6202 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 38, NO.

22, NOVEMBER 15, 2020

Comparison of Different Precoding Techniques for

Unbalanced Impairments Compensation in
Short-Reach DMT Transmission Systems
Ming Chen , Member, IEEE, Ling Wang , Dongsheng Xi, Long Zhang , Hui Zhou , and Qinghui Chen

Abstract—Channel independent precoding technique has been to enhance SNR balance. Simulated and experimental results
widely used in optical discrete multi-tone (DMT) transmission exhibit that similar BER performance can be achieved by using
systems to compensate unbalanced impairments induced by these precoding techniques together with TSPA in noise-limited
bandwidth limitations and imperfect frequency responses of scenarios. Considering the implementation complexity, WHT
electrical/optical devices and various interferences. However, the precoding may be a good option to compensate unbalanced
comparison of different precoding techniques in terms of peak- impairments in the short-reach DMT transmission system.
to-average power ratio (PAPR) reduction, nonlinear distortion
tolerance, implementation complexity, and bit error rate (BER) Index Terms—Discrete multi-tone, nonlinear distortions,
improvements has not been fully studied. In this article, we com- precoding, unbalanced impairments compensation.
paratively investigate seven most commonly used precoding tech-
niques, i.e., discrete Fourier transform (DFT), orthogonal circulant I. INTRODUCTION
matrix transform (OCT), constant amplitude zero autocorrelation
PTICAL discrete multi-tone (DMT) is a special direct-
sequence-based matrix transform (CAZACT), Zadoff-Chu matrix
transform (ZCT), discrete cosine transform (DCT), discrete Hart-
ley transform (DHT), and Walsh-Hadamard transform (WHT),
O detection optical orthogonal frequency-division multi-
plexing (DDO-OFDM). It has been widely considered as
through both numerical simulations and offline experiments. Sim- a promising physical-layer technology for future short-reach
ulations show that the ZCT can achieve the best PAPR reduction,
and the OCT cannot reduce the PAPR. Besides, DFT, CAZACT,
broadband optical communications due to its inherent advan-
ZCT, DCT, and DHT precoded DMT signals have superior error tages, such as high spectral efficiency (SE) and resistance to
vector magnitude performance after passing through nonlinear fiber dispersions. Among DDO-OFDM systems, optical DMT
models. And the corresponding precoded QPSK-DMT signals does not require up/down-conversion in the digital or analog
have better BER performance than both OCT/WHT precoded and domain and has a low implementation complexity. Therefore,
conventional ones in the distortion-limited scenarios. However, it is more suitable for cost-sensitive short-reach applications
the precoded 16/64QAM-DMT signals, excluding OCT precoded
one, are more sensitive to nonlinear distortions and provide minor
such as optical access networks [1], [2], optical interconnec-
BER improvement or even may degrade the BER performance. tions in data centers [3], [4], and visible light communications
Complexity analysis exhibits the WHT precoding does not require (VLC) [5], [6]. In the literature, numerous field programmable
multiplications and therefore has the lowest implementation gate array (FPGA) based real-time optical DMT transmission
complexity. In the inter-symbol interference-limited case, OCT demonstrations have also been reported in [7]–[11].
precoding can still achieve a good signal-to-noise ratio (SNR) In an optical DMT system, there several common impair-
balance and provide the best BER performance. A simple timing
synchronization point adjustment (TSPA) method is employed ments, such as the imperfect frequency response of opti-
cal/electrical devices, chromatic dispersion (CD) induced power
fading [12], signal-to-signal beating interference (SSBI) [13]
Manuscript received May 10, 2020; revised June 29, 2020; accepted July
14, 2020. Date of publication July 17, 2020; date of current version November
and digital-to-analog converter (DAC) clock tone leakage [14],
16, 2020. This work was supported in part by the National Natural Science can lead to the unbalanced signal-to-noise ratio (SNR) dis-
Foundation of China under Grant 61805079 and Grant 61701180, in part by the tribution over data-carrying subcarriers. We called this phe-
Hunan Provincial Natural Science Foundation of China under Grant 2020JJ4433, nomenon as unbalanced impairments. As a result, the overall
and in part by the Scientific Research Fund of the Hunan Provincial Education
Department under Grant 18B026 and Grant 18C0520. (Corresponding author: bit error rate (BER) performance may be seriously degraded
Ming Chen.) due to unbalanced impairments. One classical and effective
Ming Chen, Ling Wang, Dongsheng Xi, and Long Zhang are with the method is to use the adaptive modulation technique, includ-
School of Physics and Electronics, Hunan Normal University, Changsha 410081,
China (e-mail: [email protected]; [email protected]; ing adaptive power and/or bit loading. And this technique has
[email protected]; [email protected]). been widely studied in optical DMT systems to maximize the
Hui Zhou is with the College of Information Science and Engi- transmission rate [15], [16]. Meanwhile, a simplified method
neering, Hunan Normal University, Changsha 410081, China (e-mail:
[email protected]).
named pre-emphasis technique is also employed to enhance the
Qinghui Chen is with the College of Computer and Communica- transmission performance [17] in the bandwidth-limited system.
tion, Hunan University of Technology, Zhuzhou 412007, China (e-mail: However, the feedback of the channel state information over all
[email protected]). data-carrying subcarriers is required via the reverse link. It is
Color versions of one or more of the figures in this article are available online
at https://ieeexplore.ieee.org. timing consuming. In [13], the authors proposed a static adaptive
Digital Object Identifier 10.1109/JLT.2020.3010002 modulation scheme combined with the forward error correction
0733-8724 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.
CHEN et al.: COMPARISON OF DIFFERENT PRECODING TECHNIQUES FOR UNBALANCED IMPAIRMENTS COMPENSATION 6203

