Copyright (c) 2013 IEEE. Personal use is permitted.

For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

1

AbstractThe use of incoherent multilevel modulation formats
with high spectral efficiency (more than two bits per symbol) has
been proposed in order to enable the next generation of very
high-speed Time-Division Multiplexing Passive Optical Networks
(TDM-PONs). Incoherent multilevel modulation is attractive for
access applications since multilevel formats allow the scaling of
the bit rate with electronic and photonic components operating at
a fraction of the bit rate. On the other hand, incoherent detection
reduces the requirement for complicated Digital Signal
Processing (DSP) and crucially an additional local oscillator,
compared to coherent receivers. The modulation formats
examined are Differential 8 Phase-Shift Keying (D8PSK) and
three versions of incoherent 16 Quadrature Amplitude
Modulation (QAM), specifically Star 16QAM, coded square
16QAM and 16QAM with pre-integration. Generation and
detection of these formats is discussed, as well as the
implementation challenges associated with such advanced
modulation formats. The performance of these modulation
formats was evaluated through extensive simulation and
experimental work. Results indicate that incoherent modulation
can fulfill important requirements of networks operators, namely
increased bit rate and increased splitting ratio, and provide a
cost-effective solution for Next-Generation Optical Access
Networks.

I ndex TermsPassive optical network (PON), time-division
multiple access (TDMA), multilevel modulation, differential
detection.

I. INTRODUCTION
HE explosive growth of Internet-based services, like file
sharing, social networking, cloud computing and Internet
video have fueled the demand for broadband access among
home and business users. Forecasts indicate that these trends
will only continue, with Internet video traffic rising at around
50% per year and file sharing at around 23% per year [1].
Gigabit passive optical networks have been widely deployed to
satisfy this demand and 10 Gb/s versions have already been
standardized ([2], [3]). However, even this increased capacity

Manuscript received September 25, 2012. This work was supported by the
Netherlands Ministry of Economic Affairs and the Netherlands Ministry of
Education, Science and Culture through MEMPHIS Project.
N. Sotiropoulos, A. M. J. Koonen and H. de Waardt are with COBRA
Research Institute, Eindhoven University of Technology, Eindhoven, 5600
MB NL, The Netherlands (phone: +31402472091, e-mail:
n.sotiropoulos@tue.nl).
Copyright (c) 2013 IEEE.
might not be enough in the coming years [4]. In
addition, new requirements from network operators, such as
increased splitting ratio (more users served per PON) and
longer reach, to enable a reduction in the number of Central
Offices (CO) through node consolidation, and lower power
consumption, have emerged [5]. At the same time, maintaining
the cost-efficiency afforded by earlier PON installations is of
paramount importance.
To meet these requirements and enable cost-efficient high-
speed broadband access networks, several network
architectures and modulation formats have been proposed and
investigated. Wavelength-Division Multiplexing (WDM)
PONs have been extensively researched, since they offer
logical point-to-point connections, utilize low-speed
components and remove the need for bandwidth assignment
protocols. On the downside, typically colored components are
required in the Optical Network Unit (ONU) and port densities
in the CO can be quite high. A number of solutions have been
put forward to remove the need for colored components.
Firstly, a tunable laser can be used in the ONU [6], but this is
not a low-cost option for the time being. A second option is to
provide unmodulated carriers from the CO [7] or remodulate
the downstream signal [8]. In this option, Reflective
Semiconductor Amplifiers (RSOAs) are typically used.
However, the reach, splitting ratio and bit rate are limited due
to Rayleigh back-scattering or residual modulation [9].
Furthermore, scaling the PON size can be problematic, due to
the limited number of available wavelengths. Another
approach utilizes coherent detection, which enables frequency
selectivity by means of a tunable local oscillator and achieves
superior sensitivity, making long-reach PONs without reach
extenders possible [10]. Additionally, ultra-dense wavelength
spacing is possible, due to the wavelength-selective nature of a
coherent receiver with a tunable local oscillator. While such a
scheme has very good performance, coherent receivers with
tunable External Cavity Lasers (ECL), 90
0
hybrids, balanced
detectors, Analog-to-Digital Converters (ADCs) and extensive
DSP are currently not cost-efficient for access applications.
Hybrid configurations, also known as WDM-TDM-PONs,
stack a number of wavelengths, each of which serves a
separate TDM-PON, typically at 10 Gb/s. This allows very
high PON splitting ratios [9] without increasing the
distribution loss linearly with the number of users. In addition,
10G transceivers specified in the PON standards can be
Advanced Differential Modulation Formats for
Optical Access Networks
Nikolaos Sotiropoulos, Graduate Student Member, IEEE, Ton Koonen, Fellow, IEEE, and Huug de
Waardt, Member, IEEE
T
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

2
utilized. On the other hand, the problem of wavelength
selectivity in the ONU is as apparent as in WDM-PONs.
An alternative to moving to WDM-based architectures is to
increase the line rate of the TDM-PON to higher than 10 Gb/s,
typically 40 Gb/s. The most straightforward way would be to
employ 40 Gb/s On-Off Keying (OOK), but there are
significant technical and cost challenges (need for 40 Gb/s
burst-mode receiver, expensive 40 GHz electronics and
photonics, receiver sensitivity issues) that make it unfeasible
for access. One solution is to provide 40 Gb/s through four
multiplexed 10 Gb/s OOK signals in different wavelengths
[11], but the challenge of achieving a colorless ONU and the
inefficient use of spectrum are significant drawbacks. A
migration to multilevel modulation formats can remove the
need for wavelength multiplexing, since high bit rates can be
achieved while keeping the symbol rate at 10 GHz. A very
popular approach to multilevel modulation is multi-carrier
formats, such as Orthogonal Frequency Division Multiplexing
(OFDM). 40 Gb/s OFDM-PONs have been presented, with
both coherent [12] and direct-detection receivers [13]. For the
coherent receiver case, very good receiver sensitivity is
achieved, allowing high splitting ratios and long reach. As in
the case of coherent WDM-PONs, the coherent receiver
introduces considerable complexity in the ONU, which is
undesirable. If direct-detection receivers are used, complexity
is significantly reduced (in the optical domain), but receiver
sensitivity suffers and splitting ratio and reach are severely
limited.
A different way to take advantage of the spectral efficiency
of multilevel modulation is to utilize single-carrier, optical
multilevel formats with incoherent (differential) detection. The
differential receiver is similar to the coherent one, but the
phase reference is provided by a delayed version of the signal
itself and not by a laser source acting as a local oscillator. This
also removes the need for carrier and phase recovery, greatly
simplifying the required DSP. For certain formats and receiver
architectures, DSP may be even fully discarded. The result is
a less complex receiver, which however can provide better
sensitivity than direct detection with a single photodiode at
high bit rates. Where symmetric bit rates are required, a
multilevel transmitter located in the ONU can provide the
necessary bandwidth. Complexity can be reduced for the
upstream channel by employing OOK modulation. A Directly
Modulated Laser (DML) on a different wavelength than the
downstream channel can then provide moderately asymmetric
bit rate (e.g. 1:4 of the downstream bit rate) cost-efficiently.
Optical multilevel modulation formats with differential
detection have been extensively investigated, as a means to
increase spectral efficiency in core and long haul networks. In
that respect, the superior sensitivity of coherent detection,
along with the possibility of mitigating impairments such as
Chromatic Dispersion (CD) and performing polarization
demultiplexing in the electronic domain- since the optical field
is linearly mapped in the detected electrical field- has led to
more deployments for core networks. Nevertheless, for access
networks, where requirements are different, incoherent formats
can be an attractive option. The first differential format that
attracted attention was Differential Phase-Shift Keying
(DPSK), where information is encoded in the phase difference
between two consecutive symbols. Most popular
implementations (with higher than a bit per symbol) involve
Quadrature DPSK (DQPSK) [14] and D8PSK [15], encoding
two and three bits per symbol, respectively. DSPK has
attractive OSNR requirements and very good tolerance against
nonlinearities [16], since it is a constant-amplitude format,
for 8 M s , but for higher phase levels the distance between
symbols becomes so small that the trade-off between OSNR
requirements and spectral efficiency becomes unfavorable.
To further increase spectral efficiency without very high
OSNR requirements, an amplitude component becomes
necessary. There are several configurations of combined
amplitude-differential phase formats. A straightforward way is
to superpose Amplitude-Shift Keying (ASK) modulation over
the DPSK signal. For example, if a binary ASK signal is
superposed over a D8PSK signal, 16 Amplitude Differential
Phase-Shift Keying (ADPSK), also known as Star 16QAM,
signal is created, encoding four bits per symbol [17].
Similarly, a Quaternary ASK signal can be superposed to a
DQPSK signal, also achieving four bits per symbol [18]. With
this method, the amplitude and phase encoded bits can be
encoded in the transmitter and decoded independently in the
receiver. Depending on the transmitter and receiver structures
used, DSP may not be necessary. If DSP-based transceivers
are used, alternative 16-level formats are also possible, with
arbitrary constellation diagrams. One approach uses a
technique called phase pre-integration to encode the
information in the amplitude and phase of the signal, which
will result in a square 16QAM constellation diagram after
detection and some (simple) DSP [19]. Alternatively,
information can be encoded in such a way that transmitting
square 16QAM constellation can enable correct decoding of
the detected signals after a conventional differential receiver
[20].
The paper is organized as follows. In Section II, the chosen
modulation formats are described and suitable transmitter and
receiver structures are discussed. In Section III, the proposed
architecture of the PON is explained and the required
subsystems are analyzed. Modeling of phase pre-integration
and coded square QAM in VPI is shown in Section IV, results
are presented and implementation challenges are discussed. In
Section V, experiments with bidirectional transmission of
D8PSK and Star 16QAM in a PON are described and their
results are commented. Finally, on Section VI some
conclusions are offered.
II. INCOHERENT MULTILEVEL MODULATION FORMATS
A. D8SPK and Star 16QAM
D8PSK encodes three bits in the phase difference between
two consecutive symbols, which belong to the
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

