Professional Documents
Culture Documents
Master Thesis
Xiaojie Wang
In this thesis, channel estimation techniques are studied and investigated for a novel multi-
carrier modulation scheme, Universal Filtered Multi-Carrier (UFMC). UFMC (a.k.a. UF-
OFDM) is considered as a candidate for the 5th Generation of wireless communication sys-
tems, which aims at replacing OFDM and enhances system robustness and performance in
relaxed synchronization condition e.g. time-frequency misalignment. Thus, it may more ef-
ficiently support Machine Type Communication (MTC) and Internet of Things (IoT), which
are considered as challenging applications for next generation of wireless communication sys-
tems. There exist many methods of channel estimation, time-frequency synchronization and
equalization for classical CP-OFDM systems. Pilot-aided methods known from CP-OFDM
are adopted and applied to UFMC systems. The performance of UFMC is then compared
with CP-OFDM.
Index terms: OFDM, 5G, UFMC, Channel estimation, Channel equalization, timing offset,
carrier frequency offset, synchronization
Kurzfassung
Title page image: Spectrum of Universal Filtered Multi-Carrier with 6 Physical Resource
Blocks, which are filtered by Dolph-Chebyshev filter.
Contents
Acronyms V
Symbols VII
1. Introduction 1
2. Fundamentals 3
2.1. Wireless Channel Aspects . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
2.1.1. Tapped-delay line channel model . . . . . . . . . . . . . . . . . . . 4
2.1.2. Statistical description of the wireless channel . . . . . . . . . . . . . 5
2.1.3. Various wireless channel models . . . . . . . . . . . . . . . . . . . . 9
2.2. Multi-Carrier Modulation Techniques . . . . . . . . . . . . . . . . . . . . . 11
2.2.1. OFDM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.2. UFMC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3. Pilot sequences and LTE frame structure . . . . . . . . . . . . . . . . . . . . 20
2.3.1. LTE frame structure . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.3.2. Pilot sequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3. Time-Frequency Synchronization 23
3.1. The effect of timing and frequency offset . . . . . . . . . . . . . . . . . . . . 24
3.2. Approaches of synchronization . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2.1. Timing offset estimation . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2.2. Joint timing- and carrier-frequency-offset estimation . . . . . . . . . 27
3.3. Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.3.1. Timing offset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.3.2. Joint timing- and frequency offset . . . . . . . . . . . . . . . . . . . 35
4. Channel Estimation 40
4.1. The effect of channel-assisted ISI . . . . . . . . . . . . . . . . . . . . . . . . 40
4.2. Single user channel estimation . . . . . . . . . . . . . . . . . . . . . . . . . 42
4.3. Multi-user channel estimation . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.3.1. LS sliding window estimator . . . . . . . . . . . . . . . . . . . . . . 52
4.3.2. MMSE sliding window estimator . . . . . . . . . . . . . . . . . . . 58
5. Conclusion 61
A. Appendix 62
A.1. Derivation of noise variance in UFMC . . . . . . . . . . . . . . . . . . . . . 62
A.2. Covariance of noise in UFMC . . . . . . . . . . . . . . . . . . . . . . . . . 63
A.3. Derivation of symbol estimates in UFMC . . . . . . . . . . . . . . . . . . . 64
Bibliography 67
Acronyms
QAM Quadrature Amplitude Modulation
IFFT Inverse Fast Fourier Transformation
IDFT Inverse Discrete Fourier Transformation
FFT Fast Fourier Transformation
OFDM Orthogonal Frequency Division Multiplexing
FDM Frequency Division Multiplexing
FBMC Filter Bank Multi-Carrier
UFMC Universal Filtered Multi-Carrier
ICI Inter Carrier Interference
ISI Inter Symbol Interference
AWGN Additive White Gaussian Noise
CFO Carrier Frequency Offset
CP Cyclic Prefix
rCFO relative CFO
MIMO Multiple Input Multiple Output
FIR Finite Impulse Response
PRB Physical Resource Block
LTE Long Term Evolution
MTC Machine Type Communication
MUI Multi-User Interference
UoI User of Interest
LOS Line Of Sight
NLOS Non-Line Of Sight
MS Mobile Station
PDF Probability Density Function
CIR Channel Impulse Response
ACF Auto-Correlation Function
i.i.d independent identical distributed
CTF Channel Transfer Function
WSSUS Wide Sense Stationary Uncorrelated Scatters
PDP Power Delay Profile
rms root mean square
ITU International Telecommunication Union
PAPR Peak to Average Power Ratio
FMT Filtered Multi Tone
CSI Channel State Information
EUTRAN Evolved Universal Terrestrial Radio Access
CAZAC Constant Amplitude Zero Auto Correlation
3GPP 3rd Generation Partnership Project
MU Multi User
SNR Signal to Noise Ratio
MSE Mean Square Error
XCF Cross Correlation Function
TTI Transmission Time Interval
SER Symbol Error Rate
MUI Multi User Interference
LS Least Squares
MMSE Minimum Mean Square Error
Symbols
fc carrier frequency
I0 (·) the modified Bessel function of the first kind with order zero
∆t time difference
∆f frequency difference
BS signal bandwidth
v speed of user
c speed of light
wm raised-cosine window
∆n timing offset
ε normalized carrier frequency offset
εq quantized rCFO
Λ(∆n, ε) the log-likelihood function of OFDM systems under timing- and frequency offset
l p1 , l p2 indexes of pilots
h CIR vector
w AWGN vector
∗ convolution operator
The data transmission over wire-line is mainly corrupted by thermal noise, which is well
known as Additive White Gaussian Noise (AWGN). Usually, the thermal noise can be well
approximated by statistically independent Gauss process. This statistical process is charac-
terized by its variance. Besides thermal noise, data transmission over wireless channel also
suffers from Inter Symbol Interference (ISI), which must be dealt with to satisfy a certain
quality of transmission. The impulse response of the channel is more complicated in wireless
communication systems than that in wire-line communication, due to the multi-path propaga-
tion and fading characteristics.
The transmitted signal undergoes several shadowing, refractions and reflections by various
surrounding obstacles. This phenomenon leads to the multipath propagation nature of wireless
channel. Each path is determined by three factors: delay, attenuation and phase shift. In
general, the discrete time-variant Channel Impulse Response (CIR) h(τ,t) can be written as
[9]
h(τ,t) = ∑ an (t)e− j2π fc τn (t) δ [τ − τn (t)] (2.1)
n
where
an (t) is the attenuation factor for the signal received on the n-th path at time instant t
e− j2π fc τn (t) is the phase rotation of the signal component at delay τ at carrier frequency fc
By taking the Fourier transform of h(τ,t) with respect to the delay τ, the time-varying Channel
Transfer Function (CTF) H( f ,t) in frequency domain can be obtained. In general, the CTF
can be written as [9] Z ∞
H( f ,t) = h(τ,t)e− j2π f τ dτ (2.3)
−∞
Note that the CTF is also time-varying.
In order to investigate the effect of wireless channel, one implementation of fading wireless
channel is so called tapped-delay line channel model. The infinite channel impulse response is
modeled as an Finite Impulse Response (FIR) filter by truncating the taps below a threshold.
Without loss of generality, the first tap is always assumed to have no delay.
Depending on the impulse response of the channel, one can classify the fading channel into
two groups: frequency-flat fading channel and frequency-selective fading channel. In a frequency-
flat fading channel, the signal is almost equally weighted so the ISI can be neglected. In
a frequency-selective channel, ISI occurs due to the overlapping of consecutive transmitted
symbols.
The FIR filter description of multipath fading channel is based on truncating the taps with
small power below a threshold. Thus, (2.1) can be simplified as [10]
h(τ,t) = h0 (t)δ (τ − τ0 (t)) + h1 (t)δ (τ − τ1 (t)) + · · · + hLch −1 (t)δ (τ − τLch −1 (t)) (2.4)
where Lch denotes the number of considered fading channel paths, hi (t) = ai (t)e− j2π fc τi (t) is
the complex valued channel gain of the i-th path including attenuation factor and phase shift
and τi (t) is the delay time of i-th path at the time instant t. Usually, the delay of the first path
is set to be zero τ0 (t) = 0. The FIR filter structure model of the fading channel is illustrated in
Fig. 2.1. Both the complex valued channel gain hi (t) and delay τi (t) for the i-th path are time
varying.
If an arbitrary input signal x(n) is fed into the channel, modeled as an FIR filter, the corre-
sponding output signal y(n) is thus
where ∗ is the linear convolution operator. That results in an output signal length of N + Lch −
1, provided an input signal length of N.
2.1. Wireless Channel Aspects 5
To describe the statistical properties of the channel gain hi (t), two setups are distinguished:
Line Of Sight (LOS) and Non-Line Of Sight (NLOS). If there exists a line of sight (a direct
path, which is not reflected by obstacles) between transmitter and receiver, it is called LOS.
Otherwise, every path is reflected by obstacles at least once and received by the Mobile Sta-
tion (MS). This scenarios refers to NLOS. It is shown in [10] that the amplitude of channel
tap r = |hi (t)| follows Rician distribution and Rayleigh distribution in LOS and NLOS sce-
nario, respectively. The Probability Density Functions (PDFs) of Rician distribution [11] and
Rayleigh distribution [12] are given by
r exp −(r2 +ν 2 ) I rν for r ≥ 0
2 2σ 2 0 σ2
fRician (r|ν, σ ) = σ (2.6)
0 otherwise
and ( h 2 i
r
σ2
exp − 2σr 2 for r ≥ 0
fRayleigh (r|σ ) = (2.7)
0 otherwise
where I0 (·) denotes the modified Bessel function of the first kind with order zero. The Rician
distribution is characterized by two factors ν and σ . Since NLOS scenario is considered in this
thesis, we refer to [11] for further details about Rician distribution. The Rayleigh distribution
is a special form of Rician distribution, if ν = 0 is satisfied. In the following Fig. 2.2, the
PDFs are shown for Rayleigh distribution√ with σ = 1, 2, 3 and 4. The amplitude of a complex
numbered random variable R = |Z| = X 2 +Y 2 is Rayleigh distributed, if X ∼ N (0, σ 2 ) and
Y ∼ N (0, σ 2 ) are independent zero-mean Gaussian processes. Thus, the complex numbered
channel tap hi is the so called circularly-symmetric complex Gaussian distributed, denoted by
C N (0, Pi ). Pi is the power density of the ith channel tap.
By interpreting the CIR and CTF as stochastic process, the autocorrelation function R( f , f 0 ,t,t 0 )
is of interest. The autocorrelation function indicates the channel variation in time and fre-
2.1. Wireless Channel Aspects 6
0.7
σ =1
0.6 σ =2
σ =3
0.5
σ =4
0.4
Prob.
0.3
0.2
0.1
0
0 2 4 6 8 10
Ampltitude
• frequency-flat fading channel, if the signal bandwidth is smaller than the coherence
bandwidth, BS < Bcoh . In this case, all spectral components of the signal are approxi-
mately equally distorted.
• frequency-selective fading channel, if the signal bandwidth is larger than the coherence
bandwidth, BS > Bcoh . In this case, different spectral components of the signal are
distorted with different attenuation and phase shift.
The coherence time Tcoh gives a measure of how fast the channel varies with time. A channel
is said to experience [14]
• fast fading, if the coherence time is less than the symbol time duration, Tcoh < TS . In this
case, the channel fading characteristic changes significantly during the symbol transmis-
sion time.
• slow fading, if the coherence time is smaller than the symbol time duration, Tcoh > TS .
In this case, the channel fading characteristic remains approximately the same during
the symbol transmission time.
Obviously, the correlation function is highly dependent on the velocity v, at which the MS is
moving. Clarke derived the normalized ACF R(∆t) of a Rayleigh fading single tap channel in
[15] as
R(∆t) = J0 (2π fDoppler ∆t) (2.10)
with
v
fDoppler = (2.11)
λ
where J0 (·) is the zero-th order Bessel Function of the first kind and λ is the wavelength of the
signal carrier frequency fc . In [15], Clarke proposed an 2D isotropic scattering model in which
a large number of obstacles are randomly and independently placed around a considered MS.
The derivation in [15] is based on the assumptions:
• large amount of scatters, which are independent of each other (Thus, the central limit
theorem can be applied).
• the attenuation factor ai and phase rotation φi (when the signal is reflected by scatter i)
are independent of that of other scatters j 6= i.
• the phase rotation and scatter angles are independent identical distributed (i.i.d) and
uniformly distributed in [−π, π].
• the velocity of MS should be much smaller than the speed of light v c.
By taking Fourier transformation of the time correlation function R(∆t) with respect to the
time difference ∆t, we get the so called Doppler spectral density function s(ϑ )
where F {·} denotes the Fourier transformation operator. With (2.10) and (2.12), the Doppler
spectral density function s(ϑ ) can be thus calculated as [10]
√ 1
(
, |ϑ | < fDoppler
s(ϑ ) = π fDoppler 1−(ϑ / fDoppler )2 (2.13)
0 otherwise
In the Fig. 2.3, the time correlation function R(∆t) of a single-tap Rayleigh fading channel (in
Fig. 2.3a) and its Doppler spectral density function s(ϑ ) (in Fig 2.3b) are illustrated. It is clear
1 0.08
0.07
0.06
0.5
0.05
R(∆ t)
s(ϑ)
0.04
0.03
0
0.02
0.01
−0.5 0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1 −500 −400 −300 −200 −100 0 100 200 300 400 500
time lag frequency
(a) Time correlation function of a Rayleigh fading single- (b) Doppler spectral density function
tap channel
Figure 2.3.: Time correlation function and its Doppler spectral density function
from Fig. 2.3b that the spectrum is broadened due to the motion of the user.
