Nonuniform Bandpass Sampling in Radio Receivers

Nonuniform Bandpass Sampling in Radio Receivers
Yi-Ran Sun
Stockholm 2004
Licenciate Dissertation
Royal Institute of Technology
Department of Microelectronics and Information Technology
Sun, Yi-Ran
Nonuniform Bandpass Sampling in Radio Receivers
A dissertation submitted to Kungliga Tekniska Högskolan,

Stockholm, Sweden, in partial fulfillment of the requirements
for the degree of Teknologie Licenciate.
TRITA-IMIT-LECS AVH 04:13

ISSN 1651-4076
ISRN KTH/IMIT/LECS/AVH-04/13–SE
Copyright °
c Yi-Ran Sun, November 2004
Royal Institute of Technology

Department of Microelectronics and Information Technology
Laboratory of Electronics and Computer Systems
Electrum 229
S-164 40 Kista, Sweden
Abstract
As an interface between radio receiver front-ends and digital signal processing

blocks, sampling devices play a dominant role in digital radio communications.
Based on different sampling theorems (e.g., classic Shannon’s sampling theorem,
Papoulis’ Generalized sampling theorem, bandpass sampling theory), signals are
processed by the sampling devices and then undergo additional processing. It is a
natural goal to obtain the signals at the output of the sampling devices without
loss of information.
In conventional radio receivers, all the down-conversion and channel selection are
realized in analog hardware. The associated sampling devices in A/D converters
are based on the classic Shannon’s sampling theorem. Driven by the increased
speed of microprocessors, there is a tendency to use mixed-signal/digital hardware
and software to realize more functions (e.g., down-conversion, channel selection,
demodulation and detection) in a radio communication system. The new evolution
of radio receiver architecture is Software Defined Radio (SDR). One design goal of
SDR is to put the A/D converter as close as possible to the antenna. BandPass
Sampling (BPS) enables one to have an interface between the higher IF and the
A/D converter by a sampling rate of 2B or more (B is the information bandwidth),
and it might be a solution to SDR.
A signal can be uniquely determined from the samples by NonUniform Sam-
pling (NUS) such that NUS has the potential to suppress harmful signal spectrum
aliasing. BPS makes use of the signal spectrum aliasing to represent the signal
uniquely at any band position. A harmful aliasing of signal spectrum will cause
a performance degradation. It is of great benefit to use NUS scheme in BPS sys-
tem. However, a signal cannot be recovered from its nonuniform samples by using
only an ideal lowpass filter (or the classic Shannon’s reconstruction function). The
reconstruction of the samples by NUS is crucial for the implementation of NUS.
Besides the harmful signal spectrum aliasing, noise aliasing and timing jitter are
other two sources of performance degradation in a BPS system. Noise aliasing is
the direct consequence of lower sampling rate of subsampling. With the increase of
input frequency by directly sampling a signal at higher IF, the timing error of the
sampling clock causes large jitter effects on the sampled-data signal.
In this thesis work, first, a filter generalized by a certain Reconstruction Algo-
rithm (RA) is proposed to reconstruct the signal from its nonuniform samples. A
iii
general reconstruction formula in terms of a basis-kernel (BK) is used to describe
the algorithm. The corresponding reconstruction performance, computational com-
plexity and implementation of these RAs are discussed. Second, three sources of
performance degradation in a BPS system, harmful signal spectrum aliasing, noise
aliasing and timing jitter, are studied. In the light of noise aliasing, a Generalized
Quadrature BPS (GQBPS) algorithm is proposed to suppress the noise aliasing.
Theoretical analyses show that GQBPS might be a potential way to reduce the
noise aliasing at the cost of a more complicated reconstruction algorithm, although
it is sensitive to large timing jitter. Then, aliasing-free sampling by NUS is studied
in theory and verified by simulations. Thermal noise and timing errors are always
present in real circuit implementations. Finally, the performance of additive noise
and jitter on RAs in BPS is evaluated and discussed.
iv
Publications list:
1. Yi-Ran Sun and Svante Signell, “A Generalized Quadrature Bandpass Sam-
pling in Radio Receivers”, ASP-DAC 2005 (accepted).
2. Yi-Ran Sun and Svante Signell, “Effects of Noise and Jitter in Bandpass
Sampling”, Journal of Analog Integrated Circuits and Signal Processing –
Special Issue of Norchip’03, 42(1): 85-97, Jan. 2005.
3. Yi-Ran Sun and Svante Signell, “A Novel Quadrature Bandpass Sampling
in SDR Front-Ends”, in Proceeding of Biennial Analog Signal Processing
Conference (ASP 2004), pp. 9.1-9.6, Oxford Brookes University, Oxford, UK,
Nov. 2004.
4. Yi-Ran Sun and Svante Signell, “A Generalized Quadrature Bandpass Sam-
pling with Noise Aliasing Suppression”, Workshop on Wireless Circuits and
Systems (WoWCAS), pp. 41-42, Vancouver B.C., Canada, May 2004
5. Yi-Ran Sun and Svante Signell, “Effects of Noise and Jitter on Algorithms for
Bandpass Sampling in Radio Receiver”, in Proceedings of IEEE International
Symposium on Circuits and Systems (ISCAS), vol. I, pp. 761-764, Vancouver
B.C., Canada, May 2004.
6. Yi-Ran Sun and Svante Signell, “Effects of Noise and Jitter in Bandpass Sam-
pling”, in Proceedings of 21st Norchip Conference, Riga, Latvia, November
2003.
7. Yi-Ran Sun and Svante Signell, “Jitter Performance of Reconstruction Al-
gorithms for Nonuniform Bandpass Sampling”, in Proceedings of European
Conference of Circuit Theory and Design (ECCTD), pp. 353-356, Krakow,
Poland, September 2003.
8. Yi-Ran Sun and Svante Signell, “Algorithms for Nonuniform Bandpass Sam-
pling in Radio Receiver”, in Proceedings of IEEE International Symposium
on Circuits and Systems (ISCAS), vol. I, pp.1-4, Bangkok, Thailand, May
2003.
Non-Reviewed presentations and Technical Reports:

9. Yi-Ran Sun, ”Bandpass Sampling in the Presence of Noise and Jitter”, Swedish
System-on-Chip Conference (SSoCC), Båstad, Sweden, April, 2004.
10. Yi-Ran Sun, ”Algorithms and Performance Analysis for Nonuniform Band-
pass Sampling in Radio Receiver”, Swedish System-on-Chip Conference (SSoCC),
Sundbyholms Slott, Eskilstuna, Sweden, April, 2003.
11. Yi-Ran Sun, ”Bandpass Sampling for Radio Receiver”, SocTRic demonstrator
project, Acreo AB, Norrköping, Sweden, April, 2002.
v
Acknowledgments
My sincere thanks to my supervisor Prof. Svante Signell, for giving me the oppor-
tunity to do this interesting and also challenging research and for his mentoring,
guidance, encouragement and close collaboration. His strictness to science and
research have had a tremendous impact both on this work and on my personal
professional development.
I am extremely grateful for the harmonious environment of study and research
in LECS. Thanks to all the colleagues in LECS for always friendly help, espe-
cially Jian Liu, Xinzhong Duo and Steffen Albrecht. Thanks also to Darius Jakonis
(Linköping University) and other colleagues in the SocTRix project for many valu-
able discussions on the research. I also would like to thank all the administrators
of LECS and the system group for their excellent work. Additionally, I am greatly
thankful to Jinliang Huang and Jad Atallah for proofreading the thesis.
None of this work would have been possible without my parents’ persistent
encouragement, advise and confidence. Thank you for always supporting me to
pursue my interests.
vii
Contents
Contents viii
List of Figures xv
1 Introduction 1
1.1 Superheterodyne receivers . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Homodyne receivers . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Wideband IF Receivers . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.4 Generic Wideband Receivers . . . . . . . . . . . . . . . . . . . . . . 6
1.5 Software Defined Radio Receivers . . . . . . . . . . . . . . . . . . . . 7
1.6 Outline of Technical Problems . . . . . . . . . . . . . . . . . . . . . . 9
1.7 Overview of Previous Work on BPS . . . . . . . . . . . . . . . . . . 11
1.8 Summary of Contributions . . . . . . . . . . . . . . . . . . . . . . . . 11
2 Sampling and Reconstruction 13

2.1 Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.2 Reconstruction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.3 Basis-Kernel (BK) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.4 Reconstruction Algorithms (RAs) . . . . . . . . . . . . . . . . . . . . 23
2.5 Performance Evaluation of RAs . . . . . . . . . . . . . . . . . . . . . 30
2.6 Implementations of RAs . . . . . . . . . . . . . . . . . . . . . . . . . 36
3 Uniform Bandpass Sampling 41

3.1 Sampling Rate Selection . . . . . . . . . . . . . . . . . . . . . . . . . 43
3.2 Noise Spectrum Aliasing . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.3 Jitter Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4 Quadrature Bandpass Sampling 55

4.1 Generalized Nonuniform Sampling . . . . . . . . . . . . . . . . . . . 55
4.2 Quadrature Bandpass Sampling . . . . . . . . . . . . . . . . . . . . . 60
4.3 Implementation of Quadrature BPS . . . . . . . . . . . . . . . . . . 62
viii
5 Nonuniform Random Sampling 75
5.1 Jitter Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.2 Additive Random Sampling . . . . . . . . . . . . . . . . . . . . . . . 79
5.3 Alias-free Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
6 Noise and Jitter Performance on RAs 85

6.1 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
6.2 Sensitivity of RAs to Jitter . . . . . . . . . . . . . . . . . . . . . . . 87
6.3 Sensitivity of RAs to AWGN . . . . . . . . . . . . . . . . . . . . . . 87
6.4 Jitter Noise Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
7 Conclusions and Future Work 93
Bibliography 95
ix
List of Abbreviations
2/2.5/3 G the second/second and half/third Generation

AA Anti-Aliasing
A/D Analog-to-Digital
ARS Additive Random Sampling
AWGN Additive White Gaussian Noise
BER Bit Error Rate
BK Basis-Kernel
BPF BandPass Filter
BPS BandPass Sampling
BS cellular system Base Station
CDMA Code Division Multiple Access
CMOS Complementary Metal Oxide Semiconductor
CSMA-CA Carrier Sense Multiple Access/Collision Avoidance
CT Continuous-Time
CT-scan Computerized Tomography-scan
D/A Digital-to-Analog
DC Direct Current
DCF Density Compensation Factor
DCS Digital Cellular System
DECT Digital European Cordless Telephone
DL Down-Link
DPD Digital Product Detector
DT Discrete-Time
DTFT Discrete-Time Fourier Transform
DFT Discrete Fourier Transform
DSB Double-SideBand
DSSS Direct Sequence Spread Spectrum
EDS Energy Density Spectrum
EDGE Enhanced Data rates for Global/GSM Evolution
FHSS Frequency-Hopping Spread Spectrum
FIR Finite Impulse Response
FPO Floating Point Operation
GSM Global System for Mobile communications
xi
GQBPS Generalized Quadrature BandPass Sampling
iid independent, identically distributed
I/Q In-phase/Quadrature
IEEE Institute of Electrical and Electronics Engineers
IF Intermediate Frequency
IIR Infinite Impulse Response
IRF Image-Rejection Filter
ISI Inter-Sample Interval
ISM Industrial, Scientific and Medical
JS Jitter Sampling
LO Local Oscillator
LPF LowPass Filter
LPS LowPass Sampling
LSR Least Square Reconstruction
LTI Linear Time-Invariant
MRI Magnetic Resonance Imaging
NB Narrow Band
NUS NonUniform Sampling
OFDM Orthogonal Frequency Division Multiplexing
PDC Personal Digital Cellular
PSD Power Spectral Density
RF Radio Frequency
RA Reconstruction Algorithm
SC Switched-Capacitor
SDR Software Defined Radio
S/H Sample-and-Hold
SNDR Signal-to-Noise-and-Distortion Ratio
SNR Signal-to-Noise Ratio
SSB Single-SideBand
SVD Singular-Value Decomposition
TDMA Time Division Multiple Access
UE User Equipment for cellular terminal
UL Up-Link
UMTC Universal Mobile Telecommunication System
US Uniform Sampling
WB Wide Band
W-CDMA Wideband Code-Division Multiple-Access
WSS Wide-Sense Stationary
xii
List of Notations
B The bandwidth of lowpass information signal

fc The carrier frequency
fin The frequency of input signal
δ(t) The Dirac delta function
δ[m − n] The Kronecker delta function
fs , Fs The sampling rate
Ts The sampling interval and Ts = 1/fs
k(t, tn ) The Basis Kernel (BK) of Reconstruction Algorithm (RA)
Re{•} The real part of complex signal
< a, b > The inner product operator
b•c The floor operator
x̂(t) The reconstructed result of x(t)
∗ The complex conjugate operator
? The convolution operator
∈ An element of
στ The standard deviation of sampling jitter
E[•] The expectation operator
Mean[•] The mean value
limx→a f (x) The limit of function f (x)
rect(•) The rectangular function
p(τ ) Probability Density Function (PDF)
N n , τm )
p(τ Joint PDF
The ideal sampling operator, i.e., the process of multiplying
Bef f The effective bandwidth of noise
x(tn ) The sampled-data of NUS
x(nTs ) The sampled-data of US
x(ts (n)) The sampled-data of random sampling
rxx (γ) The autocorrelation function of x(t)
Rxx (f ) The Fourier transform of rxx (γ)
f0 @fc The notation of a signal with an information bandwidth of f0 centered at fc
fl The lower frequency of bandpass signal
fu The upper frequency of bandpass signal
U [a, b] Uniform distribution of a random variable
xiii
Q
P The symbol of product
The symbol of sum
∪ The symbol of union
|| • || The symbol of norm
6 = The symbol of inequality
≈ The symbol of approximately equal
∞ The symbol of infinity
F {•} The Fourier transform operator
µ(t) The Heaviside’s step function
xiv
List of Figures
1.1 Conventional dual-IF superheterodyne receiver architecture . . . . . . . 3

1.2 Homodyne receiver architecture . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 Wideband IF receiver architecture with double down-conversion . . . . 6
1.4 Generic wideband receiver architecture . . . . . . . . . . . . . . . . . . . 7
1.5 Software defined radio receiver architecture by BPS . . . . . . . . . . . 8
1.6 Frequency spectrum of bandpass signal . . . . . . . . . . . . . . . . . . 10
2.1 Voltage sampling process . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.2 Charge sampling process . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.3 a) Original CT band-limited signal, B = 50 and samples by US, fs =
200; b) The corresponding frequency spectrum of CT signal; c) The
corresponding frequency spectrum of sampled-data signal. . . . . . . . . 16
2.4 US and NUS sequences (left) and correspondingRnormalized ESD spec-
trum (right) based on eq. (2.11) (normalized to |Xs (f )|2 df = 1) [26]. . 18
2.5 Two-step sampling paradigm . . . . . . . . . . . . . . . . . . . . . . . . 19
2.6 Sampling patterns of nonuniform sampling [41]. (Top-left): Polar sam-
pling grid; (Top-right): Spiral sampling grid; (Bottom-left): variable-
density nonuniform sampling grid; (Bottom-right): general nonuniform
sampling grid. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.7 Cartesian uniform sampling grid. . . . . . . . . . . . . . . . . . . . . . . 22
2.8 Identity elements of (a) interpolation reconstruction with a CT filter; (b)
interpolation reconstruction with a DT filter; (c) iterative reconstruction
[44] . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2.9 Identity elements of (a) svd reconstruction with a CT filter; (b) svd
reconstruction with a DT filter. . . . . . . . . . . . . . . . . . . . . . . . 30
2.10 Sample distributions by US (Top) and NUS (Bottom), “+” shows the
sampling location and “◦” the sampled value. . . . . . . . . . . . . . . 31
2.11 Reconstruction error curves (“◦” represents the reconstruction of sam-
pled point). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
2.12 Reconstruction error curves (“◦” represents the reconstruction of sam-
pled point) (cont.). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.13 The basis-kernels of interpolations in time domain. . . . . . . . . . . . 34
2.14 The basis-kernels of interpolations in frequency domain. . . . . . . . . . 35
xv
xvi LIST OF FIGURES
2.15 Transposed FIR structure . . . . . . . . . . . . . . . . . . . . . . . . . . 37

2.16 Fractional delay filtering structure . . . . . . . . . . . . . . . . . . . . . 38
2.17 Reconstruction from generalized uniform samples using a CT filter-
bank [53]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.1 a) Original modulated bandpass signal with a bandwidth B = 50 and

fc = 500; b) The corresponding frequency spectrum of sampled-data
signal by LPS, fs = 2fc + B; c) The corresponding frequency spectrum
of sampled-data signal by BPS, Fs = 2B. . . . . . . . . . . . . . . . . . 42
3.2 An example of DSB bandpass signal with a fractional band position
located at [fl , fu ] ∪ [−fu , −fl ]. A BPS rate Fs is selected such that the
sampled-data signal is still a bandpass signal located at [−Fs , 0] ∪ [0, Fs ]. 43
3.3 An example of DSB bandpass signal with a fractional band position
located at [fl , fu ] ∪ [−fu , −fl ]. A BPS rate Fs is selected such that the
sampled-data signal is the equivalent lowpass signal located at DC. . . . 44
3.4 Examples of fractional band position and folding bands defined by dif-
ferent Fs , a) Fs = 556; b) Fs = 585; c) Fs = 592. . . . . . . . . . . . . . 45
3.5 An example of SSB bandpass signal with a fractional band position
located at [fl , fu ] ∪ [−fu , −fl ] . . . . . . . . . . . . . . . . . . . . . . . . 45
3.6 The allowed and disallowed (shaded area) uniform sampling rates ver-
sus the band position, Fs is BPS rate, B is the bandwidth, and the
information band is located at [fl , fu ] ∪ [−fu , −fl ] [16]. . . . . . . . . . . 46
3.7 (a) Heterodyning a bandpass signal y(t) to baseband in order to apply
conventional LPS; (b) The frequency down conversion and baseband
sampling by the equivalent BPS. . . . . . . . . . . . . . . . . . . . . . . 47
3.8 A switch-capacitor (SC) sampling device . . . . . . . . . . . . . . . . . . 48
3.9 Identity elements of ideal uniform BPS . . . . . . . . . . . . . . . . . . . 50
3.10 Illustration of noise aliasing in BPS . . . . . . . . . . . . . . . . . . . . . 50
3.11 Demonstration of noise aliasing. (Top): a) Decimated sampled-data
signal of LPS by factor 50 with a BPF, fs = 25000, SNR≈ 54.4 dB;
b)BPS by Fs = 500, SNR≈ 26.5 dB. (Bottom): LPS by fs = 25000,
SNR≈ 41.5 dB [24]. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3.12 Illustration of jitter on time and amplitude . . . . . . . . . . . . . . . . 52
3.13 Comparison of theoretical and simulated SNDR for y(t) = sin(2πfin t).
Left: for LPS, fin = 10 and fs = 5fin = 50; for BPS, fc = 500 and
Fs = 50. Right: fin = 500, fs = 5fin = 2500. . . . . . . . . . . . . . . . 54
4.1 Identity of signal representation by Papoulis’ generalized sampling the-

orem. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
4.2 Identity of signal representation of derivative sampling, where f <k−1>
denotes the k − 1 order derivative of f (t). . . . . . . . . . . . . . . . . . 57
4.3 Identity of signal representation of recurrent nonuniform sampling. . . . 58
4.4 Identity of signal representation of quadrature sampling, where f˜(t) rep-
resents the Hilbert transform of f (t). . . . . . . . . . . . . . . . . . . . . 58
xvii
4.5 Model of second-order BPS based on the Kohlenberg’s sampling theo-

rem, where x(t) = Re{[f (t) + j f˜(t)]ej2πfc t } is the bandpass signal with
an SSB equivalent lowpass complex signal, f (t) is the real signal and
f˜(t) represents the Hilbert transform of f (t). . . . . . . . . . . . . . . . 60
4.6 Illustration of phase shift by quadrature sampling: initial coordinate
A = [f (nT ) cos(2πfc nT ), −f˜(nT ) sin(2πfc t)], destination coordinate B =
[f (nT + α) sin(2πfc nT ), f˜(nT + α) cos(2πfc nT )], destination coordinate
C = [−f (nT + α) sin(2πfc nT ), −f˜(nT + α) cos(2πfc nT )]. . . . . . . . . 61
4.7 Reconstruction of an SSB signal. (Top): input real bandpass signal,
“×” and “◦” represent the sampling positions of two parallel uniform
samples by a quadrature bandpass sampling; (Center): reconstruction
by Kohlenberg’s interpolation function defined by eq. (4.12); (Bottom):
reconstruction by a modulated sinc function defined by eq. (4.14). . . . 69
4.8 Sequence of operations to obtain lowpass complex samples from uniform
real real samples of a bandpass signal [27] . . . . . . . . . . . . . . . . . 70
4.9 Digital baseband converter with digital product detector (DPD) [28] . . 70
4.10 Generalized Quadrature BPS architecture . . . . . . . . . . . . . . . . . 70
4.11 Illustration of SSB signal spectra. . . . . . . . . . . . . . . . . . . . . . . 71
4.12 Frequency spectra analysis of the proposed GQBPS algorithm. . . . . . 71
4.13 Demonstration of noise gain due to noise aliasing by GQBPS based on
eq. (4.30). (Top): fs = 100, fc = 700, Bef f = 5fc , M = 35; (Bottom):
fs = 700, fc = 700, Bef f = 5fc , M = 5. . . . . . . . . . . . . . . . . . . 72
4.14 Comparison of normalized sampled-data spectra, (Top): by conventional
BPS, (Bottom): by GQBPS, where fc = 700, Bef f = 10fc , B = 5,
fs1 = fc /2 = 350 (solid line) and fs2 = fc /7 = 100 (dash-dotted line) . 73
5.1 Theoretical weights Rpp (f ) based on eq. (5.6) for different jitter cases. . 79
5.2 The PSD of JS on a sinusoid input signal with f = 2 for different jitter
and fs = 5. (Top): α = 0; (Bottom): α = 0.1Ts . . . . . . . . . . . . . . . 81
5.3 The PSD of JS on a sinusoid input signal with f = 2 for different jitter
and fs = 5 (cont.). (Top): α = 0.3Ts ; (Bottom): α = 0.5Ts . . . . . . . . 82
5.4 The PSD of ARS with Poisson process. The input frequency is 2 and
the average sampling rate is 5. . . . . . . . . . . . . . . . . . . . . . . . 83
5.5 The practical p(γ) (in vertical bar) which is used for above simulation
as compared to the theoretical p(γ). . . . . . . . . . . . . . . . . . . . . 84
6.1 A simplified model of a BPS receiver. . . . . . . . . . . . . . . . . . . . 86