3
set { / 8, 3 / 8,...(2 1) / 8} ,
j
M e
|
| t t t e , where M=8 for
this case. The resulting constellation diagram is shown in
Fig.1a. The encoding procedure can be implemented either in
DSP, if DSP-based transceivers are used, or by a discrete
differential coder, if transceivers are driven by binary signals.
The encoding amounts (functionally) to multiplying the
symbol at moment k,
k
a , with the transmitted symbol at
moment k-1,
1 k
s

:
1 k k k
s a s

= . For Star 16QAM, a fourth
independent bit adds amplitude modulation, creating two rings
of D8PSK symbols, as it can be seen in the constellation
diagram in Fig. 1b. This fourth bit does not need to be encoded
and can be detected independently by an intensity detection
branch of the incoherent receiver, in addition to the D8PSK
detection part that remains unchanged.
As previously mentioned, D8PSK/Star 16QAM signals can
be generated either by binary signal-driven transmitters or by
DSP-based transmitters (which utilize Digital-to-Analog
Converters). In the first case, complexity is kept to a minimum
in the electrical domain (only binary signals are required), but
the optical part is complicated, requiring one modulator per
bit. There are two configurations possible for binary-driven
transmitters: the serial and the parallel configuration [16]. In
the serial configuration, shown in Fig. 2a, after the data are
deserialized and encoded, they drive three phase modulators
cascaded in series, creating an 8-level phase modulated signal.
The first phase modulator, responsible for creating the binary
DPS signal (0 and shift), can be replaced by an amplitude
modulator. For Star 16QAM, an amplitude modulator is
added, also in series. In the parallel configuration, the two first
phase modulators of the serial transmitter are substituted by a
nested Mach-Zehnder modulator structure with a /2 phase
shift in the lower optical path (shown in Fig. 2b), known as an
IQ modulator. Whilst the serial configuration is easier to
control, since no DC biases are required (for pure phase
modulation), the parallel configuration is preferable, because it
is more immune to the amplitude noise of the driving signals.
In a phase modulator, the amplitude ripples of the driving
signal translate into phase ripples in the optical signal, whereas
in an amplitude modulator (such as a Mach-Zehnder
modulator) this is not the case [21].
The DSP-based transmitter is shown in Fig. 2c. The optical
part consists of an IQ modulator, as the one described before.
The driving signals for the IQ modulator are no longer binary,
but multilevel. A D8PSK signal requires four-level electrical
signals, while Star 16QAM requires eight-level signals. This
transmitter configuration functions as a generic transmitter,
since it can generate arbitrary optical signals, depending on the
driving signals. In this scheme, complexity is moved on the
electrical domain. The necessary driving signals are created in
the DSP and are transferred to the analog domain through
DACs. This allows the integration of functions in the DSP
part, such as signal pre-distortion to account for the nonlinear
characteristic of the MZ modulators and CD pre-
compensation. In addition, flexible transmitters are made
possible, changing the modulation format through software
control, according to traffic requirements. For example, when
bandwidth demand is low (e.g. at night), the modulation
format can be switched from Star 16QAM to D8PSK,
decreasing energy consumption on the ONU by shutting down
the intensity detection part of the receiver.
As with the transmitters, a number of different
configurations exist for a multilevel differential receiver.
Again, the trade-off is between complexity in the optical or the
electrical domain. For binary detection schemes, M-DPSK
formats require M/2 thresholds [22]. Translated to optical
components, this requirement means that, for D8PSK, four
Mach-Zehnder Delay Interferometers (MZDIs) with phase
shifts of{ / 8, 3 / 8, 5 / 8, 7 / 8} t t t t and delay equal to one
symbol period, each followed by Balanced Photo Detector
(BPD) pairs are needed. Such a receiver is shown in Fig. 3a.
Although this receiver allows for a very simple decoding
scheme, it is very complex, power-inefficient and difficult to
control, mainly because of the four phase shifters. The optical
part of the receiver can be reduced to two MZDIs and BPDs,
shown in Fig. 3b, if multilevel thresholding or analog electrical
post-processing is employed [22], [23]. To accommodate a
Star 16QAM signal, the receiver must include an intensity
detection branch, which consists of a single photodiode, where
part of the input signal is coupled. The beating of the two
different intensity levels of the signal in the MZDIs, however,
makes the decoding of the phase levels with multilevel
thresholding difficult in the case of QAM signals.
A generic, DSP-based differential receiver that can detect
arbitrary differential multilevel formats is possible with some
slight modifications of the receiver of Fig. 3b. The optical part
of the receiver remains essentially the same, with the phase
shifts being set at 0 and 90
0
on the upper and lower MZDI,
respectively. The multilevel thresholding or the analog
electrical processing network is removed. Instead, the detected
electrical signals are sampled by ADCs and transferred to the
DSP part of the receiver. Given that the incoming optical
signal is ( ) E t , the four outputs of the two MZDIs will be the
following, not taking noise into account:


1
2
3
4
( ) ( ) ( )
( ) ( ) ( )
( ) ( ) ( )
( ) ( ) ( )
E t E t E t T
E t E t E t T
E t E t jE t T
E t E t jE t T
= +
=
= +
=
(1)

The balanced detection process will create two currents, ,
I Q
I I
(for the in-phase and quadrature component, respectively),
which will be:

2 2
1 2
2 2
3 4
( ) | ( ) | | ( ) |
( ) | ( ) | | ( ) |
I
Q
I t E t E t
I t E t E t
=
=
(2)

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

4
After expanding the right-hand terms in (2), the resulting
signal can be expressed as:


( ) Re{ ( ) ( )} | ( ) | | ( ) | cos
( ) Im{ ( ) ( )} | ( ) | | ( ) | sin ,
I
Q
I t E t E t T E t E t T
I t E t E t T E t E t T
|
|
-
-
= = A
= = A
(3)

where | A is the phase difference between the two consecutive
symbols. Equation (3) shows that the resulting signal maps the
real and imaginary part of the differential optical field,
allowing the reconstruction of the original information symbol.
That way, arbitrary constellation diagrams can be successfully
detected and decoded. An equivalent version of (3), with the
discrete notation used in the explanation of the encoding at the
transmitter side, is:

, 1 1 1
, 1 1 1
Re{ } Re{( ) } Re{ }
Im{ } Im{( ) } Im{ },
k
k
I k k k k k k
Q k k k k k k
I s s a s s a
I s s a s s a
- -

