You are on page 1of 16

IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 67, NO.

13, JULY 1, 2019 3455

Theoretical Analysis of the Peak-to-Average Power


Ratio and Optimal Pulse Shaping Filter
Design for GFDM Systems
Kaiming Liu , Member, IEEE, Weifeng Deng, and Yuan’an Liu, Member, IEEE

Abstract—Generalized frequency division multiplexing (GFDM) agility, etc.. These properties make it appropriate for future in-
has been regarded as a candidate for new multicarrier modulation novative applications, such as Machine-Type Communications
schemes in future wireless communication systems (e.g., 5G sys- (MTC), Internet of Things (IoT), Cognitive Ratio (CR), etc.
tems). However, GFDM systems still suffer from the problem of [3]–[8].
high peak-to-average power ratio (PAPR). In this paper, theoretical However, similar to other multicarrier modulation schemes,
analysis of the PAPR distribution for GFDM signals is performed.
Exact closed-form expressions for the complementary cumulative
GFDM still suffers from the problem of high Peak-to-Average
distribution function (CCDF) of PAPR are firstly derived for criti- Power Ratio (PAPR) [8], [19]–[26], and several PAPR reduction
cally sampled and oversampled GFDM signals, respectively, which schemes have been proposed, e.g., clipping with iterative recep-
tend to be sufficiently accurate as the number of subcarriers K is tion method [23], companding transforms [24], [46], and linear
large enough (e.g., K ≥ 64). With aid of theoretical analysis re- precoding techniques [25], [26].
sults, an optimization criterion for the design of pulse shaping fil- Theoretical analysis of the PAPR can be much helpful for
ters is obtained, which can nearly achieve the globally minimum understanding the PAPR properties of multicarrier signals, as
CCDF results and is appropriate for practical applications. This well as for designs of PAPR reduction schemes (e.g., for non-
criterion can also achieve the globally optimal result for minimiz- linear companding transforms [14]–[16]) and optimizations of
ing the variance of instantaneous power. Based on this criterion, a various nonlinear devices [9]–[18]. The well-known approxi-
measurement for the overall deviation of filter coefficients is pro-
posed, which can help to roughly and efficiently evaluate the ability mated expressions for the complementary cumulative distribu-
of PAPR reduction for pulse shaping filters. Although the PAPR tion function (CCDF) of PAPR have been derived in Orthogonal
reduction effect is limited with pulse shaping filters, it is found Frequency Division Multiplexing (OFDM) systems [9]–[11] and
that large deviation may also lead to higher PAPR. Additionally, Filter-Bank based Multi-Carrier with Offset Quadrature Ampli-
with the aid of closed-form CCDF expressions, PAPR character- tude Modulation (FBMC/OQAM) systems [12], [13]. Theoreti-
istics of GFDM and other multicarrier systems (e.g., OFDM and cal results in OFDM systems indicated that the CCDF of PAPR
FBMC/OQAM) are compared thoroughly. Simulation results ver- was mainly determined by the number of subchannels, and have
ify the validity and accuracy of derived theoretical results, and been employed for PAPR reduction designs [14]–[16]. Theoret-
demonstrate the effectiveness of the proposed optimization design ical results in FBMC/ OQAM systems indicated that the CCDF
criterion. of PAPR was determined by the number of subchannels and the
Index Terms—GFDM, PAPR, CCDF, Pulse shaping filter. pulse shaping filters, and have also been adopted for the design
of PAPR reduction techniques [17], [18]. However, due to the
difference of the signal structure between GFDM and the other
I. INTRODUCTION two multicarrier systems [2], [27], [30], existing theoretical re-
sults cannot be implemented to GFDM systems directly. Actu-
ENERALIZED Frequency Division Multiplexing
G (GFDM) has been extensively studied in recent years and
regarded as one of the promising candidates for new modulation
ally, different pulse shaping filters between GFDM and OFDM
systems, and different overlapping patterns of complex-valued
signals between GFDM and FBMC/OQAM systems, lead to
schemes in future wireless communication systems (e.g., 5G
different PAPR distribution properties.
systems) [1], [2], due to its good properties including low
To the best of authors’ knowledge, there have been no
latency, low out-of-band (OOB) emission, robust against carrier
theoretical analysis results for PAPR distribution properties
frequency offset (CFO) and timing offset (TO), and spectrum
of GFDM signals in the literature yet. Previous analysis work
for PAPR distribution properties of GFDM signals was mainly
through simulations [19]–[22], which might be inconvenient
Manuscript received July 11, 2018; revised February 20, 2019 and April 15, and insufficient. Meanwhile, an additional issue to be considered
2019; accepted May 6, 2019. Date of publication May 28, 2019; date of cur-
rent version June 5, 2019. The associate editor coordinating the review of this for GFDM systems may be the optimization of pulse shaping
manuscript and approving it for publication was Prof. Chandra Ramabhadra filters for minimizing the CCDF of PAPR, which has been
Murthy. This work was supported by the National Natural Science Foundations rarely addressed in the literature yet. In [44], authors researched
of China under Grants 61701041 and 61327806. (Corresponding author: Kaim- into the design of prototype filters for PAPR minimization while
ing Liu.)
The authors are with the School of Electronic Engineering, Beijing Uni-
maintaining the performance of OOB radiation and symbol
versity of Posts and Telecommunications, Beijing 100876, China (e-mail: error rate (SER), and only provided an algorithmic solution.
kmliu@bupt.edu.cn; dengweifeng@bupt.edu.cn; yuliu@bupt.edu.cn). Although the effect of PAPR reduction by pulse shaping filters
Digital Object Identifier 10.1109/TSP.2019.2919380 may be much limited compared with other PAPR reduction
1053-587X © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
3456 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 67, NO. 13, JULY 1, 2019

methods (e.g., methods in [23]–[26], [46]), it is still valuable to


address the theoretical optimization criterion, so as to evaluate
the PAPR reduction performance of candidate filters more
accurately and efficiently, make the best use of filters, and avoid
inferior design leading to serious performance degradation. On
the other hand, filter optimizations with other targets have been
performed for GFDM systmes (e.g., rate maximization and
OOB emission minimization in [31], minimizing receiver mean
square error (MSE) in [21]). Theoretical analysis results as well
Fig. 1. Frame diagram of GFDM transmitter.
as optimization criteria can also make joint optimization more
efficiently.
Additionally, it should be noted that compared with PAPR,
the variance of instantaneous power (VIP) is more related to The remainder of this paper is organized as follows. Section II
the nonlinear distortion caused by the power amplifier (PA) in describes the mathematical model of GFDM signals. Theoretical
multicarrier systems, especially with low input backoff (IBO) analysis of PAPR distributions is presented in Section III.
values [40]. Minimizing VIP has also been adopted as the opti- Section IV provides the optimization problems and solutions
mization target for designs of selected mapping (SLM) methods for the design of pulse shaping filters. The PAPR compar-
in OFDM system [40], precoding schemes in OFDMA system isons among multicarrier systems are presented in Section V.
[41] and GFDMA system [26], and prototype shaping vectors in Section VI provides simulation results, and Section VII draws
circularly pulse-shaped (CPS) precoding OFDM (CPS-OFDM) conclusions.
system [43]. It has been verified that minimizing VIP can not
only contribute to reduce PAPR statistically, but also help to re-
duce OOB emission and system bit error rate (BER). Thus VIP
may also be a valuable metric for the optimization design of II. GFDM SIGNAL MODEL
prototype filters in GFDM systems.
The main contributions of this paper can be summarized as A. GFDM Transmit Signals
follows: The GFDM signal is characterized by the block structure and
r An exact closed-form expression of CCDF for PAPRs of circular pulse shaping. In each GFDM signal block, M complex
critically sampled GFDM signals is firstly derived. This data symbols are transmitted on each of the K subcarriers, thus
expression tends to be sufficiently accurate as the number each GFDM signal block totally contains KM complex-valued
of subcarriers K is large enough (e.g., K ≥ 64). This expres- data symbols.
sion facilitates the evaluation of PAPR distribution prop- The block diagram of the GFDM transmitter is depicted in
erties of GFDM signals. It is verified that this expression Fig. 1. The serial complex-valued baseband data symbols are
can be a generalized approximation for PAPR distributions firstly converted into K parallel data symbol streams, and in
of other multicarrier signals (e.g., OFDM, FBMC/OQAM each symbol stream, the mth (m ࢠ {0, 1, …, M−1}) subsymbol
signals). By introducing an exponential modifying factor, is sampled by an oversampling factor Q (Q ࢠ Z+ ) and shaped
an extended expression is obtained, which is appropriate by a circularly time-shifted version of the prototype pulse shap-
for calculating the CCDF of PAPR for oversampled GFDM ing filter. The prototype filter has a total length of L = QKM
signals and can help to evaluate the CCDF of PAPR for and circularly shifts QK samples for each new coming subsym-
practical GFDM signals more accurately and efficiently. bol. Then, data samples are modulated onto K subcarriers and
r Two optimization problems of pulse shaping filter design added together to yield GFDM signals. Finally, a cyclic prefix is
are formulated and mathematically optimal solutions are added to each GFDM signal, and generated signals pass through
obtained. The first one targets to minimize the CCDF of the digital-to-analog (D/A) converter and high-power amplifier
PAPR directly, while the other one targets to minimize the (HPA), and are sent into the wireless channel.
variance of instantaneous power (VIP). With aid of the- For the sake of simplicity, we set the oversampling factor Q
oretical CCDF result for GFDM signals, locally optimal = 1, meaning GFDM subsymbols are critically sampled [25]
solutions for the first problem are derived and compared and the total length of prototype filter L = KM. From above de-
thoroughly. The preferred one solution is proved to be the scriptions, the nth sample of each GFDM signal corresponding
globally optimal solution for the second problem, and can to one signal block can be written as [3]
be an appropriate optimization criterion for practical appli-
cations. Based on this criterion, we also define a measure- −1 K−1
ment which can help to roughly and efficiently evaluate the 
M  2πkn

PAPR reduction ability of pulse shaping filter candidates. x [n] = dk [m] ĝm [n] ej K , n ∈ S, (1)
m=0 k=0
Additionally, we analyze and demonstrate the compatibil-
ity of the proposed optimization criterion with some of
other existing optimization criteria. where dk [m] (k ∈ {0, 1, . . . , K − 1}) denotes the mth base-
r With the aid of theoretical CCDF results, we provide band data symbol carried by the kth subcarrier in current signal
more comprehensive and universal comparison results of block, ĝm [n] = gTX [(n − mK) mod KM ] is the nth sample
PAPR distributions among common multicarrier systems, of the mth circularly time-shifted prototype filter which has a
which can help to know about PAPR distribution prop- coefficient vector of gTX = [gTX [0] gTX [1] . . . gTX [KM − 1]],
erties of these systems more clearly and guide practical and S = {0,1, …, KM−1} is the index set of samples in each sig-
applications. nal block.
LIU et al.: THEORETICAL ANALYSIS OF THE PAPR AND OPTIMAL PULSE SHAPING FILTER DESIGN FOR GFDM SYSTEMS 3457