6.2 SNDR evaluation for jitter effects on BPS and B = 50, fs = 4B. [a]
fc = 1000 JS. [b] fc = 5000 JS. [c] fc = 1000 ARS. [d] fc = 5000 ARS. . 88
6.3 Comparison of SNR responses of RAs for US and JS (top), NUS and
ARS (bottom) in BPS, B = 50, fc = 100B = 5000, fs = 4B. . . . . . . . 90
6.4 The PSD of sampled-data signal by BPS with fc = 5000, B = 50 and
fs = 10B for JS (top) and ARS (bottom). . . . . . . . . . . . . . . . . . 91
Chapter 1
Introduction
The conventional radio receiver architecture, superheterodyne, has existed for al-
most one century since Edwin H. Armstrong proposed it in the 1910s. Many
variations were proposed afterwards based on the theme, such as single-IF and
dual-IF receivers [1]. Superheterodyne receivers create a beat frequency defined by
the difference between the output of a Local Oscillator (LO) and the input signal
frequency to realize frequency down-conversion. However, the signal located at the
“image band” which is the mirror to the information band with respect to the out-
put of the LO will also be inevitably present at the beat frequency. This signal is
called image of the expected information signal. Normally an Image-Reject Filter
(IRF) is used to suppress the image prior to the mixer. In order to more efficiently
suppress the image, a special receiver family called image-reject receiver was devel-
oped based on the superheterodyne. Two typical architectures of an image-reject
receiver are the Hartley architecture and the Weaver architecture [1]. The other
way to suppress the image is to directly down-convert the RF spectrum to baseband
without IF. The corresponding receiver architecture is called homodyne, “zero-IF”
or“direct-conversion” receiver [1].
In general, a single narrow channel of the RF signal is translated to baseband
before the digitization in an A/D converter. Oversampling is normally used to
reduce the requirements on the dynamic range of the A/D converter. By moving
the A/D converter to the IF, a signal is digitized at IF, and the demodulation
and the detection are realized in the digital domain. The corresponding receiver is
called digital-IF receiver. For these two cases, the sampling in the A/D converter
is a LowPass Sampling (LPS) based on the Shannon’s sampling theorem. It is
known that the frequency translation could also be realized by subsampling (or
undersampling). Replacing the lowpass sampling with a BandPass Sampling (BPS),
the corresponding receiver architecture is the so-called subsampling receiver. In this
case, a continuous-time (CT) signal at IF will be represented in discrete-time (DT)
at a lower IF or baseband.
The receiver architectures mentioned above are mostly designed for single stan-
1
2 CHAPTER 1. INTRODUCTION
dard, narrow band radio communications. For a specific communication standard,

only a limited spectrum is allocated to each user, e.g., 200 kHz for GSM and 30
kHz for IS-54/-136 (more details are contained in Table 1.1). A single channel
Table 1.1: Wireless Radio Communication Standards
Mobile
Frequency Carrier Data
Frequency Access
Standard Band Spacing Rate
Range Method
(MHz) (MHz) (Mbps)
(MHz)
UL:824-849
IS-54/-136 TDMA 25 0.03 0.048
DL:869-894
UL:824-849
IS-95 CDMA 25 1.25 1.228
DL:869-894
UL:890-915
GSM TDMA 25 0.2 0.2708
DL:935-960
DCS 1800 UL:1710-1785
TDMA 75 0.2 0.2708
(EDGE) DL:1805-1880
PCS 1900 UL:1850-1915
TDMA 60 0.2 0.2708
(EDGE) DL:1930-1990
UL:940-956
PDC TDMA 16 0.025 0.042
DL:810-826
DECT 1880-1900 TDMA 20 1.728 1.152
IEEE 802.11a 5150-5350 CSMA-CA 200 OFDM: 20 6-54
FHSS:1 1-2
IEEE 802.11b ISM:2400-2483.5 CSMA-CA 83.5
DSSS:25 5.5-11
BluetoothTM ISM:2400-2483.5 TDMA 83.5 1 1
DCS 1800 UE:1710-1785
CDMA 75 5 3.84
(W-CDMA) BS:1805-1880
PCS 1900 UE:1850-1915
CDMA 60 5 3.84
(W-CDMA) BS:1930-1990
UL:1920-1980 CDMA
UMTS(3G) 60 5 3.84
DL:2110-2170 /TDMA
is normally selected prior to the sampling in the A/D converter. Alternatively,

the channel can be selected as late as possible. This is introduced in a family
of WideBand (WB) receivers, e.g., wide-band IF receivers and generic wide-band
receivers [2] [3].
With the evolution of radio communications, a combination of multi-band an-
tennas and RF conversions, wide-band A/D converters and their implementations
at IF results in multi-mode multi-band radio receivers. Both Software Defined Ra-
dio (SDR) receivers and homodyne receivers support this multi-mode multi-band
radio communications [4] [5]. The design key of SDR is the placement and design
techniques of A/D converters whose goal is to put the A/D converter as close as
possible to the antenna. The big revolution of SDR receivers compared to the con-
ventional receivers is that SDR technology replaces many tasks in the radio receiver
with digital processing, including down-conversion, filtering (mode- or band- selec-
tion), demodulating, decoding and converting to the desired data format. SDR
receivers make use of digital signal processing (DSP) to implement the complex
tasks in today’s communication systems, and are easy to be extended for more
complex systems. By using homodyne receivers, the mode-selection can be realized
1.1. SUPERHETERODYNE RECEIVERS 3
by an external digital controller, and the hardware in the analog part is shared by
different communication bands or modes as much as possible. However, the exten-
sion to a complex communication system is hard. In addition, all the associated
problems of the homodyne receiver are also present.
In this chapter, traditional superheterodyne and homodyne receiver architec-
tures are shown and compared with two WB receiver architectures: the wide-band
IF receiver and the generic wide-band receiver. After that, the SDR receiver is
shown and compared with the homodyne receiver with respect to multi-mode multi-
band radio communications. The SDR receiver with different data acquisition tech-
nologies are also discussed. Then an outline of technical problems in SDR receivers
by using the BPS technique is presented, and previous work on BPS is reviewed.
Finally, the contributions in the following chapters are summarized.
1.1 Superheterodyne receivers

In the literature there is usually no distinction between heterodyne and superhetero-
dyne architectures. To “heterodyne” means to mix two frequencies and produce a
beat frequency defined by either the difference or the sum of the two. To “super-
heterodyne” is only to produce a beat frequency defined by the difference of the
two frequencies. Two stages of down-conversion (dual-IF) based on the theme of
superheterodyne is mostly used in today’s RF receivers (see Fig. 1.1). This receiver
Antenna
BPF Amplifier IRF BPF BPF Amplifier
LO LO
RF IF
1
0 11
00
1
0 0
1 00
11
0
1 0
1 00
11
0
1 0
1
0
1 00
11
0
1 0
1 00
11
0
1 00
11
0
1
0
1
0
1
··· 0
1
0
1
0
1 ··· 00
11
00
11
0
1 0
1 00
11
0
1 0
1 00
11
0
1 0
1 00
11
00
11
Figure 1.1: Conventional dual-IF superheterodyne receiver architecture
translates the signal to a low frequency band by two stages of down-conversion mix-
ing and relaxes the requirement on the Q-factor of the channel-select filter. The
first IF might be between 70 and 250 MHz for 2G, 2.5G and 3G applications [6].
For narrow-channel standards, the second IF is often equal to 455 kHz, but for
wide-channel application such as DECT, it may be several megahertz. These num-
bers vary a lot in present systems. If the second IF of a dual-IF receiver is equal to
zero, the second down-conversion normally separates the signal to I (in-phase) and
Q (quadrature) components for Single-SideBand (SSB) communication systems or
frequency-/phase-modulated signals, and the corresponding demodulation and de-
tection are performed at baseband. This down-conversion is realized by two LOs
which have a 90◦ phase shift between each other. Any offset from the nominal 90◦
phase shift and the amplitude mismatches between I and Q components will raise
the Bit Error Rate (BER). If the second IF is not equal to zero, the receiver be-
comes a digital-IF receiver. The IF bandpass signal is directly processed by an A/D
converter, and the I/Q mismatch can be avoided. After that, IF demodulation and
detection are processed in the digital domain.
Both down-conversion schemes entail the image problem. The choice of two
IFs faces the trade-off between the image rejection (or sensitivity) and channel-
selection (or selectivity). If the IF is high, the image band appears far way from
the information band such that the image can be easily suppressed by an IRF.
However, the channel selection filter will require a high Q-factor to select a narrow
channel at a high IF. On the contrary, if the IF is low, the design of the chan-
nel selection filter becomes easier but the image band is so close to the information
band that it becomes difficult to achieve a proper image suppression by a BandPass
Filter (BPF). More than one stage of down-conversion makes the trade-off easily
achieved. In a dual-IF superheterodyne receiver, the first IF is selected high enough
to efficiently suppress the image, and the second IF is selected low enough to relax
the requirement on the channel selection filter. The selectivity and sensitivity of
the superheterodyne makes it a dominant choice in RF receiver architectures. Un-
fortunately, the high Q-factors of the discrete-components in the superheterodyne
receiver make it difficult to fully integrate the whole front-end on a single chip.
1.2 Homodyne receivers

In a homodyne receiver, no IF stage exists between RF and baseband. The input
of the A/D converter is located at baseband (see Fig. 1.2). The channel selection
filter is just a lowpass filter prior to the A/D converter. The homodyne receiver
has two advantages compared to superheteodyne receiver. First, the architecture is
simpler. Second, the image problem can be avoided due to zero IF (i.e., fIF = 0)
such that no IRF is needed.
The homodyne receiver allows a higher level of integration than the super-
heterodyne receiver as the number of discrete components are reduced. However,
this receiver inevitably suffers from the problems of LO leakage and DC-offset. The
output of the LO may leak to the input of the mixer or the LNA due to improper
isolation. The leaked signal will be mixed with the output of the LO (i.e., the origin
1.3. WIDEBAND IF RECEIVERS 5
Antenna
ADC
LPF
90◦
LO
BPF Amplifier
ADC
LPF
I Q
LO
RF
1
0 11
00 11
00
0
1 00
11 00
11
0
1 00
11 00
11
0
1 00
11 00
11
00
11
0
1
0
1··· 00
11
00
11··· 00
11
00
11
0
1 00
11
00
11 00
11
0
1 00
11 00
11
0
1 00
11
Figure 1.2: Homodyne receiver architecture
of the leaked signal) and produce a DC component at the output of the mixer. This
is called self-mixing. LO leakage to the antenna may result in a time-varying DC
offset due to self-mixing. The undesired DC component and offset will corrupt the
information signal that is present at the baseband. By using the quadrature down-
conversion in homodyne receivers, I/Q mismatch is another associated problem.
Because the down-converted signal is located at zero frequency, the flicker noise or
1/f noise of devices will also corrupt the information signal.
1.3 Wideband IF Receivers

A superheterodyne receiver with an RF channel-select frequency synthesizer and an
IF or baseband channel-select filter is a NarrowBand (NB) receiver. An alternative
architecture, called Wide Band (WB) IF receiver (see Fig. 1.3), postpones the
channel-select frequency synthesizer to IF and channel-select filter to baseband [2].
Antenna
ADC
BPF Amplifier LPF LPF

I Q I Q
LO LO
RF IF
1
0
0
1 11
00
00
11 11
00 11
00
0
1 00
11 00
11 00
11
0
1 00
11 00
11 00
11
00
11
0
1 00
11 00
11 00
11
0· · · 00· · · 00· · ·
0
1 00
11 00
11
11 00
11
1 11 00
11 00
11
0
1
0
1 00
11
00
11 00
11 00
11
0
1 00
11 00
11 00
11
00
11
Figure 1.3: Wideband IF receiver architecture with double down-conversion
The entire frequency band of information signal located at RF is translated to

IF by multiplying the output of the LO with a fixed frequency. The IF signal passes
through a LPF such that the frequency components above the IF band are removed.
One channel out of the entire band is first translated to DC by a tunable LO and
then fed into a LPF. The selected lowpass channel signal is processed further by an
A/D converter which is the same as in the superheterodyne and homodyne receivers.
Compared to the traditional superheterodyne receiver, this receiver architecture is
well-suited for full integration, and it has also the potential to be implementated
for multi-band multi-mode radio communications.
1.4 Generic Wideband Receivers

It is advantageous to use conventional receiver architectures for single-mode NB ra-
dio communications since the technologies are mature and it is also easy to fulfill the
system performance requirements. Nevertheless, driven by the increased speed of
microprocessors and the high performance of A/D converters, a WB radio architec-
ture has drawn more and more attention for the support of multi-band multi-mode
radio communications. From a general aspect, the wideband IF receiver mentioned
in section 1.3 is not a real WB receiver since the input of the A/D converter is
still NB. A generic wideband receiver was depicted in [3]. In this architecture, the
whole band centered at a RF corresponding to a specific communication standard
is selected by a tunable LO. Then the complete signal spectra within the band is
translated to baseband and digitized. The channel is selected in the digital domain
by a digital channelizer which is much easier to realize than in the analog domain.
1.5. SOFTWARE DEFINED RADIO RECEIVERS 7
For multi-mode WB operation, the frequency of the tunable LO is adapted to fit a

particular standard.
Antenna
ADC
BPF Amplifier BPF
I Q
LO
LO
RF IF −fs /2 fs /2
−3fs /2 3fs /2
··· ··· ··· ··· ··· ··· ··· ···
Figure 1.4: Generic wideband receiver architecture
1.5 Software Defined Radio Receivers

The concept of Software Defined Radio (SDR) was originally conceived for military
applications. It consists of a single radio receiver to communicate with different
types of military radios using different frequency bands and modulation schemes [7].
This concept is starting to be introduced into commercial applications. SDR means
a radio where functionality is extensively defined in software, and it supports multi-
band multi-mode radio communications. It constitutes the second radio evolution
since the radio systems migrated from analog to digital in 1970s and 1980s. Modern
radio designs mix analog hardware, digital hardware and software technologies.
Two key issues of SDR are the placements of the A/D converters and the per-
formance of DSP coping with the large number of samples [4]. One goal of SDR
is to put the A/D converter as close as possible to the antenna. The sampling
function block of an A/D converter can be either classic LPS (oversampling) or
BPS (undersampling). By LPS, the sampling rate of an IF bandpass signal is high,
and the performance requirements, e.g., linearity, noise floor and dynamic range, on
the A/D converter are stringent. The IF WB bandpass signal can also be sampled
by BPS with a sampling rate which is only slightly larger than twice the infor-
mation bandwidth. The lower sampling rate alleviates the requirements on the
following A/D converter. In addition, BPS can realize down-conversion through

the intentional signal spectral folding such that the input IF signal is sampled to
discrete-time (DT) at a lower IF or baseband at the output of the BPS. The con-
ventional mixer for down-conversion is redundant when BPS is used, and the A/D
converter can be moved further forward to the antenna. The receiver architecture
of SDR by BPS (thereafter called BPS receiver) is shown in Fig. 1.5. The part
of dashed-line box in Fig. 1.4 is also called the equivalent LPS system of BPS, an
ideal image-reject mixer followed by a lowpass sampler. From the view of multi-
band multi-mode communications and the placement of the A/D converter, both
the homodyne receiver and the BPS receiver are candidates for the SDR implemen-
tation.
Antenna
ADC
BPF Amplifier BPF
LO
(n − 1/2)fs nfs (n + 1/2)fs

RF IF
··· ··· ··· ··· ··· ···
Figure 1.5: Software defined radio receiver architecture by BPS
Many single-mode wideband homodyne receivers were designed for W-CDMA

(Wideband Code-Division Multiple-Access) [8] [9] [10] [11]. A design of a multi-
band multi-mode receiver for four standards in homodyne was presented in [5]. All
the problems associated with the homodyne receiver, e.g., LO leakage, DC-offset,
I/Q mismatch and flicker noise, inevitably happen and are treated in many different
ways. The selection among different standards is realized by an external digital
controller and the hardware is shared as much as possible by different standards.
Bandpass sampling theory has been studied for more than half a century. Cauchy
[12], Nyquist [13], Gabor [14] and Kohlenberg [15] did the earliest contributions.
A late comprehensive introduction on the theory of BPS can be found in [16]. In
recent years, radio receiver front-ends have been implemented by BPS [17] [18] [19]
[20] [21]. Harmful signal spectrum folding (or aliasing), noise aliasing and timing
jitter are three associated problems in the BPS receiver. The A/D converter in
1.6. OUTLINE OF TECHNICAL PROBLEMS 9
the BPS receiver could directly digitize the received RF signal and process the RF
demodulation and detection in the digital domain, provided that all the associated
problems were solved. Even though the design technology of homodyne receiver is
mature, the basic receiver architecture is fixed and designers could only find the
solutions to the associated problems at the circuit level. Orienting the design goal
toward SDR, the BPS receiver is easily extended and used for a more complicated
communication. It would be more advantageous to study the BPS receiver and
present solutions to the associated problems.
In summary, Table 1.2 provides a high level comparison among the above six
receiver architectures.
Table 1.2: High Level Comparison of Receiver Architectures
Receiver Full A/D Potential Selecting Image

Architecture Integration Converter for Multi-mode Filter Rejection
BPF
Superheterodyne low NB low IF BPF
IR mixer
Homodyne high NB/WB high LPF N/A
BPF
Digital IF high WB high IF BPF
DSP
Wideband IF high NB high LPF IR mixer
Generic wideband high WB high IF BPF IR mixer
BPS receiver high WB high IF BPF BPF or N/A
1.6 Outline of Technical Problems

BPS realizes frequency down-conversion by intentional signal spectral folding in-
stead of mixing. However, harmful signal spectrum folding, noise aliasing and
sampling timing error (or jitter) are the main causes of performance degradation
in BPS system.
A bandpass signal is expressed as
y(t) = Re{x(t)ej2πfc t }, (1.1)
where fc is the carrier frequency, x(t) is the equivalent lowpass information signal or
the complex envelope of y(t), y(t) is band-limited within [−fu , −fl ] ∪ [fl , fu ], where
fl and fu are respectively the lower and upper frequencies of the bandpass signal in
the positive frequency band. The situation of the signal spectrum aliasing depends
on the band position which is defined as the fractional number of bandwidths from
the origin at which the lower band edge resides [16]. As shown in Fig. 1.6, a
fractional number r represents the band position of y(t) and r = fl /B. The classic
bandpass sampling theory for Uniform Sampling (US) states that [16]
The signal can be reconstructed if the sampling rate is at least fsmin = 2fu /n,
where n is the largest integer within fu /B, denoted by n = Ig [ fBu ].
|Y (f )|
2B 2B
−fu −fc −fl 0 fl fc fu f

Figure 1.6: Frequency spectrum of bandpass signal
To avoid harmful signal spectrum folding by uniform BPS, the minimum sampling
rate can only be used for the special case when y(t) has an integer band position
(i.e., r = bfl /Bc = fl /B and r = n − 1, where b c denotes a floor operator). In
Jerri’s tutorial review [22], it was mentioned that
Unless the signal is band-limited to (−2πW, 2πW ) there will always be an alias-
ing error when we sample at the required Nyquist rate. So if there is any alias
free sampling it must be based on a rate different from that of the Nyquist
rate or in other words sampling at unequally spaced instants of time.
Nonuniform BPS might have also the potential to suppress harmful signal spectrum
aliasing for any band position [15] [23]. However, a signal cannot be reconstructed
from its nonuniform samples by using a conventional lowpass filtering process. Re-
construction Algorithm (RA) of NonUniform Sampling (NUS) which is extensively
used in image processing are proposed and studied for our implementation in radio
receivers.
The noise combined in each of the Nyquist bands within the effective bandwidth
Bef f of introduced thermal noise (e.g., kT /C noise) in BPS causes performance
degradation. This is the so-called noise aliasing. Noise aliasing is a consequence of
decreasing sampling rate by BPS as compared to lowpass sampling (LPS). For a
certain band position, the signal-to-noise ratio (SNR) of a sampled-data signal by
BPS depends on the ratio of Bef f and fs significantly (fs is the sampling rate of
BPS). The lower the value of fs , the worse the SNR performance [24].
Under the effects of jitter, the samples become randomly distributed. It is known
that jitter effects depend on both the standard deviation of random jitter and the
input frequency of the signal [25]. Small jitter noise can be approximately assumed
as sampled-data Additive White Gaussian Noise (AWGN) [26]. For large jitter,
this assumption is not valid anymore [24]. With the increase of input frequency
of BPS, jitter becomes a crucial problem. For the same sampling rate, the jitter
effects in BPS are larger than in the equivalent LPS system. The noise power of
BPS corresponding to the same normalized standard deviation of jitter (στ /Ts ) is
larger than for LPS due to the higher input signal frequency of BPS.
1.7. OVERVIEW OF PREVIOUS WORK ON BPS 11
1.7 Overview of Previous Work on BPS

Under the design tendency of low cost, low power consumption and full integration,
CMOS techniques have been extensively used in RF designs. Especially with the
growing of CMOS technology, it can perform well on a very high frequency signal
today and the cost is also lower than other techniques (e.g., GaAs, BiCMOS).
Recently, many successful RF CMOS designs based on the concept of SDR and
BPS have been presented.
Switched-capacitor (SC) circuits are often used in these designs and the analog
signal is processed in the DT domain. Sampling can be performed in both conven-
tional voltage-mode and current-mode. The work in [18] is based on voltage-mode
sampling. A 2.4-GHz CMOS RF sampling receiver front-end was designed for the
IEEE standard 802.11b focusing on an RF subsampling mixer. This design in-
tegrates the subsampling mixer, clock generator, DT down-conversion filter and
output buffers into a single chip. The sampling rate was chosen to be about 43
times the channel spacing, around half the input signal carrier frequency. An accu-
rate clock is generated by a specific generation scheme and the measured sampling
jitter is around 0.54 ps. In [20], a DT Bluetooth receiver using BPS is designed
in current-mode CMOS technology. The principle of corresponding multi-tap di-
rect sampling mixer is shown in [21]. The input voltage signal is first converted to
current by a transconductance amplifier. The input frequency of the direct sam-
pling mixer is 2.4 GHz and the effective data rate at the output is fo /N , where the
sampling takes place on the input signal at the rate of fo = 2.4 GHz and N is a
decimation factor.
In the existing BPS implementations in voltage-mode [18] [27] [28] [29], uniform
BPS is normally used and a thorough frequency plan is needed to avoid wrong signal
spectral folding and image problems. Nevertheless, the performance degradation
due to noise aliasing is still present. Therefore, the ratio of fs /2B is still large for
attaining a certain SNR performance. As a consequence of high input frequency of
sampling devices, jitter is still a crucial problem.
Charge sampling is a new sampling technology. It avoids the voltage settling
problem. Combining with IIR filtering performed by a cyclic charge readout, the
noise aliasing can be suppressed to a certain degree [20] [21]. Both voltage sampling
and charge sampling might be the potential way to implement BPS. This thesis work
is mostly based on voltage-mode sampling, but the basic theory of charge sampling
will be given in Chapter 2 in comparison to voltage-mode sampling.
1.8 Summary of Contributions

In this thesis work, NUS and reconstruction are mainly studied.
• In chapter 2, two kinds of deterministic sampling techniques, Uniform Sam-

pling (US) and NonUniform Sampling (NUS) are shown. A filter generalized
by a certain RA for LPS is introduced. A general reconstruction formula in
terms of a basis-kernel (BK) is used to describe the algorithms. The com-

putational complexity and implementation of these RAs are evaluated and
compared.
• In chapter 3, BPS technique is introduced. The associated problems, avail-

able sampling selection, noise aliasing and jitter are studied and numerically
analyzed.
• In chapter 4, a brief overview of the Papoulis’ generalized sampling theorem,

which is the extension of Shannon’s sampling theorem, is shown. A General-
ized Quadrature BPS (GQBPS) algorithm based on the Papoulis’ sampling
theorem is proposed for suppressing the noise aliasing. Theoretical analyses
show that GQBPS might be a potential way to reduce the noise aliasing at
the cost of a more complicated RA.
• In chapter 5, two classes of nonuniform random sampling, Jitter Sampling

(JS) and Additive Random Sampling (ARS), are studied and analyzed by
Power Spectral Density (PSD). The conditions of aliasing-free sampling are
verified in simulations.
• In chapter 6, the reconstruction performances by different RAs of lowpass

case in the presence of AWGN and jitter are evaluated and compared in a
modeled BPS system.
• In chapter 7, the thesis is concluded and some future work is proposed.

