Sampling and Multirate Techniques For Complex and Bandpass Signals

Sampling and Multirate Techniques for Complex and Bandpass Signals
M. Renfors, TUT/DCE
TLT-5806/IQ/1
21.9.2010
Sampling and Multirate Techniques for

Complex and Bandpass Signals
Markku Renfors
Department of Communications Engineering
Tampere University of Technology, Finland
markku.renfors@tut.fi
www.cs.tut.fi/tlt
DCE - Tampere University of Technology. All rights reserved.

M. Renfors, TUT/DCE
TLT-5806/IQ/2
21.9.2010
Contents
Complex signals and systems
o
o
o
o
Basic concepts and definitions

Analytic signals and Hilbert transform
Frequency translation by mixing
Complex bandpass filters
Sampling and multirate DSP with complex and bandpass

signals
o Sampling theory for complex signals
o Multirate processing of real and complex bandpass signals
o Combining mixing and multirate operations for frequency
translation
Bandpass sampling principles and related practical

issues
o
o
o
o
o
Real bandpass sampling

Quadrature sampling
Second-order sampling
Sampling jitter and quantization noise
Sigma-delta ADC
Complex and bandpass polyphase structures

o Polyphase structures for decimation and interpolation
o Real and complex filter banks

M. Renfors, TUT/DCE
TLT-5806/IQ/3
21.9.2010
Complex Signals and Systems

In telecommunications signal processing, it is common to
use the notion of complex signals.
Continuous- and
denoted here as
discrete-time
x (t ) = x R (t ) + jx I (t )
complex
signals
are
x ( k ) = x R ( k ) + jx I ( k )
In practical implementations, complex signals are nothing

more than two separate real signals carrying the real and
imaginary parts.
A complex linear time-invariant system is represented by
two real impulse responses
h( k ) = hR ( k ) + jhI ( k )
or the corresponding two real-coefficient transfer functions

H (e j ) = H R (e j ) + jH I ( e j )
In the general case, to implement a complex filter for a

complex signal, four separate real filters need to be
implemented:
y ( k ) = x ( k ) h( k ) = ( x R ( k ) + jx I ( k )) (hR ( k ) + jhI ( k ))
= x R ( k ) hR ( k ) x I ( k ) hI ( k ) + j ( x R ( k ) hI ( k ) + x I ( k ) hR ( k ))

M. Renfors, TUT/DCE
TLT-5806/IQ/4
21.9.2010
Transforming Complex to Real

Real Part -Operation
Taking the real part of the complex signal produces mirror
images of all the spectral components:
Taking real part
Aliasing takes place, if any of the mirror images overlaps

with any of the original spectral components:
Taking real part
f
0
f
0

M. Renfors, TUT/DCE
TLT-5806/IQ/5
21.9.2010
Transforming Real to Complex

Analytic Signal and Hilbert Transform
Hilbert transformer is generally defined as an allpass linear
filter which shifts the phase of its input signal by 90
degrees. The (anticausal) impulse and frequency
responses can be formulated as
discrete-time
continuous-time
n even
0,
1
hHT (n) =
hHT (t ) =
t
2 /( n), n odd
j, f 0
j, 0 <
j
H HT ( f ) =
H HT (e ) =
j
f
,
0
+
<
+ j , < 0
Hilbert filters can be used to construct analytic signals with
only positive (or negative) frequency content:
input spectrum
I
INPUT
OUTPUT
HT
output spectrum
This structure is also called as phase splitter.


M. Renfors, TUT/DCE
TLT-5806/IQ/6
21.9.2010
Analytic Signal at Baseband

Real signal spectrum, the corresponding analytic signal
spectrum, and the analytic signal spectrum translated to be
centered at 0-frequency:
X( f )
f
0
FREQUENCY
TRANSLATION
FILTERING
X( f ) + jX( f )
f
W/2
f
0
W/2
In any practical design, there is a non-zero, usually

symmetric, transition band around 0-frequency.
One practical example is VSB (vestigial side-band)
modulation, the baseband model of which is obtained by a
Hilbert filter with symmetric transition band.

M. Renfors, TUT/DCE
TLT-5806/IQ/7
21.9.2010
Analytic Bandpass Signals

Real bandpass signal and the corresponding ideal analytic
bandpass signal:
f
FILTERING
f
0
Other side of the desired signal spectrum suppressed by a

practical phase splitter with finite attenuation (e.g., for
improving
image
attenuation
in
case
of
I/Q
downconversion):
f
0
fc
To realize such a system, it is sufficient to design an

allpass filter which approximates 90 phase shift in the
passband and stopband with sufficient accuracy,
depending on the stopband attenuation requirement.