B. PAPR Definition for GFDM Signals and the variance of x[n] can be expressed by
As we known, the problem of high PAPR in multicarrier sys-   M −1

tems stems from the randomly coherent superposition of modu- σn2 = E |x[n]|2 = σd2 K |gTX [(n − mK)
lated subcarriers. In GFDM systems, there is no interference and m=0
superposition between adjacent signal blocks, due to the block × mod KM ]| , n ∈ S
2
(6)
structure and circularly time-shifted property of pulse shaping
filters. Thus the definition of PAPR for GFDM signals can be Obviously, the sampled signal will be a non-stationary com-
expressed by [26] plex Gaussian stochastic process and different samples have dif-
  ferent variances, if the sum term in (6) changes at different sam-
max0≤n≤KM −1 |x [n]|2 ple index n with certain pulse shaping filters.
PAPR =   , (2) As the amplitude of a complex Gaussian random variable
2
E |x [n]| follows the Rayleigh distribution, we can get the cumulative
distribution function (CDF) of the amplitude of x[n] (i.e., |x[n]|),
where E{ r} denotes that expectation operator, and E{|x[n]|2 } which can be written as
 
indicates the average power of GFDM symbols. Equation (2) CDF|x[n]| (x) = Prob {|x [n]| ≤ x} = 1 − exp −x2 /σn2 ,
actually examines total KM samples (with oversampling factor (7)
Q = 1) to find the peak power in each signal block, at each PAPR
To obtain the PAPR distribution result, we firstly consider the
value evaluation. We can see that by this definition, PAPR is
crest factor (CF), a measurement of the peak-to-average ampli-
evaluated in a block-by-block way. Equation (2) is reasonable
tude ratio of x[n], which is defined as
since all subsymbols in each signal block superpose together and
are generally detected together at the receiver in GFDM systems, maxn∈S {|x [n]|}
CF|x[n]| = √ , (8)
and there is no interference between adjacent signal blocks [2]. P̄
where P̄ = E{σn2 } denotes the average power of GFDM signals.
III. PAPR ANALYSIS FOR GFDM SYSTEMS The CCDF of CF can be written as
In this section, we will analyze the theoretical PAPR distri- CCDFCF (ζ) = Prob {CF > ζ}
butions of GFDM signals and derive the closed-form CCDF

expressions of PAPR for GFDM signals, which tend to be accu- = 1 − Prob max {|x [n]|} ≤ ζ P̄ , (9)
n∈S
rate enough as the subcarrier number K increases. Additionally,
we will explore the relationship among the theoretical CCDF where ζ is the target threshold of CF.
expressions for GFDM signals and those for other multicarrier It can be easily proved that |x[n]| (n ∈ S) are independent to
system (i.e., OFDM, FBMC/OQAM) signals. one another (in fact, ξm [n] are independently zero-mean com-
plex Gaussian random variables). Thus we have


A. Derivation of Theoretical CCDF Expressions of PAPR for Pr ob max {|x [n]|} ≤ ζ P̄
GFDM Signals n∈S

Firstly, equation (1) can be rewritten as −1


KM  
= Prob |x [n]| ≤ ζ P̄ . (10)

M −1
n=0
x [n] = ξm [n]gTX [(n − mK) mod KM ] , n ∈ S (3) According to (6), (9) and (10), we can obtain
−1   2 
m=0
KM
K−1 ζ P̄
where ξm [n](m = 0, 1, . . . , M − 1) denotes k=0 dk [m] CCDFCF (ζ) = 1 − 1 − exp − 2 . (11)
σn
ej2πkn/K . n=0
Assume that baseband data symbols on different subcarriers Note that CF is just the square root of PAPR. Thus we can
and different subsymbols are independent to one another, and easily obtain the CCDF of PAPR for GFDM signals, by replacing
have identical distribution. When the number of subcarriers K ζ 2 with the target PAPR threshold γ in (11), i.e.,
is large enough, according to the Central Limit Theorem (CLT), CCDFP AP R (γ) = Prob {PAPRGFDM > γ}
ξm [n] can be modeled as a complex circular symmetric Gaussian
variable [38], i.e., −1 
KM 
γ P̄
 KM −1
  =1− 1 − exp − 2 =1−
ξm [n] ∼ CN 0, Kσd2 , n ∈ S, (4) n=0
σn n=0
  
where CN (·, ·) denotes the complex Gaussian distribution and −γ P̄
σd2 denotes the variance of input data symbol dk [m] (k ࢠ {0,1, …, × 1 − exp  −1 .
K − 1}, m ࢠ {0, 1, …, M − 1}). As the linear combination of σd2 K M m=0 |gTX [(n−mK) mod KM ]|
2

independent Gaussian random variables is still Gaussian, we can (12)


obtain Let ψn (n ∈ S) be the nth superposed coefficients of M circu-
 −1
larly time-shifted pulse shaping filters in one block, i.e.,

M
2
x [n] ∼ CN 0, σd K 2
|gTX [(n − mK) mod KM ]| , 
M −1
m=0 ψn = |gTX [(n − mK) mod KM ]|2 , n ∈ S. (13)
(5) m=0
3458 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 67, NO. 13, JULY 1, 2019

Due to the circularly time-shifted superposition of the pulse The appropriate value of ε is obtained through simulations
shaping filter, we can find that ψn is a periodic sequence, i.e., and can be set as ε ≈ 2.8. We will evaluate the accuracy of this
ψn = ψn + iK , n ∈ SP ; i = 0, 1, . . . , M − 1. (14) modified CCDF expression in Section VI.
Most of previous works have assumed odd M together with
where SP = {0, 1, . . . , K − 1} is the subset of indices in each even K for analysis, due to the singularity of the transmit matrix
periodic interval. Thus (12) can be simplified to be A with even M and K [2]–[5], [33]. In [45], authors proposed an
−1 
KM 
γ P̄
 approach to support even M and K by introducing a fractional
CCDFP AP R (γ) = 1 − 1 − exp − 2 shift in frequency domain to modify pulse shaping filters. This
n=0
σd Kψn approach has little change on signal statistical property, thus all
  M
K−1
above derivations can still hold and derived theoretical CCDF
γ P̄ results can be applicable for even M and K systems.
=1− 1 − exp − .
n=0
σd2 Kψn
(15) B. Relationship of Theoretical CCDF Expressions
Equation (15) implies that the CCDF of the PAPR in GFDM As GFDM system serves as a framework for multicarrier sys-
systems is determined by system parameters (i.e., K and M), tems [29], we will explore the relationship between theoretical
the average power of input baseband data symbols (i.e., σd2 ) and CCDF expressions of PAPR for GFDM signals and those for
GFDM signals (i.e., P̄ ), and superposition of pulse shaping filter other multicarrier signals (i.e., OFDM and FBMC/ OQAM sig-
coefficients (i.e., ψn ). nals).
Furthermore, if the pulse shaping filter has normalized total Firstly, we consider the relationship between GFDM and
power, i.e., OFDM systems. As we know, GFDM signals will turn into
OFDM signals when the rectangular filter is employed and the
−1
KM
subsymbol number M equals one (i.e., there is none circularly
|gTX [n]|2 = 1 (16) time-shifting and coefficient superposition with the pulse shap-
n=0
ing filter) [2]. As generally adopted, the rectangular filter is set
we should have to have normalized total power. Then the coefficients of the rect-

K−1 M
K−1 −1
 angular filter can be expressed by
ψn = |gTX [(n − mK) mod KM ]|2 = 1,  √
1/ K, n = 0, 1, 2, . . . , K − 1,
n=0 n=0 m=0 gTX [n] = (21)
(17) 0, n = K, K+1, . . . , KM − 1.
According to (6), (13) and (17), there will be If above filter is employed, the CCDF expression (19) will be

K−1 CCDFP AP R, Rect (γ) = 1 − [1 − exp (−γ)]KM . (22)
σn2 = Kσd2 . And when M = 1 holds, (21) will turn into
n=0
CCDFP AP R,I (γ) = 1 − [1 − exp (−γ)]K , (23)
From (6), (13) and (14), we can find that σn2 (n ∈ S) also has
a periodicity of K, i.e., σn2 = σn+iK
2
(n ∈ SP ) Thus we have which is just the well-known CCDF expression of PAPR for
K−1  OFDM signals (see equation (3) in [9]).
 2  Note that (22) actually evaluates the CCDF of PAPR in con-
P̄ = E σn = E σn /K = σd2 .
2
(18) secutive M transmitted GFDM symbols, when the rectangular
n=0 filter is employed. It can be viewed as introducing a time-domain
Equation (18) indicates that the pulse shaping filter with nor- parameter into PAPR evaluations.
malized power will not change the average power of input data And then, we investigate the relationship between GFDM and
symbols. Equation (15) can further be simplified to be FBMC/OQAM systems. Comparing the transmitter structures of
  M
K−1 these two systems, we can find that GFDM signals will turn into
γ FBMC/OQAM signals if we set M = 1 and extend the length of
CCDFP AP R (γ) = 1 − 1 − exp − . (19)
n=0
Kψn pulse shaping filter to be KQ. Thus the superposition of coeffi-
cients still exists due to the overlapping of pulse shaping filters
Above theoretical CCDF results are derived under the hypoth- for consecutive transmitted symbols. Under this conditions, and
esis of critical sampling (Q = 1). The practical continuous signal considering the time-overlapping property of prototype filters,
x(t) after the D/A converter may has higher PAPR, since discrete ψn defined in (13) should be expressed by
samples may not include all peaks of x(t) [11]. Thus (19) may 
loss accuracy for continuous signals. ψn = |hFBMC [n − qK]|2 , n ∈ SP , (24)
It has been proved that if the oversampling factor Q is large q∈Z
enough (e.g., Q ≥ 4), discrete samples x[n] show almost the same where hFBMC [n] is the nth sample of the prototype filter em-
PAPR distribution to that of continuous signals [11]. However, ployed by FBMC/OQAM systems. Substituting (24) into (19)
the derivation process for (19) does not hold for oversampled and setting M = 1, we will obtain
signals due to the dependence among adjacent samples. By em-
CCDFP AP R,II (γ)
ploying similar method in [11], we introduce an exponential
modifying factor ε to extend (19) for oversampling signals. The   

K−1
γ
modified CCDF expression can be expressed by =1− 1 − exp −  ,
K · q∈Z |hFBMC [n − qK]|2

