FULLTEXT01

Download as pdf or txt
Download as pdf or txt
You are on page 1of 65

Performance and Complexity Comparison

of Doppler Spread Estimation for


WCDMA Systems

ZIQI PENG

Master’s Degree Project


Stockholm, Sweden 2014

XR-EE-KT 2014:009
Abstract

In cellular communication systems, the estimation of Doppler spread has a wide


range of applications such as handoff, channel assignment scheme, adaptive
transmission, power control, etc. A great quantity of Doppler spread estimation
algorithms have been proposed in the literature. But there has been few investi-
gations which gives a comprehensive comparison of these algorithms. Therefore,
it is of great significance to compare and evaluate the performance of the existing
algorithms in the same simulation framework.
In this report, the uplink of WCDMA is considered. Four different types of
Doppler spread estimation algorithms are evaluated and compared in a link level
baseband simulator. The performance and the ability to implement are consid-
ered as the metrics for evaluation. Both Rayleigh and Rician fading channel
model are applied, and the effect of speed, signal to noise ratio, Rician factor
and the angle of arrived line of sight component are also tested. Moreover,
the computational complexity is analysed to evaluate the practical value for
implementation.

Key words: Doppler spread estimation; WCDMA uplink systems; Rayleigh


fading model; Rician fading model; computational complexity.
Abstrakt

Estimatering av en mobils hastighet i form av Dopplerspridning har ett brett


spektrum av tillmpningar i cellulra kommunikationssystem ssom frflyttningen av
mobiler mellan celler, kanaltilldelningsschema, adaptiv sndning, effektstyrning,
etc. En stor mngd algoritmer fr estimering av Dopplerspriding har frslagits
i litteraturen, men det r ovanligt med heltckande jmfrelser mellan med dessa
algoritmer. Drfr r det av stor betydelse att jmfra och utvrdera resultaten av
befintliga algoritmer inom ramen fr samma simuleringsvertyg.
I denna rapport anvnds upplnken fr WCDMA fr utvrdering av fyra olika
typer av algoritmer fr estimering av Dopplerspridning. Metriker fr utvrderin-
gen r prestanda och implementeringsvnlighet. Bde Rayleigh och Rician fd-
ningskanal modeller har utvrderas, samt effekten av mobilens hastighet, signal
till brus frhllande, Rician faktor och infallsvinkel i ppet flt scenario. Dessutom
har den berkningsmssiga komplexiteten analyseras fr att utvrdera den praktiska
anvndbarheten i riktiga system.
Acknowledgment

First, I would like to thank my supervisor Per Löfving in Ericsson for giving me
this opportunity to work on this thesis and providing the insightful guidance for
my work. And I would also like to express my appreciation to Henrik Sahlin,
Magnus Nilsson and Lu Li for the continuous help, encouragement and feedbacks
throughout the entire thesis work.
Second, I would like to thank my thesis partner Hui Wen in KTH for the
cooperation and inspiration.
And I would also like to thank my examiner Prof. Tobias Oechtering in
KTH for his advice and time taken to supervise my thesis.
Finally I would like to thank my family for their constant support and love
during my studies in Sweden.
Contents

List of Figures ii

List of Tables v

List of Symbols vi

List of Abbrevations viii

1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

2 Background 4
2.1 Propagation Phenomena . . . . . . . . . . . . . . . . . . . . . . . 4
2.1.1 Path Loss . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.2 Shadow Fading . . . . . . . . . . . . . . . . . . . . . . . . 5
2.1.3 Multipath . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Doppler Effect and Doppler Spread . . . . . . . . . . . . . . . . . 6
2.2.1 Doppler Effect . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.2 Doppler Spread . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Channel Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3.1 Channel Response . . . . . . . . . . . . . . . . . . . . . . 7
2.3.2 Autocorrelation Functions . . . . . . . . . . . . . . . . . . 8
2.3.3 Power Spectrum . . . . . . . . . . . . . . . . . . . . . . . 9
2.4 Structure of DPCCH in WCDMA Uplink . . . . . . . . . . . . . 9

3 Estimation Algorithms 11
3.1 Level Crossing Rate Estimator . . . . . . . . . . . . . . . . . . . 11
3.2 Hybrid Estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2.1 Curvature Estimator (small Doppler spread) . . . . . . . 13
3.2.2 First Zero Detection Estimator (large Doppler spread) . . 14
3.2.3 Hybrid Estimator . . . . . . . . . . . . . . . . . . . . . . . 16
3.3 Power Integration Estimator . . . . . . . . . . . . . . . . . . . . . 17
3.4 Frequency Domain Maximum Likelihood Estimator . . . . . . . . 19

i
4 Simulation Results 21
4.1 Simulation Environment . . . . . . . . . . . . . . . . . . . . . . . 21
4.2 Simulation Results . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.2.1 Level Crossing Rate Estimator . . . . . . . . . . . . . . . 22
4.2.2 Hybrid Estimator . . . . . . . . . . . . . . . . . . . . . . . 24
4.2.3 Power Integration Estimation . . . . . . . . . . . . . . . . 26
4.2.4 Frequency Domain ML Estimator . . . . . . . . . . . . . 30

5 Performance Evaluation 34
5.1 Computational Complexity . . . . . . . . . . . . . . . . . . . . . 34
5.1.1 Level Crossing Rate Estimator . . . . . . . . . . . . . . . 34
5.1.2 Hybrid Estimator . . . . . . . . . . . . . . . . . . . . . . . 35
5.1.3 Power Integration Estimator . . . . . . . . . . . . . . . . 36
5.1.4 Frequency Domain ML Estimator . . . . . . . . . . . . . 36
5.1.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
5.2 Performance Comparison . . . . . . . . . . . . . . . . . . . . . . 38
5.2.1 Rayleigh Fading Channel . . . . . . . . . . . . . . . . . . 39
5.2.2 Rician Fading Channel . . . . . . . . . . . . . . . . . . . . 41

6 Conclusions 45
6.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6.2 Social and Ethical Aspects . . . . . . . . . . . . . . . . . . . . . . 45
6.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

Appendices 47

A Whittle Approximation Log-likelihood Goal Function Deriva-


tion 47

ii
List of Figures

2.1 The propagation environment. . . . . . . . . . . . . . . . . . . . 4


2.2 The pass loss model. . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Multipath fading. . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.4 The effect of multipath propagation. . . . . . . . . . . . . . . . . 6
2.5 Illustration of Doppler effect. . . . . . . . . . . . . . . . . . . . . 7
2.6 The structure of the uplink Dedicated Physical Data CHannel
(DPDCH) and Dedicated Physical Control CHannel (DPCCH). . 9

3.1 Level crossing example. . . . . . . . . . . . . . . . . . . . . . . . 12


3.2 Illustration of linear interpolation. . . . . . . . . . . . . . . . . . 15
3.3 Illustration of power integration estimator. . . . . . . . . . . . . . 17
3.4 The Continuous and Discrete form of Theoretical Periodogram
PSD. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

4.1 The estimated maximum Doppler spread of Level Crossing Rate


(LCR) method with different levels,versus the theoretical maxi-
mum Doppler spread. . . . . . . . . . . . . . . . . . . . . . . . . 23
4.2 The NMSE of LCR method with different levels,versus the theo-
retical maximum Doppler spread. . . . . . . . . . . . . . . . . . . 23
4.3 The NMSE of LCR method with different levels, versus SNR for
four speed values. . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.4 The comparison of two polynomial fitting algorithms. . . . . . . 25
4.5 The estimated maximum Doppler spread of Hybrid method,versus
the theoretical maximum Doppler spread. . . . . . . . . . . . . . 26
4.6 The NMSE of estimated maximum Doppler spread of Hybrid
method,versus the theoretical maximum Doppler spread. . . . . . 26
4.7 The NMSE of Hybrid method with four speed values, versus SNR. 27
4.8 The NMSE of Hybrid Estimator for different number of used
paths, versus theoretical maximum Doppler spread. . . . . . . . . 27
4.9 The NMSE of Power Integration Estimator with β = 1, 2, versus
threshold ψ. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.10 The estimated maximum Doppler spread of Power Integration
method, versus the theoretical maximum Doppler spread. . . . . 28
4.11 The NMSE of estimated maximum Doppler spread of Power In-
tegration method,versus the theoretical maximum Doppler spread. 29
4.12 The comparison of the number of frames used in simulation for
Power Integration method, versus the theoretical maximum Doppler
spread. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

iii
4.13 The NMSE of Power Integration method with four speed values,
versus SNR. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.14 The normalized bias and normalized standard deviation of Power
Integration method with four speed values, versus SNR. . . . . . 30
4.15 The estimated maximum Doppler spread of ML method,versus
the theoretical maximum Doppler spread. . . . . . . . . . . . . . 31
4.16 The comparison between estimated spectrum and theoretical spec-
trum . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.17 The NMSE of estimated maximum Doppler spread of ML method,
versus the theoretical maximum Doppler spread. . . . . . . . . . 32
4.18 The NMSE of ML method with four speed values, versus SNR. . 33

5.1 The comparison of computational complexity. . . . . . . . . . . . 38


5.2 The comparison of mean value of estimated maximum Doppler
spread in Rayleigh fading model, versus the theoretical maximum
Doppler spread. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
5.3 The comparison of NMSE of estimated maximum Doppler spread
in Rayleigh fading model, versus the theoretical maximum Doppler
spread. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
5.4 The comparison of NMSE of estimated maximum Doppler spread
in Rayleigh fading model, versus SNR. . . . . . . . . . . . . . . . 41
5.5 The comparison of the mean value and NMSE of estimated max-
imum Doppler spread, versus Rician factor. . . . . . . . . . . . . 42
5.6 The comparison of the mean value and NMSE of estimated max-
imum Doppler spread, versus angle of arrival of LOS component. 43

iv
List of Tables

1.1 Tasks division for two persons . . . . . . . . . . . . . . . . . . . . 2

4.1 Pedestrian A Channel Model . . . . . . . . . . . . . . . . . . . . 21

5.1 Parameter setting . . . . . . . . . . . . . . . . . . . . . . . . . . . 34


5.2 Computational Complexity for Level Crossing Rate Estimator . . 35
5.3 Computational Complexity for Hybrid Estimator for low velocity 36
5.4 Computational Complexity for Hybrid Estimator for high velocity 36
5.5 Computational Complexity for Power Integration Estimator . . . 37
5.6 Computational Complexity for ML Estimator . . . . . . . . . . . 37

v
List of Symbols
a Polynomial coefficient vector
am Normalized amplitude of path m
α Angle between the average scattering direction and the mo-
bile direction
β Positive exponent in Power Integration estimator
η0 First zero point of Bessel function
fd Normalized maximum Dopple spread
fˆd Estimated normalized maximum Dopple spread
fD Maximum Dopple spread
fˆD Estimated maximum Dopple spread
Fs Sampling frequency of channel estimate
fshift Doppler shift
γ SNR
Γ Goal function in ML estimator
Γc Concentrated goal function in ML estimator
h Theoretical channel response
hm Theoretical multipath component channel response
hLOS Theoretical LOS component channel response
In nth order modified Bessel function of the first kind
J0 Zero order Bessel function of the first kind
K Rician factor
kmin Smallest value make autocorrelation function is negative in
the Hybrid estimator
κ Beam width
L Maximum lag in autocorrelation functions
Lc Level in the LCR estimator
Lnc Normalized level in the LCR estimator
M Number of independent paths with same delay
N Number of channel estimate samples
Nfft Number of FFT point
Np Number of pilot symbols
Npc Number of positive crossing
P̂y Estimated PSD
φ0 Phase of the LOS component
φm Phase of path m
ψ Threshold in Power Integration estimator
Rh Autocorrelation functions of the theoretical channel
response
Ry Autocorrelation functions of the channel estimation samples
Ry 00 Second derivative of autocorrelation functions
Ryq Autocorrelation functions for finger q
Ŝy Estimated power spectrum
σh2 Power of theoretical channel response
σn2 Power of noise
σy2 Power of channel estimate
T Threshold in the Hybrid estimator
Tmeas Measurement time in the LCR algorithm

vi
Ts Sampling period of channel estimate
τ0 First zero crossing point of the autocorrelation functions
θ0 Angle of the LOS component
θm Angle of path m

vii
List of Abbrevations

AFC Automatic Frequency Control


AGC Automatic Gain Control
AOA Angle Of Arrival
BCL Baseband Core Library
DFT Discrete Fourier Transform
DPCCH Dedicated Physical Control CHannel
DPDCH Dedicated Physical Data CHannel
DSP Digital Signal Processor
FBI FeedBack Information
FFT Fast Fourier Transform
LCR Level Crossing Rate
LOS Line Of Sight
ML Maximum Likelihood
NB Normalized Bias
NMSE Normalized Mean Square Error
NSD Normalized Standard Deviation
PDF Probability Distribution Function
PSD Power Spectral Density

SNR Signal to Noise Ratio


TFCI Transport Format Combination Indicator
TPC Transmit Power Control
UE User Equipment
WCDMA Wideband Code Division Multiple Access
ZCR Zero Crossing Rate

viii
Chapter 1

Introduction

In this chapter, a brief overview of this thesis report is given. The motivation
including the general description of the research field and the previous work
are presented. The main objective and methodology are discussed prior to the
thesis outline.

