Department of Electronic Systems

Examination paper for

TTT4120 Digital Signal Processing

Academic contact during examination: Stefan Werner

Phone: +358401424948

Examination date: Saturday, August 12, 2017

Examination time (from-to): 09:00-13:00
Permitted examination support material:
D - Basic calculator allowed
No printed or handwritten materials allowed
Other information:
 Exam consists of four (4) problems.
 A few basic formulas are provided in the Appendix

Language: English.
Number of pages (front page excluded): 6
Number of pages enclosed: 2

Informasjon om trykking av eksamensoppgave

Checked by:
Originalen er:

1-sidig □ 2-sidig □ ____________________________

sort/hvit □ farger □ Date Signature

skal ha flervalgskjema □

___________________________________________________________________________________________
Merk! Studentane finn sensur i Studentweb. Har du spørsmål om sensuren må du kontakte instituttet ditt.
Eksamenskontoret vil ikkje kunne svare på slike spørsmål.
Problem 1 (2+6+5+5=18): Basics of filter theory and design

A causal filter is given by the following difference equation

𝑦[𝑛] = 𝑥[𝑛] + 𝛼 2 𝑦[𝑛 − 2]

where 𝛼 is a finite real-valued constant.

1a) Provide the system function corresponding to 𝑦[𝑛] in the form

1
𝐻(𝑧) =
1 + 𝑎1 𝑧 −1 + 𝑎2 𝑧 −2

1b) Express the filter as a cascade of two filters, i.e.

1 1
𝐻(𝑧) = 𝐻1 (𝑧) ⋅ 𝐻2 (𝑧) = ⋅
(1 − 𝑝1 𝑧 ) (1 − 𝑝2 𝑧 −1 )
−1

where 𝑝1 and 𝑝2 denote the poles of the filter. Based on your findings, determine the range
of 𝛼 that yields a stable filter. Draw the pole-zero plot and sketch the region of convergence
(ROC) for a stable filter realization.

1c) Express the filter in its parallel form

𝐴1 𝐴2
𝐻(𝑧) = 𝐻3 (𝑧) + 𝐻4 (𝑧) = +
(1 − 𝑝1 𝑧 ) (1 − 𝑝2 𝑧 −1 )
−1

and provide the unit impulse response ℎ[𝑛] = 𝒵 −1 {𝐻(𝑧)}.

1d) Provide answers to the following questions (justify the answers):

 Find the causal inverse filter 𝐻𝐼 (𝑧) such that 𝐻(𝑧)𝐻𝐼 (𝑧) = 1.
 Find the ROC for 𝐻𝐼 (𝑧) and sketch the pole-zero plot.
 For 𝛼 = 1, discuss what type of filter 𝐻𝐼 (𝑧) is (lowpass, highpass, bandpass, allpass).
 Determine the range for 𝛼 for which 𝐻𝐼 (𝑧) is a minimum-phase filter.
 Does filter 𝐻𝐼 (𝑧) have linear phase?

1 (4)
Problem 2 (6+6+6+2=20): Filter structures and implementations

The filter in Problem 1a) with 𝜶 = 𝟎. 𝟓 is implemented using fixed-point representation with
𝐵 + 1 bits and dynamic range [−1,1). Rounding is performed after each multiplication and the
rounding error 𝑒[𝑛] can be modeled as white noise with variance 𝜎𝑒2 = 2−2𝐵 /12. Consequently,
each multiplier in the fixed-point implementation is modeled as
𝑄(𝑎𝑦[𝑛 − 𝑘]) = 𝑎𝑦[𝑛 − 𝑘] + 𝑒[𝑛]

which is equivalent to adding noise sources after multipliers in the infinite-precision realization.
Rounding noise sources combine into an equivalent noise signal 𝑧[𝑛] at the filter output with
variance 𝜎𝑧2 (see hint below).
2a) Draw the direct-form structure II (DF-II) of 𝐻(𝑧) with noise sources due to rounding
included. Determine the variance of the round-off noise at the filter output.

2b) Draw the cascade-structure, 𝐻1 (𝑧)𝐻2 (𝑧), with noise sources due to rounding included.
Determine the variance of the round-off noise at the filter output.

