hw2

EE 503 - HW #2
(Due: 23:59, Jan. 5, 2025)
1. (Stochastic signal modelling) Consider a zero-mean random process 𝑥[𝑛] with the following power
spectrum:
4𝑧 , + 17𝑧 + 4
𝑃& (𝑧) =
−3𝑧 , + 10𝑧 − 3
a. Find the autocorrelation of 𝑥[𝑛]. Also find the values of 𝑟& (0), 𝑟& (1), 𝑟& (2) and 𝑟& (3).
b. What is the type of this process?
c. Find a filter which, when driven by unit variance white noise, will yield a random process with
this autocorrelation. State all the possible filters that can be used for this purpose, as well as
their properties (stability, causality, and such).
d. Find a causal and stable filter which, when excited by 𝑥[𝑛], will produce zero-mean, unit
variance, white noise.
e.
i. Suppose only the first two values of the autocorrelation (𝑟& (0), 𝑟& (1)), found in part a,
are known for the process, and your goal is to fit a first-order AR model that would
have these two autocorrelation values. Write the Yule-Walker equations for this
modeling problem.
ii. Solve these equations for the filter parameters.
iii. Find the power spectrum of this AR process. Compare with the true 𝑃& (𝑧) and
comment.
iv. The correlation function values 𝑟& (0) and 𝑟& (1) of the AR model match the given
correlation function exactly. What values for 𝑟& (2) and 𝑟& (3) would the AR model
produce? How do they differ from the true autocorrelation of 𝑥[𝑛]? Why? Comment.
v. How does the approach in part e) differ from the modeling approach in part c)?
Comment.
2. (Deterministic signal modelling) Suppose we would like to model a signal 𝑥[𝑛] as the impulse response
of a filter. The filter has the following transfer (system) function:
𝑏7 + 𝑏8 𝑧 98 + 𝑏, 𝑧 9,
𝐻(𝑧) =
(1 − 𝑧 98 )(1 + 𝑎8 𝑧 98 + 𝑎, 𝑧 9, )
a. Using the Pade approximation method, derive the set of equations that must be solved so that
the filter output 𝑥;[𝑛] is equal to the first five values of the signal 𝑥[𝑛] (i.e. for 𝑛 = 0,1,2,3,4).
b. Suppose the first seven values of 𝑥[𝑛] are given as follows: [1,0,2, −1,2,0,1]. Find the filter
coefficients in this case using the Pade's method.
c. Explain how the Prony’s method could be used for part b.
3. (Estimation basics) Consider the problem of estimating a vector 𝑥 from its noisy linear observations
that are put into the vector 𝑦. The relation is in the following form:
𝑦 = 𝐻𝑥 + 𝑤
Here 𝐻 is a known invertible square matrix and 𝑤~𝑁(0, ΛB ) is the observation noise vector with
covariance matrix ΛB (i.e. correlated noise).
a. Find the maximum likelihood estimate of the unknown vector 𝑥 given the observation vector
𝑦. Provide a closed-form expression, and state whether there are any conditions required to
obtain this closed-form. Is this ML estimate unbiased?
For the remaining parts, the following additional information is given: 𝑥~𝑁(𝜇& , Λ& ) and 𝑥 is
independent of 𝑤.
b. Find the maximum posterior estimate (MAP) of 𝑥. Provide a closed-form expression if
possible, and state whether there are any conditions required to obtain this closed-form. (Hint:
For jointly Gaussian random variables, conditional distributions are also Gaussian. For more
details, see Chapter 3 from Hajek.)
c. Find the linear MMSE estimate of 𝑥. Provide a closed-form expression if possible, and state
whether there are any conditions required to obtain this closed-form.
d. Find the MMSE estimate of 𝑥. Provide a closed-form expression if possible, and state whether
there are any conditions required to obtain this closed-form. Compare your MMSE estimator
with the one in part c, and comment.
4. (IIR Wiener filtering) Problem 7.12 from Hayes
5.
(Your codes should also be provided.)
6. (Sample moments and KL expansion) The data sets data1.dat, data2.dat and data3.dat shared in
ODTUClass represent realizations of three different random signals. In this question, you will
examine the sample moments (mean and autocorrelation) for these signals and their approximate
representations using MATLAB.
a. Plot each of these realizations. Discuss whether each random signal is positively correlated,
negatively correlated, or more-or-less uncorrelated.
b. Estimate the mean of each process using “sample mean”.
c. Compute the unbiased version of the sample autocorrelation function 𝑟E& [𝑘] for 0 ≤ k ≤100 using
H989I
1
𝑟E& [𝑘] = G 𝑥[𝑛 + 𝑘] 𝑥[𝑛] for 𝑘 ≥ 0
𝑁−𝑘
JK7
where 𝑁 is the length of the sequence (which is equal to 512 for the provided datasets). Plot
these sample autocorrelation functions over the interval −100 ≤ 𝑘 ≤ 100. Was your guess
about correlation of these sequences in part (a) correct?
d. Form a Toeplitz correlation matrix using the first ten lag values of the estimated correlation
function. For each random process, compute the eigenvalues and eigenvectors of the correlation
matrix. Plot the first ten values of each random sequence and the result of approximating each
sequence by a truncated Karhunen-Loeve expansion using one, two, four, and eight basis
functions corresponding to the eigenvectors of its correlation matrix with the largest
eigenvalues. For each truncated representation also plot the error sequence 𝑒[𝑛] = 𝑥[𝑛] −
𝑥;[𝑛]. In addition, plot the squared error ∑|𝑒[𝑛]|, as a function of 𝐿 for 1 ≤ 𝐿 ≤ 10, where 𝐿
is the number of basis functions used in the representation. Comment on your results.
(Your codes should also be provided.)

hw2

Uploaded by

Copyright:

Available Formats

hw2

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

hw2

Uploaded by

Copyright:

Available Formats

EE 503 - HW #2

(Due: 23:59, Jan. 5, 2025)

4. (IIR Wiener filtering) Problem 7.12 from Hayes

(Your codes should also be provided.)

(Your codes should also be provided.)

You might also like