6 Bit S HZ Orthogonally Polarized CSRZ D PDF

TuF1
1.6 bit/s/Hz orthogonally polarized CSRZ - DQPSK

transmission of 8u40 Gbit/s over 320 km NDSF
Y. Zhu, K. Cordina, N. Jolley, R. Feced, H. Kee, R. Rickard and A. Hadjifotiou
Nortel Networks, Harlow Laboratories, London Road, Harlow, Essex CM17 9NA, UK
Tel: +44-(0)1279 40 2499, Fax: +44-(0)1279 40 2115, e-mail: yanjun@nortelnetworks.com
Abstract: We report experimental results of 1.6 bit/s/Hz, 8u40 Gbit/s OP-CSRZ-DQPSK

transmission over 320 km NDSF and demonstration of an 8-fold extension in 40 Gbit/s chromatic
dispersion tolerance and a doubling of 40 Gbit/s PMD tolerance without any adjustable
compensator.
2004 Optical Society of America
OCIS codes: (060.2310) Fiber optics; (060.4510) Optical communications
1. Introduction
Recent advances in high speed long haul transmission have been accelerated by enabling technologies including
enhanced formats such as DPSK [1-6] and the development of strong FEC techniques [7]. On the other hand, ultra
high spectral efficiencies of up to 1.6 bit/s/Hz have been reported for 40 Gbit/s [8-11]. In [11], 8u40 Gbit/s RZ-
DQPSK signals were transmitted over a record 200 km NDSF with the highest 40 Gbit/s spectral efficiency of 1.6
bit/s/Hz demonstrated to date.
In this paper, we investigate orthogonally polarised (OP)-CSRZ-DQPSK transmission at 40 Gbit/s/ wavelength
and demonstrate transmission of 8u40 Gbit/s over 320 km NDSF at 1.6 bit/s/Hz spectral efficiency. We also
demonstrate the capability of OP-CSRZ-DQPSK for a significant extension of 40 Gbit/s chromatic dispersion
tolerance and PMD tolerance in the absence of any adjustable dispersion compensators.
2. System configuration
80 km 80 km 80 km
10 GHz 20 Gbit/s 40 Gbit/s NDSF NDSF NDSF
O1
Integrated
O2 CSRZ DQPSK 3 dB PMC DCM DCM
Pulse
: Modulator
EDFA
O8
PM Coupler
10 Gbit/s Balanced
80 km
BERT Receiver DQPSK NDSF
Decoder
Balanced 0.6 nm 0.25 nm 0.25 nm
Receiver BPF DCM DCM
PBS BPF BPF
Balanced
Receiver DQPSK
Decoder
Balanced
Receiver
Fig. 1 System configuration
The experimental set-up is shown in Fig.1. The channel wavelengths varied from 1556.4 nm to 1557.8 nm with a
spacing of 0.2 nm (25 GHz). They were combined through a PM coupler (PMC). The channels went through a
chirp-free LiNbO3 modulator driven at 5 GHz for CSRZ modulation. The CSRZ pulses had a duty cycle of
approximately 60%. The second modulator was an integrated GaAs/AlGaAs DQPSK data modulator. As a result
20 Gbit/s CSRZ-DQPSK signals were produced. The phase difference of S/2 between in-phase and quadrature
channels was maintained through an integrated two-photon-absorption monitor. The 20 Gbit/s signals were then
split, delayed and recombined to form orthogonally polarized OP-CSRZ-DQPSK signals at 40 Gbit/s per
wavelength. The time delay between dual polarizations was set at 145 ns and de-correlation was confirmed to be
sufficient. The transmission link included four 80 km spans with EDFA-only amplification. Transmission fibre loss
varied from 15.6 dB to 19 dB and DCM loss was around 9.5 dB. DCM dispersions at 1557 nm were 1369.8 ps/nm,
1375.6 ps/nm, 1371.5 ps/nm and 1542.1 ps/nm. At the receiver, the channels were wavelength and polarization
demulptiplexed to individual 20 Gbit/s signals before demodulation. The demodulator was a temperature-controlled
TuF1
interferometer with a relative delay of around 100 ps. The in-phase or quadrature demodulated channels were
received with a 10 Gbit/s balanced receiver with a bandwidth of around 14 GHz. The pattern generator made use of
pre-coded data and the error detector was programmed to receive the corresponding sequence. A data pattern length
of 215 was used for all the measurements. This pattern length was only limited by the memory of the pattern
generator.
3. Experimental results
Chromatic dispersion and PMD tolerance:
4 4
CSRZ-DQPSK CSRZ-DQPSK
CSRZ-IMDD CSRZ-IMDD
3 3
Q penalty (dB)
Q penalty (dB)
2 2
1 1
0 0
-800 -600 -400 -200 0 200 400 600 800 0 10 20 30 40 50
Chromatic dispersion (ps/nm) DGD (ps)
Fig.2 Measured chromatic dispersion tolerance Fig.3 Measured PMD tolerance
The measured chromatic dispersion and PMD tolerance of a 40 Gbit/s OP-CSRZ-DQPSK signal are presented in
Fig.2 and Fig.3, respectively. They were compared with the tolerances of a 40 Gbit/s ETDM-based CSRZ-IMDD
signal. A significant extension of the 40 Gbit/s chromatic dispersion tolerance and PMD tolerance was observed.
The 2dB (20log Q) penalty chromatic dispersion tolerance of a 40 Gbit/s OP-CSRZ-DQPSK signal was measured to
be 950 ps/nm. This represents an 8-fold increase compared to the dispersion tolerance of a 40 Gbit/s ETDM-based
CSRZ-IMDD signal, which was about 110 ps/nm. A decoded eye diagram for a CSRZ-DQPSK signal in the
presence of +680 ps/nm dispersion is shown as the inset in Fig.2. Similarly, comparison of the measured PMD
tolerance is shown in Fig.3. The 1dBQ penalty PMD tolerance of OP-CSRZ-DQPSK and CSRZ-IMDD at 40 Gbit/s
was 22 ps and around 10 ps, respectively. A doubling of the PMD tolerance (for 1dBQ penalty) was achieved
compared with 40 Gbit/s CSRZ-IMDD. Such benefits result from the use of a symbol rate of 10 Gbaud instead of 40
Gbaud for a capacity of 40 Gbit/s/wavelength. Note that no adjustable chromatic dispersion or PMD compensators
were used.
1.6 bit/s/Hz OP-CSRZ-DQPSK transmission:
10 10
40 Gbit/s CSRZ-DQPSK 10 GHz CSRZ pulse filtered 40G TE
0 0
(a) (b) filtered 40G TM
-10 -10
Power (dBm)
Power (dBm)
-20 -20
-30 -30
-40 -40
-50 -50
-60 -60
1556 1556.5 1557 1557.5 1558 1556 1556.5 1557 1557.5 1558
Wavelength (nm) Wavelength (nm)
Fig.4 Back-to-back spectra of (a) 1.6 bit/s/Hz 8u40 Gbit/s OP-CSRZ-DQPSK and the corresponding
10 GHz CSRZ intensity modulation; (b) wavelength and polarization demultiplexed Channel 4.
Fig.4(a) shows the back-to-back 1.6 bit/s/Hz OP-CSRZ-DQPSK signal spectrum, together with the spectrum of the
original 10 GHz CSRZ pulses used for the intensity modulation. The inset in Fig.4(a) is a typical eye diagram of a
40 Gbit/s OP-CSRZ-DQPSK signal before demodulation. Fig.4(b) gives the performance of the combined
TuF1
wavelength and polarization demultiplexing, showing the spectra of each polarization of channel 4 after
demultiplexing at the receiver. Despite the absence of pre-filtering at the transmitter, the crosstalk between adjacent
wavelength channels was 12.2 dB through post-filtering and the polarization extinction was more than 30 dB.
Transmission over 320 km NDSF was then performed. The launch power used was 3 dBm/channel for the
transmission fibre and 6 dBm/channel for the DCMs. Fig.5 shows the received eye diagrams for the even channels.
Clear eye opening was achieved for all wavelength channels. No phase adjustment between in-phase and quadrature
channels for different wavelengths was necessary at the integrated data transmitter. Fig.6 gives the measured Q-
factors for all the channels. The Q-factor varied from 14.8 dB to 16.3 dB, with an average Q-factor of 15.4 dB in the
absence of FEC.
18
15
12
Q-factor (dB)
9
0
1556.2 1556.6 1557 1557.4 1557.8
Fig.5 Received eye diagrams for even Wavelength (nm)
channels at 320 km Fig.6 Measured Q-factor vs. wavelength at 320 km
k
The significant performance margin at 320 km suggests a possibility of a longer reach at 1.6 bit/s/Hz with the use of
FEC techniques. With a view to future long reach applications of such high spectral efficiency 40 Gbit/s DQPSK
systems, it is expected that a polarization tracking receiver will play a critical role [12]. A polarization tracking
receiver was not employed in this experiment but in a real system a polarisation tracking subsystem will be
necessary.
4. Conclusion
In conclusion, we have successfully transmitted 8u40 Gbit/s OP-CSRZ-DQPSK signals over a distance of 320 km at
a spectral efficiency of 1.6 bit/s/Hz. This was enabled by using OP-CSRZ-DQPSK modulation format based on an
integrated DQPSK data transmitter. The high spectral efficiency of 1.6 bit/s/Hz for OP-CSRZ-DQPSK was achieved
in the absence of any pre-filtering or forward error correction. We also experimentally investigated the chromatic
dispersion and PMD tolerance of a 40 Gbit/s OP-CSRZ-DQPSK signal. A chromatic dispersion tolerance of 950
ps/nm and PMD tolerance of 25 ps at 40 Gbit/s was successfully demonstrated. This represents an 8-fold extension
of the chromatic dispersion tolerance and a doubling of PMD tolerance of a typical 40 Gbit/s ETDM system but in
the absence of any adjustable chromatic dispersion or PMD compensators. The significant increase in system
tolerance benefited from the use of a 10 Gbaud symbol rate for a capacity of 40 Gbit/s/wavelength. Polarization
tracking is expected to be required for future long haul applications of high spectral efficiency DQPSK systems.
References
[1] A. H. Gnauck, et al, OFC’2002, paper PD-FC2.
[2] C. Rasmussen, et al, OFC’2003, paper PD18.
[3] B. Zhu, et al, OFC’2003, paper PD19.
[4] T. Tsuritani, et al, OFC’2003, paper PD23.
[5] J.-X. Cai, et al, OFC’2003, paper PD22.
[6] J. Marcerou et al, OFC’2003, paper PD20.
[7] T. Mizuochi, et al, OFC’2003, paper PD21.
[8] S. Bigo, et al, OFC’2001, paper PD25.
[9] S. Sotobayashi, et al, IEEE Photon. Technol. Lett., vol.14, pp.555, (2002).
[10] I. Morita, et al, ECOC’2002, paper PD4.7.
[11] C. Wree, et al, IEEE Photon. Technol. Lett., (2003).
[12] C. Davidson, et al, OFC’2003, paper TuF3.
TuF2
Study on Optimum Pre-Filtering Condition

for 42.7 Gbit/s CS-RZ DPSK Signal
Keiji Tanaka, Itsuro Morita, and Noboru Edagawa
KDDI R&D Laboratories Inc., 2-1-15 Ohara Kamifukuoka Saitama, 356-8502, Japan
TEL:+81-49-278-7812, FAX:+81-49-278-7844, E-mail:kj-tanaka@kddi.com
Abstract: The impact of asymmetric pre-filtering for 42.7Gbit/s CS-RZ DPSK signals were
investigated experimentally and numerically, including 9000km WDM transmission experiments.
We have found that asymmetric pre-filtering is effective for quasi-linear system applications.
OCIS codes: (060.4510) Optical Communication
1. Introduction
Spectral efficiency is a key issue to enhance aggregate capacity and cost-effectiveness of optical transmission
systems. In order to achieve high spectral-efficiency, the signal bandlimitation by using optical filtering before
transmission (i.e. pre-filtering) is an attractive solution, and many successful demonstrations using this scheme have
been reported to date [1]-[9]. Regarding on-off keying (OOK) signals, detailed studies on the benefit of asymmetric
pre-filtering have been already reported for carrier-suppressed return-to-zero (CS-RZ) signals [7][10]. However,
regarding differential phase-shift keying (DPSK) signals, which is a promising format for high bit rate ULH system
applications, similar study has not been reported so far, whereas remarkable demonstrations of 40Gbit/s-based
DWDM transmission have been already reported by using symmetrically pre-filtered DPSK signals [6][8][9].
In this paper, we have experimentally and numerically investigated the impact of asymmetric pre-filtering for 42.7
Gbit/s CS-RZ DPSK signal by comparing the dispersion and nonlinear tolerance between unfiltered, asymmetrically
and symmetrically filtered CS-RZ DPSK signals. In addition, by conducting a DWDM transmission experiment up
to 9000km, we have evaluated the long-haul transmission performance of asymmetrically pre-filtered CS-RZ DPSK
signals.
2. Back-to-back performance of pre-filtered CS-RZ DPSK signals

Fig. 1 shows the schematic experimental setup employed for various evaluations. In the transmitter, we used a DFB
laser source and two LiNbO3 modulators; one is for 42.7 Gbit/s data-coding of the optical phase of the signal in
NRZ format with a true 231-1 pseudo-random binary sequence, and the other is for bit-synchronous CS-RZ
formatting. Then, the 42.7 Gbit/s CS-RZ DPSK signal was optically bandlimited by using a 50/100 GHz
interleaving device with a 3 dB/20dB bandwidth of 45 GHz/68GHz. Here, we assumed an ultra-dense WDM system
with a spectral efficiency of 80%. In the receiver, the DPSK signal was demodulated using a Mach-Zehnder delay
interferometer (MZDI) and detected by a balanced receiver. The received 42.7Gbit/s electrical signal was
demultiplexed to 10.7Gbit/s. To evaluate the signal performance, we measured a bit error rate (BER) averaged over
the four tributary 10Gbit/s data streams using a random re-synchronization technique [5] and calculated the
corresponding Q-factor.
(1) Filter Detuning ASE

direct connection OR
Source
OS 50/100
GHz-IL (2) Dispersion Compensation Error Tolerance Error
Balanced
DPSK/ DCF ATT MZDI
DFB- Optical Receiver Detector
LD CS-RZ
Filter
MOD (3) Fiber Input Power Tolerance
50km SMF DCF Random Clock
EDFA Re-synchronizer Recovery
Fig. 1. Schematic experimental setup for the evaluation of bandlimiting filter detuning characteristics, dispersion and nonlinear tolerance.
In the evaluation of the optimum pre-filtering condition, we bypassed the transmission line and the dispersion
compensation fiber at the receiver and measured the back-to-back performance by detuning the center frequency of
TuF2
the bandlimiting filter. Fig. 2 (a) shows the Q-factor penalty due to optical bandlimitation for 42.7Gbit/s CS-RZ
DPSK signal as a function of the filter detuning defined as the difference between the signal carrier frequency and
the center frequency of the bandlimiting filter. The nominal Q-factor of the unfiltered signal was set to 14.8 dB by
adding the amplified spontaneous emission noise to simulate a long-distance transmission condition. As shown in
Fig. 2 (a), the Q-factor penalty of the asymmetrically filtered signal due to bandlimitation was distinctively lower
than that of the symmetrically filtered signal, and was less than 1dB. This low penalty condition was kept for around
7.5GHz+/-2.5GHz detuning range, which indicates that the asymmetrically filtered CS-RZ DPSK signal was robust
against center frequency fluctuations of the bandlimiting filter. Next, in order to investigate the characteristics of
bandlimited CS-RZ DPSK signals further, we conducted numerical simulations and the results are summarized in
Fig. 2. In the simulation, the bandlimiting filter was assumed to 4th-order Gaussian with a 3dB bandwidth of 45GHz,
and the performance was evaluated by eye-opening penalty. As shown in Fig. 2 (a), the simulation results agree well
with the measured results. In order to find out the source of superior performance of asymmetrical filtering over
symmetrical filtering, we investigated the signal waveform after the bandlimiting filter, as shown in Fig. 2(b) and (c).
From Fig. 2(b) and (c), we found that the difference in signal power fluctuations between symmetrically and
asymmetrically filtered signals can be attributed to the additional Q-factor penalty of symmetrically filtered signal.
2.5 6
(a) Simulation
Experiment 5
2
4
1.5
3
1
2
0.5
1
0 0
0 4 8 12 16
Filter Detuning [GHz]
Fig. 2. Q-factor penalty due to bandlimitation as a function of filter detuning (a) and optical signal waveform of (b) symmetrically and (c)
asymmetrically filtered (7.5GHz detuning) 42.7 Gbit/s CS-RZ DPSK signals.
3. Tolerance evaluation of pre-filtered CS-RZ DPSK signals