2c) Draw the parallel-structure, 𝐻3 (𝑧) + 𝐻4 (𝑧), with noise sources due to rounding included.
Determine the variance of the round-off noise at the filter output. You may put the
multiplier(s) 𝐴1 and 𝐴2 , obtained from the residue calculus in Problem 1c, at the output of
the structure.

2d) Which of the three implementations above suffers the most from rounding noise? Which
implementation suffers the least? Justify your answers.

2
[Hint:] Assuming noise source 𝑒𝑖 [𝑛] with variance 𝜎𝑒𝑖 acts as input to (sub-)filter ℎ𝑖 [𝑛] that
terminates at the output. The variance of the noise signal 𝑧𝑖 [𝑛], due to 𝑒𝑖 [𝑛], is given by

2 2 2
𝜎𝑧𝑖 = 𝜎𝑒𝑖 𝑟ℎ𝑖ℎ𝑖 [0] = 𝜎𝑒𝑖 ∑ ℎ𝑖2 [𝑘]
𝑘

𝑒𝑖 [𝑛] 𝑧𝑖 [𝑛]
𝐻𝑖 (𝑧)

2
𝜎𝑒𝑖 = 𝐸{𝑒𝑖2 [𝑛]} 2
𝜎𝑧𝑖 = 𝐸{𝑧𝑖2 [𝑛]}

2 (4)
Problem 3 (2+6+6=14): Parametric modeling

𝑊[𝑛] 𝑋[𝑛] 𝑍[𝑛]

𝐻1 (𝑧) 𝐻2 (𝑧)

𝛾𝑊𝑊 [𝑙] 𝛾𝑋𝑋 [𝑙] 𝛾𝑍𝑍 [𝑙]

Γ𝑊𝑊 (𝑓) Γ𝑋𝑋 (𝑓) Γ𝑍𝑍 (𝑓)

Fig. 1: Filtering of stochastic processes

A wide-sense stationary (WSS) stochastic process 𝑋[𝑛] is generated by filtering a white noise
2
process 𝑊[𝑛], with autocorrelation sequence 𝛾𝑊𝑊 [𝑙] = 𝜎𝑊 𝛿[𝑙], through a causal and stable filter
𝐻(𝑧), as depicted in Fig. 1. The autocorrelation sequence and spectrum of 𝑋[𝑛] is obtained from
∞
2
𝜎𝑊 ∑ ℎ[𝑛]ℎ[𝑛 + 𝑙] , 𝑙≥0
𝛾𝑋𝑋 [𝑙] = {
𝑛=0
𝛾𝑋𝑋 [−𝑙], 𝑙<0

Γ𝑋𝑋 (𝑓) = |𝐻(𝑓)|2 Γ𝑊𝑊 (𝑓)

3a) Provide answers (with motivations) to the following two questions:

 What type of process, AR(𝑝), MA(𝑞), or ARMA(𝑝, 𝑞), is 𝑋[𝑛] when the noise is
filtered by 𝐻(𝑧) in Problem 1a for the case of 𝜶 = 𝟎. 𝟓? Provide the model order.

3b) For filter 𝐻(𝑧) specified in 3a) and 𝝈𝟐𝑾 = 𝟏, show that the respective autocorrelation
sequence and spectrum of 𝑋[𝑛] are given by

16 1
𝛾𝑋𝑋 [𝑙] = {15 ⋅ 2|𝑙| , 𝑙 even
0, 𝑙 odd
1
Γ𝑋𝑋 (𝑓) =
17 1
16 − 2 cos 4𝜋𝑓
3c) You are given the task to design a second-order predictor to model 𝑋[𝑛]. That is, you form
an estimate of 𝑋[𝑛], denoted 𝑋̂[𝑛], through the following linear combination 𝑋̂[𝑛] =
𝑎1 𝑋[𝑛 − 1] + 𝑎2 𝑋[𝑛 − 2].
 Find the optimal values of 𝑎1 and 𝑎2 that minimize the prediction error power.
Hint: Use the values 𝛾𝑋𝑋 [−2] through 𝛾𝑋𝑋 [2] from 3b) together with the Normal
equations, see Section H in the Table of formulas attached to the exam.
 Find the resulting prediction error power 𝜎𝑓2 when using the optimal coefficients.
 Comment on your results.