technique to compensate unbalanced impairments induced by system. It is interesting to explore which precoding technique
the SSBI and power fading. This scheme may not be suitable for is the best choice to compensate unbalanced impairments un-
the system with the time-varying channel response. der different scenarios, such as nonlinear distortion, ISI, and
Recently, channel/signal independent precoding technique noise limitations. In this paper, we comparatively investigate
has been widely employed in the optical DMT for the compensa- seven most commonly used precoding techniques, i.e., DFT,
tion of unbalanced impairments [18]–[33]. The idea of precoding OCT, CAZACT, ZCT, DCT, DHT and WHT, for unbalanced
is realized by multiplying the quadrature amplitude modulation impairments compensation in DMT transmission systems by
(QAM) mapped symbols by a precoding matrix in the trans- numerical simulations. And a simple timing synchronization
mitter. In the receiver side, the QAM symbols are recovered by point adjustment (TSPA) method is employed to combat ISI
multiplying the channel equalized data by its inverse matrix. to enhance the performance of the precoded DMT without
Theoretically, if the amplitudes of the elements in the precoding cyclic suffix (CS). Experimental verifications in a DML-based
matrix are equal, then the identical SNR distribution over all of short-reach DMT transmission system are also conducted. Both
data-carrying after decoding can be achieved [27]. Meanwhile, simulated and experimental results exhibit WHT precoding can
the precoding may improve the autocorrelation performance achieve similar BER performance in the noise-limited scenario
of input symbols for the inverse fast Fourier transform (IFFT) as other precoding techniques, but with the lowest implementa-
function and thus reduce peak-to-average power ratio (PAPR) tion complexity.
[28]. Owing to the PAPR reduction, the nonlinear distortions, The rest of the paper is structured as follows. In Section II,
introduced by optical or electrical devices such as electrical we briefly introduce the working principle of the precoding
amplifiers, directly modulated lasers (DML) and light-emitting technique and the construction of different precoding matrices.
diodes (LED), may be reduced. The analysis of implementation complexity is performed in
In [18], discrete Fourier transform (DFT) precoding tech- Section III. The simulation model used in this work is described
nique, also known as DFT-spread, was applied to improve the in Section IV. The PAPR performance of the conventional and
BER performance of the laser-based short-reach DMT transmis- the precoded DMT signals are discussed in Section V. The
sion system with a step-index plastic optical fiber link. Besides, nonlinear distortion tolerance of the precoded DMT signals is in-
DFT precoding is used to improve the BER performance of vestigated with two nonlinear distortion models in Section VI. In
the optical DMT system with clock tone leakage of DAC and Section VII, the BER performance of the precoded DMT signals
analog-to-digital converters (ADC) [22]. Moreover, a constant passing through a bandwidth-limited additive white Gaussian
amplitude zero autocorrelation sequence based matrix transform noise (AWGN) channel is measured and discussed. Also, a
(CAZACT) precoding has been proposed to reduce PAPR and simple TSPA method for combating ISI and enhancing SNR
balance SNR over data-carrying subcarriers in [23]. The similar balance is described in this section. Experimental verification is
PAPR reduction and BER improvement can be achieved by conducted in Section VIII. Finally, Section IX summarizes the
DFT and CAZACT [32]. However, inter-symbol interference conclusions drawn from this paper.
(ISI) may degrade the SNR on the edge subcarriers of the DFT
and CAZACT-precoded DMT signals. A simplified adaptive II. WORKING PRINCIPLE OF PRECODING TECHNIQUE
modulation scheme with CAZACT precoding was proposed and
In the baseband DMT transmitter, precoding is performed by
experimentally demonstrated in [24]. Y. Hong et al. proposed an
multiplying the mapped QAM symbols, x = [x1 , x2 , . . . , xM ],
orthogonal circulant matrix transform (OCT) precoding tech-
by a precoding matrix, P. After this operation, the IFFT is taken
nique to combat the severe frequency-selective fading of the
to get a real-valued time-domain DMT signal. The precoding
VLC channel [26]. It indicates that the OCT precoding has
matrix P can be expressed as
better performance in SNR balance than DFT precoding in
⎡ ⎤
the bandwidth-limited VLC system [30]. Also, a Zadoff-Chu p1,1 p1,2 · · · p1,M
sequence-based matrix transform (ZCT) has been proposed to ⎢p p2,2 · · · p2,M ⎥
1 ⎢ ⎢
2,1 ⎥
⎥
reduce the PAPR and up to 3.7 dB PAPR reduction for 1,023 P =√ ⎢ . .. . .. ⎥ (1)
subcarriers can be achieved in a compatible single-sideband M⎣ . . . . . . ⎦
DDO-OFDM system [34]. The above mentioned four types of pM,1 pM,2 · · · pM,M
precoding techniques are all realized by using complex-valued
precoding matrices. From the perspective of hardware imple- where M is the number of data-carrying subcarriers, and P is
mentation, it is preferred to use the precoding technique with the an orthogonal matrix. And the amplitude of all elements in
low complexity for optical fiber communications. In this context, the precoding matrix P is equal to 1 for DFT, CAZACT, OCT,
several real-valued matrices based precoding techniques, such as ZCT, and WHT. In this paper, vectors and matrices are denoted,
discrete cosine transform (DCT) [35], discrete Hartley transform respectively, with small bold italic letters (e.g., x, y and z) and
(DHT) [36] and Wash-Hadamard transform (WHT) [31], are capital bold italic letters (e.g., P and H). And we use small italic
also proposed to reduce PAPR or balance SNR distribution for letters with subscript indices (e.g., xk and pm,n ) to denote an
the optical OFDM transmission system. element of a vector (or a matrix). Besides, scalar constants and
Even though the precoding technique has been extensively scalar variables are denoted with capital italic letters (e.g., M
investigated in the literature, the comparison of different precod- and N) and small italic letters (e.g., n, k and i), respectively.
ing techniques in terms of PAPR reduction, nonlinear distortions The precoded symbols y = [y1 , y2 , . . . , yM ] can be written
tolerance, implementation complexity, and BER improvement as
has not been fully studied in the optical fiber DMT transmission
y T = P xT (2)
6204 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 38, NO. 22, NOVEMBER 15, 2020

where [·]T denotes the matrix transpose. Once the IFFT is done, [26]. Similarly, the real-valued precoding matrices for DCT and
the precoded DMT signal without cyclic prefix (CP) and CS can DHT are constructed by
be expressed as [37] 

1,   m=1
M
m,n =
pDCT √ (9)
1 2πkn 2 cos π(2n−1)(m−1)
, m = 1
s (n) = √ yk exp j (3) 2M
N N

k=−M
k=0 2πmn 2πmn
pDHT
m,n = cos + sin (10)
where n is an integer ranging from 0 to N−1, N is the IFFT/FFT M M
size, and y−k is the complex conjugate of yk . In general, the
Hadamard matrix, whose elements pW m,n are either +1 or
HT
average power of the transmitted QAM symbols, x, is usually
-1 and whose rows are mutually orthogonal, is used for WHT
normalized to unity with the corresponding
√ normalization factor
√ precoding. The Hadamard matrix of order M can be written as
for different QAM modulations (e.g., 2 for QPSK and 10
for 16QAM). The average power of the precoded symbols after    
H M/2 H M/2 1 1
precoding should also be 1. After the N−point IFFT operation, HM = , H2 = (11)
H M/2 −H M/2 1 −1
the average power of each data-carrying subcarrier is reduced
to 1/N, according to Eq. (3). And the average power of the
where H2 is the Hadamard matrices of order 2. If M is a
precoded DMT signal is 2M/N because 2M of such subcarriers
power of 2, the H M can be constructed from H2 . Construction
are summed.
methods for other cases can
√be found in [39]. The corresponding
precoding matric is H M / M for WHT precoding.
A. Precoding Matrices Construction
Different precoding matrices are used to realize DFT, OCT, B. SNR Balance
CAZACT, ZCT, DCT, DHT, and WHT. In DFT precoding, the
element of P in the m−th row and the n−th column is given by In the DMT receiver, we assume that the timing synchroniza-

tion is realized ideally, and there are no ISI and inter-carrier
(m − 1) (n − 1) interference (ICI). The data vector, for the M positive-frequency
pm,n = exp −j2π
DF T
(4)
M data-carrying subcarriers z = [z1 , z2 , . . . , zM ] after N−point
FFT operation, can be expressed as
where m and√n are positive integer numbers ranging from 1 to
M, and j = −1.
z = Hy T + wT (12)
For CAZACT, ZCT, and OCT precoding techniques, a Zadoff-
Chu (ZC) sequence with a length of L is used to build their where H is the channel response matrix. It is a diagonal ma-
precoding matrices with different construction methods. The trix denoted by diag(h1 , h2 , · · · , hM ) and hk is the channel
k−th element in the ZC sequence is written as [23] response of the k−th subcarrier. And w = [w1 , w2 , . . . , wM ]
⎧   2

