Transmission of Waveforms Determined by 7 Eigenvalues with

PSK-Modulated Spectral Amplitudes

Henning Bülow, Vahid Aref, Wilfried Idler
Nokia Bell Labs, Lorenzstr. 10, 70435, Stuttgart, Germany [email protected]

Abstract 2-ns waveforms with 7 eigenvalues and their QPSK-modulated spectral amplitudes were
optimized by taking constraints of link, transmitter, and receiver into account. In experiment these
signals were transmitted with a BER of 3.2E-3 over 1440-km of NZ-DSF fiber spans.

Introduction enough for reliable detection in Rx in the

In recent years the modulation of the nonlinear presence of ASE noise, and numerical limitation
spectrum or detection in this nonlinear domain of current NFT algorithms. (ii) The largest pulse-
has been re-discovered and refined for width (duration) of all pulses becomes minimum
transmission over the fiber in the nonlinear (highest transmission rate) (iii) The largest
regime. Such schemes are called nonlinear bandwidth (BW) of all pulses becomes minimum
frequency division multiplexing (NFDM)1,2. in Tx and Rx (highest spectral efficiency) (iv) No
Following the argumentation of Wahls3, inter-symbol interference (ISI) between adjacent
nonlinear Fourier transform (NFT) based pulses during the transmission (v) Pick-to-
modulation and/or detection is motivated by (1) average power ratio (PAPR) stays in a practical
replacing multi step digital back-propagation range.
(DBP) by a less complex one step NFT scheme, It is not yet fully clear how to satisfy all these
by (2) exploiting the quasi linear propagation constraints in general, while these conditions
behaviour of the nonlinear spectrum for are strongly coupled. We relax the condition (iii)
operation beyond the nonlinear threshold, or by to a given BW limit. We limit the BW in Tx to
(3) using detection schemes working on the 12.5 GHz, and BW in Rx to 20 GHz assuming
nonlinear spectrum rather than on the time the fiber link length is less than 2000 km,
domain signal. So far, only limited work has including some BW margin for more precise
been done on lab implementation of these waveform generation and detection. We define
schemes. Recently, the discrete part of the BW (pulse-width, respectively) as the frequency
nonlinear spectrum was modulated by on/off (time) interval holding 99% of pulse energy.
keying of up to 4 eigenvalues4 or by We choose a set of 7 eigenvalues as shown in
independently QPSK modulated spectral Table 1. The spectral amplitude of each
amplitudes of a 2 eigenvalue soliton5. eigenvalue, , has QPSK constellation,
Modulation in the continuous part of the . Thus, we modulate
nonlinear spectrum was also demonstrated6.
at most 14 information bits per pulse.
In this paper, we present the design of pulses
Distance between eigenvalues is chosen to fulfil
with a high count of 7 eigenvalues and QPSK
conditions (i), (ii) and (iii); on one hand, large
modulation of spectral amplitudes. The
difference in imaginary parts may lead to a large
eigenvalues are not all located on the imaginary
pulse-width and large difference in real parts
axis as it was the case in the papers before. For
leads to a large BW. On the other hand, small
the first time, we demonstrate generation,
distance between two eigenvalues may also
transmission, and detection of such 7-
increase both pulse-width and BW.
eigenvalue signals in experiment.
Since the eigenvalues have different real parts,
NFDM Signal Design the pulse will be broadened or contracted during
Our goal was to design a set of finite the propagation. To avoid ISI (condition (iv)), we
eigenvalues (carriers), , where the information set . By
bits are modulated over spectral amplitudes, choosing a positive , the pulses will first be
independently, provided that the following contracted and then eventually broadened to the
conditions are fulfilled: same pulse-width and beyond. The link length of
(i) Distance between eigenvalues is large reaching to the same pulse-width, ,

Table 1: The discrete spectrum of waveforms at Tx. is fixed and  is QPSK modulated.
 2ns
Erx(t)
1550nm, FB
=1kHz qd(i)
88GSa/s

modulator
I 1550nm,
D/A =1kHz FC
I/Q
PolC on
converter Q
i
80GSa/s
Normalization 3x 24km NZ-DSF fiber Coherent
receiver min(EDi) demap BER
INFT OBPF Offline

