Advanced_DSP_J_G_Proakis.pdf

1
Advanced Digital Signal Processing
Abdellatif Zaidi †
Department of Electrical Engineering
University of Notre Dame
azaidi@nd.edu
Outline:
1. Introduction
2. Digital processing of continuous-time signals
• Retition: Sampling and sampling theorem
• Quantization
• AD- and DA-conversion
3. DFT and FFT
• Leakage effect
• Windowing
• FFT structure
4. Digital filters
• FIR-filters: Structures, linear phase filters, least-squares
frequency domain design, Chebyshev approximation
• IIR-filters: Structures, classical analog lowpass filter
approximations, conversion to digital transfer functions
• Finite word-length effects
5. Multirate digital signal processing
• Decimation and interpolation
• Filters in sampling rate alteration systems
• Polyphase decomposition and efficient structures
• Digital filter banks
Parts of this textbook have been realized in close collaboration with Dr. Joerg Kliewer whom I warmly thank.
6. Spectral estimation
• Periodogram, Bartlett’s method, Welch’s method,
Blackman-Tukey method
• ARMA modeling, Yule-Walker equation and solution
Literature
• J. G. Proakis, D. G. Manolakis: Digital Signal Process-
ing: Principles, Algorithms, and Applications, Prentice
Hall, 2007, 4th edition
• S. K. Mitra: Digital Signal Processing: A Computer-
Based Approach, McGraw Hill Higher Education, 2006,
3rd edition
• A. V. Oppenheim, R. W. Schafer: Discrete-time signal
processing, Prentice Hall, 1999, 2nd edition
• M. H. Hayes: Statistical Signal Processing and Model-
ing, John Wiley and Sons, 1996 (chapter 6).
2

1. Introduction
1.1 Signals, systems and signal processing
What does “Digital Signal Processing” mean?
Signal:
• Physical quantity that varies with time, space or any other
independent variable
• Mathematically: Function of one or more independent
variables, s1(t) = 5 t, s2(t) = 20 t2
• Examples: Temperature over time t, brightness (luminance) of
an image over (x, y), pressure of a sound wave over (x, y, z)
or (x, y, z, t)
Speech signal:
0.1 0.2 0.3 0.4 0.5 0.6 0.7
−1
−0.5
0
0.5
1
x 10
4
t [s] →
Amplitude
→
Signal Processing:
• Passing the signal through a system
• Examples:
– Modification of the signal (filtering, interpolation, noise
reduction, equalization, . . .)
– Prediction, transformation to another domain (e.g. Fourier
transform)
3
– Numerical integration and differentiation
– Determination of mean value, correlation, p.d.f., . . .
• Properties of the system (e.g. linear/nonlinear) determine the
properties of the whole processing operation
• System: Definition also includes:
– software realizations of operations on a signal, which
are carried out on a digital computer (⇒ software
implementation of the system)
– digital hardware realizations (logic circuits) configured
such that they are able to perform the processing operation,
or
– most general definition: a combination of both
Digital signal processing: Processing of signals by digital means
(software and/or hardware)
Includes:
• Conversion from the analog to the digital domain and back
(physical signals are analog)
• Mathematical specification of the processing operations ⇒
Algorithm: method or set of rules for implementing the system
by a program that performs the corresponding mathematical
operations
• Emphasis on computationally efficient algorithms, which are
fast and easily implemented.
4

Basic elements of a digital signal processing system
Analog signal processing:
Analog
input
signal
Analog
signal
processor
Analog
output
signal
Digital signal processing:
(A/D: Analog-to-digital, D/A: Digital-to-analog)
Digital
signal
processor
Digital
input
signal
Digital
signal
output
Analog
input
signal
Analog
output
signal
converter
A/D
converter
D/A
Why has digital signal processing become so popular?
Digital signal processing has many advantages compared to
analog processing:
Property Digital Analog
Dynamics only limited by
complexity
generally limited
Precision generally unlimited
(costs, complexity ∼
precision)
generally limited (costs
increase drastically with
required precision)
Aging without problems problematic
Production
costs
low higher
Frequency
range
ωdmin ≪ ωamin, ωdmax ≪ ωamax
Linear-phase
frequency
responses
exactly realizable approximately realizable
Complex
algorithms
realizable strong limitations
5
1.2 Digital signal processors (DSPs)
• Programmable microprocessor (more flexibility), or hardwired
digital processor (ASIC, application specific integrated circuit)
(faster, cheaper)
Often programmable DSPs (simply called ”DSPs”) are used for
evaluation purposes, for prototypes and for complex algorithms:
• Fixed-point processors: Twos-complement number repre-
sentation.
• Floating-point processors: Floating point number repre-
sentation (as for example used in PC processors)
Overview over some available DSP processors see next page.
Performance example: 256-point complex FFT
(from [Evaluating DSP Processor Performance, Berkeley Design Technology, Inc., 2000])
6

Some currently available DSP processors and their properties (2006):
Data BDTImark Core Unit price
Manufacturer Family Arithmetic width (bits) 2000(TM) clock speed qty. 10000
Analog Devices ADSP-219x fixed-point 16 410 160 MHz $11-26
ADSP-2126x floating-point 32/40 1090 200 MHz $5-15
ADSP-BF5xx fixed-point 16 4190 750 MHz $5-32
ADSP-TS20x floating/fixed-point 8/16/32/40 6400 600 MHz $131-205
Freescale DSP563xx fixed-point 24 820 275 MHz $4-47
DSP568xx fixed-point 16 110 80 MHz $3-12
MSC71xx fixed-point 16 3370 300 MHz $13-35
MSC81xx fixed-point 16 5610 500 MHz $77-184
Texas-Instuments TMS320C24x fixed-point 16 n/a 40 MHz $2-8
TMS320C54x fixed-point 16 500 160 MHz $3-54
TMS320C55x fixed-point 16 1460 300 MHz $4-17
TMS320C64x fixed-point 8/16 9130 1 GHz $15-208
TMS320C67x floating-point 32 1470 300 MHz $12-31
7
8

9
2. Digital Processing of Continuous-Time Signals
Digital signal processing system from above is refined:
Digital signal
processor
A/D
D/A
Sample-and-
hold circuit
struction filter
Lowpass recon-
lowpass filter
Anti-aliasing Sample-and-
hold circuit
2.1 Sampling
⇒ Generation of discrete-time signals from continuous-time
signals
Ideal sampling
Ideally sampled signal xs(t) obtained by multiplication of the
continuous-time signal xc(t) with the periodic impulse train
s(t) =
∞
X
n=−∞
δ(t − nT ),
where δ(t) is the unit impulse function and T the sampling
period:
xs(t) = xc(t) ·
∞
X
n=−∞
δ(t − nT ) (2.1)
=
∞
X
n=−∞
xc(nT ) δ(t − nT ) (2.2)
10

(”sifting property” of the impulse function)
(a)
0 T 2T 4T 5T
(b) 1
3T
0 T 2T 4T 5T
(c)
3T
(lengths of Dirac deltas correspond to their weighting)
(lengths of Dirac deltas correspond to their weighting)
.
.
t
t
t
xc(t)
s(t)
xs(t)
How does the Fourier transform F{xs(t)} = Xs(jΩ) look
like?
Fourier transform of an impulse train:
s(t) ◦−• S(jΩ) =
2π
T
∞
X
k=−∞
δ(Ω − kΩs) (2.3)
Ωs = 2π/T : sampling frequency in radians/s.
Writing (2.1) in the Fourier domain,
Xs(jΩ) =
1
2π
Xc(jΩ) ∗ S(jΩ),
11
we finally have for the Fourier transform of xs(t)
Xs(jΩ) =
1
T
∞
X
k=−∞
Xc(j(Ω − kΩs)). (2.4)
⇒ Periodically repeated copies of Xs(jΩ)/T , shifted by integer
multiples of the sampling frequency
1
T
Ωs 2Ωs
(Ωs − ΩN )
(d) = Aliasing
3Ωs
−3Ωs −2Ωs −Ωs Ωs 2Ωs
2π
T
−ΩN
1
3Ωs
−3Ωs −2Ωs −Ωs 2Ωs
ΩN
−ΩN
1
T
0
(a)
(b)
(c)
Ωs
ΩN
0 2π 4π
−2π
−4π
−6π 6π
(Ωs − ΩN )
Xs(jΩ)
Ω
.
S(jΩ)
Xc(jΩ)
Xs(jΩ)
Ω
Ω
Ω
.
ω
ω = Ω T
12

(a) Fourier transform of a bandlimited continuous-time input
signal Xc(jΩ), highest frequency is ΩN
(b) Fourier transform of the Dirac impulse train
(c) Result of the convolution S(jΩ) ∗ Xc(jΩ). It is evident that
when
Ωs − ΩN > ΩN or Ωs > 2ΩN (2.5)
the replicas of Xc(jΩ) do not overlap.
⇒ xc(t) can be recovered with an ideal lowpass filter!
(d) If (2.5) does not hold, i.e. if Ωs ≤ 2ΩN, the copies of
Xc(jΩ) overlap and the signal xc(t) cannot be recovered by
lowpass filtering. The distortions in the gray shaded areas are
called aliasing distortions or simply aliasing.
Also in (c): Representation with the discrete (normalized) frequency
ω =ΩT =2πfT (f frequency in Hz) for the discrete signal xc(nT )=x(n)
with F∗{x(n)} = X(ejω
), F∗{·} denoting the Fourier transform for
discrete-time aperiodic signals (DTFT)
Nonideal sampling
⇒ Modeling the sampling operation with the Dirac impulse train
is not a feasible model in real life, since we always need a finite
amount of time for acquiring a signal sample.
Nonideally sampled signal xn(t) obtained by multiplication of a
continuous-time signal xc(t) with a periodic rectangular window
function an(t): xn(t) = xc(t) · an(t) where
an(t) = a0(t) ∗
∞
X
n=−∞
δ(t − n T ) =
∞
X
n=−∞
a0(t − nT )
(2.6)
13
a0(t) denotes the rectangular prototype window:
a0(t) = rect

t − αT/2
αT

(2.7)
with rect(t) :=
(
0 for |t| 1/2
1 for |t| 1/2
(2.8)
T T
1
a0(t)
t
rect(t) ◦−• sinc(Ω/2),
sinc(x) := sin(x)/x
Fourier transform of an(t):
Fourier transform of the rectangular time window in (2.7) (see
properties of the Fourier transform)
A0(jΩ) = F{a0(t)} = α T · sinc(ΩαT/2) · e
−jΩαT/2
(2.9)
Fourier transform of an(t) in (2.6) (analog to (2.3)):
An(jΩ) = A0(jΩ) ·
2π
T
∞
X
k=−∞
δ(Ω − kΩs)
= 2πα
∞
X
k=−∞
sinc(kΩsα T/2) e
−jkΩsαT/2
δ(Ω−kΩs)
= 2πα
∞
X
k=−∞
sinc(kπα) e
−jkπα
δ(Ω−kΩs)
(2.10)
14

Since
xn(t) = xc(t) an(t) ◦−• Xn(jΩ) =
1
2π
(Xc(jΩ)∗An(jΩ))
we finally have by inserting (2.10)
Xn(jΩ) = α
∞
X
k=−∞
sinc(kπα) e
−jkπα
Xc(j(Ω − kΩs)).
(2.11)
From (2.11) we can deduce the following:
• Compared to the result in the ideal sampling case (cp. (2.4))
here each repeated spectrum at the center frequency kΩs is
weighted with the term sinc(kπα) e−jkπα
.
• The energy |Xn(jΩ)|2
is proportional α2
: This is
problematic since in order to approximate the ideal case we
would like to choose the parameter α as small as possible.
15
Solution: Sampling is performed by a sample-and-hold (S/H)
circuit
(from [Proakis, Manolakis, 1996])
(a) Basic elements of an A/D converter, (b) time-domain response of an ideal
S/H circuit
• The goal is to continously sample the input signal and to hold
that value constant as long as it takes for the A/D converter to
obtain its digital representation.
• Ideal S/H circuit introduces no distortion and can be modeled
as an ideal sampler.
⇒ Drawbacks for the nonideal sampling case can be avoided
(all results for the ideal case hold here as well).
16

2.2 Sampling Theorem
Reconstruction of an ideally sampled signal by ideal lowpass
filtering:
−2Ωs −Ωs 2Ωs
ΩN
−ΩN
1
T
ΩN
−ΩN
1
Hr(jΩ)
xc(t) xs(t) xr(t)
(a)
(b)
(c)
−Ωc Ωc
ΩN
−ΩN
1
(d)
(e)
T
ΩN Ωc (Ωs − ΩN )
Ωs
(Ωs − ΩN )
s(t) =
∞
P
n=−∞
δ(t − nT)
Xs(jΩ)
Xc(jΩ)
Ω
Ω
Hr(jΩ)
Xr(jΩ)
Ω
Ω
In order to get the input signal xc(t) back after reconstruction,
i.e. Xr(jΩ) = Xc(jΩ), the conditions
ΩN
Ωs
2
and ΩN Ωc (Ωs − ΩN) (2.12)
have both to be satisfied. Then,
Xc(jΩ) = Xr(jΩ) = Xs(jΩ) · Hr(jΩ) •−◦
xc(t) = xr(t) = xs(t) ∗ hr(t). (2.13)
We now choose the cutoff frequency Ωc of the lowpass filter as
Ωc = Ωs/2 (satisfies both inequalities in (2.12)).
17
Then, with the definition of the rect(·) function in (2.8) we have
Hr(jΩ) = T rect(Ω/Ωs) •−◦ hr(t) = sinc(Ωst/2).
(2.14)
Combining (2.13), (2.14), and (2.2) yields
xc(t) =
∞
Z
−∞
∞
X
n=−∞
xc(nT ) δ(τ − nT ) sinc

1
2
Ωs(t−τ)

dτ
=
∞
X
n=−∞
xc(nT )
∞
Z
−∞
δ(τ −nT ) sinc

1
2
Ωs(t − τ)

dτ
=
∞
X
n=−∞
xc(nT ) sinc

1
2
Ωs(t−nT )

.
Sampling theorem:
Every bandlimited continuous-time signal xc(t) with
ΩN Ωs/2 can be uniquely recovered from its samples
xc(nT ) according to
xc(t) =
∞
X
n=−∞
xc(nT ) sinc

1
2
Ωs(t−nT )

. (2.15)
Remarks:
• Eq. (2.15) is called the ideal interpolation formula, and the
sinc-function is named ideal interpolation function
18

• Reconstruction of a continuous-time signal using ideal
interpolation:
Anti-aliasing lowpass filtering:
In order to avoid aliasing, the continuous-time input signal has to
be bandlimited by means of an anti-aliasing lowpass-filter with
cut-off frequency Ωc ≤ Ωs/2 prior to sampling, such that the
sampling theorem is satisfied.
2.3 Reconstruction with sample-and-hold circuit
In practice, a reconstruction is carried out by combining a D/A
converter with a S/H circuit, followed by a lowpass reconstruction
(smoothing) filter
S/H
h0(t)
digital
input signal
hr(t)
Lowpass xDA(t)
D/A
x0(t)
• D/A converter accepts electrical signals that correspond to
binary words as input, and delivers an output voltage or current
being proportional to the value of the binary word for every
clock interval nT
• Often, the application on an input code word yields a
high-amplitude transient at the output of the D/A converter
(”glitch”)
19
⇒ S/H circuit serves as a ”deglitcher”:
Output of the D/A converter is held constant at the previous
output value until the new sample at the D/A output reaches
steady state
Analysis:
The S/H circuit has the impulse response
h0(t) = rect

t − T/2
T

(2.16)
which leads to a frequency response
H0(jΩ) = T · sinc(ΩT/2) · e
−jΩT/2
(2.17)
20

• No sharp cutoff frequency response characteristics ⇒ we
have undesirable frequency components (above Ωs/2), which
can be removed by passing x0(t) through a lowpass
reconstruction filter hr(t). This operation is equivalent
to smoothing the staircase-like signal x0(t) after the S/H
operation.
• When we now suppose that the reconstruction filter hr(t) is
an ideal lowpass with cutoff frequency Ωc = Ωs/2 and an
amplification of one, the only distortion in the reconstructed
signal xDA(t) is due to the S/H operation:
|XDA(jΩ) = |Xc(jΩ)| · |sinc(ΩT/2)| (2.18)
|Xc(jΩ)| denotes the magnitude spectrum for the ideal
reconstruction case.
⇒ Additional distortion when the reconstruction filter is not ideal
(as in real life!)
⇒ Distortion due to the sinc-function may be corrected by pre-
biasing the frequency response of the reconstruction filter
Spectral interpretation of the reconstruction process (see next page):
(a) Magnitude frequency response of the ideally sampled continuous-time
signal
(b) Frequency response of the S/H circuit (phase factor e−jΩT/2
omitted)
(c) Magnitude frequency response after the S/H circuit
(d) Magnitude frequency response: lowpass reconstruction filter
(e) Magnitude frequency response of the reconstructed continuous-time signal
21
(b)
0
−Ωs Ωs 2Ωs
T sinc(ΩT/2)
(a)
0
−Ωs Ωs 2Ωs
1
T
|Xs(jΩ)|
T
(e)
0
−Ωs Ωs
|XDA(jΩ)|
(d)
0
−Ωs Ωs
1
|Hr(jΩ)|
(c)
0
−Ωs Ωs 2Ωs
|X0(jΩ)|
−Ωc Ωc
1
1
Ωc = Ωs/2
1
T |Xc(jΩ)|
Ω
Ω
Ω
Ω
Ω
22

2.4 Quantization
Conversion carried out by an A/D converter involves quantization
of the sampled input signal xs(nT ) and the encoding of the
resulting binary representation
• Quantization is a nonlinear and noninvertible process which
realizes the mapping
xs(nT ) = x(n) −→ xk ∈ I, (2.19)
where the amplitude xk is taken from a finite alphabet I.
• Signal amplitude range is divided into L intervals In using the
L+1 decision levels d1, d2, . . . , dL+1:
In = {dk x(n) ≤ dk+1}, k = 1, 2, . . . , L
Amplitude
x3 x4 xk
d3 d4 d5 dk dk+1
Ik
levels
Quantization Decision
levels
• Mapping in (2.19) is denoted as x̂(n) = Q[x(n)]
• Uniform or linear quantizers with constant quantization step
size ∆ are very often used in signal processing applications:
∆ = xk+1 − xk = const., for all k = 1, 2, . . . , L−1
(2.20)
• Midtreat quantizer: Zero is assigned a quantization level
Midrise quantizer: Zero is assigned a decision level
23
• Example: midtreat quantizer with L = 8 levels and range
R = 8∆
Amplitude
d9 =∞
x3 x4 x8
d3 d4 d5 d8
x1 d2 x2 x5 d6 x7
d7
x6
−4∆ −2∆ −∆ 0 ∆ 2∆ 3∆
−3∆
d1 =−∞
Range R of quantizer
• Quantization error e(n) with respect to the unquantized
signal
−
∆
2
e(n) ≤
∆
2
(2.21)
If the dynamic range of the input signal (xmax − xmin) is
larger than the range of the quantizer, the samples exceeding
the quantizer range are clipped, which leads to e(n) ∆/2.
• Quantization characteristic function for a midtreat quantizer
with L = 8:
24

Coding
The coding process in an A/D converter assigns a binary number
to each quantization level.
• With a wordlength of b bits we can represent 2b
L binary
numbers, which yields
b ≥ log2(L). (2.22)
• The step size or the resolution of the A/D converter is given as
∆ =
R
2b
with the range R of the quantizer. (2.23)
• Commonly used bipolar codes:
• Two’s complement representation is used in most fixed-
point DSPs: A b-bit binary fraction [β0β1β2 . . . βb−1], β0
denoting the most significant bit (MSB) and βb−1 the least
25
significant bit (LSB), has the value
x = −β0 +
b−1
X
ℓ=1
βℓ 2
−ℓ
(2.24)
• Number representation has no influence on the performance of
the quantization process!
Performance degradations in practical A/D converters:
26

Quantization errors
Quantization error is modeled as noise, which is added to the
unquantized signal:
x̂(n)
Q(x(n))
x(n)
x̂(n)=x(n) + e(n)
x(n)
e(n)
Quantizer
Mathematical model
Actual system
Assumptions:
• The quantization error e(n) is uniformly distributed over the
range −∆
2 e(n) ∆
2 .
• The error sequence e(n) is modeled as a stationary white
noise sequence.
• The error sequence e(n) is uncorrelated with the signal
sequence x(n).
• The signal sequence is assumed to have zero mean.
Assumptions do not hold in general, but they are fairly well
satisfied for large quantizer wordlengths b.
Effect of the quantization error or quantization noise on the
resulting signal x̂(n) can be evaluated in terms of the signal-to-
noise ratio (SNR) in Decibels (dB)
SNR := 10 log10
σ2
x
σ2
e
!
, (2.25)
27
where σ2
x denotes the signal power and σ2
e the power of the
quantization noise.
Quantization noise is assumed to be uniformly distributed in the
range (−∆/2, ∆/2):

2

2
1

e
p(e)
⇒ Zero mean, and a variance of
σ
2
e =
Z ∆/2
−∆/2
e
2
p(e) de =
1
∆
Z ∆/2
−∆/2
e
2
de =
∆2
12
(2.26)
Inserting (2.23) into (2.26) yields
σ
2
e =
2−2b
R2
12
, (2.27)
and by using this in (2.25) we obtain
SNR = 10 log10
σ2
x
σ2
e
!
= 10 log10
12 22b
σ2
x
R2
!
= 6.02 b + 10.8 −20 log10

R
σx

| {z }
(∗)
dB. (2.28)
Term (∗) in (2.28):
28

• σx root-mean-square (RMS) amplitude of the signal v(t)
• σx to small ⇒ decreasing SNR
• Furthermore (not directly from (∗)): σx to large ⇒ range R
is exceeded
⇒ Signal amplitude has to be carefully matched to the range of
the A/D converter
For music and speech a good choice is σx = R/4. Then the SNR
of a b-bit quantizer can be approximately determined as
SNR = 6.02 b − 1.25 dB. (2.29)
Each additional bit in the quantizer increases the signal-to-noise
ratio by 6 dB!
Examples:
Narrowband speech: b = 8 Bit ⇒ SNR = 46.9 dB
Music (CD): b = 16 Bit ⇒ SNR = 95.1 dB
Music (Studio): b = 24 Bit ⇒ SNR = 143.2 dB
29
2.5 Analog-to-digital converter realizations
Flash A/D converter
Analog
comparator:
(from [Mitra, 2000], N =b: resolution in bits)
• Analog input voltage VA is simultaneously compared with a
set of 2b
−1 separated reference voltage levels by means of a
set of 2b
−1 analog comparators ⇒ locations of the comparator
circuits indicate range of the input voltage.
• All output bits are developed simultaneously ⇒ very fast
conversion
• Hardware requirements increase exponentially with an
increase in resolution
⇒ Flash converters are used for low-resultion (b 8 bit) and
high-speed conversion applications.
Serial-to-parallel A/D converters
Here, two b/2-bit flash converters in a serial-parallel
configuration are used to reduce the hardware complextity
at the expense of a slightly higher conversion time
30

Subranging A/D converter:
Ripple A/D converter:
Advantage of both structures: Always one converter is idle while
the other one is operating
⇒ Only one b/2-bit converter is necessary
Sucessive approximation A/D converter
Iterative approach: At the k-th step of the iteration the binary
31
approximation in the shift register is converted into an (analog)
reference voltage VD by D/A conversion (binary representation
[β0β1 . . . βkβk+1 . . . βb−1], βk ∈ {0, 1}∀k):
• Case 1: Reference voltage VD VA ⇒ increase the binary
number by setting both the k-th bit and the (k+1)-th bit to 1
• Case 2: Reference voltage VD ≥ VA ⇒ decrease the binary
number by setting the k-th bit to 0 and the (k+1)-th bit to 1
⇒ High resolution and fairly high speed at moderate costs, widely
used in digital signal processing applications
Oversampling sigma-delta A/D converter, to be discussed in Section 5. . .
2.6 Digital-to-analog converter realizations
Weighted-resistor D/A converter
N 1 N 2 1 0
Output Vo of the D/A converter is given by
Vo =
N−1
X
ℓ=0
2
ℓ
βℓ
RL
(2N − 1)RL + 1
VR
VR: reference voltage
Full-scale output voltage Vo,F S is obtained when βℓ = 1 for all
ℓ = 0, . . . , N −1:
32

Vo,F S =
(2N
− 1)RL
(2N − 1)RL + 1
VR ≈ VR, since (2
N
−1)RL ≫ 1
Disadvantage: For high resolutions the spread of the resistor
values becomes very large
Resistor-ladder D/A converter
N 2 N 1
3
2
1
0
⇒ R–2R ladder D/A converter, most widely used in practice.
Output Vo of the D/A converter:
Vo =
N−1
X
ℓ=0
2
ℓ
βℓ
RL
2(RL + R)
·
VR
2N−1
In practice, often 2RL ≫ R, and thus, the full-scale output
voltage Vo,F S is given as
Vo,F S ≈
(2N
− 1)
2N
VR
Oversampling sigma-delta D/A converter, to be discussed in Section 5. . .
33
3. DFT and FFT
3.1 DFT and signal processing
Definition DFT from Signals and Systems:
DFT: v(n) ◦−• VN(k) =
N−1
X
n=0
v(n) W
kn
N (3.1)
IDFT: VN(k) •−◦ v(n) =
1
N
N−1
X
k=0
VN(k) W
−kn
N (3.2)
with WN := e−j2π/N
, N: number of DFT points
3.1.1 Linear and circular convolution
Linear convolution of two sequences v1(n) and v2(n), n ∈ Z
Z:
yl(n) = v1(n) ∗ v2(n) = v2(n) ∗ v1(n)
=
∞
X
k=−∞
v1(k) v2(n − k) =
∞
X
k=−∞
v2(k) v1(n − k) (3.3)
Circular convolution of two periodic sequences v1(n) and
v2(n), n = {0, . . . , N1,2 − 1} with the same period
N1 =N2 =N and n0 ∈ Z
Z:
34

yc(n) = v1(n) ∗ v2(n) = v2(n) ∗ v1(n)
=
n0+N−1
X
k=n0
v1(k) v2(n − k) =
n0+N−1
X
k=n0
v2(k) v1(n − k)
(3.4)
We also use the symbol N instead of ∗ .
DFT and circular convolution
Inverse transform of a finite-length sequence v(n),
n, k = 0, . . . , N −1:
v(n) ◦−• VN(k) •−◦ v(n) = v(n + λN) (3.5)
⇒ DFT of a finite-length sequence and its periodic extension is
identical!
Circular convolution property (n, k = 0, . . . , N −1)
(v1(n) and v2(n) denote finite-length sequences):
y(n) = v1(n) N v2(n) ◦−• Y (k) = V1N
(k) · V2N
(k)
(3.6)
Proof:
IDFT of Y (k): y(n) =
1
N
N−1
X
k=0
Y (k)W
−kn
N
=
1
N
N−1
X
k=0
V1N
(k)V2N
(k)W
−kn
N
35
Substitution of the DFT definition in (3.1) for v1(n) and v2(n):
y(n) =
1
N
N−1
X
k=0
N−1
X
m=0
v1(m)W
km
N
# N−1
X
l=0
v2(l)W
kl
N
#
W
−kn
N
=
1
N
N−1
X
m=0
v1(m)
N−1
X
l=0
v2(l)
N−1
X
k=0
W
−k(n−m−l)
N
#
(3.7)
Term in brackets: Summation over the unit circle
N−1
X
k=0
e
j2πk(n−m−l)/N
=
(
N for l=n−m+λN, λ ∈ Z
Z
0 otherwise
(3.8)
Substituting (3.8) into (3.7) yields the desired relation
y(n) =
N−1
X
k=0
v1(k)
∞
X
λ=−∞
v2(n − k + λN)
| {z }
=: v2((n−k))N (periodic extension)
=
N−1
X
k=0
v1(k)v2((n − k))N (3.9)
= v1(n) N v2(n)
36

