MATH11131 Part 4: Frequency analysis

⊲ MATH11131
analysis
Part 4: Frequency

Frequency analysis

Ioannis Papastathopoulos
School of Mathematics, University of Edinburgh

1 / 24
MATH11131 Part 4: Frequency
analysis

⊲ Frequency analysis
Motivation
Lutenizing hormone data
Periodogram
Example: Sine wave with white noise
Example: Sine wave with white noise
AR(1), 0.9
AR(1), -0.9
Example
Comments
Reminder: Multivariate normal
distribution
Analysis of variance
Spectral analysis of variance Frequency analysis
Properties of the periodogram
A trigonometric lemma
Reminder: Basic definitions
Spectrum
Examples
Comments
Spectral distribution
Example
Scaling
Frequency domain analysis

2 / 24
 Motivation

MATH11131 Part 4: Frequency

analysis  Many series have periodic structure (e.g., sunspots, CO2 data, . . .), but we may not
Frequency analysis
⊲ Motivation know what the frequencies are in advance of looking at the data.
Lutenizing hormone data
Periodogram
Example: Sine wave with white noise  The periodogram is a summary description based on representing the observed series
Example: Sine wave with white noise
AR(1), 0.9 as a superposition of sine and cosine waves of various frequencies.
AR(1), -0.9
Example
Comments
Reminder: Multivariate normal  The idea is that the periodogram will tell us what frequencies are most important.
distribution
Analysis of variance
Spectral analysis of variance
Properties of the periodogram
 Consider first the simple model
A trigonometric lemma
Reminder: Basic definitions
Spectrum Yt = α cos(2πωt) + β sin(2πωt) + εt , t = 1, . . . , n, (1)
Examples
Comments
Spectral distribution
Example where {εt } is white noise, ω = 1/p is the known frequency of the fluctuations, p is
Scaling
Frequency domain analysis their period, and α, β are unknown parameters to be estimated by least squares.

 As we can write

α cos(2πωt) + β sin(2πωt) = (α2 + β 2 )1/2 sin(2πωt + γ), γ = tan−1 (α/β),

the right-hand side of (1) is equivalent to any sinusoidal function with frequency ω.

3 / 24
 Lutenizing hormone data

MATH11131 Part 4: Frequency

analysis Data on lutenizing hormone in n = 48 succesive blood samples from a woman, taken at
Frequency analysis
Motivation
10-minute intervals. Fourier series for n = 48 are shown in the right panel.
⊲ Lutenizing hormone data
Periodogram
Example: Sine wave with white noise
Example: Sine wave with white noise
AR(1), 0.9
AR(1), -0.9
Example

100
Comments

3.5
Reminder: Multivariate normal
distribution
Analysis of variance
Spectral analysis of variance

80
Properties of the periodogram 3.0
A trigonometric lemma
Reminder: Basic definitions

Fourier Series

60
Spectrum
Examples
2.5
lh

Comments
Spectral distribution

40
Example
Scaling
2.0

Frequency domain analysis

20
1.5

0
0 10 20 30 40 0 10 20 30 40

Time Time

4 / 24
 Periodogram

MATH11131 Part 4: Frequency

analysis Definition 1.(a) If y1 , . . . , yn is an equally-spaced time series, its periodogram ordinate
Frequency analysis
Motivation
for ω is defined as
Lutenizing hormone data
⊲ Periodogram
( 
n
) 2 ( n
) 2
Example: Sine wave with white noise X X
Example: Sine wave with white noise
AR(1), 0.9
I(ω) = n−1  yt sin(2πωt) + yt cos(2πωt)  , 0 < ω < 1/2.
AR(1), -0.9
t=1 t=1
Example
Comments
Reminder: Multivariate normal
distribution
Analysis of variance
(b) The periodogram  is a plotof I(j/n) for the Fourier frequencies 2πj/n for
Spectral analysis of variance
Properties of the periodogram
j = 1, . . . , m = (n − 1)/2 ; I(1/2) is included only if n is even. By default R plots
A trigonometric lemma
Reminder: Basic definitions
the log-periodogram, log I(j/n), against j/n.
Spectrum
Examples
Comments (c) The cumulative periodogram
Spectral distribution
Example Pr
j=1 I(j/n)
Scaling
Frequency domain analysis
Cr = Pm ,
l=1 I(l/n)

is a plot of C1, . . . , Cm against the frequencies j/n for j = 1, . . . , m.

