Delft University of Technology

Faculty of Electrical Engineering, Mathematics, and Computer Science

Circuits and Systems Group

EE 4C03 STATISTICAL DIGITAL SIGNAL

PROCESSING AND MODELING
26 January 2017, 18:3021:30
Open book exam: copies of the book by Hayes and the course slides allowed. No other
materials allowed.

This exam has four questions (40 points)

Question 1 (10 points)

Consider a first-order AR process that is generated by the difference equation

y(n) = a y(n 1) + w(n),

2.
where |a| < 1 and w(n) is a zero mean white noise random process with variance w

(a) Find the unit sample response of the filter that generates y(n) from w(n).
(b) Find the autocorrelation sequence of y(n).
(c) What is the variance y2 of the output process?
(d) Find the power spectrum of y(n).
(e) What is the 4 4 autocorrelation matrix R y ? Give 3 properties of this matrix.
(f) In general, if the p p autocorrelation matrix R y of some WSS random process y(n) is
singular, then what can you say about that process?

Question 2 (10 points)

Consider the general second-order pole-zero model H(z) given by

b0 + b1 z 1 + b2 z 2
H(z) = .
1 + a1 z 1 + a2 z 2
We want to fit this model to the convolution of two signals x(n) = s(n) t(n), where s(n) =
0.1n u(n) and t(n) = 0.5n u(n 1).

(a) Use the property that the z-transform of n u(n) is 1/(1 z 1 ) and that of x(n k)
is z k X(z) to determine the z-transform of x(n) = s(n) t(n). Can we fit H(z) to this
model? If so, what are the parameters a 1 , a2 , b0 , b1 and b2 ?
 (b) Given the samples x(0), x(1), x(2), . . . , determine the parameters of H(z) using Pades
method. How many samples are needed?
Hint: The following expression might be useful for this:
 1  
a b 1 d b
= .
c d ad bc c a

(c) Give the equations to estimate H(z) using Pronys method? There is no need to derive
the actual solution. How many samples are needed in this case?
(d) For this problem, is Pronys model going to change if we increase the number of signal
values that we consider? Why or why not?
(e) Compare the three filter models and explain.

Question 3 (10 points)

Let x(n) = sin(0 n + ) + w(n) be a (real-valued) sinusoid in noise. Here, 0 and are
deterministic (non-random) unknown parameters. We are given data x(n), n = 0, , N 1,
and we wish to estimate the frequency 0 using the MUSIC algorithm.

(a) Consider first the noise-free case. Define a data vector

x(n) = [x(n), x(n + 1), x(n + 2), x(n + 3)] T

and a 4 (N 3) data matrix

X = [x(0), x(1), , x(N 4)].

What is the structure of this matrix in terms of the unknown parameters? What is the
rank of this matrix?
1 j
Hint: you need to use the property sin() = 2j (e ej ).

(b) Define the 4 4 data covariance matrix R = E{x(n)x(n) H }.

What is a suitable corresponding estimate R in terms of the given data x(n)?
Consider the eigenvalue decomposition of R. For the noise-free case, what is the rank
of this matrix?
What changes to the eigenvalues if we add white noise w(n) with power n2 ?

(c) Briefly explain the MUSIC algorithm for this case. How is 0 estimated?

(d) What is the smallest size of the covariance matrix that can be used (if it is known in
advance that there is only one sinusoid)?

(e) Let P be the matrix

0 0 0 1
0 0 1 0
P=
0

1 0 0
1 0 0 0

2
 When applied to a vector, this matrix reverses the entries of the vector.
In a technique called forward backward averaging, R is replaced by R+PRP. Explain
why this can work, and why this leads to a better estimate of 0 .

Question 4 (10 points)

x(n) P (z) e(n)

p(n)

W (z) y(n) S(z)

w(n) s(n)

noise headphone ear

In acoustic noise control (ANC) for headphones, we want to actively generate anti-noise
to cancel the noise such that only the useful signal remains. Suppose that x(n) is the noise we
want to cancel and which can be picked up with a microphone outside the headphone. This
noise x(n) should be canceled inside the ear so what we actually want to cancel is x(n) after
traveling to the inside ear, via what is called the primary path (modeled using the transfer
function P (z)). The anti-noise that is generated by the headphone is denoted by y(n) but this
signal also has to travel to the inside ear, via what is called the secondary path (modeled using
the transfer function S(z)). Hence, we actually want to have that S(z)Y (z) = P (z)X(z),
so that the noise is perfectly canceled inside the ear. Unless otherwise stated, it is assumed
that the primary and secondary paths are known, i.e., P (z) and S(z) are known.

(a) Suppose we aim to generate y(n) from x(n) using a linear filter w(n), i.e., y(n) =
w(n) x(n). Give the expression for the optimal transfer function of the filter w(n), i.e.,
derive the optimal W (z), where we assume no particular filter structure.
What complication do you see with this solution?

(b) Let us now use the theory of optimal filtering to compute W (z). This time we assume
that all filters (P (z), S(z), and W (z)) are causal FIR filters with a finite order. First,
derive the expression for the mean square error E{e 2 (n)} where e(n) = s(n) y(n) +
p(n) x(n) = s(n) w(n) x(n) + p(n) x(n). From this expression, derive the Wiener-
Hopf equations for w(n) by taking the derivative towards w(n) for all its taps.
Hint: Use the commutative property of the convolution and write the convolution using a
matrix-vector product. More specifically, we can write the convolution between x 1 (n) and
x2 (n) as xT1 X2 = xT2 X1 , where x1 = [x1 (0), . . . , x2 (N1 )]T , x2 = [x2 (0), . . . , x2 (N2 )]T ,
X1 is an (N2 + 1) (N1 + N2 + 1) Toeplitz matrix based on x1 , and X2 is an (N1 +
1) (N1 + N2 + 1) Toeplitz matrix based on x2 .

(c) Can you transform problem (b) into a classical optimal filtering problem with desired
signal given by d(n) and input of the filter w(n) given by u(n)? What is d(n) and u(n)
in this case?

3
(d) Based on the above, give the LMS update equations for the causal FIR filter w(n).

(e) Suppose now that the error signal e(n) is measured using a second microphone inside
the headset. What would be the advantage of such a second microphone for the LMS
algorithm in terms of the knowledge of P (z) and/or S(z)?
Give the LMS update equation for this case.