1. The document contains solutions to problems from an Advanced DSP assignment. Problem 3.1 involves determining the optimal Wiener filter and prediction filter for a noisy signal. Problem 3.2 shows that the average power of the output of a finite filter can be expressed in two alternative forms. Problem 3.3 compares filtering a signal through its own analysis filter versus a different analysis filter.
1. The document contains solutions to problems from an Advanced DSP assignment. Problem 3.1 involves determining the optimal Wiener filter and prediction filter for a noisy signal. Problem 3.2 shows that the average power of the output of a finite filter can be expressed in two alternative forms. Problem 3.3 compares filtering a signal through its own analysis filter versus a different analysis filter.
1. The document contains solutions to problems from an Advanced DSP assignment. Problem 3.1 involves determining the optimal Wiener filter and prediction filter for a noisy signal. Problem 3.2 shows that the average power of the output of a finite filter can be expressed in two alternative forms. Problem 3.3 compares filtering a signal through its own analysis filter versus a different analysis filter.
1. The document contains solutions to problems from an Advanced DSP assignment. Problem 3.1 involves determining the optimal Wiener filter and prediction filter for a noisy signal. Problem 3.2 shows that the average power of the output of a finite filter can be expressed in two alternative forms. Problem 3.3 compares filtering a signal through its own analysis filter versus a different analysis filter.
A stationary random signal 𝑥𝑛 is to be estimated on the basis of the noisy observations
𝑦𝑛 = 𝑥𝑛 + 𝜈𝑛
It is given that 1 𝑆𝑥𝑥 (𝑧) = , 𝑆𝑣𝑣 (𝑧) = 5, 𝑆𝑥𝜈 (𝑧) = 0 (1 − 0.5𝑧 −1 )(1 − 0.5𝑧)
(a) Determine the optimal realizable Wiener filter for estimating the signal 𝑥𝑛 on the basis of the observation 𝑌𝑛 = {𝑦𝑖 , 𝑖 ≤ 𝑛}. Write the difference equation of this filter and compute the mean-square estimation error.
Dr. Syed Abd. Rahman Al-Attas Advanced DSP Page - 1
1.25 1 𝑆𝑥𝑦 (𝑧) ( ) 0.2 𝐻(𝑧) = 2 [ ] = 1 − 0.5𝑧 −1 = 𝜎𝜖 𝐵(𝑧) 𝐵(𝑧 −1 ) + 1 − 0.4𝑧 −1 1 − 0.4𝑧 −1 6.25 ∙ −1 1 − 0.5𝑧 The difference equation for the estimator then is 𝑥̂𝑛/𝑛 = 0.4𝑥̂𝑛−1/𝑛−1 + 0.2𝑦𝑛 . The estimation error is computed using the contour integral (you need to refer to DSP books on this formulation).
(b) Determine the optimal realizable Wiener filter for predicting one step into the future: that is, estimate 𝑥𝑛+1 on the basis of 𝑌𝑛 .