5 / 24
 Example: Sine wave with white noise

MATH11131 Part 4: Frequency

analysis Top: data from a simulated sine wave with added white noise. Bottom: log periodogram
Frequency analysis
Motivation
with red horizontal line showing noise variance σ 2 = 0.25, and a green vertical line showing
Lutenizing hormone data
Periodogram
the signal frequency 1/200. The blue line shows the width of a 95% confidence interval for
⊲ Example: Sine wave with white
noise
the true value at each point.
Example: Sine wave with white noise
AR(1), 0.9 Sine wave with period 200 and noise variance 0.25
AR(1), -0.9
Example

2
Comments

1
Reminder: Multivariate normal
distribution

0
Analysis of variance

−2 −1
Spectral analysis of variance
Properties of the periodogram
A trigonometric lemma
Reminder: Basic definitions 0 100 200 300 400 500
Spectrum
Examples Index
Comments
Spectral distribution
Series: x
Example
Raw Periodogram
Scaling
Frequency domain analysis
1e+01
spectrum

1e−01
1e−03

0.0 0.1 0.2 0.3 0.4 0.5

frequency
bandwidth = 0.000577

6 / 24
 Example: Sine wave with white noise

MATH11131 Part 4: Frequency

analysis Top: data from a simulated sine wave with added white noise. Bottom: log periodogram
Frequency analysis
Motivation
with red horizontal line showing noise variance σ 2 = 1, and a green vertical line showing the
Lutenizing hormone data
Periodogram
signal frequency 1/20. The blue line shows the width of a 95% confidence interval for the
Example: Sine wave with white noise
Example: Sine wave with white
true value at each point.
⊲ noise
AR(1), 0.9 Sine wave with period 20 and noise variance 0.25
AR(1), -0.9
Example
Comments

2
Reminder: Multivariate normal
distribution

0
Analysis of variance
Spectral analysis of variance

−2
Properties of the periodogram
A trigonometric lemma
Reminder: Basic definitions 0 100 200 300 400 500
Spectrum
Examples Index
Comments
Spectral distribution
Series: x
Example
Raw Periodogram
Scaling

1e+02
Frequency domain analysis
spectrum

1e+00
1e−02

0.0 0.1 0.2 0.3 0.4 0.5

frequency
bandwidth = 0.000577

7 / 24
 AR(1), 0.9

MATH11131 Part 4: Frequency

analysis Data from a simulated AR(1) model with parameter 0.9, with log-periodogram and
Frequency analysis
Motivation
theoretical value (in red). The blue line shows the width of a 95% confidence interval for
Lutenizing hormone data
Periodogram
the true value at each point. The log scale on the vertical axis means there is a very large
Example: Sine wave with white noise
Example: Sine wave with white noise
change in the periodogram itself.
⊲ AR(1), 0.9 AR(1), 0.9
AR(1), -0.9
Example
Comments

4
Reminder: Multivariate normal

2
distribution
Analysis of variance

−2
Spectral analysis of variance
Properties of the periodogram
A trigonometric lemma

−6
Reminder: Basic definitions
Spectrum 0 100 200 300 400 500
Examples
Time
Comments
Spectral distribution
Example Series: x
Scaling Raw Periodogram
Frequency domain analysis

1e+01
spectrum

1e−03

0.0 0.1 0.2 0.3 0.4 0.5

frequency
bandwidth = 0.000577

8 / 24
 AR(1), -0.9

MATH11131 Part 4: Frequency

analysis Data from a simulated AR(1) model with parameter −0.9, with log-periodogram and
Frequency analysis
Motivation
theoretical value (in red).
Lutenizing hormone data
Periodogram AR(1), −0.9
Example: Sine wave with white noise

2 4 6
Example: Sine wave with white noise
AR(1), 0.9
⊲ AR(1), -0.9

Y
Example

−2
Comments
Reminder: Multivariate normal

−6
distribution
Analysis of variance
Spectral analysis of variance 0 100 200 300 400 500
Properties of the periodogram
Time
A trigonometric lemma
Reminder: Basic definitions
Spectrum Series: x
Examples Raw Periodogram
Comments
Spectral distribution

