Вы находитесь на странице: 1из 22

# EIST Pt II, 4F7 Digital Filters and Spectral Estimation Examples Sheet - Spectrum Estimation

Revision questions - DTFT, DFT and FFT, power spectrum [many of these results are required for answering later questions]
1. Determine the Discrete-time Fourier Transform (DTFT) of the following functions, which have innite time duration: (a) xn = exp(i0 n) (b) xn = sin(0 n) 2. (a) Determine and sketch the magnitude of the DTFT of the following function, xn = exp(in/5), n = 0, 1, ..., 31 0, otherwise

paying particular attention to central lobe and side lobe characteristics (b) Hence sketch the magnitude of the DTFT of xn = sin(n/5), n = 0, 1, ..., 31 0, otherwise

over the frequency range T [2, +2 ]. 3. Determine the power spectrum of the following random processes: (a) xn = A sin(0 n + ) where A and 0 are constants and is uniformly distributed between 0 and 2 . (b) xn = A sin(0 n + ) + vn where A and 0 are constants and is uniformly distributed between 0 and 2 . 2 and vn is random white Gaussian noise with variance v 1

## 4. The Discrete Fourier Transform (DFT) of a data sequence xn is dened by:

N 1

Xp =
n=0

xn ej N np
N 1

1 xn = N

Xp ej N np
p=0

xn Xp Show that: (a) xn = xN n , ie. periodic data (b) Xp = XN p , ie. periodic spectrum (c) xnq ej N nq Xp , ie. shift theorem (d) Xp Xp
(1) (2) N 1 q =0
2

## xq xnq ie. Circular convolution

(1)

(2)

5. Calculate and roughly sketch the 16 point DFT of the data sequence: xn = ej N kn for discrete frequency k = 3.0 and also k = 3.5. Comment on the signicance of the form of the spectra. You may nd it helpful to sketch the real (or imaginary) part of the signal xn and bear in mind the result proved in question 4(a). You may also nd it useful (but not necessary) to use MATLAB. 6. Most realisations of the FFT algorithm, in either software or hardware, are designed to deal with complex data so that the algorithm can be used for both forward and inverse transforms. However In many applications the time domain data are real so that there would appear to be some computational ineciency in using the complex data algorithm. Show that the DFT of 2 real data sequences xn and xn may be computed with a single complex DFT realisation by forming the complex data signal:
(2) xn = x(1) n + j xn (1) (2)
2

and that the spectra of the individual signals are given by: 1 (1) Xp = [Xp + XN p ] 2
(2) Xp =

1 [Xp XN p ] 2j

7. Show that the results from the previous question can be used to compute the DFT of a single signal with 2 N real data points. Calculate the improvement in computational eciency compared with using the direct complex-data DFT on the 2 N real data points. 2

## Windowing and Frequency Resolution

8. The Fourier transform of a continuous-time signal g (t) is to be estimated by evaluating the Fourier transform of the signal viewed through a rectangular window w(t) of duration Tw seconds. Show that the estimated spectrum Gw ( ) is given by the convolution of the spectrum G( ) of the innite duration signal and the spectrum of the window function W ( ). If the signal g (t) is given by: g (t) = a1 cos(1 t) + a2 cos(2 t) sketch the spectrum of Gw ( ) and roughly estimate the minimum window duration Tw if the two frequencies are to be resolved. You may assume that two frequency components can be approximately resolved if their centre lobes do not overlap, i.e. the 3dB central lobe bandwidth of one component does not overlap with that of the other component. Parameter values are: a1 = a2 = 1 1 = 2 1.0.103 rad.s1 2 = 2 1.1.103 rad.s1 9. Repeat the last question but using sampled signals with a sampling frequency s = 2 .104 rad.s1 and estimate the minimum duration of the window in samples. How is the answer changed if a Bartlett window is used instead? You may use the table of discrete time window properties given in question 11. 10. In some practical situations it is necessary to evaluate eciently the DFT of M data points where M is not a highly composite number so that an FFT algorithm cannot be used directly. A commonly used technique to overcome this problem is known as zero-padding. The principle is to append zero amplitude signal samples to the data to give a total of N data points where N is a highly composite number (usually N = 2K ). The DFT of the padded sequence can be computed with an FFT algorithm. The spectral spacing of the DFT components will be 2 whereas the spectral spacing N of the DFT of the unpadded data sequence will be 2 , M < N . It is sometimes M claimed that zero-padding increases the frequency resolution. By considering the DFT as evaluation of the Discrete Time Fourier Transform (DTFT) at a set of discrete frequencies, show that this claim is incorrect.