- -

= = =
= = =
(4)

from which it can be seen that the encoding is cancelled out
and information is recovered. An alternative but
mathematically identical implementation of the generic
receiver involves a 2x4 90
0
hybrid, with a delay line in one of
its inputs. This implementation is more suitable for receivers
based on optical integration, since it leads to smaller devices
and does not use controllable phase shifters.
B. 16QAM with phase pre-integration
D8PSK and Star 16QAM are the simplest formats possible
for signals of high spectral efficiency, in terms of signal
generation and detection. A drawback, however, of these
modulation formats is that they are not optimal in the sense of
minimum Euclidean distance between the symbols in the
signal space. A more optimal packing of the symbols is
achieved with square QAM constellations, which fall short 0.5
dB from the optimum arrangement [24]. The problem with
square QAM constellations is that they dont lend themselves
easily to differential detection, because while the symbols in
the inner and outer rings are evenly spaced (with /4 spacing),
as it can be seen in Fig. 4a, the symbols in the intermediate
ring are asymmetric in phase. That means that the
reconstructed differential signal, after the differential receiver,
will not be a 16QAM signal. To enable differential detection
of square 16QAM, the transmitted and received signals must
be suitably processed. Such a scheme has been first proposed
in [25], in the context of wireless communications, and it has
been successfully implemented in optical communications in
[19]. The scheme is based on a technique called phase pre-
integration, which amounts to adding the phase component of
the previous transmitted symbol to the phase of the current
one, while keeping its amplitude unchanged:


1
| | exp( [arg( ) arg( )])
k k k k
s a j a s

= + (5)

The resulting signal will have three magnitude levels, like
square 16QAM, but it will have almost continuous phase
distribution, due to the accumulation of phase after every
symbol (Fig. 4b). In the receiver, the outputs of the BPDs are
given by (3). After sampling and digitizing, the phase
difference is estimated by computing the inverse tangent, with
the two detected signals as arguments. The phase difference
amounts to the phase of the original symbol
k
a , before the
phase pre-integration operation in (5). A separate photodiode
produces an estimate of the amplitude of the signal. The two
components (phase and amplitude) are combined and the
original square 16QAM constellation can be recovered. Due to
the practically analog nature of the electrical driving and
detected signals and the DSP required, only the generic, DSP-
based transmitter and receiver described in the previous
section can be used.
C. Coded square 16QAM
An alternative way of preserving the square 16QAM
constellation is to encode the transmitted signal in such a way
that the detector can unambiguously recover the original
symbol from the outputs of the BPDs. This method requires a
group of complex integers and a function F that ensures that
the complex multiplication of any two symbols belonging to
the group will be part of the group. This property of the
function means that differential encoding and detection, two
operations that involve complex multiplications, result in
QAM symbols. F is a congruence operation that reduces the
enlarged signal set that is the result of the multiplication of two
QAM symbols into the original QAM set. This operation is
essentially a Look-Up Table (LUT), where the product of the
complex multiplication of two QAM symbols is
unambiguously mapped into a third QAM symbol. It has been
shown that a rotated by 45
0
16QAM signal set is such a group
and a function as the one described exists for this set [20].
Note that the rotation of the constellation is necessary only
during the signal processing. For transmission, the
constellation can be rotated back to the conventional, since a
phase shift common in consecutive symbols is cancelled out by
the differential detection process. The encoding in the
transmitter can be expressed as:


1
1
[ ( ) ]
k k k
s F a s
-

= (6)

In the receiver, the combination of the real and imaginary
components, ,
I Q
I I gives, as it has already been shown in (4),
the complex multiplication
1 k k
s s
-

. If this product is input to

the mapping function, it follows by its properties [20] that:


1
1 1 1
1
1 1
[ ] [ [ ( ) ] ]
[ [ [( ) ]] [ ] [ ]
k k k k k
k k k k k k
F s s F F a s s
F F a s s F a F a a c
- - -

- -

= =
= = =
(7)

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

5
Thus, the original information symbol can be recovered
from the differential constellation diagram (shown in Fig. 4c).
Note that in coded square 16QAM, a separate intensity branch
is not necessary for the detection of the signal; the differential
metrics of (4) contain sufficient information to recover the
QAM symbol. On the other hand, a DSP-based receiver is
necessary, due to the need to implement a Look-Up Table
(LUT) to map the received differential signal set to the square
QAM signal set. For the generation of the square 16QAM
signal, any of the many configurations proposed in the
literature can be used. An extensive review can be found in
[26]. The most straightforward solution, and the one typically
used in long haul, high-order QAM with coherent detection
experiments, is the IQ modulator. Four-level driving signals
are required in this case. Multilevel signals can be created
directly by DSP, where also the encoding takes place.
Alternatively, four binary signals can be encoded by a discrete
differential coder, as in D8PSK, and then can be electrically
combined to create the two four-level signals that will drive
the IQ modulator. This method of creating multilevel electrical
signals, shown in Fig. 5, is widely used in high-speed QAM
experiments [27].
III. PON ARCHITECTURE AND SUBSYSTEMS
The motivation of employing differential multilevel
modulation formats in access is to enable TDM-PONs with
higher bit rate and splitting ratio, as stated in the Introduction.
For that reason, the proposed network architecture is based on
the typical TDM-PON configuration, seen in Fig. 6. In the CO,
or Optical Line Terminal (OLT), as it is known in the PON
nomenclature, the multilevel transmitter that generates the
downstream signal and the receiver of the upstream channel
are located. The two signals are separated by a circulator or a
1550/130 nm WDM diplexer, depending on the wavelength
allocation scheme. Furthermore amplifiers may be employed,
as a booster for the downstream signal and as pre-amplifiers
for the upstream. The multilevel transmitter is more complex
than transmitters used in current PONs, but since it is a shared
resource among all the users in the PON, it is not so critical for
the overall cost-efficiency of the network. The DSP-based
configuration of Fig. 2c would be the optimal solution for the
transmitter. This configuration is generic and can
accommodate all discussed modulation formats and also
compensate for chromatic dispersion, if it is required.
Switching between formats of different spectral efficiency is
also possible, as discussed before, adding more flexibility in
the PON.
In the Remote Node (RN), a passive 1xN coupler splits the
downstream signal to all the PON users and combines the
time-multiplexed upstream signals that will reach the OLT.
Amplifiers, in particular Erbium-Doped Fiber Amplifiers
(EDFAs), can also be located in the RN, if necessary.
Traditionally, amplifiers are avoided in PONs, since they bring
up costs. However, this begins to change, with EDFAs
employed as boosters for the RF signal overlay in the XG-
PON standard [28]. As the trend towards long-reach, high-
splitting ratio PONs continues [9], [11], the utility of EDFAs
as reach extenders is becoming apparent. With splitting ratios
typically equal or higher than 256 users, the cost of EDFAs is
spread among much more users than in legacy TDM-PON,
which accommodate typically 32 or 64 users. The use of
EDFAs can be additionally justified by the cost efficiency
achieved by reducing the number of COs, made possible by
such long-reach, high splitting ratio PONs.
In the ONU, the downstream and upstream channels are
again separated by a circulator or a diplexer, as in the CO. The
ONU is the most cost-sensitive part of the network, since there
is no cost sharing involved. For the upstream channel, there
are two options: symmetric bit rate, which means mirroring the
transmitter of CO in the ONU, and asymmetric bit rate, which
utilizes simpler and more cost-efficient components. A
multilevel transmitter in the ONU is not feasible in the near
future, but it may be an attractive option when traffic demands
reach the point where very high-speed upstream channels are
necessary. In such a case, the required low-linewidth sources
for the transmitter can be located in the ONU or can be
centrally provided by the OLT. In the low-complexity,
asymmetric option, the upstream channel is implemented by a
DML operating at a different wavelength. The DML is a
simple and cost-efficient solution that can provide bit rate up
to 10 Gb/s and is compatible with existing PON standards.
The channel bit rate is set to one fourth or one third of the
downstream bit rate, which is moderately asymmetric, as is the
case historically with most broadband technologies. In
addition, note that emerging media services require extremely
high downstream bit rates e.g. IPTV, High Definition TV
(HDTV), Ultra-HDTV and 3DTV variants, hence symmetric
bit rates are not required for the upstream. The differential
receiver in itself is, however, quite complex, as described in
the previous section. Photonic integration is instrumental in
that respect. Integrated differential receivers on silicon have
already been demonstrated with very good results [29]. Since
access networks are a mass-volume market, it is conceivable
that such devices can become cost-efficient, if significant
market adoption is achieved.
As with the transmitter, the versatility of the DSP-based
configuration is preferable also for the receiver. This
configuration allows the evaluation of all proposed modulation
formats, while minimizing the optical front-end. An important
issue for the DSP-based receivers is the sensitivity, power
consumption of the DSP chipsets and ultimately cost. Due to
extensive efforts for the development of coherent, QPSK-
POLMUX based 100 Gb/s transponders for long-haul
applications, the state-of-the-art commercial ADCs have
sample rates in excess of 50 GSa/s [30]. For access
application, a sampling rate of 20 GSa/s (up sampling by a
factor of two for 10 GHz signals) is well within current
capabilities. High-speed ADCs are increasingly being
demonstrated with CMOS technology allowing single-chip
ADC and DSP solutions [31]. This configuration can enable
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

