Imperial College London: EE4-12 EE9-AO2 EE9-SC1

EE4-12
IMPERIAL COLLEGE LONDON EE9-AO2

EE9-SC1
DEPARTMENT OF ELECTRICAL AND ELECTRONIC ENGINEERING

EXAMINATIONS 2014
ExamHeader:
MSc and EEE/EIE PART IV: MEng and ACGI
DIGITAL SIGNAL PROCESSING AND DIGITAL FILTERS
Tuesday, 6 May 10:00 am
Time allowed: 3:00 hours
There are FOUR questions on this paper.
Answer Question 1 and any TWO other questions
Question 1 is worth 40% of the marks and other questions are worth 30%
Any special instructions for invigilators and information for

candidates are on page 1.
Examiners responsible First Marker(s) : D.M. Brookes

Second Marker(s) : P.T. Stathaki
© Imperial College London

D IGITAL S IGNAL P ROCESSING AND D IGITAL F ILTERS
Information for Candidates:
Notation
• All signals and filter coefficients are real-valued unless explicitly noted otherwise.
• Unless otherwise specified, upper and lower case letters are used for sequences and their z-transforms
respectively. The signal at a block diagram node V is v[n] and its z-transform is V (z).
• x[n] = [a, b, c, d, e, f ] means that x[0] = a, . . . x[5] = f and that x[n] = 0 outside this range.
• ℜ(z), ℑ(z), z∗ , |z| and ∠z denote respectively the real part, imaginary part, complex conjugate,
magnitude and argument of a complex number z.
Abbreviations
BIBO Bounded Input, Bounded Output

CTFT Continuous-Time Fourier Transform
DCT Discrete Cosine Transform
DFT Discrete Fourier Transform
DTFT Discrete-Time Fourier Transform
LTI Linear Time-Invariant
MDCT Modified Discrete Cosine Transform
SNR Signal-to-Noise Ratio
Standard Sequences
• δ [n] = 1 for n = 0 and 0 otherwise.
• δcondition [n] = 1 whenever "condition" is true and 0 otherwise.
• u[n] = 1 for n ≥ 0 and 0 otherwise.
Geometric Progression
1−α r+1 z−r−1 α q z−q −α r+1 z−r−1

• ∑rn=0 α n z−n = 1−αz−1
or, more generally, ∑rn=q α n z−n = 1−αz−1
Digital Signal Processing and Digital Filters Imperial

c College London 1/11
Forward and Inverse Transforms
−n 1
¸
z: X(z) = ∑∞
−∞ x[n]z x[n] = 2π j X(z)zn−1 dz
´∞ − jΩt dt 1
´∞ jΩt dΩ
CTFT: X( jΩ ) = −∞ x(t)e x(t) = 2π −∞ X( jΩ )e
− jωn 1
´π
DTFT: X(e jω ) = ∑∞
−∞ x[n]e x[n] = 2π −π X(e
jω )e jωn dω
kn kn
DFT: X[k] = ∑N−1
0 x[n]e− j2π N x[n] = 1
N ∑0N−1 X[k]e j2π N
2π(2n+1)k
DCT: X[k] = ∑N−1
n=0 x[n] cos 4N x[n] = X[0]
N
N−1
+ N2 ∑n=1 X[k] cos 2π(2n+1)k
4N
2π(2n+1+N)(2k+1) 2π(2n+1+N)(2k+1)
MDCT: X[k] = ∑2N−1
n=0 x[n] cos 8N y[n] = 1
N ∑N−1
0 X[k] cos 8N
Convolution
V e jω = X e jω Y e jω

DTFT: v[n] = x[n] ∗ y[n] , ∑∞
r=−∞ x[r]y[n − r] ⇔
1 1
´π
V e jω = e jθ Y e j(ω−θ ) dθ

e jω ~Y e jω =

v[n] = x[n]y[n] ⇔ 2π X 2π −π X
DFT: v[n] = x[n] ~N y[n] , ∑N−1

r=0 x[r]y[(n − r) mod N ] ⇔ V [k] = X[k]Y [k]
N−1
v[n] = x[n]y[n] ⇔ V [k] = N1 X[k] ~N Y [k] , 1
N ∑r=0 X[r]Y [(k − r) mod N ]
Group Delay
F (nh[n])
jω )

The group delay of a filter, H(z), is τH (e jω ) = − d∠H(e =ℜ −z dH(z)
=ℜ where

dω H(z) dz
z=e jω F (h[n])
F ( ) denotes the DTFT.
Order Estimation for FIR Filters
Three increasingly sophisticated formulae for estimating the minimum order of an FIR filter with unity
gain passbands:
a
1. M ≈ 3.5∆ ω
a−8
2. M ≈ 2.2∆ ω
a−1.2−20 log10 b
3. M ≈ 4.6∆ ω
where a =stop band attenuation in dB, b = peak-to-peak passband ripple in dB and ∆ ω = width of
smallest transition band in normalized rad/s.

z-plane Transformations
A lowpass filter, H(z), with cutoff frequency ω0 may be transformed into the filter H(ẑ) as follows:
Target H(ẑ) Substitute Parameters

ω −ω̂
sin 0 2 1
ẑ−1 −λ
Lowpass z−1 = 1−λ ẑ−1
λ=
ω +ω̂

sin 0 2 1
ω̂ < ω̂1
ω +ω̂
−1 cos 0 2 1
Highpass z−1 = − 1+λ
ẑ +λ
ẑ−1
λ=
ω −ω̂

cos 0 2 1
ω̂ > ω̂1
ω̂ +ω̂
(ρ−1)−2λ ρ ẑ−1 +(ρ+1)ẑ−2 cos 2 2 1
ω̂2 −ω̂1

z−1 ω0

Bandpass = − (ρ+1)−2λ ρ ẑ−1 +(ρ−1)ẑ−2
λ=
ω̂ −ω̂
, ρ = cot 2 tan 2
cos 2 2 1
ω̂1 < ω̂ < ω̂2
ω̂ +ω̂
(1−ρ)−2λ ẑ−1 +(ρ+1)ẑ−2 cos 2 2 1
ω̂2 −ω̂1

z−1 = ω0

Bandstop (ρ+1)−2λ ẑ−1 +(1−ρ)ẑ−2
λ=
ω̂ −ω̂
, ρ = tan 2 tan 2
cos 2 2 1
ω̂1 ≮ ω̂ ≮ ω̂2

1. a) i) Explain what is meant by saying that a linear time-invariant system
is “BIBO-stable”. [1]
ii) The impulse response, h[n], of a linear time-invariant system satisfies
∑∞n=−∞ |h[n]| = S where S < ∞. Prove that the system is BIBO-stable
and also that H(z) converges for |z| = 1. [2]
b) A filter, with input x[n] and output y[n], is defined by the difference equation
y[n] = αy[n − 1] + (1 − α)x[n]
where 0 < α < 1 is a real constant.

i) Determine the system function of the filter, H(z), and the impulse
response, h[n], for n = −1, 0, 1, 2. [2]
ii) State the values of z at which H(z) has a pole or zero. [ 2 ]1
iii) Determine the frequency at which the filter has a gain of −3 dB.[ 3 ]
iv) If the sample frequency is fs , show that, for n ≥ 0, the impulse re-
t
sponse, h[n], is equal to a sampled version of g(t) = Ae− τ and deter-
mine the values of the constants A and τ. [2]
c) Figure 1.1 shows the block diagram of a filter implementation comprising two
delays, five multipliers with real-valued coefficients c1 , · · · , c5 and four adder
elements.
Y (z) c3 +c4 z−1 +c5 z−2
i) Show that transfer function X(z) = 1−c1 z−1 −c2 z−2
. [3]
ii) Suppose that each multiplier introduces independent additive white
noise at its output with power spectral density S(ω) = S0 and that
the noise signals are uncorrelated with x[n]. Show that the combined
effect of the five noise sources is equivalent to two additive white
noise signals at x[n] and y[n] respectively. Hence determine the over-
all power spectral density, N(ω), of the noise at y[n]. [3]
Figure 1.1

d) The impulse response of an antisymmetric FIR filter, H(z), of order M satisfies
the relation h[n] = −h[M − n].

i) Show that the magnitude response H(e jω ) can be expressed as the
absolute value of the sum of N sine waves where N = M2 if M is even
and N = M+12 if M is odd. [3]
ii) Show that H(e jω ) is necessarily zero at ω = 0 but may be non-zero
at ω = π if M is odd. Give an example of a filter for which this is
the case. [2]
iii) Derive an expression for the phase response, ∠H(e jω ), and deter-
jω )
mine the group delay, τH e jω = − d∠H(e

dω . [2]
e) Figure 1.2 shows the analysis and synthesis sections of a subband processing
system. The input and output signals are x[n] and y[n] respectively and the
intermediate signals are vm [n], um [r] and wm [n] where m = 0 or 1 according to
the subband. The corresponding z-transforms are X(z), Y (z) etc.
i) Show that it is possible to express
the overall
transfer function in the
X(z)
form Y (z) = T (z) A(z) and determine expressions
X(−z)
for T (z) and A(z).
You may assume without proof that for m = 0 or 1, [3]
1 n 1 1 o
Um (z) = Vm z 2 +Vm −z 2
2
Wm (z) = Um z2 .

ii) Explain why it is normally desirable to have A(z) ≡ 0. [2]

iii) Suppose that H0 (z) = H1 (−z) = G0 (z) = −G1 (−z). Show that in
this case A(z) = 0 and explain how the frequency responses H1 (e jω ),
G0 (e jω ) and G1 (e jω ) are related to H0 (e jω ) assuming that H0 (z) is
an FIR or IIR filter with real coefficients. [2]
Figure 1.2

f) Figure 1.3 shows an upsampler with real-valued input x[n] and output
(
x Kr if K | r
y[r] =
0 otherwise
where K | r means K is a factor of r.

