is can be represented as
Xt =

bj Xtj + t .

(7.16)

j=1

The idea behind Hannans method is to estimate the parameters {bj }, then estimate the innovations
t , and use the estimated innovations to construct a multiple linear regression estimator of the
ARMA paramters {i } and {j }. Of course in practice we cannot estimate all parameters {bj } as
there are an infinite number of them. So instead we do a type of sieve estimation where we only
estimate a finite number and let the number of parameters to be estimated grow as the sample size
increases. We describe the estimation steps below:
n
(i) Suppose we observe {Xt }nt=1 . Recalling (7.16), will estimate {bj }pj=1
parameters. We will

suppose that pn as n and pn << n (we will state the rate below).

184

n
We use Yule-Walker to estimate {bj }pj=1
, where

b =
1 rp ,
pn
pn n
where
n|ij|
n|j|
1 X
1 X

(Xt X)(Xt+|ij| X) and (rpn )j =


(Xt X)(X
(pn )i,j =
t+|j| X).
n
n
t=1

t=1

n
(ii) Having estimated the first {bj }pj=1
coefficients we estimate the residuals with

t = Xt

pn
X

bj,n Xtj .

j=1

, where
(iii) Now use as estimates of 0 and 0
n n
n
X

, = arg min

n n

(Xt

t=pn +1

p
X

j Xtj

j=1

q
X

i ti )2 .

i=1

We note that the above can easily be minimised. In fact


, ) = R
1 sn
(
n
n n
where
n = 1
R
n

n
X

t
Y t Y

t=max(p,q)

1
and sn =
n

n
X

Y t Xt ,

t=max(p,q)

0
Y t = (Xt1 , . . . , Xtp , t1 , . . . , tq ).

7.3

The quasi-maximum likelihood for ARCH processes

In this section we consider an estimator of the parameters a0 = {aj : j = 0, . . . , p} given the


observations {Xt : t = 1, . . . , N }, where {Xt } is a ARCH(p) process. We use the conditional logP
likelihood to construct the estimator. We will assume throughout that E(Zt2 ) = 1 and pj=1 j =
< 1.

185

We now construct an estimator of the ARCH parameters based on Zt N (0, 1). It is worth
mentioning that despite the criterion being constructed under this condition it is not necessary
that the innovations Zt are normally distributed. In fact in the case that the innovations are not
normally distributed but have a finite fourth moment the estimator is still good. This is why it
is called the quasi-maximum likelihood , rather than the maximum likelihood (similar to the how
the GMLE estimates the parameters of an ARMA model regardless of whether the innovations are
Gaussian or not).
q
P
2 , E(X |X
Let us suppose that Zt is Gaussian. Since Zt = Xt / a0 + pj=1 aj Xtj
t
t1 , . . . , Xtp ) =
Pp
2 , then the log density of X given X
0 and var(Xt |Xt1 , . . . , Xtp ) = a0 + j=1 aj Xtj
t
t1 , . . . , Xtp
is
log(a0 +

p
X

2
aj Xtj
)+

j=1

a0 +

X2
Pp t

2 .
j=1 aj Xtj

Therefore the conditional log density of Xp+1 , Xp+2 , . . . , Xn given X1 , . . . , Xp is


n 
X

log(a0 +

t=p+1

p
X

2
aj Xtj
)

j=1

Xt2

a0 +

Pp

2
j=1 aj Xtj


.

This inspires the the conditional log-likelihood



p
n 
X
X
1
Xt2
2
P
Ln () =
log(0 +
.
j Xtj ) +
2
np
0 + pj=1 j Xtj
t=p+1
j=1
To obtain the estimator we define the parameter space

= { = (0 , . . . , p ) :

p
X

j 1, 0 < c1 0 c2 < , c1 j }

j=1

and assume the true parameters lie in its interior a = (a0 , . . . , ap ) Int(). We let
a
n = arg min Ln ().

(7.17)

The method for estimation of GARCH parameters parallels the approximate likelihood ARMA
estimator given in Section 7.2.1.

186

Chapter 8
Spectral Representations
Prerequisites
Knowledge of complex numbers.
Have some idea of what the covariance of a complex random variable (we do define it below).
Some idea of a Fourier transform (a review is given in Section A.3).

Objectives
Know the definition of the spectral density.
The spectral density is always non-negative and this is a way of checking that a sequence is
actually non-negative definite (is a autocovariance).
The DFT of a second order stationary time series is almost uncorrelated.
The spectral density of an ARMA time series, and how the roots of the characteristic polynomial of an AR may influence the spectral density function.
There is no need to understand the proofs of either Bochners (generalised) theorem or the
spectral representation theorem, just know what these theorems are. However, you should
P
know the proof of Bochners theorem in the simple case that r |rc(r)| < .

187

8.1

How we have used Fourier transforms so far

We recall in Section 1.2.3 that we considered models of the form


Xt = A cos (t) + B sin (t) + t

t = 1, . . . , n.

(8.1)

where t are iid random variables with mean zero and variance 2 and is unknown. We estimated
P
the frequency by taking the Fourier transform Jn () = 1n nt=1 Xt eit and using as an estimator
of , the value which maximised |Jn ()|2 . As the sample size grows the peak (which corresponds
the frequency estimator) grows in size. Besides the fact that this corresponds to the least squares
estimator of , we note that
1
Jn (k ) =
n

1 X
Xt exp(itk )
2n
t=1

n
1 X t
=
( ) exp(itk ) +
2n
n
t=1
|
{z
}
=O(1)

where k =

2k
n ,

1 X
t exp(itk )
2n
t=1
|
{z
}

1
=Op (n1/2 ) compare with n

(8.2)

Pn

t=1 t

is an estimator the the Fourier transform of the deterministic mean at frequency

k. In the case that the mean is simply the sin function, there is only one frequency which is nonzero. A plot of one realization (n = 128), periodogram of the realization, periodogram of the iid
noise and periodogram of the sin function is given in Figure 8.1. Take careful note of the scale (yaxis), observe that the periodogram of the sin function dominates the the periodogram of the noise
(magnitudes larger). We can understand why from (8.2), where the asymptotic rates are given and
we see that the periodogram of the deterministic signal is estimating nFourier coefficient, whereas
the periodgram of the noise is Op (1). However, this is an asymptotic result, for small samples sizes
you may not see such a big difference between deterministic mean and the noise. Next look at the
periodogram of the noise we see that it is very erratic (we will show later that this is because it is
an inconsistent estimator of the spectral density function), however, despite the erraticness, the
amount of variation overall frequencies seems to be same (there is just one large peak - which could
be explained by the randomness of the periodogram).
Returning again to Section 1.2.3, we now consider the case that the sin function has been

188

PS

0.0

0.2

0.4

2
0
2

signal

0.6

20

40

60

80

100

120

Time

P2

0.02
0.00

P1

0.0 0.1 0.2 0.3 0.4 0.5

0.04

frequency

frequency

frequency

Figure 8.1: Top Left: Realisation of (1.5) (2 sin( 2t


)) with iid noise, Top Right: Periodogram.
8
Bottom Left: Periodogram of just the noise. Bottom Right: Periodogram of the sin function.

corrupted by colored noise, which follows an AR(2) model


t = 1.5t1 0.75t2 + t .

(8.3)

A realisation and the corresponding periodograms are given in Figure 8.2. The results are different
to the iid case. The peak in the periodogram no longer corresponds to the period of the sin function.
From the periodogram of the just the AR(2) process we observe that it erratic, just as in the iid
case, however, there appears to be varying degrees of variation over the frequencies (though this
is not so obvious in this plot). We recall from Chapters 2 and 3, that the AR(2) process has
a pseudo-period, which means the periodogram of the colored noise will have pronounced peaks
which correspond to the frequencies around the pseudo-period. It is these pseudo-periods which
are dominating the periodogram, which is giving a peak at frequency that does not correspond to
the sin function. However, asymptotically the rates given in (8.2) still hold in this case too. In
other words, for large enough sample sizes the DFT of the signal should dominate the noise. To see
that this is the case, we increase the sample size to n = 1024, a realisation is given in Figure 8.3.
We see that the period corresponding the sin function dominates the periodogram. Studying the
periodogram of just the AR(2) noise we see that it is still erratic (despite the large sample size),

189

but we also observe that the variability clearly changes over frequency.

0.8

PS

5
0

0.0

5
0

20

40

60

80

100

120

0.4

signal2

1.2

Time

0.0

0.4

0.0 0.1 0.2 0.3 0.4 0.5

P2

0.8

P1

1.2

frequency

frequency

frequency

200

400

600

800

0.0 0.1 0.2 0.3 0.4 0.5

PS

0
10

signal3

10

Figure 8.2: Top Left: Realisation of (1.5) (2 sin( 2t


)) with AR(2) noise (n = 128), Top
8
Right: Periodogram. Bottom Left: Periodogram of just the AR(2) noise. Bottom Right:
Periodogram of the sin function.

1000

0.0 0.1 0.2 0.3 0.4 0.5

frequency

P2

0.00 0.05 0.10 0.15 0.20

P1

Time

frequency

frequency

Figure 8.3: Top Left: Realisation of (1.5) (2 sin( 2t


)) with AR(2) noise (n = 1024), Top
8
Right: Periodogram. Bottom Left: Periodogram of just the AR(2) noise. Bottom Right:
Periodogram of the sin function.
From now on we focus on the constant mean stationary time series (eg. iid noise and the AR(2))

190

(where the mean is either constant or zero). As we have observed above, the periodogram is the
absolute square of the discrete Fourier Transform (DFT), where
n

1 X

Xt exp(itk ).
2n t=1

Jn (k ) =

(8.4)

This is simply a (linear) transformation of the data, thus it easily reversible by taking the inverse
DFT

n
2 X
Jn (k ) exp(itk ).
Xt =
n t=1

(8.5)

Therefore, just as one often analyzes the log transform of data (which is also an invertible transform), one can analyze a time series through its DFT.
In Figure 8.4 we give plots of the periodogram of an iid sequence and AR(2) process defined
in equation (8.3). We recall from Chapter 3, that the periodogram is an inconsistent estimator
P
of the spectral density function f () = (2)1
r= c(r) exp(ir) and a plot of the spectral
density function corresponding to the iid and AR(2) process defined in (??). We will show later
that by inconsistent estimator we mean that E[|Jn (k )|2 ] = f (k ) + O(n1 ) but var[|Jn (k )|2 ] 9 0
as n . this explains why the general shape of |Jn (k )|2 looks like f (k ) but |Jn (k )|2 is

0.008

extremely erratic and variable.

0.25

0.15

P2

0.10

0.05

0.00

0.004
0.002

0.000

P1

0.20

0.006

frequency

frequency

Figure 8.4: Left: Periodogram of iid noise. Right: Periodogram of AR(2) process.

Remark 8.1.1 (Properties of the spectral density function) The spectral density function
was first introduced in in Section 1.6. We recall that given an autoregressive process {c(k)}, the

191

Autoregressive (2)

0.6

10

0.8

20

30

spectrum

1.0

spectrum

40

1.2

50

60

1.4

IID

0.0

0.1

0.2

0.3

0.4

0.5

0.0

frequency

0.1

0.2

0.3

0.4

0.5

frequency

Figure 8.5: Left: Spectral density of iid noise. Right: Spectral density of AR(2), note that
the interval [0, 1] corresponds to [0, 2] in Figure 8.5
spectral density is defined as
f () =

1 X
c(r) exp(2ir).
2 r=

And visa versa, given the spectral density we can recover the autocovariance via the inverse transR 2
form c(r) = 0 f () exp(2ir)d. We recall from Section 1.6 that the spectral density function
can be used to construct a valid autocovariance function since only a sequence whose Fourier transform is real and positive can be positive definite.
In Section 5.4 we used the spectral density function to define conditions under which the variance
covariance matrix of a stationary time series had minimum and maximim eigenvalues. Now from
the discussion above we observe that the variance of the DFT is approximately the spectral density
function (note that for this reason the spectral density is sometimes called the power spectrum).
We now collect some of the above observations, to summarize some of the basic properties of
the DFT:
(i) We note that Jn (k ) = Jn (nk ), therefore, all the information on the time series is contain
in the first n/2 frequencies {Jn (k ); k = 1, . . . , n/2}.
(ii) If the time series E[Xt ] = and k 6= 0 then
n

1 X
E[Jn (k )] =
exp(itk ) = 0.
n t=1

192

In other words, the mean of the DFT is zero regardless of whether the time series has a zero
mean (it just needs to have a constant mean).
(iii) However, unlike the original stationary time series, we observe that the variance of the DFT
depends on frequency (unless it is a white noise process) and that for k 6= 0, var[Jn (k )] =
E[|Jn (k )|2 ] = f (k ) + O(n1 ).
The focus of this chapter will be on properties of the spectral density function (proving some
of the results we stated previously) and on the so called Cramer representation (or spectral representation) of a second order stationary time series. However, before we go into these results (and
proofs) we give one final reason why the analysis of a time series is frequently done by transforming
to the frequency domain via the DFT. Above we showed that there is a one-to-one correspondence
between the DFT and the original time series, below we show that the DFT almost decorrelates
the stationary time series. In other words, one of the main advantages of working within the
frequency domain is that we have transformed a correlated time series into something that it almost uncorrelated (this also happens to be a heuristic reason behind the spectral representation
theorem).

8.2

The near uncorrelatedness of the Discrete Fourier


Transform
1/2

Let X n = {Xt ; t = 1, . . . , n} and n = var[X n ]. It is clear that n

X n is an uncorrelated

1/2

sequence. This means to formally decorrelate X n we need to know n

. However, if Xt is a

second order stationary time series, something curiously, remarkable happens. The DFT, almost
uncorrelates the X n . The implication of this is extremely useful in time series, and we shall be
using this transform in estimation in Chapter 9.
We start by defining the Fourier transform of {Xt }nt=1 as
n

Jn (k ) =

1 X
2t
Xt exp(ik
)
n
2n t=1

where the frequences k = 2k/n are often called the fundamental, Fourier frequencies.

193

Lemma 8.2.1 Suppose {Xt } is a second order stationary time series, where

|rc(r)| < . Then

we have

f ( 2k ) + O( 1 ) k1 = k2
2k1
2k2
n
n
cov(Jn (
), Jn (
)) =

1
n
n
O( n )
k1 =
6 k2
where f () =

1
2

r= c(r) exp(ir).

In the sections below we give two proofs for the same result.
We note that the principle reason behind both proofs is that



0 j 6= nZ
2j
exp it
=
.
n jZ
n
t=1

n
X

8.2.1

(8.6)

Seeing the decorrelation in practice

We evaluate the DFT using the following piece of code (note that we do not standardize by

2)

dft <- function(x){


n=length(x)
dft <- fft(x)/sqrt(n)
return(dft)
}
We have shown above that {Jn (k )}k are close to uncorrelated and have variance close to f (k ).
This means that the ratio Jn (k )/f (k )1/2 are close to uncorrelated with variance close to one. Let
us treat
Zk =

Jn (k )
,
f (k )1/2

as the transformed random variables, noting that {Zk } is complex, our aim is to show that the acf
corresponding to {Zk } is close to zero. Of course, in practice we do not know the spectral density
function f , therefore we estimate it using the piece of code (where test is the time series)
k<-kernel("daniell",6)
temp2 <-spec.pgram(test,k, taper=0, log = "no")$spec
n <- length(temp2)

194

temp3 <- c(temp2[c(1:n)],temp2[c(n:1)])


temp3 simply takes a local average of the periodogram about the frequency of interest (however
it is worth noting that spec.pgram does not do precisely this, which can be a bit annoying). In
Section 9.3 we explain why this is a consistent estimator of the spectral density function. Notice
that we also double the length, because the estimator temp2 only gives estimates in the interval
bk = Jn (k )/fbn (k )1/2 is
[0, ]. Thus our estimate of {Zk }, which we denote as Z
temp1 <- dft(test);

temp4 <- temp1/sqrt(temp3)

bk } over various lags


We want to evaluate the covariance of {Z
n

1Xb b
1 X Jn (k )Jn (k+r )
q
Cn (r) =
Zk Z k =
n
n
k=1
k=1
fbn (k )fbn (k+r )
to do this we use we exploit the speed of the FFT (Fast Fourier Transform)
temp5 <- Mod(dft(temp4))**2
dftcov <- fft(temp5, inverse = TRUE)/(length(temp5))
dftcov1 = dftcov[-1]
and make an ACF plot of the real and imaginary parts of the Fourier ACF:
n = length(temp5)
par(mfrow=c(2,1))
plot(sqrt(n)*Re(dftcov1[1:30]))
lines(c(1,30),c(1.96,1.96))
lines(c(1,30),c(-1.96,-1.96))
plot(sqrt(n)*Im(dftcov1[1:30]))
lines(c(1,30),c(1.96,1.96))
lines(c(1,30),c(-1.96,-1.96))
Note that the 1.96 corresponds to the 2.5% limits, however this bound only holds if the time series
is Gaussian. If it non-Gaussian some corrections have to be made (see Dwivedi and Subba Rao
(2011) and Jentsch and Subba Rao (2014)). A plot of the AR(2) model
t = 1.5t1 0.75t2 + t .

195

together with the real and imaginary parts of its DFT autocovariance is given in Figure 8.6. We

test

observe that most of the correlations lie between [1.96, 1.96]

50

100

150

200

250

Time

Real

Imaginary

sqrt(n) * Im(dftcov1[1:30])

sqrt(n) * Re(dftcov1[1:30])

10

15

20

25

30

Index

10

15

20

25

30

Index

Figure 8.6: Top: Realization. Bottom: Real and Imaginary of Cn (r) plotted against the lag
r.

Exercise 8.1

(a) Simulate an AR(2) process and run the above code using the sample size

(i) n = 64 (however use k<-kernel("daniell",3))


(ii) n = 128 (however use k<-kernel("daniell",4))
Does the near decorrelation property hold when the sample size is very small. Explain your
answer by looking at the proof of the lemma.
(b) Simulate a piecewise stationary time series (this is a simple example of a nonstationary time
series) by stringing two stationary time series together. One example is

196

ts1 = arima.sim(list(order=c(2,0,0), ar = c(1.5, -0.75)), n=128);


ts2 = arima.sim(list(order=c(1,0,0), ar = c(0.7)), n=128)
test = c(ts1/sd(ts1),ts2/sd(ts2))
Make a plot of this time series. Calculate the DFT covariance of this time series, what do
you observe in comparison to the stationary case?

8.2.2

Proof 1 of Lemma 8.2.1: By approximating Toeplitz with


Circulant matrices
(s1)(t1)

Let X 0n = (Xn , . . . , X1 ) and Fn be the Fourier transformation matrix (Fn )s,t = n1/2 n

0
n1/2 exp( 2i(s1)(t1)
) (note that n = exp( 2
n
n )). It is clear that Fn X n = (Jn (0 ), . . . , Jn (n1 )) .

We now prove that Fn X n is almost an uncorrelated sequence.


The first proof will be based on approximating the symmetric Toeplitz variance matrix of X n
with a circulant matrix, which has well know eigen values and functions. We start by considering
the variance of Fn X n , var(Fn X n ) = Fn n F n , and our aim is to show that it is almost a diagonal.
We first recall that if n were a circulant matrix, then Fn X n would be uncorrelated since Fn is the
eigenmatrix of any circulant matrix. This is not the case. However, the upper right hand side and
the lower left hand side of n can approximated by circulant matrices - this is the trick in showing
the near uncorrelatedness. Studying n

n =

c(n 1)

c(1) . . . c(n 2)
..
..
..
.
.
.
..
c(n 1) c(n 2)
.
c(1)
c(0)

c(0)
c(1)
..
.

c(1)

c(2)

...

c(0)
..
.

we observe that it can be written as the sum of two circulant matrices, plus some error, that we
will bound. That is, we define the two circulant matrices

C1n

c(n 1)

...
c(n 2)
c(n 1) c(0) c(1)
=
..
..
.
..

..
..

.
.
.
.

..
c(1)
c(2)
.
c(n 1)
c(0)

c(0)

c(1) c(2)

197

...

and

c(n 1) c(n 2)

C2n

...

c(1)

c(1)
0
c(n 1) . . . c(2)

c(2)
c(1)
0
. . . c(3)

..
..
..
..
..
.
.
.
.
.

..
c(n 1) c(n 2)
.
c(1)
0

We observe that the upper right hand sides of C1n and n match and the lower left and sides of C2n
0

and n match. As the above are circulant their eigenvector matrix is Fn (note that Fn1 = F n ).
Furthermore, the eigenvalues matrix of Cn1 is

diag

n1
X

c(j),

j=0

n1
X

c(j)jn , . . . ,

j=0

n1
X

,
c(j)(t1)j
n

j=0

whereas the eigenvalue matrix of Cn2 is

diag

= diag

n1
X

c(j),

n1
X

j=1

j=1

n1
X

n1
X

j=1

c(j),

c(n j)jn , . . . ,

n1
X

c(n j)(t1)j
n

j=1

c(j)j
n ,...,

j=1

n1
X

c(j)n(t1)j ,

j=1

Pn1
j(k1)
and k2 =
More succinctly, the kth eigenvalues of Cn1 and Cn2 are k1 =
j=0 c(j)n
P
Pn1
2j
j(k1)
. Observe that k1 + k2 = |j|(n1) c(j)e n f (j ), thus the sum of these
j=1 c(j)n
eigenvalues approximate the spectral density function.
P
We now show that under the condition r |rc(r)| < we have
0
Fn n F n

Fn Cn1 +

0
Cn2 F n

 
1
=O
I,
n

(8.7)

where I is a n n matrix of ones. To show the above we consider the differences element by
element. Since the upper right hand sides of Cn1 and n match and the lower left and sides of Cn2

198

and n match, the above difference is






 

Fn n F 0n Fn Cn1 + Cn2 F n

(s,t)
X
2 n1

1
0
0
0

= es n et es Cn1 et es Cn2 et
|rc(r)| = O( ).
n
n
r=1

Thus we have shown (8.7). Therefore, since Fn is the eigenvector matrix of Cn1 and Cn2 , altogether
we have
Fn Cn1 + Cn2
where fn () =



2
2(n 1)
F n = diag fn (0), fn ( ), . . . , fn (
) ,
n
n

Pn1

r=(n1) c(r) exp(ij).

fn (0)

var(Fn X n ) = Fn n F n =

0
..
.

...

0
fn ( 2
n ) ...
..
..
. ...
.

0
Finally, we note that since

Altogether this gives

...

...

0
..
.
fn ( 2(n1)
))
n

1
+ O( )

...

X
|r|>n

|c(r)|

1 X
|rc(r)| = O(n1 ),
n

(8.8)

|r|>n

which gives the required result.


Remark 8.2.1 Note the eigenvalues of a matrix are often called the spectrum and that above
calculation shows that spectrum of var[X n ] is close to f (n ), which may be a one reason why f ()
is called the spectral density (the reason for density probably comes from the fact that f is positive).
These ideas can also be used for inverting Toeplitz matrices (see Chen et al. (2006)).

8.2.3

Proof 2 of Lemma 8.2.1: Using brute force

2k2
1
A more hands on proof is to just calculate cov(Jn ( 2k
n ), Jn ( n )). The important aspect of this

proof is that if we can isolate the exponentials than we can use (8.6). It is this that gives rise to the
near uncorrelatedness property. Remember also that exp(i 2
n jk) = exp(ijk ) = exp(ikj ), hence

199

1 1 ... 1 1

.. .
.. . . . .
. ... .
.
.

1 ... ... 1 1

|rc(r)| <

|fn () f ()|

we can interchange between the two notations.


We note that cov(A, B) = E(AB) E(A)E(B), thus we have




n
2k1
2
2k2
1 X
cov Jn (
cov(Xt , X ) exp i(tk1 k2 )
), Jn (
) =
n
n
n
n
t, =1

Now change variables with r = t , this gives (for 0 k1 , k2 < n)




2k1
2k2
cov Jn (
), Jn (
)
n
n
 n|r|



n1
2k2 X
2it(k1 k2 )
1 X
c(r) exp ir
exp
)
=
n
n
n
t=1

r=(n1)





n
n1
X
2it(k1 k2 )
2k2 1 X
exp
+Rn ,
=
c(r) exp ir
n
n
n
r=(n1)
| t=1
{z
}
k1 (k2 )

where
1
Rn =
n
Thus |Rn |

1
n

n1
X
r=(n1)

|r|n |rc(r)|



2k2
c(r) exp ir
n

n
X


exp

t=n|r|+1

= O(n1 ) Finally by using (8.8) we obtain the result.

Exercise 8.2 The the above proof (in Section 8.2.3) uses that
P
we obtain if we relax this assumption to r |c(r)| < ?

8.2.4


2it(k1 k2 )
)
n

|rc(r)| < . What bounds do

Heuristics

In this section we summarize some spectral properties. We do this by considering the DFT of the
data {Jn (k )}nk=1 . It is worth noting that to calculate {Jn (k )}nk=1 is computationally very fast and
requires only O(n log n) computing operations (see Section A.5, where the Fast Fourier Transform
is described).

200

The spectral (Cramers) representation theorem


We observe that for any sequence {Xt }nt=1 that it can be written as the inverse transform for
1tn
n

1 X
Xt =
Jn (k ) exp(itk ),
n

(8.9)

k=1

which can be written as an integral


Xt =

n
X

1
n

exp(it)dZn (),

(8.10)

k=2

where Zn () =

exp(itk ) [Zn (k ) Zn (k1 )] =

Pb 2
nc

k=1

Jn (k ).

The second order stationary property of Xt means that the DFT Jn (k ) is close to an uncorrelated sequence or equivalently the process Zn () has near orthogonal increments, meaning that for
any two non-intersecting intervals [1 , 2 ] and [3 , 4 ] that Zn (2 ) Zn (1 ) and Zn (4 ) Zn (3 ).
The spectral representation theorem generalizes this result, it states that for any second order
stationary time series {Xt } there exists an a process {Z(); [0, 2]} where for all t Z
Z
Xt =

exp(it)dZ()

(8.11)

and Z() has orthogonal increments, meaning that for any two non-intersecting intervals [1 , 2 ]
and [3 , 4 ] E[Z(2 ) Z(1 )][Z(2 ) Z(1 )] = 0.
We now explore the relationship between the DFT with the orthogonal increment process.
Using (8.11) we see that

Jn (k ) =
=

!
Z 2 X
n
n
1 X
1

Xt exp(itk ) =
exp(it[k ]) dZ()
2n t=1
2n 0
t=1
Z 2 

1

ei(n+1)(k 0 )/2 Dn/2 (k ) dZ(),


2n 0

where Dn/2 (x) = sin[((n + 1)/2)x]/ sin(x/2) is the Dirichlet kernel (see Priestley (1983), page
419). We recall that the Dirichlet kernel limits to the Dirac-delta function, therefore very crudely
speaking we observe that the DFT is an approximation of the orthogonal increment localized about
k (though mathematically this is not strictly correct).

201

Bochners theorem
This is a closely related result that is stated in terms of the so called spectral distribution. First
the heuristics. We see that from Lemma 8.2.1 that the DFT Jn (k ), is close to uncorrelated. Using
this and inverse Fourier transforms we see that for 1 t, n we have
c(t ) = cov(Xt , X ) =

Let Fn () =

1
n

Pb 2
nc

k=1

n
n
1 X X
cov (Jn (k1 ), Jn (k2 )) exp(itk1 + i k2 )
n

1
n

k1 =1 k2 =1
n
X

var(Jn (k )) exp(i(t )k ).

(8.12)

k=1

var[Jn (k )], then the above can be written as


2

exp(i(t ))dFn (),

c(t )
0

where we observe that Fn () is a positive function which in non-decreasing over . Bochners


theorem is an extension of this is states that for any autocovariance function {c(k)} we have the
representation
Z

c(t ) =

Z
exp(i(t ))f ()d =

exp(i(t ))dF ().


0

where F () is a positive non-decreasing bounded function. Moreover, F () = E(|Z()|2 ). We note


P
that if the spectral density function exists (which is only true if r |c(r)|2 < ) then F () =
R
0 f ()d.
Remark 8.2.2 The above results hold for both linear and nonlinear time series, however, in the
case that Xt has a linear representation
Xt =

j tj ,

j=

then Xt has the particular form


Z
Xt =
where A() =

j= j

A() exp(ik)dZ(),

(8.13)

exp(ij) and Z() is an orthogonal increment process, but in addition

202

E(|dZ()|2 ) = d ie. the variance of increments do not vary over frequency (as this varying has
been absorbed by A(), since F () = |A()|2 ).
We mention that a more detailed discussion on spectral analysis in time series is give in Priestley
(1983), Chapters 4 and 6, Brockwell and Davis (1998), Chapters 4 and 10, Fuller (1995), Chapter
3, Shumway and Stoffer (2006), Chapter 4. In many of these references they also discuss tests for
periodicity etc (see also Quinn and Hannan (2001) for estimation of frequencies etc.).