6
lower power consumption, better yield compared to other
approaches, e.g. SiGe ADCs and Si DSP, and of course the
proven cost-efficiency benefits of CMOS technology [32].
CMOS Application-Specific Integrated Circuits (ASICs) costs
are volume-driven, when based on mature technologies. Again,
the relatively low speed of access applications (compared to
long haul) means no cutting-edge designs and exotic materials
are required and the potentially large volumes suggest that
ASICs for PONs fit well into this pricing model. An important
factor that determines the development effort required for the
ASIC, as well as its power consumption, is the complexity of
the signal-processing algorithms [33]. In that respect,
incoherent detection has a clear advantage over coherent
detection and OFDM. All DSP-based receivers share some
common functions, such as de-skew, timing recovery,
normalization and symbol estimation and decoding [34].
Besides these common elements, coherent receivers require
frequency and phase recovery. CD compensation is also
typically implemented. For OFDM receivers, the more
computationally intensive task is the Fast Fourier Transform
(FFT). Other tasks involve removal of the Cyclic Prefix and
frequency domain equalization. Apart from the common
algorithms, incoherent receivers for D8PSK and Star 16QAM
do not require any further DSP. The QAM with phase pre-
integration requires-besides the common algorithms described
above- the computation of the inverse tangent and the
reconstruction of the amplitude and phase component of the
signal. For the coded square QAM, an implementation of a
LUT that maps the differential signal set to the original square
QAM set is the only extra functionality needed. It is clear that
the development effort and the power consumption, which is a
significant part of operational costs, of the incoherent receivers
can be kept at significantly lower levels compared to the other
multilevel approaches.
The proposed TDM-PON architecture can be combined
with WDM techniques (as the hybrid WDM-TDM PON
proposed in [9]) to further scale the PON size and facilitate the
convergence of metro and access networks. The wavelength
stacking of independent TDM-PONs has the further advantage
of enabling cost sharing of components (such as optical
amplifiers) among an even larger number of users.
Additionally, the increased bit rate of the multilevel TDM-
PON enables significant increase on the sustainable bit rate per
user, that for 10 Gb/s PONs with splitting ratios higher than
128 users suffers.
IV. MODELING AND SIMULATION RESULTS
To evaluate the more sophisticated multilevel formats
presented, 16QAM with pre-integration and coded square
16QAM, a TDM-PON with the respective transceivers was
modeled in VPI Optical Systems simulation software. With
this approach, the performance of the modulation formats
could be examined, since analytical formulation of the error
probabilities is difficult, due to the unconventional
constellation diagrams and nonlinear operations in the
receiver.
1

A. Symmetrical 10 Gb/s 16QAM with phase pre-integration
simulations
The PON model used in the simulations is shown in Fig. 7
[35]. A symmetric configuration, with 10 Gb/s 16QAM signals
for both downstream and upstream channels and identical
transceivers in the CO and ONUs, is chosen. The low-
linewidth (200 kHz, as for the downstream signal) Continuous
Wave (CW) light for the upstream channel is generated
centrally in the CO and is distributed to the ONUs, spaced 3.2
nm away from the downstream wavelength at 1550 nm. This
removes the need for an expensive ECL in every ONU,
utilizing a single source for the entire PON. WDM
(de)multiplexers are included in the model, to accommodate
the possibility of stacking a number of TDM-PONs on
different wavelength pairs. It was not possible to include,
however, these WDM channels in the simulation, due to
computational constraints of bidirectional propagation over the
fiber model in VPI. A splitting ratio of 32 users is assumed in
the model. The amplification scheme of the PON is based on
Semiconductor Optical Amplifiers (SOAs). The SOAs are
located in the ONU, where they amplify the downstream
signals, as well as the CW carrier, and in the CO, where they
are used as pre-amplifiers for the upstream signal. Optical
amplification is generally not preferable in the ONU, but
SOAs can be integrated in the receiver design and thus do not
introduce unreasonable complexity.
Regarding the effects and impairments modeled, all optical
impairments are included in the simulations. The generation
and sampling of the multilevel electrical signals is noise and
jitter free, to keep focus in the Amplifier Spontaneous
Emission (ASE) noise-limited regime. The BER versus OSNR
curves have been obtained experimentally in [19], but in
access networks receiver sensitivity is a more important
metric, since it determines the power budget of the network
and therefore the maximum splitting ratio and reach.
The first aspect of the system design that was investigated
was the optimal power of the unmodulated carrier. As the
upstream signal is more sensitive than the downstream, the
carrier power was optimized for the upstream transmission.
With a constant downstream input power of -8.6 dBm at the
fiber, resulting to a BER for the downstream channel of around
10
-11
, the upstream BER as a function of the carrier input
power is shown in Fig. 8. The increase of the carrier power
leads to improved BER for the upstream signal, until the
nonlinear effects start to degrade the signal at around 1.5 dBm
input power. The lowest BER occurs for 0 dBm input power
and is 5 10
-7
, without Forward Error Correcting (FEC)
codes.
Using the above specified values, the BER as a function of
the received power was evaluated. Since the transmitter and
the receiver were custom-made (not part of the VPI library), it
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

7
was not possible to use the BER estimation tools of VPI and
error counting leads to prohibitive simulation times. The
approach followed was to use the Error Vector Magnitude
(EVM) of the received signal as an estimation of its SNR, and
use the SNR to obtain a BER value assuming a Gaussian
distribution of the noise. In reality, the noise is not Gaussian,
as it can be observed in the received constellation diagrams
(Fig. 9), but it affects more intensely the phase of the received
signal. This is a common characteristic of all systems
employing both interferometric and intensity detection. The
differential phase component is corrupted by noise from two
observation instances, doubling its noise variance, while the
amplitude component is affected by noise from a single
instance. This non-Gaussian noise distribution leads to an
underestimation of the error probability, if conventional,
Euclidean boundary metrics are implemented. However, if
non-Euclidean metrics for symbol detection are used, taking
into account the increased noise in the radial (phase) direction,
improved BER performance is achieved [19], justifying the
approximation.
The BER curves for back-to-back and 20 km of fiber are
shown in Fig. 10. The power penalty for the downstream
signal is negligible. For the upstream signal, however, there is
larger power penalty. It is mainly caused by the nonlinear
interactions in the fiber, with SPM and XPM being the most
important, as the signal is phase-modulated. Further
simulations, with the nonlinear effects turned off, showed that
the power penalty becomes negligible in that case. Besides the
penalty because of the fiber transmission, a significant
degradation of the performance for the upstream signal is
observed, compared to the downstream. The reason for this
degradation is the second amplifier that lies in the upstream
path. When the noise from the preamplifier at the CO is turned
off, the BER of the upstream signal is comparable to that of
the downstream. However, even with this sensitivity
degradation, error-free operation for the upstream signal can
be achieved, as the BER can be well below the FEC limit,
which is around 10
-3
, for an overhead of 7%.
B. Symmetrical 10 and 40 Gb/s coded square 16QAM
simulations
The PON model used in the simulations with the coded
square 16QAM can be seen in Fig. 11 [36]. The PON offers
symmetric bandwidth, with simulations done for bit rates of 10
and 40 Gb/s, over 20 km of SSMF. The CW carrier is again
provided by the CO, to keep the ONU source-free. The main
difference with the network architecture of the pre-integration
QAM simulations is that optical amplification is moved to the
RN and the (de)multiplexers (for WDM stacking) are not
included. Two EDFAs, located before the splitter, amplify the
downstream channel and the unmodulated carrier, and the
upstream channel, respectively. Another EDFA is used as a
pre-amplifier in the CO. The transfer of optical amplification