⎨ exp j 2πr (k−1) + q (k − 1) , Leven denotes the noise vector in the frequency domain. We assume
ck =  L  2  (5) that wk obeys Gaussian distribution with zero mean and variance
⎩ exp j 2πr k(k−1) + q (k − 1) , L σk2 .
L 2 odd
We also suppose the ideal channel estimation is obtained. The
where k is an integer number ranging from 1 to L, q is an integer recovered QAM symbols, x̂ = [x̂1 , x̂2 , · · · , x̂M ], after channel
number, r is an integer relatively prime to L which equals to M 2 equalization and decoding can be given by
for both CAZACT and ZCT, and M for OCT, respectively. The
corresponding pm,n for the precoding matrices of CAZAT, ZCT, x̂T = P −1 H −1 z
and OCT, are given below.
= xT + P −1 H −1 wT
pCAZACT
m,n = cm+(n−1)M (6)
= xT + w̄T (13)
pZCT
m,n = cn+(m−1)M (7)
 where P−1 and H −1 = diag(1/h1 , 1/h2 , . . . , 1/hM ) denote
cn−m+1 , n−m+1>0
pOCT = (8) the inverse matrices of P and H, respectively. Note that P is
m,n cn−m+1+M , n − m + 1 ≤ 0 the orthogonal matrix and its inverse matrix P-1 equals to the
The precoding matrix for ZCT can be constructed by filling the conjugate transpose of P. Thus, the noise item w̄T in Eq. (13)
ZC sequence column-wise or row-wise. However, the column- can be further expanded to
wise filling method may lead to degraded performance in PAPR
reduction [38]. In this work, we only investigate the row-wise w̄T = P −1 H −1 wT
case for ZCT, unless otherwise stated. CAZACT is a special case ⎡ ∗ ⎤
p1,1 p∗2,1 · · · p∗M,1
1 ⎢ ⎥
of ZCT, and its precoding matrix is filled with the ZC sequence, ∗ ∗ ∗
which is generated with r = 1 and q = 0, by columns. Whereas, ⎢ p1,2 p2,2 · · · pM,2 ⎥
=√ ⎢ .
the elements in the m−th row of the OCT precoding matrix is M ⎣ ..
.
.. . .. .. ⎥
.
⎦
the circular shift of the M−point ZC sequence by m-1 elements p∗1,M p∗2,M · · · p∗M,M
CHEN et al.: COMPARISON OF DIFFERENT PRECODING TECHNIQUES FOR UNBALANCED IMPAIRMENTS COMPENSATION 6205

⎡ ⎤⎡ ⎤
1
h1 0 ··· 0 w1 TABLE I
IMPLEMENTATION COMPLEXITY COMPARISON
⎢ 0 1
··· 0 ⎥⎢ w2 ⎥
⎢ h2 ⎥⎢ ⎥
×⎢ .. .. .. .. ⎥⎢ .. ⎥
⎣ . . . . ⎦⎣ . ⎦
0 0 ··· 1
hM
wM
⎡ p∗1,1 w1 p∗2,1 w2 p∗M,1 wM
⎤
+ + ··· +
⎢ h1
p∗1,2 w1
h2
p∗2,2 w2
hM
p∗M,2 wM ⎥
1 ⎢ + + ··· + ⎥
⎢ h1 h2 hM ⎥
=√ ⎢ .. ⎥
M ⎢ ⎥
⎣ . ⎦
p∗1,M w1 p∗ w2 p∗M,M wM
h1 + 2,M
h2 + ··· + hM
M T
1  p∗i,1 wi  p∗i,2 wi  p∗i,M wi
M M
=√ , ,··· ,
M i=1 hi i=1 hi i=1
hi

= [w̄1 , w̄2 , · · · , w̄M ]T (14)

∗
where [·] denotes the complex conjugate, w̄k represents the
noise on the k−th subcarrier after decoding. Here, we treat it as
an Gaussian noise with zero mean and variance σ̄k2 which can be
expressed as
 2
1  2  p∗i,k 
M
σ̄k =
2
σ (15)
M i=1 i  hi 

As discussed on Eq. (3) and the average power of the trans-

mitted QAM symbols x is normalized to 1. Thus, the SNR of Fig. 1. The required real (a) multiplications and (b) additions versus data-
the k−th data-carrying subcarrier can be defined as carrying subcarriers M for different precoding techniques.

1 M
γk = =  ∗ 2 (16)
2  pi,k 
2
σ̄k M
i=1 i  hi 
σ

We know that the amplitude of elements in the precoding ma- six precoding is given in Table I. Note that the complexity of
trices is identical, i.e., |p∗i,k | = 1, for DFT, CAZACT, OCT, ZCT, real-valued DCT was presented in [40]. However, the complex-
and WHT. Therefore, the ideal SNR balance can be achieved valued QAM symbols are used in this research. The real parts and
with these five precoding techniques. However, the SNR bal- image parts of QAM symbols can be transformed independently
ance performance of DCT and DHT precoding techniques will using two real-valued DCT functions. Therefore its complexity
depend on channel response and noise distribution. This is due will be doubled. According to [41], the fast real-valued DHT
to the unequal amplitude of elements in the precoding matrices. can be realized by simply subtracting the imaginary from the
We will further discuss this problem in the sections of numerical real portion of the DFT of a real-valued sequence. It requires
simulation and offline experiments. the same multiplications and M−2 more additions than the
The conventional DMT can also be regarded as a special real-valued DFT. Similarly, the fast complex-valued DHT will
precoded DMT with a unit or identity precoding matrix. The be implemented with the same multiplications and 2M−4 more
corresponding SNR can be derived from Eq. (16) and given by additions than the complex-valued DFT. While the complexity
|hk |2 of fast algorithms for CAZACT, OCT and ZCT is the same
γk = (17) and equivalent to one M−point complex-valued DFT and two
σk2 M−point complex-valued multiplications [30]. The multiplica-
tions and additions as a function of data-carrying subcarriers
III. COMPLEXITY COMPARISON M for their fast algorithms and WHT precoding are also shown
The implementation complexity, according to Eq. (2) and in Fig. 1(a) and (b), respectively. It shows that complex-valued
of the corresponding possible fast algorithms is analyzed. The OCT, CAZACT, and ZCT precoding techniques require more
required real multiplications and additions are listed in Table I. multiplications than real-valued DCT, DFT and DHT precoding
A straightforward implementation of precoding techniques may ones. And similar additions are needed for the precoding tech-
be very resource-intensive. These precoding techniques can be niques except for WHT. More additions but no multiplications
implemented with their fast algorithms by exploiting DFT [30], are required for the implementation of WHT precoding. Pipeline
[40], [41]. It turns out that a significant reduction in imple- multipliers are required for high-speed optical communications
mentation complexity is achieved. When M is a power of 2, and more resource intensive than adders. Therefore, WHT pre-
the implementation complexity of the fast algorithms for the coding has the lowest complexity.
6206 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 38, NO. 22, NOVEMBER 15, 2020

Fig. 2. The block diagram of simulation model.