:
5 0GHz Pin Pin Pin processing
Eref,1
EDFA
:
λ1 ... λ7 Eref,N
qd(λ1)... qd(λ7)
Fig. 2: Offline processing of received 2-ns signal
Fig. 1: Lab transmission set-up with fiber loop consisting of 3x 24-km NZ- traces Erx(t). FB: qd by a forward-backward NFT
DSF spans and EDFA amplifiers. Transmitter: 88-GSa/s DAC, Receiver: algorithm. FC: i by Fourier collocation method.
single polarization optical hybrid and 33-GHz bandwidth oscilloscope. min(ED): Decision on min. Euclidian distance.

depends on . We optimize and to detection by offline processing:

have the pulse-width (for standard nonlinear (1) On one hand the received 2-ns traces
Schrödinger equation) which scales down to Erx(t) were demultiplexed in the nonlinear
2 ns in our transmission setup (7 Gbit/s). Using spectral domain by determining the eigenvalues
NZ-DSF fiber with 2  -5.75 ps2/km and   i (FC block in Fig. 2) first and by calculating the
1.6 W-1/km, the optimized allows us to have a spectral amplitude qd(i) for each i (i=1...7)
link about Lmax  5600 km without ISI. using the forward-backward algorithm7 (FB
The optimized are given in Table 1. block). It exhibits an improved numerical stability
Using split-step Fourier method, we simulate the and precision compared to the often used
transmission of designed pulses (the random Ablowitz Ladik integration2. At FB output, we
256 symbols used in experiment) over the NZ- disregarded the magnitude of qd and determined
DSF fiber model with the power loss  = its phase qd(i) which is expected to be a
0.2 dB/km and span length of 24 km. We QPSK phase. For calculation (FC block) of the
numerically observe that we can detect the eigenvalues i, the Fourier collocation (FC)
spectrum error-free (hard-decision over phases) method2 was applied. A scatterplot of 1 ... 7 of
in the range of interest, i.e. 2000 km. 2800 symbols in the complex plain is shown in
Fig. 3 (c). We observe a clear distinction of the 7
Experimental set-up and signal processing
eigenvalues even though some of them are
As illustrated in Fig. 1, for experimental
already perturbed rather considerably at the
evaluation we randomly selected 256 different
transmitter.
2-ns waveforms offline, uploaded these 512-ns
(2) A further decision process (min(ED)) in
sequence into an 88-GSa/s digital-to-analog
the time domain was performed separately
converter (DAC) driving a Mach-Zehnder IQ
based on Euclidian distance. For each possible
modulator. This signal was launched into a NZ-
waveform i, (i=1...214), its discrete spectrum is
DSF fiber loop set-up with 24-km span length
linearly transformed to Rx. The referece
and lumped amplification by EDFA. With this
waveforms Eref,i are obtained by taking INFT
short fiber length, we have only a small penalty
from transformed spectrum. For each received
because of the mismatch between lossy multi-
pulse Erx(t), the reference Eref,i was selected
span fiber in experiment and lossless fiber
which has the least Euclidian distances EDi=
model of INFT (modified Darboux transform)
|Erx-Eref,i|2dt. Even though this min(ED) decision
and NFT. In the transmission experiments we
is sub-optimum since noise statistics is expected
measured optimum launch powers Pin of approx.
+2j
-2.2 dBm which confirms the theoretical path- |Erx(t)| |Eref(t)| (a) (b) Erx(t)
+1j
averaged value of -1.53 dBm. The transmitted Eref(t)
a.u.

signal is coherently detected and a trace of 0j

t
about 2800 2ns-symbols is sampled by a real- -1j

time oscilloscope with 80 GSa/s and 33 GHz 0 2 4 6 t [ns] 8 10 -2j

-2 -1 0 +1 +2
bandwidth. Best results were obtained if the Fig. 3: Back-to-back signals
polarization was adjusted in the optical domain (a): Five 2-ns symbol traces vs. (c)
time, black/red are |Erx| / |Eref|. 1 .... 7
by a manually adjusted polarization controller 0.5j
(b): Erx(t) / Eref(t) of first symbol in
(PolCon in Fig. 1). complex plane.
Fig. 2 shows the two different detection (c): Eigenvalues i of 2800 received
schemes were applied separately for signal symbols in complex plane. 0 j
-0.5 0 +0.5
 q d

