Bogazici University Institute of Biomedical Engineering

BM 571 – Biomedical Signal Processing

Assignment 4

Fall Term Instructor: Ahmet Ademoglu, PhD

1. Let x(n) be a random process consisting of a single complex exponential in white noise

rx (k) = P ejkω0 + σw
2
δ(k)

and let gi be the minimum variance bandpass filter

R−1
x ei
gi =
ei R−1
H
x ei

that has a center frequency ωi with Gi (ejωi ) = 1. Assuming that ωi ̸= ω0 , prove that Gi (z) has
a zero that approaches z = ejω0 as σw
2 −→ 0.

2. The autocorrelation sequence of x(n) is

rx (k) = α|k| ; |k| ≤ M

where |α| < 1. Estimate the power spectrum using

(a) The Blackman-Tukey method with a rectangular window.

(b) The minimum variance method.

(c) The maximum entropy method.

3. The estimated autocorrelation sequence of a random process x(n) for lags k = 0, 1, 2, 3, 4 are

rx (0) = 2, rx (1) = 1, rx (2) = 1, rx (3) = 0.5, rx (4) = 0

Estimate the power spectrum of x(n) for each of the following cases.

(a) x(n) is an AR(2) process.

(b) x(n) is an M A(2) process.

(c) x(n) is an ARM A(1, 1) process.

(d) x(n) contains a single sinusiod in white noise.

4. You are given the following values for the autocorrelation sequence of a wide sense stationary
process x(n)
√
rx (0) = 2, rx (1) = 3/2, rx (2) = 0.5

1
 The eigenvalues of the 3 × 3 Toeplitz autocorrelation matrix are λ1 = 3.5, λ2 = 1.5 and
λ3 = 1.0 and the corresponding normalized eigenvectors are
 √   
q 3/2 √
"
−1
#
√ 1
√
v1 = 2/5  √ 1  , v2 = 1/ 2 0 , v3 = 1/ 5  − 3 
3/2 1 1

It is known that x(n) consists of a single sinusiod in white noise.

(a) Estimate the frequency of the sinusiod using Blackman-Tukey method of frequecncy
estimation.

(b) Use the MUSIC algorithm to estimate the frequency of the sinusoid.

5. Let the signal model be

p
X
x(n) = Ak ejnωk + w(n)
k=1

where w(n) is noise that is uncorrelated with the complex exponentials.

(a) If w(n) is white noise then, the autocorrelation matrix for x(n) may be written as

Rx = EPEH + σw
2
I

where E is a matrix of complex exponentials and P is a diagonal matrix of signal powers.

If x(n) is filtered with a p−th order FIR filter a = [a(0), a(1), . . . , a(p)], then the power in
the output process is
n o
ξ = E |y(n)|2 = aH Rx a

If a is constrained to have unit norm, aH a = 1, show that the filter that minimizes ξ has p
zeros on the unit circle at the fequencies ωk of the complex exponentials, and show that
2.
the minimum value of ξ is equal to σw

(b) Now assume that w(n) has an arbitrary power spectrum, Pw (ejω ). If the autocorrelation
2 Q, then the autocorrelation matrix for x(n) becomes
matrix w(n) is σw

Rx = EPEH + σw
2
Q

Suppose that x(n) is filtered with a p−th order FIR filter a = [a(0), a(1), . . . , a(p)]T that is
normalized so that
aH Qa = 1

Show that the filter that minimizes the power in the filtered process has p zeros on the
unit circle at the frequencies ωk of the complex exponentials, and the minimum value is
2.
equal to σw

2
 (c) Show that the filter minimizing ξ = aH Rx a subject to the cosntraint aH Qa = 1 is
equivalent to solving the generalized eigenvalue problem

Rx a = λQa

for the minimum egienvalue and eigenvector. Thus the frequency of the complex
exponentials corresponds to the roots of the polynomial that is formed from the minimum
eigenvector
p
vmin (k)z −k
X
Vmin (z) =
k=0

and 2
σw corresponds to the minimum eigenvalue.

(d) A random process consists of single sinusoid in nonwhite noise,

x(n) = A sin(nω0 + ϕ) + w(n)

The first three values of the autocorrelation sequence for x(n) are

rx = [9.515, 7.758, 6.472]T

It is known that the additive noise w(n) is a moving average process that is generated by
filtering white noise v(n) as follows

w(n) = v(n) + 0.1v(n − 1)

However, the variance of v(n) is unknown. Find the frequency ω0 and the power,
P = A2 /2, of the sinusoid.