Show that Y (z) = X zK .

i) [1]
ii) The energy and average power of x[n] are defined respectively as
∞
Ex = ∑ |x[n]|2
n=−∞
N
1
Px = lim ∑ |x[n]|2 .
N→∞ 2N + 1
n=−N
Give expressions for the energy and average power of y[r] in terms
of Ex and Px . [2]
iii) Figure 1.4 shows the power spectral density of x[n] which comprises
white noise of unit magnitude together with a bandpass signal com-
ponent occupying the range 0.5 < |ω| < 1. Sketch the power spectral
density of y[r] when K = 3 and give the magnitude of its white noise
component and the magnitude and frequency range of all bandpass
components. [3]
iv) The diagram of Fig. 1.3 is followed by a lowpass filter to remove
spectral images. If K = 3 and x[n] is as specified in part iii) above,
determine the transition bandwidth and transition band centre fre-
quency of a suitable lowpass filter and explain the reasons for your
choices.
[2]
6
Power Spectral Density
0
-3 -2 -1 0 1 2 3
Frequency (rad/samp)
Figure 1.3 Figure 1.4

2. a) Suppose that G1 (z) = 1 − pz−1 and G2 (z) = 1 − qz−1 where
the constants p and
1 jω jω

q may be complex. If q = p∗ show that G1 (e ) = α G2 (e ) for all ω and

determine an expression for the constant α. [4]
−1 −2
b) H1jω(z) = 4 + 14zjω − 8z . Determine the coefficients of H2 (z)
Suppose that
such that H1 (e ) = H2 (e ) for all ω and that all the zeros of H2 (z) lie
inside the unit circle. [4]
jω
)
c) When designing an IIR filter H(e jω ) = B(e
A(e jω )
to approximate a complex target
response D(ω) two error measures that may be used are the weighted solution
error, ES (ω), and the weighted equation error, EE (ω), defined respectively by
B(e jω )

ES (ω) = WS (ω) − D(ω)
A(e jω )
EE (ω) = WE (ω) B(e jω ) − D(ω)A(e jω ) .

Explain the relative advantages of the two error measures and explain the pur-
pose of the real-valued non-negative weighting functions WS (ω) and WE (ω).
[2]
d) Suppose that 0 ≤ ω1 < ω2 < . . . < ωK ≤ π is a set of K frequencies and that
A(z) = 1 + [z−1 z−2 · · · z−N ]a and B(z) = [1 z−1 z−2 · · · z−M ]b where a and b
are real-valued coefficient column vectors.
i) Show that it is possible to express the equations EE (ωk ) = 0 for
1≤k ≤ K as a set of K simultaneous linear equations in the form
a
(P Q) = d.
b
State the dimensions of the matrices P and Q and of the vector d and
derive expressions for the elements of P, Q and d. [4]
ii) Explain how, by separating the real and imaginary parts of P, Q and
d,
it ispossible to obtain a set of simultaneous linear equations for
a
in which all coefficients are real-valued. Explain the circum-
b
stances under which some of the resultant equations will necessarily
have all-zero coefficients. [4]
iii) Explain why it may be desirable to apply the transformation of part
b) after obtaining the solution to the equations of part d) ii). [2]
iv) Assuming that ω1 = 0 and ωK = π, determine the minimum value of
K to ensure that the equations of part d) ii) are not underdetermined.
[4]
b
π

e) Suppose now that H(z) = 1+az −1 , that K = 3, that ωk = 0, 2 , π , that
(
2 for ω ≤ 0.25π
D(ω) = .
1 for ω > 0.25π
WE (ω) ≡ 1
Determine the numerical values of the elements of P, Q and d and hence deter-
mine the numerical values of a and b that minimize ∑k |EE (ωk )|2 . [6]
You may assume without proof that the least squares solution to an overdeter-
−1 T
mined set of real-valued linear equations, Rx = q, is given by x = RT R R q
assuming that R has full column rank.

3. a) Figure 3.1 shows the block diagram of a system that multiplies the input sample
rate by QP where P and Q are coprime with P < Q .
i) Explain why the cutoff frequency of the lowpass filter H(z) should
be placed at the Nyquist rate of the output signal, y[m] and give the
normalized cutoff frequency, ω0 , in rad/sample in terms of P and/or
Q.
a
Using the approximation formula M ≈ 3.5∆ ω , determine the required
filter order M in terms of P and/or Q if the stopband attenuation in dB
is a = 60 and the normalized transition bandwidth is ∆ ω = 0.1ω0 .
[4]
ii) Using the value of M from part a) i), estimate the average number
of multiplications per input sample, x[n], needed to implement the
system in the form of Figure 3.1. [2]
iii) The filter H(z) has a symmetrical impulse response h[r] = g[r]w[r]
for 0 ≤ r ≤ M where g[r] is the impulse response of an ideal lowpass
filter with cutoff frequency ω0 and w[r] is a symmetrical window
function.
Derive an expression for the ideal response, g[r], in terms of ω0 , M
and r. [4]
b) The filter H(z) is now implemented as a polyphase filter as shown in Fig. 3.2.
The filter implementation uses a single set of delays and multipliers with com-
mutated coefficients.
i) State the length of the filter impulse response h0 [n] in terms of M, P
and/or Q and give an expression for the coefficients h0 [n] in terms of
h[r]. [2]
ii) If x[n] = 0 for n < 0, give expressions for v[0], v[1], v[2P+1] in terms
of the input x[n] and the coefficients h p [n]. [2]
iii) Explain how it is possible to eliminate the output decimator by chang-
ing both the sequence and rate at which the coefficient sets, h p [n] are
accessed.
Determine the new coefficient set order for the case P = 5 and Q = 7.
[3]
iv) Determine the number of multiplications per input sample for the
system of part b) iii) and the number of distinct coefficients that must
be stored. You may assume that M + 1 is a multiple of P. [2]
c) Suppose now that the sample rate of the input, x[n], is 18 kHz and that the
system is implemented as in part b) iii) with the values of a and ∆ ω as given in
part a) i).
Determine the values of P, Q and M when the sample rate of the output, y[m],
is (i) 10 kHz and (ii) 10.1 kHz [note that 101 is a prime number].
For each of these cases estimate the number of multiplications per input sample
and the number of distinct coefficients that must be stored. [5]

d) In a Farrow filter, the coefficients, h p [n], are approximated by a low-order poly-
nomial fn (t) where t = Pp for 0 ≤ p ≤ P − 1.
i) Assuming that a rectangular window, w[r] ≡ 1, is used in the design
of H(z) and that ω0 = πP , give an expression for the target value of
f0 (t) in terms of t, M , P and Q. [3]
ii) If the polynomials, fn (v), are of order K = 5, determine the number
of coefficients that must be stored for each of the cases defined in
part c).
[3]

4. A complex-valued frequency-modulated signal, x(t) = a(t)e jφ (t) , has a 0 Hz carrier fre-
quency and a peak frequency deviation of d = 75 kHz. The´ t amplitude, a(t), is approxi-
mately constant with a(t) ≈ 1 and the phase is φ (t) = k 0 m(τ)dτ where k is a constant
and m(t) is a baseband audio signal with bandwidth b = 15 kHz. The signal x(t) is
sampled at 400 kHz to obtain the discrete-time signal x[n].
a) Carson’s rule for the bandwidth of a double-sideband FM signal is B = 2 (d + b).
Use this to determine the single-sided bandwidth, ω0 , of x[n] in radians/sample.
[2]

b) Show that m(t) = k−1 a−2 (t)ℑ x∗ (t) dx(t)
dt where ℑ ( ) denotes the imaginary
part. [4]
c) Figure 4.1 shows a block diagram that implements the equation of part b) in
discrete time. Complex-valued signals are shown as bold lines and are repre-
sented using their real and imaginary parts. The block labelled “Conj” takes the
complex conjugate of its input. The differentiation block, D(z), is designed as
an FIR filter using the window method with a target response
(
jcω for |ω| ≤ ω1
D(e jω ) =
0 otherwise
where c is a scaling constant.

i) ¯ of D(z) in simplified form.[ 4 ]
Determine the impulse response d[n]
ii) Assuming that ω1 = ω02+π , draw dimensioned sketches showing the
magnitude and phase responses of D(e jω ) over the range −π ≤ ω ≤
π.
[3]
iii) Assume that the DTFT of the window function used when designing
18
D(z) has a main lobe width of ω = ± M+1 for a window of length
ω0 +π
M + 1. If ω1 is chosen as ω1 = 2 , determine the smallest value
of M that will ensure that the transition in the response of D(e jω )
near ω = ω1 lies completely within the range (ω0 , π). [3]
iv) Stating any assumptions, determine the maximum value of c that will
ensure |s[n]| ≤ 1 where s[n] is the output of the differentiation block,
D(z), as shown in Figure 4.1. [4]
d) An alternative choice for the target response is

 − jcω (π+ω)

 π−ω1
1
for − π < ω ≤ −ω1
jω
D(e
e ) = jcω for |ω| ≤ ω1
 jcω1 (π−ω)

π−ω1 for ω1 < ω ≤ π
i) Assuming that ω1 = ω02+π , draw dimensioned sketches showing the

e jω ) over the range −π ≤ ω ≤
magnitude and phase responses of D(e
π.
[4]
ii) Outline the relative advantages and disadvantages of using D(ee jω )
rather than D(e jω ) as the target response when designing D(e jω ).
[2]

e) An alternative structure that avoids any divisions is shown in Fig. 4.2 where
the polynomial f (v) is the truncated Taylor series for v−1 expanded around
v = 1. Determine f (v) for the cases when it is (i) a linear expression and (ii) a
quadratic expression. In each case determine the gain error (expressed in dB)
resulting from the approximation when a(t) = 1.1. [4]

D IGITAL S IGNAL P ROCESSING AND D IGITAL F ILTERS
********* Solutions *********
Information for Candidates:
Notation
• All signals and filter coefficients are real-valued unless explicitly noted otherwise.
• Unless otherwise specified, upper and lower case letters are used for sequences and their z-transforms
respectively. The signal at a block diagram node V is v[n] and its z-transform is V (z).
• x[n] = [a, b, c, d, e, f ] means that x[0] = a, . . . x[5] = f and that x[n] = 0 outside this range.
• ℜ(z), ℑ(z), z∗ , |z| and ∠z denote respectively the real part, imaginary part, complex conjugate,
magnitude and argument of a complex number z.
Abbreviations
BIBO Bounded Input, Bounded Output

CTFT Continuous-Time Fourier Transform
DCT Discrete Cosine Transform
DFT Discrete Fourier Transform
DTFT Discrete-Time Fourier Transform
LTI Linear Time-Invariant
MDCT Modified Discrete Cosine Transform
SNR Signal-to-Noise Ratio
Standard Sequences
• δ [n] = 1 for n = 0 and 0 otherwise.
• δcondition [n] = 1 whenever "condition" is true and 0 otherwise.
• u[n] = 1 for n ≥ 0 and 0 otherwise.
Geometric Progression
1−α r+1 z−r−1 α q z−q −α r+1 z−r−1