Next, we evaluated the dispersion and nonlinear tolerance by measuring the penalty due to dispersion compensation
error at the receiver and the increase in fiber input power for unfiltered, symmetrically filtered and asymmetrically
filtered CS-RZ DPSK signals. The filter detuning was set to 7.5GHz for asymmetric filtering. In the evaluation of
dispersion tolerance, the signal from the transmitter was directly fed into the DCFs with various total dispersion
values. In the evaluation of nonlinear tolerance, we used a transmission line consisting of 50km of SMF followed by
a DCF, which fully compensated for the accumulated dispersion of the SMF. Fig. 3(a) and (b) show the Q-factor
penalty measured as a function of dispersion compensation error and fiber input power, respectively. The circles,
triangles, and squares show the Q-factor penalty obtained from the experiment, and the solid, dot-dashed, and
dashed lines show the eye opening penalty obtained from the numerical simulations for unfiltered, symmetrically
and asymmetrically filtered signals, respectively. The acceptable dispersion compensation error for 1dB Q-factor
penalty was found to be about 50ps/nm for the unfiltered and asymmetrically filtered signal and 140ps/nm for the
symmetrically filtered signal. Note that this large dispersion error tolerance of symmetrically filtered CS-RZ-DPSK
signal was not observed for the CS-RZ OOK signal [10] and seems a specific feature of optically pre-filtered CS-RZ
DPSK signal. Regarding the nonlinear tolerance, for all the filtering conditions, almost no performance degradation
was observed up to the fiber input power of +10dBm. However, as the fiber input power exceeded +10dBm, distinct
penalty appeared in the order of asymmetrically, symmetrically filtered and unfiltered signal. These phenomena
were the same as the case for CS-RZ OOK signal [10], and this indicates that there is a trade-off between the
back-to-back performance and nonlinear tolerance for asymmetric optical bandlimitation, as in the case of CS-RZ
OOK signal [7][10].
TuF2
4 4 4 8
3 (a) 3 3 (b) 6
2 2 2 4
1 1 1 2
0 0 0 0
-1 -1 -1 -2
-200 -150 -100 -50 0 50 100 150 200 -5 0 5 10 15 20 25
Dispersion Comensation Error [ps/nm] Fiber Input Power [dBm]
Fig. 3. Tolerance against (a) dispersion compensation error and (b) fiber input power of unfiltered (circles and solid lines), asymmetrically filtered
(7.5GHz detuning, squares and dot-dashed lines) and symmetrically filtered (triangles and dashed lines) 42.7 Gbit/s CS-RZ DPSK signals. Points
and lines show the experimental and numerical simulation results, respectively.
Finally, in order to confirm the long-haul transmission that asymmetric filtering is effective when the fiber
performance of symmetrically and asymmetrically nonlinearity per repeater span is well suppressed.
pre-filtered CS-RZ DPSK signals, we conducted a
50GHz-spaced 64 x 42.7Gbit/s transmission
12.5
experiment up to 9000km by using a 360km-long
-6dBm/ch
re-circulating loop with 43km-long dispersion-flattened 12
fiber spans [7] and 980nm-pumped EDFA repeaters. -5dBm/ch
The odd and even channels were orthogonally 11.5
polarization multiplexed. We measured the 11

transmission performance of the channel 32, the center
channel, as a function of transmission distance. Fig. 4 10.5
shows the results for -6dBm/ch and -5dBm/ch repeater
10
output power with symmetric and 3.8GHz-detuning
Symmetrically filtered signal
asymmetric filtering. For 6000km transmission at 9.5 Asymmetrically filtered signal
–5dBm/ch repeater out-put, the Q-factor calculated
from the obtained BER was 11.1dB and 11.7dB for 9
5000 6000 7000 8000 9000 10000
7.5GHz and 3.8GHz detuning asymmetrically filtered
signal, 12dB for symmetrically filtered signal. On the Distance [km]
other hand, for 9000km transmission at –6dBm/ch Fig. 4. Long-haul transmission performance of 42.7 Gbit/s
repeater out-put, the Q-factor calculated from the symmetrically and asymmetrically (3.8GHz detuning) filtered CS-RZ
obtained BER was 9.6dB and 9.7dB for 7.5GHz and DPSK WDM signals. The solid and dashed lines show the results for
3.8GHz detuning asymmetrically filtered signal, 9.3dB -6dBm/ch and -5dBm/ch repeater output, and the circles and triangles
show the results of symmetrically and asymmetrically filtered
for symmetrically filtered signal. These results show signals.
4. Conclusion
We have experimentally and numerically investigated the impact of asymmetric pre-filtering for 42.7 Gbit/s CS-RZ
DPSK signals. By detuning the bandlimiting optical filter, the back-to-back performance was distinctively improved
with sufficient robust against center frequency fluctuations of the bandlimiting filter, although the nonlinear
tolerance was decreased. Through this study, we have found that asymmetrically pre-filtering is effective for
quasi-linear DWDM system applications.
5. References
[1] T. Tsuritani et al., ECOC2000, PD1.5, (2000). [6] H. Gnauck et al., Photon. Tech. Lett., 15, pp467-469, (2003).
[2] G. Charlet, et al., OAA2002, PDP1, (2002) [7] T. Tsuritani et al., OFC2003, FE4, (2003).
[3] T. Tsuritani et. ai., ECOC2002, 9.1.4, (2002) [8] B. Zhu et al., OFC2003, PD19, (2003).
[4] G. Charlet, et al., ECOC2002, PD4.1, (2002). [9] T. Tsuritani et al., OFC2003, PD23, (2003).
[5] I. Morita, et al., ECOC2002, PD4.7, (2002). [10]A. Agata et al., OFC2003, MF78, (2003).
TuF3
Nonlinear tolerance of differential phase shift keying

modulated signals reduced by XPM
B. Spinnler
Siemens AG, Corporate Technology, Otto-Hahn-Ring 6, D-81739 Munich Germany
E-mail: Bernhard.Spinnler@siemens.com
N. Hecker-Denschlag, S. Calabrò, M. Herz, C.-J. Weiske, E.-D. Schmidt

Siemens AG, ICN Optical Solutions, Hofmannstrasse 51, D-81359 Munich Germany
D. van den Borne, G.-D. Khoe, H. de Waardt

COBRA Institute, Eindhoven University of Technology, The Netherlands
R. Griffin, S. Wadsworth,
Bookham Technology PLC., Caswell, Towcester, Northants, NN12 8EQ England
Abstract: We show that in order to maintain the high nonlinear tolerance of the DQPSK
modulation format, XPM from neighboring OOK-modulated channels must be avoided. The
negative impact on 10Gb/s DQPSK channels is higher than at 20Gb/s.
© 2003 Optical Society of America
OCIS codes: (060.4510) Optical communications; (060.5060) Phase modulation
1. Introduction
Advanced phase modulation formats have been very successful in extending the reach and capacity of ultra long
haul, high capacity, optical transmission systems [1,2]. This success is not only due to the use of balanced receivers
for a gain of 3dB at the receiver, but also that D(Q)PSK modulation is less sensitive to non-optimum dispersion
compensation [3] and also allows for a larger input power before nonlinear effects degrade the signal quality [4,5].
While the high nonlinear tolerance of systems running solely with D(Q)PSK can be exploited for extending the
reach of high capacity systems or increasing amplifier spacing, the effect of mixed usage, i.e. combining OOK and
D(Q)PSK modulation formats at different wavelengths, on the individual channel tolerances still needs to be
determined. For instance mixed usage can occur when installed systems are upgraded with new modulation formats
or when meshed networks come into operation. In such scenarios, wavelength channels with both D(Q)PSK and
OOK modulation formats can become nearest neighbors and the nonlinear interactions must be understood. The
robustness of binary DPSK in presence of NRZ modulated channels with 100 GHz spacing for transmission over
SSMF has been discussed in [6]. In this paper we discuss the reduction of the nonlinear tolerance of NRZ-DQPSK
because of XPM from NRZ-OOK neighboring channels at a 50 GHz spacing. 50 GHz spacing is used instead of 100
GHz spacing as it represents the standard today for OOK systems and is more likely to be limited by XPM effects.
2. Experimental setup
The two experimental setups shown in fig. 1a and 1b were used to compare the transmission performance of five
wavelength channels on a 50 GHz grid over 100 km SSMF. The middle channel was always modulated with a 20
Gb/s DQPSK (10 Gsymbol/s). In the first setup (fig. 1a) the four remaining channels were also 20 Gb/s DQPSK
modulated. In the second setup (fig. 1b) these channels were 10 Gb/s OOK modulated. We chose half the bit rate for
the OOK channels in order to have similar spectral width of the two types of formats. Therefore, this bit rate seems a
“natural” choice considering 10 Gb/s OOK systems in operation today. The DQPSK modulation was achieved with a
GaAs-based modulator [4] which is composed of two Mach-Zehnder interferometers (MZI) offset biased around the
zero transmission point so that with a 10 Gb/s modulation, a phase difference between bits of ʌ is obtained. Each of
the MZI are modulated with the same pseudo-random bit sequence with a length 211-1, but a delay of 22 bits between
the two allows for a decorrelation of the bits, since a precoder was unavailable. The two MZIs are combined with an
added ʌ/2 phase shifter in a ‘super’ MZI structure. We used a tunable dispersion compensator in order to optimize
performance. Full compensation was found to be optimal for DQPSK. The total input power of the five wavelength
channels was varied before the transmission fiber to study the effect of the XPM. At the end of line, we used a 0.2
nm passband filter to extract the channel under consideration. Differential demodulation was performed by an optical
one-symbol delayed MZI. The two outputs of the demodulator were differentially detected with a balanced receiver
and a limiting differential amplifier. In order to measure a BER, the receiver is given the corresponding bit sequence
TuF3
for detection of either the I or Q channel. At the output of the differential amplifier, a noise loading experiment was
performed to determine the BERs of the middle channel as a function of the per channel power and OSNR.
In the mixed modulation setup we additionally adjusted the polarization of the neighboring channels so that they
were either parallel or orthogonal to the middle DQPSK channel under consideration in order to better observe the
XPM effect. In the DQPSK-only setup all channels had an identical polarization. All channels have equal power
independent of the modulation format.
a) O1
O1 100 km DCF b)
O2
VOA SSMF 100% MZI
O2 100 km DCF
VOA SSMF 100%
O3 O4
O4 QPSK Mod O5 3 dB
Coupler
O5 2.8 nm
O3 2.8 nm
QPSK Mod
Integrated device
MZ-Demod 0.2 nm
+ Integrated device
MZ-Demod 0.2 nm
-
+
-
Fig. 1. Experimental setup with a) 5 DQPSK channels, b) one DQPSK channel and 4 OOK channels.
3. Results
In order to assess the effect of XPM on the middle DQPSK channel, we made measurements and simulations for
both systems shown in fig 1. We evaluated the BER of the DQPSK channel for various values of the transmitted
power. Fig. 2 shows the measured BER versus OSNR for total input powers ranging from 7 to 18.5 dBm. The figures
in the first row show results obtained by measurements, the figures in the second row show the corresponding
simulation results. The figures in the left column (a and d) present the BER for the system with five DQPSK
channels, the figures in the middle column (b and e) show the BER for the system with one DQPSK channel and four
OOK neighbors (all polarizations aligned parallel), and the figures in the right column (c and f) show the results for
the system with one DQPSK channel and four OOK neighbors (polarizations of the OOK channels orthogonal to that
of the middle DQPSK channel).
Let us first discuss the measurement results. It can be seen that in all cases the higher the launch power the more
the performance is degraded by XPM from the neighboring channels. However, the results differ widely if we
compare them quantitatively. When we increase the launch power from 7 to 16 dBm the OSNR penalty for the
DQPSK-only system is about 2.5 dB at BER= 10–9. If the polarization of the OOK neighboring channels is parallel to
the DQPSK channel the penalty is 2 dB for 7 dBm launch power and the penalty for 16 dBm launch power could not
be determined at 10–9 because the signal is significantly degraded. When the polarization of the OOK neighboring
channels is orthogonal to the DQPSK channel, the performance is equal to that of the DQPSK-only system for 7
dBm launch power, and the OSNR penalty for 16 dBm launch power is about 6 dB.
The corresponding simulation results are shown in the second row of fig. 2. BERs above 10-5 were measured
directly using Monte-Carlo simulation. For lower BERs we employed the tail extrapolation technique. While the
results differ quantitatively from the measurements because we did not include all relevant impairments into the
simulation (e.g. in the simulation we did not include laser phase noise and used ideal filters, de/modulators and clock
recovery), the general trend in the measurements and the simulations is the same. We observe only a slight
degradation for an increase in the launch power when we use DQPSK for the neighboring channels. If we use OOK
with parallel polarization with respect to the middle DQPSK channel for the neighboring channels, the degradation is
dramatic. If we use OOK with orthogonal polarization with respect to the middle DQPSK channel, the degradation is
still larger than in the DQPSK-only system, but far less than in the case with parallel polarized OOK neighbors.
These results are in agreement with our measurements and support our claim that the performance of DQPSK
depends very much on the type of the modulation format and polarization of the DQPSK’s neighboring channels.
A prevalent measure to further enhance the dispersion tolerance of DQPSK is the reduction of the data rate. In
order to investigate the influence of the data rate we repeated the simulations corresponding to the setup shown in
fig. 1b with the DQPSK data rate reduced to 10 Gb/s. The neighboring channels again use 10 Gb/s OOK with the
same polarization (worst case). Fig. 3 compares the simulation results for the cases 20 Gb/s (fig. 3a) and 10 Gb/s
(fig. 3b). While for low input power there is the usual 3 dB gain in favour of the 10 Gb/s system, we observe that for
higher input power the 10 Gb/s system performs even worse than the 20 Gb/s system. Hence, the XPM from OOK
neighbors is more detrimental for the 10 Gb/s channel than for the 20 Gb/s channel. This effect has to be taken into
account when an OOK channel is to be replaced by a “more tolerant” DQPSK channel with the same data rate.
TuF3
a) -4
10
P = 7 dBm
P = 12 dBm
b) -4
c) -4
10
P = 16 dBm 10
-5 P = 18.5 dBm -5
10 10
-5
10
-6 -6
BER
10
BER
10
BER
-6
10
-7 -7
10 P = 7 dBm 10 P = 7 dBm
-7
10 P = 12 dBm P = 12 dBm
-8
10
P = 14 dBm -8 P = 14 dBm
10
-8
P = 15 dBm P = 16 dBm
10 P = 16 dBm P = 18 dBm
-9 -9
10 10
12 14 16 18 20 22 24 26 12 14 16 18 20 22 24 26 12 14 16 18 20 22 24 26
10
0
10
0
OSNR 10
0
OSNR (dBm)
d) 10
Ŧ1
OSNR (dBm)
P = 7 dBm
P = 14 dBm
P = 18 dBm
e) 10
Ŧ1
f) 10
Ŧ1
Ŧ2 Ŧ2 Ŧ2
10 10 10
Ŧ3 Ŧ3 Ŧ3
10 10 10
Ŧ4 Ŧ4 Ŧ4
10 10 10
BER
BER
BER
Ŧ5 Ŧ5 Ŧ5
10 10 10
Ŧ6 Ŧ6 Ŧ6
10 10 P = 7 dBm 10 P = 7 dBm
P = 10 dBm P = 10 dBm
Ŧ7 Ŧ7 Ŧ7
10 10 P = 12 dBm 10 P = 12 dBm
P = 14 dBm P = 14 dBm
Ŧ8 Ŧ8 Ŧ8
10 10 P = 16 dBm 10 P = 16 dBm
P = 18 dBm P = 18 dBm
Ŧ9 Ŧ9 Ŧ9
10 10 10
5 10 15 20 25 5 10 15 20 25 5 10 15 20 25
OSNR [dB] OSNR [dB] OSNR [dB]
Fig. 2. BER of middle DQPSK channel versus OSNR. a) and d) 5 DQPSK channels, b) and e) DQPSK and 4 OOK
channels (polarization of all channels parallel), c) and f) DQPSK and 4 OOK channels (polarization of OOK
channels orthogonal to middle DQPSK channel). a) – c) Measurements, d) – f) Simulations.
0 0
10 10
20 Gb/s P = 7 dBm 10 Gb/s P = 7 dBm
a) 10
Ŧ1
Ŧ2
P = 10 dBm
P = 12 dBm b) Ŧ1
10
Ŧ2
P = 10 dBm
P = 12 dBm
10 P = 14 dBm 10 P = 14 dBm
Ŧ3 Ŧ3
10 10
Ŧ4 Ŧ4
10 10
BER
BER
Ŧ5 Ŧ5
10 10
Ŧ6 Ŧ6
10 10
Ŧ7 Ŧ7
10 10
Ŧ8 Ŧ8
10 10
Ŧ9 Ŧ9
10 10
5 10 15 20 25 5 10 15 20 25
OSNR [dB] OSNR [dB]
Fig. 3. Performance of DQPSK with 4 10 Gb/s OOK neighboring channels (polarization of all channels parallel). The
DQPSK channel runs at 20 Gb/s (a) and 10 Gb/s (b), respectively.
4. Conclusions
Since practically all systems today use OOK modulation, the tolerance to interference generated by neighboring
OOK channels will become a major criterion for the deployment of optical alternative modulation schemes. In this
paper we addressed the, for practical reasons, very important case of 10 and 20 Gb/s DQPSK channels being turned
on in a 50 GHz WDM grid nearby 10 Gb/s OOK channels. We showed by measurements and simulations that a
middle 20 Gb/s DQPSK channel surrounded by four OOK channels suffers from a very large OSNR penalty even for
moderate launch powers when the polarization of the neighboring channels is aligned parallel to the middle channel.
Part of this degradation is recovered when the polarization of the neighboring channels is orthogonal to the middle
DQPSK channel. But this cannot be guaranteed unless polarization interleaving is employed at the transmitter site.
Furthermore, we showed by means of simulations that a 10 Gb/s DQPSK channel is even more vulnerable to the
effect of XPM from OOK neighboring channels for moderate to high launch power. However, elaborate dispersion
maps can be chosen such that XPM interference is minimized. A verification by measurements and thorough
investigation of this issue represent an open field for future work.
5. References
[1] C. Rasmussen et al. “DWDM 40G transmission over trans-Pacific distances (10,000km) using CSRZ-DPSK, enhanced FEC and all-Raman
amplified 100 km UltraWave fiber spans”, OFC 2003, PD18.
[2] B. Zhu et al. “6.4-Tb/s (160 x 42.7 Gb/s) transmission with 0.8 bit/s/Hz spectral efficiency over 32 x 100 km of fiber using CSRZ-DPSK
format “, OFC 2003, PD19.
[3] H. Bissessur et al. “1.6 Tb/s (40 x 40 Gb/s) DPSK transmission with direct detection”, ECOC 2003, paper 8.1.2.
[4] R.A. Griffin et al. “10 Gb/s optical differential quadrature phase shift key (DQPSK) transmission using GaAs/AlGaAs integration” OFC 2002
FD6.
[5] C. Wree et al. “RZ-DQPSK format with high spectral efficiency and high robustness towards fiber nonlinearities”, ECOC 2002, paper 9.6.6.
[6] M. Rohde et al., “Robustness of DPSK direct detection transmission format in standard fibre WDM systems” , Electron. Lett. Vol. 36, No.
17, Aug. 2000, pp. 1483-4.
TuF4
Experimental comparison of non linear threshold and

optimum pre dispersion of 43 Gb/s ASK and DPSK formats
R. Dischler, A. Klekamp, J. Lazaro, W. Idler
Alcatel Research and Innovation, Holderaeckerstrasse 35, D-70499 Stuttgart, Germany
Roman.Dischler@alcatel.de
Abstract: 43Gb/s ASK and DPSK formats, each with 4 duty-cycles, are compared on single
channel nonlinear impairments. We measured the nonlinear threshold, pre-dispersion optimum
and tolerance and found a NLT benefit for DPSK up to 2dB.
2003 Optical Society of America