8.3
8.3.1

The spectral density and spectral distribution


The spectral density and some of its properties

We start by showing that under certain strong conditions the spectral density function is nonnegative. We later weaken these conditions (and this is often called Bochners theorem).
Theorem 8.3.1 (Positiveness of the spectral density) Suppose the coefficients {c(k)} are abP
solutely summable (that is k |c(k)| < ). Then the sequence {c(k)} is positive semi-definite if an
only if the function f (), where

1 X
f () =
c(k) exp(ik)
2
k=

is nonnegative. Moreover
Z

c(k) =

exp(ik)f ()d.

(8.14)

It is worth noting that f is called the spectral density corresponding to the covariances {c(k)}.
PROOF. We first show that if {c(k)} is a non-negative definite sequence, then f () is a nonnegative
function. We recall that since {c(k)} is non-negative then for any sequence x = (x1 , . . . , xN ) (real
P
or complex) we have ns,t=1 xs c(s t)
xs 0 (where x
s is the complex conjugate of xs ). Now we
consider the above for the particular case x = (exp(i), . . . , exp(in)). Define the function

fn () =

n
1 X
exp(is)c(s t) exp(it).
2n
s,t=1

203

Thus by definition fn () 0. We note that fn () can be rewritten as

fn () =

1
2

Comparing f () =

1
2

(n1)

X
k=(n1)

k= c(k) exp(ik)


f () fn ()

n |k|
n


c(k) exp(ik).

with fn () we see that


1 X
1
c(k) exp(ik) +
2
2
|k|n

(n1)

X
k=(n1)


|k|
c(k) exp(ik)
n

:= In + IIn .
Since

k= |c(k)|

< it is clear that In 0 as n . Using Lemma A.1.1 we have IIn 0

as n . Altogether the above implies




f () fn () 0

as n .

(8.15)

Now it is clear that since for all n, fn () are nonnegative functions, the limit f must be nonnegative
(if we suppose the contrary, then there must exist a sequence of functions {fnk ()} which are not
necessarily nonnegative, which is not true). Therefore we have shown that if {c(k)} is a nonnegative
definite sequence, then f () is a nonnegative function.

1
2

We now show the converse, that is the Fourier coefficients of any non-negative `2 function f () =
R 2
P
k= c(k) exp(ik), is a positive semi-definite sequence. Writing c(k) = 0 f () exp(ik)d

we substitute this into Definition 1.6.1 to give


n
X
s,t=1

Z
xs c(s t)
xs =

f ()
0

n
X

xs exp(i(s t))
xs d =
0

s,t=1

Hence we obtain the desired result.


2
n
X



xs exp(is) d 0.
f ()


s=1

The above theorem is very useful. It basically gives a simple way to check whether a sequence
{c(k)} is non-negative definite or not (hence whether it is a covariance function - recall Theorem 1.6.1). See Brockwell and Davis (1998), Corollary 4.3.2 or Fuller (1995), Theorem 3.1.9, for
alternative explanations.
Example 8.3.1 Consider the empirical covariances (here we gives an alternative proof to Remark

204

6.2.1) defined in Chapter 6

cn (k) =

1
n

Pn|k|
t=1

Xt Xt+|k| |k| n 1

otherwise

we give an alternative proof to Lemma 6.2.1 to show that {


cn (k)} is non-negative definite sequence.
To show that the sequence we take the Fourier transform of cn (k) and use Theorem 8.3.1. The
Fourier transform of {
cn (k)} is
(n1)

(n1)

exp(ik)
cn (k) =

k=(n1)

exp(ik)

n|k|
n

1 X
1 X
Xt Xt+|k| =
Xt exp(it) 0.
n
n
t=1

k=(n1)

t=1

Since the above is non-negative, this means that {


cn (k)} is a non-negative definite sequence.
We now state a useful result which relates the largest and smallest eigenvalue of the variance
of a stationary process to the smallest and largest values of the spectral density (we recall we used
this in Lemma 5.4.1).
Lemma 8.3.1 Suppose that {Xk } is a stationary process with covariance function {c(k)} and
spectral density f (). Let n = var(X n ), where X n = (X1 , . . . , Xn ). Suppose inf f () m > 0
and sup f () M < Then for all n we have
min (n ) inf f ()

and

max (n ) sup f ().

PROOF. Let e1 be the eigenvector with smallest eigenvalue 1 corresponding to n . Then using
R
c(s t) = f () exp(i(s t))d we have
min (n ) = e01 n e1 =

n
X

Z
es,1 c(s t)et,1 =

f ()

s,t=1

n
X

es,1 exp(i(s t))et,1 d =

s,t=1

2
2

Z 2 X
n
n

X

2




e
exp(is)
es,1 exp(is) d inf f ()
f ()

d = inf f (),
s,1





0
0

Z
=

s=1

s=1

since by definition

R Pn
P
| s=1 es,1 exp(is)|2 d = ns=1 |es,1 |2 = 1 (using Parsevals identity). Using

a similar method we can show that max (n ) sup f ().

We now state a version of the above result which requires weaker conditions on the autocovariance function (only that they decay to zero).

205

Lemma 8.3.2 Suppose the covariance {c(k)} decays to zero as k , then for all n, n =
var(X n ) is a non-singular matrix (Note we do not require the stronger condition the covariances
are absolutely summable).
PROOF. See Brockwell and Davis (1998), Proposition 5.1.1.

8.3.2

The spectral distribution and Bochners theorem

Theorem 8.3.1 only holds when the sequence {c(k)} is absolutely summable. Of course this may not
always be the case. An extreme example is the time series Xt = Z. Clearly this is a stationary time
series and its covariance is c(k) = var(Z) = 1 for all k. In this case the autocovariances {c(k) = 1},
is not absolutely summable, hence the representation of the covariance in Theorem 8.3.1 does not
apply in this case. The reason is because the Fourier transform of the infinite sequence {c(k) = 1}k
is not well defined (clearly {c(k) = 1}k does not belong to `2 ).
However, we now show that Theorem 8.3.1 can be generalised to include all non-negative definite
sequences and stationary processes, by considering the spectral distribution rather than the spectral
density.
Theorem 8.3.2 A function {c(k)} is non-negative definite sequence if and only if
Z
c(k) =

exp(ik)dF (),

(8.16)

where F () is a right-continuous (this means that F (x + h) F (x) as 0 < h 0), non-decreasing,


non-negative bounded function on [, ] (hence it has all the properties of a distribution and it
can be consider as a distribution - it is usually called the spectral distribution). This representation
is unique.
This is a very constructive result. It shows that the Fourier coefficients of any distribution function
form a non-negative definite sequence, and thus, if c(k) = c(k) (hence is symmetric) correspond
to the covarance function of a random process. In Figure 8.7 we give two distribution functions.
the top plot is continuous and smooth, therefore its derivative will exist, be positive and belong
to `2 . So it is clear that its Fourier coefficients form a non-negative definite sequence. The interesting aspect of Thereom 8.3.2 is that the Fourier coefficients corresponding to the distribution
function in the second plot also forms a non-negative definite sequence even though the derivative

206

of this distribution function does not exist. However, this sequence will not belong to `2 (ie. the
correlations function will not decay to zero as the lag grows).

Figure 8.7: Both plots are of non-decreasing functions, hence are valid distribution functions.
The top plot is continuous and smooth, thus its derivative (the spectral density function)
exists. Whereas the bottom plot is not (spectral density does not exist).
PROOF of Theorem 8.3.2. We first show that if {c(k)} is non-negative definite sequence,
R 2
then we can write c(k) = 0 exp(ik)dF (), where F () is a distribution function.
To prove the result we adapt some of the ideas used to prove Theorem 8.3.1. As in the proof
of Theorem 8.3.1 define the (nonnegative) function
n
1 X
1
exp(is)c(s t) exp(it) =
fn () = var[Jn ()] =
2n
2
s,t=1

(n1)

X
k=(n1)

n |k|
n


c(k) exp(ik).

If {c(k)} is not absolutely summable, the limit of fn () is no longer be well defined. Instead we consider its integral, which will always be a distribution function (in the sense that it is nondecreasing
and bounded). Let us define the function Fn () whose derivative is fn (), that is
Z
Fn () =
0

n1
2 X
r

sin(r)
fn ()d =
c(0) +
1
c(r)
2
2
n
r
r=1

207

0 2.

Since fn () is nonnegative, Fn () is a nondecreasing function. Furthermore it is bounded since


Z

fn ()d = c(0).

Fn (2) =
0

Hence Fn satisfies all properties of a distribution and can be treated as a distribution function. This
means that we can use Hellys theorem which states that for any sequence of distributions {Gn }
defined on [0, 2], were Gn (0) = 0 and supn Gn (2) < M < , there exist a subsequence {nm }m
where Gnm (x) G(x) as m for each x [0, 2] at which G is continuous. Furthermore, since
Gnm (x) G(x) (pointwise as m ), this implies that for any bounded sequence h we have that
Z

Z
h(x)dGnm (x)

h(x)dG(x)

as m

(a very nice proof is given in Varadhan, Theorem 4.1).


We now apply this result to Fn . Using Hellys theorem there exists a subsequence of distributions
{Fnm }m which has a limit F , thus
Z

Z
h(x)dFnm (x)

h(x)dF (x)

as m .

We focus on the function h(x) = exp(ik). It is clear that for every k and n we have
Z

Z
exp(ik)dFn () =
0

(1 |k| )c(k) |k| n


n
exp(ik)fn ()d =

0
|k| n

(8.17)

Fixing k and letting n we see that


Z
dn,k =
0



|k|
exp(ik)dFn () = 1
c(k)
n

is a Cauchy sequence, where


dn,k dk = c(k)

(8.18)

as n . Thus
Z
dnm ,k =

Z
exp(ikx)dFnm (x)

exp(ikx)dF (x)

208

as m

and by (8.18) we have


Z
c(k) =

exp(ikx)dF (x),

where F (x) is a well defined distribution. This gives the first part of the assertion.
To show the converse, that is {c(k)} is a non-negative definite sequence when c(k) is defined as
R
c(k) = exp(ik)dF (), we use the same method given in the proof of Theorem 8.3.1, that is
n
X

Z
xs c(s t)
xs =
0

s,t=1

n
X


xs exp(i(s t))
xs dF ()

s,t=1


2
n

2 X


x
exp(is)

dF () 0,
s


0

Z
=

s=1

since F () is a distribution.

R 2
0

Finally, if {c(k)} were absolutely summable, then we can use Theorem 8.3.1 to write c(k) =
R
1 P
exp(ik)dF (), where F () = 0 f ()d and f () = 2
k= c(k) exp(ik). By using

Theorem 8.3.1 we know that f () is nonnegative, hence F () is a distribution, and we have the
result.

Example 8.3.2 Using the above we can construct the spectral distribution for the (rather silly)
time series Xt = Z. Let F () = 0 for < 0 and F () = var(Z) for 0 (hence F is the step
function). Then we have
Z
cov(Xt , Xt+k ) = var(Z) =

exp(ik)dF ().

Example 8.3.3 Consider the second order stationary time series


Xt = U1 cos(t) + U2 sin(t),
where U1 and U2 are iid random variables with mean zero and variance 2 and the frequency. It
can be shown that
cov(Xt , Xt+k ) =

2
[exp(ik) + exp(ik)] .
2

209

Observe that this covariance does not decay with the lag k. Then
Z

exp(ik)dF ().

cov(Xt , Xt+k ) = var(Z) =


0

where

F () =

8.4

<

2 /2 <

2
.

The spectral representation theorem

We now state the spectral representation theorem and give a rough outline of the proof.
Theorem 8.4.1 If {Xt } is a second order stationary time series with mean zero, and spectral distribution F (), and the spectral distribution function is F (), then there exists a right continuous,
orthogonal increment process {Z()} (that is E[(Z(1 ) Z(2 )(Z(3 ) Z(4 ))] = 0, when the
intervals [1 , 2 ] and [3 , 4 ] do not overlap) such that
Z
Xt =

exp(it)dZ(),

(8.19)

where for 1 2 , E|Z(1 ) Z(2 )|2 = F (1 ) F (2 ) (noting that F (0) = 0). (One example
of a right continuous, orthogonal increment process is Brownian motion, though this is just one
example, and usually Z() will be far more general than Brownian motion).
Heuristically we see that (8.19) is the decomposition of Xt in terms of frequencies, whose
amplitudes are orthogonal. In other words Xt is decomposed in terms of frequencies exp(it)
which have the orthogonal amplitudes dZ() (Z( + ) Z()).
Remark 8.4.1 Note that so far we have not defined the integral on the right hand side of (8.19). It
is known as a stochastic integral. Unlike many deterministic functions (functions whose derivative
exists), one cannot really suppose dZ() Z 0 ()d, because usually a typical realisation of Z()
will not be smooth enough to differentiate. For example, it is well known that Brownian is quite
rough, that is a typical realisation of Brownian motion satisfies |B(t1 ,
) B(t2 ,
)| K(
)|t1
tt | , where
is a realisation and 1/2, but in general will not be larger. The integral

210

g()dZ() is well defined if it is defined as the limit (in the mean squared sense) of discrete
Pbn/2c
P
[Z(k ) Z(k1 )], then
sums. More precisely, let Zn () = nk=1 Z(k )Ink 1 ,nk () = k=1
Z
g()dZn () =

n
X

g(k ){Z(k ) Z(k1 )}.

k=1

R
R
R
The limit of g()dZn () as n is g()dZ() (in the mean squared sense, that is E[ g()dZ()
R
g()dZn ()]2 ). Compare this with our heuristics in equation (8.10).
For a more precise explanation, see Parzen (1959), Priestley (1983), Sections 3.6.3 and Section
4.11, page 254, and Brockwell and Davis (1998), Section 4.7. For a very good review of elementary
stochastic calculus see Mikosch (1999).
A very elegant explanation on the different proofs of the spectral representation theorem is given
in Priestley (1983), Section 4.11. We now give a rough outline of the proof using the functional
theory approach.
Rough PROOF of the Spectral Representation Theorem To prove the result we first
define two Hilbert spaces H1 and H2 , where H1 one contains deterministic functions and H2 contains
random variables.
First we define the space
H1 = sp{eit ; t Z}
with inner-product
Z
hf, gi =

f (x)g(x)dF (x)

(8.20)

(and of course distance hf g, f gi =

R 2
0

|f (x) g(x)|2 dF (x)) it is clear that this inner product

is well defined because hf, f i 0 (since F is a measure). It can be shown (see Brockwell and Davis
n R
o
2
(1998), page 144) that H1 = g; 0 |g()|2 dF () < 1 . We also define the space
H2 = sp{Xt ; t Z}
1

Roughly speaking it is because all continuous functions on [0, 2] are dense in L2 ([0, 2], B, F ) (using
the metric kf gk = hf g, f gi and the limit of Cauchy sequences). Since all continuous function can be
written as linear combinations of the Fourier basis, this gives the result.

211

with inner-product cov(X, Y ) = E[XY ] E[X]E[Y ].


Now let us define the linear mapping T : H1 H2
n
n
X
X
T(
aj exp(ik)) =
aj Xk ,
j=1

(8.21)

j=1

for any n (it is necessary to show that this can be extended to infinite n, but we wont do so here).
We will shown that T defines an isomorphism (ie. it is a one-to-one linear mapping that preserves
norm). To show that it is a one-to-one mapping see Brockwell and Davis (1998), Section 4.7. It is
clear that it is linear, there all that remains is to show that the mapping preserves inner-product.
P
Suppose f, g H1 , then there exists coefficients {fj } and {gj } such that f (x) = j fj exp(ij)
P
and g(x) = j gj exp(ij). Hence by definition of T in (8.21) we have
hT f, T gi = cov(

X
j

fj Xj ,

gj Xj ) =

fj1 gj2 cov(Xj1 , Xj2 )

(8.22)

j1 ,j2

Now by using Bochners theorem (see Theorem 8.3.2) we have


Z
hT f, T gi =
0


fj1 gj2 exp(i(j1 j2 )) dF () =

f (x)g(x)dF (x) = hf, gi.


0

j1 ,j2

(8.23)
Hence < T f, T g >=< f, g >, so the inner product is preserved (hence T is an isometry).
Altogether this means that T defines an isomorphism betwen H1 and H2 . Therefore all functions
which are in H1 have a corresponding random variable in H2 which has similar properties.
For all [0, 2], it is clear that the identity functions I[0,] (x) H1 . Thus we define the
random function {Z(); 0 2}, where T (I[0,] ()) = Z() H2 (since T is an isomorphism).
Since that mapping T is linear we observe that
T (I[1 ,2 ] ) = T (I[0,1 ] I[0,2 ] ) = T (I[0,1 ] ) T (I[0,2 ] ) = Z(1 ) Z(2 ).
Moreover, since T preserves the norm for any non-intersecting intervals [1 , 2 ] and [3 , 4 ] we have
cov ((Z(1 ) Z(2 ), (Z(3 ) Z(4 )) = hT (I[1 ,2 ] ), T (I[3 ,4 ] )i = hI[1 ,2 ] , I[3 ,4 ] i
Z
=
I[1 ,2 ] ()I[3 ,4 ] ()dF () = 0.

212

Therefore by construction {Z(); 0 2} is an orthogonal increment process, where


E|Z(2 ) Z(1 )|2 = < T (I[1 ,2 ] ), T (I[1 ,2 ] ) >=< I[1 ,2 ] , I[1 ,2 ] >
Z 2
Z 2
dF () = F (2 ) F (1 ).
I[1 ,2 ] dF () =
=
1

Having defined the two spaces which are isomorphic and the random function {Z(); 0
2} and function I[0,] (x) which have orthogonal increments, we can now prove the result. Since
dI[0,] (s) = (s)ds, where (s) is the dirac delta function, any function g L2 [0, 2] can be
represented as
2

g(s)dI[,2] (s).

g() =
0

Thus for g() = exp(it) we have


2

Z
exp(it) =

exp(its)dI[,2] (s).
0

Therefore
2

Z
T (exp(it)) = T
Z
=

 Z
exp(its)dI[,2] (s) =

exp(its)T [dI[,2] (s)]

exp(its)dT [I[,2] (s)],

where the mapping goes inside the integral due to the linearity of the isomorphism. Using that
I[,2] (s) = I[0,s] () we have
2

Z
T (exp(it)) =

exp(its)dT [I[0,s] ()].


0

By definition we have T (I[0,s] ()) = Z(s) which we substitute into the above to give
Z
Xt =

exp(its)dZ(s),
0

which gives the required result.


Note that there are several different ways to prove this result.

213

It is worth taking a step back from the proof and see where the assumption of stationarity crept
in. By Bochners theorem we have that
Z
c(t ) =

exp(i(t ))dF (),

where F is a distribution. We use F to define the space H1 , the mapping T (through {exp(ik)}k ),
the inner-product and thus the isomorphism. However, it was the construction of the orthogonal
random functions {Z()} that was instrumental. The main idea of the proof was that there are
functions {k ()} and a distribution H such that all the covariances of the stochastic process {Xt }
can be written as
2

t () ()dH(),

E(Xt X ) = c(t, ) =
0

where H is a measure. As long as the above representation exists, then we can define two spaces
H1 and H2 where {k } is the basis of the functional space H1 and it contains all functions f such
R
that |f ()|2 dH() < and H2 is the random space defined by sp(Xt ; t Z). From here we can
P
define an isomorphism T : H1 H2 , where for all functions f () = k fk k () H1
T (f ) =

fk Xk H2 .

An important example is T (k ) = Xk . Now by using the same arguments as those in the proof
above we have
Z
Xt =

t ()dZ()

where {Z()} are orthogonal random functions and E|Z()|2 = H(). We state this result in the
theorem below (see Priestley (1983), Section 4.11).
Theorem 8.4.2 (General orthogonal expansions) Let {Xt } be a time series (not necessarily
second order stationary) with covariance {E(Xt X ) = c(t, s)}. If there exists a sequence of functions
{k ()} which satisfy for all k
Z

|k ()|2 dH() <

214

and the covariance admits the representation


2

t ()s ()dH(),

c(t, s) =

(8.24)

where H is a distribution then for all t we have the representation


Z
Xt =

t ()dZ()

(8.25)

where {Z()} are orthogonal random functions and E|Z()|2 = H(). On the other hand if Xt
has the representation (8.25), then c(s, t) admits the representation (8.24).
Remark 8.4.2 We mention that the above representation applies to both stationary and nonstationary time series. What makes the exponential functions {exp(ik)} special is if a process is
stationary then the representation of c(k) := cov(Xt , Xt+k ) in terms of exponentials is guaranteed:
Z
c(k) =

exp(ik)dF ().

(8.26)

Therefore there always exists an orthogonal random function {Z()} such that
Z
Xt =

exp(it)dZ().

Indeed, whenever the exponential basis is used in the definition of either the covariance or the
process {Xt }, the resulting process will always be second order stationary.
We mention that it is not always guaranteed that for any basis {t } we can represent the
covariance {c(k)} as (8.24). However (8.25) is a very useful starting point for characterising
nonstationary processes.

215

8.5

The spectral density functions of MA, AR and


ARMA models

We obtain the spectral density function for MA() processes. Using this we can easily obtain the
spectral density for ARMA processes. Let us suppose that {Xt } satisfies the representation

Xt =

j tj

(8.27)

j=

where {t } are iid random variables with mean zero and variance 2 and

j= |j |

< . We

recall that the covariance of above is

c(k) = E(Xt Xt+k ) =

j j+k .

(8.28)

j=

Since

j= |j |

< , it can be seen that

|c(k)|

X X

|j | |j+k | < .

k j=

Hence by using Theorem 8.3.1, the spectral density function of {Xt } is well defined.

There

are several ways to derive the spectral density of {Xt }, we can either use (8.28) and f () =
1 P
k c(k) exp(ik) or obtain the spectral representation of {Xt } and derive f () from the spec2
tral representation. We prove the results using the latter method.

8.5.1

The spectral representation of linear processes

Since {t } are iid random variables, using Theorem 8.4.1 there exists an orthogonal random function
{Z()} such that
Z
t =

exp(it)dZ().
0

Since E(t ) = 0 and E(2t ) = 2 multiplying the above by t , taking expectations and noting that
due to the orthogonality of {Z()} we have E(dZ(1 )dZ(2 )) = 0 unless 1 = 2 we have that
E(|dZ()|2 ) = 2 d, hence f () = (2)1 2 .

216

Using the above we obtain the following spectral representation for {Xt }
2

Z
Xt =
0

j=

j exp(ij) exp(it)dZ().

Hence
Z

A() exp(it)dZ(),

Xt =

(8.29)

where A() =

j= j

exp(ij), noting that this is the unique spectral representation of Xt .

Definition 8.5.1 (The Cramer Representation) We mention that the representation in (8.29)
of a stationary process is usually called the Cramer representation of a stationary process, where
Z
Xt =

A() exp(it)dZ(),
0

where {Z() : 0 2} are orthogonal functions.


Exercise 8.3

(i) Suppose that {Xt } has an MA(1) representation Xt = t + t1 . What is its

Cramers representation?
(ii) Suppose that {Xt } has a causal AR(1) representation Xt = Xt1 + t . What is its Cramers
representation?

8.5.2

The spectral density of a linear process

Multiplying (8.29) by Xt+k and taking expectations gives


2

A(1 )A(2 ) exp(i(t + k)1 + it2 )E(dZ(1 )dZ(2 )).

E(Xt Xt+k ) = c(k) =


0

Due to the orthogonality of {Z()} we have E(dZ(1 )dZ(2 )) = 0 unless 1 = 2 , altogether this
gives
Z
E(Xt Xt+k ) = c(k) =

|A()|2 exp(ik)E(|dZ()|2 ) =

where f () =

2
2
2 |A()| .

f () exp(ik)d,
0

Comparing the above with (8.14) we see that f () is the spectral density

function.

217

The spectral density function corresponding to the linear process defined in (8.27) is

2 X
|
j exp(ij)|2 .
f () =
2
j=

Remark 8.5.1 (An alternative, more hands on proof ) An alternative proof which avoids the
P
Cramer representation is to use that the acf of a linear time series is c(r) = 2 k j j+r (see
Lemma 3.1.1). Thus by definition the spectral density function is

f () =
=

1 X
c(r) exp(ir)
2 r=

2 X X
j j+r exp(ir).
2 r=
j=

Now make a change of variables s = j + r this gives


f () =

2
2 X X
j s exp(i(s j)) =
|A()|2 .
2 r= s=
2

Example 8.5.1 Let us suppose that {Xt } is a stationary ARM A(p, q) time series (not necessarily
invertible or causal), where

Xt

p
X

j Xtj =

j=1

q
X

j tj ,

j=1

{t } are iid random variables with E(t ) = 0 and E(2t ) = 2 . Then the spectral density of {Xt } is
Pq
2
2 |1 + j=1 j exp(ij)|
Pq
f () =
2 |1 j=1 j exp(ij)|2
We note that because the ARMA is the ratio of trignometric polynomials, this is known as a rational
spectral density.
Remark 8.5.2 The roots of the characteristic function of an AR process will have an influence on
the location of peaks in its corresponding spectral density function. To see why consider the AR(2)
model
Xt = 1 Xt1 + 2 Xt2 + t ,

218

where {t } are iid random variables with zero mean and E(2 ) = 2 . Suppose the roots of the
characteristic polynomial (B) = 1 1 B 2 B 2 lie outside the unit circle and are complex
conjugates where 1 = r exp(i) and 2 = r exp(i). Then the spectral density function is
2

f () =

|1 r exp(i(

))|2 |1

r exp(i( )|2

2
.
[1 + r2 2r cos( )][1 + r2 2r cos( + )]

If r > 0, the f () is maximum when = , on the other hand if, r < 0 then the above is maximum
when = . Thus the peaks in f () correspond to peaks in the pseudo periodicities of the
time series and covariance structure (which one would expect), see Section 3.1.2. How pronounced
these peaks are depend on how close r is to one. The close r is to one the larger the peak. We
can generalise the above argument to higher order Autoregressive models, in this case there may be
multiple peaks. In fact, this suggests that the larger the number of peaks, the higher the order of the
AR model that should be fitted.

8.5.3

Approximations of the spectral density to AR and MA spectral densities

In this section we show that the spectral density


f () =

1 X
c(r) exp(ir)
2 r=

can be approximated to any order by the spectral density of an AR(p) or MA(q) process.
We do this by truncating the infinite number of covariances by a finite number, however, this
does not necessarily lead to a positive definite spectral density. This can easily be proven by noting
that
fem () =

m
X

f ()Dm ( )d,

c(r) exp(ir) =
0

r=m

where Dm () = sin[(n + 1/2)]/ sin(/2). Observe that Dm () can be negative, which means that
fem () can be negative despite f being positive.
Example 8.5.2 Consider the AR(1) process Xt = 0.75Xt1 + t where var[t ] = 1. In Lemma

219

3.1.1 we showed that the autcovariance corresponding to this model is c(r) = [1 0.752 ]1 0.75|r| .
Let us define a process whose autocorrelation is c(0) = [1 0.752 ]1 , c(1) = c(1) = [1
0.752 ]1 0.75 and c(r) = 0 for |r| > 1. The spectral density of this process is
1
fem () =
1 0.752


3
1 + 2 cos[] .
4

It is clear that this function can be zero for some values of . This means that {
c(r)} is not a well
defined covariance function, hence there does not exist a time series with this covariance structure.
In other words, simply truncating an autocovariance is not enough to guarantee that it positive
definite sequence.
Instead we consider a slight variant on this and define

m 
1 X
|r|
1
c(r) exp(ir)
2 r=m
m
which is positive.
Remark 8.5.3 We note that fm is known as a Ces
aro sum because it can be written as
fm () =


m 
m
1 X
|r|
1 Xe
1
c(r) exp(ir) =
fn (),
2 r=m
m
m

(8.30)

n=0

where fen () =

1
2

Pn

r=n c(r) exp(ir).

Strangely, there is no guarantee that the truncated Fourier

transform fen is not negative, however fn () is definitely positive. There are are a few ways to prove
this:
(i) The first method we came across previously, var[Jn ()] = fn (), it is clear that using this
construction inf fn () 0.
(ii) By using (8.30) we can write fm () as
2

f ()Fm ( )d,

fm () =
0

where Fm () =

1
m

Pm

r=m Dr () =

1
m

sin(n/2)
sin(/2)

2

and Dr () =

Pr

j=r

exp(ij) (these are

the Fejer and Dirichlet kernels respectively). Since both f and Fm are positive, then fm has
to be positive.