M. Renfors, TUT/DCE
TLT-5806/IQ/8
21.9.2010
Frequency Translation
One key operation in communications signal
processing is the frequency translation of a signal
spectrum from one center frequency to another.
Conversions between baseband and bandpass
representations (modulation and demodulation) are
special cases of this.
We consider two different ways to do the frequency
translation: mixing and multirate operations, i.e.,
decimation and interpolation.
In case of multirate operations, we assume for
simplicity that the following two sampling rates are
used:
low sampling rate:
fs
high sampling rate:
fs = 1
T
=1
NT

M. Renfors, TUT/DCE
TLT-5806/IQ/9
21.9.2010
Mixing for Complex Discrete-Time Signals

y (k ) = e j 2 f LO kT x(k ) = e jLO k x(k )
Special case with
real input signal:
jLOk
I
cos(Lok)
fc
sin(Lok)
fc+fLO
This produces a pure frequency translation of the spectrum

by f LO .
Important special cases are:
f LO = f s / 2 = 1
2T
in which case the multiplying sequence is +1, -1, +1, -1, ...
This case can be applied to a real signal without
producing a complex result. Converts a lowpass signal
to a highpass signal, and vice versa.
f LO = f s / 4 = 1
4T
in which case the multiplying sequence is

+1, j, -1, -j, +1, j, ...

M. Renfors, TUT/DCE
TLT-5806/IQ/10
21.9.2010
Complex Bandpass Filters

Certain types of complex filters based on Hilbert
transformers can be designed using standard filter design
packages, like Parks-McClellan routine for FIR filters.
Another way to get complex bandpass filters is through
frequency translations:
Real
prototype
filter:
f
0
Complex
bandpass
filter:
f
fc
Transformation for frequency response and transfer

function:
H e j H e j ( c )
H ( z ) H ze j 2f cT
( )
Generic transformation for block diagram:

T
e j2f T
c
If 1/T is an integer multiple of fc, this might be much easier

than in the general case, see the special cases of the
previous page.

M. Renfors, TUT/DCE
TLT-5806/IQ/11
21.9.2010
Example of a Complex Bandpass Filters:

Frequency Translated FIR
Frequency translation by fs/4 => Analytic bandpass
filter with passband around fs/4.
fs/4
fs/4
Impulse response translated as:

h0, h1, h2, h3, h4, , hN
h0, jh1, -h2, -jh3, h4, , (j)NhN
fs/2
fs/2

M. Renfors, TUT/DCE
TLT-5806/IQ/12
21.9.2010
FIR Filter with Frequency Translation by fs/4

(i) Real input signal
I
h2
h0
h4
h3
h1
...
T
h5
(ii) Complex input signal

I
T
h0
h1
h2
...
h3 h4
h5
I
Q
h0
Q
h1
h2
h3
h4
h5
...
There are possibilities to exploit the possible

coefficient symmetry (of linear phase FIR) in both
cases.

M. Renfors, TUT/DCE
TLT-5806/IQ/13
21.9.2010
Contents
o
o
o
o


signals
translation

issues
o
o
o
o
o

Quadrature sampling
Sigma-delta ADC


M. Renfors, TUT/DCE
TLT-5806/IQ/14
21.9.2010
Sampling Theorem
The sampling theorem says that a (real or complex)
lowpass signal limited to the frequency band [-W, W] can
represented completely by discrete-time samples if the
sampling rate (1/T) is at least 2W.
In case of a complex signal, each sample is, of course, a
complex number.
In general, discrete-time signals have periodic spectra,
where the continuous-time spectrum is repeated around
frequencies 1 T , 2 T , 3 T ,
2fs
fs
f
0
fs
2fs
In case of complex signals, it is not required that the

original signal is located symmetrically around 0.
Any part of the periodic signal can be considered as the
useful part. This allows many possibilities for multirate
processing of bandpass signals.
In general, the key criterion is that no destructive aliasing
effect occur on top of the desired part of the spectrum.

M. Renfors, TUT/DCE
TLT-5806/IQ/15
21.9.2010
Real vs. Complex Discrete-Time Signals

Real signal:
fs
fs
Here 2W real samples per second are sufficient to

represent the signal.
Complex signal:
fs
f
0 fs=W
Here W complex samples per second are sufficient.

The resulting rates of real-valued samples are the
same.
However, the quantization effects may be quite
different. (Recall from the standard treatment of
SSB that Hilbert-transformed signals may be
difficult.)

M. Renfors, TUT/DCE
TLT-5806/IQ/16
21.9.2010
Interpolation for Complex Signal

Sampling rate increase produces a periodic
spectrum, and the desired part of the spectrum is
then separated by an (analytic) bandpass filter.
N
a)
f
0
n/NT1/T
1/NT
COMPLEX
BP-FILTER
RESPONSE
f
0
n/NT
b)
f
0
1/NT
f
0
n/NT

M. Renfors, TUT/DCE
TLT-5806/IQ/17
21.9.2010
Decimation for Complex Signal

Sampling rate decrease produces aliasing, such that
the original band is at one of the image bands of the
resulting final band.
The signal has to be band-limited to a bandwidth of
1 / NT before this operation can be done without
severe aliasing effects.
N
COMPLEX
BP-FILTER
RESPONSE
n/NT1/T
f
0
1/NT
n/NT
f
0

M. Renfors, TUT/DCE
TLT-5806/IQ/18
21.9.2010
Combined Multirate Operations for

Complex Signal
Combining decimation and interpolation, a frequency shift
by n / NT can be realized, where n is an arbitrary integer.
N
f
n1 fS /N- fS
n1 fS /N
f
0
f
0
n2 fS /N
It can be seen that the low sampling rate, limited to be

higher than the signal bandwidth, determines the resolution
of the frequency translations based on multirate
operations.
If, for example, a bandpass signal is desired to be
translated to the baseband form, this can be done using
multirate operations if and only if the carrier frequency is a
multiple of the low sampling rate.
Using also simple frequency translations (with coefficients
+1, -1, +j, -j), the resolution is 1/(4NT).