Example: Circular convolution y(n) = v1(n) N v2(n):
0 N
0 N
0 N
0 N
0 N
n
v2(n)
n
v1(n) = δ(n−1)
k
v2((0 − k))N , k = 0, . . . , N −1, n=0
n
k
v2((1 − k))N , k = 0, . . . , N −1, n=1
y(n) = v1(n)°
N v2(n)
3.1.2 Use of the DFT in linear filtering
• Filtering operation can also be carried out in the frequency
domain using the DFT ⇒ attractive since fast algorithms (fast
Fourier transforms) exist
• DFT only realizes circular convolution, however, the desired
37
operation for linear filtering is linear convolution. How can
this be achieved by means of the DFT?
Given: Finite-length sequences v1(n) with length N1 and v2(n)
with length N2
• Linear convolution:
y(n) =
N1−1
X
k=0
v1(k) v2(n − k)
Length of the convolution result y(n): N1 + N2 − 1
• Frequency domain equivalent: Y (ejω
) = V1(ejω
) V2(ejω
)
• In order to represent the sequence y(n) uniquely in the
frequency domain by samples of its spectrum Y (ejω
), the
number of samples must be equal or exceed N1 + N2 − 1
⇒ DFT of size N ≥ N1 + N2 − 1 is required.
• Then, the DFT of the linear convolution y(n) = v1(n) ∗
v2(n) is Y (k) = V1(k) · V2(k), k = 0, . . . , N −1.
This result can be summarized as follows:
The circular convolution of two sequences v1(n) with length N1
and v2(n) with length N2 leads to the same result as the linear
convolution v1(n)∗v2(n) when the lengths of v1(n) and v2(n)
are increased to N =N1+N2−1 points by zero padding.
Other interpretation: Circular convolution as linear
convolution with aliasing
IDFT leads to periodic sequence in the time-domain:
yp(n) =





∞
P
λ=−∞
y(n − λN) n = 0, . . . , N −1,
0 otherwise
38

with Y (k) = DFTN{y(n)} = DFTN{yp(n)}
⇒ For N N1 + N2 − 1: Circular convolution equivalent to
linear convolution followed by time domain aliasing
Example:
0 2N1 −1
0 N1
0
−N1
0 N1 =N2
0 2N1 −1
0 N1 =N2
y(n)=x1(n) ∗ x2(n)
n
N1
n
N1
n
y(n+N1), N1 =6
N1
n
N1
n
N1
N =12
x1(n)°
N x2(n)
y(n−N1), N1 =6
N1 =N2 =6
x1(n)°
N x2(n)
x1(n)=x2(n)
n
1 N1 =N2 =6
39
3.1.3 Filtering of long data sequences
Filtering of a long input signal v(n) with the finite impulse
response h(n) of length N2
Overlap-add method
1. Input signal is segmented into separate blocks:
vν(n) = v(n + νN1), n ∈ {0, 1, . . . , N1−1}, ν ∈ Z
Z.
2. Zero-padding for the signal blocks vν(n) → ṽν(n) and the
impulse response h(n) → h̃(n) to the length N = N1 +
N2−1.
Input signal can be reconstructed according to
v(n) =
∞
X
ν=−∞
ṽν(n − νN1)
since ṽν(n) = 0 for n = N1+1, . . . , N.
3. The two N-point DFTs are multiplied together to form
Yν(k) = Ṽν(k) · H̃(k), k = 0, . . . , N −1.
4. The N-point IDFT yields data blocks that are free from
aliasing due to the zero-padding in step 2.
5. Since each input data block vν(n) is terminated with N−N1
zeros the last N −N1 points from each output block yν(n)
must be overlapped and added to the first N−N1 points of the
succeeding block (linearity property of convolution):
y(n) =
∞
X
ν=−∞
yν(n − νN1)
⇒ Overlap-add method
40

Linear FIR (finite impulse response) filtering by the overlap-add
method:
00
00
11
11
00
00
11
11
00
00
11
11
00
00
11
11
00
00
11
11
00
00
11
11
00
00
11
11
00
00
11
11
Output signal:
y0(n)
y1(n)
y2(n)
added together
N −N1 samples
added together
N −N1 samples
L L L
N −N1 zeros
x̂2(n)
N −N1 zeros
x̂1(n)
N −N1 zeros
Input signal:
x̂0(n)
Overlap-save method
1. Input signal is segmented into by N−N1 samples overlapping
blocks:
vν(n) = v(n + νN1), n ∈ {0, 1, . . . , N −1}, ν ∈ Z
Z.
2. Zero-padding of the filter impulse response h(n) → h̃(n) to
the length N =N1+N2−1.
3. The two N-point DFTs are multiplied together to form
Yν(k) = Vν(k) · H̃(k), k = 0, . . . , N −1.
41
4. Since the input signal block is of length N the first N −N1
points are corrupted by aliasing and must be discarded. The
last N2 =N −N1−1 samples in yν(n) are exactly the same
as the result from linear convolution.
5. In order to avoid the loss of samples due to aliasing the last
N −N1 samples are saved and appended at the beginning of
the next block. The processing is started by setting the first
N −N1 samples of the first block to zero.
Linear FIR filtering by the overlap-save method:
00
00
11
11
00
00
11
11
00
00
11
11
00
00
11
11 00
00
11
11
00
00
11
11
00
00
11
11
00
00
11
11
00
00
11
11
00
00
11
11
Discard
N −N1 samples
Discard
N −N1 samples
Discard
N −N1 samples
L L L
N −N1 zeros
x0(n)
x1(n)
N−L
Input signal:
x2(n)
y0(n)
y1(n)
y2(n)
Output signal:
Nore computationally efficient than linear convolution? Yes, in
combination with very efficient algorithms for DFT computation.
42

3.1.4 Frequency analysis of stationary signals
Leakage effect
Spectral analysis of an analog signal v(t):
• Antialiasing lowpass filtering and sampling with Ωs ≥ 2Ωb,
Ωb denoting the cut-off frequency of the signal
• For practical purposes (delay, complexity): Limitation of the
signal duration to the time interval T0 = L T (L: number of
samples under consideration, T : sampling interval)
Limitation to a signal duration of T0 can be modeled as
multiplication of the sampled input signal v(n) with a rectangular
window w(n)
v̂(n) = v(n) w(n) with w(n)=
(
1 for 0 ≤ n ≤ L−1,
0 otherwise.
(3.10)
Suppose that the input sequence just consists of a single sinusoid,
that is v(n) = cos(ω0n). The Fourier transform is
V (e
jω
) = π(δ(ω − ω0) + δ(ω + ω0)). (3.11)
For the window w(n) the Fourier transform can be obtained as
W (e
jω
) =
L−1
X
n=0
e
−jωn
=
1 − e−jωL
1 − e−jω
= e
−jω(L−1)/2sin(ωL/2)
sin(ω/2)
.
(3.12)
43
We finally have
V̂ (e
jω
) =
1
2π
h
V (e
jω
) ∗ W (e
jω
)
i
=
1
2
h
W (e
j(ω−ω0)
) + W (e
j(ω+ω0)
)
i
(3.13)
Nagnitude frequency response |V̂ (ejω
)| for L=25:
(from [Proakis, Nanolakis, 1996])
Windowed spectrum V̂ (ejω
) is not localized to one frequency,
instead it is spread out over the whole frequency range
⇒ spectral leaking
First zero crossing of W (ejω
) at ωz = ±2π/L:
• The larger the number of sampling points L (and thus also the
width of the rectangular window) the smaller becomes ωz (and
thus also the main lobe of the frequency response).
• ⇒ Decreasing the frequency resolution leads to an increase
of the time resolution and vice versa (duality of time and
frequency domain).
In practice we use the DFT in order to obtain a sampled
representation of the spectrum according to V̂ (ejωk),
k =0, . . . , N −1.
44

Special case: If
N = L and ω0 =
2π
N
ν, ν =0, . . . , N −1
then the Fourier transform is exactly zero at the sampled
frequencies ωk, except for k =ν.
Example: (N = 64, n = 0, . . . , N − 1, rectangular window
w(n))
v0(n) = cos

5
2π
N
n

, v1(n) = cos

5
2π
N
+
π
N

n

0 20 40 60
−1
−0.5
0
0.5
1
v0
(n)=cos(2π/N⋅5⋅n)
n
0 20 40 60
−1
−0.5
0
0.5
1
v1
(n)=cos((2π/N⋅5+π/N)⋅n)
n
0 10 20 30
0
0.2
0.4
0.6
0.8
1
DFT(v0
(n) w(n)), rect. window
k
0 10 20 30
0
0.2
0.4
0.6
0.8
1
DFT(v1
(n) w(n)), rect. window
k
• Left hand side: V̂0(ejωk) = V0(ejωk) ∗ W (ejωk) = 0 for
45
ωk = k2π/N except for k =5, since ω0 is exactly an integer
multiple of 2π/N
⇒ periodic repetition of v0(n) leads to a pure cosine
sequence
• Right hand side: slight increase of ω0 6= ν2π/N for ν ∈ Z
Z
⇒ V̂1(ejωk) 6= 0 for ωk = k2π/N, periodic repetition is
not a cosine sequence anymore
Windowing and different window functions
Windowing not only distorts the spectral estimate due to leakage
effects, it also reduces the spectral resolution.
Consider a sequence of two frequency components
v(n)=cos(ω1n)+cos(ω2n) with the Fourier transform
V (e
jω
) =
1
2
h
W (e
j(ω−ω1)
) + W (e
j(ω−ω2)
) +
+ W (e
j(ω+ω1)
) + W (e
j(ω+ω2)
)
i
where W (ejω
) is the Fourier transform of the rectangular
window from (3.12).
• 2π/L|ω1−ω2|: Two maxima, main lobes for both window
spectra W (ej(ω−ω1)
) and W (ej(ω−ω2)
) can be separated
• |ω1 −ω2| = 2π/L: Correct values of the spectral samples,
but main lobes cannot be separated anymore
• |ω1 − ω2| 2π/L: Nain lobes of W (ej(ω−ω1)
) and
W (ej(ω−ω2)
) overlap
⇒ Ability to resolve spectral lines of different frequencies is
limited by the main lobe width, which also depends on the length
of the window impulse response L.
46

Example: Magnitude frequency response |V (ejω
)| for
v(n) = cos(ω0n) + cos(ω1n) + cos(ω2n) (3.14)
with ω0 = 0.2 π, ω1 =0.22 π, ω2 = 0.6 π and (a) L = 25,
(b) L=50, (c) L=100
The cosines with the frequencies ω0 and ω1 are only resolvable
for L=100.
To reduce leakage, we can choose a different window function
with lower side lobes (however, this comes with an increase of
the width of the main lobe). One choice could be the Hanning
47
window, specified as
wHan(n) =



1
2
h
1 − cos

2π
L−1n
i
for 0 ≤ n ≤ L − 1,
0 otherwise.
(3.15)
Nagnitude frequency response |V̂ (ejω
)| from (3.13), where
W (ejω
) is replaced by WHan(ejω
) (L=25):
Spectrum of the signal in (3.14) after it is windowed with wHan(n)
in (3.15):
48

The reduction of the sidelobes and the reduced resolution
compared to the rectangular window can be clearly observed.
Alternative: Hamming window
wHam(n) =
(
0.54 − 0.46 cos

2π
L−1n

for 0 ≤ n ≤ L − 1,
0 otherwise.
(3.16)
Comparison of rectangular, Hanning and Hamming window
(L=64):
0 20 40 60
0
0.5
1
Rectangular window wrect
(n)
n →
0 0.2 0.4 0.6 0.8 1
−100
−50
0
dB
ω / π →
20 log
10
|Wrect
(e
jω
)|
0 20 40 60
0
0.5
1
Hanning window w
Han
(n)
n →
0 0.2 0.4 0.6 0.8 1
−100
−50
0
dB
ω / π →
20 log
10
|W
Han
(ejω
)|
0 20 40 60
0
0.5
1
Hamming window wHam
(n)
n →
0 0.2 0.4 0.6 0.8 1
−100
−50
0
dB
ω / π →
20 log10
|WHam
(e
jω
)|
Remark: Spectral analysis with DFT
• Sampling grid can be made arbitrarily fine by increasing
the length of the windowed signal with zero padding (that
49
is increasing N). However, the spectral resolution is not
increased (separation of two closely adjacent sine spectra
lines).
• An increase in resolution can only be obtained by increasing
the length of the input signal to be analyzed (that is
increasing L), which also results in a longer window (see
examples above).
3.2 Fast computation of the DFT: The FFT
Complexity of the DFT calculation in (3.1) for v(n) ∈ I
C,
VN(k) ∈ I
C:
VN(k) =
N−1
X
n=0
v(n)W
kn
N
| {z }
1 complex multiplication
| {z }
N compl. mult., N compl. add.
for k = 0, . . . , N −1
| {z }
N results
⇒ Overall N2
complex multiplications and additions.
Remarks:
• 1 complex multiplication → 4 real-valued mult. + 2 real-
valued additions
1 complex addition → 2 real valued additions
• A closer evaluation reveals that there are slightly less than N2
operations:
– N values have to be added up ⇒ (N −1) additions.
– Factors ej0
, ejπλ
, e±j π
2 λ
⇒ no real multiplications.
– For k =0 no multiplication at all.
• Complexity of the the DFT becomes extremely large for
large values of N (i.e. N = 1024) ⇒ efficient algorithms
advantageous.
50

Fast algorithms for DFT calculation (as the fast Fourier transform,
FFT) exploit symmetry and periodicity properties of W kn
N as
• complex conjugate symmetry: W
k(N−n)
N =W −kn
N =(W kn
N )∗
• periodicity in k and n: W kn
N =W
k(n+N)
N =W
(k+N)n
N .
3.2.1 Radix-2 decimation-in-time FFT algorithms
Principle:
Decomposing the DFT computation into DFT computations
of smaller size by means of decomposing the N-point input
sequence v(n) into smaller sequences ⇒ ”decimation-in-time”
Prerequisite:
N integer power of two, i.e. N = 2m
, m=log2(N) ∈ I
N ⇒
”radix-2”
Decomposing a N-point DFT into two N/2-point transforms
DFT V (k) (we drop the subscript N for clarity) can be written
as
V (k) =
N−1
X
n=0
v(n) W
kn
N , k = 0, . . . , N −1
=
N/2−1
X
n=0
v(2n) W
2kn
N +
N/2−1
X
n=0
v(2n + 1) W
k(2n+1)
N ,
(3.17)
where in the last step v(n) is separated into two N/2-point
sequences consisting of the even- and odd-numbered points in
v(n).
51
Since
W
2
N = e
−2j·2π/N
= e
−j2π/(N/2)
= WN/2
we can rewrite (3.17) as
V (k) =
N/2−1
X
n=0
v(2n) W
kn
N/2 + W
k
N
N/2−1
X
n=0
v(2n + 1) W
kn
N/2
(3.18)
= G(k) + W
k
NH(k), k = 0, . . . , N − 1 (3.19)
• Each of the sums in (3.18) represents a N/2 DFT over the
even- and odd-numbered points of v(n), respectively.
• G(k) and H(k) need only to be computed for 0, . . . , N/2−
1 since both are periodic with period N/2.
Signal flow graph for N =8 (v → x, V → X):
(from [Oppenheim, Schafer, 1999])
Complexity:
2 DFTs of length N/2 → 2 · N2
/4 = N2
/2 operations +
52

N operations for the combination of both DFTs
⇒ N + N2
/2 operations (less than N2
for N 2)
Decomposition into 4 N/4-point DFTs
G(k) and H(k) from (3.19) can also be written as
G(k) =
N/4−1
X
n=0
g(2n) W
kn
N/4 + W
k
N/2
N/4−1
X
n=0
g(2n+1) W
kn
N/4,
(3.20)
H(k) =
N/4−1
X
n=0
h(2n) W
kn
N/4 + W
k
N/2
N/4−1
X
n=0
h(2n+1) W
kn
N/4
(3.21)
where k = 0, . . . , N/2−1.
53
The overall flow graph now looks like (v → x, V → X):
Complexity:
4 DFTs of length N/4 → N2
/4 operations
+ 2 · N/2 + N operations for the reconstruction
⇒ N2
/4 + 2N complex multiplications and additions
Final step: Decomposition into 2-point DFTs
DFT of length 2:
V
′
2 (0) = v
′
(0) + W
0
2 v
′
(1) = v
′
(0) + v
′
(1) (3.22)
V
′
2 (1) = v
′
(0) + W
1
2 v
′
(1) = v
′
(0) − v
′
(1) (3.23)
Flow graph:
1
1
1
1
v0
(0)
v0
(1)
V 0
2(0)
V 0
2(1)
Inserting this in the resulting structure from the last step yields the
overall flow graph for (N =8)-point FFT: (v → x, V → X):
54

In general, our decomposition requires m = log2(N) = ld N
stages and for N ≫ 1 we have
N · m = N ld N complex multiplications and additions.
(instead of N2
)
Examples:
N =32 → N2
≈ 1000, N ld N ≈ 160 → factor 6
N =1024 → N2
≈ 106
, N ld N ≈ 104
→ factor 100
Butterfly computations
Basic building block of the above flow graph is called butterfly
(ρ ∈ {0, . . . , N/2−1}):
1
1
Wρ
N
W
(ρ+N/2)
N
1
1
1
−1
Wρ
N
⇒
Simpli£cation
The simplification is due to the fact that
55
W
N/2
N = e−j(2π/N)N/2
= e−jπ
= −1. Therefore we have
W
ρ+N/2
N = W
ρ
N W
N/2
N = −W
ρ
N.
Using this modification, the resulting flow graph for N = 8 is
given as (v → x, V → X):
In-place computations
• The intermediate results V
(ℓ)
N (k1,2) in the ℓ-th stage,
ℓ = 0, . . . , m−1 are obtained as
V
(ℓ)
N (k1) = V
(ℓ−1)
N (k1) + W
ρ
N V
(ℓ−1)
N (k2),
V
(ℓ)
N (k2) = V
(ℓ−1)
N (k1) − W
ρ
N V
(ℓ−1)
N (k2)
(butterfly computations) where k1, k2, ρ ∈ {0, . . . , N−1}
vary from stage to stage.
56

• ⇒ Only N storage cells are needed, which first contain the
values v(n), then the results from the individual stages and
finally the values VN(k) ⇒ In-place algorithm
Bit-reversal
• v(n)-values at the input of the decimation-in-time flow graph
in permuted order
• Example for N = 8, where the indices are written in binary
notation:
# flow graph input 000 001 010 011
v(n) v(000) v(100) v(010) v(110)
# flow graph input 100 101 110 111
v(n) v(001) v(101) v(011) v(111)
⇒ Input data is stored in bit-reversed order
Bit-reversed order is due to the sorting in even and odd indices in
every stage, and thus is also necessary for in-place computation:
(v → x):
57
Inverse FFT
According to (3.2) we have for the inverse DFT
v(n) =
1
N
N−1
X
k=0
VN(k) W
−kn
N ,
that is
v(−n) =
1
N
N−1
X
k=0
VN(k) W
kn
N , ⇐⇒
v(N − n) =
1
N
DFT{VN(k)} (3.24)
⇒ With additional scaling and index permutations the IDFT can
be calculated with the same FFT algorithms as the DFT!
3.2.2 FFT alternatives
Alternative DIT structures
Rearranging of the nodes in the signal flow graphs lead to
FFTs with almost arbitrary permutations of the input and output
sequence. Reasonable approaches are structures
(a) without bitreversal, or
(b) bit-reversal in the frequency domain
58

(a)
(b)
(from [Oppenheim, Schafer, 1999], v → x, V → X )
The flow graph in (a) has the disadvantage, that it is a non-inplace
algorithm, because the butterfly-structure does not continue past
the first stage.
Decimation-in-frequency algorithms
Instead of applying the decomposition to time domain, we could
also start the decomposition in the frequency domain, where
the sequence of DFT coefficients VN(k) is decomposed into
smaller sequences. The resulting algorithm is called decimation-
in-frequency (DIF) FFT.
59
Radix r and mixed-radix FFTs
When we gerally use
N = r
m
for r ≥ 2, r, m ∈ I
N (3.25)
we obtain DIF or DIT decompositions with a radix r. Besides
r =2, r =3 and r =4 are commonly used.
Radix-4 butterfly
(q = 0, . . . , N/4−1)
(N → N):
For special lengths, which can not be expressed as N = rm
,
so called mixed-radix FFT algorithms can be used (i.e. for
N =576=26
·32
).
60

3.3 Transformation of real-valued sequences
v(n) ∈ I
R → FFT program/hardware: vR(n) + j vI(n)
| {z }
= 0
∈ I
C
⇒ Inefficient due to performing arithmetic calculations with zero
values
In the following we will discuss methods for the efficient usage of
a complex FFT for real-valued data.
3.3.1 DFT of two real sequences
Given: v1(n), v2(n) ∈ I
R, n = 0, . . . , N −1
How can we obtain VN1
(k) •−◦ v1(n), VN2
(k) •−◦ v2(n)?
Define
v(n) = v1(n) + j v2(n) (3.26)
leading to the DFT
VN(k) = DFT{v(n)} = VN1
(k)
| {z }
∈ I
C
+j VN2
(k)
| {z }
∈ I
C
. (3.27)
Separation of VN(k) into VN1
(k), VN2
(k)?
Symmetry relations of the DFT:
v(n) = vRe(n) + vRo(n)
| {z }
= v1(n)
+ j vIe(n) + j vIo(n)
| {z }
= v2(n)
(3.28)
Corresponding DFTs:
vRe(n) ◦−• VNRe
(k), vRo(n) ◦−• j VNIo
(k), (3.29)
j vIe(n) ◦−• j VNIe
(k), j vIo(n) ◦−• VNRo
(k). (3.30)
61
Thus, we have
VN1
(k) =
1
2
VNR
(k) + VNR
(N − k)

+
+
j
2
VNI
(k) − VNI
(N − k)

, (3.31)
where
VNRe
(k) =
1
2
VNR
(k) + VNR
(N − k)

,
VNIo
(k) =
1
2
VNI
(k) − VNI
(N − k)

.
Likewise, we have for VN2
(k) the relation
VN2
(k) =
1
2
VNI
(k) + VNI
(N − k)

−
−
j
2
VNR
(k) − VNR
(N − k)

, (3.32)
with
VNIe
(k) =
1
2
VNI
(k) + VNI
(N − k)

,
VNRo
(k) =
1
2
VNR
(k) − VNR
(N − k)

.
62

Rearranging (3.31) and (3.32) finally yields
VN1
(k) =
1
2
VN(k) + V
∗
N(N − k)

,
VN2
(k) = −
j
2
VN(k) − V
∗
N(N − k)

.
(3.33)
Due to the Hermitian symmetry of real-valued sequences
VN(1,2)
(k) = V
∗
N(1,2)
(N − k) (3.34)
the values VN(1,2)
(k) for k ∈ {N/2+1, . . . , N −1} can be
obtained from those for k ∈ {0, . . . , N/2}, such that only a
calculation for N/2+1 values is necessary.
Application: Fast convolution of two real-values sequences with
the DFT/FFT
3.3.2 DFT of a 2N-point real sequence
Given: v(n) ∈ I
R, n = 0, . . . , 2 N −1
Wanted:
V2N(k) = DFT{v(n)} =
2N−1
X
n=0
v(n) W
kn
2N
with k = 0, . . . , 2N −1.
Hermitian symmetry analog to (3.34) since v(n) ∈ I
R for all n:
V2N(2N − k) = V
∗
2N(k), k = 0, . . . , N
63
Define
ṽ(n) := v(2n) + j v(2n + 1), n = 0, . . . , N − 1,
(3.35)
=: v1(n) + j v2(n),
where the even and odd samples of v(n) are written alternatively
into the real and imaginary part of ṽ(n). Thus we have a complex
sequence consisting of two real-valued sequences of length N
with the DFT
ṼN(k
′
) = VN1
(k
′
) + j VN2
(k
′
), k
′
= 0, . . . , N − 1.
(3.36)
VN1
(k′
) and VN2
(k′
) can easily be obtained with (3.33) as
VN1
(k
′
) =
1
2

ṼN(k
′
) + Ṽ
∗
N(N − k
′
)

,
VN2
(k
′
) = −
j
2

ṼN(k
′
) − Ṽ
∗
N(N − k
′
)

for k′
= 0, . . . , N −1.
In order to calculate V2N(k) from VN1
(k′
) and VN2
(k′
) we
rearrange the expression for DFT{v(n)} according to
V2N(k) =
N−1
X
n=0
v(2n)
| {z }
= v1(n)
W
2kn
2N +
N−1
X
n=0
v(2n + 1)
| {z }
= v2(n)
W
(2n+1)k
2N ,
=
N−1
X
n=0
v1(n) W
kn
N + W
k
2N ·
N−1
X
n=0
v2(n) W
nk
N ,
64

Finally we have
V2N(k)=VN1
(k)+W
k
2N VN2
(k), k =0, . . . , 2N −1
(3.37)
Due to the Hermitian symmetry V2N(k)=V ∗
2N(2N−k), k only
needs to be evaluated from 0 to N with VN1/2
(N)=VN1/2
(0).
Signal flow graph:
j
1/2
ṼN (k)
Ṽ ∗
N (N − k)
V2N (k)
Wk
2N
VN1
(k)
1/2
k = 0, . . . , N
−j
k = N → k = 0
VN2
(k)
⇒ Computational savings by a factor of two compared to the
complex-valued case since for real-valued input sequences only
an N point DFT is needed
65
4. Digital Filters
Digital filter = linear-time-invariant (LTI) causal system with a
rational transfer function (without loss of generality: numerator
degree N = denominator degree)
H(z) =
N
P
i=0
bN−i zi
N
P
i=0
aN−i zi
=
N
P
i=0
bi z−i
1 +
N
P
i=1
ai z−i
(4.1)
where a0 =1 without loss of generality.
ai, bi: parameters of the LTI system (⇒ coefficients of the digital
filter), N is said to be the filter order
Product notation of (4.1):
H(z) = b0
N
Q
i=1
(z − z0i)
N
Q
i=1
(z − z∞i)
(4.2)
where the z0i are the zeros, and the z∞i are the poles of the
transfer function (the latter are responsible for stability).
Difference equation:
y(n) =
N
X
i=0
bi v(n − i) −
N
X
i=1
ai y(n − i), (4.3)
66

with v(n) denoting the input signal and y(n) the resulting signal
after the filtering
Remarks
• Generally, (4.3) describes a recursive filter with an infinite
impulse response (IIR filter):
y(n) is calculated from v(n), v(n−1), . . . , v(n−N) and
recursively from y(n−1), y(n−2), . . . , y(n−N).
• The calculation of y(n) requires memory elements in order to
store v(n−1), . . . , v(n−N) and
y(n−1), y(n−2), . . . , y(n−N) ⇒ dynamical system.
• bi ≡ 0 for all i 6= 0:
H(z) =
b0 zN
N
P
i=0
aN−i zi
=
b0 zN
N
Q
i=1
(z − z∞i)
(4.4)
⇒ Filter has no zeros ⇒ All-pole or autoregressive (AR-)
filter.
Transfer function is purely recursive:
y(n) = b0 v(n) −
N
X
i=1
ai y(n − i) (4.5)
• ai ≡ 0 for all i 6= 0, a0 = 1 (causal filter required!):
Difference equation is purely non-recursive:
y(n) =
N
X
i=0
bi v(n − i) (4.6)
⇒ Non-recursive filter
67
Transfer function:
H(z) =
1
zN
N
X
i=0
bN−i z
i
=
N
X
i=0
bi z
−i
(4.7)
– Poles z∞i = 0, i = 1, . . . , N, but not relevant for
stability ⇒ all-zero filter
– According to (4.6): y(n) obtained by a weighted average
of the last N +1 input values ⇒ Moving average (MA)
filter (as opposite to the AR filter from above)
– From (4.7) it can be seen that the impulse response has
finite length ⇒ Finite impulse response (FIR) filter of
length L = N + 1 and order N
4.1 Structures for FIR systems
• Difference equation given by (4.6)
• Transfer function given by (4.7)
• Unit sample response is equal to the coefficients bi
h(n) =
(
bn for 0 ≤ n ≤ L−1
0 otherwise
4.1.1 Direct form structures
The direct form structure follows immediately from the
nonrecursive difference equation given in (4.6), which is
equivalent to the linear convolution sum
y(n) =
L−1
X
k=0
h(k) v(n − k).
68