220

The Cesaro sum is special in the sense that


sup |fm () f ()| 0,

as m .

(8.31)

Thus for a large enough m, fm () will be within of the spectral density f . Using this we can
prove the results below.
Lemma 8.5.1 Suppose that

|c(r)| < and f is the spectral density of the covariances. Then

for every > 0, there exists a m such that |f () fm ()| < and fm () = 2 |()|2 , where
Pm
() =
j=0 j exp(ij). Thus we can approximate the spectral density of f with the spectral
density of a MA.
PROOF. We show that there exists an MA(m) which has the spectral density fm (), where fm is
defined in (8.30). Thus by (8.31) we have the result.
Before proving the result we note that if a polynomial is of the form

p(z) = a0 +

m1
X

ar z + am z +

r=1

then it has the factorization p(z) = C

2m
X

arm z r ,

r=m+1
1
j=1 [1 j z][1 j z],

Qm

where j is such that |j | < 1. The

proof is clear by factorising the polynomial and using contradiction. Therefore, if a(r) = a(r),
then we have
m
X

a(r) exp(ir) = exp(im)

r=m

2m
X

a(r m) exp(ir)

r=0
m
Y

= C exp(im)
= C(1)m

m
Y
j=1

h
i
[1 j exp(i)] 1 1
exp(i)
j

j=1
m h
Y

ih
i
1
1 1
exp(i)
1

exp(i)
,
j
j

j=1

for some finite constant C. Using the above, with a(r) = [1 |r|n1 ]c(r), we can write fm as

fm () = K

m
Y

(1 1
j exp(i))

j=1

m
Y

(1 1
j exp(i))

j=1
2

= A()A() = |A()| ,

221

where
A(z) =

m
Y

(1 1
j z).

j=1

Since A(z) is an mth order polynomial where all the roots are greater than 1, we can always
construct an MA(m) process which has A(z) as its transfer function. Thus there exists an MA(m)
process which has fm () as its spectral density function.

Lemma 8.5.2 Suppose that

|c(r)| < and f is corresponding the spectral density function

where inf f () > 0. Then for every > 0, there exists a m such that |f () gm ()| < and
P
gm () = 2 |()1 |2 , where () = m
j=0 j exp(ij) and the roots of (z) lie outside the unit
circle. Thus we can approximate the spectral density of f with the spectral density of a causal
autoregressive process.
PROOF. We first note that we can write


f () gm () = f ()|gm ()1 f ()1 |gm ().
Since f () L2 and is bounded away from zero, then f 1 L2 and we can write f 1 as
f 1 () =

dr exp(ir),

r=

where dr are the Fourier coefficients of f 1 . Since f is positive and symmetric, then f 1 is posP
ir and {d } is a positive definite symmetric
itive and symmetric such that f 1 () =
r
r= dr e
sequence. Thus we can define the positive function gm where
1
gm
()

X 
|r|m

|r|
1
m


dr exp(ir)

1 () f 1 ()| < , which implies


and is such that |gm

X


f () gm () [
|c(r)|]2 .
r
1 can be
Now we can apply the same arguments to prove to Lemma 8.5.1 we can show that gm
1 () = C| ()|2 (where
factorised as gm
m
m is an mth order polynomial whose roots lie outside

222

the unit circle). Thus gm () = C|m ()|2 and we obtain the desired result.

8.6

Higher order spectrums

We recall that the covariance is a measure of linear dependence between two random variables.
Higher order cumulants are a measure of higher order dependence. For example, the third order
cumulant for the zero mean random variables X1 , X2 , X3 is
cum(X1 , X2 , X3 ) = E(X1 X2 X3 )
and the fourth order cumulant for the zero mean random variables X1 , X2 , X3 , X4 is
cum(X1 , X2 , X3 , X4 ) = E(X1 X2 X3 X4 ) E(X1 X2 )E(X3 X4 ) E(X1 X3 )E(X2 X4 ) E(X1 X4 )E(X2 X3 ).
From the definition we see that if X1 , X2 , X3 , X4 are independent then cum(X1 , X2 , X3 ) = 0 and
cum(X1 , X2 , X3 , X4 ) = 0.
Moreover, if X1 , X2 , X3 , X4 are Gaussian random variables then cum(X1 , X2 , X3 ) = 0 and
cum(X1 , X2 , X3 , X4 ) = 0. Indeed all cumulants higher than order two is zero. This comes from
the fact that cumulants are the coefficients of the power series expansion of the logarithm of the
characteristic function of {Xt }, which is
1
gX (t) = i 0 t t0 |{z}
t.
2
|{z}
cumulant

mean

Since the spectral density is the Fourier transform of the covariance it is natural to ask whether
one can define the higher order spectral density as the fourier transform of the higher order cumulants. This turns out to be the case, and the higher order spectra have several interesting properties.
Let us suppose that {Xt } is a stationary time series (notice that we are assuming it is strictly stationary and not second order). Let 3 (t, s) = cum(X0 , Xt , Xs ), 3 (t, s, r) = cum(X0 , Xt , Xs , Xr )
and q (t1 , . . . , tq1 ) = cum(X0 , Xt1 , . . . , Xtq ) (noting that like the covariance the higher order cumulants are invariant to shift). The third, fourth and the general qth order spectras is defined

223

as
f3 (1 , 2 ) =
f4 (1 , 2 , 3 ) =
fq (1 , 2 , . . . , q1 ) =

3 (s, t) exp(is1 + it2 )

s= t=

X
X

4 (s, t, r) exp(is1 + it2 + ir3 )

s= t= r=

q (t1 , t2 , . . . , tq1 ) exp(it1 1 + it2 2 + . . . + itq1 q1 ).

t1 ,...,tq1 =

Example 8.6.1 (Third and Fourth order spectral density of a linear process) Let us suppose that {Xt } satisfies

Xt =

j tj

j=

where

j= |j |

< , E(t ) = 0 and E(4t ) < . Let A() =

j= j

exp(ij). Then it is

straightforward to show that


f () = 2 |A()|2
f3 (1 , 2 ) = 3 A(1 )A(2 )A(1 2 )
f4 (1 , 2 , 3 ) = 4 A(1 )A(2 )A(3 )A(1 2 3 ),
where 3 = cum(t , t , t ) and 4 = cum(t , t , t , t ).
We see from the example, that unlike the spectral density, the higher order spectras are not
necessarily positive or even real.
A review of higher order spectra can be found in Brillinger (2001). Higher order spectras have
several applications especially in nonlinear processes, see Subba Rao and Gabr (1984). We will
consider one such application in a later chapter.
Using the definition of the higher order spectrum we can now generalise Lemma 8.2.1 to higher
order cumulants (see Brillinger (2001), Theorem 4.3.4).
Proposition 8.6.1 {Xt } is a strictly stationary time series, where for all 1 i q 1 we
P
have
t1 ,...,tq1 = |(1 + ti )q (t1 , . . . , tq1 )| < (note that this is simply a generalization of the

224

covariance assumption

|rc(r)| < ). Then we have

cum(Jn (k1 ), . . . , Jn (kq )) =

fq (k2 , . . . , kq )
nq/2

where ki =

8.7
8.7.1

n
X

exp(ij(k1 . . . kq )) + O(

j=1

1
f ( , . . . , kq )
n(q1)/2 q k2
1
O( nq/2
)

1
+ O( nq/2
)

Pq

i=1 ki

1
nq/2

= nZ

otherwise

2ki
n .

Extensions
The spectral density of a time series with randomly missing
observations

Let us suppose that {Xt } is a second order stationary time series. However {Xt } is not observed at
everytime point and there are observations missing, thus we only observe Xt at {k }k . Thus what is
observed is {Xk }. The question is how to deal with this type of data. One method was suggested
in ?. He suggested that the missingness mechanism {k } be modelled stochastically. That is define
the random process {Yt } which only takes the values {0, 1}, where Yt = 1 if Xt is observed, but
Yt = 0 if Xt is not observed. Thus we observe {Xt Yt }t = {Xtk } and also {Yt } (which is the time
points the process is observed). He also suggests modelling {Yt } as a stationary process, which is
independent of {Xt } (thus the missingness mechanism and the time series are independent).
The spectral densities of {Xt Yt }, {Xt } and {Yt } have an interest relationship, which can be
exploited to estimate the spectral density of {Xt } given estimators of the spectral densities of {Xt Yt }
and {Xt } (which we recall are observed). We first note that since {Xt } and {Yt } are stationary,
then {Xt Yt } is stationary, furthermore
cov(Xt Yt , X Y ) = cov(Xt , X )cov(Yt , Y ) + cov(Xt , Y )cov(Yt , X ) + cum(Xt , Yt , X , Y )
= cov(Xt , X )cov(Yt , Y ) = cX (t )cY (t )

225

where the above is due to independence of {Xt } and {Yt }. Thus the spectral density of {Xt Yt } is

fXY () =

1 X
cov(X0 Y0 , Xr Yr ) exp(ir)
2 r=

1 X
=
cX (r)cY (r) exp(ir)
2 r=
Z
=
fX ()fY ( )d,
where fX () =

1
2

r= cX (r) exp(ir)

and fY () =

densities of the observations and the missing process.

226

1
2

r= cY (r) exp(ir)

are the spectral

Chapter 9
Spectral Analysis
Prerequisites
The Gaussian likelihood.
The approximation of a Toeplitz by a Circulant (covered in previous chapters).

Objectives
The DFTs are close to uncorrelated but have a frequency dependent variance (under stationarity).
The DFTs are asymptotically Gaussian.
For a linear time series the DFT is almost equal to the transfer function times the DFT of
the innovations.
The periodograms is the square of the DFT, whose expectation is approximately equal to the
spectral density. Smoothing the periodogram leads to an estimator of the spectral density as
does truncating the covariances.
The Whittle likelihood and how it is related to the Gaussian likelihood.
Understand that many estimator can be written in the frequency domain.
Calculating the variance of an estimator.

227

9.1

The DFT and the periodogram

In the previous section we motivated transforming the stationary time series {Xt } into its discrete
Fourier transform
n

Jn (k ) =

2t
1 X
Xt exp(ik
)
n
2n t=1


n
n
2t
2t
1 X
1 X

Xt cos(k
Xt sin(k
) + i
)
n
n
2n t=1
2n t=1

k = 0, . . . , n/2

(frequency series) as an alternative way of analysing the time series. Since there is a one-to-one
mapping between the two, nothing is lost by making this transformation. Our principle reason for
n/2

using this transformation is given in Lemma 8.2.1, where we showed that {Jn (k )}n=1 is an almost
uncorrelated series. However, there is a cost to the uncorrelatedness property, that is unlike the
original stationary time series {Xt }, the variance of the DFT varies over the frequencies, and the
variance is the spectral density at that frequency. We summarise this result below, but first we
recall the definition of the spectral density function

1 X
c(r) exp(ir)
f () =
2 r=

[0, 2].

(9.1)

We summarize some of the results derived in Chapter 8 here.


Lemma 9.1.1 Suppose that {Xt } is a zero second order stationary time series, where cov(X0 , Xr ) =
P
c(r) and r |c(r)| < . Define k = 2k
n . Then
(i)

|Jn ()|2 =

1
2

n1
X

cn (r) exp(ir),

(9.2)

r=(n1)

where cn (r) is the sample autocovariance.


(ii) for k 6= 0 we have E[Jn (k )] = 0,
X



1 X
E(|Jn ()|2 ) f () 1
|c(r)| +
|rc(r)| 0
2
n
|r|n

as n ,

228

|r|n

(9.3)

(iii)


f ( 2k ) + o(1) k1 = k2
2k1
2k2
n
cov Jn (
), Jn (
) =

n
n
o(1)
k1 =
6 k2


where f () is the spectral density function defined in (9.1). Under the stronger condition

|rc(r)| <

the o(1) above is replaced with O(n1 ).


In addition if we have higher order stationarity (or strict stationarity), then we also can find
expressions for the higher order cumulants of the DFT (see Proposition 8.6.1).
It should be noted that even if the mean of the stationary time series {Xt } is not zero (ie. E(Xt ) =

6= 0), so long as k 6= 0 E(Jn (k )) = 0 (even without centering Xt , with Xt X).


Since there is a one-to-one mapping between the observations and the DFT, it is not surprising
that classical estimators can be written in terms of the DFT. For example, the sample covariance
can be rewritten in terms of the DFT
n

1X
b
cn (r) + b
cn (n r) =
|Jn (k )|2 exp(irk ).
n

(9.4)

k=1

(see Appendix A.3(iv)). Since b


cn (n r) =

1
n

Pn

t=|nr| Xt Xt+|nr| ,

for small r (relative to T ) this

term is negligible, and gives


n

b
cn (r)

1X
|Jn (k )|2 exp(irk ).
n

(9.5)

k=1

The modulo square of the DFT plays such an important role in time series analysis that it has
its own name, the periodogram, which is defined as
In () = |Jn ()|2 =

1
2

n1
X

cn (r) exp(ir).

(9.6)

r=(n1)

By using Lemma 9.1.1 or Theorem 8.6.1 we have E(In ()) = f () + O( n1 ). Moreover, (9.4) belongs
to a general class of integrated mean periodogram estimators which have the form
n

A(, In ) =

1X
In (k )(k ).
n
k=1

229

(9.7)

Replacing the sum by an integral and the periodogram by its limit, it is clear that these are
estimators of the integrated spectral density
Z

f ()()d.

A(f, ) =
0

Before we consider these estimators (in Section 9.5). We analyse some of the properties of the
DFT.

9.2

Distribution of the DFT and Periodogram under


linearity

An interesting aspect of the DFT, is that under certain conditions the DFT is asymptotically
normal. We can heuristically justify this by noting that the DFT is a (weighted) sample mean. In
pn
In this section we prove this result,
X).
fact at frequency zero, it is the sample mean (Jn (0) = 2
and a similar result for the periodogram. We do the proof under linearity of the time series, that is
Xt =

j tj ,

j=

however the result also holds for nonlinear time series (but is beyond this course).
1 Pn
itk is a very simple object to deal with it.
The DFT of the innovations J (k ) = 2n
t=1 t e
First the DFT is an orthogonal transformation and the orthogonal transformation of iid random
variables leads to uncorrelated random variables. In other words, {J (k )} is completely uncorrelated as are its real and imaginary parts. Secondly, if {t } are Gaussian, then {J (k )} are
independent and Gaussian. Thus we start by showing the DFT of a linear time series is approximately equal to the DFT of the innovations multiplied by the transfer function. This allows us to
transfer results regarding J (k ) to Jn (k ).
P
We will use the assumption that j |j 1/2 j | < , this is a slightly stronger assumption than
P
j |j | < (which we worked under in Chapter 2).
Lemma 9.2.1 Let us suppose that {Xt } satisfy Xt =

230

j= j t ,

where

j= |j

1/2 |
j

< ,

and {t } are iid random variables with mean zero and variance 2 . Let
n

J () =

1 X
t exp(it).
2n t=1

Then we have
X

Jn () =


j exp(ij) J () + Yn (),

(9.8)

j
1
j j exp(ij)Un,j , with Un,j =
2n
P

1
1/2 )2 = O( 1 ).
( n1/2
j= |j | min(|j|, n)
n

where Yn () =
E(Yn ())2

Pnj

t=1j

exp(it)t

Pn

t=1 exp(it)t

and

PROOF. We note that

1 X
Jn () =
j exp(ij)
tj exp(it)
2n t=1
j=
nj
X
1
s exp(is)
=
j exp(ij)
2n s=1j
j=

nj
n
X
X
X
X
1
=
j exp(ij) J () +
j exp(ij)
exp(it)t
exp(it)t .
2n j
t=1
j
t=1j
{z
}
|

=Yn ()

We will show that Yn () is negligible with respect to the first term. We decompose Yn () into
three terms

Yn () =

n
X

nj
X

n
X

1
j eij
exp(it)t
exp(it)t +
2n j=
t=1
t=1j
{z
}
|
no terms in common

nj
n
n
X
X
X
1

j eij
exp(it)t
exp(it)t +
2n j=n
t=1
t=1j
|
{z
}

(nj) terms in common, 2j terms not in common

nj
X

1
j eij
exp(it)t
2n j=n+1
t=1j
|
{z

n
X

exp(it)t

t=1

no terms in common

= I + II + II.

231

If we took the expectation of the absolute of Yn () we find that we require the condition

|jj | <

(and we dont exploit independence of the innovations). However, by evaluating E|Yn ()|2 we
exploit to independence of {t }, ie.

[E(I 2 )]1/2

n
X

nj
X

|j | E
exp(it)t
2n j=
t=1j

n
X

2 1/2
exp(it)t

t=1

n
n

X
X
X

1/2
1
1
1

|j | 2n 2

|j 1/2 j |
|j 1/2 j |
2n j=
2n j=
2n j=

similarly, III = O(n1/2 ) and

[E(I 2 )]1/2

n
X

nj
X

1
|j | E
exp(it)t
2n j=n
t=1j
n
X

n
X

2 1/2
exp(it)t

t=1

n
X

X


1
1
1
2 1/2
1/2

|j | 2j

|j j |
|j 1/2 j |.
2n j=n
2n j=
2n j=

Thus we obtain the desired result.

The above shows that under linearity and the condition


Jn () =

X
j

|j 1/2 j | < we have


1
j exp(ij) J () + Op ( ).
n

(9.9)

This implies that the distribution of Jn () is determined by the DFT of the innovations J (). We
generalise the above result to the periodogram.
P
Lemma 9.2.2 Let us suppose that {Xt } is a linear time series Xt =
j= j tj , where
P
1/2 | < , and { } are iid random variables with mean zero, variance 2 E(4 ) < .
j
t
t
j= |j
Then we have
2




X
In () =
j exp(ij) |J ()|2 + Rn (),

j

(9.10)

where E(sup |Rn ()|) = O( n1 ).


PROOF. See Priestley (1983), Theorem 6.2.1 or Brockwell and Davis (1998), Theorem 10.3.1. 

232

To summarise the above result, for a general linear process Xt =


In () = |

X
j

j= j tj

we have

1
1
j exp(ij)|2 |J ()|2 + Op ( ) = 2f ()I () + Op ( ),
n
n
1 P
j
2 |

where we assume w.l.o.g. that var(t ) = 1 and f () =

(9.11)

j exp(ij)|2 is the spectral density

of {Xt }.
The asymptotic normality of Jn () follows from asymptotic normality of J (), which we prove
in the following proposition.
Proposition 9.2.1 Suppse {t } are iid random variables with mean zero and variance 2 . We

2
1 Pn
1 Pn

define J () = 2n
t=1 t exp(it) and I () = 2n
t=1 t exp(it) . Then we have

J () =

<J ()
=J ()



2
0, I2 ,
2

(9.12)

where I2 is the identity matrix. Furthermore, for any finite m




2
(J (k1 ) , . . . , J (km ) ) N 0, I2m ,
2
0

(9.13)

I ()/ 2 2 (2) (which is equivalent to the exponential distribution with mean one) and

cov(|J (j )| , |J (k )| ) =

4
(2)2 n
4
(2)2 n

2 4
(2)2

j 6= k

(9.14)

j=k

where j = 2j/n and k = 2k/n (and j, k 6= 0 or n).


1 Pn
PROOF. We first show (9.15). We note that <(J (k )) = 2n
t=1 t,n and =(J (k )) =
P
n
1
t=1 t,n where t,n = t cos(2kt/n) and t,n = t sin(2kt/n). We note that <(J (k )) =
2n
P
Pn
n
1
1
t=1 t,n and =(J (k )) =
t=1 t,n are the weighted sum of iid random variables,
2n
2n

hence {t,n } and {t,n } are martingale differences. Therefore, to show asymptotic normality, we
will use the martingale central limit theorem with the Cramer-Wold device to show that (9.15). We
note that since {t,n } and {t,n } are independent random variables we an prove the same result
using a CLT for independent, non-identically distributed variables. However, for practice we will
use a martingale CLT. To prove the result we need to verify the three conditions of the martingale

233

CLT. First we consider the conditional variances


n


1 X
E |t,n |2 t1 , t2 , . . . , 1 =
2n

2
1 X
P
cos(2kt/n)2 2t
2n
2

1
2n

1
2n

1
2n

t=1
n
X



E |t,n |2 t1 , t2 , . . . , 1 =

t=1
n
X



E t,n t,n t1 , t2 , . . . , 1 =

t=1

1
2n

t=1
n
X
t=1
n
X

sin(2kt/n)2 2t

2
2
P

cos(2kt/n) sin(2kt/n)2t 0,

t=1

where the above follows from basic calculations using the mean and variance of the above. Finally
1 Pn
we need to verify the Lindeberg condition, we only verify it for 2n
t=1 t,n , the same argument
P
n
1
holds true for 2n
t=1 t,n . We note that for every  > 0 we have
n
n



1 X 
1 X
2

E |t,n | I(|t,n | 2 n) t1 , t2 , . . . =
E |t,n |2 I(|t,n | 2 n) .
2n
2n
t=1

t=1

By using |t,n | = | cos(2t/n)t | |t | the above can be bounded by


n

1 X 
E |t,n |2 I(|t,n | 2 n)]
2n
t=1

n


 P
1 X  2
E |t | I(|t | 2 n) = E |t |2 I(|t | 2 n) 0
2n

as

n ,

t=1

the above is true because E(2t ) < . Hence we have verified Lindeberg condition and we obtain
(9.15). The proof of (9.13) is similar, hence we omit the details. Because I () = <(J ())2 +
=(J ())2 , from (9.15) we have I ()/ 2 2 (2).
To prove (9.14) we can either derive it from first principles or by using Proposition 8.6.1. Here
we do it from first principles. We observe
cov(I (j ), I (k )) =

XXXX
1
cov(t1 t1 +k1 , t2 t2 +k2 ).
(2)2 n2
t
t
k1

k2

Expanding the covariance gives


cov(t1 t1 +k1 , t2 t2 +k2 ) = cov(t1 , t2 +k2 )cov(t2 , t1 +k1 ) + cov(t1 , t2 )cov(t1 +k1 , t2 +k2 ) +
cum(t1 , t1 +k1 , t2 , t2 +k2 ).

234

Since {t } are iid random variables, for most t1 , t2 , k1 and k2 the above covariance is zero. The
exceptions are when t1 = t2 and k1 = k2 or t1 = t2 and k1 = k2 = 0 or t1 t2 = k1 = k2 . Counting
all these combinations we have
cov(|J (j )|2 , |J (k )|2 ) =

X
1
2 4 X X X
exp(ik(j k )) +
4
2
2
2
2
(2) n
(2)
n
t
t
t
k

P
where 2 = var(t ) and 4 = cum4 () = cum(t , t , t , t ). We note that for j 6= k, t exp(ik(j
P
k )) = 0 and for j = k, t exp(ik(j k )) = n, substutiting this into cov(|J (j )|2 , |J (k )|2 )
gives us the desired result.

By using (9.9) the following result follows immediately from Lemma 9.2.1, equation (9.15).
P
Corollary 9.2.1 Let us suppose that {Xt } is a linear time series Xt =
j= j tj , where
P
1/2 | < , and { } are iid random variables with mean zero, variance 2 E(4 ) < .
j
t
t
j= |j
Then we have

<Jn ()
=Jn ()

N (0, f ()I2 ) ,

(9.15)

Using (9.11) we see that In () f ()|J ()|2 . This suggest that most of the properties which
apply to |J ()2 also apply to In (). Indeed in the following theorem we show that the asympototic
distribution of In () is exponential with asymptotic mean f () and variance f ()2 (unless = 0
in which case it is 2f ()2 ).
By using Lemma 9.2.1 we now generalise Proposition 9.2.1 to linear processes. We show that
just like the DFT the Periodogram is also near uncorrelated at different frequencies. This result
will be useful when motivating and deriving the sampling of the spectral density estimator in
Section 9.3.
Theorem 9.2.1 Suppose {Xt } is a linear time series Xt =

j= j tj ,

where

j= |j

1/2 |
j

<

with E[t ] = 0, var[t ] = 2 and E[4t ] < . Let In () denote the periodogram associated with
{X1 , . . . , Xn } and f () be the spectral density. Then
(i) If f () > 0 for all [0, 2] and 0 < 1 , . . . , m < , then

In (1 )/f (1 ), . . . , In (m )/f (m )

235

converges in distribution (as n ) to a vector of independent exponential distributions


with mean one.
(ii) Furthermore, for j =

2j
n

and k =

2k
n

we have

2f (k )2 + O(n1/2 ) j = k = 0 or

cov(In (k ), In (j )) =
f (k )2 + O(n1/2 ) 0 < j = k <

O(n1 )
j 6= k
where the bound is uniform in j and k .

Remark 9.2.1 (Summary of properties of the periodogram)

(i) The periodogram is nonP


negative and is an asymptotically an unbiased estimator of the spectral density (when j |j | <

).
(ii) It symmetric about zero, In () = In ( + ), like the spectral density function.
(iii) At the fundemental frequencies {In (j )} are asymptotically uncorrelated.
(iv) If 0 < < , In () is asymptotically exponentially distributed with mean f ().
It should be mentioned that Theorem 9.2.1 also holds for several nonlinear time series too.

9.3

Estimating the spectral density function

There are several explanations as to why the raw periodogram can not be used as an estimator of
the spectral density function, despite its mean being approximately equal to the spectral density.
One explanation is a direct consequence of Theorem 9.2.1, where we showed that the distribution of
the periodogram standardized with the spectral density function is a chi-squared with two degrees
of freedom, from here it is clear it will not converge to the mean, however large the sample size.
An alternative, explanation is that the periodogram is the Fourier transform of the autocovariances
estimators at n different lags. Typically the variance for each covariance cn (k) will be about O(n1 ),
thus, roughly speaking, the variance of In () will be the sum of these n O(n1 ) variances which
leads to a variance of O(1), this clearly does not converge to zero.
Both these explanation motivate estimators of the spectral density function, which turn out to
be the same. It is worth noting that Parzen (1957) first proposed a consistent estimator of the

236

spectral density. These results not only lead to a revolution in spectral density estimation but also
the usual density estimation that you may have encountered in nonparametric statistics (one of the
first papers on density estimation is Parzen (1962)).
We recall that Jn (k ) are zero mean uncorrelated random variables whose variance is almost
equal to f (k ). This means that E|Jn (k )|2 = E[In (n )] f (k ).
Remark 9.3.1 (Smoothness of the spectral density) We observe that
f (s) () =

1 X
(ir)s c(r) exp(ir).
(2)
rZ

Therefore, the smoothness of the spectral density function is determined by finiteness of

|rs c(r)|,

in other words how fast the autocovariance function converges to zero. We recall that the acf of
ARMA processes decay exponential fast to zero, thus f is extremely smooth (all derivatives exist).
Assuming that the autocovariance function converges to zero sufficiently fast f will slowly vary over
frequency. Furthermore, using Theorem 9.2.1, we know that {In (k )} are close to uncorrelated and
In (k )/f (k ) is exponentially distributed. Therefore we can write In (k ) as
In (k ) = E(In (k )) + [In (k ) E(In (k ))]
f (k ) + f (k )Uk ,

k = 1, . . . , n,

(9.16)

where {Uk } is sequence of mean zero and constant variance almost uncorrelated random variables.
We recall (9.16) resembles the usual nonparametric equation (function plus noise) often considered in nonparametric statistics.
Remark 9.3.2 (Nonparametric Kernel estimation) Let us suppose that we observe Yi where
 
i
+ i
Yi = g
n

1 i n,

and {i } are iid random variables and g() is a smooth function. The kernel density estimator of
gbn ( ni )
  X


j
1
ji
gbn
=
W
Yi ,
n
bn
bn
i

where W () is a smooth kernel function of your choosing, such as the Gaussian kernel, etc.

237

This suggest that to estimate the spectral density we could use a local weighted average of {In (k )}.
Equation (9.16) motivates the following nonparametric estimator of f ()


X 1
j

k
In (k ),
W
fbn (j ) =
bn
bn

(9.17)

where W () is a spectral window which satisfies

W (x)dx = 1 and

W (x)2 dx < .