M. Renfors, TUT/DCE
TLT-5806/IQ/19
21.9.2010
Combining Mixing and Multirate Operations

for Complex Signals
A general frequency shift of f O = n NT + f can be done
in the following two ways:
(1) Direct frequency conversion by f O using mixing.
(2) Conversion using multirate operations by n NT
followed by a mixing with f (or vice versa).
The differences in these two approaches are due to
the possible filtering operations associated with the
multirate operations, and aliasing/reconstruction
filters in case of mixed continuous-time/discrete-time
processing.
Assuming ideal filtering, these two ways would be
equivalent.

M. Renfors, TUT/DCE
TLT-5806/IQ/20
21.9.2010
Example of Combining Mixing and Multirate

Operations
Conversion from bandpass to baseband representation
and decimation to symbol rate, i.e., I/Q-demodulation.
Assume that
- N=6, f0=4/(6T)+f.
- The required complex bandpass filter is obtained from
an FIR filter of length 50 by frequency translation.
The following three ways are equivalent but lead to
different computational requirements (the required real
multiplication rates at input rate are shown, not exploiting
possible coefficient symmetry):
BPF
LPF
(ii)
(i)
N
N
e j
(iii)
e j
BPF
e j
Filter
Mixer
Total
Case (i)
100
4
104
Case (ii)
100/6
2
18.7
Case (iii)
100/6
4/6
17.3

M. Renfors, TUT/DCE
TLT-5806/IQ/21
21.9.2010
Example of Combining Mixing and Multirate

Operations (continued)
Notes:
(i) Complex bandpass filter, real inputs
=> 100 real multipliers needed for filter
(ii) Real lowpass filter, complex input to filter
=>100 real multipliers needed for filter
- Decimation can be combined efficiently with the
filter. Utilizing coefficient symmetry is easiest in
this case.
(iii) As (i) but decimation can be included efficiently
with the filter.
- Mixing and LO generation done at lower rate
and thus easier to implement.
Here we have not taken use of the possible
coefficient symmetry, which may reduce the
multiplication rates by 1/2 in all cases.
In general, mixing is a memoryless operation, so upsampling and down-sampling operations can be
commuted with it in block diagram manipulations.

M. Renfors, TUT/DCE
TLT-5806/IQ/22
21.9.2010
Frequency Translation for Real Signals

Mixing and multirate operations can be done in similar way
for real signals. The difference is that the two parts of the
spectrum, on the positive and negative frequency axis, and
their images, must be accommodated in the spectrum.
(1) Mixing
Mixing produces two translated spectral components (note
that cos(t ) = (e jt + e jt ) / 2 ). The image band appearing
on top of the desired band after mixing must be
suppressed before mixing.
fc
fcfLO
fc
fc+fLO 0 fcfLO
fc+fLO
cos(LOt)
(2) Multirate operations

In case of decimation, to avoid destructive aliasing effects,
the signal to be translated must be within one of the
intervals
n
,
NT
NT
+1
2 NT
or
n
1
,
2 NT
NT
NT
Otherwise destructive aliasing will occur. In the latter case,

the spectrum will be inverted.

M. Renfors, TUT/DCE
TLT-5806/IQ/23
21.9.2010
Interpolation for Real Bandpass Signal

w(m)
x(n)
y(m)
N
Nf s
fs
X( f )
fs
fs/2
fs/2
fs
W( f)
k=3
k=2
k=1
k=0
k=0
k=1
k=2
k=3
f
0
Y ( f )| k= 2
k=2
k=2
f
0
Y ( f )| k= 3
k=3
k=3
f
0

M. Renfors, TUT/DCE
TLT-5806/IQ/24
21.9.2010
Decimation for Real Bandpass Signal

xBP(n)
x(n)
fs
k=3
k=2
y(m)
k=1
f s /N
k=0
k=0
k=1
k=2
k=3
0
X BP( f )
3fs/(2N) fs/N
fs /N
3fs/(2N)
Y( f)
fs/N
fs /N
X BP( f )
2fs/N 3fs/(2N)
3fs/(2N) 2fs/N
Y( f)
fs/N
f
0
fs / N

M. Renfors, TUT/DCE
TLT-5806/IQ/25
21.9.2010
Analytic Filtering and Taking Real Part as

Multirate Operations
The transformations between real and complex signal
formats can be seen as multirate operations:
Taking the real part effectively reduces the rate of realvalued samples by two. It produces mirror images, in
contrast to the periodic images produced by
decimation by two. In both cases, the new spectral
components may fall on top of the existing spectral
components.
If this operation follows, e.g., an FIR filter, considerable
computational simplifications can be made by
combining the real part- operation with the filter in a
cleaver way. There is no sense to compute samples
that are thrown away by the real part operation!!.
Analytic filtering (in any form baseband, bandpass,

filter bank) increases the rate of real-valued samples
by two. Mirror images are removed form the spectrum,
in contrast to the periodic images that are removed in
interpolation.

