Convolutional Codes for MGFDM system

Shravan Kumar Bandari V.V. Mani A. Drosopoulos

Electronics & Communication Engineering Electronics & Communication Engineering Electrical Engineering
National Institute of Technology National Institute of Technology TEI Western Greece
Warangal, India-506004 Warangal, India-506004 Patras, Greece-26334
Email: shravnbandari@nitw.ac.in Email: vvmani@nitw.ac.in Email: drosop@teiwest.gr

Abstract—In order to address a wide variety of future require- which reduces spectral efficiency. All these make OFDM
ments, Generalized Frequency Division Multiplexing (GFDM), unattractive for future wireless needs.
a non-orthogonal multicarrier scheme, is considered as one
of the most promising techniques available today. To improve Recently, Generalized Frequency Division Multiplexing
the orthogonality of the conventional GFDM system, a multi- (GFDM) has received a lot of attention as one of the leading
taper implementation of GFDM (MGFDM) using discrete prolate waveform candidates for 5G [2]. GFDM is a flexible two
spheroidal sequences (DPSSs) or multi-tapers, which have lower dimensional block based structure, dividing the data into a
out of band (OOB) radiation, can be exploited for pulse shaping number of subcarriers (K) and subsymbols (M ) with addi-
the individual subcarriers. For reliable data communications, tional degrees of freedom in choosing the prototype filters for
we propose in this article a combination of a convolution
code with an MGFDM system to improve the Bit Error Rate
low OOB emission. In addition, GFDM includes a CP for every
(BER) performance. The standard convolution code (7, [171,133]), GFDM block instead of each symbol like in OFDM, which in
commonly used in most wireless systems is employed. Simulated turn enhances spectral efficiency. In recent years, a lot of work
BER performance of coded MGFDM (CMGFDM) along with has been carried out on GFDM performance evaluation [3], [4],
conventional GFDM is investigated. The results show an im- [5]. With the inherent non-orthogonal nature of conventional
provement in the performance of CMGFDM when compared GFDM systems, a degradation in performance occurs. Hence,
to MGFDM and GFDM systems. Also, an analysis on different a need for orthogonal filters with low OOB emission is
receivers namely, Zero Forcing (ZF) and Matched Filter (MF) desirable as a crucial part of a 5G communication system. [6]
is studied and the Peak-to-average power ratio (PAPR) of the investigates the advantage of using Discrete Prolate Spheroidal
proposed CMGFDM system is evaluated. Sequence (DPSS) or multi-tapers as a pulse shaping filter
Keywords—MGFDM, Convolution Code, Viterbi Decoding, which has lower OOB radiation compared to a Root Raised
BER, PAPR. Cosine (RRC) filter, used in conventional GFDM systems. In
fact, the choice of using multi-tapers makes the GFDM system
I. I NTRODUCTION a bit more orthogonal in nature.
The rapid growth of hand held devices among peo- During data transmission, random bit errors may occur due
ple today, demands a high data volume capability with to environmental interference and physical defects. For reliable
low round trip latency and ultra reliable communications. communications, Forward Error Correction (FEC) codes are
Additional capabilities such as the support of more de- most commonly used in order to overcome the adverse effects
manding applications/services (e.g. smart cities including of the Additive White Gaussian Noise (AWGN) channel.
smart home/buildings), massive machine type communica- FEC codes are classified into convolution and block codes.
tions, MTC, and Internet of Things, IoT (self driving cars, Most of the standards employ a combination of OFDM, as
industry automation), will truly revolutionize connectivity ev- the multicarrier framework and a convolution code using a
erywhere. 5G should be E 4 connect- Everything, Everywhere, Trellis structure with a constraint length of 7 and generating
Every time and Every path connected. In summary, 5G gives polynomials 1718 and 1338 . The usefulness of convolutional
us a new user experience. coded OFDM to Digital Audio Broadcasting (DAB) system in
Europe is demonstrated in 1995 [7].
Orthogonal Frequency Division Multiplexing (OFDM), the
most widely adopted multicarrier scheme in most present However, the investigation of these convolution codes in
standards offers many advantages [1]. An elegant way of multi-taper GFDM (MGFDM) systems has not been reported
executing fast Fourier transform (FFT) blocks in order to anywhere. Convolution codes improve performance in OFDM
avoid complex oscillators/modulators and robustness against multicarrier systems tremendously and the same follows in
frequency selective multipath fading channels are two of the conventional GFDM and MGFDM systems. In this article, we
properties of OFDM that make it the method of choice in consider an MGFDM communication system with a convolu-
current wireless standards like IEEE 802.11a/g, Digital Au- tion code at the transmitter side for efficient data transmission
dio/Video Broadcasting (DAB/DVB) and Long Term Evolu- and this adds redundancy to the overall transmitted data in the
tion (LTE) implementations. Despite these advantages, OFDM form of parity information. At the receiver side, the detection
suffers from some critical performance issues, such as carrier scheme is carried out by using Viterbi decoding [8] and a
frequency offset (CFO), Peak-to-average power ratio (PAPR), fair amount of channel errors can be corrected by exploiting
out of band (OOB) emission, due to the rectangular filters the redundant data that are transmitted. These coded MGFDM
employed and the cyclic prefix (CP) usage for every symbol, (CMGFDM) systems can tolerate more interference and noise