## Non-parametric Power spectrum Estimation

Window Sidelobe level, dB 3dB Bandwidth Rectangular -13 0.89(2/N ) Bartlett -27 1.28(2/N ) Hanning -32 1.44(2/N ) Hamming -43 1.30(2/N ) Blackman -58 1.68(2/N ) 6dB bandwidth 1.78(2/N )

Table 1: Table of properties for discrete time windows. 3dB Bandwidths are measured in normalised frequencies T from the middle of the central lobe to the half power point. 11. The table above gives the basic properties of a few standard window functions, with bandwidths stated in terms of normalised frequency T . Consider a random process {Xn } composed of two random phase sine-waves: xn = A sin(1 n + 1 ) + B sin(2 n + 2 ) + vn where A and B are constants, 1 and 2 are independent and uniformly distributed 2 between 0 and 2 , and vn is white noise with variance v . (a) Determine and sketch the power spectrum for the process. Hence sketch approximately the expected value of the periodogram for N data points measured from such a random process. (b) Determine the approximate data length required for the periodogram to resolve the two frequencies reliably if 2 1 0.01 , where T is the sampling period. You may assume that two frequency components can be approximately resolved if their centre lobes in the expected value of the periodogram do not overlap, i.e. the 6dB central lobe bandwidth of one component does not overlap with that of the other component in the periodogram. (c) For the data length calculated in the previous part, and 2 1 0.01 , determine approximately the smallest ratio of amplitudes B/A such that the second component can reliably be detected in the presence of the rst (assume 2 v is small). 12. The modied periodogram applies a window to the data before computing the DTFT: 2 N 1 1 jnT jT M (e ) = wn xn e S N U n=0 where U =
1 N N 1 n=0

|wn |2 .

(a) Show that the expected value of the modied periodogram is: M (ejT )] = E [S 1 NU 4
+N 1

wM k RXX [k ]ejkT
k=(N 1)

where wM k = wk wk i.e. the convolution of wk with itself time-reversed. Hence show that M (ejT )] = E [S 1 SX (ejT ) |W (ejT )|2 2N U

where W (ejT ) is the DTFT of the window function wn (b) Comment on the relationship of this result with the expected value of the standard periodogram and discuss how the modied periodogram might achieve a dierent trade-o between frequency resolution and variance of the estimate. 13. A stationary random phase complex exponential is given by xn = exp(i(n0 T + )) where is uniformly distributed between 0 and 2 . (a) What is the power spectrum for this process? [Note that for a complex process, the autocorrelation function is dened as RXX [k ] = E [x n xn+k ], and the power spectrum is the DTFT of RXX .] (b) Write an expression for the periodogram estimate for a sample of N data points measured from the process. (c) Hence determine the mean and variance of the periodogram for this process. Does this tally with the rule of thumb that the variance of the periodogram is approximately equal to the true power spectrum squared? If not, why is it that this process could be dierent from the rule? (d) Comment on how these results impinge on the periodogram estimate for random phase sine waves buried in noise. 14. (a) State the variance of periodogram power spectral estimates of white Gaussian noise having variance 2 . Comment on the signicance of this result for power spectrum estimation of noise-like processes. (b) The Bartlett procedure segments the available data into K contiguous subsequences of length NB and computes a spectral estimate from: X (ejT ) = 1 S K
(k ) K 1