• ∑rn=0 α n z−n = 1−αz−1
or, more generally, ∑rn=q α n z−n = 1−αz−1

Forward and Inverse Transforms
−n 1
¸
z: X(z) = ∑∞
−∞ x[n]z x[n] = 2π j X(z)zn−1 dz
´∞ − jΩt dt 1
´∞ jΩt dΩ
CTFT: X( jΩ ) = −∞ x(t)e x(t) = 2π −∞ X( jΩ )e
− jωn 1
´π
DTFT: X(e jω ) = ∑∞
−∞ x[n]e x[n] = 2π −π X(e
jω )e jωn dω
kn kn
DFT: X[k] = ∑N−1
0 x[n]e− j2π N x[n] = 1
N ∑0N−1 X[k]e j2π N
2π(2n+1)k
DCT: X[k] = ∑N−1
n=0 x[n] cos 4N x[n] = X[0]
N
N−1
+ N2 ∑n=1 X[k] cos 2π(2n+1)k
4N
2π(2n+1+N)(2k+1) 2π(2n+1+N)(2k+1)
MDCT: X[k] = ∑2N−1
n=0 x[n] cos 8N y[n] = 1
N ∑N−1
0 X[k] cos 8N
Convolution
V e jω = X e jω Y e jω

DTFT: v[n] = x[n] ∗ y[n] , ∑∞
r=−∞ x[r]y[n − r] ⇔
1 1
´π
V e jω = e jθ Y e j(ω−θ ) dθ

e jω ~Y e jω =

v[n] = x[n]y[n] ⇔ 2π X 2π −π X
DFT: v[n] = x[n] ~N y[n] , ∑N−1

r=0 x[r]y[(n − r) mod N ] ⇔ V [k] = X[k]Y [k]
N−1
v[n] = x[n]y[n] ⇔ V [k] = N1 X[k] ~N Y [k] , 1
N ∑r=0 X[r]Y [(k − r) mod N ]
Group Delay
F (nh[n])
jω )

The group delay of a filter, H(z), is τH (e jω ) = − d∠H(e =ℜ −z dH(z)
=ℜ where

dω H(z) dz
z=e jω F (h[n])
F ( ) denotes the DTFT.
Order Estimation for FIR Filters
Three increasingly sophisticated formulae for estimating the minimum order of an FIR filter with unity
gain passbands:
a
1. M ≈ 3.5∆ ω
a−8
2. M ≈ 2.2∆ ω
a−1.2−20 log10 b
3. M ≈ 4.6∆ ω
where a =stop band attenuation in dB, b = peak-to-peak passband ripple in dB and ∆ ω = width of
smallest transition band in normalized rad/s.

z-plane Transformations
A lowpass filter, H(z), with cutoff frequency ω0 may be transformed into the filter H(ẑ) as follows:
Target H(ẑ) Substitute Parameters

ω −ω̂
sin 0 2 1
ẑ−1 −λ
Lowpass z−1 = 1−λ ẑ−1
λ=
ω +ω̂

sin 0 2 1
ω̂ < ω̂1
ω +ω̂
−1 cos 0 2 1
Highpass z−1 = − 1+λ
ẑ +λ
ẑ−1
λ=
ω −ω̂

cos 0 2 1
ω̂ > ω̂1
ω̂ +ω̂
(ρ−1)−2λ ρ ẑ−1 +(ρ+1)ẑ−2 cos 2 2 1
ω̂2 −ω̂1

z−1 ω0

Bandpass = − (ρ+1)−2λ ρ ẑ−1 +(ρ−1)ẑ−2
λ=
ω̂ −ω̂
, ρ = cot 2 tan 2
cos 2 2 1
ω̂1 < ω̂ < ω̂2
ω̂ +ω̂
(1−ρ)−2λ ẑ−1 +(ρ+1)ẑ−2 cos 2 2 1
ω̂2 −ω̂1

z−1 = ω0

Bandstop (ρ+1)−2λ ẑ−1 +(1−ρ)ẑ−2
λ=
ω̂ −ω̂
, ρ = tan 2 tan 2
cos 2 2 1
ω̂1 ≮ ω̂ ≮ ω̂2

********* Questions and Solutions *********
1. a) i) Explain what is meant by saying that a linear time-invariant system

is “BIBO-stable”. [1]
An LTI system is BIBO-stable if any bounded input sequence x[n]

always results in an output sequence y[n] that is also bounded. A
sequence x[n] is bounded iff ∃B < ∞ such that |x[n]| < B ∀n.
Several people though that “bounded” meant ∑ x[n] < B instead of
x[n] < B ∀n. A few thought it meant finite energy: ∑ |x[n]|2 < B. Even
those that knew what bounded meant were often imprecise about
their definition, for example defining BIBO as “x[n] < ∞ ⇒ y[n] <
∞”. Others said that a BIBO-stable system implied that the input
was always bounded; no property of a system can ever impose a
condition on its input signal.
ii) The impulse response, h[n], of a linear time-invariant system satisfies

∑∞n=−∞ |h[n]| = S where S < ∞. Prove that the system is BIBO-stable
and also that H(z) converges for |z| = 1. [2]
Suppose that x[n] is any bounded sequence with |x[n]| < B < ∞ ∀n.
We need to show that the output, y[n] = ∑∞
r=−∞ h[r]x[n − r] is also
bounded. We have

∞
|y[n]| = ∑ h[r]x[n − r]

r=−∞
∞
≤ ∑ |h[r]x[n − r]|
r=−∞
∞
= ∑ |h[r]| |x[n − r]|
r=−∞
∞
< B ∑ |h[r]| = BS
r=−∞
Many people proved the converse of what was asked in the question:
i.e. BIBO ⇒ ∑∞ n=−∞ |h[n]| < ∞ . If you need to show something is
true for any input x[n], you cannot start by assuming one particular
x[n].
We have H(z) = ∑∞ −n
n=−∞ h[n]z . If |z| = 1 we can write

∞
|H(z)| = ∑ h[n]z−n

n=−∞
∞
|h[n]| z−n

≤ ∑
n=−∞
∞
= ∑ |h[n]| = S
n=−∞

Hence H(z) is convergent for |z| = 1.
Some wrote H(z) ≤ . . . instead of |H(z)| ≤ . . .; inequalities make no
sense for complex numbers. Quite a few people proved the result
for z = 1 and thought this was the same thing as proving it for |z| =
1 a few even assumed |z| = 1 ⇒ z−n = 1. Others just restated the
question, e.g. “Since ∑∞n=−∞ |h[n]| = S it follows that H(z) converges
for |z| = 1”; this does not constitute a proof.
Both of the proofs given above depend on the triangle inequality:
|x + y| ≤ |x| + |y| (also called the “subadditivity” property); some
people implicitly assumed that |x + y| = |x| + |y| (or, equivalently
|∑ h[n]| = ∑ |h[n]|) which is not true in general.
b) A filter, with input x[n] and output y[n], is defined by the difference equation
y[n] = αy[n − 1] + (1 − α)x[n]
where 0 < α < 1 is a real constant.
i) Determine the system function of the filter, H(z), and the impulse
response, h[n], for n = −1, 0, 1, 2. [2]
Taking z-transforms: Y (z) = αz−1Y (z) + (1 − α)X(z) from which

(z)
H(z) = YX(z) 1−α
= 1−αz −1 .
h[−1] = 0, h[0] = 1 − α, h[1] = (1 − α) α, h[2] = (1 − α) α 2 .

You can easily find the impulse response directly from the difference
equation by setting x[n] = δ [n]. Working out by taking the inverse
z-transform or IDTFT is much more laborious and error prone since
the inverse z-transform is not unique.
Surprisingly, very many people made the response non-causal, usu-
ally saying h[−1] = (1 − α)α −1 . You can tell from the difference
equation that h[n] is causal since the output y[n] depends only on
past or present inputs and outputs. In contrast, since the inverse z-
transform of H(z) is not unique, you cannot tell from H(z) whether
h[n] = (1 − α)α n u[n] (bounded, causal) or h[n] = (α − 1)α n u[−n −
1] (unbounded, anti-causal) unless you also known the region of con-
vergence.
ii) State the values of z at which H(z) has a pole or zero. [ 2 ]1
1−α (1−α)z
The system function is H(z) = 1−αz−1
= z−α . This has a pole at
z = α and a zero at z = 0.
The zero at z = 0 is easily overlooked. It is easiest to find all the
poles and zeros if you express the transfer function as a rational
polynomial in z rather than in z−1 . Even when H(z) is written in the
1−α
form H(z) = 1−αz −1 , it can easily be seen that H(0) = 0.

iii) Determine the frequency at which the filter has a gain of −3 dB.[ 3 ]
2
We want H(e jω ) = H(e jω ) × H ∗ (e jω ) = 0.5 . Thus

1−α 1−α
0.5 = − jω
×
1 − αe 1 − αe jω
(1 − α)2
=
1 − 2α cos ω + α 2
2 2 (1−α)2
From this, cosω = 1+α−2(1−α) 4α−1−α 2
2α =
2α = 1 − 2α and so
−1 4α−1−α 2 −1 (1−α)2
ω3dB = cos 2α = cos 1 − 2α .
Quite
a few √ people wrote down an equation with modulus signs:
H(e jω ) = 0.5 but did not known how to solve it. The stan-
dard procedure is to multiply both sides by the complex conjugate
as we do above. Some people just quietly ignored the modulus signs
instead and consequently got an imaginary √ value for ω. Several
thought −3 dB meant a gain of
√0.5 instead
of 0.5.
Several took the
jω j0
) = 0.5 × H(e ); Luckily for them,
question to mean H(e
3dB