1
A theoretical calculation of the BER performance of phase pre-
integration QAM was carried out in [25], but a coherent receiver was
assumed and the phase of each symbol was estimated individually.
in the RN further simplifies the ONU hardware and shares the
cost among all the users, since the amplifiers are placed before
the splitter. Since there is no need for integration in this
amplification scheme, EDFAs, which have a lower noise figure
than SOAs, are used.
A more realistic modeling of the driving signals is
implemented, including noise and filtering. The Bit Error Rate
(BER) curves as a function of the OSNR for 10 and 40 Gb/s
downstream and upstream transmission are shown in Fig. 12
and 13. For the upstream case, coherent detection was also
included in the simulations, for comparison reasons. Carrier
and phase recovery was implemented in the coherent receiver,
using an algorithm provided in VPI, but no CD compensation
was performed. In the CO, the receiver is shared by a number
of users (32 in this case), so a more expensive solution can be
justified. The linewidth of the lasers involved was 100 kHz
and fiber input powers were 0 dBm for the carrier and -4 dBm
for the downstream signal. The received power was -6 dBm
(right before the receiver) for both channels. The results
indicate that for the downstream channel, error-free (with
FEC) transmission can be achieved for both 10 and 40 Gb/s. In
the upstream channel, error-free incoherent detection is
possible for 10 Gb/s, but for 40 Gb/s there could be no error-
free operation within the OSNR limits of the amplification
scheme. If coherent detection is used, error-free operation is
possible for 10 and 40 Gb/s. For incoherent receivers, BER
was obtained by direct error-counting. For the coherent
receivers, BER was estimated using VPIs built-in tools.
C. Simulation results discussion and implementation
challenges
Simulations showed that the DSP-based, square QAM
modulation formats can be attractive options for future TDM-
PONs. 10 Gb/s symmetric operation over 20 km fiber and at
least 32 users is shown for both formats. For the 40 Gb/s
coded 16QAM PON, error-free downstream transmission is
possible. For the upstream channel, coherent detection can
provide the required sensitivity or a lower bit rate signal can
be used. For the QAM with pre-integration, significant margin
is available, which indicates that error-free operation at 40
Gb/s with FEC and a high splitting ratio are possible.
Regarding OSNR requirements, the pre-integration technique
offers superior performance [19], but on the other hand, the
simpler receiver and processing of the coded QAM can be
advantageous in access applications. To enable comparison of
the proposed advanced modulation formats with conventional
OOK, the receiver sensitivities for 10 Gb/s OOK that are
specified in the relevant standards, along with the
sensitivities obtained through simulations and
experimental work are shown in Table 1. Comparing the
obtained results with OOK-based systems, for the 10 Gb/s
phase pre-integration 16QAM it can be seen that considerably
higher splitting ratio and longer reach can be achieved, due to
the formats high sensitivity, while operating at one fourth of
the bit rate. An implementation power penalty, for reasons
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

8
extensively discussed in the following paragraphs, will reduce
the receiver sensitivity, but it can be compensated for by an
increased input power to maintain the high power budget. For
the coded square QAM, receiver sensitivity and as a result
power budget has not been estimated, but due to smaller
symbol spacing it is expected to be lower than the phase pre-
integration results.
The main motivation for the simulation work on these
advanced modulation formats was to identify the theoretical
limits of the sensitivity of the incoherent receivers, assuming
perfect driving signals, modulation and detection processes.
For an experimental implementation of the systems explored in
simulations, a number of non-idealities occur. The first issue is
the generation of the multilevel signals. For the signals
required for these two formats, the only option for laboratory
demonstrations is a high-speed Arbitrary Waveform Generator
(AWG). However, the signal quality of the AWG is not (yet)
comparable with the quality achieved in Binary Pattern
Generators (BPGs). A four-level signal (that generates the I or
Q component of a square 16QAM signal) from the available
AWG, at 2.5 GHz, is shown in Fig. 14. There are considerable
ringing oscillations in the signal. Taking into account that for
the phase pre-integration technique, the driving signals have
tens of levels, it is obvious that there is significant degradation
compared to the ideal case in simulation work. A further
related problem is the linearity of the electrical amplifiers,
both of the drivers in the transmitter side and of the Trans-
Impedance Amplifiers (TIAs) and pre-amplifiers in the
receiver side. As the dominant modulation format in high-
speed optical communication systems so far has been OOK,
amplifiers are optimized for binary signals. The introduction of
commercial (D)QPSK transceivers has not changed the
situation, since they also operate with binary signals. This
means that the amplifiers are another source of distortion for
multilevel signals. The introduction of coherent 16QAM
transceivers requires improvements in this respect and will
lead to the availability of suitable linear amplifiers. It should
be noted that these kinds of distortions are not easy to be
modeled in VPI, so their impact cannot be estimated before the
experimental validation.
The sensitivity of the advanced modulation formats under
discussion to deviations from the optimal operating points of
the transmitter and the receiver is also a source of potential
degradation in system performance. Deviation from the
optimal amplitude of the driving signals, the proper biasing of
the IQ modulator, the optimal phase shifts in the MZDIs and
imbalances in the receiver can affect signal quality and
receiver performance. The best signal quality is achieved when
the driving signals have amplitude of 2 V
t
, where V
t
is the
voltage required for an 180
0
phase shift in the optical signal.
For lower driving voltages, the signal is distorted. For
example, the simulated received constellation diagrams for the
coded 16QAM format for signals with amplitude at the
optimal value and at 2/3 of it are shown in Fig. 15 a and b
respectively, where the reduction of the inter-symbol distances
for the reduced amplitude case can be clearly observed. The
biasing of the two MZ modulators of the IQ superstructure in
the null transmission point and the biasing of the phase
shifter in the quadrature point are also critical. After initial
setting of the DC biases, environmental changes can lead to
drifting and consequently to signal degradation. Fig. 15c
shows the constellation diagram when both lower driving
voltage (2/3 of the optimal, as before) and a 10
0
deviation for
the phase shifter are simulated. For commercial
implementations, the operating points for the DC biases can be
maintained with control circuits using feedback from
photodiodes located in the modulator structures.
In the receiver side, the possible impairments include
deviations from the optimal values of the phase shifters in the
MZDIs (or the 90
0
hybrid, depending on the implementation),
deviations from the nominal delay, non-ideal extinction ratio
of the MZDI and amplitude and timing mismatch in the BPD.
It has been shown for DQPSK in [37] that the last four have
small impact on the achievable BER for reasonable deviations.
Similar conclusions have been reached for 16QAM with phase
pre-integration in [38]. On the other hand, the phase detuning
can be detrimental for the performance of the system, if left
uncompensated, because it results in a rotated constellation
diagram. It is possible, however, to recover the correct
alignment of the in-phase and quadrature components through
orthonormalization algorithms in the DSP, such as Gram-
Schmidt or Lwdin [34].
V. EXPERIMENTAL WORK AND RESULTS
For the reasons described extensively in the previous
section, experimental work on PONs with differential
multilevel formats was focused on D8PSK and Star 16QAM.
These two modulation formats put considerably less strain in
electrical components (AWG and amplifiers) when using
multilevel signals and can also be generated by binary driving
signals, trading complexity on the optical domain with simpler
electronics.
A. 10 Gb/s Star 16QAM/OOK PON
The experimental set-up of a TDM-PON based on a 10
Gb/s, DSP-based Star 16QAM downstream channel and 2.5
Gb/s OOK upstream channel is shown in Fig. 16 [39]. The
differential encoding of the downstream data, as well as the
pre-distortion to compensate for the IQ modulators non-linear
transfer function, is performed offline in Matlab and the
resulting waveforms are loaded to the AWG. The transmitted
sequence is 100k symbols long and the ECL source operates at
1550 nm, with a linewidth of 100 kHz. The created Star
16QAM signal is then launched from the Central Office (CO)
unamplified into 25 km of SSMF through a circulator, with an
input power of -4.5 dBm. The RN consists of two circulators
that separate the downstream and the upstream channel, two
EDFAs and an attenuator that introduces 15 dB attenuation,
modeling the loss of 1:32 splitter. The EDFA amplifies the
downstream signal from -12 dBm to +10 dBm. In the ONU, an
EDFA is used as a pre-amplifier to boost the received power to
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

