Scribd
Upload
59 views9 pages

9 Linear Filters: (X (t), t ∈ Z) (β, j ∈ Z) (Y (t), t ∈ Z) − j), t ∈ Z,

The document discusses linear time series models where one time series is a filtered version of another. It defines linear filters and provides examples such as moving averages and differencing. It then presents a filter theorem that describes how the properties of the output series, such as its spectral density and autocovariance function, relate to the properties of the input series and the filter coefficients. The document explores applications of the filter theorem as well as different types of linear filters.

Uploaded by

souvik5000
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
59 views9 pages

9 Linear Filters: (X (t), t ∈ Z) (β, j ∈ Z) (Y (t), t ∈ Z) − j), t ∈ Z,

The document discusses linear time series models where one time series is a filtered version of another. It defines linear filters and provides examples such as moving averages and differencing. It then presents a filter theorem that describes how the properties of the output series, such as its spectral density and autocovariance function, relate to the properties of the input series and the filter coefficients. The document explores applications of the filter theorem as well as different types of linear filters.

Uploaded by

souvik5000
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 9

'

Statistics 626

9 Linear Filters
Time series models are often of the form of one series being a ltered version of another.

Z } be a time series and let {j , j Z } be a set of real numbers. If the time series {Y (t), t Z } is dened by

Def: Let {X (t), t

Y (t) =
j =

j X (t j ), t Z,

then we say that Y is a linear ltered version of X having the s as the lter coefcients.

Ex1: The moving average smoother


K

Y (t) =
j =K

1 X (t j ), 2K + 1

has Y as a ltered version of X with coefcients

j =

&
Topic 9: Linear Filters

1/(2K + 1), if |j | K if |j | > K . 0,

Ex2: The dth differencing operator

Y (t) = X (t) X (t d), t Z

Copyright c 1999 by H.J. Newton

Slide 1

'

Statistics 626

has Y as a ltered version of X with coefcients

if j = 0 1, j = 1, if j = d 0, otherwise.

Ex3: The M A(1, , 2 ) process

X (t) = (t) + (t 1), t Z


has X as a ltered version of white noise with coefcients

1, if j = 0 j = , if j = 1 0, otherwise

&
Topic 9: Linear Filters

Copyright c 1999 by H.J. Newton

Slide 2

'

Statistics 626

9.1 The Filter Theorem


It is easy to prove a theorem that lets us nd the properties of the output of a lter from the properties of the input to the lter and the coefcients of the lter:

Filter Theorem: If Y is a ltered version of X with lter coefcients

{j , j Z }, then if X is covariance stationary, so is Y and


1. The spectral density fY of the output is related to the spectral density fX of the input by

fY ( ) = g(e2i ) fX ( ), [0, 1],


where the complex valued polynomial g (z ) is given by

g(z ) =
j =

j z j , z C.

Note that g (e2i ) is called the frequency transfer function of the lter.

&
Topic 9: Linear Filters

2. The autocovariance function RY of the output is related to the autocovariance function RX of the input by

RY (v ) =
k=

R (k )RX (v k ), v Z,

Copyright c 1999 by H.J. Newton

Slide 3

'

Statistics 626

where the sequence {R (k ), k

Z } is given by j j +k , k Z.

R (k ) =
j =

&
Topic 9: Linear Filters

Copyright c 1999 by H.J. Newton

Slide 4

'

Statistics 626

9.2 Using the Filter Theorem


We can derive a large number of facts from this simple theorem: 1. Since white noise is covariance stationary with spectral density

$
2

( ) = 2 and an MA process X is a ltered version of white noise, we know immediately that X is covariance stationary
function f with spectral density function

fX ( ) = 1 + e2i

2 ,

which a little algebra will show is the same as what we found in the previous topic by rst principles. We can use the lter theorem also to nd RX . Further, we can easily nd RX and fX for an MA process of order q