H(e j0 ) = 1 in this case.
iv) If the sample frequency is fs , show that, for n ≥ 0, the impulse re-
t
sponse, h[n], is equal to a sampled version of g(t) = Ae− τ and deter-
mine the values of the constants A and τ. [2]
For n ≥ 0, we have from part i) that h[n] = (1 − α) α n . This can also

be proved by induction from the given recurrence relation assuming
that x[n] is an impulse at n = 0. If we now substitute t = fns (or,
equivalently t = nT where the sample period is T = f1s ) into g(t) =
n
Ae− τ we obtain h[n] = g fns = Ae− τ fs from which A = 1 − α and
t
1
α = e− τ fs . Rearranging the later equation gives τ = −1
fs ln α .
Surprisingly many people couldn’t do this. It really just involves
making the substitution t = nT = fns and then matching the coeffi-
cients of the resultant expressions. Some said t = n fs instead; this is
1
not dimensionally consistent since fs has dimensions time .
c) Figure 1.1 shows the block diagram of a filter implementation comprising two
delays, five multipliers with real-valued coefficients c1 , · · · , c5 and four adder
elements.
Y (z) c3 +c4 z−1 +c5 z−2
i) Show that transfer function X(z) = 1−c1 z−1 −c2 z−2
. [3]
The two delay elements form a shift register whose input is u[n].
Hence the inputs to the c1 and c2 multipliers are u[n − 1] and u[n − 2]
respectively. It follows directly from the diagram that u[n] = x[n] +
c1 u[n − 1] + c2 u[n − 2]. Taking z-transforms gives U(z) = X(z) +

U(z)
c1 z−1U(z) + c2 z−1U(z) from which X(z) = 1
1−c1 z−1 −c2 z−2
. From the
Y (z)
diagram y[n] = c3 u[n] + c4 u[n − 1] + c5 u[n − 2] from which U(z) =
−1 −2
c3 + c4 z + c5 z . Combining this with the previous result gives
Y (z) Y (z) U(z) c3 +c4 z−1 +c5 z−2
X(z) = U(z) × X(z) = 1−c z−1 −c z−2 .
1 2
Most got this right. A few had sign errors in the coefficients c1 and/or
c2 even thought the correct answer was given in the question. Some
just stated that it was a Direct Form II implementation of B(z)/A(z);
this does not constitute a proof.
ii) Suppose that each multiplier introduces independent additive white

noise at its output with power spectral density S(ω) = S0 and that
the noise signals are uncorrelated with x[n]. Show that the combined
effect of the five noise sources is equivalent to two additive white
noise signals at x[n] and y[n] respectively. Hence determine the over-
all power spectral density, N(ω), of the noise at y[n]. [3]
The noise components added by multipliers c1 and c2 merely add

onto the input and so add noise with power spectral density 2S0 onto
the input (note that, since the noise sources are assumed indepen-
dent, their power spectral densities add). Similarly, the remaining
noise sources add noise with power spectral density 3S0 onto the
output. Hence the overall noise power spectral density at y[n] is
2
N(ω) = 3S0 + 2S0 × H(e jω )

!
c3 + c4 z−1 + c5 z−2 c3 + c4 z1 + c5 z2

= S0 3 + 2
(1 − c1 z−1 − c2 z−2 ) (1 − c1 z1 − c2 z2 )
c23 + c24 + c25 + 2 (c3 + c5 ) c4 cos ω + 2c3 c5 cos 2ω

= S0 3 + 2 .
1 + c21 + c22 + 2 (c2 − 1) c1 cos ω − 2c2 cos 2ω
Note that the expression in the first line gets full marks.
Some mixed up the time domain and z-transform domain: you never
get x[n] and z in the same equation. Some just assumed the input
noise was multiplied by c3 instead of by H(e jω ) and most people
forgot that the noise power spectrum was multiplied by the gain
squared.
Figure 1.1

d) The impulse response of an antisymmetric FIR filter, H(z), of order M satisfies
the relation h[n] = −h[M − n].

i) Show that the magnitude response H(e jω ) can be expressed as the
absolute value of the sum of N sine waves where N = M2 if M is even
and N = M+12 if M is odd. [3]
If M is odd, there is an even number of coefficients so we can write

M−1
2 M
H(z) = ∑ h[n]z−n + ∑ h[n]z−n
n=0 n= M+1
2
M−1 M−1
2 2
−n
= ∑ h[n]z + ∑ h[M − r]z−(M−r)
n=0 r=0
M−1
2
z−n − zn−M

= ∑ h[n]
n=0
M−1
2
−0.5M
z−n+0.5M − zn−0.5M

= z ∑ h[n]
n=0
M−1
2
jω − j0.5Mω
H(e ) = − je ∑ 2h[n] sin ((n − 0.5M) ω)
n=0

M−1
2

H(e jω ) = ∑ 2h[n] sin ((n − 0.5M) ω)

n=0
M+1
The right side is the absolute value of the sum of 2 sine waves as
required.
If M is even, there is an odd number of coefficients but the central
coefficient, h[0.5M], must be zero since h[0.5M] = −h[M − 0.5M] =
−h[0.5M]. Hence
M
2 −1 M

−n M −0.5M
H(z) = ∑ h[n]z +h z + ∑ h[n]z−n
n=0 2 n= M +1 2
M−2
2
z−n − zn−M

= 0+ ∑ h[n]
n=0
M−2
2

jω

⇒ H(e ) = ∑ 2h[n] sin ((n − 0.5M) ω)

n=0
The derivation is identical to that for odd M except for the upper
summation limit. The right side is the absolute value of the sum of
M
2 sine waves as required.
The coefficients, h[n] go from n = 0 to n = M and so if M is odd,
there is an even number of coefficients. Many got the summation
limits slightly wrong, sometimes because they assumed there were M
coefficients rather than M + 1 and sometimes because they assumed

the upper summation limit to be 0.5M. Quite often the summation
limits were not even integers. An easy way to check that the limits
are correct is to try a very small value for M such as M = 1 or M =
2.Several did not notice that if M was even, h[0.5M] must equal zero.
for the even M case, a few people extracted the h[0] term from the
summation rather than the h[0.5M] term; this is incorrect because
h[0] = −h[M] is a symmetric pair. Not everyone seemed completely
familiar with e jθ − e− jθ = 2 j sin θ which can easily be derived from
e jθ = cos θ + j sin θ .
ii) Show that H(e jω ) is necessarily zero at ω = 0 but may be non-zero

at ω = π if M is odd. Give an example of a filter for which this is
the case. [2]
When ω = 0 then sin αω = 0 for any α. Hence all the summation

terms are zero for both odd and even M.
When ω = π then sin αω = 0 if α is an integer. So, if M is even,
sin ((n − 0.5M) ω) will always be 0 at ω = π. However, if M is odd
then this is not necessarily true.
An example, for M = 1, is H(z) = 1 − z−1 . For this case, H(e jω ) =
1 − e− jω = 2 je−0.5 jω sin 0.5ω. When ω = π, H(e jω ) = 2 which is
non-zero.
Most people correctly explained the case of ω = 0. However the
arguments for ω = π were often not precisely correct and, in deed,
many people omitted this part entirely.
iii) Derive an expression for the phase response, ∠H(e jω ), and deter-
jω )
mine the group delay, τH e jω = − d∠H(e

dω . [2]
From part d), the phase response is

  M−1  
π 2
∠H(e jω ) = ∠ − je− j0.5Mω + sgn  ∑ 2h[n] sin ((n − 0.5M) ω) − 1
2 n=0
  M−1  
2
π
= −0.5 (Mω + π) + sgn  ∑ 2h[n] sin ((n − 0.5M) ω) − 1
2 n=0
jω )
Differentiating this gives τH e jω = − d∠H(e

dω = 0.5M. Note that
the sgn() function is piecewise constant and so its derivative is zero.
Almost everyone omitted the π2 (sgn( ) − 1) term completely and many
omitted the −0.5π term arising from the factor − j. Somejωtried
jω

to derive τH e from the formula sheet expression τH e =
−z dH(z)
ℜ H(z) dz jω which, although correct, is not a good approach.
z=e

e) Figure 1.2 shows the analysis and synthesis sections of a subband processing
system. The input and output signals are x[n] and y[n] respectively and the
intermediate signals are vm [n], um [r] and wm [n] where m = 0 or 1 according to
the subband. The corresponding z-transforms are X(z), Y (z) etc.
i) Show that it is possible to express
the overall
transfer function in the
X(z)
form Y (z) = T (z) A(z) and determine expressions
X(−z)
for T (z) and A(z).
You may assume without proof that for m = 0 or 1, [3]
1 n 1 1 o
Um (z) = Vm z 2 +Vm −z 2
2
Wm (z) = Um z2 .

Combining the two given equations we get

1
Wm (z) = Um z2

= {Vm (z) +Vm (−z)}
2
1
= = {Hm (z)X (z) + Hm (−z)X (−z)} .
2
Hence
Y (z) = G0 (z)W0 (z) + G1 (z)W1 (z)

1
= {(G0 (z)H0 (z) + G1 (z)H1 (z)) X (z) + (G0 (z)H0 (−z) + G1 (z)H1 (−z)) X (−z)}
2
X(z)
= T (z) A(z)
X(−z)
where T (z) = 21 (G0 (z)H0 (z) + G1 (z)H1 (z)) and A(z) = 21 (G0 (z)H0 (−z) + G1 (z)H1 (−z)).
Most people got this right.
ii) Explain why it is normally desirable to have A(z) ≡ 0. [2]
The factor A(z) multiplies X(−z). The spectrum of X(−z) is X(−e jω ) =

X(e j(ω+π) ) is an aliased version of the spectrum X(e jω ) in which the
spectrum is shifted by π (or equivalently conjugated and reflected
around π2 for the case of a conjugate symmetric spectrum). These
aliased spectral images are normally unwanted and so we would
like A(e jω ) ≡ 0 ⇔ A(z) ≡ 0 (assuming A(z) is analytic).
Most people got this right although few described the relationship
between X(−z) and X(z) beyond saying that it was “aliased” and
therefore undesirable.
iii) Suppose that H0 (z) = H1 (−z) = G0 (z) = −G1 (−z). Show that in
this case A(z) = 0 and explain how the frequency responses H1 (e jω ),

G0 (e jω ) and G1 (e jω ) are related to H0 (e jω ) assuming that H0 (z) is
an FIR or IIR filter with real coefficients. [2]
From the relations given, we can express all the blocks in terms of
H0 (z) to obtain
A(z) = G0 (z)H0 (−z)+G1 (z)H1 (−z) = H0 (z)H0 (−z)−H0 (−z)H0 (z) =
0.
Our no-alias condition is therefore met.
Some people correctly derived A(z) = G0 (z)H0 (−z) − G0 (−z)H0 (z)
but went on to say that this must be zero because, they wrongly as-
serted, G0 (−z) = G0 (z) and H0 (z) = H0 (−z). This is the wrong
pairing: in fact G0 (−z) = H0 (−z) and H0 (z) = G0 (z).
Clearly G0 (e jω ) is identical to H0 (e jω ).
H1 (e jω ) = H0 (−e jω ) = H0 (e j(ω−π) ) = H0∗ (e j(π−ω) ).
H1 (e jω ) is therefore the frequency response of H0 (e jω ) but conju-
gated and reflected around ω = π2 .
G1 (e jω ) is the same as H1 (e jω ) but negated or, equivalently, with π
added onto the phase response.
Some people only discussed the magnitude responses.
Figure 1.2

f) Figure 1.3 shows an upsampler with real-valued input x[n] and output
(
x Kr if K | r
y[r] =
0 otherwise
where K | r means K is a factor of r.