1.1 Motivation
In mobile cellular systems, the movement of the User Equipment (UE) or the
surrounding objects leads to time variations of the channel. Having an accu-
rate estimation of the maximum Doppler spread, which indicates the rate of
channel variation, is of great importance in many applications such as adaptive
transmission, channel assignment scheme, handover, power control, etc [1] [2].
Therefore the Doppler spread estimation has draw much attention in the
research communities. Up until now, many different Doppler spread estimation
algorithms have been proposed in the literature. Based on their principle, they
can be classified into four types: crossing rate based estimators [1] [3] [4], cor-
relation based methods [5] [6] [7], power spectrum based approaches [8] [9], and
maximum likelihood techniques [10] [11] [12].
Most of the references have been focusing on designing new algorithms and
give the simulation results with an ideal channel model. Whereas in [13], several
algorithms are compared, but only for two types of estimators. However, it is
very interesting and practical to give a comparison study to evaluate four classes
of algorithms with the same simulation environment. Moreover, the possibility
of implementation is another important aspect that should be considered.

1.2 Objective
This thesis aims at evaluating and comparing several algorithms for estimating
the Doppler spread in Wideband Code Division Multiple Access (WCDMA)
uplink systems. The investigation shall take both the performance and the
ability of implementation into consideration. The thesis is done by two persons
during 2014 at Ericsson Lindholmen, Gothenburg.
From the investigation, the following three research questions can be an-
swered:

1
1. Which algorithm is the most accurate one in the same simulation
framework?
2. If the simulation results are inaccurate, what is the reason causing the
inaccuracy for each investigated algorithms?
3. Considering the computational complexity of each algorithm, which
one has the largest practical value for implementation?

1.3 Methodology
In order to achieve the objective of the thesis, the whole work is divided into five
steps: literature investigation, algorithm selection, implementation of selected
algorithms, performance comparison, and the analysis of computational com-
plexity. Since the thesis is for two persons, the detailed work division is shown
in Table 1.1. Noting that the implementation of algorithms are done parallel
for two persons, and for the complexity analysis, each person focuses on one’s
own investigated algorithms.

Ziqi Peng Hui Wen


(This Report) (See [14])
Literature Investigation X X
Algorithm Selection X X
LCR Estimator X
Hybrid Estimator X
Power Integration Estimator X
Frequency Domain ML Estimator X
ZCR Estimator X
Moser’s Estimator X
PSD Slope Estimator X
Time Domain ML Estimator X
Complexity Analysis X X
Performance Comparison X X

Table 1.1: Tasks division for two persons

After the literature investigation, we found that the existing Doppler spread
estimation algorithms can be classified into four categories as described in Sec-
tion 1.1. In order to do the research in a wide range of algorithm types, we
decided to choose two algorithms from each type of Doppler spread estimator,
and four algorithms are allocated to each person. The benefits of this selection
scheme is that the algorithms can be compared through not only different types
but also different approaches in the same type. Thus four types of estimators
can be compared with each other, and the two algorithms of same type can be
compared as well. Moreover, the metric of algorithm selection is based on the
possibility of implementation in Baseband Core Library (BCL) and WCDMA,
and the reliability of the algorithms, such as how many times the article has
been citied or referenced by other articles, how is the comment on the method
from other articles. The selected algorithms are listed in the Table 1.1. In
this report, the LCR estimator, the hybrid estimator, the power integration
estimator and the frequency domain Maximum Likelihood (ML) estimator are

2
presented.
Moreover, this thesis research focuses on the evaluation and comparison from
two perspectives: the performance and the computational complexity. In order
to answer the first two research questions, the algorithms should be implemented
in the same simulation framework, with same parameter setting, such as Signal
to Noise Ratio (SNR), how many simulated frames are tested, etc. For the
evaluation part, the Rayleigh fading channel model is applied, and only affected
by two main factors, i.e. the velocity and the SNR. Here the velocity is consid-
ered from walk speed to high vehicle speed, i.e. 2.7-162 km/h, corresponding
to theoretical maximum Doppler spread of 5-300 Hz. The effect of noise is also
taken into account with SNR setting of 0-15 dB for various level of speeds.
For the comparison part, another four estimation algorithms implemented
by my partner [14] are also introduced. All the algorithms are implemented in
BCL by C++. In this part, both Rayleigh fading and Rician fading channel
are applied. The reason is that some previous research only consider the ideal
Rayleigh fading channel, thus it is interesting to evaluate these estimators in
Rician fading model as well. The Rician fading model is closer to the real radio
channel and it is helpful to test the effect of more factors. In the Rician fading
case, we concentrate on assessing the effect of Rician factor K with K from 0
to 8 and the Angle Of Arrival (AOA) of Line Of Sight (LOS) component with
θ0 = 0◦ , 20◦ , 40◦ , · · · , 180◦ .
Finally, the computational complexity is analysed as the criteria to evaluate
the ability of implementation. The number of multiplication operations required
for each algorithm is considered as main metric to evaluate the algorithmic
complexity.

1.4 Outline
The rest of report is organized as follows: Chapter 2 gives a background in-
cluding some concepts related to the propagation and Doppler spread. Channel
models for Rayleigh and Rician fading are given as well. In Chapter 3, several
Doppler spread estimators are described. The simulation environment and the
estimation results of each algorithm are presented individually in Chapter 4.
In Chapter 5, we give the analysis of the computational complexity of all al-
gorithms, along with the performance comparison based on the four described
estimators together with another four estimators implemented by my thesis
partner [14]. Finally, we summarize some conclusions and discuss the social and
ethical aspect and the future work directions in Chapter 6.

3
Chapter 2

Background

In this chapter, the background for the thesis research is presented. First,
the propagation in a radio channel is introduced. Then the concepts of the
Doppler effect and Doppler spread are described. Next, the channel models
of Rayleigh fading and Rician fading as well as some channel properties are
presented. Finally, we provide the structure of the uplink DPCCH in WCDMA.

2.1 Propagation Phenomena


In general, the transmitted signal is affected by the propagation phenomena
during its travel from the transmitter antenna to the receiver antenna. These
effects depend on the distance between the transmitter and receiver and the
environment around the path, such as buildings and trees, see Figure 2.1. The
three major propagation phenomena are path loss, shadow fading, and multi-
path as introduced below.

Figure 2.1: The propagation environment.

4
2.1.1 Path Loss
Since the transmitted signal is spread spherically, the energy attenuates with
respect to the distance. This loss of energy is called path loss. In the free space
LOS channel, where no obstacles lie between the transmitter and the receiver
or around the path between them, the power falls off proportional to square of
the range. But due to the reflection from the ground, there is not only the LOS
path for long term propagation. In this case, the reflected waves may reduce
the received power and result that the power falls off in proportional to fourth
power of the distance [15]. These two cases are shown in Figure 2.2.

Free Space Loss Plane Earth Loss

Figure 2.2: The pass loss model.

2.1.2 Shadow Fading


Shadow fading is caused by the obstacles between or around the transmission
paths. These obstacles change the direction of the transmitted signal or even
lose the signal through absorption, reflection, scattering, and diffraction.

2.1.3 Multipath

Direct Signal

Reflected Signal

Combined Signal

Figure 2.3: Multipath fading.

5
Multipath phenomenon is caused by the reflection or refraction from the
terrestrial objects as mentioned above. As a result, the receiver receives signals
from different path, with different delay and angles of arrival. In addition,
the combination of the direct wave and the out-of-phase reflected wave yields
attenuated signals as shown in Figure 2.3, this may result in the well known
Rayleigh fading or Rician fading.
See Figure 2.4 for a presentation of the effect of multipath. Assuming a
Dirac pulse is transmitted, the received signal is a sum of pulses with different
amplitudes at different times. The channel response can be expressed as
N
X
h(t) = ρn ejφn δ(t − τn ), (2.1)
n=1

note that ρn ejφn represents the signals received from different paths have dif-
ferent amplitude and phase.

ℎ(�) ℎ(�)
�1

�2

��

� �1 �2 �� �
transmitted signal received signal

Figure 2.4: The effect of multipath propagation.

2.2 Doppler Effect and Doppler Spread


2.2.1 Doppler Effect
The Doppler effect is the change in frequency of the signal due to the relative
motion between the transmitter and the receiver. It means when a UE (or
reflectors in its environment) is moving, a frequency shift occurs at the received
signal. This shift is known as the Doppler shift. The relationship between the
Doppler shift and the velocity v of transmitter or receiver is given by
v
fshif t = fc cos θ, (2.2)
c
in Hz, where c is the speed of light, fc is the carrier frequency, and θ is the
angle difference between direction of the transmitted signal and the speed, as
illustrated in Figure 2.5. When the movement has the same direction as the
direction towards the receiving antenna, then Equation (2.2) reduced to
v
fshift = fc . (2.3)
c

6
θ
v

Figure 2.5: Illustration of Doppler effect.

2.2.2 Doppler Spread


As described in Section 2.1.3, signals from different transmission paths are re-
ceived at the receiver due to the multipath phenomenon. The signals propagate
through different paths with different angles which give rise to different Doppler
shifts. This difference in Doppler shifts contributes to the spread in the spec-
trum. The spectrum is spread to the range of [fc − fD , fc + fD ], where fD
is known as the maximum Doppler spread, which is also called the maximum
Doppler frequency in some papers, corresponding the maximum speed v with
θ = 0◦ :
v
fD = fc . (2.4)
c

2.3 Channel Model


2.3.1 Channel Response
The complex base band channel impulse response h(t) can be modelled as the
combination of a diffuse component and a LOS component [16] [17]
s s
σh2 Kσh2
h(t) = hm (t) + hLOS (t), (2.5)
K +1 K +1

where σh2 is the average power of received signal, K is Rician factor describing
the ratio of the LOS power to the scattering power.
Here, the term finger is used to denote a sum of the path responses with
the same delay. In a finger, the channel response for the scattered and the LOS
component are given by [18] [19]
M
1 X
hm (t) = lim √ am ej(2πfD t cos θm +φm ) (2.6)
M →∞ M m=1
and
hLOS (t) = ej(2πfD t cos θ0 +φ0 ) , (2.7)
where M is the number of independent paths, fD is the maximum Doppler
spread, {am }Mm=1 are the normalized amplitude of path that satisfy limM →∞
PM
√1 2 M
M m=1 m | = 1, {θm }m=1 are identically distributed arrival angles of the
|a
paths, {φm }M m=1 are the uniformly distributed path phase, and θ0 and φ0 are
deterministic constants representing the angle and phase of the LOS component,
respectively.

7
Note that when M is large, hm (t) is a zero-mean unit-power complex Gaus-
sian process by the central limit theorem. The envelop |h(t)| has a Rician
Probability Distribution Function (PDF), which reduces to the Rayleigh distri-
bution when there is no LOS component (K = 0).
In wireless transmission systems, a frequency offset f0 between the terminal
oscillator and the base station oscillator affects the channel response. It can
have a great impact to the Doppler spread estimation. The channel response
including the frequency offset is given by
s s 
σ 2 Kσ 2
h(t) =  h
hm (t) + h
hLOS (t) ej2πf0 t . (2.8)
K +1 K +1

To mitigate the effect of the frequency offset, the Automatic Frequency Control
(AFC) and the frequency offset compensation are used. The AFC responsible
for keeping the frequency offset between ±50 Hz, and the remaining frequency
offset will be reduced by frequency offset compensation.
Moreover, in order to capture the effects of directional scattering, a para-
metric model for the non uniform AOA distribution can be applied. It is called
von Mises distribution:
1
p(θ) = eκ cos(θ−α) , (2.9)
2πI0 (κ)

where In (κ) is the nth order modified Bessel function of the first kind, κ is
the beam width and α denotes the angle between the average scattering direc-
tion and the mobile direction. More detailed explanation can be found in [16]
[20] [21].