Fig. 4. The frequency response of the 3rd order FIR LPF.

as [9]

CRdB = 20 · log10 (C/rms [s (n)]) (19)

where rms[s(n)] denotes the root mean square of the DMT signal
s(n).
The other is a simple solid-state power amplifier (SSPA)
model. The transfer function of SSPA can be defined as [42]
g0 · A · sgn (x (t))
y (t) = 1/s
(20)
(1 + (A/|x (t)|)s )
Fig. 3. The normalized transfer function of two nonlinear models.
where g0 is the amplifier gain, A denotes the limited level or
the input saturation amplitude, and the parameter S controls the
sharpness of the saturation region. The input and output levels
IV. SIMULATION MODEL
are x(t) and y(t), respectively. We take S = 4 in our simulation
In the numerical simulation, a DMT transmission system as a good approximation of the commercial amplifiers [43]. The
model with nonlinear distortion, bandwidth limitation and Gaus- normalized transfer function curve is also plotted in Fig. 3.
sian noise is established, as depicted in Fig. 2. The random In the simulation, unbalanced impairments are emulated by
binary bits are first mapped to QAM symbols and followed by a 3rd order finite impulse response (FIR) low-pass filter (LPF).
precoding. And the precoded symbols are constrained to have The frequency response as a function of the normalized fre-
Hermitian symmetry (HS) to offer a real-valued time-domain quency is shown in Fig. 4. The maximum positive-frequency
signal after IFFT operation. Afterward, a CP is inserted in front subcarrier index of the DMT signal is 384. The corresponding
of each IFFT output to combat the ISI caused by the bandwidth normalized frequency is 0.75. Therefore the power fading over
limitation. Besides, a training symbol (TS) is also added for the data-carrying subcarriers can be up to about 20 dB after the
each DMT frame to realize both timing synchronization and LPF. The average power of the k−th data-carrying subcarrier of
channel equalization [9]. In this work, the IFFT size (N), CP the filtered signal will be |hk |2 /N , according to the discussion
and CS lengths are 1024, 32 and 0, respectively. Among 1024 of Eq. (3). Once the filtered DMT signal passes through an
subcarriers, direct current (DC) subcarrier is reserved, 768 low- AWGN channel and the noise is evenly distributed to all N
frequency subcarriers are used for data transmission and other subcarriers, the unbalanced SNR over data-carrying subcarriers
high-frequency subcarriers are filled with zeros to achieve an will be observed. If no nonlinear distortions are considered, then
oversampling factor of 1.33. Three QAM modulation formats, the noise power σ 2 on each subcarrier before N−point FFT
i.e., 4, 16 and 64QAM, are discussed. There are one TS and operation can be written as
1,000 DMT symbols in each DMT frame.  
Two types of nonlinear distortion models are considered. One 2 M |hi |2 N 
i=1 2 M i=1 |hi |
2
is the well-known peak clipping model and its transfer function σ2 = = (21)
γN γN 2
for the DMT signal can be described by
 where γ is the SNR of the AWGN channel. The noise power
x (t) , |x (t)| < C after N−point FFT operation, i.e., σk2 , on each subcarrier is also
y (t) = (18)
sgn (x (t)) · C, |x (t)| ≥ C identical and equal to Nσ 2 , according to the definition of FFT in
[37]. Subsequently, the noise power σk2 can be substituted into
where C is the clipping threshold value, and sgn(·) represents the Eqs. (16) and (17) to obtain the SNR on the each data-carrying
sign function. The normalized transfer function curve is given subcarrier for the precoded DMT signals and the conventional
in Fig. 3. Here, the clipping ratio (CR) in decibel is defined signals, respectively.
CHEN et al.: COMPARISON OF DIFFERENT PRECODING TECHNIQUES FOR UNBALANCED IMPAIRMENTS COMPENSATION 6207

Fig. 6. The EVM performance versus CR using peak clipping model:

(a) QPSK, (b) 16QAM and (c) 64QAM.

is 384, and r and q for the CAZAC sequence generation are

1 and 0 [23], [24], respectively. The precoding matrices for
OCT and ZCT are generated with r = 11 and q = 1. We can
see that the ZCT precoding offers superior PAPR reduction
capability, and a similar PAPR performance can be observed
for CAZACT and DFT precoding techniques. Besides, OCT
precoding has no ability to reduce the PAPR. At a CCDF of
1e-3, the PAPRs of the QPSK-DMT signals with ZCT, DFT,
CAZACT, DCT, DHT, and WHT precoding techniques can
be reduced by 3.7, 3.5, 3.5, 3.2, 2.1 and 0.4 dB, respectively,
compared to the conventional DMT signal. However, we also
observe that the PAPR performance with precoding is degraded
for higher-order modulation formats. And there are only 2.7 and
2.5 dB PAPR reductions with ZCT precoding, at the CCDF of
1e-3, for 16QAM and 64QAM, respectively. It is concluded from
Fig. 5 that the PAPR of the precoded DMT signals, except for
OCT and WHT cases, is sensitive to QAM modulation formats.
The PAPR of the precoded DMT signals increases with the order
of modulation format. One possible reason for this phenomenon
is that the high-order QAM symbols have higher PAPRs than
that of the low-order ones.
Fig. 5. The PAPR of different precoded DMT signals with (a) QPSK,
(b) 16 QAM and (c) 64 QAM mapping. VI. NONLINEAR DISTORTIONS
It is interesting to explore the impacts of nonlinear distortions
In this paper, we define the error vector magnitude (EVM) in on the precoded DMT transmission system. Firstly, we measure
decibel as the EVM performance under the different CRs using the peak
⎛K K ⎞ clipping model, as shown in Fig. 6. It seems that better EVM
  performance is obtained with larger PAPR reduction after peak
EV MdB = 20 · log10 ⎝ |ek |2 |xk |2 ⎠ (22) clipping. A similar EVM performance is observed for ZCT,
k=1 k=1
DFT, and CAZACT precoding techniques. There is no difference
where xk denotes the reference QAM symbol, x̂k is the recovered in EVM performance between the OCT precoded DMT signal
version, K is the total number of the transmitted QAM symbols and the conventional DMT signal. The EVM performance is
and ek = x̂k − xk is the error signal. In this work, K = 1,000 × slightly improved by using WHT precoding due to its low PAPR
M = 384,000. reduction. It also exhibits that the DHT precoded DMT signal
is more sensitive to clipping noise than the DCT precoded one.
Although there is a difference in PAPR between DCT and DHT
V. PAPR OF THE PRECODED DMT SIGNALS precoding, a similar EVM performance can be found, especially
The PAPR reduction performance of different precoding tech- for the 16QAM and 64QAM modulation formats.
niques is investigated by numerical simulations. The comple- The BER performance as a function of CR for the three
mentary cumulative distribution function (CCDF) as a function modulation formats is also investigated and presented in Fig. 7.
of PAPR for the precoded DMT signals is shown in Fig. 5. With QPSK modulation format, DHT precoding outperforms
Each CCDF curve is calculated from 100,000 DMT symbols. other precoding techniques in combating clipping distortions.
And three QAM modulation formats, i.e., QPSK, 16QAM and An improvement in BER performance can be achieved with
64QAM, are discussed. The corresponding CCDF curves are precoding techniques, except for WHT and OCT precoding
presented in Fig. 5(a), (b), and (c), respectively. Note that M techniques. Besides, a slight BER performance degradation
6208 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 38, NO. 22, NOVEMBER 15, 2020

Fig. 7. The measured BER performance versus CR using peak clipping model. Fig. 9. The measured BER performance versus normalized limited level using
SSPA model.

Fig. 8. The EVM performance versus the normalized limited level using SSPA
model: (a) QPSK, (b) 16QAM and (c) 64QAM.