(from [Proakis, Manolakis, 1996], v → x, L → M)
⇒ Tapped-delay-line or transversal filter in the first direct form
If the unit impulse v(n) = δ(n) is chosen as input signal, all
samples of the impulse response h(n) appear successively at the
output of the structure.
In the following we mainly use the more compact signal flow
graph notation:
h(0) h(1) h(2) h(3) h(L−2) h(L−1)
z−1
v(n)
z−1
z−1
z−1
z−1
z−1
y(n)
The second direct form can be obtained by transposing the flow
graph:
• Reversing the direction of all branches,
• exchanging the input and output of the flow graph
• exchanging summation points with branching points and vice
versa.
69
Transversal filter in the second direct form:
h(L−1) h(L−2) h(1) h(0)
v(n)
y(n)
z−1
z−1
z−1
z−1
z−1
h(2)
h(3)
When the FIR filter has linear phase (see below) the impulse
response of the system satisfies either the symmetry or asymmetry
condition
h(n) = ±h(L − 1 − n). (4.8)
Thus, the number of multiplications can be reduced from L to
L/2 for L even, and from L to (L+1)/2 for L odd.
Signal flow graph for odd L:
z−1
z−1
z−1
z−1
z−1
v(n)
z−1
z−1
z−1
z−1
z−1
z−1
z−1
z−1
h(0) h(1) h(2) h(3)
y(n)
h
¡L−1
2
¢
h
¡L−3
2
¢
4.1.2 Cascade-form structures
By factorizing the transfer function
H(z) = H0
P
Y
p=1
Hp(z) (4.9)
we obtain a cascade realization. The Hp(z) are normally second-
order, since in order to obtain real coefficients, conjugate complex
70

zeros z0i and z∗
0i have to be grouped:
Hp(z) = (1 − z0iz
−1
)(1 − z
∗
0iz
−1
)
= 1 + b1 z
−1
+ b2 z
−2
For linear-phase filters due to the special symmetry (4.8) the zeros
appear in quadruples: Both z0i and z∗
0i, and z−1
0i and (z∗
0i)−1
are
a pair of complex-conjugate zeros. Consequently, we have
Hp(z) = (1−z0iz
−1
)(1−z
∗
0iz
−1
)(1−z
−1
0i z
−1
)(1−(z
∗
0i)
−1
z
−1
),
= 1 + b1 z
−1
+ b2 z
−2
+ b1 z
−3
+ z
−4
.
4.1.3 Lattice structures
Lattice structures are mainly used as predictor filter (i.e. in digital
speech processing) due to their robustness against coefficient
quantization:
2-th stage
(L−1)-th
stage
xL−1(n)
yL−1(n)
x1(n)
y1(n)
x2(n)
1-th stage
v(n) y2(n)
z−1
qi
qi
xi(n)
yi(n)
xi−1(n)
yi−1(n)
1 1 1
1
1
General structure:
i-th stage lattice filter:
71
The behavior of the i-th stage can be written in matrix notation as

Xi(z)
Yi(z)

=

1 qi z−1
qi z−1

·

Xi−1(z)
Yi−1(z)

. (4.10)
After the first stage we have
X1(z) = V (z) + q1 z
−1
V (z),
Y1(z) = q1 V (z) + z
−1
V (z).
(4.11)
It follows
H1(z) =
X1(z)
V (z)
= 1 + q1 z
−1
= α01 + α11 z
−1
,
G1(z) =
Y1(z)
V (z)
= q1 + z
−1
= β01 + β11 z
−1
.
Second stage:
X2(z) =X1(z) + q2 z
−1
Y1(z),
Y2(z) =q2 X1(z) + z
−1
Y1(z).
(4.12)
X2(z) = V (z) + q1z
−1
V (z) + q2q1z
−1
V (z) + q2z
−2
V (z),
Y2(z) = q2V (z) + q1q2z
−1
V (z) + q1z
−1
V (z) + z
−2
V (z),
72

which finally leads to the transfer functions
H2(z) =
X2(z)
V (z)
= 1 + (q1 + q1 q2)z
−1
+ q2 z
−2
, (4.13)
= α02 + α12 z
−1
+ α22z
−2
,
G2(z) =
Y2(z)
V (z)
= q2 + (q1 + q1 q2)z
−1
+ z
−2
, (4.14)
= β02 + β12 z
−1
+ β22z
−2
.
By comparing (4.13) and (4.14) we can see that
H2(z) = z
−2
G2(z
−1
),
that is, the zeros of H2(z) can be obtained by reflecting the zeros
of G2(z) at the unit circle. Generally, it can be shown that
Hi(z) = z
−i
Gi(z
−1
), for i = 1, . . . , L − 1. (4.15)
4.2 Structures for IIR systems
4.2.1 Direct form structures
Rational system function (4.1) can be viewed as two systems in
cascade: H(z) = N(z)/D(z) = H1(z) · H2(z) with
H1(z) =
N
X
i=0
bi z
−i
, H2(z) =
1
1 +
N
P
i=1
ai z−i
The all-zero H1(z) can be realized with the direct form from
Section 3.1.1. By attaching the all-pole system H2(z) in cascade,
73
we obtain the direct form I realization:
z−1
z−1
z−1
z−1
z−1
b0
b1
b2
b3
bN−1
bN
z−1
v(n)
z−1
z−1
z−1
z−1
z−1
−a1
−a2
−a3
−aN−1
−aN
z−1
y(n)
All-zero system N(z) All-pole system 1/D(z)
Another realization can be obtained by exchanging the order of
the all-pole and all-zero filter. Then, the difference equation for
the all-pole section is
w(n) = −
N
X
i=1
ai w(n − i) + v(n),
where the sequence w(n) is an intermediate result and represents
74

the input to the all-zero section:
y(n) =
N
X
i=0
bn w(n − i).
The resulting structure is given as follows:
z−1
z−1
z−1
z−1
z−1
−a1
−a2
−a3
−aN−1
−aN
z−1
v(n)
b1
b2
b3
bN−1
bN
b0
y(n)
⇒ Only one single delay line is required for storing the delayed
versions of the sequence w(n). The resulting structure is called
a direct form II realization. Furthermore, it is said to be canonic,
since it minimizes the number of memory locations (among other
structures).
Transposing the direct form II realization leads to the following
75
structure, which requires the same number of multiplications,
additions, and memory locations as the original structure:
z−1
z−1
z−1
z−1
z−1
b1
b2
b3
bN−1
bN
z−1
v(n)
−a1
−a2
−a3
−aN−1
−aN
y(n)
b0
4.2.2 Cascade-form structures
Analog to Section 4.1.2 we can also factor an IIR system H(z)
into first and second order subsystems Hp(z) according to
H(z) =
P
Y
p=1
Hp(z). (4.16)
⇒ Degrees of freedom in grouping the poles and the zeros
76

First order subsystems:
Canonical direct form for N =1:
z−1
−a1
v(n)
b1
b0
y(n)
Corresponding transfer function:
H(z) =
Y (z)
V (z)
=
b0 + b1 z−1
1 + a1 z−1
(4.17)
Every first order transfer function can be realized with the above
flow graph:
H(z)=
b′
0 + b′
1 z−1
a′
0 + a′
1 z−1
=
(b′
0/a′
0) + (b′
1/a′
0) z−1
1 + (a′
1/a′
0) z−1
=
b0 + b1 z−1
1 + a1 z−1
Second order subsystems:
Canonical direct form for N =2:
z−1
−a1
−a2
z−1
v(n)
b1
b2
b0
y(n)
Corresponding transfer function:
H(z) =
Y (z)
V (z)
=
b0 + b1 z−1
+ b2 z−2
1 + a1 z−1 + a2 z−2
(4.18)
77
Example:
A so called Chebyshev lowpass filter of 5-th order and the cut-off frequency
fco =0.25 fs (fs denoting the sampling frequency) is realized. A filter design
approach (we will discuss the corresponding algorithms later on) yields the
transfer function
H(z) = 0.03217·
·
1 + 5 z−1
+ 10 z−2
+ 10 z−3
+ 5 z−4
+ z−5
1 − 0.782 z−1 + 1.2872 z−2 − 0.7822 z−3 + 0.4297 z−4 − 0.1234 z−5
(4.19)
• The zeros are all at z =−1: z0i = −1 for i = 1, 2, . . . , 5.
The location of the poles are:
z∞1,2 = −0.0336 ± j 0.8821,
z∞3,4 = 0.219 ± j 0.5804, z∞5 = 0.4113.
(4.20)
Grouping of poles z∞1,2:
Ĥ1,2(z) =
1 + 2 z−1
+ z−2
1 + 0.0672 z−1 + 0.7793 z−2
Grouping of poles z∞3,4:
Ĥ3,4(z) =
1 + 2 z−1
+ z−2
1 − 0.4379 z−1 + 0.3849 z−2
Real-valued pole z∞5 leads to a first-order subsystem
Ĥ5(z) =
1 + z−1
1 − 0.4113 z−1
.
• For the implementation on a fixed-point DSP it is advantageous to ensure
that all stages have similar amplification in order to avoid numerical
problems. Therefore, all subsystems are scaled such that they have
78

approximately the same amplification for low frequencies:
H1(z) =
Ĥ5(z)
Ĥ5(z = 1)
=
0.2943 + 0.2943 z−1
1 − 0.4113 z−1
H2(z) =
Ĥ3,4(z)
Ĥ3,4(z = 1)
=
0.2367 + 0.4735 z−1
+ 0.2367 z−2
1 − 0.4379 z−1 + 0.3849 z−2
H3(z) =
Ĥ1,2(z)
Ĥ1,2(z = 1)
=
0.4616 + 0.9233 z−1
+ 0.4616 z−2
1 − 0.4379 z−1 + 0.3849 z−2
Remark: The order of the subsystems is in principle arbitrary. However,
here we know from the pole analysis in (4.20) that the poles of Ĥ1,2(z)
are closest to the unit circle. Thus, using a fixed-point DSP may lead more
likely to numerical overflow compared to Ĥ3,4(z) and Ĥ5(z). Therefore,
it is advisable to realize the most sensible filter as the last subsystem.
• Frequency responses:
0 0.5 1
−40
−20
0
20
20⋅ log
10
|H
1
(e
jω
)|
ω / π
dB
0 0.5 1
−40
−20
0
20
20⋅ log
10
|H
2
(e
jω
)|
ω / π
dB
0 0.5 1
−40
−20
0
20
20⋅ log
10
|H
3
(ejω
)|
ω / π
dB
0 0.5 1
−40
−20
0
20
20⋅ log
10
|H(ejω
)| (overall filter)
ω / π
dB
79
• Resulting signal flow graph (V → U):
(from [Fliege: ”Analoge und digitale Filter”, Hamburg University of Technology, 1990])
4.2.3 Parallel-form structures
⇒ An alternative to the factorization of a general transfer function
is to use a partial-fraction expansion, which leads to a parallel-
form structure
• In the following we assume that we have only distinct poles
(which is quite well satisfied in practice). The partial-fraction
expansion of a transfer function H(z) with numerator and
denominator degree N is then given as
H(z) = A0 +
N
X
i=1
Ai
1 − z∞iz−1
. (4.21)
The Ai, i ∈ {1, . . . , N} are the coefficients (residues) in
the partial-fraction expansion, A0 = bN/aN.
• We furthermore assume that we have only real-valued
coefficients, such that we can combine pairs of complex-
conjugate poles to form a second order subsystem
(i ∈ {1, . . . , N}):
Ai
1 − z∞iz−1
+
A∗
i
1 − z∗
∞iz−1
=
2 ℜ{Ai} − 2 ℜ{Ai z∗
∞i}z−1
1 − 2 ℜ{z∞i}z−1 + |z∞i|2z−2
=:
b0 + b1 z−1
1 + a1 z−1 + a2 z−2
(4.22)
80

• Two real-valued poles can also be combined to a second order
transfer function (i, j ∈ {1, . . . , N}):
Ai
1 − z∞iz−1
+
Aj
1 − z∞jz−1
=
(Ai + Aj) − (Ai z∞j + Aj z∞i) z−1
1 − (z∞j + z∞i) z−1 + (z∞j z∞i) z−2
=:
b0 + b1 z−1
1 + a1 z−1 + a2 z−2
(4.23)
• If N is odd, there is one real-valued pole left, which leads to
one first order partial fraction (see example).
Parallel structure:
H0
H1(z)
H2(z)
HP (z)
P: number of parallel subsystems
V(z)
Y(z)
Signal flow graph of a second order section:
z−1
−ap1
−ap2
z−1
v(n)
bp1
bp0
yp(n)
p = 1, . . . , P
81
Example:
Consider again the 5-th order Chebyshev lowpass filter with the transfer
function from (4.19). The partial fraction expansion can be given as
H(z) = −0.2607 +
A1
1 − z∞1 z−1
+
A∗
1
1 − z∗
∞1 z−1
+
+
A3
1 − z∞3 z−1
+
A∗
3
1 − z∗
∞3 z−1
+
A5
1 − z∞5 z−1
with the poles from (4.20) and the residues
z∞1 = −0.0336 + j 0.8821, A1 = 0.1 + j 0.0941,
z∞3 = 0.219 + j 0.5804, A3 = −0.5533 + j 0.00926,
z∞5 = 0.4114, A5 = 1.1996.
With (4.22) the resulting transfer function writes
H(z) = −0.2607 +
0.2 − 0.1592 z−1
1 + 0.0673 z−1 + 0.7793 z−2
+
−1.1066 + 0.3498 z−1
1 − 0.4379 z−1 + 0.3849 z−2
+
1.1996
1 − 0.4114 z−1
.
Resulting signal flow graph (V → U):
(from [Fliege: ”Analoge und digitale Filter”, Hamburg University of Technology, 1990])
82

4.3 Coefficient quantization and round-off effects
In this section we discuss the effects of a fixed-point digital filter
implementation on the system performance.
4.3.1 Errors resulting from rounding and truncation
Number representation in fixed-point format:
A real number v can be represented as
v = [β−A, . . . , β−1, β0, β1, . . . , βB] =
B
X
ℓ=−A
βℓr
−ℓ
,
(4.24)
where βℓ is the digit, r is the radix (base), A the number of integer
digits, B the number of fractional digits.
Example: [101.01]2 = 1·22
+0·21
+1·20
+0·2−1
+1·2−2
Most important in digital signal processing:
• Binary representation with r =2 and βℓ ∈{0, 1}, β−A MSB,
βB LSB.
• b-bit fraction format: A=0, B =b−1, binary point between
β0 and β1 → numbers between 0 and 2−2−b+1
are possible.
Positive numbers are represented as
v = [0.β1β2 . . . βb−1] =
b−1
X
ℓ=1
βℓ2
−ℓ
. (4.25)
Negative fraction:
v = [−0.β1β2 . . . βb−1] = −
b−1
X
ℓ=1
βℓ2
−ℓ
, (4.26)
83
can be represented with one of the three following formats
• Sign-magnitude format:
vSM = [1.β1β2 . . . βb−1] for v 0. (4.27)
• One’s-complement format:
v1C = [1.β̄1β̄2 . . . β̄b−1] for v 0, (4.28)
with β̄ℓ = 1 − βℓ denoting the one’s complement of βℓ.
Alternative definition:
v1C = 1·2
0
+
b−1
X
ℓ=1
(1−βℓ)·2
−ℓ
= 2−2
−b+1
−|v| (4.29)
• Two’s-complement format:
v2C = [1.β̄1β̄2 . . . β̄b−1] + [00 . . . 01] for v 0,
(4.30)
where + denotes a binary addition. We thus have by using
(4.29)
v2C = v1C + 2
−b+1
= 2 − |v|. (4.31)
Does (4.30) really represent a negative number? Using the
identity
1 =
b−1
X
ℓ=1
2
−ℓ
+ 2
−b+1
84

we can express a negative number as
v = −
b−1
X
ℓ=1
βℓ2
−ℓ
+ 1 − 1
= −1 +
b−1
X
ℓ=1
(1 − βℓ)
| {z }
= β̄ℓ
2
−ℓ
+ 2
−b+1
= v2C − 2.
Example:
Express the fractions 7/8 and −7/8 in sign-magnitude, two’s-complement,
and one’s-complement format.
v =7/8 can be represented as 2−1
+ 2−2
+ 2−3
, such that v =[0.111]. In
sign-magnitude format, v =−7/8 is represented as vSM =[1.111], in one’s
complement we have v1C = [1.000], and in two’s-complement the result is
v2C =[1.000] + [0.001]=[1.001].
(For further examples see also the table in Section 2.4.)
Remarks:
• Most DSPs use two’s-complement arithmetic. Thus any b-bit
number v has the number range
v ∈ {−1, −1 + 2
−b+1
, . . . , 1 − 2
−b+1
}.
• Two’s-complement arithmetic with b bits can be viewed as
arithmetic modulo 2b
(Example for b = 3):
85
• Important property: If the sum of numbers is within the range
it will be computed correctly, even if individual partial sums
result in overflow.
Truncation and rounding:
Problem: Multiplication of two b-bit numbers yields a result of
length (2b−1) → truncation/rounding necessary → can again be
regarded as quantization of the (filter) coefficient v
Suppose that we have a fixed-point realization in which a number
v is quantized from bu to b bits.
We first discuss the truncation case. Let the truncation error be
defined as Et = Qt[v] − v.
• For positive numbers the error is
−(2
−b+1
− 2
−bu+1
) ≤ Et ≤ 0
(truncation leads to a number smaller than the unquantized
number)
• For negative numbers and the sign-magnitude representation
the error is
0 ≤ Et ≤ (2
−b+1
− 2
−bu+1
).
(truncation reduces the magnitude of the number)
• For negative numbers in the two’s-complement case the error
is
−(2
−b+1
− 2
−bu+1
) ≤ Et ≤ 0
(negative of a number is obtained by subtracting the
corresponding positive number from 2, see (4.31))
86

• Quantization characteristic functions for a continuous input
signal v (v → x):
Sign-magnitude representation:
+1
+1
Two’s-complement representation:
+1
+1
Rounding case, rounding error is defined as Er = Qr[v] − v:
• Rounding affects only the magnitude of the number and is thus
independent from the type of fixed-point realization.
• Rounding error is symmetric around zero and falls in the range
−
1
2
(2
−b+1
− 2
−bu+1
) ≤ Er ≤
1
2
(2
−b+1
− 2
−bu+1
).
• Quantization characteristic function, bu = ∞ (v → x):
+1
+1
87
4.3.2 Numerical overflow
If a number is larger/smaller than the maximal/minimal possible
number representation
• ±(1 − 2−b+1
) for sign magnitude and ones’s-complement
arithmetic;
• −1 and 1 − 2−b+1
, resp., for two’s-complement arithmetic,
we speak of an overflow/underflow condition.
Overflow example in two’s-complement arithmetic
(range: −8, . . . , 7)
[0.111]
| {z }
7
+ [0.001]
| {z }
1
= [1.000]
| {z }
−8
⇒ resulting error can be very large when overflow/underflow
occurs
Two’s-complement quantizer for b=3, ∆ = 2−b
(v → x):
Alternative: saturation or clipping, error does not increase
abruptly in magnitude when overflow/underflow occurs:
88

Disadvantage: ”Summation property” of the two’s-complement
representation is violated
4.3.3 Coefficient quantization errors
• In a DSP/hardware realization of an FIR/IIR filter the accuracy
is limited by the wordlength of the computer ⇒ Coefficients
obtained from a design algorithm have to be quantized
• Wordlength reduction of the coefficients leads to different
poles and zeros compared to the desired ones. This may lead
to
– modified frequency response with decreased selectivity
– stability problems
Sensitivity to quantization of filter coefficients
Direct form realization, quantized coefficients:
āi = ai + ∆ai, i = 1, . . . , N,
b̄i = bi + ∆bi, i = 0, . . . , N,
∆ai and ∆bi represent the quantization errors.
89
As an example, we are interested in the deviation
∆z∞i = z∞i − z̄∞i, when the denominator coefficients ai
are quantized (z̄∞i denotes the resulting pole after quantization).
It can be shown [Proakis, Manolakis, 1996, pp. 569] that this
deviation can be expressed as:
∆z∞i = −
N
X
n=1
zN−n
∞i
N
Q
ℓ=1, ℓ6=i
(z∞i − z∞ℓ)
∆an, i = 1, . . . , N.
(4.32)
From (4.32) we can observe the following:
• By using the direct form, each single pole deviation ∆z∞i
depends on all quantized denominator coefficients āi.
• The error ∆z∞i can be minimized by maximizing the distance
|z∞i − z∞ℓ| between the poles z∞i and z∞ℓ.
⇒ Splitting the filter into single or double pole sections (first or
second order transfer functions):
• Combining the poles z∞i and z∗
∞i into a second order section
leads to a small perturbation error ∆z∞i, since complex
conjugate poles are normally sufficiently far apart.
• ⇒ Realization in cascade or parallel form: The error of a
particular pole pair z∞i and z∗
∞i is independent of its distance
from the other poles of the transfer function.
Example: Effects of coefficient quantization
Elliptic filter of order N = 12 (Example taken from [Oppenheim, Schafer,
1999]):
90

Unquantized:
(a) Magnitude frequency response 20 · log10 |H(ejω
)|
(b) Passband details
Quantized with b=16 bits:
(c) Passband details for cascade structure
(d) Passband details for parallel structure
(e) Magnitude frequency response (log) for direct structure
Pole locations of quantized second order sections
Consider a two-pole filter with the transfer function
H(z) =
1
1 − (2r cos θ) z−1 + r2 z−2
.
91
Poles: z∞1,2 = r e±jθ
,
Coefficients:
a1 = −2r cos θ, a2 = r2
,
Stability condition: |r| ≤ 1
2r cos θ
−r2
z−1
v(n) y(n)
z−1
Quantization of a1 and a2 with b = 4 bits: → possible pole
positions:
• Nonuniformity of the pole position is due to the fact that
a2 =r2
is quantized, while the pole locations z∞1,2 =r e±jθ
are proportional r.
• Sparse set of possible pole locations around θ =0 and θ =π.
Disadvantage for realizing lowpass filters where the poles are
normally clustered near θ =0 and θ =π.
Alternative: Coupled-form realization
y1(n) = v(n) + r cos θ y1(n − 1) − r sin θ y(n − 1),
y(n) = r sin θ y1(n − 1) + r cos θ y(n − 1),
(4.33)
92

which corresponds to the following signal flow graph:
v(n)
z−1
z−1
−r sin θ r sin θ
r cos θ
r cos θ
y(n)
y1(n)
By transforming (4.33) into the z-domain, the transfer function of
the filter can be obtained as
H(z) =
Y (z)
V (z)
=
(r sin θ) z−1
1 − (2r cos θ) z−1 + r2 z−2
.
• We can see from the signal flow graph that the two coefficients
r sin θ and r cos θ are now linear in r, such that a
quantization of these parameters lead to equally spaced pole
locations in the z-plane:
93
• Disadvantage: Increased computational complexity compared
to the direct form.
Cascade or parallel form?
Cascade form: H(z) =
P
Y
p=1
bp0 + bp1 z−1
+ bp2 z−2
1 + ap1 z−1 + ap2 z−2
Parallel form: H(z) = A0 +
P
X
p=1
cp0 + cp1 z−1
1 + ap1 z−1 + ap2 z−2
• Cascade form: Only the numerator coefficients bpi of
an individual section determine the perturbation of the
corresponding zero locations (an equation similar to (4.32) can
be derived) → direct control over the poles and zeros
• Parallel form: A particular zero is affected by quantization
errors in the numerator and denominator coefficients of all
individual sections → numerator coefficients cp0 and cp1 do
not specify the position of a zero directly, direct control over
the poles only
⇒ Cascade structure is more robust against coefficient
quantization and should be used in most cases
Example:
Elliptic filter of order N =7, frequency and phase response
(taken from [Proakis, Manolakis, 1996]):
94

Cascade form (3 digits b
= b≈10 bits, 4 digits b
= b≈14 bits)
Parallel form (b=10 bits)
95
Coefficient quantization in FIR systems
In FIR systems we only have to deal with the locations of the
zeros, since for causal filters all poles are at z =0.
Remarks:
• For FIR filters an expression analogous to (4.32) can be
derived for the zeros ⇒ FIR filters should also be realized
in cascade form according to
H(z) = H0
P
Y
p=1
(1 + bp1 z
−1
+ bp2 z
−2
)
with second order subsections, in order to limit the effects of
coefficient quantization to the zeros of the actual subsection
only.
• However, since the zeros are more or less uniformly spread in
the z-plane, in many cases the direct form is also used with
quantized coefficients.
• For a linear-phase filter with the symmetry (4.8) in the
impulse response, quantization does not affect the phase
characteristics, but only the magnitude.
4.3.4 Round-off effects
Direct-form I IIR structure
Starting point for the analysis: Direct-form I structure with the
difference equation
y(n) =
N
X
i=0
bi v(n − i) −
N
X
i=1
ai y(n − i).
96

All signal values and coefficients are represented by b-bit
binary fixed-point numbers (for example in two’s-complement
representation):
• truncation or rounding of the (2b − 1)-bit products to b bit
necessary
• modelling as a constant multiplication followed by a quantizer
−
−
−
−
−
−
−
(from [Oppenheim, Schafer, 1999], v → x)
This can be described with the difference equation
y(n) =
N
X
i=0
Q[bi v(n − i)] −
N
X
i=1
Q[ai y(n − i)]. (4.34)
97
As already stated in Section 2.4 the result of each single
quantization stage can be modeled by adding a noise source ei(n)
with the following properties:
• Each ei(n) corresponds to a wide-sense-stationary white-
noise process.
• Each ei(n) has an uniform distribution of amplitudes over one
quantization interval (uniform p.d.f.).
• Each ei(n) is uncorrelated with the quantizer input, all other
quantization noise sources and the input signal of the filter.
We have shown above that for b-bit quantization the rounding
error falls in the range
−
1
2
(2
−b+1
) ≤ ei(n) ≤
1
2
(2
−b+1
),
and for two’s-complement truncation we have
−2
−b+1
≤ ei(n) ≤ 0.
Mean and variance for rounding
µe = 0, σ
2
e =
2−2b+2
12
, (4.35)
and for truncation
µe = −
2−b+1
2
, σ
2
e =
2−2b+2
12
. (4.36)
Autocorrelation (white noise process):
ϕee(n) = σ
2
eδ(n) + µ
2
e. (4.37)
98