1e+01
Example

spectrum
Scaling
Frequency domain analysis

1e−03

0.0 0.1 0.2 0.3 0.4 0.5

frequency
bandwidth = 0.000577

9 / 24
 Example

MATH11131 Part 4: Frequency

analysis Data on lutenizing hormone in n = 48 blood samples at 10-minute intervals from a human
Frequency analysis
Motivation
female. Top left: data; top right: periodogram; bottom left: possible Fourier series for
Lutenizing hormone data
Periodogram
n = 48; bottom right: cumulative periodogram
Example: Sine wave with white noise Series: x
Example: Sine wave with white noise Raw Periodogram
AR(1), 0.9

3.5

1.00
AR(1), -0.9
⊲ Example

3.0
Comments

0.20
spectrum
Reminder: Multivariate normal

2.5
lh
distribution
Analysis of variance

0.05
2.0
Spectral analysis of variance
Properties of the periodogram

0.01
1.5
A trigonometric lemma
Reminder: Basic definitions
Spectrum 0 10 20 30 40 0.1 0.2 0.3 0.4 0.5
Examples Time frequency
Comments bandwidth = 0.00601
Spectral distribution
Example Series: lh
Scaling
100

1.0
Frequency domain analysis

0.8
80
Fourier Series

0.6
60

0.4
40

0.2
20

0.0
0

0 10 20 30 40 0.0 0.1 0.2 0.3 0.4 0.5

Time frequency

10 / 24
 Comments
 Low frequency variation (trend) appears at the left of the periodogram, and high
MATH11131 Part 4: Frequency
analysis
frequency variation (rapid oscillations) appears at the right.
Frequency analysis
Motivation
 The rationale for considering only the frequencies ω = 2πj/n is that
Lutenizing hormone data
Periodogram
Example: Sine wave with white noise
n
X m
X
Example: Sine wave with white noise
AR(1), 0.9
yt2 = I(0) + 2 I(j/n) + I(1/2), (2)
AR(1), -0.9
Example
t=1 j=1
⊲Reminder:
Comments
Multivariate normal
distribution with I(1/2) Pincluded only if n is even. Thus the periodogram decomposes the total
Analysis of variance
Spectral analysis of variance variability yt2 of the data into components associated with each of these frequencies,
Properties of the periodogram
A trigonometric lemma plus one for the grand mean I(0) = nȳ 2 which we ignore because it is not periodic.
Reminder: Basic definitions
Spectrum
Examples  The rationale for plotting the log periodogram is that the periodogram ordinates are
Comments
Spectral distribution (roughly) exponentially distributed, and the log-transformation is variance-stabilising for
Example
Scaling the exponential distribution. A rough significance scale for the log-periodogram is shown
Frequency domain analysis
by the vertical line on its right.
 The cumulative periodogram provides a visual test of whether the series white noise.
We compare Cr with its expected value r/m: a large value of the
Kolmogorov-Smirnov statistic D = max|Cr − r/m| suggests that the underlying
series is not white noise. The test involves seeing whether the cumulative periodogram
falls outside a diagonal band, whose width determines the size of the test.

11 / 24
 Reminder: Multivariate normal distribution
Definition 2. Let µ = (µ1 , . . . , µn )T ∈ Rn and Σ be n × n positive definite matrix with
MATH11131 Part 4: Frequency
analysis
elements σjk . Then we say that the random vector X = (X1 , . . . , Xn )T with density
Frequency analysis  
Motivation 1 1 T −1 n
Lutenizing hormone data
f (x; µ, Σ) = n/2 1/2
exp − (x − µ) Σ (x − µ) , x, µ ∈ R ;
Periodogram
Example: Sine wave with white noise
(2π) |Σ| 2
Example: Sine wave with white noise
AR(1), 0.9
AR(1), -0.9 has a multivariate normal distribution with expectation vector µn×1 and covariance
Example
Comments matrix Σn×n .
Reminder: Multivariate normal
⊲ distribution
Analysis of variance Some facts:
Spectral analysis of variance
Properties of the periodogram
A trigonometric lemma
 the distribution is determined by its mean and covariances
Reminder: Basic definitions
Spectrum
Examples
Comments
E(Xj ) = µj , var(Xj ) = σjj , cov(Xj , Xk ) = σjk , j 6= k
Spectral distribution
Example
Scaling and Xi independent of Xj ⇒ σjk = 0;
Frequency domain analysis
iid
 if X1 , . . . , Xn ∼ N (µ, σ 2 ), then Xn×1 = (X1 , . . . , Xn )T ∼ Nn (µ1n , σ 2 In );
 linear combinations of normal variables are normal:

ar×1 + Qr×n X ∼ Nr (a + Qµ, QΣQT )

where we suppose that |QΣQT |> 0, so r ≤ n.