9
7 dBm, followed by a filter to remove excess ASE noise. Pre-
amplification is necessary in this experiment, since the BPDs
used do not have integrated TIAs and their output voltage is
very low. BPDs with integrated TIAs are commercially
available, but typically with a limiting amplifier configuration.
This configuration is optimal for DPSK and DQSPK, which
produce binary signals, but is detrimental when multilevel
signals need to be amplified. A portion of the input power,
through a 70:30 splitter, is sent to a single photodiode for
amplitude detection. The electrical signals are amplified
(except for the amplitude component), sampled by a Digital
Phosphor Oscilloscope (DPO) at 50 GSa/s and transferred to
Matlab for offline processing.
The 2.5 Gb/s NRZ driving electrical signal for the upstream
channel is created by a 12.5 Gb/s Bit Pattern Generator (BPG).
It then modulates a DML in the ONU, which has an output
power of 0 dBm and operates at 1554 nm. The signal is
introduced through a circulator into the drop fiber and then
into the RN, where it is attenuated by the splitter and
forwarded into the upstream-path EDFA. The signal is then
amplified to +3 dBm (minimum output power of the EDFA)
and then attenuated, so it does not affect the downstream
signal. The EDFA in the upstream path was necessary in this
experiment, as the sensitivity of the PIN-TIA receiver used is -
23 dBm, while the total losses in the system are 27 dB.
However, if a better receiver (e.g. with an APD) or a
transmitter with higher output power is used, the EDFA in the
upstream channel may be removed.
Detected constellation diagrams of the Star 16QAM signal
for back to back and transmission over 25 km of fiber are
shown in Fig. 17. The 16 signal points are clearly separated
and the predominance of phase noise can be observed. The
differential phase component is corrupted by noise from two
observation instances (as it depends on the current as well as
the previous symbol), doubling its noise variance, while the
amplitude component is affected by noise from a single
instance. From the constellation, it is obvious that an important
parameter in the system design is the radius ratio between the
two D8PSK circles. Since phase noise is twice as large as
amplitude noise, the inner circle should have as large radius as
possible, to exploit the larger distance between its symbols,
which is proportional to its radius. However, if the two circles
are spaced too close to each other, amplitude errors begin to
dominate. To find the optimal operating conditions, the BER
of the Star 16QAM signal as function of the radius ratio has
been evaluated and the results are shown in Fig. 18. The
minimum BER is found for values between 0.65 and 0.75
(normalized to the outer ring radius of 1), so the inner ring
radius was chosen to be 0.7.
The BER of the Star 16QAM signal as a function of the
input power at the receiver was measured by tuning the optical
attenuator before the EDFA. The results can be seen in Fig.
19. Error-free transmission with FEC, introducing 7%
overhead, is possible. The sensitivity for the FEC threshold of
10
-3
is around -31 dBm, which indicates very good transmitted
signal quality and detection process. There is negligible power
penalty observed due to fiber transmission. The corresponding
BER curve for the upstream OOK channel is shown in Fig. 20.
The sensitivity of the signal at BER of 10
-9
is around -22 dBm.
Very small power penalty due to fiber transmission is observed
and the sensitivity measured is quite close to the reference
sensitivity of the particular receiver, indicating minimal
interference of the downstream channel.
B. 30 Gb/s D8PSK/OOK PON
For the very-high speed PON experiment [40], with a
symbol rate of 10 GSa/s, binary-signal generated D8PSK was
chosen. The available AWG has a bandwidth of around 6
GHz, so the DSP-based approach was not possible. Again, a
DML provides an OOK upstream signal, this time at 10 Gb/s
and at 1310 nm. The experimental set-up can be seen in Fig.
21. In the CO, a BPG creates the binary signals with bit pattern
of 2
13
-1. The signals are decorrelated through different cable
lengths, aligned and amplified, and then drive the D8PSK
serial modulator. An ECL set at 1550 nm functions as the low-
linewidth CW light source. The D8PSK signal passes through
4km of Dispersion Compensating Fiber (DCF) that
compensates for the dispersion over the transmission fiber and
is then amplified. After a 1550/1310 nm diplexer, the signal is
launched into 25 km of SSMF with an input power of around 0
dBm. The RN is entirely passive in this case. An attenuator
introduces 15 dB of splitting losses (that corresponds to a
splitting ratio of 32 users). In the ONU, the downstream signal
is separated through a diplexer and is pre-amplified, before
being detected and sampled by a Digital Phosphor
Oscilloscope (DPO), with a captured length of 100k symbols.
Optical pre-amplification was again necessary, as the same
BPDs without TIAs as in the Star 16AM experiment were
used. Detected constellation diagrams for back to back and
transmission over 25 km of fiber are shown in Fig. 22a and b.
No degradations from imperfect generation, detection
processes or fiber transmission are observed.
For the upstream channel, a 12.5 Gb/s BPG creates a 10
Gb/s NRZ signal with a pattern of 2
23
-1 that drives the DML.
The OOK upstream signal, after passing through the splitter
and 25 km of fiber, is pre-amplified in the CO by an SOA to
boost the signal before being detected by a PIN-TIA receiver,
which has a nominal sensitivity of -19 dBm. If a more sensitive
APD-TIA were available at the CO, the SOA would be
unnecessary. Detected OOK eyes for back to back and
transmission over 25 km of fiber are shown in Fig. 23a and b,
also exhibiting no sign of degradation.
The BER curves as a function of the received power for
bidirectional transmission of the downstream and upstream
channels can be seen in Fig. 24 and Fig. 25. For the D8PSK
signal, error-free transmission with FEC, introducing 7%
overhead, is possible. Observed sensitivity is -33.5 dBm and a
small power penalty less than 1 dB is measured. The results
validate the very good generation and detection of the
multilevel signal. For the OOK upstream channel, receiver
sensitivity is -24 dBm, with very low power penalty due to
fiber transmission and no error floor. If a commercial APD
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