QK−1 
γ
M ε n=0
CCDFP AP R (γ) = 1 − 1 − exp − . (25)
n=0
Kψn
(20)
LIU et al.: THEORETICAL ANALYSIS OF THE PAPR AND OPTIMAL PULSE SHAPING FILTER DESIGN FOR GFDM SYSTEMS 3459

which is just same to the CCDF expression of PAPR for 0, thus the exponential function exp(−1/x) is right-continuous
FBMC/OQAM signals (see equation (20) in [13]). at the point of x = 0.
From above descriptions, we can see that the derived CCDF As the target function defined in (28) is not a convex function
expression (15) can be viewed as a generalized approximation entirely in whole feasible region defined by (29) and (30), this
for the CCDF of PAPR for above multicarrier systems. This ex- optimization problem may have none global minimizer. In fact,
pression can help us evaluate the PAPR characteristics of multi- we can find that this optimization problem may mathematically
carrier systems more efficiently. Moreover, it can also guide us have two different forms of solution vector Ω with variant
to explore solutions for PAPR reductions. number of subcarriers and target PAPR thresholds, and so does
the desired Ψ. More specifically, for a given K and relatively
IV. OPTIMIZATION DESIGN FOR PULSE SHAPING FILTERS large γ satisfying (A-31), the locally optimal coefficient vector
should be
From (15) and (19), we can see that the coefficients of
pulse shaping filters actually influence the PAPR distribution ΨM inP D = (1/K, 1/K, . . . , 1/K )T , (31)
of GFDM signals, which makes it meaningful to seek optimiza-   
K
tion design criteria for pulse shaping filters in terms of PAPR
reduction. and the nth superposed coefficients should be expressed by
Based on theoretical CCDF results derived above, the opti-
mization target can be minimizing the PAPR distribution di- 
M −1
rectly, and we call this problem as Minimization of PAPR Dis- ψn = |gTX [(n − mK) mod KM ]|2 = 1/K, n ∈ SP .
tribution (MinPD). Additionally, as mentioned in Section I, it m=0
is valuable to adopt the target of minimizing VIP for optimiza- (32)
tion design, and we call this problem as Minimization of VIP
(MinVIP). While for a given K and relatively small γ dissatisfying
Without loss of generality, hereafter we still assume that pulse (A-31), the locally optimal coefficient vector should be in the
shaping filters have normalized total power. Additionally, for the form of
sake of simplicity, we still set the oversampling factor Q = 1.
The derived results can easily be extended to other scenarios by Ψ̂M inP D = (ψM ax , ψ̄, . . . , ψ̄ )T , (33)
  
some minor modifications. K−1

A. Minimization of PAPR Distribution (MinPD) Method where ψM ax is the largest element in Ψ̂M inP D and determined
Firstly, we seek the optimization design criterion with the by K and γ, and can be at any position in Ψ̂M inP D . All other ele-
target of minimizing the PAPR distribution directly, based on the ments are equal to be ψ̄ = (M − ψM ax )/(K − 1). More deriva-
relationship between theoretical CCDF results and coefficients tion details for (31)−(33) can be found in Appendix A.
of pulse shaping filters. We should compare (31) with (33) to determine which one is
According to (19), we can find that the PAPR distribution more appropriate for practical applications.
of GFDM signals is actually influenced by K superposed filter Firstly, we can see that the optimal coefficient vector defined
coefficients in one periodic interval. We define the vector of by (31) is only determined by K and has no relation to γ, which
these coefficients to be means that the optimality can universally hold for different tar-
get PAPR thresholds γ (relatively large enough). The particu-
Ψ = (ψ0 , ψ1 , . . . , ψK−1 )T , (26) lar solution result defined by (33) needs to be calculated with
where ( r)T denotes the transpose. Then the optimization target each given parameter group of {K, γ}, as indicated by (33) and
for the design of pulse shaping filter can be stated as finding (A-21), and will lose the optimality when γ increases to be large
the optimal Ψ to minimize the CCDF for certain target PAPR enough.
thresholds. For the sake of simplicity, we can also define an Secondly, it can be verified that even though the particular
auxiliary parameter of ωn = Kψn /γ(n ∈ SP ), and (19) will be solution result defined by (33) may lead to mathematically lower
CCDF for a given K and some relatively small γ when compared

K−1 
1
M
to (31), the superiority is so small that it can be ignored for
CCDFP AP R (γ) = 1 − 1 − exp − . (27) practical applications. Actually, the PAPR reduction ability of
n=0
ωn
the result defined by (33) is too weak for all γ and nearly useless
The optimization problem can then be formulated as for practical applications (see Appendix A and corresponding
 M 

K−1 
1
simulation results in Section VI).
Min 1 − 1 − exp − , (28) Consequently, (31) may be the preferred optimal criterion for
Ω
n=0
ωn the design of pulse shaping filters for practical implementations.
In fact, this criterion indicates that in order to reduce the CCDF

K−1
for certain PAPR thresholds as much as possible, the overlapped
s.t. ωn = K/γ, (29) pulse shaping filters tend to distribute the signal power uniformly
n=0 in the time domain.
ωn ≥ 0, ∀n ∈ SP , (30) Although the PAPR reduction may be limited only by the
optimized pulse shaping filter, the optimized filter can still help
where Ω = (ω0 , ω1 , . . . , ωK−1 ) is the desired solution vector
T
to reduce the PAPR to some extent and can relieve the burden of
of above optimization problem. Considering that the optimal other PAPR reduction algorithms to achieve the desired PAPR
Δ
solution result may lead to ωn = 0, we define exp(−1/x)|x=0 = distribution results.
3460 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 67, NO. 13, JULY 1, 2019

B. Minimization of VIP (MinVIP) Method


According to (3) and (18), the VIP averaged over a GFDM
signal block can be defined as [26], [42]

1 −1

KM
2
2
Pvar = E |x[n]| − P̄ . (34)
KM n=0

Based on (6), (13), (14) and (17), we can obtain

1 −1  
KM  
Pvar = E |x[n]|4 − 2σn2 σd2 + σd4
KM n=0

1 −1 
KM 
= E |x[n]|4 −σd4 . (35)
KM n=0 Fig. 2. CRM SE results for different filters with different roll-off factors.

According to (5), we can find that |x[n]|2 (n ∈ S) follows


the central chi-squared distribution with two degrees of freedom C. Measurement for Evaluation of PAPR Reduction Ability for
[43], i.e., |x[n]|2 ∼ χ2 (2). Note that the real part and imaginary Pulse Shaping Filters
part of x[n] have identical variance which equals to σn2 /2. Thus The optimization criterion of (30) can also be adopted as
the variance of |x[n]|2 can be calculated by a benchmark for evaluating the PAPR reduction performance
 
 2  2 2 of pulse shaping filter candidates. Here we define the normal-
σn ized root-mean-square-error (RMSE) of coefficients, CRM SE ,
Var |x[n]|2 = E |x[n]|2 − σn2 =4· = σn4 .
2 to evaluate the average deviation of actual ψn (n ∈ SP ) from the
(36) desired optimal value of 1/K for pulse shaping filter candidates,
which can be expressed by
According to (35) and (36), we obtain E{|x[n]|4 } = 2σn4 , and √
then (35) will be D F / K √
CRM SE = = K · D F , (41)
 1/K
2
KM−1 
K−1
Pvar = σ −σ = σd 2K
4 4 4
ψn − 1 . (37)
2
where || • ||F means the Frobenius norm, and D = [δ0 δ1 . . .
KM n=0 n d n=0 δK−1 ] is the deviation vector with each element δn defined as
Considering the constraint of (17), the optimization problem 
M −1
2 1
for MinVIP method can then be formulated as δn = gTX,C [(n − mK) mod KM ] − , n ∈ SP ,
m=0
K
Min Pvar (Ψ) , (38)
Ψ (42)

K−1 where gTX,C (n) denotes the nth coefficient sample of the filter
s.t. ψn = 1, (39) candidate. CRM SE = 0 indicates that the optimization criterion
n=0 of (31) is fulfilled and the filter candidate can maintain the low-
ψn ≥ 0, ∀n ∈ SP (40) est CCDF of PAPRs. As CRM SE increases, the filter candidate
might lead to higher PAPRs. A noteworthy feature of CRM SE is
We can find that this optimization problem has only one glob- that the increment of CRM SE is nearly linear to that of PAPR (in
ally optimal solution ΨM inV IP which is just same to the locally dB) at certain CCDF with a fixed M and different K (see simu-
optimal solution defined by (31), i.e., ΨM inV IP = ΨM inP D , lation results in Section VI). Thus it can be adopted as a general
and the minimum VIP should be σd4 . Thus we can achieve the measurement to roughly and efficiently evaluate the ability of
following corollary. PAPR reduction for pulse shaping filters, and be a benchmark
Corollary 1: The solution of ΨM inP D defined by (31) is the for filter design.
globally optimal solution for the MinVIP problem, and filters As an implementation example, we calculate correspond-
with coefficients satisfying the optimization criterion of (31) ing CRM SE for some common pulse shaping filters, including
can achieve the minimum VIP in GFDM systems. Dirichlet, Rectangular, 1st Xia, 4th Xia, Raised Cosine (RC)
Proof: See Appendix B. and Root Raised Cosine (RRC), to evaluate their potentialities of
This result indicates that the optimization criterion of (31) PAPR reduction. We can find that values of CRM SE for Dirichlet
actually minimizes the fluctuation of instantaneous power in and Rectangular filter always equal zero, indicating the optimal-
time domain, and so reduces the PAPR as much as possible. ity in term of PAPR reduction. Meanwhile, CRM SE values of
Note that the optimization criterion of (31) only relates to other four types of filters vary with roll-off factor, K and M (see
the amplitudes of filter coefficients in time domain. The frac- Fig. 2). From Fig. 2, we can see when roll-off factor is small
tional shifting approach proposed in [45] will not change the enough, some filters have small CRM SE near to zero, thus may
amplitudes of filter coefficients in time domain, since shifting have nearly optimal PAPR reduction capability. Larger roll-off
frequency response samples only lead to phase rotation with factor tends to increase CRM SE and degrade the PAPR reduc-
time response samples, thus the optimization criterion of (31) tion capability. CRM SE values can help us evaluate the PAPR
can be applicable for even M and K systems. reduction capabilities of different filters efficiently.
LIU et al.: THEORETICAL ANALYSIS OF THE PAPR AND OPTIMAL PULSE SHAPING FILTER DESIGN FOR GFDM SYSTEMS 3461

