Family Name: Given Name:

Student Number: Degree Program:

Universität Paderborn
Signal and System Theory Group

Written Examination in
Statistical Signal Processing

Winter Semester 2017/18

20 February 2018

Please note:

• Have your student ID ready for inspection.

• Clearly write your name and student ID number on every sheet of paper.

• Attempt all questions. You have 120 minutes available.

• This is a closed-book exam. You may, however, use one handwritten, double-sided, A4-
size formula sheet.

• You may NOT use any electronic devices (including calculators).

• Justify all your answers with appropriate arguments and/or algebraic or numerical devel-
opments.

• Put all formal results on the single white graph sheets (not the outside folder). The blue
sheets are intended to be used only as scratch paper and will not be graded.

Good luck!

Question 1 2 3 4 Sum Grade

Points
Statistical Signal Processing Page 2 of 5

Question 1 Estimation (25 pts.)

Consider the following model where N independent and identically distributed (i.i.d.) samples
of a random variable x are generated as

x[n] = a + w[n], n = 1, 2, . . . , N .

Here, a is an unknown parameter to be estimated, and w[n] are i.i.d. samples of zero-mean
Gaussian noise with variance σ2 .

(a) (5 pts.) Derive the maximum likelihood estimator, â, of a. Is â biased?

(b) (5 pts.) Derive the probability density function (pdf ) of â.

(c) (5 pts.) Show that â is a consistent estimator of a.

(d) (10 pts.) Now assume that the pdf of w[n] is not Gaussian but unknown. Can you still find
a linear minimum-variance unbiased (MVUB) estimator of a? If yes, find this estimator
and its variance. If not, explain why.

Written Examination Winter Semester 2017/18

Statistical Signal Processing Page 3 of 5

Question 2 Complex random variables (25 pts.)

(a) (6 pts.) Show that a zero-mean complex Gaussian random vector s is proper if and only if
s and e j α s have the same probability density function for all α ∈ IR.

(b) (6 pts.) Consider an m-dimensional complex random vector x and a complex n×m matrix
T. Show that if x is proper, then Tx is also proper. Is the converse also true? That is, does
this mean that if x is improper, then Tx is also improper?

(c) Now we would like to estimate a zero-mean complex random variable x from its complex
conjugate x ∗ .

i. (7 pts.) Derive the linear MMSE estimator. Note that the distribution of x has not
been specified.

ii. (6 pts.) Find the MSE of the linear MMSE estimator. When is the MSE minimal and
when is it maximal?

Written Examination Winter Semester 2017/18

Statistical Signal Processing Page 4 of 5

Question 3 Detection (25 pts.)

(a) In general, only God can build “perfect” detectors with probability of detection P D = 1 and
probability of false alarm P F = 0. But for some very special problems, perfect detectors
may exist. Consider the binary hypothesis test

H0 : x is uniform on [−c, c]
H1 : x is uniform on [1 − c, 1 + c],

where c > 0, based on a single observation of x.

i. (4 pts.) For which value(s) of c does there exist a perfect detector for the above test?

ii. (5 pts.) For this part consider c = 2. Can you obtain a detector for the above test
with P D > 0 for P F = 0? If yes, find the corresponding value of P D . If not, justify your
answer.

(b) Consider the following detection problem based on N i.i.d. samples:

H0 : x[n] = w 0 [n]
H1 : x[n] = w 1 [n]

where n = 1, 2, . . . , N , and w 0 [n] and w 1 [n] are i.i.d. samples of zero-mean Gaussian with
variances σ20 and σ21 , respectively, with σ20 < σ21 .

i. (7 pts.) Derive the Neyman-Pearson detector for this detection problem. Simplify
the test statistic such that it is independent of σ20 and σ21 .

ii. (5 pts.) Draw the decision regions of this detector for N = 1. Note that the test statis-
tic can be further simplified for N = 1.

(c) (4 pts.) Explain in words the meaning of a “sufficient statistic” in the context of a Neyman-
Pearson detector.

Written Examination Winter Semester 2017/18

Statistical Signal Processing Page 5 of 5

Question 4 Correlation analysis (25 pts.)

In this question, we will look at how to define a correlation coefficient between random vectors.
For that, we consider two zero-mean random vectors x ∈ IRm and y ∈ IRn with known covariance
matrices Rxx , R y y , and cross-covariance matrix Rx y . We also define the coherence matrix

Cx y = R−1/2 −1/2
xx Rx y R y y

and its singular value decomposition (SVD)

Cx y = FKGT .

The singular values k 1 , ..., k p , p = min(m, n), are called the canonical correlations between x
and y. It can be shown (and you can use that fact without proof ) that the canonical correlations
satisfy 0 ≤ k i ≤ 1, i = 1, ..., p.

(a) (4 pts.) In the scalar case m = n = 1, show how the canonical correlation k 1 is related to
the correlation coefficient between scalar x and y.

(b) (3 pts.) Show that the squared canonical correlations are equal to the eigenvalues of
Cx y CTx y .

(c) (6 pts.) Show that

³ ´ Xp
−1 T
tr R−1
xx R x y R R
yy xy = k i2
i =1

(d) (8 pts.) Show that the canonical correlations are invariant under linear nonsingular trans-
formation. That is, show that for square nonsingular Ax and A y , the canonical correlations
between x and y are equal to the canonical correlations between Ax x and A y y.

(e) (4 pts.) Combining all of these results, argue why the quantity

p
1X
ρ= k2
p i =1 i

would be a good candidate for defining a correlation coefficient between random vectors
x and y.

Written Examination Winter Semester 2017/18