(k) (ejT ) S X
k=0

where SX (ejT ) is the periodogram of the k th subsequence. Show that the Bartlett procedure reduces the variance of the spectral estimate of white noise by K times. (c) Show, for general signals, that the Bartlett procedure is biased as for the periodogram, but asymptotically unbiased. (d) Show that the frequency resolution of the Bartlett method is K times worse than that of the periodogram applied to the same data overall length. 5

Parametric Methods
15. Show that if input of linear time invariant discrete-time system is wide-sense stationary white noise then the output is also wide-sense stationary, provided the linear system is stable. Does the result still apply if the linear system is unstable? Use this result to derive the autocorrelation function for a stable ARMA process, carefully stating any assumptions required. 16. Estimates are made of the correlation function of a particular signal and the values obtained are: RXX  = 7.24 RXX  = 3.6 Determine the parameter values of the 1st order MA model: H (z ) = b0 + b1 z 1 which matches these correlation values using: (a) Direct solution of of the MA equations: RXX  RXX  . = . . RXX [Q] where: cr = (b) By spectral factorisation Sketch the power spectral estimate obtained using this MA model. 17. Fit a 1st order AR model 1 a0 + a1 z 1 to the correlation data given in the previous question and sketch the resulting spectral estimate. Do you have any reason to suppose that this estimate is better than that obtained using the MA model? H (z ) = 0
Q q =r b q b q r

c0 c1 . . . cQ

, rQ , r>Q

Suitable past tripos questions: Most questions from the old I7 course and all questions from current 4F7, including: 4F7 2004 - all questions 4F7 2003 - all questions I7 2002 - all questions (Q4. quite a challenge!) I7 2001 - all questions I7 pre-2001 - most questions, but not questions on the MUSIC algorithm for frequency estimation.

1. (a)
+

2
m=

(T + 2m 0 )

(b)
+

/j
m=

(T + 2m 0 ) (T + 2m + 0 )

+

## sin((0.2 T )N/2) sin((0.2 T )/2)

SX (e ) = /2A

2 n=

[ ( 0 2n ) + ( + 0 2n )]

## (expressed in terms of normalised frequency = T ) (b)

+

SX (e

= /2A

2 n=

2 [ ( 0 2n ) + ( + 0 2n )] + v

(expressed in terms of normalised frequency = T ). 4. 5. 6. 7. 8. 11.2 ms 9. 178 samples. With Bartlett window 256 samples 10.

11. (a)
+

SX (e ) = /2A +B
2

2 n= +

[ ( 1 2n ) + ( + 1 2n )]
2 [ ( 2 2n ) + ( + 2 2n )] + v

n=

(expressed in terms of normalised frequency = T ). (b) 356 samples (c) 0.2 . 12. 13. 14. 15. 16. b0 = 2.0, b1 = 1.8 by either method. 17. b0 = 2.3345, a0 = 1.0, a1 = 0.4972 Suitable questions from past papers: 4F7 2006 - 3 (Frequency estimation),4 (autoregression/correlogram) 4F7 2005 - 3 (Periodogram/MA model), 4 (ARMA models) 4F7 2004 - 3 (MA models), 4 (Windowing/periodogram) 4F7 2003 - 3 (Periodogram), 4 (AR models) Old I7 questions: I7 2002 - 1 (AR Model), 4 (DFT - a bit o syllabus, but a challenging question - try it!) I7 I7 I7 I7 2001 2000 1999 1998 1 2 2 1 (ARMA/MA model), 3 (DFT/windowing) (Periodogram), 4 (AR Models) (MA model) (Windowing), 3 (AR Models)

Worked solutions
1. (a)
+

X (eiT ) =
n= +

n= +

= =
n=

exp(in(0 T ))
+

= 2
m=

(T + 2m 0 )

## since + n= exp(in(0 T )) is the Fourier series representation of a periodic train of impulses.

(b) Using sin(0 n) = 0.5/j (exp(+i0 n) exp(i0 n)) we have, by superposition: X (e 2. (a)
+ iT

1 )= (T + 2m 0 ) (T + 2m + 0 ) j m=

X (e

iT

)=
n= N 1

n=0 N 1

= =
n=0

## sin((0 T )N/2) sin((0 T )/2)

(b) Get this by superposition of two complex exponentials as in part a). See gure.