978-1-5090-4032-2/17/$31.00 ©2017 IEEE 545 ICTC 2017


     



    
 

     


     



   
 ! " 

 % &
   

" 
'
 

" ! "
  #
$

Fig. 1. Transceiver schematic of Coded MGFDM system



effects.




 
The goal of this article is to develop a robust convolu- 

tion coded MGFDM system which outperforms conventional

GFDM system in terms of BER. Furthermore, we investigated
the performance of Peak-to-average power ratio (PAPR) of 
CMGFDM, MGFDM and conventional GFDM systems. The
paper is organised as follows. In Section II, the system model
of a Coded MGFDM system is outlined, along with the con- Fig. 2. Convolutional encoder with shift registers
volution encoder, Code Trellis diagram and Viterbi Decoding.


  
The results are discussed in Section III. Finally, section IV    
concludes the paper.


II. S YSTEM MODEL
   




In this section we will first briefly outline the details
about the transceiver structure of a Coded MGFDM system 
 





model along with Viterbi decoding. The baseband transceiver
structure of CMGFDM is shown in Fig. 1.

   


A. CMGFDM Transmitter Fig. 3. Trellis Diagram

The data source generates a binary vector b that is supplied
to the convolution encoder as shown in Fig. 1. Convolutional
codes which are widely used as channel codes for error bits for p number of message bits. A widely used decoding
correction are incorporated in this system model. In essence, algorithm for convolutional codes is the Viterbi algorithm.
convolution coding is performed by appropriately combining Before starting with the Viterbi decoding algorithm expla-
the fixed number of input bits. These bits are stored in the nation, a simple mathematical analysis has been developed
fixed length shift registers and are combined with the help of starting with the states of the encoder. Referring to Fig. 2 the
modulo-2 adders [9]. This operation is equivalent to binary previous two successive message bits c1 and c2 represent the
convolution and hence is called convolution coding. In order state of the encoder. The input message bit c effects the ’state’
to have an idea of the encoding process we start with the of the encoder as well as the outputs s1 and s2 during that
simple block diagram of a convolutional encoder with code rate state. Whenever a new message bit is shifted to c, the contents
1/2. Fig. 2 illustrates that the information bits are fed in small of c1 and c2 define the new state. Also, the outputs s1 and s2
groups of p-bits at a time to a shift register. The output encoded are changed according to the new state c1 , c2 and message bit
bits are obtained by modulo-2 addition (EXCLUSIVE-OR c. Here, for p = 1 the number of encoded bits for one message
operation) of the input information bits and the contents of bit is q = 2. Also the outputs are expressed as,
the shift registers which contain a few previous information
bits. In brief, the encoder generates a group of q encoded bits s1 = c ⊕ c1 ⊕ c2 (2)
per group of p information bits, where the code rate R has the
following representation, s 2 = c ⊕ c2 (3)

p
R= (1) TABLE I. S TATE TABLE
q
C1 C2 State of the encoder
0 0 a
where p < q. Generally, q and p are significant in determining 0 1 b
the dimension of the code. Here, the dimension of the code 1 0 c
is denoted by (p, q) where, q represents the encoded output 1 1 d

546
 

 





 









 








 






 