12 / 24
 Analysis of variance

 Suppose Yn×1 ∼ Nn (µ, σ 2 In )

MATH11131 Part 4: Frequency
analysis  Let the first column of the orthogonal matrix Qn×n consist of ones; hence QT Q = D,
Frequency analysis
Motivation
say, is diagonal and positive definite.
Lutenizing hormone data
Periodogram
 Now I = D−1/2 QT QD−1/2 = AAT , say, where A = D−1/2 QT is orthonormal, giving the
Example: Sine wave with white noise
Example: Sine wave with white noise
analysis of variance decomposition
AR(1), 0.9
AR(1), -0.9 n
X n
X
Yt2 = Y T Y = Y T IY = Y T AT AY = Zj2 ,
Example
Comments
Reminder: Multivariate normal
distribution
t=1 j=1
⊲ Analysis of variance
Spectral analysis of variance
Properties of the periodogram
A trigonometric lemma where Z = (Z1 , . . . , Zn )T = AY = D−1/2 QT Y ∼ Nn (D−1/2 QT µ, σ 2 In ), so
Reminder: Basic definitions
Spectrum
Examples n
−1/2 T −1/2 T −1/2 T
X
Comments −1/2
Spectral distribution Z1 = d 1 q1 Y =n Yt , Z2 = d2 q2 Y, Z3 = d3 q3 Y, ...,
Example
Scaling t=1
Frequency domain analysis

ind −1/2 T
in an obvious notation, and Zj ∼ N (dj qj µ, σ 2 ).
 This implies that Zj2 have independent non-central σ 2 χ21 distributions, with
non-centrality parameters δj = (qjT µ)2 /(dj σ 2 ). In particular, if qjT µ = 0, i.e., the jth
column of Q is orthogonal to the mean vector µ, then Zj2 ∼ σ 2 χ21 .
 Different choices of Q will give different decompositions of the variation in Y .

13 / 24
 Spectral analysis of variance
 
 Write stj = sin(tωj ), ctj = cos(tωj ), ωj = 2πj/n, for j = 1, . . . , m = n/2 ,
MATH11131 Part 4: Frequency
analysis
t = 1, . . . , n. Then Lemma 1 tells us that the columns of the n × n matrix
Frequency analysis
1 s11 c11 s12 c12 · · · s1m c1m
 
Motivation
Lutenizing hormone data
Periodogram  1 s21 c21 s22 c22 · · · s2m c2m 
Example: Sine wave with white noise
Example: Sine wave with white noise
Q= .
 1 .. .
.. .
.. .
.. .
..
,

AR(1), 0.9
AR(1), -0.9
Example
1 sn1 cn1 sn2 cn2 · · · snm cnm
Comments
Reminder: Multivariate normal
distribution
Analysis of variance are orthogonal (and therefore span Rn ), and QT Q = diag(n, n/2, . . . , n/2) = D, say.
⊲ Spectral analysis of variance  With this choice of Q, we have
Properties of the periodogram
A trigonometric lemma
Reminder: Basic definitions ( n )2 ( n )2
Spectrum
2 X 2 X
Examples
Comments Z12 = nȲ 2 , Z2j
2
= Yt sin(ωj t) 2
Z2j+1 = Yt cos(ωj t)
Spectral distribution n t=1 n t=1
Example
Scaling
Frequency domain analysis 2 2
so I(j/n) = (Z2j + Z2j+1 )/2, etc., and this leads to the decomposition (2), i.e.,
n
X m
X
Yt2 = I(0) + 2 I(j/n) + I(1/2),
t=1 j=1

with I(1/2) included only if n is even.