Analogously, Power Delay Profile (PDP) Ph (τ) can be obtained by taking inverse Fourier
transformation of frequency correlation function R(∆ f ). It is given by
where F −1 {·} denotes the inverse Fourier transformation operator. The PDP Ph (τ) describes
the amount of signal power arrives at the receiver with the delay of τ. Thus, it is straightfor-
ward for the tapped-delay line channel model in section 2.1.1.
where Pi is the expected power of signal at the receiver with the corresponding delay τi . With
(2.14), the frequency correlation function can be easily obtained as
Lch −1
R(∆ f ) = ∑ Pi e j2π∆ f τi (2.16)
i=0
Furthermore,the mean delay τm and root mean square (rms) delay spread τrms can be defined
2.1. Wireless Channel Aspects 9
as the normalized first-order moment and normalized second-oder central moment of the PDP
and they are given by [13] Z ∞
Ph (τ)τdτ
τm = Z−∞
∞ (2.17)
Ph (τ)dτ
−∞
and vZ
u ∞ 2
u −∞ Ph (τ)τ dτ
u
τrms = t Z ∞
u − τm2 (2.18)
Ph (τ)dτ
−∞
Throughout this thesis, a various of wireless channel models are investigated. Thus, we discuss
all the channel models that are investigated in this thesis now.
A simple single-tap Rayleigh fading channel is first of all discussed. The properties of this
single-tap Rayleigh fading channel are thoroughly studied in the previous section. The corre-
sponding CIR is thus simply
h(τ,t) = h0 (t)δ (τ) (2.19)
Obviously, it is a frequency-flat fading channel and it does not cause any ISI. Hence, it is
refer to as flat fading channel throughout this thesis. The complex channel gain h0 follows
Rayleigh distribution as described in (2.7) and its time correlation function is given by (2.10).
Furthermore, its Doppler spectral density function follows the classical Jake’s spectrum, which
is shown in (2.13) and Fig. 2.3.
Second of all, a two-path channel model is introduced. Its CIR is given by
We assume that the complex channel gain h0 and h1 are i.i.d Rayleigh distribution. Further-
more, the expected power of the two paths is the same, i.e. P0 = P1 .
For the purpose of evaluating wireless channels for vehicular users, International Telecommu-
nication Union (ITU) has standardized several channel models for vehicular test environment.
In the following table 2.1, the PDPs are shown for Vehicular A (VEHA) and Vehicular B
(VEHB) channel models [16]. Both of VEHA and VEHB channels consist of 6 taps. But
the PDPs are quite different. VEHA has a significantly shorter delay than VEHB, thus the
coherence bandwidth of VEHB is expected to be smaller. Additionally, a larger ISI will also
be introduced in VEHB channel. Furthermore, the received signal power decreases for VEHA
while the delay increases and the strongest path is with zero delay. However, the strongest
path for VEHB channel has a delay of 300 ns.
2.1. Wireless Channel Aspects 10
Vehicular A Vehicular B
Delay (ns) Relative Path Power (dB) Delay (ns) Relative Path Power (dB)
0 0 0 -2.5
310 -1.0 300 0
710 -9.0 8900 -12.8
1090 -10.0 12900 -10.0
1730 -15.0 17100 -25.2
2510 -20.0 20000 -16.0
Another commonly used channel model is so called exponentially decaying impulse response
model. The power of channel taps Pi decays exponentially with increasing delay. If a total
number of considered path is Lch , then the expected power of each taps is given by [17]
1−α i
Pi = α , i = 0, · · · , Lch − 1 (2.21)
1 − α Lch
where 0 < α < 1 determines the rate of power decaying. The mean delay τm and rms delay
spread τrms are also investigated in [17]. They can be written as
Lch −1
∑i=0 iα i α
τm = Lch −1 i
≈ , Lch α Lch 1 (2.22)
∑i=0 α 1−α
and
Lch −1
∑i=0 (i − τm )α i α 2 Lch
τrms = Lch −1 i
≈ , Lch α 1 (2.23)
∑i=0 α (1 − α)2
Thus, both the mean delay and rms delay spread are determined by the decaying factor α.
All the aforementioned channel models are considered in this thesis. The single-tap Rayleigh
fading is frequency-flat, while all the other channel models are frequency-selective. For the
two-path channel model and exponentially decaying impulse response model, the mean delay
and rms delay spread can be varied by changing the delay time of second path τ1 and the
decaying factor α respectively. Vehicular A and Vehicular B on the other hand have standard-
ized and fixed mean delay and rms delay spread values. Furthermore, the total power of all
the considered channel models is to be normalized in the following.
Lch −1
∑ Pi = 1 (2.24)
i=0
The reason is to keep the total channel gain of all paths constant for fair comparison purpose.
2.2. Multi-Carrier Modulation Techniques 11
Instead of transmitting data over the whole available bandwidth in single carrier systems, the
wideband channel is divided into several subchannels with narrow bandwidth in multi carrier
systems. Each information-bearing subchannel transmits with a lower bit rate over separate
carrier signals [9]. Multi carrier systems have several advantages and disadvantages compared
with single carrier systems. The advantages of data transmission over multi carrier are
• higher spectral efficiency since the spectrum of each subchannel can be overlapped.
Thus, it can be used to achieve higher data rate.
• immunity to frequency-selective fading channels, because the bandwidth of each sub-
channel is so small that CTF can be assumed to be constant. Thus, each subchannel
experiences a flat fading. That allows a simpler equalization at the receiver.
• adaptive power allocation and data modulation per subchannel is possible. If some
subchannels experiences deep fading, the power or data modulation format in that sub-
channel can anticipate accordingly.
The disadvantages include
• higher Peak to Average Power Ratio (PAPR) is to be expected, which raises difficulty
for amplifier.
• sensitive to synchronization errors because the subchannal data must be separated at the
receiver.
The focus of this work is to investigate multi carrier systems because of its advantages, thus
multi carrier systems are discussed in the following.
In this section some basics of two kinds of multi-carrier modulation techniques are introduced.
In the first section 2.2.1, we can get an overview into classical CP-OFDM system. In the last
section 2.2.2, the properties of a novel modulation technique UFMC is analyzed.
2.2.1. OFDM
OFDM stands for Orthogonal Frequency Division Multiplexing, which is a special form of
Frequency Division Multiplexing (FDM). It has been today’s most prominent multi-carrier
modulation technique for broadband communications, being used in standards such as LTE
and WiFi. OFDM is very simple and efficient with Fast Fourier Transformation (FFT) based
modulation and Inverse Fast Fourier Transformation (IFFT) based demodulation. It can rely
on per-subcarrier scalar equalization, which allows to reduce the receiver complexity. Fur-
thermore, a higher spectral efficiency can be achieved by overlapping the partial spectra of the
subcarriers. In OFDM systems, the subchannels are orthogonal to each other. Its main draw-
backs are the high spectral side-lobe levels, due to the rectangular symbol shape in time. This
2.2. Multi-Carrier Modulation Techniques 12
mapping, the frequency domain signal Xk is converted into time domain signal xk by IFFT.
The relation between Xk and xk can be written as follows [18]:
1 N−1
x(k) = √ ∑ X(l)e j2πlk/N , k = 0, 1, . . . , N − 1; (2.26)
N l=0
where N is the total number of subcarriers. Because of multi-path propagation in the reality,
2.2. Multi-Carrier Modulation Techniques 13
a guard interval or Cyclic Prefix (CP), which is a copy of a certain length LCP of the time
domain signal xk , is inserted at the beginning of symbol after IFFT. Thus, the transmitted
OFDM signal yO is of the length of N + LCP and it is given by
(
x(N − LCP + k) if k < LCP
yO (k) = (2.27)
x(k − LCP ) otherwise
The cost for insertion of guard interval is a reduction of spectral efficiency. Afterwards, the
parallel signal is again converted into serial signal and transmitted over a channel. Hence, the
received signal rO , which is also corrupted by thermal AWGN, can be written as
where w(n) denotes the AWGN with zero mean and variance of σn2 . The OFDM receiver
firstly splits the serial signal to parallel and then removes the inserted guard interval, which
may be overlapped with previous OFDM symbol because of the linear convolution between
CIR and the transmitted signal. It is shown that ISI can be perfectly mitigated [9], given that
the maximum time delay of the channel is smaller than the guard interval duration i.e.
In order to get the frequency domain signal, FFT operation is thus performed.
1 N+LCP −1
YO (k) = √ ∑ rO (l)e− j2π(l−LCP )k/N , k = 0, 1, · · · , N − 1 (2.30)
N l=LCP
Given that (2.29) is satisfied, (2.30) can be further simplified using (2.27) and (2.26) as
with WO (k) is the N-point FFT of w(n) [14]. Moreover, the N-point FFT does not change the
statistical properties of w(n). Thus, WO (k) is also zero mean AWGN with the same variance
σn2 . With the above equation (2.31), it is clear that ISI is fully mitigated and the subchannels
experience a flat fading channel thus only a one-tap equalizer is needed at the receiver. If the
CTF at each subcarrier k is known by the receiver, the transmitted symbol X̂(k) can be easily
obtained by
YO (k) WO (k)
X̂(k) = = X(k) + (2.32)
H(k) H(k)
Thus, the estimation of H(k) plays a very important role in data communication systems. In
the end, the estimated symbols X̂(k) are decoded by demapper.
2.2. Multi-Carrier Modulation Techniques 14
2.2.2. UFMC
The system model of UFMC is shown in Fig. 2.5 [6, 8]. The overall bandwidth is divided into
B sub-bands. Each sub-band can be allocated with NB consecutive subcarriers and the sub-
band may correspond to Physical Resource Block (PRB) in LTE. The total number of subcar-
riers is N. A N-point Inverse Discrete Fourier Transformation (IDFT) operation is performed
for every sub-band i to transform the frequency domain signal into time domain. Data sym-
bols are modulated in the allocated subcarrier positions for sub-band i and zeros are padded in
frequency domain in the unallocated subcarrier positions to perform IDFT. Thus, the output
signal after IDFT is N. Then, the output signal xi is filtered by a FIR-filter fi with the length
of L. Hence, the output signal of the subband i is given by
L−1
yi (k) = xi ∗ fi = ∑ fi (l)xi (k − l), k = 0, · · · , N + L − 1. (2.33)
l=0
That results in a symbol length of N + L − 1, due to the linear convolution between xi and
fi . The purpose of introducing a FIR filter to filter each subband is to reduce the out-of-
band radiation. It is well known that the rectangular symbol shape in OFDM is neither well
localized in time nor in frequency. The comparatively high spectral side lobe level of the sinc-
function, i.e. the Fourier transformation of the rectangular symbol shape, causes high out-of-
2.2. Multi-Carrier Modulation Techniques 15
radiation by filtering in UFMC, the waveforms are compared between OFDM and UFMC
with the Dolph-Chebyshev filter mentioned above for one subband in Fig. 2.7. Compared to
OFDM systems, the out-of-band radiation is very significantly reduced in UFMC. Thus, users
2.2. Multi-Carrier Modulation Techniques 16
or devices with a relaxed synchronization condition, i.e. timing and frequency misalignment,
cause less interference to other well-synchronized users in UFMC.
Filtering approach is also used in Filter Bank Multi-Carrier (FBMC) systems. In FBMC sys-
tems, each subcarrier is individually filtered, strongly enhancing robustness against Inter Car-
rier Interference (ICI) effects. However, typical FBMC systems utilize filters, whose length is
multiple times of the number of total subcarriers. Hence, its drawback is its long filter length,
making it disadvantageous for communication in short uplink bursts, as required in potential
application scenarios of 5G systems [1], like low latency communication or energy-efficient
Machine Type Communication (MTC). UFMC can be seen as a generalization of filtered
OFDM and FBMC (in its Filtered Multi Tone (FMT) variant). While the former filters the
entire band and the latter filters each subcarrier individually, UFMC filters subband-blocks,
thus groups of subcarriers. This allows reducing the filter length considerably, compared to
FBMC. Furthermore, QAM is still efficient (in contrast to the FBMC case [3]), making UFMC
compatible to all kinds of Multiple Input Multiple Output (MIMO). UFMC can also rely on
FFT-based receive processing with per-subcarrier equalization. As UFMC is very close in
nature to OFDM, it is also known as Universal Filtered OFDM (UF-OFDM) [5].
After every subband is filtered, all the sub-band signals are added together and transmitted.
The transmitted signal yU is given by
B
yU (k) = ∑ yi (k) (2.34)
i=1
After passing a wireless channel with the the CIR h(k), the received signal rU is then
!
B
rU (k) = h(k) ∗ yU (k) = h(k) ∗ ∑ xi ∗ fi + w(k) (2.35)
i=1
The received signal can be considered firstly to be weighted with a window wm . As an exam-
ple, a possible window can be with the raised-cosine shape
h i
1 k
1 − cos L π m ∈ [0, L2 − 1]
2 2 −1
wm (k) = 1 m ∈ [ L2 , L2 + N] (2.36)
h L i
1 1 + cos k−N− 2 +1 π
m ∈ [ L + N + 1, N + L − 1]
2 L 2
2 −1
The windowing process is not considered in this thesis. We refer to [7] for further results with
windowing approach, advantages of windowing are shown there considering time-frequency
misalignment.