OCIS codes: (060.2330) Fiber optics communications, (060.4080) Modulation
1. Introduction
Comprehensive experimental analysis of modulation formats regarding system performance and tolerances, linear
and nonlinear transmission impairments are needed for cost effective system design. Comparisons of linear fiber
impairments at on-off keying (OOK) and differential phase shift keying (DPSK) formats at 43 Gb/s have been
reported [1,2] showing the advantages of DPSK formats: 3 dB of OSNR gain, enhanced tolerances on chromatic
dispersion and differential group delay by balanced detection. The duty cycle has been identified as beneficial
parameter for NRZ format and RZ formats differentiation [2]. On one hand, simulations show that 43Gb/s DPSK
formats benefit also from reduced nonlinear effect [3,4], but on the other hand, nonlinear phase noise has been
identified to effect at DPSK formats adversely, potentially reducing the gain achieved by the use of a balanced
receiver, at 10 Gb/s even more severe than at 40 Gb/s [5]. However, DWDM experiments clearly show the benefits
of DPSK formats at 40 Gb/s [6,7] and also at 10 Gb/s [8,9] achieving impressive submarine transmission distances.
Concerning single channel nonlinear transmission impairments key system design parameters are the nonlinear
phase shift [10,11] or the nonlinear threshold (NLT), in other words: the maximum transmission fibre input power,
varying inversely proportionally with span number [11].
In this paper we present an experimental survey of NLT, including the optimum pre dispersion value and pre
dispersion tolerance for single channel 43 Gb/s formats. RZ and NRZ modulation at ASK and DPSK formats by
transmission experiments over 1 and 3 spans of SMF. We compare systematically the modulation formats across
the different RZ duty cycles ranging from 33%, 50% to 66% (CSRZ) and finally 100% is applied for NRZ.
2. Experimental set up
Fig. 1 shows the experimental set up for the measurement of NLT of the 8 modulations formats [2] for the
transmission over 1 and 3 spans of SMF, with a span length of 80km.
Pre Post
80km
DCF SMF DCF ASE-
Filter
TX ~ RX
~
~
Att. Att.
P OSNR RZ-33% RZ-50%

Pre Inline Inline Post
80km 80km 80km
DCF SMF DCF SMF DCF SMF DCF ASE-
Filter
TX ~ RX
~
~
Att. Att. Att. Att.
RZ-66% NRZ (100%)
P1 P2 P3 OSNR
Fig. 2: 43Gb/s DPSK eye-pattern (experimental,
Fig. 1. Set up for NLT evaluation over 1 span (top) and 3 spans(bottom) of SMF back-to-back, PRBS 2 23 -1, balanced detection), all 4
duty cycles; hor: 10ps/div, vert.: different scaling
With 6 Mach-Zehnder modulator configurations at the transmitter (Tx) we realized 8 modulation formats:
4xASK and 4xDPSK [2], each with 3 different RZ duty cycles of 33%, 50%, 66% (CSRZ) and NRZ modulation,
which we will denote as 100% duty cycle. Examples of 4xDPSK format eye-pattern are shown in Fig. 2 at back-to-
back configuration, showing a good quality of the DPSK Tx. All experiments were carried out with one single
optical channel at 1552nm.
TuF4
The Tx is attached to a double stage booster EDFA with interstage dispersion compensating fiber (DCF) for pre
compensation. This booster EDFA has a fixed optical output power of 18.5 dBm. The input power to the SMF, can
be varied with optical attenuators. The launched power is controlled via 20dB tap by an optical spectrum analyzer
(OSA) at the span input. For the 3 span experiment this configuration is repeated 3 times. The inline dispersion
compensation is set to 105%. At the end of the last span, a 4th attenuator is used to vary the received optical power,
i.e. to vary the OSNR at the receiver (Rx), measured with an OSA of 1 nm resolution bandwidth. The DCF of the
post compensation could be varied in steps of about 10 ps/nm to optimize the residual dispersion at the Rx-side. The
input power into all DCF was maintained below 0 dBm, to ensure linear propagation through the DCFs. For the
DPSK formats a Mach-Zehnder DPSK-demodulator was inserted in front of the receiver. BER measurements are
performed here with single ended detection for the DPSK formats.
From the measured OSNR value a corrected OSNR for 0.1nm bandwidth was calculated. OSNR penalty was
measured with respect to BER of 10-5 using PRBS length of 223-1 for all measurements. As non linear threshold we
denote the fiber input power, which gives a penalty of 1dB with respect to OSNR sensitivity at linear transmission.
To find the optimum pre compensation for each format , the DCF in the booster was varied from 0 to -420
ps/nm in steps of about 100ps/nm. For each pre compensation the post compensation was optimized at an SMF
input power of 0dBm, i.e. in the linear regime. When the fiber input power was increased, only the decision
threshold and the sample phase at the receiver was adapted.
3. Results and discussion

Fig. 3 show the results of the NLT measurements for ASK and DPSK, e.g. with RZ modulation of 50% duty cycle.
The measurement uncertainty for the NLT value is about 0.5dB. To determine the optimum pre dispersion, the
maximum NLT and the pre dispersion tolerance with respect to NLT, we used a parabolic fit over the measured
NLT data.
16 16 DPSK - RZ50%
15 ASK - RZ50%
1 span 15
1 span
14 14
13 13
NLT (dBm)
NLT (dBm)
12 12
11 11
10 10
3 spans
9 9
8 8 3 spans
7 7
6 6
-500 -400 -300 -200 -100 0 -500 -400 -300 -200 -100 0
Precompensation (ps/nm) Precompensation (ps/nm)
Fig. 3. Non linear threshold (NLT) vs. variation of p re dispersion for 1 and 3 spans over SMF, ASK at left, DPSK at right side
16 16
15 ASK DPSK
15
14 14
13 13 1 span
1 span
NLT (dBm)
NLT (dBm)
12 12
11 11
10 10
9 9
8 8
3 spans 3 spans
7 7
6 6
33 50 66 83 100 33 50 66 83 100
duty cycle (%) duty cycle
Fig.4:.Maximum non linear threshold vs. duty-cycle for 1 and 3 spans over SMF, ASK at left, DPSK at right side
Fig. 4 summarizes the maximum NLT data from parabolic fit, we obtained for 4xASK and 4xDPSK formats by 1
and 3 span measurements. The difference for NLT data between 1 and 3 spans is about 5 dB, indicating that 105%
of inline compensation was a suitable value for all formats, since the expected difference between 1 and 3 span
NLT data is 4.8dB according to a log(1/N)-correlation [10], where N is the number of spans.
DPSK formats show higher NLT than the corresponding ASK formats, except at RZ 66%: DPSK benefits by
2dB at NRZ format, and up to 1 dB at RZ formats. The highest NLT of 15dBm at 1 span configuration was
observed with DPSK-RZ33%.
TuF4
Further measurements under post dispersion optimisation at each fiber input power yield even higher NLT
values than shown in fig. 4, e.g. NLT of ASK-NRZ could be increased by 2 dB.
0 500 0 500
ASK 3 spans
DPSK
-50 -50
400 400
pre-dispersion tolerance
1 span
pre-dispersion tolerance
pre-dispersion (ps/nm)
pre-dispersion (ps/nm)
-100 -100
1 span 300 3 spans 300
(ps/nm)
(ps/nm)
-150 -150
3 spans
200 200
-200 -200
1 span 3 spans
100 100
-250 -250
1 span
-300 0 -300 0
33 50 66 83 100 33 50 66 83 100
duty cycle (%) duty cycle (%)
Fig.5. Optimum pre- dispersion (solid lines) and pre dispersion tolerance (dashed lines) vs. duty -cycle for 1 and 3 spans over SMF,
ASK at left, DPSK at right side
Fig. 5 shows the optimum pre dispersion and the pre dispersion tolerance of NLT for all 8 modulations formats,
4xASK at left and 4xDPSK at right side. The optimum pre dispersion of ASK formats is about -200 ps/nm with low
variation over the duty-cycles. The positive shift of the optimum pre dispersion value for 3 spans is attributed to the
residual inline dispersion of about –70ps/nm [12]. At DPSK formats we observed a more pronounced shift of
optimum pre dispersion for 3 spans and a higher variation across the duty cycles than at ASK formats.
The pre dispersion tolerance, i.e. the full width 1dB down of the parabolic fit in Fig. 1, where NLT is 1dB
below maximum, is within 300 and 400 ps/nm for all ASK and DPSK formats. We received no clear trend within
the ASK (but small increase with duty cycle) and the DPSK formats (but small decrease with duty cycle). Note, that
the NLT measurement accuracy of 0.5dB yield an pre dispersion tolerance accuracy of 50ps/nm.
4. Summary
We have experimentally evaluated single channel nonlinear threshold for transmission over SMF with ASK and
DPSK formats with 4 different duty cycles from 33% to 100% (NRZ). At NRZ format, DPSK show 2dB higher
NLT. At RZ formats, the NLT gain by DPSK is up to 1dB. Highest NLT of 15dBm for 1 span configuration was
observed with DPSK-RZ33%. We confirmed a log(1/N) correlation of NLT from 1 and 3 spans measurement result.
The optimum pre dispersion was found around -200 ps/nm across all formats, with a higher variation over duty
cycles for DPSK formats than for ASK formats. The pre dispersion tolerance is comparable for all formats between
300 and 400ps/nm with no clear trend over the duty cycles.
5. References
[1] F. An et al, “Comparison of linear fiber impairments tolerance among 40Gb/s modulation formats”,
in Proc. OFC 2003, vol.2, paper FE2, pp. 657-658
[2] W.Idler et al, “System performance and tolerances of 43Gb/s ASK and DPSK modulation formats”, in Proc. ECOC 2003, paper Th2.6.3
[3] O. Vassilieva et al, “Numerical comparison of NRZ, CS-RZ and IM-DPSK formats at 43Gb/s WDM transmission”, in
Proc LEOS 2001, pp 673-674
[4] T. Hoshida et al, “Optimal 40Gb/s modulation formats for spectrally efficient long -haul DWDM systems”, Journal of Lightwave
Technology, vol. 20, 2002, pp. 1989-1995
[5] H. Kim, A. H. Gnauck, “Experimental investigation of the performance limit ation of DPSK systems due to nonlinear phase noise”
Phot. Techn. Lett. Vol.15 -2, 2003, pp 320-322
[6] A. H. Gnauck et al, “2.5 Tb/s (64x42.7) transmission over 4x100km NZDSF using RZ -DPSK formats and all Raman amplified spans”,
in Proc. OFC 2002, postdeadline paper FC2
[7] C. Ramussen et al., “DWDM 40G transmission over trans-Pacific distance (10,000 km) using CSRZ-DPSK, enhanced FEC and all-Raman
amplified 100 km UltraWave fiber span”, in Proc. OFC 2003, postdeadline paper PD18
[8] G. Vareille et al., “8370km with 22dB spans ULH transmission of 185x10.709 Gb/s RZ -DPSK channels”, in Proc. OFC 2003, postdeadline
paper PD20
[9] J.X. Cai et al, “A DWDM demonstration of 3.73 Tb/s over 11,000km using 373 RZ -DPSK channels at 10 Gb/s”, in
Proc. OFC 2003, postdea dline paper PD22
[10] G. P. Agrawal, “Nonlinear fibre optics”, second edition, Academic Press
[11] J.-C. Antona et al., “Nonlinear cumulated phase as a criterion to assess performance of terrestrial WDM systems”, in
Proc. OFC 2002, paper WX5
[12] Y. Frignac et al, “ Numerical optimization of pre- and in-line dispersion compensation in dispersion-managed systems at 40 Gbit/s”, in
Proc. OFC 2002, paper ThFF5
TuG1
Electronic Dispersion Compensation for Extended Reach