Through the thesis, all the signals involved in the theoretical analyses are as-
sumed ideal band-limited, which means that their Fourier transforms are zero for
|f | > B, although this is not ideally realizable in practice.
Chapter 2
Sampling and Reconstruction
With the launch of digital radio communications, A/D and D/A converters become
important devices as the interface between RF conversions and digital signal pro-
cessing. A natural signal, such as speech, music, image and electromagnetic wave,
is generally an analogue signal in a continuous-time (CT) domain. To process a
signal digitally, it has to be represented as a digital format in a discrete-time (DT)
domain. It is required that this digital format is fixed, and uniquely represents all
the features of the original analogue signal. The reconstructed CT signal from this
digital format may not be exactly the same as the original analogue signal, but it
is a goal to decrease the difference as much as possible.
The two basic operations of an A/D converter are sampling and quantization.
Sampling is to convert a CT analogue information signal into a DT representa-
tion by measuring the value of the analogue signal at regular or irregular intervals.
Quantization is to convert a value or range of values into a digital value. The
quantization level determines the resolution of the A/D converter (in bits per sam-
ple). In this chapter, two ideal sampling methods, voltage sampling and charge
sampling, are introduced. Regular sampling and irregular sampling are compared.
A filter generalized by a Reconstruction Algorithm (RA) is proposed and studied
in terms of a Basis-Kernel (BK). Nine RAs are evaluated and compared based on
their performance, computational complexity and hardware implementation.
2.1 Sampling
Nowadays the sampling theorem plays a crucial role in signal processing and com-
munications. The selecting of a time sequence x(tn ) to represent a CT function
x(t) is known as sampling.
Sampling methods in electrical unit include voltage sampling and charge sam-
pling. Voltage sampling is a conventional method that is realized by the sample-
and-hold (S/H) circuit. It tracks an analog signal and stores its value as a voltage
across a sampling capacitor for some length of time. Charge sampling does not
13
14 CHAPTER 2. SAMPLING AND RECONSTRUCTION
track the signal voltage but integrates the signal current within a given time win-
dow [30]. An analog signal in voltage mode is first converted to current mode by a
transconductor before charge sampling. As compared to voltage sampling, charge
sampling has the advantage that the bandwidth of the charge sampling device only
relies on the sampling duration but not on the switch-on resistance so that a wide-
band sampler design is more feasible [31]. BPS can also be performed by a charge
sampling [17] [20] [21] besides a voltage sampling.
Whether the sampled-data signal uniquely represents the original signal or not
depends on the sampling patterns and their implementations. Referring to the
sampling period (or interval), sampling can be ideally divided into two categories,
Uniform Sampling (US) and NonUniform Sampling (NUS). It is justified to assume
that the sampling set is uniformly distributed in many applications, i.e., the samples
are acquired at the regular time instants. However, in many realistic situations,
the data is known only in a irregularly spaced sampled set. This irregularity is a
fact of life and prevents the standard methods of Fourier analysis. For example in
communication systems, when data from a uniformly distributed samples is lost,
the obtained result is generally nonuniformly distributed, the so-called missing data
problem. Scratching a CD is also such kind of a problem. On the contrary, it may
be of advantage to use NUS patterns for some special cases (e.g., an aliasing-free
sampling) [32] [22]. For NUS, there are four general sampling scenarios: generalized
nonuniform sampling [33], Jitter sampling [34], Additive random sampling [32],
and Predetermined nonuniform sampling. Without any specifications, the NUS
mentioned in this chapter is predetermined and each sampling instant is known
with high precision.
2.1.1 Voltage sampling and Charge sampling

Voltage sampling A voltage sampling process can be modeled as an input CT
signal x(t) multiplied by a sampling function s(t) (see Fig. 2.1). The CT
x(t) xs(t)
s(t)
Figure 2.1: Voltage sampling process
sampled-data signal xs (t) is given by
xs (t) = x(t)s(t). (2.1)

P∞
For ideal voltage sampling process, s(t) = n=−∞ δ(t − tn ), where {tn }
represents the set of sampling instants. The multiplication in time domain
corresponds to the convolution in frequency domain, i.e.,
Xs (f ) = (X ? S)(f ), (2.2)
2.1. SAMPLING 15
where ? represents an convolution operation, X(f ) and S(f ) are the Fourier
transforms of x(t) and s(t), respectively.
Charge sampling Charge sampling integrates charge within a time window [tn , tn +
∆t] instead of storing the voltage value across a sampling capacitor. It is mod-
eled as an input CT signal x(t) convolved with a sampling function s(t) (see
Fig. 2.2). The sampled-data signal xs (t) is given by
x(t) xs(t)
?
s(t)
Figure 2.2: Charge sampling process
∞ Z
X tn +∆t
xs (t) = (x ? s)(t) = x(ξ)sn (t − ξ)dξ. (2.3)
n=−∞ tn
In frequency domain,
Xs (f ) = X(f )S(f ). (2.4)
The sampling function s(t)

P is a series
P of pulses with a duration of ∆t. For ideal
charge sampling, s(t) = n sn (t) = n µ(t − tn ) − µ(t − tn − ∆t), where µ(t) is a
Heaviside’s step function.
S/H circuit by voltage sampling has been extensively applied in common data
acquisition systems (e.g., speech communication, music, image processing) and it
has a good performance. Due to the tendency of moving the A/D converter as
close as possible to the antenna for SDR in wireless radio communications, both
sampling methods might be the potential way to implement BPS. Without any
specification, the sampling process in the following will use the voltage sampling.
2.1.2 Uniform sampling and Nonuniform sampling

Uniform Sampling (US) For an ideal US process, tn = nTs . Starting from
eq. (2.1), the sampled-data signal is given by
∞
X
xs (t) = x(t) δ(t − nTs ), (2.5)
n=−∞
and the Fourier transform of xs (t) can be expressed as

Z ∞
Xs (f ) = xs (t)e−j2πf t dt
−∞
∞
X
= x(nTs )e−j2πf nTs . (2.6)
n=−∞
This is the well-known Discrete-Time Fourier Transform (DTFT). The Dis-

crete Fourier Transform (DFT) is a special case of DTFT, which is defined
to be the DTFT evaluated at equally spaced frequencies over the Nyquist
interval [0, 2π). The N -point DFT of a length M signal is defined as
X
M −1
X(k) = x(m)e−j2πkm/N , k = 0, 1, · · · , N. (2.7)
m=0
By using Poisson summation formula [35], eq. (2.6) can be written as
∞
X
Xs (f ) = fs X(f − mfs ), (2.8)
m=−∞
where fs = 1/Ts . Obviously, the frequency spectrum of a sampled-data signal

is a series of copies of the original CT signal and Xs (f ) is a periodic function
with period fs (see Fig. 2.3).
0.1
0.05
x(t) 0 a)
−0.05
−0.1 t
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.015
0.01
X(f) b)
0.005
0 f
−400 −300 −200 −100 0 100 200 300 400
0.015
0.01
TsXs(f)
c)
0.005
0 f
−400 −300 −200 −100 0 100 200 300 400
Figure 2.3: a) Original CT band-limited signal, B = 50 and samples by US, fs =

200; b) The corresponding frequency spectrum of CT signal; c) The corresponding
frequency spectrum of sampled-data signal.
2.1. SAMPLING 17
A band-limited signal can be completely determined by a US sequence with

the sampling rate of at least twice the maximum frequency B (critical- or
over-sampling) according to the Shannon’s sampling theorem [36]:
Theorem 1: If a function f (t) contains no frequencies higher than W cps

(in cycles per second), it is completely determined by giving its ordinates
at a series of points spaced 1/2W seconds apart.
When fs < 2B (undersampling), the frequency components above B will be

aliased back into the Nyquist band [−fs /2, fs /2] such that the original signal
cannot be uniquely reconstructed from the sampled-data signal. For a LPS
process, the input signal is regarded as a lowpass signal with a bandwidth con-
sisting of the maximum frequency component, and critical- or over-sampling is
normally used to avoid the harmful signal spectrum aliasing. The input signal
of BPS is, however, always a bandpass signal such that BPS can make use of
a harmless signal spectrum aliasing by tactically selecting the undersampling
rate. More discussion about BPS will be presented in Chapter 3.
Nonuniform Sampling (NUS) For an ideal NUS process,
∞
X
xs (t) = x(t) δ(t − tn ), (2.9)
n=−∞
and tn 6= nTs . The frequency spectrum of xs (t) is not necessarily periodic

and the Fourier transform becomes
∞
X
Xs (f ) = x(tn )e−j2πf tn . (2.10)
n=−∞
The corresponding Energy Density Spectrum (EDS) of an ideal NUS is the

magnitude squared of the Fourier transform [37], i.e.,
Ps (f ) = |Xs (f )|2
= [· · · + x(t0 ) cos 2πf t0 + x(t1 ) cos 2πf t1
+ x(t2 ) cos 2πf t2 + · · · + x(tn ) cos 2πf tn + · · · ]2
+ [· · · + x(t0 ) sin 2πf t0 + x(t1 ) sin 2πf t1
+ x(t2 ) sin 2πf t2 + · · · + x(tn ) sin 2πf tn + · · · ]2 . (2.11)
NUS is well applied for obtaining oscillograms in oscilloscopes and spectro-

grams for spectral analysis [38]. The aperiodic property of the frequency
spectrum enables NUS to suppress harmful signal spectrum aliasing.
As shown in Fig. 2.4, given a wanted signal s(t) = cos(2π · 2t) (solid line)
and an interference signal i(t) = cos(2π · 3t) (dashed line), when fs = 5, the
component of f = 3 is larger than fs /2 and will be folded back to f = 2 (see

a)). By intentionally introducing a random shift with a uniform distribution
U (−αTs , αTs ) (α is a scale factor) on the equidistant US time instants, for
instance α = 0.3 for b) and 0.7 for c) in Fig. 2.4, the aliasing effect is reduced
to a certain degree. Obviously, this aliasing can be overcome by using NUS.
NUS relaxes the requirements on the anti-aliasing (AA) filter.
1 1
0.5
|X (f)|2
A(t)
0 0.5 a)
s
−0.5
−1 0 f
t
0 0.2 0.4 0.6 0.8 1 0 1 2 3 4 5
1 1
0.5
|X (f)|2
A(t)
0 0.5 b)
s
−0.5
−1 0 f
0 0.2 0.4 0.6 0.8 1t 0 1 2 3 4 5
1 1
0.5
2
|Xs(f)|
A(t)
0 0.5 c)
−0.5
−1 t 0 f
0 0.2 0.4 0.6 0.8 1 0 1 2 3 4 5
Figure 2.4: US and NUS sequences (left) and corresponding

R normalized ESD spec-
trum (right) based on eq. (2.11) (normalized to |Xs (f )|2 df = 1) [26].
2.1.3 Nonideal sampling process
In an A/D converter, when the input signal is band-limited, the sampling process
is realized by a two-step sampling, an ideal sampling followed by a pulse shaping
filtering, as shown in Fig. 2.5. When the band is not limited, an AA filter prior to
the sampling is needed to restrict the bandwidth of the input signal.
For the two-step sampling case with an ideal arbitrary sampling scheme, the
2.2. RECONSTRUCTION 19
x(t) xs(t) y(t)

h(t)
P∞
n=−∞ δ(t − tn)
Figure 2.5: Two-step sampling paradigm
output in the time domain is given by

y(t) = xs ? h(t)
Z ∞
= xs (τ )h(t − τ )dτ. (2.12)
−∞
Substituting eq.(2.9) into eq.(2.12),

Z ∞ X ∞
y(t) = x(tn )δ(τ − tn )h(t − τ )dτ
−∞ n=−∞
∞
X
= x(tn )h(t − tn ). (2.13)
n=−∞
The corresponding frequency domain expression of y(t) is expressed as

Z ∞ X ∞
Y (f ) = x(tn )h(t − tn )e−j2πf t dt
−∞ n=−∞
∞
X
= x(tn )H(f )e−j2πf tn
n=−∞
= H(f )Xs (f ), (2.14)
where Xs (f ) is given by eq.(2.8) and eq.(2.10) for US and NUS, respectively. The
form of eq. (2.13) may be interpreted as the convolution summation of the mod-
ulated delta-function series x(tn ) with the filter weighting function h(t). A step
function (or zero-order holding) is mostly used in the A/D converter as a filter,
½
1, tn−1 ≤ t < tn
h(t) = , (2.15)
0, otherwise
which is equivalent to the sample-and-hold interpolant. Eq. (2.14) also indicates
a complete reconstruction of signal x(t) as long as the sampling rate satisfies the
Nyquist criterion and H(f ) is well designed.
2.2 Reconstruction
Depending on the context, “reconstruction” has different definitions. Image re-
construction is defined in imaging technology wherein data is gathered through
methods such as computerized tomography-scan (CT-scan) and magnetic resonance

imaging (MRI), and then reconstructed into viewable images [39]. In analog signal
processing, reconstruction mostly means that a continuous-time signal is obtained
from the DT data by an interpolation filter or some other filtering processes.
Although modern data processing always uses a DT version of the original sig-
nal that is obtained by a certain sampling pattern on a discrete set, reconstruction
to a continuous version of sampled data is also needed for some specific applica-
tions. In Hi-Fi applications such as digital audio, to maintain high quality in the
resulting reconstructed analog signal, a very high quality analog reconstruction fil-
ter (postfilter) is required. Reconstruction from one discrete set to another is also
useful in the non-fractional sampling rate alternation in digital signal processing.
Additionally, in radio receiver front-ends, if the output of the sampling process is
not uniformly distributed, a reconstruction process is needed to reconstruct the
nonuniform samples to uniform distributed samples prior to the quantizer.
According to the Shannon’s sampling theorem [36], a band-limited signal can be
exactly reconstructed from its samples by US. The perfect reconstruction formula
derived by Whittaker [40] for a critical uniform sampling is given by
∞
X
x(t) = x(nTs )sinc[2B(t − nTs )], (2.16)
n=−∞
where x(nTs ) represents samples at the series of equidistant sample instants, Ts =

1/(2B), and sinc(x) = sin(πx)/(πx). The reconstruction of the input signal is
realized by a convolution summation of uniform distributed samples x(nTs ) with a
sinc function which is equivalent to ideal low-pass filtering.
In practice, the CT signal reconstruction is enhanced by first passing the sampled-
data signal xs (t) through a holding circuit with the function of eq. (2.15), and then
feeding into a LPF or other RAs. The reconstruction discussed in the thesis is only
realized by a certain RA without any enhancement from the zero-order holding.
For NUS, even if there is a large number of samples, only few of them possess
a uniform distribution property with respect to the average sampling rate. The
expansion of X(f ) does not consist of periodic replicas of the fundamental spectrum.
Consequently, the signal cannot be determined uniquely by the samples with only
a lowpass filter. Based on the Fourier series expansion, X(f ) can be generally
expanded as
∞
X
X(f ) = cn e−j2πf τn , (2.17)
n=−∞
where τn is the set of sampling instants either uniformly or nonuniformly dis-

tributed. Using inverse Fourier transform, we obtain the general reconstruction
2.2. RECONSTRUCTION 21
formula:
Z Ã ∞
!
B X
−j2πf τn
x(t) = cn e ej2πf t df
−B n=−∞
∞
X
= 2B cn sinc[2B(t − τn )]. (2.18)
n=−∞
For US τn = nTs , cn = x(nTs )/2B and eq. (2.18) is exactly the same as
eq. (2.16). However, for NUS, since τn = tn and cn 6= x(tn )/2B except when
tn = nTs , the reconstruction formula of eq. (2.18) cannot directly represent the
original signal x(t) unless cn is determined. RAs are expected to accurately predict
the original signal x(t) from the nonuniform samples x(tn ).
In biomedical image processing, CT-scan and MRI frequently use the NUS pat-
tern in the frequency domain. Four sampling patterns are shown in Fig. 2.6. The
0.6
0.4
0.2
−0.2
−0.4
−0.5 0 0.5
0.6 0.6
0.4 0.4
0.2 0.2
0 0
−0.2 −0.2
−0.4 −0.4
−0.5 0 0.5 −0.5 0 0.5
Figure 2.6: Sampling patterns of nonuniform sampling [41]. (Top-left): Polar

sampling grid; (Top-right): Spiral sampling grid; (Bottom-left): variable-density
nonuniform sampling grid; (Bottom-right): general nonuniform sampling grid.
sampled data of CT-scan and MRI are measured in the Fourier frequency domain.
The RA is needed to derive the Cartesian US grid (see Fig. 2.7) from the acquired
data prior to the inverse Fourier transform operation. Inspired by the applications
0.6
0.4
0.2
−0.2
−0.4
−0.5 0 0.5
Figure 2.7: Cartesian uniform sampling grid.
in biomedical image processing, some RAs extensively used in image reconstructions

are proposed for the applications of radio communications.
However, the reconstruction process in radio communications is different from
that in biomedical image processing. In radio communications, both sampling and
reconstruction are in the time domain while they are in the frequency domain in
image processing. Additionally, in radio communications, the RA can be used to
reconstruct a set of unknown data at a regular time set from the NUS sequence.
Then the reconstructed result can be directly fed into the following digital signal
processing block (e.g., A/D converter). It is also possible to convert the samples
by NUS to a CT signal when an analog signal is needed (e.g., in Hi-Fi) in the
processing steam, which is different from the reconstruction in image processing.
2.3 Basis-Kernel (BK)
It is known that {ej2πf tn } is a complete basis for X(f ) within the bandwidth
[−B, B] and that {sinc[2B(t − tn )]} forms a complete basis for x(t) in t ∈ (−∞, ∞),
given in eq. (2.17) and eq. (2.18). In [42], another sampling basis k(t, tn ) which
is the unique reciprocal basis of {g(t, tn )} = sinc[2B(t − tn )] was introduced. An
expression in terms of Kronecker delta function δ[m − n] is given by
hk(t, tm ), g(t, tn )i = δ[m − n], (2.19)
where
R∞ ha, bi denotes the inner product of a and b which is given by ha, bi =
−∞
a(t)b(t)dt. In t ∈ (−∞, ∞), {k(t, tn )} is also a complete basis-kernel for x(t).
2.4. RECONSTRUCTION ALGORITHMS (RAS) 23
Therefore, x(t) can be given either by

∞
X
x(t) = hx(•), k(•, tn )ig(t, tn )
n=−∞
X∞
= 2B cn g(t, tn ) (2.20)
n=−∞
in terms of cn or by
∞
X
x(t) = hx(•), g(•, tn )ik(t, tn )
n=−∞
X∞
= x(tn )k(t, tn ) (2.21)
n=−∞
in terms of the nonuniform samples x(tn ). It was also mentioned in [42] that this
method is appropriateR ∞in the case of L2 signals only. In other words, the CT function
x(t) has to satisfy −∞ |x(t)| dt < ∞ [43]. According to Parseval’s equation [35],
2
RB
−B
|X(f )|2 df < ∞. This CT function has a finite energy, and this method is
only suitable for a band-limited signal. However, the only complete orthonormal
sampling basis for χ are of the form {g(t, tn )} = {g(t, nTs )} (where χ is a subspace
of L2 -space in the time domain). Obviously, Higgins sampling theorem includes the
Shannon’s sampling theorem as a special case: For US tn = nTs ,
k(t, mTs ) = g(t, mTs )

= sinc[2B(t − mTs )]
hk(t, mTs ), g(t, nTs )i = sinc[2B(n − m)Ts ]
= δ[m − n].
A close form of the basis kernel (BK) k(t, tn ) is needed for the reconstruction of
NUS, and k(t, tn ) 6= g(t, tn ).
2.4 Reconstruction Algorithms (RAs)

A filter generalized by a certain RA is expected to reconstruct the signal as close as
possible to the original from the nonuniformly distributed samples. The selection
of the BK k(t, tn ) determines the reconstruction performance. The reconstruction
filter can be in either CT or DT. Based on eq. (2.21), a new sampling paradigm
with RAs is proposed as shown in Fig. 2.8.
Frequently used RAs for NUS include
• Low-pass Filtering (LPF) [interpolation],

x(t) x̂(t)
k(t, tn) a)
P∞
n=−∞ δ(t − tn)
x(t) x̂(nTs) Ts x̂(t)
k(nTs, tn) b)
−B B
P∞
n=−∞ δ(t − tn)
Θ(t, tn) P (tn → nTs ) c)
−B B
P∞
n=−∞ δ(t − tn)
Figure 2.8: Identity elements of (a) interpolation reconstruction with a CT filter;
(b) interpolation reconstruction with a DT filter; (c) iterative reconstruction [44]
• Lagrange Interpolating Polynomial [interpolation],
• Spline Interpolating [interpolation],
• Gridding Algorithm [interpolation],
• Least Square Reconstruction (LSR) Algorithm [svd],
• Iterative Algorithms [iterative],
• Yen’s Interpolations [interpolation],
• Coefficient cn Determination Reconstruction Algorithm [svd],
• “Minimum-Energy” Signals [svd].
These methods can be simply classified into three types: interpolation, iterative
and svd methods. The conventional FIR filter design with a constant data rate
is normally based on Interpolation. Iterative methods are extensively used in im-
age processing. They consist of three steps: orthogonal projection, iteration and
procedure convergence. svd is an important element of many numerical matrix
algorithms. If the matrix of eigenvectors of a given matrix is not a square matrix,
the matrix of eigenvectors has no matrix inverse, and the given matrix does not
have an eigen decomposition. The standard definition for the matrix inverse fails.
By svd, it is possible to obtain a pseudoinverse which is defined as
A−1 = (A∗ A)−1 A∗ = VDUT , (2.22)

where A = UDVT is a given m × n real matrix, U and V are m × m and n × n

unitary matrices (i.e., U∗ = U−1 , V∗ = V−1 ), D is a m×n diagonal matrix and the
elements in the diagonal consist of the singular values of A and zeros, {•}T denotes
a matrix transpose operator. All the RAs involving matrix inverse operations are
classified within the family of svd methods, e.g., LSR algorithm and coefficient cn
determination.
As we discussed in section 2.3, a signal can be reconstructed either in terms of
the sampled-data signal by a BK k(t, tn ) (see eq. (2.21)) or in terms of the coefficient
cn by a sinc function (see eq. (2.20)). Most of the RAs start from eq. (2.20) except
for the coefficient cn determination method.
Note that all these RAs are directly applicable for LPS but not for BPS. In
order to be used in BPS implementations, a carrier-modulated BK is needed (see
chapter 4). In this thesis work, the above nine RAs are studied and compared from
the aspects of reconstruction performance, computational complexity and hardware
implementation. The BKs of four among these nine which are based on interpolation
will be discussed and compared in both the time domain and the frequency domain.
Low-pass Filtering (LPF)

LPF technique directly treats nonuniform samples x(tn ) with the reconstruction
method to uniform sampling by using a lowpass filter. The algorithm can be written
as eq. (2.21) with
k(t, tn ) ≈ sinc[2B(t − tn )] (2.23)
Lagrange Interpolating Polynomial [42]

When nonuniform sampling instants deviate from the equivalent uniform Nyquist
sample instants by no more than Ts /4 (Ts is the equivalent US interval), k(t, tn )
can be approximated by the Lagrange interpolation function
P (t)
k(t, tn ) ≈ , (2.24)
P 0 (tn )(t − tn )
Y t
where P (t) = (t − t0 ) (1 −
), m ∈ (−∞, +∞)
tm
m6=0
Y tn t0 − t n Y tn
P 0 (tn ) = (1 − )+ (1 − ).
tm tn tm
m6=0 m6=0,m6=n
In eq. (2.24), k(t, tn ) is equal to zero at every sampling point except for the nth
where it is equal to one. If t0 = 0, eq. (2.24) can be simplified to
Y t − tm
k(t, tn ) ≈ . (2.25)
tn − tm
m6=n
Spline Interpolation
A spline function is a piecewise polynomial that has a simple form locally but is
flexible globally. Cubic spline is one kind of spline with a third-order polynomial
passing through a series of mesh points between any two fixed points. Assuming
that x(tn ) is the ordinate of tn , a cubic spline is given by
Sc (t) = an + bn t + cn t2 + dn t3 (2.26)
in t ∈ (tn−1 , tn ], where the 4n − 4 coefficients an , bn , cn and dn are determined by

the following conditions: i) Sc (tn ) = x(tn ); ii) the first and second derivatives of
Sc (t) are continuous at tn ; iii) the second derivative of Sc (t) is zero at the endpoints
(for ”natural” cubic spline) [45]. Therefore, the reconstructed signal when using a
cubic spline can be expressed as
∞
X
x̂(t) = Sc (t)[µ(t − tn−1 ) − µ(t − tn )], (2.27)
n=−∞
where µ(t) is a Heaviside’s step function, being 1 for t ≥ 0 and 0 otherwise. It is

difficult to write the spline interpolation in the form of eq. (2.21) in terms of x(tn )
and k(t, tn ) directly.
Gridding Reconstruction Algorithm [41]

Gridding reconstruction algorithm is suitable for a nonuniform sampling pattern
but with a uniform measurement space. A Density Compensation Factor (DCF) is
used, which is the same as in the adaptive weights method [46]. DCF is inversely
proportional to the local sampling density.
The samples are weighted by DCF and then convolved with a kernel. For one-
dimensional case, ωn = (tn+1 − tn−1 )/(2∆t) is a simple and rather good definition
for DCF, where ∆t is the average sampling interval. The algorithm can be written
as eq. (2.21) with
k(nTs , tn ) ≈ h(nTs , tn )ωn (2.28)
where nTs represents the uniform measurement space with an interval of Ts , h(t, tn )
is a kernel which can be either Gaussian, a sinc or some other small finite windows
[41]. As shown in Fig. 2.8 (b), the reconstruction filter comprising k(nTs , tn ) is a
DT filter. According to eq. (2.16), a reconstructed CT signal x̂(t) will be obtained
if the reconstructed DT signal x̂(nTs ) is filtered by an ideal lowpass filter.
Least Square Reconstruction (LSR) Algorithm

Assume that χ = [x̂(tm )] denotes an M × 1 vector of the reconstructed signal,
A = [h(tm , tn )] is an N × M sampled approximation matrix of a kernel {h(t, tn )}
and x = [x(tn )] is an N × 1 vector of sampled data. Then there exists a linear

expression between x and χ,
x = Aχ, (2.29)
which can be regarded as reconstructing the known samples x in terms of the
unknown signal χ. However, we can only find an approximate basis-kernel matrix
A and consequently introduce ² = |x − Aχ|. The least-square error is minimized
by solving for χ in
minχ ||x − Aχ||2 .
Here, x consists of the sampled data and χ = (A? A)−1 (A? )x (see also eq. (2.21)).
A? denotes the Hermitian transpose (or complex conjugate transpose) of the ma-
trix. This linear system should be overdetermined for getting a unique recon-
structed signal. It implies that the number of samples x(tn ) should be greater or
equal than the number of reconstructed points. For χ = [x̂(mTs )], where m ∈ [1, M ]
and M ≤ N , x̂(mTs ) can also be filtered by an ideal lowpass filter according to the
same rules as the gridding algorithm, and a CT signal reconstruction is achieved.
Iterative Methods [46] [47]

P∞
The iterative method first gets an initial reconstruction function n=−∞ x(tn )Θn
from {x(tn )} by using an indicator function Θn which could be sample-and-hold
interpolation function within [tn−1 , tn ]:
Θn (t) = µ(t − tn−1 ) − µ(t − tn ) (2.30)
or the midpoints of subsequent intervals, or some other interpolation function, and
then goes through projection, correction and accumulation iteratively, expressed by
the following equations
∞
X
x0 (t) = P{ x(tn )Θn } (2.31)
n=−∞
∞
X
∆1 (t) = P {x0 (t) − x0 (tn )Θn }, · · · ,
n=−∞
∞
X
∆i (t) = P {∆i−1 (t) − ∆i−1 (tn )Θn } (2.32)
n=−∞
and finally x̂(t) = x0 (t)+∆1 +· · ·+∆i (i = 1, 2, ..., ∞). P {•} denotes an orthogonal
projection which projects a given signal onto the space of band-limited signals in
the frequency range f ∈ [−B, B] (see Fig. 2.8 (c)). The projector could be a lowpass
filter or a convolution using a sinc kernel. It is obvious that the filtering process
destroys the pointwise interpolation property of the approximation procedure. The
number of iterative procedures i depends on the convergence rate of ∆i .
Yen’s Interpolation (THEOREM I in [48])

As mentioned in [49], the interpolation function for the reconstruction from the
samples by NUS should: i) be band-limited to [−B, B]; ii) take on correct values
at sampling instants. LPF based on the sinc kernel is band-limited, but it cannot
take on correct values at nonuniform sampling instants (see Fig. 2.13 Bottom). A
modification interpolation function based on LPF was proposed by Yen such that
both requirements for the reconstruction of NUS are attainable.
theorem I in [48] states that if a finite number of uniform sample points
in a uniform distribution are migrated to new distinct positions thus forming a
new distribution denoted by t = tn , the band-limited signal x(t) remains uniquely
defined. The reconstruction of the signal is obtained as
∞
X
x̂(t) = x(tn )Ψn (t), (2.33)
n=−∞
with the composing function Ψn (t) given by