Example 9.3.1 (Spectral windows) Here we give examples of spectral windows (see Section
6.2.3, page 437 in Priestley (1983)).
(i) The Daniell spectral Window is the local average

1/2 |x| 1
W (x) =
0 |x| > 1
This window leads to the estimator

fbn (j ) =

1
bn

j+bn/2

In (k ).

k=jbn/2

A plot of the periodgram, spectral density and different estimators (using Daniell kernel with
bn = 2 and bn = 10) of the AR(2) process Xt = 1.5Xt1 0.75Xt2 +t is given in Figure 9.1.
We observe that too small b leads to undersmoothing but too large b leads to over smoothing
of features. There are various methods for selecting the bandwidth, one commonly method
based on the Kullbach-Leibler criterion is proposed in Beltrao and Bloomfield (1987).
(ii) The Bartlett-Priestley spectral Window

W (x) =

3
4

1 x2

|x| 1
|x| > 1

This spectral window was designed to reduce the mean squared error of the spectral density
estimator (under certain smoothness conditions).
The above estimator was constructed within the frequency domain. We now consider a spectral
density estimator constructed within the time domain. We do this by considering the periodogram

238

Autoregressive

0 10 20 30 40 50 60

spectrum

0.0 0.2 0.4 0.6 0.8 1.0

P2[c(1:128)]

0.1

0.2

0.3

0.4

0.5

0.0

0.2

0.3

0.4

frequency

Series: ar2
Smoothed Periodogram

Series: ar2
Smoothed Periodogram

30

40

0.5

10

20

spectrum

60
40

20

spectrum

0.1

frequency[c(1:128)]

80 100

0.0

0.0

0.1

0.2

0.3

0.4

0.5

0.0

frequency
bandwidth = 0.00564

0.1

0.2

0.3

0.4

0.5

frequency
bandwidth = 0.0237

Figure 9.1: Using a realisation of the AR(2): Xt = 1.5Xt1 0.75Xt2 + t where n = 256.
Top left: Periodogram, Top Right: True spectral density function. Bottom left: Spectral
density estimator with bn = 2 and Bottom right: Spectral density estimator with bn = 10.
from an alternative angle. We recall that
1
2

In () =

n1
X

cn (k) exp(ik),

k=(n1)

thus it is the sum of n autocovariance estimators. This is a type of sieve estimator (a nonparametric
function estimator which estimates the coefficients/covariances in a series expansion). But as we
explained above, this estimator is not viable because it uses too many coefficient estimators. Since
the true coefficients/covariances decay to zero for large lags, this suggests that we do not use all the
sample covariances in the estimator, just some of them. Hence a viable estimator of the spectral
density is the truncated autocovariance estimator
fen () =

m
1 X
cn (k) exp(ik),
2

(9.18)

k=m

or a generalised version of this which down weights the sample autocovariances at larger lags

fen () =

1
2

n1
X

k=(n1)

239

k
m


cn (k) exp(ik),

(9.19)

where () is the so called lag window. The estimators (9.17) and (9.19) are very conceptionally simiR 2
lar, this can be understood if we rewrite cn (r) in terms the periodogram cn (r) = 0 In () exp(ir)d,
and transforming (9.19) back into the frequency domain

fen () =

where Wm () =

1
2
1
2

n1
X

Z
In ()

k=(n1)

k
1
( ) exp(ik( ))d =
m
2

Z
In ()Wm ( )d, (9.20)

Pn1

k
k=(n1) ( m ) exp(ik).

Example 9.3.2 (Examples of Lag windows) Here we detail examples of lag windows.
(i) Truncated Periodogram lag Window (u) = I[1,1] (u), where {(k/m)} corresponds to
m
1 sin[(m + 1/2)x]
1 X ik
e
=
,
Wm (x) =
2
2
sin(x/2)
k=m

which is the Dirichlet kernel.


Note that the Dirchlet kernel can be negative, thus we can see from (9.20) that fen can be
negative. Which is one potential drawback of this estimator (see Example 8.5.2).
(ii) The Bartlett lag Window (x) = (1 |x|)I[1,1] (x), where {(k/m)} corresponds to



m 
1 X
||k
1
sin(nx/2) 2
ik
Wm (x) =
1
e
=
2
m
2n sin(x/2)
k=m

which is the Fejer kernel. We can immediately see that one advantage of the Bartlett window
is that it corresponds to a spectral density estimator which is positive.
Note that in the case that m = n (the sample size), the truncated periodogram window estimaP
P
ir and the Bartlett window estimator corresponds to
tor corresponds to
|r|n c(r)e
|r|n [1
|r|/n]c(r)eir .
Wm () and 1b W ( b ) (defined in (9.17)) cannot not be the same function, but they share many of the
same characteristics. In particular,
n1
X

Wm () =

k=(n1)

= m

1
m

k
m

m1
X


exp(ik) = m

m1
X
k=(m1)

(k ) exp (ik (m)) ,

k=(m1)

240

k
m



k
exp i m
m

where k = k/n. By using (A.2) and (A.3) (in the appendix), we can approximate the sum by the
integral and obtain
Z
Wm () = mW (m) + O(1),

where W () =

(x) exp(i)dx.

Therefore
Z
fen () m

In ()K(m( ))d.

Comparing with fbn and fen () we see that m plays the same role as b1 . Furthermore, we observe
P 1
jk
k bn W ( bn )I(k ) is the sum of about nb I(k ) terms. The equivalent for Wm (), is that it
k
1 Pn1
has the spectral width n/m. In other words since fn () = 2
cn (k) exp(ik) =
k=(n1) ( M )
R
1
mIn ()W (M ( ))d, it is the sum of about n/m terms.
2
We now analyze the sampling properties of the spectral density estimator. It is worth noting
that the analysis is very similar to the analysis of nonparametric kernel regression estimator gbn ( nj ) =
ji
1 P
i
i
i W ( bn )Yi , where Yi = g( n ) + g( n )i and {i } are iid random variables. This is because the
bn
periodogram {In ()}k is near uncorrelated. However, still some care needs to be taken in the
proof to ensure that the errors in the near uncorrelated term does not build up.
Theorem 9.3.1 Suppose {Xt } satisfy Xt =

j= j tj ,

where

j= |jj |

< and E(4t ) <

. Let fn () be the spectral estimator defined in (9.17). Then




E(fn (j )) f (j ) C


1
+b
n

(9.21)

and

var[fn (j )]

1
2
bn f (j )

0 < j <

2
2
bn f (j )

j = 0 or

(9.22)

bn , b 0 as n .
PROOF. The proof of both (9.21) and (9.22) are based on the spectral window W (x/b) becoming narrower as b 0, hence there is increasing localisation as the sample size grows (just like
nonparametric regression).
We first note that by using Lemma 3.1.1(ii) we have

241

|rc(r)| < , thus |f 0 ()|

|rc(r)| <

. Hence f is continuous with a bounded first derivative.


To prove (9.21) we take expections




E(fn (j )) f (j )



 
X 1


k  

W
E I(jk ) f (j )



bn
bn
k
 
 
X 1
X 1



k
W
E I(jk ) f (jk ) +
W k |f (j ) f (jk )|



bn
bn
bn
bn

=
=

:= I + II.
Using Lemma 9.1.1 we have


X 1  k 


W
| E I(jk ) f (jk )
I =

bn
bn
k

!
X
X
X
1
k
1
1
C
|W ( )|
|c(k)| +
|kc(k)| = O( ).
bn
bn
n
n
|k|n

|k|n

To bound II we use that |f (1 ) f (2 )| sup |f 0 ()| |1 2 |, this gives


X 1

k 
II =
K( ) f (j ) f (jk ) = O(b).
bn
bn
k

Altogether this gives I = O(n1 ) and II = O(b) as bn , b 0 and n . The above two
bounds mean give (9.21).
We will use Theorem 9.2.1 to prove (9.22). We first assume that j 6= 0 or n. To prove the result
we use that
cov(|Jn (k1 )|2 , |Jn (k2 )|2 ) =
1
1
1
[f (k1 )I(k1 = k2 ) + O( )]2 + [f (k1 )I(k1 = n k2 ) + O( )][f (k1 )I(n k1 = k2 ) + O( )]
n
n
n
1
1
+[ f4 (1 , 1 , 2 ) + O( 2 )].
n
n

242

where the above follows from Proposition 8.6.1. This gives


var(fbn (j ))

 

X 1
j k1
j k2
W
W
cov(I(k1 ), I(k1 ))
=
(bn)2
bn
bn
k1 ,k2

 

X 1
j k1
j k2
=
W
W
(bn)2
bn
bn
k1 ,k2


1 2 
1 
1
f (k1 )I(k1 = k2 ) + O( ) + f (k1 )I(k1 = n k2 ) + O( )][f (k1 )I(n k1 = k2 ) + O( )
n
n
n


1
1
+ f4 (1 , 1 , 2 ) + O( 2 )
n
n


n
X
1
j k1 2
=
W
f (k2 )
(bn)2
bn
k=1

 

n
X
1
1
j k1
j (n k1 )
+
W
W
f (k2 ) + O( )
2
(bn)
bn
bn
n
k=1

2

 

Z
Z
j
j
j 2 +
1
1
1
1
1
2
=
W
f () d +
W
W
f ()d +O( )
2nb
b
b
2nb
b
b
b
n
|
{z
}
0
Z
 2
1
1
1
f (j )2
W
d + O( )
=
2nb
b
b
n
where the above is using the Riemann integral. A similar proof can be used to prove the case j = 0
or n.

The above result means that the mean squared error of the estimator

2
E fn (j ) f (j ) 0,
where bn and b 0 as n . Moreover

2
E fn (j ) f (j )
= O


1
+b .
bn

Remark 9.3.3 (The distribution of the spectral density estimator) Using that the periodogram
In ()/f () is asymptotically 2 (2) distributed and uncorrelated at the fundemental frequencies, we
can heuristically deduce the limiting distribution of fn (). Here we consider the distribution with

243

the rectangular spectral window

fn (j ) =

1
bn

j+bn/2

I(k ).

k=jbn/2

Pj+bn/2
Since I(k )/f (k ) are approximately 2 (2), then since the sum k=jbn/2 I(k ) is taken over a
Pj+bn/2
local neighbourhood of j , we have that f (j )1 k=jbn/2 I(k ) is approximately 2 (2bn).
We note that when bn is large, then 2 (2bn) is close to normal. Hence

bnfn (j ) N (f (j ), f (j )2 ).

Using this these asymptotic results, we can construct confidence intervals for f (j ).
In general, to prove normality of fn we rewrite it as a quadratic form, from this asymptotic
normality can be derived, where



Z
2
2

W (u) du .
bnfn (j ) N f (j ), f (j )

The variance of the spectral density estimator is simple to derive by using Proposition 8.6.1. The
remarkable aspect is that the variance of the spectral density does not involve (asymptotically) the
fourth order cumulant (as it is off lower order).

9.4

The Whittle Likelihood

In Chapter 6 we considered various methods for estimating the parameters of an ARMA process.
The most efficient method (in terms of Fisher efficiency), when the errors are Gaussian is the
Gaussian maximum likelihood estimator. This estimator was defined in the time domain, but it is
interesting to note that a very similar estimator which is asymptotically equivalent to the GMLE
estimator can be defined within the frequency domain. We start by using heuristics to define the
Whittle likelihood. We then show how it is related to the Gaussian maximum likelihood.
To motivate the method let us return to the Sunspot data considered in Exercise 5.1. The

244

Periodogram and the spectral density corresponding to the best fitting autoregressive model,


f () = (2) 1 1.1584ei 0.3890ei2 0.1674ei3 0.1385ei4 0.1054ei5 0.0559ei6
2

i7
i8

0.0049e 0.0572e 0.2378e ,
1

is given in Figure 9.2. We see that the spectral density of the best fitting AR process closely
follows the shape of the periodogram (the DFT modulo square). This means that indirectly the
autoregressive estimator (Yule-Walker) chose the AR parameters which best fitted the shape of
the periodogram. The Whittle likelihood estimator, that we describe below, does this directly. By

80 120

40

P2[c(2:n3)]

selecting the parametric spectral density function which best fits the periodogram. The Whittle

0.0

0.1

0.2

0.3

0.4

0.5

frequency2[c(2:n3)]

50 100
0

spectrum

Autoregressive

0.0

0.1

0.2

0.3

0.4

0.5

frequency

Figure 9.2: The periodogram of sunspot data (with the mean removed, which is necessary
to prevent a huge peak at zero) and the spectral density of the best fitting AR model.
likelihood measures the distance between In () and the parametric spectral density function using
the Kullbach-Leibler criterion
Lw
n ()


n 
X
In (k )
=
,
log f (k ) +
f (k )
k=1

k =

2k
,
n

and the parametric model which minimises this distance is used as the estimated model. The choice
of this criterion over the other distance criterions may appear to be a little arbitrary, however there
are several reasons why this is considered a good choice. Below we give some justifications as to

245

why this criterion is the prefered one.


First let us suppose that we observe {Xt }nt=1 , where Xt satisfies the ARMA representation
Xt =

p
X

j Xtj +

j=1

q
X

j tj + t ,

j=1

and {t } are iid random variables. We will assume that {j } and {j } are such that the roots
of their corresponding characteristic polynomial are greater than 1 + . Let = (, ). As we
P
mentioned in Section 8.2 if r |rc(r)| < , then

cov(Jn (k1 ), Jn (k2 )) =

f (k1 ) + O( n1 )

k1 = k2

O( n1 )

k1 6= k2 ,

where
f () =

2 |1 +
2|1 +

Pq

Ppj=1

j exp(ij)|2

j=1 j

exp(ij)|2

In other words, if the time series satisfies an ARMA presentation the DFT is near uncorrelated,
its mean is zero and its variance has a well specified parametric form. Using this information
we can define a criterion for estimating the parameters. We motivate this criterion through the
likelihood, however there are various other methods for motivating the criterion for example the
Kullbach-Leibler criterion is an alternative motivation, we comment on this later on.
If the innovations are Gaussian then <Jn () and =Jn () are also Gaussian, thus by using above
we approximately have

<Jn (1 )

=Jn (1 )

..

Jn =
.

<Jn (n/2 )

=Jn (n/2 )

N (0, diag(f (1 ), f (1 ), . . . , f (n/2 ), f (n/2 ))).

In the case that the innovations are not normal then, by Corollary 9.2.1, the above holds asymptotically for a finite number of frequencies. Here we construct the likelihood under normality of the
innovations, however, this assumption is not required and is only used to motivate the construction.
Since Jn is normally distributed random vector with mean zero and approximate diagonal

246

matrix variance matrix diag(f (1 ), . . . , f (n )), the negative log-likelihood of Jn is approximately


Lw
n ()

n 
X
k=1


|JX (k )|2
.
log |f (k )| +
f (k )

To estimate the parameter we would choose the which minimises the above criterion, that is
bw = arg min Lw (),

n
n

(9.23)

where consists of all parameters where the roots of the corresponding characteristic polynomial
have absolute value greater than (1 + ) (note that under this assumption all spectral densities
corresponding to these parameters will be bounded away from zero).
Example 9.4.1 Fitting an ARMA(1, 1) model to the data To fit an ARMA model to the data using
the Whittle likelihood we use the criterion
Lw
n ()

n/2 
X
k=1

2 |1 + eik |2
2|1 eik |2
log
+
I
(
)
n
k
2|1 eik |
2 |1 + eik |2


.

By differentiating Ln with respect to , 2 and we solve these three equations (usually numerically), this gives us the Whittle likelihood estimators.
Whittle (1962) showed that the above criterion is an approximation of the GMLE. The correct
proof is quite complicated and uses several matrix approximations due to Grenander and Szeg
o
(1958). Instead we give a heuristic proof which is quite enlightening.
Returning the the Gaussian likelihood for the ARMA process, defined in (7.10), we rewrite it
as


Ln () = det |Rn ()| + X0n Rn ()1 Xn = det |Rn (f )| + X0n Rn (f )1 Xn ,
where Rn (f )s,t =

(9.24)

f () exp(i(s t))d and X0n = (X1 , . . . , Xn ). We now show that Ln ()

Lw
n ().
Lemma 9.4.1 Suppose that {Xt } is a stationary ARMA time series with absolutely summable

247

covariances and f () is the corresponding spectral density function. Then


det |Rn (f )| +

X0n Rn (f )1 Xn

n 
X
k=1

|Jn (k )|2
log |f (k )| +
f (k )


+ O(1),

for large n.
PROOF. There are various ways to precisely prove this result. All of them show that the Toeplitz
matrix can in some sense be approximated by a circulant matrix. This result uses Szegos identity
(Grenander and Szeg
o (1958)). The main difficulty in the proof is showing that Rn (f )1
R
Un (f1 ), where Un (f1 )s,t = f ()1 exp(i(s t))d. An interesting derivation is given in
Brockwell and Davis (1998), Section 10.8. The main ingredients in the proof are:
1. For a sufficiently large m, Rn (f )1 can be approximated by Rn (gm )1 , where gm is the
spectral density of an mth order autoregressive process (this follows from Lemma 8.5.2), and
showing that


X 0n Rn (f )1 X n X 0n Rn (gm )1 X n = X 0n Rn (f )1 Rn (gm )1 X n
= X 0n Rn (gm )1 [Rn (gm ) Rn (f )] Rn (f1 )X n 0.

2. From Section 3.2.3, we recall if gm is the spectral density of an AR(m) process, then for
n >> m, Rn (gm )1 will be bandlimited with most of its rows a shift of the other (thus with
the exception of the first m and last m rows it is close to circulant).
3. We approximate Rn (gm )1 with a circulant matrix, showing that


1
X 0n Rn (gm )1 Cn (gm
) X n 0,
2 ) is the corresponding circulant matrix (where for 0 < |ij| m and either i or
where Cn (gm
P
j is greater than m, (Cn (g 1 ))ij = 2 m
k=|ij| m,k m,k|ij|+1 m,|ij| ) with the eigenvalues

{gm (k )1 }nk=1 .
4. These steps show that


1
) Xn 0
X 0n Rn (f )1 Un (gm
as m as n , which gives the result.

248

Remark 9.4.1 (A heuristic derivation) We give a heuristic proof. Using the results in Section
8.2 we have see that Rn (f ) can be approximately written in terms of the eigenvalue and eigenvectors
of the circulant matrix associated with Rn (f ), that is
Rn (f ) Fn (f )Fn
(n)

Rn (f )1 Fn (f )1 Fn ,

thus

(n)

(n)

where (f ) = diag(f (1 ), . . . , f (n )), f () =

P(n1)

j=(n1) c (k) exp(ik)

(9.25)
f () and

k = 2k/n. Basic calculations give


Xn Fn = (Jn (1 ), . . . , Jn (n )).

(9.26)

Substituting (9.26) and (9.25) into (9.27) yields


n

1X
1
Ln ()
n
n

k=1

|Jn (k )|2
log f (k ) +
f (k )


=

1 w
L ().
n

(9.27)

Hence using the approximation in (9.25) leads to a heuristic equivalence between the Whittle and
Gaussian likelihood.
Lemma 9.4.2 (Consistency) Suppose that {Xt } is a causal ARMA process with parameters
bw be defined as in (9.23)
whose roots lie outside the (1 + )-circle (where > 0 is arbitrary). Let
and suppose that E(4t ) < . Then we have
P
bw

PROOF. To show consistency we need to show pointwise convergence and equicontinuity of

1
n Ln .

Let
1
L () =
2
w

Z
0



f0 ()
log f () +
d.
f ()

It is straightforward to show that E( n1 Lw


n ()) Ln (). Next we evaluate the variance, to do this

249

we use Proposition 8.6.1 and obtain




1 w
var
L ()
n n

n
1 X
1
1
cov(|Jn (k1 )|2 , |Jn (k2 )|2 ) = O( ).
2
n
f (k1 )f (k2 )
n


=

k1 ,k2 =1

Thus we have
1 w
P
L () Lw ().
n n
To show equicontinuity we apply the mean value theorem to

1 w
n Ln .

We note that because the

parameters (, ) , have characteristic polynomial whose roots are greater than (1 + ) then
f () is bounded away from zero (there exists a > 0 where inf , f () ). Hence it can be
1 w
shown that there exists a random sequence {Kn } such that | n1 Lw
n ( 1 ) n Ln ( 2 ))| Kn (k 1 2 k)

and Kn converges almost surely to a finite constant as n . Therefore

1
n Ln

is stochastically

equicontinuous. Since the parameter space is compact, the three standard conditions are satisfied
and we have consistency of the Whittle estimator.
To show asymptotic normality we note that
1 w
L () =
n n

log f (k ) +
0

1
n

1 w
n Ln ()
n1
X

r=(n1)


can be written as a quadratic form
n|r|

dn (r; )

Xk Xk+r

k=1

where
n

1X
dn (r; ) =
f (k )1 exp(irk ).
n
k=1

Using the above quadratic form and its derivatives wrt one can show normality of the Whittle
likelihood under various dependence conditions on the time series. Using this result, in the following
theorem we show asymptotic normality of the Whittle estimator. Note, this result not only applies
to linear time series, but several types of nonlinear time series too.
Theorem 9.4.1 Let us suppose that {Xt } is a strictly stationary time series with a sufficient
dependence structure (such as linearity, mixing at a certain rate, etc.) with spectral density function

250

f () and E|Xt4 | < . Let


Lw
n ()

n 
X
k=1


|Jn (k )|2
,
log |f (k )| +
f (k )

bn = arg min Lw ()

= arg min Lw ()

Then we have

 D
bn
n
N (0, 2V 1 + V 1 W V 1 )

where
V



Z 2 
1
f ()
f () 0
d
2 0
f ()
f ()
Z 2 Z 2

0
2
f (1 )1 f (2 )1 f4,0 (1 , 1 , 2 )d1 d2 ,
2
(2) 0
0

and f4,0 (1 , 2 , 3 ) is the fourth order spectrum of {Xt }.


We now apply the above result to the case of linear time series. We now show that in this case,
in the fourth order cumulant term, W , falls out. This is due to the following lemma.
Lemma 9.4.3 Suppose that the spectral density has the form f () = 2 |1 +

j=1 j

exp(ij)|2

and inf f () > 0. Then we have


1
2

log f ()d = log 2

PROOF. Since f (z) is non-zero for |z| 1, then log f (z) has no poles in {z; |z| 1}. Thus we have
Z 2
Z 2

X
1
1
2
log f ()d =
log d +
log |1 +
j exp(ij)|2 d
2
2
0
0
0
j=1
Z 2
Z

X
1
1
log 2 d +
log |1 +
j z|2 dz
2 0
2 |z|=1
j=1
Z 2
1
log 2 d.
2 0
1
2
=
=

An alternative proof is that since f (z) is analytic and does not have any poles for |z| 1, then

251

log f (z) is also analytic in the region |z| 1, thus for |z| 1 we have the power series expansion
P
P
j
2
log |1 +
j=1 bj z (a Taylor expansion about log 1). Using this we have
j=1 j exp(ij)| =
1
2
=

log |1 +
0

X
j=1

1
j exp(ij)| d =
2
2

Z
0

2 X

bj exp(ij)d

j=1

Z 2

1 X
exp(ij)d = 0,
bj
2
0
j=1

and we obtain the desired result.

Lemma 9.4.4 Suppose that {Xt } is a linear ARMA time series Xt

Pp

j=1 j Xtj

Pq

i=1 i ti +

t , where E[t ] = 0, var[t ] = 2 and E[4t ] < . Let = ({j , j }), then we have W = 0 and

 w
D
b
n
N (0, 2V 1 ).
n
PROOF. The result follows from Theorem 9.4.1, however we need to show that in the case of
linearity that W = 0.
We use Example 8.6.1 for linear processes to give f4, (1 , 1 , 2 ) = 4 |A(1 )|2 |A(2 )|2 =
4
f (1 )f (2 ).
4

Substituting this into W gives

1
2

4
4

=
=



0
1
1
f (1 )
f (2 )
f4,0 (1 , 1 , 2 )d1 d2

0
2

2
 Z 2
2
f ()
4 1
f ()
f ()d = 4
d
f ()2
2 0
f ()
0
 Z 2
2
4 1
log f ()d
4 2 0

2

2
Z 2
4
2
4 1

log f ()d = 4 log


= 0,
4 2

2
0


1
2

where by using Lemma 9.4.3 we have

R 2
0

and since does not include


log f ()d = 2 log 2

2 we obtain the above. Hence for linear processes the higher order cumulant does not play an
asymptotic role in the variance thus giving the result.

On first appearances there does not seem to be a connection between the Whittle likelihood
and the sample autocorrelation estimator defined in Section 6.2.1. However, we observe that the
variance of both estimators, under linearity, do not contain the fourth order cumulant (even for
non-Gaussian linear time series). In Section 9.5 we explain there is a connection between the two,

252

and it is this connection that explains away this fourth order cumulant term.
Remark 9.4.2 Under linearity, the GMLE and the Whittle likelihood are asymptotically equivalent, therefore they have the same asymptotic distributions. The GMLE has the asymptotic distri
D
bution n(
, n ) N (0, 1 ), where
n

E(Ut Ut0 ) E(Vt Ut0 )

E(Ut Vt0 ) E(Vt Vt0 )

and {Ut } and {Vt } are autoregressive processes which satisfy (B)Ut = t and (B)Vt = t .
By using the similar derivatives to those given in (7.11) we can show that

E(Ut Ut0 ) E(Vt Ut0 )

= 1
2
E(Ut Vt0 ) E(Vt Vt0 )

9.5

f ()
f ()



f ()
f ()

0
d.

Ratio statistics in Time Series

We recall from (9.4) that the covariance can be written as a general periodogram mean which has
the form
n

1X
In (k )(k ).
A(, In ) =
n

(9.28)

k=1

The variance of this statistic is


var(A(, In )) =

n
1 X
(k1 )(k1 )cov(|Jn (k1 )|2 , |Jn (k2 )|2 )
n2
k1 ,k2 =1

n
1 X
(k1 )(k1 ) cov(Jn (k1 ), Jn (k2 ))cov(Jn (k1 ), Jn (k2 ))
n2
k1 ,k2 =1

+cov(Jn (k1 ), Jn (k2 ))cov(Jn (k1 ), Jn (k2 ))



+cum(Jn (k1 ), Jn (k2 ), Jn (k2 ), Jn (k2 ) .

253

(9.29)

By using Proposition 8.6.1 we have


cov(|Jn (k1 )|2 , |Jn (k2 )|2 ) =
  
 

 2 
1
1
1
+ f (k1 )I(k1 = n k2 ) + O
f (k1 )I(n k1 = k2 ) + O
f (k1 )I(k1 = k2 ) + O
n
n
n
 
1
1
+ f4 (1 , 1 , 2 ) + O
.
(9.30)
n
n2
Substituting (9.30) into (9.29) the above gives
var(A(, In ))
n
n
1 X
1 X
2
2
=
|(
)|
f
(
)
+
(k )(nk )f (k )2
k
k
n2
n2
k=1

k=1

n
X

1
1
(k1 )(k2 )f4 (k1 , k1 , k2 ) + O( 2 )
3
n
n
k1 ,k2 =1
Z 2
Z 2
1
1
|()|2 f ()2 d +
()(2 )f ()2 d
=
n 0
n 0
Z
Z
1 2 2
1
+
(1 )(2 )f4 (1 , 1 , 2 )d1 d2 + O( 2 ),
n 0
n
0
+

(9.31)

where f4 is the fourth order cumulant of {Xt }. From above we see that unless satisfies some
special conditions, var(A(, In )) contains the fourth order spectrum, which can be difficult to
estimate. There are bootstrap methods which can be used to estimate the variance or finite sample
distribution, but simple bootstrap methods, such as the frequency domain bootstrap, cannot be
applied to A(, In ), since it is unable to capture the fourth order cumulant structure. However, in
special cases the fourth order structure is disappears, we consider this case below and then discuss
how this case can be generalised.
Lemma 9.5.1 Suppose {Xt } is a linear time series, with spectral density f (). Let A(, In ) be
defined as in (9.28) and suppose the condition
Z
A(, f ) =

()f ()d = 0

(9.32)

holds, then
1
var(A(, In )) =
n

Z
0

1
|()| f () d +
n
2

254

Z
0

()(2 )f ()2 d.

PROOF. By using (9.31) we have


var(A(, In ))
Z
Z
1 2
1 2
2
2
|()| f () d +
()(2 )f ()2 d
=
n 0
n 0
Z
Z
1 2 2
1
(1 )(2 )f4 (1 , 1 , 2 ) + O( 2 ).
n 0
n
0
But under linearity f4 (1 , 1 , 2 ) =

4
f (1 )f (2 ),
4

substituting this into the above gives

var(A(, In ))
Z
Z
1 2
1 2
2
2
|()| f () d +
()(2 )f ()2 d
=
n 0
n 0
 
Z
Z
4 1 2 2
1
(1 )(2 )f (1 )f (2 )d1 d2 + O
2
4 n 0
n
0
Z 2
Z 2
1
1
=
|()|2 f ()2 d +
()(2 )f ()2 d
n 0
n 0
Z
2
 

4 1 2
1

+ 4
()f ()d + O
.
n 0
n2
Since

()f ()d = 0 we have the desired result.

