Introduction to Time Series Analysis. Lecture 15.

Spectral Analysis
1. Spectral density: Facts and examples.
2. Spectral distribution function.
3. Wolds decomposition.

Spectral Analysis
Idea: decompose a stationary time series {Xt } into a combination of
sinusoids, with random (and uncorrelated) coefficients.
Just as in Fourier analysis, where we decompose (deterministic) functions
into combinations of sinusoids.
This is referred to as spectral analysis or analysis in the frequency
domain, in contrast to the time domain approach we have considered so far.
The frequency domain approach considers regression on sinusoids; the time
domain approach considers regression on past values of the time series.

A periodic time series

Consider
Xt = A sin(2t) + B cos(2t)
= C sin(2t + ),
where A, B are uncorrelated, mean zero, variance 2 = 1, and
C 2 = A2 + B 2 , tan = B/A. Then
t = E[Xt ] = 0
(t, t + h) = cos(2h).
So {Xt } is stationary.

An aside: Some trigonometric identities

sin
,
cos
sin2 + cos2 = 1,
tan =

sin(a + b) = sin a cos b + cos a sin b,

cos(a + b) = cos a cos b sin a sin b.

A periodic time series

For Xt = A sin(2t) + B cos(2t), with uncorrelated A, B
(mean 0, variance 2 ), (h) = 2 cos(2h).
The autocovariance of the sum of two uncorrelated time series is the sum of
their autocovariances. Thus, the autocovariance of a sum of random
sinusoids is a sum of sinusoids with the corresponding frequencies:
Xt =

k
X

(Aj sin(2j t) + Bj cos(2j t)) ,

j=1

(h) =

k
X

j2 cos(2j h),

j=1

where Aj , Bj are uncorrelated, mean zero, and Var(Aj ) = Var(Bj ) = j2 .

5

A periodic time series

Xt =

k
X

(Aj sin(2j t) + Bj cos(2j t)) ,

(h) =

k
X

j2 cos(2j h).

j=1

j=1

Thus, we can represent (h) using a Fourier series. The coefficients are the
variances of the sinusoidal components.
The spectral density is the continuous analog: the Fourier transform of .
(The analogous spectral representation of a stationary process Xt involves
a stochastic integrala sum of discrete components at a finite number of
frequencies is a special case. We wont consider this representation in this
course.)

Spectral density
If a time series {Xt } has autocovariance satisfying
P
h= |(h)| < , then we define its spectral density as
f () =

(h)e2ih

h=

for < < .

Spectral density: Some facts

P



(h)e2ih  < .

1. We have h=
This is because |ei | = | cos + i sin | = (cos2 + sin2 )1/2 = 1,
and because of the absolute summability of .
2. f is periodic, with period 1.
This is true since e2ih is a periodic function of with period 1.
Thus, we can restrict the domain of f to 1/2 1/2. (The text
does this.)

Spectral density: Some facts

3. f is even (that is, f () = f ()).
To see this, write
f () =

1
X

(h)e2ih + (0) +

h=

f () =

1
X

(h)e2ih ,

h=1

(h)e2i(h) + (0) +

(h)e2i(h) ,

h=1

h=

(h)e2ih + (0) +

h=1

1
X

h=

= f ().
4. f () 0.
9

(h)e2ih

Spectral density: Some facts

5. (h) =
Z

1/2

e2ih f () d.
1/2

1/2

e2ih f () d =

1/2

1/2

e2i(jh) (j) d

1/2 j=

(j)

j6=h

= (h) +

1/2

e2i(jh) d

1/2

j=

= (h) +



(j)
ei(jh) ei(jh)
2i(j h)

X (j) sin((j h))

j6=h

10

(j h)

= (h).

Example: White noise

2
For white noise {Wt }, we have seen that (0) = w
and (h) = 0 for
h 6= 0.

Thus,
f () =

(h)e2ih

h=
2
= (0) = w
.

That is, the spectral density is constant across all frequencies: each
frequency in the spectrum contributes equally to the variance. This is the
origin of the name white noise: it is like white light, which is a uniform
mixture of all frequencies in the visible spectrum.
11

Example: AR(1)
|h|

2
For Xt = 1 Xt1 + Wt , we have seen that (h) = w
1 /(1 21 ). Thus,

2
X

|h| 2ih
w
f () =
(h)e2ih =

1 e
1 21
h=
h=
!

2
X


w
h
2ih
2ih
1+
1 e
=
+e
2
1 1
h=1


2i
2i
2
1 e
1 e
w
1+
+
=
1 21
1 1 e2i
1 1 e2i
2
w
1 1 e2i 1 e2i
=
(1 21 ) (1 1 e2i )(1 1 e2i )

2
w
.
=
2
1 21 cos(2) + 1

12

Examples
2
White noise: {Wt }, (0) = w
and (h) = 0 for h 6= 0.
2
f () = (0) = w
.
|h|

2
1 /(1 21 ).
AR(1): Xt = 1 Xt1 + Wt , (h) = w

f () =

2
w
.
121 cos(2)+21

If 1 > 0 (positive autocorrelation), spectrum is dominated by low

frequency componentssmooth in the time domain.
If 1 < 0 (negative autocorrelation), spectrum is dominated by high
frequency componentsrough in the time domain.

13

Example: AR(1)
Spectral density of AR(1): Xt = +0.9 Xt1 + Wt
100

90

80

70

f()

60

50

40

30

20

10

0.05

0.1

0.15

0.2

0.25

14

0.3

0.35

0.4

0.45

0.5

Example: AR(1)
Spectral density of AR(1): Xt = 0.9 Xt1 + Wt
100

90

80

70

f()

60

50

40

30

20

10

0.05

0.1

0.15

0.2

0.25

15

0.3

0.35

0.4

0.45

0.5