The prediction part is handled by defining 𝑥1 (𝑛) = 𝑥𝑛+1 . The problem of predicting 𝑥𝑛+1 on the basis of 𝑌𝑛 is equivalent to the problem of estimating 𝑥1 (𝑛) on the basis of 𝑌𝑛 . Using the filtering equation 𝑋1 (𝑧) = 𝑧𝑋(𝑧), we find 𝑧 𝑆𝑥1 𝑦 (𝑧) = 𝑧𝑆𝑥𝑦 (𝑧) = (1 − 0.5𝑧 −1 )(1 − 0.5𝑧)
where denoted 𝑥̂𝑛+1/𝑛 = 𝑥̂1 (𝑛). We note that the prediction filter 𝐻1 (𝑧)is related to the estimation filter 𝐻(𝑧) by 𝐻1 (𝑧) = 0.5𝐻(𝑧). Since both filters have the same input, that is, 𝑦𝑛 , it follows that
𝑥̂𝑛+1/𝑛 = 0.5𝑥̂𝑛/𝑛
Dr. Syed Abd. Rahman Al-Attas Advanced DSP Page - 2
Noting that 𝑆𝑥1 𝑥1 (𝑧) = 𝑧𝑆𝑥𝑥 (𝑧)𝑧 −1 = 𝑆𝑥𝑥 (𝑧), and 𝑆𝑦𝑥1 (𝑧) = 𝑆𝑦𝑥 (𝑧)𝑧 −1 , we find for the prediction error
𝑑𝑧 1 𝑑𝑧 𝜀1 = 𝐸 [𝑒 2 1] = ∫ [𝑆𝑥1 𝑥1 (𝑧) ∙ 𝐻1 (𝑧)𝑆𝑦𝑥1 (𝑧)]𝑧 𝑛−1 = ∫ 𝑧 𝑛−1 = 1.25 𝑛+𝑛 2𝜋𝑗𝑧 (𝑧 − 0.4)(1 − 0.5𝑧) 2𝜋𝑗 𝑢.𝑐. 𝑢.𝑐. This is slightly worse than the estimation error 𝜀 because the estimation filter uses one more observation than the prediction filter, and thus, makes a better estimate.
Problem 3.2
A stationary random signal 𝑦𝑛 is sent through a finite filter 𝐴(𝑧) = 𝑎0 + 𝑎1 𝑧 −1 + ⋯ + 𝑎𝑀 𝑧 −𝑀 to
where we used eqn (10) from Modul 2a – Random Process as well as 𝑆𝑒𝑒 (𝜔) = |𝐴(𝜔)|2 𝑆𝑦𝑦 (𝜔), which follows from the fact that 𝑒𝑛 is the output of the linear filter 𝐴(𝑧) when the input is 𝑦𝑛 .
Problem 3.3
Consider the two autoregressive random signals 𝑦𝑛 and 𝑦𝑛′ generated by the two signal models:
(a) Suppose 𝑦𝑛 is filtered through the analysis filter 𝐴′ (𝑧) of 𝑦𝑛′ producing the output signal 𝑒𝑛 ; that is,
𝑀 𝑦𝑛 𝑒𝑛 𝐴′ (𝑧) ′ 𝑒𝑛 = ∑ 𝑎𝑚 𝑦𝑛−𝑚 𝑚=0
If 𝑦𝑛 were to be filtered through its own analysis filter 𝐴(𝑧), it would produce the innovations sequence 𝜖𝑛 . Show that the average power of 𝑒𝑛 compared to the average power of 𝜖𝑛 is given by:
where 𝒂, 𝒂′ and 𝑅𝑦𝑦 have the same meaning as in Problem 2.2 above. This ratio can be taken as a measure of similarity between the two signal models. The log of this ratio is
Dr. Syed Abd. Rahman Al-Attas Advanced DSP Page - 4
Itakura’s Linear Predictive Coefficient (LPC) distance measure used in speech recognition.
Solution
Because 𝐵(𝑧) = 1/𝐴(𝑧) is the signal model of 𝑦𝑛 , we must have:
𝜎𝜖2 𝑆𝑦𝑦 (𝜔) = 𝜎𝜖2 |𝐵(𝜔)|2 = |𝐴(𝜔)|2
Using the results of Problem 2.3 above, applied to the filter 𝐴′ (𝑧), we find 𝜋 𝑑𝑤 𝜎𝑒2 = 𝐸[𝑒𝑛2 ] = ∫ |𝐴′ (𝜔)|2 𝑆𝑦𝑦 (𝜔) = 𝐚′𝑻 𝑹𝒚𝒚 𝐚′ 𝑜𝑟, 2𝜋 −𝜋
Dr. Syed Abd. Rahman Al-Attas Advanced DSP Page - 5
Generate 1500 samples of a unit-variance, zero-mean, white-noise sequence 𝑥𝑛 𝑛 = 0,1, … ,1499 and filter them through the filter 𝐻 to obtain the output sequence 𝑦𝑛 . Compute the sample cross- correlation 𝑅̂𝑦𝑥 (𝑘) for 𝑘 = 0,1, … , 50 to obtain estimates of the impulse response ℎ𝑘 . On the same graph, plot the estimated impulse response versus time, together with the actual impulse response. Repeat, using a different realization of 𝑥𝑛 .