10
receiver is used, improved receiver sensitivity can be achieved.
C. Discussion
Experimental results show that very good receiver sensitivity
is achieved for both D8PSK and Star 16QAM, for the pre-
amplified receiver. For the Star 16QAM experiment, it is
shown that, for conventional fiber lengths, there is no need for
amplification in the RN. Comparing to the 10 Gb/s OOK
standards (Table 1), superior sensitivity is achieved, in
addition to increased spectral efficiency (by a factor of four).
The available power budget is 31 dB, assuming 0 dBm input
power, which is in line with 10 Gb/s standards [2], [3]. For a
fiber length of 25 km, up to 256 users can be accommodated.
If both long reach and large splitting ratio are required, the
EDFA in the RN can function as a reach extender, increasing
the available power budget. By moving to lower-order
modulation (D8PSK) and binary-signal generation, even
higher sensitivity is obtained, despite the fourfold increase in
the symbol rate. As a result, for the D8PSK experiment the
power budget is even higher, around 34 dB (assuming 0.5
dBm input power). This result represents a 3 dB increase over
existing PON standards for comparable input powers, while
providing three times higher bit rate in the downstream
channel. This extra power budget can be translated into 15 km
longer reach or double splitting ratio and clearly showcases the
potential of advanced modulation formats as an enabling
technology for future PONs. This value allows a splitting ratio
of 512 users, again for 25 km-long PONs. Alternative
allocations of the budget are naturally possible, e.g. 60 km
fiber length and 128 users. Even longer distances and higher
PON sizes are possible, if an in-line amplification scheme, like
the one used in the Star 16QAM experiment, is implemented.
Regarding power penalties, no degradation due to the
bidirectional transmission over a single fiber with the upstream
channel is observed, as well as no impact of Rayleigh back-
scattering. In general, the suitability of the differential formats
for access applications has been demonstrated and the
performance of the generic, DSP-based receiver has been
proved. The splitting ratios and fiber lengths afforded by the
achieved power budgets enable the reduction of the number of
COs required and the upgraded system rate allows increased
available bandwidth per user. The subsequent reduction of
capital and operating expenses for the operators can justify the
introduction of more complex, DSP-based transceivers in
access networks.
The complexity of the set-ups used for the experimental
proof-of-concept can be considerably reduced, making the
proposed systems more suitable for future commercial
deployments. DCF can be removed, if DSP-based transmitters,
which enable CD compensation by pre-distorting the signal
before transmission, are utilized. CD compensation can also be
performed in the receiver side [41]; however it is not
preferable since it introduces additional complexity in the
ONU. In actual PON deployments, it is not desirable to
employ optical pre-amplification in the ONU. However, if
balanced photo detectors with linear TIAs are used, un-
amplified receivers with satisfactory sensitivity will be
feasible. Any power penalty due to the lack of optical pre-
amplification can be offset by increasing the transmitted
power. Another option is to perform single-ended detection on
each output of the demodulator with PIN/TIA, as photodiodes
with linear TIAs are available, and perform the balancing in
the digital domain, after sampling. Theoretically, the two
aforementioned detection schemes have the same performance,
but the limited resolution of ADCs means that balanced
detection is preferable in practice [34]. The proposed system
can be further simplified by employing more cost-efficient
laser sources, such as DFB lasers. It has been shown [42] that
a linewidth of around 1.5 MHz (at 10 Gsym/s) incurs modest
penalty of 1 dB for D8PSK, for BER close to the FEC limit.
Regarding the required ADC rates, for 10 Gsym/s signals a
sampling rate of 20 Gsa/s is sufficient for the reconstruction of
the signal, which can ease the requirements for high-speed
electronics in the ONU.
VI. CONCLUSION
In this paper, a review of our proposals to use differential
multilevel modulation formats in access networks is given. It
has been shown through simulations and experimental results
that differential formats with high spectral efficiency can
enable Next-Generation PONs, with increased bit rate (30
Gb/s shown experimentally), splitting ratio (up to 512 users)
and potentially reach. These results represent the highest
combination of system bit rate and splitting ratio achieved so
far for PONs without coherent receivers on a single
wavelength and polarization. The cost aspects of such systems
has been explored, with the analysis showing that the progress
in optical integration and silicon technology can allow cost-
efficient receivers, if mass-market criteria are met.
Implementation challenges have been discussed and the
bottlenecks in terms of components for the introduction of
advanced, DSP-based incoherent 16QAM formats have been
identified. The presented results indicate a promising future
for competitively priced incoherent receiver based ONUs to
support high spectral efficiency in access networks.
REFERENCES
[1] Cisco, Cisco Visual Networking Index: Forecast and Methodology,
2010-2015, white paper, Jun. 2011.
[2] F. J. Effenberger, "The XG-PON System: Cost Effective 10 Gb/s
Access," J. Lightw.Technol., vol. 29, no. 4, p.p. 403-409, Feb. 10, 2011.
[3] K Tanaka, A. Agata, Y. Horiuchi, "IEEE 802.3av 10G-EPON
Standardization and Its Research and Development Status," J. Lightw.
Technol., vol.28, no.4, pp.651-661, Feb.15, 2010.
[4] Analysis Mason, Fibre Capacity Limitations in Access Networks,
report for OFCOM, Jan. 2010.
[5] R. Heron, "Next Generation Optical Access Networks," presented at
Access Netw. and In-house Comm., Toronto, Canada, paper AMA2,
2011.
[6] J. J. Zhang, N. Ansari, "Design of WDM PON with tunable lasers: The
upstream scenario," J. Lightw. Technol., vol.28, no. 2228-236, Jan. 15,
2010.
[7] P.J. Urban, E.Pluk, M.M. de Laat, F.M. Huijskens, G.D. Khoe,
A.Koonen, H. de Waardt, "1.25-Gb/s Transmission Over an Access
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