G. S. Kanter, A. K. Samal, and A. Gandhi
Santel Networks, 39899 Balentine Dr. #350, Newark, CA. 94560
510-252-4301 (phone), 510-498-4215 (fax), agandhi@santelnetworks.com
Abstract: We investigate the benefits of using post-detection electronic processing in high-speed optical links. In
particular, electronic post-processing is studied in conjunction with other advanced methods, such as optical duo-
binary encoding and chirped modulation.
2003 Optical Society of America
OCIS codes: (060.2330, 060.2340, 230.0250, 230.5160)
1. Introduction
Electronic dispersion compensation, or more generally electronic post-detection processing (EPP), which is typically
based on feed-forward equalizers (FFE) and/or decision feedback equalizers (DFE), are known to increase the
inherent tolerance of a receiver to various types of impairments such as chromatic dispersion (CD), polarization
mode dispersion (PMD), differential mode dispersion (DMD), and low-pass filtering [1-4]. EPP has no optical
insertion loss and it promises to be both compact and inexpensive. Because EPP is a post-detection process, it can be
used in addition to optical techniques [5,6]. For example, EPP may be useful for combating chromatic dispersion
even in long-haul links, which also require periodic optical methods of dispersion compensation, in order to relax
the specifications of the optical methods. EPP may also replace optical methods completely in some cases. For
instance, although optical PMD compensation can be very effective [7], it is too bulky and expensive to find
applications in current systems. Thus EPP will likely be the first line of defense for PMD troubled links.
Other performance enhancing techniques such as chirped modulation, optical duo-binary (ODB) modulation, and
forward-error correction (FEC) coding can be used in conjunction with EPP to form more robust and effective
solutions. We will analyze the performance of EPP using chirped modulators, ODB modulation, and FEC encoding.
Additionally, we will highlight the potential of using EPP with low-cost components, such as an integrated
laser/electro-absorption modulator (EAM) designed for use in lower speed systems, in order to allow cost savings in
shorter-reach applications.
2. Discussion
Mach-Zehnder interferometer (MZI) modulators and electro-absorption modulators can be designed to have various
chirp characteristics [8,9]. While chirping will not enhance the dispersion tolerance (amount of variation in
dispersion which is tolerable), it can be chosen to partially compensate for the known dispersion of the transmission
fiber. Simply replacing an un-chirped modulator with a chirped design can increase the acceptable amount of mean
dispersion by more than 25%. We have simulated typical links using modulators of various chirps with and without
EPP as shown in Fig 1. The simulation parameters are described in [4]. All the EPP simulations use an eight tap
(two taps per symbol) FFE and one tap DFE architecture, with a least-mean square updating algorithm [4]. The
benefit of adding EPP, when measured as the increase of positive dispersion allowable for a given power penalty,
decreases as the chirping factor is optimized. However, there is still a clear benefit to adding EPP to chirped
modulation.
ODB is a more complex solution than chirping because pre-coding and filtering is required before the MZI [10].
Additionally, higher driving voltages are typically needed. However, a very substantial increase in CD tolerance
(over a factor of 2) can be expected, as well as other potential benefits such as higher spectral efficiency. This makes
ODB interesting for extending the metro-scale distance which can be covered without adding costly optical
dispersion compensation, as well as for reducing issues associated with dispersion slope compensation in long-haul
systems. The duo-binary data was simulated by placing a 2.5GHz (10th order) electrical Bessel filter after the
modulator driver. The simulations were done using a PRBS input sequence without any pre-coder. Such a
configuration is acceptable for Q-factor analysis, and even for laboratory measurements of BER, while a pre-coder
is needed for transmitting actual data [10]. Our simulations suggest (see Fig. 1) adding EPP leads to about a 15%
reach increase (at the 2dB penalty point), in addition to reducing the back-to-back penalty usually observed when
using ODB modulation [11]. Moreover, the EPP also provides much more robustness to the ODB system, which is
known to be very sensitive to modulator asymmetries [11].
TuG1
Forward error correction (FEC) coding is a technique used to reduce the required signal-to-noise ratio for a given
BER. The extra margin supplied by FEC can offset penalties from dispersion. FEC requires overhead bits, thus
increasing the effective bit-rate. However, the amount of dispersion typically allowable in an optical link is
quadratically related to the inverse bitrate. Thus, the FEC coding gain can be significantly reduced, or even
eliminated, depending on the dispersion in the system. We simulated RS(255,239) encoded data and passed it
through an impaired channel. The signal was then detected at the optimal sample and threshold points with hard
decision decoding. For lower bit error rates, the decoder performance was calculated by the ITU specified intrinsic
performance equation [12]. Figure 2 shows simulated performance of uncompensated and EPP compensated links
either at 10Gb/s or with FEC overhead. Using EPP with FEC encoding allows the coding gain to be largely
maintained even with significant amounts of added dispersion. In fact, there is no penalty even after 2000ps/nm
when EPP and FEC are used together (compared to not using either FEC or EPP at the native data-rate).
Additionally, it can be seen that EPP at the native data rate will outperform FEC encoding without EPP after a
certain amount of dispersion (>1550 ps/nm). We expect EPP to be less expensive than FEC encoding, as it is a
simpler component which is only used on the receive side.
Previous work [4] has demonstrated that reasonable performance can be expected from components typically used at
2.5Gb/s rates even at 10Gb/s when EPP is employed. We expand on that work here by experimentally demonstrating
a 50km 9.95Gb/s link using an EAM designed for 2.5Gb/s applications in an un-amplified link. We use an integrated
laser/EAM (Fujitsu FLD5F14CN) which is interfaced to the high-speed data input pins though a hand-made circuit
board. We believe better performance, in particular a reduction of the error floor, could be expected by using a
properly designed high-speed interface. Our equalizer is a 4 tap (1 tap per symbol) FFE with a 1 tap DFE employing
a LMS algorithm. As can be seen in figure 2, we can achieve a bit error rate less than 1e-10 after 50km of SMF-28
fiber using a PIN detector specified for –20dBm (typical) receiver sensitivity (BER=1e-10) near –16.5 dBm input
power. This represents an approximately 3.5dB penalty with respect to our measured performance using a 10Gb/s
MZI modulator in the back-to-back condition. Considering that many commercial 10Gb/s EAMs specified for
intermediate reach have 2dB dispersion penalties near this level of dispersion and that integrated laser/EAMs
typically have poorer intrinsic performance than MZIs (due to, for instance, lower extinction ratios), the penalty for
replacing a typical 10G laser/EAM with a 2.5G version followed by EPP is on the order of only 1dB.
The possibility also exists for increasing the sensitivity of a receiver by combining more sensitive, low bandwidth
receivers with EPP [4]. More sensitive receivers allow for longer un-amplified links and low-bandwidth components
allow for cost savings. Figure 3 shows the performance of a 10 Gb/s MZI transmitter with an APD fabricated for
2.5Gb/s applications followed by our EPP equalizer as a function of data-rate. Also shown is a curve which gives a
rough estimate of how we might expect the sensitivity to vary with data-rate if we had various receivers with the
same bandwidth-to-data-rate ratio as our experimental receiver. The curve assumes that the sensitivity of such
receivers is proportional to the number of received photons-per-bit for all data-rates. For instance, at 5Gb/s we
expect the required optical power for a given BER to increase by 3dB over that required at 2.5Gb/s. The equalized
receiver shows a significant improvement in sensitivity over the constant photons-per-bit model up to about 8Gb/s.
We estimate our receiver (mounted on a homemade circuit board) had a bandwidth of about 2GHz. We expect that
good performance will be observed beyond 10 Gb/s if a 2.5GHz bandwidth receiver is used.
3. Conclusion
Electronic post-detection processing can be used in conjunction with other performance enhancing techniques, such
as chirped modulation, ODB modulation, and FEC encoding to improve system performance. Through simulations
we show that EPP further improves performance in each of these cases. EPP is particularly important when FEC is
employed due to the effectively higher data rate imposed by the FEC overhead. We anticipate that EPP can be used
to extend the reach of short, un-amplified links by increasing the sensitivity of the receiver as well as to reduce cost
by allowing the use of inexpensive, low-bandwidth transmitters. Our experimental data suggests that appropriately
chosen components devised for 2.5Gb/s systems can successfully be used at 10Gb/s if combined with EPP.
REFERENCES
[1] F. Buchali et al., “Reduction of the chromatic dispersion penalty at 10Gbit/s by integerated electronic equalizers,” OFC 2000 vol 3,
268-270 (2000).
[2] H. Bulow et al., “PMD mitigation at 10Gbit/s using linear and nonlinear integrated electronic equaliser circuits,” Electron. Lett. 36,
163-164 (2000).
TuG1
[3] X. Zhao and F.S.Choa, “Demonstration of a 10Gb/s transmissions over a 1.5 km long multimode fiber using equalization techniques,”
IEEE Photon. Tech. Lett. 1187-1189 (2002).
[4] G. S. Kanter et al., "Electronic equalization for enabling communications at OC-192 rates using OC-48 components," Opt. Express 11,
2019-2029 (2003).
[5] J. C. Cartledge and R. G. McKay, “Performance of 10 Gb/s lightwave systems using an adjustable chirp optical modulator and linear
equalization,” IEEE Photon Tech. Lett. 4, 1394-1397 (1992).
[6] H. Bulow, F. Buchali, and G. Thielecke, “Electrinically enhanced optical PMD compensation,” ECOC 2000, 4.2.4 (2000).
[7] R. Noe et al., “Polarization mode dispersion compensation at 10, 20, and 40 Gb/s with various optical equalizers,” IEEE J. of Lightwave
Tech. 7, 1602-1616 (1999).
[8] A. H. Gnauch et al., “Dispersion penalty reduction using an optical modulator with adjustable chirp,” IEEE J. Photon. Tech. Lett. 10,
916-918 (1991).
[9] J. Cartledge and B. Christensen, “Optimum operating points for electroabsorption modulators in 10 Gb/s transmission systems using
nondispersion shifted fiber,” IEEE Photon. Tech. Lett. 16, 349-357 (1998).
[10]W. Kaiser and W. Rosencranz, “Simple precoder for high-speed optical duobinary transmission,” J. of Opt. Commun. 22, 741 (2001).
[11] A. Royset and D. R. Hjelme, “Symmetry requirements for 10-Gb/s optical duobinary transmitters,” IEEE J. of Photon. Tech. Lett. 10,
273-275 (1998).
[12] ITU G.975, “Forward error correction for submarine systems,” Section 7 (2000).
Binary MZI, Chirp = 0

Uncomp at 10Gb/s
Binary MZI, Chirp = - 0.9
With FEC at 10.67 Gb/s
Binary EAM 40 km
Relative Received Power Penalty (dB)
6 Duobinary 25
With EPP at 10Gb/s
With EPP + FEC at 10.67Gb/s
BER = 1e-9
4
OSNR (dB)
20
BER = 1e-12
2
15
10
-2
0 40 80 120 160 200 240 0 500 1000 1500 2000
Distance (km) Chromatic Dispersion (ps/nm)
Figure 1: (a) Left - Sensitivity results for various modulators and receivers at a BER of 1E-12. SMF fiber has a
dispersion coefficient of 17ps/nm.km. Solid (dashed) lines are with (without) EPP. (b) Right- OSNR (measured in
0.1nm) required for a BER of 1E-9 as a function of dispersion for systems with and without FEC and EPP.
10-4 -18
10G MZI (0 km) Constant BER = 1e-9
10-5 2.5G EAM (0 km) Constant photons/bit
2.5G EAM (50 km) -20
10-6
Received Power (dBm)
10-7 -22
BER
10-8
-24
10-9
10-10
-26
10-11
10-12 -28
10-13
-30
-23 -21 -19 -17 -15 2 4 6 8 10
Received Power (dBm) Data Rate (Gb/s)
Figure 2: (a) Left -BER vs. received power for a 10 Gb/s MZI transmitter (diamonds) with no fiber, and an OC-48
specified EAM transmitter after 0km (triangles) and 50km of SMF-28 fiber (squares). (b) Right - Power (dBm)
required to reach a 1E-9 BER as a function of bit-rate for the APD/Equalizer combination (diamonds). Also shown
is a curve with the same sensitivity at 2.5 Gb/s which assumes a constant number of photons per bit are required to
reach 1E-9 at each data rate.
TuG2
Electronic Signal Processing for Differential Phase

Modulation Formats
M. Cavallari, C.R.S. Fludger, P.J. Anslow
Nortel Networks, London Rd, Harlow, CM17 9NA, UK
Tel: +44-(0)1279-402426, Fax: +44-(0)1279-402115, Email: mcav@nortelnetworks.com
Abstract: We present simulations of novel electronic signal processing techniques that may be
used to greatly improve the dispersion tolerance of DBPSK and DQPSK modulation formats by at
least a factor of two.
OCIS codes: (060.2330) Fiber optic communications, (060.2380) Fiber optics sources and detectors.
1. Introduction
In recent years, there has been growing interest in the use of electrical signal processing (ESP) techniques for the
mitigation of optical penalties, such as chromatic dispersion [1] or PMD [2], as they offer the potential for
significant cost savings compared to all-optical approaches. Although this has been a very active research topic,
little has been published on the application of these techniques to the D(B/Q)PSK format, which is currently being
considered as a competitive alternative to ASK for long-haul links [3] or uncompensated metro systems.
In this article, we present simulations of novel ESP techniques that may be used to greatly improve (by a factor
of 2 at least) the dispersion tolerance of DBPSK and DQPSK modulation formats. These techniques include a
Maximum Likelihood Sequence Estimator (MLSE) using balanced detection, an MLSE using dual input detection
(separate inputs from photo-detectors) and an MLSE using joint symbol estimation for DQPSK that makes
decisions on I&Q symbols, taking into account any cross-coupling caused, for example, by chromatic dispersion.
2. Electrical signal processing techniques
The technique at the heart of this work is the MLSE algorithm [4,5], which can be used in different ways, according
to the modulation format. Rather than making a decision based on a single instance using a D-type, the MLSE
searches through a whole sequence of bits, selecting the ‘most likely’ sequence. Mathematically, this is expressed
as selecting the sequence (S) which maximises the probability p(x|S) of generating the received data (x).
For DBPSK, the simplest approach is to use an MLSE with a single input in conjunction with a balanced
detector. The A/D converters may operate at 10 or 20 Gsamples/s, providing either one or two samples/bit (for 10
Gbit/s data). The MLSE algorithm maximises the total log-likelihood probability which may be expressed as the
sum of probabilities over successive bits (k):
ª º
(1)
max log p x | S max ¦ log p x k | S «¦ log p x k | S »
S S
k T
« k T »
¬ 2 ¼
where the term in square brackets may be omitted when using only 1 sample/bit. Rather than using a balanced
detector, the MLSE may process data from the constructive and destructive photodetectors (indicated as d1 and d2):
ª º
max log p x | S max ¦ log p x k | S ¦ log p x k | S «¦ log p x k | S ¦ log p x k | S » (2)
S S
k T ;d1 k T ;d 2
«k T
;d1 k T
;d 2
»
¬ 2 2 ¼
where, once again, the term in square brackets may be omitted when using only 1 sample/bit.
In the case of DQPSK the situation is more complex because there are two independent channels and four
photodiodes. Two independent MLSEs can be applied to the two outputs of the balanced detectors (Fig.1.a). These
would each process the data channel using equation 1. Alternatively, two independent MLSEs can process the
outputs from the constructive and destructive photodiodes (Fig.1.b) using equation 2. The last approach examined
here uses an MLSE that makes joint decisions on the noisy data from the I channel (x) and from the Q channel (y):
ª º
(3)
max log p x, y | S max ¦ log p x k , y k | S «¦ log p x k , y k | S »
S S
k T
« k T »
¬ 2 ¼
where p(xk,yk|S) represents the joint probability of receiving noisy sample xk and yk for a sequence of symbols S.
This takes into account any cross-coupling or distortion between the I and Q channels.
TuG2
Bessel
f ilter Bessel
Bessel f ilter
T f ilter
MLSE
S
MLSE
Joint Symbol MLSE

I data
I data I data
MLSE
Optical T
MLSE
S
f ilter Q data
Q data Q data
Balanced Balanced
5 bit A/D 5 bit A/D 5 bit A/D
detectors 1 or 2 sample/bit detectors
1 or 2 sample/bit 1or 2 sample/bit
DQPSK
Interf erometer a b c
Fig.1. Receiver architecture: DQPSK interferometer followed by (a) single input MLSE using balanced detectors, (b) dual
input MLSE using separate photodiodes, and (c) joint symbol MLSE using balanced detectors.
Here, we use Monte-Carlo simulations to assess the performance of these equalisers and their tolerance to
chromatic dispersion. The DQPSK or DBPSK waveforms, consisting of about 100,000 bits are dispersed, noise
loaded with ASE using optical amplifiers and then detected using the receivers shown in Fig.1. Non-linearity was
not included in these simulations to assess the performance of the equaliser in mitigating the effect of chromatic
dispersion alone. For 10 GBaud data, the optical filter has a Gaussian transfer function with a FWHM of 0.11 nm,
and the electrical low pass filter is a 5th order Bessel filter with a 3 dB bandwidth of 7.5 GHz. The performance is
evaluated in terms of the required OSNR (ROSNR) to achieve an output bit error ratio of 10-3, which may then be
corrected to 10-15 using strong FEC. Due to simulation limitations, the performance of the FEC is not included in
this study.
The A/D converters have 5 bits quantisation resolution (32 levels), and the MLSE algorithms use either 5 bit
state sequences, or 3 symbol state sequences (for joint symbol MLSE). Note that the complexity of the MLSE
scales exponentially with the number of bits or symbols: a 3 symbol joint MLSE has 22x3=64 states, and a 5 symbol
joint MLSE has 22x5=1024 states! In order to keep the computation complexity tractable, we consider a 3 symbol
joint MLSE in the rest of this work.
Probability tables are represented by a lookup table and are trained using separate data from that used for
measurement. The search for the best match is done with the Viterbi algorithm. Finally, MLSEs can operate on data
sampled at the Baud rate or at a multiple of the Baud rate, for higher performance.
3. Numerical results
The results for DBPSK and DQPSK with the different receiver types are shown in figures 2 and 3, which report
only positive values of chromatic dispersion for simplicity. ASK results are reported as well, as a reference.
ASK, D-type ASK, MLSE (2S/B)
DBPSK, dual input MLSE (1S/B) DBPSK, dual input MLSE (2S/B)
DBPSK, single input MLSE (1S/B) DBPSK, single input MLSE (2S/B)
DBPSK, D-type
20
18
ROSNR for BER=1E-3 (dB)
16
14
12
10
4
0 1000 2000 3000 4000 5000 6000 7000
D (ps/nm)
Fig. 2. Required OSNR for a BER of 10-3 vs. chromatic dispersion for 10 GBaud DBPSK and ASK data.
For DBPSK data (see figure 2), the improvements achieved by the MLSE algorithm depend both on the type of
receiver and the number of samples per bit used. The single input MLSE using 1 sample/bit is not particularly
TuG2
effective until the ISI spreads the dispersion over a neighbouring timeslot, at about 2000 ps/nm. The 3 dB tolerance
of 1700 ps/nm when a D-type is used increases by a factor of 2 when a single input MLSE (2 samples/bit) is used,
and by a factor of 3 with a dual input MLSE, which is superior to the achievable tolerance with ASK data and a
2 samples/bit MLSE. The phase distortion due to chromatic dispersion creates some independence in the data
presented at the constructive and destructive photodiodes and results in incomplete extinction and eye closure when
using a balanced detector. The independent data that is lost when using balanced detection, may be used by a dual
input MLSE to provide a better chromatic dispersion tolerance.
3 Symbol MLSE (1S/B) dual input MLSE (1S/B) single input MLSE (1S/B) D-type
3 Symbol MLSE (2S/B) dual input MLSE (2S/B) single input MLSE (2S/B)
26
24
ROSNR for BER=1E-3 (dB)
22
20
18
16
14
12
10
8
0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
D (ps/nm)
Fig 3. Required OSNR for a BER of 10-3 vs. chromatic dispersion for 10 GBaud DQPSK data.
It can be seen in figure 3 that for DQPSK, the MLSE using only a single input is not particularly effective in
compensating for chromatic dispersion. Since the phase difference between symbols is only half that of the DBPSK,
the tolerance for a 3 dB penalty is approximately half that of DBPSK when using a D-type at the receiver. For the
same number of samples per bit, the best results are obtained with a joint symbol estimator, followed by a dual
input MLSE and then by the single input MLSE. This figure also shows that the MLSE with 2 samples/bit always
performs better than with 1 sample/bit. Compared to the simplest form of receiver (a D-type flip-flop) a 3 symbol
joint MLSE with 2 samples/bit can improve the 3 dB chromatic dispersion tolerance by a factor of 3. Even with a
simpler version of this device, operating on 1 sample/bit, an improvement of the 3 dB tolerance by nearly a factor of
2 can be achieved. This improvement can be explained in the following way. Without any distortion in the
transmission path, the data in the I and Q channels of a DQPSK system are independent. Chromatic dispersion
spreads the data over their neighbours, and modifies their phase causing coupling between the I and Q channels. A
joint symbol estimator makes decisions on both I and Q channels, taking into account any cross-coupling due to
chromatic dispersion.
Similar calculations performed at 5 GBaud (10 Gbit/s) indicate that the 3 dB chromatic dispersion tolerance is
as high as 13000 ps/nm when a 3 joint symbol MLSE is used. This is equivalent to 760 km of NDSF or about
4640 km of NZDSF, in the absence of other system limitations such as non-linearities.
4. Conclusion
We have shown that the MLSE algorithm can be successfully applied to D(B/Q)PSK data, for an improvement of
chromatic dispersion tolerance by a factor of 3. This performance is obtained with a dual input MLSE in the
DBPSK case, and with a joint symbol estimation scheme in the DQPSK case. At 10 Gbit/s, the best results are
obtained with DQPSK and a 3 symbol joint MLSE, for a 3 dB tolerance of 13000 ps/nm. This opens the door to
differential PSK links without in-line compensation for both medium and long haul applications.
References
[1] F.Buchali, H.Bülow et al., “Reduction of the Chromatic Dispersion Penalty at 10 Gbit/s by Integrated Electronic Equalisers”, OFC2000,
paper ThS1.
[2] H.F.Haunstein et al., “Design of near optimum electrical equalisers for optical transmission in the presence of PMD”, OFC2001, Paper
WAA4.
[3] A.Gnauck, “40-Gb/s RZ-Differential Phase Shift Keyed Transmission”, OFC 2003, paper ThE1, and references therein.
[4] G.D.Forney, “Maximum-Likelihood Sequence Estimations of Digital Sequences in the Presence of Intersymbol Interference”, IEEE
Transactions on Information Theory, Vol.IT-18, No.3, 363-377 (May 1972).
[5] G.D.Forney, “The Viterbi Algorithm”, Proceedings of the IEEE, Vol.61, No.3, 268-278 (March 1973).
TuG3
Experimental verification of combined adaptive PMD and

GVD compensation in a 40Gb/s transmission using
integrated optical FIR-filters and spectrum monitoring
Marc Bohn, Werner Rosenkranz
Chair for Communications, University of Kiel, Kaiserstr. 2, 24143 Kiel, Germany
e-mail: mbo@tf.uni-kiel.de
Peter M. Krummrich
Siemens AG, Optical Networks, Advanced Technology, Munich, Germany
e-mail: peter.krummrich@siemens.com
Folkert Horst, Bert Jan Offrein, Gian Luca Bona

IBM Zürich Research Laboratory, 8803 Rüschlikon, Switzerland
e-mail: fho@zurich.ibm.com
Abstract: We prove the capability of adaptive PMD compensation with integrated optical FIR-
filters in lattice structure and an electrical spectrum monitoring based feedback signal in 40Gb/s
experiments. Moreover we present the combined PMD and GVD compensation.