3 (4)
Problem 4 (6+2+6=14): Sampling and rate-conversion

𝑋𝑎 (𝐹)

𝐹 [Hz]
−8000 8000

Fig. 2: Spectrum 𝑋𝑎 (𝐹) of continuous-time signal 𝑥𝑎 (𝑡)

Let 𝑥𝑎 (𝑡) be a continuous-time signal whose spectrum 𝑋𝑎 (𝐹) is shown in Fig. 2. Signal 𝑥𝑎 (𝑡) is
1
sampled at rate 𝐹𝑥 = 𝑇 = 16 kHz to generate sequence 𝑥[𝑛] = 𝑥𝑎 (𝑡)|𝑡 =𝑛𝑇𝑥 . We would now like
𝑥
to design a system that changes the sampling frequency of signal 𝑥[𝑛] in digital domain from 𝐹𝑥
1
to 𝐹𝑦 = 𝑇 = 12 kHz. Let 𝑦[𝑚] be the resulting output signal. The conversion should not
𝑦
introduce any distortion due to aliasing.

4a) Sketch the block diagram of the digital system that implements the sampling rate conversion.
Explain the function of each individual block along with the specifications that arise from
the rate conversion at hand (e.g., upsampling/downsampling factors, filter bandwidths, etc.).

4b) Will the rate conversion above incur any loss of information? That is, can 𝑥𝑎 (𝑡) be perfectly
reconstructed from 𝑦[𝑚]?

4c) Sketch the spectra of all the signals in the rate-conversion system.

4 (4)
Appendix: TTT4120 Table of formulas, 2017

A. Sequences:

1−𝛼𝑁
∑𝑁−1 𝑛
𝑛=0 𝛼 = 1−𝛼

1 1
|𝛼| < 1 ⇒ ∑∞ 𝑛
𝑛=0 𝛼 = 1−𝛼 and − ∑−∞ 𝑛
𝑛=−1 𝛼 = 1−𝛼

1−𝛼𝑁 𝑁𝛼𝑁
∑𝑁−1 𝑛
𝑛=0 (𝑛 + 1)𝛼 = (1−𝛼)2 − 1−𝛼 ; 𝛼≠1

1
|𝛼| < 1 ⇒ ∑∞ 𝑛
𝑛=0(𝑛 + 1)𝛼 = (1−𝛼)2

B. Linear convolution:

𝑦[𝑛] = ℎ[𝑛] ∗ 𝑥[𝑛] = ∑∞ ∞

𝑘=−∞ ℎ[𝑘]𝑥[𝑛 − 𝑘] = ∑𝑘=−∞ 𝑥[𝑘]ℎ[𝑛 − 𝑘]

𝑌(𝑧) = 𝐻(𝑧)𝑋(𝑧)

𝑌(𝑓) = 𝐻(𝑓)𝑋(𝑓)

𝑌(𝑘) = 𝐻(𝑘)𝑋(𝑘), 𝑘 = 0,1, … , 𝑁 − 1 where 𝑌(𝑘) = 𝑌(𝑓𝑘 ) with 𝑓𝑘 = 𝑘/𝑁

C. Transforms:

Z-transform: 𝐻(𝑧) = ∑∞
𝑛=−∞ ℎ[𝑛]𝑧
−𝑛

DTFT: 𝐻(𝑓) = ∑∞
𝑛=−∞ ℎ[𝑛]𝑒
−𝑗2𝜋𝑓𝑛

DFT: 𝐻(𝑘) = ∑𝑁−1

𝑛=0 ℎ[𝑛]𝑒
−𝑗2𝜋𝑓𝑛𝑘/𝑁
𝑘 = 0,1, … , 𝑁 − 1
1
IDFT: ℎ[𝑛] = 𝑁 ∑𝑁−1
𝑘=0 𝐻(𝑘)𝑒
𝑗2𝜋𝑓𝑛𝑘/𝑁
𝑛 = 0,1, … , 𝑁 − 1