After windowing preprocessing, the received time domain signal is to be transformed into
frequency domain. Since the UFMC symbol has the length of N + L − 1, zeros are padded in
order to perform the 2N-point FFT. Thus, the corresponding frequency domain signal YU is
1 N+L−2
YU (k) = √ ∑ rU(l)e− j2πlk/2N ,
N l=0
k = 0, 1, · · · , 2N − 1 (2.37)
It should be noted that all N + L − 1 samples are used for detection rather than only N sam-
ples are used in OFDM. Lacking of CP, ISI because of multipath fading channel cannot be
completely mitigated like OFDM. If the channel has only a single-tap or ISI and ICI is negli-
gible,the frequency domain signal YU is given as
B
YU (k) = H(k) ∑ X̃i (k)Fi (k) +WU (k) (2.38)
i=1
where X̃i and Fi (k) are 2N-point FFT of xi and fi respectively. If Xi is the N-point FFT of xi ,
X̃i can be written as
k
Xi 2
if k is even
π
X̃i (k) = sin 2 (2m − k) j π (2m−k)(1− N1 ) (2.39)
∑ Xi (m) N sin π (2m − k) · e 2
if k is odd
m∈Si 2N
where Si is a set which contains the subcarrier indexes that are allocated to subband i. Further-
more, Xi can be formulated as
(
0 if k ∈
/ Si
Xi (k) = (2.40)
Xi (k) if k ∈ Si
It is clear from (2.39) that on one hand all even subcarriers contain data symbols and on the
other hand all odd subcarriers contain interferences. Thus, all the signals in odd subcarriers are
dropped as indicated in Fig. 2.5. Furthermore, WU is the 2N-point FFT of AWGN w(k) with
variance σn2 . Thereby, the variance σU2 of WU is slightly enhanced with the factor of N+L−1
N
N +L−1 2
σU2 = σn (2.41)
N
The reason for the noise enhancement is that
√ the scaling factor of 2N-point FFT at the receiver
remains the same as in OFDM, namely 1/ N, while the time domain samples within a receive
window is increased from N to N + L − 1. The mathematical derivation of (2.41) is shown in
A.1. Moreover, the white time domain noise is transformed into a colored noise, i.e. noise is
correlated between different subcarriers. In A.2, the covariance of WU is discussed and it is
2.2. Multi-Carrier Modulation Techniques 18
given by
π∆k
1 sin N (N + L − 1) π∆k
c(l 0 , k0 ) = ·ej N (N+L−2) σn2 (2.42)
N sin π∆k
N
0 0
where l 0 and k0 are even subcarrier indexes and ∆k = l −k
2 . The reason for correlation in
frequency domain is so called Leakage-Effect of FFT, since 2N-point FFT is performed at the
receiver using N + L − 1 time domain samples. In the following Fig. 2.8, the auto-covariance
of the frequency domain noise is shown for OFDM and UFMC. From Fig. 2.8a, it is clear that
sim−UFMC
1 analytical−UFMC
0.8
0.6
0.4
0.2
0
100 200 300 400 500 600 700 800 900 1000
(a) Simulated auto-covariance of noise in frequency do-(b) Simulated and analytical covariance of noise for
main for UFMC and OFDM UFMC
the noise in OFDM has approximately no correlation between different subcarriers, while a
small correlation can be observed in UFMC. Furthermore, the analytical covariance in (A.8)
is compared with simulated covariance for UFMC in Fig. 2.8b. It can be concluded that
analytical results match the simulated covariance quite well.
Now, we come back to the assumption of negligible ISI and ICI. In general, ISI and ICI arise
in UFMC systems if the channel has multi-taps. However, the filter ramp-up and ramp-down
act as soft protection against delay spread of the channel, shown in Fig. 2.9a [8]. The signal
energy of ramp-up and ramp-down due to filtering approach in UFMC is relatively small.
Thus, it can be expected that if the delay spread of channel is less than a certain percentage of
filter length L, the resulting ISI and ICI is not very significant and can be neglected. In contrast
to UFMC, CP is inserted in OFDM systems to provide robustness against delay spread and
timing offset, shown in Fig. 2.9b. As long as the inserted CP length is larger than the channel
delay spread, ISI can be completely mitigated. In chapter 4, further discussion regarding ISI
and ICI under different channel models is provided. As an example, the energy of ISI and
ICI is shown to be below -40 dB in Vehicular A channel model. Practically, it can thus be
neglected since the noise power and other inference power is dominant. Furthermore, the
inserted CP is again removed at the receiver since it contains ISI from previous symbol while
all N +L −1 samples are used for symbol detection. So, UFMC is expected to be slightly more
2.2. Multi-Carrier Modulation Techniques 19
Figure 2.9.: Comparison of signal temporal property between OFDM and UFMC [8]
advantageous in reducing the transmission power of signal and have slightly better spectral
efficiency, compared with CP-OFDM. Since every sub-band is filtered to reduce the out-
of-band radiation, guard subcarriers between sub-bands (used in OFDM to protect against
interference from neighboring sub-bands) can be reduced or are not required any more in
UFMC systems. This will further increase spectral efficiency of UFMC.
If the assumption of single tap channel or negligible ICI and ISI is not hold, the received
frequency domain signal can be derived as (see A.3)
YU (k) = H(k)Fi (k)X̃i (k) − αS (k)X̃i (k) + IICI (k) + IISI (k) +WU (k) (2.43)
Without loss of generality the considered subcarrier k is assumed to be allocated to the sub-
band i. Furthermore,
2.3. Pilot sequences and LTE frame structure 20
IICI (k) is the ICI caused by all other subcarriers at the subcarrier k
The transmitted signal passes a time-varying channel with discussed properties (in section 2.1).
In order to recover the transmitted symbols at the receiver, Channel State Information (CSI)
i.e. CIR or CTF have to be estimated. Besides channel estimation, the synchronization of
timing and frequency mismatch at the receiver is another issue. Transmitter and receiver usu-
ally have no common sense to the timing, at which a transmission starts. Additionally, Car-
rier Frequency Offset (CFO) occurs because of Doppler effect and local oscillator frequency
mismatch. The time-frequency misalignment causes system performance degradation. Thus,
known sequences called pilot sequences or training sequences are transmitted. With these
known sequences, the receiver can estimate the CSI and perform time- frequency synchro-
nization between transmitter and receiver.
In LTE, a radio frame with 10 ms contains 20 slots, shown in Fig 2.10 [20]. Each slot consists
of 7 multi carrier symbols with the time duration of 0.5 ms. A sub-frame with transmission
duration of 1 ms contains two slots i.e. 14 symbols, illustrated in Fig. 2.11 [21]. Every 4-th
multi carrier symbol within a slot is used as pilot sequence and all other 6 symbols are data-
bearing symbols. The insertion of pilot sequences enables the receiver to perform channel
estimation and synchronization, while it reduces the data-rate. There exist various types of
pilot arrangements, in EUTRAN the so called time-spaced all frequency pilot arrangement
is applied. In this arrangement, all the available subcarriers or subchannels are modulated
2.3. Pilot sequences and LTE frame structure 21
with known pilot sequences. No subcarrier is used to transmit data i.e. the symbol is purely
pilot sequences. In other pilot arrangement, information-bearing symbols are superimposed
with pilot sequences, since some subcarriers are used to transmit date while others are used
to transmit pilot sequences. This kind of pilot arrangement is not discussed in this thesis.
Furthermore, the subcarrier spacing is set to be 15 kHz and one PRB contains 12 consecutive
subcarriers.
Pilot sequences, also known as training sequences and reference signals, can be used for chan-
nel estimation and time-frequency synchronization. A well-known pilot sequence is Zadoff-
Chu sequences. The u-th root Zadoff-Chu sequence is defined by
um(m+1)
− jπ NZC
xZC,u (m) = e (2.44)
where NZC is a prime number and denotes the length of the sequence [22]. Zadoff-Chu se-
quences are Constant Amplitude Zero Auto Correlation (CAZAC) sequences, since the cycli-
cal shifted version of this sequence is orthogonal to itself. Furthermore, the cross correlation
2.3. Pilot sequences and LTE frame structure 22
And the length NZC of Zadoff-Chu sequence is chosen to be the largest prime number such that
NZC < NAlloc [22]. Users can thus be separated by different roots of Zadoff-Chu sequences. In
[22], the pilot sequences are divided into 30 groups using different roots. Intended for Multi
User (MU)-MIMO, 3rd Generation Partnership Project (3GPP) also introduces cyclical shift
of the pilot sequence s̄u with the same root u.
(α)
s̄u (n) = e jαn s̄u (n) (2.46)
The linear phase shift in frequency domain creates a cyclic shift in time domain. By doing so,
the orthogonality of Zadoff-Chu sequences is preserved [21].
3. Time-Frequency Synchronization
Before the actual channel estimation starts, the start position of FFT window should be deter-
mined because there is a transmission delay between transmitter and receiver. Besides timing
error, carrier frequency offset is another issue caused by Doppler Effect of moving user and
mismatch of local oscillator at the transmitter and receiver. In the Fig. 3.1, the transmitted
(black and solid box) and received (green and dashed box) signal with corresponding time and
frequency lag are illustrated. Because of the time-frequency mismatch, ICI and ISI are to be
expected which results in a lower value of Signal to Noise Ratio (SNR). Furthermore, it also
causes out-of-band radiation which is an interference to all the neighboring users.
OFDM is very sensitive to CFO because of the comparatively high side-lobe levels of the rect-
angular symbol shape. It is very robust against positive timing offset (signal arrives later than
receiver expected) due to CP as long as the timing offset is within the CP duration. Compared
to OFDM, UFMC can achieve more robustness against time-frequency misalignment with the
FIR filter [7]. The reason for robustness against CFO is because of the reduced out-of-band
radiation. Robustness against timing offset is achieved by soft protection of filter ramp-up and
ramp-down.
Receive side synchronization becomes more important when discarding the closed-loop syn-
chronized operation mode of LTE with the timing advance adjustments. In this chapter,
3.1. The effect of timing and frequency offset 24
pilot-based time and frequency synchronization approach is investigated for both OFDM and
UFMC systems and compared with CP-based synchronization approach (only applicable in
OFDM systems, since no CP is inserted in UFMC systems). In section 3.1, the effect of tim-
ing and frequency offset is discussed for OFDM and UFMC systems. Then, two approaches,
CP-based and pilot-based synchronization, are introduced in section 3.2. The last section 3.3
shows simulation results and compares the performance of synchronization.
The effect of timing and frequency offset in OFDM systems have been thoroughly studied
in [24, 25]. In OFDM systems, positive timing offset (signal arrives later than the receiver
expected), has no effect on the SIR as long as it is within the CP duration. Once the timing
offset exceeds the CP-length or it is negative (signal arrives earlier than the receiver expected),
it starts to degrade the system performance. This non-symmetric effect of timing errors is
shown in [26]. In contrast to OFDM, any timing offset in UFMC systems affect the system
performance, since no CP is inserted. Even with a small timing offset, the orthogonality
between subcarriers is destroyed, which causes ICI. Additionally, the previous (positive delay)
or subsequent (negative delay) symbol introduces ISI. But the filter ramp-up and ramp-down
indicate a soft protection against timing offset, since relatively small energy is contained.
In Fig. 3.2, the Mean Square Error (MSE) of symbol estimates are plotted over relative timing
offset with relative CFO (rCFO) 0.1 (Fig. 3.2a) and 0.5 (Fig. 3.2b) subcarrier spacing. The
−10 −10
−12 −10.5
−14
−11
−16
−11.5
MSE in dB
MSE in dB
−18
−12
−20
−12.5
−22
−13
−24
OFDM
−26 OFDM −13.5 UFMC
UFMC
−28 −14
−0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5 −0.5 −0.4 −0.3 −0.2 −0.1 0 0.1 0.2 0.3 0.4 0.5
relative timing offset relative timing offset
(a) Effect of timing offset in UFMC and OFDM under(b) Effect of timing offset in UFMC and OFDM under
rCFO 0.1 subcarrier spacing rCFO 0.5 subcarrier spacing
Figure 3.2.: System performance comparison between OFDM and UFMC under timing and
frequency offset
relative timing offset and rCFO is normalized to total number of samples per symbol and
subcarrier spacing respectively. The simulation settings are shown in table 3.1. We considered
3.2. Approaches of synchronization 25
Table 3.1.: Simulation settings to evaluate the effect of timing and frequency offset in OFDM
and UFMC systems
a scenario in the simulation that one subband ,User of Interest (UoI), is perfect synchronized
at the receiver, while all 9 other subbands interfere the UoI. From the results, it is clear
that UFMC has a symmetric effect of timing offset rather than the non symmetric effect in
OFDM systems. For further results regarding the effect of timing and frequency offset and the
comparison between OFDM and UFMC, we refer to [7, 27, 28].
To recover the transmitted symbols by performing FFT at the receiver, starting position of FFT
window have to be determined to avoid the effect of ISI. The procedure of finding start position
of FFT window refers to as timing synchronization. In addition, CFO should be compensated
as well to reduce the effect of ICI. Thus, the inserted pilot sequences, intended for channel
estimation, are also utilized for timing and frequency synchronization. It is also known as
pilot-assisted synchronization. Another approach uses the redundant information of CP to
perform synchronization, known as blind synchronization. However, it is only applicable in
OFDM systems since CP is not available in UFMC systems.