OCIS codes: (230.7390) Planar Waveguides, (060.4510) Optical Communications
1. Introduction
In current and next generation high speed optical transmission systems solutions for adaptive polarization mode
dispersion (PMD) and group velocity dispersion (GVD) compensation are of high interest. Increasing bit rates, path
rerouting, temperature changes, power variations and stress induced birefringence lead to GVD and PMD values that
exceed the system tolerances. With the increasing complexity of the optical layer the tolerances are even getting
smaller. Therefore, an adaptive solution is necessary to meet the tolerances and to compensate for these time varying
distortions.
The adaptive equalizer should consist of an easy to handle PMD and GVD compensating device and a fast, robust,
and easy to generate feedback signal. Integrated optical FIR-filters can compensate for both, GVD [1] and PMD [2].
These devices are well understood and their transfer function can be easily and fast tuned (∆t<100µs) by the
thermooptic effect.
A number of adaptive control criterions have been proposed [3-6]. Because of its capability of detecting both - PMD
and GVD - and its speed and ease of implementation, an adaptive feedback strategy of monitoring frequencies
within the electrical spectrum of the received signal is chosen.
Signal distortions caused by the transmission fiber can be determined very well in the electrical spectrum of the
received signal. A fast, inexpensive and easy to implement solution is to monitor frequencies within the electrical
spectrum by electrical bandpass filtering. The monitoring of the power of a single frequency component f0 leads to a
power response with well defined minima and maxima for different values of PMD and GVD. Combining the PMD
and GVD response we get a 2 dimensional space with a global maximum at zero GVD and PMD. Monitoring two or
more frequencies prevent from tracking a local maximum. The adaptive GVD equalization with integrated optical
FIR filters in lattice structure and electrical spectrum monitoring as feedback has been demonstrated by our group,
[6].
In this paper we report on the experimental adaptive PMD compensation with integrated optical FIR filters in lattice
structure and electrical spectrum monitoring as feedback at 40Gb/s and show for the first time the promising results
for simultaneous PMD and GVD compensation with this kind of device.
TuG3
2. System
The setup in Fig.1 was used to demonstrate experimentally the adaptive PMD and GVD compensation with optical
FIR-filters and a feedback strategy using the electrical spectrum of the received signal.
pattern
generator
integrated optical 6th order FIR
101 lattice filter (double filter setup )
SSMF, DCF Rx BER
with variable
laser EDFA length polarization VOA EDFA BP-filter 3dB
scope
PMD PBS
emulator
PD BP-filter
transmission
channel
adaptive filter adaptive feedback signal
control
Fig.1: Experimental system setup: Simultaneous adaptive PMD and GVD compensation at 40Gb/s with integrated optical FIR -filters in lattice
structure and electrical spectrum monitoring as feedback signal
The transmitter provides a 42.46Gb/s NRZ signal (PRBS sequence= 231-1) at 193.1THz. The signal is distorted by
different values of GVD from SSMF and DCF of variable length. A 1st order PMD emulator sets the desired
differential group delay (DGD) value. The compensating devices, the 6th order optical FIR-filters, are based on a
lattice structure, which is implemented by cascading symmetrical and asymmetrical Mach-Zehnder Interferometers
(MZI) in a planar lightwave circuit (PLC). The FIR-filters are designed and fabricated by IBM Research using the
high index contrast SiON technology [7]. An arbitrary transfer function is achieved by controlling the coupling
ratios and the phase differences of the interfering signals. Both orthogonal polarization modes of the transmission
signal are separated by a polarization beam splitter (PBS) and each mode is equalized with a separate filter. At the
receiver, the electrical signal is bandpass filtered, fFWHM=300MHz, at a center frequency f0=20GHz and the detected
power is used as feedback signal for the adaptive control and equalization. As we operate in the range between
±25ps, this single frequency is sufficient for the feedback signal.
3. Results and Discussion

For adaptive PMD and GVD compensation, the compensating device should be able to compensate for both effects
and the feedback signal should provide a unique error signal in the desired range of operation. To verify the adaptive
feedback signal for PMD compensation at 40Gb/s, the DGD is varied from 0ps to +35ps and the power of the
bandpass filtered (center frequency f0=20GHz) received electrical signal is measured, Fig.2a. The feedback signal
has a maximum at zero PMD and a minimum at Tbit=25ps. For increasing the range with a unique response of the
feedback signal, the bandpass center frequency f0 has to be decreased [6].
-36 10
a.) b.)
group delay (ps)
measured 20GHz power ( dBm)
-38 0
differential
-40 -10
T bit =
-42 -20
25ps
-44 -30
-46 10.5
-48
c.) 10
-log(BER)
9.5
-50
9
-52 8.5
8
-54
0 10 20 30 40 7.5 1 2 3 4 5 6 7 8 9 10 11
differential group delay ( ps) iteration
Fig.2: a.) measured feedback signal at 40Gb/s for varying differential group delay values from 0ps to +35ps; b.) adaptive PMD compensation
results: differential group delay values=channel+equalizer, eye pattern; c.) Bit Error Rate measurements
TuG3
From a starting point of an approximate differential group delay of 23ps and equal power splitting between the
orthogonal modes (γ=0.5) (almost closed eye, no measurable BER), the adaptive control algorithm varies the overall
differential group delay (transmission channel+equalizer) slightly to detect the optimization direction. The actual
step is compared with the previous one. If the power of the BP-filtered electrical signal is decreasing, the
optimization direction is changed and the step size is halved until the optimum value is reached (minimum BER).
During the adaptive operation the overall DGD is decreased to 0ps and the eye pattern is very well opened up to an
error free transmission, Fig.2b,c.
The capability of compensating for both, PMD and GVD, with integrated optical FIR filters in the double filter setup
is demonstrated by measurements, Fig.3. The worst case with equal power splitting between the orthogonal modes
(γ=0.5) is considered, the filter transfer functions are set without the adaptive feedback signal.
a.) b.) GVD=100ps/nm DGD=25ps
-4
back to back
GVD=100ps/nm
-5 GVD=100ps/nm equalized
PMD=25ps equalized
GVD=100ps/nm &
-6
DGD=25ps/nm equalized
log(BER)
-7
-8 DGD=25ps & DGD=25ps & GVD=100ps/nm

GVD=100ps/nm equalized
-9
-10
-11
-29 -28 -27 -26 -25 -24 -23 -22 -21 -20
input power (dBm)
Fig. 3: a.) BER measurement results: back to back (star), GVD=100ps/nm (circle), GVD=100ps/nm equalized (square), DGD=25ps equalized
(right triangle), DGD=25ps & GVD=100ps/nm equalized (upper triangle); b.) measured eye pattern: GVD=100ps/nm, DGD=25ps, DGD=25ps &
GVD=100ps/nm, DGD=25ps & GVD=100ps/nm equalized
At a first step the transmission channel is set to a GVD value of 100ps/nm. While compensating for GVD only, we
realize a sensitivity gain of 4.5dB and a sensitivity penalty of less than 1dB in comparison to the back to back case at
a BER=10-9. Next, a PMD setting of the transmission channel of DGD=25ps is compensated. The initially closed
eye pattern is clearly opened and the resulting sensitivity penalty is approximately 1dB. Finally the transmission
channel is set to GVD and PMD values of GVD=100ps/nm and DGD=25ps. Equalizing the combination of PMD
and GVD, the totally closed eye pattern is clearly opened and the sensitivity penalty is less than 1.5dB.
4. Conclusions
The experimental results prove that two integrated optical FIR-filters in the polarization separated setup, where each
filter compensates for the distortion of one polarization mode, are very well suitable to compensate for PMD, GVD
and combinations of both. Equalizing a distortion of DGD=25ps and GVD=100ps/nm regenerates a totally closed
eye pattern to a clearly open one and a sensitivity penalty of less than 1.5dB. A feedback criterion generated by
electrical spectrum monitoring in terms of bandpass filtering the received electrical signal makes this filter setup
adaptive to compensate for distortions due to PMD and GVD.
5. References
[1] M.Bohn et al., “Tunable Dispersion Compensation in a 40Gb/s System using a compact FIR lattice filter in SiON technology”, European
Conference on Optical Communications proceedings (ECOC 2002), Tu.4.2.3
[2] M.Bohn et al., “Adaptive polarization mode dispersion compensation @ 40Gb/s with integrated optical FIR-filters”, National Fiber Optics
Engineers Conference (2002), We D6
[3] Q.Yu et al., “Chromatic dispersion monitoring technique u sing sideband optical filtering and clock phase-shift detection”, Optical Fiber
Communication Conference (OFC 2002), We 3
[4] P.Westbrook et al., “Measurement of residual chromatic dispersion of a 40-Gb/s RZ signal via spectral broadening”, IEEE Photonics
Technology Letters, Vol.14, No.3, pp. 346-48 (2002)
[5] M.Petersen et al., “Dispersion monitoring and compensation using a single in -band subcarrier tone”, Optical Fiber Communication
Conference (OFC 2000), WH4-1
[6] M.Bohn et al., “Simultaneous adaptive equalization of group velocity and polarization mode dispersion at 40Gb/s with integrated optical
FIR-filters and electrical spectrum monitoring as feedback”, European Conference on Optical Communications proceedings (ECOC 2003), Th2
[7] R. Germann et al., "SiON high-refractive-index waveguide and planar lightwave circuits", IBM J. Res. & Dev. vol 47 No 2/3, March 2003
TuG4
Electrical PMD equalizer ICs for a 40-Gbit/s transmission
M. Nakamura, H. Nosaka, M. Ida, K. Kurishima, and M. Tokumitsu

NTT Photonics Laboratories, NTT Corporation, 3-1, Morinosato Wakamiya, Atsugi, Kanagawa, 243-0198 Japan
Phone: +81-46-240-2744, Fax.: +81-46-240-4383, E-mail: nakamura@aecl.ntt.co.jp
Abstract: Electrical PMD equalizer ICs for 40-Gbit/s transmission were developed using
InP/InGaAs HBT technology. In experiments using a PMD emulator, these ICs exhibit good
compensation for DGD of over 20 psec at 40 Gbit/s.
OCIS codes: (060.2330, 060.2340)
1. Introduction
In optical communications systems operating at high-data rates over 40 Gbit/s, signal distortion caused by
polarization mode dispersion (PMD) limits transmission distance. PMD is regarded as birefringence with a
differential group delay (DGD) caused by dual path propagation in the fiber, and it produces intersymbol
interference (ISI), which causes an increase of the bit error rate (BER). As the DGD is statistically distributed, an
adaptive equalizer is needed. To remove the system degradation caused by PMD, several optical and electrical
equalization techniques have been proposed and demonstrated [1-4].
Electrical compensation has advantages in cost and size over optical compensation. In addition, it can also mitigate
chromatic dispersion (CD), which can be handled as ISI by an electrical equalizer. However, the speed of electrical
equalization is limited by the speed of the electrical components. Improving the operation data rate requires
high-speed ICs that can handle large DGD. In this paper, we describe the design and performance of 40-Gbit/s
equalizer ICs using InP/InGaAs HBT technology to reduce PMD-induced degradation. This is the first report of an
electrical PMD equalizer for 40-Gbit/s transmissions.
2. FFE and DFE Equalizers

A linear equalizer as a feed-forward equalizer (FFE) and a nonlinear equalizer as a decision feedback equalizer
(DFE) were developed. It is known that the optimum equalizer structure combines both FFE and DFE components.
A FFE can remove both pre- and post-cursor ISI, and a DFE can remove post-cursor PMD without increments of
noise. These filters can therefore reduce the ISI in the received signal.
The FFE circuit has a transversal filter structure with three taps using two delay stages. It consists of tap delays,
mixers, adders, and buffer amplifiers, as shown in Fig. 1(a). Each component with unity gain is used for linear filter
operation. Mixer cells adjust the tap weights Ci in the range between “–1” and “1”. The tap delay is given as the
difference between delay1 Td1 and delay2 Td2. These delayed and weighted signals are superimposed at the output.
Delay1 Delay Subtractor

Buffer Buffer
Td1 Td1 Mixer In Td1 DEC
In + - Out
Mixer Buffer
Buffer
Out Buffer DEC Td2
Td2 Td2
CLK
Delay2 Adder
Tap1 Tap2 Tap3 Tap
(a) (b)
Fig. 1 Block diagram of equalizer circuits. (a) FFE circuit, (b) DFE circuit.
TuG4
The noise generated at each tap is added together and the accumulated noise degrades eye diagram at output. In
addition, the noise is enlarged by increasing bandwidth. To achieve high-speed performance, small numbers of three
tap stages are effective in reducing the influence of noise. The tap delay is designed as an optimal value for
40-Gbit/s operations. A differential configuration is used for stable operation at high speed.
The DFE circuit is based on a feed-forward structure with one tap, as shown in Fig. 1(b). A conventional DFE
circuit has a feedback loop and is difficult to speed up because of the parasitic delay in the loop. To achieve
high-speed operation of over 40 Gbit/s, we devised a feed-forward circuit that works as well as a conventional DFE.
This circuit has two signal paths. One is the main signal path, and the other is the weighted signal path with a
decision. In this circuit, the tap delay is designed to be a period of the data rate. The weighted signal is subtracted
from the main signal. Using this circuit, the tap delay is the difference between the two signal paths. It can be set to
an arbitrary value, which makes it possible to increase operation speed.
The ICs were fabricated using InP/InGaAs HBTs, which have fT of 150 GHz and fmax of 240 GHz [5]. The die size
of the FFE IC is 3 x 3 mm2, and that of the DFE IC is 2 x 2 mm2. For high-speed analog filters, InP-base devices
have an advantage in linearity over Si devices due to their high-breakdown voltage, which enables enlargement of
the linear operation region. In addition, the circuits are configured using lumped circuits instead of the distributed
circuits [6] used in conventional equalizers for high-frequency applications. The lumped components improve the
operation of addition and subtraction in the analog filter and reduce the chip size.
3. Experiment and Results