QN QN
q=1 (t − tq ) q=1 [n/(2B) − nq /(2B)] (−1)n sin 2πBt
Ψn (t) = Ψn1 (t) = QN QN ·
q=1 [t − nq /(2B)] q=1 [n/(2B) − tq ] π(2Bt − n)
for tn = n/(2B) 6= nq /(2B), while

QN QN
q=1,6=p (t − tq ) q=1 [tp − nq /(2B)] sin 2πBt
Ψn (t) = Ψn2 (t) = QN QN ·
q=1 [t − nq /(2B)] q=1,6=p (tp − tq )
sin 2πBtq
for tn = tq .
Starting from eq. (2.21), Ψn (t) corresponds to k(t, tn ). These nonuniformly
sampled points consist of two sets: tn = nTs and tn = tq (tq /Ts is not an integer).
Then the reconstruction of x(t) can be written in another way:
∞
X X
N
x̂(t) = x(nTs )Ψn1 (t) + x(tq )Ψn2 (t). (2.34)
n=−∞,n6=nq q=1
Sinc kernel sinc[2B(t − tn )] has a zero at every t = mTs but n 6= m for tn = nTs .
However, it does notQvanish at tn = tq (see Fig. 2.13 Bottom). We must provide
zeros for each tq by q (t − tq ) in Ψn (t). In addition,
Q it is also necessary to remove
the zeros from the denominator factor of Ψn (t), q [t−nq /(2B)], when t = nq /(2B)
for which the samples are unknown.
Coefficient cn Determination Reconstruction Algorithm

As shown in eq. (2.20), a signal can also be perfectly reconstructed from nonuni-
formly sampled points by using sinc kernel provided that the coefficient cn is de-
termined. For a special case of US, eq. (2.20) is written as

∞
X
x̂(t) = 2B cm g(t, mTs ). (2.35)
m=−∞
Substituting the time instants tn of NUS and its corresponding sampled data x(tn )
by t and x(t) in eq. (2.35) respectively, there exists a linear expression:
X = CA, (2.36)
where X = [x(tn )] (n = 1, 2, · · · , N ) denotes an 1×N vector of nonuniform sampled

data, A = [g(tn , mTs )] is an M × N (M ≤ N ) sampled approximation matrix of
kernel g(t, mTs ) and mT represents a set of uniform sampling time instants which
satisfies the Nyquist criterion (1/Ts ≥ 2B), and C = [2Bcm ] represents a 1 × M
vector of coefficients. C can be easily obtained by XA−1 , where A−1 is obtained
by svd. By substituting C into [2Bcm ] in eq. (2.35), the signal reconstruction is
achieved.
“Minimum-Energy” Signals (THEOREM IV in [48])

As we have known, {sinc[2B(t−tn )]} form a complete basis for x(t) in t ∈ (−∞, ∞)
based on eq. (2.18). However, in most cases, the time variable t is not infinite but
limited over an interval To . When one does not wish to specify the time interval
explicitly, for critical sampling, 2BTo arbitrarily-located samples can be used to
define uniquely a “minimum-energy” signal according to Yen’s theorem IV.
theorem IV in [48] states that if the sampled values at a finite set of arbitrarily
distributed sample points t = tn , n = 1, 2, · · · , N are given, a signal x(t) with no
frequency component above R ∞ B is defined uniquely under the condition that the
“energy” of the signal −∞ |f (t)|2 dt is a minimum. Moreover, the reconstruction
of the signal is
X
N
x̂(t) = x(tn )Ψn (t), (2.37)
n=1
where
X
N
sin 2πB(t − tm )
Ψn (t) = amn . (2.38)
m=1
2πB(t − tm )
The coefficients amn are the coefficients of the inverse of a matrix whose elements
are
sin 2πB(tn − tm )
, n, m = 1, 2, · · · , N. (2.39)
2πB(tn − tm )
It is observed that amn has the same form as matrix A in eq. (2.36) with the only
exception that tm = mTs in matrix A. As shown in Fig. 2.9, both CT and DT
signal reconstruction can be achieved by algorithms based on svd depending on
the specification of the reconstruction filter.
x(t) x̂(t)
g(t, tn) ● a)
P∞
n=−∞ δ(t − tn) SVD

g(nT, tn) ● b)
−B B
P∞
n=−∞ δ(t − tn) SVD
Figure 2.9: Identity elements of (a) svd reconstruction with a CT filter; (b) svd
reconstruction with a DT filter.
2.5 Performance Evaluation of RAs

To evaluate the properties of these RAs, a sinusoidal signal with two periods is
sampled. Then the sampled data is reconstructed by using these RAs. The sample
distributions by US and NUS are shown in Fig. 2.10.
2.5.1 Reconstruction performance

A reconstruction error curve e(t) can characterize the reconstruction performance,
where
e(t) = x(t) − x̂(t), (2.40)
x(t) and x̂(t) are the original input signal and the reconstructed signal, respectively.
The error curves of nine RAs are evaluated for NUS and shown in Fig. 2.11 and
Fig. 2.12.
It is observed that for Lagrange interpolating polynomial, Spline interpolation,
LSR algorithm, Yen’s interpolation, cn determination and “Minimum-energy” sig-
nals, correct values are taken at the sampled points by these RAs. Lagrange Poly-
nomial Interpolating shows the best reconstruction performance. LPF has a good
reconstruction performance for US except at the ends due to the truncation error
of the sinc function [44]. The finite convolution of the sinc function will contribute
to sidelobes and consequently cause a poor reconstruction performance. The gen-
eral truncation error was defined by theorem 5 in [50]. Using the BK of LPF
defined by eq. (2.23), we obtain the expression for reconstructing one point x(tn )
by LPF [44]:
∞
X
x̂(tn ) = x(tn ) + x(tm )sinc[2B(tn − tm )], (2.41)
m=−∞,m6=n
2.5. PERFORMANCE EVALUATION OF RAS 31
1.5
0.5
−0.5
−1
−1.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t−−>
1.5
0.5
−0.5
−1
−1.5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t−−>
Figure 2.10: Sample distributions by US (Top) and NUS (Bottom), “+” shows the
sampling location and “◦” the sampled value.
where the second term represents the sidelobe effects of truncated sinc function. For
NUS, however, the sinc kernel cannot pick up correct values at the sampled points
such that the reconstruction performance by a LPF is degraded. The sinc kernel is
also used in Gridding algorithm and the corresponding reconstruction performance
is predetermined by the sinc kernel. It is observed that introducing DCF in the
Gridding algorithm does not improve the reconstruction performance significantly
in this simulation. In the iterative algorithm, a convolution by a sinc kernel is used
as a projector and the number of iterative procedures is 10. The reconstruction
performance of sampled points is improved to a certain degree by the iterative
algorithm compared to LPF.
The BK of RAs based on interpolation are shown in Fig. 2.13 and Fig. 2.14.
They are symmetric at the origin for US but asymmetric for NUS.
The SNDR is normally used to numerically evaluate the accuracy of reconstruc-
tion which is defined as [51]
PL
i=1 x2i
SN DR = PL , (2.42)
i=1 (xi − x̂i )2
LPF Lagrange interpolating polynomial

0.5 0.05
e(t)
e(t)
0 0
−0.5 −0.05
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
t−−> t−−>
Spline interpolation Gridding algorithm

0.05 0.5
e(t)
e(t)
0 0
−0.05 −0.5
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
t−−> t−−>
Figure 2.11: Reconstruction error curves (“◦” represents the reconstruction of sam-
pled point).
where i = [1, L] denotes the evaluated points, normally L > N , xi and x̂i represent
the points from the original and reconstructed signal, respectively. The SNDR in
dB is evaluated for the reconstruction performance of sampled and interpolated
points of NUS respectively by different RAs (see Table 2.1).
2.5.2 Computational complexity

The computational complexity of a reconstruction filter depends on the accuracy
requirement of the simulation model. For the interpolation reconstruction filter, the
approximation error between x(t) and x̂(t) can be decreased by increasing the length
of the filters or the degree of the interpolation. For the iterative reconstruction,
increasing the order of iteration is also helpful for reducing the error when the
repeating procedure is convergent. Based on the simulations that we have done
and the approximation errors shown in Fig. 2.11 and Fig. 2.12, the computational
LSR algorithm Iterative algorithm

0.05 0.5
e(t)
0 0
−0.05 −0.5
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
t−−> t−−>
Yens interpolation cn determination

0.05 0.05
e(t)
e(t)
0 0
−0.05 −0.05
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
t−−> t−−>
‘‘Minimum−energy’’ signals
0.05
e(t)
−0.05
0 0.2 0.4 0.6 0.8 1
t−−>
Figure 2.12: Reconstruction error curves (“◦” represents the reconstruction of sam-
pled point) (cont.).
1 US
0.5
−0.5
−1
LPF
spline
Lagrange
−1.5
−6 −4 −2 0 2 4 6
t−−>
1 NUS
0.5
−0.5
−1 LPF
Yen I
spline
Lagrange
−1.5
−6 −4 −2 0 2 4 6
t−−>
Figure 2.13: The basis-kernels of interpolations in time domain.

US
−20
−40
−60
Magnitude (dB)
−80
−100
−120
−140
sinc
−160 spline
Lagrange
−180
−5 −4 −3 −2 −1 0 1 2 3 4 5
f−−>
NUS
−20
−40
−60
Magnitude (dB)
−80
−100
−120
−140
Yen I
−160 spline
Lagrange
−180
−5 −4 −3 −2 −1 0 1 2 3 4 5
f−−>
Figure 2.14: The basis-kernels of interpolations in frequency domain.

Table 2.1: SNDR (in dB) comparison of different algorithms (N=34, L=201) for
the NUS pattern shown in Fig. 2.10
Algorithm Nonuniform sampling
Sampled points Interpolated points
Lowpass filtering (LPF) 17.33 17.97
Lagrange interpolating polynomial ∞ ∞†
†
Spline interpolation ∞ 69.22
Gridding algorithm 18.49 19.22
LSR algorithm ∞† 39.54
Iterative algorithm 18.63 13.89
†
Yen’s interpolation ∞ 40.22
cn determination ∞† 39.54
”Minimum-energy” signals ∞† 37.30
†
It is a reasonable assumption that SNDR is approximated by ∞ when SNDR> 100.
complexity is only determined by the different BK functions. Here all the RAs
were divided into two groups, sinc-based and nonsinc-based, and the number of
floating point operations (FPOs) was evaluated by using matlab 5.3 for each RA
(see Table 2.2). We find that spline interpolation is the most expensive technique.
Repeating the procedure many times causes a large number of FPOs for the iterative
algorithm.
2.6 Implementations of RAs
Those RAs (i.e., gridding and iterative algorithm) which are extensively used in
image processing cannot immediately be used for radio communications. One im-
portant difference between radio communication and image processing is that the
former requires data processing on-line but the latter does not. These RAs to-
gether with those based on svd (i.e., LSR algorithm, coefficient determination and
“minimum-energy” signals) have to be applied to blocks of data while the other
methods (i.e., Lagrange interpolating polynomial, spline interpolation, Yen’s inter-
polation) can be applied on a sample-by-sample basis.
For NUS, Lagrange interpolating polynomial has a rather good reconstruction
performance, the computation is also not very complex (see Table 2.2). However,
it is observed from eq. (2.25) that the input samples intercept the interpolating
filter impulse response at different time instants. This implies that it has a time-
varying characteristic. Currently, Lagrange fractional delay filtering [52] and time-
invariant filterbank [53] with synthesis filters generalized to Lagrange interpolating
polynomial are two feasible methods for implementation of Lagrange interpolating
polynomial.
2.6. IMPLEMENTATIONS OF RAS 37
Table 2.2: Computational complexity comparison of different algorithms by evalu-

ating the number of FPOs (floating point operations) [44]
Algorithm Sinc-based Nonsinc-based
LPF 55,141 –
Lagrange – 697,072
Spline – 8.088 × 109
Gridding 64,307 –
LSR – 1,131,518
Iterative 1,160,311 –
Yen’s – 1,376,353
“Minimum-energy” 610,676 –
Coefficient 996,532 –
2.6.3 Fractional delay filtering [54]

Discrete-time linear time-invariant (LTI) systems can be classified into FIR (finite
impulse response) and IIR (infinite impulse response) systems. For an M th-order
interpolation by FIR filtering, the output y(n) is given by
X
M
y(n) = h(m)x(n − m) (2.43)
m=0
in a direct form of convolution, where h(m) represents the impulse response coef-
ficients (or filter taps), and the set of {x(n − m)} represents the past M samples.
The corresponding transposed FIR structure is shown in Fig. 2.15 [37]. For a LTI
x(n)
···
hT (M ) hT (M − 1) hT (2) hT (1) hT (0)
+ z −1 + ··· + z −1 + z −1 + y(n)
Figure 2.15: Transposed FIR structure
system HT , the set of filter taps {hT (m)} which are the response of the system to
a series of unit pulse {δ[n − m]} are constants. The input delay will cause the same
time shift at the output, and normally this delay is an integer. The z transform of
hT (m) is given by
XM
HT (z) = hT (m)z −m . (2.44)
m=0
For a time-varying system H∆ , for instance, a filter generalized to Lagrange inter-

polating polynomial based on eq. (2.25), each coefficient h∆ (n) is not a constant
but a polynomial in terms of the delay parameter ∆ and
X
M Y
N
∆ − tp
h∆ (n) = cm (n)∆m = , n = 0, 1, 2, · · · , N. (2.45)
m=0
tn − tp
p=0,6=n
The z transform of h∆ (n) is given by

Ã !
X
N X
N X
M
H∆ (z) = h∆ (n)z −n = cm (n)∆m z −n
n=0 n=0 m=0
Ã !
X
M X
N X
M
= cm (n)z −n ∆m = Cm (z)∆m . (2.46)
m=0 n=0 m=0
As compared to eq. (2.44) and referred to Fig. 2.15, we can get the structure of
fractional delay filtering (see Fig. 2.16). For example if the output y(n) = x̂(n−∆),
then
XM
Cm (z)∆m = z −∆ , ∆ = 1, 2, · · · , M. (2.47)
m=0
and consequently the set of coefficients {Cm (z)} (or the N th order FIR transfer
function) can be obtained by solving M equations.
x(n)
···
CM (z) CM −1(z) C2(z) C1(z) C0(z)
+ ∆ + ··· + ∆ + ∆ + y(n)
Figure 2.16: Fractional delay filtering structure
2.6.4 Filterbank processing

Reconstruction from nonuniform samples is considerably more complex than re-
construction from uniform samples. Based on the Papoulis’ generalized sampling
theorem [33], a band-limited signal can be uniquely determined from uniformly
distributed samples of the outputs of N LTI systems with the signal as their input
sampled at one-N th of the Nyquist rate. Any set of nonuniform samples of a band-
limited signal can be divided into N groups such that the complete set of sampling
points can be expressed as
tp + nT, p = 0, 1, · · · , N − 1, n = 0, ±1, ±2, · · · (2.48)
2.6. IMPLEMENTATIONS OF RAS 39
where T = N/2B and B is the bandwidth of the band-limited input signal. For
the general case, there is only one sample in each group and N is very large.
As shown in Fig. 2.17, a CT band-limited signal x(t) can be reconstructed by
a CT reconstruction filterbank with the synthesis filters generalized to a suitable
reconstruction algorithm. An alternative implementation using a bank of DT filters
can be obtained by using an interpolation identity defined in [53].
xc(nT ) s0(t)
H0(Ω)
P∞
n=−∞ δ(t − tn)
xc(nT + t1) s1(t) xc(t)
H1(Ω) +
P∞
n=−∞ δ(t − nT − t1)
... ... ...
xc(nT + tN −1) sN −1(t)

HN −1(Ω)
P∞
n=−∞ δ(t − nT − tN −1)
Figure 2.17: Reconstruction from generalized uniform samples using a CT filter-

bank [53].
The above two methods of implementations are useful for all the RAs. For some
RAs, an approximation of the algorithm is necessary for generating a suitable filter.
Chapter 3
Uniform Bandpass Sampling
Signals can be categorized as lowpass versus bandpass in terms of the center fre-
quency. In the transmission of signal information over a communication channel, we
always encounter bandpass signals. The modulation at the transmitter generates
the bandpass signal, and the corresponding center frequency is not equal to zero.
The demodulation at the receiver recovers the information-bearing signal located at
DC (baseband) from the bandpass signal through frequency down conversion. With
respect to the bandwidth of the equivalent lowpass signal, carrier-modulated signals
can be classified into Single-SideBand (SSB) signals and Double-SideBand (DSB)
signals. A modulated bandpass signal with an SSB equivalent complex lowpass
signal can be expressed as
y(t) = Re{x(t)ej2πfc t } = a(t) cos(2πfc t) − b(t) sin(2πfc t), (3.1)
where fc is the carrier frequency, x(t) is the equivalent complex lowpass signal,
x(t) = a(t) + jb(t), a(t), b(t) are called the quadrature (I/Q) components of the
bandpass signal, and b(t) is the Hilbert transform of a(t) [55].
The modulated signals that satisfy the condition that their bandwidth is much
smaller than the carrier frequency are termed narrowband bandpass signals and
otherwise wideband bandpass signals. For a bandpass signal, it could be sampled
either by LowPass Sampling (LPS) process or BandPass Sampling (BPS). BPS
is a technique for undersampling a modulated signal to realize frequency down
conversion through intentional aliasing with the sampling rate of being down to
only twice the information bandwidth B (B << fc ), i.e., Fs ≥ 2B. LPS is based
on the Shannon’s sampling theorem, and fs ≥ 2fc + B (see Fig. 2.3). When
fc >> B, fs >> Fs . An example of sampled-data signal spectrum of LPS and BPS
is shown in Fig. 3.1. The randomly generated real band-limited signal is modulated
by a sinusoidal signal cos(2πfc t) (i.e., y(t) = a(t) cos(2πfc t)) and fc = 500. The
minimum sampling rate, fs = 2fc + B (B = 50) and Fs = 2B, are used for
LPS and BPS, respectively. Obviously, the output signal spectrum of LPS is the
periodic replica of the original modulated bandpass signal with period of fs , which
41
42 CHAPTER 3. UNIFORM BANDPASS SAMPLING
0.1
0.05
0 a)
−0.05
−0.1 t
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
0.015
−f /2 fs/2
s
0.01
b)
0.005
0 f
−2000 −1500 −1000 −500 0 500 1000 1500 2000
0.015
−F /2 F /2
s s
0.01
c)
0.005
0 f
−2000 −1500 −1000 −500 0 500 1000 1500 2000
Figure 3.1: a) Original modulated bandpass signal with a bandwidth B = 50 and

fc = 500; b) The corresponding frequency spectrum of sampled-data signal by LPS,
fs = 2fc + B; c) The corresponding frequency spectrum of sampled-data signal by
BPS, Fs = 2B.
is similar as shown in Fig. 2.3. The output spectrum of BPS is equivalent to a

periodic replica of a lowpass signal with a bandwidth of B in the period of Fs .
The sampled-data signal at the output is at baseband. The BPS technique shrinks
the Nyquist interval [−fs /2, fs /2] based on the first Nyquist criterion to the new
narrow interval [−Fs /2, Fs /2] and realizes a frequency down conversion at the same
time. With a goal of low power consumption, BPS is more and more attractive to
mixed-signal system design. It has been extensively studied in optics, radar, sonar,
communications and general instrumentation, etc.
The concept of software defined radio (SDR) has been paid more and more
attention for its support of multi-mode wideband radio communications. One key
technology of SDR is the placement and design technique of A/D converter in the
channel processing stream, and it is a goal to put the A/D converter as close as
possible to the antenna (see Fig. 1.5). By using conventional LPS, the sampling
rate would be too high to be achieved by current design technology. BPS may be
a solution for SDR by using a much lower sampling rate.
Besides the advantage of lower sampling rate, BPS has also limitations in real
3.1. SAMPLING RATE SELECTION 43
implementations. The BPS rate has to be carefully chosen in order to avoid harmful
signal spectrum aliasing. Noise aliasing is a direct consequence of lower sampling
rate as compared to the highest frequency component of the input bandpass signal.
The input signal frequency of a BPS is still high even though the sampling rate is
low. It was shown in [56] that the jitter effects depend on both the variance of the
random jitter and the input frequency such that the performance is degraded at the
output of BPS as compared to the equivalent LPS system, an ideal image-rejecting
mixer followed by an ideal lowpass sampler.
3.1 Sampling Rate Selection
The classic bandpass sampling theory states that for uniform sampling the sig-
nal can be reconstructed if the sampling rate is at least twice the information
bandwidth. Feldman & Bennett [57] and Kohlenberg [15] showed that for uniform
sampling, the minimum BPS rate is only valid for integer band position [16] [23],
where r = bfl /Bc = fl /B (see Fig. 1.6). The definition of band position has been
given in section 1.6.
For uniform BPS, the determination of Fs depends significantly on the band
position which represents how far away the information band is from DC. To min-
imize the transmission bandwidth, it is important to know the minimum sampling
rate for different band positions.
Assume that a DSB bandpass signal is located at [fl , fu ] ∪ [−fu , −fl ] as shown
in Fig. 3.2 with a fractional band position, i.e., bfl /Bc 6= fl /B, where fl and fu are
|Y (f )|
Fs
2B
−fu −fc −fl −Fs/2 0 Fs/2 fl fc fu f

x
Figure 3.2: An example of DSB bandpass signal with a fractional band position
located at [fl , fu ]∪[−fu , −fl ]. A BPS rate Fs is selected such that the sampled-data
signal is still a bandpass signal located at [−Fs , 0] ∪ [0, Fs ].
given by fl = fc − B, fu = fc + B. The minimum sampling rate corresponds to the

maximum number of foldings. For easily representing the signal spectral folding, we
introduce the folding triangle in the width of Fs (by dotted line in figures below).
For the positive frequency components, assume that the distance between fc and
the left boundary of the closest folding triangle is x. Then we obtain
(fc − x) − nFs = 0,
fc − x
Fs = (3.2)
n
It is also true that
½
B + x ≤ Fs
(3.3)
x≥B
Substituting eq. (3.3) in eq. (3.2), the acceptable minimum sampling rate for frac-
tional band position is obtained in the range of
fu fl
≤ Fs ≤ , (3.4)
n+1 n
where n is the maximum number of folding triangles in [0, fl ], n = bfl /2Bc. The
sampled data is still a bandpass signal located at [−Fs , 0] ∪ [0, Fs ]. A demodulation
is needed to get the equivalent lowpass signal.
The minimal Fs for directly getting the equivalent lowpas signal can be obtained
by the same way.
|Y (f )|
Fs
2B
−fu −fc −fl −Fs/2 0 Fs/2 fl fc fu f

Fs/2
Figure 3.3: An example of DSB bandpass signal with a fractional band position
located at [fl , fu ]∪[−fu , −fl ]. A BPS rate Fs is selected such that the sampled-data
signal is the equivalent lowpass signal located at DC.
Fs Fs
(fc − ) − nFs =
2 2
fc
Fsmin = (3.5)
n+1
where n = maxm∈Z + {fc /(m + 1) ≥ 2B}. As an example of a bandpass sig-
nal with B = 267 and fc = 5000, it has a fractional band position and n =
maxm∈Z + {5000/(m + 1) ≥ 534} = 8, then Fsmin ≈ 556 (see Fig. 3.4 a)). The
acceptable sampling rate for getting a bandpass signal in [−Fs , 0] ∪ [0, Fs ] is
3.1. SAMPLING RATE SELECTION 45
1
1
0.5 a)
0.5
0
0 f −200 −100 0 100 200 F /2=278
−6000 −4000 −2000 0 2000 4000 6000 −F /2 s
s
1
1
0.5 b)
0.5
0
0 f −Fs −500 0 500
−6000 −4000 −2000 0 2000 4000 6000 F =585
s
1
1
0.5 c)
0.5
0
0 f −500 0 500 F =592
−6000 −4000 −2000 0 2000 4000 6000 −Fs s
Figure 3.4: Examples of fractional band position and folding bands defined by
different Fs , a) Fs = 556; b) Fs = 585; c) Fs = 592.
585 < Fs < 592 and n = 8 (see Fig. 3.4 b) and c)). Eq. (3.5) is also useful
for a signal with an integer band position where Fsmin is equal to 2B.
The above discussion is based on a DSB bandpass signal. For an SSB bandpass
signal with fl = fc − B/2 and fu = fc + B/2 (see Fig. 3.5), the acceptable uniform
BPS rates have been obtained as [58] [16]
Fs
|Y (f )| B
−fu −fl −Fs/2 0 Fs/2 fl fu f