Fig. 10. The recovered 16QAM constellation diagrams (Anorm = 2.4). The
is also observed for WHT precoding. However, the precoded ideal constellation points are marked with red circles.
DMT signals with 16QAM and 64QAM modulation formats,
except for the OCT precoding, are more sensitive to the peak that we only discuss the effects of the nonlinear distortions on
clipping at the large CR. The BER performance of the precoded the performance of the DMT signals without taking other factors
16/64QAM-DMT signals, with a CR of larger than 4/6.5 dB, is (e.g., noises and bandwidth limitations) to account. It does
worse than that of the conventional and OCT precoded DMT not mean the conventional and OCT-precoded DMT systems
signals. The improvement in the BER performance is obtained always outperforms the other precoded DMT ones suffering
by using precoding techniques, excluding WHT, only when the noises or/and other impairments under relatively low nonlinear
16/64QAM-DMT signals suffer heavy clipping distortions. distortions. We will discuss this point in detail in Section VII.
Similarly, we also investigate the EVM and BER perfor- In an SSPA model with a normalized limited level of 2.4,
mances of the precoded DMT signals using the SSPA model. the recovered constellation diagrams are plotted in Fig. 10.
Instead of changing the amplitude of the signal, we adjust Although there is an improvement in EVM performance by
the limited level A of the SSPA model to explore the effect using precoding techniques except for OCT, a small number of
of different nonlinear distortions on the precoded signals. We recovered 16QAM symbols show a more significant deviation
define the normalized limited level as Anorm = A/rms[s(n)]. from their ideal constellation points than that of the conventional
The corresponding EVM and BER results are shown in Figs. 8 one. This case may lead to the degraded BER performance
and 9, respectively. in nonlinear distortion-limited high-order QAM encoded DMT
As we can see from the EVM and BER performance curves, systems. However, the QPSK modulation format, with larger
similar conclusions can be obtained for the SSPA model, com- minimum Euclidean distance (MED), is less sensitive to the non-
pared to the peaking clipping model. OCT precoded signals have linear distortion. And the BER performance of the QPSK-DMT
the same EVM and BER performances as the conventional DMT system with distortion limitations can be improved by using
signals with different modulation formats. Precoding techniques DFT, CAZACT, ZCT, DCT, and DHT precoding techniques.
may be employed to combat severe nonlinear distortions and
improve the BER performance except for OCT and WHT cases.
VII. UNBALANCED IMPAIRMENTS COMPENSATION
However, the BER performance of the precoded DMT signals
with high-level modulation formats, excluding OCT, are more In this section, we first discuss the unbalanced impairments
sensitive to the nonlinear distortion. It leads to the degraded BER emulated by an LPF with high power fading in the
performance under relatively low nonlinear distortions. Note high-frequency subcarriers. This phenomenon often can
CHEN et al.: COMPARISON OF DIFFERENT PRECODING TECHNIQUES FOR UNBALANCED IMPAIRMENTS COMPENSATION 6209

response of LPF. Also, we calculate the corresponding theoret-

ical SNR in the absence of ISI, according to Eqs. (16), (17) and
(21). The estimated SNR distributions for different precoding
techniques agree well with theoretical ones. The unbalanced
SNR distributions for DHT and DCT precoding techniques are
attributed to their precoding matrices which consist of elements
with unequal amplitudes. It should be noted that the estimated
average SNR values of all precoded signals are identical to
about 24 dB. Therefore, high SNRs on the edge subcarriers
can be observed for the DCT precoding, but SNRs on the other
Fig. 11. The normalized timing metric versus sample index (SNR = 30 dB).
subcarriers are slightly degraded.
In the presence of ISI, the overall SNR performance of the
conventional DMT signal is only slightly degraded, which is
benefited from the use of large-size FFT. In [21], it exhibits that
large-size FFT can be used to combat ISI more effectively. The
identical average SNR values of 23.7 dB can also be obtained for
both the conventional DMT and precoded ones. However, the be-
haviors of SNR distribution on the data-carrying subcarriers are,
to some extent, different even if the precoding matrices consist of
elements with equal amplitudes. Among precoding techniques,
OCT precoding shows good robustness against ISI and achieves
the best SNR balance. As we can see from Fig. 12(c), slight
SNR degradation occurs on the tens of consecutive subcarriers.
However, severe SNR degradation centrally occurs on a small
number of subcarriers for other precoding techniques. In the
literature, we can find similar situations in [24]–[26]. In this case,
the BER performance of the precoded DMT transmission system
may be limited by the ISI. One may use CS to combat this type of
ISI [30], however, which results in reduced SE. Alternatively, an
accurate timing synchronization method, which is robust against
the low-pass filtering effect and SCPO, can be employed to avoid
the effect of ISI on the precoding. In this paper, we use a simple
Fig. 12. The estimated SNR over different data-carrying subcarriers for w/
and w/o ISI cases (SNR = 30 dB). TSPA method to reduce the effect of ISI on the SNR balance
performance of the precoded DMT signals. By using TSPA,
the final estimated start point of the DMT frame is adjusted to
dmax − Q, where Q is a non-zero positive integer and dmax is
be found in the bandwidth-limited optical/electrical devices
the sample index when the timing metric is maximized. As we
(e.g., DAC, DML and LED) and CD-induced frequency
can see from Fig. 11, the Q value can be set to 2, which is enough
selective power fading. The low-pass filtering effect may
to combat this type of ISI.
degrade the performance of the TS-based autocorrelation
We also investigate the BER performance of the precoded sig-
timing synchronization [9], which may lead to ISI if no CS is
nals which suffer from both nonlinear distortions and low-pass
applied. Fig. 11 shows the normalized timing metrics of the
filtering in an AWGN channel. Here, we use the SSPA model
timing synchronization where two cases, i.e., only the AWGN
with a normalized limited level of 4.2. Intra-symbol frequency
channel and LPF+AWGN channel, are considered.
averaging method [45] is employed to enhance the accuracy
The accurate start position of the DMT frame is at the
of channel estimation. The BER performance as a function of
sample index of 1001 in our simulation. It can be accurately
SNR ranging from 10 to 40 dB is depicted in Fig. 13. As we
estimated under AWGN channel. However, there two similar
can see from Fig. 11, the estimated start position of the DMT
peaks at sample indices of 1002 and 1003 can be observed for
frame with TS maybe 1002 or 1003. Here, we fix it at 1002 to
the LPF+AWGN case. This inaccurate timing synchronization
keep the consistency under different SNRs. We investigate the
will lead to ISI and degrade the SNR balance after precoding.
BER performance under two cases, i.e., without TSPA and with
Note that sampling clock phase offsets (SCPO) between DAC
TSPA, and the corresponding results are presented in Fig. 13(a)
and ADC will also influence the timing performance, and a
and (b), respectively.
similar timing metric can also be obtained as presented in our
In general, the BER performance can be improved by using
previous works [9], [44]. The SNR as a function of the subcarrier
the precoding technique. However, it may be limited by the
index is estimated for different precoding techniques in the
ISI. Fig. 13(a) exhibits that OCT precoding with high-order
presence/absence of ISI, as shown in Fig. 12. Here, the SNR of
modulation formats can achieve the best BER performance than
the AWGN channel is set as 30 dB. Note that no nonlinear dis-
other precoding techniques. This fact is due to its more rubout
tortions are considered, and zero-forcing channel equalization
against the ISI. Also, similar BER performance can be observed
is performed with accurate channel estimation from the channel
6210 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 38, NO. 22, NOVEMBER 15, 2020

Fig. 14. Experimental setup.