Fig. 4. Details of MGFDM modulator

The Trellis diagram of the convolutional code is obtained to the circular convolution which leads to tail biting process
from its state diagram. All state transitions at each time step [2].
are explicitly shown in Fig. 3. Usually, supporting descriptions
on state transitions, corresponding input and output bits etc., An alternate matrix representation is written as,
are labeled in the Trellis diagram. It is interesting to note
that the Trellis diagram, which describes the operation of the x = Ad (6)
encoder, is very convenient for also describing the behavior of
the corresponding ’Viterbi Algorithm’ decoder.
where A denotes the (M)GFDM modulation matrix incorpo-
The output of the encoded data is interleaved. The in- rating all the signal processing steps of upconversion, pulse
terleaved data is passed through the P −QAM mapper to shaping and frequency shifting and can be given in matrix
map groups of μ bits, where μ denotes the modulation order notation as [5],
representing the number of bits per symbol of a chosen
modulation scheme. Pilot symbols are later inserted at precise
A = [G E1 G...EK−1 G]
locations that are known at the receiver side. The resultant
data block d is given by, d = {dk (m)} ∀ k  (0, K − 1)
and m  (0, M − 1). It is to be noted that each data symbol with G a KM × M transmitter filter matrix containing the
dk (m) corresponds to the information transmitted on the k th first M tapers each of length KM for MGFDM system and
subcarrier and mth subsymbol. The data vector of length KM G is a KM × M filter matrix where the first column of the
generated from the mapper unit along with the pilot symbols matrix consists of the KM samples of the prototype filter and
are passed through the MGFDM modulator. the subsequent columns of the matrix are the K times shifted
versions of the previous columns.
A detailed schematic of MGFDM modulator is depicted  
in Fig. 4. Initially, in the MGFDM modulator, the symbols Ei = diag [ei T , ..., ei T ]T is a KM ×KM diagonal ma-
are upsampled by a factor of K. Unlike conventional GFDM trix whose diagonal elements are comprised of M concatenated
2πi 2πi
where the filter coefficients are circularly shifted, here, in copies of the vector ei = [1, ej K , ..., ej K (K−1) ]T . Lastly,
MGFDM, the prototype filter is designed to use DPSS tapers a cyclic extension is performed in order to combat channel
where each taper is orthogonal to every other. Finally, subcar- effects.
rier upconvertion is done accordingly as shown in Fig. 4.
In a later part, a cyclic extension is transmitted along with B. CMGFDM Receiver Structure
the output of the MGFDM modulator for better protection The received signal after adding AWGN is given by,
against channel distortions. The transmitted signal x(n) of
each MGFDM block is given by [6],
R=x+W (7)

K−1
M −1
 2
where R is the noise affected signal, W ∼ CN (0, σw IKM )
x(n) = dk (m)gk,m (n) n = 0, 1, .., KM − 1 (4) are the complex additive white Gaussian noise samples with
k=0 m=0 2
noise variance σw and IKM is the identity matrix of order
where KM .
At the receiver side, the initial cyclic prefix is removed
j2πkn form the corrupted signal. Then, in the MGFDM demodulator,
MGFDM: gk,m (n) = gm (n)e K (5) the signal will undergo a series of operations that are opposite
j2πkn
GFDM: gk,m (n) = g[(n − mK) mod KM ]e K to that of the MGFDM modulator at the transmitter side. The
demodulated output can be expressed as,
with gm (l) being the lth coefficient of the mth taper. Observe
that the modulo operation to the filter coefficients corresponds Y = BR (8)

547
 
 
 
 


 










  
   








 

 
 









 

Fig. 5. Viterbi Decoding

TABLE II. S IMULATION PARAMETERS

where B under Zero Forcing (ZF) and Matched Filter (MF)
receivers are given by, Description Parameter Value
Number of Subcarriers K 32
Number of time slots M 2
Pulse shape filter g DPSS, RRC
 −1 H Roll-off factor α {0.1, 0.9}
ZF: BZF = AH A A (9) Length of CP Ncp 16
H Number of pilots Np 4
MF: BMF = A Modulation order μ 4

Later, pilot synchronization is performed, followed by

100
symbol demapping to bits. The data is de-interleaved and
passed to the Viterbi decoder where the convolutional encoded
data is decoded.
10-1
We perform maximum likelihood or Viterbi decoding to
Bit Error Rate (BER)

regenerate the sequence of data bits that are actually the

convolution encoder input used for encoding the message
or information bits. The input to the Viterbi decoder is the 10-2
received bit stream and now the main purpose of the Viterbi
decoder is to decode the received stream of bits [9].