2.3.2 Autocorrelation Functions


Without loss of generality, the following derivation will be based on the Rician
fading channel model with frequency offset and non isotropic distribution of
AOA. The autocorrelation functions of the channel respond is given by [20]:

Rh (τ ) = E[h(t)h∗ (t − τ )]
σh 2 Kσh 2 (2.10)
 
= Rhm (τ ) + RhLOS (τ ) ej2πf0 τ ,
K +1 K +1

with p 
J0 −κ2 + (2πfD τ )2 − 4jκ cos(α)πfD τ
Rhm (τ ) = (2.11)
I0 (κ)
and
RhLOS (τ ) = exp(j2πfD τ cos(θ0 )), (2.12)
where J0 (·) is the zeroth order Bessel function of the first kind. Note that
Equation (2.10) is complex valued in general, but it is real when K = κ = f0 = 0,
in which case the AOA of scatted component is omnidirectional and uniform,
no frequency offset, and results in the Rayleigh model with [16] [8]

Rh (τ ) = σh 2 J0 (2πfD τ ). (2.13)

8
Moreover, the zeroth order Bessel function of the first kind J0 (z) is given by
Z 2π ∞
1 ix cos τ
X (−1)k  x 2k
J0 (x) = e dτ = . (2.14)
2π 0 (k!)2 2
k=0

2.3.3 Power Spectrum


The Power Spectral Density (PSD) of channel response can be calculated by the
Fourier transformation of the correlation functions, which is also called Doppler
spectrum, which is given in [17] and [20]:

σh 2 Kσh 2
Ph (f ) = Phm (f − f0 ) + Ph (f − f0 ) (2.15)
K +1 K + 1 LOS
with q
exp( κf cos(α)
fD ) cosh(κ sin(α) 1 − ( ffD )2 )
Phm (f ) = q (2.16)
πI0 (k) 1 − ( ffD )2
and
PhLOS (f ) = δ(f − fD cos(θ0 )). (2.17)
When K = κ = f0 = 0, the spectrum has a high density at plus and minus
the maximum Doppler spread, which can be written as
1
Ph (f ) = σh 2 q . (2.18)
π 1 − ( ffD )2

It is called the Klarke’s or Jakes’ spectrum.

2.4 Structure of DPCCH in WCDMA Uplink


Since the investigation in this report focuses on the WCDMA uplink, the back-
ground about the uplink dedicated channel is presented in this section.

Figure 2.6: The structure of the uplink DPDCH and DPCCH.

The structure of the uplink dedicated channels are shown in Figure 2.6. It
consists of the DPDCH and the DPCCH, which are in-phase (I)/ quadrature

9
(Q) code multiplexed. The DPDCH carries higher layer information includ-
ing the user data, whereas the DPCCH is used to carry physical layer control
information.
Each frame of length 10 ms is divided into 15 slots. There are 10 bits
per DPCCH slot and 10 to 640 bits per DPDCH slot. Each DPCCH slot has
four fields to be used for pilots bits, the Transport Format Combination Indi-
cator (TFCI), the FeedBack Information (FBI) and the Transmit Power Con-
trol (TPC). The pilot bits are used for the channel estimation in the receiver and
the TFCI indicates the transport format combination of the transport channels
mapped to DPDCH. The FBI bits are used for transmit diversity and the TPC
bits carry the power control commands. The exact number of bits of different
fields depends on the slot format, which can be found in [22].
In this thesis, the pilot bits are used to estimate the channel response as
the input of all the selected algorithms, the detailed estimation is given in next
chapter.

10
Chapter 3

Estimation Algorithms

In this chapter, four Doppler spread estimation algorithms will be described.


All of these methods require the estimation of the channel response as input,
thus the following assumption is applied.
As described in Chapter 2, the pilot symbols in uplink DPCCH are used
for the channel estimation in the receiver. However, it is difficult to use the
continuous pilot symbols with more than one slot length in the Doppler spread
estimation since the pilot bits are not continuously, and only 3 to 8 pilot bits
exist in each slot. Therefore the average received pilot symbols over one slot is
used as one sample of channel response estimation, denoted by
Np
1 X
y[n] = ĥ[n] = yp [i], (3.1)
Np i=1

where yp are the pilot bits in each slot and Np is the number of pilot symbols in
one slot. Then the sampling period Ts for the channel impulse response estima-
tion is equal to the slot length, Ts = 23 ms≈ 6.7×10−4 s. Moreover, assuming the
obtained channel estimation is a block of N samples, i.e. y[0], y[1], · · · , y[N − 1].
And the value of maximum Doppler spread is assumed to be stable for these N
samples. Noting that [·] is used to represent time discrete symbols.
One thing needs to mention here is that all the estimators are designed based
on Rayleigh fading channel.

3.1 Level Crossing Rate Estimator


The level crossing rate estimator calculates the Doppler spread by using the
LCR of the channel response. The LCR is defined as the number of times
that the envelope of channel response crosses a certain level Lc with positive or
negative slope during a time interval Tmeas . It is well known [1] [3] that LCR
is proportional to the maximum Doppler spread. In case of a Rayleigh fading
channel, the relation between LCR and the maximum Doppler spread is given
by [16]
Npc √ 2
LCR = = fD 2πLnc e−Lnc , (3.2)
Tmeas

11
where Npc is the number of positive crossing of the channel response envelope,
Lnc ∈ [0, 1] is the normalized level. The above equation is a function of normal-
ized level Lnc , which is normalized to the root of the average power, i.e.
Lc
Lnc = q PN −1 , (3.3)
n=0 |y[n]|2
N

where y[n] is the channel response estimation and N is the number of channel
estimation samples.

x x x x x x x x x

Figure 3.1: Level crossing example.

It can be calculated from Equation (3.2), the maximum level crossing rate
is obtained for Lnc = √12 . And the author in [3] proposed that the normalized
level equals to one is roughly independent of Rician factor. So both of these two
levels tested for comparison in this report, i.e. one level is to use the square root
of the half average power as level, the other one uses the root mean square value
of the received signal as level. Figure 3.1 shows the envelop of channel response
with two different levels. If take the red line as an example, the green circles
marked as the crossings with positive slope, while the orange crossed represent
the crossings with negative slope.
Then the maximum Doppler spread can be estimated by
( N p
e
pc
Lnc = √12
fˆD = TN meas π
, (3.4)
pc
√e Lnc = 1
Tmeas 2π

the corresponding level Lc is given by


q PN −1
2
n=0 |y[n]| √1

2N Lnc = 2
Lc = q −1
. (3.5)
 PN n=0 |y[n]|
2

N Lnc = 1

12
Here counting the number of crossing is achieved by compare the absolute value
of the channel response samples with the level. If the value of sample smaller
than the level, and the value of next sample is larger than the level, it can be
seen as one time crossing with positive slope. On the contrary, if the value of
sample larger than the level, and the value of next sample is smaller than the
level, then it is considered as a crossing with negative slope.

3.2 Hybrid Estimator


A hybrid Doppler spread estimation algorithm is proposed in [23]. It is based
on autocorrelation functions of the channel estimate, and consists of two steps.
The first step is to determine whether the Doppler spread is “small” or “large”.
The second step is to estimate the exact value of the maximum Doppler spread
using either a curvature method (for small Doppler spread) as proposed in [20]
or a first zero detection method (for large Doppler spread) as proposed in [24].
The reason behind using a two step approach depends on the fact that the
curvature estimator is accurate only at low speeds but cannot achieve enough
accuracy at high speeds, and the first zero detection estimator is more suitable
for high speeds than low speeds.
The following sections will begin with the description of the two estimation
methods in Section 3.2.1 and 3.2.2, then the hybrid estimation method will be
presented in Section 3.2.3.

3.2.1 Curvature Estimator (small Doppler spread)


In [20], the author proposes that the Doppler spread can be calculated as:
s
1 −2Ry 00 [0]
fˆD = . (3.6)
2π Ry [0]
It requires the knowledge of autocorrelation functions Ry and the second
derivative of autocorrelation functions Ry 00 at lag zero, i.e. Ry [0] and Ry 00 [0].
Here Ry [0] can be estimated directly from the channel estimate, and Ry 00 [0] can
be obtained by the polynomial fitting of Ry . The polynomial fitting algorithm
is described as below.

Polynomial Fitting Algorithm


Polynomial fitting algorithm models variables as an nth order polynomial. The
polynomial coefficients can be solved by using the least-squares method.
In the curvature estimator [23], both quadratic and fourth order fitting are
considered. The goal is to minimize
L
X X
arg min |Ry [k] − ai k i |2 , (3.7)
ai
k=1 i

with respect to ai , where i = 0, 1, 2 for a second order polynomial and i = 0, 2, 4


for a fourth order polynomial, and L is the maximum lag of autocorrelation
functions. More detailed explanation can be found in [20].

13
The polynomial fitting algorithm can also be described by a matrix. As-
suming the autocorrelation vector r = [r1 , r2 , · · · , rL ]T , where rk = Ry [k], (·)T
denotes matrix transpose,
P a is coefficient vector with elements ai and H is a
matrix such that Ha = i ai k i for all k from 1 to L. Then Equation (3.7) can
be written as kr − Hak2 .
When the second degree polynomial is applied, define the vector a = [a0 , a1 , a2 ]T ,
H is given by
 
1 1 1
 1 2 22 
H= . . ..  .
 
 .. .. . 
1 L L2
When the fourth degree polynomial is applied, the vector a becomes [a0 , a2 , a4 ]T ,
and H should have the following form instead [23]:
 
1 1 1
 1 2 2 24 
H= . ..  .
 
. ..
 . . . 
1 L2 L4

For these two approximations, the polynomial coefficients vector â is calcu-


lated by [25]

â = (HT H)−1 HT r. (3.8)


Since the nth order derivative of autocorrelation functions can be obtained
by [20]
Ry(n) [0] = n!ân /Tsn , n = 0, 2, (3.9)
only â0 and â2 are needed to estimate the maximum Doppler spread, such that
(3.6) can be simplified to
r
ˆ 1 −â2
fD = . (3.10)
Ts π â0

3.2.2 First Zero Detection Estimator (large Doppler spread)


In [24], first zero detection estimator detects the first zero crossing point of
the autocorrelation functions, and relates this point with the Bessel function to
estimate the maximum Doppler spread.
As described in Section 2.3.2, in Rayleigh fading model, the autocorrelation
functions can be expressed as

Ry [k] = σy2 J0 [2πfD k], (3.11)

where σy2 is the power of channel estimate, J0 is the zeroth order Bessel function
of the first kind. If the Equation (3.2.2) is setted equal to zero, i.e.

Ry [τ0 ] = J0 [2πfd τ0 ] = J0 [η0 ] = 0, (3.12)

14
400

200
Ryy [k]

−200

−400

0 10 20 30 40 50 60 70
k
150

100

( k1 , Ryy [k1 ] )
50

τ 0 /Ts
Ryy [k]

−50
( k2 , Ryy [k2 ] )

−100

−150
8.5 9 9.5 10 10.5 11 11.5 12 12.5
k

Figure 3.2: Illustration of linear interpolation.

where τ0 is the first zero crossing of autocorrelation functions and η0 is defined


as the first zero point of Bessel function, therefore the relation between the
maximum Doppler spread fD and τ0 can be represented by

2πfD τ0 = η0 , (3.13)

η0
. fD =
2πτ0
Substituting η0 = 2.40 and π = 3.14 into Equation 3.2.2 yields
0.38
fˆD ≈ . (3.14)
τ0
Since the autocorrelation function estimated from simulation is discrete in
time, τ0 can be obtained by the linear interpolation as described in the following
section.

15
Linear Interpolation Method
Linear interpolation calculates the zero-crossing point of autocorrelation func-
tions using first two adjacent samples of the Ry that have positive and negative
sign, respectively. A realization of the autocorrelation functions is shown in the
upper figure of Figure 3.2 and the lower one is the enlarged samples around
first zero crossing point (in the red rectangular), where the two blue points
are samples of the autocorrelation function with coordinates (k1 , Ry [k1 ]) and
(k2 , Ry [k2 ]). The straight line represents the linear interpolation between these
two points, and the red dot in the middle corresponds to the first zero crossing
with coordinates (τ0 , Ry [τ0 /Ts ]). With a straight line approximation, the value
of τ0 can be calculated from
Ry [k1 ] − Ry [k2 ]) Ry [τ0 /Ts ] − Ry [k2 ]
= . (3.15)
k2 − k1 k2 − τ0 /Ts

Substituting Ry [τ0 /Ts ] = 0 into (3.15) and solve the equation, gives

(k2 − k1 )Ry [k2 ]Ts


τ0 = k2 Ts + . (3.16)
Ry [k1 ] − Ry [k2 ]

It shows that τ0 can be more accurate if the resolution of autocorrelation


samples around zero crossing points is increased.

3.2.3 Hybrid Estimator


The hybrid estimation algorithm is based on estimated autocorrelation func-
tions to decide which method will be used, i.e. curvature estimator or first zero
detection estimator. So the first step of this method is to estimate the autocor-
relation functions. The real part of the autocorrelation functions is considered,
since the correlation function in Equation (3.2.2) is real [23]. In order to im-
prove the accuracy, the author in [23] suggests to estimate the autocorrelation
functions using all detected fingers instead of only one finger:
N −k
1 X
Ryq [k] = R{yq∗ [n]yq [n + k]}, (3.17)
N − k n=1

where Ryq denotes the autocorrelation functions for path q, k = 0, 1, · · · , L, and


R{·} denotes the operation for taking the real part.
Then sum Ryq [k] for all fingers to form a composite autocorrelation func-
tions: X
Ry [k] = Ryq [k]. (3.18)
q

In order to decide which estimator should be used, the smallest kmin is


needed to obtain such that Ry [kmin ] is negative, and compare it with a prede-
fined threshold T , where T ≤ L.
If kmin > T or if Ry [k] is positive for all k, then choose curvature estimator
as described in Section 3.2.1.
If kmin < T , then choose first zero detection estimator as described in
Section 3.2.2.