M. Renfors, TUT/DCE
TLT-5806/IQ/26
21.9.2010
Example of Down-Conversion:
I/Q-Demodulation
N
4
PHASE
SPLITTER
MATCHED
FILTER
I
Q
f
0
f
0
2/T
f
0
1/T
It is usually a good idea to keep the signal as a real signal

as long as possible, because after converting to complex
form, all subsequent signal processing operations require
(at least) double computational capacity compared to the
corresponding real algorithms.

M. Renfors, TUT/DCE
TLT-5806/IQ/27
21.9.2010
Contents
o
o
o
o


signals
translation

issues
o
o
o
o
o

Quadrature sampling
Sigma-delta ADC


M. Renfors, TUT/DCE
TLT-5806/IQ/28
21.9.2010
Real Bandpass Sampling

Down-conversion can also be implemented by sampling a
bandpass signal. Any part of the periodic spectrum can be
selected for further processing.
fs
fcW/2
T/H
f
fc
kfsfs/2
f
kfs
kfs+fs/2
Concerning the sampling frequency, it is sufficient that no

aliasing appears on top of the desired band.
In general, the feasible sampling frequencies are
determined from W, B (useful signal bandwidth), and fs.
Minimum sampling frequency is B+W, which is adequate in
the case where the center frequency of the desired signal
is kfs fs/4:
W+B=fs
kfsfs/4 kfs kfs+fs/4


M. Renfors, TUT/DCE
TLT-5806/IQ/29
21.9.2010
Quadrature Sampling
In this case we are sampling the complex analytic signal
obtained by a phase-splitter:
I
T/H
fs
90
fcW/2
T/H
ANALYTIC
BANDPASS
SIGNAL
fc
kfs
(k+1)fs
Of course, in practise the image bands in between the

shown periodic replicas are not completely attenuated but
only suppressed to a level that is determined by the
amplitude and phase imbalances in the phase splitter &
sampler & ADC blocks.
The gain and phase imbalance analysis of quadrature
down-conversion applies also to this case.

M. Renfors, TUT/DCE
TLT-5806/IQ/30
21.9.2010
Second-Order Sampling
Quadrature sampling can be approximated by the following
structure:
I
T/H
fs
=1/4fc
T/H
At the carrier frequency, the sampling time offset

corresponds exactly to the 90o phase shift. Farther away
from the center frequency this is only approximative, but
for relatively narrowband signals, it works. The nonideality
can be evaluated using the phase imbalance analysis.

M. Renfors, TUT/DCE
TLT-5806/IQ/31
21.9.2010
Analysis of Second-Order Sampling

The second-order sampling concept works perfectly at the
carrier frequency (ignoring the other sources of I/Q
imbalance) but only approximately at other frequencies. At
frequency f c + f , a time-shift of 1 (4 f c ) corresponds to a
phase shift of
1
4 fc
f
2 = 1 +
rads
1
fc 2
( fc + f )
We are actually dealing with phase imbalance and the
image rejection formula for quadrature mixing can be
utilized. The resulting image rejection is:
f
1 cos
1 cos
fc 2
R=
=
1 + cos
f
1 + cos

fc 2
Example: fc=1 GHz

f
0.1 MHz
1 MHz
10 MHz
100 MHz
fc
0.0001
0.001
0.01
0.1
Phase
imbalance
0.009o
0.09o
0.9o
9o
Image rejection
82.1dB
62.1 dB
42.1 dB
22.1 dB

M. Renfors, TUT/DCE
TLT-5806/IQ/32
21.9.2010
Problems with Wideband Sampling

Analog to Digital Converter (ADC)
Sampling a wideband signal, containing several channels
is a tempting approach for designing a flexible radio
receiver. However, there are some great challenges to do
this.
The strongest signal in the ADC input signal band should
be in the linear range of the ADC. When the desired signal
is weak, a large ADC dynamic range is needed, the
resolution of the converter has to be many bits, e.g., 14 ...
17 bits.
MAGNITUDE
REQUIRED
DYNAMIC
RANGE
f
STRONG NEIGHBORING
CHANNELS
WEAK DESIRED
CHANNEL
Sampling
The sampling to get a discrete time signal is done usually
with a track-and-hold circuit (T/H).
In practical sampling clocks and sampling circuits, there
are unavoidable random variations in the sampling
instants, sampling aperture jitter. In bandpass sampling,
the requirements for aperture jitter become very hard.

M. Renfors, TUT/DCE
TLT-5806/IQ/33
21.9.2010
Quantization Noise in ADCs - 1

In general, the maximum S/N-ratio for an A/D-converter is
SNR = 6.02n + 4.76 CFdB + 10log10 [ f s / kB ]
(dB)
where n
is the number of bits
CFdB is the Crest Factor in dB
B
is the useful signal bandwidth
f s is the sampling rate
k
2 for baseband, 1 for bandpass sampling.
The crest factor for a voltage signal is defined as the ratio of the
peak absolute value and the RMS-value. The maximum SNR is
achieved when the signal utilizes the A/D-converters full voltage
range, which is assumed to be symmetric around DC. For a
sinusoidal signal, the crest factor is 3 dB, and for a bandpass
signal, the crest factor is 3 dB higher than that of the equivalent
baseband signal.
The last term takes into account the processing gain due to
oversampling in relation to the useful signal band. When the
quantization noise outside that useful signal band is filtered away,
the overall quantization noise power is reduced by the factor fs/kB.
Considering the choice between (real) baseband and bandpass
sampling, the difference is due to the 3 dB higher crest factor in
the bandpass sampling case. The k parameter is just due to the
definition of bandwidth. The bandwidths of B in baseband model
2B in bandpass model contain the same amount of signal power
and quantization noise power.