11
Network Link With Tunable OADM and a Reflective SOA," IEEE
Photon. Technol. Lett., vol.21, no.6, pp.380-382, Mar. 15, 2009.
[8] S.-B. Park, D.K. Jung, D.J. Shin, H.S. Shin, I.K. Yun, J.S. Lee, Y.K.
Oh, Y.J. Oh, "Demonstration of WDM-PON with 50 GHz channel
spacing employing spectrum-sliced reflective semiconductor optical
amplifiers, " Electron. Lett., vol.42, no.20, pp.1172-1173, Sept. 28,
2006.
[9] P. Ossieur et al., "Demonstration of a 32 512 Split, 100 km Reach, 2
32 10 Gb/s Hybrid DWDM-TDMA PON Using Tunable External
Cavity Lasers in the ONUs," J. Lightw. Technol., vol. 29, no. 4, p.p.
3705-3718, Dec. 15, 2011.
[10] D. Lavery, E. Torrengo, and S. Savory, "Bidirectional 10 Gbit/s long-
reach WDM-PON using digital coherent receivers," presented at Optical
Fiber Conf. (OFC), 2011, paper OTuB4.
[11] P. Iannone et al., "Bi-Directionally Amplified Extended Reach 40Gb/s
CWDM-TDM PON with Burst-Mode Upstream Transmission,"
presented at Optical Fiber Conf. (OFC), 2011, paper PDPD6.
[12] N. Cvijetic, M. Huang, E. Ip, Y. Shao, Y. Huang, M. Cvijetic, and T.
Wang, "Coherent 40Gb/s OFDMA-PON for Long-Reach (100
+
km)
High-Split Ratio (>1:64) Optical Access/Metro Networks," presented at
Optical Fiber Conf. (OFC), 2012, paper OW4B.8.
[13] D. Qian, S. Fan, N. Cvijetic, J. Hu, and T. Wang, "64/32/16QAM-
OFDM using Direct-Detection for 40G-OFDMA-PON Downstream,"
presented at Optical Fiber Conf. (OFC), 2011, paper OMG4.
[14] R.A. Griffin, A.C.Carter, "Optical differential quadrature phase-shift
key (oDQPSK) for high capacity optical transmission," presented at
Optical Fiber Conf. (OFC), 2002, paper WX6.
[15] M.Serbay, C. Wree, and W. Rosenkranz, "Experimental investigation of
RZ-8DPSK at 310.7 Gb/s," presented at The 18th Ann.Meeting of the
IEEE Lasers and Electro-Optics Society, pp. 483- 484, 22-28 Oct.
2005.
[16] M. Seimetz, M. Noelle, and E. Patzak, "Optical Systems With High-
Order DPSK and Star QAM Modulation Based on Interferometric
Direct Detection," J. of Lightw. Technol., vol.25, no.6, pp.1515-1530,
Jun. 2007.
[17] Y. Takushima, H. Y. Choi, and Y. C. Chung, "Transmission of 108-
Gb/s PDM 16ADPSK signal on 25-GHz grid using non-coherent
receivers," Opt. Express, vol.17, issue 16, p.p. 13458-13466, Aug. 3,
2009.
[18] M. Serbay, T. Tokle, P. Jeppesen, and W. Rosenkranz, 42.8 Gbit/s, 4
bits per symbol 16-ary inverse-RZ-QASK-DQPSK transmission
experiment without Polmux, presented at Optical Fiber Conf. (OFC),
2006, paper OThL2.
[19] N. Kikuchi, S. Sasaki, "Highly Sensitive Optical Multi-level
Transmission of Arbitrary Quadrature-Amplitude Modulation (QAM)
Signals With Direct Detection," J. Lightw. Technol. vol.28, no.1,
pp.123-130, Jan.1, 2010.
[20] R.G. Egri, F.A. Horrigan, "A finite group of complex integers and its
application to differentially coherent detection of QAM signals," IEEE
Trans. Inf. Theory, vol.40, no.1, pp.216-219, Jan 1994.
[21] K. Ho, Phase-Modulated Optical Communication Systems, Berlin,
Springer, pp. 301-316, 2005.
[22] H. Yoon, D. Lee, and N. Park, Performance comparison of optical 8-
ary differential phase-shift keying systems with different electrical
decision schemes, Opt. Express, vol.13, is. 2, p.p. 371-376, Jan. 24,
2005.
[23] Y. Han, C. Kim, G. Li, Simplified receiver implementation for optical
differential 8-level phase-shift keying, Electron. Lett., vol.40, no.21,
pp. 1372- 1373, Oct. 14, 2004.
[24] G. Foschini, R. Gitlin, S. Weinstein, "Optimization of Two-Dimensional
Signal Constellations in the Presence of Gaussian Noise," IEEE Trans.
Commun. vol.22, no.1, pp. 28-38, Jan. 1974.
[25] M. Simon, G. Huth, A. Polydoros, "Differentially Coherent Detection of
QASK for Frequency-Hopping Systems--Part I: Performance in the
Presence of a Gaussian Noise Environment," IEEE Trans. Commun.,
vol.30, no.1, pp. 158- 164, Jan. 1982.
[26] M. Seimetz, High Order Modulation for Optical Fiber Transmission,
Berlin, Springer, pp. 37-57, 2009.
[27] M. S. Alfiad, M. Kuschnerov, S. L. Jansen, T. Wuth, D. van den Borne,
and H. de Waardt, 11 x 224-Gb/s POLMUX-RZ-16QAM Transmission
Over 670 km of SSMF With 50-GHz Channel Spacing, IEEE Photon.
Technol. Lett., vol.22, no. 15, pp. 11501152, Aug. 1, 2010.
[28] S. Jain, F. Effenberger, A. Szabo, Feng Zhishan, A. Forcucci, Guo Wei,
Luo Yuanqiu, R. Mapes, Zhang Yixin, V. O'Byrne, "World's First XG-
PON Field Trial," J. Lightw. Technol., vol.29, no.4, pp.524-528, Feb.15,
2011.
[29] C.R. Doerr, L. Chen, Monolithic PDM-DQPSK receiver in silicon,
presented at Europ. Conf. on Opt. Commun. (ECOC), 2010, paper
PDP3.6.
[30] N.Cvijetic, "OFDM for Next-Generation Optical Access Networks," J.
Lightw. Technol., vol.30, no.4, pp.384-398, Feb.15, 2012.
[31] I.Dedic, High-speed CMOS DSP and data converters, presented at
Optical Fiber Conf. (OFC), 2011, paper OTuN1.
[32] I.Dedic, "56Gs/s ADC: Enabling 100GbE," presented at Optical Fiber
Conf. (OFC), 2010, paper OThT6.
[33] S.J. Savory, Digital Signal Processing Options in Long Haul
Transmission, presented at Optical Fiber Conf. (OFC), 2008, paper
OTuO3.
[34] S.J. Savory, "Digital Coherent Optical Receivers: Algorithms and
Subsystems," IEEE J. Sel. Topics in Quantum Electron., vol.16, no.5,
pp.1164-1179, Sept.-Oct. 2010.
[35] N. Sotiropoulos, T. Koonen, H. de Waardt, "Bidirectional incoherent
16QAM transmission over hybrid WDM/TDM passive optical
network," presented at 12th Intern. Conf. on Transparent Opt. Netw.
(ICTON), 2010, paper Th.A2.4.
[36] N. Sotiropoulos, H. de Waardt, T. Koonen, "Novel 16QAM Detection
Scheme for Optical Access networks," presented at Access Netw. and
In-house Comm., Toronto, Canada, paper AMC5, 2011.
[37] G. Bosco, P. Poggiolini, "On the joint effect of receiver impairments on
direct-detection DQPSK systems," J. of Lightw. Technol., vol.24, no.3,
pp.1323-1333, March 2006.
[38] N. Sotiropoulos, A.M.J. Koonen, H. de Waardt, On the effect of
receiver impairments on incoherent QAM systems, Presented at the
15th Annual Symposium of the IEEE Photonics Benelux Chapter, 2010,
Delft, The Netherlands, pp. 29-32.
[39] N. Sotiropoulos, A.M.J. Koonen, H. de Waardt, Star 16QAM/OOK
Bidirectional Transmission over a TDM-PON, in Microw. and Opt.
Technol. Lett., accepted for publication
[40] N. Sotiropoulos, A.M.J. Koonen, H. de Waardt, Experimental
Demonstration of an Incoherent TDM-PON with 30 Gb/s D8PSK
Downstream and 10 Gb/s OOK Upstream Data, presented at Europ.
Conf. on Opt. Commun. (ECOC), 2012, paper Tu.3.B.2.
[41] X. Liu, X. Wei, "Electronic Dispersion Compensation Based on Optical
Field Reconstruction with Orthogonal Differential Direct-Detection and
Digital Signal Processing," presented at Optical Fiber Conf. (OFC),
2007, paper OTuA6.
[42] Y. Han, and G. Li, "Sensitivity limits and degradations in OD8PSK,"
IEEE Photon. Technol. Lett., vol.17, no.3, pp.720-722, March 2005.
Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

12
Table 1 Modulation formats and receiver sensitivity
Modulation format/ bit rate Receiver sensitivity (dBm)
OOK/ 10 Gb/s (XG-PON) [2] -22.5 (PIN)/ -28 (APD)
OOK/ 10 Gb/s (10G-EPON) [3] -20.5 (PIN)/ -28.5 (APD)
Phase pre-integration 16QAM/ 10 Gb/s -39 (pre-amplifier)
Star 16QAM/ 10Gb/s -31 (pre-amplifier, no TIA)
D8PSK/ 30 Gb/s -33.5 (pre-amplifier, no TIA)


Fig. 1a D8PSK constellation diagram


Fig. 1b Star 16QAM constellation diagram


Fig. 2a Serial transmitter

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

13

Fig. 2b Parallel transmitter


Fig. 2c DSP-based transmitter


Fig. 3a Binary differential receiver


Fig. 3b Multilevel differential receiver

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

14

Fig. 4a 16QAM constellation diagram


Fig. 4b Transmitted constellation diagram with phase pre-integration



Fig. 4c Differential constellation diagram, coded 16QAM


Fig. 5 Binary square 16QAM transmitter

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

15

Fig. 6 TDM-PON architecture


Fig. 7 Phase pre-integration 16QAM simulation model


Fig. 8 Upstream channel BER vs. carrier input power


Fig. 9 Phase pre-integration 16QAM received constellation diagram

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

16

Fig. 10 BER vs. received power, phase pre-integration 16QAM PON


Fig. 11 Coded 16QAM simulation model


Fig.12 BER vs. OSNR downstream channel, coded 16QAM PON


Fig. 13 BER vs. OSNR upstream channel, coded 16QAM PON


Fig.14 4-level driving signal at the output of the AWG at 2.5 Gsym/s

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

17

Fig. 15a: Simulated constellation diagram coded 16QAM, ideal case

Fig. 15b Simulated constellation diagram coded 16QAM, 2/3 driving voltage


Fig. 15c Simulated constellation diagram coded 16QAM, 2/3 driving voltage and 80
0
phase shift


Fig. 16 Experimental set-up, Star 16QAM/OOK PON

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

18

Fig. 17 Star 16QAM received constellation diagrams (a) back to back (b) 25 km fiber

Fig. 18 BER vs. inner ring radius (normalized to 1 for outer ring)


Fig. 19 BER vs. received power, 10 Gb/s Star 16QAM


Figure 20 BER vs. received power, 2.5 Gb/s OOK

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

19

Fig. 21 Experimental set-up, D8PSK/OOK PON


Fig. 22 D8PSK received constellation diagrams (a) back to back (b) 25 km fiber


Fig. 23 OOK received eyes (a) back to back (b) 25 km fiber


Fig. 24 BER vs. received power, 30 Gb/s D8PSK

Copyright (c) 2013 IEEE. Personal use is permitted. For any other purposes, permission must be obtained from the IEEE by emailing pubs-permissions@ieee.org.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication.
JLT-14584-2012

20

Fig. 25 BER vs. received power, 10 Gb/s OOK