Show that Y (z) = X zK .

i) [1]
−r −r −nK = ∞
−n
Y (z) = ∑∞ ∞
r=−∞ y[r]z = ∑{r:K|r} y[r]z = ∑n=−∞ y[nK]z ∑n=−∞ x[n] zK =
X zK .

In order to prove this you need to change the summation variable

using, in the expression above, r = nK. “Proofs” that did not make
such a substitution were usually not correct. Many people omit-
ted the stage ∑{r:K|r} y[r]z−r in the above proof but instead wrote
r −r
∑∞r=−∞ x[ K ]z which is invalid/meaningless when Kr is not an inte-
ger.
ii) The energy and average power of x[n] are defined respectively as
∞
Ex = ∑ |x[n]|2
n=−∞
N
1
Px = lim
N→∞ 2N + 1
∑ |x[n]|2 .
n=−N
Give expressions for the energy and average power of y[r] in terms
of Ex and Px . [2]
The non-zero samples of y[n] are identical to the samples of x[n] and
so the energy of the two signals is the same: EY = EX . However y[n]
has K times as many samples, so its power is PY = K1 PX .
To show this algebraically (rather laborious), we write
N
1
Px , lim
N→∞ 2N + 1
∑ |x[n]|2
n=−N
N
1
= lim
N→∞ 2N + 1
∑ |y[Kn]|2
n=−N
KN
2KN + 1 1
= lim ∑ |y[m]|2
N→∞ 2N + 1 2KN + 1 m=−KN
!
KN
K −1 1 2
= lim K +
N→∞ 2N + 1
× lim
N→∞ 2KN + 1
∑ |y[m]|
m=−KN
= K × Py .
Most people correctly said Ey = Ex but quite a few also said Py =

Px . Many people got this question wrong because they wrote down

complicated algebraic expressions which they were then unable to
evaluate.
iii) Figure 1.4 shows the power spectral density of x[n] which comprises
white noise of unit magnitude together with a bandpass signal com-
ponent occupying the range 0.5 < |ω| < 1. Sketch the power spectral
density of y[r] when K = 3 and give the magnitude of its white noise
component and the magnitude and frequency range of all bandpass
components. [3]
The upsampling horizontally compresses the 2-sided spectrum by a

factor of K and replicates it K times at increments of ∆ ω = 2π
K . For
the DTFT and the energy spectrum, each image is the same magni-
tude as the original. However, for the power spectrum, the magni-
tudes will be reduced by a factor of K.
1.5
0.5
0
-3 -2 -1 0 1 2 3
Thus the magnitude of the white component will be 0.33 and that of
the bandpass components will be 2.
The frequency ranges of the bandpass components will be
2π
− 3 , 0, 2π
1 1 2π 2π
1 1
3 ± 4 ± 12 = − 3 , 0, 3 ± 6 , 3 .
That is, each will be of width 16 and centered ± 14 either side of the
image centre. The band edges are therefore at
ω = ± {0.167, 0.333, 1.761, 1.928, 2.261, 2.428}.
Most people got the overall spectrum shape correct although a few
expanded the spectrum horizontally by a factor of 3 instead of com-
pressing it. Often the precise limits of the bandpass components
were incorrectly calculated. Some made the bandpass regions even
spaced apart. Very few reduced the magnitude by a factor of 3 even
those that correctly said that Py = K −1 Px in the previous part.
iv) The diagram of Fig. 1.3 is followed by a lowpass filter to remove

spectral images. If K = 3 and x[n] is as specified in part iii) above,
determine the transition bandwidth and transition band centre fre-
quency of a suitable lowpass filter and explain the reasons for your
choices.
[2]
The required filter order is inversely proportional to the transition

width, ∆ ω which we therefore wish to make as wide as possible.
We therefore make the transition width from the edge of the wanted

signal component to the start of the first image component: 13 =
0.333 to 2π−1
3 = 1.761 which makes the total width equal to 3 =
2π−2
1.428. The centre of the transition band is the average of these two
values and corresponds to the old Nyquist frequency, i.e. π3 = 1.047.

2
1.5
0.5
0
-3 -2 -1 0 1 2 3
Mostly OK.
Power Spectral Density 4
0
-3 -2 -1 0 1 2 3

2. a) Suppose that G1 (z) = 1 − pz−1 and G2 (z) = 1 − qz−1 where
the constants p and
1 jω jω

q may be complex. If q = p∗ show that G1 (e ) = α G2 (e ) for all ω and

determine an expression for the constant α. [4]
For z = e jω , we have z∗ = z−1 and so we can write

|G2 (z)| = 1 − (p∗ )−1 z−1

|G2 (z)| = |G2 (z)∗ | = 1 − p−1 z

= p−1 z pz−1 − 1

= |p|−1 |G1 (z)|
Thus the ratio |G1 (z)|

|G2 (z)| = |p| is independent of ω and α = |p|. Note the second
line is the complex conjugate of the first and conjugation does not affect the
magnitude of a complex number.
One of several alternative derivations is
|G1 (z)|2 = G1 (z)G∗1 (z) = 1 − pz−1 (1 − p∗ z)

= 1 − pz−1 − p∗ z + pp∗
= 1 − (q∗ )−1 z−1 − q−1 z + (qq∗ )−1
= (qq∗ )−1 qq∗ − qz−1 − q∗ z + 1

= (qq∗ )−1 1 − qz−1 (1 − q∗ z)

= |q|−2 |G2 (z)|
q
from which α = |q|−2 = |q|−1 = |p|.
The question proved much harder than I expected; in part this was because of
numerous algebraic errors involving complex numbers. Many people had diffi-
culty with the algebra because of the modulus signs; some people just omitted
the modulus signs completely which gives easy algebra but the wrong answer.
Equations containing modulus signs are fine if you restrict yourself to multi-
plication and conjugation (as in the first method above) since |xy| = |x| |y| and
|x∗ | = |x|. Alternatively, as in the second method, it can be a good idea to
eliminate modulus signs from algebraic equations by squaring them (i.e. mul-
tiplying by the
p complex conjugate). This is usually a much better idea than
using |x| = ℜ(x)2 + ℑ(x)2 which is inevitably very messy. Some people de-
composed z = e jω = cos ω + j sin ω and/or decomposed p = a + jb which led
to lots of algebra and occasionally the right answer. Several implicitly assumed
q
−1 = (1 − p cos ω)2 + p2 sin2 ω.

p was real by writing 1 − pz

−1 −2
b) H1jω(z) = 4 + 14zjω − 8z . Determine the coefficients of H2 (z)
Suppose that
such that H1 (e ) = H2 (e ) for all ω and that all the zeros of H2 (z) lie

inside the unit circle. [4]
z−1 or else
To find the roots of H1 (z) you can either treat it as a polynomial in √
142 +128
treat z2 H(z) as a polynomial in z. These alternatives give z−1 = −14± −16 =
√
−14±18 −14± 142 +128 −14±18
−16 = {2, −0.25} or else z = 8 = 8 = {0.5, −0.4}. The

roots of H1 (z) = 4 1 − 0.5z−1 1 + 4z−1 are z = {0.5, −4} and so the roots

−1
of
1 + 0.25z−1 .

H2 (z) must be z = {0.5, −0.25} which implies that H2 (z) = α 1 − 0.5z
To determine alpha, we substitute a suitable value of z that lies on the unit
circle, e.g. z = 1, to obtain |H2 (1)| = |α (1 − 0.5) (1 + 0.25)| = 0.625 |α| =
10
|H1 (1)| = |4 + 14 − 8| = 10. From this we get α = ± 0.625 = ±16 and so
−1
H2 (z) = 16 1 − 0.25z − 0.125z −2 = 16−4z −2z or −16+4z−1 +2z−2 .
−1 −2
We can check the scaling by evaluating at z = 1: H1 (1) = 4 + 14 − 8 = 10 and

H2 (1) = 16 − 4 − 2 = 10. These have the same magnitude so all is well.
Several people just reversed all the coefficients (as in an allpass filter) to get
H2 (z) = −8 + 14z−1 + 4z−2 . This inverts all the roots including the ones that
were inside the unit circle to start with. Thus the roots of this polynomial are 2
and −0.25.
jω )
c) When designing an IIR filter H(e jω ) = B(e
A(e jω )
to approximate a complex target
response D(ω) two error measures that may be used are the weighted solution
error, ES (ω), and the weighted equation error, EE (ω), defined respectively by
B(e jω )

ES (ω) = WS (ω) − D(ω)
A(e jω )
EE (ω) = WE (ω) B(e jω ) − D(ω)A(e jω ) .