30

25

20

15

10

0.1

0.2 0/

0.3 Frequency T/

0.4

0.5

0.6

10

0 2

1.5

0.5

0 T/

0.5

1.5

20

15

10

0 2

1.5

0.5

0 T/

0.5

1.5

## Figure 2: plot of |X (eiT )| for sin(0 n)

10

3. (a) Standard material - see e.g. 3F3 lecture notes. (b) Get this by noting that the sine and noise terms are uncorrelated. Hence you can calculate the power spectrum of each term and add them together to get the result. 4. (a) x(n) = 1 N
N 1

X (p) ej N np
p=0 N 1
2

1 x(n) = N 1 = N
N 1

X (p) ej N np
p=0
2

X (p) ej N (N n)p
p=0

x(n) = x(N n) Similarly x(n + kN ) = x(n) k = integer. Thus the data are periodic. One way of viewing this is that the DFT is derived by sampling the Discrete Time Fourier Transform (DTFT) at a set of uniformly spaced frequencies. Sampling in the frequency domain leads to periodic repetition of time domain data in just the same way as sampling in the time domain leads to periodic repetition of spectra. (b)
N 1

X (p) =
n=0

x(n) ej N np
N 1
2

X (p) =
n=0 N 1

x(n) e+j N np
2

=
n=0

x(n) ej N n(N p)

## X (p) = X (N p) Similarly X (p + kN ) = X (p) k = integer. (c) Let: x(n q ) X (p)

N 1

X (p) =
n=0

x(n q ) ej N np

Let m = n q then:
N 1q

X (p) =
m=q

x(m) ej N (m+q)p

11

N 1q

=e
j 2 mp N

j 2 qp N

x(m) ej N mp
m=q

are periodic in m with period N so that the summation Both x(m) and e can be carried out over m = 0, 1, . . . , N 1 which gives the desired result; this observation can be justied more formally if desired. A more straightforward method is to approach the problem from the other end. Let: x (n) ej N qp X (p) 1 x (n) = N 1 = N
N 1
2

[ej N qp X (p)] ej N np
p=0

N 1

X (p) ej N (nq)p
p=0

= x(n q ) (d) Let x (n) X1 (p) X2 (p). Then: x (n) = Substitute X1 (p) =
q =0

1 N

N 1

X1 (p) X2 (p)ej N np
p=0

N 1

x1 (q ) ej N qp
2 2

1 x (n) = N =

N 1 N 1

{
p=0 q =0

x1 (q ) ej N qp }X2 (p) ej N np
N 1
2

N 1

q =0

1 x1 (q ) N
N 1

X2 (p)ej N (nq)p
p=0

=
q =0

x1 (q ) x2 (n q )

This is called a circular convolution in that x1 (n) and x2 (n) are periodic. 5. X (p) =
n=0 N 1

N 1

ej N nk ej N np
2

=
n=0

ej N n(kp)

12

This is a geometric progression and may be summed as: X (p) = |X (p)| = 1 ej 2(kp) 1 ej N (kp) sin (k p) sin N (k p)
2

From this expression and from gure 3 it can be seen that if k is an integer then the spectrum has a single component at frequency p = k . If, however, k is not an integer then the spectrum is smeared. One way of looking at this is that the DFT is eectively the Fourier series of the periodic repetition of the N data points. Now if k is an integer, the periodic repetition will be continuous. If k is not an integer then the periodic repetition is discontinuous and the Fourier series will contain many frequency components.
Time Domain k=3.0 1 0.5 0 -0.5 -1 1 0.5 0 -0.5 -1 Time Domain k=3.5

10

15

10

15

## Frequency Domain k=3.5

10

15

Figure 3: 6. X (p) =
n=0 N 1

N 1

n=0
2 2

X (N p) =
N 1

n=0

13

N 1

X (N p) =
n=0

## [x1 (n) j x2 (n)] ej N np

N 1

X (p) + X (N p) = 2
n=0

x1 (n) ej N np = 2 X1 (p)
N 1

andX (p) X (N p) = 2j
n=0

x2 (n) ej N np = 2j X2 (p)