D. Comparisons of Optimization Criteria for Design of Pulse criterion of (31) can achieve statistically optimal performance
Shaping Filters in term of PAPR reduction, since it is entirely based on the
statistical property of transmit signals. Additionally, it declare
It should be noted that some other optimization criteria for
the relationship between filter coefficients and the CCDF result
the design of pulse shaping filters in GFDM systems have also
mathematically, which makes the filter design, as well as joint
been proposed in the literature [21], [31]. We can find that the optimization with OOB and SER, more definitely and efficiently.
proposed optimization criterion in terms of PAPR reduction can
be compatible with some of other criteria.
In [21], authors designed low complexity minimum mean V. PAPR COMPARISONS
square error (MMSE) receivers and proposed optimal filters For practical applications, one generally concerned is-
in terms of minimizing receiver MSE, based on the matrix sue may be the comparison of PAPR distribution property
characterization method. The optimal filters were found to be among different multicarrier systems (e.g., GFDM, OFDM, and
constant-magnitude-characteristic-matrix (CMCM) filters with FBMC/OQAM).
constant magnitude entries in corresponding characteristic ma- Some comparisons have already been performed between
trix of GFDM transmitter matrix. Authors also presented three GFDM system and other multicarrier systems [3], [19]–[22],
examples for CMCM filters, i.e., Dirichlet filter, a frequency- mainly through simulations with particular system parameters,
shifted version of Dirichlet filter, and a modified Dirichlet filter. certain pulse shaping filters and assumed conditions (e.g., equal
We can verify that the latter two filters also lead to CRM SE = 0. bandwidth in [3], [19]–[21], or equal subcarrier numbers in
Actually, the following corollary states that all CMCM filters can [22]). Due to the limitation of simulations, these comparison
satisfy the optimization criterion of (30) simultaneously. This results might be insufficient or lack of universality. With the aid
corollary can guide us to design pulse shaping filters achieving of theoretical analysis results presented in Section III, we could
lowest PAPRs while maintain minimized receiver MSE. perform comparisons more efficiently and obtain more universal
Corollary 2: If a prototype filter is a CMCM filter as de- conclusions.
fined in [21], it can satisfy the optimization criterion of (30) and Considering most general cases in practical applications, we
achieve the minimum PAPR distribution in GFDM systems. still assume that pulse shaping filters adopted by all multicarrier
Proof: See Appendix C. systems have normalized average power, and subcarrier numbers
In [31], authors researched into filter design problems focus- are large enough (i.e., ≥ 64). And all multicarrier signals are
ing on two optimization targets: rate maximization and OOB critical sampled for PAPR evaluations, which indicates that there
radiation minimization. Authors presented five optimization will be KM samples in each signal block of GFDM system,
problems (see Problem 1∼Problem 5 in [31]) and provided nu- and N samples in each symbol of OFDM and FBMC/OQAM
merical or algorithmic solutions respectively. Some of designed system. There is none particular constraint on the bandwidth or
filters can satisfy the optimization criterion of (30), while others subcarrier numbers.
cannot. Firstly, authors clarified that any filters satisfying the Firstly, we focus our comparisons on GFDM system and
power constraint were rate-optimal for GFDM systems with OFDM system. When optimal pulse shaping filters satisfying
the ideal nonlinear MF (matched filter)/SIC (successive inter- (31) are employed by GFDM system, the CCDF of PAPR will
ference cancellation) receiver in the AWGN channel. Secondly, just be same to (21), and can be rewritten as
authors proved that only the Dirichlet filter was rate-optimal for
GFDM systems with the zero-forcing (ZF) or MMSE receiver CCDFP AP R,GF DM (γ) = 1 − [1 − exp (−γ)]KM . (43)
in the AWGN channel. And then, authors provided numerical While according to (23), the CCDF of PAPR for OFDM sys-
examples of OOB radiation-optimal filters for GFDM-MF/SIC tem can be expressed by
and GFDM-ZF systems, respectively (see Fig. 5 in [31]). We
can verify that the proposed two filters lead to CRM SE = 0.70 CCDFP AP R,OF DM (γ) = 1 − [1 − exp (−γ)]N . (44)
and CRM SE = 0.35 respectively, thus cannot satisfy (30). Comparing (43) to (44), we can obtain following conclusions:
Other optimization problems (i.e., rate-optimal with carrier i) When KM > N, the PAPR of GFDM system should be
frequency offset, and joint design of filter and window) might be higher than that of OFDM system;
too complicated and need more elaborative computations and ii) When KM = N, the PAPR of GFDM system should be
tradeoffs, thus none explicit designed result was provided and equal to that of OFDM system;
it is untraceable for direct evaluating PAPR reduction abilities. iii) When KM < N, the PAPR of GFDM system should be
In [44], authors proposed a prototype filter optimization al- lower than that of OFDM system.
gorithm for GFDM system with the PAPR minimization toward From above results, we can see that when GFDM system em-
GFDM data blocks, jointly considering maintaining the OOB ploys optimal filters, the relationship of PAPR between GFDM
radiation under the maximum allowable level and minimizing and OFDM system is mainly determined by KM and N, or actu-
receiver MSE. This algorithm can actually be considered as an ally the total number of samples in each signal block of GFDM
approximate PAPR minimization, since it designs the prototype systems and that in each symbol of OFDM system.
filter only minimizing the maximum peak power of transmit When non-optimal pulse shaping filters are employed by
samples of given randomly generated blocks, with a constraint GFDM system, the PAPR should increase. Thus when KM ≥
that the total average power of transmit samples of generated N holds, the PAPR of GFDM system should be definitely higher
blocks is within a small range. Obviously, the performance of than that of OFDM system. But if KM < N holds, the relationship
this algorithm is closely related to the pre-generated data blocks, of PAPR should be uncertain, and be determined by particular
and this offline training model may degrade the optimality for filters employed by GFDM system. We can adopt (19) and (44)
long term and massive data block transmission. Joint optimiza- to calculate CCDF values and get comparison result directly.
tion may also degrade PAPR reduction performance to some ex- Secondly, consider comparisons between GFDM system and
tent. Compared with this algorithm, the proposed optimization FBMC/OQAM system. According to (19) and (25), we can find
3462 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 67, NO. 13, JULY 1, 2019

TABLE I
PARAMETERS OF SIMULATION SYSTEMS

Fig. 3. Theoretical and simulation results of PAPR distributions with different


K (M = 5, Q = 1, QPSK, RC filter with α= 0.1).

that CCDF results for both of these two systems are determined
by particular filters and the relationship is not straightforward.
We can also adopt (19) and (25) to calculate CCDF values and
obtain comparison result easily and directly.
However, for some application cases that subcarrier numbers
are relatively small, all theoretical CCDF expressions will be
not accurate enough, and simulations with particular system pa-
rameters have to be performed to obtain comparison results.

VI. SIMULATION RESULTS


A. Parameters of Simulation Systems
Parameters for simulation systems are listed in Table I. As
generally employed, subcarrier numbers (i.e., K for GFDM sys- Fig. 4. Theoretical and simulation results of PAPR distributions with different
tems and N for OFDM and FBMC/OQAM systems) are set to M (K = 256, Q = 1, QPSK, RC filter with α= 0.1).
be the power of 2 for conveniently adopting the Fast Fourier
Transform (FFT). Several types of common filters are consid-
ered for GFDM systems, and the typical PHYDYAS filter [28]
is adopted for FBMC/OQAM systems. All pulse shaping filters
have normalized average power in simulations.

B. Verifications for Theoretical Analysis Results in GFDM


Systems
In this subsection, we verify the accuracy of the theoretical
CCDF expressions derived in Section III. Additionally, we will
explore the influence of subcarrier number, subsymbol number,
pulse shaping filters and modulation order on the PAPR distri-
butions in GFDM systems.
Fig. 3 presents theoretical and simulation results of PAPR dis-
tribution in GFDM systems with different K and given M, pulse Fig. 5. Theoretical and simulation results of PAPR distributions with different
shaping filter and modulation scheme. From this figure, we can pulse shaping filters (M = 5, K = 256, Q = 1, QPSK).
see that increasing K will result in higher CCDF of PAPR. And
when K increases, the accuracy of the theoretical CCDF expres-
sions improves quickly, e.g., the PAPR gap between theoreti- Fig. 4 presents theoretical and simulation results of PAPR dis-
cal and simulation results for CCDF being 10−3 is only about tribution for M = 1, 5, 8, 16, and only theoretical results for M =
0.18 dB with K = 64, and reduces to be about 0.06 dB with K = 32, 45, 55 and 64, with given K, pulse shaping filter and modula-
128, and even much less when K increases further. These results tion scheme. All simulation results coincide with corresponding
suggest that the derived theoretical CCDF expressions could be theoretical results well. And from all theoretical results, we can
accurate enough and applicable for practical applications when see that increasing M will also result in higher CCDF of PAPR,
subcarrier number K is large enough (i.e., K ≥ 64). While for which can be attributed to the more superposition of complex-
relatively small K, we can find that simulated CCDF curves are valued signal components and more evaluated samples. But the
lower than those by theoretical CCDF expressions, thus theoret- CCDF increment rate tends to be slower as M increases.
ical results could be regarded as rough upper bounds for actual Fig. 5 and Fig. 6 provide theoretical and simulation results
CCDF values. of PAPR distributions, with different pulse shaping filters and
LIU et al.: THEORETICAL ANALYSIS OF THE PAPR AND OPTIMAL PULSE SHAPING FILTER DESIGN FOR GFDM SYSTEMS 3463

Fig. 6. Theoretical and simulation results of PAPR distributions with different Fig. 8. Theoretical and simulation results of PAPR distributions with different
modulation orders (K = 256, Q = 1, RC filter with α= 0.1, RRC filter with α= pulse shaping filters (M = 5, K = 256, Q = 4, QPSK).
0.9).

Fig. 9. Theoretical results of PAPR distributions for pulse shaping filters de-
Fig. 7. Theoretical and simulation results of PAPR distributions with different fined by Theorem 1 and Theorem 2. (M = 3, Q = 1).
oversampling factor Q (QPSK, RC filter with α = 0.1).

oversampling case. We can find that simulation results coin-


modulation schemes respectively. We can also find that all sim- cide with theoretical results well, which indicates that theoretical
ulation results coincide with corresponding theoretical results CCDF expression (20) is accurate with different types of filters.
well (PAPR gaps between theoretical and corresponding simu- Additionally, Dirichlet and Rectangular filters still have the low-
lation results are less than 0.06 dB in two figures when CCDF est CCDF values, which actually indicates that the optimality of
is about 10−3 ). Thus the accuracy of derived theoretical CCDF these two filters can still hold in oversampling scenario, and the
expressions can hold well for different pulse shaping filters and optimization criterion of (31) can be extended to oversampling
modulation schemes. From Fig. 5, we can find that different case, as described in Section IV.
types of pulse shaping filters lead to different PAPR distribu-
tions, which accords with theoretical analysis results. While in
C. Evaluation of the Optimization Criterion for the Design of
Fig. 6, we can see that with given system parameters and fil-
ters, CCDF results with different modulation schemes are almost Pulse Shaping Filters
identical. These results may confirm that modulation schemes Firstly, we compare the theoretical CCDF results of PAPR
have no influence on PAPR distributions, and also coincide with for pulse shaping filters defined by Theorem 1 (i.e., equa-
theoretical analysis results. tion (31)) and Theorem 2 (i.e., equation (33)) respectively (see
Fig. 7 presents simulation and theoretical results of PAPR dis- Appendix A). In Fig. 9 and Fig. 10, ‘PAPRthreshold ’ indicates
tributions with different oversampling factor Q and system pa- the PAPR threshold γ for calculating the coefficient vector de-
rameters. From this figure, we can find that CCDF values with fined by Theorem 2. Fig. 9 presents theoretical CCDF results
critical sampling are truly lower than those with oversampling. for K = 64, 512 and 4096, with M = 3.
For oversampling case, with different Q, M and K, simulation From Fig. 9, we can see that all CCDF results for pulse
results coincide with theoretical results well. These results con- shaping filters defined by Theorem 2 almost maintain constant
firm the accuracy of (20). With Q = 4, 8, 16 and given M and K, and are very close to 1 at different PAPR values, regardless
all CCDF curves almost overlap together. This result is desired, of γ. Meanwhile, although CCDF results for filters defined by
since the PAPR distribution of discrete samples will be close to Theorem 2 are less than those of filters defined by Theorem1 at
that of actually continuous signals when the oversample factor relatively small PAPR values, we can see that the superiority is
Q is large enough (Q ≥ 4). too small and can be ignored for practical applications.
Fig. 8 presents simulation and theoretical results of PAPR Fig. 10 confirms this result further and more clearly. Fig. 10
distributions with different types of pulse shaping filters for the presents the difference between theoretical CCDF results for
3464 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 67, NO. 13, JULY 1, 2019