Example 9.5.1 (The Whittle likelihood) Let us return to the Whittle likelihood in the case of
linearity. In Lemma 9.4.4 we showed that the fourth order cumulant term does not play a role in
the variance of the ARMA estimator. We now show that condition (9.32) holds.
Consider the partial derivative of the Whittle likelihood
Lw
n ()

n 
X
f (k )
k=1

f (k )


In (k )

f (k ) .
f (k )2

To show normality we consider the above at the true parameter , this gives
Lw
n ()

n 
X
f (k )
k=1

f (k )


In (k )

f (k ) .
f (k )2

Only the second term of the above is random, therefore it is only this term that yields the variance.
Let
n

A(f2 f , In ) =

1 X In (k )
f (k ).
n
f (k )2
k=1

255

To see whether this term satisfies the conditions of Lemma 9.5.1 we evaluate
A(f2 f , f )

f ()
f ()
f ()2

=
0

log f ()

=
0

Z
=
0

1
log f () =
2

log f ()d = 0,
0

by using Lemma 9.4.3. Thus we see that the derivative of the Whittle likelhood satisfies the condition
(9.32). Therefore the zero cumulant term is really due to this property.

The Whittle likelihood is a rather special example. However we now show that any statistic of
the form A(, In ) can be transformed such that the resulting transformed statistic satisfies condition
(9.32). To find the suitable transformation we recall from Section 6.2.1 that the variance of b
cn (r)
involves the fourth order cumulant, but under linearity the sample correlation bn (r) = b
cn (r)/b
cn (0)
does given not. Returning to the frequency representation of the autocovariance given in (9.5) we
observe that

bn (r) =

n/2

k=1

k=1

1 1X
1 1X
In (k ) exp(irk ),
In (k ) exp(irk )
b
cn (0) n
b
cn (0) n

(it does not matter whether we sum over n or n/2 for the remainder of this section we choose the
case of summing over n). Motivated by this example we define the so called ratio statistic
n

k=1

k=1

X In (k )(k )
1 X In (k )(k )
e In ) = 1
A(,
=
,
n
b
cn (0)
n
Fn (2)
where Fn (2) =

1
n

Pn

k=1 In (k )

1
n

Pn

2
t=1 Xt

(9.33)

= b
cn (0). We show in the following lemma that

e In ) can be written in a form that almost satisfies condition (9.32).


A(,
e In ) satisfies (9.33) and
Lemma 9.5.2 Let us suppose that A(,
n

X f (k )(k )
e f) = 1
A(,
,
n
Fn (2)
k=1

256

where Fn (2) =

1
n

Pn

In ) as
Then we can represent A(,

j=1 f (k ).

e In ) A(,
e f) =
A(,

1
1X
n (k )In (k ),
F (2)Fn (2) n
k=1

where
n

n (k ) = (k )Fn (2)

1X
(j )f (j )
n

1X
(k )f (k ) = 0.
n

and

j=1

(9.34)

k=1

PROOF. Basic algebra gives


n

In ) A(,
f) =
A(,

1X
n
k=1

=
=

(k )In (k ) (k )f (k )

Fn (2)
Fn (2)


n 
1 X (k )Fn (2)In (k ) (k )Fn (2)f (k )
n
Fn (2)Fn (2)
k=1

n 
n
1X
In (k )
1X
(k )Fn (2)
(k )f (k )
n
n
Fn (2)Fn (2)
k=1
k=1
n

1 X (k )In (k )
,
n
Fn (2)Fn (2)
k=1

where Fn (2) and are defined as above. To show (9.34), again we use basic algebra to give
n

1X
(k )f (k ) =
n
k=1


n 
n
1X
1X
()Fn (2)
(j )f (j ) f (k )
n
n
j=1

k=1

n
n
n
1X
1X
1X
(k )f (k )Fn (2)
(k )f (k )
f (j ) = 0.
n
n
n
k=1

k=1

j=1


In ) A(,
f ) almost seems to satisfy the conditions in
From the lemma above we see that A(,
Lemma 9.5.1, the only difference is the random term b
cn (0) = Fbn (2) in the denominator. We now
show that that we can replace Fbn (2) with its limit and that error is asymptotically negligible.
Let
n

In ) A(,
f) =
A(,

1
1X
e
n (k )In (k ) := B(,
In )
Fn (2)Fn (2) n k=1

257

and
n

B(n , In ) =

1X
1
(k )In (k ).
2
Fn (2) n
k=1

e n , In ) about
By using the mean value theorem (basically the Delta method) and expanding B(
B(n , In ) (noting that B(n , f ) = 0) gives
e
B(,
In ) B(, In )
n

1X
1
1
= Fn (2) Fn (2)
n (k )In (k ) = Op ( ),
3 n
n
{z
} Fn (2)
|
}
| k=1 {z
Op (n1/2 )
Op (n1/2 )

where F n (2) lies between Fn (2) and Fbn (2). Therefore the limiting distribution variance of
In ) A(,
f ) is determined by
A(,
e In ) A(,
e f ) = B(n , In ) + Op (n1/2 ).
A(,
B(n , In ) does satisfy the conditions in (9.32) and the lemma below immediately follows.
Lemma 9.5.3 Suppose that {Xt } is a linear time series, then
var(B(n , In )) =

1
n

1
n

|()|2 f ()2 d +

()(2 )f ()2 d + O(

1
),
n2

where
1
() = ()F (2)
2

()f ()d.
0

e In ) is
Therefore, the limiting variance of A(,
1
n

Z
0

1
|()| f () d +
n
2

()(2 )f ()2 d + O(

1
).
n2

This is a more elegant explanation as to why under linearity the limiting variance of the correlation
estimator does not contain the fourth order cumulant term. It also allows for a general class of
statistics.

258

Remark 9.5.1 (Applications) As we remarked above, many statistics can be written as a ratio statistic. The advantage of this is that the variance of the limiting distribution is only in
terms of the spectral densities, and not any other higher order terms (which are difficult to estimate). Another perk is that simple schemes such as the frequency domain bootstrap can be used
to estimate the finite sample distributions of statistics which satisfy the assumptions in Lemma
9.5.1 or is a ratio statistic (so long as the underlying process is linear), see Dahlhaus and Janas
(1996) for the details. The frequency domain bootstrap works by constructing the DFT from the
data {Jn ()} and dividing by the square root of either the nonparametric estimator of f or a
q
parametric estimator, ie. {Jn ()/ fn ()}, these are close to constant variance random variables.
q
q

{J (k ) = Jn (k )/ fn (k )} is bootstrapped, thus Jn (k ) = J (k ) fn (k ) is used as the bootstrap


DFT. This is used to construct the bootstrap estimator, for example
The Whittle likelihood estimator.
The sample correlation.
With these bootstrap estimators we can construct an estimator of the finite sample distribution.
The nature of frequency domain bootstrap means that the higher order dependence structure is
destroyed, eg. cum (Jn (k1 ), Jn (k2 ), . . . , Jn (kr )) = 0 (where cum is the cumulant with respect
to the bootstrap measure) if all the ki s that are not the same. However, we know from Proposition
8.6.1 that for the actual DFT this is not the case, there is still some small dependence, which
can add up. Therefore, the frequency domain bootstrap is unable to capture any structure beyond
the second order. This means for a linear time series which is not Gaussian the frequency domain
bootstrap cannot approximate the distribution of the sample covariance (since it is asymptotically
with normal with a variance which contains the forth order cumulant), but it can approximate the
finite sample distribution of the correlation.
Remark 9.5.2 (Estimating 4 in the case of linearity) Suppose that {Xt } is a linear time
series
Xt =

j tj ,

j=

with E(t ) = 0, var(t ) = 2 and cum4 (t ) = 4 . Then we can use the spectral density estimator to
estimate 4 without any additional assumptions on {Xt } (besides linearity). Let f () denote the

259

spectral density of {Xt } and g2 () the spectral density of {Xt2 }, then it can be shown that
R 2
2g2 (0) 4 0 f ()2 d
.
4 =
2
R 2
0 f ()d
Thus by estimating f and g2 we can estimate 4 .
Alternatively, we can use the fact that for linear time series, the fourth order spectral density
f4 (1 , 2 , 3 ) = 4 A(1 )A(2 )A(3 )A(1 2 3 ). Thus we have
4 =

4 f4 (1 , 1 , 2 )
.
f (1 )f (2 )

This just demonstrates, there is no unique way to solve a statistical problem!

9.6

Goodness of fit tests for linear time series models

As with many other areas in statistics, we often want to test the appropriateness of a model. In this
section we briefly consider methods for validating whether, say an ARMA(p, q), is the appropriate
model to fit to a time series. One method is to fit the model to the data and the estimate the
residuals and conduct a Portmanteau test (see Section 3, equation (6.9)) on the estimated residuals.
It can be shown that if model fitted to the data is the correct one, the estimated residuals behave
almost like the true residuals in the model and the Portmanteau test statistic
Sh = n

h
X

|
n (r)|2 ,

r=1

where n (r) = cn (r)/


cn (0)
n|r|
1 X
t t+r
cn (r) =
n
t=1

should be asymptotically a chi-squared. An alternative (but somehow equivalent) way to do the


test, is through the DFTs. We recall if the time series is linear then (9.11) is true, thus
IX ()
= |J ()|2 + op (1).
f ()

260

Therefore, if we fit the correct model to the data we would expect that
IX ()
= |J ()|2 + op (1).
f ()
are the model parameter estimators. Now |J ()|2 has the special property that not only is
where
it almost uncorrelated at various frequencies, but it is constant over all the frequencies. Therefore,
we would expect that
n/2

 D
1 X IX ()

2 N (0, 1)
f ()
2 n
k=1

Thus, as an alternative to the goodness fit test based on the portmanteau test statistic we can use
the above as a test statistic, noting that under the alternative the mean would be different.

261

Chapter 10
Consistency and and asymptotic
normality of estimators
In the previous chapter we considered estimators of several different parameters. The hope is that
as the sample size increases the estimator should get closer to the parameter of interest. When we
say closer we mean to converge. In the classical sense the sequence {xk } converges to x (xk x),
if |xk x| 0 as k (or for every > 0, there exists an n where for all k > n, |xk x| < ).
Of course the estimators we have considered are random, that is for every (set of all out
comes) we have an different estimate. The natural question to ask is what does convergence mean
for random sequences.

10.1

Modes of convergence

We start by defining different modes of convergence.


Definition 10.1.1 (Convergence)

Almost sure convergence We say that the sequence

{Xt } converges almost sure to , if there exists a set M , such that P(M ) = 1 and for
every N we have
Xt () .

262

In other words for every > 0, there exists an N () such that


|Xt () | < ,

(10.1)

for all t > N (). Note that the above definition is very close to classical convergence. We
a.s.

denote Xt almost surely, as Xt .


a.s.

An equivalent definition, in terms of probabilities, is for every > 0 Xt if

P (;
m=1 t=m {|Xt () | > }) = 0.

It is worth considering briefly what


m=1 t=m {|Xt () | > } means. If m=1 t=m

{|Xt () | > } =
6 , then there exists an
m=1 t=m {|Xt () | > } such that for

some infinite sequence {kj }, we have |Xkj ( ) | > , this means Xt ( ) does not converge

to . Now let
m=1 t=m {|Xt () | > } = A, if P (A) = 0, then for most the sequence

{Xt ()} converges.


Convergence in mean square
We say Xt in mean square (or L2 convergence), if E(Xt )2 0 as t .
Convergence in probability
Convergence in probability cannot be stated in terms of realisations Xt () but only in terms
P

of probabilities. Xt is said to converge to in probability (written Xt ) if


P (|Xt | > ) 0,

t .

Often we write this as |Xt | = op (1).


If for any 1 we have
E(Xt ) 0

t ,

then it implies convergence in probability (to see this, use Markovs inequality).
Rates of convergence:
(i) Suppose at 0 as t . We say the stochastic process {Xt } is |Xt | = Op (at ),

263

if the sequence {a1


t |Xt |} is bounded in probability (this is defined below). We see
from the definition of boundedness, that for all t, the distribution of a1
t |Xt | should
mainly lie within a certain interval.
(ii) We say the stochastic process {Xt } is |Xt | = op (at ), if the sequence {a1
t |Xt |}
converges in probability to zero.
Definition 10.1.2 (Boundedness)

(i) Almost surely bounded If the random variable X

is almost surely bounded, then for a positive sequence {ek }, such that ek as k
(typically ek = 2k is used), we have
P (; {
k=1 {|X()| ek }}) = 1.
Usually to prove the above we consider the complement
c
P ((; {
k=1 {|X| ek }}) ) = 0.

Since (
k=1 {|X| ek }) = k=1 {|X| > ek } k=1 m=k {|X| > ek }, to show the above we

show

P ( : {
k=1 m=k {|X()| > ek }}) = 0.

(10.2)

We note that if ( : {
k=1 m=k {|X()| > ek }}) 6= , then there exists a and an

infinite subsequence kj , where |X( )| > ekj , hence X( ) is not bounded (since ek ). To
P
prove (10.2) we usually use the Borel Cantelli Lemma. This states that if
k=1 P (Ak ) < ,
the events {Ak } occur only finitely often with probability one. Applying this to our case,
P
if we can show that
m=1 P ( : {|X()| > em |}) < , then {|X()| > em |} happens
P
only finitely often with probability one. Hence if
m=1 P ( : {|X()| > em |}) < , then

P ( : {
k=1 m=k {|X()| > ek }}) = 0 and X is a bounded random variable.

P
It is worth noting that often we choose the sequence ek = 2k , in this case
m=1 P ( :
P
{|X()| > em |}) = m=1 P ( : {log |X()| > log 2k |}) CE(log |X|). Hence if we can
show that E(log |X|) < , then X is bounded almost surely.
b
(ii) Sequences which are bounded in probability A sequence is bounded in probability,

264

written Xt = Op (1), if for every > 0, there exists a () < such that P (|Xt | ()) < .
Roughly speaking this means that the sequence is only extremely large with a very small
probability. And as the largeness grows the probability declines.

10.2

Sampling properties

Often we will estimate the parameters by maximising (or minimising) a criterion. Suppose we have
the criterion Ln (a) (eg. likelihood, quasi-likelihood, Kullback-Leibler etc) we use as an estimator
of a0 , a
n where
a
n = arg max Ln (a)
a

and is the parameter space we do the maximisation (minimisation) over. Typically the true
parameter a should maximise (minimise) the limiting criterion L.
If this is to be a good estimator, as the sample size grows the estimator should converge (in
some sense) to the parameter we are interesting in estimating. As we discussed above, there are
various modes in which we can measure this convergence (i) almost surely (ii) in probability and
(iii) in mean squared error. Usually we show either (i) or (ii) (noting that (i) implies (ii)), in time
series its usually quite difficult to show (iii).
Definition 10.2.1

(i) An estimator a
n is said to be almost surely consistent estimator of a0 , if

there exists a set M , where P(M ) = 1 and for all M we have


a
n () a.

(ii) An estimator a
n is said to converge in probability to a0 , if for every > 0
P (|
an a| > ) 0

T .

To prove either (i) or (ii) usually involves verifying two main things, pointwise convergence and
equicontinuity.

265

10.3

Showing almost sure convergence of an estimator

We now consider the general case where Ln (a) is a criterion which we maximise. Let us suppose
we can write Ln as
n

Ln (a) =

1X
`t (a),
n

(10.3)

t=1

where for each a , {`t (a)}t is a ergodic sequence. Let


L(a) = E(`t (a)),

(10.4)

we assume that L(a) is continuous and has a unique maximum in . We define the estimator
n
where
n = arg mina Ln (a).
Definition 10.3.1 (Uniform convergence) Ln (a) is said to almost surely converge uniformly
to L(a), if
a.s.

sup |Ln (a) L(a)| 0.


a

In other words there exists a set M where P (M ) = 1 and for every M ,


sup |Ln (, a) L(a)| 0.
a

Theorem 10.3.1 (Consistency) Suppose that a


n = arg maxa Ln (a) and a0 = arg maxa L(a)
a.s.

is the unique maximum. If supa |Ln (a) L(a)| 0 as n and L(a) has a unique maximum.
a.s.

Then Then a
n a0 as n .
PROOF. We note that by definition we have Ln (a0 ) Ln (
an ) and L(
an ) L(a0 ). Using this
inequality we have
Ln (a0 ) L(a0 ) Ln (
an ) L(a0 ) Ln (
an ) L(
an ).
Therefore from the above we have
|Ln (
aT ) L(a0 )| max {|Ln (a0 ) L(a0 )|, |Ln (
aT ) L(
an )|} sup |Ln (a) L(a)|.
a

266

a.s.

Hence since we have uniform converge we have |Ln (


an ) L(a0 )| 0 as n . Now since L(a)
a.s.

a.s.

has a unique maximum, we see that |Ln (


an ) L(a0 )| 0 implies a
n a0 .

We note that directly establishing uniform convergence is not easy. Usually it is done by assuming the parameter space is compact and showing point wise convergence and stochastic equicontinuity, these three facts imply uniform convergence. Below we define stochastic equicontinuity and
show consistency under these conditions.
Definition 10.3.2 The sequence of stochastic functions {fn (a)}n is said to be stochastically equicontinuous if there exists a set M where P (M ) = 1 and for every M and and > 0, there
exists a and such that for every M
sup

|fn (, a1 ) fn (, a2 )| ,

|a1 a2 |

for all n > N ().


A sufficient condition for stochastic equicontinuity of fn (a) (which is usually used to prove
equicontinuity), is that fn (a) is in some sense Lipschitz continuous. In other words,
sup |fn (a1 ) fn (a2 )| < Kn ka1 a2 k,
a1 ,a2
a.s.

where kn is a random variable which converges to a finite constant as n (Kn K0 as


a.s.

n ). To show that this implies equicontinuity we note that Kn K0 means that for every
M (P (M ) = 1) and > 0, we have |Kn () K0 | < for all n > N (). Therefore if we
choose = /(K0 + ) we have
sup

|fn (, a1 ) fn (, a2 )| < ,

|a1 a2 |/(K0 +)

for all n > N ().


In the following theorem we state sufficient conditions for almost sure uniform convergence. It
is worth noting this is the Arzela-Ascoli theorem for random variables.
Theorem 10.3.2 (The stochastic Ascoli Lemma) Suppose the parameter space is compact,
a.s.

for every a we have Ln (a) L(a) and Ln (a) is stochastic equicontinuous. Then supa |Ln (a)
a.s.

L(a)| 0 as n .

267

We use the theorem below.


Corollary 10.3.1 Suppose that a
n = arg maxa Ln (a) and a0 = arg maxa L(a), moreover L(a)
has a unique maximum. If
a.s.

(i) We have point wise convergence, that is for every a we have Ln (a) L(a).
(ii) The parameter space is compact.
(iii) Ln (a) is stochastic equicontinuous.
a.s.

then a
n a0 as n .
PROOF. By using Theorem 10.3.2 three assumptions imply that | sup ||Ln () L()| 0, thus
by using Theorem 10.3.1 we obtain the result.
We prove Theorem 10.3.2 in the section below, but it can be omitted on first reading.

10.3.1

Proof of Theorem 10.3.2 (The stochastic Ascoli theorem)

We now show that stochastic equicontinuity and almost pointwise convergence imply uniform convergence. We note that on its own, pointwise convergence is a much weaker condition than uniform
convergence, since for pointwise convergence the rate of convergence can be different for each parameter.
Before we continue a few technical points. We recall that we are assuming almost pointwise
convergence. This means for each parameter a there exists a set Na (with P (Na ) = 1)
such that for all Na Ln (, a) L(a). In the following lemma we unify this set. That is show
(using stochastic equicontinuity) that there exists a set N (with P (N ) = 1) such that for all
N Ln (, a) L(a).
Lemma 10.3.1 Suppose the sequence {Ln (a)}n is stochastically equicontinuous and also pointwise
convergent (that is Ln (a) converges almost surely to L(a)), then there exists a set M where
) = 1 and for every M
and a we have
P (M
|Ln (, a) L(a)| 0.
PROOF. Enumerate all the rationals in the set and call this sequence {ai }i . Since we have almost
sure convergence, this implies for every ai there exists a set Mai where P (Mai ) = 1 and for every

268

Mai we have |LT (, ai ) L(ai )| 0. Define M = Mai , since the number of sets is countable
P (M ) = 1 and for every M and ai we have Ln (, ai ) L(ai ).
where P (M
) = 1 and for every
Since we have stochastic equicontinuity, there exists a set M
, {Ln (, )} is equicontinuous. Let M
=M
{Ma }, we will show that for all a and
M
i
we have Ln (, a) L(a). By stochastic equicontinuity for every M
and /3 > 0, there
M
exists a > 0 such that
|Ln (, b1 ) Ln (, b2 )| /3,

sup

(10.5)

|b1 b2 |

for every rational aj and N


we have
for all n > N (). Furthermore by definition of M
|Ln (, ai ) L(ai )| /3,

(10.6)

where n > N 0 (). Now for any given a , there exists a rational ai such that ka aj k . Using
this, (10.5) and (10.6) we have
|Ln (, a) L(a)| |Ln (, a) Ln (, ai )| + |Ln (, ai ) L(ai )| + |L(a) L(ai )| ,
and a , we have |Ln (, a) L(a)|
for n > max(N (), N 0 ()). To summarise for every M
.
0. Hence we have pointwise covergence for every realisation in M

We now show that equicontinuity implies uniform convergence.


with
Proof of Theorem 10.3.2. Using Lemma 10.3.1 we see that there exists a set M
) = 1, where Ln is equicontinuous and also pointwise convergent. We now show uniform
P (M
we have
convergence on this set. Choose /3 > 0 and let be such that for every M
sup

|LT (, a1 ) LT (, a2 )| /3,

(10.7)

|a1 a2 |

for all n > n(). Since is compact it can be divided into a finite number of open sets. Construct
the sets {Oi }pi=1 , such that pi=1 Oi and supx,y,i kx yk . Let {ai }pi=1 be such that ai Oi .
we have Ln (, ai ) L(ai ), hence for every /3, there exists an
We note that for every M
ni () such that for all n > ni () we have |LT (, ai ) L(ai )| /3. Therefore, since p is finite (due

269

to compactness), there exists a n


() such that
max |Ln (, ai ) L(ai )| /3,

1ip

for all n > n


() = max1ip (ni ()). For any a , choose the i, such that open set Oi such that
a Oi . Using (10.7) we have
|LT (, a) LT (, ai )| /3,
for all n > n(). Altogether this gives
|LT (, a) L(a)| |LT (, a) LT (, ai )| + |LT (, ai ) L(ai )| + |L(a) L(ai )| ,
for all n max(n(), n
()). We observe that max(n(), n
()) and /3 does not depend on a,
therefore for all n max(n(), n
()) and we have supa |Ln (, a) L(a)| < . This gives for every
(P(M
) = 1), supa |Ln (, a) L(a)| 0, thus we have almost sure uniform convergence. 
M

10.4

Toy Example: Almost sure convergence of the


least squares estimator for an AR(p) process

In Chapter ?? we will consider the sampling properties of many of the estimators defined in Chapter
6. However to illustrate the consistency result above we apply it to the least squares estimator of
the autoregressive parameters.
To simply notation we only consider estimator for AR(1) models. Suppose that Xt satisfies
Xt = Xt1 + t (where || < 1). To estimate we use the least squares estimator defined below.
Let
n

Ln (a) =

1 X
(Xt aXt1 )2 ,
n1

(10.8)

t=2

we use n as an estimator of , where


n = arg min LT (a),
a

270

(10.9)

where = [1, 1].


How can we show that this is consistent?
In the case of least squares for AR processes, a
T has the explicit form
n =

1 Pn
t=2 Xt Xt1
n1
.
P
T 1
1
2
t=1 Xt
n1
a.s.

By just applying the ergodic theorem to the numerator and denominator we get n .
1 Pnt=2 Xt Xt1
< 1 is not necesIt is worth noting, that unlike the Yule-Walker estimator n11 P
n1
X2
n1

t=1

sarily true.
Here we will tackle the problem in a rather artifical way and assume that it does not have an
explicit form and instead assume that n is obtained by minimising Ln (a) using a numerical
routine.
In order to derive the sampling properties of n we need to directly study the least squares
criterion Ln (a). We will do this now in the least squares case.
We will first show almost sure convergence, which will involve repeated use of the ergodic
theorem. We will then demonstrate how to show convergence in probability. We look at almost
sure convergence as its easier to follow. Note that almost sure convergence implies convergence in
probability (but the converse is not necessarily true).
The first thing to do it let
`t (a) = (Xt aXt1 )2 .
Since {Xt } is an ergodic process (recall Example ??(ii)) by using Theorem ?? we have for a, that
{`t (a)}t is an ergodic process. Therefore by using the ergodic theorem we have
n

1 X
a.s.
Ln (a) =
`t (a) E(`0 (a)).
n1
t=2

a.s.

In other words for every a [1, 1] we have that Ln (a) E(`0 (a)) (almost sure pointwise convergence).
Since the parameter space = [1, 1] is compact and a is the unique minimum of `() in the

271

parameter space, all that remains is to show show stochastic equicontinuity. From this we deduce
almost sure uniform convergence.
To show stochastic equicontinuity we expand LT (a) and use the mean value theorem to obtain
Ln (a1 ) Ln (a2 ) = LT (
a)(a1 a2 ),

(10.10)

where a
[min[a1 , a2 ], max[a1 , a2 ]] and
n

Ln (
a) =

2 X
Xt1 (Xt a
Xt1 ).
n1
t=2

Because a
[1, 1] we have
n

2 X
2
|Ln (
a)| Dn , where Dn =
(|Xt1 Xt | + Xt1
).
n1
t=2

2 } is an ergodic process. Therefore, if


Since {Xt }t is an ergodic process, then {|Xt1 Xt | + Xt1

var(t ) < , by using the ergodic theorem we have


a.s.

2
Dn 2E(|Xt1 Xt | + Xt1
).

2 ). Therefore there exists a set M , where P(M ) = 1 and for every


Let D := 2E(|Xt1 Xt | + Xt1

M and > 0 we have


|DT () D| ,
for all n > N (). Substituting the above into (10.10) we have
|Ln (, a1 ) Ln (, a2 )| Dn ()|a1 a2 | (D + )|a1 a2 |,
for all n N (). Therefore for every > 0, there exists a := /(D + ) such that
sup
|a1 a2 |/(D+ )

|Ln (, a1 ) Ln (, a2 )| ,

for all n N (). Since this is true for all M we see that {Ln (a)} is stochastically equicontinuous.

272

a.s.
Theorem 10.4.1 Let n be defined as in (10.9). Then we have n .

PROOF. Since {Ln (a)} is almost sure equicontinuous, the parameter space [1, 1] is compact and
a.s.
a.s.
we have pointwise convergence of Ln (a) L(a), by using Theorem 10.3.1 we have that n a,

where a = mina L(a). Finally we need to show that a = . Since


L(a) = E(`0 (a)) = E(X1 aX0 )2 ,
we see by differentiating L(a) with respect to a, that it is minimised at a = E(X0 X1 )/E(X02 ), hence
a = E(X0 X1 )/E(X02 ). To show that this is , we note that by the Yule-Walker equations
Xt = Xt1 + t

2
E(Xt Xt1 ) = E(Xt1
) + E(t Xt1 ) .
| {z }
=0

a.s.
Therefore = E(X0 X1 )/E(X02 ), hence n .

We note that by using a very similar methods we can show strong consistency of the least
squares estimator of the parameters in an AR(p) model.

10.5

Convergence in probability of an estimator


a.s.

We described above almost sure (strong) consistency (


aT a0 ). Sometimes its not possible to show
strong consistency (eg. when ergodicity cannot be verified). As an alternative, weak consistency
P

where a
T a0 (convergence in probability), is shown. This requires a weaker set of conditions,
which we now describe:
(i) The parameter space should be compact.
P

(ii) Probability pointwise convergence: for every a Ln (a) L(a).


(iii) The sequence {Ln (a)} is equicontinuous in probability. That is for every  > 0 and > 0
there exists a such that
!
lim P

sup

|Ln (a1 ) Ln (a2 )| > 

|a1 a2 |

If the above conditions are satisified we have a


T a0 .

273

< .

(10.11)

Verifying conditions (ii) and (iii) may look a little daunting but by using Chebyshevs (or
Markovs) inequality it can be quite straightforward. For example if we can show that for every
a
E(Ln (a) L(a))2 0

T .

Therefore by applying Chebyshevs inequality we have for every > 0 that


E(Ln (a) L(a))2
0
2

P (|Ln (a) L(a)| > )

T .

Thus for every a we have Ln (a) L(a).