7. The solution really amounts to doing the rst stage of the FFT algorithm to convert the DFT of 2N data points into 2 DFTs of N data points and using the result from question 5 to evaluate the 2 DFTs.
2N 1

X (p) =
n=0 N 1

x(n) ej 2N np
N 1

p {0, 2N 1}

=
n=0

x(2n) e

2 j 2 2np N

+
n=0

x(2n + 1) ej 2N (2n+1)p
2

## = X1 (p) + ej 2N p X2 (p) where:

N 1 N 1

(1)

X1 (p) =
n=0

x(2n) e

j 2 np N

X2 (p) =
n=0

x(2n + 1) ej N np

p {0, N 1}

and Direct evaluation of the 2N point DFT requires evaluation of equation 1 requires: [

X (p + N ) = X1 (p) ej 2N p X2 (p)
2N 2

## N N log2 (N ) + log2 (N ) + N ] 2 2 = N [log2 (N ) + 1] = N [log2 (N ) + log2 (2)] = N log2 (2N )

However X1 (p) and X2 (p) are the DFTs of real data sequences so can be evaluated as in question 5 thus giving an approximate halving of the computation for large N.

8. The rst part is standard bookwork, see lecture notes pp. 35-36. The spectrum of the rectangle window is :
+Tw /2

W ( ) =
Tw /2

exp(jt)dt

= Tw sinc(Tw /2) 14

[Note the window can be centered on any time value; t = 0 is a conveient choice] The 3dB point is i.e. Tw /2 0.44 (nd this by trial and error or by plotting in Matlab) We require that the two central lobes do not overlap. Hence the gap between the two frequencies must be greater than 2 2 0.44/Tw . Now 2 1 = 2 0.1 103 rad.s1 . Hence 2 0.88/Tw = 2 0.1 103 i.e. Tw = 11.2ms. Note that this is a highly approximate estimate - a longer window length would be needed to ensure resolvability in all cases. The complex sidelobe structure means that the resolvability depends on the precise value of 2 1 . Also, any noise in measurements can have an adverse eect on resolvability. Comment: the extra factor of 2 in 2 0.88/Tw arises so as to make sure the 3dB point of the 1 component exactly matches the 3dB point of the 2 component. 9. The table give the 3dB bandwidth for the rectangular window as T = 0.89 2/N Hence, with the same frequency spacing as before, we require: 2 0.89 2/N = 2 0.1 103 T T is the sampling period, given by T = 2/s = 1/104 s Hence N = 2 0.89 104 /(0.1 103 ) = 178 Again, this result is very approximate. With the Bartlett window the 3dB point is 1.28 2/N and hence the window length must be increased to N = 2 1.28 104 /(0.1 103 ) = 256samples 10. Let the signal be g (n), n {0, M 1} and this is padded with zeros to give: gpad = g (n) n = 0, 1, . . . , M 1 0 n = M, M + 1, . . . , N 1 |sinc(Tw /2)| = 1/ 2

15

## Now the DTFT of the unpadded sequence is given by:

M 1

G(e

jT

)=
n=0

g (n) ejnT

(Note that we are implicitly assuming that g (n) is zero outside the interval n = 0, 1, . . . , M 1). Sampling the DTFT at T = 2 p gives the normal M point DFT. However if the M DTFT is sampled at frequencies T = 2 p then: N
M 1

G(e

j2 p N

)=
n=0

g (n) ej N np

N 1

=
n=0

which is the DFT of the padded sequence so we see that the N-point DFT of the padded sequence and the M-point DFT of the unpadded sequence correspond to sampling the DTFT at N frequency points and M frequency points respectively. Thus the padded DFT gives no greater frequency resolution (say between 2 closely spaced sinusoids) but simply evaluates the spectrum at more frequencies; the eects of leakage and smearing will be the same. 11. (a) From question 3 we have the power spectrum of a single random phase sine wave in noise. To get the two-sine version, notice that both sine terms and the noise term are mutually uncorrelated (check this if you are unsure). Hence to get overall power spectrum, just add together the power spectra of the sine waves with that of the noise (white). (b) Expected value of the periodogram is (see lecture notes): X (ejT )] = E [S 1 2