Fig. 10. Difference between theoretical CCDF results of PAPR for filters de-
fined by Theorem 1 and Theorem 2. (M = 3, Q = 1). Fig. 11. Comparisons of theoretical CCDF results of PAPR for different pulse
shaping filters (M = 5, K = 256, Q = 1, QPSK).

filters defined by Theorem 1 and Theorem 2. The x-label indi-


cates the PAPR thresholds, as well as the actual PAPR values
for CCDF evaluations. We can find that when PAPR is relatively
small, filters defined by Theorem 2 are indeed superior to filters
defined by Theorem 1, and the superiority enlarges as PAPR in-
creases and K decreases. But the superiority is truly too small at
different PAPR values. It should be noted that when γ is large
enough, the optimal solution of coefficient vector and the corre-
sponding filter defined by Theorem 2 will not exist, so all curves
terminate at certain PAPR values in Fig. 10. Additionally, the
optimality of filter defined by Theorem 2 can only hold when
PAPR is large enough, and filter defined by Theorem 1 will be
much better, which can also be seen from Fig. 9.
And then, we compare the theoretical CCDF results of optimal Fig. 12. Theoretical results of PAPR distributions for different pulse shaping
pulse shaping filters defined by Theorem 1 (i.e., equation (31)) filters with different CRM SE .
to those of other filters.
Here we provide two filter examples and evaluate CCDF re-
sults of them to validate the optimization criteria proposed in This result verifies the validity of Corollary 2 in term of CCDF
this paper. The two filters are designed based on Corollary 2, minimization. RC and RRC filters (with α = 0.1, 0.5, 0.9) have
targeting lowest CCDF results while guaranteeing the receiver higher CCDF results. We can also find that larger α lead to higher
performance. These two filters are modified from Dirichlet filter, CCDF, which can be attributed to larger CRM SE . RRC filters
since it is a CMCM filter. According to Corollary 2, a CMCM fil- have the worst PAPR performance. The difference of PAPR at
ter after any phase rotating only in frequency domain can still be CCDF = 10−3 is about 1.5 dB between the RRC filter with α =
a CMCM filter. This criterion provide flexibility for optimization 0.9 and the optimal filters.
the phase-frequency characteristic of filters. These two filters are We also provide simulation results to explore the relation be-
designed to rotate the phases of frequency samples of Dirich- tween CRM SE defined in Section IV and theoretical CCDF re-
let filter with a definite and linear increasing phase-rotation se- sults with given K and M, as depicted in Fig. 12. From Fig. 12, we
quences and a random phase-rotation sequences, respectively. can find that with a given M, the same increment of CRM SE may
The frequency response samples of these two filters are defined nearly lead to the same increment of PAPR for certain CCDF
as follows (here we set Q = 1): with different K. And a step length of 0.1 for increasing CRM SE
Example-I: GExam−I [n] = GDirichlet [n] · exp(jnθ/KM ), may nearly lead to the PAPR increment of about 0.4 dB at CCDF
n ∈ S, = 10−3 (except for an approximate 0.25 dB with CRM SE in-
Example-II: GExam−II [n] = GDirichlet [n] · exp(jρπ/KM ), creasing from zero to 0.1). These results indicate that CRM SE
n∈S can be employed as a measurement for approximately and effi-
where GDirichlet [n] denotes the nth frequency response sample ciently evaluating PAPR distribution properties of pulse shaping
of Dirichlet filter, θ denotes a constant phase (e.g., θ = π/3), and filter candidates in practical applications.
ρ ∼ N(0, 1). Filter coefficients√in time domain can be obtained In Fig. 13, theoretical PAPR distributions results with 4th Xia
by gExam = WKM H
· GExam / KM , with WKM H
be the Her- filter and Dirichlet filter are presented, where KM is fixed and
mitian transpose of KM-point DFT matrix, gExam and GExam set to be 2048. We can find that Dirichlet filter can maintain
be the filter sample vector in time domain and frequency domain, the optimality of minimum CCDF with different K and M. The
respectively. CCDF results of 4th Xia filter vary with different K and M, which
Theoretical CCDF results of different pulse shaping filters can be attributed to the different CRM SE , as illustrated in Fig. 2.
are illustrated in Fig. 11. From Fig. 11, we can find that the Although fixed KM indicates same length of filter samples in
designed Example-I / II filters just have the same CCDF values each block, different filter shapes will result in different PAPR
to those of optimal filters (i.e., Rectangular and Dirichlet filter). distribution results.
LIU et al.: THEORETICAL ANALYSIS OF THE PAPR AND OPTIMAL PULSE SHAPING FILTER DESIGN FOR GFDM SYSTEMS 3465

Fig. 13. Theoretical and simulation results of PAPR distributions with fixed Fig. 15. Comparisons of theoretical results of PAPR distributions in OFDM,
KM (KM = 2048, Q = 1, QPSK). FBMC/OQAM, and GFDM system (Q = 1).

optimization is still valuable, since appropriate filters can fur-


ther reduce PAPR effectively. With NPC method, Dirichlet and
Rectangular filters have nearly identical and the lowest CCDF
values, and reduce PAPR about 0.29 dB at CCDF = 10−3 , against
RRC filter with α = 0.9. Although the further PAPR reduction
is much less, these results also indicate that optimized filters can
still help to reduce PAPR further even when the employed PAPR
reduction methods have excellent PAPR reduction performance.

D. PAPR Comparisons of Different Multicarrier Systems


In this subsection, we compare PAPR distributions of different
multicarrier systems to verify analysis results in Section V.
Fig. 14. Simulation results of PAPR distributions with LFDMA, IFDMA and Fig. 15 provides theoretical results of PAPR distributions of
NPC method (M = 5, K = 128, Q = 1, QPSK). OFDM, FBMC/OQAM, and GFDM systems. The subcarrier
number for OFDM and FBMC/OQAM system is fixed to be N
= 1024, while several values of KM and pulse shaping filters
To evaluate PAPR reduction performance of different fil- are employed by GFDM system for comparisons. From this fig-
ters combined with PAPR reduction schemes, we present simu- ure, it can be easily found that when GFDM system adopt the
lated CCDF results for different filters combined with the local- optimal filter (here is the Rectangular filter), smaller KM (i.e.,
ized frequency division multiple access (LFDMA) method [8], K = 256 and M = 2) leads to better CCDF results of PAPR
[25], interleaved frequency division multiple access (IFDMA) and larger KM (i.e., K = 256 and M = 7) leads to worse CCDF
method [8], [25] and the novel piecewise companding (NPC) results of PAPR, compared with those of OFDM system and
method [46], respectively. Simulation results are illustrated in FBMC/OQAM system. The CCDF results of FBMC/OQAM
Fig. 14. When LFDMA method is employed, Rectangular filter system are a little better, but very close to those of OFDM sys-
still has the lowest CCDF values, while PAPR reduction perfor- tem. When non-optimal filters are adopted by GFDM system,
mance of Dirichlet filter degrades about 1.5 dB at CCDF = 10−3 , the relationship will be uncertain. For example, when KM < N
compared with that of Rectangular filter, and is even inferior to holds (e.g., K = 128, M = 5; or K = 256, M = 3), GFDM sys-
that of RRC filter with α = 0.1, 0.5. These results may conflict tems have better CCDF results of PAPR compared with OFDM
with previous conclusions. The reason is that LFDMA is actually system if the RRC filter with α = 0.1 or RC filter with α = 0.1
a precoding method which changes the transmitted multicarrier are employed, but worse CCDF results of PAPR if the RRC filter
signals nearly to be single-carrier signals. The statistical prop- with α = 0.9 is employed. And when KM = N holds (e.g., K =
erty of transmitted signals is changed, thus previous theoretical 256, M = 4), GFDM system employing the non-optimal filter
analysis results as well as the optimization design criteria are not (i.e., RC filter with α = 0.9) has worse CCDF results, compared
applicable. Similar results can be found with IFDMA method. with OFDM system. This result is desired, since only optimal
Note that with IFDMA method and QPSK modulation, Rect- filters can make GFDM system have the same PAPR distribu-
angular filter leads to zero PAPR and this result is not drawn in tions to those of OFDM system when KM = N holds. All these
Fig. 14. Here we can see with IFDMA algorithm, RRC filter with results match the analysis results in Section V.
α = 0.5 has the lowest CCDF values, and Dirichlet filter has the
worst performance with a degradation about 2.8 dB at CCDF
= 10−3 , compared with the former. IFDMA method achieves E. Bit Error Rate (BER) Performance Comparisons
lower PAPR results than LFDMA method, since it is also a pre- Lastly, we examine the transmission performance of dif-
coding method and can equivalently change transmitted multi- ferent type of filters, as well as the proposed Example-I and
carrier signals into exact single-carrier signals. Simulation re- Example-II filters. Simulation results of BER performance in
sults with LFDMA and IFDMA methods also indicate that filter AWGN channel are illustrated in Fig. 16, and zero forcing
3466 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 67, NO. 13, JULY 1, 2019

APPENDIX A
Here we will prove that the optimization problem formulated
by (28)−(30) may have two different forms of locally optimal
solution vectors mathematically, and compare them to obtain the
ultimate solution vector which is more appropriate for practical
implementations.
By exploiting the monotonicity of the natural logarithm func-
tion, the optimization target (28) can equivalently be expressed
by
 K−1   
 1
Min − ln 1 − exp − . (A-1)
Ω
n=0
ωn
Fig. 16. Bit error rate of GFDM systems with different filters (M = 5, K = According to (A-1), (29) and (30), we can see the desired
256, Q = 1, QPSK, AWGN, ZF receiver, θ = π/3). solution vector Ω should be determined only by the subcarrier
number K and the target PAPR threshold γ.
Define the target function as
(ZF) detection [21] is employed. As desired, we can find that  
K−1 
1