To show (iii) we often use the mean value theorem Ln (a). Using the mean value theorem we
have
|Ln (a1 ) Ln (a2 )| sup ka Ln (a)k2 ka1 a2 k.
a

Now if we can show that supn E supa ka Ln (a)k2 < (in other words it is uniformly bounded in
probability over n) then we have the result. To see this observe that
!
sup

|Ln (a1 ) Ln (a2 )| > 



P sup ka Ln (a)k2 |a1 a2 | > 
a

|a1 a2 |

supn E(|a1 a2 | supa ka Ln (a)k2 )


.


Therefore by a careful choice of > 0 we see that (10.11) is satisfied (and we have equicontinuity
in probability).

10.6

Asymptotic normality of an estimator

Once consistency of an estimator has been shown this paves the way to showing normality. To
make the derivations simple we will assume that is univariate (this allows to easily use Taylor
expansion) . We will assume that that the third derivative of the contrast function, Ln (), exists,
its expectation is bounded and its variance converges to zero as n . If this is the case we have
have the following result

274

Lemma 10.6.1 Suppose that the third derivative of the contrast function Ln () exists, for k =
k

Ln ()
0, 1, 2 E( theta
k ) =

k L
k

Ln ()
and var( theta
k ) 0 as n and

3 Ln ()
theta3

is bounded by a random

variable Zn which is independent of n where E(Zn ) < and var(Zn ) 0. Then we have
(n 0 ) = V ()1

where V (0 ) =

Ln ()
Ln ()
+ op (1)
,
=0
=0

2 L()
.
2 0

PROOF. By the mean value theorem we have


Ln ()
=0

2
Ln ()
2 Ln ()
n 0 ) Ln ()
=
(

(10.12)
(n 0 )
=n
2 =n
2 =n

where n lies between 0 and n . We first study

2 Ln ()
.
2 =n

By using the man value theorem we

have
2
2 Ln ()
2 Ln ()
n 0 ) Ln ()
=
+
(

2 =n
2 0
2 =n

where n lies between 0 and n . Since

2 Ln ()
2 0

2 L()
2 0

= V (0 ), under the stated assumptions

we have

2

2L
|n 0 | Ln ()
|n 0 |Wn .

V
(
)
0
2
2
=n

=n
Therefore, by consistency of the estimator it is clear that

P
2L

2 =n

V (0 ). Substituting this into

(10.12) we have
L
=0

= (n 0 )(V (0 ) + o(1)),
2

since V (0 ) is bounded away from zero we have [ L2 = ]1 = V (0 )1 + op (1) and we obtain the
n

desired result.

The above result means that the distribution of (n 0 ) is determined by


following section we show to show asymptotic normality of

275

L
=0 .

L
=0 .

In the

10.6.1

Martingale central limit theorem

The first central limit theorm goes back to the asymptotic distribution of sums of binary random
variables (these have a binomial distribution and Bernoulli showed that they could be approximated
to a normal distribution). This result was later generalised to sums of iid random variables. However
from mid 20th century to late 20th century several advances have been made for generalisating the
results to dependent random variables. These include generalisations to random variables which
have n-dependence, mixing properties, cumulant properties, near-epoch dependence etc (see, for
example, Billingsley (1995) and Davidson (1994)). In this section we will concentrate on a central
limit theore for martingales. Our reason for choosing this flavour of CLT is that it can be applied
in various estimation settings - as it can often be shown that the derivative of a criterion at the
true parameter is a martingale.
Let us suppose that
n

1X
Sn =
Zt ,
n
t=1

we shall show asymptotic normality of

n(Sn E(Sn )). The reason for normalising by

n, is

a.s.

that (Sn E(Sn )) 0 as n , hence in terms of distributions it converges towards the point
mass at zero. Therefore we need to increase the magnitude of the difference. It it can show that

var(Sn ) = O(n1 ), then n(Sn E(S0 ) = O(1).


Definition 10.6.1 The random variables {Zt } are called martingale differences if
E(Zt |Zt1 , Zt2 , . . .) = 0.
The sequence {ST }T , where

ST =

T
X

Zt

k=1

are called martingales if {Zt } are martingale differences.


Remark 10.6.1 (Martingales and covariances) We observe that if {Zt } are martingale dif-

276

ferences then if t > s and Fs = (Zs , Zs1 , . . .)




cov(Zs , Zt ) = E(Zs Zt ) = E E(Zs Zt |Fs ) = E Zs E(Zt |Fs ) = E(Zs 0) = 0.
Hence martingale differences are uncorrelated.
Example 10.6.1 Suppose that Xt = Xt1 + t , where {t } are idd rv with E(t ) = 0 and || < 1.
Then {t Xt1 }t are martingale differences.
Let us define ST as

ST =

T
X

Zt ,

(10.13)

t=1

where Ft = (Zt , Zt1 , . . .), E(Zt |Ft1 ) = 0 and E(Zt2 ) < . In the following theorem adapted
from Hall and Heyde (1980), Theorem 3.2 and Corollary 3.1, we show that ST is asymptotically
normal.
Theorem 10.6.1 Let {ST }T be defined as in (10.62). Further suppose
T
1X 2 P 2
Zt ,
T

(10.14)

t=1

where 2 is a finite constant, for all > 0,


T

1X
P
E(Zt2 I(|Zt | > T )|Ft1 ) 0,
T

(10.15)

t=1

(this is known as the conditional Lindeberg condition) and


T
1X
P
E(Zt2 |Ft1 ) 2 .
T

(10.16)

t=1

Then we have
D

T 1/2 ST N (0, 2 ).

277

(10.17)

10.6.2

Example: Asymptotic normality of the least squares estimator

In this section we show asymptotic normality of the least squares estimator of the AR(1) (Xt =
Xt1 + t , with var(t ) = 2 ) defined in (10.8).
We call that the least squares estimator is n = arg maxa[1,1] Ln (a). Recalling the criterion
n

1 X
Ln (a) =
(Xt aXt1 )2 ,
n1
t=2

the first and the second derivative is


Ln (a) =
and 2 Ln (a) =

t=2

t=2

2 X
2 X
Xt1 (Xt aXt1 ) =
Xt1 t
n1
n1
2
n1

n
X

2
Xt1
.

t=2

Therefore by using (??) we have


(n ) = 2 Ln

1

Ln ().

(10.18)
a.s.

Since {Xt2 } are ergodic random variables, by using the ergodic theorem we have 2 Ln 2E(X02 ).
This with (10.18) implies

n(n ) = 2 Ln
| {z

1

nLn ().

(10.19)

a.s.

(2E(X02 ))1

To show asymptotic normality of

n(n ), will show asymptotic normality of

nLn ().

We observe that
n

2 X
Ln () =
Xt1 t ,
n1
t=2

is the sum of martingale differences, since E(Xt1 t |Xt1 ) = Xt1 E(t |Xt1 ) = Xt1 E(t ) = 0
(here we used Definition 10.6.1). In order to show asymptotic of Ln () we will use the martingale
central limit theorem.
We now use Theorem 10.6.1 to show that

nLn () is asymptotically normal, which means we

278

have to verify conditions (10.14)-(10.16). We note in our example that Zt := Xt1 t , and that the
series {Xt1 t }t is an ergodic process. Furthermore, since for any function g, E(g(Xt1 t )|Ft1 ) =
E(g(Xt1 t )|Xt1 ), where Ft = (Xt , Xt1 , . . .) we need only to condition on Xt1 rather than the
entire sigma-algebra Ft1 .
C1 : By using the ergodicity of {Xt1 t }t we have
n

t=1

t=1

1X 2
1X 2 2 P
2
Zt =
Xt1 t E(Xt1
) E(2t ) = 2 c(0).
| {z }
n
n
=1

C2 : We now verify the conditional Lindeberg condition.


n
n

1X
1X
2
E(Zt2 I(|Zt | > n)|Ft1 ) =
E(Xt1
2t I(|Xt1 t | > n)|Xt1 )
n
n
t=1

t=1

2 2 and
We now use the Cauchy-Schwartz inequality for conditional expectations to split Xt1
t

I(|Xt1 t | > ). We recall that the Cauchy-Schwartz inequality for conditional expectations
is E(Xt Zt |G) [E(Xt2 |G)E(Zt2 |G)]1/2 almost surely. Therefore
n

1X
E(Zt2 I(|Zt | > n)|Ft1 )
n
t=1

n
1/2

1 X
4
E(Xt1
4t |Xt1 )E(I(|Xt1 t | > n)2 |Xt1 )
n

n

1/2

1X 2
Xt1 E(4t )1/2 E(I(|Xt1 t | > n)2 |Xt1 )
.
n

t=1

(10.20)

t=1

We note that rather than use the Cauchy-Schwartz inequality we can use a generalisation
of it called the H
older inequality. The Holder inequality states that if p1 + q 1 = 1, then
E(XY ) {E(X p )}1/p {E(Y q )}1/q (the conditional version also exists). The advantage of
using this inequality is that one can reduce the moment assumptions on Xt .
Returning to (10.20), and studying E(I(|Xt1 t | > )2 |Xt1 ) we use that E(I(A)) = P(A)
and the Chebyshev inequality to show

E(I(|Xt1 t | > n)2 |Xt1 ) = E(I(|Xt1 t | > n)|Xt1 )

= E(I(|t | > n/Xt1 )|Xt1 )

X 2 var(t )
n
= P (|t | >
)) t1 2
.
Xt1
n

279

(10.21)

Substituting (10.21) into (10.20) we have

1X
E(Zt2 I(|Zt | > n)|Ft1 )
n
t=1
 2
1/2
n
X
1
2
4 1/2 Xt1 var(t )
Xt1 E(t )
n
2 n
t=1
E(4t )1/2
n3/2

n
X

|Xt1 |3 E(2t )1/2

t=1
4
1/2
E(t ) E(2t )1/2
n1/2

1X
|Xt1 |3 .
n
t=1

If E(4t ) < , then E(Xt4 ) < , therefore by using the ergodic theorem we have

1
n

3 a.s.
t=1 |Xt1 |

Pn

E(|X0 |3 ). Since almost sure convergence implies convergence in probability we have


n
n

E(4t )1/2 E(2t )1/2 1 X


1X
E(Zt2 I(|Zt | > n)|Ft1 )
|Xt1 |3
1/2
n
n
n
|
{z
} t=1
t=1
|
{z
}
0
P

E(|X0 |3 )
P

0.
Hence condition (10.15) is satisfied.
C3 : We need to verify that
n

1X
P
E(Zt2 |Ft1 ) 2 .
n
t=1

Since {Xt }t is an ergodic sequence we have


n

t=1

t=1

1X
1X
2
E(Zt2 |Ft1 ) =
E(Xt1
2 |Xt1 )
n
n
=

n
n
X
1X 2
2
2 1
2
Xt1 E( |Xt1 ) = E( )
Xt1
n
n
t=1
| t=1{z }
a.s.

E(X02 )

E(2 )E(X02 ) = 2 c(0),


hence we have verified condition (10.16).

280

Altogether conditions C1-C3 imply that

1 X
D
nLn () =
Xt1 t N (0, 2 c(0)).
n t=1

Recalling (10.19) and that

(10.22)

D
nLn () N (0, 2 ) we have

n(n ) = 2 Ln
| {z
a.s.

1
nLn () .
{z
}
} |

(10.23)

(2E(X02 ))1 N (0, 2 c(0))

Using that E(X02 ) = c(0), this implies that

1
D
n(n ) N (0, 2 c(0)1 ).
4

(10.24)

Thus we have derived the limiting distribution of n .


Remark 10.6.2 We recall that
(n ) = 2 Ln
2
and that var( n1

Pn

t=2 t Xt1 )

2
n1

1

Ln () =

Pn

t=2 var(t Xt1 )

2 Pn
t=2 t Xt1
n1
,
2 Pn
2
t=2 Xt1
n1

(10.25)

= O( n1 ). This implies

(n ) = Op (n1/2 ).
Indeed the results also holds almost surely
(n ) = O(n1/2 ).

(10.26)

The same result is true for autoregressive processes of arbitrary finite order. That is

10.6.3

) N (0, E(p )1 2 ).
n(
n

(10.27)

Example: Asymptotic normality of the weighted periodogram

Previously we have discussed the weight peiodogram, here we show normality of it, in the case
P
that the time series Xt is zero mean linear time series (has the representation Xt = j j tj ).

281

Recalling Lemma 9.2.2 we have


n

A(, In ) =

1X
(k )In (k )
n
k=1
n

1
1X
(k )|A(k )2 |I (k ) + o( ).
n
n
k=1

Therefore we will show asymptotic normality of

1
n

Pn

2
k=1 (k )|A(k ) |I (k ),

which will give asymp-

totic normality of A(, In ). Expanding |I (k ) and substituting this into


1 Pn
2
k=1 (k )|A(k ) |I (k ) gives
n
n

1X
(k )|A(k )2 |I (k ) =
n
k=1

n
n
n
1 X
1X
1 X
2
t
(k )|A(k ) exp(ik (t )) =
t gn (t )
n
n
n
t, =1

t, =1

k=1

where
n

1X
1
gn (t ) =
(k )|A(k )2 exp(ik (t )) =
n
2
k=1

()|A()|2 exp(i(t ))d + O(

1
),
n2

(the rate for the derivative exchange is based on assuming that the second derivatives of A() and
P
exist and (0) = (2)). We can rewrite n1 nt, =1 t gn (t ) as

n
1 X
[t E(t )]gn (t )
n
t, =1

n 
X

1X
2
2
[(t E(t )]gn (0) + t
[gn (t ) gn ( t)]
n
<t
t=1

:=

n
1X
Zt,n
n
t=1

where it is straightforward to show that {Zt,n } are the sum of martingale differences. Thus we can
show that
n
n
n

1 X
1 X
1 X

t gn (t ) E
t gn (t ) =
Zt,n
n t, =1
n t, =1
n t=1

satisfies the conditions of the martingale central limit theorem, which gives asymptotic normality
P
of n1 nt, =1 t gn (t ) and thus A(, In ).
In the remainder of this chapter we obtain the sampling properties of the ARMA estimators

282

defined in Sections 7.2.1 and 7.2.2.

10.7

Asymptotic properties of the Hannan and Rissanen estimation method

In this section we will derive the sampling properties of the Hannan-Rissanen estimator. We will
obtain an almost sure rate of convergence (this will be the only estimator where we obtain an almost
sure rate). Typically obtaining only sure rates can be more difficult than obtaining probabilistic
rates, moreover the rates can be different (worse in the almost sure case). We now illustrate why
that is with a small example. Suppose {Xt } are iid random variables with mean zero and variance
P
one. Let Sn = nt=1 Xt . It can easily be shown that
var(Sn ) =

1
1
therefore Sn = Op ( ).
n
n

(10.28)

However, from the law of iterated logarithm we have for any > 0
p
p
1.
P (Sn (1 + ) 2n log log n infinitely often) = 0P (Sn (1 ) 2n log log n infinitely often) =(10.29)
Comparing (10.28) and (10.29) we see that for any given trajectory (realisation) most of the time
1
n Sn

n
will be within the O( 1n ) bound but there will be excursions above when it to the O( loglog
n

bound. In other words we cannot say that

1
n Sn

= ( 1n ) almost surely, but we can say that This

basically means that

1
2 log log n

Sn = O(
) almost surely.
n
n
Hence the probabilistic and the almost sure rates are (slightly) different. Given this result is true
for the average of iid random variables, it is likely that similar results will hold true for various
estimators.
In this section we derive an almost sure rate for Hannan-Rissanen estimator, this rate will
be determined by a few factors (a) an almost sure bound similar to the one derived above (b)
the increasing number of parameters pn (c) the bias due to estimating only a finite number of
parameters when there are an infinite number in the model.

283

We first recall the algorithm:


n
(i) Use least squares to estimate {bj }pj=1
and define

n = R
1 rn ,
b
n

(10.30)

0 = (b1,n , . . . , bp ,n ),
where b
n
n
n =
R

n
X

Xt1 X0t1

T
X

rn =

t=pn +1

Xt Xt1

t=pn +1

and X0t1 = (Xt1 , . . . , Xtpn ).


(ii) Estimate the residuals with

t = Xt

pn
X

bj,n Xtj .

j=1

, where
(iii) Now use as estimates of 0 and 0
n n
, = arg min

n n

n
X

(Xt

t=pn +1

p
X

j Xtj

j=1

q
X

i ti )2 .

(10.31)

i=1

We note that the above can easily be minimised. In fact


, ) = R
1
sn
(
n
n n
where
n
X
tY
t
n = 1
R
Y
n
t=pn +1

sn =

1
T

n
X

t Xt ,
Y

t=pn +1

0 = (Xt1 , . . . , Xtp , t1 , . . . , tq ). Let n = (n , n ).


Y
t
We observe that in the second stage of the scheme where the estimation of the ARMA parameters
are done, it is important to show that the empirical residuals are close to the true residuals. That
is t = t + o(1). We observe that from the definition of t , this depends on the rate of convergence

284

of the AR estimators bj,n

t = Xt

= t +

pn
X

bj,n Xtj

j=1
p
n
X

j=1

j=pn +1

(bj,n bj )Xtj

bj Xtj .

(10.32)

Hence
pn


X
X



t t
bj Xtj .
(bj,n bj )Xtj +

(10.33)

j=pn +1

j=1

, we need to
=
Therefore to study the asymptotic properties of
n n
Obtain a rate of convergence for supj |bj,n bj |.
Obtain a rate for |
t t |.
, ).
n = (
Use the above to obtain a rate for
n n
We first want to obtain the uniform rate of convergence for supj |bj,n bj |. Deriving this is
technically quite challanging. We state the rate in the following theorem, an outline of the proof
can be found in Section 10.7.1. The proofs uses results from mixingale theory which can be found
in Chapter B.
Theorem 10.7.1 Suppose that {Xt } is from an ARMA process where the roots of the true char n be defined
acteristic polynomials (z) and (z) both have absolute value greater than 1 + . Let b
as in (10.30), then we have almost surely
r
n bn k2 = O p2
kb
n

(log log n)1+ log n p3n


+
+ pn pn
n
n

for any > 0.


PROOF. See Section 10.7.1.

Corollary 10.7.1 Suppose the conditions in Theorem 10.7.1 are satisfied. Then we have


t t pn max |bj,n bj |Zt,pn + Kpn Ytpn ,
1jpn

285

(10.34)

where Zt,pn =

1
pn

Ppn

t=1 |Xtj |

and Yt =

Ppn

j
t=1 |Xt |,

n

1 X
ti Xtj ti Xtj = O(pn Q(n) + pn )
n

(10.35)

n

1 X
ti tj ti tj = O(pn Q(n) + pn )
n

(10.36)

t=pn +1

t=pn +1

where Q(n) = p2n

(log log n)1+ log n


n

p3n
n

+ pn pn .

PROOF. Using (10.33) we immediately obtain (10.34).


To obtain (10.35) we use (10.33) to obtain
n
n


1 X
1 X

|Xtj | ti ti
ti Xtj ti Xtj
n
n
t=pn +1

O(pn Q(n))

t=pn +1

1
n

n
X

|Xt ||Zt,pn | + O(pn )

t=pn +1
pn

n
1 X
|Xt ||Ytpn |
n
t=pn +1

= O(pn Q(n) + ).

To prove (10.36) we use a similar method, hence we omit the details.

We apply the above result in the theorem below.


Theorem 10.7.2 Suppose the assumptions in Theorem 10.7.1 are satisfied. Then
r



n 0 2 = O p3n

(log log n)1+ log n p4n


+
+ p2n pn
n
n

!
.

, ) and = ( , ).
n = (
for any > 0, where
0
n n
0 0
n that
PROOF. We note from the definition of
n 0


1
n
0 .
= R
sn R
n

n and sn we replace the estimated residuals n with the true unobserved residuals.
Now in the R

286

This gives us
n 0

Rn =

1
n


1
1 sn )
= R1
n sn Rn 0 + (Rn sn Rn

n
X

Yt Yt

sn =

t=max(p,q)

1
n

n
X

(10.37)

Yt Xt ,

t=max(p,q)

Yt0 = (Xt1 , . . . , Xtp , t1 , . . . , tq ) (recalling that Yt0 = (Xt1 , . . . , Xtp , t1 , . . . , tq ). The


error term is
1 sn ) = R1

1
1
(R1
sn ).
n sn Rn
n (Rn Rn )Rn sn + Rn (sn
1
Now, almost surely R1
n , Rn = O(1) (if E(Rn ) is non-singular). Hence we only need to obtain a
n Rn and sn sn . We recall that
bound for R
X
tY
0 Yt Y0 ),
n Rn = 1
(Y
R
t
t
n
t=pn +1

hence the terms differ where we replace the estimated t with the true t , hence by using (10.35)
and (10.36) we have almost surely
n Rn | = O(pn Q(n) + pn ) and |sn sn | = O(pn Q(n) + pn ).
|R
Therefore by substituting the above into (10.38) we obtain
n 0


pn
= R1
n sn Rn 0 + O(pn Q(n) + ).

(10.38)

Finally using straightforward algebra it can be shown that


1
sn R n n =
n

n
X

t Y t .

t=max(p,q)

By using Theorem 10.7.3, below, we have sn Rn n = O((p + q)

287

(log log n)1+ log n


).
n

Substituting

the above bound into (??), and noting that O(Q(n)) dominates O(
r



n n 2 = O p3n

(log log n)1+ log n


)
n

(log log n)1+ log n p4n


+
+ p2n pn
n
n

and the required result.

10.7.1

gives

T bT k 2 )
Proof of Theorem 10.7.1 (A rate for kb

We observe that



n bn = R1 rn R
n bn + R
n1 Rn1 rn R
n bn
b
n
where b, Rn and rn are deterministic, with bn = (b1 . . . , bpn ), (Rn )i,j = E(Xi Xj ) and (rn )i =
E(X0 Xi ). Evaluating the Euclidean distance we have





n bn k2 kR1 kspec rn R
n bn + kR1 kspec kR
1 kspec R
n Rn rn R
n bn , (10.39)
kb
n
n
n
2
2
2
1 R
1 = R
1 (Rn R
n )R1 and the norm inequalities. Now by using
where we used that R
n
n
n
n
Lemma 5.4.1 we have min (Rn1 ) > /2 for all T . Thus our aim is to obtain almost sure bounds
n bn k2 and kR
n Rn k2 , which requires the lemma below.
for krn R
Theorem 10.7.3 Let us suppose that {Xt } has an ARMA representation where the roots of the
characteristic polynomials (z) and (z) lie are greater than 1 + . Then
(i)
n
1 X
t Xtr = O(
n

t=r+1

(log log n)1+ log n


)
n

(10.40)

(ii)
1
n

n
X

r
Xti Xtj = O(

t=max(i,j)

for any > 0.

288

(log log n)1+ log n


).
n

(10.41)

PROOF. The result is proved in Chapter B.2.

To obtain the bounds we first note that if the there wasnt an MA component in the ARMA pro n bn = 1 Pn
cess, in other words {Xt } was an AR(p) process with pn p, then rn R
t=pn +1 t Xtr ,
n
which has a mean zero. However because an ARMA process has an AR() representation and we
n bn . Therefore we obtain
are only estimating the first pn parameters, there exists a bias in rn R
the decomposition
n bn )r =
(rn R
=

n
1 X
n

Xt

t=pn +1
n
X

bj Xtj

n
1 X
Xtr +
n

bj Xtj Xtr (10.42)

t=pn +1 j=pn +1

j=1

1
1 X X
t Xtr +
bj Xtj Xtr
n
n
t=pn +1
t=pn +1 j=pn +1
{z
} |
|
{z
}
stochastic term

(10.43)

bias

Therefore we can bound the bias with




n
n

X
X
X


pn 1
n bn ) r 1
(rn R

K
|X
|
j |Xtpn j |.
t tr
tr

n
n
t=pn +1

Let Yt =

j
j=1 |Xtj

and Sn,k,r =

1
n

t=1

Pn

t=1 |Xtr |

j=1

(10.44)

j=1

j |X
tkj |.

We note that {Yt } and {Xt }

are ergodic sequences. By applying the ergodic theorm we can show that for a fixed k and r,
a.s.

Sn,k,r E(Xtr Ytk ). Hence Sn,k,r are almost surely bounded sequences and
pn

t=1

j=1

X
1X
|Xtr |
j |Xtpn j | = O(pn ).
n

Therefore almost surely we have


n bn k 2 = k
krn R

n
1 X
t Xt1 k2 + O(pn pn ).
n
t=pn +1

Now by using (10.40) we have


(r
n bn k2 = O pn
krn R

(log log n)1+ log n


+ pn
n

289

)!
.

(10.45)

n bn . Next we consider R
n . It is clear from the definition of R
n that
This gives us a rate for rn R
almost surely we have
n )i,j E(Xi Xj ) =
(R

n
1 X
Xti Xtj E(Xi Xj )
n

1
n
1
n

t=pn +1
n
X

[Xti Xtj

t=min(i,j)
T
X

1
E(Xi Xj )]
n

pn
X
t=min(i,j)

[Xti Xtj E(Xi Xj )] + O(

t=min(i,j)

Xti Xtj +

min(i, j)
E(Xi Xj )
n

pn
).
n

Now by using (10.41) we have almost surely


n )i,j
|(R

pn
E(Xi Xj )| = O( +
n

(log log n)1+ log n


).
n

Therefore we have almost surely


(
n Rn k2 = O
kR

p2n

pn
+
n

(log log n)1+ log n


n

)!
.

(10.46)

We note that by using (10.39), (10.45) and (10.46) we have


r
n bn k2 kR1 kspec kR
1 kspec O p2
kb
n
n
n

(log log n)1+ log n p2n


+
+ pn pn
n
n

!
.

As we mentioned previously, because the spectrum of Xt is bounded away from zero, min (Rn )
n ) min (Rn ) max (R
n Rn )
is bounded away from zero for all T . Moreover, since min (R
n Rn )2 ), which for a large enough n is bounded away from zero. Hence we obtain
min (Rn )tr((R
almost surely
r
n bn k2 = O
kb

p2n

(log log n)1+ log n p3n


+
+ pn pn
n
n

thus proving Theorem 10.7.1 for any > 0.

290

!
,

(10.47)

10.8

Asymptotic properties of the GMLE

Let us suppose that {Xt } satisfies the ARMA representation

Xt

p
X

(0)

i Xti = t +

i=1
(0)

(0)

(0)

q
X

(0)

j tj ,

(10.48)

j=1
(0)

and 0 = (1 , . . . , q ), 0 = (1 , . . . , p ) and 02 = var(t ). In this section we consider the


sampling properties of the GML estimator, defined in Section 7.2.1. We first recall the estimator.
= ( ,
n ) = arg min(,) Ln (, , ), where
We use as an estimator of (0 , 0 ),
n
n n,
1
Ln (, , ) =
n

n1

n1

t=1

t=1

(,) 2

1X
1 X (Xt+1 Xt+1|t )
.
log rt+1 (, , ) +
n
n
rt+1 (, , )

(10.49)

To show consistency and asymptotic normality we will use the following assumptions.
Assumption 10.8.1

(i) Xt is both invertible and causal.

(ii) The parameter space should be such that all (z) and (z) in the parameter space have roots
whose absolute value is greater than 1 + . 0 (z) and 0 (z) belong to this space.
Assumption 10.8.1 means for for some finite constant K and
P
P
j
j
j
j
1
K
j=0 | ||Z |.
j=0 | ||z | and |(z) | K

1
1+

< 1, we have |(z)1 |

To prove the result, we require the following approximations of the GML. Let
(,) =
X
t+1|t,...

t
X

bj (, )Xt+1j .

(10.50)

j=1

This is an approximation of the one-step ahead predictor. Since the likelihood is constructed from
the one-step ahead predictors, we can approximated the likelihood n1 Ln (, , ) with the above and
define
T 1
1
1 X
(,) )2 .
Ln (, , ) = log 2 +
(Xt+1 X
t+1|t,...
2
n
n
t=1

291

(10.51)

(,) was derived from X (,) which is the one-step ahead predictor of Xt+1
We recall that X
t+1|t,...
t+1|t,...
given Xt , Xt1 , . . ., this is
(,)
Xt+1|t,...

bj (, )Xt+1j .