Explain the relative advantages of the two error measures and explain the pur-
pose of the real-valued non-negative weighting functions WS (ω) and WE (ω).
[2]
The solution error is a direct measure of the error in the frequency response
but has the disadvantage that it gives rise to a set of non-linear simultane-
ous equations which do not have a closed form solution. Although the equa-
tion error gives rise to a set of linear equations which are straightforward to
solve, it multiplies the true frequency response errors by A(e jω ) and so, un-
less WE (ω)
isjωadjusted
accordingly, gives a higher weight to spectral regions
in which A(e ) is large. Although the two errors are equivalent if there hap-

pens to be a solution that makes them zero, minimizing the equation error will
not generally give the same solution as minimizing the solution error.
The weighting functions, WS (ω) and WE (ω) are used to control the relative
importance of errors in different parts of the spectrum. A high weight will
result in a lower error.
Most understood the benefit of the equation error over the solution error. How-
ever, many people did not understand that the weighting functions allow you
to decide that you would like higher accuracy in one part of the spectrum than
another. For example, you might well want lower errors in the stop band than
in the pass band.
d) Suppose that 0 ≤ ω1 < ω2 < . . . < ωK ≤ π is a set of K frequencies and that

A(z) = 1 + [z−1 z−2 · · · z−N ]a and B(z) = [1 z−1 z−2 · · · z−M ]b where a and b
are real-valued coefficient column vectors.
i) Show that it is possible to express the equations EE (ωk ) = 0 for
1 ≤ k ≤ K as a set of K simultaneous linear equations in the form


a
(P Q) = d.
b
State the dimensions of the matrices P and Q and of the vector d and
derive expressions for the elements of P, Q and d. [4]
If EE (ωk ) = 0 then WE (ωk ) B(e jωk ) − D(ωk )A(e jωk ) = 0. Substi-

tuting in the expressions for A(z) and B(z) gives
WE (ωk ) [1 e− jωk e− j2ωk · · · e− jMωk ]b

−D(ωk ) − D(ωk )[e− jωk e− j2ωk · · · e− jNωk ]a = 0.

a
pTk qTk

Rearranging this equation gives = WE (ωk )D(ωk )
b
where pTk = −WE (ωk )D(ωk )[e− jωk e− j2ωk · · · e− jNωk ]
and qTk = WE (ωk )[1 e− jωk e− j2ωk · · · e− jMωk ]. Note that there are
M + 1 coefficients in b but only N coefficients in a because the first
coefficient of A(z) is always equal to 1.
Thus the dimensions of P, Q and d are K × N, K × (M + 1) and
K × 1. and their kth rows are pTk , qTk and WE (ωk )D(ωk ) respectively.
Note that finding a least squares solution to an equation is different
from finding an exact solution. Thus if f (x)g(x) = 0 has an exact
solution, it must satisfy f (x) = 0 or g(x) = 0. However, if the original
equation has no exact solution, its least-squares solution is not the
same as the least squares solution to either f(x)=0 or g(x)=0. Thus
jω jω
in the equation WE (ωk ) B(e ) − D(ωk )A(e ) = 0, you cannot
k k
just ignore the WE (ωk ) term even if it is non-zero (several people did
this). Many people got the dimensions wrong; some used M instead
of M + 1 and others said the number of columns of P and/or Q was
max (M, N) or something similar. The total number of columns of
of P and Q must equal the number of unknowns (N + M + 1) and
the number of rows of P, Q and d must be equal to the number of
equations (K). Some assumed k = 0, · · · , K even though the question
says k = 1, · · · , K.
ii) Explain how, by separating the real and imaginary parts of P, Q and
d,
it ispossible to obtain a set of simultaneous linear equations for
a
in which all coefficients are real-valued. Explain the circum-
b
stances under which some of the resultant equations will necessarily
have all-zero coefficients. [4]
By treating the real and imaginary parts ofeach

row astwo separate
ℜ (P) ℜ (Q) a ℜ (WE (ωk )D(ωk ))
equations, we can write =
ℑ (P) ℑ (Q) b ℑ (WE (ωk )D(ωk ))
in which all coefficients are real.
If WE (ωk ) = 0 for any particular k, then the corresponding pair of
equations will have all-zero coefficients. In addition, if ωk = 0 or π

then pTk and qTk will both be real-valued if, as is likely, D(ωk ) is real,
so the equation formed from their imaginary parts will have all-zero
coefficients.
Not everyone realized that this means that we will now have 2K
equations (which is good: the more equations the better when you
are trying to minimize errors). Some people suggested using only
the real parts of the equations, but this is a bad idea as it leaves the
imaginary parts unconstrained. Only a few people understood why
the imaginary parts might be structurally zero.
iii) Explain why it may be desirable to apply the transformation of part

b) after obtaining the solution to the equations of part d) ii). [2]
To obtain a stable filter, the zeros of A(z) must lie inside the unit cir-
cle. There is no guarantee that the solutions of the equations will
meet this requirement so the transformation of part b) may be ap-
plied to A(z) to enforce stability. This will change the phase response
but not the magnitude response of the resultant filter.
Some suggested that the transformation should also be applied to
B(z) in order to make it minimum phase. This is a possible goal but
is not nearly as important as ensuring a stable filter.
iv) Assuming that ω1 = 0 and ωK = π, determine the minimum value of

K to ensure that the equations of part d) ii) are not underdetermined.
[4]
From part (ii) we get 2K real-valued equations. However, since pTk

and qTk are real for k = 1 and k = K, we have a total of 2K − 2
equations with M + N + 1 unknowns.
Therefore we require 2K − 2 ≥ M + N + 1 which implies that K ≥
M+N+3
2 .
Several said there were only max (M, N) unknowns but without any
explanation.
b

e) Suppose now that H(z) = that K = 3, that ωk = 0, π2 , π , that
1+az−1
,
(
2 for ω ≤ 0.25π
D(ω) = .
1 for ω > 0.25π
WE (ω) ≡ 1
Determine the numerical values of the elements of P, Q and d and hence deter-
mine the numerical values of a and b that minimize ∑k |EE (ωk )|2 . [6]
You may assume without proof that the least squares solution to an overdeter-
−1 T
mined set of real-valued linear equations, Rx = q, is given by x = RT R R q
assuming that R has full column rank.

Since K = 3 the values of ωk are {0 0.5π π} and hence e− jωk = {1 − j − 1}.
 k ) = {2
At these frequencies D(ω 1 1}. The complex
 equations from part d)i)
−2 1 2
a
are therefore given by  j 1  =  1  where the vectors a and
b
1 1 1
   
−2 1
b each contain only a single element. Thus P =  j  ,Q=  1  and
1 1
 
2
d =  1 .
1
From
 this, applying
 the 
technique
 of d) ii), we get a set of four real equations
−2 1 2
 0 1  a  1 
 b =  1 .
   
 1 1
1 0 0
Hence, using the formula given in the question, we can write
  −1  
−2 1 2
a  −2 0 1 1  0 1 
   −2 0 1 1 
 1 

= 
b  1 1 1 0  1 1  1 1 1 0  1 
1 0 0
−1
6 −1 −3
=
−1 3 4

1 3 1 −3 1 −5
= =
17 1 6 4 17 21
Where, in the last line, we have used the formula for the inverse of a 2 × 2
−1
p q 1 s −q
matrix , = ps−rq .
r s −r p
From this a = −0.294 and b = 1.235.
Equivalently, you can just solve the simultaneous equations 6a − b = 3 and
−a + 3b = 4.
Surprisingly, many people did not convert the 3 complex-valued equations into
4 real-valued equations despite having given a correct answer to part (ii). As a
result, they ended up with complex values for a and b.

3. a) Figure 3.1 shows the block diagram of a system that multiplies the input sample
rate by QP where P and Q are coprime with P < Q .
i) Explain why the cutoff frequency of the lowpass filter H(z) should
be placed at the Nyquist rate of the output signal, y[m] and give the
normalized cutoff frequency, ω0 , in rad/sample in terms of P and/or
Q.
a
Using the approximation formula M ≈ 3.5∆ ω , determine the required
filter order M in terms of P and/or Q if the stopband attenuation in dB
is a = 60 and the normalized transition bandwidth is ∆ ω = 0.1ω0 .
[4]
The lowpass filter must eliminate the images introduced by the up-
sampler and the alias components introduced by the down sampler
and must therefore eliminate all frequencies above the lower of the
input and output Nyquist frequencies. Since Q > P,the output Nyquist
frequency is Qπ . The normalized cutoff frequency is therefore ω0 =
π π
max(P, Q) = Q .
60 60Q
We have M = 3.5×0.1ω0 = 3.5×0.1π = 54.6Q.
Most correctly said that the filter cutoff should be the lower of the two
Nyquist frequencies but several people calculated its value incor-
rectly. Some gave the Nyquist frequency as 2π Q , some used min(P, Q)
instead of max(P, Q), and some normalized ω0 by the input sample
rate of x[n] rather than by the sample rate of v[r] to get πP
Q . The cutoff
frequency of a filter is always normalized by the sampling frequency
at the input and output of the filter itself. Some left the answer as
π
ω0 = max(P, Q) even though the question explicitly said that Q > P;
this makes the answers to the remaining parts of the question much
messier.
ii) Using the value of M from part a) i), estimate the average number
of multiplications per input sample, x[n], needed to implement the
system in the form of Figure 3.1. [2]
The filter requires M + 1 multiplications per filter output sample,

v[r], which equals (M + 1) P per input sample, x[n] (since there are
P times as many samples of v[r] as there are of x[n]). Substituting
M = 54.6Q gives (54.6Q + 1) P ≈ 54.6PQ multiplications per input
sample.
Many people did not multiply by P and gave the answer as (M + 1).
A few said (M+1) P instead of (M + 1) × P. In the diagram of Figure
3.1, the lowpass filter operates at P times times the sample rate of the
input x[n] and requires M + 1 multiplications for each value of v[r]
even though most of the samples at the input of the filter are zero;
it is this inefficiency that allows polyphase decomposition to gain an
advantage. Some people included the number of additions as well
(giving 2M + 1 instead of M + 1 per v[r]) even though the question
explicitly asked only for multiplications.

