' 4 The Periodogram

Statistics 626
$
4.1 Motivation

A basic idea in mathematics and statistics is to take a complicated object

(such as a time series) and break it up into the sum of simple objects
that can be studied separately, see which ones can be thrown away as
being unimportant, and then adding what’s left back together again to
obtain an approximation to the original object. The periodogram of a time
series is the result of such a procedure.

4.2 Sinusoidal Decomposition of a Time Series

Given a time series data set of length n (assumed to be even for now for
convenience), it is possible to find cosines and sines of periods
n, n/2, . . . , n/(n/2) = 2 that when added together along with x̄
gives the data set back again.

Mathematically, we have

X
n/2+1
γk [ak cos(2π(t − 1)ωk ) + bk sin(2π(t − 1)ωk )] ,

& %
x(t)−x̄ =
k=2

where ωk = (k − 1)/n, γk is two unless k = n/2 + 1 when it is one,

and ak and bk are the real and imaginary parts of the discrete Fourier

Topic 4: The Periodogram Copyright

c 1999 by H.J. Newton Slide 1
' Statistics 626

transform z(1), . . . , z(n) of the data:

$
X
n
z(k) = x(t)e2πi(t−1)ωk , k = 1, . . . , n,
j=1

If we define

Ck2 = a2k + b2k , φk = arctan(bk /ak ),

we can also write

X
n/2+1
x(t) − x̄ = γk Ck cos(2π(t − 1)ωk − φk ).
k=2

Further,
1X X
n n/2+1
R̂(0) = (x(t) − x̄) =
2
γk Ck2 .
n t=1
k=2

There are similar formulas for when n is odd (see Theorem 1.4.2 of the
text).

&
Topic 4: The Periodogram Copyright
c 1999 by H.J. Newton
%
Slide 2
' 4.3 Definition of Periodogram
Statistics 626
$
Def: A plot of nCk2 versus ωk = (k − 1)/n for k = 1, . . . , [n/2] + 1
is called the periodogram of a data set. The function
 2
 X
n 
1  
fˆ(ω) =  x(t)e2πi(t−1)ω
 , ω ∈ [0, .5],
n t=1


and fˆ(ω) = fˆ(1 − ω) for ω ∈ [0.5, 1] is called the sample spectral

density of the data set.

Note: It is customary to plot log(nCk2 /R̂(0)) vs. ωk rather than nCk2 .

&
Topic 4: The Periodogram Copyright
c 1999 by H.J. Newton
%
Slide 3
' 4.4 Interpreting the Periodogram
Statistics 626
$
1. The periodogram is a very useful tool for describing a time series
data set. We will see it is much more useful than the correlogram but
it does require some training to interpret properly.

2. We will use the terms low frequency and high frequency extensively.
The basic idea is that sinusoids of low frequency (or equivalently
long period) are smooth in appearance whereas those of high
frequency (or short period) are very wiggly. Thus if a time series
appears to be very smooth (wiggly), then the values of the
periodogram for low (high) frequencies will be large relative to its
other values and we will say that the data set has an excess of low
(high) frequency. For a purely random series, all of the sinusoids
should be of equal importance and thus the periodogram will vary
randomly around a constant.

3. If a time series has a strong sinusoidal signal for some frequency,

then there will be a peak in the periodogram at that frequency.

4. If a time series has a strong nonsinusoidal signal for some frequency,

then there will be a peak in the periodogram at that frequency but
also peaks at some multiples of that frequency. The first frequency is

& %
called the fundamental frequency and the others called harmonics.

5. One of the aims of this course is to have you describe what a data
set must look like when you can only see its periodogram (see the
old exams for the course for examples).

Topic 4: The Periodogram Copyright

c 1999 by H.J. Newton Slide 4
' Statistics 626

4.5 Examples of the Periodogram

$

&
Topic 4: The Periodogram Copyright
c 1999 by H.J. Newton
%
Slide 5
' 4.6 Cumulative Periodogram
Statistics 626
$
A useful tool for describing the overall behavior of the periodogram (and
thus the data set) is the cumulative periodogram
Pk
j=1 fˆ(ωj )
F̂ (ωk ) = Pq ˆ
, k = 1, . . . , q = [n/2] + 1.
j=1 f (ωj )

Note that F̂ starts out near zero at ω1 = 0 and must grow to be equal to
one at ωq = 0.5 since we are accumulating quantities that are
nonnegative.

Note that for a purely random series, F̂ should follow along a line from
(0, 0) to (0.5, 1). For a series having excess of low (high) frequency, F̂
will start out above (below) that y = 2x line, while there will be a jump in
F̂ at any frequency where fˆ has a peak.

&
Topic 4: The Periodogram Copyright
c 1999 by H.J. Newton
%
Slide 6
' Statistics 626

4.7 Examples of Cumulative Periodogram

$

&
Topic 4: The Periodogram Copyright
c 1999 by H.J. Newton
%
Slide 7