16
3.3 Power Integration Estimator
The method proposed in [8], referred to as power integration estimator in the
rest of the paper, is based on a nonparametric estimation of the power spectrum
density. The Doppler spectrum is bandlimited in principle, since the maximum
Doppler frequency is proportional to the finite mobile velocity. However, the
total spectrum also contains noise spectrum in the simulator. Therefore, if
the PSD of the received signal is obtained, it can be expected that a high
percentage of the total power would lie in the frequency range [−fD , fD ]. It
means the integral of the estimated PSD within the frequency range [−fD , fD ]
should occupy a certain percentage of the total power, as shown in Figure 3.3.
In [8], fD is calculated as the minimum frequency fˆD that satisfies
R fˆD
P̂ (f )β df
−fˆD yy
R 1/2Ts > ψ, (3.19)
−1/2Ts yy
P̂ (f )β df

where P̂yy (f ) is the estimated PSD of received signal, β is a positive exponent,


ψ is a threshold that satisfies 0 < ψ < 1. The practical implementation of
the algorithm is described in the following. A periodogram-based method is

-1/2Ts -fD fD 1/2Ts

Figure 3.3: Illustration of power integration estimator.

proposed in [8] to estimate PSD based on a block of observed channel samples.


For a length-N block, this estimator is given by

N −1 2
Ts X
P̂y (f ) = y[n]exp(−j2πf nTs ) . (3.20)
N n=0

In practice, P̂yy (f ) can be evaluated by using a finite number of frequency


points, which can be seen as a sampling of P̂yy (f ). The periodogram estimator

17
350
Continuous PSD
Discrete PSD
300

250

200

150

100

50

0
0 200 400 600 800 1,000 1,200 1,400
Frequency

Figure 3.4: The Continuous and Discrete form of Theoretical Periodogram PSD.

considered in the report is Nfft -point PSD estimate


2
Nfft −1
Ts X
P̂yy [fk ] = y[n]exp(−j2πfk nTs )
N n=0
2
Nfft −1
Ts X
= y[n]exp(−j2πnk/M ) , (3.21)
N n=0

where fk = k/(Nfft Ts ), k = 0, 1, · · · , Nfft − 1. The continuous and discrete form


of periodogram estimated PSD is shown in Figure 3.4.
The Discrete Fourier Transform (DFT) of N -point y[n] can be given by
N
X −1
Y [k] = DF T {y[n]} = y[n]exp(−j2πnk/N ), (3.22)
n=0

which can be computed efficiently using an Nfft -point Fast Fourier Transform
(FFT), and where Nfft is a power of two and larger than the block size N .
In this case, y[n] must be zero-padded with Nfft − N zeros prior to the FFT
processing. Equation (3.21) can be seen as the squared modulus of a Nfft -point
FFT
Ts 2
P̂yy [fk ] = |F F T {y[n], Nfft }| . (3.23)
N
Considering the periodicity property of the periodogram estimator and using
the discrete samples of PSD, Equation (3.19) can be transformed to the following
form. The goal is to find the minimum value of the index p that satisfies
Pp h i
P̂yy [f0 ] + k=1 P̂yy [fk ] + P̂yy [fNfft −k ]
PNfft −1 > ψ, (3.24)
k=0 P̂yy [fk ]

18
and the estimation of Doppler spread fˆD can be obtained by
p
fˆD = . (3.25)
Nfft Ts

3.4 Frequency Domain Maximum Likelihood Es-


timator
An ML Doppler spread estimation algorithm in frequency domain is described
in [10]. The basic idea is to use an estimated spectrum to approximate the
shape of theoretical Jakes’ spectrum. Furthermore, the Whittle approximation
is applied in the likelihood function to estimate the maximum Doppler spread.
For better explanation, the estimated power spectrum from the fading chan-
nel and ideal Jakes’ spectrum model are given as follow.
Based on the channel estimate, power spectrum Sy [fn ] can be estimated by
applying periodogram approach as described in Section 3.3:

P̂yy [fn ]
Ŝy [fn ] = P̂yy [fn ] × Fs =
Ts
1 2
= |F F T {y[n], Nfft }| , (3.26)
N

where fn = Ffks = Nnfft , n = 0, 1, · · · , Nfft − 1 is the normalized frequency with


sampling frequency Fs for an Nfft -point FFT, Ts is the sampling period.
The ideal Jakes’ spectrum is given by [10]

 √ 21 2 0 ≤ f < fD
fD −f
c(f ; fD ) = 1
. (3.27)
√ 2 2
1 − fD < f ≤ 1
fD −(1−f )

Taking the power of channel estimate and noise into consideration, the power
spectrum can be calculated from

SJ [fn ; fd ] = σy2 c[fn ; fd ] + σn2


= σn2 (γc[fn ; fd ] + 1), (3.28)

where fd = fFDs is the normalized maximum Doppler spread, σy2 and σn2 denote
the power of received signal and noise respectively, and the SNR is given by
σ2
γ = σy2 .
n
According to [10], Whittle approximation can be applied to the log-likelihood
when Nfft is large. The ML function proposed in [10] is given by:
Nfft −1 " #
X Ŝy [fn ]
Γ ≈ −Nfft ln π − ln(SJ [fn ; fd ]) + . (3.29)
n=0
SJ [fn ; fd ]

One can see that except for the estimated power spectrum of channel esti-
mates, the knowledge of received SNR γ and noise variance σn2 are also needed

19
in Equation (3.29). From (3.28), the noise variance σn2 can be written as

SJ [fn ; fd ]
σn2 =
γc[fn ; fd ] + 1
Nfft −1 ! (3.30)
1 X Ŝy [fn ]
≈ .
Nfft n=0 γc[fn ; fd ] + 1

Substituting the above equation into the goal function (3.29) and neglecting
constant terms yields the concentrated likelihood function (the detailed deriva-
tion is shown in Appendix A)
 
Nfft −1 Nfft −1
X Ŝy [f n ] X
Γc = −Nfft ln   − ln [γc[fn ; fd ] + 1] (3.31)
n=0
γc[fn ; fd ] + 1 n=0

to be maximized with respect to fˆd , where fˆd ∈ (0, 1/2] is the normalized
frequency.
Finally, the estimation of maximum Doppler spread is obtained by

fˆD = fˆd Fs . (3.32)

The selection of the suitable choice for the parameters in all the algorithms
are discussed in next chapter, and the analysis of their effects are presented as
well.

20
Chapter 4

Simulation Results

In this chapter, the simulation results for four Doppler spread estimation algo-
rithms will be presented. Section 4.1 gives a general description of the simulation
environment. In Section 4.2, the simulation results for each algorithm will be
provided and analysed individually.

4.1 Simulation Environment


The simulation framework is implemented by Ericsson. It consists of user equip-
ment, radio channel and antennas, demodulator and decoder. Since the inves-
tigation in this report focuses on the uplink in WCDMA systems, the proposed
algorithms are implemented in the BCL which is a reference model of WCDMA
uplink baseband processing. The simulation is executed by Linux servers.
Both Rayleigh fading and Rician fading channel model are considered in
this report. The common configuration for these two scenarios is to turn off the
TPC and the Automatic Gain Control (AGC). The number of channel samples
N = 150 is used in each algorithm, and 8000 frames are simulated, the length of
FFT is set to 512 points. As described in the previous chapter, the maximum
value of Doppler spread is assumed to be stable for each 150 slots, which implies
the channel is also stable for each 150 slots. The empirical channel model used
to model the Rayleigh fading channel in the simulation is Pedestrian A” model,
which is specified in ITU-R recommendation M.1225 [26]. The specific delay
and power for each multipath component are listed in Table 4.1. Most of the
simulations in this report only depend on the first tap.

Relative delay Average power


Tap
(ns) (dB)
1 0 0
2 110 -9.7
3 190 -19.2
4 410 -22.8

Table 4.1: Pedestrian A Channel Model

Noting that in this chapter, only the simulation results based on Rayleigh
fading channel are presented. The results under Rician fading scenario will be

21
discussed in the next chapter for comparison.

4.2 Simulation Results


The four estimation algorithms described in this report are evaluated in this
section. The evaluation concentrates on the performance of every estimator and
the selection of optimum parameters used in further comparisons. To investigate
the accuracy of the estimation results over the velocity range of 2.7-162 km/h
(corresponding the theoretical Doppler spread is 5-300 Hz), the mean value and
the Normalized Mean Square Error (NMSE) of estimated maximum Doppler
spread are shown as a function of theoretical Doppler spread with SNR = 15
dB. The reason of choosing mean value and NMSE as the metrics is that the
mean value shows an intuitive estimation result compared to the true value of
maximum Doppler spread, while the NMSE not only indicates the bias and
variation of the estimation results, but also shows the relative error compared
to the theoretical value.

fˆD − fD 2
N M SE{fˆD } = E{( ) } (4.1)
fD
In addition, the effect of the noise is also evaluated for several level of veloc-
ities. Here the value of SNR is from 0 to 15 dB, and fD = 10, 80, 160, 240 Hz is
used in the simulation. The quality of fˆD is mainly measured by NMSE.

4.2.1 Level Crossing Rate Estimator


The simulation result of LCR estimator is presented in this section. In the
simulation, two levels are tested, i.e. Lnc = √12 and Lnc = 1. Since the
estimation result of the Doppler spread is not only related to the number of
crossing, it is also related to the measurement time. It means if the measurement
time is very short, even though the different between the number of cross with
positive slope and negative slope is only one or zero, one different crossing will
results in a large difference in the estimation result. So for each level, not
only the crossing with positive slope, but also the crossing with negative slope,
and the average number of crossings of the two previous cases are considered.
Therefore, there are totally six estimators to estimate the maximum Doppler
spread.

Effect of Speed
Figure 4.1 demonstrates the mean value of the estimated maximum Doppler
spread. There should be six lines in the figure, but they overlap with each other
for the same level, so only two curves can be observed. It implies there is no
big difference between different crossing count methods when the large number
of simulation applies, which shows an agreement of the theory. Moreover, the
curve with normalized level Lnc = √12 is more accurate than the other one,
especially for the low speed values as shown in Figure 4.1 (b).
However, both two estimators have larger bias at low velocities. In this
report, each crossing increased the estimate by 9.3 Hz for Lnc = √12 and 10.8
Hz for Lnc = 1. Comparing to high speed, one more miscount will cause larger

22
percentage of error for low speed. This is one of the reason that why this
method is biased at low velocities. Another reason is that the channel response
for the low velocity will change slowly, it is easier influenced by noise and then
introduces additional crossing.

300

Lnc =1/ 2 posi

100 Lnc =1/ 2 nega

Lnc =1/ 2 mean
Mean estimate maximum doppler spread [Hz]

Mean estimate maximum doppler spread [Hz]


250 Lnc =1 posi
Lnc =1 nega
90 Lnc =1 mean

200

80

150


Lnc =1/ 2 posi 70

Lnc =1/ 2 nega

100 Lnc =1/ 2 mean
Lnc =1 posi
Lnc =1 nega 60
Lnc =1 mean
50
0 50 100 150 200 250 300 6 8 10 12 14 16 18 20 22 24
True maximum Doppler spread [Hz] True maximum Doppler spread [Hz]

(a) Regular speed (b) Low speed

Figure 4.1: The estimated maximum Doppler spread of LCR method with dif-
ferent levels,versus the theoretical maximum Doppler spread.

Figure 4.2 illustrates the NMSE of fˆD over the range of 5-300 Hz. It can be
seen that estimator with Lnc = √12 has lower NMSE, providing more reliable
Doppler spread estimate than the estimator with Lnc = 1. Consequently, only
the mean crossing count method with Lnc = √12 is chosen for comparison.

103 103
√ √
Lnc =1/ 2 posi Lnc =1/ 2 posi
√ √
Lnc =1/ 2 nega Lnc =1/ 2 nega
√ √
Lnc =1/ 2 mean Lnc =1/ 2 mean
102 Lnc =1 posi Lnc =1 posi
Lnc =1 nega Lnc =1 nega
Lnc =1 mean Lnc =1 mean
102
Normalized MSE

Normalized MSE

101

100
101

−1
10

10−2 100
0 50 100 150 200 250 300 6 8 10 12 14 16 18 20 22 24
True maximum Doppler spread [Hz] True maximum Doppler spread [Hz]

(a) Regular speed (b) Low speed

Figure 4.2: The NMSE of LCR method with different levels,versus the theoret-
ical maximum Doppler spread.

Effect of Noise
Figure 4.3 shows the NMSE of estimated maximum Doppler spread with two
different levels of various Rayleigh fading channel. As expected, it is found that
for all scenarios the NMSE becomes lower as the SNR increases, which means
LCR method is sensitive to the noise. In addition, one can draw the same
conclusion with the previous part that LCR method is biased at low velocities.
Moreover, the results with Lnc = √12 always provides better performance than

23
the other one with Lnc = 1 when the SNR is above 11 dB. It can explain why
normalized level to √12 is chosen in the following comparison.