iii) The filter H(z) has a symmetrical impulse response h[r] = g[r]w[r]
for 0 ≤ r ≤ M where g[r] is the impulse response of an ideal lowpass
filter with cutoff frequency ω0 and w[r] is a symmetrical window
function.
Derive an expression for the ideal response, g[r], in terms of ω0 , M
and r. [4]
The ideal response (centered on r = 0) is G(e jω ) = 1 for |ω| < ω0

and zero otherwise. To this ideal response, we need to add a delay
of M2 samples which corresponds to a phase shift of e− j0.5Mω (or,
equivalently, we can design a centred filter and then delay the coeffi-
cients by 0.5M samples to make it causal). Using the inverse DTFT
(available in the formula sheet)
ˆ π
1
g[r] = H(e jω )e− j0.5Mω e jωr dω
2π −π
ˆ ω0
1 1 h iω0
= e jω(r−0.5M) dω = e jω(r−0.5M)
2π −ω0 j2 (r − 0.5M) π −ω0
1 sin ((r − 0.5M) ω0 )
= × 2 j sin ((r − 0.5M) ω0 ) =
j2 (r − 0.5M) π π (r − 0.5M)
We can also write this as g[r] = ωπ0 sinc ((r − 0.5M) ω0 ) where sinc(x) =
sin(x)
x .
Many people omitted the offset of 0.5M. Instead of calculating it ex-
plicitly, some just stated g[r] = ω0 sinc(rω0 ) which is correct except
for the offset and the factor of π1 .
(
1 |r| < ω0
A few people said that g[r] = which mixes up the
0 otherwise
time and frequency domains.
´ jωr A few wrote the correct
´ integral but
were unable to integrate e dω; converting it to (cos ωr + j sin ωr) dω
makes the integral slightly harder if anything. Quite a few people
tried to calculate H(e jω ) instead of g[r] which was what the ques-
tion asked for.
b) The filter H(z) is now implemented as a polyphase filter as shown in Fig. 3.2.
The filter implementation uses a single set of delays and multipliers with com-
mutated coefficients.
i) State the length of the filter impulse response h0 [n] in terms of M, P
and/or Q and give an expression for the coefficients h0 [n] in terms of
h[r]. [2]
The length of the filter h0 [n] is M+1

P rounded up to the nearest integer
(the order of h0 [n] is one less than this). The coefficients are given
by h0 [n] = h[nP].
M+1 M+1
Some said Q instead of P . From the diagram, the sub-filters

are labelled h0 [n], · · · , hP−1 [n] and so, since there are P of them and
the total number of coefficients stays the same, each one must be of
length M+1 M
P . A few said P because they forgot that an FIR filer of
order M has M + 1 coefficients.
ii) If x[n] = 0 for n < 0, give expressions for v[0], v[1], v[2P+1] in terms
of the input x[n] and the coefficients h p [n]. [2]
Since x[n] is causal, v[0] = h0 [0]x[0], v[1] = h1 [0]x[0] and v[2P + 1] =

h1 [0]x[2] + h1 [1]x[1] + h1 [2]x[0].
For each input sample, n, there are P samples of v[r], namely
{v[Pn + 0], v[Pn + 1], · · · , v[Pn + P − 1},
one coming from each of the filters {h0 [n], h1 [n], · · · , hP−1 [n]}. If
we just consider one of the sub-filters, the outputs from filter h0 [n]
are v[Pn] = ∑Ss=0 h0 [s]x[n − s] and, in general, the outputs from filter
h p [n] are v[Pn + p] = ∑Ss=0 h p [s]x[n − s] where S = M+1

P − 1 is the
order of the sub-filters.
Because x[n] is causal, v[Pn + p], has only n + 1 non-zero terms. So,
for example, v[0], · · · , v[P − 1] all depend only on x[0] and
v[P], · · · , v[2P − 1] depend on x[0] and x[1] only.
Some said v[1] = h1 [0]x[1] or else v[1] = h1 [0]x[1] + h1 [1]x[0], the
second expression is actually the formula for v[P + 1]. A few wrote
things like v[0] = h0 [n]x[n]; an expression like this makes no sense
since “n” appears only on the right hand side.
iii) Explain how it is possible to eliminate the output decimator by chang-

ing both the sequence and rate at which the coefficient sets, h p [n] are
accessed.
Determine the new coefficient set order for the case P = 5 and Q = 7.
[3]
The output decimator selects every Qth sample of v[r] and discards
the others. Therefore if we access the coefficient sets in the order
p = (mQ) mod P for m = 0, 1, · · · and reduce the rate by a factor
of Q will will generate only the wanted output samples.
For the specific values P = 5 and Q = 7, so the coefficient set se-
quence becomes
h0 [n], h7 mod 5 [n], h14 mod 5 [n], h21 mod 5 [n], h28 mod 5 [n]
which equals h0 [n], h2 [n], h4 [n], h1 [n], h3 [n].

Since Q mod P = 2 the value of p increments each time either by 2
or by 2 − P = −3
Quite well done on the whole although not everyone got the correct
sequence. The original question used “order” instead of “sequence”
which several people interpreted as the order (i.e. length) of the

subfilters. Quite a few people suggested applying the Noble identity
to swap H p (z) and Q : 1 but this is only possible if the filter is Hp (zQ )
which is not the case.
iv) Determine the number of multiplications per input sample for the
system of part b) iii) and the number of distinct coefficients that must
be stored. You may assume that M + 1 is a multiple of P. [2]
We require M+1 P multiplications per output sample, y[m], and there-

fore M+1P × P M+1
Q = Q multiplications per input sample, x[n]. Because
of symmetry, the number of distinct coefficients is only M+1
2 although
it is not necessarily easy to take advantage of this symmetry to reduce
the storage requirements.
Although not requested in the question, it is interesting to note that
since M ≈ 54.6Q, the number of multiplications per input sample is
approximately 54.6 independently of P or Q.
Although most got the number of multiplications per output sample
correct (at M+1
P ), not everyone multiplied by the correct factor to get
the number of multiplications per input sample. Since the overall
system reduces the sample rate, there must be fewer multiplications
per input sample than per output sample; hence we multiply by QP
which is < 1.
c) Suppose now that the sample rate of the input, x[n], is 18 kHz and that the
system is implemented as in part b) iii) with the values of a and ∆ ω as given in
part a) i).
Determine the values of P, Q and M when the sample rate of the output, y[m],
is (i) 10 kHz and (ii) 10.1 kHz [note that 101 is a prime number].
For each of these cases estimate the number of multiplications per input sample
and the number of distinct coefficients that must be stored. [5]
(i) For an output sample rate of 10 kHz = 95 × 18 kHz, P = 5 and Q = 9.

M = 54.6Q = 491. The number of multiplications per input sample is there-
fore M+1 M+1
Q = 54.7. The total number of distinct coefficients is 2 = 246.
(ii) For an output sample rate of 10.1 kHz = 101

180 × 18 kHz, P = 101 and Q =
180. M = 54.6Q = 9822. The number of multiplications per input sample
is therefore M+1Q = 54.6 (virtually unchanged). The total number of distinct
coefficients is M+1
2 = 4912 (much increased).
A small number of people did not know that “coprime” means “having no
common factors” and therefore used P = 10 and Q = 18 for the first example.

d) In a Farrow filter, the coefficients, h p [n], are approximated by a low-order poly-
nomial fn (t) where t = Pp for 0 ≤ p ≤ P − 1.
i) Assuming that a rectangular window, w[r] ≡ 1, is used in the design
of H(z) and that ω0 = πP , give an expression for the target value of
f0 (t) in terms of t, M , P and Q. [3]
From the answer to part a) iii) h[r] = g[r] = sin(r−0.5M)ω 0

π(r−0.5M) so h p [0] =
0)
g[p] = sin((p−0.5M)ω = f0 Pp . We now substitute p → Pt and ω0 =

π(p−0.5M)
π sin((Pt−0.5M) Q
π
)
Q to get f 0 (t) = g[Pt] = π(Pt−0.5M) for 0 ≤ t ≤ 1.
Although not requested in the question, a more general formula is
0)
h p [n] = g[Pn + p] = sin((Pn+p−0.5M)ω = fn Pp from which fn (t) =

π(Pn+p−0.5M)
sin((P(t+n)−0.5M) π )
g[P (n + t)] = π(P(t+n)−0.5M)Q for 0 ≤ t ≤ 1 and 0 ≤ n ≤ M+1

P −1 .
Very few attempted this part.
ii) If the polynomials, fn (v), are of order K = 5, determine the number

of coefficients that must be stored for each of the cases defined in
part c).
[3]
The M+1
P polynomials f n (t) each require K = 6 coefficients, so we
require a total of 6(M+1)
P coefficients. For the two cases, this gives
(i) 5 = 590 (somewhat larger than before) and (ii) 6×9823
6×492
101 = 584
(much less than before).
Very few attempted this part.

4. A complex-valued frequency-modulated signal, x(t) = a(t)e jφ (t) , has a 0 Hz carrier fre-
quency and a peak frequency deviation of d = 75 kHz. The´ t amplitude, a(t), is approxi-
mately constant with a(t) ≈ 1 and the phase is φ (t) = k 0 m(τ)dτ where k is a constant
and m(t) is a baseband audio signal with bandwidth b = 15 kHz. The signal x(t) is
sampled at 400 kHz to obtain the discrete-time signal x[n].
a) Carson’s rule for the bandwidth of a double-sideband FM signal is B = 2 (d + b).
Use this to determine the single-sided bandwidth, ω0 , of x[n] in radians/sample.
[2]
From Carson’s rule, B = 180 kHz. This bandwidth includes both sidebands, so
90 kHz
ω0 = 2π × 400 kHz = 0.45π = 1.41 rad/samp.
Mostly correct. A few people did not divide by 400 kHz which gave an answer
of 5.7 × 105 rad/samp; except for very rare circumstances, frequencies in units
of rad/samp always lie in the range ±π.

b) Show that m(t) = k−1 a−2 (t)ℑ x∗ (t) dx(t)
dt where ℑ ( ) denotes the imaginary
part. [4]
From the definition of φ (t), m(t) = k−1 dφ dt . Differentiating x(t) gives dt =

dx
da jφ (t)
+ ja(t) dφ jφ (t) from which (since a(t) is real-valued), x∗ (t) dx(t) = a(t) da +
dt e dt e dt dt
dx(t) dx(t)
ja2 (t) dφ ∗
dt and hence ℑ x (t) dt = a2 (t) dφ −1 −2 ∗
dt . It follows that k a (t)ℑ x (t) dt =
m(t) as required.
Note that since x(t) = a(t)e jφ (t) , we could alternatively ∗ − jφ (t) =
write x (t) = a(t)e
dx(t)
a2 (t)x−1 (t) which results in m(t) = k−1 ℑ x−1 (t) dt . However this expres-
sion is harder to implement because it involves taking the reciprocal, x−1 (t), of
a rapidly varying complex number instead of the reciprocal, a−2 (t), of a slowly
varying real number whose value is always close to 1.
Some people derived instead the alternative expression involving x−1 (t) by dif-
ferentiating log x(t). Several assumed that a(t) ≡ 1 ∀t which stronger than the
statement in the question.
c) Figure 4.1 shows a block diagram that implements the equation of part b) in
discrete time. Complex-valued signals are shown as bold lines and are repre-
sented using their real and imaginary parts. The block labelled “Conj” takes the
complex conjugate of its input. The differentiation block, D(z), is designed as
an FIR filter using the window method with a target response
(
jω jcω for |ω| ≤ ω1
D(e ) =
0 otherwise
where c is a scaling constant.