In the following we will restrict ourselves to the rounding case,
where ϕee(n) = σ2
eδ(n) and thus, for the power spectral density
we have Φ(ejω
) = σ2
e:
• The following structure can be obtained by summing up all the
noise sources:
e(n) =
4
X
i=0
ei(n),
−
−
• ⇒ Overall noise variance in the special case from above:
σ
2
e =
4
X
i=0
σ
2
ei
= 5 ·
2−2b+2
12
Overall noise variance in the general case:
σ
2
e = (2N + 1) ·
2−2b+2
12
(4.38)
• Due to linearity the output of the whole filter is ŷ(n) =
y(n) + f(n). Thus, the difference equation for the
99
quantization noise e(n) now is given as
f(n) =
N
X
i=1
aif(n − i) + e(n), (4.39)
since e(n) can be regarded as the input to an all-pole system
with output f(n).
• Suppose that the allpole-filter has the transfer function
Hef (z) with
Hef (z) =
1
D(z)
, H(z) =
N(z)
D(z)
analog to (4.1).
– Mean of f(n): µf = µe Hef (ej0
) = 0 for rounding.
– Power spectral density (power spectrum)
Φff (ejω
) = σ2
e |Hef (ejω
)|2
.
– Variance σ2
f = F−1
∗ {Φff (ejω
)}|n=0:
σ
2
f =
σ2
e
2π
Z π
−π
|Hef (e
jω
)|
2
dω = σ
2
e
∞
X
n=−∞
|hef (n)|
2
,
(4.40)
where the last expression is obtained by applying Parseval’s
theorem.
By combining (4.38) and (4.40) we can now state the total output
100

variance due to internal round-off as
σ
2
f = (2N +1)
2−2b+2
2π · 12
π
Z
−π
dω
|D(ejω)|2
= (2N +1)
2−2b+2
12
∞
X
n=−∞
|hef (n)|
2
(4.41)
Round-off noise in a first-order system
Given: Transfer function
H(z) =
b0
1 − a1z−1
, |a1| 1
→ Impulse response: h(n) = b0 an
1 u(n) (u(n): unit step
function)
a1
y(n)
z−1
b0
v(n)
e(n) = ea1
(n) + eb0
(n)
Considering (4.41) and the two error sources ea1
(n), eb0
(n), we
have
σ
2
f = 2
2−2b+2
12
∞
X
n=0
|a1|
2n
=
2−2b+2
6

1
1 − |a1|2

.
(4.42)
The output noise variance increases when the pole z∞ = a1
101
approaches the unit circle ⇒ In order to neep the noise variance
below a certain level, the wordlength b has to be increased.
Round-off noise in a second-order system
Second-order direct form I system:
H(z) =
b0 + b1 z−1
+ b2 z−2
(1 − r ejθ z−1)(1 − r e−jθ z−1)
Thus we have
σ
2
f = 5
2−2b+2
12
1
2π
π
Z
−π
dω
|(1 − r ejθ e−jω)|2 |(1 − r e−jθ e−jω)|2
.
With a1 = −2r cos θ and a2 = r2
it can be shown via a partial
fraction expansion that
σ
2
f = 5
2−2b+2
12
1 + r2
1 − r2
!
1
r4 + 1 − 2r2 cos(2θ)
. (4.43)
As in the first-order case we can see that the total variance
increases if the poles approach the unit circle (r → 1).
Direct-form II structure
In this case, the nonlinear difference equations are of the form
w(n) = −
N
X
i=1
Q[ai w(n − i)] + v(n),
y(n) =
N
X
i=0
Q[bi w(n − i)].
(4.44)
102

Signal flow graph:
−
−
−
−
For rounding (µe = 0) the power spectrum of the output noise is:
Φff (e
jω
) = N
2−2b+2
12
|H(e
jω
)|
2
+ (N + 1)
2−2b+2
12
.
(4.45)
We then have:
σ
2
f = N
2−2b+2
12
1
2π
π
Z
−π
|H(e
jω
)|
2
dω + (N + 1)
2−2b+2
12
,
(4.46)
103
and by applying a relation similar to (4.40)
σ
2
f = N
2−2b+2
12
∞
X
n=−∞
|h(n)|
2
+ (N + 1)
2−2b+2
12
.
(4.47)
• White noise produced in implementing the poles is filtered
by the entire system, whereas the white noise produced in
implementing the zeros is added directly to the output of the
system.
• A comparison with (4.41) for the direct form I structure shows
that both structures are affected differently by the quantization
of products.
4.3.5 Zero-input limit cycles
• Stable IIR filters implemented with infinite-precision
arithmetic: If the excitation becomes zero and remains
zero for n n0 then the output of the filter will decay
asymptotically towards zero.
• Same system implemented with fixed-point arithmetic: Output
may oscillate indefinitely with a periodic pattern while the
input remains equal to zero: ⇒ Zero-input limit cycle
behavior, due to nonlinear quantizers in the feedback loop or
overflow of additions.
No general treatment, but two examples are given
Limit cycles due to round-off and truncation
Given: First-order system with the difference equation
y(n) = a y(n − 1) + v(n), |a| 1.
104

Register length for storing a and the intermediate results: 4 bits
(sign bit plus 3 fractional digits) ⇒ product a y(n − 1) must be
rounded or truncated to 4 bits, before being added to v(n)
Signal flow graphs:
a
z−1
v(n) y(n)
z−1
v(n)
a
Q[·]
ŷ(n)
Infinite-precision linear system
Nonlinear system due to quantization
Nonlinear difference equation (Q[·] represents two’s-complement
rounding):
ŷ(n) = Q[a ŷ(n − 1)] + v(n).
Suppose we have a = 1/2 = [0.100], v(n) = 7/8 δ(n) =
[0.111] δ(n):
ŷ(0) = 7/8 = [0.111]
ŷ(1) = Q[a ŷ(0)] = Q [ [0.100] × [0.111] ]
= Q [ [0.011100] ] = Q[7/16] = [0.100] = 1/2
ŷ(2) = Q[a ŷ(1)] = [0.010] = 1/4
ŷ(3) = Q[a ŷ(2)] = [0.001] = 1/8
ŷ(4) = Q[a ŷ(3)] = Q [ [0.000100] ] = [0.001] = 1/8
105
⇒ A constant steady value is obtained for n ≥ 3.
For a = −1/2 we have a periodic steady-state oscillation
between −1/8 and 1/8.
⇒ Such periodic outputs are called limit cycles.
Limit cycles due to overflow
Consider a second-order system realized by the difference
equation
ŷ(n) = v(n) + Q [a1 ŷ(n − 1)] + Q [a2 ŷ(n − 2)] ,
(4.48)
Q[·] represents two’s-complement rounding with one sign and
3 fractional digits.
106

Overflow can occur with the two’s-complement addition of the
products in (4.48).
Suppose that a1 = 3/4 = [0.110], a2 = −3/4 = [1.010],
ŷ(−1) = 3/4 = [0.110], ŷ(−2) = −3/4 = [1.010],
v(n)=0 for all n≥0:
ŷ(0) = Q[ [0.110] × [0.110] ] + Q[ [1.010] × [1.010] ]
= Q[ [0.100100] ] + Q[ [0.100100] ] = [0.101] + [0.101]
= [1.010] = −3/4
ŷ(1) = [1.011] + [1.011] = [0.110] = 3/4
⇒ ŷ(n) continues to oscillate unless an input is applied.
Remarks
Some solutions for avoiding limit cycles:
• Use of structures which do not support limit-cycle oscillations.
• Increasing the computational wordlength.
• Use of a double-length accumulator and quantization after the
accumulation of products.
FIR-filters are limit-cycle free since there is no feedback involved
in its signal flow graph.
107
4.4 Design of FIR filters
General remarks (IIR and FIR filters)
• Ideal filters are noncausal, and thus physically unrealizable for
real-time signal processing applications.
• Causality implies that the a filter frequency response H(ejω
)
cannot have an infinitely sharp cutoff from passband to
stopband, and that the stopband amplification can only be zero
for a finite number of frequencies ω.
Magnitude characteristics of physically realizable filters:
δ1: passband ripple, δ2: stopband ripple,
ωp: passband edge frequency, ωs: stopband edge frequency
Filter design problem:
• Specify δ1, δ2, ωp, and ωs corresponding to the desired
application.
• Select the coefficients ai and bi (free parameters), such that
the resulting frequency response H(ejω
) best satisfies the
requirements for δ1, δ2, ωp, and ωs.
• The degree to which H(ejω
) approximates the specifications
depends on the criterion used for selecting the ai and bi and
108

also on the numerator and denominator degree N (number of
coefficients)
4.4.1 Linear phase filters
Important class of FIR filters, which we will mainly consider in
the following. Definition: A filter is said to be a linear-phase
filter, if its impulse response satisfies the condition (L=N +1):
h(n) = ±h(L − 1 − n). (4.49)
With the definition S := (L − 1)/2 and for L odd this leads to
a z-transform
H(z) =
L−1
X
n=0
h(n)z
−n
(4.50)
= z
−S

h(S) +
S−1
X
n=0
h(n)

z
(S−n)
± z
−(S−n)

#
,
(4.51)
for L even we have
H(z) = z
−S
L/2−1
X
n=0
h(n)

z
(S−n)
± z
−(S−n)

. (4.52)
When we now substitute z−1
for z in (4.50) and multiply both
sides by z−(L−1)
we obtain with (4.49)
z
−(L−1)
H(z
−1
) = ±H(z), (4.53)
109
which is the z-transform equivalent to (4.49). Consequences:
• The roots of the polynomial H(z) are identical to the roots of
the polynomial H(z−1
): If z0i is a zero of H(z) then z−1
0i is
also a zero.
• If additionally the impulse response h(n) is real-valued, the
roots must occur in complex-conjugate pairs: If z0i is a zero
of H(z) then z∗
0i is also a zero.
⇒ The zeros of a real-valued linear-phase filter occur in
quadruples in the z-plane (exception: zeros on the real axis, zeros
on the unit circle)
(a) Type-1 linear phase system
Definition: Odd length L, even symmetry h(n)=h(L−1−n)
Frequency response from (4.51):
H(e
jω
) = e
−jSω

h(S) + 2
S−1
X
n=0
h(n) cos ((S − n)ω)
#
(4.54)
= e
−jSω
H01(ω)
110

H01(ω): amplitude frequency response, real-valued (generally
denoted with H0(ω))
• linear phase ϕ1(ω) = −arg H(ejω
) = Sω
0 5 10 15 20 25 30
0
type 1, odd length (L=31)
n
h(n)
0 0.5 1 1.5 2 2.5 3 3.5 4
0
0.5
1
type 1, amplitude frequency response
ω / π
H
01
(ω)
(b) Type-3 linear phase system
Definition: Odd length L, odd symmetry h(n)=−h(L−1−n)
H(e
jω
) = e
−jSω
j

h(S) + 2
S−1
X
n=0
h(n) sin ((S − n)ω)
#
(4.55)
= e
−jSω
j H03(ω) = e
−jSω+jπ/2
H03(ω)
) = Sω − π/2
• H(ej0
) = 0, S ∈ I
N ⇒ H(ejπ
) = 0
111
0 5 10 15 20 25 30
−0.5
0
0.5
type 3, odd length (L=31)
n
h(n)
0 0.5 1 1.5 2 2.5 3 3.5 4
−1
−0.5
0
0.5
1
ω / π
H
03
(ω)
(c) Type-2 linear phase system
Definition: Even length L, even symmetry h(n)=h(L−1−n)
H(e
jω
) = e
−jSω

2
L/2−1
X
n=0

 (4.56)
= e
−jSω
H02(ω)
) = Sω
• S = (2λ−1)/2, λ ∈ I
N ⇒ H(ejπ
) = 0
112

• Amplitude frequency response has 4π-periodicity:
H02(ω) = 2
L/2−1
X
n=0
H02(ω + 2π) = 2
L/2−1
X
n=0
h(n) cos ((S − n)(ω + 2π))
= 2
L/2−1
X
n=0
h(n) cos ((S − n)ω) ·
· cos ((S − n)2π)
| {z }
= cos((L−1−2n)π) = −1
= −H02(ω)
• Remark: H(ejω
) is 2π-periodic again due to the e−jSω
phase
factor
0 5 10 15 20 25 30
0
type 2, even length (L=30)
n
h(n)
0 0.5 1 1.5 2 2.5 3 3.5 4
−1
−0.5
0
0.5
1
ω / π
H
02
(ω)
113
(d) Type-4 linear phase system
Definition: Even length L, odd symmetry h(n)=−h(L−1−n)
H(e
jω
) = e
−jSω
j

2
L/2−1
X
n=0
h(n) sin ((S − n)ω)


(4.57)
= e
−jSω
j H04(ω) = e
−jSω+jπ/2
H04(ω)
) = Sω − π/2
• H(ej0
) = 0
• Similar to the type-2 filter the amplitude frequency response
has 4π-periodicity: H04(ω + 2π) = −H04(ω). The
multiplication with the exponential factor e−jSω+jπ/2
then
again leads to 2π-periodicity for H(ejω
).
0 5 10 15 20 25 30
−0.5
0
0.5
type 4, even length (L=30)
n
h(n) 0 0.5 1 1.5 2 2.5 3 3.5 4
−1
−0.5
0
0.5
1
ω / π
H
04
(ω)
114

Applications:
• Type-1 and Type-2 filters used for ”ordinary” filtering,
however type 2 filters are unsuitable for highpass filtering
• Type-3 and Type-4 filters for example used for 90 degree phase
shifters and so called Hilbert transformers
4.4.2 Design of linear-phase filters using a window function
Given: Desired frequency response
Hd(e
jω
) =
∞
X
n=−∞
hd(n) e
−jωn
(4.58)
⇒ Impulse response hd(n) can be obtained with the inverse
Fourier-Transform
hd(n) =
1
2π
π
Z
−π
Hd(e
jω
) e
jωn
dω. (4.59)
Impulse response has generally infinite length ⇒ truncation
to length L by multiplication with a window function w(n)
necessary: h(n) = hd(n) · w(n).
Rectangular window:
w(n) =
(
1 n = 0, . . . , L − 1
0 otherwise
⇒ h(n) =
(
hd(n) n = 0, . . . , L − 1
0 otherwise
Frequency response of the rectangular window: see (3.12)
Suppose, we want to design a linear-phase lowpass filter of length
115
L with the desired frequency response
Hd(e
jω
) =
(
e−jω(L−1)/2
for 0 ≤ |ω| ωc,
0 otherwise,
(4.60)
where ωc is denoting the cut-off frequency.
Corresponding unit sample response:
hd(n) =
1
2π
ωc
Z
−ωc
e
jω(n−(L−1)/2)
dω.
=
sin

ωc n − L−1
2

π n − L−1
2
, n 6=
L − 1
2
(4.61)
Multiplication with a rectangular window of length L leads to
h(n) =
sin
h
ωc

n − L−1
2
i
π

n − L−1
2
, n 6=
L − 1
2
, n = 0, . . . , L − 1
For L odd: h

L − 1
2

=
ωc
π
Example for ωc =0.2π, L=61 and L=101:
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
1.2
|H(e
jω
)|
→
ω / π →
L=61, ωc
=0.2⋅π
0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
1.2
|H(e
jω
)|
→
ω / π →
L=101, ωc
=0.2⋅π
Disadvantage of using an rectangular window:
116

Large sidelobes lead to in undesirable ringing effects (overshoot
at the boundary between pass- and stopband) in the frequency
response of the resulting FIR filter
⇒ Gibbs phenomenon:
• Result of approximating a discontinuity in the frequency
response with a finite number of filter coefficients and a mean
square error criterion
• Eq. (4.58) can be interpreted as a Fourier series representation
for Hd(ejω
) with the Fourier coefficients hd(n) → Gibbs
phenomenon resulting from a Fourier series approximation
• Squared integral error E =
π
R
−π
(Hd(ejω
) − H(ejω
))2
dω
approaches zero with increasing length of h(n). However, the
maximum value of the error |Hd(ejω
)−H(ejω
)| approaches
a constant value.
⇒ Use of other appropriate window functions with lower
sidelobes in their frequency responses
117
(n → n, L → M, I0: Bessel function of the first kind of order zero)
Frequency-domain characteristics of some window functions
(taken from [Proakis, Manolakis, 1996]):
Approximate transition Peak
Type of window width of main lobe sidelobe [dB]
Rectangular 4π/L -13
Bartlett 8π/L -27
Hanning 8π/L -32
Hamming 8π/L -43
Blackman 12π/L -58
Parameter α in the Kaiser window allows to adjust the width of
the main lobe, and thus also to adjust the compromise between
overshoot reduction and increased transition bandwidth in the
118

resulting FIR filter.
Magnitude frequency response 20 log10 |H(ejω
)| of the
resulting linear-phase FIR filter, when different window functions
are used to truncate the infinite-length impulse response from
(4.61) (desired frequency response Hd(ejω
) in (4.60)):
0 0.2 0.4 0.6 0.8 1
−100
−80
−60
−40
−20
0
20
log
10
|H(e
jω
)|
→
ω / π →
Rectangular window
L=61, ω
c
=0.2⋅π
0 0.2 0.4 0.6 0.8 1
−100
−80
−60
−40
−20
0
20
log
10
|H(e
jω
)|
→
ω / π →
Hamming window
L=61, ω
c
=0.2⋅π
0 0.2 0.4 0.6 0.8 1
−100
−80
−60
−40
−20
0
20
log
10
|H(e
jω
)|
→
ω / π →
Blackman window
L=61, ωc
=0.2⋅π
0 0.2 0.4 0.6 0.8 1
−100
−80
−60
−40
−20
0
20
log
10
|H(e
jω
)|
→
ω / π →
Kaiser window, β=4
L=61, ωc
=0.2⋅π
MATLAB-command for windowed linear-phase FIR design:
fir1
4.4.3 Frequency sampling design
Desired frequency response Hd(ejω
) is specified as a set of
equally spaced frequencies:
ωk =
2π
L
(k + α), k = 0, 1, . . . , L − 1, α ∈

0,
1
2

.
(4.62)
119
Frequency response of an FIR filter (requirement):
Hd(e
jω
)
!
=
L−1
X
n=0
h(n)e
−jωn
With (4.62) we obtain for k = 0, . . . , L−1:
Hd(e
j 2π
L
(k+α)
) =
L−1
X
n=0
h(n)e
−j2π(k+α)n/L
. (4.63)
Multiplication of (4.63) with ej2πkℓ/L
, ℓ = 0, . . . , L−1 and
summation over k = 0, . . . , L−1 yields to
L−1
X
k=0
e
j2πkℓ/L
Hd(e
j 2π
L
(k+α)
) =
=
L−1
X
k=0
L−1
X
n=0
h(n)e
−j2π(k+α)n/L
e
j2πkℓ/L
=
L−1
X
n=0
h(n)e
−j2παn/L
L−1
X
k=0
e
−j2π(n−ℓ)k/L
= L h(ℓ) e
−j2παℓ/L
.
Thus, the impulse response h(n) can be obtained from the
sampled desired frequency response as (n = 0, . . . , L−1):
h(n) =
1
L
L−1
X
k=0
Hd(e
j 2π
L
(k+α)
) e
j2π(k+α)n/L
(4.64)
120

Remarks:
• For α = 0 (4.64) is identical to the IDFT ⇒ fast evaluation
with IFFT
• In general Hd(ej 2π
L
(k+α)
) has to be specified in amplitude
and phase for every k.
Exception: Type 1/3 linear-phase filters, where Hd(·) can be
real-valued with an additional shift of the obtained impulse
response (see below).
If h(n) is a real-valued sequence, the frequency response and thus
also its sampled version have the Hermitian symmetry
Hd(e
j 2π
L
(k+α)
) = H
∗
d(e
j 2π
L
(L−k−α)
).
⇒ The number of frequency specifications can be reduced, and
(4.62) becomes
ωk =
2π
L
(k + α),







k = 0, 1, . . . , L−1
2 L odd,
k = 0, 1, . . . , L
2 − 1 L even,
α ∈

0, 1
2 .
(4.65)
Linear-phase FIR filters:
• Symmetry in h(n) can be additionally exploited such that
(4.64) only has to be evaluated for n = 0, . . . , L
2 − 1 for L
even, and n = 0, . . . , L−1
2 for L odd, resp.
• Linear-phase property may be included by specifying real-
valued frequency response samples Hd(ej 2π
L
(k+α)
) →
Application of (4.64) leads to a zero-phase response which has
to be shifted to the right by L−1
2 samples.
121
Example:
Determine the coefficients of a type 2 linear-phase filter with length L = 32,
and the desired sampled frequency response
Hd(e
j 2π
L
(k+α)
) = e
−j L−1
2
2π
L
(k+α)
·





1 k = 0, 1, . . . , 5,
T1 k = 6,
0 k = 7, 8, . . . , 15.
The parameter T1 is responsible for the transition band behavior and is
obtained via numerical optimization in order to reduce the magnitude of the
sidelobes. The corresponding values are tabulated in the literature ([Rabiner,
Gold, McGonegal, 1970], [Proakis, Manolakis, 1996]).
For L=32 we obtain T1 =0.3789795 for α=0, and T1 =0.3570496 for
α=1/2.
Coefficients of the impulse response h(n):
122

Magnitude frequency responses (20 log10 |H(ejω
)|):
L=32, α=0: L=32, α=0.5:
MATLAB-command for the frequency sampling design of
linear-phase FIR filters: fir2
4.4.4 Optimum equiripple design (Chebyshev approximation)
• Windowing design techniques (section 4.4.2) try to reduce
the difference between the desired and the actual frequency
response (error function) by choosing suitable windows
• How far can the maximum error be reduced?
⇒ Theory of Chebyshev approximation answers this question
and provides us with algorithms to find the coefficients
of linear-phase FIR filters, where the maximum frequency
response error is minimized
• Chebyshev approximation: Approximation that minimizes the
maximum errors over a set of frequencies
• Resulting filters exhibit an equiripple behavior in their
frequency responses ⇒ equiripple filters
123
Linear-phase filters revisited
As we have shown in section 4.4.1, every linear-phase filter has a
frequency response of the form
H(e
jω
) = (j)
m
·A(ω)·e
−jω(L−1)/2
, m ∈ {0, 1}, (4.66)
where A(ω) is a real-valued positive or negative function (ampli-
tude frequency response) (cmp. (4.54), (4.55), (4.56), (4.57)).
It can be shown that for all types of linear-phase symmetry A(ω)
can always be written as a weighted sum of cosines. For example
for type 1 linear-phase filters we have
A(ω) =
(L−1)/2
X
n=0
an cos(nω) (4.67)
with a0 =h

L − 1
2

, an =2 h

− n +
L − 1
2

. (4.68)
Remaining linear-phase filters (ω → 2πf):
(from [Parks, Burrus: Digital Filter Design, 1987])
124

Problem definition
Acceptable frequency response for the resulting FIR filter:
• Linear phase,
• transition bandwidth ∆ω between pass- and stopband,
• passband deviation ±δ1 from unity,
• stopband deviation ±δ2 from zero.
(Multiple bands are possible as well.)
We will restrict ourselves in the following to lowpass type 1 linear-
phase filters.
Approximation problem: Given
• a compact subset F of [0, π] in the frequency domain
(consisting of pass- and stopband in the lowpass filter case),
• a desired real-valued frequency response D(ω), defined on
F,
• a positive weight function W (ω), defined on F, and
• the form of A(ω), here (type-1 linear-phase)
A(ω) =
Pr−1
n=0 an cos(nω).
Goal: Minimization of the error
Emax = max
ω∈F
W (ω) · |D(ω) − A(ω)| (4.69)
over an by the choice of A(ω).
Alternation theorem (without proof):
If A(ω) is a linear combination of r cosine functions,
A(ω) =
r−1
X
n=0
an cos(nω),
125
then a necessary and sufficient condition that A(ω) be the
unique, best weighted Chebyshev approximation to a given
continuous function D(ω) on F is that the weighted error
function E(ω) = W (ω) · (D(ω) − A(ω)) exhibit at least
r + 1 extremal frequencies in F. These frequencies are points
for which ω1 · · · ωr ωr+1,
E(ωm) = −E(ωm+1), m = 1, 2, . . . , r,
and
|E(ωi)| = max
ω∈F
E(ω), i = 1, . . . , r + 1.
• Consequences from the alternation theorem: Best Chebyshev
approximation must have an equiripple error function E(ω),
and is unique.
• Example: Amplitude frequency response of an optimum
type 1 linear-phase filter with L=13 → r =7
ω
π
L = 13
ωs
ωp
A(ω) =
(L−1)/2
P
n=0
an cos(nω)
• If the r+1 extremal frequencies were known, we could use the
frequency-sampling design from above to specify the desired
126

values 1 ± δ1 at the extremal frequencies in the passband, and
±δ2 in the stopband, respectively.
How to find the set of extremal frequencies?
Remez exchange algorithm (Parks, McClellan, 1972)
• It can be shown that the error function
E(ω) = D(ω) −
r−1
X
n=0
an cos(nω) (4.70)
can always be forced to take on some values ±δ for any given
set of r+1 frequency points ωi, i = 1, . . . , r + 1.
Simplification: Restriction to W (ω) = 1, leading to δ1 =
δ2 =δ.
This can be written as a set of linear equations according to
D(ωi) =
r−1
X
n=0
an cos(nωi)+(−1)
i
δ, i = 1, . . . , r+1,
(4.71)
from which we obtain a unique solution for the coefficients
an, n = 0, . . . , r−1 and the error magnitude δ.
• In the Remez exchange algorithm {F} is usually chosen as
an equally spaced grid of frequency points with the number of
frequency points being approximately 10 · L. The algorithm
is initialized with a trial set of arbitrarily chosen frequencies
T = {ω1, ω2, . . . , ωr+1}.
• The Remez algorithm now consists of the following basic
computations:
1. Solve the linear equations in (4.71), yielding an error
magnitude δk in the k-th iteration.
127
2. Interpolate to find the frequency response on the entire grid
of frequencies.
3. Search over the entire grid of frequencies for a larger error
magnitude than δk obtained in step 1.
4. Stop, if no larger magnitude error has been found.
Otherwise, take the r + 1 frequencies, where the error
attains its maximum magnitude as a new trial set of
extremal frequencies and go to step 1.
Flowchart (ω → f):
Example:
Choose the two coefficients d0 and d1 such that they minimize the Chebyshev
error
max
x∈[0,1]
|x
2
− (d0 + d1x1)|
(approximation of a parabola by a straight line) → three extremal points →
128

resulting linear equations to be solved:
x
2
i = d0 + d1 xi + (−1)
i
δ, i = 0, 1, 2 (4.72)
1. Arbitrarily chosen trial set: T0 = [0.25, 0.5, 1.0]
Matrix version of the linear equations in (4.72):


1 0.25 1
1 0.5 −1
1 1.0 1




d0
d1
δ0

 =


0.0625
0.25
1.0


→ δ0 = 0.0625
2. Next trial set chosen as those three points, where the error
E(x) = x
2
− (d0 + d1x)
129
achieves its maximum magnitude: → T1 = [0.0, 0.625, 1.0]
Linear equations to solve:


1 0.0 1
1 0.625 −1
1 1.0 1




d0
d1
δ0

 =


0.0
0.390625
1.0


→ δ1 = 0.1171875
3. Next trial set: T2 = [0.0, 0.5, 1.0]
Linear equations to solve:


1 0.0 1
1 0.5 −1
1 1.0 1




d0
d1
δ0

 =


0.0
0.25
1.0


→ δ1 = 0.125 b
= maximum error → T2 is the extremal point set
After the extremal points ωi are found, the coefficients an from
Step 1 of the above algorithm are used to obtain the filter
coefficients with (4.68).
MATLAB-command for optimum equiripple design: remez
Estimation of the filter length
Given the stop- / passband ripple δ1, δ2, and the transition
bandwidth ∆ω = ωs − ωp the necessary filter order N can
be estimated as (Kaiser, 1974)
N =
−10 log10(δ1 δ2) − 13
2.324∆ω
. (4.73)
MATLAB-command for estimating the filter order: remezord
130

Design example
Design a linear-phase lowpass filter with the specifications
δ1 = 0.01, δ2 = 0.001, ωp = 0.4π, ωs = 0.6π.
→ weighting δ2/δ1 = 10 in the stopband.
Inserting these values into (4.73) leads to N ≈ 25.34 and
rounding up (to be on the save side) yields a filter length
L= N +1=27.
Impulse response
0 5 10 15 20 25
−0.2
0
0.2
0.4
0.6
n →
h(n)
→
Magnitude frequency response
0 0.2 0.4 0.6 0.8 1
−100
−80
−60
−40
−20
0
20
log
10
|H(e
jω
)|
[dB]
→
ω / π →
Passband detail
0 0.2 0.4
0.985
0.99
0.995
1
1.005
1.01
|H(e
jω
)|
→
ω / π →
In the passband the specifications are not satisfied → Increasing
131
the filter-length by one, L=28:
Impulse response
0 5 10 15 20 25
−0.2
0
0.2
0.4
0.6
n →
h(n)
→
0 0.2 0.4 0.6 0.8 1
−100
−80
−60
−40
−20
0
20
log
10
|H(e
jω
)|
[dB]
→
ω / π →
Passband detail
0 0.2 0.4
0.985
0.99
0.995
1
1.005
1.01
|H(e
jω
)|
→
ω / π →
4.5 Design of IIR filters
• In the following only design algorithms are discussed which
convert an analog into a digital filter, however, there are also
numerous algorithms for directly designing an IIR filter in the
z-domain (frequency sampling method, least-squares design).
• Why starting point analog filter? Analog filter design is a well
developed field (lots of existing design catalogs). Problem can
be defined in the z-domain, transformed into the s-domain and
solved there, and finally transformed back into the z-domain.
132