The experimental setup for PMD compensation measurement is shown in Fig. 2. A 40-Gbit/s NRZ optical signal was
generated by an external modulator driven with 27–1 pseudo-random-bit-stream (PRBS) signal from a pulse-pattern
generator (PPG) and fed to a polarization controller (PC). The polarized signal generated by the PC was fed to a
PMD emulator that generates DGD. This PMD-emulated signal was provided to a PD through an erbium-doped
fiber amplifier (EDFA). The output of the PD was directly connected to the equalizer ICs. The bit error rate was
measured with a 40-Gbit/s BER test set. The performances of the ICs were evaluated using high-speed wafer probes.
DGD
1-J
J
“1” “0” “1” “1” “1” “0”
40 Gbit/s PPG
PRBS 27-1
LN PMD Equalizer
IC BER
Modulator Emulator
PC
LD EDFA PD
PC: polarization controller, PPG: pulse-pattern generator, BER: bit-error rate detector
Fig. 2 Experimental setup for PMD equalization using an emulator.
Measurement results of PMD equalization using the FFE IC are shown in Fig. 3. Fig. 3(a) is the input waveform
with DGD of 20 psec at the power splitting ratio J of around 0.5. The eye is almost completely closed due to the ISI
caused by the PMD. The eye diagram equalized using the developed equalizer IC is shown in Fig. 3(b). It is clearly
open, when the normalized tap weight is Ci = [0.5, 0.5, –1]. Fig. 3(c) shows measured BER characteristics with and
without the FFE IC. The IC has immunity from DGD of over 20 psec. Power supply voltage of the FFE IC is –4.3 V,
and its power dissipation is 0.82 W.
Measured input and output waveforms of DFE IC are shown in Fig. 4, where (a) is the input waveform and (b) the
output one. The input data contains DGD of 25 psec at J of 0.3. The DFE IC can also mitigate the PMD of over 50
psec at the power splitting ratio J of around 0.3. Power supply voltage is –4.5 V, and the power dissipation is 1.3 W.
TuG4
100 mV/div
40 Gbit/s NRZ
PRBS 27-1 with FFE
-2 J= ~0.5
w/o FFE
Log10 BER
J= ~0.5 J= ~0.34
5ps/div -4
(a)
-6 J= ~0.28
70 mV/div
-8
-10
-12
0 10 20 30 40
DGD [ps]
5ps/div
(c)
(b)
Fig. 3 40-Gbit/s PMD equalization of the FFE IC. (a) Input waveform with DGD of 20 psec at J of 0.5, (b) output waveform,
(c) BER performance.
100 mV/div 100 mV/div
(a)
5ps/div
5ps/div
(a) (b)
Fig. 4 Measured waveforms of the DFE IC. (a) Input waveform with DGD of 25 psec at J of 0.3, (b) output at 40 Gbit/s.
4. Conclusions
Electrical PMD equalizer ICs for 40-Gbit/s transmissions were developed, and they exhibit good performance for
PMD equalization. To achieve high-speed operation, we designed a FFE circuit with minimized tap stages and
optimized delay time and devised a feed-forward DFE circuit. These ICs were fabricated using InP/InGaAs HBTs.
The ICs can compensate DGD of over 20 psec for 40-Gbit/s NRZ signal. Theses chips permit adaptive equalization.
REFERENCES
[1] S. Lanne, et al., “Extension of polarization-mode dispersion limit using optical mitigation and phase-shaped binary transmission”, ThH3-1,
OFC2000.
[2] H. Bulow, et al., “PMD mitigation at 10 Gbit/s using linear and nonlinear integrated electronic equaliser”, IEE Electron. Lett., vol. 36, pp.
163-164, Jan. 2000.
[3] K. Azadet, et al., “Equalization and FEC Techniques for Optical Transceivers”, IEEE J. Solid-State Circuits, vol. 37, pp. 317-327, Mar. 2002.
[4] G. Kanter, et al., “Electric equalization for extending the reach of electro-absorption modulator based transponder”, ThG6, OFC2003.
[5] M. Ida, et al., “Undoped-Emitter InP/InGaAs HBTs for High-Speed and Low-Power Applications”, Tech. Dig. IEDM2000, pp. 854-856.
[6] J. Lee, et al., “MMIC Adaptive Transversal Filtering Using Gilbert Cells and is Suitable for High-Speed Lightwave Systems”, IEEE Photon.
Technol. Lett., vol. 12, pp. 196-198, Feb. 2000.
TuG5
Mitigation of Nonlinear Impairments from Semiconductor

Optical Amplifiers using an Optical Equalizer
Ashish Bhardwaj, Christopher R. Doerr, S. Chandrasekhar and Lawrence W. Stulz
Bell Laboratories, Lucent Technologies, 791 Holmdel-Keyport Road, Holmdel, NJ 07733
e-mail: ashishb@bell-labs.com
Abstract: We demonstrate and study significant improvements in the measured bit-error

rates of 40-Gb/s signals distorted after passing through a single and a cascade of two
semiconductor optical amplifiers by using a simple colorless optical equalizer.
©2003 Optical Society of America
OCIS codes: (250.5980) Semiconductor optical amplifiers; (130.3120) Integrated optics devices; (350.2460)
Filters, Interference
1. Introduction
Optical equalizers have been proposed and demonstrated recently for the mitigation of inter-symbol
interference (ISI) [1] arising from impairments such as chromatic dispersion and polarization mode
dispersion. However, these previously studied impairments are linear in nature. In this paper, we
investigate the improvement in the bit-error rates (BER) by using the equalizer to mitigate the intra-channel
distortions arising from ultra-fast nonlinear gain dynamics from a semiconductor optical amplifier (SOA).
It is well known that the nonlinearities in an SOA, primarily inter-band transitions, can cause a significant
power penalty to amplitude-shift-keyed signals passing through the SOA, with the penalty worsening the
closer the SOA output power is to saturation [2]. We investigate the interesting property that a linear
device, the optical equalizer, can significantly reduce this penalty. We also study the improvement caused
by the equalizer on nonlinear distortions caused by a cascade of two SOAs.
2. Optical equalizer
The optical equalizer consists of two single-mode-connected Mach-Zehnder interferometers (MZIs) each
with a relative optical path delay of 20 ps [1]. At each instant of time, the optical equalizer takes a
controllable portion of the energy present and adds it at ± 20 ps with a controllable phase. Each MZI is on a
separate silica waveguide chip, but both are mounted on a single thermoelectric cooler. The waveguides
have a 0.8% index contrast and are on a silicon substrate. The fiber-to-fiber insertion loss for both MZIs in
series is 4.2 dB, and the polarization-dependent loss is less than 0.5 dB.
Each MZI has two tunable couplers and a thermo-optic phase shifter in one arm. Each tunable coupler
consists of a small MZI with a thermo-optic phase shifter in one arm and a quarter-wavelength length
increase in the other arm. Both couplers in each MZI are adjusted to be at the same value for minimum
insertion loss, so the coupler drives within an MZI are wired together. The coupler drives control the
impulse response satellite amplitudes, and the phase shifters control the satellite phases.
3. Experiment
The SOA used in this study was a commercially available one with a fiber-to-fiber small-signal gain of 18
dB at 1550 nm at a bias current of 200 mA. Non-return-to-zero (NRZ) data at 40-Gb/s was modulated on
the light from an external cavity laser at 193.4 THz (1550.12 nm) with a chirp-free LiNbO3 modulator and
launched into the SOA. The input power to the SOA was varied from –12 dBm to –3 dBm to study the
degradation of the BER resulting from the nonlinearity of the SOA. The BER is measured right after the
SOA without any equalization, as shown in Figure 1 (a). The BER is also measured after equalization, as
shown in Figure 1 (b).
TuG5
Pattern Generator
40 Gb/s NRZ
PRBS 231-1
Tunable Error
Laser MOD SOA Pre-amp PD Detector
(a)
Pattern Generator
40 Gb/s NRZ
PRBS 231-1
Tunable Optical Error

Laser MOD SOA Pre-amp Equalizer PD Detector
(b)
Figure 1: Measurement set-up: (a) without and (b) with the optical equalizer. PD: photo-detector.
Figure 2 (a) shows the measured BER after transmission through the SOA both with and without
equalization. The input power to the SOA is –3 dBm. The SOA penalizes the sensitivity in the BER at the
receiver. The equalizer substantially reduces this penalty. For a BER of 10-9 an improvement of nearly 5
dB is observed in the receiver sensitivity for a pseudo-random bit stream (PRBS) of length 231-1. Figure 2
(b) shows the receiver sensitivity for a BER of 10-9 both with and without equalization for different input
powers to the SOA. The equalizer settings were readjusted for each power level.
3 -20
31 unequalized
40-Gb/s, NRZ, PRBS 2 -1 unequalized 31
equalized -21 40-Gb/s, NRZ, PRBS 2 -1 equalized
back-to-back back-to-back
4 -22
Receiver Sensitivity (dBm)
-23
5 -24
-Log(BER)
-25
6 unequalized
-26
7
-27
8
-28
9 equalized
-29
10
11 -30
-40 -38 -36 -34 -32 -30 -28 -26 -24 -22 -20 -18 -16 -14 -12 -10 -8 -6 -4 -2 0
Received Power (dBm) Input Power to SOA (dBm)
(a) (b)
Figure 2: (a) Measured BER after transmission through one SOA both with and without equalization after the SOA. Input power to the
SOA is –3 dBm. Eye diagrams for the unequalized and equalized case are added as inserts. (b) Receiver Sensitivity for a BER of 10-9
versus input power to the SOA measured with and without equalization at 1550.12 nm.
The overshoots from the SOA lead to an increase in the average power of the data-stream, resulting in a
power penalty. However, this effect is small compared to the fact that the overshoots on the rising edges
can cause a significant enhancement of ISI in the receiver [2]. Ref. [2] showed how an ideal integrate-and-
dump receiver would be relatively immune to SOA-induced distortions, whereas the usual filtered receiver
exhibits significant eye closure from the overshoots.
TuG5
The measurements were repeated on a cascade of two SOAs using the same set-up as shown in Figure 1,
except that one SOA is replaced with two SOAs in series. The power launched into each SOA was the
same and was controlled by variable attenuators placed before each SOA. Figure 3 (b) shows the receiver
sensitivity both with and without the equalizer for a BER of 10-9 for different input powers to a cascade of
two SOAs. For input powers higher than –9 dBm into each SOA, the nonlinear distortion after passing
through the SOAs is so high that a BER of 10-9 could not be measured without equalization. Figure 3 (a)
shows the quality factor, Q, measured both with and without equalization after the cascade of two SOAs.
The improvements caused by the equalizer after one SOA versus two SOAs can be seen by comparing
Figures 2 (b) and 3 (b).
24 -16
unequalized 31 unequalized
31 40-Gb/s, NRZ, PRBS 2 -1 equalized
40-Gb/s, NRZ, PRBS 2 -1 equalized -18
22
back-to-back
-20
Receiver Sensitivity (dBm)

20
-22
Quality Q (dB)
18
-24
16
-26
14
-28
12
-30
10 -32
-10 -9 -8 -7 -6 -5 -4 -3 -2 -10 -9 -8 -7 -6 -5 -4 -3 -2
Input Power to both SOAs (dBm) Input Power to both SOAs (dBm)
(a) (b)
Figure 3: (a) Quality factor Q versus input power into each SOA measured with and without equalization at 1550.12 nm. (b) Receiver
sensitivity for a BER of 10-9 versus input power into each SOA measured with and without equalization at 1550.12 nm.
4. Conclusions
We have shown that an optical equalizer can be used to significantly reduce the power penalty in the
measured BER for NRZ data at 40-Gb/s due to the fast nonlinear gain saturation of a SOA. Significant
improvements in the receiver sensitivity for a BER of 10-9 are observed with the use of the equalizer after
the SOA. The use of such optical equalizers could enhance the performance of SOAs in a variety of
applications as power boosters on the transmitter side [3], in-line amplifiers [4] and pre-amplifiers on the
receiver side [5].
5. References
[1] C. R. Doerr, S. Chandrasekhar, P. J. Winzer, L. Stultz, A. R. Chraplyvy, and R. Pafchek, “Simple Multi-Channel Optical
Equalizer for Mitigating Intersymbol Interference”, Post-deadline Paper PD 11, Optical Fiber Communication Conference, OFC
2003, Atlanta, GA.
[2] A. A. M. Saleh, and I. M. I. Habbab, “Effects of Semiconductor-Optical-Amplifier Nonlinearity on the Performance of High-
Speed Intensity-Modulation Lightwave Systems”, IEEE Trans. Commun., vol. 38, no. 6, pp. 839-846, June 1990.
[3] Y. Kim, H. Jang, Y. Kim, J. Lee, D. Jang, and J. Jeong, “Transmission Performance of 10-Gb/s 1550-nm Transmitters using
Semiconductor Optical Amplifiers as Booster Amplifiers”, J. Lightwave Technol., vol. 21, no. 2, pp. 476-481, February 2003.
[4] J. Jennen, H. de Waardt, and G. Acket, “Modeling and Performance Analysis of WDM Transmission Links Employing
Semiconductor Optical Amplifiers”, J. Lightwave Technol., vol. 19, no. 8, pp. 1116-1124, August 2001.
[5] B. Mikkelsen, C. G. Jorgensen, N. Jensen, T. Durhuus, K. E. Stubkjaer, P. Doussiere, and B. Fernier, “High-performance
Semiconductor Optical Preamplifier Receiver at 10 Gb/s”, IEEE Photon. Technol. Lett., vol. 5, no. 9, pp. 1096-1097, September
1993.
TuG6
The impact of nonlinearity on electronic dispersion

compensation of optical channels
Oscar E. Agazzi and Venugopal Gopinathan
Broadcom Corporation, 16215 Alton Parkway, Irvine, California 92618.
E-mail: oea@broadcom.com, venu@broadcom.com
Abstract: We show that the nonlinear, non-minimum phase nature of the optical channel is
largely responsible for the marked performance difference between maximum likelihood sequence
estimation (MLSE) and decision-feedback equalizer (DFE) approaches to dispersion
compensation .
OCIS code: (060.4510) Optical communications; (060.2330) Fiber optics communications
1. Introduction
Electrical signal processing techniques have been proposed to compensate impairments of optical channels such as
chromatic and polarization mode dispersion (PMD) [1-2]. Advances in integrated circuit technology are now
beginning to make the implementation of these techniques feasible at the speeds required in today’s applications
(10Gb/s or higher) [3]. As a result, interest in electronic dispersion compensation is growing in industry, and some
products have been announced. However, since this technology has not been proven in real world applications yet, it
is important to clarify the potential of the different approaches proposed. Among the techniques proposed are the
decision-feedback equalizer (DFE) [4], and maximum likelihood sequence estimation (MLSE)[1,5]. These
techniques have already been applied in electrical channels such as voiceband modems, twisted pairs, digital radio,
etc. However optical channels have unique properties, such as strong nonlinearities and non-Gaussian noise, that
require that these techniques be revisited.
In this paper we show that MLSE is robust in the presence of nonlinearity. In Section 3, we show that an MLSE
receiver based on an 8-state Viterbi decoder with a nonlinear channel estimator can compensate for the linear and
nonlinear intersymbol interference (ISI) of up to 5000ps/nm of dispersion. We also show that the DFE is severely
hampered not only by the nonlinearity but also by the fact that the nonlinear ISI created by a bit is observed before
the bit is actually detected.
2. Model of the optical channel