103

102
Normalized MSE

101

100 √
Lnc =1/ 2 fD =10 Hz

Lnc =1/ 2 fD =80 Hz

Lnc =1/ 2 fD =160 Hz

Lnc =1/ 2 fD =240 Hz
10−1 Lnc =1 fD =10 Hz
Lnc =1 fD =80 Hz
Lnc =1 fD =160 Hz
Lnc =1 fD =240 Hz
10−2
0 2 4 6 8 10 12 14
SNR [dB]

Figure 4.3: The NMSE of LCR method with different levels, versus SNR for
four speed values.

4.2.2 Hybrid Estimator


The hybrid estimator uses either curvature method or first zero detection method
for estimation. In the implementation process, some modification for these two
methods are made.
For curvature estimator, when the value under the square root is negative
in Equation (3.10), it’s invalid to calculate the estimated Doppler spread. So in
this case, the result is setted to equal to the previous calculated value. If the
value under the square root is negative at the first time calculation, the initial
estimated maximum Doppler spread value is set to zero. This method is applied
to avoid the potential error.
For the first zero detection estimator, the first zero crossing is required, i.e.
the first negative value point of autocorrelation functions. However, the first
zero crossing point will be large when the speed becomes low, which implies
that if the maximum lag L is too small, the crossing point can not be found. In
order to evaluate the hybrid estimator, the performance of curvature method
and first zero detection method are also presented in the report, over the entire
velocity range. Theoretically when the speed under 30 km/h, corresponding
maximum Doppler spread is 56 Hz, the first negative point of autocorrelation
functions is larger than 10, which is the maximum lag used in the simulation.
In this case, kmin can not be found, and it is impossible to calculate τ0 , then
the estimated maximum Doppler spread by first zero detection is set to be equal
to the previous estimated value. Moreover, if kmin can not be found at the first
time, then set fˆD = 0.
In this report, we fix the parameters to L = 10, T = 7. Applying the
theoretical derivation, if τ0 = 7Ts , the maximum Doppler spread fD ≈ 81 Hz,

24
which implies the switching of algorithms at 81 Hz. In order to compare this
method with the others, we consider only one finger in this algorithm, which is
different compared to [23].
As described in Section 3.2.1, there are two kinds of polynomial fitting.
Therefore, these two approximation approach are discussed as shown in Fig-
ure 4.4. Figure 4.4 (a) shows that the fourth degree approach has a larger
range than the second degree approach. When the maximum Doppler spread
is smaller than 100 Hz, fourth degree gives a good estimation, while both of
these two lines are accurate only when fD < 40 Hz. However, if we focus on
Figure 4.4 (b), where the results for low speed are presented. We find that the
second order approximation is more accurate than the fourth order one at low
velocities. Since the threshold is set to around 81 Hz, fourth degree polynomial
fitting is chosen to use as curvature estimator.

300
degree2
degree4

250
Estimate maximum Doppler spread [Hz]

Estimate maximum Doppler spread [Hz]


20

200

15
150

100
10

50
degree2
5 degree4
0
0 50 100 150 200 250 300 6 8 10 12 14 16 18 20 22 24
True maximum Doppler spread [Hz] True maximum Doppler spread [Hz]

(a) Regular speed (b) Low speed

Figure 4.4: The comparison of two polynomial fitting algorithms.

Effect of Speed
Figure 4.5 illustrates the mean value of estimated maximum Doppler spread
with different speeds for previous described three algorithms. It can be observed
that the hybrid estimator gives an unbiased estimation result. We can see from
Figure 4.5, it chooses curvature estimator when fD ≤ 60 Hz, otherwise it chooses
first zero detection estimator. It confirms switching algorithms in a proper way.
It is worth mentioning that the first zero detection method is not able to give
reasonable result at low velocities, since it results in the initial value as described
earlier.
Again, as shown in Figure 4.6, where the NMSE of estimated maximum
Doppler spread is presented, the Hybrid estimator is a good combination of
these two methods since it always results in lower NMSE.

Effect of Noise
The estimated results for different SNR are shown in Figure 4.7. It shows the
robustness against noise especially for the higher velocity values. As the speed
increases, the performance goes better. However, there is no big difference
between the NMSEs of the medium and high speeds.

25
300 40
degree4
first zero
hybrid
Mean estimate maximum Doppler spread [Hz]

Mean estimate maximum Doppler spread [Hz]


250

30

200

150 20

100

10

50 degree4
first zero
hybrid
0 0
0 50 100 150 200 250 300 6 8 10 12 14 16 18 20 22 24
True maximum Doppler spread [Hz] True maximum Doppler spread [Hz]

(a) Regular speed (b) Low speed

Figure 4.5: The estimated maximum Doppler spread of Hybrid method,versus


the theoretical maximum Doppler spread.

101 102
degree4 degree4
first zero first zero
hybrid hybrid

100 101
Normalized MSE

Normalized MSE

10−1 100

10−2 10−1

10−3 10−2
0 50 100 150 200 250 300 6 8 10 12 14 16 18 20 22 24
True maximum Doppler spread [Hz] True maximum Doppler spread [Hz]

(a) Regular speed (b) Low speed

Figure 4.6: The NMSE of estimated maximum Doppler spread of Hybrid


method,versus the theoretical maximum Doppler spread.

Figure 4.8 shows the comparison between different numbers of received paths
used in the calculation of the autocorrelation functions. It can be observed that
the usage of two paths gives a better performance. It is more obvious for the
high speed values. However, in order to have the same parameter setting with
other algorithms, the single path case is considered in the following report.

4.2.3 Power Integration Estimation


In order to demonstrate the effect of the value of exponent β and threshold ψ,
Figure 4.9 is presented. It shows the NMSE of four speed values with β = 1, 2.
On the one hand, β = 2 provides better estimation performance. The difference
between β = 1 and β = 2 for the same speed becomes smaller as speed increases.
On the other hand, the optimum choice of threshold varies with the actual
Doppler spread. For instance, ψ = 0.75 is the best when fD = 10 Hz, ψ = 0.9
is the best when fD = 240 Hz. For the rest of the simulations, the parameters
are fixed to β = 2, ψ = 0.9 as suggested in [8].

26
100

10−1
Normalized MSE

10−2

fD = 10 Hz
fD = 80 Hz
fD = 160 Hz
fD = 240 Hz
10−3
0 2 4 6 8 10 12 14
SNR [dB]

Figure 4.7: The NMSE of Hybrid method with four speed values, versus SNR.

101
one path
two paths

100
Normalized MSE

10−1

10−2

10−3
0 50 100 150 200 250 300
True maximum Doppler spread [Hz]

Figure 4.8: The NMSE of Hybrid Estimator for different number of used paths,
versus theoretical maximum Doppler spread.

Effect of Speed
Figure 4.10 illustrates the mean value of fˆD under various Doppler spreads. It
provides a reliable and unbiased estimation result. In theory, the high speed
part should have the better performance than low speeds since the chosen value
of parameter is more suitable to high speeds, Figure 4.10 shows good agreement
with it. The enlarged figure for low velocities has a correct tendency but with
variance. It is presented more clearly in Figure 4.11, considering the fact of
accuracy mean value, the reason for the fluctuation of the curves in two figures
is their high variance.
In order to find out the reason of this high variance, the estimation results

27
102

101
Normalized MSE

100

10−1
fD = 10Hz, β =1
fD = 10Hz, β =2
fD = 80Hz, β =1
fD = 80Hz, β =2
10−2
fD = 160Hz, β =1
fD = 160Hz, β =2
fD = 240Hz, β =1
fD = 240Hz, β =2
10−3
0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95
Threshold

Figure 4.9: The NMSE of Power Integration Estimator with β = 1, 2, versus


threshold ψ.

300
Mean estimate maximum doppler spread [Hz]

Mean estimate maximum doppler spread [Hz]

250
20

200

150 15

100

10
50

0
0 50 100 150 200 250 300 6 8 10 12 14 16 18 20 22 24
True maximum Doppler spread [Hz] True maximum Doppler spread [Hz]

(a) Regular speed (b) Low speed

Figure 4.10: The estimated maximum Doppler spread of Power Integration


method, versus the theoretical maximum Doppler spread.

with different number of frames are simulated, as shown in Figure 4.12. As


expected, the usage of longer simulation provides better performance in terms
of mean value. But the NMSE does not give a reasonable result. That is because
of the outlier during the simulation. However, we haven’t been able to give a
reasonable explanation from analysis due to the time constraint. It will be put
into the future work of the investigation.

Effect of Noise
Based on Figure 4.13, one can see that the curves are unstable for four speed
values, but it becomes better as the velocity increases.
If we investigate the Normalized Bias (NB) and the Normalized Standard

28
100 101

10−1 100
Normalized MSE

Normalized MSE
10−2 10−1

10−3 10−2
0 50 100 150 200 250 300 6 8 10 12 14 16 18 20 22 24
True maximum Doppler spread [Hz] True maximum Doppler spread [Hz]

(a) Regular speed (b) Low speed

Figure 4.11: The NMSE of estimated maximum Doppler spread of Power Inte-
gration method,versus the theoretical maximum Doppler spread.

300 101
8000 frames
1000 frames
Mean estimate maximum doppler spread [Hz]

250

100

200
Normalized MSE

150 10−1

100

10−2

50
8000 frames
1000 frames
0 10−3
0 50 100 150 200 250 300 0 50 100 150 200 250 300
True maximum Doppler spread [Hz] True maximum Doppler spread [Hz]

(a) Regular speed (b) Low speed

Figure 4.12: The comparison of the number of frames used in simulation for
Power Integration method, versus the theoretical maximum Doppler spread.

Deviation (NSD), which are given by

NB = E[fˆD /fD − 1] (4.2)


1
2 2
NSD = E[(fˆD /E[fˆD ] − 1) ] , (4.3)

as shown in Figure 4.14. NB reflects the error between the estimated result and
the theoretical value, and NSD indicates the variance or dispersion from the
average value of the estimation. It can be seen that the NB for the low Doppler
spread is degraded when the SNR is below 3 dB, whereas it is approximately
stable for SNR > 3 dB. While the NBs for the medium and high Doppler
spread are close to zero, which implies the errors for this area are at a low level.
Moreover, it is not hard to observe that the shape of curves in NSD is very
similar to that in Figure 4.13. Thus, the dominant factor for the unacceptable
results of Figure 4.13 should be the high variance of the estimation. Comparing
to this high variance, noise does not give a significant effect to the results.

29
102
fD = 10 Hz
fD = 80 Hz
fD = 160 Hz
101 fD = 240 Hz

Normalized MSE
100

10−1

10−2

10−3
0 2 4 6 8 10 12 14
SNR [dB]

Figure 4.13: The NMSE of Power Integration method with four speed values,
versus SNR.

0.4 101
fD = 10 Hz fD = 10 Hz
fD = 80 Hz fD = 80 Hz
Normalized Standard deviation error

fD = 160 Hz fD = 160 Hz
0.3 fD = 240 Hz fD = 240 Hz

100
Normalized Bias

0.2

0.1 10−1

10−2
0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14
SNR [dB] SNR [dB]

(a) Normalized Bias (b) Normalized Standard Deviation

Figure 4.14: The normalized bias and normalized standard deviation of Power
Integration method with four speed values, versus SNR.

4.2.4 Frequency Domain ML Estimator


Among all the factors, the accuracy of estimation results of this algorithm
mainly depends on two factors. The first one is the shape of the estimated
Doppler spectrum. The other factor is the resolution of frequency estimation,
i.e. the interval between two hypothesis values of Doppler spread fd . Theo-
retically more accurate result can be obtained by increasing the resolution, but
the computational complexity will be increased as well. There is a trade-off
between accuracy and computational cost. For instance, if assuming the sam-
pling frequency is set to be 1000 Hz, and 5 Hz is selected as the interval length,
the algorithm needs to calculate the concentrated likelihood function for 200
times. Whereas if 20 Hz is selected, only 50 times calculation are required.
Thus, several resolution values are tested in the simulation in order to get rel-
atively reliable estimation result without increasing too much computational

30
complexity.

Effect of Speed
As mentioned in Section 4.2.1, the estimation resolution of the LCR method is
around 10 Hz for both of two different levels. And the resolution of the power
integration estimator can be calculated from Equation 3.25. The minimum
possible interval can be distinguished by the method is 2.93 Hz. Moreover,
in order to consider a wider range, four values between 2 Hz and 20 Hz are
chosen for evaluation, which are 2 Hz, 5 Hz, 10 Hz and 20 Hz. Figure 4.15
demonstrates the estimation results with four resolutions. It is shown that the
smaller interval does not result in a significant improvement of performance.
One can also observe that the simulation results have a bias around 40 Hz.
Since the error is even larger than the maximum tested resolution value, i.e. 20
Hz, then the conclusion can be drawn that the overestimate is not caused by
the inappropriate selection of resolution.