M. Renfors, TUT/DCE
TLT-5806/IQ/34
21.9.2010
Quantization Noise in ADCs - 2

Lets consider now the case of complex I/Q sampling. In most
cases, half of the overall signal power is captured by both of the
ADCs in the I and Q branches. Thus the signal to quantization
noise power ratio is the same in both ADCs as that of the overall
complex signal. Also the processing gain is the same as in the
real case, if the bandwidths are defined in the same way.
However, the full peak value of the waveform may appear in both
of them. Thus the crest factor should be defined as the ratio of
the peak complex signal magnitude divided by the rms value of
the I or Q signal.
The number of additional bits needed to quantize a wideband
signal, containing strong spectral components in addition to the
wanted signal, can be estimated by:
CF 2 P
B / 6 bits
10log10 B
CFd 2 Pd
where PB is the worst case power in the full band
Pd is the minimum useful signal power
CFB is the crest factor in the worst case test situation
CFd is the crest factor of the desired signal.
Usually, in communications receivers, the worst case power is
determined from the adjacent channel or blocking specifications.
For many types of communications signals (e.g., CDMA, OFDM),
the crest factor is much higher than that of the sinusoid, which
might be a valid assumption in the blocking test situation.
B

M. Renfors, TUT/DCE
TLT-5806/IQ/35
21.9.2010
Spurious-Free Dynamic Range

Practical ADC's have also discrete spectral frequency
components, spurious signals (or spurs), in addition to the
flat quantization noise.
In many applications, the spurious-free dynamic range,
SFDR, is the primary measure of the dynamic range of the
converter.

M. Renfors, TUT/DCE
TLT-5806/IQ/36
21.9.2010
Track&Hold Circuit Nonidealities

Advanced bandpass sampling approaches could
mean that we are sampling a tens-of-MHz to GHzrange signal with a relatively low sampling rate.
Noise Aliasing
Wideband noise at the sampling circuitry will be

aliased to the signal band. In case of bandpass
sampling, aliasing increases with increasing
subsampling (fc/fs) factor. Basically, the noise figure
depends on the subsampling factor.
Therefore, it is important to have a good noise figure
for the track&hold circuit and/or to have sufficient
amplification in the analog front-end.

M. Renfors, TUT/DCE
TLT-5806/IQ/37
21.9.2010
Aperture Jitter
Aperture jitter is the variation in time of the exact
sampling instant, that causes phase modulation and
results in an additional noise component in the
sampled signal.
Aperture jitter is caused both by the sampling clock
and the sampling circuit.

M. Renfors, TUT/DCE
TLT-5806/IQ/38
21.9.2010
SNR Due to Sampling Jitter

The noise produced by aperture jitter is usually modelled
as white noise, which results in a signal-to-noise ratio of
SNRaj = 20 log10
2 f maxTa
where fmax is the maximum frequency in the sampler input
and Ta is the RMS value of the aperture jitter.
This model is derived for a sinusoidal input signal, but
applied also more generally. In critical test cases of the
wideband sampling receiver application, the blocking signal
is often defined as a sinusoidal signal, and the model is
expected to work reasonably well.
The processing gain due to oversampling effects in the
same way as in case of quantization noise.
Example of Sampling Jitter Effects
110
- 14 bits
- 1 ps RMS jitter
Aperture jitter effect

Quantization effect
Joint effect
100
SNR in dB
90
80
70
60
50
40
0
10
10
10
Frequency in MHz
10

M. Renfors, TUT/DCE
TLT-5806/IQ/39
21.9.2010
About A/D-Conversion for SW Radio

It is obvious that the requirements for the T/H-circuit and
A/D-converter are the main bottlenecks for implementing
receiver selectivity with DSP.
One promising A/D-converter technology in this context is
the sigma-delta () principle.
- This principle involves low-resolution, high-speed
conversion in a noise-shaping configuration, together
with decimating noise filtering.
- In case of lowpass and bandpass sampling with suitable
fixed center frequency, this principle can be combined
nicely with the selectivity filtering part of the receiver.
Noise filtering in basic ADC:
f
fc
Noise filtering in sigma-delta converter:
f
fc