Rectangular, Dirichlet, Example-I/II filters have nearly identical T (Ω) = − ln 1 − exp − , (A-2)
ωn
BER performance which is better than that of RC and RRC n=0
filters. Since Rectangular and Dirichlet are CMCM filters, and an auxiliary function to be
Example-I/II filters also achieve the optimal BER performance.  
This result verify the validity of Corollary 2, in term of achieving fA (x) = − ln 1 − e−1/x . (A-3)
minimization receiver MSE. The 1st Xia and 4th Xia filters with Δ
α = 0.1 are close to CMCM filter, thus have BER performance By the definition of exp(−1/x)|x=0 = 0 in Section (IV),
close to the optimal result. fA (x) is continuous and differentiable on the interval of
[0, +∞). The first-order and second-order derivative of fA (x)
can respectively be expressed by
VII. CONCLUSIONS
1 e−1/x
An exact closed-form expression of CCDF for the PAPR fA1 (x) = 2 · , (A-4)
x 1 − e−1/x
is firstly derived for critically sampled GFDM signals, and is  
modified to be suitable for oversampled GFDM signals. These 1 e−1/x 1 1
theoretical expressions reveal main factors influencing PAPR fA2 (x) = 3 · · · − 2 . (A-5)
x 1 − e−1/x x 1 − e−1/x
distribution of GFDM signals. These expressions tend to be
And then, the target function T (Ω) can also be expressed by
sufficiently accurate as K increases to be large enough (e.g., K
≥ 64). The derived theoretical CCDF expressions can be much   K−1

helpful for evaluating the PAPR characteristics of GFDM sig- T (Ω) = T (ω0 , ω1 , . . . , ωKM −1 )T = fA (ωn ). (A-6)
nals. With aid of theoretical analysis results, two optimization n=0
problems, i.e., MinPD and MinVIP, are formulated and solved. The target function T (Ω) is not a convex function entirely in
The preferred solution for MinPD problem can nearly achieve whole feasible region defined by (29) and (30), and it is difficult
globally minimum CCDF and is proved to be the globally to find the globally optimal solution vector directly. So we seek
optimal solution for MinVIP problem, and can minimize the the locally optimal solution vectors when ωn (n ∈ SP ) locates
fluctuation of instantaneous power. Based on this criterion, a in different intervals, which can be obtained through following
metric named CRM SE is proposed to evaluate PAPR reduction two theorems.
ability of filters efficiently, and increment of 0.1 with CRM SE
Theorem 1: If all ωn (ωn ∈ Ω) locate in the interval of [0, tC ]
may nearly lead to PAPR increment of 0.4 dB at CCDF = 10−3 .
The effect of PAPR reduction with common pulse shaping filters (here tC ≈ 0.6275 is the upper bound guaranteeing fA2 (x) larger
is much limited compared with special PAPR reduction algo- than or equal to zero), there is only one optimal solution vector
rithms. Filters with Large CRM SE may also lead to relatively Ω which has equal elements and can be expressed by
high PAPRs and degrade system performance. Theoretical ΩOptimal = (1/γ, 1/γ, . . . , 1/γ )T . (A-7)
CCDF results also provide efficient way for comparisons of   
K
PAPR distribution properties among multicarrier systems and
help to achieve more comprehensive and universal comparison Theorem 2: If one or more ωn (ωn ∈ Ω) locate in the interval
conclusions. of (tC , K/γ), the optimal solution vector Ω should have only
Based on analysis and discussions in this paper, we can see one ωn larger than tC and all of other elements should be equal
that there may be some valuable and suitable research direc- to one another. The optimal solution vector Ω should be in the
tions for future works, for example, joint optimization for the form of
design of filters, with targets of PAPR reduction, rate maximiza- Ω̂Optimal = (ω̂M ax , ω̄, . . . , ω̄ )T , (A-8)
tion, OOB radiation minimization, etc.; joint optimization de-   
sign of filters and PAPR reduction algorithms (e.g., clipping, K−1
compading, precoding); joint optimization design of filters and where ω̂M ax is the largest element and can be at any position in
multiple-input-multiple-output (MIMO) receivers, and so on. Ω̂Optimal , and ω̄ = (K/γ − ω̂M ax )/(K − 1).
LIU et al.: THEORETICAL ANALYSIS OF THE PAPR AND OPTIMAL PULSE SHAPING FILTER DESIGN FOR GFDM SYSTEMS 3467

A. Proof of Theorem 1 any ωn ∈ (0, tC ], we have ∂ 2 T /∂ωn2 ≥ 0 and F(Ω) is positive


Firstly, we prove that If all ωn (ωn ∈ Ω) locates in the interval semidefinite. And so, for any vector Ωa , Ωb ∈ Φ, we will have
of [0, tC ], any vector containing zero-value elements cannot be (Ωb − Ωa )T F (Ωa ) (Ωb − Ωa ) ≥ 0, (A-15)
the optimal solution vector.
which indicates that the target function T (Ω) is a convex func-
Assume that there is at least one zero-value element in Ω.
tion over the convex set Φ (refer to Theorem 22.5 and its exten-
Without loss of generality, we set ωv = 0 (v ∈ SP ). Obviously,
sion to nonopen sets in [40]).
Ω also has at least one element larger than zero, and we set
We now adopt the Lagrange Multiplier Method (LMM) to
ωμ > 0(μ ∈ SP , μ = v). Then vector Ω can be expressed by
obtain the critical point of T (Ω). By introducing a Lagrange
Ω = (. . . , ωμ , . . . , ωv , . . .)T . multiplier λ, the Lagrange function is defined as
Define a new vector Ω∗ = (. . . , ωμ − δ, . . . , ωv + δ, . . .)T ,  
K−1 
1

where δ is a positive real parameter and locates in the interval L (Ω, λ) = − ln 1 − exp −
ωn
of (0, ωμ /2). Obviously, Ω∗ still satisfies constraint condition n=0

of (28) and (29), and all elements still locate in the interval of K−1
 K
[0, tC ]. According to (A-6), we will have −λ ωn − . (A-16)
γ
T (Ω) − T (Ω∗ ) = fA (ωμ )−fA (ωμ − δ)−[fA (δ) − fA (0)] . n=0

(A-9) Let ∂L/∂ωn = 0 and ∂L/∂λ = 0, we can get


According to the Mean Value Theorem (MVT) [39], we have ∂L 1 e−1/ωn
= fA1 (ωn ) − λ = 2 · − λ = 0, (A-17)
fA (ωμ ) − fA (ωμ − δ) = fA1 (α) · δ, (A-10) ∂ωn ωn 1 − e−1/ωn

fA (δ) − fA (0) = fA1 (β) · δ, (A-11) ∂L 


K−1
K
= ωn − = 0. (A-18)
where α ∈ (ωμ − δ, ωμ ) and β ∈ (0, δ) are real parameters. ∂λ n=0
γ
Set fA2 (x) = 0 and solve this equation numerically, we can Equation (A-17) is a transcendental equation and difficult to
obtain the solution tC ≈ 0.6275. It can be easily verified that be solved directly. Due to the monotonicity of fA1 (x) on the
1/[x(1 − e−1/x )] is monotonically decreasing with x, thus for interval of (0, tC ], it is reasonable to infer that (A-17) will has
x ∈ (0, tC ], we have fA2 (x) ≥ 0 and fA1 (x) is monotonically in- unique solution (if it does exist) and all ωn should be equal to
creasing with x; for x ∈ (tC , K/γ), fA2 (x) < 0 and fA1 (x) is one another. Taking (A-18) into account further, we can obtain
monotonically decreasing with x. the solution result of
As 0 < β < α < tC , we will have fA1 (α) > fA1 (β). Accord-
ing to (A-9), (A-10) and (A-11), we can obtain T (Ω) > T (Ω∗ ). ωn = 1/γ, n ∈ SP ,
Due to the generality of ωv , we can conclude that for any Ω  
λ = γ 2 e−γ / 1 − e−γ
containing zero-value elements, we could find a corresponding
new vector Ω∗ which can make the value of the target function Thus the Lagrange function (A-16) has unique critical point
reduced further. Thus the optimal solution vector Ω should have of
none zero element. ΩC = (1/γ, 1/γ, . . . , 1/γ )T . (A-19)
Secondly, we will seek the optimal solution vector Ω with   
K
all elements locate in the interval of (0, tC ]. Define the new
Note that the necessary condition for ΩC ∈ Φ should be
constraint condition as
1/γ ≤ tC , or γ ≥ 1/tC ≈ 2.02 dB equivalently. This condition
0 < ωn ≤ tC , ∀n ∈ SP . (A-12) can easily be satisfied in practical applications, since we need
And define the new feasible region restrained by (A-12) and not to consider the CCDF of too small PAPR thresholds. When
(28) to be Φ (Φ ⊂ RK + ,, R+ denotes the K–dimensional positive
K this condition holds, ΩC will be the optimal solution vector (i.e.,
real vector space). Obviously, both of feasible regions defined ΩOptimal = ΩC ) for the optimization problem formulated by
by (A-12) and (28) are convex sets respectively, thus Φ is also (27), (35) and (A-12) (refer to Theorem 22.8 in [40]). And the
a convex set (refer to Theorem 4.1 in [40]). Note that Φ is a corresponding optimal coefficient vector for pulse shaping filters
nonopen set, while RK + is an open set, and the target function
will be expressed by (30). Thus Theorem 1 is proved.
T (Ω) is twice differentiable over RK
+.
We can calculate elements of Hessian matrix F(Ω) for the B. Proof of Theorem 2
target function T (Ω) as follows
Actually, for the cases of one or more ωn (ωn ∈ Ω) locate
∂2T within (tC , K/γ), we can firstly employ the MVT to prove that
= 0, ∀i, j ∈ SP , and i = j, (A-13)
∂ωi ∂ωj if we reduce all ωn larger than tC (except the maximum ele-
  ment ωM ax ) to be tC , and increase ωM ax successively so as to
∂2T 1 e−1/ωn 1 1
= · · · − 2 , ∀n ∈ SP maintain the constraint condition of (29), the value of the tar-
∂ωn2 ωn3 1 − e−1/ωn ωn 1 − e−1/ωn get function can be reduced successively, and ultimately there
(A-14) is only one element larger than tC (i.e., the maximum element).
Thus F(Ω) is a diagonal matrix, and main diagonal elements Let U = {ωn |ωn > tC , n ∈ SP } be the set of all elements
are just its eigenvalues. According to analysis results above, for larger than tC in vector Ω, and assume that the total number
3468 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 67, NO. 13, JULY 1, 2019

of elements is V (V ∈ Z+ , 1 ≤ V ≤ K). If V = 1, the first step where ω̄ = (K/γ − ω̂M ax )/(K − 1) is the average value of all
establishes naturally. So we begin with V ≥ 2. K−1 elements locating on (0, tC ]. According to the analysis
For the largest element ωM ax and any other element ωv in Proof of Theorem 1, we can find that the upper bound for
(v ∈ SP ) in U, the value of target function T (Ω) can be cal- guaranteeing the target function to be convex is a constant (i.e.,
culated by tC ≈ 0.6275) and has no relation to the summation of all ele-
 ments in corresponding solution vectors. Thus Theorem 1 is still
T (Ω) = fA (ωn ) + fA (ωM ax ) + fA (ωv ) .
applicable here, and (A-28) will hold.
n∈SP ,n =v
According to (A-27) and (A-28), we can see if there is at
(A-20)
least one ωn (ωn ∈ Ω) locating in the interval of (tC , K/γ), the
Now we reduce ωv to be tC and increase ωM ax to be optimal solution vector Ω should have only one ωn larger than
ω̂M ax = ωM ax + ωv − tC in Ω, we can obtain a new vector tC and all other elements in Ω should be equal to each other.
of Ω̂ = (. . . , tC , . . . , ω̂M ax , . . .)T . Obviously, vector Ω̂ sill sat- The optimal solution vector Ω should be in the form of (A-8),
isfies constraint condition of (29) and (30). And the value of and the corresponding optimal coefficient vector should be in
target function T (Ω) should update to be the form of (33). Note that there is no limitation for the position
  