Figure 3.5: An example of SSB bandpass signal with a fractional band position
located at [fl , fu ] ∪ [−fu , −fl ]
2fu 2fl
≤ fs ≤ , (3.6)
n n−1
where n is the integer given by
fu
1≤n≤b c. (3.7)
B
The minimum acceptable sampling rate corresponds to n = bfu /Bc. The DSB
signal requests twice the channel bandwidth of the equivalent lowpass signal for
transmission. The transmission bandwidth of the SSB signal is only half of the
DSB signal.
For uniform BPS, the selection of acceptable sampling rate depends on the
band position. The conditions of acceptable uniform BPS rate for SSB signals was
depicted graphically by [57] and [16], where [57] only showed the minimum sampling
frequency and [16]’s was an extension of [57]’s for all the cases. Brown [59] pointed
out that for symmetric DSB signals, the spectra can be “folded over” each other
without loss of information. This is exactly the case shown in Fig. 3.4 a), and
the DSB of sampled-data signal is overlapped at the baseband. The corresponding
sampled rates falls over the dashed-line within the disallowed area in the Fig. 3.6.
However, it is difficult to adjust the BPS rate exactly to Fsmin . Any small sampling
rate variation will cause Fs move into the disallowed area such that an incorrect
folding of signal spectrum happens.
7
n=1
6
n=2
5
u
n=3
2f
fu
=
=
s
4
F
s
F 3)f u
Fs /B = (2/ n=4
F s
3
0
1 2 3 4 5 6 7
fu /B
Figure 3.6: The allowed and disallowed (shaded area) uniform sampling rates versus
the band position, Fs is BPS rate, B is the bandwidth, and the information band
is located at [fl , fu ] ∪ [−fu , −fl ] [16].
It is observed from Fig. 3.6 that the set of allowable BPS rates consists of n
disconnected segments within [2B, ∞). To do sampling efficiently, a lower sampling
3.2. NOISE SPECTRUM ALIASING 47
rate is more attractive. With the increase of fu /B (or increasing fc of information

band), n is increased and consequently the gap between any two segments in the
area of lower Fs becomes narrower and narrower. Even a small error in the sampling
rate might cause Fs to fall into a disallowed area. The efficient selection of sampling
rate becomes more and more difficult.
3.2 Noise Spectrum Aliasing

Additive noise is one common cause of performance degradation in telecommuni-
cation systems. Physically, the additive noise process may arise from electronic
components and amplifiers at the receiver of the communication system or from
the interference encountered in transmission. The noise at the input signal of BPS
can be assumed as an Additive White Gaussian Noise (AWGN), i.e., having a
delta-function autocorrelation with a flat power spectral density (PSD).
BPS technique can perform frequency down conversion by sampling a signal
at an IF stage and shift the frequency to a lower IF stage or baseband through
intentional signal spectral folding (also called downsampling mixer). However, the
resulting SNR of BPS will be lower than that of the equivalent LPS system in the
presence of thermal noise in sampling devices.
As shown in Fig. 3.7, the model of heterodyning a bandpass signal y(t) to
baseband in order to apply conventional LPS is shown in (a) as compared to the
equivalent BPS shown in (b), where y(t) is a bandpass signal and x(t) is the equiv-
alent lowpass signal of y(t). In the LPS system shown in (a), a lowpass filter
y(t) x(t) x(nT )

(a)
LPF LPS
cos(2πfct)
y(t) x(nT )
(b)
BPF BPS
Figure 3.7: (a) Heterodyning a bandpass signal y(t) to baseband in order to apply
conventional LPS; (b) The frequency down conversion and baseband sampling by
the equivalent BPS.
is used as an Anti-Aliasing (AA) filter prior to the lowpass sampler. Generally, a

BandPass Filter (BPF) is needed as an AA filter prior to the bandpass sampler in
the BPS system shown in (b). These AA filters can only reduce the out-of-band
noise prior to the sampler.
It is known that a resistor charging a capacitor gives rise to a total thermal
Vin Vout
Ron C
Figure 3.8: A switch-capacitor (SC) sampling device
noise with power kT /C [1], where k is Boltzmann constant, T is the absolute

temperature and C is the capacitance. The on-resistance of the switch will introduce
thermal noise at the output. The noise is stored on the capacitor along with the
instantaneous value of the input voltage when the switch turns off. As shown in
Fig. 3.8, the resistor Ron and sampling capacitor C is an LPF with a transfer
function of
1
H(f ) = . (3.8)
1 + j2πf Ron C
The PSD of thermal noise introduced by a resister can be given by Sin (f ) =
4kT Ron . The corresponding PSD of noise at the output of LPF is given by
1
Sout (f ) = Sin (f )|H(f )|2 = 2kT Ron (3.9)
1 + 4π 2 f 2 Ron
2 C2
by a two-sided representation, and the total noise power is obtained as

Z ∞
kT
Pout = Sout (f )df = . (3.10)
−∞ C
The effective noise bandwidth of the sampling device Bef f depends on the on-
resistance of the switch and the sampling capacitance, and it is normally larger
than the maximum frequency of the input signal. Besides the capacitor switching
noise (kT /C noise), op-amp wide-band noise and op-amp 1/f noise are two other
noise sources with minor weights in practical SC circuits [60]. To simplify the
following analysis, the dominant capacitor switching noise is regarded as the only
noise source in the sampling device.
Based on Fig. 3.7 (b), assume that a bandpass signal is first fed into an ideal
AA filter whose passband is located at [−f0 −B/2, −f0 +B/2]∪[f0 −B/2, f0 +B/2]
(f0 is the center frequency of the bandpass signal), and then sampled by critical
sampling (i.e., Fs = 2B). The sampled-data signal is located at baseband. For
the equivalent system (ES) as shown in Fig. 3.7 (a), all the noise at the output is
in-band, including the introduced thermal noise. Under the assumption that the
introduced thermal noise in a sampling device is an AWGN with zero-mean and

the PSD of the noise is a constant N0 , the corresponding SNR is given by
Ps Ps
SN RES = = , (3.11)
PNs + PNT h PNi
where Ps is the signal power, PNs is the input signal noise power after the AA filter,
PNT h is the introduced thermal noise power and PNT h = N0 · B, PNi denotes the
total in-band noise power which is the sum of PNs and PNT h . For BPS, the SNR
is given by
Ps
SN RBP S = , (3.12)
PNi + (M − 1)PNT h
where M = Bef f /B is the total number of Fs bands within [−Bef f , Bef f ], (M −
1)PNT h represents the total out-of-band thermal noise power. This SNR of BPS
is consistent with the result obtained by Vaughan (see eq. (63) in [16]). When
PNi >> PNT h , SNR degradation is only loosely dependent on the effects of noise
aliasing. However, when PNi ≈ PNT h ,
Ps
SN RBP S ≈ , (3.13)
M · PNi
and the SNR degradation in dB between BPS system and the equivalent LPS system
is expressed as
Bef f
SN Rdeg ≈ 10 log10 M = 10 log10 . (3.14)
B
For more general case when Fs > 2B,
Ps
SN RBP S ≈ , (3.15)
M· B
Fs /2 · PNi
and the SNR degradation in dB becomes

Bef f B 2Bef f
SN Rdeg ≈ 10 log10 · = 10 log10 . (3.16)
B Fs /2 Fs
Obviously, all out-of-band noise in BPS will be combined into each of the bands of
width Fs . The higher the BPS rate, the lower 2Bef f /Fs and hence the lower SNR
degradation.
It is known that an ideal uniform BPS is equivalent to an ideal uniform LPS
followed by a decimation operation [24] (see Fig. 3.9) provided that the BPS rate
Fs = 1/Ts ≥ 2B and the LPS rate M/Ts ≥ 2fc + B, where M is the decimation
factor.
The effects of noise aliasing can be graphically interpreted by the PSD spectrum.
As shown in Fig. 3.10, a BPS is replaced by a LPS followed by a decimation and
the noise aliasing in BPS is illustrated step by step. To avoid the noise aliasing
y(t) y(nTs) y(t) y(n · M

Ts
) y(nTs)
BPS LPS ↓M
x[n] yd[n]
Figure 3.9: Identity elements of ideal uniform BPS
LPS
X(f )
(a)
2kT R
−Bef f 0 B 2B ··· (M − 1)B Bef f f
↓M M=
Bef f
B
Yd(fN )
2M · 2kT R
11
00
00
11
M
00
11
... ... ... ... 00
11
... ... ...
··· ··· 00
11 ··· ... ... ... ... ···
00
11 (b)
00
112
00
11
00
111
−Bef f ··· 0 11
00B 2B ··· (M − 1)B Bef f f
Figure 3.10: Illustration of noise aliasing in BPS
in LPS, the LPS rate is larger than or equal to 2Bef f . The PSD of LPS from
−fs /2 to fs /2 is shown in Fig. 3.10 (a). Assume that the minimum sampling rate
2B is used for BPS and Bef f is an integer multiple M of B. By doing an M -fold
decimation on the output of LPS, the sampling rate will be reduced to the rate
of BPS. Decimation is one of the most basic operations in multirate digital signal
processing. It is also called decimator, downsampler or sampling rate compressor.
For the M -fold decimation, the expression of the output PSD Yd (fN ) in terms of
the input PSD X(fN ) is given by [61]
M −1
1 X
Yd (f /fs ) = X((f /fs − k)/M ), (3.17)
M
k=0
where
½
2kT R, −Bef f ≤ f ≤ Bef f
X(f ) = (3.18)
0, others,
and f /fs is the normalized frequency. It can be interpreted as three steps: (i)
stretch X(f /fs ) by a factor M to obtain X(f /(M · fs )), (ii) create M − 1 copies of
this stretched version by shifting it uniformly in successive amounts of 1, and (iii)

add all these shifted stretched versions to the unshifted stretched version X(f /fs ),
and divided by M . The final PSD spectrum of the M -fold decimation is shown in
Fig. 3.10 (b). The noise PSD from −Fs /2 to Fs /2 is increased by M such that the
output SNR by BPS is degraded as compared to the equivalent system shown in
Fig. 3.7 (a).
0
Ri evaluation band
−20
−40
PSD (in dB)
−60 Ro evaluation band
−80
a) BPS
−100 b) LPS + Decimator
−120
0 50 100 150 200 250
f (Hz)
−20
PSD (in dB)
−40
−60
−80
0 2000 4000 6000 8000 10000 12000
f (Hz)
Figure 3.11: Demonstration of noise aliasing. (Top): a) Decimated sampled-data

signal of LPS by factor 50 with a BPF, fs = 25000, SNR≈ 54.4 dB; b)BPS by
Fs = 500, SNR≈ 26.5 dB. (Bottom): LPS by fs = 25000, SNR≈ 41.5 dB [24].
The effects of noise aliasing can also be demonstrated by simulations using

the matlab psd function. A sinusoidal carrier signal with a carrier frequency of
5000 is modulated by a randomly generated band-limited signal with B = 50. A
band-limited AWGN is added as the introduced thermal noise and Bef f = 12500.
Assume that PNs = 0 such that PNi = PNT h . To see the effects of noise aliasing,
we choose Fs equal to 10B for uniform BPS (undersampling). As a reference, we
choose fs = 2Bef f for uniform LPS (oversampling). The sampled-data output
signal by BPS is directly available at baseband. The resulting PSD spectra for
these two cases are shown in Fig. 3.11. To avoid the effects of the transition band
of the AA filter, the passband and stopband frequencies are tactically selected as
shown in Fig. 3.11. The SNR is evaluated by
(Avg[Ri ] − Avg[Ro ]) · B
SN R = , (3.19)
Avg[Ro ] · fs /2
where Ri and Ro represents the in-band and out-of-band PSD, Avg[•] denotes
the average value in a given band of frequencies. The periodogram spectrum by
averaging 20 power spectra is shown in Fig. 3.11. The SNR of BPS and LPS are
about 26.8 dB and 42.5 dB, respectively (see Fig. (3.11)) and hence the degradation
of SNR is about 15.7 dB. By eq. (3.16), SNRdeg ≈ 17 dB. The difference of 1.3 dB
between the theoretical and simulated result is probably due to the transition band
of the AA filter, a forth order Butterworth BPF. Sampling rate can be converted
by either decimation (↓ fs ) or interpolation (↑ fs ). The sampling rate of LPS
fs = 2Bef f can be converted to 10B (the same as Fs of BPS) by decimation with
a factor of 50. The resulting PSD spectrum is exactly the same as that of BPS due
to the same noise aliasing (see Fig. 3.11 Top a)). If the sampled-data signal of LPS
is first fed into a BPF and then decimated, the out-of-band noise is suppressed by
the BPF and hence SNR is increased as compared to that of BPS. Note that when
using BPS, the out-of-band noise cannot be suppressed by a filter.
3.3 Jitter Effects

The intention of sampling systems is to obtain a sample value at the corresponding
time instant for an input signal. Based on sampling theorems, it is expected to
uniquely determine the input signal by the sampled data information. The effects
of random errors on the nominal sampling time instant are commonly called timing
jitter. As shown in Fig. 3.12, the random error τn which is a time offset from the
y(t)
y(tn + τn ) ●
ετ (n)
y(tn ) ●
τn
tn tn + τ n t
Figure 3.12: Illustration of jitter on time and amplitude
nominal time instant tn causes a random error ετ (n) in the amplitude. The effect
3.3. JITTER EFFECTS 53
of jitter on the spectrum of the signal may give rise to new discrete components
and produce frequency selective attenuation [34].
The noise power due to jitter is given by
Nτ (n) = E[ε2τ (n)] − E[ετ (n)]2 = E[ε2τ (n)] (3.20)
under the assumption that jitter noise has a zero-mean, where ετ (n) = y(tn +
τn ) − y(tn ). The approximate normalized average noise power in the time domain
is expressed as
1 X 2
K−1
N τ = E{ lim [ετ (n)]}. (3.21)
K→∞ K
n=0
For a sinusoidal signal y(t) = A sin(2πfin t) and 2πfin τn << 1, the error between
input and output of a sampling system is given by
dy(t)
ετ (n) ≈ τn = 2πfin τn A cos(2πfin tn ) (3.22)
dt
and the corresponding average noise power is approximately given by
Nτ ≈ 2π 2 fin
2 2 2
στ A , (3.23)
where στ2 = E[τn2 ]−E[τn ]2 . However, when the jitter is larger such that 2πfin τn <<
1 is not satisfied, the average noise power becomes [25]
2 2
στ2
Nτ = A2 (1 − e−2π fin
). (3.24)
Note that N τ is independent of the sampling sequence tn but depends on fin . The
higher the value of fin , the more noise power N τ and hence the larger jitter effects.
Under the assumption of 2πfin στ << 1, eq. (3.24) reduces to eq. (3.23). Eq. (3.23)
applies to all jitter distributions while eq. (3.24) assumes a Gaussian distributed
jitter.
To study the jitter effects, a random jitter with Gaussian distribution N (0, στ )
is applied to the real sinusoidal signal y(t) = sin(2πfin t) with fin = 10 and 500,
respectively, where στ is the standard deviation of jitter and στ = αTs (α is a scale
factor, α = [1.15×10−3 , 12×10−3 ]). The LPS rate is 5fin . The theoretical Signal-to-
Noise-and-Distortion Ratio SNDRt and simulated SNDR [44] are calculated using
½
A2 /2 1/(4π 2 fin
2 2
στ ), fin = 10
SN DRt = = −2π 2 fin
2
στ2 (3.25)
Nτ 1/[2(1 − e )], fin = 500
and PL
x2
SN DR = PL i=1 i . (3.26)
i=1 (xi − x̂i )
2
SNDR is normally used to measure the signal reconstruction error where xi and x̂i
denote the points from the original and reconstructed signal, respectively, i = [1, L]
is the index of evaluated points and normally L > N (N is the number of sampled
points). In this simulation, xi and x̂i are used to represent the sampled points
without jitter and with jitter respectively, and L = N .
It is observed from Fig. 3.13 that the theoretical SNDR is in agreement with
the simulation result by LPS very well for both fin = 10 and 500. However, when
the real sinusoidal signal with fin = 10 is frequency-shifted to fc = 500 and then
sampled by BPS with Fs = 50, the corresponding simulated SNDR is lower than
that of the equivalent LPS system (see Fig. 3.13 Left). With the increase of στ
from 0.23 × 10−4 to 2.4 × 10−4 , the SNDR difference varies from 13.5 dB to 27.6
dB. This difference is only due to the large jitter in BPS which is different from the
SNRdeg due to noise aliasing discussed in section 3.2. Jitter effects depend on both
the standard deviation of random jitter and the input frequency (see eq. (3.25)).
With the increase of input frequency by using BPS, jitter becomes a more crucial
problem than in the equivalent LPS system.
60 60
LPS LPS
BPS Theoretical
50 Theoretical 50
40 40
SNDR (in dB)
SNDR (in dB)
30 30
20 20
10 10
0 0
0.5 1 1.5 2 2.5 1 2 3 4 5
σ −4 σ −6
τ x 10 τ x 10
Figure 3.13: Comparison of theoretical and simulated SNDR for y(t) = sin(2πfin t).
Left: for LPS, fin = 10 and fs = 5fin = 50; for BPS, fc = 500 and Fs = 50. Right:
fin = 500, fs = 5fin = 2500.
Additionally, jitter effects for a general input signal was also discussed in [25].
Time skewing problem in A/D converter system which is very similar to the jitter
problem was also analyzed and compared in [25].
Chapter 4
Quadrature Bandpass Sampling
Shannon (1949) mentioned in [36] that any function limited to the bandwidth B
and the time interval T can be specified by giving 2BT samples. These samples
are unnecessarily evenly spaced, and the samples from the signal and its derivative
at half the Nyquist rate at least can also uniquely determine the signal without
loss of information. Later Papoulis (1977) established the generalized nonuniform
sampling theorem [33] which is an expansion of classic Shannon’s sampling theorem.
It states that a band-limited signal is uniquely determined by the samples on the
outputs of M linear systems with input of the signal at one-M th of the Nyquist
rate at least for each. The Papoulis’ generalization of sampling theorem treats
extensively the representation of the signal from (i) the samples of the signal and
its derivatives, (ii) Recurrent nonuniform sampling [48], (iii) the samples of the
signal and its Hilbert transform (e.g. quadrature sampling) [59], and some other
functions.
In digital communications, the modulated signal is always expressed in terms
of I/Q formats or in quadrature. The main advantage of I/Q modulation is the
symmetric case of combining independent signal components into a single composite
signal and later splitting such a composite signal into its independent component
parts [62]. It is more attractive to use quadrature mixers or quadrature BPS to
separate the signal to I and Q parts before baseband.
4.1 Generalized Nonuniform Sampling
The generalized sampling expansion was first introduced in [33]. As shown in

Fig. 4.1, given M linear systems with transfer functions of {Hk (ω)}, k = 1, 2, · · · , M ,
the responses of the linear systems to an input band-limited signal f (t) is given by
Z ω0
gk (t) = F (ω)Hk (ω)dω, (4.1)
−ω0
55
56 CHAPTER 4. QUADRATURE BANDPASS SAMPLING
where ω0 = 2πB (B is the bandwidth of the signal), and F (ω) is the Fourier
transform of f (t). Each of the M responses is sampled at least in one-M th Nyquist
rate. Define M Linear Time-Invariant (LTI) functions {yk (t)} such that the input
signal f (t) can be obtained at the output in terms of the samples {gk (nT )} and the
LTI functions {yk (t)}:
∞
X
f (t) = [g1 (nT )y1 (t−nT )+g2 (nT )y2 (t−nT )+· · ·+gM (nT )yM (t−nT )]. (4.2)
n=−∞
where T = 1/fs and {yk (t)} is given by

Z −ω0 +∆ω
1
yk (t) = Yk (ω, t)ejωt dω,
∆ω −ω0
and ∆ω = 2ω0 /M , T = 2π/∆ω, M unknown functions {Yk (ω, t)} are determined
by M linear expressions:
H1 (ω)Y1 (ω, t)+ · · · +HM (ω)YM (ω, t) = 1;

H1 (ω + ∆ω)Y1 (ω, t)+ · · · +HM (ω + ∆ω)YM (ω, t) = ej∆ωt
··· ··· ···
H1 [ω + (M − 1)∆ω]Y1 (ω, t)+ ··· +HM [ω + (M − 1)∆ω]YM (ω, t) = ej(M −1)∆ωt .
(4.3)
Eq. 4.3 can be easily written in matrix form.
g1 (t)
H1 (ω) Y1 (ω, t)
P
n δ(t − nT )
f (t) g2 (t) fˆ(t)
H2 (ω) Y2 (ω, t)
P
n δ(t − nT )
.. .. ..
. . .
gM (t)
HM (ω) YM (ω, t)
P
n δ(t − nT )
Figure 4.1: Identity of signal representation by Papoulis’ generalized sampling the-

orem.
4.1. GENERALIZED NONUNIFORM SAMPLING 57
4.1.1 Example – Derivative sampling

Starting from the generalized sampling theorem, if
Hk (ω) = (jω)k−1 ,
Z ω0
gk (t) = F (ω)(jω)k−1 dω
−ω0
= f <k−1> (t) (4.4)
based on the property of derivative of Fourier transform. The responses of the
linear system Hk (ω) are derivatives of input signal. Starting from eq. (4.3), the M
linear expressions for determining the M unknowns {Yk (ω, t)} can be expressed in
matrix form as
2 32 3 2 3
1 jω ··· (jω)M −1 Y1 (ω, t) 1
6 1 j(ω + ∆ω) ··· [j(ω + ∆ω)]M −1 76 Y2 (ω, t) 7 6 e j∆ωt 7
6 76 7 6 7
6 .. .. .. .. 7·6 .. 7=6 .. 7
4 . . . . 54 . 5 4 . 5
1 j[ω + (M − 1)∆ω] ··· {j[ω + (M − 1)∆ω]}M −1 YM (ω, t) ej(M −1)∆ωt
(4.5)
It can be solved in a closed form, using Cramer’s rule and the Vandermonde
determinant [63].
Generalized sampling with M branches is also called M th-order sampling. It
could be either US or NUS depending on the distribution of sampling time instants
from all the branches. For a special case of M = 1, the generalized sampling
theorem is reduced to the classic Shannon’s sampling theorem.
f (t)
1 Y1 (ω.t)
P
n δ(t − nT )
f (t) f˙(t) fˆ(t)
jω Y2 (ω, t)
P
n δ(t − nT )
.. .. ..
. . .
f <k−1> (t)
(jω)k−1 YM (ω, t)
P
n δ(t − nT )
Figure 4.2: Identity of signal representation of derivative sampling, where f <k−1>

denotes the k − 1 order derivative of f (t).
f (t)
1 Y1 (ω, t)
P
n δ(t − nT )
f (t) f (t + α1 ) fˆ(t)
ejα1 ω Y2 (ω, t)
P
n δ(t − nT )
.. .. ..
. . .
f (t + αM −1 )
ejαM −1 ω YM (ω, t)
P
n δ(t − nT )
Figure 4.3: Identity of signal representation of recurrent nonuniform sampling.
f (t)
1 Y1 (ω, t)
f (t) fˆ(t)
P
n δ(t − nT )
f˜(t)
Htr (ω) Y2 (ω, t)
P
n δ(t − nT )
Figure 4.4: Identity of signal representation of quadrature sampling, where f˜(t)

represents the Hilbert transform of f (t).
4.1.2 Example – Recurrent nonuniform sampling
Starting from the generalized sampling theorem and Fig. 4.1, suppose that the
sampling time instant of one branch lags behind the previous one by αk and |αk | <
T /2, then we have
gk (t) = f (t + αk )
and
Hk (ω) = ejαk ω .
4.1. GENERALIZED NONUNIFORM SAMPLING 59
Such sampling is also referred to as bunched or interlaced sampling [63]. The M

unknowns Yk (ω, t) are given by
    
1 ejα1 ω ··· ejαM ω Y1 (ω, t) 1
 1 ejα2 (ω+∆ω) ··· ejαM (ω+∆ω)   Y2 (ω, t)   ej∆ωt 
    
 .. .. .. .. · .. = .. 
 . . . .  .   . 
1 ejα2 [ω+(M −1)∆ω] · · · ejαM [ω+(M −1)∆ω] YM (ω, t) ej(M −1)∆ωt
(4.6)
By using Vandermonde determinant, the closed form of the reconstructing func-
tion yk (t) is given by [eq. (4.47) in [63]]
· ¸ YN
2B N (t − αl )]
sin[ 2πB
yk (t) = sinc (t − αk ) . (4.7)
N (αk − αl )]
N 2πB
l=1,l6=k
sin[
Yen also showed an expression of reconstruction function for recurrent£ nonuniform¤

N (t − αk )
sampling which is similar to eq. (4.7) with only the replacement of sinc 2B
(−1)mN
by 2πB mN (−∞ < m < ∞ is the index of the sampled points) (eq. (9)
N (t−αk − 2B )
in [48]).
4.1.3 Example – Quadrature sampling

A quadrature local oscillators (LOs) is normally used to split an input IF signal
into I and Q parts and then fed into a quadrature sampler at a half-Nyquist rate
(M=2). The quadrature LO realizes both frequency down-conversion and phase
shift by 90◦ . As shown in Fig. 4.4, the filter Htr (ω) is called Hilbert transformer.
The frequency response of this filter is given by [55]