## W (ejT )SX (ej ( )T ) dT

(2)

i.e. the convolution of the true power spectrum with the Bartlett window (assuming biased form for the autocorrelation function estimate). The convolution is easy to sketch since the power spectrum is a train of delta functions plus a noise oor. (c) 6dB bandwidth for Bartlett window is 1.78 2/N , where N is the window length. Hence, frequency resolution is approximately 2 1.78 2/(2N ) = 0.01 [note Bartlett window is of length 2N for the periodogram estimate]. Hence N = 356. (d) Sidelobes are at -27dB for the Bartlett window. Hence can roughly detect a component 27dB lower than the main component. Hence 20 log10 (B 2 /A2 ) 27, B/A 0.2.

16

12. (a) M (e E [S
jT

1 )] = E[ NU

N 1

wn xn e
n=0 N 1

jnT

] wm xm e+jmT ]

1 = E[ wn xn ejnT N U n=0 = = = 1 NU 1 NU 1 NU
N 1 N 1

N 1

m=0

n=0 m=0 N 1 N 1

n=0 m=0 N 1 N 1

## wn wnk RXX [k ]ejkT with n m = k

n=0 k=(N 1) N 1 N 1

1 = NU 1 = NU

{
k=(N 1) n=0 N 1

## wn wnk }RXX [k ]ejkT

wM k RXX [k ]ejkT
k=(N 1)

where wM k =

N 1

wn wnk = wk wk
n=0

as required. Hence, by Fourier transform convolution theorem: M (ejT )] = E [S 1 SX (ejT ) |W (ejT )|2 2N U

(b) Modied periodogram allows choice of a window function with suitable spectral leakage and spectral smearing properties to the application. This contrasts with the periodogram, in which the windowing function is xed as the rectangular window - narrow central lobe but very severe sidelobes. 13. (a) Power spectrum is a train of delta functions centred at frequency 0 :
+

S (e ) = 2
n=

( 0 T + 2n)

17

(b) Assuming the biased autocorrelation estimator, we have the periodogram as:
N 1

(ei ) = | S
n=0 N 1

## xn exp(in)|2 exp(i(n0 T + )) exp(in)|2

n=0 N 1

=|

= | exp(i)
n=0 N 1

exp(i(n0 T + )) exp(in)|2

=|
n=0

exp(i(n0 T )) exp(in)|2

[This can be simplied further, but unnecessary for this question] The important point here for the next parts is that the periodogram estimate does not depend on the value of the random variable . Hence the variance of the periodogram estimate is zero, see next part. (c) The mean is the Bartlett window spectrum shifted across in frequency to center frequency 0 . The variance is, however, zero. Thus in this case the variance is not the rule of thumb. i.e. for single complex exponentials the periodogram gives no variability. This is because the periodogram of a complex exponential is constant whatever the phase of the exponential (see last part) (d) The result carries through more or less for estimation of random phase sine waves since these can be written as a superposition of complex exponentials. Note however that the phase term does have an eect here for short data lengths since the spectrum of the two complex exponentials sum together differently in the frequency domain depending on their phases, but much less signicant variability than the variability of the periodogram of noise-like signals. Hence the practically observed property that sine waves in noise become well estimated (with little variance) by the periodogram for long data lengths. 14. (a) Variance of periodogram is approximately 4 for all data lengths, becoming exact as data length goes to innity. This means that the variance does not decrease as data length increases and therefore periodogram will have unacceptable variability for many noise like processes. (b) Each of the K subsequences of data are statistically independent so that the periodogram estimates for each subsequence are also statistically independent. k (eji T from each of the K periConsider a particular frequency component P XX odograms. These K components may be regarded as statistically independent random numbers with variance 4 . In order to ease the notation, let:
k XX (eji T ) Zi S