T Ω̂ = fA (ωn ) + fA (ω̂M ax ) + fA (tC ) . of the maximum element in optimal vectors. Thus Theorem 2 is
n∈SP ,n =v
proved.
(A-21)
C. Remarks for Theorem 1 and Theorem 2
The different between T (Ω) and T (Ω̂) is
  To obtain the particular solution vector defined by Theorem 2
T (Ω) − T Ω̂ = fA (ωM ax ) − fA (ω̂M ax ) with given K and γ, we should seek the optimal ω̂M ax . According
to (A-6) and (A-8), the target function T (Ω) can be expressed
+ fA (ωv ) − fA (tC ) (A-22) by
 
As fA (x) is continuous and differentiable on the interval of T Ω̂Optimal = fA (ω̂M ax ) + (K − 1) · fA (ω̄)
[0, +∞), we can employ the MVT to obtain
 
fA (ωv ) − fA (tC ) = fA1 (α ) · (ωv − tC ) , (A-23) K/γ − ω̂M ax
= fA (ω̂M ax ) + (K − 1) · fA . (A-29)
K −1
fA (ωM ax ) − fA (ω̂M ax ) = −fA1 (β  ) · (ωv − tC ) , (A-24)
Equation (A-29) has only one variable of ω̂M ax . By set-
with parameter α ∈ (tC , ωv ) and β  ∈ (ωM ax , ωM ax + ωv ting dT (Ω̂Optimal )/dω̂M ax = 0 and solve this equation numer-
− tC ). ically, we can get the optimal ω̂M ax . This equation can be ex-
Then (A-22) should be pressed by
  ! "  
T (Ω) − T Ω̂ = fA1 (α ) − fA1 (β  ) · (ωv − tC ) . (A-25) dT Ω̂Optimal
= fA1 (ω̂M ax ) − fA1 (ω̄) = 0. (A-30)
As ωv ≤ ωM ax , there will be tC < α < β  . Considering that dω̂M ax
fA1 (x)
is monotonically decreasing with x in the interval of Equation (A-30) contains nonlinear function and may have
x ∈ (tC , K/γ), thus we have fA1 (α ) > fA1 (β  ). Additionally, non-unique solutions. We can get all critical points by solv-
we have ωv > tC . According to (25), there will be ing (A-30) numerically and determine the one which minimizes
  T (Ω̂Optimal ) to be ω̂M ax . We have calculated and obtained
T (Ω) > T Ω̂ . (A-26)
ω̂M ax for some K and γ. For example, when K = 64, ω̂M ax
This result indicate that reduce ωv to be tC and add the decre- equals to 42.87, 32.00, 23.18, 15.88, 9.42, for γ (in dB) = 1, 2,
ment to ωM ax , the value of the target function can be reduced, …, 5; when K = 512, ω̂M ax equals to 358.52, 273.30, 205.06,
while constraint conditions can be maintained. We can repeat 150.17, 105.61, 68.64, for γ (in dB) = 1, 2, …, 6; when K
this operation in set U successively, until there remains only one = 4096, ω̂M ax equals to 2941.88, 2265.20, 1725.42, 1293.96,
element (i.e., the updated largest element of ω̂M ax ) larger than 947.92, 668.63, 440.12, 245.18, for γ (in dB) = 1, 2, …, 8. We
tC in set U, and the value of the target function will be reduced also find that with a given K, (A-30) will has none solution when
continuously. And so, if one or more ωn (ωn ∈ Ω) locate in the γ is large enough.
interval of (tC , K/γ), the optimal solution vector Ω should have And then, we will compare the actual values of target function
only one element larger than tC . with different solution results respectively defined by Theorem 1
And then, let us consider the ultimate vector Ω̂. The value of and Theorem 2.
the target function can accordingly be expressed by According to the derivation of ω̄ in Proof of Theorem 2, we
   have ω̄ ≤ tC . For generally practical implementations, we only
T Ω̂ = fA (ωn ) + fA (ω̂M ax ) , (A-27) consider γ ≥ 1/tC , thus we have (K/γ − tC )/(K − 1) < tC .
ωn ∈Ω̂,ωn =ω̂M ax As ω̂M ax > tC , we have ω̄ < (K/γ − tC )/(K − 1). Due to the
Obviously, all elements except for ω̂M ax in Ω̂ locate in the monotonically increasing property of fA1 (x) within (0, tC ], we
interval of (0, tC ]. According to Theorem 1, we will have obtain fA1 (ω̄) < fA1 ((K/γ − tC )/(K − 1)). In another hand,
 due to K/γ > ω̂M ax > tC and the monotonically decreasing
fA (ωn ) ≥ (K − 1) fA (ω̄) , (A-28) property of fA1 (x) within (tC , K/γ), we obtain fA1 (ω̂M ax ) >
ωn ∈Ω̂,ωn =ω̂M ax fA1 (K/γ).
LIU et al.: THEORETICAL ANALYSIS OF THE PAPR AND OPTIMAL PULSE SHAPING FILTER DESIGN FOR GFDM SYSTEMS 3469

TABLE II to (36) and (38), the Lagrange function can be defined as


CCDF VALUES WITH SOLUTION RESULTS OF THEOREM 1 K−1


L (Ψ) = Pvar (Ψ) − λ ψn − 1
n=0
 K−1

K−1 

= σd4 2K ψn2 −1 −λ ψn − 1 ,
n=0 n=0
Let ∂L(Ψ)/∂ψn = 0(n ∈ SP ), we get 4Kσd4 ψn = λ∗ ,
which indicates that all ψn should be equal. According to (38),
we can obtain ψn = 1/K(n ∈ SP ), and the critical point of
According to analysis above, we can find that if above Lagrange function is just equal to ΨM inP D defined by
(31). Obviously, the target function Pvar (Ψ) is convex over
fA1 (K/γ) > fA1 ((K/γ − tC ) / (K − 1)) , (A-31) whole feasible region defined by (39) and (40), and the feasible
fA1 (ω̂M ax ) fA1 (ω̄),
will be definitely larger than and (A-30) will region is a convex set, thus ΨM inP D is the globally optimal for
have none solution. Actually, when (A-31) holds, from (A-30) the MinVIP problem. Corollary 1 is proved.
we can find that T (Ω̂Optimal ) will be monotonically increas-
ing with ω̂M ax , and ω̂M ax = tC will result in the minimum APPENDIX C
T (Ω̂Optimal ). If ω̂M ax = tC holds, all ωn follow into the scope PROOF OF COROLLARY 2
of Theorem 1, and the solution result of Theorem 1 will be
Firstly, we give the concept of characteristic matrix as defined
better. Thus (A-31) can be viewed as a sufficient condition for
in [21]. Consider a KM × KM GFDM transmitter matrix A in
guaranteeing the solution result of Theorem 1 superior to that of
(3) with a prototype filter gTX . The characteristic matrix G of
Theorem 2. It should also be noted that (A-31) is not a necessary
the GFDM transmitter matrix A is defined as
condition, since it is a relatively broader restriction. √
By changing the greater-than sign to be equal sign in (A-31) G = N · Reshape (gTX , K, M ) · WM , (C-1)
and solving this equation numerically, we can obtain the critical where Reshape(gTX , K, M ) is a K × M matrix gener-
values of γ with certain K. We have calculated these values ated from gTX and the (k, m)-element is gTX (k + mK)
for K = 64, 128, 256, 512, 1024, 2048 and 4096, which are (k ∈ [0, K − 1], m ∈ [0, M − 1]).
about 7.63, 8.23, 8.75, 9.20, 9.60, 9.96 and 10.28, respectively Here we also denote that
(all in dB). And it can be easily verified that for any K ≥ 2, ĝTX = Reshape (gTX , K, M )
these values always exist. More importantly, according to the
monotonicity of fA1 (x) in different intervals, we can find that = [ĝTX (0) ; ĝTX (1) ; . . . , ĝTX (K − 1)] , (C-2)
when K is fixed and γ increases to be larger than the critical where the 1 × M vector ĝTX (k) (k ∈ [0, K − 1]) is just the kth
value, (A-31) will definitely hold (note that increasing γ will row of ĝTX (k).
reduce both of K/γ and (K/γ − tC )/(K − 1), and the former According to the definition of characteristic matrix in (C-1),
is always larger than the latter). So we can conclude that for we will have
any given K in practical implementations, the solution result of √
ĝTX = G · WM H
/ N. (C-3)
Theorem 1 will be definitely superior to that of Theorem 2 when
γ increases to be large enough. And for any row vector ĝTX (k), we will have

Additionally, we compare the actual CCDF performance of ĝTX (k) = G(k) · WM H
/ N , k ∈ [0, K − 1], (C-4)
pulse shaping filters with coefficient vector results by Theorem 1
where the 1 × M vector G(k) (k ∈ [0, K − 1]) is the kth row
and Theorem 2. We have calculated the CCDF by employing (19)
of G.
and (31) with different K, M and critical values of γ to evaluate
According to (C-4) and the Parseval’s Theorem [41], we can
the result of Theorem 1, as listed in Table II. From Table II, we
obtain
can see that all CCDF values are relatively high (larger than 0.2)
and some of them are nearly to be 1. It should be noted that with ĝTX (k) 2F = G(k) 2F /N, k ∈ [0, K − 1]. (C-5)
parameters of {K, M, γ} in Table II, solution results of The- where || • ||2F
indicates the vector energy.
orem 1 are superior to those of Theorem 2. Thus we can infer According to the definition of CMCM filters, the correspond-
that even if results of Theorem 2 have better CCDF performance ing characteristic matrix contains constant magnitude entries,
than those of Theorem 1 when γ is small and less than critical thus we will have
values listed in Table II, the superiority can be ignored for prac-
ĝTX (0) 2F = ĝTX (1) 2F = · · · = ĝTX (K − 1) 2F , (C-6)
tical implementations, since the corresponding values of CCDF
are too high. which indicates that all row vectors in ĝTX have equal energy.
According to (16) and (C-2), we just have

M −1
APPENDIX B
PROOF OF COROLLARY 1 ψn = |gTX [(n − mK) mod M K]|2
m=0
The MinVIP problem defined by (38)–(40) can also be solved
by the LMM. We introduce a Lagrange multiplier λ∗ . According = ĝTX (n) 2F , n ∈ SP . (C-7)
3470 IEEE TRANSACTIONS ON SIGNAL PROCESSING, VOL. 67, NO. 13, JULY 1, 2019