350
Mean estimate maximum Doppler spread [Hz]

Mean estimate maximum Doppler spread [Hz]

64
300

250 62

200
60

150
resolution=20Hz resolution=20Hz
58
resolution=10Hz resolution=10Hz
100
resolution=5Hz resolution=5Hz
resolution=2Hz resolution=2Hz
50 56
0 50 100 150 200 250 300 6 8 10 12 14 16 18 20 22 24
True maximum Doppler spread [Hz] True maximum Doppler spread [Hz]

(a) Regular speed (b) Low speed

Figure 4.15: The estimated maximum Doppler spread of ML method,versus the


theoretical maximum Doppler spread.

If we turn to consider another factor, see Figure 4.16, where the left plot
shows the estimated spectrum based on the channel response and the right plot
represents the ideal Jakes’ spectrum when fD = 210 Hz. It is clear to see
that the estimated spectrum does not have the ideal shape, which is affected
by noise seriously. Therefore, there are two possible reasons for the bias of
ML estimation result. The first one is the inaccuracy of the estimation for
the Doppler spectrum. This can be improved by increasing the number of
obtained channel estimate or increasing the FFT length. The second reason
is that BCL models the practical environment, and it is not suitable to use
practical spectrum approximate the theoretical one. If the ideal channel model
is used in the simulation, it will result in a more accurate performance.
Figure 4.17 shows the NMSE of the ML estimation with four resolution val-
ues. It provides almost the same results for different resolutions, which implies
that not only the mean value but also the variance of estimation results are
closed to each other. It also confirms the conclusion that the inaccuracy of es-
timation is independent of current selection of resolution. Consequently, 20 Hz
is used in the following simulations.

31
·108
5,000
1.4
Periodogram estimated spectrum

1.2 4,000

Model Jakes spectrum


1
3,000
0.8

0.6 2,000

0.4
1,000
0.2

0 0
0 200 400 600 800 1,000 1,200 1,400 0 200 400 600 800 1,000 1,200 1,400
Frequency [Hz] Frequency [Hz]

(a) Periodogram estimated spectrum (b) Jakes’ spectrum

Figure 4.16: The comparison between estimated spectrum and theoretical spec-
trum

103 103
resolution=20Hz resolution=20Hz
resolution=10Hz resolution=10Hz
2 resolution=5Hz resolution=5Hz
10
resolution=2Hz resolution=2Hz

102
Normalized MSE

Normalized MSE

101

100
101

10−1

10−2 100
0 50 100 150 200 250 300 6 8 10 12 14 16 18 20 22 24
True maximum Doppler spread [Hz] True maximum Doppler spread [Hz]

(a) Regular speed (b) Low speed

Figure 4.17: The NMSE of estimated maximum Doppler spread of ML method,


versus the theoretical maximum Doppler spread.

Effect of Noise
In Figure 4.18, we consider the effect of noise and show the NMSEs of the ML
estimator as functions of SNR. The NMSEs slowly decrease with the increasing
of the SNR except the result at low speed. Specifically, the value of NMSEs tends
to be stable when the SNR is above 8 dB. Thus, this method does not require
very high SNR to guarantee the performance for medium and high velocities.
On the other hand, the NMSE is fluctuate within 20 to 30 when the velocity
is low, the corresponding error is 45 Hz to 55 Hz. It means the estimator is
seriously biased, and the noise is not the dominant reason for the bias.

32
102

101
Normalized MSE

100

10−1
fD = 10 Hz
fD = 80 Hz
fD = 160 Hz
fD = 240 Hz
10−2
0 2 4 6 8 10 12 14
SNR [dB]

Figure 4.18: The NMSE of ML method with four speed values, versus SNR.

33
Chapter 5

Performance Evaluation

In this chapter, the evaluation and comparison of proposed algorithms are car-
ried out . In order to implement the estimator in practical applications, it is
of great importance to consider the computational complexity for each method.
In Digital Signal Processor (DSP), if a certain number of multiplications are
calculated, the same number of accumulations can be computed synchronously
in each clock cycle. In this case, only the number of multiplications or ac-
cumulations is needed to represent the computational complexity. Since the
multiplication is the limitation factor, the number of multiplication operations
is considered as the main metric to access the algorithmic complexity. More-
over, performance comparison under both Rayleigh and Rician fading channel
model are given.

5.1 Computational Complexity


In this section, the computational complexity of four types of Doppler spread
estimator algorithms will be discussed. After a detailed analysis, a summary
table for each method will be presented. The table includes the theoretical as-
sessment and an example case with parameter setting as shown in Table 5.1(the
same as Chapter 4). In order to analyse without loss of generality, the worst
case scenario is considered in this section, i.e. the maximum complexity cost
required for each algorithm.

Symbol Description Value


N Number of channel estimate samples 150
L Maximum lag of Ry , i.e. the length of Ry is L + 1 10
Nfft Number of FFT point 512

Table 5.1: Parameter setting

5.1.1 Level Crossing Rate Estimator


The LCR estimator is directly based on the channel estimate, as it counts the
number of times that the channel estimate crosses a given level. The value of
the level is calculated by Equation (3.5). One thing worth mentioning is that

34
one multiplication operation between two complex valued (i.e. consists of real
part and image part) equals to the complexity of four real valued multiplica-
tion. However, one multiplication operation between the complex value and its
conjugate value costs the complexity of two real valued multiplication. There-
fore in order to calculate the level, 2N + 1 multiplications and one square root
are required. The next step is to compare the channel response to the level
to detect the crossing. Note that in this report, one comparison is considered
as one multiplication. After the number of crossing is obtained, the maximum
Doppler spread can be estimated according to Equation (3.4). Assuming the
constant term is already pre-calculated, the cost for the final estimation is only
one multiplication.
Consequently, the computational complexity for the LCR algorithm consists
of the calculation of level, comparison between the channel response and the level
and the final Doppler spread estimation. The summary is shown in Table 5.2.

Description Multiplication Example Others


Level Calculation 2N+1 301 Square root: 1
Comparison N 150
fˆD 1 1
Sum 3N+2 452 Square root: 1

Table 5.2: Computational Complexity for Level Crossing Rate Estimator

5.1.2 Hybrid Estimator


As described in Section 3.2.3, the hybrid estimator uses either the curvature
method (for “low” Doppler spread) or the first zero detection method (for “high”
Doppler spread). Thus, in this section the computational complexity for two
cases will be analysed: low velocity and high velocity.
However, there are some common cost for these two cases since there is a
judgement to decide which method will be used. The judgement is based on
the estimated autocorrelation functions as shown in Equation (3.17). For lag
k, it requires [4 ∗ (N − k) + 1] times of multiplication operations. If consider
PL
k from 0 to L, it costs k=0 [4 ∗ (N − k) + 1] = (L + 1)(4N − 2L + 1) times
of multiplications for one finger. In [23] the author suggests to combine all
detected fingers (for instance, totally q fingers are detected) to calculate the
autocorrelation functions to increase the accuracy. It will results in q times
as much as the cost for only one finger. But note that only one finger case is
considered in this part in order to compare the complexity with other algorithms
with the same parameter setting. To make the judgement, the smallest kmin
needs to be detected such that Ry [kmin ] is negative, and then compare it with
the predefined threshold T . The worst case is that Ry [k] is positive for all k, it
costs L + 1 comparisons between Ry [k] and 0, and one additional comparison
between kmin and T is required.
For the curvature estimator, Equations (3.8) and (3.10) are applied to esti-
mate the maximum Doppler spread. Assuming (HT H)−1 HT is pre-calculated
in Equation (3.8), then this equation becomes matrix multiplication with size
(3 × 1) = (3 × L) · (L × 1), which costs 3L multiplications to solve this equation.

35
Note the elements in matrix are real valued. After obtaining the vector a, two
multiplications and one square root operation are needed in (3.10).
For the first zero detection estimator, the value of τ0 can be calculated
by using linear interpolation method according to Equation (3.16).One scaling
and one division are required to get τ0 and one multiplication to estimate the
maximum Doppler spread in Equation (3.14). Here one division is considered
as having the same cost as one multiplication. The results are presented in
Table 5.3 and Table 5.4.

Description Multiplication Example Others


Aucocorrelation Function (L+1)(4N-2L+1) 6391
Comparison L+2 12
Matrix Multiplication 3L 30
fˆD 2 2 Square root: 1
Sum (L+1)(4N-2L+5) 6435 Square root: 1

Table 5.3: Computational Complexity for Hybrid Estimator for low velocity

Description Multiplication Example Others


Aucocorrelation Function (L+1)(4N-2L+1) 6391
Comparison L+2 12
Linear Interpolation 2 2
fˆD 1 1
Sum (L+1)(4N-2L+2)+4 6406

Table 5.4: Computational Complexity for Hybrid Estimator for high velocity

5.1.3 Power Integration Estimator


The power integration estimator based on PSD is described in Section 3.3. Ac-
cording to Equation (3.23), PSD is estimated by periodogram method and FFT
operation is applied in this approach. Since FFT can be efficiently calculated by
some DSP chips, the number of FFT operation is listed separately. In order to
calculate one point of P̂yy [fk ], only two multiplications are required for square
operation, since the scaling is cancelled out in Equation (3.24). Thus for Nfft
point PSD, it results in 2Nfft multiplication operations in total. Furthermore, to
find the index p, one division and one comparison with threshold are needed for
one value of p. Considering the worst case, the selection of p should begin from
N −1
1 to b fft2 c, where b·c denotes to find the nearest smaller integer. After the
index p is obtained, only one operation is needed in Equation (3.25) to estimate
the maximum Doppler spread. See Table 5.5 for a summary of the result.

5.1.4 Frequency Domain ML Estimator


The goal of the ML estimator is to maximized Equation (3.31) with respect to
fˆD . Assuming the ideal Jakes’ spectrum is pre-calculated, periodogram esti-
mated Power Spectrum Ŝyy [fn ] has the same cost with estimated PSD as pre-
sented in the previous section. For one loop, the calculation of the concentrated

36
Description Multiplication Example Others
Periodogram Estimation 2Nfft 1024 FFT: Nfft
N −1
Division b fft2 c 255
N −1
Comparison(with φ) b fft2 c 255
fˆD 1 1
N −1
Sum 2 + Nfft + 2b fft2 c + 1 1535 FFT: Nfft

Table 5.5: Computational Complexity for Power Integration Estimator

likelihood function in Equation (3.31) requires (Nfft +1) multiplications and one
logarithm in the first term. Note that the second term can be pre-calculated.
Moreover, the whole computational complexity depends on the resolution
of selected fˆd , i.e. the interval between the two adjacent frequency. If the
normalized frequency resolution is defined as ∆f , since the range of normalized
Doppler spread is fˆd ∈ (0, 1/2], the tested hypothesis of normalized Doppler
spread is fˆd = ∆f, 2∆f, · · · , b 2∆f
1 1
c∆f , the number of loops is b 2∆f c. For
20 1 Fs
instance, if we set ∆f = Fs , totally b 2∆f c = b 2×20 c = 37 times of calculation
are needed. Furthermore, the complexity also depends on another two factors:
the sampling period of the channel estimation Ts and the number of FFT Nfft .
First, the sampling period of the channel estimation directly decides the period
of the spectrum, larger period means larger frequency range in spectrum, which
results in more points need to be calculated under the same resolution. Second,
the number of FFT points can be seen as how many points are used to sample
the PSD from the continuous PSD. For instance, if Nfft = 128, on the one hand,
only 128 points need to be calculated in Equation (3.31) for one loop. On the
other hand, a smaller Nfft may reduce the number of loops. Thus, if we only
consider the perspective of the computational cost, increasing of the sampling
period, decreasing the number of FFT points or increasing the interval between
two adjacent hypothesis frequency can lower the complexity of the frequencies
domain ML estimator.

Description Multiplication Example Others


Periodogram 2Nfft 1024 FFT: Nfft
1 1
Goal Function b 2∆f c (Nfft +1) 18981 Logarithm: b 2∆f c
1
Comparison b 2∆f c−1 36
fˆD 1 1
1 1
Sum b 2∆f + 2c(Nfft + 1) − 1 20042 Logarithm: b 2∆f c
FFT: Nfft

Table 5.6: Computational Complexity for ML Estimator

5.1.5 Summary
In order to give an intuitive comparison, the number of multiplication is used to
represent the total cost of each algorithm. Here (4 × N2 log N ) times multiplica-
tion are used to approximate an N points complex valued FFT operation. One
time comparison, square root, or logarithm is considered to have the same cost

37
as one multiplication. Based on these assumptions, the comparison are shown
in Figure 5.1.

Figure 5.1: The comparison of computational complexity.

As one can see, the LCR estimator has the lowest computational cost among
all algorithms. In contrast, the frequency domain ML estimator has the high-
est complexity of computation due to the large number of loops required to
test the hypothesis values in ML algorithm. Moreover, for the hybrid and the
power integration estimators, the computational cost are relatively low com-
pared to others, with the hybrid estimator showing lower complexity. However,
if the FFT can be calculated efficiently, then the cost for the power integration
estimator can be further reduced.