M. Renfors, TUT/DCE
TLT-5806/IQ/40
21.9.2010
Sigma-Delta Modulator - 1
e[n]
u[n]
x[n]
y[n]
DAC
This is a second-order sigma-delta modulator. The

sequence e[n] models the quantization noise of a single-bit
(or few-bit) ADC included in the loop. In an Lth-order
modulator, the first integrator is repeated L-1 times.
The system has different transfer functions for input signal
and quantization noise. They can be analyzed as follows:
z 1 1
U ( z) =
1
1 z 1 z 1
Y ( z) = U ( z) + E ( z)
L 1
z 1
( X ( z) Y ( z))
Y ( z)
1
1 z
Now Y(z) can be solved as
Y ( z ) = z X ( z ) + (1 z
1
) E (z)
L
The signal and noise transfer functions can be identified

from this expression as:
STF = z 1
NTF = (1 z 1 ) L

M. Renfors, TUT/DCE
TLT-5806/IQ/41
21.9.2010
Sigma-Delta Modulator - 2
NTF attenuates noise from the desired signal band. The inband quantization noise variance becomes
e2 =
SQ ( f ) NTF ( f ) df
2
2
=
12 f s
2
B sin ( f / f s ) df 12 f s
B
2L
2 2L 2B
=
12 ( 2 L + 1) f s
( f / f s )
2L
df
2 L +1
Oversampling ratio has great effect:

Doubling the oversampling ratio Fs/2B, the noise is
decreased by the factor 3(2L+1) in dB.
The number of quantization bits can be reduced by
increasing the oversampling ratio.
Noise can be filtered out by decimating digital filters.

M. Renfors, TUT/DCE
TLT-5806/IQ/42
21.9.2010
Contents
o
o
o
o


signals
translation

issues
o
o
o
o
o

Quadrature sampling
Sigma-delta ADC


M. Renfors, TUT/DCE
TLT-5806/IQ/43
21.9.2010
Polyphase Decomposition
Any FIR filter transfer function can be expressed as
L 1
H ( z ) = z i H i ( z L )
i =0
In this case, the polyphase branch filters are FIR filters of

length M / L or M / L , where M is the length of the
original FIR filter.
Example: Polyphase structure with L=2:
=
H0(z2)
z 1
H1(z2)
Also certain types of IIR filters can be expressed in the

above form. Especially, the Nth-band IIR filters have
polyphase decomposition where the polyphase branches
are allpass filters.

M. Renfors, TUT/DCE
TLT-5806/IQ/44
21.9.2010
Applications of Polyphase Structure 1

Complex Bandpass Filters
Including a frequency shift by kfs/L, the polyphase filter can
be represented as
L 1
H ( z ) = ( z e jk 2 / L ) i H i ( z L )
i =0
So we can see that the complex multipliers for frequency

shifting do not appear in the polyphase branches, but only
in the separate unit delay chain.
If this part of the structure is on the output side, then just
replicating this structure with different values of k, we
obtain bandpass filters with the same input but different
center frequencies, and sharing the main part of the
computations.
In this way we can implement filter banks that split the
input signal into a number of narrower signal bands.
Polyphase structures are mostly considered in multirate
signal processing applications, as discussed on the
following pages. Here we want to point out that they can
be useful even in cases where sampling rate conversion is
not involved, e.g., for implementing complex bandpass
filters and filter banks.

M. Renfors, TUT/DCE
TLT-5806/IQ/45
21.9.2010
Applications of Polyphase Structure 2

Decimation and Interpolation Filters
The polyphase structure with sampling rate reduction by
two looks as follows:
2
H0(z2)
z 1
H1(z2)
Using the so-called noble identities, the structure can be

developed as follows:
z 1
H0(z)
H1(z)
The unit delay and sampling rate reduction blocks just

model the operation of connecting even-indexed input
samples to the upper branch and odd-indexed ones two
the lower branch.

M. Renfors, TUT/DCE
TLT-5806/IQ/46
21.9.2010
General Polyphase Structures

Decimation by L:
H0(z)
H1(z)
...
fS
H(z)
fS
fS
fS /L
HL-1(z)
fS /L
(b)
(a)
Interpolation by L:
H0(z)
H1(z)
fS
fS
(a)
LfS
H(z)
...
HL-1(z)
LfS
(b)
LfS

M. Renfors, TUT/DCE
TLT-5806/IQ/47
21.9.2010
Bandpass Polyphase Structures and

Filter Banks
As mentioned already, we can easily embed in the
polyphase structure the complex coefficients needed for
frequency translation. In the decimation case, the
bandpass polyphase structure looks as follows:
H0(z)
fS
HL-1(z)
with wi,k = e jk (2 )i / L ,
...
wL-1, k
fS /L
H1(z)
w1, k
i = 1, 2, ..., L 1.
A uniform filter bank can be obtained by repeating the

complex multipliers and output adder with different values
of k. If all (or even several) channels are need, this can be
done very efficiently using FFT operation. Such a structure
is generally called DFT filter bank.
The decimation case corresponds to analysis bank, in
which the wideband signal is divided into L subchannels
with sample rate fs/L.
Likewise, the interpolation case corresponds to a synthesis
bank, in which several narrowband signals are combined
into a single wideband signal.

M. Renfors, TUT/DCE
TLT-5806/IQ/48
21.9.2010
Real Critically-Sampled Filter Banks

In critically sampled L-channel filter banks, the sampling
rate conversion factor is L. For every L input samples of a
(decimating) analysis bank, one sample is generated to
each of the low-rate subbands.
An analysis-synthesis system includes analysis and
synthesis banks, connected by a subband processing
stage.
L
L-channel
Analysis
Filter
Bank
L-channel
Synthesis
Filter
Bank
Processing
x(n)
The perfect reconstruction property means that, in the

absence of subband processing, the synthesis bank output
is an exact delayed replica of the analysis bank input. With
proper filter bank design, perfect reconstruction can be
achieved even in the critically sampled case, even though
aliasing takes place in the subband signals. These must
have considerably overlapping transition bands in practical
designs.
Cosine-modulated filter banks (CMFBs) are a common and
efficient solution for real critically-sampled uniform filter
banks.