 −j2π, ω > 0
Htr (ω) = 0, ω = 0, (4.8)

j2π, ω<0
This filter basically realizes a 90◦ phase shift for all frequencies in the input sig-
nal. The response of Hk (ω) = {1, Htr (ω)} for an input signal has the same phase
response as the output of a quadrature LO [1].
Starting from eq. (4.3), the corresponding reconstruction function for represent-
ing a signal by quadrature sampling is given by
· ¸ · ¸ · ¸
1 Htr (ω) Y1 (ω, t) 1
· = (4.9)
1 Htr (ω + ωo ) Y2 (ω, t) ejω0 t
in matrix form, where ω0 = 2πB. Hence,
ejω0 t − 1
Y1 (ω, t) = ,
Htr (ω + ω0 ) − Htr (ω)
ejω0 t − 1
Y2 (ω, t) = 1 − Htr (ω) · (4.10)
Htr (ω + ω0 ) − Htr (ω).
x(t) xs (t)
y(t)
x̂(t)
P
n δ(t − nT )
x(t + α) xs (t + α)
y(−t)
P
n δ(t − nT )
Figure 4.5: Model of second-order BPS based on the Kohlenberg’s sampling the-
orem, where x(t) = Re{[f (t) + j f˜(t)]ej2πfc t } is the bandpass signal with an SSB
equivalent lowpass complex signal, f (t) is the real signal and f˜(t) represents the
Hilbert transform of f (t).
Quadrature sampling is a special case of second-order sampling.
4.2 Quadrature Bandpass Sampling

As we discussed in section 3.1 that for uniform BPS, a minimum sampling rate 2B
is only valid for a bandpass signal with an integer band position. Kohlenberg [15]
showed that the minimum sampling rate in the form of an average can be applied
for a bandpass signal independent of the band position by a second-order bandpass
sampling. It was also stated that an SSB bandpass signal x(t) located at (fl , fu ) ∪
(−fu , −fl ) can be exactly represented by
X
N −1
x(t) = [x(nT )y(t − nT ) + x(nT + α)y(nT + α − t)], (4.11)
n=0
where T = 1/B for the minimum BPS rate, y(t) is given by
cos[2πfu t − (r + 1)πBα] − cos[2π(rB − fl )t − (r + 1)πBα]

y(t) =
2πBt sin[(r + 1)πBα]
cos[2π(rB − fl )t − rπBα] − cos[2πBt − rπBα]
+ , (4.12)
2πBt sin(rπBα)
α is the time lag between two sets of samples from the first and the second branches
and it is arbitrarily selected except for the values which would make y(t) infinite,
r is in the range of [2fl /B, 2fl /B + 1) (r = 2fl /B for integer band position and
r = 2fl /B+1/2 for half integer band position). The model of second-order sampling
for a bandpass signal based on the Kohlenberg’s sampling theorem is shown in
Fig. 4.5. For each branch, the sampling rate is half of the equivalent BPS rate.
Assume that x(t) is an SSB bandpass signal and x(t) = Re{x(t)ej2πfc t } =
f (t) cos(2πfc t) − f˜(t) sin(2πfc t), where f˜(t) is the Hilbert transform of f (t). The
4.2. QUADRATURE BANDPASS SAMPLING 61
j f˜(t)
0 f (t)
Figure 4.6: Illustration of phase shift by quadrature sampling: initial coordinate

A = [f (nT ) cos(2πfc nT ), −f˜(nT ) sin(2πfc t)], destination coordinate B = [f (nT +
α) sin(2πfc nT ), f˜(nT + α) cos(2πfc nT )], destination coordinate C = [−f (nT +
α) sin(2πfc nT ), −f˜(nT + α) cos(2πfc nT )].
input signals of samplers x(t) and x(t + α) with an arbitrary time lag α are sampled
at the same rate 1/T . When α = 1/(4fc ) + m/(2fc )(m = 0, ±1, ±2, · · · ) [16], the
corresponding sampled-data signals are given by
∞
X
xs (t) = x(t) δ(t − nT )
n=−∞
∞ h
X i
= f (nT ) cos(2πfc nT ) − f˜(nT ) sin(2πfc nT )
n=−∞
∞
X
xs (t + α) = x(t + α) δ(t − nT )
n=−∞
∞
X 1 m
= [f (nT + α) cos(2πfc nT + 2πfc · + 2πfc · )
n=−∞
4fc 2fc
1 m
−f˜(nT + α) sin(2πfc nT + 2πfc · + 2πfc · )]
4fc 2fc
∞
X
= (−1)m+1 [f (nT + α) sin(2πfc nT ) + f˜(nT + α) cos(2πfc nT )]
n=−∞
(4.13)
and there is a 90◦ phase shift between xs (t) and xs (t+α) (see Fig. 4.6). Therefore,
this second-order sampling becomes the so-called quadrature BPS. For a special case
when α = T /2, the second-order BPS becomes the conventional uniform BPS. The
sampled-data signal at the output of each sampler is called I (in-phase) and Q
(quadrature) component, respectively.
Without loss of generality, we assume that α = 1/(4fc ). It has been verified
by simulations that two parallel uniform samples with a time shift of 1/(4fc ) are
distinguished by the interpolation function y(t) (see eq. (4.12)) such that the input
signal can be reconstructed by these samples at the output. A band-limited SSB
signal x(t) is randomly generated as shown in Fig. 4.7, where fc = 100, (fl , fu ) =
(95, 105) with a half integer band position, B = 10, α = 1/4fc = 0.0025. It is
sampled at 1/T = B. The corresponding reconstructed result by eq. (4.11) and
eq. (4.12) is shown in Fig. 4.7 (Center) and it is consistent with the original signal
very well except the ends.
Besides the exact interpolation developed by Kohlenberg for a high-frequency
band-limited function (see eq. (4.12)), Ries [64] also suggested a form of general
reconstruction function derived from a lowpass reconstruction kernel. Shannon’s
lowpass sampling theorem shows that a lowpass band-limited signal can be exactly
reconstructed from its uniform samples by a sinc kernel. An alternative way to
represent the bandpass signal by the samples is to use a carrier-modulated sinc
function based on theorem 4.2 in [64]:
∞
X
x(t) = x(tn )s(t − tn )
n=−∞
where
s(t) = Re{sinc(2Bt)ej2πfc t } (4.14)
and the set of {tn } consists of the samples from both I and Q branches. The corre-
sponding reconstructed result is shown in Fig. 4.7 (Bottom). In general, eq. (4.14)
could be extended to
s(t) = Re{k(t)e2πfc t }, (4.15)
for any BK discussed in Chapter 2 provided that the expression of k(t) could be
found.
It is observed that the reconstructed signal by eq. (4.11) is obtained at the orig-
inal band position (fl , fu ). The frequency is not down-converted by BPS. An extra
resampler is needed to digitize the reconstructed results before the A/D converter.
By conventional LPS technique, quadrature lowpass signals are obtained by using a
pair of analog multipliers or mixers prior to the sampler, for instance in homodyne
architecture (see Fig. 1.2). The main advantage of quadrature BPS compared to the
homodyne architecture is that the effect of DC-offset that occurs with quadrature
mixers is removed.
4.3 Implementation of Quadrature BPS

Quadrature BPS as a special case of second-order BPS has been extensively stud-
ied. Rice and Wu [27] applied a single uniform sampling on a real bandpass analog
4.3. IMPLEMENTATION OF QUADRATURE BPS 63
input signal and then obtained the equivalent low-pass quadrature components by
a digital Hilbert transform and frequency translation. The corresponding block
diagram is shown in Fig. 4.8. The innovations of this architecture include that (i)
only a single sampling device is used for both I and Q branches instead of conven-
tional double sampling devices; (ii) the quadrature component is computed via a
digital Hilbert transform, which eliminates many problems of analog methods, such
as temperature sensitivity and drifts in component values; (iii) the decimation by 2
on the samples is applied before the Hilbert transform by separating even samples
and odd samples by a commutator. However, the bandpass signal is uniformly sam-
pled such that a thorough receiver frequency plan is needed for avoiding harmful
signal spectrum aliasing. The input bandpass analog signal is first frequency shifted
by BPS to a lower IF, and then frequency translated to baseband by multiplying
e−jπn . It can be also done by decimating by 2 and modulating by (−1)n which is
used in Pellon’s architecture [28].
Pellon also proposed an architecture of quadrature sampling with a double
Nyquist Digital Product Detector (DPD) as shown in Fig. 4.9. The primary advan-
tage of this architecture is the digital separation of the I and Q components by DPD
such that the mismatch of I and Q parts from the analog devices can be avoided.
The other property is the double Nyquist property. The input signal of DPD can
be digitized into its I and Q components at the Nyquist rate. The sampling rate
in an A/D converter is also the operation rate of digital filters. However, an extra
mixer is needed to frequency down-convert to fs1 /4, where fs1 is the sampling rate
in the single A/D converter.
As a specific effect in BPS system, the noise aliasing cannot be avoided in a
uniform BPS system. As discussed in chapter 2, NUS has the potential to sup-
press harmful signal spectrum aliasing by using a lower sampling rate. Consider-
ing the advantages and drawbacks of these two architectures, making use of NUS
and RAs, a new Generalized Quadrature BPS (GQBPS) algorithm is proposed
for suppressing the noise aliasing. As shown in Fig. 4.10, a real IF bandpass signal
x(t) = Re{a(t)e2πfc t } with an arbitrary band position is sampled by a second-order
bandpass sampling, where
a(t) = i(t) + jq(t), (4.16)
i(t) and q(t) represents the I and Q components of the equivalent lowpass complex
signal a(t). The uniform sampling period for each branch is Ts and Ts ≤ 1/B such
that the equivalent BPS rate is greater than or equal to 2B. The samples from the
second sampling branch lags behind those from the first by α. A carrier-modulated
sinc function s(t) defined by eq. (4.14) is expected to obtain the reconstructed
signal x̂(t) at the same band position as the input. It is well-known that the
sinc function performs a lowpass filtering such that s(t) has a property of BPF.
This eliminates the need for a extra BPF to remove the out-of-band noise and
unwanted image bands. A resampler is used to obtain the sampled-data signal of
quadrature components î(mTs0 ) and q̂(mTs0 ) from x̂(t). Finally, î(mTs0 ) and q̂(mTs0 )
are quantized by corresponding A/D converters. The combined system consisting
of s(t) and the resampler can be realized digitally:

∞
X
c(t) = [s(mTs0 )δ(t − mTs0 ) + s(mTs0 + α)δ(t − mTs0 − α)], (4.17)
m=−∞
where 1/Ts0 = fc /N is a new data rate, N is a decimation factor and N ≤ fc /(2B)

based on the Nyquist criterion. The resampler can also be built in the following
A/D converters sampled at {mTs0 } and {mTs0 + α}, respectively.
4.3.1 Frequency domain analysis

In a previous study, it was found that the samples by NUS can determine the input
signal uniquely. By using GQBPS algorithm, the input high frequency signal is
unambiguously determined by the undersampled data and the selection of sampling
rate is simple. It is also observed that the noise aliasing by GQBPS is unequally
weighted and the noise gain due to aliasing is lower than by conventional BPS
at some specific frequency bands. In the following subsections, the sampling rate
selection, signal reconstruction and noise aliasing suppression are analyzed and
illustrated in the frequency domain.
Deterministic Input Signal

It is known that the carrier-modulated bandpass signal can be represented either as
a double sideband (DSB) signal or a single sideband (SSB) signal depending on the
definition of the equivalent lowpass signal a(t) [55]. A DSB signal requires twice
the channel bandwidth of the equivalent lowpass signal for transmission. For saving
transmission bandwidth, SSB signal is generally used in radio communications, and
a(t) is defined as eq. (4.16). The Fourier transforms of the equivalent complex
lowpass signal a(t) and its complex conjugate a∗ (t) are shown in Fig. 4.11, where
I(f ) and Q(f ) is the Fourier transform of i(t) and q(t), respectively. The spectrum
of the corresponding bandpass signal x(t) is illustrated in Fig. 4.12 a).
Ideal quadrature sampling is equal to the continuous-time (CT) input signal
multiplied by two infinite sequences of Dirac delta functions where the second se-
quence lags behind the first by α:
" ∞ ∞
#
X X
xs (t) = x(t) δ(t − nTs ) + δ(t − nTs − α) . (4.18)
n=−∞ n=−∞
The corresponding Fourier transform of xs (t) is

∞
1 X
Xs (f ) = (1 + e−j2πkfs α )X(f − kfs ), (4.19)
Ts
k=−∞
where X(f ) is the Fourier transform of x(t) and

1 ∗
X(f ) = [A (f + fc ) + A(f − fc )] , (4.20)
2
Then Xs (f ) is given by
Xs (f ) = Al (f ) + Ar (f ), (4.21)
where
∞
1 X
Al (f ) = (1 + e−j2πkfs α )A∗ (f + fc − kfs )
2Ts
k=−∞
X∞
1
Ar (f ) = (1 + e−j2πkfs α )A(f − fc − kfs )
2Ts
k=−∞
as shown in Fig. 4.12 b) and c). The carrier modulated sinc function is a rectangular
function centered at ±fc (see Fig.4.12 d)). A convolution between s(t) and xs (t)
in the time domain is equivalent to a multiplication between S(f ) and Xs (f ) in
frequency domain, where
Z ∞
S(f ) = s(t)e−j2πf t dt
−∞
· µ ¶ µ ¶¸
1 f − fc f + fc
= rect + rect (4.22)
2fs fs fs
and
½
1, |f | < 1/2
rect(f ) = (4.23)
0, otherwise.
It is observed that the spectrum of Al (f ) and Ar (f ) is the periodic replica of
A∗ (f + fc ) and A(f − fc ) in the period of fs , respectively. Two criteria for sampling
rate selection which are independent of the band position of x(t) are used such that
x(t) can be reconstructed from the two sets of samples {x(nTs )}, {x(nTs + α)} by
s(t) in the proposed algorithm:
1. To avoid overlap between the adjacent folding spectra within the set of Al (f )
and Ar (f ), the sampling rate has to satisfy fs ≥ B.
2. To avoid overlap between the set of Al (f ) and Ar (f ), the ratio of fc to fs
should be an integer or a half integer, and fs ≥ 2B.
In the analyses below, it is assumed that the ratio fc to fs is an integer, i.e.,
fs = fc /i, i = 1, 2, · · · . It is seen from eq. (4.19) that there is a phase shift due to
time-lag α. Without loss of generality, it is assumed that α = 1/(4fc ):
fc
)( 4f1c )
1 + e−j2πkfs α = 1 + e−j2πk( i = 1 + e−jπk/(2i) . (4.24)
When k/i is even, e−jπk/(2i) = ±1 such that
½
2, k/i = 4l,
1 + e−j2πkfs /(4fc ) = l = 0, 1, 2, · · · (4.25)
0, k/i = 4l + 2,
The frequency spectra analysis is shown in Fig. 4.12. We expect that all other
frequency bands are filtered out by s(t) except for two located at [−fc − fs /2, −fc +
fs /2] and [fc −fs /2, fc +fs /2]. To obtain the reconstruction with s(t), it is expected
that the spectra located at [−fc − fs /2, −fc ] and [fc , fc + fs /2] are the same as
the CT input signal spectra multiplied by a gain factor while the spectra located
at [−fc , −fc + fs /2] and [fc − fs /2, fc ] are zero. It is always the case as long as
the above two criteria are used. For the spectrum located at [−fc , −fc + fs /2]
and [fc − fs /2, fc ] which is the copy of A(f − fc ) and A∗ (f + fc ) respectively by
k = 2fc /fs = 2i foldings, they are always zero, based on the second condition in
eq. (4.25) for l = 0. However, the spectrum at [−fc − fs /2, −fc ] and [fc , fc + fs /2]
is just the copy of A∗ (f + fc ) and A(f − fc ) with zero folding (k = 0). Based on the
first condition in eq. (4.25) for l = 0, the weight factor 1 +e−j2πkfs /α in eq. (4.19) is
equal to 2 and the gain factor 2fs in Xs (f ) will be balanced by S(f ) (see eq (4.22))
such that the signal reconstruction is realized.
Stochastic Input Signal

For a randomly generated input signal x(t), the Power Spectral Density (PSD) of
xs (t) is given by
∞
1 X
Rss (f ) = 4 cos2 (πkfs α)Rxx (f − kfs ), (4.26)
Ts2
k=−∞
where Rxx (f ) is the PSD of x(t). It is observed that it is always the case that
½
1, k/i = 4l,
2
cos (πkfs α) = l = 0, 1, 2, · · · (4.27)
0, k/i = 4l + 2,
Based on the same process for deterministic signal reconstruction, the PSD spec-
trum of the stochastic x(t) can be obtained by s(t) without loss of information.
All the above analysis is based on an ideal sampling case, i.e., the input band-
pass signal is sampled without any noise from the sampling device. However, the
introduced thermal noise can never be avoided during the sampling process in real
applications. Assume that the introduced noise e(t) is Gaussian distributed with a
zero mean and a constant PSD N0 /2. It is band-limited into [−Bef f , Bef f ]. The
time-varying autocorrelation function of xs (t) is given by
rss (t + τ, t) = E[es (t + τ )e∗s (t)], (4.28)
where es (t) is the sampled-data signal of e(t), τ is a time lag and E[•] represents
an expectation operation. The time-average of rss (τ ) over a single sampling period
is defined as [65]
Z Ts /2
1
rss (τ ) = rss (t + τ, t)dτ. (4.29)
Ts −Ts /2
Using this definition, we obtain

∞
1 X
rss (τ ) = ree (τ ) 4 cos2 (πkfs α)ej2πkfs τ
Ts2
k=−∞
The total noise power of es (t) within the fundamental noise bandwidth [−Bef f , Bef f ]
is obtained when τ = 0:
N0 2 X
M
Pe = · fs 4 cos2 (πkfs α), (4.30)
2
k=−(M −1)
where M = Bef f /fs and k represents the different order of Nyquist bands in the
period of fs . As compared to the Pe of conventional uniform BPS that is given by
N0 2 X
M
N0 2
Pe = · fs 2= · fs · 4M, (4.31)
2 2
k=−(M −1)
the noise aliasing by conventional uniform BPS is equally weighted but not by
GQBPS. It is observed that neither of the cases can avoid noise aliasing. M rep-
resents the number of noise spectral foldings. The lower the sampling rate fs , the
larger the value of M such that Pe is increased at the output of BPS system. The
noise power within each of Nyquist bands is the same for conventional uniform
BPS. However, for GQBPS, it is not constant but varying around a mean value
 
 X M 
Mean 4 cos2 (πkfs α) = 4M. (4.32)
 
k=−(M −1)
The factor of 4fs2 in both eq. (4.30) and (4.31) will be balanced by the gain of signal
power (see eq. (4.26)). The gain of noise due to aliasing only depends on M , which
is consistent with eq. (3.16). As shown in Fig. 4.13, it is observed that the noise
gain due to aliasing by GQBPS has a certain shape depending on the sampling
parameters. It is the same as conventional uniform BPS at ±fc but lower in the
range R = ((4n+1)fc , (4n+3)fc ), with the lowest at (4n+2)fc (n = 0, ±1, ±2, · · · ).
With the increase of fs or M , the noise gain by GQBPS approaches to a constant
value M .
Obviously, the amplified in-band noise still survives from the process of the
BPF s(t), but the out-of-band noise and other uninteresting image bands are fil-
tered out. However, if we shift the passband of s(t) to the frequency bands with
the lowest noise gain, the SNR of GQBPS algorithm will be larger than that by
conventional uniform BPS. This advantage of performance improvement becomes
more significant with the decrease of fs . As a consequence, the reconstruction be-
comes more complicated since the spectra located in R are not real anymore. The
GQBPS might be a potential way to reduce the noise aliasing at the cost of a more
complicated reconstruction algorithm. We can combine this filter s(t) with the
following resampler, and both the reconstruction and resampling can be performed
by DSP (see eq. (4.17)). The thermal noise introduced in analog sampling devices
will not present in the resampler realized digitally and consequently no noise alias-
ing happens. The output SNR of GQBPS algorithm is on average the same as by
conventional uniform BPS if s(t) is defined by eq. (4.22).
4.3.2 Simulation results and discussions

The signal reconstruction and noise aliasing can be demonstrated by simulations
using the matlab psd function. Assume that a 2.11 GHz RF signal is received at the
antenna based on W-CDMA standard. The selected channel with 5 MHz bandwidth
is centered at 700 MHz after the first mixer. Scaling down by 106 , a randomly
generated band-limited SSB signal a(t) with B = 5 is frequency-translated to fc =
700 by multiplying with a sinusoidal carrier. A band-limited white Gaussian noise is
added into such that PNi = PNT h (see eq. (3.12)) and Bef f = 10fc . Oversampling
with respect to the BPS theorem is used to see the effects of noise aliasing and
fs1 = fc /2 = 350, fs2 = fc /7 = 100. The simulation results are shown in Fig. 4.14.
It is observed that by GQBPS, the folded spectrum located at [695, 700] dis-
appears while a copy of the input signal spectrum corrupted by noise is located
at [700, 705]. The signal reconstruction can be realized by the BPF s(t). It is in
agreement with the analyses in section 4.3.1. An alternative s(t) with a narrower
passband centered at ±(fc + B/2) can be also used. The SNR of the interesting
band is decreased by around 5 dB for both conventional BPS and GQBPS when
the sampling rate is decreased from fs1 to fs2 . It is consistent with the theoretical
evaluation result 10 log10 (fs2 /fs1 ) ≈ 5.4 dB.
Quadrature processing inevitably encounters the I/Q mismatches in real im-
plementations. The I/Q mismatches cause a signal component related to a∗ (t) to
appear also in the band of a(t) [66], and vice versa. By GQBPS, both the wanted
signal bands and “self-images” are sampled and processed. Due to the zero value
of 1 + e−j2πkfs α or cos2 (πkfs α), both the wanted information bands and the as-
sociated “self-images” are transmitted to zero within the passband of s(t). The
folded spectra of “self-images” will not overlapped with the wanted information
bands. The “self-image” problem due to I/Q mismatches can be overcome by the
proposed GQBPS algorithm. However, the presented approach is sensitive to phase
shift of the sampling clock, e.g., by jitter in the sampling device. This limits the
performance of GQBPS.
Original SSB bandpass signal

0.4
0.3
0.2
0.1
−0.1
−0.2
−0.3
−0.4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t−−>
Reconstructed by Kohlenberg interpolation function

0.4
0.3
0.2
0.1
−0.1
−0.2
−0.3
−0.4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t−−>
Reconstructed by modulated sinc function

0.4
0.3
0.2
0.1
−0.1
−0.2
−0.3
−0.4
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
t−−>
Figure 4.7: Reconstruction of an SSB signal. (Top): input real bandpass signal,
“×” and “◦” represent the sampling positions of two parallel uniform samples by
a quadrature bandpass sampling; (Center): reconstruction by Kohlenberg’s inter-
polation function defined by eq. (4.12); (Bottom): reconstruction by a modulated
sinc function defined by eq. (4.14).
Sample
and
digitize
Even samples
Delay
Bandpass analog input
Frequency Complex
translate lowpass
−jπn
e output
Odd samples Hilbert
transform
Figure 4.8: Sequence of operations to obtain lowpass complex samples from uniform
real real samples of a bandpass signal [27]
(−1)n
Interpolation
x(2n) I(n)
fo = fs1 /4 3/4 Sample
Bandpass signal x(n) Delay
BPF LPF A/D fs2 = fs1 /2
A(t) cos(2πfc t + θ(t)) x(2n − 1)
Interpolation
Q(n)
1/4 Sample
Delay
fLO = fc − fs1 /4 fs1
(−1)n
DPD
Figure 4.9: Digital baseband converter with digital product detector (DPD) [28]
P P
n δ(t − nTs ) m δ(t − mTs0 )
î(mTs0 ) î(m)
A/D
Bandpass Signal xs (t) s(t) x̂(t)
Re{a(t)ej2πfc t }
q̂(mTs0 ) q̂(m)
A/D
P P
n δ(t − nTs − α) m δ(t − mTs0 − α)
c(t)
Figure 4.10: Generalized Quadrature BPS architecture
I(f ) A(f ) = I(f ) + jQ(f )
1111
0000
00001111
11110000
0000
1111 11111
00000
00000
11111
−B 0 B f a)
00000
11111
00000
11111 0 B f c)
Q(f ) A∗ (f ) = I(f ) − jQ(f )
1111
0000
j
11111
00000
00000
11111
00001111
11110000
B
00000
11111
00000
11111
−B 0 f b)
0000
1111 −j 0 B f d)
Figure 4.11: Illustration of SSB signal spectra.
X(f )
111
000 111
000
000
111
000
111 000
111
000
111
a)
000
111 0 000
111
−fc fc f
Al (f )
k=0 k =i−1 k=i k =i+1 k =i+2 k = 2fc /fs k = 2fc /fs + 1
1111
0000 111
000 111
000 111
000 111
000 111
000
0000
1111 000
111
000
111 000
111
000
111 000
111
000
111 000
111
000
111 000
111
0000
1111 000
111 000
111 000
111 000
111 000
111
0000
1111
0000
1111 000
111 000
111 000
111 000
111 000
111
000
111
b)
0000
1111 000
111 0000
111 000
111 000
111 000
111
−fc fc f
Ar (f )
k = −2fc /fs − 1 k = −2fc /fs k = −(i + 2) k = −(i + 1) k = −i k = −i + 1 k=0
111
000 1111
0000 111
000 111
000 111
000 111
000
000
111
000
111 0000
1111
0000
1111 000
111
000
111 000
111
000
111 000
111
000
111 000
111
000
111 0000
1111 000
111 000
111 000
111 000
111
000
111 0000
1111 000
111 000
111 000
111 000
111
000
111
c)
000
111 0000
1111 000
111 000
111
0
000
111 000
111
−fc fc f
S(f )
fs /2
Ts /2
d)
−fc 0 fc f
[Al (f ) + Ar (f )] · S(f )
000
111
111
000 000
111
111
000 e)
000
111 000
111
000
111 0 000
111
−fc fc f
Figure 4.12: Frequency spectra analysis of the proposed GQBPS algorithm.