14 / 24
 Properties of the periodogram

MATH11131 Part 4: Frequency

iid
analysis

Frequency analysis
Theorem 1. If Y1 , . . . , Yn ∼ N (µ, σ 2 ), then all the periodogram ordinates are independent,
Motivation
Lutenizing hormone data
and
Periodogram
Example: Sine wave with white noise (a) I(1/n), . . . , I(m/n) are exponential random variables with mean σ 2 ;
Example: Sine wave with white noise
AR(1), 0.9
AR(1), -0.9
Example
(b) if n is even, I(1/2) ∼ σ 2 χ21 ; finally,
Comments
Reminder: Multivariate normal
distribution
Analysis of variance
(c) the cumulative periodogram ordinate
Spectral analysis of variance
⊲ Properties of the periodogram Pr
A trigonometric lemma j=1 I(j/n)
Reminder: Basic definitions Cr = Pm , r = 1, . . . , m,
l=1 I(l/n)
Spectrum
Examples
Comments
Spectral distribution
Example
has a beta(r, m − r) distribution, and so has mean and variance r/m, r(m − r)/m
Scaling
Frequency domain analysis
Part (a) says that Gaussian white noise has a flat spectrum. In fact the assumption of
Gaussianity is needed only for the independence, and the spectrum of non-Gaussian white
noise is flat.

15 / 24
 A trigonometric lemma
Lemma 1.(a) Let ωj = 2πj/n, for n ∈ N and positive integer j < n/2, and write
MATH11131 Part 4: Frequency
ct = cos(tωj ), st = sin(tωj ). Then
analysis

Frequency analysis n n n n n
Motivation
X X X X X
Lutenizing hormone data
Periodogram
ct = st = st ct = 0, s2t = c2t = n/2.
Example: Sine wave with white noise t=1 t=1 t=1 t=1 t=1
Example: Sine wave with white noise
AR(1), 0.9
AR(1), -0.9
Example
(b) If ω1 = 2πj1/n, ω2 = 2πj2 /n for positive integers j1 6= j2 < n/2, and we write
Comments
Reminder: Multivariate normal
ckt = cos(tωk ), skt = sin(tωk ) for k = 1, 2, then
distribution
Analysis of variance
Spectral analysis of variance
n
X n
X n
X
Properties of the periodogram
⊲ A trigonometric lemma s1t s2t = s1t c2t = c1t c2t = 0.
Reminder: Basic definitions
t=1 t=1 t=1
Spectrum
Examples
Comments
Spectral distribution
(c) If n is odd, and we write
 stj = sin(tωj ), ctj = cos(tωj ), ωj = 2πj/n, for
Example
Scaling
j = 1, . . . , m = n/2 , t = 1, 2, . . . , n, then the columns of the n × n matrix
Frequency domain analysis

···
 
1 s11 c11 s12 c12 s1m c1m
1 s21 c21 s22 c22 ··· s2m c2m 
Q=
1 .. .. .. .. .. ,
. . . . . 
1 sn1 cn1 sn2 cn2 ··· snm cnm

are orthogonal, QT Q = diag(n, n/2, . . . , n/2) = D, say.

16 / 24
 Reminder: Basic definitions

MATH11131 Part 4: Frequency

analysis Definition 3.(a) A random function is a random variable {Y (t) : t ∈ R}.
Frequency analysis
Motivation
Lutenizing hormone data (b) The trend of {Yt } is the non-random function µ(t) = E{Y (t)}, and its
Periodogram
Example: Sine wave with white noise autocovariance function is
Example: Sine wave with white noise
AR(1), 0.9
AR(1), -0.9
Example γ(t, s) = cov{Y (t), Y (s)} = E [{Y (t) − µ(t)}{Y (s) − µ(s)}]
Comments
Reminder: Multivariate normal
distribution
Analysis of variance (c) The random function is stationary if µ(t) = µ and γ(t, s) = γ(|t − s|): the trend is
Spectral analysis of variance
Properties of the periodogram constant and the covariance between Y (t) and Y (s) depends only on their time
A trigonometric lemma
⊲ Reminder: Basic definitions separation |t − s|. In this case we can define the autocorrelation function
Spectrum
Examples
ρ(t) = γ(t, 0)/γ(0, 0).
Comments
Spectral distribution
Example (d) A random sequence is a collection of random variables {Yt } in which the time index
Scaling
Frequency domain analysis takes only integer values: t ∈ Z. We use a subscript notation to make this clear.

(e) A white noise sequence is a random sequence consisting of mutually independent

random variables each with mean zero and variance σ 2 .