(10.52)

j=1

Using the above we define a approximation of

1
n Ln (, , )

which in practice cannot be obtained

(since the infinite past of {Xt } is not observed). Let us define the criterion
T 1
1
1 X
(,)
Ln (, , ) = log 2 +
(Xt+1 Xt+1|t,... )2 .
n
n 2

(10.53)

t=1

In practice

1
n Ln (, , )

can not be evaluated, but it proves to be a convenient tool in obtaining

. The main reason is because 1 Ln (, , ) is a function of {Xt } and


the sampling properties of
n
n
P
(,)
{Xt+1|t,... = j=1 bj (, )Xt+1j } both of these are ergodic (since the ARMA process is ergodic
when its roots lie outside the unit circle and the roots of , are such that they lie outside
the unit circle). In contrast looking at Ln (, , ), which is comprised of {Xt+1|t }, which not an
ergodic random variable because Xt+1 is the best linear predictor of Xt+1 given Xt , . . . , X1 (see the
number of elements in the prediction changes with t). Using this approximation really simplifies
the proof, though it is possible to prove the result without using these approximations.
n ) = arg min(,) Ln (, , ) and
First we obtain the result for the estimators n = (n , n ,
then show the same result can be applied to n .
Proposition 10.8.1 Suppose {Xt } is an ARMA process which satisfies (10.48), and Assumption
(,) (,)
(,)
10.8.1 is satisfied. Let Xt+1|t , X
t+1|t,... and Xt+1|t,... be the predictors defined in (??), (10.50) and

(10.52), obtained using the parameters = {j } and = {i }, where the roots the corresponding
characteristic polynomial (z) and (z) have absolute value greater than 1 + . Then
t

(,)

(,)
X

X
t+1|t
t+1|t,...

t X i
|Xi |,
1

(10.54)

i=1

(,)
(,) )2 Kt ,
E(Xt+1|t X
t+1|t,...

292

(10.55)

X


X
t
t+1|t,... (1) Xt+1|t,... =

X
bj (, )Xt+1j K
j |Xj |,
j=t+1

(10.56)

j=0

(,)
(,) )2 Kt
E(Xt+1|t,... X
t+1|t,...

(10.57)

|rt (, , ) 2 | Kt

(10.58)

and

for any 1/(1 + ) < < 1 and K is some finite constant.


PROOF. The proof follows closely the proof of Proposition 10.8.1. First we define a separate
ARMA process {Yt }, which is driven by the parameters and (recall that {Xt } is drive by the
P
P
parameters 0 and 0 ). That is Yt satisfies Yt pj=1 j Ytj = t + qj=1 j tj . Recalling that
,
Xt+1|t
is the best linear predictor of Xt+1 given Xt , . . . , X1 and the variances of {Yt } (noting that

it is the process driven by and ), we have


,
Xt+1|t

t
X

bj (, )Xt+1j +

j=1


bj (, )r 0t,j (, )t (, )1 X t ,

(10.59)

j=t+1

where t (, )s,t = E(Ys Yt ), (r t,j )i = E(Yti Yj ) and X 0t = (Xt , . . . , X1 ). Therefore


,
t+1|t,... =
Xt+1|t
X


bj r 0t,j t (, )1 X t .

j=t+1

Since the largest eigenvalue of t (, )1 is bounded (see Lemma 5.4.1) and |(r t,j )i | = |E(Yti Yj )|
K|ti+j| we obtain the bound in (10.54). Taking expectations, we have
,
2
,
E(Xt+1|t
X
t+1|t,... ) =


bj r 0t,j t (, )1 t (0 , 0 )t (, )1

j=t+1


bt+j r t,j .

j=t+1

Now by using the same arguments given in the proof of (5.29) we obtain (10.55).
To prove (10.57) we note that
t+1|t,... (1) Xt+1|t,... )2 = E(
E(X

X
bj (, )Xt+1j )2 = E(
bt+j (, )Xj )2 ,

j=t+1

j=1

293

now by using (2.20), we have |bt+j (, )| Kt+j , for

1
1+

< < 1, and the bound in (10.56).

t+1|t,... (1) Xt+1|t,... )2 Kt , which proves the result.


Using this we have E(X
Using t = Xt

j=1 bj (0 , 0 )Xtj

and substituting this into Ln (, , ) gives

X
2
1
1
Ln (, , ) = log 2 +
Xt
bj (, )Xt+1j
2
n
n
j=1

T 1


1
1 X
Ln (, , ) log 2 +
(B)1 (B)Xt (B)1 (B)Xt
2
n
n
t=1

n
n

X

1 X 2 2X
= log 2 +

bj (, )Xtj
t
t
2
n
n
t=1

t=1

j=1

2
1X X
(bj (, ) bj (0 , 0 ))Xtj .
+
n
t=1

j=1

Remark 10.8.1 (Derivatives involving the Backshift operator) Consider the transformation

j=0

j=0

X
X
1
Xt =
j B j Xt =
j Xtj .
1 B
Suppose we want to differentiate the above with respect to , there are two ways this can be done.
P
1
j
Either differentiate
j=0 Xtj with respect to or differentiate 1B with respect to . In other
words

X
d
1
B
Xt =
Xt =
jj1 Xtj .
2
d 1 B
(1 B)
j=0

Often it is easier to differentiate the operator. Suppose that (B) = 1 +


P
1 qj=1 j B j , then we have

Pp

j=1 j B

and (B) =

d (B)
B j (B)
(B)
Xt =
Xt =
Xtj
2
dj (B)
(B)
(B)2
d (B)
Bj
1
Xt =
Xt =
Xtj .
dj (B)
(B)2
(B)2
Moreover in the case of squares we have
(B)
(B)
d (B)
(
Xt )2 = 2(
Xt )(
Xtj ),
dj (B)
(B)
(B)2

d (B)
(B)
1
(
Xt )2 = 2(
Xt )(
Xtj ).
dj (B)
(B)
(B)2

294

Using the above we can easily evaluate the gradient of

1
n Ln

n
1
(B)
2 X
i Ln (, , ) = 2
((B)1 (B)Xt )
Xti
n

(B)2
t=1
n
X

1
2
j Ln (, , ) = 2
n
n

((B)1 (B)Xt )

t=1

1
Xtj
(B)

n
X
2
1 X
1

bj (, )Xtj .
t
2
4

1
2 Ln (, , ) =
n

t=1

(10.60)

j=1

Let = (i , j , 2 ). We note that the second derivative 2 Ln can be defined similarly.


Lemma 10.8.1 Suppose Assumption 10.8.1 holds. Then
1
sup k Ln k2 KSn
, n

1
sup k 3 Ln k2 KSn
, n

(10.61)

for some constant K,


1
Sn =
n

max(p,q) n
X X

Ytr1 Ytr2

(10.62)

r1 ,r2 =0 t=1

where
Yt = K

j |Xtj |.

j=0

for any

1
(1+)

< < 1.

PROOF. The proof follows from the the roots of (z) and (z) having absolute value greater than
1 + .

Define the expectation of the likelihood L(, , )) = E( n1 Ln (, , )). We observe


L(, , )) = log 2 +

02
1
+ 2 E(Zt (, )2 )
2

where

X
Zt (, ) =
(bj (, ) bj (0 , 0 ))Xtj
j=1

295

Lemma 10.8.2 Suppose that Assumption 10.8.1 are satisfied. Then for all , , we have
(i)

a.s.
1 i
n Ln (, , ))

i L(, , )) for i = 0, 1, 2, 3.

Pmax(p,q) Pn
a.s.
(ii) Let Sn defined in (10.62), then Sn E( r1 ,r2 =0
t=1 Ytr1 Ytr2 ).
PROOF. Noting that the ARMA process {Xt } are ergodic random variables, then {Zt (, )} and
{Yt } are ergodic random variables, the result follows immediately from the Ergodic theorem.
We use these results in the proofs below.

,
Theorem 10.8.1 Suppose that Assumption 10.8.1 is satisfied. Let (n ,
) = arg min Ln (, , )
n n

(noting the practice that this cannot be evaluated). Then we have

a.s.
,
(i) (n ,
) (0 , 0 , 0 ).
n n

(ii)


D
)
n(n 0 ,
N (0, 02 1 ), where
0
n

E(Ut Ut0 ) E(Vt Ut0 )

E(Ut Vt0 ) E(Vt Vt0 )

and {Ut } and {Vt } are autoregressive processes which satisfy 0 (B)Ut = t and 0 (B)Vt = t .
PROOF. We prove the result in two stages below.

PROOF of Theorem 10.8.1(i) We will first prove Theorem 10.8.1(i). Noting the results in
Section 10.3, to prove consistency we recall that we must show (a) the (0 , 0 , 0 ) is the unique
minimum of L() (b) pointwise convergence

a.s.
1
T L(, , ))

L(, , )) and (b) stochastic equicon-

tinuity (as defined in Definition 10.3.2). To show that (0 , 0 , 0 ) is the minimum we note that
L(, , )) L(0 , 0 , 0 )) = log(

2
2
)
+
1 + E(Zt (, )2 ).
02
02

P
Since for all positive x, log x + x 1 is a positive function and E(Zt (, )2 ) = E(
j=1 (bj (, )
bj (0 , 0 ))Xtj )2 is positive and zero at (0 , 0 , 0 ) it is clear that 0 , 0 , 0 is the minimum of
L. We will assume for now it is the unique minimum. Pointwise convergence is an immediate
consequence of Lemma 10.8.2(i). To show stochastic equicontinuity we note that for any 1 =
(1 , 1 , 1 ) and 2 = (2 , 2 , 2 ) we have by the mean value theorem
,

Ln (1 , 1 , 1 ) Ln (2 , 2 , 2 )) = (1 2 )Ln (,
).

296

Now by using (10.61) we have


Ln (1 , 1 , 1 ) Ln (2 , 2 , 2 )) ST k(1 2 ), ( 1 2 ), (1 2 )k2 .
Pmax(p,q) Pn
a.s.
By using Lemma 10.8.2(ii) we have Sn E( r1 ,r2 =0
t=1 Ytr1 Ytr2 ), hence {Sn } is almost surely
bounded. This implies that Ln is equicontinuous. Since we have shown pointwise convergence and
equicontinuity of Ln , by using Corollary 10.3.1, we almost sure convergence of the estimator. Thu
proving (i).

PROOF of Theorem 10.8.1(ii) We now prove Theorem 10.8.1(i) using the Martingale central
limit theorem (see Billingsley (1995) and Hall and Heyde (1980)) in conjunction with the CramerWold device (see Theorem 10.6.1).
Using the mean value theorem we have

n 0 = 2 Ln (
n )1 Ln (0 , 0 , 0 )

,

,
), = ( , 0 , 0 ) and
n = ,
n and 0 .
n = (
lies between
where
0
0
n n,
n
Using the same techniques given in Theorem 10.8.1(i) and Lemma 10.8.2 we have pointwise
a.s.

convergence and equicontinuity of 2 Ln . This means that 2 Ln (n ) E(2 Ln (0 , 0 , 0 )) =


1

a.s.

n 0 ). Therefore by applying Slutskys theorem (since is


(since by definition of n

nonsingular) we have
a.s.

n )1 2 1 .
2 Ln (

(10.63)

Now we show that Ln (0 ) is asymptotically normal. By using (10.60) and replacing Xti =
0 (B)1 0 (B)ti we have
n

t=1
n
X

t=1
T
X

1
Ln (0 , 0 , 0 ) =
n i

2 X (1)
2 X

=
t Vti
t
ti
2n
0 (B)
2n

1
Ln (0 , 0 , 0 ) =
n j

2
2n

1
2 Ln (0 , 0 , 0 ) =
n

T
T
1
1 X 2
1 X 2

=
( 2 ),
2 4n
4n

t=1

1
2
tj = 2
0 (B)
n

t=1

297

t=1

t Utj

t=1

i = 1, . . . , q

j = 1, . . . , p

where Ut =

1
0 (B) t

and Vt =

1
0 (B) t .

We observe that

1
n Ln

is the sum of vector martingale

differences. If E(4t ) < , it is clear that E((t Utj )4 ) = E((4t )E(Utj )4 ) < , E((t Vti )4 ) =
E((4t )E(Vti )4 ) < and E(( 2 2t )2 ) < . Hence Lindebergs condition is satisfied (see the
proof given in Section 10.6.2, for why this is true). Hence we have

nLn (0 , 0 , 0 ) N (0, ).

Now by using the above and (10.63) we have


n 0 = n2 Ln (
n )1 Ln (0 )
n

 D
n 0 N (0, 4 1 ).
n

Thus we obtain the required result.

,
The above result proves consistency and asymptotically normality of (n ,
), which is based
n n

on Ln (, , ), which in practice is impossible to evaluate. However we will show below that the
gaussian likelihood, Ln (, , ) and is derivatives are sufficiently close to Ln (, , ) such that

,
,
the estimators (n ,
) and the GMLE, (n ,
) = arg min Ln (, , ) are asymptotically
n n
n n

equivalent. We use Lemma 10.8.1 to prove the below result.


Proposition 10.8.2 Suppose that Assumption 10.8.1 hold and Ln (, , ), Ln (, , ) and Ln (, , )
are defined as in (10.49), (10.51) and (10.53) respectively. Then we have for all (, ) T heta we
have almost surely
1 (k)
1
| L(, , ) k Ln (, , )| = O( )
n
(,,) n
sup

1
1
|Ln (, , ) L(, , )| = O( ),
n
(,,) n
sup

for k = 0, 1, 2, 3.
PROOF. The proof of the result follows from (10.54) and (10.56).

We show that result for

, ) Ln (, , )|, a similar proof can be used for the rest of the result.
sup(,,) n1 |L(,
Let us consider the difference
Ln (, ) Ln (, ) =

1
(In + IIn + IIIn ),
n

298

where
In =
IIIn =

n1
X
t=1
n1
X
t=1


rt (, , ) 2 ,

IIn =

n1
X
t=1

1
(,)
(,)
(X
Xt+1|t )2
rt (, , ) t+1


1
(,)
(,) ) + ((X (,) )2 (X
(,) )2 ) .
2X
(X

X
t+1
t+1|t
t+1|t,...
t+1|t
t+1|t,...
2

Now we recall from Proposition 10.8.1 that


(,)

(,) K Vt
X

X
t+1|t
t+1|t,...
where Vt =

Pt

i
i=1 |Xi |.

t
(1 )

Hence since E(Xt2 ) < and E(Vt2 ) < we have that supn E|In | < ,

supn E|IIn | < and supn E|IIIn | < . Hence the sequence {|In + IIn + IIIn |}n is almost surely
bounded. This means that almost surely


1
sup Ln (, ) Ln (, ) = O( ).
n
,,
Thus giving the required result.

Now by using the above proposition the result below immediately follows.
= arg min LT (, , ) and (, )
= arg min L
T (, , )
Theorem 10.8.2 Let (, )
a.s.
a.s.
(i) (, )
(0 , 0 ) and (, )
(0 , 0 ).
(ii)

D
)
T (T 0 ,
N (0, 04 1 )
0
T

D
)
and T (T 0 ,
N (0, 04 1 ).
0
T

PROOF. The proof follows immediately from Proposition 10.8.1.

299

Appendix A
Background
A.1

Some definitions and inequalities

Some norm definitions.


The norm of an object, is a postive numbers which measure the magnitude of that object.
P
P
Suppose x = (x1 , . . . , xn ) Rn , then we define kxk1 = nj=1 |xj | and kxk2 = ( nj=1 |x2j )1/2
(this is known as the Euclidean norm). There are various norms for matrices, the most
popular is the spectral norm k kspec : let A be a matrix, then kAkspec = max (AA0 ), where
max denotes the largest eigenvalue.
Z denotes the set of a integers {. . . , 1, 0, 1, 2, . . .}. R denotes the real line (, ).
Complex variables.

i = 1 and the complex variable z = x + iy, where x and y are real.


Often the radians representation of a complex variable is useful. If z = x + iy, then it can
p
also be written as r exp(i), where r = x2 + y 2 and = tan1 (y/x).
If z = x + iy, its complex conjugate is z = x iy.
The roots of a rth order polynomial a(z), are those values 1 , . . . , r where a(i ) = 0 for
i = 1, . . . , r.
Let (A) denote the spectral radius of the the matrix A (the largest eigenvalue in absolute
terms). Then for any matrix norm kAk we have limj kAj k1/j = (A) (see Gelfands

300

formula). Suppose (A) < 1, then Gelfands formula implies that for any (A) < < 1,
there exists a constant, C, (which only depends A and ), such that kAj k CA, j .
The mean value theorem.
This basically states that if the partial derivative of the function f (x1 , x2 , . . . , xn ) has a
bounded in the domiain , then for x = (x1 , . . . , xn ) and y = (y1 , . . . , yn )
n
X
f
cx=x
f (x1 , x2 , . . . , xn ) f (y1 , y2 , . . . , yn ) =
(xi yi )
xi
i=1

where x lies somewhere between x and y.


The Taylor series expansion.
This is closely related to the mean value theorem and a second order expansion is
f (x1 , x2 , . . . , xn ) f (y1 , y2 , . . . , yn ) =

n
X

(xi yi )

n
X
f 2
f
+
(xi yi )(xj yj )
cx=x
xi
xi xj
i,j=1

i=1

Partial Fractions.
We use the following result mainly for obtaining the MA() expansion of an AR process.
Suppose that |gi | > 1 for 1 i n. Then if g(z) =

Qn

i=1 (1

z/gi )ri , the inverse of g(z)

satisfies
n

i=1

j=1

i
XX
gi,j
1
=
,
g(z)
(1 gzi )j

where gi,j = ..... Now we can make a polynomial series expansion of (1 gzi )j which is valid
for all |z| 1.
Dominated convergence.
Suppose a sequence of functions fn (x) is such that pointwise fn (x) f (x) and for all n and
R
R
x, |fn (x)| g(x), then fn (x)dx f (x)dx as n .
We use this result all over the place to exchange infinite sums and expectations. For example,

301

if

j=1 |aj |E(|Zj |)

< , then by using dominated convergence we have

X
X
E(
aj Zj ) =
aj E(Zj ).
j=1

j=1

Dominated convergence can be used to prove the following lemma. A more hands on proof
is given below the lemma.
Lemma A.1.1 Suppose

k= |c(k)|

1
n

as n . Moreover, if

< , then we have

(n1)

|kc(k)| 0

k=(n1)

k= |kc(k)|

< , then

1
n

P(n1)

k=(n1) |kc(k)|

= O( n1 ).

P
PROOF. The proof is straightforward in the case that
k= |kc(k)| < (the second asserP(n1)
1
tion), in this case k=(n1) |k|
n |c(k)| = O( n ). The proof is slightly more tricky in the case
P
P
that
k= |c(k)| < for every > 0 there
k= |c(k)| < . First we note that since
P
exists a N such that for all n N , |k|n |c(k)| < . Let us suppose that n > N , then we
have the bound
1
n

(n1)

|kc(k)|

k=(n1)

1
n

(N 1)

X
k=(N 1)

1
2n

Hence if we keep N fixed we see that

|kc(k)| +

1
n

1
n

|kc(k)|

N |k|n

(N 1)

|kc(k)| + .

k=(N 1)

P(N 1)

k=(N 1) |kc(k)|

0 as n . Since this is

true for all (for different thresholds N ) we obtain the required result.
Cauchy Schwarz inequality.
In terms of sequences it is
|

X
j=1

X
X
aj bj | (
a2j )1/2 (
b2j )1/2
j=1

302

j=1

. For integrals and expectations it is


E|XY | E(X 2 )1/2 E(Y 2 )1/2

Holders inequality.
This is a generalisation of the Cauchy Schwarz inequality. It states that if 1 p, q and
p + q = 1, then
E|XY | E(|X|p )1/p E(|Y |q )1/q
. A similar results is true for sequences too.
Martingale differences. Let Ft be a sigma-algebra, where Xt , Xt1 , . . . Ft . Then {Xt } is a
sequence of martingale differences if E(Xt |Ft1 ) = 0.
Minkowskis inequality.
If 1 < p < , then
n
n
X
X
(E(
Xi )p )1/p
(E(|Xi |p ))1/p .
i=1

i=1

Doobs inequality.
This inequality concerns martingale differences. Let Sn =

Pn

t=1 Xt ,

then

2
E( sup |Sn |2 ) E(SN
).
nN

Burkh
olders inequality.
Suppose that {Xt } are martingale differences and define Sn =

Pn

k=1 Xt .

For any p 2 we

have
{E(Snp )}1/p

2p

n
X

E(Xkp )2/p

1/2

k=1

An application, is to the case that {Xt } are identically distributed random variables, then
we have the bound E(Snp ) E(X0p )2 (2p)p/2 np/2 .
It is worthing noting that the Burkholder inequality can also be defined for p < 2 (see

303

Davidson (1994), pages 242). It can also be generalised to random variables {Xt } which are
not necessarily martingale differences (see Dedecker and Doukhan (2003)).
Riemann-Stieltjes Integrals.
R
In basic calculus we often use the basic definition of the Riemann integral, g(x)f (x)dx, and if
R
R
the function F (x) is continuous and F 0 (x) = f (x), we can write g(x)f (x)dx = g(x)dF (x).
There are several instances where we need to broaden this definition to include functions F
which are not continuous everywhere. To do this we define the Riemann-Stieltjes integral,
which coincides with the Riemann integral in the case that F (x) is continuous.
R
R
g(x)dF (x) is defined in a slightly different way to the Riemann integral g(x)f (x)dx.
P
Let us first consider the case that F (x) is the step function F (x) = ni=1 ai I[xi1 ,xi ] , then
R
R
Pn
g(x)dF (x) is defined as g(x)dF (x) =
i=1 (ai ai1 )g(xi ) (with a1 = 0). Already
we see the advantage of this definition, since the derivative of the step function is not
well defined at the jumps. As most functions can be written as the limit of step funcR
P k
ai,nk I[xi 1,x ] ), we define g(x)dF (x) =
tions (F (x) = limk Fk (x), where Fk (x) = ni=1
ik
k1
Pnk
limk i=1 (ai,nk ai1,nk )g(xik ).
In statistics, the function F will usually be non-decreasing and bounded. We call such
functions distributions.
Theorem A.1.1 (Hellys Theorem) Suppose that {Fn } are a sequence of distributions with
Fn () = 0 and supn Fn () M < . There exists a distribution F , and a subsequence
Fnk such that for each x R Fnk F and F is right continuous.

A.2

Martingales

Definition A.2.1 A sequence {Xt } is said to be a martingale difference if E[Xt |Ft1 ], where
Ft=1 = (Xt1 , Xt2 , . . .). In other words, the best predictor of Xt given the past is simply zero.
Martingales are very useful when proving several results, including central limit theorems.
Martingales arise naturally in several situations. We now show that if correct likelihood is
used (not the quasi-case), then the gradient of the conditional log likelihood evaluated at the true
P
parameter is the sum of martingale differences. To see why, let BT = Tt=2 log f (Xt |Xt1 , . . . , X1 )

304

be the conditonal log likelihood and CT () its derivative, where

CT () =

T
X
log f (Xt |Xt1 , . . . , X1 )

t=2

We want to show that CT (0 ) is the sum of martingale differences. By definition if CT (0 ) is the


sum of martingale differences then

E




log f (Xt |Xt1 , . . . , X1 )
c=0 Xt1 , Xt2 , . . . , X1 = 0,

we will show this. Rewriting the above in terms of integrals and exchanging derivative with integral
we have



log f (Xt |Xt1 , . . . , X1 )

c=0 Xt1 , Xt2 , . . . , X1

log f (xt |Xt1 , . . . , X1 )


c=0 f0 (xt |Xt1 , . . . , X1 )dxt

1
f (xt |Xt1 , . . . , X1 )
c=0 f0 (xt |Xt1 , . . . , X1 )dxt
f0 (xt |Xt1 , . . . , X1 )

Z

f (xt |Xt1 , . . . , X1 )dxt c=0 = 0.


E
Z
=
Z
=
=

|Xt1 ,...,X1 )
Therefore { log f (Xt
c=0 }t are a sequence of martingale differences and Ct (0 ) is the sum

of martingale differences (hence it is a martingale).

A.3

The Fourier series

The Fourier transform is a commonly used tool. We recall that {exp(2ij); j Z} is an orthogonal
R2
basis of the space L2 [0, 1]. In other words, if f L2 [0, 1] (ie, 0 f ()2 d < ) then
fn (u) =

n
X

cj e

iju2

Z
cj =

f (u) exp(i2ju)du,
0

j=n

where

|f (u) fn (u)|2 du 0 as n . Roughly speaking, if the function is continuous then we

can say that


f (u) =

cj eiju .

jZ

305

An important property is that f (u) constant iff cj = 0 for all j 6= 0. Moreover, for all n Z
f (u + n) = f (u) (hence f is periodic).
Some relations:
(i) Discrete Fourier transforms of finite sequences
It is straightforward to show (by using the property

Pn

j=1 exp(i2k/n)

= 0 for k 6= 0) that if

1 X
dk =
xj exp(i2jk/n),
n
j=1

then {xr } can be recovered by inverting this transformation


n

1 X
xr =
dk exp(i2rk/n),
n
k=1

(ii) Fourier sums and integrals


Of course the above only has meaning when {xk } is a finite sequence. However suppose that
P
{xk } is a sequence which belongs to `2 (that is k x2k < ), then we can define the function

1 X
xk exp(ik),
f () =
2 k=

where

R 2
0

f ()2 d =

x2k , and we we can recover {xk } from f (), through


1
xk =
2

(iii) Convolutions. Let us suppose that

f () exp(ik).
0

|ak |2 < and

Fourier transform of the sequences {ak } and {bk } as A() =


P
1
k bk exp(ik) respectively. Then
2

Z
aj bkj

j=

1 ak
2

exp(ik) and B() =

A()B() exp(ik)d

=
0

j=

|bk |2 < and we define the

A()B( )d.

aj bj exp(ij) =
0

306

(A.1)

The proof of the above follows from

aj bj exp(ij) =

Z
X
r= 0

j=

A(1 )B(2 ) exp(ir(1 + 2 )) exp(ij)


0

Z Z
=

A(1 )B(2 )

exp(ir( 1 2 )) d1 d2

r=

|
Z

{z

= (1 +2 )

A()B( )d.

=
0

(iv) Using the DFT to calculate convolutions. Our objective is calculate

Pn

j=k

aj bjs for

all s = 0, . . . , n 1 in as few computing computing operations. This is typically done via the
DFT. Examples in time series where this is useful is in calculating the sample autocovariance
function.
Suppose we have two sequences a = (a1 , . . . , an ) and b = (b1 , . . . , bn ). Let An (k,n ) =
Pn
Pn
j=1 bj exp(ijk,n ) where k,n = 2k/n. It is straightj=1 aj exp(ijk,n ) and Bn (k,n ) =
forward to show that
n

s1

k=1

j=s

j=1

X
X
1X
An (k,n )Bn (k,n ) exp(isk,n ) =
aj bjs +
aj bjs+n ,
n
this is very fast to compute (requiring only O(n log n) operations using first the FFT and
then inverse FFT). The only problem is that we dont want the second term.
By padding the sequences and defining An (k,2n ) =

Pn

j=1 aj

exp(ijk,2n ) =

P2n

j=1 aj

exp(ijk,2n ),

with k,2n = 2k/2n (where we set aj = 0 for j > 0) and analogously Bn (k,2n ) =
Pn
j=1 bj exp(ijk,2n ), we are able to remove the second term. Using the same calculations
we have
2n

s1

X
X
1X
An (k,2n )Bn (k,2n ) exp(isk,2n ) =
aj bjs +
aj bjs+2n .
n
j=s
j=1
k=1
|
{z
}
=0

This only requires O(2n log(2n)) operations to compute the convolution for all 0 k n 1.
(v) The Poisson Summation Formula Suppose we do not observe the entire function and
observe a sample from it, say ft,n = f ( nt ) we can use this to estimate the Fourier coefficient

307

cj via the Discrete Fourier Transform:


n

cj,n

2t
1X t
f ( ) exp(ij
).
=
n
n
n
t=1

The Poisson Summation formula is


cj,n = cj +

cj+kn +

k=1

cjkn ,

k=1

which we can prove by replacing f ( nt ) with

jZ cj e

ij2t/n .

In other words, cj,n cannot

disentangle frequency eij from its harmonics ei(j+n) (this is aliasing).


(vi) Error in the DFT By using the Poisson summation formula we can see that

|cj,n cj |

|cj+kn | +

k=1

|cjkn |

k=1

It can be shown that if a function f () is (p + 1) times differentiable with bounded derivatives


or that f p () is bounded and piecewise montonic then the corresponding Fourier coefficients
satisfy
|cj | C|j|(p+1) .
Using this result and the Poisson summation formula we can show that for |j| n/2 that
if if a function f () is (p + 1) times differentiable with bounded derivatives or that f p () is
piecewise montonic and p 1 then
|cj,n cj | Cn(p+1) ,

(A.2)

where C is some finite constant. However, we cannot use this result in the case that f is
bounded and piecewise monotone, however it can still be shown that
|cj,n cj | Cn1 ,
see Section 6.3, page 189, Briggs and Henson (1997).