In the following Fig. 3.3, we illustrate the timing offset between transmitted and received sig-
nal due to propagation delay. The underlaying timing offset degrades system performance
by introducing ISI and destroying orthogonality between subcarriers. Thus, it should be esti-
mated and corrected at the receiver. The inserted pilot symbols at the transmitter can be also
utilized to estimate the timing offset. Since the pilots are already known by the receiver, the
Cross Correlation Function (XCF) A(k) between the known pilots in time domain s(n) and the
received signal r(n) can be calculated.
where (·)∗ is conjugate complex. The peak position k of the XCF indicates the timing offset.
Consider a Transmission Time Interval (TTI) of one sub-frame in LTE (in section 2.3.1), it
contains two pilot symbols and the time lag between these two pilots are known. In ideal
case, the XCF A(k) should have two peaks with the known sample distance ld . However, the
peak position can be wrongly estimated due to noise and ISI. In the Fig. 3.4, we show the
XCF for a TTI of 2 slots in a UFMC system with SNR 0 dB. We are able to observe two
0.45
0.4
0.35
0.3
Correlation
0.25
0.2
0.15
0.1
0.05
0
0 2000 4000 6000 8000 10000 12000 14000
Sample index
Figure 3.4.: Exemplary XCF in UFMC system with SNR 0 dB and TTI 2 slots
peaks corresponding to the pilot position. In order to increase robustness of the algorithm, we
consider the sum of peak amplitude instead of a single peak for one pilot. The final timing
offset estimator can be written as
In general, if Npilot pilot symbols are contained in the TTI, the estimator is given by
Npilot −1
∆n = arg max ∑ |A (k + (n − 1)ld ) |. (3.3)
k n=0
Obviously, the performance of this method highly depends on the auto-correlation (AC) prop-
erties of the pilot symbol. The well-known Zadoff-Chu sequence (in section 2.3.2) has an
optimal AC-properties, whereby the cyclically shifted versions of the sequence results in a
CAZAC sequence, provided that the length of sequence NZC is a prime number.
After the procedure of timing offset estimation is discussed, a further step is joint timing- and
frequency-offset estimation. Because of the Doppler-Effect and frequency mismatch of local
oscillator, frequency error occurs. In order to extend the timing offset estimator to a joint
timing- and carrier-frequency-offset estimator, hypotheses tests for various frequency offsets
can be performed, shown in Fig. 3.5. First of all, we assume that the normalized CFO ε is
Figure 3.5.: Procedure of joint timing- and carrier-frequency offset estimation via pilots
within the interval -0.5 and 0.5 subcarrier spacing. We quantize this normalized CFO between
-0.5 and 0.5 subcarrier spacing into lq levels. These quantized discrete normalized CFOs form
a set SqCFO . For every CFO from this set εq ∈ SqCFO , the timing offset estimation algorithm
described in previous section is performed and the peak amplitude and estimated timing offset
is recorded. Instead of using the conjugate complex original known pilot sequence in time
domain s∗ (k), the CFO-compensating version of the known pilot s∗ (k) · e− j2πεq k/N is used
3.2. Approaches of synchronization 28
to perform timing offset estimation. Among all the peak values, the largest peak is selected.
Finally, the frequency and timing offset can be roughly estimated by the corresponding indexes
of the largest peak. The estimator can be formulated as
Npilot −1
(ε, ∆n) = arg max ∑ |A (k + (n − 1)ld ) | (3.4)
k,εq n=0
with
A(k) = ∑ s̃∗ (n − k)r(n) (3.5)
n
m+L−1
∗
A(k) = ∑ rO (m)rO (m + N). (3.7)
m=k
Under timing and frequency offset, the log-likelihood function Λ(∆n, ε) can be formulated to
be [29]
Λ(∆n, ε) = |A(∆n)| cos(2πε + ∠A(∆n)) − ρΦ(∆n) (3.8)
where ∠ denotes the argument of a complex number and they are given by
1 m+L−1
|r(k)|2 + |r(k + N)|2
Φ(m) = ∑ (3.9)
2 k=m
3.3. Results 29
SNR
ρ= (3.10)
SNR + 1
They act as a normalization term of the correlation magnitude, since Φ(m) computes the
energy of the CP and this energy is weighted by ρ due to the presence of noise. Furthermore,
they are independent of rCFO ε. Additionally, the maximum of the log-likelihood function
Λ(∆n, ε) is achieved, if
cos(2πε + ∠A(∆n)) = 1 (3.11)
is satisfied. The other terms do not dependent on the rCFO, thus the joint maximization of
log-likelihood function can be separated into two steps, maximizing with respect to ∆n and ε.
Thus the joint Maximum-Likelihood (ML) estimation of ∆n and ε becomes
1
ε̂ML = − ∠A(∆n̂ML ) (3.13)
2π
Notice that the frequency offset is contained in the phase shift of the peak value, so the com-
putational complexity is reduced compared to the two dimensional search approach of pilot-
based method. However, CP is not available in UFMC so that this CP-based ML estimation
can only be used in OFDM systems.
3.3. Results
• the Vehicular A channel model (refer to as VEHA). In this channel model, we take all
effects into account including fading, delay spread and noise.
Furthermore, the performance of pilot-based synchronization is compared between conven-
tional CP-OFDM systems and the novel UFMC systems. Additionally, the CP-based synchro-
nization is also simulated for OFDM systems as a reference to investigate the performance of
pilot-based synchronization approach.
First of all, we evaluate the performance of the pilot-based timing offset estimator, discussed
in section 3.2.1. The basic simulation settings are shown in table 3.2. The used FIR filters are
FFT size Filter/CP Length PRB size No. of PRBs Timing Offset (TO) TTI
1024 74/73 12 10/2 10 samples 2 slots
Dolph-Chebyshev filters with its center frequency shifted to that of every PRB. The side-lobe
attenuation of Dolph-Chebyshev filters are set to be 40 dB. Timing offset is set to be of the
length 10 samples. At the transmitter, one sub-frame, which contains 12 data symbols and 2
pilot symbols, is transmitted. At the receiver, a timing synchronization algorithm described
above is to be performed. The number of allocated PRB for users affects the system perfor-
mance in such a way that it affects the SNR. The larger the number of allocated PRB for a
user, the larger the SNR is for constant power spectral density of data symbols. In order to
show the influence of pilot sequences on the performance of the estimator, we introduce a “all-
ones” pilot sequence in the simulation as a reference. The all-ones pilot refer to a pilot, which
contains only ones in all the allocated subcarrier positions. For comparison purpose, LTE stan-
dard compliant pilots [22] are also evaluated. The user is assumed to move at the speed of 50
kmh, which results in a Jake’s Doppler spectral power density described in section 2.1.2 (not
valid for AWGN-channel). To evaluate the performance of this pilot-based timing estimator,
the PDF of the estimated timing offset is calculated.
In Fig. 3.7, we investigate the effect of PRB allocation size and pilot type. The SNR is fixed to
be 10 dB for both UFMC and CP-OFDM systems. The used pilots are all-ones vector and pi-
lots according to LTE standard. The PDF of estimated timing offset are plotted for UFMC (left
side) and OFDM (right side) under the three considered channel models separately. It is clear
from the simulation results that the method works well under AWGN and flat fading channel
model with quite satisfying detection probability. The all-ones and LTE standard compliant
pilot sequences have almost the same performance in UFMC systems, while the LTE stan-
dard compliant pilot sequences outperform the simple all-ones pilot in OFDM systems. The
reason is that the auto-correlation properties of the pilot standardized in LTE is optimized for
OFDM systems and these properties are again destroyed in UFMC systems. The optimal auto-
correlation properties of Zadoff-Chu sequences is destroyed by the filtering approach, since
3.3. Results 31
Prob.
0.4 0.4
0.2 0.2
0 0
7 8 9 10 11 12 13 2 4 6 8 10 12 14
TO/Samples TO/Samples
0.6 0.6
Prob.
Prob.
0.4 0.4
0.2 0.2
0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
TO/Samples TO/Samples
UFMC VEHA 50 kmh SNR 10 dB TO 10 Samples OFDM VEHA 50kmh SNR 10 dB TO 10 Samples
0.4 0.4
2PRB−Allones
0.35
2PRB−LTE 2PRB−Allones
0.3 10PRB−Allones 0.3 2PRB−LTE
10PRB−LTE 10PRB−Allones
0.25 10PRB−LTE
Prob.
Prob.
0.2 0.2
0.15
0.1 0.1
0.05
0
0 5 10 15 20 0
0 2 4 6 8 10 12 14 16 18 20
TO/Samples TO/Samples
the symbols are not equally weighted in different subcarrier positions in UFMC. Furthermore,
it is noteworthy to mention that UFMC has more tolerance on negative timing errors than
OFDM if residual timing offset remains after synchronization. Under the Vehicular A channel
model, the performance of the estimator becomes worst because of the delay spread effect of
the channel. The estimator shows large variance for both OFDM and UFMC. While still the
pilots do not affect the performance much in UFMC systems, an obvious difference can be
observed for different pilots in OFDM systems. Moreover, the more the allocated PRB is, the
more accurate the estimation is.
Now, we study the performance behavior with respect to SNR of the estimator. For the sake
of simplicity, we compare the results for different SNRs in the range of -10 dB to 30 dB for a
settings of 10 PRB allocation with standard compliant pilot sequence. The results are shown in
Fig. 3.8. It can be easily observed that this method works already well for -10 dB SNR for both
OFDM and UFMC systems. Starting from SNR of the value 0 dB, no further improvement can
be achieved any more. Furthermore, the results are almost the same for OFDM and UFMC.
Because the estimation method is based on correlation between the known pilot and received
signal, another aspect that can affect the performance of the estimator could be the length of
received signal. The more slots that are included in the received signal, the more correlation
peaks exist. In the Fig. 3.9, PDFs are shown for different settings. In each figure, the PDFs are
calculated for 10 PRBs with 2 slots, 2 PRBs each with 2, 5 and 10 slots scenarios. In all the
scenarios, the standard compliant pilot sequences are used. For the considered three different
channels, it can be concluded based on the simulation results that the larger the correlation
size is, the better the estimation performance is. Moreover, increasing the allocation size is
more advantageous than increasing the correlation sequence length.
A performance comparison between CP-based and pilot-based timing offset estimation is dis-
cussion in the following section.
3.3. Results 33
UFMC AWGN 10 PRB LTE pilot TO 10 Samples OFDM AWGN 10 PRB LTE pilot TO 10 Samples
1 1
−10dB −10dB
0.9 0dB 0.9 0dB
10dB 10dB
0.8 0.8
20dB 20dB
0.7 30dB 0.7 30dB
0.6 0.6
Prob.
Prob.
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
TO/Samples TO/Samples
UFMC flat fading 10 PRB LTE pilot TO 10 Samples OFDM flat fading 10 PRB LTE pilot TO 10 Samples
1 1
−10dB −10dB
0.9 0dB 0.9 0dB
10dB 10dB
0.8 0.8
20dB 20dB
0.7 30dB 0.7 30dB
0.6 0.6
Prob.
Prob.
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
TO/Samples TO/Samples
UFMC VEHA 50 kmh 10 PRB LTE pilot TO 10 Samples OFDM VEHA 50 kmh 10 PRB LTE pilot TO 10 Samples
0.14 0.14
−10dB −10dB
0dB 0dB
0.12 0.12
10dB 10dB
20dB 20dB
0.1 30dB 0.1 30dB
0.08 0.08
Prob.
Prob.
0.06 0.06
0.04 0.04
0.02 0.02
0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
TO/Samples TO/Samples
Prob.
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
7 8 9 10 11 12 13 7 8 9 10 11 12 13 14
TO/Samples TO/Samples
0.6 0.6
Prob.
Prob.
0.4 0.4
0.2 0.2
0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
TO/Samples TO/Samples
0.08 0.08
Prob.
Prob.
0.06 0.06
0.04 0.04
0.02 0.02
0
0 5 10 15 20 25 30 5 10 15 20 25 30
TO/Samples TO/Samples
Besides timing offset, carrier frequency offset is also to be estimated using the joint timing
and frequency estimator based on pilot and CP. For OFDM systems, both CP-based and pilot-
based method can be used to jointly estimate timing and frequency offset. But only pilot-based
method can be used for UFMC systems, lacking of CP. Thus, the performance is discussed
and compared in this section.
In the pilot-based method, the normalized CFO is quantized into lq = 21 levels from -0.5
to 0.5 with quantization accuracy of 0.05. Because of quantization error, the estimation is
expected to be biased. However, the larger the quantization level lq is, the more accurate the
estimation can be. But it also results in higher computational complexity, since the number
of hypotheses increases. Thus, there is a compromise between the estimation accuracy and
computational complexity. The quantization accuracy of 0.05 with lq = 21 is sufficient to
neglect the quantization errors. The rCFO is set to be 0.1 subcarrier spacing and timing offset
of 10 samples. The allocation size is 10 PRBs and standard compliant pilot sequences are used
in pilot-based estimation method. In the Fig. 3.10, the resulting PDFs of the estimated timing
offset are shown for OFDM with CP-based and pilot-based method for different SNRs. From
the results for timing offset estimation, it is obvious that pilot-based method is more robust for
frequency unselective channel (AWGN and flat fading channel) . Satisfying performance can
be achieved under -10 dB SNR and good performance already from 0 dB SNR for the pilot-
based synchronization. But the CP-based method requires SNR larger than 10 dB to achieve
the same performance of pilot-based synchronization. The obtained results for UFMC systems
using pilot-based approach are similar than that for OFDM, since they are not again plotted
in Fig. 3.10. But under Vehicular A channel, CP-based method is more probable to estimate
the real timing offset. The corresponding frequency offset estimation performance are shown
in the Fig. 3.11. The MSE of estimated CFO and the real CFO is computed for UFMC with
pilot-based method and OFDM with pilot- and CP-based method. Thus, the MSE id defined
as
MSE = (ε̂ − ε)2 (3.14)
For AWGN channel, pilot-based method outperforms CP-based method especially for SNRs
which are higher than 5 dB. For flat fading, CP-based method is more robust for SNRs under
0 dB, while the MSE decreases fast for pilot-based method with increasing SNR. In Vehicular-
A channel model, CP-based method is more advantageous for SNRs under -5 dB and also
for SNRs higher than 10 dB. Furthermore, the performance of pilot-based method is almost
the same for OFDM and UFMC. Although the received signal is impaired with CFO, the
performance of timing offset seems not to be affected by CFO.