signal is suppressed by an LPF. A 14 GHz electrical amplifier

(EA, Mini-Circuit ZX60-14012L-S+) is employed to boost the
filtered signal and drives a low-cost 10 Gbit/s DML (NEL,
NLK1551SSC) biased at its optimal linear regime. The central
wavelength and optical power of the optical DMT signal are
1558 nm and 2.3 dBm, respectively. An erbium-doped fiber
amplifier (EDFA) manufactured by Amonics and followed by
a variable optical attenuator (VOA) is used to adjust the launch
power. The maximum output power of the EDFA is about 13
dBm, and the insertion loss of VOA is about 1 dB. Therefore, the
maximum launch power in our experiment is 12 dBm. It should
Fig. 13. The measured BER performance versus SNR: (a) w/o TSPA and be noted that the launch power can also be tuned by adjusting
(b) w/ TSPA. the pump current of the EDFA. However, the greater noise figure
(NF) can be found in [46] when the pump power is relatively
low. Therefore, we use the VOA to change the launch power
for the precoded signal with QPSK modulation format due and maintain the same NF for different launch powers. After
to its large MED. Furthermore, there are slight degradations 20 km of single-mode fiber (SMF) transmission, the second
in the BER performance in relative low SNRs for both w/o VOA and an optical coupler (OC) placed in front of a 10 GHz
TSPA and w/ TSPA cases, compared to the conventional one. photodiode (PD) and a power meter (PM) to change and measure
One possible reason may be the large SNR difference over the received optical power (ROP) simultaneously. The power
data-carrying subcarriers before precoding. By using the TSPA ratio of the OC is 9:1. The recovered signal is amplified by the
method, the BER performance of the precoded signals can be second EA and captured by a Lecory digital storage oscilloscope
obviously improved under high SNRs, as shown in Fig. 13(b). (DSO, Wavemaster 820Zi-A). The analog bandwidth, vertical
We also observe the error floor in BER performance for the three resolution, and maximum sampling rate are 20 GHz, 8-bit, and
real-valued precoding techniques due to nonlinear distortions. 40 GSa/s, respectively. The sampling rate is set to 20 GSa/s in
our experiment. The captured samples are stored and resampled
to 12 GSa/s. The following data processing procedure is shown
VIII. EXPERIMENTAL VERIFICATIONS
in Fig. 2.
A proof-of-concept short-reach optical DMT transmission The estimated SNR from the recovered QPSK symbols after
experiment is conducted to fully investigate the performance 20 km SMF transmission as a function of the subcarrier (SC)
of different precoding techniques. The block diagram of the index is presented in Fig. 15. The peak-to-peak voltage (Vpp)
experimental setup is illustrated in Fig. 14. In the transmitter, of the output signal from the AWG is set to 500 mV. There
the digital DMT signals are generated offline using the same are unbalanced impairments over data-carrying subcarriers of
DSP algorithms and parameters as simulations and uploaded to the conventional DMT signal. Low SNRs can be found on both
a Tektronix arbitrary waveform generator (AWG, AWG7122C) low- and high-frequency subcarriers. The imperfect frequency
operating at 10-bit resolution and the sampling rate of 12 GSa/s. response of the DML is the main reason for low SNRs on the sub-
It should be pointed out that the average power of the pre- carriers around DC. The degraded SNR on the high-frequency
coded DMT signals and the conventional signal remains the subcarriers is mainly attributed to the bandwidth limitations of
same by inserting the same peak amplitude in front of TS in AWG and DML, and CD-induced power fading. Even though
our experiment. The high-frequency image of the converted the bandwidth of DMT signals and the transmission distance are
CHEN et al.: COMPARISON OF DIFFERENT PRECODING TECHNIQUES FOR UNBALANCED IMPAIRMENTS COMPENSATION 6211

Fig. 17. The measured BER performance with different signal Vpp. (LP =
6 dBm and ROP = 0 dBm).
Fig. 15. The estimated SNR versus subcarrier (SC) index for different DMT
signals. (Vpp = 500 mV, ROP = 0 dBm and LP = 6 dBm).

Fig. 18. The recovered 16QAM constellation diagrams and the corresponding
EVM values. (Vpp = 500 mV, LP = 6 dBm and ROP = 0 dBm).

leads to slight BER performance degradation. We also can see

that the nonlinear effect induced by the optical fiber is small
and can be ignored even though the LP is set up to 12 dBm.
Fig. 16. The measured BER performance versus LP. (Vpp = 500 mV and The error-free transmission is observed for the QPSK modula-
ROP = 0 dBm). tion format. And the BER performance of the 16QAM-DMT
signal can be improved by more than one order of magnitude
by using precoding techniques, compared to the conventional
only 4.5 GHz and 20 km, respectively, the CD-induced power DMT case. According to the relationship between BER and
fading is intensified after fiber transmission due to the strong SNR, the required SNR is 19.8 dB when the desired BER is
positive transient chirp of the DML [47], [48]. The effect of 1e-2 for 64QAM modulation format [15]. The measured BER
ISI on the conventional DMT signal with large-size FFT is ig- of the precoded 64QAM-DMT signals is well consistent with
nored. It further indicates that the SNR balance of the precoding the theoretical value. The slight BER improvement is obtained
techniques can be enhanced by employing the proposed simple after precoding for 64QAM modulation. This fact is mainly due
TSPA method. Only the case of with TSPA is considered in the to the limited average SNR. It is concluded from Fig. 16 that
following discussions. precoding exhibits superior BER improvement for low-order
We adjust the launch power (LP) to investigate the effect of QAM modulation format due to its larger Euclidean distance.
optical fiber nonlinearity on the BER performance. The related We set the LP to 6 dBm in the following discussions.
results are shown in Fig. 16. The signal Vpp is set to 500 We also investigate the different output signal Vpp from the
mV and ROP is fixed at 0 dBm for the BER measurements AWG to further explore the effects of the nonlinear distortions
when LP is higher than 6 dBm. Due to the fiber attenuation induced by the EA and DML on the BER performance. The
and insertion losses of OC and two VOAs, the corresponding related results are given in Fig. 17. Since the minimum signal
ROP is less than 0 dBm when LP is set to 2 or 4 dBm. This Vpp from the AWG is 500 mV, we use a fixed 3 dB attenuator
6212 JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 38, NO. 22, NOVEMBER 15, 2020