You can use the Scilab coding given in the System Identification example. You should get something that looks like below:
Problem 3.5 (Scilab Assignment)
(a) Generate 1000 samples of zero-mean, unit-variance, white Gaussian noise sequence 𝑥𝑛 , 𝑛 = 0,1, … , 999, and filter them through the filter defined by the difference equation:
𝑦[𝑛] = 𝑎𝑦[𝑛 − 1] + (1 − 𝑎)𝑥[𝑛]
with 𝑎 = 0.95. to avoid the transient effects introduced by the filter, discard the first 900 output samples and save the last 100 samples of 𝑦𝑛 . Compute the sample autocorrelation of 𝑦𝑛 from the length-100 block samples.
(b) Determine the theoretical autocorrelation 𝑅𝑦𝑦 (𝑘), and on the same graph, plot the theoretical and sample autocorrelations versus k. Do they agree?
We can compute this using the notes in module 2
0.05 From (a), we compute 𝐻(𝑧) = and taking the inverse z-xform gives us the 1−0.95𝑧 −1 impulse response ℎ[𝑛] = 0.05(0.95)𝑛 𝑢[𝑛].
Dr. Syed Abd. Rahman Al-Attas Advanced DSP Page - 6
The sample autocorrelation to some extend is a rough estimation of the theoretical
autocorrelation.
Problem 3.6
A filter is defined by 𝑦[𝑛] = −0.64𝑦[𝑛 − 2] + 0.36𝑥[𝑛].
(a) Suppose the input is zero-mean, unit-variance, white noise. Compute the output spectral density 𝑆𝑦𝑦 (𝑧) and power spectrum 𝑆𝑦𝑦 (𝜔) and plot it roughly versus frequency.
(b) Compute the output autocorrelation 𝑅𝑦𝑦 (𝑘) for all lags k.
(c) Compute the noise reduction ratio of this filter.
(d) What signal 𝑠[𝑛] can pass through this filter and remain entirely unaffected (at least in the steady-state region)?
(e) How can the filter coefficients be changed so that (i) the noise reduction capability of the filter is improved, while at the same time (ii) the above signal 𝑠[𝑛] still goes through unchanged? Explain any tradeoffs
Dr. Syed Abd. Rahman Al-Attas Advanced DSP Page - 7
(b) The autocorrelation function is obtained by performing the partial expansion fraction and then select the poles inside the u.c. only. Doing this we should get the following
𝜋𝑘 𝑅𝑦𝑦 (𝑘) = 0.2195(0.8)𝑘 cos ( ) 2 for 𝑘 ≥ 0
(c) Since 𝜎𝑦2 = 𝑅𝑦𝑦 (0), we find for the noise reduction ratio
𝜎𝑦2 𝑅𝑦𝑦 (0)
= = 0.2195 𝜎𝑥2 1 𝜋 (d) Since the filter has unity gain at the frequencies 𝜔 = ± , 𝐻(𝜔) = 1, it follows that any 2 linear combination of 𝜋𝑛 𝜋𝑛 sin ( ) 𝑎𝑛𝑑 cos ( ) 2 2 will go through the filter completely unchanged (in the steady state).
(e) By letting the poles closer to the unit circle, the steady state is reached more slowly, but the noise reduction ratio will become smaller. This is the basic tradeoff between speed of response and the effective noise reduction.
Dr. Syed Abd. Rahman Al-Attas Advanced DSP Page - 8