5.2 Performance Comparison


In this section, the four Doppler spread estimators will be compared together
with another four algorithms, which are implemented in the same simulator
by my partner [14]. The first one is the Zero Crossing Rate (ZCR) estimator.
It is similar to the LCR method since both of them are based on the channel
response directly. The second one is the Moser’s estimator, it depends on the
autocorrelation functions of channel estimate. The third one is the PSD slope
estimator, it detects the slope of PSD which is estimated by Periodogram es-
timator. The last one is the time domain ML estimator, with the objective to
maximize the likelihood function based on the approximation of the correlation
functions. More detailed description can be found in Hui Wen’s report or see
[1] [27] [9] [12], respectively.
For a more comprehensive performance evaluation, both Rayleigh and Rician
fading channel model are considered. Section 5.2.1 discusses the assessment for
the effect of speed and noise. Section 5.2.2 focuses on the effect of the Rician
factor K and the AOA of the LOS component.
Note that in the following figures, four different colours of plots refer to four
types of estimator introduced in this report, and the curves with triangle denote
the estimators implemented by Hui Wen [14].

38
5.2.1 Rayleigh Fading Channel
Effect of Speed
Figure 5.2 illustrates the mean value of fˆD of eight algorithms over the speed
range of 2.7-162 km/h (corresponding theoretical Doppler spread is 5-300 Hz).
Generally, most of the estimators provide reliable estimate for the whole range
of the Doppler spread, except for the frequency domain ML and crossing-based
methods.

350 70
LCR
Hybrid
300 60 Power Integration
Mean estimate maximum Doppler spread [Hz]

Mean estimate maximum Doppler spread [Hz]


ML in Frequency Domain
ZCR
250 50 Moser
PSD Slope
ML in Time Domain
200 40

150 30
LCR
Hybrid
100 Power Integration 20
ML in Frequency Domain
ZCR
50 Moser 10
PSD Slope
ML in Time Domain
0 0
0 50 100 150 200 250 300 6 8 10 12 14 16 18 20 22 24
True maximum Doppler spread [Hz] True maximum Doppler spread [Hz]

(a) Regular speed (b) Low speed

Figure 5.2: The comparison of mean value of estimated maximum Doppler


spread in Rayleigh fading model, versus the theoretical maximum Doppler
spread.

It is clearly seen that both LCR and ZCR algorithms have a high bias at
low velocities, as shown in the lower plot in Figure 5.2. The LCR estimator
gives even worse results than the ZCR estimator in terms of error. However,
Figure 5.3 demonstrates that LCR has a lower NMSE at both low and high
speeds, and the curve of LCR estimator is more smooth, which means LCR is
more stable. The reason is that the results of LCR are based on the number of
crossing count directly, and the count depends on the value of level which is a
function of the received signal’s power. Thus, the level can adapt to the change
of the channel such that provide a stable result. Moreover, it is believed that a
more suitable level can be found to reduce the error after statistical analysis.
Now the autocorrelation functions based algorithms are compared, i.e. the
Hybrid and the Moser’s estimator. It is hard to distinguish the two curves of
the mean estimation values over the entire speeds, since the performance is very
close to each other. So the analysis concentrates on the comparison of NMSEs.
As one can clearly see that the hybrid estimator performs better at low velocities,
it confirms that the hybrid estimator has a wide detection range. And both of
these two algorithms performs good when the speed is high. Furthermore, the
autocorrelation based algorithms provide the best estimation results among all
methods in the current scenario.
Next, the performance of the PSD based algorithms is taken into considera-
tion. The difference between the power integration estimator and the PSD slope
estimator is that the former one integrates over entire PSD and the latter one
focus on the peak slope of PSD. Even though the curve of NMSE for the power
integration method at low speed values is very unstable, it is still lower than

39
103 103
LCR
Hybrid
Power Integration
102 ML in Frequency Domain 102
ZCR
Moser
101 PSD Slope
ML in Time Domain
Normalized MSE

Normalized MSE
101

100

100 LCR
10−1 Hybrid
Power Integration
ML in Frequency Domain
10−1 ZCR
10−2 Moser
PSD Slope
ML in Time Domain
10−3 10−2
0 50 100 150 200 250 300 6 8 10 12 14 16 18 20 22 24
True maximum Doppler spread [Hz] True maximum Doppler spread [Hz]

(a) Regular speed (b) Low speed

Figure 5.3: The comparison of NMSE of estimated maximum Doppler spread


in Rayleigh fading model, versus the theoretical maximum Doppler spread.

the NMSE of the PSD slope method. For the whole velocity range, the PSD
slope provides a smooth and relatively stable result comparing to the power
integration estimator.
Finally, two ML based estimators are compared. It is not difficult to see
that the time domain ML method gives much more accurate results than the
frequency domain ML method. As mentioned earlier, the possible reason are
the mismatch between the received spectrum and the Jakes’ spectrum or the
instability of received spectrum. Whereas, the time domain ML estimator is
based on the approximation of the autocorrelation functions. Hence, it is can
be seen that the autocorrelation is more reliable for the estimation in practice.

Effect of SNR
Figure 5.4 demonstrates the NMSE of all the algorithms with different received
SNRs for four Doppler spread values. In general, for the low and medium
Doppler spread, autocorrelation functions based methods are robust to noise,
whereas the crossing rate based methods are relatively more sensitive to noise.
The power integration estimator flucturates with the increase of SNR due to
the outlier of simulation. And the PSD slope is relatively independent to noise.
In addition, frequency domain ML is robust to the effect of noise even though
it suffers high bias. On the contrary, the time domain ML shows its sensitivity
to the noise.
When the Doppler spread is high, see (c) and (d) in Figure 5.4, crossing
based estimators are still sensitive to noise, but it can be improved with the
increase of speed. For the autocorrelation based estimator, PSD slope estimator
and frequency domain ML estimator, the performance for low velocities and
high velocities are similar. Whereas, for the power integration method, the
variation of the NMSE decreases when the velocity increases, specifically when
fD = 240 Hz, the NMSE is even lower than that of the hybrid method. It also
demonstrates the robustness against noise. And the NMSEs of the time domain
ML estimator are relatively stable over all values of SNR.
Therefore, from the perspective of robustness to noise, the Hybrid estimator
shows the best performance among all the algorithms.

40
103 101
LCR LCR
Hybrid Hybrid
Power Integration Power Integration
102 ML in Frequency Domain ML in Frequency Domain
ZCR ZCR
Moser Moser
PSD Slope 100 PSD Slope
ML in Time Domain ML in Time Domain
Normalized MSE

Normalized MSE
101

100
10−1

10−1

10−2 10−2
0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14
SNR [dB] SNR [dB]

(a) fD = 10 Hz (b) fD = 80 Hz

100 100
LCR LCR
Hybrid Hybrid
Power Integration Power Integration
ML in Frequency Domain ML in Frequency Domain
ZCR ZCR
Moser Moser
10−1 PSD Slope 10−1 PSD Slope
ML in Time Domain ML in Time Domain
Normalized MSE

Normalized MSE

10−2 10−2

10−3 10−3
0 2 4 6 8 10 12 14 0 2 4 6 8 10 12 14
SNR [dB] SNR [dB]

(c) fD = 160 Hz (d) fD = 240 Hz

Figure 5.4: The comparison of NMSE of estimated maximum Doppler spread


in Rayleigh fading model, versus SNR.

5.2.2 Rician Fading Channel


In this section, the simulation results in Rician fading channel are presented.
Here two theoretical Doppler spread values are used in the simulation, i.e. fD =
20 Hz and fD = 120 Hz to represent the low speed scenario and high speed
scenario. The criterion for performance evaluation is the same as above, i.e.the
mean estimation value and the NMSE.

Effect of Rician Factor


Figure 5.5 demonstrates the impact of Rician factor K. K represents the ratio
of the LOS power to the diffuse power. The upper plots are the mean value and
lower plots are the NMSE for 20 Hz (left) and 120 Hz (right), and the black
dashed line represents the theoretical value of Doppler spread.
We will start with focusing on the analysis for the low speed. As one can
see that the frequency domain ML estimator and the LCR estimator are fairly
insensitive to the Rician factor K, with the frequency domain ML estimator
showing much smaller errors within 5 Hz. However, both of these two methods
have very high NMSEs compare to other algorithms due to their large bias to the
actual value. Moreover, the PSD slope estimator also results in a convex curve,
which reflects that the estimation results changes slowly when K is samll, i.e. 0 <
K ≤ 4 and changes rapidly when K is large, i.e. 4 < K ≤ 8. On the contrary,

41
80 250
LCR
LCR
Hybrid
Hybrid
Power Integration
Mean estimate maximum Doppler spread [Hz]

Mean estimate maximum Doppler spread [Hz]


Power Integration
200 ML in Frequency Domain
ML in Frequency Domain
ZCR
60 ZCR
Moser
Moser
PSD Slope
PSD Slope
ML in Time Domain
ML in Time Domain 150

40

100

20
50

0 0
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
Rician factor Rician factor

(a) fD = 20 Hz, Mean value (b) fD = 120 Hz, Mean value

101 100

100 10−1
Normailzed MSE

Normailzed MSE

LCR LCR
10−1 Hybrid 10−2 Hybrid
Power Integration Power Integration
ML in Frequency Domain ML in Frequency Domain
ZCR ZCR
Moser Moser
PSD Slope PSD Slope
ML in Time Domain ML in Time Domain
10−2 10−3
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
Rician factor Rician factor

(c) fD = 20 Hz, NMSE (d) fD = 120 Hz, NMSE

Figure 5.5: The comparison of the mean value and NMSE of estimated maxi-
mum Doppler spread, versus Rician factor.

the rest five algorithms show concave behaviour as Rician factor K increases,
and the estimation results of them are all have a tendency of convergence. The
bias for the autocorrelation based algorithms is relatively most less. Whereas,
the ZCR estimation error change significantly. One thing worth mentioning is
that considering the high bias for the ZCR estimator in Rayleigh fading channel
(K = 0), the ZCR estimation result shows a trend to the true value as the
increasing of K mainly because of the attenuation in Rician fading channel.
Overall, Hybrid estimator has the best performance for the low speed.
On the other hand, see Figure (b) and (d) for the simulation result at high
speed. Frequency domain ML estimator is overestimated since the estimated
power spectrum consists LOS component, which is different from the ideal Jakes’
spectrum. Whereas, all the rest algorithms are underestimated. It also can be
observed that the LOS component affects the simulation results significantly.
Among all the algorithms, the Moser’s estimator performs relatively best due
to its lowest NMSE.

Effect of Angle of Arrival


Figure 5.6 depicts the mean value and the NMSE of eight algorithms over the
AOA range of 0◦ − 180◦ , with Rician factor K = 2, SNR = 15 dB.
As one can observe that almost all the algorithms present the similar per-
formance, which shows that they are affected by the AOA but with different

42
80 250
LCR LCR
Hybrid Hybrid
Power Integration Power Integration
Mean estimate maximum Doppler spread [Hz]

Mean estimate maximum Doppler spread [Hz]


ML in Frequency Domain 200 ML in Frequency Domain
ZCR ZCR
60
Moser Moser
PSD Slope PSD Slope
ML in Time Domain ML in Time Domain
150

40

100

20
50

0 0
0 20 40 60 80 100 120 140 160 180 0 20 40 60 80 100 120 140 160 180
Angle of arrival (degree) Angle of arrival (degree)

(a) fD = 20 Hz, Mean value (b) fD = 120 Hz, Mean value

100
LCR LCR
Hybrid Hybrid
Power Integration Power Integration
ML in Frequency Domain ML in Frequency Domain
100.5 ZCR ZCR
Moser Moser
PSD Slope PSD Slope
ML in Time Domain ML in Time Domain
Normailzed MSE

100 Normailzed MSE 10−1

10−0.5

10−2
0 20 40 60 80 100 120 140 160 180 0 20 40 60 80 100 120 140 160 180
Angle of arrival (degree) Angle of arrival (degree)

(c) fD = 20 Hz, NMSE (d) fD = 120 Hz, NMSE

Figure 5.6: The comparison of the mean value and NMSE of estimated maxi-
mum Doppler spread, versus angle of arrival of LOS component.

levels of effect. The reason is that most of the algorithms are proposed based on
the Rayleigh fading channel model, and are not suitable for the Rician fading
channel. Moreover, it is difficult to design a method that can always perform
accurate in Rician channel, since there are different kinds of variation in Rician
channel.
It can also be seen that the results are symmetrical about 90◦ , where the
worst performance is given by most of estimators. Noting that the NMSE
of frequency domain ML at high speed shows a opposite trend to the other
algorithms, since it has a high bias at 0◦ and the mean value is reduced as the
angle increases to 90◦ , which results in less bias to the actual value, and decrease
of NMSE. The symmetry can be explained by Figure 2.5 and Equation (2.2).
For example, if the car in the figure drive to the opposite direction, the angle
between the drive direction and the signal transmission is larger than 90◦ , which
results in negative frequency shift. However, the absolute values of the Doppler
shift with angle θ and (180◦ − θ) are the same, which means the frequency shift
in the spectrum are with the same value but with opposite direction.
Moreover, according to Equation (2.7), the AOA of LOS component gives
rise to the frequency offset to the channel response except for the case with
θ0 = 90◦ . When θ0 = 90◦ , Equation (2.7) is a real value, it only results in
a Dirac pulse in the spectrum without frequency offset. Theoretically, when
θ0 = 90◦ , more accurate results can be given compared to other values of angle.