M. Renfors, TUT/DCE
TLT-5806/IQ/49
21.9.2010
Filter Bank vs. FFT/IFFT

Probably the computationally most efficient filter banks are
FFT (as analysis bank) and IFFT (as synthesis bank). The
benefit of filter banks is better frequency selectivity of the
subbands, at the cost of somewhat higher complexity.
Example: Subband frequency responses for L=16
channels
(a) FFT:
Amplitude in dB
20
40
60
0.1
0.2
0.3
0.4
0.5
0.6
Frequency /
0.7
0.8
0.9
(b) CMFB, prototype filter length=8*L, 100% roll-off:

Amplitude in dB
20
40
60
0.1
0.2
0.3
0.4
0.5
0.6
Frequency /
0.7
0.8
0.9
The uniform, modulation-based filter banks considered

here consist of the prototype filter in polyphase or lattice
form, and a transform block, e.g., discrete cosine/sine
transform, or DFT/FFT.

M. Renfors, TUT/DCE
TLT-5806/IQ/50
21.9.2010
Complex Critically-Sampled Filter Banks - 1

The structure of the real critically sampled filter bank
cannot be generalized to the complex case as such. Using
L-channel banks and sampling rate conversion by L,
together with complex subband signals leads to perfect
reconstruction only in case of the DFT-IDFT pair.
The solution is to use 2M-channel banks, sampling rate
conversion by M, and real subband signals.
Exponentially modulated filter bank (EMFB):
x(n)
H0 (z)
Re[]
F0 (z)
H1 (z)
Re[]
F1 (z)
H2M 1 (z)
Re[]
F2M 1 (z)
x
(n)

M. Renfors, TUT/DCE
TLT-5806/IQ/51
21.9.2010
Complex Critically-Sampled Filter Banks 2

In the EMFB case, the sub-channel impulse responses of
the synthesis bank are obtained from a low-pass prototype
filter hp(n) as follows:
f k ( n ) = f kc ( n ) j f ks ( n ) =
2
L + 1
1
+
h p ( n ) exp j n +
k

L
2
2
Here the +/- signs are for the subbands on the

positive/negative sides of the spectrum, respectively. The
analysis bank impulse responses are related to the
synthesis bank as follows:
hk ( n ) = f kc ( N 1 n ) j f ks ( N 1 n ) = hkc ( n ) j hks ( n )
where N is the length of the prototype filter.

Critically-sampled perfect-reconstruction EMFB systems
can be obtained by using the same prototype filters as for
CMFBs.

M. Renfors, TUT/DCE
TLT-5806/IQ/52
21.9.2010
Complex Critically-Sampled Filter Banks - 3

EMFBs can be implemented efficiently using CMFBs and
SMFBs as building blocks (also an efficient FFT-based
structure is available):
x0 (m)
Synthesis
Analysis
CMFB
CMFB
x
I (n)
xM 1 (m)
xI (n)
x2M 1 (m)
x
Q (n)
xQ (n)
Synthesis
Analysis
SMFB
SMFB
xM (m)

M. Renfors, TUT/DCE
TLT-5806/IQ/53
21.9.2010
Complex Critically-Sampled Filter Banks 4

EMFBs use so-called odd channel stacking arrangement:
FM 1 ( )
F1 ( ) F0 ( ) F0+ ( ) F1+ ( )
FM+ 1 ( )
After decimation, the subband signals are centered at /2

or /2, and the other side of the spectrum is empty. This is
also important for perfect reconstruction, since now there is
room for the mirror image produced by the real-part
operation. Then, only the transition bands of the two
spectral components are overlapping. What basically
happens is that a real subband signal is constructed from
its analytic version.
Also even channel stacking arrangement can be used in
complex filter banks. In this case, the subbands are
centered at k/M, and after decimation they are centered at
0 or . In this case, the real part operation cannot be done
as such for the subband signals, because the mirror image
would fall on top of the decimated subband signal.
One well known class of complex filter banks using even
channel stacking are the modified DFT (MDFT) filter banks.
MDFT banks use slightly more complicated processing of
the subband signals to achieve perfect reconstruction. This
leads to more complicated relation between the original
spectrum and the subband signal spectra.

M. Renfors, TUT/DCE
TLT-5806/IQ/54
21.9.2010
Transmultiplexers
A transmultiplexer system includes also analysis and
synthesis banks, but in the reverse order. They are mostly
applied in communication systems for TDM FDM
transmultiplexing and for multicarrier modulation, as an
alternative to the widely adopted IFFT/FFT based OFDM
system.
In a transmultiplexer, a number of low-rate symbol
sequences are modulated to subcarriers and combined
into a wideband signal for transmission. In the analysis
bank of the receiver, the subcarrier signals are again
separated. In an ideal transmultiplexer, in case of an ideal
noise-free channel, the output subcarrier sequences are
exactly the same as the input sequences.
Perfect-reconstruction CMFBs, EMFBs, and MDFT banks
can directly be used to get ideal transmultiplexers.
The main benefits of filter bank based multicarrier
modulation over OFDM are due to the high frequency
selectivity: the system is very robust to narrowband
interferences and the overall spectrum has very sharp
transition bands, allowing to use very narrow guard-bands
in frequency domain. Also, the receiver filter bank can
implement a considerable part of the channel selection
filtering in a flexible manner.