308

(A.3)

A.4

Application of Burkholders inequality

There are two inequalities (one for 1 < p 2). Which is the following:
Theorem A.4.1 Suppose that Yk are martingale differences and that Sn =

Pn

j=1 Yk ,

then for

0<q2
E|Sn |q 2

n
X

E(Xkq ),

(A.4)

j=1

See for example Davidson (p. 242, Theorem 15.17).


And one for (p 2), this is the statement for the Burkolder inequality:
Theorem A.4.2 Suppose {Si : Fi } is a martingale and 1 < p < . Then there exists constants
C1 , C2 depending only on p such that

C1 E

m
X

!p/2
Xi2

E|Sn |p C2 E

i=1

m
X

!p/2
Xi2

(A.5)

i=1

An immediately consequence of the above for p 2 is the following corollary (by using Holders
inequality):
Corollary A.4.1 Suppose {Si : Fi } is a martingale and 2 p < . Then there exists constants
C1 , C2 depending only on p such that

kSn kE
p

2/p
C2

m
X

!1/2
kXi2 kE
p/2

(A.6)

i=1

PROOF. By using the right hand side of (A.5) we have


1/p
E|Sn |p

!p/2 2/p 1/2


m
X


C2 E
Xi2

i=1

E 1/2

m

X

2/p
= C2
Xi2 .

i=1

309

p/2

(A.7)

By using Holder inequality we have


"



p 1/p

E|Sn |

2/p
C2

m
X

#1/2
kXi2 kE
p/2

(A.8)

i=1

Thus we have the desired result.

We see the value of the above result in the following application. Suppose Sn =

1
n

Pn

k=1 Xk

and kXk kE
p K. Then we have
n

1X
Xk
n

!p

#p/2
n
1 2/p X
2 E
C
kXk kp/2
n 2
k=1
#p/2
" n
" n
#p/2
C2 X 2
1
C2 X
2 E
p
kXk kp/2
K
= O( p/2 ).
np
n
n

"

k=1

k=1

(A.9)

k=1

Below is the result that that Moulines et al (2004) use (they call it the generalised Burkholder
inequality) the proof can be found in Dedecker and Doukhan (2003). Note that it is for p 2,
which I forgot to state in what I gave you.
Lemma A.4.1 Suppose {k : k = 1, 2, . . .} is a stochastic process which satisfies E(k ) = 0 and
E(pk ) < for some p 2. Let Fk = (k , k1 , . . .). Then we have that

1/2
s

s
s
X E
X
X


.
k
2p
kk kE
kE(j |Fk )kE

p
p


k=1

We note if

Ps

j=k

k=1

(A.10)

j=k

kE(j |Fk )kE


p < , then we (A.11) is very similar to (A.6), and gives the same

rate as (A.9).
But I think one can obtain something similar for 1 p 2. I think the below is correct.
Lemma A.4.2 Suppose {k : k = 1, 2, . . .} is a stochastic process which satisfies E(k ) = 0 and
E(qk ) < for some 1 < q 2. Let Fk = (k , k1 , . . .). Further, we suppose that there exists a
0 < < 1, and 0 < K < such that kE(t |Ftj )kq < Kj . Then we have that

E
s
X



ak k




k=1

K
1

310

s
X
k=1

!1/q
|ak |q

(A.11)

where K is a finite constant.


PROOF. Let Ej (k ) = E(k |Fkj ). We note that by definition {k } is a mixingale (see, for
example, Davidson (1997), chapter 16), therefore amost surely k satisfies the representation

k =

[Ekj (k ) Ekj1 (k )].

(A.12)

j=0

By substituting the above into the sum


s
X

ak k =

s X

Ps

k=1 ak k

[Ekj (k ) Ekj1 (k )] =

k=1 j=0

k=1

we obtain

s
X
X
j=0

!
[Ekj (k ) Ekj1 (k )] .

(A.13)

k=1

Keeping j constant, we see that {Ekj (k ) Ekj1 (k )}k is a martingale sequence. Hence
Ps
k=1 [Ekj (k ) Ekj1 (k )] is the sum of martingale differences. This implies we can apply
(A.4) to (A.13), and get
s
E
s
E
X
X


X




ak k

|ak |[Ekj (k ) Ekj1 (k )]








k=1

j=0

X
j=0

k=1

s
X

!1/q
|ak |(kEkj (k )

q
Ekj1 (k )kE
q )

k=1

j
Under the stated assumption kEkj (k ) Ekj1 (k )kE
q 2K . Substituting this inequality into

the above gives



E
!1/q
s

X

X
X
X


q
j q
1+1/q
ak k

2
|ak | (2K )
2
K
j



k=1

j=0

j=0

k=1

Thus we obtain the desired result.

A.5

s
X

!1/q
|ak |q

k=1

The Fast Fourier Transform (FFT)

The Discrete Fourier transform is used widely in several disciplines. Even in areas its use may
not be immediately obvious (such as inverting Toeplitz matrices) it is still used because it can be
evalated in a speedy fashion using what is commonly called the fast fourier transform (FFT). It is
an algorithm which simplifies the number of computing operations required to compute the Fourier

311

transform of a sequence of data. Given that we are in the age of big data it is useful to learn what
one of most popular computing algorithms since the 60s actually does.
Recalling the notation in Section 8.2.2 the Fourier transform is the linear transformation
Fn X n = (Jn (0 ), . . . , Jn (n1 )).
If this was done without any using any tricks this requires O(n2 ) computing operations. By using
some neat factorizations, the fft reduces this to n log n computing operations.
To prove this result we will ignore the standardization factor (2n)1/2 and consider just the
Fourier transform
d(k,n ) =

n
X

xt exp (itk,n ),

t=1

{z

k different frequencies

where k,n =

2k
n .

Here we consider the proof for general n, later in Example A.5.1 we consider

the specific case that n = 2m . Let us assume that n is not a prime (if it is then we simply pad the
vector with one zero and increase the length to n + 1), then it can be factorized as n = pq. Using
these factors we write t as t = t1 p + tmodp where t1 is some integer value that lies between 0 to
q 1 and t0 = tmodp lies between 0 to p 1. Substituting this into d(k ) gives

d(k ) =

n
X

xt exp [i(t1 p
t=1
p1 X
q1
X

+ tmodp)k,n ]

xt1 p+t0 exp [i(t1 p + t0 )k,n ] =

t0 =0 t1 =0

It is straightforward to see that t1 pk,n =

p1
X

exp [it0 k,n ]

t0 =0
2t1 pk
n

312

2t1 k
q

q1
X

xt1 p+t0 exp [it1 pk,n ]

t1 =0

= t1 k,q and that exp(it1 pk,n ) =

exp(it1 k,q ) = exp(it1 kmodq,q ). This means d(k ) can be simplified as

d(k ) =

p1
X

exp [it0 k,n ]

xt1 p+t0 exp [it1 kmodq,q ]

t1 =0

t0 =0
p1
X

q1
X

exp [it0 k,n ]

q1
X

xt1 p+t0 exp [it1 k0 ,q ]

t1 =0

t0 =0

{z

embedded Fourier transform

p1
X
t0 =0

exp [it0 k,n ] A(t0 , kmodq),


|
{z
}
q frequencies

where k0 = kmodq can take values from 0, . . . , q 1. Thus to evaluate d(k ) we need to evaluate
A(t0 , kmodq) for 0 t0 p 1, 0 k0 q 1. To evaluate A(t0 , kmodq) requires q computing
operations, to evaluate it for all t0 and kmodq requires pq 2 operations. Note, the key is that less
frequencies need to be evaluated when calculating A(t0 , kmodq), in particular q frequencies rather
than N . After evaluating {A(t0 , k0 ); 0 t0 p 1, 0 k0 q 1} we then need to take the
Fourier transform of this over t0 to evaluate d(k ) which is p operations and this needs to be done
n times (to get all {d(k )}k ) this leads to np. Thus in total this leads to
p2 q
|{z}

evaluation of all A

= pq 2 + pn = n(q + p).

np
|{z}

(A.14)

evaluation of the transforms of A

Observe that n(p + q) is a lot smaller than n2 .


Looking back at the above calculation we observe that q 2 operations were required to calculate
A(t0 , kmodq) = A(t0 , k0 ) for all 0 k0 q 1. However A(t0 , k0 ) is a Fourier transform

A(t0 , k0 ) =

q1
X

xt1 p+t0 exp [it1 k0 ,q ] .

t1 =0

Therefore, we can use the same method as was used above to reduce this number. To do this we

313

need to factorize q into p = p1 q1 and using the above method we can write this as

A(t0 , k0 ) =

pX
1 1
1 1 qX

x(t2 +t3 p1 )p+t0 exp [i(t2 + t3 p1 )k0 ,q ]

t2 =0 t3 =0

pX
1 1

exp [it2 k0 ,q ]

qX
1 1

t2 =0

t3 =0

pX
1 1

qX
1 1

exp [it2 k0 ,q ]

x(t2 +t3 p1 )p+t0 exp [it3 p1 k0 ,q ]

x(t2 +t3 p1 )p+t0 exp [it3 k0 modq1 ,q1 ] .

t3 =0

t2 =0

We note that k0 modq1 = (kmod(p1 q1 )modq1 ) = kmodq1 , substituting this into the above we have

A(t0 , k0 ) =

pX
1 1
t2 =0

pX
1 1
t2 =0

exp [it2 k0 ,q ]

qX
1 1

x(t2 +t3 p1 )p+t0 exp [it3 k0 modq1 ,q1 ]

t3 =0

exp [it2 k0 ,q ] A(t0 , t2 , k0 modq1 ) .


|
{z
}
q1 frequencies

Thus we see that q1 computing operations are required to calculate A(t0 , t2 , k0 modq1 ) and to calculate A(t0 , t2 , kmodq1 ) for all 0 t2 p1 1 and 0 kmodq1 q1 1 requires in total q12 p1
computing operations. After evaluating {A(t0 , t2 , k0 modq1 ); 0 t2 q2 1, 0 kmodq1 q1 1}
we then need to take its Fourier transform over t2 to evaluate A(t0 , k0 ), which is p1 operations.
Thus in total to evaluate A(t0 , k0 ) over all k0 we require q12 p1 + p1 q operations. Thus we have
reduced the number of computing operations for A(t0 , k0 ) from q 2 to q(p1 + q1 ), substituting this
into (A.14) gives the total number of computing operations to calculate {d(k )}
pq(p1 + q1 ) + np = n(p + p1 + q1 ).
In general the same idea can be used to show that given the prime factorization of n =
P
then the number of computing operations to calculate the DFT is n( m
s=1 ps ).

314

Qm

s=1 ps ,

Example A.5.1 Let us suppose that n = 2m then we can write d(k ) as

d(k ) =

n
X

xt exp(itk ) =

n/2
X

t=1

(n/2)1

X2t exp(i2tk ) +

t=1

X2t+1 exp(i(2t + 1)k )

t=0

n/2

(n/2)1

X2t exp(i2tk ) + exp(ik )

t=1

X2t+1 exp(i2tk )

t=0

= A(0, kmod(n/2)) + exp(ik )A(1, kmod(n/2)),


since

Pn/2

t=1 X2t exp(i2tk )

and

Pn/2

t=1 X2t+1 exp(i2tk )

are the Fourier transforms of {Xt } on a

coarser scale, therefore we can only identify the frequencies on a coarser scale. It is clear from the
above that the evaluation of A(0, kmod(n/2)) for 0 kmod(n/2) n/2 requires (n/2)2 operations
and same for A(1, kmod(n/2)). Thus to evaluate both A(0, kmod(n/2)) and A(1, kmod(n/2)) requires 2(n/2)2 operations. Then taking the Fourier transform of these two terms over all 0 k
n 1 is an additional 2n operations leading to
2(n/2)2 + 2n = n2 /2 + 2n operations < n2 .
We can continue this argument and partition

A(0, kmod(n/2)) =

n/2
X

X2t exp(i2tk )

t=1

n/4
X

(n/4)1

X4t exp(i4tk ) + exp(i2k )

t=1

X4t+2 exp(i4tk ).

t=0

Using the same argument as above the calculation of this term over all k requires 2(n/4)2 +2(n/2) =
n2 /8 + n operations. The same decomposition applies to A(1, kmod(n/2)). Thus calculation of both
terms over all k requires 2[n2 /8 + n] = n2 /4 + 2n operations. In total this gives
(n2 /4 + 2n + 2n)operations.
Continuing this argument gives mn = n log2 n operations, which is the often cited rate.
Typically, if the sample size is not of order 2m zeros are added to the end of the sequence (called
padding) to increase the length to 2m .

315

Appendix B
Mixingales
In this section we prove some of the results stated in the previous sections using mixingales.
We first define a mixingale, noting that the definition we give is not the most general definition.
Definition B.0.1 (Mixingale) Let Ft = (Xt , Xt1 , . . .), {Xt } is called a mixingale if it satisfies
 
2 1/2
= E E(Xt |Ftk ) E(Xt )
,

t,k

where t,k 0 as k . We note if {Xt } is a stationary process then t,k = k .


Lemma B.0.1 Suppose {Xt } is a mixingale. Then {Xt } almost surely satisfies the decomposition


X
Xt =
E(Xt |Ftj1 ) E(Xt |Ftj1 ) .

(B.1)

j=0

PROOF. We first note that by using a telescoping argument that


Xt E(Xt ) =

m
X




E(Xt |Ftk ) E(Xt |Ftk1 ) + E(Xt |Ftm1 ) E(Xt ) .

k=0

By definition of a martingale E E(Xt |Ftm1 ) E(Xt )

2

0 as k , hence the remainder term

in the above expansion becomes negligable as m and we have almost surely


Xt E(Xt )

X


=
E(Xt |Ftk ) E(Xt |Ftk1 ) .
k=0

316

Thus giving the required result.

We observe that (B.1) resembles the Wold decomposition. The difference is that the Wolds
decomposition decomposes a stationary process into elements which are the errors in the best linear
predictors. Whereas the result above decomposes a process into sums of martingale differences.
It can be shown that functions of several ARCH-type processes are mixingales (where t,k Kk
(rho < 1)), and Subba Rao (2006) and Dahlhaus and Subba Rao (2007) used these properties to
obtain the rate of convergence for various types of ARCH parameter estimators. In a series of
papers, Wei Biao Wu considered properties of a general class of stationary processes which satisfied
P
Definition B.0.1, where
k=1 k < .
In Section B.2 we use the mixingale property to prove Theorem 10.7.3. This is a simple illustration of how useful mixingales can be. In the following section we give a result on the rate of
convergence of some random variables.

B.1

Obtaining almost sure rates of convergence for


some sums

The following lemma is a simple variant on a result proved in Moricz (1976), Theorem 6.
Lemma B.1.1 Let {ST } be a random sequence where E(sup1tT |St |2 ) (T ) and {phi(t)} is a
monotonically increasing sequence where (2j )/(2j1 ) K < for all j. Then we have almost
surely
1
ST = O
T

p
(T )(log T )(log log T )1+ 
.
T

PROOF. The idea behind the proof is to that we find a subsequence of the natural numbers and
define a random variables on this subsequence. This random variable, should dominate (in some
sense) ST . We then obtain a rate of convergence for the subsequence (you will see that for the
subsequence its quite easy by using the Borel-Cantelli lemma), which, due to the dominance, can
be transfered over to ST . We make this argument precise below.
Define the sequence Vj = supt2j |St |. Using Chebyshevs inequality we have
P (Vj > )

(2j )
.

317

Let (t) =

p
(t)(log log t)1+ log t. It is clear that

P (Vj > (2 ))

j=1

X
j=1

C(2j )
< ,
(2j )(log j)1+ j

where C is a finite constant. Now by Borel Cantelli, this means that almost surely Vj (2j ). Let
us now return to the orginal sequence ST . Suppose 2j1 T 2j , then by definition of Vj we have
a.s (2j )
Vj
ST

<

(T )
(2j1 )
(2j1 )

under the stated assumptions. Therefore almost surely we have ST = O((T )), which gives us the
required result.

We observe that the above result resembles the law of iterated logarithms. The above result
is very simple and nice way of obtaining an almost sure rate of convergence. The main problem
is obtaining bounds for E(sup1tT |St |2 ). There is on exception to this, when St is the sum
of martingale differences then one can simply apply Doobs inequality, where E(sup1tT |St |2 )
E(|ST |2 ). In the case that ST is not the sum of martingale differences then its not so straightforward.
However if we can show that ST is the sum of mixingales then with some modifications a bound
for E(sup1tT |St |2 ) can be obtained. We will use this result in the section below.

B.2

Proof of Theorem 10.7.3

We summarise Theorem 10.7.3 below.


Theorem 1 Let us suppose that {Xt } has an ARMA representation where the roots of the characteristic polynomials (z) and (z) lie are greater than 1 + . Then
(i)
n
1 X
t Xtr = O(
n

t=r+1

(log log n)1+ log n


)
n

(B.2)

(ii)
1
n

n
X

r
Xti Xtj = O(

t=max(i,j)

318

(log log n)1+ log n


).
n

(B.3)

for any > 0.


By using Lemma ??, and that

Pn

t=r+1 t Xtr

is the sum of martingale differences, we prove

Theorem 10.7.3(i) below.


PROOF of Theorem 10.7.3. We first observe that {t Xtr } are martingale differences,
P
hence we can use Doobs inequality to give E(supr+1sT ( st=r+1 t Xtr )2 ) (T r)E(2t )E(Xt2 ).
Now we can apply Lemma ?? to obtain the result.

We now show that


1
T

T
X

r
Xti Xtj = O(

t=max(i,j)

(log log T )1+ log T


).
T

However the proof is more complex, since {Xti Xtj } are not martingale differences and we cannot
directly use Doobs inequality. However by showing that {Xti Xtj } is a mixingale we can still
show the result.
To prove the result let Ft = (Xt , Xt1 , . . .) and Gt = (Xti Xtj , Xt1i Xtji , . . .). We
observe that if i > j, then Gt Fti .
Lemma B.2.1 Let Ft = (Xt , Xt1 , . . .) and suppose Xt comes from an ARMA process, where
the roots are greater than 1 + . Then if E(4t ) < we have
E E(Xti Xtj |Ftmin(i,j)k ) E(Xti Xtj )

2

Ck .

PROOF. By expanding Xt as an MA() process we have


E(Xti Xtj |Ftmin(i,j)k ) E(Xti Xtj )

X


=
aj1 aj2 E(tij1 tjj2 |Ftkmin(i,j) ) E(tij1 tjj2 ) .
j1 ,j2 =0

Now in the case that tij1 > tkmin(i, j) and tjj2 > tkmin(i, j), E(tij1 tjj2 |Ftkmin(i,j) ) =
E(tij1 tjj2 ). Now by considering when tij1 tkmin(i, j) or tj j2 tkmin(i, j)
we have have the result.

Lemma B.2.2 Suppose {Xt } comes from an ARMA process. Then

319

(i) The sequence {Xti Xtj }t satisfies the mixingale property


2
E E(Xti Xtj |Ftmin(i,j)k ) E(Xti Xtj |Ftk1 ) Kk ,

(B.4)

and almost surely we can write Xti Xtj as

Xti Xtj E(Xti Xtj ) =

n
X

Vt,k

(B.5)

k=0 t=min(i,j)

where Vt,k = E(Xti Xtj |Ftkmin(i,j) ) E(Xti Xtj |Ftkmin(i,j)1 ), are martingale differences.
2 ) Kk and
(ii) Furthermore E(Vt,k


E

s
X

sup
min(i,j)sn


{Xti Xtj E(Xti Xtj )})2 Kn,

(B.6)

t=min(i,j)

where K is some finite constant.


PROOF. To prove (i) we note that by using Lemma B.2.1 we have (B.4). To prove (B.5) we use
the same telescoping argument used to prove Lemma B.0.1.
To prove (ii) we use the above expansion to give

E

s
X

sup
min(i,j)sn

= E

sup
min(i,j)sn

= E

{Xti Xtj E(Xti Xtj )})2

t=min(i,j)

s
X
X

Vt,k

k=0

(B.7)

2

k=0 t=min(i,j)

sup

s
X


Vt,k1

k1 =0 k2 =0 min(i,j)sn t=min(i,j)

X
 
=
E

sup

s
X

2
Vt,k1

s
X


Vt,k2

t=min(i,j)

1/2 2

min(i,j)sn t=min(i,j)

Now we see that {Vt,k }t = {E(Xti Xtj |Ftkmin(i,j) ) E(Xti Xtj |Ftkmin(i,j)1 )}t , therefore


Ps
{Vt,k }t are also martingale differences. Hence we can apply Doobs inequality to E supmin(i,j)sn
t=min(i,j) Vt,k

320

and by using (B.4) we have



E

s
X

sup
min(i,j)sn

Vt,k

2

t=min(i,j)

n
X

Vt,k

t=min(i,j)

2

n
X

2
E(Vt,k
) K nk .

t=min(i,j)

Therefore now by using (B.7) we have



E

sup
min(i,j)sn

s
X


{Xti Xtj E(Xti Xtj )})2 Kn.

t=min(i,j)

Thus giving (B.6).

We now use the above to prove Theorem 10.7.3(ii).


PROOF of Theorem 10.7.3(ii). To prove the result we use (B.6) and Lemma B.1.1.

321

Bibliography
Hong-Zhi An, Zhao-Guo. Chen, and E.J. Hannan. Autocorrelation, autoregression and autoregressive approximation. Ann. Statist., 10:926936, 1982.
A. Aue, L. Horvath, and J. Steinbach. Estimation in random coefficient autoregressive models.
Journal of Time Series Analysis, 27:6176, 2006.
K. I. Beltrao and P. Bloomfield. Determining the bandwidth of a kernel spectrum estimate. Journal
of Time Series Analysis, 8:2338, 1987.
I. Berkes, L. Horv
ath, and P. Kokoskza. GARCH processes: Structure and estimation. Bernoulli,
9:20012007, 2003.
I. Berkes, L. Horvath, P. Kokoszka, and Q. Shao. On discriminating between long range dependence
and changes in mean. Ann. Statist., 34:11401165, 2006.
R.N. Bhattacharya, V.K. Gupta, and E. Waymire. The hurst effect under trend. J. Appl. Probab.,
20:649662, 1983.
P. Billingsley. Probability and Measure. Wiley, New York, 1995.
T Bollerslev. Generalized autoregressive conditional heteroscedasticity. J. Econometrics, 31:301
327, 1986.
P. Bougerol and N. Picard. Stationarity of GARCH processes and some nonnegative time series.
J. Econometrics, 52:115127, 1992a.
P. Bougerol and N Picard. Strict stationarity of generalised autoregressive processes. Ann. Probab.,
20:17141730, 1992b.

322

G. E. P. Box and G. M. Jenkins. Time Series Analysis, Forecasting and Control. Cambridge
University Press, Oakland, 1970.
A. Brandt. The stochastic equation Yn+1 = An Yn + Bn with stationary coefficients. Adv. in Appl.
Probab., 18:211220, 1986.
W.L. Briggs and V. E. Henson. The DFT: An Owners manual for the Discrete Fourier Transform.
SIAM, Philadelphia, 1997.
D.R. Brillinger. Time Series: Data Analysis and Theory. SIAM Classics, 2001.
P. Brockwell and R. Davis. Time Series: Theory and Methods. Springer, New York, 1998.
W. W. Chen, C. Hurvich, and Y. Lu. On the correlation matrix of the discrete Fourier Transform
and the fast solution of large toeplitz systems for long memory time series. Journal of the
American Statistical Association, 101:812821, 2006.
R. Dahlhaus and D. Janas. A frequency domain bootstrap for ratio statistics in time series analysis.
Ann. Statistic., 24:19341963, 1996.
R. Dahlhaus and S. Subba Rao. A recursive online algorithm for the estimation of time-varying
arch parameters. Bernoulli, 13:389422, 2007.
J Davidson. Stochastic Limit Theory. Oxford University Press, Oxford, 1994.
J. Dedecker and P. Doukhan. A new covariance inequality. Stochastic Processes and their applications, 106:6380, 2003.
R. Douc, E. Moulines, and D. Stoffer. Nonlinear Time Series: Theory, Methods and Applications
with R Examples. Chapman and Hall, 2014.
Y. Dwivedi and S. Subba Rao. A test for second order stationarity based on the discrete fourier
transform. Journal of Time Series Analysis, 32:6891, 2011.
R. Engle. Autoregressive conditional heteroscedasticity with estimates of the variance of the United
Kingdom inflation. Econometrica, 50:9871006, 1982.
J. C. Escanciano and I. N Lobato. An automatic Portmanteau test for serial correlation. Journal
of Econometrics, 151:140149, 2009.

323

J. Fan and Q. Yao. Nonlinear time series: Nonparametric and parametric methods. Springer,
Berlin, 2003.
W. Fuller. Introduction to Statistical Time Series. Wiley, New York, 1995.
C. W. J. Granger and A. P. Andersen. An introduction to Bilinear Time Series models. Vandenhoek
and Ruprecht, G
ottingen, 1978.
U. Grenander and G. Szeg
o. Toeplitz forms and Their applications. Univ. California Press, Berkeley,
1958.
P Hall and C.C. Heyde. Martingale Limit Theory and its Application. Academic Press, New York,
1980.
E.J. Hannan and Rissanen. Recursive estimation of ARMA order. Biometrika, 69:8194, 1982.
J. Hart. Kernel regression estimation with time series errors. Journal of the Royal Statistical
Society, 53:173187, 1991.
C. Jentsch and S. Subba Rao. A test for second order stationarity of multivariate time series.
Journal of Econometrics, 2014.
D. A. Jones. Nonlinear autoregressive processes. Proceedings of the Royal Society (A), 360:7195,
1978.
Rosenblatt. M. and U. Grenander. Statistical Analysis of Stationary Time Series. Chelsea Publishing Co, 1997.
T. Mikosch. Elementary Stochastic Calculus With Finance in View. World Scientific, 1999.
T. Mikosch and C. St
aric
a. Is it really long memory we see in financial returns? In P. Embrechts,
editor, Extremes and Integrated Risk Management, pages 149168. Risk Books, London, 2000.
T. Mikosch and C. St
aric
a. Long-range dependence effects and arch modelling. In P. Doukhan,
G. Oppenheim, and M.S. Taqqu, editors, Theory and Applications of Long Range Dependence,
pages 439459. Birkh
auser, Boston, 2003.
F. Moricz. Moment inequalities and the strong law of large numbers. Z. Wahrsch. verw. Gebiete,
35:298314, 1976.

324

D.F. Nicholls and B.G. Quinn.

Random Coefficient Autoregressive Models, An Introduction.

Springer-Verlag, New York, 1982.


E. Parzen. On consistent estimates of the spectrum of a stationary process. Ann. Math. Statist.,
1957.
E. Parzen. On estimation of the probability density function and the mode. Ann. Math. Statist.,
1962.
M. Pourahmadi. Foundations of Time Series Analysis and Prediction Theory. Wiley, 2001.
M. B. Priestley. Spectral Analysis and Time Series: Volumes I and II. Academic Press, London,
1983.
B.G. Quinn and E.J. Hannan. The Estimation and Tracking of Frequency. Cambridge University
Press, 2001.
X. Shao. A self-normalized approacj to confidence interval construction in time series. Journal of
the Royal Statistical Society (B), 72:343366, 2010.
R. Shumway and D. Stoffer. Time Series Analysis and Its applications: With R examples. Springer,
New York, 2006.
D. Straumann. Estimation in Conditionally Heteroscedastic Time Series Models. Springer, Berlin,
2005.
S. Subba Rao. A note on uniform convergence of an arch() estimator. Sankhya, pages 600620,
2006.
T. Subba Rao. On the estimation of bilinear time series models. In Bull. Inst. Internat. Statist.
(paper presented at 41st session of ISI, New Delhi, India), volume 41, 1977.
T. Subba Rao. On the theory of bilinear time series models. Journal of the Royal Statistical
Society(B), 43:244255, 1981.
T. Subba Rao and M. M. Gabr. An Introduction to Bispectral Analysis and Bilinear Time Series
Models. Lecture Notes in Statistics (24). Springer, New York, 1984.
S. C. Taylor. Modelling Financial Time Series. John Wiley and Sons, Chichester, 1986.

325

Gy. Terdik. Bilinear Stochastic Models and Related Problems of Nonlinear Time Series Analysis;
A Frequency Domain Approach, volume 142 of Lecture Notes in Statistics. Springer Verlag, New
York, 1999.
M. Vogt. Nonparametric regression for locally stationary time series. Annals of Statistics, 40:
26012633, 2013.
A. M. Walker. On the estimation of a harmonic component in a time series with stationary
independent residuals. Biometrika, 58:2136, 1971.
P. Whittle. Gaussian estimation in stationary time series. Bulletin of the International Statistical
Institute, 39:105129, 1962.

326