Here, the Viterbi algorithm utilizes the Trellis diagram to 10-3

compute the path metrics. The decoding operation starts from Coded MGFDM
state  00 , i.e. with the assumption that the initial state of Coded OFDM
MGFDM
the encoder is  00 . With the receipt of one noisy codeword, -4 GFDM, α=0.5
10
the decoding operation progresses by one step deeper into 0 2 4 6 8 10 12 14 16 18
the Trellis diagram. The branches, associated with a state of Signal-to-Noise Ratio (SNR) in [dB]
the Trellis tell us about the corresponding codewords that the
encoder may generate starting from this state. Hence, upon Fig. 6. BER comparison under AWGN channel with K=32, M =2, μ = 4
receipt of a codeword, it is possible to note the ’branch metric’ using ZF receiver
of each branch by determining the Hamming distance of the
received codeword from the valid codeword associated with
that branch. Path metrics of all branches, associated with all III. R ESULTS AND D ISCUSSIONS
the states are calculated similarly. And now, at each level of
the Trellis, each state also carries some ’accumulated path The simulation results in terms of BER and PAPR, which
metric’, which is the addition of all the metrics of all branches compare the coded MGFDM system to that of the conventional
that construct the ’most likely path’ to that state. At any GFDM and MGFDM systems are reported in this section.
depth of the Trellis, each state can be reached through two The simulation parameters that are used for investigation are
paths, one from the previous stage and one from the Viterbi displayed in Table II with pilot locations at 1, 14, 27, 40. The
Algorithm. The path with lower accumulated path metric is channel is assumed to be Additive White Gaussian Noise
chosen. Finally, the decoded binary bits are collected at the (AWGN). At the receiver, we have used ZF and MF receiver
data sink as shown in Fig. 1. structures in the (M)GFDM demodulator to retrieve the data.

548
 100
0
10

-1
10-1
CCDF (Pr[PAPR>PAPR0])

10

Bit Error Rate (BER)

(i)

-2 10-2
10

10-3 (ii)
10-3 GFDM, α=0.9
MGFDM with Convolution CMGFDM-MF Receiver

GFDM, α=0.5 CMGFDM-ZF Receiver

MGFDM 10-4
-4 0 2 4 6 8 10 12 14 16 18
10
4 6 8 10 12 SNR (dB)
PAPR 0 [dB]
Fig. 8. BER performance of CMGFDM with ZF and MF Receivers for
Fig. 7. Comparison plot of PAPR analysis K=32, M =2, 16-QAM