A value of CFO, which exactly matches the quantized CFO, does not exist in the practice. The
largest estimation error due to the quantization of CFO occurs, if the real CFO value is located
exactly between two quantized CFOs. Thus, we also simulate and evaluate the performance of
frequency estimation for a CFO of 0.075. The MSE of this pilot-based frequency estimation
are shown in fig. 3.12 and compared between OFDM and UFMC. It is shown that OFDM is
more sensitive to quantization errors than UFMC.
3.3. Results 36
CP−based OFDM AWGN 10 PRB LTE pilot TO 10 Samples rCFO 0.1 Pilot−based OFDM AWGN 10 PRB LTE pilot TO 10 Samples rCFO 0.1
1 1
−10dB −10dB
0.9 0dB 0.9 0dB
10dB 10dB
0.8 0.8
20dB 20dB
0.7 30dB 0.7 30dB
0.6 0.6
Prob.
Prob.
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
TO/Samples TO/Samples
(a) (b)
CP−based OFDM flat fading 10 PRB LTE pilot TO 10 Samples Pilot−based OFDM flat fading 10 PRB LTE pilot TO 10 Samples rCFO 0.1
1 1
−10dB −10dB
0.9 0dB 0.9 0dB
10dB 10dB
0.8 0.8
20dB 20dB
0.7 30dB 0.7 30dB
0.6 0.6
Prob.
Prob.
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
TO/Samples TO/Samples
(c) (d)
CP−based OFDM VEHA 10 PRB LTE pilot TO 10 Samples Pilot−based OFDM VEHA 10 PRB LTE pilot TO 10 Samples rCFO 0.1
0.45 0.14
−10dB −10dB
0.4 0dB 0dB
0.12
10dB 10dB
0.35 20dB
20dB
0.1 30dB
0.3 30dB
0.25 0.08
Prob.
Prob.
0.2 0.06
0.15
0.04
0.1
0.02
0.05
0 0
0 5 10 15 20 25 30 0 5 10 15 20 25 30
TO/Samples TO/Samples
(e) (f)
OFDM−CP
−2 OFDM−Pilot
10
UFMC−Pilot
MSE
−4
10
−6
10
−10 −5 0 5 10 15 20 25 30
SNR/dB
(a)
−2
10
MSE
−4
10
OFDM−CP
OFDM−Pilot
UFMC−Pilot
−6
10
−10 −5 0 5 10 15 20 25 30
SNR/dB
(b)
−3
10
−4
10
−10 −5 0 5 10 15 20 25 30
SNR/dB
(c)
AWGN Channel
−1
10
OFDM−2PRB
OFDM−10PRB
UFMC−2PRB
UFMC−10PRB
MSE 10
−2
−3
10
−10 −5 0 5 10 15 20 25 30
SNR/dB
(a)
−2
10
−3
10
−10 −5 0 5 10 15 20 25 30
SNR/dB
(b)
VEHA Channel
−1
10
OFDM−2PRB
OFDM−10PRB
UFMC−2PRB
UFMC−10PRB
MSE
−2
10
−3
10
−10 −5 0 5 10 15 20 25 30
SNR/dB
(c)
From the simulation results, it can be concluded that OFDM and UFMC have almost same per-
formance of timing and frequency synchronization via pilot. Furthermore, CP-based timing
synchronization requires higher SNR than pilot-based. For the investigated two kinds of pilot
sequences, UFMC systems exhibit approximately the same performance while the standard-
ized pilot shows some advantages in OFDM systems. The reason is that the standardized pilot
(e.g. Zadoff-Chu sequences) is designed with good auto-correlation properties for OFDM,
while these properties are not hold in UFMC.
4. Channel Estimation
After timing and carrier frequency offset are corrected, the CSI have to be estimated to get the
symbol estimates. In this chapter, we investigate pilot-assisted channel estimation for the novel
multi-carrier modulation technique UFMC and compare the performance with conventional
CP-OFDM systems. UFMC is expected to have several advantages compared with OFDM
due to its reduced out-of-band radiation and reduced guard subcarriers between sub-bands [5].
The channel estimation in UFMC systems is also interesting because of its absence of CP.
Because of the insertion of CP in OFDM systems at the transmitter, the overlapping part which
contains ISI is within CP, if the channel impulse length is smaller than the CP duration. The
OFDM receiver can efficiently mitigate channel-assisted ISI by removing the CP part without
causing any distortion. As we know, CP is not available in UFMC systems. This fact can lead
to higher ISI and destroy the orthogonality between subcarriers in a frequency-selective fading
channel, compared with CP-OFDM. However, the filter ramp-up and ramp-down act as a soft
protection against the channel delay spread since the signal energy contained within this part
is comparatively small.
To investigate the effect of channel-assisted ISI, simulations are carried out for three different
channels
• Vehicular-A channel model with the Power Delay Profile (PDP) according to C802.20
[16]
• A fading channel with two equal power density fading taps, but the delay between the
two taps can be varied
• Exponentially decaying PDP channel [17] with various asymptotic rms delay spread
Simulation parameters are shown in the following table 4.1. Every PRB is filtered by Dolph-
Chebyshev filter with a side lobe attenuation of 40 dB and the filters are frequency-shifted to
4.1. The effect of channel-assisted ISI 41
the center frequencies of each PRB. Furthermore, perfect channel knowledge is available at
the receiver and the noise is neglected, since we only interest on the effect of channel-assisted
ISI. The MSE is then calculated for the symbol estimates obtained at the receiver.
According to [16], the Vehicular-A channel model with the settings in 4.1 has 40 taps in sample
and this channel impulse length is smaller than the CP length in OFDM. In Fig. 4.1, MSE of
the received symbol estimates are plotted over different Vehicular-A channel realizations. The
UFMC and OFDM with VEHA channel
0
−10
−30
MSE/dB
−40
−50
−60
−70
−80
10 20 30 40 50 60 70 80 90 100
VEHA channel realization
mean MSE for 100 channel realizations is -42.7 dB for UFMC. This mean MSE indicates
that the channel ISI is negligible for this channel model. In CP-OFDM, the channel-assisted
ISI is completely mitigated because CP length is so designed that it is greater than channel
impulse length. A channel with exponentially decaying PDP is determined by asymptotic rms
delay spread, which describes the decaying speed of power density. In Fig. 4.2, the MSEs are
calculated for various rms delay spread, which is already normalized to the sample time. As
Exponentially Decaying PDP channel, Path Number 1000
0
−10
−20
MSE in dB
−30
−40
OFDM
UFMC
−50
0 20 40 60 80 100 120 140 160
asymptotic rms delay spread/samples
the results show that CP-OFDM can tolerate larger rms delay spread than UFMC for a same
target MSE. For the two fading taps case, shown in Fig. 4.3, CP-OFDM is also more robust
than UFMC. Furthermore, the x axis is delay between the two taps and normalized to FIR-
Two Same Fading Taps Channel
10
0
CP−OFDM
−10 UFMC−CW40
36% filter length
MSE in dB
−20
21% filter length
−30
−40
−50
0 2 4 6 8 10 12 14
τ /L
filter length, which is used in UFMC. As aforementioned, ISI can be completely mitigated in
CP-OFDM as long as the channel impulse length is smaller than CP duration. Compared to
CP-OFDM, the MSE increases while the delay increases. If a target MSE of -40 dB should
be achieved, then one might choose a filter with a appropriate length Lch /21%. If a target
MSE of -30 dB should be achieved, then one might choose a filter with a appropriate length
Lch /36%.
In the previous section, we show that the channel-assisted ICI is negligible in UFMC and
OFDM systems. By appropriately designing the FIR-filter and CP length in UFMC and
OFDM systems respectively, the channel-assisted ISI can be limited to below a certain value
of MSE such that it does not affect the system performance significantly. If UFMC and CP-
OFDM system are so designed that the channel-assisted ISI is negligible, we can write the
symbol estimates at the receiver for OFDM as (equation (2.31))
Without loss of generality, we can assume that the considered subcarrier k is allocated to the
PRB i and we also drop all odd subcarrier outputs of the 2N-point FFT in UFMC. Thus, (4.2)
can be further simplified as
Since the pilot symbol X(k) or Xi (k) is already known at the receiver, so we are able to estimate
H(k). It is straightforward that the channel estimates for OFDM and UFMC at subcarrier k
can be written as
YO (k) N(k)
Ĥ(k) = = H(k) + . (4.4)
X(k) X(k)
YU (k) N(k)
Ĥ(k) = = H(k) + . (4.5)
Xi (k)Fi (k) Xi (k)Fi (k)
Thus, the channel estimation differs only in the equalization of the filter shape in UFMC from
OFDM. In Fig. 4.4, the scheme of equalization of filter shape before channel estimation is
shown for UFMC. At the output of 2N-point FFT, the symbols are firstly divided by its filter
response in the corresponding subcarrier position in UFMC. Once the filter shape is equalized,
UFMC is then equivalent to OFDM systems. After obtaining all CTF at known pilot positions,
an interpolation algorithm can be performed to estimate CTF at other data symbol positions.
A linear interpolation algorithm is used to obtain CTF estimates for data symbols due to its
simplicity and low complexity. Consider two nearest pilot symbols with index l p1 and l p2
(l p2 > l p1 ) to the data symbol with time index l, the CTF estimates Ĥ(k, l) can be formulated
as
Ĥ(k, l) = β (l)Ĥ(k, l p1 ) + (1 − β (l))Ĥ(k, l p2 ) (4.6)
4.2. Single user channel estimation 44
l −l
where β (l) = l p2p2−l p1 is a weighting factor. The underlaying linear interpolation algorithm aims
at estimating the channel for other time instances, since the used pilot is with all frequency
placement. Thus, the performance of this approach highly depends on the velocity (i.e. the
coherence time of channel), at which the user is moving.
Noticed the fact that the initial CTF estimates of pilot symbols are corrupted by noise and
the CTF can be assumed to be constant within a certain length of consecutive subcarriers,
the variance of AWGN can be reduced by taking the mean of several consecutive channel
estimates as the final estimates. This algorithm is called sliding window. The procedure
of sliding window algorithm is illustrated in the Fig. 4.5. Mathematically, the final channel
estimates after applying sliding window algorithm can be described as
!
(LSW −1)/2
1
ĤSW (k, l) = Ĥ(k, l) + ∑ Ĥ(k − n, l) + Ĥ(k + n, l) (4.7)
LSW n=1
A selection window of size LSW is shifted by one subcarrier each time to obtain the final
estimate of the central subcarrier ĤSW within the selection window. Special treatment should
be taken for the “edge subcarriers”, which are located at the edge of an allocation because
channel estimates outside of allocation is not available. Whereas the noise variance is reduced
by applying sliding window algorithm, the channel estimates can be very inaccurate if the
4.2. Single user channel estimation 45
sliding window is chosen to be too large in a frequency-selective fading channel (e.g. sliding
window size larger than the coherence bandwidth). However, if the channel is frequency-flat,
a larger sliding window size is then advantageous. Normally, the size of sliding window LSW
is chosen to be odd so that the estimator is unbiased.
For our evaluation of this basic pilot-based channel estimator with the above described linear
interpolation and sliding window algorithm, we consider three types of channel:
• AWGN, only noise is present
• flat fading channel, the user moves at the speed of 50 kmh
• vehicular-A channel, the user moves at the speed of 50 kmh
In our simulation, a all-ones vector with 10 PRB allocation is used as pilot sequence and this
pilot is repeated in another slot. And other simulation parameters are shown in table 4.1.
The Symbol Error Rate (SER) are plotted in the following Fig. 4.6. As expected, the larger
the size of sliding window is, the smaller the SER is, if the channel is frequency unselective
e.g. AWGN (in Fig. 4.6a and 4.6b) or flat fading (in Fig. 4.6c and 4.6d). The reason is that
the averaging over frequencies reduces the noise variance and the CTF remains the same in
frequency domain. The unbiased effect of sliding window can be observed in a frequency-
selective channel such as Vehicular-A channel model, shown in Fig. 4.6e and 4.6f. For this
frequency-selective channel, an irreducible error floor occurs for high SNR per bit. With a
increasing size of sliding window (from sliding window size 5), the SER also increases for
high SNR per bit. If the noise dominates i.e. small SNR per bit values, a larger size of
sliding window is then advantageous. According to the expected SNR per bit in the system
under test, an optimal sliding window size can be selected. With this single user sliding
window channel estimator, a good performance can be achieved both for UFMC and CP-
OFDM system. According to the SER obtained for different channel models, a sliding window
with the size of 7 or 9 can be a choice for both frequency-flat and frequency-selective channel.