17 / 24
 Spectrum

Definition 4.(a) The autocovariance generating function of a stationary random

MATH11131 Part 4: Frequency
analysis
sequence {Yt} with autocovariances γk = cov(Yt , Yt+k ) is
Frequency analysis
Motivation ∞
X
Lutenizing hormone data
Periodogram G(z) = γk z k , z ∈ C.
Example: Sine wave with white noise
Example: Sine wave with white noise k=−∞
AR(1), 0.9
AR(1), -0.9 P
(b) If h |γh |< ∞, then the spectrum of {Yt } is defined as
Example
Comments
Reminder: Multivariate normal
distribution
Analysis of variance ∞
X
Spectral analysis of variance
Properties of the periodogram f (ω) = G(e−2πiω ) = γk e−2πikω , ω ∈ R,
A trigonometric lemma
Reminder: Basic definitions k=−∞
⊲ Spectrum
Examples
Comments where i2 = −1; this may also be written as the real-valued function
Spectral distribution
Example
Scaling ∞
X
Frequency domain analysis
f (ω) = γ0 + 2 γk cos(2πkω), ω ∈ R.
k=1

The normalised spectrum is f ∗ (ω) = f (ω)/γ0 .

The spectrum provides a convenient summary of the second-order properties of the process
in a single function, and also shows the effect of linear operations on the series very simply.

18 / 24
 Examples

MATH11131 Part 4: Frequency

analysis Example 1. Find the spectrum and normalised spectrum for white noise
Frequency analysis
Motivation
Lutenizing hormone data Example 2. Show that the spectrum of the AR(1) process Yt = αYt−1 + εt with |α|< 1 is
Periodogram
Example: Sine wave with white noise
Example: Sine wave with white noise σ2
AR(1), 0.9
f (ω) = 2
,
AR(1), -0.9
Example
1 − 2α cos(2πω) + α
Comments
Reminder: Multivariate normal
distribution
Analysis of variance
and that the normalised spectrum is
Spectral analysis of variance
2
Properties of the periodogram
1 − α
f ∗ (ω) =
A trigonometric lemma
Reminder: Basic definitions
2
.
Spectrum 1 − 2α cos(2πω) + α
⊲ Examples
Comments
Spectral distribution
Example
Scaling
Frequency domain analysis

19 / 24
 Comments
 The form of the spectrum suggests that
MATH11131 Part 4: Frequency
analysis

Frequency analysis f (ω) = f (−ω), f (ω) = f (ω + k), k ∈ Z,

Motivation
Lutenizing hormone data
Periodogram
Example: Sine wave with white noise
so f need only be defined in 0 < ω < 1/2—recall the definition of the periodogram.
Example: Sine wave with white noise
AR(1), 0.9
 The variance of the average Ȳ of data Y1 , . . . , Yn from a stationary process satisfies
AR(1), -0.9
n−1 ∞
Example
( )
Comments X X
Reminder: Multivariate normal
distribution lim n var(Ȳ ) = lim γ0 + 2 (1 − h/n)γh = γ0 + 2 γh = f (0)
Analysis of variance n→∞ n→∞
Spectral analysis of variance h=1 h=1
Properties of the periodogram
A trigonometric lemma
Reminder: Basic definitions for large n. Thus the spectral density at the origin, if finite, equals the (rescaled)
Spectrum
Examples variance of Ȳ
⊲ Comments  The spectrum is the discrete Fourier transform of the autocovariance sequence.
Spectral distribution
Example
Scaling  The inverse Fourier transform gives the autocovariance function by
Frequency domain analysis
Z 1/2 Z 1/2
γk = e2πikω f (ω)dω = 2 cos(2πkω)f (ω)dω, k ∈ Z.
−1/2 0

R 1/2
 We easily see that −1/2 f (ω)dω = γ0 .
 The spectrum need not exist.

20 / 24
 Spectral distribution
Theorem 2 (No proof).(a) The numbers {γh }h∈Z are the autocovariance function of a
MATH11131 Part 4: Frequency stationary random sequence iff there exists a unique bounded function F defined on
analysis