The Bit error rate performance of CMGFDM with ZF and

Fig. 6 compares the proposed coded MGFDM system MF receivers is depicted in Fig. 8 under AWGN channel. At
with an uncoded MGFDM system using the ZF receiver. As low SNRs both ZF and MF receivers are achieving the same
an example, for comparison purposes, we have utilized an error rates while at the higher ones, the error rate observed for
RRC pulse shaping filter, with a roll-off factor α = 0.5 MF is considerably increased as compared to the ZF receiver.
for uncoded GFDM. From the results, we can observe the A huge difference between the BER curves can be observed.
benefit of using a convolution code to an MGFDM system For example at an SNR of 14dB an error rate of 0.006 and
as the BER performance is much improved compared to the 0.05 is achieved and is shown in Fig. 8.
corresponding uncoded MGFDM and conventional GFDM
systems. For instance, at an SNR of 14dB, the BER for IV. C ONCLUSION
CMGFDM, MGFDM and conventional GFDM is given by
0.006, 0.06 and 0.6 respectively. This improvement in system In this manuscript, a novel convolution coded MGFDM
performance comes at a cost of adding redundant bits to system is proposed. Simulation results suggest that the pro-
the original MGFDM system and hence there is a trade off posed CMGFDM BER performance is almost similar to that of
involved. Lastly, for comparison purposes, we have plotted the coded OFDM scheme in an AWGN channel. Both exhibit
the corresponding performance curves for an OFDM structure an increased robustness to noise interference compared to
with 64 subcarriers using convolution code as a reference. As uncoded MGFDM and GFDM systems, thereby demonstrating
can be observed from the figure, coded OFDM and CMGFDM the advantage of the proposed system. The receiver structure is
have similar performance but it is to be noted that CMGFDM also seen to play a key role. CMGFDM was tested with both
adds one cyclic prefix for a whole block consisting of M the ZF and MF receivers. At low SNRs the error behavior was
subsymbols, whereas in coded OFDM the cyclic extension is similar while at higher SNRs the difference was dramatic, with
added to every symbol. Lower OOB radiation in CMGFDM ZF strongly outperforming MF.
using DPSSs is also observed, increasing the spectral efficiency Besides BER performance comparison, a comparative anal-
compared to the high OOB emission in OFDM. ysis of PAPR of the proposed system to that of MGFDM
and conventional GFDM was also investigated. In this case
For evaluating PAPR performance, we have considered 105 uncoded MGFDM displayed better performance that the coded
symbols and the PAPR CCDF is shown in Fig. 7. In this figure version, probably due to the effect of adding redundant bits
we compare the PAPR of the proposed CMGFDM system to for data encoding. This indicates that PAPR performance still
that of an MGFDM and a conventional GFDM system with remains an issue even with convolutional coding and requires
α = 0.5, 0.9. Interestingly, the CMGFDM PAPR is increased further investigation.
when compared to the MGFDM system, which may be due to
the effect of adding the redundant bits for encoding the data. Although the performance of CMGFDM is comparable
The PAPR of an MGFDM system is decreased in comparison to Coded OFDM, the advantage of lower OOB radiation
to a conventional GFDM system for any roll-off factors which in CMGFDM is to be noted. Hence for future 5G systems
is due to using orthogonal filters with DPSSs in the MGFDM coded MGFDM can be used effectively in applications such
system. Also note that, as we increase the roll-off factor from as machine to machine communication.
0.5 to 0.9 the PAPR of conventional GFDM is considerably
increased. For example, a CCDF of 10−3 is achieved at a PAPR R EFERENCES
of 10.58, 10.84, 11.17 and 11.4 respectively for MGFDM, [1] D. Tse and P. Viswanath, Fundamentals of wireless communication.
GFDM (α = 0.5), CMGFDM and GFDM with α = 0.9. Cambridge university press, 2005.

549
[2] G. Fettweis, M. Krondorf, and S. Bittner, “GFDM - Generalized Fre- pp. 1507–1518, March 2016.
quency Division Multiplexing,” in Vehicular Technology Conference, [6] S. K. Bandari, V. V. Mani, and A. Drosopoulos, “Multi-taper implemen-
2009. VTC Spring 2009. IEEE 69th, April 2009, pp. 1–4. tation of gfdm,” in 2016 IEEE Wireless Communications and Networking
[3] N. Michailow, S. Krone, M. Lentmaier, and G. Fettweis, “Bit error Conference, April 2016, pp. 1–5.
rate performance of Generalized Frequency Division Multiplexing,” in [7] P. Shelswell, “The cofdm modulation system: the heart of digital audio
Vehicular Technology Conference (VTC Fall), 2012 IEEE, Sept 2012, broadcasting,” Electronics Communication Engineering Journal, vol. 7,
pp. 1–5. no. 3, pp. 127–136, Jun 1995.
[4] M. Matthe, N. Michailow, I. Gaspar, and G. Fettweis, “Influence of [8] A. Viterbi, “Error bounds for convolutional codes and an asymptotically
pulse shaping on bit error rate performance and out of band radiation optimum decoding algorithm,” IEEE transactions on Information Theory,
of Generalized Frequency Division Multiplexing,” in Communications vol. 13, no. 2, pp. 260–269, 1967.
Workshops (ICC), 2014 IEEE International Conference on, June 2014,
[9] K. Moon Todd, “Error correction coding: mathematical methods and
pp. 43–48.
algorithms. 2005 by john wiley & sons,” ISBN 0-471-64800-0, Tech.
[5] A. Farhang, N. Marchetti, and L. E. Doyle, “Low-complexity modem Rep.
design for gfdm,” IEEE Transactions on Signal Processing, vol. 64, no. 6,

550