K 1

Zk E {
k=0

1 K

K 1

Zk }]2 }
k=0

18

## 1 2 + Z 2 = 2 E {[ Zk ]2 } 2Z K k=0 where: 2 = = E {Zk } Z 1 E{ K2 2 Zk Zj } Z

k j

K 1

we can Now remembering that Zk is a random variable with mean value Z write: 2 k = j 4 + Z E { Zk Zj } = 2 Z k=j 2 = 1 K2 2) + ( 4 + Z
k

1 2 Z 2 (K 2 K ) Z 2 K

4 K i.e. the variance of the spectrum estimate has been reduced by a factor of K. 2 = (c) X (ejT )] = E [ 1 E [S K = 1 K
K 1

## (k) (ejT )] S X (ejT )] E [S X

(k)

k=0 K 1

k=0

Now each expectation term is an expectation of a periodogram estimate for each sub-block. Hence, as for the periodogram the method is biased but asymptotically unbiased. (d) Clearly each periodogram in the summation corresponds to a window length NB = N/K , where N is the total length. Hence each periodogram estimate in the Bartlett summation has K times poorer resolution, since central lobe of window spectrum is K times wider than that of the full perdiodogram estimate for all N data points. 15. If impulse response of the ARMA lter is hn , then we know from standard linear systems theory that the lter is stable if and only if

|h n | <
n=0

Now ARMA signal can be written as the output of a linear system with white noise input {wn }:

xn =
i=0

hi wni

## First the mean

X [n] = E [xn ] = E [
i=0

hi wni ] =
i=0

hi E [wni ] =
i=0

h i w

i.e. mean is constant since mean of stationary input process is constant w . Then the autocorrelation function:

## RXX [n, m] = E [xn xm ] = E [ =

hi wni
j =0

hj wnj ]

i=0

hi hj E [wni wmj ]
i=0 j =0

=
i=0 j =0

hi hj RW W [(n m) (i j )]

since {wn } is stationary. Hence autocorrelation function depends only on time dierence n m, as required for WSS. Finally check variance of process. Here we use the white noise property of the input 2 process, i.e. cW W [k ] = W [k ]:
2 X = RXX 2 X = i=0 j =0 2 2 h2 i (RW W W ) = (RW W W ) i=0 i=0

hi hj RW W [(ij )]2 X =

## Now, for a stable system,

h2 i <
i=0

since:

>(
i=0

|hi |)2 =
i,j

|hi ||hj | |h i |2 +
i i,j, i=j

## = and since both terms on right are positive,

|hi ||hj |

|h i |2 <
i

Hence the variance is nite. This will not be the case for an unstable process. Now, derivation of autocorrelation function for ARMA model follows as in the lecture notes.

20

MA model Spectrum 15

10

0 0

0.05

0.1

0.15

## 0.2 0.25 0.3 Normalised frequency

0.35

0.4

0.45

0.5

Figure 4: 16. Either method should give: b0 = 2.0 b1 = 1 . Remember that the spectral factorisation method only gives the roots of the polynomial and the scaling must be calculated separately. The MA spectrum corresponding to the model is shown in gure 4

10. The AR coeecients are: b0 = 2.3345 a0 = 1 a1 = 0.4972 The AR spectrum corresponding to the model is shown in gure 5 Without any prior knowledge of the physical system which produced the signals, one spectral estimate should not be preferred over the other. However, the 1st order MA model assumes that the signal correlation is zero for lags greater than 1 whereas the AR model assumes that the correlation function satises the AR difference equation so that the correlation function is not zero for lags greater than 1. It might be argued that this is a more reasonable reection of what might be the case in the system which generated the signal. Simon Godsill, November 2003, Peter Rayner 1999 21

AR model Spectrum 22 20 18 16 14 12 10 8 6 4 2 0

0.05

0.1

0.15

0.35

0.4

0.45

0.5

Figure 5:

22