We are interested in intensity-modulated single-mode fiber links without optical dispersion compensation (this could
apply, for example, to metro links). However, this analysis is also applicable to residual dispersion in long-haul links
that have optical dispersion compensating fibers (DCFs). For sufficiently long fibers, optical amplification is used to
overcome fiber loss.
Nonlinearity of the optical channel stems predominantly from the photodetector. Current in photodetectors is
proportional to the received optical power, which is a quadratic function of the electromagnetic field amplitude in
the fiber. Another source of nonlinearity is the intensity dependence of the refractive index of the fiber. This is
important in DWDM links, where the optical power is the aggregate of many channels. Although for simplicity we
only consider the nonlinearity in the photodetector, our conclusions remain valid when other forms of nonlinearity
are also present.
It can be shown [6], that the optical channel behaves approximately linearly in spite of the square law behavior
of the photodetector when there are multiple modes of propagation and/or when the coherence time of the source is
small compared to the bit-interval. These conditions are often satisfied by multimode fibers, but not by single-mode
fibers driven by highly coherent sources [7]. The quantity that determines the degree of nonlinearity present in a
signal that has suffered chromatic dispersion is the ratio of the coherence time of the source to the bit-period. The
coherence time of modern DFB lasers is in the tens of nanoseconds range, which is much larger than the symbol
period at rates of 10Gb/s or higher. Under these conditions, the nonlinearity of the photodetector plays a major role,
and should be included in the analysis of electrical equalization techniques to realistically assess their effectiveness.
To model the channel we use the formulation of sections II and III-B of [7]. We assume that the transmitted
f
power is Pi t ¦ k f
a k p t kT , where a k ^0,1` are the transmitted symbols, and p(t) is the transmitted
pulse, assumed to be confined to a symbol interval T. As shown in [7], for a coherent source the received optical
power can be expressed as
TuG6
Po t ¦ a ht kT ¦¦ a a v t nT
k n k k n (1)
k n k
where h(t) = |\c(t)|2 is the impulse response for the received optical power, \c(t) is the complex envelope of the
electromagnetic field for the received pulse, and functions
^
v m t 2 Re \ c t \ c t mT ` (2)
are second order Volterra kernels corresponding to nonlinear interactions between symbols m bit periods apart. The
Volterra kernels provide a complete description of the channel nonlinearity. It is interesting to mention that the
second term of (1) vanishes if the source is incoherent [6,7].
3. Simulation Results
In the following we assume that the transmitted pulse has an unchirped Gaussian envelope exp(-t2/2T02) with unit
amplitude and T0 = 36ps, and we compute \c(t) solving the dispersion equation (equation (2.4.15) of [8]). Although
in this example we consider chromatic dispersion only, our conclusions are still valid if PMD is also present. Figure
1 shows the impulse response and Volterra kernels v1(t) and v2(t) for the case of a single-mode fiber with a Dl
product of 1700ps/nm, where D is the dispersion parameter of the fiber and l is its length. For example, it could
correspond to an unchirped link of 100km of standard fiber with D=17ps/nm.km. Figure 2 shows the corresponding
eye pattern at the receiver. In this figure the sampling point should be located at about 100ps with a threshold of
about 0.4 (this is shown as a cross in the figure). The large sidelobes at 50ps and 150ps are caused by the v1(t)
kernel.
Fig. 1. Fig.2.
Three receivers, a non-equalized clock/data recovery (CDR) detector, an 8-tap DFE, and an 8-state MLSE
receiver have been simulated using this channel model. Amplified stimulated emission (ASE) noise [9], is added to
the received signal. Since this noise has a non-Gaussian probability density function (pdf), the thresholds for the
CDR and the DFE are adjusted for minimum bit error rate. In the case of the MLSE receiver, the nonlinearities in
the channel are captured in the nonlinear channel estimator, that automatically adapts itself to the channel response
(1) during actual operation of the link. It also adapts automatically to the pdf of the noise. Figure 3 shows the optical
signal to noise ratio (OSNR) required for a constant bit error rate of 10-6 for the three receivers as a function of the
dispersion product Dl (to make this discussion independent of fiber properties, we use the product Dl instead of fiber
length l). It can be seen that the CDR reaches its limit at about Dl=1430ps/nm at an OSNR of 20dB. The DFE
tolerates an extra 360ps/nm of dispersion at the same OSNR. Finally, the MLSE receiver reaches well over
5000ps/nm under the same conditions.
TuG6
Tolerable dispersion
at OSNR=20dB
CDR 1430ps/nm
DFE 1790ps/nm
MLSE >5000ps/nm
Fig. 3. Required OSNR vs Dispersion
The performance limitations of the DFE result from the fact that the signal it creates to cancel the channel ISI is
a purely linear combination of previously received bits, therefore it cannot cancel the nonlinear ISI generated by the
Volterra kernels of Fig.1. It is interesting to observe that a nonlinear version of the DFE exists [10], where the ISI
replica can be a nonlinear function of previously received bits. Unfortunately this structure is fundamentally unable
to compensate for nonlinearites excited by the bit currently being detected or by future bits (precursor type nonlinear
ISI) which are present in (1). As can be seen in Fig. 1, at least part of the nonlinear ISI is of precursor type. On the
other hand, the nonlinear channel estimator of an MLSE receiver can accurately model nonlinear ISI regardless of
whether it is of precursor or postcursor type. It can also provide accurate modeling of ASE or any other non-
Gaussian noise probability density function.
4. Conclusions
As a result of its inability to compensate nonlinear ISI, particularly of the precursor type, the DFE suffers from
severe performance limitations in moderate to high dispersion single-mode fiber links driven by highly coherent
sources. The MLSE receiver is the appropriate electronic dispersion compensation solution for these channels.
5. References
[1] J.H.Winters and R.D.Gitlin, “Electrical Signal Processing Techniques in Long-Haul Fiber-Optic Systems,” IEEE Trans. Commun., Vol.38,
No.9, Sept.1990, pp.1439-1453.
[2] J.H.Winters, R.D.Gitlin, and S.Kasturia, “Reducing the Effects of Transmission Impairments in Digital Fiber Optic Systems,” IEEE
Communications Magazine, June 1993, pp.68-76.
[3] H.Bulow, “Electronic Equalization of Transmission Impairments,” Proceedings of the 2002 Optical Fiber Communications Conference,
March 17-22, 2002, pp.24-25.
[4] B.L.Kasper, “Equalization of Multimode Optical Fiber Systems,” Bell Syst. Tech. J., Vol.61, No.7, Sept.1982, p.1367.
[5] H.F.Haunstein, K.Sticht, A.Dittrich, W.Sauer-Greff, and R.Urbansky, “Design of near optimum equalizers for optical transmission in the
presence of PMD,” Proceedings of the Optical Fiber Communication Conference and Exhibit, Vol.3, 2001, pp.558-560
[6] S.D.Personick, “Baseband linearity and equalization in fiber optic digital communication systems,” Bell System Technical Journal, Vol.52,
No.7, Sept.1973, pp1175-1194.
[7] B.E.A.Saleh and M.I.Irshid, “Coherence and intersymbol interference in digital fiber optic communication systems,” IEEE Journal of
Quantum Electronics, Vol. QE-18, No.6, June 1982, pp.944-951.
[8] G.P. Agrawal, Fiber –Optic Communication Systems, (John Wiley and Sons, 1997).
[9] D.Marcuse, “Calculation of Bit-Error Probability for a Lightwave System with Optical Amplifiers and Post-Detection Gaussian Noise,”
Jour. of Lightwave Technology, Vol.9, No.4, April 1991, pp.505-513.
[10] S.Kasturia and J.H.Winters, “Techniques for high-speed implementation of nonlinear cancellation,” IEEE Journal on Selected Areas in
Communications, Vol.9, No.5, June 1991, pp.711-717
TuH1
Efficient Physical Topologies for Regular WDM Networks

Chi (Kyle) Guan and Vincent Chan
MIT Laboratory for Information and Decision Systems, Cambridge, MA 02139 ckguan@mit.edu, chan@mit.edu
Abstract: In this paper, we study the design of efficient physical topology of WDM network
by building an analytical framework to find optimal connectivity of the regular networks as
network sizes and traffic volume scale.
¤2003 Optical Society of America
OCIS Code: (060.4250) Networks; (060.2330) Fiber optics communications.
1. Introduction
With the maturity of enabling technologies such as micro-electromechanical system (MEMS), optical
cross-connect (OXC) switched wavelength division multiplexing (WDM) networks become attractive
transports for both the backbone and metropolitan area networks [1, 2]. This work focuses on finding the
efficient topological architectures of such networks. An important question that repeatedly surfaces in the
network topology design is: given the locations of the candidate nodes, enabling technologies and traffic
demands (logical topologies), how to find the classes of underlying physical topologies that achieve the
minimum cost. Most literatures approach this problem by formulating the design processes as cost
minimization problems using integer linear or integer nonlinear programming (ILP or INLP) techniques.
Since the design processes solely rely on ILP or INLP techniques, they provide limited insight into what
constitutes good network architectures. We find analytical results for the various classes of regular
topologies so that their behaviors can help us to better understand the network architecture trends as
network parameters (e.g. number of nodes) change. We first focus on an important subset of the mesh
topology - symmetrically connected regular topologies and study their optimal connectivity for simplified
albeit representative all-to-all uniform traffic. Our cost model includes both the fiber cost which is
proportional to the number of fibers in the network, and the OXC cost which is a convex function of the
number of transit and add/drop lightpaths at the nodes. One important aspect of our work is that we
quantify how the required switching resources are related to the network routing strategies, which had been
treated as decoupled by most previous researches. We show that the minimum hop distance is more than a
parameter for the physical size and optical reach of the lightpaths. It is also a key parameter in
dimensioning the network switching resources. By assigning parametric cost structures to network fiber
plants and switching resources and analyzing their tradeoffs, we are able to specify the optimal node
degrees analytically for interesting classes of good regular topologies as the sizes and traffic volumes of the
network scale.
2. Optimal Network Connectivity Under the Uniform Traffic

We study the optimal connectivity of the topologies for the uniform traffic using minimum hop
routing. We adopt the minimum hop routing because it allows each lightpath will pass through the least
number of intermediate nodes and tie up the least amount of switching resources. In order to make the
problem analytically tractable, we restrict our investigation on symmetrically connected regular graph
topologies. Irregular graph structures will require numerical analysis. We consider two symmetrically
connected regular topologies the '-nearest neighbor and symmetric Hamilton graph as shown in Fig. 1. We
also include Moore Graphs in our analysis, since topologies such as the Shuffle Net and de Brujin Graph
have properties that their minimum hop distances are close to the Moore Bound [3]. We proceed to show
that the switching resource required at each node is proportional to network size and the minimum hop
distance. By assigning the first order cost structures to fiber plants as well as three most representative
types of switching fabrics, namely two-dimensional (2D) single-MEMS, multi-stages and three-
dimensional (3D) single-MEMS optical switches, we can express the cost explicitly as functions of the
network size, traffic volume units of t wavelengths, routing strategy and switching technologies. We show
that with parametric cost ratio between fiber and switching resources, the cost function is convex with
respect to node degree ' and there is an optimal node degree '* for a regular network with N nodes. Table
1 shows in closed-form how the network connectivity changes as the networks size and traffic demand
scale for different topologies and OXC switching fabrics. We also plot how optimal connectivity and cost
scale as network sizes and traffic volumes in Fig.2 to Fig.5.
1
TuH1
3. Dynamic Traffic
We assume that each node has exactly t wavelengths of traffic to send to every other node in the network.
We use static or quasi-static traffic on the fibers and nodes to approximate the mean of the dynamic traffic
and dimension the capacities of the fibers and the nodes accordingly. For simplicity of analysis, we also
assume that active switching will handle every unit of the pass through and add-drop traffic at nodes. In
the more realistic scenarios, fixed routing and wavelength assignment (RWA) algorithm implies a fixed
setting of switches along a lightpath. If the traffic does not change, it can be supported without active
optical switches. Besides restoration and protection switching, the main advantage of active optical
switching lies in its capability to adapt to the fluctuating/stochastic traffic. When the arrivals of traffic
exceed the capacities dimensioned, the network will encounter blocking. Thus for a network to achieve a
given requirement on blocking probability, the network designers need to dimension the capacity of the
fibers (or nodes) with the mean l plus some margin kV, where V is the estimated variance of the traffic
distribution and k t 1 is the amount of extra provisioning in addition to the average. To be economical, the
low cost switching equipment such as fiber-patched panels can handle the static or the quasi-static portions
of the traffic with fixed routing and passive switching, while the more expensive active switching
accommodate the fluctuating/stochastic portions. The blocking probability of the network will depend on
the detailed traffic statistics (which may not be available a priori) and the policy of reconfiguring (or not)
existing connections. In this paper, we will provide some insight into the problem, by providing an upper
bound on the worst case blocking probability over the set of all possible stochastic uniform all-to-all traffic
distributions. We will show that for any given k the maximum blocking probability is 1/(1+k2). The
distribution (with mean l and variance V) as function of k that has the maximum blocking probability is
given by,
1
°°1 k 2 if l l kV ; (1)
Pr(l ) ® 2
° k V
if l l .
°¯ 1 k 2 k
Similarly for any given the blocking probability p, the maximum margin is k (1 p)/ p . The resulting
distribution as function of p is given by,
1 p
° p if l l V; (2)
° p
Pr(l ) ®
°1- p if l l p
° V.
¯ 1 p
I.V. Conclusion
An important result of our research is that switching technologies have huge impacts on the final
architectures. Networks with 3-D switching fabrics have the best scalability. While 2-D architectures do
not scale well in the sense that the topologies having the minimum cost are fully connected as the networks
and their traffic volume exceed certain sizes. The other important result is that topologies close to Moore
Graphs can achieve near optimum cost. A Moore Graph of 100 Nodes with 3-D fabric has optimal
normalized node degree '*/(N-1) | 5%.
Table 1. Optimal node degree '* as functions of N and t.
Topologies/ Optimal node degree as the function of N and t

Switching '-Nearest Symmetric Moore Graphs
Fabrics Neighbors Hamilton Graph (asymptotic)
Et 1 Et Et N
( N 1) N 1
3-D 2D 2 D D ln( N )
Multi- Et
N
Stages Numerical Numerical D
Et2 N 2
3 4 4 5 3 4 4 5
2-D v at 2 N 3 bt 3 N 3 v ct 2 N 3 dt 3 N 3 4D ln( N )
2
TuH1
Reference:
[1] V.W.S. Chan, K. L. Hall, E. Modiano, and K. A. Rauschenbach, “ Architectures and Technologies for High-Speed
Optical Data Networks”, J. Lightwave Technol. vol.16, no.12, pp 2146-2168, Dec. 1998.
[2] A. A. M. Saleh and J. M. Simmons, “Architectural Principles of Regional and Metropolitan Access Networks”, J.
Lightwave Technol. vol. 17, pp. 2431-2448, Dec. 1999.
[3] K. Sivarajan, and R. Ramaswami, “Lightwave Networks Based on de Bruijn Graphs”, IEEE/ACM Trans.
Networking, vol.2, no.1, Feb. 1994.
1 2 1 2
6 3 6 3
5 4 5 4
(a) (b)
Fig.1 '-nearest neighbors (a) and symmetric Hamilton Graph (b) regular topologies with '=3 and N=6.
6
10
3D Moore.
( */(N-1))
1 3D '-nearest neighbors.
Multi-stage Moore.
Multi-stage '-nearest neigbors.
3D Moore Graphs. 2D Moore graphs.
Normalized Optimal Node Degree '
3D '-nearest neighbors.
10
5 2D '-nearest neighbors
Normalized Optimal Cost
0.8 Multi-stage Moore Graphs.

Multi-stage '-nearest neighbord
2D Moore Graphs
2D '-nearest neighbors
0.6
4
10
0.4
3
10
0.2
2
0 10
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Units of traffic t Units of traffic t
Fig.2 (Left) Analytical results of normalized optimal node degree as a function of traffic unit t. The fiber to switching cost ratio D/E is
set at 40. The network size is N=50.
Fig.3 (Right) Analytical results of normalized optimal cost as a function of traffic unit t. The fiber to switching cost ratio D/E is set at
40. The network size is N=50.
0
10
Normalized Optimal Node Degree,'*/(N-1)
10
3 '-Nearest Neighbors
Symmetric Hamilton Graph
Hypercube
Shufflenet
Swtich Size at Each Node
DeBrujin Graph
Moore Bound
2 -1
10 10
'-Nearest Neighbor 2D
Symmetric Hamilton 2D
Moore Bound 2D
'-Nearest Neighbor Multi-stage
Symmetric Hamilton Multi-stage
Moore Bound Multi-stage
'-Nearest Neighbor 3D
Symmetric Hamilton 3D
1
10 Moore Bound 3D
-2
1 2
10 1 2
10 10 10 10
Number of Nodes ('=3) N, No. of the Nodes
Fig.4 (Left) Analytical results of switch sizes as functions of networks size N, '=3.
Fig.5 (Right) Analytical results of normalized optimal node degrees as functions of N with traffic unit t=1. The fiber to switching cost
ratio D/E is set at 40.
3
TuH2
Using tunable optical transceivers for reducing the number

of ports in WDM/TDM networks
Eytan Modiano
Lab. for Information and Decision Systems, MIT, Cambridge, MA 02139, USA
modiano@mit.edu
Randall Berry
Dept. of ECE, Northwestern Univ., Evanston, IL 60208, USA
rberry@ece.northwestern.edu
Abstract: We consider the benefits of using tunable transceivers for reducing the required number
of ports in WDM/TDM optical networks. We show that tunable transceivers can be used to effi-
ciently “groom” subwavelength traffic and significantly reduce the number of ports compared to the
fixed tuned case. We provide a new formulation for this “tunable grooming” problem and develop
algorithms for designing such networks.
2003
c Optical Society of America
OCIS codes: (060.4250) Networks; (060.0060) Fiber optics and optical communications
1. Introduction
In a WDM/TDM optical network each fiber link supports multiple wavelength channels operating at a given bit rate,
e.g., 2.5 Gbps (OC-48) and sub-wavelength traffic is time-division multiplexed onto a wavelength. A significant design
consideration for such networks is reducing the number of ports required at each node in the network, where a port
refers to the combination of optical transceivers and electronic terminal equipment needed to access a wavelength.
There has been much interest in the reducing this requirement by efficiently grooming the low rate traffic so that only
a subset of the available wavelengths must be electronically processed at any node, while the remaining wavelengths
optically bypass the node. Most of the work on grooming has focused on the case where optical transceivers are
fixed tuned and so a fixed subset of wavelengths are dropped at a each node; each dropped wavelength requiring
an electronic port (e.g. a SONET ADM). The basic traffic grooming problem, as in [1-5], is then to assign a given
traffic requirement to wavelengths so that the total number of needed ports is minimized. The general traffic grooming
problem has been shown to be NP-complete [1]; however, optimal algorithms have been found for several special
cases, and a variety of heuristic algorithms have also been presented.
We consider a different approach to designing WDM/TDM networks based on using tunable optical transceivers,
where these transceivers can be tuned from TDM time-slot to time-slot. This complements work on reconfigurable
WDM networks, where tunable components are used to change the virtual topology in response to traffic variations
or for protection purposes. One goal of this work is to highlight a different advantage of tunable components - they
can be used to to significantly reduce the required number of ports over an architecture with fixed tuned transceivers,
even with a static traffic requirement. This approach requires transceivers to be tunable from time-slot to time-slot.
Hence, with time-slots on the order of µs, these devices must be able to tune in sub-µs time. Presently fast-tunable
transceivers are much more costly than their fixed tuned counterparts. It is reasonable to expect that as demand for
tunable components increases their cost will continue to drop. Also, as we show in this paper, the use of tunable
transceivers can reduce the total required amount of hardware in the network (both optical and electronic); this savings
may justify their use.
2. Network Model
Consider a network with N nodes numbered 1, . . . , N . On each wavelength in the network, up to g low-rate circuits
can be time division multiplexed; the parameter g is referred to as the traffic granularity. A static traffic requirement
for the network is given by an N × N matrix R = [Ri,j ], where Ri,j indicates the number of circuits required from
node i to node j. For simplicity, we assume that all traffic requirements are symmetric , i.e., Ri,j = Rj,i for all i, j;
this represents the case where all connections are bi-directional. Each node i generates Wi = j Ri,j /g (fractional)
wavelengths of traffic. Hence, to support the traffic requirement, node i must have at least Wi optical transceivers.
Also for simplicity, we focus in this paper on the case of unidirectional rings; although, many of our results are
applicable to general network topologies. Let Wmin denote the minimum number of (fractional) wavelengths needed
to support the given traffic requirement. In a unidirectional ring with symmetric
traffic, each
Nsymmetric traffic demand
Ri,j = Rj,i uses exactly Ri,j circuits around the ring, and so Wmin = i=j Ri,j /2g = i=1 Wi /2.
TuH2
300
Tuning, no wavelength limit
Tuning, wavelength limit
Fixed−tuned
250
λ1…λw T unable λ1…λw
Optical ADM
200
Drop Add
ports
150
T unable T unable
receiver Laser
100
O/E E/O 50
User equipment 0
6 8 10 12 14 16 18 20 22 24
N
Fig. 1: An example node with tunable transceiver. Fig. 2: Number of ports vs. N for a ring with uniform demand of r = 1
circuits and g = 4.
Each node in the network is assumed to have a set of tunable transceivers. An example of such a node is shown
in Fig. 1; however, many different implementations are possible. As shown, each tunable transceiver consists of a
tunable optical ADM, a tunable receiver and a tunable laser. In addition, the node must also be equipped with optical-
to-electrical (OE) and electrical-to-optical (EO) converters. The specific implementation of such tunable transceivers
are not of interest in this paper. However, we require the nodes to be synchronized at the slot level. Moreover, in a ring
network as a slot propagates around the ring it should return to its source on a slot boundary. This can be taken care of
by adding extra delay with fiber delay lines or by framing transmissions so that synchronization is maintained. In the
following we use the term “port” to refer to all of the optical and electronic equipment required to receive and transmit
on one wavelength. From the above discussion, for symmetric traffic, each node requires at least Wi tunable ports.
To illustrate the potential advantages of tunablity consider the following simple example of a unidirectional ring
with N = 4 nodes, g = 3, and assume there is a uniform demand of one circuit between every pair of nodes,
i.e. Ri,j = 1 for all i = j. In this case the minimum number of wavelengths, Wmin = 2 and there is a total of
N (N − 1) = 12 circuits that need to be assigned to the wavelength. With g = 3, as many as 6 circuits can be
assigned to each wavelength; this can be accomplished by assigning both circuits for each duplex connection to the
same time-slot. The traffic demand can then be supported by finding an assignment of each duplex connection to
one of the g time-slots in the TDM frame, on one of the wavelengths in the ring. Without the possibility of tunable
transceivers the assignment of circuits to wavelengths corresponds to the standard traffic grooming problem; for which
an optimal solution is given in Table 1 requiring a total of 7 transceivers. However, if nodes are equipped with tunable
transceivers, then the number of transceivers can be further reduced. For example, notice in the traffic assignment in
Table 1 node 3 only transmits and receives on one wavelength at any time (i.e. λ2 in slots 1 and 3 and λ1 in slot 2).
Hence if node 3 were equipped with a tunable transceiver, it would only need one transceiver, rather than two. An
optimal traffic assignment using tunable transceivers is shown in Table 2; this requires each node to only transmit on
one wavelength during each slot and hence each node need be equipped with a single tunable transceiver.
Table 1: Optimal traffic assignment for fixed tuned Table 2: Optimal traffic assignment with tunable
transceivers. transceivers.
λ1 λ2 λ1 λ2
Slot 1 (1-2) (2-3) Slot 1 (1-2) (3-4)
Slot 2 (1-3) (2-4) Slot 2 (1-3) (2-4)
Slot 3 (1-4) (3-4) Slot 3 (1-4) (2-3)
This example shows that the number of transceivers can be reduced from 7 to 4 by proper slot assignment. In this
case the optimal assignment can be found by inspection; however, as we show next, in larger networks this can be a
non-trivial combinatorial problem.
3. Minimum tunable transceiver provisioning