According to (C-6) and (C-7), we can find that all ψn [20] N. Michailow and G. Fettweis, “Low peak-to-average power ratio for next
(n ∈ [0, K − 1]) have equal value. Without loss of generality, generation cellular systems with generalized frequency division multi-
plexing,” in Proc. Int. Symp. Intell. Signal Process. Commun. Syst., 2013,
we still assume that the prototype filter has normalized total pp. 651–655.
power. Then, according to the constraint condition (21), we will [21] P. C. Chen, B. Su, and Y. Huang, “Matrix characterization for GFDM:
have ψn = 1/K (n ∈ [0, K − 1]), which is just the result of Low complexity MMSE receivers and optimal filters,” IEEE Trans. Signal
Process., vol. 65, no. 18, pp. 4940–4955, Sep. 2017.
(31). Consequently, all CMCM filters satisfy the optimization [22] R. Gerzaguet et al., “The 5G candidate waveform race: A comparison of
criterion of (30) and can achieve the minimum PAPR distribu- complexity and performance,” EURASIP J. Wireless Commun. Network-
tion in GFDM system. Thus Corollary 2 is proved. ing, vol. 2017, 2017, Art. no. 13.
[23] L. Sendrei, S. Marchevský, N. Michailow, and G. Fettweis, “Iterative re-
ceiver for clipped GFDM signals,” in Proc. Int. Conf. Radio Elektronika,
2014, pp. 1–4.
REFERENCES [24] Z. Sharifian, M. J. Omidi, A. Farhang, and H. Saeedi-Sourck, “Polynomial-
based compressing and iterative expanding for PAPR reduction in GFDM,”
[1] J. G. Andrews et al., “What will 5G be?,” IEEE J. Sel. Areas Commun., in Proc. 23rd Iranian Conf. Elect. Eng., Tehran, 2015, pp. 518–523.
vol. 32, no. 6, pp. 1065–1082, Jun. 2014. [25] S. S. Das and S. Tiwari, “Discrete Fourier transform spreading-based gen-
[2] N. Michailow et al., “Generalized frequency division multiplexing for eralized frequency division multiplexing,” Electron. Lett., vol. 51, no. 10,
5th generation cellular networks,” IEEE Trans. Commun., vol. 62, no. 9, pp. 789–791, May 2015.
pp. 3045–3061, Sep. 2014. [26] Z. Sharifian, M. J. Omidi, H. Saeedi-Sourck, and A. Farhang, “Linear
[3] G. Fettweis, M. Krondorf, and S. Bittner, “GFDM—generalized frequency precoding for PAPR reduction of GFDMA,” IEEE Wireless Commun. Lett.,
division multiplexing,” in Proc. 69th IEEE Veh. Technol. Conf. Spring, vol. 5, no. 5, pp. 520–523, Oct. 2016.
Barcelona, Spain, 2009, pp. 1–4. [27] A. Farhang, N. Marchetti, and L. E. Doyle, “Low-complexity modem
[4] G. Wunder et al., “5GNOW: Non-orthogonal asynchronous waveforms for design for GFDM,” IEEE Trans. Signal Process., vol. 64, no. 6, pp. 1507–
future mobile applications,” IEEE Commun. Mag., vol. 52, no. 2, pp. 97– 1518, Mar. 2016.
105, Feb. 2014. [28] B. Farhang-Boroujeny, “OFDM versus filter bank multicarrier,” IEEE Sig-
[5] B. Lim and Y. C. Ko, “SIR analysis of OFDM and GFDM Waveforms with nal Process. Mag., vol. 28, no. 3, pp. 92–112, May 2011.
timing offset, CFO, and phase noise,” IEEE Trans. Wireless Commun., [29] M. Matthé et al., “Generalized frequency division multiplexing: A flexible
vol. 16, no. 10, pp. 6979–6990, Oct. 2017. multi-carrier waveform for 5G,” in 5G Mobile Communications. Cham,
[6] D. Zhang, M. Matthé, L. L. Mendes, and G. Fettweis, “A study on the link Switzerland: Springer Int. Publ., 2017, pp. 223–259.
level performance of advanced multicarrier waveforms under MIMO wire- [30] S. B. Slimane, “Peak to average power ratio reduction of OFDM signals
less communication channels,” IEEE Trans. Wireless Commun., vol. 16, using pulse shape,” in Proc. Global Commun., San Francisco, CA, USA,
no. 4, pp. 2350–2365, Apr. 2017. 2000, pp. 1412–1416.
[7] P. Wei, X. G. Xia, Y. Xiao, and S. Li, “Fast DGT-based receivers for [31] S. Han, Y. Sung, and Y. H. Lee, “Filter design for generalized frequency-
GFDM in broadband channels,” IEEE Trans. Commun., vol. 64, no. 10, division multiplexing,” IEEE Trans. Signal Process., vol. 65, no. 7,
pp. 4331–4345, Oct. 2016. pp. 1644–1659, Apr. 2017.
[8] S. Tiwari, S. S. Das, and K. K. Bandyopadhyay, “Precoded generalized [32] X.-G. Xia, “A family of pulse-shaping filters with ISI-free matched and
frequency division multiplexing system to combat inter-carrier interfer- unmatched filter properties,” IEEE Trans. Commun., vol. 45, no. 10,
ence: Performance analysis,” IET Commun., vol. 9, no. 15, pp. 1829–1841, pp. 1157–1158, Oct. 1997.
2015. [33] M. Matthé, L. L. Mendes, and G. Fettweis, “Generalized frequency di-
[9] R. Van Nee and A. De Wild, “Reducing the peak-to-average power ratio of vision multiplexing in a Gabor transform setting,” IEEE Commun. Lett.,
OFDM,” in Proc. IEEE Veh. Technol. Conf., Ottawa, ON, Canada, 1998, vol. 18, no. 8, pp. 1379–1382, Aug. 2014.
pp. 2072–2076. [34] R. M. Corless, G. H. Gonnet, D. E. G. Hare, D. J. Jeffrey, and D. H. Knuth,
[10] H. Ochiai and H. Imai, “On the distribution of the peak-to-average power “On the Lambert w function,” Adv. Comput. Math., vol. 5, pp. 329–359,
ratio in OFDM signals,” IEEE Trans. Commun., vol. 49, no. 2, pp. 282– 1996.
289, Feb. 2001. [35] S. Mazahir and S. A. Sheikh, “On companding schemes for PAPR re-
[11] S. H. Han and J. H. Lee, “An overview of peak-to-average power ratio duction in OFDM systems employing higher order QAM,” IEEE Trans.
reduction techniques for multicarrier transmission,” IEEE Wireless Com- Broadcast., vol. 62, no. 3, pp. 716–726, Sep. 2016.
mun., vol. 12, no. 2, pp. 56–65, Apr. 2005. [36] O. Johnson, Information Theory and the Central Limit Theorem. London,
[12] A. Skrzypczak, P. Siohan, and J. P. Javaudin, “Analysis of the peak- U.K.: Imperial College Press, 2004.
to-average power ratio for OFDM/OQAM,” in Proc. IEEE 7th Work- [37] W. Rudin, Principles of Mathematical Analysis, 3rd ed. New York, NY,
shop Signal Process. Advances Wireless Commun., Cannes, France, 2006, USA: McGraw-Hill, 1976.
pp. 1–5. [38] E. K. P. Chong and S. H. Zak, An Introduction to Optimization, 4th ed.
[13] M. Chafii, J. Palicot, and R. Gribonval, “Closed-form approximations of New York, NY, USA: Wiley, 2013.
the PAPR distribution for multi-carrier modulation systems,” in Proc. 22nd [39] John G. Proakis, Digital Signal Processing. 4th ed. Englewood Cliffs, NJ,
Eur. Signal Process. Conf., Lisbon, Portugal, 2014, pp. 1920–1924. USA: Prentice-Hall, 2006.
[14] J. Hou, J. Ge, D. Zhai, and J. Li, “Peak-to-average power ratio reduc- [40] A. Behravan and T. Eriksson, “Some statistical properties of multicarrier
tion of OFDM signals with nonlinear companding scheme,” IEEE Trans. signals and related measures,” in Proc. IEEE Veh. Technol. Conf.-Spring,
Broadcast., vol. 56, no. 2, pp. 258–262, Jun. 2010. 2006, pp. 1854–1858.
[15] S. S. Jeng and J. M. Chen, “Efficient PAPR reduction in OFDM sys- [41] D. Falconer, “Linear precoding of OFDMA signals to minimize their
tems based on a companding technique with trapezium distribution,” IEEE instantaneous power variance,” IEEE Trans. Commun., vol. 59, no. 4,
Trans. Broadcast., vol. 57, no. 2, pp. 291–298, Jun. 2011. pp. 1154–1162, Apr. 2011.
[16] Y. Wang, L. H. Wang, J. H. Ge, and B. Ai, “An efficient nonlinear com- [42] Y. Huang and B. Su, “Circularly pulse-shaped precoding for OFDM: A new
panding transform for reducing PAPR of OFDM signal,” IEEE Trans. waveform and its optimization design for 5G New Radio,” IEEE Access,
Broadcast., vol. 58, no. 4, pp. 677–684, Dec. 2012. vol. 6, pp. 44129–44146, 2018.
[17] B. Elmaroud, A. Faqihi, M. Abbad, and D. Aboutajdine, “PAPR re- [43] S. M. Kay, Fundamentals of Statistical Signal Processing. Volume II: De-
duction of FBMC signals by combining exponential companding and tection Theory. Englewood Cliffs, NJ, USA: Prentice-Hall, 1998.
Hadamard transforms,” in Proc. Int. Symp. Netw. Comput. Commun., 2014, [44] C.-L. Tai, B. Su, and P.-C. Chen, “Optimal filter design for GFDM that
pp. 1–4. minimizes PAPR under performance constraints,” in Proc. IEEE Wireless
[18] Z. You, I. T. Lu, R. Yang, and J. Li, “Flexible companding design for Commun. Netw. Conf., Barcelona, Spain, Apr. 2018, pp. 1–6.
PAPR reduction in OFDM and FBMC systems,” in Proc. IEEE Int. Conf. [45] A. Nimr, M. Matthé, D. Zhang, and G. Fettweis, “Optimal radix-2 FFT
Comput. Netw. Commun., San Diego, CA, USA, 2013, pp. 408–412. compatible filters for GFDM,” IEEE Commun. Lett., vol. 21, no. 7,
[19] J. Wu, X. Ma, X. Qi, Z. Babar, and W. Zheng, “Influence of pulse shap- pp. 1497–1500, Jul. 2017.
ing filters on PAPR performance of underwater 5G communication sys- [46] Y. Li and W. Deng, “A novel piecewise nonlinear companding transform
tem technique: GFDM,” Wireless Commun. Mobile Comput., vol. 2017, for PAPR reduction in GFDM,” in Proc. 10th Int. Conf. Wireless Commun.
Jan. 2017, Art. no. 4361589. Signal Process., Hangzhou, China, 2018, pp. 1–5.

You might also like