M. Renfors, TUT/DCE
TLT-5806/IQ/55
21.9.2010
EMFBs as Transmultiplexer
Real transmultiplexers have real baseband signal, whereas
in the complex case, the wideband baseband signal is
complex, and it can be directly modulated to the RF carrier
in a spectrally efficient way.
CMFBs and SMFBs can again be used as building blocks
for critically-sampled complex transmultiplexers:
X
+
0
(z )
X 1+ ( z )
Synthesis:
Cosine
Modulated
Filter Bank
X 0+ ( z )
X 1+ ( z )
Analysis:
Cosine
Modulated
Filter Bank
I
+
M 1
(z )
X 1 ( z )
Channel
X 0 ( z )
M
M
+
M
Synthesis:
Sine
Modulated
Filter Bank
XM
1 ( z )
+
XM
1 ( z )
X 0 ( z )
X 1 ( z )
Analysis:
Sine
Modulated
Filter Bank
XM
1 ( z )
Here we have again 2M subchannels using sampling rate

conversion factor M. In the critically-sampled case, the
subchannel signals are real. In the modulation sense, VSB
modulation or offset-QAM modulation take place for the
subcarrier signals.

M. Renfors, TUT/DCE
TLT-5806/IQ/56
21.9.2010
Complex Oversampled Transmultiplexer

In a analysis-synthesis or transmultiplexer system, aliasing
and imaging are taking place at different processing
stages, but the effects cancel each other and perfect
reconstruction is achieved in ideal conditions.
However, in case of transmultiplexers, non-ideal,
frequency-selective transmission channel destroys the
orthogonality of the subcarrier signals, introducing
intercarrier interference (ICI, crosstalk) between adjacent
subchannels and intersymbol interference (ISI) within each
subchannel. These effects cannot be compensated
effectively by subband-wise processing. Thus, in the
critically sampled case, the channel equalizers after
analysis bank have to include connections between
adjacent subchannels.
Usually in the filter bank design, the roll-off is no more than
100 %, i.e., the overall subchannel bandwidth is twice the
channel spacing, or less. If the subchannel signals can be
obtained in 2x -oversampled form, then they are essentially
alias free (the alias components are attenuated at least by
the stopband attenuation of the channel filters). Then, in
principle, it becomes possible to do the channel
equalization for each subcarrier independently of the
others.
In the EMFB, there is a simple way to obtain 2x
oversampling for the subchannel signals: using the
complex subchannel signals instead of the real ones that
are sufficient in the critically sampled case.

M. Renfors, TUT/DCE
TLT-5806/IQ/57
21.9.2010
EMFB-Based Transmultiplexer with

2x-Oversampled Analysis Bank and PerSubcarrier Equalization
CMFB
Analysis
I
Re{ . }
X k+ (m )
1
2
CMFB
SMFB
Synthesis
Analysis
ck+,0 , ck+,1 ,..., ck++ L
Re{ . }
ck,0 , ck,1 ,..., ck+ L
Re{ . }
X k+ (m )
Hlp(z)
X k (m )
SMFB
1
2
Synthesis
CMFB
Equivalent lowpass channel
Analysis
Q
Im{ . }
SMFB
X k (m )
Analysis
Here the synthesis bank is critically sampled, analysis

bank is oversampled by 2, and after subchannel
equalizers, symbol-rate sequences are obtained by taking
the real part of the equalizer output (usually, the real part
operation can be efficiently combined with the equalizer).

M. Renfors, TUT/DCE
TLT-5806/IQ/58
21.9.2010
Complex Oversampled Analysis-Synthesis

System
One good application of analysis-synthesis FB systems is
in frequency-domain channel equalization (FDE) in
communication systems.
FFT-IFFT -based FDEs have received a lot of attention
recently because, in wideband transmission, they result in
significantly
lower
complexity
than
time-domain
equalization. FFT-based FDEs use commonly a cyclic
prefix, in the same way as OFDM, to achieve simple and
robust system.
Using analysis-synthesis FB system in FDE, instead of
FFT-IFFT, has a number of benefits due to the good
frequency selectivity, as mentioned already in the
multicarrier context. Also here, in the critically sampled
case, subband-wise equalization results in poor
performance, and cross-connections between subbands
would be needed.
A 2x -oversampled EMFB-type analysis bank can be
realized in the same manner as in the transmultiplexer
case. This makes it feasible to do the channel equalization
subband-wise. Then the real parts of the subband
equalizer outputs are connected to a critically sampled
synthesis bank.
Naturally, analysis-synthesis systems with oversampled
subband processing could be a useful tool for various other
applications.

Sampling and Multirate Techniques For Complex and Bandpass Signals

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Sampling and Multirate Techniques For Complex and Bandpass Signals

Uploaded by

Copyright:

Available Formats

Sampling and Multirate Techniques for Complex and Bandpass Signals