account, the four complex-valued precoding techniques, i.e.,

ZCT, DFT, CAZACT and OCT, and one real-valued precoding
technique WHT can achieve the best SNR balance. In the view
of hardware implementation, no multipliers but only additions
are required for WHT precoding; therefore, it has the lowest
implementation complexity. And OCT precoding has the highest
implementation complexity. We comparably investigated these
precoding techniques in terms of PAPR reduction and nonlinear
distortion tolerance by numerical simulations. It showed that
three complex-valued precoding techniques (ZCT, DFT, and
CAZACT) could achieve better PAPR reduction than the other
three real-valued precoding techniques (DCT, DHT, and WHT)
for different QAM modulation formats. Up to 3.7 dB PAPR
reduction can be obtained by using ZCT precoding with QPSK
modulation. However, less PAPR reduction is observed for
higher-order modulation formats. And OCT precoding cannot
reduce PAPR. With higher PAPR reduction, more improvements
Fig. 19. The measured BER performance after 20 km transmission under in EVM performance can be obtained under both the peak
different ROPs. (Vpp = 500 mV and LP = 6 dBm).
clipping and SSPA models. It was indicated that the precoded
QPSK-DMT signals benefited from high PAPR reduction and
or two cascaded fixed 3-dB attenuators followed by the AWG outperformed the precoded ones with less or without PAPR
with the signal Vpp of 500 mV to further reduce the signal Vpp. reduction as well as the conventional one in the distortion-
It indicates that the optimal signal Vpp is about 500 mV. For the limited case. But the precoded 16/64QAM-DMT signals are
small signal Vpp, the various noises (e.g., PD-induced noises) more sensitive to nonlinear distortions and provide minor BER
may be the dominant factor for the degraded BER performance. improvement or even may degrade the BER, compared to the
With the increase of signal Vpp, the BER performances for all OCT precoded and the conventional ones.
three modulation formats are gradually degraded. It shows the The estimated SNR distributions after decoding agree well
complex-valued DFT/CAZACT/ZCT precoding techniques are with the theoretical ones. In the ISI-limited scenario, OCT
more robust against these nonlinear distortions than the other precoding can achieve the best SNR balance and obtain the
precoding ones for 16 and 64QAM modulation formats. The best BER performance. Besides, the SNR balance and the BER
experimental results agree with the results obtained by using performance of the precoded DMT signals can be enhanced by
nonlinear distortion models, as shown in Figs. 7 and 9. The the simple TSPA method in the presence of the ISI induced by
BER performance is the noise-limited in small signal Vpp cases. channel low-pass filtering or/and SCPO. In the noise-limited
Therefore, there is no obvious difference in the BER perfor- case, the unbalanced impairments can be well compensated
mance between OCT and the other precoded 16QAM DMT by every precoding technique. And all precoding techniques
signals. The recovered 16QAM constellation diagrams and their exhibited a similar improvement in BER performance.
EVM values are shown in Fig. 18. The slight improvements in Finally, the proof-of-concept experimental demonstrations
the EVM performance, compared to the conventional signal, are also performed in a short-reach DML-based DMT transmis-
are mainly obtained from the PAPR reduction with precoding sion system. Under the average power constraint, the unbalance
techniques. impairments caused by the imperfect frequency response of the
The BER performance of the DMT signals after 20-km DAC and DML and CD-induced power fading were well com-
SMF transmission versus ROP is also investigated, as shown pensated. And the receiver sensitivities of QPSK- and 16QAM-
in Fig. 19. At the BER of 1e-3, the receiver sensitivities of DMT were improved by 1.4 and 4 dB, respectively. A similar
QPSK and 16QAM-DMT transmission systems with precoding, improvement in BER performance can be observed for these
in terms of ROP, can be improved by 1.4 and 4 dB, respectively, precoding techniques. Considering both the BER performance
compared to the conventional DMT ones. The error floor at the improvement and implementation complexity, WHT precoding
high ROPs higher than −6 dBm is mainly due to that the PD is may be a good option to compensate unbalanced impairments
close to saturation. Noise limitation is the main reason for the in the noise-limited optical DMT transmission system.
slight improvement in BER performance of the 64QAM-DMT
signals. In view of BER performance improvement and imple-
mentation complexity, WHT precoding may be a good option for REFERENCES
the unbalanced impairments compensation in the noise-limited [1] T. Takahara et al., “Discrete multi-tone for 100 Gb/s optical ac-
optical DMT transmission system. cess networks,” in Proc. Opt. Fiber Commun. Conf. Exhib., 2014,
Paper M2I.1.
[2] C. Qin, V. Houtsma, D. Van Veen, J. Lee, H. Chow, and P. Vetter, “40 Gbps
IX. CONCLUSION PON with 23 dB power budget using 10 Gbps optics and DMT,” in Proc.
Opt. Fiber Commun. Conf. Exhib., 2017, Paper M3H.5.
In this work, the SNR balance and implementation complexity [3] A. Dochhan, H. Grießer, M. Eiselt, and J. Elbers, “Flexible bandwidth
were theoretically analyzed and discussed for seven commonly 448 Gb/s DMT transmission for next generation data center inter-
used precoding techniques in detail. Without taking ISI/ICI into connects,” in Proc. Eur. Conf. Opt. Commun., 2014, Paper P.4.10.
CHEN et al.: COMPARISON OF DIFFERENT PRECODING TECHNIQUES FOR UNBALANCED IMPAIRMENTS COMPENSATION 6213