1 2 3 4 5 6 7
Fig. 4: Histograms (2800 symbols) of spectral amplitude phases qd(i) of eigenvalues i (i=1…7) measured at transmitter.
Gray scale denotes relative occurrence.
to be signal dependent, it can indicate on such as separation of eigenvalues, minimum
whether the waveforms Erx contain sufficient pulse width, maximum link length, limited
information for reliable decision on the symbols. bandwidth at transmitter and receiver, avoiding
BER after symbol to bit demapping are provided inter-symbol interference. Under realistic
as output for min(ED) decision (see Fig. 2). constraints for the lab set-up, we successfully
generated 2-ns long waveforms of 7 eigenvalue
Experimental results
solitons whose discrete spectra were modulated
The magnitude of the first five measured traces
in phase (PSK).
Erx(t) of the signal at modulator output are
We transmitted them over up to 1440-km NZ-
shown in Fig. 3 (a) as black lines. They are
DSF fiber in lab. A low BER of 3.210-3 was
compared to the ideal (calculated) signal Eref,i
obtained by searching for the minimum
(i=1...5) plotted as red lines. We observe slight
Euclidian distance of the time domain signal.
deviations at sharp peaks and at zeros between
Even though the low BER confirmed that
the symbols. Slight but not paramount
sufficient information for low BER detection is
deviations can also be observed when plotting
present in the received symbol waveform, low
the first symbol trace Erx in the complex plane,
error-rate decision was not possible when
as shown in Fig. 3 (b). We attribute these field
demultiplexing the signal by nonlinear Fourier
perturbations mainly to imperfections of the
transform (NFT) and deciding on the phases of
transmitter5,8, i.e. limited bandwidth of DAC and
the 7 discrete spectral amplitudes qd(i)
modulator drive amplifier, and degraded DAC
separately. These observations confirm that 7
resolution at high frequencies (ENOB5.5bit).
eigenvalue solitons were successfully generated
Nevertheless the eigenvalues appear in 7
and transmitted, but it also shows that the
distinct clusters as shown in Fig. 3 (c). The
simple concept of demultiplexing eigenvalues by
histogram of the measured phases qd(i) at
NFT processing in its current form remains
transmitter output are shown as differently
challenging for PSK modulated eigenvalues.
shaded gray points on a circle in Fig. 4. The
subsets of the 4 QPSK phase states used for References
signal design are clearly visible. However, the [1] J.E. Prilepsky et al., "Nonlinear inverse synthesis and
non-vanishing gray points between 90° eigenvalue division multiplexing in optical fiber
separated phases (see e.g. histograms 1, 3 and channels," Physical review letters, 113, 013901, (2014).
6 in Fig. 4) already indicate large variance of the [2] M.I. Yousefi et al., “Information Transmission using the
phases which we attribute to the lab set-up Nonlinear Fourier Transform, Part I-III”, IEEE Trans. Inf.
imperfections and which leads to a high error- Theory, vol. 60, no. 7, pp. 4329–4369, (2014).
floor. [3] S. Wahls, “Fiber-Optic Communication using Fast
Nevertheless, as summarized in Tab. 2, with Nonlinear Fourier Transforms,” OFC, W3A.1 (2016).
min(ED) detection of 2800 waveforms we [4] Z. Dong et al., "Nonlinear Frequency Division
obtained no error after up to 864-km Multiplexed Transmissions based on NFT," Photon.
transmission (12x72km) and BER = 3.210-3 Technol. Lett., Vol. 27, No. 15, (2015).
after 1440-km (20x72km). We also observed [5] V. Aref et al., “Experimental Demonstration of Nonlinear
that the (best) launch power Pin (bottom row) Frequency Division Multiplexed Transmission,” Proc.
has the tendency to increase with increasing link ECOC, Tu.1.1.2, (2015)
length. [6] Son T. Le et al., “First Experimental Demonstration of
Table 2: BER and optimum launch power for min(ED) Nonlinear Inverse Synthesis Transmission over
detection vs.link length. Transoceanic Distances,” OFC, Tu2A.1 (2016).
864 km 1152 km 1440 km 1800 km [7] V. Aref, ”Control and Detection of Discrete Spectral
-4 -3 -2
BER 0 9.110 3.210 110 Amplitudes in Nonlinear Fourier Transform,” Arxiv.org
Pin -2.3dBm -2.1dBm -2.1dBm -1.8dBm
preprint, (2016).
[8] H. Buelow et al., “Experimental Nonlinear Frequency
Conclusions Domain Equalization of QPSK Modulated 2-Eigenvalue
For the design of waveforms defined by many Soliton,” OFC, Tu2A.3 (2016).
eigenvalues, several conditions have to be met,