43
However, in this thesis, the frequency offset compensation is turned on, which
can reduce the effect of the frequency offset. It might be one of the reasons that
can explain why the estimators have better performance when θ is low.
Furthermore, it can be seen that the hybrid estimator and the LCR estimator
are least sensitive to the effect of AOA of LOS component in term of NMSE at
low speed and high speed, respectively.

44
Chapter 6

Conclusions

6.1 Summary
This report investigates four Doppler spread estimation algorithms. The evalu-
ation and conclusion is presented from the perspective of the performance and
the computational complexity.
The simulation shows that the hybrid estimator demonstrates the best per-
formance among all the estimators under both Rayleigh and Rician models.
The inaccuracy of LCR estimator is caused by two reasons: First, the rapid
variation of the channel leads to the variance of the energy, which will influence
the value of threshold level. As a result, the current level used may not be
optimal. Second, due to the property of the level crossing rate, the number of
crossing is easily affect by noise. For the power integration estimator, the bias
of estimation results in that it is difficult to find the optimal parameters for the
entire range of velocity. Furthermore, the simulation results of the frequency
domain ML estimator show a large bias compared to the theoretical value. This
is because of the mismatch between the estimated power spectrum and the ideal
Jakes’ spectrum.
On the other hand, the analysis and discussion of computational complexity
show that the LCR estimator has the lowest cost. Whereas, due to the large
number of loops required, the computational cost is heavy for the frequency
domain ML. In addition, the hybrid and the power integration have the relatively
lower computational complexity compared to the other estimators.
In summary, with relatively less complexity and high accuracy, the hybrid
estimation algorithm has the largest practical value for implementation in real
applications.

6.2 Social and Ethical Aspects


It is important to consider the possible influence of a research from the social
and ethical aspects. In this thesis, the objective is to compare several Doppler
spread estimation algorithms to find the best algorithm for practical implemen-
tation. The increasing of accuracy can bring benefits to many applications. For
instance, it can be applied to the optimal tuning of system parameters with
respect to changing channel conditions in adaptive transmission systems. It can

45
potentially increase the energy efficiency of data transmission. Economically,
less energy means less cost for operators. On the other hand, consider from the
ethical point of view, the Doppler spread estimation will not involve any ethical
issues if used to improve the estimate of WCDMA baseband algorithms. How-
ever, if the focus is on the estimation value of the Doppler spread, which directly
indicates the speed of the user, then it can be used to collect information about
the behaviour of each individual.

6.3 Future Work


The results of this thesis serves as motivation to continue the research, and
the topic can be further investigated in several directions. First, the simulation
environment can be further extended in order to test more influencing factors.
For instance, applied the von Mises’s distribution to investigate the effect of
nonisotropic scattering, and test different length of channel estimation samples.
Second, one can attempt to increase the accuracy of the proposed algorithms,
such as finding optimum level for the LCR estimator, fixing the problem with
the simulation length of the power integration estimator, or reducing the bias
in the frequency domain ML estimator by modifying the shape of theoretical
Doppler spectrum. Moreover, since the aim of this thesis is to compare and
evaluate several algorithms, the selection of parameter setting has to be rela-
tively suitable to all algorithms. However, the optimum setting can be selected
for each algorithm, such as to turn on the function of TPC, AGC, to adjust
the length of the channel estimation, to combine all the detected fingers, etc.
All this will help to further improve the accuracy of the estimation result and
reduce the computational complexity.

46
Appendix A

Whittle Approximation
Log-likelihood Goal
Function Derivation

In this section, the derivation from Equation (3.29) to Equation (3.31) of the
ML estimator is presented.
To clearly present the derivation, here Equations (3.29) and (3.30) are given
again, respectively:
Nfft −1 " #
X Ŝy [fn ]
Γ ≈ −Nfft ln π − ln(SJ [fn ; fd ]) + , (3.29)
n=0
SJ [fn ; fd ]

and

SJ [fn ; fd ]
σn2 =
γc[fn ; fd ] + 1
Nfft −1 !
1 X Ŝy [fn ]
≈ . (3.30)
Nfft n=0 γc[fn ; fd ] + 1

Expanding Equation (3.29) with the first line of Equation (3.30) to replace
SJ [fn ; fd ] by σn2 (γc[fn ; fd ] + 1) results in
Nfft −1 ( )
X Ŝy [fn ]
Γ ≈ −Nfft ln π − ln[σn2 (γc[fn ; fd ] + 1)] + 2 , (A.1)
n=0
σn (γc[fn ; fd ] + 1)

and then replacing σn2 in the above equation by the relation in second line of

47
Equation (3.30), gives

Γ ≈ − Nfft ln π
  
Nfft −1  Nfft −1 !
X 1 X Ŝy [fm ] 
− ln  (γc[fn ; fd ] + 1)
Nfft m=0 γc[fm ; fd ] + 1
(A.2)
 
n=0
 
Nfft −1 
X Ŝy [fn ] 
−   .
 1 PNfft −1 Ŝy [fm ]
γc[fm ;fd ]+1 (γc[fn ; fd ] + 1)

n=0 Nfft m=0

Use Γ1 , Γ2 , Γ3 to present the three terms in the above equation for simplicity,

Γ1 = −Nfft ln π
  
Nfft −1  Nfft −1 !
X 1 X Ŝy [fm ] 
Γ2 = − ln  (γc[fn ; fd ] + 1)
n=0
 Nfft m=0
γc[fm ; fd ] + 1 
  !  Nfft −1

Nfft −1    NX fft −1  Nfft −1
X 1 X Ŝ [f
y m ]  X
=− ln − ln   − [ln(γc[fn ; fd ] + 1)]
n=0
Nfft n=0

m=0
γc[fm ; fd ] + 1 
n=0
  ! Nfft −1
   Nfft −1
1  X Ŝ [f
y m ]  X
= −Nfft ln − Nfft ln   − [ln(γc[fn ; fd ] + 1)]
Nfft 
m=0
γc[fm ; fd ] + 1 
n=0
 
Nfft −1 
X Ŝy [fn ] 
Γ3 = −  
 1 PNfft −1 Ŝy [fm ]
(γc[f ; f ] + 1) 
n=0 Nfft m=0 γc[fm ;fd ]+1 n d

Nfft −1 " #  Nfft −1 !


X Ŝy [fn ] 1 X Ŝ [f
y m ]
=− / 
n=0
(γc[f n ; f d ] + 1) N fft m=0
γc[fm ; fd ] + 1
 ! Nfft −1 !
Nfft −1
X Ŝy n[f ] X Ŝ [f
y m ]
= −Nfft  / 
n=0
γc[f n ; f d ] + 1 m=0
γc[f m ; fd ] + 1

= −Nfft .
(A.3)

Neglect the constant terms, (A.2) results the concentrated likelihood func-
tion:
 
Nfft −1 Nfft −1
X Ŝy [f n ] X
Γc = −Nfft ln   − ln [γc[fn ; fd ] + 1]
n=0
γc[fn ; fd ] + 1 n=0

as shown in Equation (3.31).

48
Bibliography

[1] Gordon L Stüber. Principles of mobile communication. Springer, 2011.


[2] Michael J Chu and Wayne E Stark. Effect of mobile velocity on communi-
cations in fading channels. Vehicular Technology, IEEE Transactions on,
49(1):202–210, 2000.

[3] Mark D Austin and GL Stüber. Velocity adaptive handoff algorithms


for microcellular systems. IEEE Transactions on Vehicular Technology,
43(3):549–561, 1994.
[4] Hong Zhang and Ali Abdi. Mobile speed estimation using diversity com-
bining in fading channels. In Global Telecommunications Conference, 2004.
GLOBECOM’04. IEEE, volume 6, pages 3685–3689. IEEE, 2004.
[5] Jack M Holtzman and Ashwin Sampath. Adaptive averaging methodology
for handoffs in cellular systems. Vehicular Technology, IEEE Transactions
on, 44(1):59–66, 1995.

[6] Kofi D Anim-Appiah. On generalized covariance-based velocity estimation.


Vehicular Technology, IEEE Transactions on, 48(5):1546–1557, 1999.
[7] Wei Sheng and SD Blostein. Snr-independent velocity estimation in a
rayleigh fading channel. Proc. IEEE ICASSP, Orlando, Florida, USA,
pages 2469–2472, 2002.

[8] Kareem E Baddour and Norman C Beaulieu. Nonparametric doppler spread


estimation for flat fading channels. In Wireless Communications and Net-
working, 2003. WCNC 2003. 2003 IEEE, volume 2, pages 953–958. IEEE,
2003.
[9] Shantidev Mohanty. Vepsd: a novel velocity estimation algorithm for next-
generation wireless systems. Wireless Communications, IEEE Transactions
on, 4(6):2655–2660, 2005.
[10] Aleksandar Dogandzic and Benhong Zhang. Estimating jakes’ doppler
power spectrum parameters using the whittle approximation. Signal Pro-
cessing, IEEE Transactions on, 53(3):987–1005, 2005.

[11] Yuh-Ren Tsai and Kai-Jie Yang. Approximate ml doppler spread estima-
tion over flat rayleigh fading channels. Signal Processing Letters, IEEE,
16(11):1007–1010, 2009.

49
[12] Leonid Krasny, Hüseyin Arslan, David Koilpillai, and Sandeep Chen-
nakeshu. Doppler spread estimation in mobile radio systems. IEEE com-
munications letters, 5(5):197–199, 2001.
[13] Ali Abdi, Hong Zhang, and Cihan Tepedelenlioglu. A unified approach to
the performance analysis of speed estimation techniques in mobile commu-
nication. Communications, IEEE Transactions on, 56(1):126–135, 2008.
[14] Hui Wen. Comparison and evaluation of doppler spread estimation algo-
rithms in wcdma. 2014.
[15] AG Williamson, B Egan, and JW Chester. Mobile radio propagation in
auckland at 927 mhz. Electronics Letters, 20(12):517–518, 1984.
[16] Cihan Tepedelenlioğlu, Ali Abdi, Georgios B Giannakis, and Mostafa
Kaveh. Estimation of doppler spread and signal strength in mobile commu-
nications with applications to handoff and adaptive transmission. Wireless
Communications and Mobile Computing, 1(2):221–242, 2001.
[17] Mingzheng Cao, Hongya Ge, Hong Zhang, and Ali Abdi. Parametric
doppler spread estimation in mobile fading channels. In Military Com-
munications Conference, 2007. MILCOM 2007. IEEE, pages 1–5. IEEE,
2007.
[18] Chengshan Xiao, Yahong R Zheng, and Norman C Beaulieu. Statistical
simulation models for rayleigh and rician fading. In Communications, 2003.
ICC’03. IEEE International Conference on, volume 5, pages 3524–3529.
IEEE, 2003.
[19] Li-Chun Wang and Yun-Huai Cheng. A statistical mobile-to-mobile rician
fading channel model. In Vehicular Technology Conference, 2005. VTC
2005-Spring. 2005 IEEE 61st, volume 1, pages 63–67. IEEE, 2005.
[20] Cihan Tepedelenlioglu and Georgios B Giannakis. On velocity estimation
and correlation properties of narrow-band mobile communication channels.
Vehicular Technology, IEEE Transactions on, 50(4):1039–1052, 2001.
[21] Ali Abdi, Janet A Barger, and Mostafa Kaveh. A parametric model for the
distribution of the angle of arrival and the associated correlation function
and power spectrum at the mobile station. Vehicular Technology, IEEE
Transactions on, 51(3):425–434, 2002.
[22] ETSI Specification. 3gpp ts 25.211 version 7.9.0:physical channels and
mapping of transport channels onto physical channels (fdd). Physical layer
procedures (FDD), 2009.
[23] Oskar Mauritz. A hybrid method for doppler spread estimation [mobile
radio systems]. In Vehicular Technology Conference, 2004. VTC 2004-
Spring. 2004 IEEE 59th, volume 2, pages 962–965. IEEE, 2004.
[24] David Sandberg. Method and apparatus for estimating doppler spread,
2001.
[25] LL Scharf. Statistical signal processing: Detection, estimation, and time
series analysis. 1991.

50
[26] ITU Recommendation. R m. 1225. Guidelines for evaluation of radio
transmission technologies (RTT) for IMT-2000, 1997.

[27] Mario Moser. Doppler shift and spread estimation method and apparatus,
May 30 2006. US Patent 7,054,394.

51

You might also like