• Analog filter: Transfer function
Ha(s) =
N(s)
D(s)
=
M
P
n=0
βn sn
N
P
n=0
αn sn
(4.74)
with the filter coefficients αn, βn, and the filter order N.
Furthermore,
Ha(s) =
∞
Z
−∞
h(t) e
−st
dt (4.75)
(Laplace transform).
• Note that linear-phase designs are not possible for causal and
stable IIR filters, since the condition
H(z) = ±z
−N
H(z
−1
)
has to be satisfied (compare (4.53)) → mirror-image pole
outside the unit-circle for every pole inside the unit-circle →
unstable filter.
4.5.1 Filter design by impulse invariance
Goal: Design an IIR filter with an impulse response h(n) being
the sampled version of the impulse response ha(t) of the analog
filter:
h(n) = ha(nT ), n = 0, 1, 2, . . . ,
where T is the sampling interval.
133
Frequency response (ideal sampling assumed, compare (2.4)):
H(jΩ) =
1
T
∞
X
n=−∞
Ha

jΩ − j
2πn
T

(4.76)
• T should be selected sufficiently small to avoid aliasing.
• Method is not suitable to design highpass filters due to the
large amount of possible aliasing.
Suppose that the poles of the analog filter are distinct. Then the
partial-fraction expansion of Ha(s) writes
Ha(s) =
N
X
i=1
Ai
s − s∞i
, (4.77)
the Ai are the coefficients of the partial-fraction expansion, and
the s∞i denote the poles of the analog filter. Inverse Laplace
transform of (4.77) yields
ha(t) =
N
X
i=1
Ai e
s∞it
, t ≥ 0.
Periodical sampling:
h(n) = ha(nT ) =
N
X
i=1
Ai e
s∞inT
.
Transfer function H(z) •−◦ h(n):
H(z) =
∞
X
n=0
h(n)z
−n
=
∞
X
n=0
N
X
i=1
Ai e
s∞inT
!
z
−n
.
134

We then have
H(z) =
N
X
i=1
Ai
∞
X
n=0

e
s∞iT
z
−1
n
=
N
X
i=1
Ai
1 − es∞iT z−1
.
Thus, given an analog filter Ha(s) with poles s∞i, the transfer
function of the corresponding digital filter using the impulse
invariant transform is
H(z) =
N
X
i=1
Ai
1 − es∞iT z−1
(4.78)
with poles at z∞i = es∞iT
, i = 1, . . . , N. Note: (4.78)
holds only for distinct poles, generalization to multiple-order
poles possible.
Example:
Convert the analog filter with transfer function
Ha(s) =
s + 0.1
(s + 0.1)2 + 9
into a digital filter using the impulse invariant method.
Poles of Ha(s): s∞0,1 = −0.1 ± j3
Partial-fraction expansion:
Ha(s) =
0.5
s + 0.1 − j3
+
0.5
s + 0.1 + j3
135
From (4.78) we then finally have
H(z) =
0.5
1 − e−(0.1−j3)T z−1
+
0.5
1 − e−(0.1+j3)T z−1
=
1 − (e−0.1T
cos(3T )) z−1
1 − (2 e−0.1T cos(3T )) z−1 + e−0.2T z−2
.
Magnitude frequency responses:
Digital filter: 20 log10 |H(ejω
)| Analog filter: 20 log10 |Ha(jΩ)|
4.5.2 Bilinear transform
Algebraic transformation between the variables s and z, mapping
of the entire jΩ-axis of the s-plane to one revolution of the unit
circle in the z-plane.
Definition:
s =
2
T
1 − z−1
1 + z−1
!
, (4.79)
T denoting the sampling interval.
The transfer function of the corresponding digital filter can be
obtained from the transfer function of the analog filter Ha(s)
136

according to
H(z) := Ha

2
T
1 − z−1
1 + z−1
!#
= Ha(s).
Properties
• Solving (4.79) for z yields
z =
1 + (T/2)s
1 − (T/2)s
, (4.80)
and by substituting s = σ + jΩ we obtain
z =
1 + σT/2 + jΩT/2
1 − σT/2 − jΩT/2
.
σ 0 → |z| 1, σ 0 → |z| 1 for all Ω
⇒ causal, stable continuous-time filters map into causal stable
discrete-time filters
• By inserting s=jΩ into (4.80), it can be seen that |z|=1 for
all values of s on the jΩ-axis ⇒ jΩ-axis maps onto the unit
circle.
• Relationship between ω and Ω: From (4.79) we obtain with
s = jΩ and z = ejω
jΩ =
2
T
1 − e−jω
1 + e−jω
!
,
=
2
T

j sin(ω/2)
cos(ω/2)

=
2j
T
tan(ω/2)
137
⇒ Nonlinear mapping between ω and Ω (warping of the
frequency axis) according to
Ω =
2
T
tan(ω/2), ω = 2arctan(ΩT/2). (4.81)
Remarks:
• Design of a digital filter often begins with frequency
specifications in the digital domain, which are converted
to the analog domain by (4.81). The analog filter is
then designed considering these specifications (i.e. using the
classical approaches from the following section) and converted
back into the digital domain using the bilinear transform.
• When using this procedure, the parameter T cancels out and
can thus be set to an arbitrary value (T = 1).
• Example:
Design a digital single-pole lowpass filter with a −3 dB frequency (cutoff
frequency) of ωc = 0.2π, using the bilinear transform applied to the
138

analog filter with the transfer function
Ha(s) =
Ωc
s + Ωc
,
Ωc denoting the analog cutoff frequency.
Ωc is obtained from ωc using (4.81)
Ωc =
2
T
tan(ωc/2) =
0.65
T
.
The analog filter now has the transfer function
Ha(s) =
0.65/T
s + 0.65/T
,
which is transformed back into the digital domain using the bilinear
transform in (4.79), leading to the transfer function of our desired digital
filter:
H(z) =
0.245 (1 + z−1
)
1 − 0.509 z−1
.
Note that the parameter T has been divided out. The frequency response is
H(e
jω
) =
0.245 (1 + e−jω
)
1 − 0.509 e−jω
,
especially we have H(ej0
) = 1, and |H(ej0.2π
)| = 0.707, which is
the desired response.
4.5.3 Characteristics of commonly used analog filters
• Design of a digital filter can be reduced to design an
appropriate analog filter and then performing the conversion
from H(s) to H(z).
• In the following we briefly discuss the characteristics of
commonly used analog (lowpass) filters.
139
Butterworth filters
Lowpass Butterworth filters are allpole-filters characterized by the
squared magnitude frequency response
|H(Ω)|
2
=
1
1 + (Ω/Ωc)2N
, (4.82)
N is the order of the filter, Ωc is the −3 dB frequency (cutoff
frequency).
Since H(s) · H(−s)|s=jΩ = |H(jΩ)|2
we have from (4.82)
by analytic continuation into the whole s-plane
H(s) · H(−s) =
1
1 + (−s2/Ω2
c)N
.
→ Poles of H(s) H(−s):
−s2
Ω2
c
= (−1)
1/N
= e
j(2n+1)π/N
→ s∞,n = Ωc e
jπ/2
e
j(2n+1)π/(2N)
, n = 0, . . . , 2N − 1
(4.83)
• From (4.83): The 2N poles of H(s) H(−s) occur on a
circle of radius Ωc at equally spaced points in the s-plane.
• The N poles for n = 0, . . . , N − 1 in (4.83) are located in
the left half of the s-plane and belong to H(s).
• The N remaining poles lie in the right half of the s-plane and
belong to H(−s) (stability!).
• Furthermore, a Butterworth filter has N zeros at Ω → ∞.
140

Pole locations in the s-plane:
Frequency responses (ω → Ω, |H(Ωp)|2
= 1/(1 + ǫ2
))
Estimation of the required filter order N:
At the stopband edge frequency Ωs (4.82) can be written as
1
1 + (Ωs/Ωc)2N
= δ
2
2,
which leads to
N =
log((1/δ2
2) − 1)
2 log(Ωs/Ωc)
. (4.84)
MATLAB commands: buttord for order estimation, butter
141
for the design of the corresponding digital filter obtained via
bilinear transform.
Example:
Determine the order and the poles of a lowpass Butterworth filter that has a
−3 dB bandwidth of 500 Hz and an attenuation of 40 dB at 1000 Hz.
• −3 dB frequency Ωc = 2 π · fc = 1000 π,
• stopband frequency Ωs = 2 π · fs = 2000 π,
• attenuation of 40 dB → δ2 = 0.01.
From (4.84) we then obtain
N =
log10(104
− 1)
2 log10 2
= 6.64
In order to be on the safe side we choose N = 7.
Properties of the resulting digital filter designed with butter for N = 7,
fsamp = 8000 Hz, and the above parameters (continuous-time filter
transformed by bilinear transform into the discrete-time domain):
0 0.1 0.2 0.3 0.4 0.5 0.6
−100
−80
−60
−40
−20
0
ω / π →
20
log
10
|H(e
jω
)|
[dB]
→
Transition band
0 0.05 0.1 0.15 0.2 0.25
0
0.2
0.4
0.6
0.8
1
ω / π →
|H(e
jω
)|
→
142

Phase response
0 0.2 0.4 0.6 0.8 1
−15
−10
−5
0
ω / π →
arg
H(e
jω
)
[rad]
→
Pole/zero locations
−1 −0.5 0 0.5 1
−1
−0.5
0
0.5
1
Re{z}
Im{z}
Chebyshev filters
Two types of Chebyshev filters:
• Type 1 filters are all-pole filters with equiripple behavior
in the passband and monotonic characteristic (similar to a
Butterworth filter) in the stopband.
• Type 2 filters have poles and zeros (for finite s), and equiripple
behavoir in the stopband, but a monotonic characteristic in the
passband.
Type 1 Chebyshev filter:
Squared magnitude frequency response:
|H(Ω)|
2
=
1
1 + ǫ2 T 2
N(Ω/Ωp)
, (4.85)
where ǫ is a parameter related to the passband ripple, and TN(x)
is the N-th order Chebyshev polynomial defined as
TN(x) =
(
cos(Ncos−1
(x)) for |x| ≤ 1,
cosh(Ncosh−1
(x)) for |x| 1.
(4.86)
143
The Chebyshev polynomials can be obtained by the recursive
equation
TN+1(x) = 2 x TN(x) − TN−1(x), N = 1, 2, . . .
Examples:
• T0(x) = 1, T1(x) = cos(cos−1
(x)) = x
• T2(x) = cos(2 cos−1
(x)) = 2cos2
(cos−1
x) − 1
= 2 x2
− 1
• T3(x) = cos(3 cos−1
(x))
= 4 cos3
(cos−1
x)−3 cos(cos−1
(x)) = 4 x3
−3 x
• .
.
.
⇒ TN(x) represents a polynom of degree N in x.
⇒ Chebyshev behavior (minimizing the maximal error) in the
passband (or in the stopband for type 2 filters).
The filter parameter ǫ in (4.85) is related to the passband ripple:
For N odd, TN(0) = 0 → |H(0)|2
= 1,
for N even, TN(0) = 1 → |H(0)|2
= 1
1+ǫ2
At the passband edge frequency Ω = Ωp we have TN(1) = 1,
such that
1
√
1 + ǫ2
= 1−δ1 ⇐⇒ ǫ =
s
1
(1 − δ1)2
− 1, (4.87)
which establishes a relation between the passband ripple δ1 and
the parameter ǫ.
Typical squared magnitude frequency response for a Chebyshev
type 1 filter:
144

Type 2 Chebyshev filter:
Squared magnitude frequency response:
|H(Ω)|
2
=
1
1 + ǫ2 [T 2
N(Ωs/Ωp)/T 2
N(Ωs/Ω)]
(4.88)
⇒ contains zeros at s ∞ and poles
Typical squared magnitude frequency response for a Chebyshev
type 2 filter:
145
Estimation of the filter order:
Chebyshev filter only depend on the parameters N, ǫ, δ2, and the
ratio Ωs/Ωp. Using these values, it can be shown that the required
order can be estimated as
N =
log
q
1 − δ2
2 +
q
1 − δ2
2(1 + ǫ2)

/(ǫ δ2)

log
h
Ωs/Ωp +
p
(Ωs/Ωp)2 − 1
i .
(4.89)
MATLAB commands:
• Order estimation: cheb1ord for type 1, cheb2ord for
type 2.
• Design of the corresponding digital filter, obtained from the
analog version by bilinear transform: cheby1 for type 1,
cheby2 for type 2.
Elliptic (Cauer) filters
• Elliptic filters have equiripple (Chebyshev) behavior in both
pass- and stopband.
• Transfer function contains both poles and zeros, where the
zeros are located on the jΩ-axis.
• Squared magnitude frequency response
|H(Ω)|
2
=
1
1 + ǫ2 UN(Ω/Ωp)
, (4.90)
where UN(x) denotes the Jacobian elliptic function of order
N, and the parameter ǫ controls the passband ripple.
• Characteristic squared magnitude frequency response for a
elliptic filter:
146

Æ2
2 Æ2
2
• Filter design is optimal in pass- and stopband in the equiripple
sense: However, other types of filters may be preferred due to
their better phase response characteristics (i.e. approximately
linear-phase), for example the Butterworth filter.
Estimation of the filter order:
Required order to achieve the specifications with the parameters
δ1, δ2 and Ωp/Ωs, (1 − δ1 = 1/
√
1 + ǫ2, 1 − δ2 =
1/
√
1 + δ2):
N =
K(Ωp/Ωs) K(
p
1 − (ǫ/δ)2)
K(ǫ/δ) K(
p
1 − (Ωp/Ωs)2)
(4.91)
where K(x) denotes the complete elliptic integral of the first kind
(tabulated)
K(x) =
π/2
Z
0
dθ
p
1 − x2 sin2
θ
.
MATLAB commands: ellipord for order estimation, ellip
for the design of the corresponding digital filter obtained via
bilinear transform.
147
5. Multirate Digital Signal Processing
• In many practical signal processing applications different
sampling rates are present, corresponding to different
bandwidths of the individual signals → multirate systems.
• Often, a signal has to be converted from one rate to another.
This process is called sampling rate conversion.
– Sampling rate conversion can be carried out by analog
means, that is D/A conversion followed by A/D conversion
using a different sampling rate → D/A converter
introduces signal distortion, and the A/D converter leads
to quantization effects.
– Sampling rate conversion can also be carried out
completely in the digital domain: Less signal distortions,
more elegant and efficient approach.
• ⇒ Topic of this chapter is multirate signal processing and
sampling rate conversion in the digital domain.
5.1 Basic multirate operations
5.1.1 Sampling rate reduction
Reduction of the sampling rate (downsampling) by factor M:
Only every M-th value of the signal x(n) is used for further
processing, i.e. y(m) = x(m · M)
x(n) y(m)
.
.
↓ M
Example: Sampling rate reduction by factor 4
148

(from [Fliege: Multiraten-Signalverarbeitung, 1993])
In the z-domain we have
X0(z) = X0(z
M
) =
∞
X
m=−∞
x(mM)z
−mM
=
∞
X
m=−∞
y(m)(z
M
)
−m
= Y (z
M
) = Y (z
′
) •−◦ y(m) (5.1)
Frequency response after downsampling
Starting point: orthogonality of the complex exponential sequence
1
M
M−1
X
k=0
e
j2πkm/M
=
(
1 for m = λM, λ ∈ Z
Z,
0 otherwise.
(5.2)
With x0(mM) = x(mM) it follows
x0(m) = x(m)
1
M
M−1
X
k=0
W
−km
M , WM := e
−j2π/M
(5.3)
149
With (5.3) the z-transform X0(z) can be obtained as
X0(z) =
∞
X
m=−∞
x0(m)z
−m
=
1
M
M−1
X
k=0
∞
X
m=−∞
x(m)(W
k
Mz)
−m
. (5.4)
By replacing Y (zM
) = X0(z) in (5.4) we have for the z-
transform of the downsampled sequence y(m)
Y (z
M
) =
1
M
M−1
X
k=0
X(zW
k
M). (5.5)
With z = ejω
and ω′
= ωM the corresponding frequency
response can be derived from (5.5):
Y (e
jω′
) =
1
M
M−1
X
k=0
X(e
j(ω′−k2π)/M)
) (5.6)
⇒ Downsampling by factor M leads to a periodic repetition of
the spectrum X(ejω
) at intervals of 2π/M (related to the high
sampling frequency).
Example: Sampling rate reduction of a bandpass signal
by M = 16 (Ω → ω)
150

|Q(ejΩ)|
|Q(ejΩ)| ⋅ |H(ejΩ)| = |X(ejΩ)|
π
8
π
0
π
16 π
Ω
a)
|X(ej(Ω+Ωm))|
π Ω
b)
2π
0
Ω
Ω´
c)
2π
2π
32π
2π
0
0
3π
8
Ωm=
π
8
3π
8
Ωm=
π
8
3π
8
Ωm=
(from [Vary, Heute, Hess: Digitale Sprachsignalverarbeitung, 1998])
(a) Bandpass spectrum X(ejω
) is obtained by filtering.
(b) Shift to the baseband, followed by decimation with M = 16.
(c) Magnitude frequency response |X(ejω′
)| at the lower sampling rate.
Remark: Shifted versions of X(ejω
) are weighted with the factor
1/M according to (5.6).
Decimation and aliasing
If the sampling theorem is violated in the lower clock rate,
we obtain spectral overlapping between the repeated spectra ⇒
Aliasing
How to avoid aliasing? Bandlimitation of the input signal v(n)
prior to the sampling rate reduction with an antialiasing filter
h(n) (lowpass filter).
.
h(n)
v(n)
.
y(m)
↓ M
x(n)
151
⇒ Antialiasing filtering followed by downsampling is often
called decimation.
Specification for the desired magnitude frequency response of the
lowpass antialiasing (or decimation) filter:
|Hd(e
jω
)| =
(
1 for |ω| ≤ ωc/M,
0 for π/M ≤ |ω| ≤ π,
(5.7)
where ωc π denotes the highest frequency that needs to be
preserved in the decimated signal.
Downsampling in the frequency domain, illustration for M = 2:
(a) input and filter spectra, (b) output of the decimator, (c) no
filtering, only downsampling (V → X):
(from [Mitra, 2000])
152

More general approach: Sampling rate reduction with phase
offset
Up to now we have always used y(0) = x(0), now we introduce
an additional phase offset ℓ into the decimation process.
Example for ℓ = 2
=m `=M
0 1 3
2
` ` `
Note that y2(m) in (c) is a formal description for the output signal of the
downsampler with non-integer sample indices. The real output signal y2(κ) is
obtained by assuming integer sample locations.
Derivation of the Fourier transform of the output signal y(m):
Orthogonality relation of the complex exponential sequence:
1
M
M−1
X
k=0
e
j2πk(m−ℓ)/M
=
(
1 for m = λM + ℓ, λ ∈ Z
Z,
0 otherwise.
(5.8)
153
Using (5.8) we have
xℓ(m) = x(m)
1
M
M−1
X
k=0
W
−k(m−ℓ)
M , (5.9)
and transforming (5.9) into the z-domain yields
Xℓ(z) =
1
M
M−1
X
k=0
∞
X
m=−∞
x(m)(W
k
Mz)
−m
W
kℓ
M
=
1
M
M−1
X
k=0
X(zW
k
M) W
kℓ
M . (5.10)
The frequency response can be obtained from (5.10) by
substituting z = ejω
and ω′
= Mω as
Yℓ(e
jω′
) =
1
M
M−1
X
k=0
X(e
j(ω′−2πk)/M
) W
kℓ
M , (5.11)
Yℓ(e
jMω
) =
1
M
M−1
X
k=0
X(e
jω−j2πk/M
) W
kℓ
M . (5.12)
⇒ We can see that each repeated spectrum is weighted with a
complex exponential (rotation) factor.
5.1.2 Sampling rate increase
Increase of the sampling rate by factor L (upsampling): Insertion
of L − 1 zero samples between all samples of y(m)
u(n) =
(
y(n/L) for n = λL, λ ∈ Z
Z,
0 otherwise.
(5.13)
154

↑ L
y(m) u(n)
.
.
Notation: Since the upsampling factor is named with L in conformance with
the majority of the technical literature in the following we will denote the length
of an FIR filter with LF .
Example: Sampling rate increase by factor 4
In the z-domain the input/output relation is
U(z) = Y (z
L
). (5.14)
Frequency response after upsampling
From (5.14) we obtain with z = ejω
U(e
jω
) = Y (e
jLω
). (5.15)
⇒ The frequency response of y(m) does not change by
upsampling, however the frequency axis is scaled differently. The
new sampling frequency is now (in terms of ω′
for the lower
sampling rate) equal to L · 2π.
155
ω′
ω
Interpolation
The inserted zero values are interpolated with suitable values,
which corresponds to the suppression of the L − 1 imaging
spectra in the frequency domain by a suitable lowpass
interpolation filter.
↑ L
y(m)
.
.
g(n) v(n)
u(n)
g(n): Interpolation or antiimaging lowpass filter
Specifications for the interpolation filter:
Suppose y(m) is obtained by sampling a bandlimited continuous-
time signal ya(t) at the Nyquist rate (such that the sampling
theorem is just satisfied). The Fourier transform Y (ejω
) can thus
be written with (2.4) and Ω = ω/T0 as
Y (e
jω
) =
1
T0
∞
X
k=−∞
Ya

j(ω − 2πk)
T0

,
156

where T0 denotes the sampling period. If we instead sample ya(t)
at a much higher rate T = T0/L we have
V (e
jω
) =
1
T
∞
X
k=−∞
Ya

j(ω − 2πk)
T

, (5.16)
=
L
T0
∞
X
k=−∞
Ya

j(ω − 2πk)
(T0/L)

.
On the other hand by upsampling of y(m) with factor L we
obtain the Fourier transform of the upsampled sequence u(n)
analog to (5.15) as
U(e
jω
) = Y (e
jωL
).
⇒ If u(n) is passed through an ideal lowpass filter with cutoff
frequency at π/L and a gain of L, the output of the filter will be
precisely v(n) = F−1
{V (ejω
)} in (5.16).
Therefore, we can now state our specifications for the lowpass
interpolation filter:
|Gd(e
jω
)| =
(
L for |ω| ≤ ωc/L,
0 for π/L ≤ |ω| ≤ π,
(5.17)
where ωc denotes the highest frequency that needs to be
preserved in the interpolated signal (related to the lower sampling
frequency).
157
Upsampling in the frequency domain, illustration for L = 2:
(a) input spectrum, (b) output of the upsampler, (c) output after
interpolation with the filter h(n):
5.1.3 Example: Decimation and interpolation
Consider the following structure:
.
.
h(n) ↓ M ↑ M f(n)
u(n) v(n)
x(n) x̂(n)
y(m)
Input-output relation?
Relation between Y (z) and U(z) (see (5.5)), where z is replaced
158

by z1/M
:
Y (z) =
1
M
M−1
X
k=0
U(z
1/M
W
k
M),
which by using U(z) = H(z)X(z) leads to
Y (z) =
1
M
M−1
X
k=0
H(z
1/M
W
k
M) X(z
1/M
W
k
M). (5.18)
With V (z) = Y (zM
) it follows
V (z) =
1
M
M−1
X
k=0
H(zW
k
M) X(zW
k
M), (5.19)
and we finally have
X̂(z) = F (z)Y (z
M
) =
1
M
M−1
X
k=0
F (z) H(zW
k
M) X(zW
k
M).
(5.20)
159
Example:
M = 4, no aliasing:
F
with aliasing:
(from [Mertins: Signal Analysis, 1999])
5.1.4 Polyphase decomposition
• A polyphase decomposition of a sequence x(n) leads to M
subsequences xℓ(m), ℓ = 0, . . . , M − 1, which contain
160

only every M-th value of x(n). Example for M = 2:
Decomposition into an even and odd subsequence.
• Important tool for the derivation of efficient multirate filtering
structures later on.
Three different decomposition types:
• Type-1 polyphase components:
Decomposition of x(n) into xℓ(m), ℓ = 0, 1, . . . , M − 1
with
xℓ(m) = x(mM + ℓ), n = mM + ℓ. (5.21)
With xℓ(m) ◦−• Xℓ(z) the z-transform X(z) can be
obtained as
X(z) =
M−1
X
ℓ=0
z
−ℓ
Xℓ(z
M
) (5.22)
Example for M = 3:
n
m
m
0
0
0
0
1
1
1
1
2
2
2
2
m
x(n)
x0(m)
x1(m)
x2(m)
161
x0(0) = x(0), x0(1) = x(3), . . .
x1(0) = x(1), x1(1) = x(4), . . .
x2(0) = x(2), x2(1) = x(5), . . .
X(z) =
M−1
X
ℓ=0
z
−(M−1−ℓ)
X
′
ℓ(z
M
) (5.23)
with X
′
ℓ(z) •−◦ x
′
ℓ(n) = x(nM + M − 1 − ℓ) (5.24)
Example for M = 3:
x
′
0(0) = x(2), x
′
0(1) = x(5), . . .
x
′
1(0) = x(1), x
′
1(1) = x(4), . . .
x
′
2(0) = x(0), x
′
2(1) = x(3), . . .
X(z) =
M−1
X
ℓ=0
z
ℓ
X̄ℓ(z
M
) (5.25)
with X̄ℓ(z) •−◦ x̄ℓ(n) = x(nM − ℓ) (5.26)
162

5.1.5 Nyquist-Filters
Nyquist- or L-band filters:
• Used as interpolator filters since they preserve the nonzero
samples at the output of the upsampler also at the interpolator
output.
• Computationally more efficient since they contain zero
coefficients.
• Preferred in interpolator and decimator designs.
↑ L
y(m)
.
.
g(n) v(n)
u(n)
Using (5.14) the input-output relation of the interpolator can be
stated as V (z) = G(z) Y (zL
).
The filter G(z) can be written in polyphase notation according to
G(z) = G0(z
L
) + z
−1
G1(z
L
) + · · · + z
−(L−1)
GL−1(z
L
),
where the Gℓ(z), ℓ = 0, . . . , L−1 denote the type 1 polyphase
components of the filter G(z).
Suppose now that the m-th polyphase component of G(z) is a
constant, i.e. Gm(z) = α. Then the interpolator output V (z)
can be expressed as
V (z) = αz
−m
Y (z
L
) +
L−1
X
ℓ=0, ℓ6=m
z
−ℓ
Gℓ(z
L
)Y (z
L
).
(5.27)
⇒ v(Ln+m) = αy(n); the input samples appear at the output
of the system without any distortion for all n. All in-between
(L − 1) samples are determined by interpolation.
163
Properties:
• Impulse response of a zero-phase L-th band filter:
g(Ln) =
(
α for n = 0,
0 otherwise.
(5.28)
⇒ every L-th coefficient is zero (except for n = 0) →
computationally attractive
g(n)
n
• It can be shown for α = 1/L that for a zero-phase L-th band
filter
L−1
X
ℓ=0
G(zW
ℓ
L) = Lα = 1. (5.29)
⇒ The sum of all L uniformly shifted versions of G(ejω
)
add up to a constant.
G(z) G(zWL) G(zWL
−
1
L )
G(zW2
L)
z = ejω
ω
Half-band filters
Special case of L-band filters for L = 2:
164