[4] F. Li, J. Yu, Z. Cao, J. Zhang, M. Chen, and X. Li, “Experimental [26] Y. Hong, X. Guan, L. Chen, and J. Zhao, “Experimental demonstration
demonstration of four-channel WDM 560 Gbit/s 128QAM-DMT using of an OCT-based precoding scheme for visible light communications,” in
IM/DD for 2-km optical interconnect,” J. Lightw. Technol., vol. 35, no. 4, Proc. Opt. Fiber Commun. Conf. Exhib., 2016, Paper M3A.6.
pp. 941–948, Feb. 2017. [27] X. Chen, Z. Feng, M. Tang, S. Fu, and D. Liu, “Performance enhanced
[5] F. Hu et al., “20.09-Gbit/s underwater WDM-VLC transmission based on DDO-OFDM system with adaptively partitioned precoding and single
a single Si/GaAs-substrate multichromatic LED array chip,” in Proc. Opt. sideband modulation,” Opt. Express, vol. 25, no. 19, pp. 23093–23108,
Fiber Commun. Conf. Exhib., 2020, Paper M3I.4. 2017.
[6] H. Qian, S. Cai, S. Yao, T. Zhou, Y. Yang, and X. Wang, “On the benefit [28] T. Jiang et al., “Investigation of DC-biased optical OFDM with precoding
of DMT modulation in nonlinear VLC systems,” Opt. Express, vol. 23, matrix for visible light communications: Theory, simulations, and exper-
no. 3, pp. 2618–2632, 2015. iments,” IEEE Photon. J., vol. 10, no. 5, Oct. 2018, Art. no. 7906916.
[7] Y. Benlachtar et al., “Generation of optical OFDM signals using 21.4 [29] Y. Hong, L. Chen, and J. Zhao, “Performance-enhanced gigabit/s MIMO-
GS/s real time digital signal processing,” Opt. Express, vol. 17, no. 29, OFDM visible light communications using CSI-free/dependent precoding
pp. 17658–17668, 2009. techniques,” Opt. Express, vol. 27, no. 9, pp. 12806–12816, 2019.
[8] X. Xiao, F. Li, J. Yu, Y. Xia, and Y. Chen, “Real-time demonstration of 100 [30] J. Zhao, Y. Hong and L. Chen, “Discrete-circulant-transform spread
Gbps class dual-carrier DDO-16QAM-DMT transmission with directly OFDM for bandwidth-limited VLC systems,” J. Lightw. Technol., vol. 37,
modulated laser,” in Proc. Opt. Fiber Commun. Conf. Exhib., 2014, Paper no. 20, pp. 5340–5353, Oct. 2019.
M2E.6. [31] R. Deng, J. He, M. Chen, and Y. Zhou, “Experimental demonstration of
[9] M. Chen, J. He, Q. Fan, Z. Dong, and L. Chen, “Experimental demonstra- a real-time gigabit OFDM-VLC system with a cost-efficient precoding
tion of real-time high-level QAM-encoded direct-detection optical OFDM scheme,” Opt. Commun., vol. 423, pp. 69–73, 2018.
systems,” J. Lightw. Technol., vol. 33, no. 22, pp. 4632–4639, Nov. 2015. [32] J. Ma, M. Chen, K. Wu, and J. He, “Performance enhancement of proba-
[10] J. Zhang et al., “Stage-dependent DSP operation range clipping-induced bilistically shaped OFDM enabled by precoding technique in an IM-DD
bit resolution reductions of full parallel 64-point FFTs incorporated in system,” J. Lightw. Technol., vol. 37, no. 24, pp. 6063–6071, Dec. 2019.
FPGA-based optical OFDM receivers,” J. Lightw. Technol., vol. 34, no. [33] J. Ma et al., “Performance enhanced 256-QAM BIPCM-DMT system
16, pp. 3752–3760, Aug. 2016. enabled by CAZAC precoding,” J. Lightw. Technol., vol. 38, no. 3,
[11] M. Chen, P. Zou, L. Zhang, and N. Chi, “Demonstration of a 2.34 Gbit/s pp. 557–563, Feb. 2020.
real-time single silicon-substrate blue LED-based underwater VLC sys- [34] Ö. Bulakçı, M. Schuster, C. Bunge, and B. Spinnler, “Precoding based
tem,” IEEE Photon. J., vol. 12, no. 1, Feb. 2020, Art. no. 7900211. peak-to-average power ratio reduction for optical OFDM demonstrated
[12] H. Dar-Zu et al., “2.1-Tb/s/km OFDM long-reach PON transmission on compatible single-sideband modulation with direct detection,” in Proc.
using a cost-effective electro-absorption modulator,” in Proc. Opt. Fiber Opt. Fiber Commun. Conf. Exhib., 2008, Paper JThA56.
Commun. Conf. Exhib., 2011, Paper OMG2. [35] O. Bulakci, M. Schuster, C. Bunge, and B. Spinnler, “Reduced complexity
[13] Z. Cao, G. Wen, F. Li, Q. Shu, J. Yu, and L. Chen, “Unbalanced impair- precoding based peak-to-average power ratio reduction applied to opti-
ments compensation for low cost direct detection OFDM-PON systems,” cal direct-detection OFDM,” in Proc. Eur. Conf. Opt. Commun., 2008,
Opt. Commun., vol. 310, pp. 35–41, 2014. Paper P.4.11.
[14] D. Zou et al., “Comparison of null-subcarriers reservation and adaptive [36] X. Ouyang, J. Jin, G. Jin, and Z. Wang, “Low complexity discrete Hartley
notch filter for narrowband interference cancellation in intra-data center transform precoded OFDM for peak power reduction,” Electron. Lett.,
interconnect with DMT signal transmission,” Opt. Express, vol. 27, no. 4, vol. 48, no. 2, pp. 90–91, 2012.
pp. 5696–5702, 2019. [37] J. Armstrong, “OFDM for optical communications,” J. Lightw. Technol.,
[15] L. Peng, M. Hélard, and S. Haese, “On bit-loading for discrete multi-tone vol. 27, no. 3, pp. 189–204, Feb. 2009.
transmission over short range POF systems,” J. Lightw. Technol., vol. 31, [38] I. Baig and V. Jeoti, “A new ZCT precoded OFDM system with pulse
no. 24, pp. 4155–4165, Dec. 2013. shaping: PAPR analysis,” in Proc. IEEE Asia Pacific Conf. Circuits Syst.,
[16] C. Li, R. Hu, Q. Yang, M. Luo, W. Li, and S. Yu, “Fading-free transmission 2010, pp. 1131–1134.
of 124-Gb/s PDM-DMT signal over 100-km SSMF using digital carrier [39] J. H. van Lint and R. M. Wilson, A Course in Combinatorics. Cambridge,
regeneration,” Opt. Express, vol. 24, no. 2, pp. 817–824, 2016. U.K.: Cambridge Univ. Press, 2001.
[17] Y. Gao, J. Yu, J. Xiao, Z. Cao, F. Li, and L. Chen, “Direct-detection optical [40] M. Vetterli and H. J. Nussbaumer, “Simple FFT and DCT algorithms with
OFDM transmission system with pre-emphasis technique,” J. Lightw. reduced number of operations,” Signal Process., vol. 6, pp. 267–278, 1984.
Technol., vol. 29, no. 17, pp. 2138–2145, Jul. 2011. [41] H. Sorensen, D. Jones, C. Burrus, and M. Heideman, “On computing the
[18] S. Karabetsos, E. Pikasis, T. Nikas, A. Nassiopoulos, and D. Syvridis, discrete Hartley transform,” IEEE Trans. Acoust., Speech, Signal Process.,
“DFT-spread DMT modulation for 1-Gb/s transmission rate over 100 m of vol. ASSP-33, no. 5, pp. 1231–1238, Oct. 1985.
1-mm SI-POF,” IEEE Photon. Technol. Lett., vol. 24, no. 10, pp. 836–838, [42] A. Cann, “Nonlinearity model with variable knee sharpness,” IEEE Trans.
May 2012. Aerosp. Electron. Syst., vol. AES-16, no. 6, pp. 874–877, Nov. 1980.
[19] T. Truong, M. Arzel, H. Lin, B. Jahan and M. Jézéquel, “DFT precoded [43] S. Amiralizadeh, A. Nguyen, and L. Rusch, “Modeling and compensation
OFDM—An alternative candidate for next generation PONs,” J. Lightw. of transmitter nonlinearity in coherent optical OFDM,” Opt. Express,
Technol., vol. 32, no. 6, pp. 1228–1238, Mar. 2014. vol. 23, no. 20, pp. 26192–26207, 2015.
[20] F. Li, J. Yu, Y. Fang, Z. Dong, X. Li, and L. Chen, “Demonstration of DFT- [44] M. Chen, J. He, Z. Cao, J. Tang, L. Chen, and X. Wu, “Symbol synchro-
spread 256QAM-OFDM signal transmission with cost-effective directly nization and sampling frequency synchronization techniques in real-time
modulated laser,” Opt. Express, vol. 22, no. 7, pp. 8742–8748, 2014. DDO-OFDM systems,” Opt. Commun., vol. 326, pp. 80–87, 2014.
[21] F. Li, X. Li, L. Chen, Y. Xia, C. Ge, and Y. Chen, “High-level QAM OFDM [45] X. Liu and F. Buchali, “Intra-symbol frequency-domain averaging based
system using DML for low-cost short reach optical communications,” channel estimation for coherent optical OFDM,” Opt. Express, vol. 16,
IEEE Photon. Technol. Lett., vol. 26, no. 9, pp. 941–944, May 2014. no. 26, pp. 21944–21957, 2008.
[22] F. Li, X. Li, J. Zhang, and J. Yu, “Transmission of 100-Gb/s VSB DFT- [46] R. I. Laming, M. N. Zervas, and D. N. Payne, “Erbium-doped fibre
spread DMT signal in short-reach optical communication systems,” IEEE amplifier with 54 dB gain and 3.1 dB noise figure,” IEEE Photon. Technol.
Photon. J., vol. 7, no. 5, Oct. 2015, Art. no. 7904307. Lett., vol. 4, no. 12, pp. 1345–1347, Dec. 1992.
[23] Z. Feng et al., “Performance-enhanced direct detection optical OFDM [47] I. Tomkos, B. Hallock, I. Roudas, R. Hesse, A. Boskovic, and R. Vodhanel,
transmission with CAZAC equalization,” IEEE Photon. Technol. Lett., “Transmission of 1550 nm 10 Gb/s directly modulated signal over 100 km
vol. 27, no. 14, pp. 1507–1510, Jul. 2015. of negative dispersion fiber without any dispersion compensation,” in Proc.
[24] Z. Feng et al., “Dispersion-tolerant DDO-OFDM system and simplified Opt. Fiber Commun. Conf. Exhib., 2001, Paper TuU6-1.
adaptive modulation scheme using CAZAC precoding,” J. Lightw. Tech- [48] C. C. Wei, “Small-signal analysis of OOFDM signal transmission with
nol., vol. 34, no. 11, pp. 2743–2751, Jun. 2016. directly modulated laser and direct detection,” Opt. Lett., vol. 36, no. 2,
[25] M. Chen et al., “Experimental demonstration of an IFFT/FFT size effi- pp. 151–153, 2011.
cient DFT-spread OFDM for short reach optical transmission systems,” J.
Lightw. Technol., vol. 34, no. 9, pp. 2100–2105, May 2016.