By choosing the sliding window with size of 7, a SNR lost per bit with respect to perfect
channel state information of 0.5 dB (0.7 dB) can be observed for OFDM (UFMC) with AWGN
channel for the SER of 10−2 . For a flat fading channel, the corresponding SNR lost per bit
are 0.7 dB for OFDM and 0.8 dB for UFMC. Furthermore, 0.8 dB and 1 dB SNR per bit
are lost for OFDM and UFMC respectively in the Vehicular-A channel with the speed of 50
kmh for the SER of 10−2 . UFMC has higher SNR per bit lost than OFDM, applying the
discussed sliding window algorithm for channel estimation. The reason can be explained by
the properties of the noise, since the noise is uncorrelated in OFDM systems and slightly
correlated in UFMC systems (the auto covariance properties was shown in Fig. 2.8 and A.2).
Furthermore, the noise variance is slightly enhanced in UFMC systems with the factor N+L−1 N
(see A.1). In our setting with N = 1024 and L = 74, the noise variance is about 0.3 dB larger
in UFMC systems.
A performance comparison between OFDM and UFMC is interesting. The results are shown
in Fig. 4.7 and they have almost the same performance for both frequency-flat and -selective
channel. The required SNR per bit for SER of 10−2 under AWGN channel is 11.6 dB and
12.1 dB for perfect CSI knowledge and sliding window size of 7 in OFDM systems, while
4.2. Single user channel estimation 46
SER
−3 −3
10 10
−4 −4
10 10
−5 −5
10 10
−6 −6
10 10
5 10 15 20 25 30 5 10 15 20 25 30
Eb/N0 in dB Eb/N0 in dB
(a) (b)
SER
−2 −2
10 10
−3 −3
10 10
5 10 15 20 25 30 5 10 15 20 25 30
Eb/N0 in dB Eb/N0 in dB
(c) (d)
SER
−2 −2
10 10
−3 −3
10 10
5 10 15 20 25 30 5 10 15 20 25 30
Eb/N0 in dB Eb/N0 in dB
(e) (f)
Figure 4.6.: Performance of the single user sliding window channel estimator
4.2. Single user channel estimation 47
AWGN channel
0
10
UFMC−perfectCSI
OFDM−perfectCSI
−1 UFMC−SW7
10
OFDM−SW7
−2
10
SER
−3
10
−4
10
−5
10
−6
10
5 10 15 20 25 30
Eb/N0 in dB
(a)
−1
10
SER
−2
10
−3
10
5 10 15 20 25 30
Eb/N0 in dB
(b)
VEHA channel
0
10
−1
10
SER
−2
10
−3
10
5 10 15 20 25 30
Eb/N0 in dB
(c)
they are 11.4 dB and 12.1 dB in UFMC systems respectively. If a single-tap Rayleigh fading
channel is considered, the required SNR per bit is 22.5 dB and 22.7 dB for UFMC and OFDM
correspondingly if perfect CSI is known. Furthermore, 23.3 dB (UFMC) and 23.4 dB (OFDM)
SNR per bit is required to achieved SER of 10−2 using sliding window with size of 7 respec-
tively. For the frequency-selective channel Vehicular-A, 22.75 dB and 23 dB SNR per bit is
required to achieve the SER 10−2 for UFMC and OFDM. Using the sliding window algorithm
with the size 7, the required SNR per bit become 23.7 dB and 23.8 dB respectively.
Another aspect of the estimator is the interpolation error. Since a linear interpolation according
to (4.6) is used, it causes interpolation error due to the linear model assumption. Thus, we
simulate the MSE of channel estimates for different speed, with which the user is moving,
shown in Fig. 4.8. The simulation is carried out using sliding window size of 5 with the linear
MSE of CIR in dB
−16 −16
−18 −18
−20 −20
−22 −22
−24 −24
−26 −26
0 50 100 150 200 250 300 0 50 100 150 200 250 300
speed in kmh speed in kmh
(a) (b)
interpolation approach for OFDM and UFMC systems. The SNR is set to be 20 dB because
the focus is to evaluate the interpolation error with respect to the speed of user. In Fig. 4.8a,
the MSE is computed for a single-tap Rayleigh fading channel with the speed ranging from
10 to 300 kmh. The results, shown in Fig. 4.8b are simulated under a Vehicular-A channel
model with the same speed range. From the simulation results, the channel estimates are more
accurate for OFDM systems if the speed of user is below 150 kmh, while it is more accurate
in UFMC systems for speed above 200 kmh. The same tendency can be observed for both flat
fading and Vehicular-A channel model.
In a frequency-selective channel, a received UFMC symbol lost its orthogonality on one hand
and on the other hand ISI is introduced due to the overlapping with previous symbol. As
discussed in A.3, three effects arise, namely signal amplitude reduction, ICI and ISI. By
extending the receiver window length (not exceed the channel impulse length), the energy of
ISI increases since the signal within the extended window also includes ISI from subsequent
symbol. However, the effect of loss of orthogonality i.e. ICI and signal amplitude reduction
can be reduced. Thus, extending the receiver window length might be considered for UFMC
to reduce the total energy of interferences, shown in Fig. 4.9. It should be noted that the ramps
4.2. Single user channel estimation 49
include both the waveform filter and the channel for UFMC. To evaluate the effect of window
extension, SER and MSE of symbol estimates are plotted in Fig. 4.10 for Vehicular-A channel
model. In the simulation, a user is allocated with 10 PRB and perfect channel state information
UFMC VEHA with Velocity 50 kmh UFMC VEHA with Velocity 50 kmh
0
10 30
25
20
−1 15
10
10 0dB
0dB
MSE/dB
10dB
SER
10dB 5
20dB
20dB
0 30dB
−2 30dB
10
−5
−15
−3
10 −20
0 20 40 60 80 100 0 20 40 60 80 100
Window extension length Window extension length
(a) (b)
Figure 4.10.: SER and MSE of extended receiver window for UFMC
is available at the receiver. Moreover, no windowing preprocessing is used. Within the channel
impulse length, no gains can be achieved for the considered Vehicular-A channel model. The
reason behind is the energy of included ISI is the same as the reduced energy of ICI and signal
amplitude reduction by extending the window length (see A.3).
The results till now are based on the assumption that the channel-assisted ISI is negligible for
UFMC and it can be completely mitigated via removing the CP for OFDM. In the following
Fig. 4.11, we also show the SER for OFDM and UFMC under Vehicular-B channel model. As
discussed in section 2.1.3, the Vehicular-B channel exhibits a very large delay spread, which
is larger than the CP and the filter ramp-up duration for OFDM and UFMC. This fact leads
to high ISI an ICI in both systems. First of all, (4.3) is not hold anymore for UFMC. Instead
of (4.3), we should take (2.43). It requires more advanced channel estimator to mitigate the
ICI and ISI. Second of all, ISI can also not be completely mitigated in OFDM systems in our
settings, since the delay spread is larger than CP duration. Even if perfect CSI is available at
the receiver, the system performance is not good enough for practical usage due to ICI and ISI.
Furthermore, OFDM has better performance than UFMC in this channel model. However, this
performance is also not good for practical usage.
4.2. Single user channel estimation 50
0
VEHB channel
10
O−PerfCSI
U−PerfCSI
O−SW3
U−SW3
O−SW7
U−SW7
SER
−1
10
−2
10
5 10 15 20 25 30
SNR per Bit in dB
carrier symbol contains NMC = N + L − 1 samples, the received signal is a linear convolution
between CIR and the transmitted symbol. Thus, it can be written in matrix form as
y = s̃ h + w (4.8)
[(NMC +Lch −1)×1] [(NMC +Lch −1)×Lch ][Lch ×1] [(NMC +Lch −1)×1]
where s̃ = [x̃(0)T , x̃(1)T , · · · , x̃(Lch −1)T ]T . x̃(l) is the linearly l samples shifted time domain pilot
signal s with Lch − 1 zeros padded at the end. However, the first Lch − 1 and the last Lch − 1
samples contain ISI from previous and subsequent symbols. Thus, they are dropped. The
ISI-free part of received signal thus can be written as
y = s̃ h + w (4.9)
[(NMC −Lch +1)×1] [(NMC −Lch +1)×Lch ][Lch ×1] [(NMC −Lch +1)×1]
4.3. Multi-user channel estimation 51
ĥ = s̃+ y (4.10)
Compared to the frequency domain estimation of CTF, ICI and ISI are excluded in this
method. However, this method causes noise enhancement due to the pseudo-inverse opera-
tion. According to our evaluation with this method, it works well without noise. Negligible
MSE of CIR can be achieved without noise. For an extreme high SNR of 100 dB, the MSE is
-21.1 dB. But if the SNR is 10 dB, the MSE becomes 23 dB. The sensitivity against noise is be-
cause the filter ramp-up and ramp-down signal, which have comparatively low SNR, are used
for estimation. Thus, in order to make use of the advantages of this method, pilot sequence
have to be carefully designed according to the following criterion
−1
MSE = σn2 tr{ s̃H s̃ } (4.11)
In this section, we investigate and compare channel estimation performance for OFDM and
UFMC in multi-user scenario. In a multi-user scenario, users transmit their signals using
same time and frequency slots. Thus, the received signal is the superimposed signal of all the
transmitting users. In general, if the users are allocated with NAlloc subcarriers, the received
frequency domain signal inOFDM is given by
YO = S̃O H + WO (4.12)
[NAlloc ×1] [NAlloc ×NUser NAlloc ][NUser NAlloc ×1] [NAlloc ×1]
YU = S̃U H + WU (4.15)
[NAlloc ×1] [NAlloc ×NUser NAlloc ][NUser NAlloc ×1] [NAlloc ×1]
4.3. Multi-user channel estimation 52
where the filter shape is also taken into account in the matrix S̃U with
It contains the filter response in frequency domain for each user at the corresponding subcar-
rier positions. The number of variables to estimate is NUser NAlloc , while we only have NAlloc
equations. It is not possible to estimate all the parameters.
By choosing a sliding window size which is smaller than the coherence bandwidth of the
channel, it can be assumed that the CTF is flat within the sliding window. By doing so, the
number of variables to be estimated is reduced from LSW NUser to NUser . Mathematically, (4.12)
and (4.15) can be simplified as
YO = SO H + WO (4.18)
[LSW ×1] [LSW ×NUser ][NUser ×1] [LSW ×1]
and
YU = SU H + WU (4.19)
[LSW ×1] [LSW ×NUser ][NUser ×1] [LSW ×1]
with
S0,i ··· SNUser −1,i
S0,i+1 ··· SNUser −1,i+1
SO = (4.20)
.. .. ..
. . .
S0,i+LSW −1 · · · SNUser −1,i+LSW −1
where i is a starting subcarrier index within the sliding window. Similarly, taking the filter in
subcarrier into account SU is give by
F0,i S0,i ··· FNUser −1,i SNUser −1,i
F0,i+1 S0,i+1 ··· FNUser −1,i+1 SNUser −1,i+1
SU = (4.21)
.. ... ..
. .
F0,i+LSW −1 S0,i+LSW −1 · · · FNUser −1,i+LSW −1 SNUser −1,i+LSW −1
4.3. Multi-user channel estimation 53
From (4.18) and (4.19), there exists a left inverse matrix of SO and SU respectively, provided
LSW ≥ NUser and they have full rank NUser . The channel estimator is then
Ĥ = S+ +
O,U YO,U = H + SO,U WO,U (4.22)
where S+
O,U is the Moore–Penrose pseudo inverse matrix of SO for OFDM systems and SU for
UFMC systems, respectively. For simplicity reason, we drop the index O and U for discussing
OFDM and UFMC systems consistently, i.e. without subindex means it can apply both to
OFDM and UFMC. In this case, it can be written as
−1
S+ = SH SSH (4.23)
= tr {E S+ RWW (S+ )H }
(4.24)
where tr{·} is the trace operator and RWW denotes the auto-covariance matrix of the noise in
OFDM and UFMC respectively. For OFDM systems, the auto-covariance matrix of the noise
is given by
RWW,O = σn2 I (4.25)
where I is identity matrix, which contains 1 for all the diagonal elements and 0 for off-diagonal
elements. For UFMC systems, the auto-covariance matrix of the noise RWW,U is derived in
A.1 and A.2. Thus, (4.24) can be further simplified for OFDM as
The corresponding minimum MSE per user is thus σn2 /Eav , providing that (4.27) is satisfied.
The condition in (4.27) implies that the pilot sequences between different users should be or-
thogonal to each other in OFDM. In contrast to OFDM, because the noise auto-covariance ma-
trix in UFMC is not a identity matrix, the MSE is expected to be larger than that of OFDM.
For two users scenario, the DFT-sequences can achieve the minimum MSE due to orthogo-
nality property. For a given allocation size and sliding window size, one of the two users uses
4.3. Multi-user channel estimation 54
all ones vector as pilot, S0 = [1, 1, · · · , 1]. Another user is then allocated with the the cyclical
extended pilots
[S1 ]m = [S01 ]q
(4.30)
where q = m mod LSW
with
[S1 ]k = e j2πk/LSW , k = 0, ..., LSW − 1 (4.31)
Hence, a minimum MSE can be achieved in every sliding window for OFDM.