• Transfer function G(z) = α + z−1
G1(z2
)
• For α = 1/2 we have from (5.29) for the zero-phase filter
g(n)
G(z) + G(−z) = 1. (5.30)
• If g(n) is real-valued then G(−ejω
) = G(ej(π−ω)
) and by
using (5.30) it follows
G(e
jω
) + G(e
j(π−ω)
) = 1. (5.31)
⇒ G(ejω
) exhibits a symmetry with respect to the half-band
frequency π/2 → halfband filter.
ω
ωp ωs
G(ejω
)
• FIR linear-phase halfband filter: Length is restricted to
LF = 4λ − 1, λ ∈ I
N
5.2 Structures for decimation and interpolation
5.2.1 FIR direct form realization for decimation
.
h(n)
v(n)
.
y(m)
↓ M
x(n)
165
The convolution with the length LF FIR filter h(n) can be
described as
x(n) =
LF −1
X
k=0
h(k) · v(n − k),
and the downsampling with y(m) = x(m M). Combining both
equations we can write the decimation operation according to
y(m) =
LF −1
X
k=0
h(k) · v(m M − k). (5.32)
Visualization (M = 3):
⇒ x(0) → y(0)
⇒ x(1) is removed
⇒ x(2) is removed
⇒ x(3) → y(1)
v(0)
h(k)
k
k
k
k
k
v(−k)
v(1 − k)
v(2 − k)
v(3 − k)
v(0)
v(0)
v(0)
h(0) h(LF −1)
166

⇒ Multiplication of h(n) with v(1 − n) and v(2 − n) leads to
the results x(1) and x(2) which are discarded in the decimation
process → these computations are not necessary.
More efficient implementation (v(n) → u(n), LF → N):
(a) Antialiasing FIR filter in first direct form followed by
downsampling.
(b) Efficient structure obtained from shifting the downsampler
before the multipliers:
• Multiplications and additions are now performed at the lower
sampling rate.
• Additional reductions can be obtained by exploiting the
symmetry of h(n) (linear-phase).
5.2.2 FIR direct form realization for interpolation
↑ L
y(m)
.
.
g(n) v(n)
u(n)
The output v(n) of the interpolation filter can be obtained as
v(n) =
LF −1
X
k=0
g(k) · u(n − k),
167
which is depicted in the following:
g(k)
y(−m)
u(−k)
k
m
k
⇒ v(0)
⇒ The output sample v(0) is obtained by multiplication of g(n)
with u(−n), where a lot of zero multiplications are involved,
which are inserted by the upsampling operation.
More efficient implementation (v(n) → x(n), LF → N):
(a) Upsampling followed by interpolation FIR filter in second
direct form
(b) Efficient structure obtained from shifting the upsampler behind
the multipliers:
• Multiplications are now performed at the lower sampling
168

rate, however the output delay chain still runs in the higher
sampling rate.
• Zero multiplications are avoided.
• Additional reductions can be obtained by exploiting the
symmetry of h(n) (linear-phase).
5.3 Decimation and interpolation with polyphase filters
5.3.1 Decimation
• From Section 5.1.4 we know that a sequence can be
decomposed into polyphase components. Here type-1
polyphase components (5.21) are considered in the following.
• Type-1 polyphase decomposition of the decimation filter
h(n): The z-transform H(z) can be written according to
(5.22) as
H(z) =
M−1
X
ℓ=0
z
−ℓ
Hℓ(z
M
), (5.33)
M denoting the downsampling factor and
Hℓ(z′
) •−◦ hℓ(m) the z-transform of the type-1 polyphase
components hℓ(m), ℓ = 0, . . . , M − 1.
Resulting decimator structure (V (z) → U(z)):
(z0
)
(z0
)
(z0
)
(z0
)
Y (z0
) Y (z0
)
169
(a): Decimator with decimation filter in polyphase representation
(b): Efficient version of (a) with M times reduced complexity
Remark: The structure in (b) has the same complexity as the direct
form structure from Section 5.2.1, therefore no further advantage.
However, the polyphase structures are important for digital filter
banks which will be discussed later on.
Structure (b) in time domain (v(n) → u(n)):
5.3.2 Interpolation
Transfer function of the interpolation filter can be written analog
170

to (5.33) for the decimation filter as
G(z) =
L−1
X
ℓ=0
z
−ℓ
Gℓ(z
L
),
L denoting the upsampling factor, and gℓ(m) the type-1
polyphase components of g(n) with gℓ(m) ◦−• Gℓ(z′
).
Resulting interpolator structure (V (z) → X(z)):
Y (z0
) Y (z0
)
(z0
)
(z0
)
(z0
)
(z0
)
(a): Interpolator with interpolation filter in polyphase
representation
(b): Efficient version of (a) with L times reduced complexity
As in the decimator case the computational complexity of the
efficient structure in (b) is the same as for the direct form
interpolator structure from Section 5.2.2.
5.4 Noninteger sampling rate conversion
Notation: For simplicity a delay by one sample will be generally denoted with
171
z−1
for every sampling rate in a multirate system in the following (instead of
introducing a special z for each sampling rate as in the sections before).
• In practice often there are applications where data has to be
converted between different sampling rates with a rational
ratio.
• Noninteger (synchronous) sampling rate conversion by factor
L/M: Interpolation by factor L, followed by a decimation
by factor M; decimation and interpolation filter can be
combined:
Y(z)
.
Y1(z)
G(z) # M
L
X1(z)
X(z)
.
• Magnitude frequency responses:
ω
ω
ω
ω
ω
Efficient conversion structure
In the following derivation of the conversion structure we assume
172

a ratio L/M 1. However, a ratio L/M 1 can also be used
with the dual structures.
1. Implementation of the filter G(z) in polyphase structure,
shifting of all subsamplers into the polyphase branches:
2. Application of the following structural simplifications:
(a) It is known that if L and M are coprime (that is they have
no common divider except one) we can find ℓ0, m0 ∈ I
N
such that
ℓ0 L − m0 M = −1 (diophantic equation) (5.34)
⇒ delay z−λ
in one branch of the polyphase structure can
be replaced with the delay zλ(ℓ0 L−m0 M)
L
.
G(z) # M
.
z(`0 L m0 M)
(b) The factor zλℓ0L
can be shifted before the upsampler, and
the factor z−λm0M
behind the downsampler:
.
.
z`0 L # M z m0
G(z)
173
(c) Finally, if M and L are coprime, it can be shown that up-
and downsampler may be exchanged in their order:
.
.
z m0
z`0 # M L
G(z)
(d) In every branch we now have a decimator (marked with the
dashed box), which can again be efficiently realized using
the polyphase structure from Section 5.3.1. Thus, each
type-1 polyphase component gλ(n) is itself decomposed
again in M polyphase components gλµ(n) ◦−• Gλµ(z),
λ = 0, . . . , L − 1, µ = 0, . . . , M − 1.
Resulting structure:
• Delays z−λm0 are realized with the output delay chain.
174

• The terms zλℓ0 are noncausal elements: In order to obtain a
causal representation, we have to insert the extra delay block
z−(L−1)ℓ0 at the input of the whole system, which cancels out
the ”negative” delays zλℓ0.
• Polyphase filters are calculated with the lowest possible
sampling rate.
• L/M 1 is realizable using the dual structure (exchange:
input ↔ output, downsamplers ↔ upsamplers, summation
points ↔ branching points, reverse all branching directions)
Example for L = 2 and M = 3:
Application: Sampling rate conversion for digital audio signals
from 48 kHz to 32 kHz sampling rate
Polyphase filters are calculated with 16 kHz sampling rate
compared to 96 kHz sampling rate in the original structure.
Rate conversion from 32 kHz to 48 kHz: Exercise!
5.5 Efficient multirate filtering
In the following we only consider lowpass filtering, however, the
175
presented methods can easily be extended to band- or highpass
filters.
5.5.1 Filtering with lower sampling rate
If the stopband edge frequency of a lowpass filter is substantially
smaller than half of the sampling frequency, it may be advisable
to perform the filtering at a lower sampling rate:
H (z) M
# M Hi(z)
.
fs1 =fs=M fs1
Hd(z)
V(z)
.
H(z)
fs
Y(z)
fs
Hd(z) is the (input) decimation filter, the actual filtering is carried
out with the core filter Hc(z) in the M-times lower sampling rate
with sampling frequency fs1
= fs/M, and after upsampling
the output signal is interpolated with Hi(z) ⇒ Single-stage
implementation
Stopband- and passband edge frequencies of the decimation and
interpolation filters have to be adjusted to the filter specifications
of the core filter:
fs1
f
fs1 =2 fs1 fstop
fstop
fpass
jHdj; jHij
jH j
• Stop- and passband edge frequencies fstop and fpass of the
core filter Hc(z) are identical with those for the overall filter
H(z) = Y (z)/V (z).
176

• Stopband edge frequency for the decimation filter then has to
be chosen less or equal than (fs1
−fstop).
• The interpolation filter can be chosen identical to the
decimation filter, since then it is guaranteed that all imaging
components are in the stopband region of the interpolation
filter.
• Transition bandwidth for H(z) is M-times smaller than for
Hc(z) ⇒ design with a fairly small number of coefficients
for Hc(z) possible (compared to a direct design of H(z)).
• Stopband ripple δ2 for the overall filter H(z):
δ2 =
(
δ2,c(1 + δ1,i)(1 + δ1,d) ≈ δ2,c, fstop ≤ f ≤ (fs1 −fstop),
δ2,c δ2,d δ2,i, (fs1 −fstop) f ≤ fs
(5.35)
where the approximation for fstop ≤ f ≤ (fs1
−fstop) holds
for small decimation and interpolation filter passband ripples
δ1,d and δ1,i.
• Passband ripple δ1 for H(z):
1 + δ1 = (1 + δ1,c) (1 + δ1,d) (1 + δ1,i),
(5.36)
approximation
=⇒ δ1 ≈ δ1,c + δ1,d + δ1,i, (5.37)
where the last approximation is valid for small passband
ripples δ1,c, δ1,d, and δ1,i.
• Complexity savings (#multiplications and #additions) can be
obtained by roughly a factor of 100. An even higher gain can
be achieved by multistage implementations.
177
5.5.2 Interpolating FIR (IFIR) filters
Alternative to multirate filters with decimation and interpolation,
also suitable for very narrowband filters.
Principle:
V (z)
.
.
Y (z)
H(z)
G(zK) F(z)
• No real multirate filter since both filters are calculated with the
same (input) sampling rate. Multirate technique is applied to
the coefficients of the impulse response h(n).
• Realization of G(zK
) in the first direct structure:
g(0) g(1) g(2) g(3) g(LF −2) g(LF −1)
z−K
v(n)
z−K
z−K
z−K
z−K
z−K
y(n)
G(zK
) is a function where all z−1
are replaced by z−K
,
which is equivalent to inserting K − 1 zeros between the
coefficients of G(z).
• G(ejω
) → G(ejKω
): Frequency response G(ejω
) is
”compressed” by factor K, K−1 imaging spectra are present:
ω
ω
|G(ejω
)|
|G(ejKω
)|
178

Furthermore, the transition bandwidth and the width of
the passband for G(ejKω
) are K-times smaller than for
the original filter G(ejω
) with the same number of filter
coefficients.
• The filter F (z) removes the imaging spectra, and
H(z) = G(zK
) · F (z) only consists of the baseband
part of G(ejKω
).
Design: Starting point for the design: Passband and stopband
edge frequencies ωs, ωp for the overall filter H(z) → search for
a suitable factor K leading to a less complex interpolation filter
f(n).
Filter specifications (H, F, G are all magnitude frequency responses)
ω
ω
Kωp Kωs
ωs
ωp
• Requirements for passband and stopband edge frequency of
the prototype G(z):
ωp,G = K · ωp, ωs,G = K · ωs. (5.38)
• Requirements for passband and stopband edge frequency of
179
the interpolation filter F (z):
ωp,F = ωp, ωs,F =
2π
K
− ωs. (5.39)
• Passband ripple δ1 for H(z):
1 + δ1 = (1 + δ1,G) (1 + δ1,F ). (5.40)
Small passband ripples δ1,G for G(z) and δ1,F for F (z),
resp., lead to the simplification
δ1 ≈ δ1,G + δ1,F . (5.41)
• Stopband ripple δ2 for H(z):
δ2 =
(
δ2,G (1 + δ1,F ) for ωs ≤ ω ≤ ωs,F ,
δ2,F (1 + δ1,G) for ωs,F ω ≤ π.
(5.42)
For small passband ripples δ1,G, δ1,F we have approximately
δ2 ≈ δ2,F = δ2,G. (5.43)
Example:
Design a lowpass IFIR filter with the following specifications:
ωd = 0.05π, ωs = 0.1π,
20 log10(1 + δ1) = 0.2 dB → δ1 ≈ 0.023, 20 log10(|δ2|) = −40 dB
1. We select a factor K =4: Prototype G(z) has the parameters
ωp,G = 0.2π, ωs,G = 0.4π,
δ1G
≈ 0.0116 → 20 log10(1+δ1,G)≈0.1 dB,
20 log10(|δ2,G|) = −40 dB,
180

ripple is equally distributed between G(z) and F (z), see (5.41).
2. We use an linear-phase FIR Chebyshev design and insert these values for
G(z) into (4.73), yielding a filter order N = 19. However, several test
designs show that the above specifications are only met for N = 21,
leading to a filter length of LF =N +1=22.
Impulse response
0 5 10 15 20
−0.2
−0.1
0
0.1
0.2
0.3
0.4
n →
g(n)
→
0 0.2 0.4 0.6 0.8 1
−80
−60
−40
−20
0
20
log
10
|G(e
jΩ
)|
[dB]
→
ω / π →
3. Specifications for the interpolation filter F (z):
ωp,F = ωp = 0.05π, ωs,F =
2π
K
− ωs = 0.4π,
20 log10(1 + δ1,F ) ≈ 0.1 dB, 20 log10(|δ2,F |) = −40 dB.
⇒ resulting filter order ((4.73)) N = 12:
Impulse response
0 2 4 6 8 10 12
−0.1
0
0.1
0.2
0.3
n →
f(n)
→
0 0.2 0.4 0.6 0.8 1
−80
−60
−40
−20
0
20
log
10
|F(e
jΩ
)|
[dB]
→
ω / π →
4. Upsampling of the impulse response g(n) by factor K = 4 →
g1(n) ◦−• G(ejKω
):
181
Impulse response
0 20 40 60 80
−0.2
−0.1
0
0.1
0.2
0.3
0.4
n →
g
1
(n)
→
0 0.2 0.4 0.6 0.8 1
−80
−60
−40
−20
0
20
log
10
|G(e
jKΩ
)|
[dB]
→
ω / π →
5. Final IFIR filter h(n) = g1(n) ∗ f(n):
0 0.2 0.4 0.6 0.8 1
−80
−60
−40
−20
0
20
log
10
|H(e
jω
)|
[dB]
→
ω / π →
Passband detail
0 0.02 0.04 0.06
0.98
0.99
1
1.01
1.02
|H(e
jω
)|
→
ω / π →
6. H(z) in the IFIR realization has an overall length of 35 coefficients. On
the other hand we obtain from (4.73) an estimated length of 65 coefficients
for the direct form implementation.
5.6 Application: Oversampled A/D and D/A converter
5.6.1 Oversampled A/D converter
Extension of the A/D and D/A conversion concepts from
Section 2.
Structure:
C
R
x(t)
h(n) ↓ 2L
Q[xL(n)]
x(nT/2L
) = xL(n)
T/2L
y(n)
ŷ(n)
b bit
Q
A/D converter
182

• Continuous-time input signal is band-limited by the analog
lowpass such that the frequency ωu represents the maximum
frequency in the interesting frequency range for the input
signal:
Sampling with a sampling frequency 2L
· ωs ≥ 2 · ωu,
L ∈ {0, 1, 2, . . . }, after the analog filtering. Here ωs
denotes the lowest possible sampling frequency in order not
to violate the sampling theorem: ωs = 2 · ωu = 2π/T .
• A/D converter here is idealized as concatenation of sampler
and quantizer.
• After A/D conversion a lowpass filtering is carried out where
the lowpass has the idealized frequency response
|H(e
jω
)| =
(
1 for |ω| π/2L
,
0 otherwise.
The resulting bandlimited signal can then be downsampled by
factor 2L
.
Quantization noise variance of a b-bit midtreat quantizer
according to (2.27), where the range R of the quantizer is chosen
as R = 2:
σ
2
e =
2−2b+2
12
.
As an alternative σ2
e can also be obtained via the power spectral
density (power spectrum) Φee(ejω
) as
σ
2
e =
1
2π
π
Z
−π
Φee(e
jω
) dω.
183
The filtering with the lowpass h(n) now reduces the quantization
noise variance by factor 2L
, since
σ
2
e(L,b)
=
1
2π
π/2L
Z
−π/2L
Φ
(L)
ee (e
jω
) dω =
2−2b+2
12 · 2L
. (5.44)
⇒ Reduction of the noise variance due to oversampling by
factor 2L
:
Gain = 10 log10


σ2
e
σ2
e(L,b)

 = 10 log10(2
L
) = L·3.01 dB.
(5.45)
⇒ An increase of the oversampling factor 2L
by factor 2 leads to
an increase in quantizer resolution by half a bit (compare (2.29))
and to an increase of the signal-to-noise ratio (SNR) in (2.29) by
3 dB!
Visualization:
00000000000000
00000000000000
00000000000000
00000000000000
00000000000000
11111111111111
11111111111111
11111111111111
11111111111111
11111111111111
00
00
00
00
00
11
11
11
11
11
00
00
00
00
00
11
11
11
11
11
ω
π
σ2
e
−π 0
(a)
Φee(ejω
)
ω
π
π
2L
− π
2L 0
σ2
e
(b)
Φ
(L)
ee (ejω
)
Gain in (5.45) can also be used for reducing the quantizer
wordlength while keeping the SNR constant:
184

• Reduction of the hardware complexity of the core A/D
converter
• Extreme case: Quantizer wordlength is reduced to b = 1 bit
→ only a simple comparator is necessary
Requirements:
σ2
e(L,bL)
σ2
e(0,b0)
!
= 1 =
2−2bL+2
2L · 2−2b0+2
→ bL = b0 −
L
2
, (5.46)
with bL denoting the quantizer wordlength for a given L leading
to the same quantization noise variance as b0 for L=0, bL ≤ b0.
Example:
Given a b0 = 16 bit A/D converter, where the analog input signal is sampled
at the Nyquist rate. Choose the parameter L in the oversampled A/D converter
from above such that the same quantization noise variance for bL = 1 bit is
achieved.
From (5.46) we obtain L = 30, leading to an oversampling factor of
2L
≈ 109
.
Improvement: Shaping of the quantization noise
The quantizer in the above block is now replaced by the following
structure:
Q
[xL(n)]Q
z−1
− +
−
+
Q̂
ê(n)
xL(n) x̂L(n)
185
Analysis:
y(n)=[xL(n)]Q = Q̂[xL(n) − ê(n − 1)],
= xL(n) − ê(n − 1) + ê(n),
= xL(n)+ê(n) ∗ (1−δ(n−1)) (5.47)
◦−• Y (z) = XL(z) + Ê(z)(1 − z
−1
). (5.48)
Therefore, the z-transform of the overall quantization error
sequence can be expressed as
E(z) = Ê(z)

1 − z
−1

,
which leads to the quantization noise power spectrum
Φee(e
jω
) = Φêê(e
jω
)|1 − e
−jω
|
2
.
With Φêê(ejω
) = 2−2b+2
12 (noise power spectrum of a b-bit
midtreat quantizer with range R = 2) we have
Φee(e
jω
) =
2−2b+2
6
(1 − cos(ω)). (5.49)
0 0.2 0.4 0.6 0.8 1
0
1
2
3
4
5
6
x 10
−3
ω / π →
|Φ
ee
(e
jω
)|
→
1/3⋅ 2−2b+2
→
186

⇒ Quantization noise power spectrum now has highpass
character → noiseshaping
The noise variance after lowpass filtering with h(n) in the above
oversampled A/D converter structure is now given with (5.49) as
σ
2
e(L,b)
=
1
2π
π/2L
Z
−π/2L
Φee(e
jω
) dω,
=
2−2b+2
12

2
−L+1
−
2
π
sin(2
−L
π)

. (5.50)
Reduction of the noise variance due to oversampling by factor 2L
:
Gain = −10 log10

2
−L+1
−
2
π
sin(2
−L
π)

. (5.51)
For L = 4: Gain ≈ 31 dB (compared to 12 dB without noise
shaping, see (5.45)).
Reducing the quantizer wordlength for constant SNR:
σ2
e(0,b0)
σ2
e(L,bL)
!
= 1 → bL =b0+
1
2
log2

2
−L+1
−
2
π
sin(2
−L
π)

.
(5.52)
Example:
The above example is again carried out for the noiseshaping case: For
b0 =16 bit and bL = 1 bit we obtain from (5.52) via a computer search
(fix-point iteration) L ≈ 10.47 → When the input signal is sampled with
fs = 44.1 kHz and b0 = 16 the new sampling frequency in the oversampled
A/D converter would be fsover = 211
· fs ≈ 90 MHz for bL =1.
187
⇒ Improved noiseshaping by other techniques, where the
quantization noise power spectrum shows even stronger highpass
character (sigma-delta modulation, more selective shaping filters).
5.6.2 Oversampled D/A converter
Structure:
C
R
y(t)
DAC
b1 bit
b bit b bit
h(n) Q
b1 b bit
h1(n)
Q[xL(n)]
xL(n)
x̂L(n)
x(n)
↑ 2L
1. Input signal sampled with b bits is upsampled by factor 2L
and
then interpolated with h(n).
2. The resulting signal xL(n) is requantized to a wordlength
of b1 b bits, leading to a worse SNR due to higher
quantization noise.
3. Filtering by h1(n) removes the quantization noise in the
unused spectral regions, which increases the SNR again.
4. The b1 bit DAC in combination with a simple analog lowpass
converts the signal back into the analog domain.
Reason for performing a requantization: Use of a cheap low
resolution D/A converter is possible, often with b1 = 1 bit.
In combination with a noiseshaping approach the same SNR is
obtained as when a b bit converter would be used directly on the
input signal (but with higher costs).
⇒ Favored converter principle in CD players (→ ”bitstream”
conversion for b1 =1 bit)
188

5.7 Digital filter banks
• A digital filter bank consists of a set of filters (normally
lowpass, bandpass and highpass filters) arranged in a parallel
bank.
• The filters split the input signal x(n) into a number of subband
signals yk(n), k = 0, . . . , K − 1 (analysis filter bank).
• Subband signals are processed and finally combined in a
synthesis filter bank leading to the output signal x̂(n).
• If the bandwidth of the subband signal is smaller than the
bandwidth of the original signal, they can be downsampled
before processing → processing is carried out more efficiently.
.
.
↓ M0
↓ M1
u1(n)
u0(n)
uK−1(n)
h1(n)
h0(n)
hK−1(n)
x(n)
analysis filter bank
subband
processing
↑ M1
↑ M0
↑MK−1
v0(n)
v1(n)
vK−1(n)
synthesis filter bank
x̂(n)
fK−1(n)
f0(n)
f1(n)
y0(m0)
y1(m1)
yK−1(mK−1)
↓MK−1
Subband processing: Quantization (wordlength reduction) →
coding, (adaptive) filtering → equalization, ...
5.7.1 Two-channel filter banks: Structure and analysis
Basic two-channel subband coder:
↓ 2
↓ 2
↑ 2
↑ 2
Demultiplexer
Multiplexer
Channel
Decoder
Decoder
Encoder
Encoder
.
.
x(n) H0(z)
H1(z)
x̂(n)
F0(z)
F1(z)
Only the errors in the above system related to the filter bank are
189
investigated in the following → simplified structure:
↓ 2
↓ 2
↑ 2
↑ 2
.
.
x(n)
y0(m)
y1(m)
H0(z)
H1(z)
x̂(n)
F0(z)
F1(z)
⇒ Critically subsampled filter bank: The number of subband
equals the subsampling factor in every subband.
Typical frequency responses for the analysis filters:
Frequency responses of the subband signals (for ideal filters)
(Ω → ω) :
0 π 2π Ω
Ω
0 2π
π 3π 4π
Ω´
0 π 2π
0 π 2π
Ω
0 2π
π 3π 4π
Ω´
b)
a)
c)
0 0
0 1
1
1
1 1
1
0
0 0
X(ejΩ)
(From [Vary, Heute, Hess: Digitale Sprachsignalverarbeitung, 1998])
(a): Magnitude frequency response of the input signal, (b) magnitude frequency
response of the lowpass, and (c) highpass branch after analysis filtering and
downsampling
⇒ Highpass subband: Baseband spectra in frequency reversed
order
190

Analysis
How do we have to design the filters such that
x̂(n) = x(n − D) holds?
(D denotes the delay of the overall analysis-synthesis system in
samples)
We have the following relations:
Y0(z
2
) =
1
2
[H0(z) X(z) + H0(−z)X(−z)] (5.53)
Y1(z
2
) =
1
2
[H1(z) X(z) + H1(−z)X(−z)] (5.54)
Proof:
These relations can be obtained from (5.18)
Y (z) =
1
M
M−1
X
k=0
H(z
1/M
W
k
M) X(z
1/M
W
k
M)
for M = 2. With
W
k
2 = e
−j2πk/2
= e
−jπk
=
(
1 for n even,
−1 for n odd,
we have
Y (z) =
1
2
1
X
k=0
H((−1)
k
z
1/2
) X((−1)
k
z
1/2
)
=
1
2
h
H(+z
1/2
) X(+z
1/2
) + H(−z
1/2
) X(−z
1/2
)
i
.
(5.55)
191
Replacing z by z2
then leads to
Y (z
2
) =
1
2
[H(z) X(z) + H(−z) X(−z)] ,
where (5.53), (5.54) can be obtained by replacing H with H0 and
H1, resp. ✷
The connection between the subband signals and the reconstructed
signal is
X̂(z) =
h
Y0(z
2
) F0(z) + Y1(z
2
) F1(z)
i
, (5.56)
and finally by combining (5.56), (5.53), and (5.54) the input-
output relation for the two-channel filter bank writes
X̂(z) =
1
2
[H0(z) F0(z) + H1(z) F1(z)] X(z)+
+
1
2
[H0(−z) F0(z) + H1(−z) F1(z)] X(−z).
(5.57)
(5.57) consists of two parts:
S(z) = [H0(z) F0(z) + H1(z) F1(z)] , (5.58)
G(z) = [H0(−z) F0(z) + H1(−z) F1(z)] (5.59)
• S(z): Transfer function for the input signal X(z) through
the filter bank, desirable is
S(z)
!
= 2z
−D
. (5.60)
192

• G(z): Transfer function for the aliasing component X(−z),
desirable is (no aliasing!)
G(z)
!
= 0. (5.61)
Two cases have now to be distinguished:
1. If G(z) = 0, but S(z) 6= c z−D
, then the reconstructed
signal x̂(n) is free of aliasing, however, linear distortions are
present.
2. If G(z) = 0 and S(z) = c z−D
, then we have a perfect
reconstruction (PR) system, except of a scaling factor c/2 and
an additional delay of D samples.
5.7.2 Two-channel quadrature-mirror filter (QMF) banks
(Crosier, Esteban, Galand, 1976)
Quadrature-mirror filter (QMF) banks allow the cancelation of
all aliasing components, but generally lead to linear distortions
(i.e. phase and amplitude distortions)
Starting point: Given (lowpass) prototype H0(z), all other filters
are chosen as
F0(z)=H0(z), H1(z)=H0(−z), F1(z)=−H1(z)
(5.62)
Thus we have from (5.59) for the aliasing transfer function
G(z) = H0(−z) F0(z) + H1(−z) F1(z)
= H0(−z) H0(z) + H0(z)(−H0(−z)) = 0
(5.63)
193
⇒ Cancelation of all aliasing components
For the linear distortion transfer function S(z) one obtains by
inserting (5.62) into (5.58)
S(z) = H
2
0 (z) − H
2
0 (−z),
that is, the prototype H0(z) has to be designed such that
S(z) = H
2
0 (z) − H
2
0 (−z)
!
≈ 2z
−D
(5.64)
is satisfied as good as possible → requirement for an ideal
(constant) overall frequency response for the whole analysis-
synthesis system
• Unfortunately, (5.64) can not be satisfied exactly with FIR
filters, but it can be approximated with an arbitrarily small
error
⇒ Linear distortions can be kept small
• Exception: (5.64) is satisfied exactly by using the prototype
H0(z) =
1
√
2
(1 + z
−1
)
(Haar filter):
1
2
(1 + 2z
−1
+ z
−2
) −
1
2
(1 − 2z
−1
+ z
−2
) = 2z
−1
• The magnitude frequency responses of highpass and lowpass
filter have for real-valued filter coefficients the mirror image
property (therefore the name QMF):
|H1(e
j(π
2 −ω)
)| = |H0(e
j(π
2 +ω)
)| (5.65)
194