40
38
Magnitude
36
34
32
30
−3000 −2000 −1000 0 1000 2000 3000
f−−>
5.5
Magnitude
4.5
4
−3000 −2000 −1000 0 1000 2000 3000
f−−>
Figure 4.13: Demonstration of noise gain due to noise aliasing by GQBPS based on
eq. (4.30). (Top): fs = 100, fc = 700, Bef f = 5fc , M = 35; (Bottom): fs = 700,
fc = 700, Bef f = 5fc , M = 5.
−20
fs1=fc/2
fs2=fc/7
−40
PSD (in dB)
−60
−80
−100
0 100 200 300 400 500 600 700

f−−>
−20
−40
PSD (in dB)
−60
−80
−100
0 100 200 300 400 500 600 700

f−−>
Figure 4.14: Comparison of normalized sampled-data spectra, (Top): by con-

ventional BPS, (Bottom): by GQBPS, where fc = 700, Bef f = 10fc , B = 5,
fs1 = fc /2 = 350 (solid line) and fs2 = fc /7 = 100 (dash-dotted line) .
Chapter 5
Nonuniform Random Sampling
Nonuniform random sampling is very close to the realistic implementation. Under

the effects of jitter, deterministic sampling becomes nonuniform random sampling.
The NUS can be further classified into ideal nonuniform sampling (ideal NUS),
jitter sampling (JS), and additive random sampling (ARS) [32]. Ideal NUS is a
deterministic sampling. The other two are random sampling and the sampling
time is not predetermined but is defined by the stochastic process of random jitter.
JS is a common form in real life since the intentional US is generally used. ARS is
equivalent to a nominal ideal NUS under the effects of jitter.
The sampling theorem defines the minimum sampling frequency 2B, below
which the reconstruction of a band-limited signal x(t) is impossible (B is the band-
width of x(t)). From a practical point of view, the sampling rate fs must be many
times greater than 2B to avoid aliasing. However, we neither encounter band-
limited signals nor ideal filters in the real world. In addition, we have no possibility
to obtain an infinite set of samples of the function. Finding a way to unambigu-
ously determine a signal (either ideal or nonideal band-limited) by a finite number
of samples is more of a challenge and interesting.
The NUS mentioned in section 2.1 is deterministic. It is known that NUS has
the potential to suppress the harmful spectrum aliasing of sampled-data signal by
a lower sampling rate even though the input signal is not ideal band-limited (see
Fig. 2.4), and hence the requirements on the AA filter prior to the sampler is relaxed
by using NUS. However, it is still a mystery to select the nonuniformly distributed
sampling scheme such that the input nonideal band-limited signal is uniquely de-
termined by the samples without the harmful effects of aliasing (i.e., alias-free
sampling). Shapiro and Silverman [32], Beutler [67] and Marsy [68] successively
gave or extended the definition and conditions for alias-free sampling. Wojtiuk [69]
highlighted that alias terms can be suppressed by increasing jitter variance, and
also showed that a jitter with a uniform distribution over [−0.5Ts , 0.5Ts ] has the
potential to eliminate the discrete frequency components of the sampled-data sig-
nal except for the information signal. Shapiro [32] also showed that some random
75
76 CHAPTER 5. NONUNIFORM RANDOM SAMPLING
sampling schemes (e.g., Poisson sampling) can eliminate aliasing and lead to an
unambiguous determination of the PSD spectrum.
Random sampling and reconstruction are never a pair of things. Reconstruction
requires a set of exact sampling times, but random sampling can never provide it.
RA∞stochastic process xs (t) by random sampling is not absolutely integrable, i.e.,
−∞
|xs (t)| <
6 ∞, and the Fourier transform of xs (t) does not exist [37]. The study
of PSD is a normal way to analyze random sampling.
5.1 Jitter Sampling
Jitter sampling is also called jittered periodic sampling [32] which is an ideal periodic
sampling affected by timing jitter. The set of sampling time instants is of the form
ts (n) = nTs + τn , n = 0, ±1, ±2, · · · , (5.1)
where Ts is an ideal US interval, τn are a family of independent, identically dis-

tributed (iid) Gaussian random variables with a zero-mean and a standard deviation
στ . Normally στ << Ts . In [70], timing jitter are classified into readin jitters and
readout jitters depending on the way to be introduced in the system. Readin jitters
are introduced when the analog signal is being sampled, whereas readout jitter when
the samples of the output of the digital filter are being read out for reconstruction
back to an analog signal. In the present thesis work, only the case of readin jitters is
considered. We assume that the effects of jitter are unknown. If they were known,
the sampling theory for deterministic NUS could be used [63] [71].
Starting from eq. (2.1), the CT sampled-data signal by JS is given by
∞
X
x̃js (t) = x(t) δ(t − nTs − τn )
n=−∞
∞
X
= x(ts (n))δ(t − nTs − τn ), (5.2)
n=−∞
where the input signal x(t) could be either deterministic or stochastic process and
x̃js (t) is a stochastic process. The statistic process τn and x(t) are independent.
5.1. JITTER SAMPLING 77
The autocorrelation function of xjs (t) is given by

£ ¤
rx̃x̃ (γ, t) = Ex,γ x̃js (t + γ)x̃∗js (t)
X∞ X∞
= Eγ [ Ex [x(t + γ)x∗ (t)] · δ(t + γ − mTs − τm )δ(t − nTs − τn )]
m=−∞ n=−∞
∞
X X∞
= Eγ [ rxx (γ) · δ(t + γ − mTs − τm )δ(t − nTs − τn )]
m=−∞ n=−∞
∞
X X ∞ Z ∞ Z ∞
= rxx (γ) · δ(t + γ − mTs − τm )δ(t − nTs − τn )
m=−∞ n=−∞ −∞ −∞
p(τm , τn )dτm dτn

X∞ X ∞ Z ∞ Z ∞
= rxx (γ) · δ(t + γ − mTs − τm )δ(t − nTs − τn )
m=−∞ n=−∞ −∞ −∞
p(τm )p(τn )dτm dτn

X∞ X∞
= rxx (γ)p(t − mTs − γ)p(t − nTs ), (5.3)
m=−∞ n=−∞
for m 6= n, where E[•] represents an expectation operator, Ex,γ is the average over
the product statistics of τn and x(t), Eγ is over the statistics of τn and Ex over
the statistics of x(t), rxx (γ) is the autocorrelation function of x(t), γ is a time-
lag between any two variables of stochastic process xjs (t), p(τm , τn ) is the joint
probability density function (PDF) of {τn } and {τm }. The random variables τn
and τm are assumed independent such that p(τm , τn ) = p(τn )p(τm ), where p(x) is
the PDF of stochastic process x. When m = n, τm = τn ,
rx̃x̃ (γ) = rxx (0)δ(γ), (5.4)
where rxx (0) corresponds to the total input signal power. Assuming that xjs (t) is
a wide-sense stationary (WSS) process and xjs (t), xjs (t + γ) are jointly ergodic,
the time average may be used to replace the ensemble average. The autocorrelation
function of xjs (t) is simplified by time-average over a single sampling period [65]:
Z Ts /2
1
rx̃x̃ (γ) = rx̃x̃ (γ, t)dt
Ts −Ts /2
Ã ∞
!
1 X
= rxx (γ) rpp (lTs + γ) − rpp (γ) + δ(γ) , (5.5)
Ts
l=−∞
where rpp (lTs + γ) is the convolution of two PDF functions. Based on Wiener-
Khintchine Theorem, the PSD of the WSS process xjs (t) can be obtained from the
Fourier transform of the autocorrelation function rx̃x̃ (γ) [65],

Z ∞" Ã ∞
X
! #
1 −j2πf γ
Rx̃x̃ (f ) = rxx (γ) rpp (lTs + γ) − rpp (γ) + δ(γ) e dγ
Ts −∞
l=−∞
( ∞ )
1 X
= F {rxx (γ)} ? F rpp (lTs + γ) − rpp (γ) + δ(γ)
Ts
l=−∞
Ã ∞ !
1 X
= Rxx (f ) ? Rpp (f )e j2πlTs
− Rpp (f ) + 1
Ts
l=−∞
Ã ∞
!
1 1 X
= Rxx (f ) ? Rpp (f ) δ(f − kfs ) + [1 − Rpp (f )]
Ts Ts
k=−∞
∞
X
1 1
= Rpp (kfs )Rxx (f − kfs ) + Rxx (f ) ? (1 − Rpp (f )) , (5.6)
Ts2 Ts
k=−∞
where ? denotes the convolution operator, F {•} is the Fourier transform operator,
Rxx (f ) and Rpp (f ) is the Fourier transform of rxx (γ) and rpp (γ), respectively and
fs = 1/Ts .
It is observed that the PSD of JS is equivalent to the power spectrum of the
original signal plus an “additive uncorrelated noise”. The first term of eq. (5.6) can
be regarded as a discrete component while the second term a continuous component.
The discrete component is a weighted sum of the periodically shifted copies of input
spectrum Rxx (f ) in the period of average sampling rate fs . It is not necessarily
a periodic function except when Rpp (kfs ) is periodic. When the jitter is small,
Rpp (kfs ) decreases slowly and the discrete component is almost periodic. For a
special case where jitter τn is zero, p(τn ) reduces to δ(τn ) and then Rpp (f ) = 1,
eq. (5.6) reduces to the average PSD of US:
∞
1 X
Rx̃x̃ (f ) = Rxx (f − kfs ). (5.7)
Ts2
k=−∞
However, for ARS ts (n) 6= nTs + τn , the corresponding PSD still consists of the
power spectrum of the original signal plus an “additive uncorrelated noise”, but
the additive part could be arbitrary.
The PSD of JS on a sinusoidal input signal with a random phase is shown in
Fig. 5.2 and Fig. 5.3 for different jitter. The corresponding theoretical weights
Rpp (f ) and theoretical PSD evaluation based on eq. (5.6) are superimposed. The
input frequency is 2 and the average sampling rate is 5. The jitter is assumed to
have a uniform distribution U (−αTs , αTs ) where α = 0, 0.1, 0.3, 0.5 is a scale factor
defined by jitter and 1/Ts is the average sampling rate. All the theoretical weights
Rpp (f ) are shown in Fig. 5.1 for different jitter cases.
Without care of the continuous component (or the bias), the PSD of JS is
a periodically shifted copies of input spectrum Rxx (f ) in the period of average
5.2. ADDITIVE RANDOM SAMPLING 79
10
−10
−20
PSD (in dB)
−30
−40
α=0
α=0.1T
s
α=0.3T
s
α=0.5T
−50 s
−60
0 5 10 15
f−−>
Figure 5.1: Theoretical weights Rpp (f ) based on eq. (5.6) for different jitter cases.
sampling rate fs = 5 shaped by the weight of Rpp (f ). From the simulation result
(see Fig. 5.2 and Fig. 5.3), it is observed that the input spectrum is weighted, and
it matches with the theoretical estimation very well (see Fig. 5.1). When α = 0 (or
ideal US), image spectra appear at higher order Nyquist bands (2nd order [2.5, 7.5],
3rd order [7.5, 12.5], · · · ). The corresponding weight is a flat straight line since the
PDF p(τ ) = δ(τ ) in time domain. With the increase of α, the amplitude of image
spectra decreased with the increase of frequency and the peak level is shaped by
the weight function. When α is increased to 0.5, all image spectra in higher order
Nyquist bands disappear and the spectrum uniquely identifies the input signal.
This simulation result is also consistent with the conclusion given by Wojtiuk [69].
The corresponding sampling scenario is one kind of alias-free sampling.
5.2 Additive Random Sampling
Due to the contribution of nTs , the PSD of JS at ts (n) = nTs + τn still retains the
periodic property in the period of 1/Ts such that the aliasing is still presented in JS.
It is also observed that when jitter has a uniform distribution over [−0.5Ts , 0.5Ts ],
i.e., the samples of JS get rid of the characteristics of US and distribute completely
irregularly, aliasing from higher order Nyquist bands are significantly suppressed.
Shapiro [32] first noticed this and introduced Additive Random Sampling which
breaks up the regular property from JS. It was defined that the samples are located
at
tn = tn−1 + γn , (5.8)
where tn−1 and tn are two successive sampling time instants, γn is an iid stochastic
process with a certain distribution. There exists an average Ts such that E[γn ] = Ts
but tn − tn−1 6= Ts . The PDF of {γn } is equal to zero (i.e., p(γn ) = 0) for γn < 0.
This condition corresponds to the requirement that a set of samples in a given set
of indices should come successively in the time order.
This is equivalent to a nominal ideal NUS under the effects of jitter, since
ts (n) = tn + τn (5.9)
= tn−1 + γn−1 + τn
= tn−1 + γn0 ,
where {tn } is the set of sampling time instants of nominal ideal NUS, E[τn ] = 0
and E[γn0 ] = E[γn ] = Ts .
5.3 Alias-free Sampling

In [32], it was shown that the aliasing can be avoided if the sampling occurs in a
Poisson process with an average rate of ρ. For the given Poisson process {γn }, the
corresponding Poisson distribution in terms of the average rate is given by [32]
p(γ) = ρe−ργ . (5.10)
The same sinusoidal input signal with a random phase that is used for presenting
the PSD of JS is also used for simulating the PSD of ARS. The input frequency is 2
and the average sampling rate is 5. The inter-sample intervals (ISI) {γn } satisfy the
Poisson process defined by eq. (5.10). The corresponding simulated PSD is shown
in Fig. 5.4. The PDF of {γn } used in the simulation is shown and compared with
the theoretical PDF of Poisson process, see Fig. 5.5. It is observed that only the
frequency component of f = 2 in the PSD exists and aliasing effects from other
Nyquist bands are completely avoided. However, the noise floor is significantly
increased such that SNR is degraded. Compared to Fig. 5.3 (Bottom), the in-band
noise power by this ARS is higher than that by JS with the jitter distribution
U [−0.5Ts , 0.5Ts ].
Random sampling which is under the effects of jitter usually causes performance
degradation in radio communications. However, by making use of the random sam-
pling, aliasing can be suppressed efficiently while the signal reconstruction becomes
hardly achievable.
5.3. ALIAS-FREE SAMPLING 81
10
−10
−20
PSD (in dB)
−30
−40
−50
−60
0 5 10 15
f−−>
10
−10
−20
PSD (in dB)
−30
−40
−50
−60
0 5 10 15
f−−>
Figure 5.2: The PSD of JS on a sinusoid input signal with f = 2 for different jitter
and fs = 5. (Top): α = 0; (Bottom): α = 0.1Ts .
10
−10
−20
PSD (in dB)
−30
−40
−50
−60
0 5 10 15
f−−>
10
−10
−20
PSD (in dB)
−30
−40
−50
−60
0 5 10 15
f−−>
Figure 5.3: The PSD of JS on a sinusoid input signal with f = 2 for different jitter
and fs = 5 (cont.). (Top): α = 0.3Ts ; (Bottom): α = 0.5Ts .
5.3. ALIAS-FREE SAMPLING 83
10
−10
PSD (in dB)
−20
−30
−40
−50
−60
0 5 10 15
f (Hz)
Figure 5.4: The PSD of ARS with Poisson process. The input frequency is 2 and
the average sampling rate is 5.
Figure 5.5: The practical p(γ) (in vertical bar) which is used for above simulation
as compared to the theoretical p(γ).
Chapter 6
Noise and Jitter Performance on

RAs
In chapter 2, nine RAs are studied. Three among the nine RAs based on interpola-
tion are possibly used by sample-by-sample basis for online radio communications.
The BKs of the three RAs are studied and presented in both time and frequency
domain. Although it is shown that jitter errors cannot be canceled by using the
RAs but amplified by a large kernel, it is still of interest to study the sensitivity
to jitter and SNR responses for these RAs. As also discussed in chapter 3 that the
jitter effects depend on both the input frequency and standard deviation of random
jitter. BPS provides an interface to a higher input frequency signal. The SNR by
BPS is degraded as compared to the equivalent LPS system in the presence of same
random jitter. However, this difference of SNR is due to the larger jitter effects in
BPS but not noise aliasing as discussed in section 3.2.
In this chapter, a concise model of radio receiver front-end based on BPS is
modeled in matlab. The sensitivity to jitter of these three RAs are studied by
SNDR evaluation based on eq. (2.42). As we have shown in chapter 5 that the
sampling under the effects of jitter can be classified into jitter sampling and additive
random sampling based on the nominal US and NUS, respectively. The jitter effects
are studied for JS and ARS, respectively. The corresponding simulation results are
shown and compared in this chapter.
6.1 Modeling
For studying the signal reconstruction by RAs in the presence of AWGN and jitter
in BPS, a simple system model of a radio receiver front-end by BPS is simulated
in matlab. Considering a current wireless communication standard, Wideband
Code-Division Multiple-Access (W-CDMA) is one of the main technologies of 3G
cellular systems. The required frequency band is located at 1920 MHz - 1980 MHz
and 2110 MHz - 2170 MHz for uplink and downlink, respectively. The channel
85
86 CHAPTER 6. NOISE AND JITTER PERFORMANCE ON RAS
spacing is 5 MHz [72]. Most traditionally used radio receiver architecture is the
conventional superheterodyne receiver architecture, which normally includes two
mixers by using two local oscillators (LOs). With respect to the BPS technique
and the concept of SDR, the first IF stage is directly followed by a bandpass sampler
and the output of BPS is at baseband. We assume that a 2.11 GHz RF signal is
received at the antenna based on W-CDMA standard. The selected channel with
5 MHz bandwidth is centered at 500 MHz after the first mixer. For conveniently
analyzing AWGN and jitter effects on RAs performance, the sampled data signal
is fed into a reconstruction filter which is generalized by a RA.
As shown in Fig. 6.1, scaling by 105 , a sinusoidal carrier signal cos(2πfc t) is
modulated by a randomly generated band-limited signal x(t), where fc = 5000
and B = 50 (50@5000). The passband of the bandpass signal is located at
e(t)
ts(n)
0
x (t) x(t) y(t) x̂(t)
+ k(t, nT )
40001 201 40001
cos(2πfct)
Figure 6.1: A simplified model of a BPS receiver.
[−fc − B, fc + B] ∪ [fc − B, fc + B] (a DSB signal). The AA filter is a fourth-order

Butterworth BPF whose passband is exactly the same as the passband of the input
bandpass signal. The modulated signal y(t) = Re{x(t)ej2πfc t } is either uniformly
or nonuniformly sampled with an average rate of fs = 4B. For JS and ARS, jitter
{τn } has a Gaussian distribution N (0, στ ) (στ is the standard deviation of {τn }).
The reconstruction error is measured by SNDR (see eq. (2.42)), For easily tracking
the discussion, the definition of SNDR is given here again:
PL
i=1 x2i
SN DR = PL , (6.1)
i=1 (xi − x̂i )2
where xi and x̂i denote the points from the original and reconstructed signal, re-
spectively, and L > N (N is the number of sampled points). Three RAs based on
eq. (2.21) and three approximate expressions of kernel k(t, tn ) are evaluated in the
simulation: Low-pass filtering (LPF), Lagrange interpolating polynomial and Spline
interpolation (see the algorithm description in section 2.2).
Based on the above specification of modeling and simulation, we sampled 201
points within a unit time period out of an approximate CT signal consisting of 40001
points. Only the middle range of t ∈ [0.4, 0.6] is evaluated due to the divergence
of Lagrange interpolating polynomial at the interval ends. A sliding window at
6.2. SENSITIVITY OF RAS TO JITTER 87
the input combined with a low order polynomial is a way to implement Lagrange
interpolating polynomial in real applications.
6.2 Sensitivity of RAs to Jitter

In the presence of only jitter, Fig. 6.2 shows that when the band position is close to
DC ([a] and [c] 50@1000), the reconstruction performance of all the RAs is good and
stable. With the shift of band position to a higher frequency ([b] and [d] 50@5000),
the SNDR has a dramatic degradation with the increase of στ for both JS and ARS.
It is known that a sinc kernel has a bad reconstruction performance for NUS [44],
although it is good for US. For small jitter, the jitter error is additive (see eq. (3.22))
such that the frequency spectrum of JS still retains the periodic property of US
spectrum with only spreading at the tone-frequency and higher power levels at
neighboring frequencies. This periodic nature does not exist for ARS such that
LPF based on a sinc kernel always performs worse for ARS than for JS. The La-
grange interpolating polynomial has good reconstruction performance for both US
and ideal NUS [44]. Comparing [a] and [c], the reconstruction performance of the
Lagrange interpolating polynomial is not sensitive to jitter and is also independent
of the sampling distribution (JS or ARS) when jitter is small. When jitter is large,
all the RAs perform badly (see [b] and [d]). It is concluded that the RAs based on
interpolation do not provide any immunity to a large jitter.
6.3 Sensitivity of RAs to AWGN

The SNR response to an input disturbed by noise is defined as
2
SN Rout Pout /σout
= 2 , (6.2)
SN Rin Pin /σin
which is the inverse of the noise reduction ratio [35], where Pin and Pout are the
2 2
power of noise-free input and output signal, σin and σout are the noise variance
of input and output. In this simulation model, Pout = Pin /2. The SNRin and
SNRout is evaluated at the input of BPS after the AA filter and the output of RAs,
respectively.
Based on the same model, the SNR response of RAs is studied and two different
noise effects are considered in BPS device, with only band-limited AWGN e(t) and
with both the band-limited AWGN and random jitter. As shown in Fig. 6.3, with
only AWGN effects (in solid line), the SNR response at the output approximates
the input SNR in the range of SNRin ∈ [0, 10] dB. With the increase of SNRin ,
the performance of each RA shows no significant improvement but approaches a
constant that is defined by the current RA. Under the effects of both AWGN and
jitter (in dash-dotted line), SNRout of each RA decreases by a larger value compared
to the corresponding jitter free case. With the increase of SNRin , SNRout of each
[a] [b]
30 30
LPF
25 25 Spline
Lagrange
20 20
SNDR(in dB)
SNDR(in dB)
15 15
10 10
5 5
0 0
−5 −5
−10 −10
1 2 3 4 5 1 2 3 4 5
σ −5 σ −5
τ x 10 τ x 10
[c] [d]
30 30
LPF
25 25 Spline
Lagrange
20 20
SNDR(in dB)
SNDR(in dB)
15 15
10 10
5 5
0 0
−5 −5
−10 −10
1 2 3 4 5 1 2 3 4 5
σ −5 σ −5
τ x 10 τ x 10
Figure 6.2: SNDR evaluation for jitter effects on BPS and B = 50, fs = 4B. [a]
fc = 1000 JS. [b] fc = 5000 JS. [c] fc = 1000 ARS. [d] fc = 5000 ARS.
6.4. JITTER NOISE EFFECTS 89
RA approaches a constant that is defined by the current RA under the effects of

only jitter.
Comparing Fig. 6.3 (top) and (bottom), LPF shows the best SNR response while
Lagrange interpolating polynomial shows the worst at the output of the reconstruc-
tion filter for both US and JS. For JS, all RAs have equal bad performance. Spline
interpolation shows the highest performance while LPF has the worst for ideal NUS.
For ARS, spline interpolation and Lagrange interpolating polynomial are equally
bad, but the LPF is even worse than theirs. It is also observed that spline inter-
polation algorithm has around 1 dB enhancement on the SNR at the output for
SNRin ∈ [0, 10] dB for both US and ideal NUS. It is also seen that the RAs can not
provide any immunity to noise aliasing. On the contrary, when the SNR is high at
the input, the performance of the RAs limits the SNR performance at the output.
6.4 Jitter Noise Effects