6. In the MUSIC algorithm, finding the peaks of the frequency estimation function

1
P̂M U (ejω ) = M
|eH vi |2
P
i=p+1

is equivalent to finding the minima of the denominator. Show that finding the minima of the
denominator is equivalent to finding the maxima of
M
X
|eH vi |2
i=p+1

Hint: Use the fact that

M
X
I= vi viH
i=1

MATLAB Questions

3
7. Consider the broad band AR(4) process x(n) that is produced by filtering unit variance white
Gaussian noise with the filter

1
H(z) =
(1 − 0.5z −1 + 0.5z −2 )(1 + 0.5z −2 )

(a) Generate N = 256 samples of this process and estimate the power spectrum using the
autocorrelation method with p = 4. Make a plot of the estimate and compare it to the
true power spectrum.

(b) Repeat part a) for 20 different realizations of the process x(n). Generate an overlay plot of
the 20 estimates and plot the ensemble average. Is autocorrelation method able to
acccurately estimate the power spectrum?

(c) Repeat part b) using model orders of p = 6, 8 and 12. What happens when the order
becomes too large?

(d) Repeat part b) and c) for the covariance and modified covariance and maximum entropy
methods. Which approach seems to be the best for a broadband AR process?

8. Consider the process x(n) consisting of two sinusiods in noise,

x(n) = 2 cos(nω1 + ϕ1 ) + 2 cos(nω2 + ϕ2 ) + w(n)

where ϕ1 and ϕ2 are uncorrelated random variables thar are uniformly distributed over the
interval [0, 2π], and w(n) is a fourth order moving average process that is generated by filtering
unit variance white noise with a linear shift-invariant filter that has a system function

H(z) = (1 − z −1 )(1 + z −1 )3

Let ω1 = π/2 and ω2 = 1.1π/2.

(a) Plot the power spectrum of x(n).

(b) Find the variance of the moving average process and compare the power in w(n) to the
power in each of the sinusoids.

(c) How many values of x(n) are necessary in order to resolve the frequencies of the two
sinusoids using the periodogram?

(d) Compare the expected value of the periodogram at the sinusoid frequencies ω1 and ω2
with the power spectrum of the noise Pw (ejω ). What record length N is required in order
for the expected value of the periodogram at ω1 and ω2 to be twice the value of P (ejω )?

4
 (e) Generate a sample realization of x(n) length N = 256 and plot the periodogram. Are the
two sinusoidal frequencies clearly visible in the spectrum estimate? Repeat for 20 different
realizations of the process and discuss your findings.

(f) Estimate the expected value of the periodograms that are formed from an ensemble of 50
realizations of x(n). Plot the estimate and comment on its resolution and on how closely it
estimates the moving average part of the spectrum.

(g) Repeat parts e) and f) using Bartlett’s method with K = 2, 4 and 8 sections.

9. It has been said that the minimum variance spectrum estimate exhibits for long data records.
Consider the autocorrelation sequence

rx (k) = cos(0.35πk) + δ(k) + w(k) k = 0, 1, . . . , p

2 . This sequence
where w(k) is uniformly distributed white noise with zero mean and variance σw
represents a noisy estimate of the autocorrelation sequence of a random phase sinusoid of
frequency ω = 0.35π in unit variance white noise.

2 = 0.
(a) Compute the minimum variance spectrum estimate with p = 10 and σw
2 = 0.1 and
(b) Generate an overlay plot of 25 minimum variance spectrum estimates with σw
compare these estimates to that generated in part a).

(c) Repeat parts a) and b) using the maximum entropy method and compare your results
with the minimum variance estimates.

10. Let x(n) be the harmonic process

3
X
x(n) = Ai ejnωi + w(n)
i=1

where w(n) is unit variance white Gaussian noise. Let A1 = 4ejϕ1 , A2 = 3ejϕ2 and A3 = ejϕ3
where ϕi are uncorrelated random variables that are unformly distributed between −π and π.
In addition, let ω1 = 0.4π, ω2 = 0.45π and ω3 = 0.8π.

(a) Assuming that it is known that x(n) contains three complex exponentials, use MUSIC
algorithm to estimate the frequencies. Repeat for 20 different realizations of x(n). How
accurate are the frequency estimates, on the average? How much variance is there in the
frequency estimates?

(b) Estimate the powers using the frequency estimates obtained in part a).