0 0.2 0.4 0.6 0.8 1
0
0.2
0.4
0.6
0.8
1
1.2
|H0
(e
jω
)| |H1
(e
jω
)|
ω / π →
|H(e
jω
)|
→
Design of QMF banks
• Usually the design is carried out by minimization of an error
measure
E = Er + αEs
!
= min. (5.66)
Er refers to the linear distortion error energy
Er = 2
π
Z
ω=0

|H0(e
jω
)|
2
+ |H0(e
j(ω−π)
)|
2
− 1

dω,
(5.67)
and Es to the energy in the stopband region of the filter
Es =
π
Z
ω=ωs
|H0(e
jω
)|
2
dω, (5.68)
with the stopband edge ωs = (π + ∆ω)/2. ∆ω denotes the
width of the transition band, which is symmetrically centered
around ω = π/2.
• Minimization of (5.66) can be carried out via a numerical
195
minimization approach for a given ∆ω and given prototype
length LF .
• Catalogs with optimized filter coefficients for h0(n) may
for example be found in [Akansu, Haddad: Multirate signal
decomposition, 1992].
• Once a good prototype H0(z) has been found, the remaining
filters can be obtained from (5.62).
Example: (LF = 20, ∆ω = 0.2π)
Impulse response h0(n)
0 5 10 15 20
−0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
n →
h
0
(n)
→
0 5 10 15 20
−0.5
0
0.5
n →
h
1
(n)
→
Frequency responses for H0, H1
0 0.2 0.4 0.6 0.8 1
−100
−80
−60
−40
−20
0
20
|H
0
(ejω
)| |H
1
(ejω
)|
ω / π →
20
log
10
|H(e
jω
)|
[dB]
→
Frequency response of the QMF bank
0 0.2 0.4 0.6 0.8 1
−0.04
−0.03
−0.02
−0.01
0
0.01
0.02
0.03
0.04
20
log
10
|S(e
jω
)|
[dB]
→
ω / π →
196

Efficient realization using polyphase components
From (5.62) we know that H1(z) = H0(−z) holds for the
analysis highpass filter. Then, the type-1 polyphase components
H0ℓ(z) and H1ℓ(z), ℓ ∈ {0, 1}, are related according to
H10(z) = H00(z) and H11(z) = −H01(z). (5.69)
This leads to an efficient realization of the analysis and synthesis
filter bank, where the number of multiplications and additions can
be reduced by factor four:
0
0
(From [Mertins: Signal analysis, 1999])
5.7.3 Two-channel perfect reconstruction filter banks
In order to obtain a perfect reconstruction filter bank the analysis
and synthesis filters are chosen as follows (ℓ ∈ {0, 1, 2, . . . }):
F0(z) = z
−ℓ
H1(−z), F1(z) = −z
−ℓ
H0(−z)
(5.70)
Aliasing transfer function: Inserting (5.70) into (5.59) yields
G(z) = H0(−z) F0(z) + H1(−z) F1(z)
= H0(−z) z
−ℓ
H1(−z) + H1(−z)

−z
−ℓ
H0(−z)

= 0 ⇒ No aliasing components in the reconstructed signal!
Transfer function for the input signal: Inserting (5.70) into (5.58)
197
yields
S(z) = H0(z) F0(z) + H1(z) F1(z)
= H0(z) F0(z) + (−1)
ℓ+1
H0(−z)F0(−z) (5.71)
Condition for a linear distortion free system: S(z)
!
= 2z−D
With the abbreviation
T (z) := F0(z) H0(z) (5.72)
the PR condition in (5.71) becomes
T (z) + (−1)
ℓ+1
T (−z) = 2 z
−D
. (5.73)
Interpretation:
• [T (z) + T (−z)] refers to the z-transform of a sequence
whose odd coefficients are zero.
• [T (z) − T (−z)]: All coefficients with an even index are
zero.
• The PR condition in (5.71) now states that the corresponding
sequences with z-transforms [T (z) + T (−z)] or
[T (z) − T (−z)] are allowed to have only one non-zero
coefficient. Hence, for t(n) ◦−• T (z) holds
t(n) =







1 for n = D,
0 for n = D + 2λ, λ 6= 0,
arbitrary otherwise.
(5.74)
198

⇒ Half-band filter (Nyquist(2) filter), see Section 5.1.5.
A half-band filter has 4λ − 1 coefficients (λ ∈ I
N).
Example:
(a): linear-phase half-band filter
(b): half-band filter with lower system delay
Filter design via spectral factorization
A given half-band filter T (z) can be factorized according to
T (z) = F0(z) H0(z),
i.e. one looks for the zeros of T (z), and distributes them among
the polynomials F0(z) and H0(z).
The missing filters F1(z) and H1(z) can then be obtained from
(5.70).
⇒ General approach for the PR design of two-channel banks
Example:
A half-band filter is given as
{t(n)} = {−1, 0, 9, 16, 9, 0, −1}.
199
The zeros are {3.7321, −1.0, −1.0, 0.2679}
| {z }
zeros of H0(z)
, {−1.0, −1.0}
| {z }
zeros of F0(z)
,
such that (linear-phase filter).
H0(z) = α(−1 + 2z
−1
+ 6z
−2
+ 2z
−3
− z
−4
),
F0(z) = β(1 + 2z
−1
+ z
−2
)
Orthogonal filter banks
• In the above example: Still two filters F0(z) and H0(z) to
design in order to construct the whole PR filter bank.
• In an orthogonal two-channel filter bank (Smith, Barnwell,
1984), (Mintzer, 1985) it suffices to design the lowpass
analysis filter H0(z) for a PR system.
For an orthogonal two-channel bank the filter H1(z) is chosen as
H1(z) = z
−(LF −1)
H0(−z
−1
), (5.75)
LF denoting the length of h0(n). In the following we will only
consider even lengths LF . Then, using (5.70) with ℓ = 0, the
remaining synthesis filter can be obtained as
F̂0(z)=(−z)
−(LF −1)
H0(z
−1
), F̂1(z)=−H0(−z).
Note that (−z)−(LF −1)
= (−1)−(LF −1)
z−(LF −1)
=
(−1) z−(LF −1)
since (LF −1) odd.
The factor (−1) in F0(z), F1(z) can be regarded as a common
factor multiplied to the output of the synthesis bank and can thus
be removed for simplicity reasons:
F0(z)=z
−(LF −1)
H0(z
−1
), F1(z)=H0(−z). (5.76)
200

Removing the factor (−1) does not change anything at the
aliasing cancelation property: Inserting F0(z) and F1(z) into
(5.59) still yields G(z) = 0.
In order to obtain the condition for a analysis-synthesis system
free of linear distortions we now insert (5.75) and (5.76) into
(5.58), leading to
S(z) = z
−(LF −1)
(H0(z) H0(z
−1
) + H0(−z) H0(−z
−1
)),
!
= 2 z
−D
.
Hence, the PR condition is
H0(z) H0(z
−1
) + H0(−z) H0(−z
−1
)
!
= 2. (5.77)
With z = ejω
(5.77) can be written for real-valued h0(n)
according to
|H0(e
jω
)|
2
+ |H0(e
j(ω+π)
)|
2
= 2 (5.78)
⇒ power-complementary property for H0(ejω
).
0 0.2 0.4 0.6 0.8 1
0
0.5
1
1.5
2
|H
0
(e
jω
)|
2
|H
0
(e
j(ω+π)
)|
2
ω / π →
201
→ It can be easily shown that the power-complementary property
also holds for h1(n), f0(n), and f1(n).
Design of orthogonal filter banks
1. Search a prototype H0(z) which satisfies the power-
complementary property (5.78) or (for the interesting real-
valued filters) the corresponding equation (5.77) in the z-
domain.
With the abbreviation TZ(z) = H0(z) H0(z−1
)
TZ(z) + TZ(−z) = 2 (5.79)
is satisfied, where TZ(z) denotes a zero-phase half-band filter.
Notation: In the following zero-phase filters and amplitude responses are
denoted with the subscript ”Z” instead of ”0” to avoid confusion with the
lowpass analysis filter h0(n).
Valid half-band filter: TZ(z) is a valid half-band filter if it can
be factorized into H0(z) and H0(z−1
).
⇒ Design goal: Find a valid half-band filter and factorize it
into H0(z) and H0(z−1
).
2. When a suitable filter H0(z) has been found, the remaining
filter can be obtained from
H1(z) = z
−(LF −1)
H0(−z
−1
),
F0(z) = z
−(LF −1)
H0(z
−1
),
F1(z) = z
−(LF −1)
H1(z
−1
) = H0(−z).
(5.80)
⇒ special case of the conditions in (5.70).
202

How to design a valid half-band filter?
Spectral factorization may be problematic:
With TZ(z) = H0(z) H0(z−1
) and z = ejω
:
TZ(e
jω
) = H0(e
jω
) H0(e
−jω
) = |H0(e
jω
)|
2
!
≥ 0 for all ω.
(5.81)
Design approach due to Smith and Barnwell, 1986
Starting point is an arbitrary linear-phase half-band filter (for
example designed with the Remez algorithm from Section 4.4.4):
A(e
jω
) = AZ(ω)e
−jω(LF −1)
AZ(ω): Real-valued amplitude frequency response of the half-
band filter a(n), AZ(ω) •−◦ aZ(n), LF : length of h0(n).
If the value
δ = min
ω∈[0,2π]
AZ(ω) 0, (5.82)
a non-negative amplitude frequency response can be generated
with
AZ+(ω) = AZ(ω) + |δ|. (5.83)
In time-domain this corresponds with
AZ+(ω) •−◦ aZ+(n) =
(
aZ(n) for n 6= 0,
aZ(n) + |δ| for n = 0.
(5.84)
(aZ(n) denotes the zero-phase impulse response with the center
of symmetry located at n = 0).
203
Visualization:
ω
(From [Fliege: Multiraten-Signalverarbeitung, 1993])
⇒ Single zeros on the unit-circle are converted to double zeros
→ factorization into two filters is now possible
A valid zero-phase half-band filter TZ(z) can finally be
obtained by scaling of the resulting transfer function such that
TZ(ejπ/2
)=1 (note that TZ(z) + TZ(−z)
!
= 2 has to hold),
leading to the expression
TZ(z) =
1
1 + |δ|
AZ+(z). (5.85)
Remark:
In practice for double zeros on the unit circle the separation
process is numerically very sensitive. As a solution, the parameter
|δ| in (5.82) and (5.83) can be enlarged by a small value ǫ →
zeros move pairwise off the unit-circle where due to the linear-
phase property they are mirrored at the unit-circle
204

Example
Parameter for the design of TZ(z): LF = 16, ωs = 0.6π,
∆ω = 0.2π, ǫ = 10−4
Pole-zero pattern TZ(z)
−1 0 1 2
−1
−0.5
0
0.5
1
Re{z} →
Im{z}
→
30
Pole-zero pattern H0(z)
−1 0 1 2
−1
−0.5
0
0.5
1
Re{z} →
Im{z}
→
15
Amplitude frequency response TZ(ω)
0 0.2 0.4 0.6 0.8 1
0
0.5
1
1.5
2
2.5
Ω / π →
T
Z
(ω)
→
Frequency responses analysis filters
0 0.2 0.4 0.6 0.8 1
0
0.5
1
1.5
|H0
(ejω
)| |H1
(ejω
)|=|H0
(−e−jω
)|
ω / π →
After h0(n) has been constructed, the remaining impulse
responses h1(n), f0(n), f1(n) are obtained with (5.80).
205
0 5 10 15
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
n →
h
0
(n)
→
0 5 10 15
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
n →
h
1
(n)
→
Impulse response f0(n)
0 5 10 15
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
n →
f
0
(n)
→
Impulse response f1(n)
0 5 10 15
−0.8
−0.6
−0.4
−0.2
0
0.2
0.4
0.6
0.8
n →
f
1
(n)
→
• Filter in orthogonal two-channel filter banks have an even
number of coefficients since TZ(z) = H0(z) H0(z−1
)
Example:
TZ(z) with 7 coefficients can be factorized into two filters of
length 4.
The next feasible length is 11 coefficients
206

which leads to two filters of length 6.
• Filter in orthogonal two-channel filter banks can not be made
linear-phase (except two trivial cases) (Vaidyanathan, 1985).
5.8 Tree filter banks
K channel filter banks can be developed by iterating the two-
channel systems from above.
F
F
F
F
F
F
F
F
F
F
207
(a): Fully iterated tree filter bank, (b): Octave filter bank
If the two band filter bank is of the perfect reconstruction
type, the generated multiband structure also exhibits the perfect
reconstruction property.
5.9 K-band filter banks
In the following we briefly consider K-band systems with equal
subband widths and equal subsampling factors in every subband.
.
.
↓ M
↓ M ↑ M
↑ M
↑ M
u1(n)
u0(n)
uK−1(n)
h1(n)
h0(n)
hK−1(n)
v0(n)
v1(n)
vK−1(n)
x(n)
analysis filter bank synthesis filter bank
y0(m)
y1(m)
x̂(n)
fK−1(n)
f0(n)
f1(n)
yK−1(m)
↓ M
• If K = M the analysis-synthesis system is called critically
subsampled, for K M we speak of an oversampled
system.
• The case K M (undersampled system) is infeasible since
the sampling theorem is violated in the subbands also for ideal
(brickwall) filters → no reasonable reconstruction is possible.
Subset of general K-band filter banks:
→ Modulated filter banks
All K analysis and synthesis filters are obtained from one single
analysis and synthesis prototype, resp.
208

Advantages:
• More efficient implementation compared to general K-band
systems.
• Less design complexity since only the prototype filters have to
be designed.
• (Less storage required for the filter coefficients.)
Two important types of modulated filter banks: DFT and cosine-
modulated filter banks
In the following only DFT filter banks are briefly considered.
DFT filter banks
• Analysis and synthesis filters (k = 0, 1, . . . , K − 1):
hk(n) = p(n)
| {z }
analysis prototype
· e
j 2π
K
k(n−D
2 )
| {z }
modulation and phase shifting
(5.86)
fk(n) = q(n)
|{z}
synthesis prototype
· e
j 2π
K
k(n−D
2 )
| {z }
modulation and phase shifting
(5.87)
D denotes the overall analysis-synthesis system delay, for
linear phase filters we have D = LF − 1.
• Frequency responses:
Hk(e
jω
) = P (e
j(ω−2π
K
k)
)
| {z }
frequency shift by k 2π/K
· e
−j 2π
K
kD
2
| {z }
phase factor
(5.88)
Fk(e
jω
) = Q(e
j(ω−2π
K
k)
)
| {z }
frequency shift by k 2π/K
· e
−j 2π
K
kD
2
| {z }
phase factor
(5.89)
209
2π
K
2π
π
K
ω
π
|HK−1(ejω
)|
|H0(ejω
)|
|Hn(ejω
)| K = 8
• z-transforms:
Hk(z) = P (zW
k
K)W
kD
2
K , Fk(z) = Q(zW
k
K)W
nD
2
K
(5.90)
• Perfect reconstruction in the critical subsampled case is only
possible for filter lengths LF = K → not very good
frequency selectivity. Therefore, the DFT filter bank is mainly
used with M K.
Why the name DFT filter bank?
Type 1 polyphase components of the analysis prototype,
P (z) =
K−1
X
ℓ=0
z
−ℓ
Pℓ(z
K
). (5.91)
Inserting (5.91) into (5.90) then yields
Hk(z) = W
kD
2
K
K−1
X
ℓ=0
z
−ℓ
Pℓ(z
K
) W
−kℓ
K . (5.92)
Analogy to the IDFT: x(n) =
1
K
K−1
X
k=0
X(k) W
−kn
K
⇒ The subband filter of an DFT filter bank can be calculated
with the IDFT (and thus also with the fast IFFT), where the input
210

signals of the IDFT are obtained by filtering with the delayed
polyphase components
The n-th output of the IDFT has finally to be multiplied with the
rotation factors W
nD
2
K .
Structure of the analysis filter bank: .
x(k)
x(k 1)
x(k M+1)
.
PK 1(zK)
P1(zK)
P0(zK)
# N
# N
# N
x(k)
z 1
z 1
z 1
1
W 1
K
W (K 1)
K
KIDFT
yK 1(m)
y1(m)
y0(m)
WD=2
K
WD=2(K 1)
K
1
(dual structure for the synthesis)
If K = cM, c ∈ I
N, then besides the IDFT also the polyphase
filtering can be calculated in the lower sampling rate.
211
6. Spectral Estimation
In this chapter we consider the problem of estimating the power
spectral density (power spectrum) of a wide-sense stationary
(WSS) random process.
Applications for power spectrum estimation:
Signal detection and tracking, frequency estimation (e.g. for sonar
or radar signals), harmonic analysis and prediction, beamforming
and direction finding,...
Problems in power spectrum estimation:
• The amount of data is generally limited, for example, a random
process may be stationary only for a short time (e.g. speech
signal). On the other hand, as we have seen in Section 3.1.4 for
the frequency analysis of non-random signals using the DFT,
the longer the input data vector, the higher the frequency
resolution of the spectral analysis.
• Often, only one representation of a stochastic process may be
available. Therefore, an ergodic process is often assumed.
• Additionally, the input data may be corrupted by noise.
Estimation approaches can be classified into two classes:
1. Classical or nonparametric methods: Estimation of the
autocorrelation sequence, where the power spectrum is then
obtained by Fourier transform of the autocorrelation sequence.
2. Nonclassical or parametric approaches, which use a model for
the random process in order to estimate the power spectrum.
6.1 Periodogram-based nonparametric methods
6.1.1 The periodogram
Power spectrum Φvv(ejω
) of a WSS random process v(n) is the
212

Fourier transform of the autocorrelation function ϕvv(κ):
Φvv(e
jω
) =
∞
X
κ=−∞
ϕvv(κ) e
−jκω
.
⇒ Spectral estimation is also an autocorrelation estimation
problem
From Signals and Systems we know that for a stationary random
process v(n) which is ergodic in the first and second moments
the autocorrelation sequence can be theoretically obtained from
the time-average
ϕvv(κ) = lim
N→∞
1
2N + 1
N
X
k=−N
v(k + κ) v
∗
(k). (6.1)
If v(n) is only measured over a finite interval of N samples,
n = 0, . . . , N −1, the autocorrelation sequence is estimated as
ϕ̂vv(κ) =
1
N
N−1−κ
X
k=0
v(k + κ) v
∗
(k) (6.2)
with the values of ϕ̂vv(κ) for κ 0 defined via the symmetry
relation ϕ̂vv(−κ) = ϕ̂∗
vv(κ), and with ϕ̂vv(κ) = 0 for
|κ| ≥ N.
The discrete Fourier transform of ϕ̂vv(κ) from (6.2) is called
periodogram (Schuster, 1898) and leads to an estimate of the
power spectrum according to
Φ̂
(per)
vv (e
jω
) =
N−1
X
κ=−N+1
ϕ̂vv(κ)e
−jκω
. (6.3)
213
With the rectangular window
wr(n) =
(
1 n = 0, . . . , N − 1,
0 otherwise,
we can describe the input sequence being analyzed also as
vN(n) = v(n) · wr(n). (6.4)
Then, the estimated autocorrelation sequence may be written as
ϕ̂vv(κ) =
1
N
∞
X
k=−∞
vN(k+κ) v
∗
N(k) =
1
N
vN(κ)∗v
∗
N(−κ).
(6.5)
Fourier transform of the rightmost equation in (6.5)
finally yields the following expression for the periodogram
(VN(ejω
) •−◦ vN(n)):
Φ̂
(per)
vv (e
jω
) =
1
N
VN(e
jω
) V
∗
N(e
jω
) =
1
N
|VN(e
jω
)|
2
.
(6.6)
MATLAB-command: periodogram
The sampled version Φ̂(per)
vv (ejk2π/M
) = Φ̂(per)
vv (k) can be
obtained from the M1-point DFT VN(k) of vN(n):
Φ̂
(per)
vv (k) =
1
N
|VN(k)|
2
, k = 0, . . . , M1 − 1. (6.7)
Thus, the periodogram can be calculated very efficiently using the
FFT.
214

Example: Periodogram of white noise
If v(n) is white noise with a variance of σ2
v, then ϕvv(κ) = σ2
v δ(κ) with a
constant power spectrum Φvv(ejω
) = σ2
v.
Sample realization for N =32
0 5 10 15 20 25 30
−3
−2
−1
0
1
2
n →
v(n)
→
Autocorrelation sequence ϕ̂vv(κ)
−40 −20 0 20 40
−0.2
0
0.2
0.4
0.6
0.8
1
λ →
Periodogram Φ̂
(per)
vv (ejω
) (solid), power spectrum Φvv(ejω
) (dashed)
0 0.2 0.4 0.6 0.8 1
−15
−10
−5
0
5
10
ω / π →
Magnitude
[dB]
→
Definition: Bias and consistency
Desirable:
• Convergence of the periodogram to the exact power spectrum
in the mean-square sense
lim
N→∞
E

Φ̂
(per)
vv (e
jω
) − Φvv(e
jω
)
2

!
= 0. (6.8)
• In order to achieve this it is necessary that the periodogram is
215
asymptotically unbiased, which means that for N → ∞ the
expectation value of the estimated power spectrum is equal to
the true power spectrum:
lim
N→∞
E
n
Φ̂
(per)
vv (e
jω
)
o
!
= Φvv(e
jω
). (6.9)
On the other hand for a biased estimator there would be a
difference between the expectation value and the true result.
• Furthermore, the estimation variance should go to zero as the
data length N goes to infinity:
lim
N→∞
Var
n
Φ̂
(per)
vv (e
jω
)
o
!
= 0. (6.10)
• If (6.9) and (6.10) are satisfied we say that the periodogram
Φ̂(per)
vv (ejω
) is a consistent estimate of the power spectrum.
Note that there are different definitions of consistency in the literature.
Bias of the periodogram
First step: Calculation of the expected value of the autocorrelation
ϕ̂vv(κ). From (6.2) we have
E{ϕ̂vv(κ)} =
1
N
N−1−κ
X
k=0
E{v(k + κ) v
∗
(k)}
=
1
N
N−1−κ
X
k=0
ϕvv(κ) =
N − κ
N
ϕvv(κ) (6.11)
for κ = 0, . . . , N − 1, and for κ ≥ N it follows
E{ϕ̂vv(κ)} = 0.
By using the symmetry relation ϕ̂vv(−κ) = ϕ̂∗
vv(κ) (6.11) can
be written as E{ϕ̂vv(κ)} = wB(κ) ϕvv(κ) with the Bartlett
216

(triangular) window
wB(κ) =
( N−|κ|
N for |κ| ≤ N,
0 for |κ| N.
(6.12)
The expected value of the periodogram can now be obtained as
E
n
Φ̂
(per)
vv (e
jω
)
o
=
N−1
X
κ=−N+1
E{ϕ̂vv(κ)}e
−jκω
,
=
∞
X
κ=−∞
wB(κ) ϕvv(κ) e
−jκω
,
which finally yields
E
n
Φ̂
(per)
vv (e
jω
)
o
=
1
2π
Φvv(e
jω
) ∗ WB(e
jω
) (6.13)
with WB(ejω
) denoting the Fourier transform of the Bartlett
window
WB(e
jω
) =
1
N

sin(Nω/2)
sin(ω/2)
2
.
⇒ Periodogram is a biased estimate, since the expected value is
the convolution between the true power spectrum and the Fourier
transform of the Bartlett window.
Since WB(ejω
) converges to an impulse for N → ∞ the
periodogram is asymptotically unbiased:
lim
N→∞
E
n
Φ̂
(per)
vv (e
jω
)
o
= Φvv(e
jω
) (6.14)
217
Spectral resolution
We know from the discussion in Section 3.1.4 that the convolution
with the frequency response of a window may lead to
• spectral smoothing,
• spectral leaking,
• the loss of the ability to resolve two nearby spectral lines.
Similarly, this also holds for the convolution between the power
spectrum and the Bartlett window frequency response in (6.13).
Example: (a) Power spectrum of two sinusoids in white noise,
(b) expected value of the periodogram
(from [Hayes, 1996])
• Width of the main lobe of WB(ejω
) increases as the data
record length decreases.
• ⇒ For a given length N there is a limit on how closely two
sinusoids or narrowband processes may be located before they
no longer can be resolved.
218

• One way to define this frequency resolution limit is to set ∆ω
equal to the width of the main lobe of the Bartlett window at
its −6 dB point:
∆ω = 0.89
2π
N
, (6.15)
which is also the frequency resolution of the periodogram.
Variance of the periodogram
White Gaussian random processes:
It can be shown that for a white Gaussian random process v(n)
the variance of the periodogram is equal to the square of the power
spectrum Φvv(ejω
) (see [Hayes, 1996]):
Var
n
Φ̂
(per)
vv (e
jω
)
o
= Φ
2
vv(e
jω
). (6.16)
Non-white Gaussian random processes:
For non-white Gaussian processes, which are more important
in practice, we derive an approximation for the variance of the
periodogram in the following.
A random process v(n) with power spectrum Φvv(ejω
) may be
generated by filtering white noise x(n) with variance σ2
x = 1
with a linear filter h(n) ◦−• H(ejω
) and
|H(e
jω
)|
2
= Φvv(e
jω
). (6.17)
The sequences vN(n) and xN(n) are now formed by windowing
219
analog to (6.4). The periodograms of these processes are
Φ̂
(per)
vv (e
jω
) =
1
N
|VN(e
jω
)|
2
, Φ̂
(per)
xx (e
jω
) =
1
N
|XN(e
jω
)|
2
.
(6.18)
If N is large compared to the length of h(n), vN(n) can be
described as vN(n) ≈ h(n)∗xN(n), since the transition effects
can be neglected. Thus, the magnitude square frequency response
|VN(ejω
)|2
of vN(n) can be with (6.17) expressed as
|VN(e
jω
)|
2
≈|H(e
jω
)|
2
|XN(e
jω
)|
2
=Φvv(e
jω
) |XN(e
jω
)|
2
.
(6.19)
Φ̂
(per)
vv (e
jω
) ≈ Φvv(e
jω
) Φ̂
(per)
xx (e
jω
).
Applying the variances on both sides results in
Var
n
Φ̂
(per)
vv (e
jω
)
o
≈ Φ
2
vv(e
jω
) Var
n
Φ̂
(per)
xx (e
jω
)
o
,
and, since Var{Φ̂(per)
xx (ejω
)}=1 according to (6.16), the variance
for large N can be obtained as
Var
n
Φ̂
(per)
vv (e
jω
)
o
≈ Φ
2
vv(e
jω
). (6.20)
⇒ Periodogram is not a consistent estimator
Example:
For a white Gaussian noise process v(n) with σ2
v = 1 and Φvv(ejω
) = 1 it
follows from (6.13) and (6.16), resp., that
E
n
Φ̂
(per)
vv (e
jω
)
o
= 1 and Var
n
Φ̂
(per)
vv (e
jω
)
o
= 1.
220