Using the DFT-sequence mentioned above in conjunction with the sliding window approach,
the performance of single user and two users are compared for AWGN and flat-fading channel
in the Fig. 4.13. Furthermore, the edge subcarriers are discarded since it is not possible to
perform sliding window with the given size for them. The MSE of single user and two users
scenario are in good agreement, since the pilots of the two users are orthogonal to each other in
OFDM. Furthermore, the MSE inUFMC systems are about 0.7 dB larger than that in OFDM
systems, if sliding window of size 11 is used. The reason is that the theoretical minimum MSE
O−2U−SW11 O−2U−SW11
−10 U−1U−SW3 −10 U−1U−SW3
U−2U−SW3 U−2U−SW3
U−1U−SW11 U−1U−SW11
−15 U−2U−SW11 −15 U−2U−SW11
−20 −20
−25 −25
−30 −30
−35 −35
−40 −40
5 10 15 20 25 30 5 10 15 20 25 30
SNR per bit in dB SNR per bit in dB
(a) (b)
Figure 4.13.: Performance of the basic MU-channel estimator with DFT sequence for fre-
quency flat channel without edge subcarriers
can not be achieved with the DFT-sequence for UFMC on one hand, whereas it is achieved
in OFDM. On the other hand, the noise variance is about 0.3 dB larger than in OFDM. For
VEHA channel, the MSE of single user and two users case is shown in Fig. 4.14. Compared to
single user scenario, larger MSE in multi-user scenario appears for high SNR for both OFDM
and UFMC if the size of sliding window is chosen to be very small, namely 3. The reason
is that if the frequency-flat assumption within sliding window is not hold, it also introduces
Multi User Interference (MUI) compared to single user case. If a very large sliding window
size of 11 is used, an irreducible error floor occurs from SNR per bit of the value 15 dB. This
fact indicates that the frequency-flat assumption does not hold any more.
4.3. Multi-user channel estimation 55
VEHA channel
0
O−1U−SW3
O−2U−SW3
−5 O−1U−SW11
−20
−25
−30
−35
5 10 15 20 25 30
SNR per bit in dB
Figure 4.14.: Performance of the sliding window MU-channel estimator with DFT sequence
for VEHA channel without edge subcarriers
For the subcarriers at the edge of the allocation, there exists not enough data to perform sliding
window algorithm for the given size of LSW . We discarded the edge subcarriers in above to
evaluate the performance of sliding window algorithm in multi user scenario. Now, we also
take the edge subcarriers into account. If the edge subcarriers are not well treated, MUI in
theses subcarrier positions might be very significant. Possible approaches to handle the edge
subcarriers might be
• Set the channel estimates at the edge subcarriers as the channel estimates in the near-
est subcarrier position, which is possible to perform sliding window algorithm. That
is ∀k ≤ (LSW + 1)/2, Ĥ(k) = Ĥ((LSW + 1)/2) and ∀k ≥ NAlloc − (LSW + 1)/2, Ĥ(k) =
Ĥ((NAlloc − (LSW + 1)/2 − 1). By using this approach, the performance for flat fad-
ing channel can be expected to be good. However, it might have negative aspects for
frequency-selective channel, since the estimates is biased. This approach is refer to as
EgCont.
• We can also decrease the size of sliding window to perform channel estimation at the
edge subcarrier positions. However, due to the pseudo inverse matrix, the noise can be
enhanced by using this approach. Furthermore, the edge subcarriers 0 and NAlloc − 1
can still not be treated properly. This approach is refer to as Eg– –.
• Using the nearest two channel estimates and perform linear extrapolation to get the
channel estimates in the edge subcarrier positions. That is ∀k < (LSW + 1)/2, Ĥ(k) =
α Ĥ( LSW2+1 + 1) + (1 − α)Ĥ((LSW + 1)/2) and similarly for k ≥ NAlloc − (LSW + 1)/2.
This approach is refer to as EgExpl.
4.3. Multi-user channel estimation 56
Simulations are carried out to evaluate the performance between the three aforementioned
different edge subcarrier treatments, shown in Fig. 4.15. In order to show the performance
OFDM AWGN SW11 with different Edge subcarrier treatment UFMC AWGN SW11 with different Edge subcarrier treatment
−5 −5
O−noEg U−noEg
O−Eg−− U−Eg−−
−10 O−EgCont −10 U−EgCont
O−EgExpl U−EgExpl
−15 −15
MSE in dB
MSE in dB
−20 −20
−25 −25
−30 −30
−35 −35
−40 −40
5 10 15 20 25 30 5 10 15 20 25 30
SNR per bit in dB SNR per bit in dB
(a) (b)
OFDM FF SW11 with different Edge subcarrier treatment UFMC FF SW11 with different Edge subcarrier treatment
−5 −5
O−noEg U−noEg
O−Eg−− U−Eg−−
−10 O−EgCont −10 U−EgCont
O−EgExpl U−EgExpl
−15 −15
MSE in dB
MSE in dB
−20 −20
−25 −25
−30 −30
−35 −35
−40 −40
5 10 15 20 25 30 5 10 15 20 25 30
SNR per bit in dB SNR per bit in dB
(c) (d)
Figure 4.15.: Comparison between different edge subcarrier treatments in AWGN and flat-
fading channel model
differences between the approaches, a large sliding window size of 11 is chosen for AWGN
and single-tap Rayleigh fading channel with speed of 50 kmh. noEg denotes that the MSE
is calculated without taking the edge subcarriers into account. From the simulation results,
EgCont has the best performance among all the three approaches and it does not cause any
additional performance degradation compared to noEg, because frequency-flat channel is con-
sidered and this approach simply repeats its channel estimates in the edge subcarriers. The
approach EgExpl, based on linear extrapolation, causes about 0.8 dB lost of SNR per bit. And
the approach Eg– –, based on decreasing sliding window size, causes about 2 dB lost of SNR
per bit. If a frequency-selective channel model e.g. Vehicular-A is considered, the results are
plotted in Fig. 4.16. A sliding window size of 5 is chosen in the simulation, since a larger size
of sliding window is not practical for the considered Vehicular-A channel model. In this case,
the approach Eg– – has best performance especially for high SNR. For example, if a MSE of
25 dB should be achieved, the lost of SNR per bit compared to noEg is 0.3 dB, 0.5 dB and
4.3. Multi-user channel estimation 57
OFDM VEHA SW5 with different Edge subcarrier treatment UFMC VEHA SW5 with different Edge subcarrier treatment
−5 −5
U−noEg
U−Eg−−
U−EgCont
U−EgExpl
−10 −10
−15 −15
MSE in dB
MSE in dB
−20 −20
−25 −25
−30 −30
5 10 15 20 25 30 5 10 15 20 25 30
SNR per bit in dB SNR per bit in dB
(a) (b)
Figure 4.16.: Comparison between different edge subcarrier treatments inVEHA channel
model
0.9 dB for Eg– –, EgCont and EgExpl respectively both for UFMC. The corresponding value
of SNR per bit lost in OFDM is 0.3 dB, 0.4 dB and 0.8 dB.
Till now, we used the DFT-sequences to investigate the performance of channel estimation.
There are however also other pilot sequences. Users can be allocated with Zadoff-Chu se-
quences with different roots and different shifts as well. In Fig. 4.17, we compare the per-
formance between different pilots for AWGN and single-tap Rayleigh fading channel model.
The sliding window size is chosen to be 3 in the simulations. It is clear that the orthogonal
−5 −5
MSE in dB
MSE in dB
−10 −10
−15 −15
−20 −20
−25 −25
−30 −30
−35 −35
5 10 15 20 25 30 5 10 15 20 25 30
SNR per bit in dB SNR per bit in dB
(a) (b)
Figure 4.17.: Comparison of performance between different pilots in AWGN and flat-fading
channel model
DFT-sequence has the best performance due to its orthogonality. Furthermore, OFDM sys-
tems slightly outperform UFMC systems with about 0.2 dB lower value of MSE. Because
4.3. Multi-user channel estimation 58
the orthogonality is obtained for OFDM but not hold for UFMC. The performance of Zadoff-
Chu sequences with different shifts is better than that with different roots. Using Zadoff-Chu
sequences, UFMC systems slightly outperform OFDM systems. The same tendency can be
observe for Vehicular-A channel model, shown in Fig. 4.18.
−5
MSE in dB
−10
−15
−20
−25
−30
5 10 15 20 25 30
SNR per bit in dB
Figure 4.18.: Comparison of performance between different pilots in VEHA channel model
In the previous section, the Least Squares (LS) estimator in conjunction with sliding window
approach is discussed. In this section, a Minimum Mean Square Error (MMSE) estimator
in conjunction with sliding window is introduced to reduced the effect of potential noise en-
hancement of LS estimator. In (4.18) and (4.19), the signal model is given for OFDM and
UFMC within a considered sliding window. In [32], it is shown that the MMSE estimator is
given by
−1
Cmmse = RHH SH SRHH SH + RWW (4.32)
where RHH is auto-correlation matrix of the channel coefficients between different users. It is
assumed to be identity matrix
RHH = I (4.33)
[NUser ×NUser ]
since the channel coefficients are not correlate between users. Thus, (4.32) can be simplified
for OFDM as
2 −1
Cmmse,O = SH H
O SO SO + σn I (4.34)
and for UFMC as −1
Cmmse,U = SH H
U SU SU + RWW,U (4.35)
4.3. Multi-user channel estimation 59
where RWW,U is the auto-covariance matrix of noise in UFMC, given in A.1 and A.2.
In the following Fig. 4.19, the performance of LS and MMSE estimator in conjunction with
sliding window is compared for different channel models. The results show that MMSE es-
timator provides some gain for low SNR. While the SNR increases, the gain decreases for
both OFDM and UFMC. Furthermore, for both the LS and MMSE approach, OFDM slightly
outperforms UFMC.
4.3. Multi-user channel estimation 60
−15
MSE in dB −20
−25
−30
−35
−40
0 5 10 15 20 25 30 35
SNR per Bit in dB
(a)
−15
−20
MSE in dB
−25
−30
−35
−40
0 5 10 15 20 25 30 35
SNR per Bit in dB
(b)
−12
−14
MSE in dB
−16
−18
−20
−22
−24
−26
0 5 10 15 20 25 30 35
SNR per Bit in dB
(c)
Figure 4.19.: Performance Comparison between LS- and MMSE estimator in conjunction with
sliding window for different channel models
5. Conclusion
The aspects of time-frequency synchronization and channel estimation are analyzed and in-
vestigated in this thesis for both OFDM and UFMC systems. We compared the performance
between OFDM and UFMC systems in different point of views. These methods will be im-
plemented in a demonstrator to test the performance under real world radio propagation.
Regarding the time-frequency estimation, both pilot-based and cyclic prefix-based synchro-
nization methods are investigated. However, the cyclic prefix-based method can only be used
in OFDM systems. The simulation results show that the pilot-based method outperform cyclic
prefix-based method. It’s main shortcoming is higher computational complexity, compared to
cyclic prefix-based method. Furthermore, this method has a very good accuracy even under
low SNRs in frequency-flat fading channels.
Although cyclic prefix is not available in UFMC, the filter ramp-up and ramp-down act as
inherent soft protection against channel delay spread. It was shown that the Vehicular-A chan-
nel causes a negligible MSE of -40 dB for UFMC. Furthermore, almost same performance
compared with OFDM can be achieved for UFMC for the investigated channel models.
In this thesis, we used existing synchronization and channel estimation methods and also
pilot sequences, that are originally designed and optimized for OFDM systems. Thus, many
potential methods can be done to further enhance the performance for UFMC like
• pilot design for UFMC, which will have optimal or suboptimal auto-correlation prop-
erties (like Zadoff-Chu sequences for OFDM) to enhance the performance of synchro-
nization.
• pilot design for the time-domain estimator in chapter 4 in UFMC to reduce the effect of
noise enhancement
• more advanced channel estimation method for UFMC
• taking the auto-covariance of noise in UFMC systems into account to design pilot se-
quences
A. Appendix
The time domain zero mean AWGN w(k) with variance σn2 is firstly padded with zeros to
perform 2N-point FFT at the receiver. The 2N-point FFT of w(k) can be thus written as
1 N+L−2
WU (k) = √ ∑ w(l)e− j2πlk/2N , k = 0, 1, · · · , 2N − 1 (A.1)
N l=0
where (·)∗ denotes conjugate complex operator [30]. Insert (A.1) into (A.3), we get
" ! !#
N+L−2 N+L−2
1
σU2 (k) = E ∑ w(l)e− j2πlk/2N ∑ w∗ (l)e j2πlk/2N
N l=0 l=0
" # " #
N+L−2 N+L−2 N+L−2
1 1
= E
N ∑ w(l)w∗(l) + N E ∑ ∑ w(l)w∗ (m)e j2π(m−l)k/2N
(A.3)
l=0 m=0,m6=l l=0
| {z } | {z }
N+L−1 2 =0
= N σn
N +L−1 2
= σn
N
The second term is equal zero since w is a white noise, which satisfies
(
σn2 , if l = k
E [w(l)w∗ (k)] = (A.4)
0, otherwise
The derivation holds for all subcarriers. Thus, the noise in every subcarrier k is a zero mean
Gaussian-distributed process with the variance of N+L−1
N σn .