Frequency analysis
[− 12 , 12 ] such that F (−1/2) = 0, F is right-continuous and non-decreasing, with
Motivation
Lutenizing hormone data
symmetric increments about zero, and
Periodogram
Example: Sine wave with white noise
Z
e2πihudF (u), h ∈ Z.
Example: Sine wave with white noise
AR(1), 0.9 γh =
AR(1), -0.9
(−1/2,1/2]
Example
Comments
Reminder: Multivariate normal
distribution The function F is called the spectral distribution of γh , and its derivative, if it exists,
Analysis of variance
Spectral analysis of variance called the spectral density function, or spectrum.
is P
Properties of the periodogram
A trigonometric lemma (b) If h |γh |< ∞, then f exists.
Reminder: Basic definitions
Spectrum (c) A function f (ω) defined on [−1/2, 1/2] is the spectrum of a stationary process if and
Examples R 1/2
Comments
⊲ Spectral distribution
only if f (ω) = f (−ω), f (ω) ≥ 0, and −1/2 f (ω)dω < ∞.
Example
Scaling
Frequency domain analysis  The symmetric increments property means that if 0 ≤ a < b ≤ 1/2, then
F (b) − F (a) = F (−a) − F (−b).
 The interpretation of F is that F (ω2) − F (ω1) measures the variation accounted for by
fluctuations in frequency in the interval (ω1 , ω2 ).
 Part (c) of the theorem suggests how we may construct covariance functions with
desired properties—by choosing f , which may be any scaled density function.

21 / 24
 Example

MATH11131 Part 4: Frequency

analysis Example 3. Show that the covariance function of the stationary random sequence given by
Frequency analysis
Motivation
iid
Lutenizing hormone data
Periodogram
Yt = U1 cos(2πω0 t) + U2 sin(2πω0t), U1 , U2 ∼ N (0, σ 2 )
Example: Sine wave with white noise
Example: Sine wave with white noise
AR(1), 0.9 may be written as
AR(1), -0.9
Example 
Comments
Reminder: Multivariate normal Z 1/2 0,
 ω < −ω0 ,
distribution
Analysis of variance
Spectral analysis of variance
γh = e2πiωh dF (ω), F (ω) = σ 2 /2, −ω0 ≤ ω < ω0 ,
Properties of the periodogram −1/2 σ 2 ,

A trigonometric lemma ω ≥ ω0
Reminder: Basic definitions
Spectrum
Examples
Comments
Spectral distribution
⊲ Example
Scaling
Frequency domain analysis

22 / 24
 Scaling

MATH11131 Part 4: Frequency

analysis  Books and computer programs are inconsistent about the scaling of the periodogram, or
Frequency analysis
Motivation
equivalently,
P about the choice of ζ in the definition of the spectrum
Lutenizing hormone data
Periodogram
f (ω) = k γk e−ζikω ,
Example: Sine wave with white noise
Example: Sine wave with white noise
AR(1), 0.9  Some possibilities are:
AR(1), -0.9
Example
Comments
– ζ = 2π, used in Shumway and Stoffer, so that the periodogram is graphed for
Reminder: Multivariate normal
distribution ω ∈ (0, 1/2);
Analysis of variance
Spectral analysis of variance
Properties of the periodogram
A trigonometric lemma
– ζ = 1, so that the periodogram is graphed for ω ∈ (0, π);
Reminder: Basic definitions
Spectrum
Examples
Comments
– ζ = 2π/c, where c is the frequency of the data per unit of time, used in R when
Spectral distribution c > 1, so that the periodogram is graphed for ω ∈ (0, c/2).
Example
⊲ Scaling
Frequency domain analysis
– Sometimes the spectral density is divided by the frequency c, so that its integral
remains the total variance.

 Conclusion: Read the documentation very carefully.

23 / 24
 Frequency domain analysis

MATH11131 Part 4: Frequency

analysis  Excellent introduction is Bloomfield (2000) Fourier Analysis of Time Series,
Frequency analysis
Motivation
Second edition, Wiley
Lutenizing hormone data
Periodogram
Example: Sine wave with white noise  The bible is Priestley (1981) Spectral Analysis and Time Series, Volumes 1 and
Example: Sine wave with white noise
AR(1), 0.9 2, Academic Press which also deals with spectra for non-stationary series.
AR(1), -0.9
Example
Comments
Reminder: Multivariate normal  See also Persival and Walden (1993) Spectral Analysis for Physical
distribution
Analysis of variance Applications: Multitape and Conventional Univariate Techniques, CUP
Spectral analysis of variance
Properties of the periodogram
A trigonometric lemma
Reminder: Basic definitions  More modern approach is use of wavelets, which decompose series in orthogonal
Spectrum
Examples functions that are local in time: see Percival and Walden (2000) Wavelet Methods
Comments
Spectral distribution for Time Series Analysis, CUP.
Example
Scaling
⊲ Frequency domain analysis

24 / 24