Thus, although the periodogram for white Gaussian noise is unbiased, the
variance is independent of the data record length N.
N = 64, overlay of 30 periodo-
grams Φ̂
(per)
vv (ejω
)
0 0.2 0.4 0.6 0.8 1
−50
−40
−30
−20
−10
0
10
ω / π →
Magnitude
[dB]
→
N = 64, approximated periodo-
gram variance
0 0.2 0.4 0.6 0.8 1
0
0.5
1
1.5
2
2.5
3
ω / π →
Variance
→
N = 256, overlay of 30 periodo-
grams Φ̂
(per)
vv (ejω
)
0 0.2 0.4 0.6 0.8 1
−50
−40
−30
−20
−10
0
10
ω / π →
Magnitude
[dB]
→
N = 256, approximated periodo-
gram variance
0 0.2 0.4 0.6 0.8 1
0
0.5
1
1.5
2
2.5
3
ω / π →
Variance
→
N =64, periodogram average
0 0.2 0.4 0.6 0.8 1
−50
−40
−30
−20
−10
0
10
ω / π →
Magnitude
[dB]
→
N =256, periodogram average
0 0.2 0.4 0.6 0.8 1
−50
−40
−30
−20
−10
0
10
ω / π →
Magnitude
[dB]
→
221
6.1.2 Bartlett’s method: Periodogram averaging
In contrast to the periodogram, Bartlett’s method (1948) provides
a consistent estimate of the power spectrum.
The motivation for this method comes from the fact that by
averaging a set of uncorrelated measurements for a random
variable v one obtains a consistent estimate of the mean E{v}.
Since the periodogram is asymptotically unbiased
lim
N→∞
E
n
Φ̂
(per)
vv (e
jω
)
o
= Φvv(e
jω
),
it obviously suffices to find a consistent estimate of the
periodogram E{Φ̂(per)
vv (ejω
)} in order to find a consistent
estimate for the true power spectrum Φvv(ejω
).
⇒ Estimation of the power spectrum by periodogram averaging!
Let vi(n) for i = 0, . . . , K − 1 denote K uncorrelated
realizations of a random process v(n) for n = 0, . . . , L−1.
The periodogram of each single realization is obtained from (6.6)
as
Φ̂
(per)
vivi
(e
jω
) =
1
L
L−1
X
n=0
vi(n)e
−jnω
2
. (6.21)
The average of these periodograms is
Φ̂vv(e
jω
) =
1
K
K−1
X
i=0
Φ̂
(per)
vivi
(e
jω
). (6.22)
For the expected value of Φ̂vv(ejω
) we have with (6.22) and
222

(6.13)
E
n
Φ̂vv(e
jω
)
o
= E
n
Φ̂
(per)
vivi
(e
jω
)
o
=
1
2π
Φvv(e
jω
) ∗ WB(e
jω
). (6.23)
As with the periodogram, the estimate Φ̂vv(ejω
) is asymptotically
unbiased, i.e. for L → ∞.
For uncorrelated data records vi(n) the variance Var{Φ̂vv(ejω
)}
can be obtained in the same way from (6.22) and (6.20) as
Var
n
Φ̂vv(e
jω
)
o
=
1
K
Var
n
Φ̂
(per)
vivi
(e
jω
)
o
≈
1
K
Φ
2
vv(e
jω
).
(6.24)
We can observe that the variance goes to zero if K goes to infinity
⇒ Φ̂vv(ejω
) is a consistent estimate of the power spectrum if
both L and K tend to infinity.
In practice:
Uncorrelated realizations of a random process are generally not
available, instead we have only one single realization of length N.
Alternative: v(n) of length N is divided into K nonoverlapping
sequences of length L with N = L · K, that is
vi(n) = v(n + iL), n = 0, . . . , L−1, i = 0, . . . , K−1.
Thus, the Bartlett estimate is
Φ̂
(B)
vv (e
jω
) =
1
N
K−1
X
i=0
L−1
X
n=0
v(n + iL) e
−jnω
2
. (6.25)
223
Properties
From (6.23) the expected value is
E
n
Φ̂
(B)
vv (e
jω
)
o
=
1
2π
Φvv(e
jω
) ∗ WB(e
jω
) (6.26)
As the periodogram the Bartlett estimate is asymptotically
unbiased.
The spectral resolution of the Bartlett estimate can be obtained
from the resolution of the periodogram in (6.15). Since we now
use sequences of length L the resolution becomes
∆ω = 0.89
2π
L
= 0.89 K
2π
N
, (6.27)
which is K times larger (worse!) than that of the periodogram.
Variance: Assuming that the data sequences vi(n) are
approximately uncorrelated (this is generally not the case!) the
variance of the Bartlett estimate is for large N
Var
n
Φ̂
(B)
vv (e
jω
)
o
≈
1
K
Φ
2
vv(e
jω
). (6.28)
• Φ(B)
vv (ejω
) is a consistent estimate for K, L → ∞.
• The Bartlett estimate allows to trade spectral resolution for a
reduction in variance by adjusting the parameters L and K.
Examples:
• The power spectrum of a white noise Gaussian process with σ2
v = 1 of
length N =256 is estimated with Bartlett’s method.
224

K = 4, L = 64, overlay of 30
Bartlett estimates Φ̂
(B)
vv (ejω
)
0 0.2 0.4 0.6 0.8 1
−15
−10
−5
0
5
10
ω / π →
Magnitude
[dB]
→ K = 4, L = 64, approximated
variance of Bartlett estimates
0 0.2 0.4 0.6 0.8 1
0
0.1
0.2
0.3
0.4
0.5
ω / π →
Variance
→
K = 8, L = 32, overlay of 30
Bartlett estimates Φ̂
(B)
vv (ejω
)
0 0.2 0.4 0.6 0.8 1
−15
−10
−5
0
5
10
ω / π →
Magnitude
[dB]
→
K = 8, L = 32, approximated
variance of Bartlett estimates
0 0.2 0.4 0.6 0.8 1
0
0.1
0.2
0.3
0.4
0.5
ω / π →
Variance
→
K = 4, L = 64, Bartlett estimate
average
0 0.2 0.4 0.6 0.8 1
−15
−10
−5
0
5
10
ω / π →
Magnitude
[dB]
→
K = 8, L = 32, Bartlett estimate
average
0 0.2 0.4 0.6 0.8 1
−15
−10
−5
0
5
10
ω / π →
Magnitude
[dB]
→
• Here, the input signal consists of two sinusoids in white Gaussian noise
225
η(n) of variance σ2
η = 1,
v(n) = 3 · sin(nω1) + sin(nω2) + η(n) (6.29)
with ω1 = 0.2π, ω2 = 0.25π, and length N = 512 samples.
The following figures show the average power spectrum estimate over
30 realizations, and demonstrate the reduced spectral resolution of the
Bartlett estimate compared to the periodogram.
K =4, L=128, Bartlett estimate
0 0.2 0.4 0.6 0.8 1
−10
0
10
20
30
ω / π →
Magnitude
[dB]
→
K =8, L=64, Bartlett estimate
0 0.2 0.4 0.6 0.8 1
−10
0
10
20
30
ω / π →
Magnitude
[dB]
→
Periodogram
0 0.2 0.4 0.6 0.8 1
−10
0
10
20
30
ω / π →
Magnitude
[dB]
→
6.1.3 Welch’s method: Averaging modified periodograms
In 1967, Welch proposed two modifications to Bartlett’s method:
1. The data segments vi(n) of length L are allowed to overlap,
where D denotes the offset between successive sequences:
vi(n) = v(n+iD), n = 0, . . . , L−1, i = 0, . . . , K−1.
(6.30)
226

The amount of overlap between vi(n) and vi+1(n) is L−D
samples, and if K sequences cover the entire N data points
we have N = L + D(K − 1). If D = L the segments do
not overlap as in Bartlett’s method with K = N/L.
⇒ By allowing the sequences to overlap it is possible to
increase the number and/or length of the sequences that are
averaged. Reduction of the variance (for larger K) can thus
be traded in for a reduction in resolution (for smaller L) and
vice versa.
2. The second modification is to window the data segments prior
to computing the periodogram. This leads to a so called
modified periodogram
Φ̂
(mod)
vivi
(e
jω
) =
1
LU
L−1
X
n=0
vi(n) w(n)e
−jnω
2
(6.31)
with a general window w(n) of length L, and U denoting
a normalization factor for the power in the window function
according to
U =
1
L
L−1
X
n=0
|w(n)|
2
. (6.32)
Welch’s method may explicitly be written as
Φ̂
(W )
vv (e
jω
) =
1
KLU
K−1
X
i=0
L−1
X
n=0
v(n + iD) w(n)e
−jnω
2
.
(6.33)
MATLAB-command: pwelch
227
Properties
• It can be shown that the expected value of Welch’s estimate is
E
n
Φ̂
(W )
vv (e
jω
)
o
=
1
2πLU
Φvv(e
jω
) ∗ |W (e
jω
)|
2
,
(6.34)
where W (ejω
) denotes the Fourier transform of the general
L-point window sequence w(n). Thus, Welch’s method is an
asymptotically unbiased estimate of the power spectrum.
• The spectral resolution of Welch’s estimate depends on the
used window sequence and is specified as the 3 dB width
∆ω3 dB of the main lobe of the spectral window. ∆ω3 dB is
specified for some commonly used windows in the following
table.
Sidelobe 3 dB bandwidth
Type of window level [dB] ∆ω3 dB
Rectangular -13 0.89 2π
L
Bartlett -27 1.28 2π
L
Hanning -32 1.44 2π
L
Hamming -43 1.30 2π
L
Blackman -58 1.68 2π
L
Remark: In (6.15) we stated the frequency resolution of the peridogram
as the 6 dB main lobe width of the Bartlett window. Since WB(ejω
) =
|WR(ejω
)|2
this is equivalent to the 3 dB bandwidth of the frequency
response WR(ejω
) of the rectangular window.
• The variance of Welch’s estimate highly depends on the
228

amount of overlapping. For a Bartlett window and a 50%
overlap the variance is approximately
Var
n
Φ̂
(W )
vv (e
jω
)
o
≈
9
8K
Φ
2
vv(e
jω
) (6.35)
(→ consistent estimate). A comparison with (6.28) shows that
the variance for Welch’s method seems to be larger than for
Bartlett’s method. However, for fixed amount of data N and a
fixed resolution L here twice as many sections are averaged
compared to Bartlett’s method. With K = 2N/L (50%
overlap) (6.35) becomes
Var
n
Φ̂
(W )
vv (e
jω
)
o
≈
9L
16N
Φ
2
vv(e
jω
). (6.36)
A comparison with (6.28) and K = N/L for the Bartlett
estimate we have
Var
n
Φ̂
(W )
vv (e
jω
)
o
≈
9
16
Var
n
Φ̂
(B)
vv (e
jω
)
o
. (6.37)
Increasing the amount of overlap yields higher computational
complexity and also the correlation between the subsequences
vi(n) → amount of overlap is typically chosen as 50% or
75%.
Example:
As an input signal we again use (6.29) which contains two sinusoids in white
Gaussian noise η(n) of variance σ2
η = 1, with ω1 = 0.2π, ω2 = 0.25π,
and a signal length of N = 512 samples. The section length is chosen as
L = 128, the amount of overlapping is 50%, and for the window we use a
Hamming window.
229
Overlay plot of 30 Welch estimates
0 0.2 0.4 0.6 0.8 1
−10
0
10
20
30
ω / π →
Magnitude
[dB]
→
Welch estimate ensemble average
0 0.2 0.4 0.6 0.8 1
−10
0
10
20
30
ω / π →
Magnitude
[dB]
→
Compared to the Bartlett estimate for the same example above the use of
the Hamming window reduces the spectral leakage in the estimated power
spectrum.
Since the number of sections (7) are about the same to those in the above
example for the Bartlett estimate with K = 8, L = 64 (8 sections) both
variances are also approximately the same.
6.1.4 Blackman-Tukey method: Periodogram smoothing
Recall that the periodogram is obtained by a Fourier transform
from the estimated autocorrelation sequence. However, for any
finite data record of length N the variance of ϕ̂vv(κ) will be
large for values of κ, which are close to N. For example for
lag κ = N − 1 we have from (6.2)
ϕ̂vv(N − 1) =
1
N
v(N − 1) v(0).
Two approaches for reducing the variance of ϕ̂vv(κ) and thus also
the variance of the peridogram:
1. Averaging periodograms and modified periodograms, resp., as
utilized in the methods of Bartlett and Welch.
2. Periodogram smoothing → Blackman-Tukey method (1958)
230

Blackman-Tukey method: Variance of the autocorrelation
function is reduced by applying a window to ϕ̂vv(κ) to decrease
the contribution of the undesired estimates to the periodogram.
The Blackman-Tukey estimate is given as
Φ̂
(BT )
vv (e
jω
) =
M
X
κ=−M
ϕ̂vv(κ) w(κ) e
−jκω
, (6.38)
where w(κ) is a lag window being applied to the autocorrelation
estimate and extending from −M to M for M N − 1.
⇒ Estimates of ϕvv(κ) having the largest variance are set to zero
by the lag window → the power spectrum estimate will have a
smaller variance.
The Blackman-Tukey power spectrum estimate from (6.38) may
also be written as
Φ̂
(BT )
vv (e
jω
) =
1
2π
Φ̂
(per)
vv (e
jω
) ∗ W (e
jω
), (6.39)
=
1
2π
π
Z
−π
Φ̂
(per)
vv (e
ju
) W (e
j(ω−u)
) du (6.40)
with W (ejω
) denoting the Fourier transform of the lag window.
⇒ Blackman-Tukey estimate smooths the periodogram by
convolution with W (ejω
).
Choice of a suitable window:
• w(κ) should be conjugate symmetric, such that W (ejω
) (and
also the power spectrum) is real-valued.
231
• W (ejω
) ≥ 0, such that Φ̂(BT )
vv (ejω
) is nonnegative for all ω.
Note that some of the window functions we have introduced
do not satisfy this condition, for example, the Hamming and
Hanning windows.
Properties
• The expected value of the Blackman-Tukey estimate can be
derived for M ≪ N as
E
n
Φ̂
(BT )
vv (e
jω
)
o
=
1
2π
Φvv(e
jω
) ∗ W (e
jω
)
(6.41)
where W (ejω
) is the Fourier transform of the lag window.
• The spectral resolution of the Blackman-Tukey estimate
depends on the used window.
• It can be shown that the variance can be approximated as
Var
n
Φ̂
(BT )
vv (e
jω
)
o
≈ Φ
2
vv(e
jω
)
1
N
M
X
κ=−M
w
2
(κ).
(6.42)
• From (6.41) and (6.42) we again see the trade-off between bias
and variance: For a small bias, M should be large in order to
minimize the width of the main lobe of W (ejω
), whereas M
should be small to minimize the sum term in (6.42). As a
general rule of thumb, M is often chosen as M = N/5.
Examples:
• The power spectrum of a white noise Gaussian process with σ2
v = 1 of
length N = 256 is estimated with the Blackman-Tukey method, where a
Bartlett window with M = 51 is used.
232

Overlay of 30 Blackman-Tukey
estimates Φ̂
(BT )
vv (ejω
)
0 0.2 0.4 0.6 0.8 1
−15
−10
−5
0
5
10
ω / π →
Magnitude
[dB]
→
Approximated variance of the
Blackman-Tukey estimates
0 0.2 0.4 0.6 0.8 1
0
0.1
0.2
0.3
0.4
0.5
ω / π →
Variance
→
Blackman-Tukey estimate ensemble average
0 0.2 0.4 0.6 0.8 1
−15
−10
−5
0
5
10
ω / π →
Magnitude
[dB]
→
• In a second example we use two sinusoids in white Gaussian noise ((6.29),
σ2
n = 1, ω1 = 0.2π, ω2 = 0.25π) for N = 512 samples. The
window is a Bartlett window with M = 102.
Overlay plot of 30 Blackman-
Tukey estimates
0 0.2 0.4 0.6 0.8 1
−10
0
10
20
30
ω / π →
Magnitude
[dB]
→
Blackman-Tukey estimate ensem-
ble average
0 0.2 0.4 0.6 0.8 1
−10
0
10
20
30
ω / π →
Magnitude
[dB]
→
233
6.1.5 Performance comparisons
Performance of the discussed estimators is assessed in terms of
two criteria:
1. Variability V of the estimate,
V =
Var
n
Φ̂vv(ejω
)
o
E2
n
Φ̂vv(ejω)
o , (6.43)
which can be regarded as a normalized variance.
2. Overall figure of merit, which is defined as the product of the
variability and the spectral resolution ∆ω,
M = V ∆ω. (6.44)
Results for the periodogram-based spectrum estimation tech-
niques:
Variability Resolution Figure of merit
V ∆ω M
Periodogram 1 0.89
2π
N
0.89
2π
N
Bartlett
1
K
0.89 K
2π
N
0.89
2π
N
Welch (50% overlap,
Bartlett window)
9
8 K
1.28
2π
L
0.72
2π
N
Blackman-Tukey
(Bartlett window of length
2M, 1 ≪ M ≪ N)
2 M
3 N
0.64
2π
M
0.43
2π
N
234

• Each technique has a figure of merit being approximately the
same, figures of merit are inversely proportional to the length
N of the data sequence.
• ⇒ Overall performance is fundamentally limited by the
amount of data being available!
6.2 Parametric methods for power spectrum estimation
Disadvantages of the periodogram-based (nonparametric)
methods:
• Long data records are required for sufficient performance,
windowing of the autocorrelation sequence limits spectral
resolution.
• Spectral leakage effects due to windowing.
• A-priori information about the generating process is not
exploited.
Disadvantages are removed by using parametric estimation
approaches, where an appropriate model for the input process is
applied.
Parametric methods are based on modeling the data sequence as
the output of a linear system with the transfer function (IIR filter!)
H(z) =
q
P
i=0
bi z−i
1 +
p
P
i=1
ai z−i
, (6.45)
where the ai and bi are the model parameters. The corresponding
235
difference equation is
v(n) =
q
X
i=0
bi w(n − i) −
p
X
i=1
ai v(n − i), (6.46)
where v(n) denotes the output and w(n) the input sequence
If w(n) represents a stationary random process, then v(n) is also
stationary and the power spectrum can be given as (Wiener-Lee
relation)
Φvv(e
jω
) = |H(e
jω
)|
2
Φww(e
jω
). (6.47)
Parametric power spectrum estimation:
H(ejω
)
w(n)
source, µ=0
white noise
ai, bi
Φvv(ejω
) = σ2
w · |H(ejω
)|2
stationary random process σw
v(n)
!
= received process
model process
In order to estimate the power spectrum Φvv(ejω
) we assume in
our model that Φww(ejω
) comes from a zero-mean white noise
process with variance σ2
w. By inserting Φww(ejω
) = σ2
w into
(6.47) the power spectrum of the observed data is
Φvv(e
jω
) = σ
2
w |H(e
jω
)|
2
= σ
2
w
|B(ejω
)|2
|A(ejω)|2
. (6.48)
Goal: Make the model process v(n) as similar as possible to
the unknown received process in the mean-square error sense by
adjusting the parameters ai, bi, and σw ⇒ the power spectrum
Φvv(ejω
) can then be obtained via (6.48).
236

In the following we distinguish among three specific cases for
H(z) leading to three different models:
Autoregressive (AR) process
The linear filter H(z) = 1/A(z) is an all-pole filter, leading to
b0 = 1, bi = 0 for n 0 in the difference equation (6.46):
v(n) = w(n) −
p
X
i=1
ai v(n − i). (6.49)
Moving average (MA) process
Here, H(z) = B(z) is an all-zero (FIR!) filter, and the
difference equation becomes with an = 0 for n ≥ 0
v(n) =
q
X
i=0
bi w(n − i). (6.50)
Autoregressive, moving average (ARMA) process
In this case the filter H(z) = B(z)/A(z) has both finite poles
and zeros in the z-plane and the corresponding difference equation
is given by (6.46).
Remarks:
• The AR model is most widely used, since the AR model is
suitable of modeling spectra with narrow peaks (resonances)
by proper placement of the poles close to the unit circle.
• MA model requires more coefficients for modeling a narrow
spectrum, therefore it is rarely used as a model for spectrum
estimation.
• By combining poles and zeros the ARMA model has a more
efficient spectrum representation as the MA model concerning
the number of model parameters.
237
6.2.1 Relationship between the model parameters and the
autocorrelation sequence
In the following it is shown that the model parameters ai, bi can
be obtained from the autocorrelation sequence of the observed
process v(n). These values are then inserted into (6.45) yielding
H(ejω
), which is then inserted into (6.48) leading to the power
spectrum Φvv(ejω
) of our observed process.
In a first step the difference equation (6.46) is multiplied by
v∗
(n − κ) and the expected value is taken on both sides
E{v(n) v
∗
(n − κ)} =
q
X
i=0
bi E{w(n − i) v
∗
(n − κ)}−
−
p
X
i=1
ai E{v(n − i) v
∗
(n − κ)},
which leads to
ϕvv(κ) =
q
X
i=0
bi ϕwv(κ − i) −
p
X
i=1
ai ϕvv(κ − i). (6.51)
The crosscorrelation sequence ϕwv(κ) depends on the filter
impulse response:
ϕwv(κ) = E{v
∗
(n) w(n + κ)},
= E
( ∞
X
k=0
h(k) w
∗
(n − k) w(n + κ)
)
= σ
2
w h(−κ).
(6.52)
238

In the last step we have used our prerequisite from above that the
process w(n) is assumed to be a zero-mean white random process
with E{w(n − k)w∗
(n + κ)} = δ(κ + k)σ2
w and known
variance σ2
w. Thus we have from (6.52)
ϕwv(κ) =
(
0 for κ 0,
σ2
w h(−κ) for κ ≤ 0.
(6.53)
By combining (6.51) and (6.53) we obtain the desired relationship
for the general ARMA case:
ϕvv(κ) =













−
p
P
i=1
ai ϕvv(κ − i) for κ q,
σ2
w
q−κ
P
i=0
h(i) bi+κ −
p
P
i=1
ai ϕvv(κ − i) for 0 ≤ κ ≤ q
ϕ∗
vv(−κ) for κ 0.
(6.54)
→ nonlinear relationship between the parameters ϕvv(κ) and
ai, bi
In the following we only consider the AR model case, where
(6.54) simplifies to
ϕvv(κ) =













−
p
P
i=1
ai ϕvv(κ − i) for κ 0,
σ2
w −
p
P
i=1
ai ϕvv(κ − i) for κ = 0
ϕ∗
vv(−κ) for κ 0.
(6.55)
These equations are also called Yule-Walker equations and denote
a system of linear equations for the parameters ai. Equation (6.55)
239
may also be expressed in matrix notation according to





ϕvv(0) ϕ∗
vv(1) . . . ϕ∗
vv(p − 1)
ϕvv(1) ϕvv(0) . . . ϕ∗
vv(p − 2)
.
.
.
.
.
. ... .
.
.
ϕvv(p − 1) ϕvv(p − 2) . . . ϕvv(0)










a1
a2
.
.
.
ap





= −





ϕvv(1)
ϕvv(2)
.
.
.
ϕvv(p)





(6.56)
which is in short-hand notation R a = r. Once the ai have been
obtained by solving for the coefficient vector a, the variance σ2
w
can be calculated from
σ
2
w = ϕvv(0) +
p
X
i=1
aiϕvv(−i). (6.57)
Since the matrix R has a special structure (Toeplitz structure) there
exist efficient algorithms for solving the system of linear equations
in (6.55) and (6.56), respectively (Levison-Durbin algorithm
(1947, 1959), Schür recursion (1917)).
6.2.2 Yule-Walker method
In the Yule-Walker method (also called autocorrelation method)
we simply estimate the autocorrelation sequence from the
observed data v(n), where the autocorrelation estimate in (6.2)
is used:
ϕ̂vv(κ) =
1
N
N−1−κ
X
n=0
v(n + κ) v
∗
(n), κ = 0, . . . , p.
(6.58)
In the matrix version of the Yule-Walker equations (6.56) we
replace ϕvv(κ) with ϕ̂vv(κ). The resulting linear equation
system is solved for the parameter vector a, which now contains
the estimated AR parameters âi, i = 1, . . . , p. Finally, we
240

obtain σ2
w via (6.57) from the âi and the estimated autocorrelation
sequence ϕ̂vv(κ).
The corresponding power spectrum estimate can now be stated
from (6.48) as
Φ̂
(AR)
vv (e
jω
) =
σ2
w
1 +
p
P
k=1
âk e−jkω
2
. (6.59)
MATLAB-command: paryule
6.2.3 Examples and comparisons to nonparametric methods
In the following we compute power spectrum estimates obtained
with the Yule-Walker method and for comparison purposes also
with the Blackman-Tukey method (Bartlett window of length
LB).
The input process consists of two sinusoids in white Gaussian noise η(n) of
variance σ2
η = 1 according to (6.29) with ω1 = 0.2π, ω2 = 0.25π. The
model order of the Yule-Walker method is chosen as p = 50.
• Length of observed process N =512 (Blackman-Tukey: LB =205):
overlay of 30 Yule-Walker estimates
0 0.2 0.4 0.6 0.8 1
−10
0
10
20
30
40
ω / π →
Magnitude
[dB]
→
Yule-Walker ensemble average
0 0.2 0.4 0.6 0.8 1
−10
0
10
20
30
40
ω / π →
Magnitude
[dB]
→
241
overlay of 30 Blackman-Tukey estimates
0 0.2 0.4 0.6 0.8 1
−10
0
10
20
30
40
ω / π →
Magnitude
[dB]
→
Blackman-Tukey ensemble average
0 0.2 0.4 0.6 0.8 1
−10
0
10
20
30
40
ω / π →
Magnitude
[dB]
→
• Length of observed process N = 100 (Blackman-Tukey: LB = 41):
overlay of 30 Yule-Walker estimates
0 0.2 0.4 0.6 0.8 1
−10
0
10
20
30
40
ω / π →
Magnitude
[dB]
→
Yule-Walker ensemble average
0 0.2 0.4 0.6 0.8 1
−10
0
10
20
30
40
ω / π →
Magnitude
[dB]
→
overlay of 30 Blackman-Tukey estimates
0 0.2 0.4 0.6 0.8 1
−10
0
10
20
30
40
ω / π →
Magnitude
[dB]
→
Blackman-Tukey ensemble average
0 0.2 0.4 0.6 0.8 1
−10
0
10
20
30
40
ω / π →
Magnitude
[dB]
→
• We can see that only for the longer data sequence with N = 512 the
resolution of the estimates are comparable. Clearly, for N = 100 the
estimate based on an AR-model provides much better frequency resolution
for the sinusoidal components than the Blackman-Tukey method.
242

Remark: Use of a certain model generally requires a-priori
knowledge about the process. In case of a model mismatch
(e.g. MA process and AR model) using a nonparametric approach
may lead to a more accurate estimate.
Example:
Consider the MA process (length N = 512)
v(n) = w(n) − w(n − 2),
where w(n) is again a white-noise zero-mean process with variance σ2
w = 1.
The power spectrum of v(n) is
Φvv(e
jω
) = 2 − 2 cos(2ω).
Ensemble average over 30 power spectrum estimates for the Yule-Walker
method (AR model of order p = 4) and the Blackman-Tukey method (Bartlett
window, LB = 205):
0 0.2 0.4 0.6 0.8 1
−20
−15
−10
−5
0
5
10
ω / π →
Magnitude
[dB]
→
Yule−Walker (AR(4) model)
Blackman−Tukey
exact power spectrum
→ Blackman-Tukey estimate, which makes no assumption about the process,
yields a better estimate of the power spectrum compared to the model-based
Yule-Walker approach.
243

Advanced_DSP_J_G_Proakis.pdf

More Related Content

Similar to Advanced_DSP_J_G_Proakis.pdf (20)

Recently uploaded (20)

Advanced_DSP_J_G_Proakis.pdf