2
A.2. Covariance of noise in UFMC 63
The covariance between WU (l) and WU (k), i.e. noise process in subcarrier l and k, is defined
as [33]
CWU (l, k) =E (WU (l) − E[WU (l)]) (WU (k) − E[WU (l)])∗
(A.5)
=E [WU (l)WU∗ (k)]
since WU (l) is zero mean and Gaussian-distributed for all subcarriers. Similarly, insert (A.1)
into (A.6), we get
" ! !#
N+L−2 N+L−2
1
CWU (l, k) = E
N ∑ w(m)e− j2πlm/2N ∑ w∗(m)e j2πmk/2N
m=0 m=0
" # " #
N+L−2 N+L−2 N+L−2
1 ∗ j2π(k−l)m/2N 1 ∗ j2π(mk−ln)/2N
= E ∑ w(m)w (m)e + E ∑ ∑ w(l)w (m)e
N m=0 N n=0,n6=m m=0
| {z
} | {z }
π(k−l)
sin 2N (N+L−1) π(k−l) =0
1
N
π(k−l)
· e j 2N (N+L−2) σn2
sin 2N
= c(l, k)σn2 if l 6= k
(A.6)
Obviously, the covariance only depends on the differences between l and k, hence ∆k = k − l is
introduced. Furthermore, all odd subcarriers are dropped in OFDM. Thus, only the covariance
between even subcarriers are of interest. We introduce l 0 = 2l and k0 = 2k, such that (A.7) the
covariance between subcarriers of interest can be formulated as
π∆k
1 sin (N + L − 1) π∆k
c(l 0 , k0 ) = N
· e j N (N+L−2) σn2 (A.8)
N sin Nπ∆k
It can be observed from (A.8), c(l 0 , k0 ) = 0 if the filter length is L = 1, which corresponds
to OFDM systems. The correlation between different subcarriers in UFMC results from the
Leakage effect of FFT if the FFT size is not an integer multiple of the signal length.
A.3. Derivation of symbol estimates in UFMC 64
Lacking of CP, delay spread of a wireless channel can destroy the orthogonality of UFMC
signals and introduce ISI resulting from previous symbol, illustrated in the Fig. A.1. Consider
a symbol of interest is with the index of m, the received signal rU,m within the detection
window can be written as
where circshift(·, −(N + L − 1)) denotes a cyclical shift of signal to the left of N + L − 1
samples. wr and wch are two rectangular windows and they are given by
(
1 if 0 ≤ k ≤ N + L − 2
wr (k) = (A.10)
0 if N + L − 1 ≤ k ≤ 2N − 1
and (
1 if 0 ≤ k ≤ Lch − 1
wch (k) = (A.11)
0 if Lch ≤ k ≤ 2N − 1
Applying 2N-point FFT to (A.9), we get
ỸU,m (k) = Wr (k) ∗ (Hm (k)YU,m (k)) + eφ (k)Wch (k) ∗ (Hm+1 (k)YU,m+1 (k)) +WU (k) (A.12)
| {z } | {z }
=ỸU1,m (k) =IISI (k)
where φ (k) denotes a linear phase shift in frequency domain due to circular shift in time
domain, it is given by
φ (k) = jπ(N + L − 1)/N (A.13)
A.3. Derivation of symbol estimates in UFMC 65
Furthermore, Wr , Wch , YU,m and YU,m+1 are 2N-point FFT of wr , wch , yU,m and yU,m+1 respec-
tively. They are straightforward
πk
sin 2N (N + L − 1) πk
− j 2N (N+L−2)
Wr (k) = πk
· e (A.14)
sin 2N
πk
sin 2N Lch πk
Wch (k) = πk
· e− j 2N (Lch −1) (A.15)
sin 2N
B
YU,m (k) = ∑ X̃i,m (k)Fi (k) (A.16)
i=1
B
YU,m+1 (k) = ∑ X̃i,m+1 (k)Fi (k) (A.17)
i=1
where X̃i is given in (2.39). Observe the term ỸU1,m (k) in (A.12), it causes ICI because the
tail of complete m-th symbol is lost. Thus, it is desired to separate the ICI term from signal
term. Without loss of generality, it is assumed that the subcarrier k is allocated to the subband
i. Hence, we reformulate ỸU1,m (k) as
ỸU1,m (k) = (Wch (k) +Wr (k)) ∗ (Hm (k)YU,m (k)) −Wch (k) ∗ (Hm (k)YU,m (k))
| {z }
=Hm (k)YU,m (k))
(A.18)
1 2N−1
= Hm (k)X̃i,m (k)Fi (k) − ∑ Wch(k − l)Hm(l)YU,m(l)
2N l=0
| {z }
ỸU2,m (k)
The second term ỸU2,m (k) still contains a part of signal, since YU,m contains signal part X̃i,m (k).
Thus, for a considered subcarrier k belonging to subband i, we decompose YU,m into
B B
YU,m (k) = ∑ X̃i,m (k)Fi (k) = X̃i,m,Sk (k)Fi (k) + ∑ X̃i,m,ICIk (k)Fi (k) (A.19)
i=1 i=1
with
k
X i,m 2 if l = k
k sin( 2 (k−l))
π 1
Xi,m,Sk (l) = Xi,m ( 2 ) N sin π (k−l) e j 2 (k−l)(1− N )
π
if l odd (A.20)
( 2N )
0 otherwise
and
X̃i,m,ICIk (l) = X̃i,m (l) − Xi,m,Sk (l) (A.21)
A.3. Derivation of symbol estimates in UFMC 66
1 2N−1
ỸU1,m (k) =Hm (k)X̃i,m (k)Fi (k) − ∑ Wch(k − l)Hm(l)Fi(l)X̃i,m,Sk (l)
2N l=0
| {z }
= αS (k)X̃i,m (k) :Amplitude reduction
(A.22)
1 2N−1 B
− ∑ ∑ Wch(k − l)Hm(l)Fi(l)X̃i,m,ICIk (l)
2N l=0 i=1
| {z }
= IICI (k) : ICI within subband and between subband
ỸU,m (k) = Hm (k)X̃i,m (k)Fi (k) − αS (k)X̃i,m (k) − IICI (k) + IISI (k) +W (k) (A.23)
Furthermore, we observed
2
E{IISI 2
(k)} = E{ỸU2,m (k)} = E{(αS (k) + IICI (k))2 } (A.24)
since
1 2N−1
ỸU2,m (k) = ∑ Wch(k − l)Hm(l)YU,m(l)
2N l=0
(A.25)
and
1 φ (k) 2N−1
IISI (k) =
2N
e ∑ Wch(k − l)Hm(l)YU,m+1(l) (A.26)
l=0
only differs in a phase shift.
Bibliography
[1] G. Wunder, M. Kasparick, S. ten Brink, F. Schaich, T. Wild, I. Gaspar, E. Ohlmer,
S. Krone, N. Michailow, A. Navarro, G. Fettweis, D. Ktenas, V. Berg, M. Dryjanski,
S. Pietrzyk, and B. Eged, “5gnow: Challenging the lte design paradigms of orthogonal-
ity and synchronicity,” in Vehicular Technology Conference (VTC Spring), 2013 IEEE
77th, June 2013, pp. 1–5.
[2] G. Wunder, P. Jung, M. Kasparick, T. Wild, F. Schaich, Y. Chen, S. Brink, I. Gaspar,
N. Michailow, A. Festag, L. Mendes, N. Cassiau, D. Ktenas, M. Dryjanski, S. Pietrzyk,
B. Eged, P. Vago, and F. Wiedmann, “5gnow: non-orthogonal, asynchronous waveforms
for future mobile applications,” Communications Magazine, IEEE, vol. 52, no. 2, pp.
97–105, February 2014.
[3] B. Farhang-Boroujeny, “Ofdm versus filter bank multicarrier,” Signal Processing Maga-
zine, IEEE, vol. 28, no. 3, pp. 92–112, May 2011.
[4] V. Vakilian, T. Wild, F. Schaich, S. ten Brink, and J.-F. Frigon, “Universal-filtered multi-
carrier technique for wireless systems beyond LTE,” in 2013 IEEE Globecom Workshops
(GC Wkshps), Dec 2013, pp. 223–228.
[5] T. Wild, F. Schaich, and Y. Chen, “5G Air Interface Design based on Universal Filtered
(UF-)OFDM,” in Int. Conf. on Dig. Sig. Processing (DSP’14), August 2014.
[6] F. Schaich and T. Wild, “Waveform contenders for 5G – OFDM vs. FBMC vs. UFMC,”
in 2014 6th International Symposium on Communications, Control and Signal Process-
ing (ISCCSP), May 2014, pp. 457–460.
[7] F. Schaich and T. Wild, “Relaxed Synchronization Support of Universal Filtered
Multi-Carrier including Autonomous Timing Advance,” in 11th Int. Symp. on Wireless
Comm.Sys.(ISWCS’14), August 2014.
[8] F. Schaich, T. Wild, and Y. Chen, “Waveform contenders for 5G – suitability for short
packet and low latency transmissions,” in IEEE Veh. Technol. Conf. Spring (VTC’14
Spring), April 2014.
[9] J. G. Proakis, Digital Communications, 4th ed. McGraw-Hill, 2001.
[10] G. L. Stuber, Principles of Mobile Communications, 2nd ed. Kluwer Academic Pub-
lishers, 2001.
Bibliography 68
[11] T. S. Rappaport, Wireless Communications: Principles and Practice, 2nd ed. Prentice
Hall, 2002.
[12] B. Sklar, “Rayleigh fading channels in mobile digital communication systems .i. charac-
terization,” Communications Magazine, IEEE, vol. 35, no. 7, pp. 90–100, Jul 1997.
[13] A. F. Molisch, Wireless Communications, 2nd ed. Wiley & Sons Ltd., 2011.
[14] M. S. Akram, “Pilot-based Channel Estimation in OFDM Systems,” 2007.
[15] R. Clarke, “A statistical theory of mobile-radio reception,” Bell System Technical Jour-
nal, The, vol. 47, no. 6, pp. 957–1000, July 1968.
[16] A. Kogiantis, “Proposed Text for the SCM Evaluation Process,” IEEE 802.20 Working
Group on Mobile Broadband Wireless Access, 2005.
[17] D. Morgan, “Analysis and realization of an exponentially-decaying impulse response
model for frequency-selective fading channels,” Signal Processing Letters, IEEE, vol. 15,
pp. 441–444, 2008.
[18] J. Speidel, “Übertragungstechnik II,” University Lecture, Institut für Nachrichtenübertra-
gung , University of Stuttgart, 2011.
[19] T. Wild, “5G Waveform Candidate Selection,” The 5GNOW Project Consortium groups,
2013.
[20] 3GPP TS 36.201, “LTE Physical Layer – General Description,” Technical Specification
Group Radio Access Network (Release 8), 2011.
[21] T. Wild, L.-H. Nguyen, and S. ten Brink, “Joint channel estimation across multiple cells
in coordinated multi-point,” in 2012 International Symposium on Wireless Communica-
tion Systems (ISWCS), Aug 2012, pp. 845–849.
[22] 3GPP TS 36.211, “Evolved Universal Terrestrial Radio Access (E-UTRA) Physical
Channels and Modulation,” Technical Specification Group Radio Access Network (Re-
lease 10), 2011.
[23] B. Popovic, “Generalized chirp-like polyphase sequences with optimum correlation
properties,” Information Theory, IEEE Transactions on, vol. 38, no. 4, pp. 1406–1409,
Jul 1992.
[24] J. Lee, H.-L. Lou, D. Toumpakaris, and J. Cioffi, “Effect of carrier frequency offset on
ofdm systems for multipath fading channels,” in Global Telecommunications Confer-
ence, 2004. GLOBECOM ’04. IEEE, vol. 6, Nov 2004, pp. 3721–3725 Vol.6.
[25] M. Hasan, F. Nabita, A. Khandakar, I. Ahmed, and F. Ahmed, “Analytical evaluation of
timing offset error in ofdm system,” in Communication Software and Networks, 2010.
ICCSN ’10. Second International Conference on, Feb 2010, pp. 3–7.
Bibliography 69
[26] Y. Mostofi and D. Cox, “Mathematical analysis of the impact of timing synchronization
errors on the performance of an ofdm system,” Communications, IEEE Transactions on,
vol. 54, no. 2, pp. 226–230, Feb 2006.
[27] X. Wang, T. Wild, F. Schaich, and A. Fonseca dos Santos, “Universal filtered multi-
carrier with leakage-based filter optimization,” in European Wireless 2014; Proceedings
of 20th European Wireless Conference;, May 2014, pp. 1–5.
[28] X. Wang, T. Wild, and F. Schaich, “Filter optimization for carrier frequency- and tim-
ing offset in universal filtered multi-carrier systems,” in to be published in IEEE Veh.
Technol. Conf. Spring (VTC’15 Spring), May 2015.
[29] J.-J. van de Beek, M. Sandell, and P. Borjesson, “Ml estimation of time and frequency
offset in ofdm systems,” Signal Processing, IEEE Transactions on, vol. 45, no. 7, pp.
1800–1805, Jul 1997.
[30] S. M. Kay, Fundamentals of Statistical Signal Procesing Volume I: Estimation Theory.
Prentice Hall PTR, 1993.
[31] H. Minn and N. Al-Dhahir, “Optimal training signals for mimo ofdm channel estima-
tion,” Wireless Communications, IEEE Transactions on, vol. 5, no. 5, pp. 1158–1168,
May 2006.
[32] J. Speidel, “Space time wireless communications,” University Lecture, Institut für
Nachrichtenübertragung , University of Stuttgart, 2011.
[33] A. Papoulis and S. U. Pillai, Probability, Random Variables and Stochastic Processes,
4th ed. McGraw-Hill, 2002.
Erklärung
Hiermit erkläre ich, dass ich diese Arbeit selbstständig verfasst und keine anderen als die
angegebenen Quellen und Hilfsmittel benutzt sowie alle wörtlich oder sinngemäß aus anderen
Werken übernommenen Aussagen als solche gekennzeichnet habe.
Die Arbeit wurde bisher keiner anderen Prüfungsbehörde vorgelegt und auch noch nicht ver-
öffentlicht. Das elektronische Exemplar stimmt mit den gedruckten Exemplaren überein.
Xiaojie Wang