X (t) = (t) + 1 (t 1) + + q (t q ),
where

W N ( 2 ) as fX ( ) = 1 + 1 e2i + + q e2iq
2

2 ,

and

&
Topic 9: Linear Filters

RX (v ) =

q |v |

2
k=0

k k+|v| , if |v | q
if |v |

0,

>q

2. If we think of each realization of a time series as a sum of sinusoids and f ( ) as the average value of the amplitude of the sinusoid of

Copyright c 1999 by H.J. Newton

Slide 5

'

Statistics 626

frequency in the various realizations, then the function g (e2i ) is

useful in determining what the lter does to the sinusoid of frequency

in the input series. For example, for dth differencing, we have g(e2i ) = 1 e2i(d) , [0, 1], = 1 for any integer k . Thus at any frequency that is a multiple of 1/d, we have that g is zero and thus so is |g |2 and so is the spectral density of the output series. Thus the dth difference operator annhilates the sinusoids of frequencies that are multiple of 1/d in the input (which
is what we intend it to do). 3. Another example of what g tells us is the moving average smoother, for which which is zero whenever d is an integer since e2ik

gK (e

2i

1 )= 2K + 1

e2ij ,
j =K

where we placed the subscript K on g to indicate the moving average smoother parameter. This summation is a famous one and is called the Dirichlet kernel
K 2ij

&
Topic 9: Linear Filters

DK ( ) =
j =K

sin[(K + 1 2 )2 ] = . sin
2

Thus

fY ( ) =

1 2K + 1

|DK ( )|2 fX ( ).

Copyright c 1999 by H.J. Newton

Slide 6

'

Statistics 626

Graphs of |gK |2 are given in the gure below for

K = 2, 4, 6, . . . , 20 (note that g gets narrower as K increases).

&
Topic 9: Linear Filters

Copyright c 1999 by H.J. Newton

Slide 7

'

Statistics 626

9.3 Low Pass, High Pass, and Band Pass Filters


We can categorize lters by their effect in the frequency domain. A lter whose frequency transfer function is zero for frequencies greater than some frequency 0 only allows sinusoids of frequencies less than 0 to

pass through the lter. Thus the lter is called a low pass lter. A lter that allows only high frequencies to pass through is called a high pass lter, while if only frequencies in the band of frequencies 0

are

allowed to pass through the lter, it is called a band pass lter.

Since g is the Fourier transform of the s, we can can nd the s for a given g (such as a band pass lter) by
1

j =
0

g(e2i )e2ij d.

we must integrate g(e2i ) = 1 for [0 , 0 + ] or [1 (0 + ), 1 (0 )] and 0 otherwise since f ( ) = f (1 ) for [0.5, 1]. If you do the integral, you nd
Ex: If we want a band pass lter at 0

&
Topic 9: Linear Filters

j =

if j = 0 4, 2 j cos 2j0 sin 2j, if j = 0

Copyright c 1999 by H.J. Newton

Slide 8

'

Statistics 626

9.4 Truncating Filters


If we want to actually apply a lter to a data set we have a problem since we must truncate the lter
M

Y (t) =
j =M

j X (t j ),

since we dont have an innitely long stretch of data. So how do the true frequency transfer function g (e2i ) and that of the truncated lter
M

gM (e2i ) =
j =M

j e2ij

compare? One example is the band pass lter example above where basically what we are doing is approximating the function that is 1 for

0 and zero otherwise (and thus has a jump discontinuity at the end points of the band) by the rst M terms of its Fourier series
representation. A famous result in Fourier series is that the representation converges to the average of the two points at the discontinuities and that the representation becomes very wiggly within the frequency band. The usual solution to this problem is to apply a set

&
Topic 9: Linear Filters

of standard weights to the s that improve how well the resulting approximate transfer function ts the desired g (see page 83 of the text for more details).

Copyright c 1999 by H.J. Newton

Slide 9

You might also like