i) ¯ of D(z) in simplified form.[ 4 ]
Determine the impulse response d[n]
From the inverse DTFT formula (included in the formula sheet) we

use integration by parts to obtain
ˆ π
¯ 1
D e jω e jωn dω

d[n] =
2π −π
ˆ
jc ω1
= ωe jωn dω
2π −ω1
" #ω1
jc ω jωn 1 jωn
= e − e
2π jn ( jn)2 −ω1
!
jc ω1 1
= 2 cos nω1 − 2 j sin nω1
2π jn ( jn)2
c
= (nω1 cos nω1 − sin nω1 ) .
πn2
¯
For interest only, the above formula gives d[−5 : 5] = [−0.07, 0.18, −0.19, −0.02, 0.71, 0, −0
This is fairly similar to the simplest zero-phase differentiator which
¯
would be d[−1 : 1] = [0.5, 0, −0.5].
Some tried to use the inverse z-transform instead of the inverse DTFT
but not everyone who chose this method realized that the inverse z-
transform involves a contour integration in the complex plane. To
use it, you have to choose an integration contour; if you choose the
unit circle as your contour (a sensible choice), then the inverse z-
transform is identical to the inverse DTFT.
ii) Assuming that ω1 = ω02+π , draw dimensioned sketches showing the

magnitude and phase responses of D(e jω ) over the range −π ≤ ω ≤
π.
[3]
Since ω0 = 0.45π = 1.41, ω1 = ω02+π = 2.28. For ω > ω1 , D(e jω ) =

0 and so the phase
is indeterminate
(shown here as zero). For |ω| ≤
ω1 , we can write D(e ) = | jcω| = c×|ω| and ∠D(e jω ) = ∠ jcω =
jω
π π
2 [−π i f ω < 0] = 2 sgn(ω) assuming c > 0.
2
1
Magnitude/c
Phase (rad)
1.5
0
1
0.5 -1
0
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
Frequency (rad/samp) Frequency (rad/samp)
Surprisingly
a large number of people made the magnitude gain,
D(e jω ), negative for ω < 0. Some thought that the phase too was
∝ ω or that it was always + π2 .
iii) Assume that the DTFT of the window function used when designing
18
D(z) has a main lobe width of ω = ± M+1 for a window of length

M + 1. If ω1 is chosen as ω1 = ω02+π , determine the smallest value
of M that will ensure that the transition in the response of D(e jω )
near ω = ω1 lies completely within the range (ω0 , π). [3]
The transition in the response of D(e jω ) near the discontinuity in

18
D(e jω ) near ω = ω1 = 2.28 will extend for M+1 either side of the
36 36
discontinuity for a total width of M+1 . So we need M+1 ≤ π − ω0 =
36 36 36
0.55π from which M ≥ π−ω0 − 1 = 0.55π − 1 = 1.728 − 1 = 20.83 −
1 = 19.83 ≈ 20.
18 36
Some took the width to be M+1 rather than M+1 .
iv) Stating any assumptions, determine the maximum value of c that will
ensure |s[n]| ≤ 1 where s[n] is the output of the differentiation block,
D(z), as shown in Figure 4.1. [4]
We assume that a(t) ≡ 1 and that D(e jω ) = D(e jω ). Then, at the

maximum frequency deviation of 75 kHz, x[n] =e jω f n where ωf =
75 jω

2π × 400 = 1.18. To ensure |s[n]| ≤ 1, we require D(e ) = cω f ≤
f
1. Hence c ≤ ω1f = 2π×75
400
= 0.849.
1
Many people chose instead to make c ≤ ω1 = 0.44 which is over
conservative.
d) An alternative choice for the target response is

 − jcω (π+ω)

 π−ω1
1
for − π < ω ≤ −ω1
e jω ) = jcω
D(e for |ω| ≤ ω1
 jcω1 (π−ω)

π−ω1 for ω1 < ω ≤ π
i) Assuming that ω1 = ω02+π , draw dimensioned sketches showing the

e jω ) over the range −π ≤ ω ≤
magnitude and phase responses of D(e
π.
[4]
2
1
Magnitude/c
Phase (rad)
1.5
0
1
0.5 -1
0
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
Frequency (rad/samp) Frequency (rad/samp)
Again, quite a few people made the magnitude gain negative for ω <
0.
ii) e jω )
Outline the relative advantages and disadvantages of using D(e

rather than D(e jω ) as the target response when designing D(e jω ).
[2]
e jω ) is a continuous function of ω
The advantage is that since the D(e
it will not be affected by Gibbs phenomenon and the coefficients will
decay ∝ n−2 instead of ∝ n−1 . For any given filter length, M + 1, the
e jω ) than for D(e jω ).
errors will be much smaller for D(e
The disadvantage is that the gain is no longer approximately zero
for frequencies above ω1 and so it may be necessary to include ad-
ditional filtering in the channel selection process in order to remove
frequency components between ω0 and π.
Many people said correctly that a smooth target response was good,
but few were precise about why this was so.

e) An alternative structure that avoids any divisions is shown in Fig. 4.2 where
the polynomial f (v) is the truncated Taylor series for v−1 expanded around
v = 1. Determine f (v) for the cases when it is (i) a linear expression and (ii) a
quadratic expression. In each case determine the gain error (expressed in dB)
resulting from the approximation when a(t) = 1.1. [4]
If we write f (v) = v−1 ,we can write f (v) = f (1)+(v−1) f 0 (1)+ 2!1 (v−1)2 f 00 (1)+
. . . = 1 + (v − 1) × −1 + 2!1 (v − 1)2 × 2 + . . . The Taylor series can also be ob-
tained easily from the geometric progression formula by writing w = 1 − v.
Then v−1 = (1 − w)−1 = 1 + w + w2 + w3 + . . . = 1 + (1 − v) + (1 − v)2 + . . ..
(i) Linear case: f (v) = 1 + (1 − v) = 2 − v. Thus we multiply by 2 − a2 instead
of a−2 . The error is therefore found by substituting a = 1.1 in the ratio of the
2
actual gain (2−a2 ) to the ideal gain (a−2 ) to obtain 2−a
a−2
= 0.959 = −0.392 dB.
(ii) Quadratic case: f (v) = 1 + (1 − v) + (1 − v)2 = 3 − 3v + v2 . The er-
2 +a4
ror is therefore found by substituting a = 1.1 in the ratio 3−3a
a−2
= 1.009 =
+0.0801 dB.
Few attempted this part. Many of those who did attempt it, had difficulty in
writing down the Taylor series even for the linear case.


Imperial College London: EE4-12 EE9-AO2 EE9-SC1

Uploaded by

Copyright:

Available Formats

Imperial College London: EE4-12 EE9-AO2 EE9-SC1

Uploaded by

Document Information

Original Title

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

Imperial College London: EE4-12 EE9-AO2 EE9-SC1

Uploaded by

Copyright:

Available Formats

EE4-12

IMPERIAL COLLEGE LONDON EE9-AO2

DEPARTMENT OF ELECTRICAL AND ELECTRONIC ENGINEERING

DIGITAL SIGNAL PROCESSING AND DIGITAL FILTERS

Tuesday, 6 May 10:00 am

Time allowed: 3:00 hours

There are FOUR questions on this paper.

Answer Question 1 and any TWO other questions

Any special instructions for invigilators and information for

Examiners responsible First Marker(s) : D.M. Brookes

© Imperial College London

Information for Candidates:

BIBO Bounded Input, Bounded Output

• δ [n] = 1 for n = 0 and 0 otherwise.

• δcondition [n] = 1 whenever "condition" is true and 0 otherwise.

• u[n] = 1 for n ≥ 0 and 0 otherwise.

1−α r+1 z−r−1 α q z−q −α r+1 z−r−1

Digital Signal Processing and Digital Filters Imperial

DFT: v[n] = x[n] ~N y[n] , ∑N−1

Order Estimation for FIR Filters

Digital Signal Processing and Digital Filters Imperial

Target H(ẑ) Substitute Parameters

Digital Signal Processing and Digital Filters Imperial

y[n] = αy[n − 1] + (1 − α)x[n]

where 0 < α < 1 is a real constant.

Digital Signal Processing and Digital Filters Imperial

ii) Explain why it is normally desirable to have A(z) ≡ 0. [2]

Digital Signal Processing and Digital Filters Imperial

where K | r means K is a factor of r.

Figure 1.3 Figure 1.4

Digital Signal Processing and Digital Filters Imperial

Digital Signal Processing and Digital Filters Imperial

Digital Signal Processing and Digital Filters Imperial

Figure 3.1 Figure 3.2

Digital Signal Processing and Digital Filters Imperial

where c is a scaling constant.

d) An alternative choice for the target response is

i) Assuming that ω1 = ω02+π , draw dimensioned sketches showing the

Digital Signal Processing and Digital Filters Imperial

Figure 4.1 Figure 4.2

Digital Signal Processing and Digital Filters Imperial

********* Solutions *********

Information for Candidates:

BIBO Bounded Input, Bounded Output

• δ [n] = 1 for n = 0 and 0 otherwise.

• δcondition [n] = 1 whenever "condition" is true and 0 otherwise.

• u[n] = 1 for n ≥ 0 and 0 otherwise.

1−α r+1 z−r−1 α q z−q −α r+1 z−r−1

Digital Signal Processing and Digital Filters Imperial

DFT: v[n] = x[n] ~N y[n] , ∑N−1

Order Estimation for FIR Filters

Digital Signal Processing and Digital Filters Imperial

Target H(ẑ) Substitute Parameters

Digital Signal Processing and Digital Filters Imperial

1. a) i) Explain what is meant by saying that a linear time-invariant system

An LTI system is BIBO-stable if any bounded input sequence x[n]

ii) The impulse response, h[n], of a linear time-invariant system satisfies

Digital Signal Processing and Digital Filters Imperial

Taking z-transforms: Y (z) = αz−1Y (z) + (1 − α)X(z) from which

* Solutions *

By treating the real and imaginary parts ofeach