D. Sampling theorem:

Given an analog signal 𝑥𝑎 (𝑡) sampled at 𝐹𝑠 = 1/𝑇. The DTFT of the resulting discrete-time
sequence 𝑥[𝑛] = 𝑥𝑎 (𝑡)|𝑡=𝑛𝑇 is given by

𝑋(𝑓) = 𝑋(𝐹/𝐹𝑠) = 𝐹𝑠 ∑∞
𝑘=−∞ 𝑋([𝑓 − 𝑘]𝐹𝑠 )

E. Autocorrelation, energy spectrum and Parseval:

Given a sequence ℎ[𝑛] with finite energy 𝐸ℎ

Autocorrelation: 𝑟ℎℎ [𝑙] = ∑∞

𝑛=−∞ ℎ[𝑛]ℎ[𝑛 + 𝑙] 𝑙∈ℤ

Energy spectrum: 𝑆ℎℎ (𝑧) = 𝐻(𝑧)𝐻(𝑧 −1 ) ⇒ 𝑆ℎℎ (𝑓) = |𝐻(𝑓)|2

2𝜋
Parseval’s theorem: 𝐸ℎ = 𝑟ℎℎ [0] = ∑∞ 2 2
𝑛=−∞ ℎ [𝑛] = ∫0 |𝐻(𝑓)| 𝑑𝑓

Appendix: TTT4120 Table of formulas A1 (A2)

F. Multirate:

Decimation (downsampling) where 𝑇𝑦 = 𝐷𝑇𝑥

𝑣(𝑚𝑇𝑦 ) = ∑∞
𝑘=−∞ ℎ[(𝑚𝐷 − 𝑘)𝑇𝑥 ]𝑥(𝑘𝑇𝑥 ) 𝑚∈ℤ

Interpolation (upsampling) where 𝑇𝑦 = 𝑇𝑥 /𝐼

𝑦(𝑙𝑇𝑦 ) = ∑∞
𝑛=−∞ ℎ[(𝑙 − 𝑛𝐼)𝑇𝑦 ]𝑥(𝑛𝑇𝑥 ) 𝑙∈ℤ
𝐷
Rate coversion where 𝑇𝑦 = 𝐷𝑇𝑣 = 𝐼 𝑇𝑥

𝑦(𝑙𝑇𝑦 ) = ∑∞
𝑚=−∞ ℎ[(𝑙𝐷 − 𝑚𝐼)𝑇𝑣 ]𝑥(𝑚𝑇𝑥 ) 𝑙∈ℤ

G. Autocorrelation, power density spectrum and Wiener-Khintchin:

Given a wide-sense stationary and ergodic sequence 𝑋[𝑛] with infinite energy

Autocorrelation: 𝛾𝑋𝑋 [𝑙] = 𝐸{𝑋[𝑛]𝑋[𝑛 + 𝑙]} 𝑙∈ℤ

Power spectrum: 𝛤𝑋𝑋 (𝑧) = 𝒵{𝛾𝑋𝑋 [𝑙]} ⇒

Wiener-Khintchin: 𝛤𝑋𝑋 (𝑓) = DTFT{𝛾𝑋𝑋 [𝑙]} = ∑∞

𝑙=−∞ 𝛾𝑋𝑋 [𝑙]𝑒
−𝑗2𝜋𝑓𝑙

H. Yule-Walker and Normal equations where 𝒂𝟎 = 𝟏:

Autocorrelation: ∑𝑃𝑘=0 𝑎𝑘 𝛾𝑋𝑋 [𝑛 − 𝑘] = 𝜎𝑓2 𝛿[𝑛] 𝑛 = 0, … , 𝑝

Normal equations: ∑𝑃𝑘=1 𝑎𝑘 𝛾𝑋𝑋 [𝑛 − 𝑘] = −𝛾𝑋𝑋 [𝑛] 𝑛 = 1, … , 𝑝

I. Some common z-transform pairs:

Appendix: TTT4120 Table of formulas A2 (A2)