To study the jitter noise effects in the BPS, a random jitter with Gaussian dis-
tribution N (0, στ ) is applied to the randomly generated bandpass signal, where
στ = 5.8 × 10−5 . The center frequency is 5000 (50@5000). The bandpass signal
is sampled by uniform and nonuniform BPS in the presence of jitter (i.e., JS and
ARS). An oversampling rate of 10B is used to see the effects of jitter on both in-
band and out-of-band. It is better to discuss noise effects by PSD rather than by a
signal reconstruction performance. The PSD of the sampled-data signal by BPS is
obtained by matlab psd function.
As shown in Fig. 6.4, the out-of-band noise power PNo is increased due to the
large jitter effects. PNo is evaluated for f > 100 to avoid the effects from the
transition band of the AA filter. The PNo of ARS and JS are about 11.1 dB and
23.6 dB higher than that of ideal NUS and US, respectively. The jitter mainly affects
the out-of-band noise power, and the effects of in-band are negligible. Comparing
the dotted lines in Fig. 6.4 (top) and (bottom), the noise power by ideal NUS is
larger than for ideal US. Even though the PNo increment of ARS is much lower
than that of JS, the output performance by ARS and JS is still in the same order.
40
LPF: στ/Ts=0
Spline: σ /T =0
τ s
35 SNRin=SNRout
Lagrange: σ /T =0
τ s
LPF: σ /T =0.0036
τ s
Spline: σ /T =0.0036
30 τ s
Lagrange: σ /T =0.0036
τ s
25
(in dB)
20
out
SNR
15
10
−5
0 5 10 15 20 25 30 35 40
SNR (in dB)
in
40
LPF: στ/Ts=0
Spline: σ /T =0
τ s SNRin=SNRout
35 Lagrange: σ /T =0
τ s
LPF: σ /T =0.0036
τ s
30 Spline: στ/Ts=0.0036
Lagrange: σ /T =0.0036
τ s
25
(in dB)
20
out
SNR
15
10
−5
0 5 10 15 20 25 30 35 40
SNRin (in dB)
Figure 6.3: Comparison of SNR responses of RAs for US and JS (top), NUS and
ARS (bottom) in BPS, B = 50, fc = 100B = 5000, fs = 4B.
6.4. JITTER NOISE EFFECTS 91
0
σ =5.8 × 10−5
τ
−10 jitter free
−20
PSD (in dB)
−30
−40
−50
−60
0 50 100 150 200 250
f (in Hz)
−10
PSD (in dB)
−20
−30
−40
−50
0 50 100 150 200 250
f (in Hz)
Figure 6.4: The PSD of sampled-data signal by BPS with fc = 5000, B = 50 and
fs = 10B for JS (top) and ARS (bottom).
Chapter 7
Conclusions and Future Work
In this thesis work, the current existing receiver architectures are reviewed and
compared to SDR receiver. Basic sampling and reconstruction theory are studied.
In practice the samples by US is never uniformly distributed due to the effects of
clock jitter or power supply noise in sampling devices. It is of great benefit to study
the NUS. However, a single ideal lowpass filter based on the Shannon’s sampling
theory is not good enough to reconstruct the signal from the samples by NUS.
Starting from a general reconstruction formula in terms of the nonuniform samples
and a BK, nine RAs are investigated for reconstructing the input signal from the
nonuniform samples. The performance of this RAs are evaluated and compared
by simulations. Most of them are extensively used in off-line image processing,
but some of them based on interpolation are also possibly used in on-line radio
communications.
The design goal of SDR is to put the A/D converter as close as possible to the
antenna. BPS realizes frequency down conversion on a modulated bandpass signal
by undersampling. It enables one to have an interface between the higher IF and
the A/D converter and might be a solution to SDR. Three main aspects in the
BPS technique, the allowable uniform BPS rate selection, noise aliasing and timing
jitter, are reviewed and studied in this thesis work as compared to the conventional
LPS cases.
It is noticed that noise aliasing plays an important role in the BPS applica-
tions. Starting from the Papoulis’ generalized sampling theorem, a Generalized
Quadrature BPS (GQBPS) algorithm is proposed to suppress the noise aliasing. It
is shown that the out-of-band noise aliasing is suppressed significantly but in-band
noise aliasing is still present in the process of GQBPS algorithm. However, the
GQBPS might be a potential way to reduce the noise aliasing at both the in-band
and out-of-band at the cost of a more complicated reconstruction algorithm.
BPS makes use of signal spectral folding (or aliasing) by undersampling. Harm-
ful signal spectrum aliasing due to careless uniform BPS rate selection will cause loss
of information. It was shown in literature that NUS has the potential to suppress
93
94 CHAPTER 7. CONCLUSIONS AND FUTURE WORK
harmful signal spectrum aliasing. In this thesis work, the PSDs of sampled-data
signal by two random samplings (JS and ARS) which are special cases of NUS
are studied. The definitions and conditions of alias-free sampling are verified by
simulations.
It is of more interest to see the performance of noise and jitter in real BPS
applications. A simplified model of a BPS receiver is modeled and simulated. The
sensitivity of RAs to jitter and AWGN are studied based on the model.
Future work is still needed to establish a more efficient sampling architecture
or algorithm that possesses the properties of i) flexible sampling rate selection
for avoiding harmful signal spectrum aliasing; ii) noise aliasing suppression; iii)
tolerance or correction to jitter. A CMOS chip is expected after the idea becomes
mature.
Bibliography
[1] B. Razavi, RF Microelectronics. Prentice Hall, 1998.
[2] J. C. Rudel et al, “A 1.9-GHz wide-band IF double conversion CMOS re-

ceiver for cordless telephone applications,” IEEE Journal of Solid-State Cir-
cuits, vol. 32, pp. 2071–2088, Dec. 1997.
[3] K. C. Zangi and R. D. Koilpillai, “Software radio issues in cellular base sta-
tions,” IEEE Journal of Solid-State Circuits, vol. 17, pp. 561–573, April 1999.
[4] J. Mitola, “The software radio architecture,” IEEE Communications Maga-

zine, pp. 26–38, May 1995.
[5] J. Ryynänen et al, “A single-chip multimode receiver for GSM900, DCS1800,

PCS1900, and WCDMA,” IEEE Journal of Solid-State Circuits, vol. 38,
pp. 594–602, April 2003.
[6] B. Brannon, “Fast and hot - data converters for tomorrow’s software-defined
radios,” RF mixed signal, 2002.
[7] R. J. Lackey and D. W. Upmal, “Speakeasy: The military software radio,”

IEEE Commun. Mag., vol. 33, pp. 56–61, May 1995.
[8] K.-Y. Lee et al, “Full-CMOS 2-GHz WCDMA direct conversion transmitter
and receiver,” IEEE Journal of Solid-State Circuits, vol. 38, pp. 43–53, January
2003.
[9] K.-J. Koh et al, “Subharmonically pumped CMOS frequency conversion (up
and down)circuits for 2-GHz WCDMA direct-conversion transceiver,” IEEE
Journal of Solid-State Circuits, vol. 39, pp. 871–884, June 2004.
[10] S. K. Reynolds et al, “A direct-conversion receiver IC for WCDMA mobile

systems,” IEEE Journal of Solid-State Circuits, vol. 38, pp. 1555–1560, Sept.
2003.
[11] R. Gharpurey et al, “A direct-conversion receiver for the 3G WCDMA stan-

dard,” IEEE Journal of Solid-State Circuits, vol. 38, pp. 556–560, March 2003.
95
96 BIBLIOGRAPHY
[12] H. S. Black, Modulation Theory. New York, Van Nostrand, 1953.
[13] H. Nyquist, “Certain topics in telegraph transmission theory,” AIEE Trans.,

vol. 47, pp. 617–644, 1928.
[14] D. Gabor, “Theory of communication,” J. Inst. Elec. Eng., vol. 93, 1946.
[15] A. Kohlenberg, “Exact interpolation of band-limited functions,” J. Appl.

Phys., vol. 24, pp. 1432–1436, Dec. 1953.
[16] R. G. Vaughan, N. L. Scott, and D. R. White, “The theorem of bandpass

sampling,” IEEE Trans. on Signal Processing, vol. 39, pp. 1973–1984, Sept.
1991.
[17] J. Yuan, “Accurate sampling of radio signals beyond GHz in CMOS,” GHz
2000 Symposium, pp. 277–280, March 2000.
[18] D. Jakonis, Direct RF Sampling Receivers for Wireless Systems in CMOS Tech-
nology. PhD thesis, Linköping University, Sweden, 2004.
[19] T. O. Salo et al, “80-MHz bandpass Σ∆ modulators for multimode digital IF

receivers,” IEEE Journal of Solid-State Circuits, vol. 38, pp. 464–474, March
2003.
[20] K. Muhammad et al, “A discrete-time bluetooth receiver in a 0.13µm digital

CMOS process,” Proceeding of ISSCC, pp. 268–269, 2004.
[21] K. Muhammad and R. B. Staszewski, “Direct RF sampling mixer with recur-

sive filtering in charge domain,” Proc. of IEEE International Symposium on
Circuits and System (ISCAS), vol. I, pp. 577–580, 2004.
[22] A. J. Jerri, “The shannon sampling theory – its various extensions and appli-
cations: A tutorial review,” Proc. of the IEEE, vol. 65, no. 11, pp. 1565–1598,
1977.
[23] A. J. Coulson, “A generalization of nonuniform bandpass sampling,” IEEE

Trans. on Signal processing, vol. 43, pp. 694–704, Mar. 1995.
[24] Y.-R. Sun and S. Signell, “Effects of noise and jitter on algorithms for band-
pass sampling in radio receivers,” Proc. of IEEE International Symposium on
Circuits and System (ISCAS), May 2004.
[25] H. Kobayashi et al, “Aperture jitter effects in wideband sampling systems,”

Proc. of 16th IEEE Instrumentation and Measurement Technology Conference,
vol. II, pp. 880–885, 1999.
[26] Y.-R. Sun and S. Signell, “Effects of noise and jitter in bandpass sampling,”
Proc. 21st Norchip conference, Nov. 2003.
97
[27] D. W. Rice and K. H. Wu, “Quadrature sampling with high dynamic range,”
IEEE Trans. Aerosp. Electron. Syst., vol. AES-18, pp. 736–739, Nov. 1982.
[28] L. E. Pellon, “A double nyquist digital product detector for quadrature sam-
pling,” IEEE Trans. on Signal Processing., vol. 40, pp. 1670–1681, July 1992.
[29] S. Levantino, “A CMOS GSM IF-sampling circuit with reduced in-channel

aliasing,” IEEE Journal of Solid-State Circuits, vol. 38, no. 6, pp. 895–904,
2003.
[30] L. R. Carley and T. Mukherjee, “High-speed low-power integrating CMOS

sample-and-hold amplifier architecture,” IEEE Custom Integrated Circuits
Conference, pp. 543–546, 1995.
[31] G. Xu, Charge Sampling Circuits and A/D Converters–Theory and Experi-
ments. PhD thesis, Lund University, Sweden, 2004.
[32] H. S. Shapiro and R. A. Silverman, “Aliasing free sampling of random noise,”

J. Soc. Indust. Appl. Math., vol. 8, pp. 225–248, June 1960.
[33] A. Papoulis, “Generalized sampling expansion,” IEEE Trans. on Circuits and

Systems, vol. CAS-24, pp. 652–654, Nov. 1977.
[34] A. V. Balakrishnan, “On the problem of time jitter in sampling,” IRE Trans.
on Information Theory, pp. 226–236, April 1961.
[35] S. J. Orfanidis, Introduction to signal Processing. Prentice Hall, 1996.
[36] C. E. Shannon, “Communication in the presence of noise,” Proc. IRE, vol. 37,
pp. 10–21, 1949.
[37] J. G. Proakis and D. G. Manolakis, Digital Signal Processing: principles, al-

gorithms, and applications. Prentice Hall, 1996.
[38] Digital Alias-free Signal Processing–Application Notes: Nonuniform sampling

(AN1), 2001.
[39] Online, http://docs.sun.com/db/doc/805-4368, URL.
[40] E. T. Whittaker, “On the functions which are represented by the expansions
of the interpolation theory,” Proc. Royal Soc., Edinburgh, vol. 35, pp. 181–194,
1914.
[41] H. Sedarat and D. G. Nishimura, “On the optimality of the gridding recon-
struction algorithm,” IEEE Trans. on Medical Imaging, vol. 19, pp. 306–317,
April 2000.
[42] J. R. Higgins, “A sampling theorem for irregularly spaced sample points,”

IEEE Trans. Info. Theory, vol. 22, pp. 621–622, 1976.
98 BIBLIOGRAPHY
[43] L. Mate, Hilbert space methods in science and engineering. Adam Hilger,
Bristol and New York, 1989.
[44] Y.-R. Sun and S. Signell, “Algorithms for nonuniform bandpass sampling in
radio receiver,” Proc. of International Symposium on Circuits and Systems
(ISCAS), vol. I, pp. 1–4, May 2003.
[45] R. H. Bartels, J. C. Beatty, and B. A. Barsky, Hermite and Cubic Spline

Interpolation. San Francisco, CA: Morgan Kaufmann, 1998.
[46] C. Cenker, H. G. Feichtinger, and M. Herrmann, “Iterative algorithms in ir-

regular sampling a first comparison of methods,” IEEE, 1991.
[47] A. Aldroubi and K. Gröchenig, “Nonuniform sampling and reconstruction in
shift-invariant spaces,” SIAM Rev., vol. 43, no. 4, pp. 585–620, 2001.
[48] J. L. Yen, “On nonuniform sampling of bandwidth-limited signals,” IRE Trans.

on circuit theory, vol. CT-3, pp. 251–257, Dec. 1956.
[49] H. Freeman, Discrete-time systems: an introduction to the theory. John Wiley

& Sons, INC, 1965.
[50] K. Yao, “Applications of reproducing kernel hilbert spaces-bandlimited signal

model,” Information and control, vol. 11, pp. 429–444, 1967.
[51] B. Sankur and L. A. Gerhardt, “Reconstruction of signals from nonuniform

samples,” IEEE international conference on communications, pp. 1513–1518,
1973.
[52] Online, http://ccrma-www.stanford.edu/ jos/Interpolation/ Interpola-

tion.html, URL.
[53] Y. C. Eldar and A. V. Oppenheim, “Filterbank reconstruction of bandlimited
signals from nonuniform and generalized samples,” IEEE Trans. Sig. Proc.,
vol. 48, no. 10, pp. 2864–2875, 2000.
[54] V. Välimäki, Discrete-Time Modeling of Acoustic Tubes Using Fractional Delay

Filters. PhD thesis, Helsinki University of Technology, Finland, 1995.
[55] J. G. Proakis, Digital Communications, Fourth Edition. McGraw-Hill, 2001.
[56] S. S. Awad, “Analysis of accumulated timing-jitter in the time domain,” IEEE

Trans. on Instrumentation and Measurement, vol. 47, no. 1, pp. 69–73, 1998.
[57] C. B. Feldman and W. R. Bennett, “Bandwidth and transmission perfor-

mance,” Bell Syst. Tech. J., vol. 28, pp. 490–595, 1949.
[58] J. D. Gaskell, “Linear systems, Fourier transforms, and Optics,” Wiley, New
York, 1978.
99
[59] J. L. J. Brown, “On uniform sampling of amplitude modulated signals,” IEEE

Trans. Aerosp. Electron. Syst., vol. AES-19, July 1983.
[60] J. H. Fischer, “Noise sources and calculation techniques for switched capacitor
filters,” IEEE Journal of Solid-state Circuits, vol. SC-17, pp. 742–752, August
1982.
[61] P. P. Vaidynanathan, Multirate Systems and Filter Banks. Prentice Hall, 1993.
[62] Digital Modulation in Communications Systems–An Introduction, 2001.
[63] R. J. Marks II, Introduction to Shannon Sampling and Interpolation Theory.

Springer-Verlag, 1991.
[64] S. Ries, “Reconstruction of real and analytic band-pass signals from a finite
number of samples,” Signal Processing, vol. 33, no. 3, pp. 237–257, 1993.
[65] L. W. Couch II, Digital and Analog Communication Systems, Fifth Edition.
Prentice-Hall International, 1987.
[66] M. Valkama and M. Renfors, “A novel image rejection architecture for quadra-
ture radio receivers,” IEEE Trans. on Circuits and Systems-II: Express Briefs,
vol. 51, pp. 61–68, Feb. 2004.
[67] F. J. Beutler, “Alias-free randomly timed sampling of stochastic processes,”

IEEE Trans. Info. Theory, vol. 16, pp. 147–152, 1970.
[68] E. Marsy, “Alias-free sampling: An alternative conceptualization and its ap-
plications,” IEEE Trans. Info. Theory, vol. 24, pp. 317–324, 1978.
[69] J. J. Wojtiuk, Randomised Sampling for Radio Design. PhD thesis, Univ. of
South Australia, 2000.
[70] B. Liu, “Timing jitter in digital filtering of analog signals,” IEEE Trans. on
Circuits and Systems, vol. CAS-22, pp. 218–223, March 1975.
[71] F. A. Marvasti, “A unified approach to zero-crossing and nonuniform sam-

pling,” Oak Park Ill., 1987.
[72] Online, http://www.umtsworld.com/technology/wcdma.htm, URL.

Other Publications from ESD Lab
Licentiate Theses
• Johnny Öberg. An Adaptable Environment for Improved High-level Syn-
thesis, ISRN KTH/ESD/AVH–99/14-SE, 1996.
• Bengt Oelmann. Design and Performance Evaluation of Asynchrounus Mi-
cropipline Circuits for Digital Radio, ISRN KTH/ESD/R–96/17–SE, 1996.
• Henrik Olson. ASIC Implementable Synchronization and Detection Meth-
ods for Direct Sequence Spread Spectrum Wideband Radio Receivers,
ISRN KTH/ESD/R–97/11-SE, 1997.
• Mattias O’Nils. Hardware/Software Partitioning of Embedded Computer
Systems, ISRN KTH/ESD/R–96/17–SE, 1996.
• Daniel Kerek. Design of A Wideband Direct Sequence Spread Spectrum
Radio Transceiver ASIC, ISRN KTH/ESD/AVH–99/5-SE, 1999.
• Bengt Svantesson. Dynamic Process Management in SDL to VHDL Syn-
thesis, ISRN KTH/ESD/AVH–2000/4-SE, 2000.
• Johnny Holmberg. On Design and Analysis of LDI and LDD Lattice Filters,
ISRN KTH/ESD/AVH–2000/8-SE, 2000.
• Bingxin Li. Design of Sigma Delta Modulators for Digital Wireless Com-
munications, ISRN KTH/ESD/AVH–2001/2-SE, 2001.
• Chuansu Chen. Reusable Macro Based Synthesis for Digital ASIC Design,
ISRN KTH/IMIT/LECS/AVH–2002/1-SE, 2002.
• Li Li. Noise Analysis of Mixers for RF Receivers,
ISRN KTH/IMIT/LECS/AVH–2002/2-SE, 2002.
• Tomas Bengtsson. Boolean Decomposition in Combinational Logic Synthe-
sis, ISRN KTH/IMIT/LECS/AVH-03/08–SE, 2003.
• Meigen Shen. Chip and Package Co-Design for Mixed-Signal Systems: SoC
versus SoP, ISRN KTH/IMIT/LECS/AVH-03/11-SE, 2004.
• Jimson Mathew. Design and Evaluation of Fault Tolerant VLSI Architec-
tures, ISRN, KTH/IMIT/LECS/AVH-04/03–SE, 2004.
• Petra Färm. Advanced Algorithms for Logic Synthesis,
ISRN KTH/IMIT/LECS/AVH-04/01–SE, 2004.
• René Krenz. Graph Dominators in Logic Synthesis and Verification, ISRN,
KTH/IMIT/LECS/AVH-04/04–SE, 2004.
• Andrés Martinelli. Advanced Algorithms for Boolean Decomposition,
• Tarvo Raudvere. Verification of Local Design Refinements in a System
Design Methodology, ISRN KTH/IMIT/LECS/AVH-04/09–SE, 2004.
• Maxim Teslenko. Decomposition Algorithms for Efficient Logic Synthesis,
• Xinzhong Duo. Design and Implementation of RF System-on-Package
Modules for Short-Range Wireless Communications,
ISRN KTH/IMIT/LECS/AVH-03/12-SE, 2004.
• Wim Michielsen. Chip and Package Co-Design for Colpitts Oscillators,
Doctoral Theses
• Tawfik Lazraq. Design Techniques and Structures for ATM Switches, ISBN
91-7170-703-4, 1995.
• Bengt Jonsson. Switched-Current Circuits: from Building Blocks to Mixed
Analog-Digital Systems, ISRN KTH/ESD/AVH–99/1–SE, 1999.
• Johnny Öberg. ProGram: A Grammar-Based Method for Specification
and Hardware Synthesis of Communication Protocols,
ISRN KTH/ESD/AVH–99/3–SE, 1999.
• Mattias O’Nils. Specification, Synthesis and Validation of
Hardware/Software Interfaces, ISRN KTH/ESD/AVH–99/4–SE, 1999.
• Peeter Ellervee. High-Level Synthesis of Control and Memory Intensive
Applications, ISRN KTH/ESD/AVH–2000/1–SE, 2000.
• Henrik Olson. Algorithm-to-Architecture Refinement for Digital Baseband
Radio Receivers, ISRN KTH/ESD/AVH–2000/2–SE, 2000.
• Bengt Oelmann. Asynchronous and Mixed Synchronous/Asynchronous
Design Techniques for Low Power, ISRN KTH/ESD/AVH–2000/6–SE, 2000.
• Yonghong Gao. Architecture and Implementation of Comb Filters and
Digital Modulators for Oversupplying A/D and D/A Converters, ISRN
KTH/ESD/AVH–2001/1–SE, 2001.
• Lirong Zheng. Design, Analysis and Integration of Mixed-Signal Systems
for Signal and Power Integrity, ISRN KTH/ESD/AVH-2001/3-SE, 2001.
• Per Bjuréus. High-Level Modeling and Evaluation of Embedded Real-Time
Systems, ISRN KTH/IMIT/LECS/AVH–02/3–SE, 2002.
• Imed Ben Dhaou. Low Power Design Techniques for Deep Submicron Tech-
nology with Application to Wireless Transceiver Design,
ISRN KTH/IMIT/LECS/AVH–02/4–SE, 2002.
• Ingo Sander. System Modeling and Design Refinement in ForSyDe, ISRN
KTH/IMIT/LECS/AVH-03/03–SE, 2003.
• Andreas Göthenberg. Modeling and Analysis of Wideband Sigma-Delta
Noise Shapers, ISRN KTH/IMIT/LECS/AVH-03/04–SE, 2003.
• Dinesh Pamunuwa. Modelling and Analysis of Interconnects for Deep Sub-
micron Systems-on-Chip, ISRN KTH/IMIT/LECS/AVH-03/07–SE, 2003.
• Bingxin Li. Design of Multi-bit Sigma-Delta Modulators for Digital Wire-
less Communications, ISRN KTH/IMIT/LECS/AVH-03/10–SE, 2003.
• Li Li. Modelling, Analysis and Design of RF Mixed-Signal Mixer for
Wireless Communications, ISRN KTH/IMIT/LECS/AVH-04/12–SE, 2004.
• Abhijit Kumar Deb. System Design for DSP Applications with the MASIC
Methodology, ISRN KTH/IMIT/LECS/AVH-04/10–SE, 2004.

Nonuniform Bandpass Sampling in Radio Receivers

Uploaded by

Copyright:

Available Formats

Nonuniform Bandpass Sampling in Radio Receivers

Uploaded by

Document Information

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

Nonuniform Bandpass Sampling in Radio Receivers

Uploaded by

Copyright:

Available Formats

Nonuniform Bandpass Sampling in Radio Receivers

Nonuniform Bandpass Sampling in Radio Receivers

A dissertation submitted to Kungliga Tekniska Högskolan,

TRITA-IMIT-LECS AVH 04:13

Royal Institute of Technology

As an interface between radio receiver front-ends and digital signal processing

Non-Reviewed presentations and Technical Reports:

2 Sampling and Reconstruction 13

3 Uniform Bandpass Sampling 41

4 Quadrature Bandpass Sampling 55

6 Noise and Jitter Performance on RAs 85

7 Conclusions and Future Work 93

2/2.5/3 G the second/second and half/third Generation

B The bandwidth of lowpass information signal

1.1 Conventional dual-IF superheterodyne receiver architecture . . . . . . . 3

2.1 Voltage sampling process . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.15 Transposed FIR structure . . . . . . . . . . . . . . . . . . . . . . . . . . 37

3.1 a) Original modulated bandpass signal with a bandwidth B = 50 and

4.1 Identity of signal representation by Papoulis’ generalized sampling the-

4.5 Model of second-order BPS based on the Kohlenberg’s sampling theo-

6.1 A simplified model of a BPS receiver. . . . . . . . . . . . . . . . . . . . 86

dard, narrow band radio communications. For a specific communication standard,

Table 1.1: Wireless Radio Communication Standards

is normally selected prior to the sampling in the A/D converter. Alternatively,

1.1 Superheterodyne receivers

BPF Amplifier IRF BPF BPF Amplifier

Figure 1.1: Conventional dual-IF superheterodyne receiver architecture

1.2 Homodyne receivers

Figure 1.2: Homodyne receiver architecture

1.3 Wideband IF Receivers

BPF Amplifier LPF LPF

Figure 1.3: Wideband IF receiver architecture with double down-conversion

The entire frequency band of information signal located at RF is translated to

1.4 Generic Wideband Receivers

For multi-mode WB operation, the frequency of the tunable LO is adapted to fit a

BPF Amplifier BPF

··· ··· ··· ··· ··· ··· ··· ···

Figure 1.4: Generic wideband receiver architecture

1.5 Software Defined Radio Receivers

following A/D converter. In addition, BPS can realize down-conversion through

BPF Amplifier BPF

(n − 1/2)fs nfs (n + 1/2)fs

··· ··· ··· ··· ··· ···

Figure 1.5: Software defined radio receiver architecture by BPS

Many single-mode wideband homodyne receivers were designed for W-CDMA

Table 1.2: High Level Comparison of Receiver Architectures

Receiver Full A/D Potential Selecting Image

1.6 Outline of Technical Problems

−fu −fc −fl 0 fl fc fu f

1.7 Overview of Previous Work on BPS

1.8 Summary of Contributions

• In chapter 2, two kinds of deterministic sampling techniques, Uniform Sam-

terms of a basis-kernel (BK) is used to describe the algorithms. The com-

• In chapter 3, BPS technique is introduced. The associated problems, avail-

• In chapter 4, a brief overview of the Papoulis’ generalized sampling theorem,

• In chapter 5, two classes of nonuniform random sampling, Jitter Sampling

• In chapter 6, the reconstruction performances by different RAs of lowpass

• In chapter 7, the thesis is concluded and some future work is proposed.

Sampling and Reconstruction

2.1.1 Voltage sampling and Charge sampling