We consider finding a time-slot assignment that minimizes the number of tunable ports needed for a ring with a given
traffic requirement R = [Ri,j ] and W ≥ Wmin available wavelengths. We refer to this as the minimum tunable
TuH2
port (MTP) problem. A solution to this requires specifying the number of ports at each node and specifying which
wavelength each port must be tuned to during each time-slot, as in table 2. This problem can be formulated as an ILP
N
where the objective is to minimize the i=1 Xi , where Xi is an integer variable indicating the number of transceivers
at node i. The constraints are given by a set of linear relations representing the conditions that (1) the traffic demand
is satisfied, (2) no node can transmit or receive on more wavelengths than it has ports, and (3) each time-slot on each
wavelength is not used more than once on any link in the ring. As stated next, this problem is in general NP-complete.
Theorem 1 The MTP problem with W = Wmin available wavelengths is NP-complete.
This can be shown by transforming the MTP problem into a graph edge-coloring problem [6]. Though the general
MTP problem is NP-complete, the optimal solution can be found in several important cases. First we consider the case
where there are sufficient available wavelengths so that the wavelength limitation is not binding.
Theorem 2 If a network has no wavelength limitations, then each node requires Wi tunable ports. Moreover, a
time-slot allocation that achieves this can be found in polynomial time.
Notice that Wi tunable ports is the minimum number of ports required to support node i’s traffic. Thus, removing the
wavelength limitation allows each node to use the minimum possible number of ports and lets us solve an otherwise
NP-complete problem in polynomial time. The proof of this also uses a correspondence with an edge coloring problem,
however here the coloring is performed on a bipartite graph representing unidirectional circuits. In a bipartite graph an
optimal edge coloring can be found in polynomial time [7]. The above solution applies to arbitrary network topologies
(not necessarily rings).
When the wavelengths are limited, the above time-slot allocation will no longer be feasible and the circuits must
be packed more efficiently onto the available wavelengths. In this case the network topology is important. First
we consider a unidirectional ring with a uniform traffic requirement of r circuits between each pair of nodes and a
wavelength limit of Wmin wavelengths. In this case if the number of nodes is even, then each node again need only
be equipped with the minimum number of transceivers.
Theorem 3 In a ring with uniform traffic, N even, and Wmin wavelengths, each node requires Wi tunable ports.
Moreover, an optimal time-slot allocation can be found in polynomial time.
For a general traffic requirement, including the case of uniform traffic with N odd, we have developed heuristic
algorithms for approximating the solution to the MTP problem. Moreover, these heuristics have bounded approxi-
mation error and exhibit good performance. An example of the performance of these algorithms is shown in Fig. 2.
This figure shows the number of ports in a ring with g = 4 and a uniform demand of r = 1 circuits for different
values of N . Three curves are shown. The top curve is a lower bound on the number of ports required in a ring with
fixed-tuned transceivers given in [1]. The middle curve is the number of ports needed with tunable transceivers and
Wmin wavelengths. When N is even this is given by Theorem 3; when N is odd, our heuristic algorithms are used.
The bottom curve is the number of tunable ports needed without any wavelength restrictions, as in Theorem 2. In
the case with tunablity, the number of ports can be reduced by over 40% as compared to the lower-bound from [1]
with fixed tuned transceivers. Also note that there is little difference between the case with wavelength limitation and
without. Similar performance is attained for other parameters.
4. Conclusions
We show that using tunable transceivers can significantly reduce the required hardware in a WDM/TDM network. As
the cost and capabilities of optical hardware improve, the ability to trade-off additional complexity in optical hardware
for a significant reduction in electronic hardware may become extremely beneficial.
References
[1] E. Modiano and A. Chiu, ”Traffic Grooming Algorithms for Minimizing Electronic Multiplexing Costs in Unidirectional SONET/WDM Ring
Networks,” CISS 98, Princeton, NJ, Feb., 1998. Extended version appeared in IEEE J.L.T., Jan., 2000.
[2] J. Simmons, A. Saleh, “Quantifying the benefit of wavelength add-drop in WDM rings with distance-independent and dependent traffic,” IEEE
JLT, vol. 17, pp. 48–57, Jan. 1999.
[3] O. Gerstel, P. Lin, and G. Sasaki, “Combined WDM and SONET network design,” Proc. Infocom, New York, Mar. 1999.
[4] K. Zhu and B. Mukherjee, ”Traffic grooming in an optical WDM mesh network”, IEEE JSAC, January, 2002.
[5] E. Modiano and P. Lin, ”Traffic grooming in WDM networks”, IEEE Communications Magazine, July, 2001.
[6] I. Holyar, “The NP-completeness of edge-coloring,” SIAM J. on Computing, vol. 10, no. 4, 718-720, 1981.
[7] H. N. Gabow, O. Kariv, “Algorithms for edge coloring bipartite graphs and multigraphs,” SIAM J. on Comp. 11(1982),117-129.
TuH3
Asymmetric Reconfigurable OADMs for next generation

Metro-DWDM networks
Valerio Viscardi and Gianpaolo Barozzi
Photonics Technology Unit, Cisco Systems, Inc., Via Philips 12, 20052 Monza, Italy
vviscard@cisco.com - gbarozzi@cisco.com
Ori Gerstel
Advanced planning and Technology, Cisco Systems, Inc., 275 East Tasman Drive, San Jose, California, U.S.A
ogerstel@cisco.com
Abstract: Reconfigurable OADMs allow responding to unpredictable traffic demand. We propose

ROADM structures with asymmetric flexibility at the add vs. drop section and a metric to evaluate
their flexibility and their applications in Metro-DWDM optical networks.
OCIS codes: (060.4250) Networks; (060.1810) Couplers, switches, and multiplexers.
Introduction
Wavelength Division Multiplexing (WDM) technologies has proven to be a successful choice for the deployment of
a high speed communication backbone. The optical layer is suitable for carrying different kinds of traffic: from
traditional voice services to IP traffic, but its main purpose in current installed networks is to provide physical media
connectivity. The next step in the evolution of the optical layer is the capability to provide advanced functionalities,
such as automatic wavelength provisioning and optical restoration.
It has been shown that transparent pass-through (e.g. by means of OADMs) provides cost benefits when
compared to nodes in which the entire WDM is processed electrically [1]. But this advantage, if obtained with
current fixed optical Add/Drops, is counter-balanced by limitations in overall network flexibility: a wrong traffic
growth estimate or a deviation from the estimated path can put the network in a blocking state and require a
complete connections re-shuffle. These limitations can be removed by inserting reconfigurable Optical Network
Elements (ONEs) into the network.
Network Growth and ROADM Design

It is important to understand how optical networks evolve in time in order to effectively design Optical Network
Elements (ROADMs or OXCs). To evaluate network flexibility for architectures and technologies that are very
different we need first a metric that allows defining and quantifying ONEs flexibility irrespective of the architectural
details. A good starting point is Routing Power as defined in [2]. For this analysis we only need the Routing Power
“R” and the Segmented Routing Power on a single port “R1”. R is a parameter that expresses the ONE capacity to
Add/Drop any wavelength combination at node level. In a similar way R1 is related to the combination of
wavelengths that can be dropped on a given physical port.
R is a parameter that is related at network level to node-to-node connectivity, i.e. the capability to reach a node
on a given wavelength. In a similar way, R1 is a measure of port-to-port connectivity, i.e. the capability to reach
client equipment (e.g. a SONET ADM or an IP Router) connected to a given physical ONE port on a given
wavelength.
In order to understand how these two parameters are related to network traffic growth it is useful to consider
two possible scenarios:
1. Network growth without churn: new traffic is simply added;
2. Network with churn: new traffic is composed by a re-routing of old connections and by new ones.
Case 1 requires the network to maximize the probability to find a free wavelength to interconnect two nodes.
This requires that at least one wavelength is free and that that wavelength can be dropped by both the nodes.
ROADM architecture has to maximize node-to-node interconnection: therefore it has to be designed in order to
maximize R. Case 2 is similar to 1, but since there is wavelength re-routing, it is important to maximize the re-use of
installed equipment, without changing its physical interconnection to the WDM layer. In order to do so, the
ROADM architecture has to maximize R1, or port-to-port connectivity.
Of course, an ideal ROADM that maximizes both R and R1 would be the answer to any network scenario, but
that would come with high costs. The real push for re-configurability comes from the requirement to adapt the
network to unforeseen traffic demands. Since much of the network churn is typically handled in the electrical layer,
TuH3
the bandwidth required in the optical layer grows more moderately and predictably – especially in Metro-Regional
Networks. Therefore, when a new connection is provisioned, it is very unlikely that it will be removed or rerouted
(not referring to optical protection). As a result, maximizing the number of wavelengths accessible from a given
physical port (R1) is not as important as maximizing the Add/Drop combinations at ONE level (R).
Transmit vs. Receive Tunability

R and R1 as defined in [2] are based on the hypothesis that the ROADM drop section and add section have the same
flexibility. But this is not always the case. It is possible to design asymmetric architectures that exploit the
functionality of a given technology, such as tunable filter (see Fig. 1a) or tunable lasers (see Fig. 1b) without having
to use both. These architectures could provide enough overall network flexibility for certain applications. For
example, in an access network, where all traffic flows to a hub, one could dedicate a wavelength to every non-hub
node and use tunable lasers in the hub to reach the various nodes, without tunability on the drop side (architecture
(b) in the Fig.1).
Fig.1. Block diagram of: asymmetric ROADM with Drop tunability a); asymmetric ROADM with Add tunability b);
ROADM with Add/Drop tunability
Fig.2. Block diagram of an asymmetric ROADM; second band filter (i.e. Band B) is optional and required if the A/D
capacity is up to 8 channels.
Asymmetric ROADM can still be analyzed using the metric presented in [2], but with minor changes. First, we
introduce R1*, which is a R1 normalization in order to have R1* = 1 for a ROADM able to access any wavelength
to/from a given physical port (1),
* R1
R1 =
log 10 ( N + 1) (1)
N ⋅ log 10 (2)
This normalization provides more readable values that simplify architectures comparisons. By definition [2], R1
< R2 < Ri < R and therefore R1 can be very small even for a fully reconfigurable OADM. As the latter can drop any
TuH3
of the N wavelengths of a DWDM system or none, the maximum number of Add/Drop states on a single physical
port is N+1.
Second, R and R1* are now referred either to the Add section (RADD, R1*ADD) or to the Drop section (RDROP,
R1*DROP): the overall parameters are obtained as an average of the two (2),
R ADD + RDROP R1*ADD + R1*DROP

RTOT = R1*TOT = (2)
2 2
With these new derived parameters, it is possible to analyze asymmetric architectures. Flexibility performances
of ROADM architectures of Fig. 1 and 2 are shown in Table 1. The table reports also the number of dropped
wavelengths (K) over total wavelengths (N) ratio. It is worth noting that Fig. 1a and 1b provides the same overall
amount of flexibility, by exploiting flexibility either in the Drop section or in the Add section. The intuition behind
this equivalence is that one can dedicate a wavelength to the source (in the 1b case) or to the destination (in the 1a
case) and as long as the peer can tune to that wavelength, the solution is fully flexible. Fig. 1c is a composition of
the two and shows, of course, better overall performances. This is because it can redistribute how wavelengths are
assigned to nodes, whereas solutions 1a and 1b have fixed assignments of wavelengths to sites. Solutions 1a and 1b
cannot in fact guarantee full node interconnections if some of the wavelengths are used by intermediate nodes.
Solution 1c can overcome this condition, as it provides the capability to shift both the nodes on an unused
wavelength. Fig. 2 shows a scalable asymmetric architecture that can grow up to 8 channels (K/N = 0.25)
Table1. R and R1* values for the architectures shown in Fig. 1 and 2; N=32.
R R1*
K/N
RADD RDROP RTOT R1*ADD R1*DROP R1*TOT
Fig. 1a 0 0.48 0.24 0 1 0.5 0.125
Fig. 1b 0.48 0 0.24 1 0 0.5 0.125
Fig. 1c 0.48 0.48 0.48 1 1 1 0.125
Fig. 2 0.74 0 0.37 1 0 0.5 0.250
Flexibility vs. Remote Reconfigurability

It is worth noting that an optical network based on fixed OADM can still provide network flexibility. This flexibility
involves manual installation of patch-cords, but is hitless on existing traffic. It is therefore a cost effective way to
answer to unpredicted traffic demands, as long as the timescales for such changes are long enough (which is the case
today). Remote configurability is on the contrary the capability to reconfigure the network from a remote
management station or automatically. One obvious difference is the need to pre-deploy OEOs in the second case in
order to exploit the automation. Another difference is related to the relative importance of R and R1. While R1 is
important in reconfigurable nodes with pre-deployed OEO, it becomes irrelevant for flexible nodes, since client
equipment is manually connected to the port on which the wavelength is add/dropped.
Conclusions
In this paper we have shown the importance of network traffic growth scenarios in the architecture design of
Reconfigurable OADM. The major push for reconfigurable ONEs is the capacity of answer, at node level, to
unpredictable traffic demands and this can be obtained maximizing R parameter. R1 has minor importance due to the
low (or zero) churn level of DWDM networks and due to the fact that manual flexibility is typically sufficient. We
have also shown a method to analyze and evaluate asymmetric ROADM architectures. Using this metric we have
demonstrated the equivalence between Add section flexibility and Drop section flexibility. It should be noted that
Drop section flexibility has an advantage due to the relative cost of tunable filters vs. tunable lasers. These
asymmetric ROADM structures provide a good trade-off between flexibility requirements and ONE overall cost.
References
[1] Simmons, J.M.; Saleh, A.A.M, “The value of optical bypass in reducing router size in gigabit networks”, in ICC '99 (1999 IEEE International
Conference on Communications), Volume: 1, (6-10 June 1999), Page(s): 591 -596
[2] Feuer, M.D.; Al-Salameh, D., “Routing power: a metric for reconfigurable wavelength add/drops”, in OFC 2002 (Optical Fiber
Communication Conference and Exhibit), (17-22 March 2002), Page(s): 156 -158
TuH4
A New Optical Network Architecture that Exploits Joint Time and Wavelength
Interleaving
Iraj Saniee and Indra Widjaja

Bell Labs, Lucent Technologies
600 Mountain Avenue, Murray Hill, NJ 07974
E-mail: {iis, iwidjaja}@research.bell-labs.com
Abstract: This paper presents an optical network that provides fractional wavelength services without
optical-to-electronic conversion. The architecture emulates fast switching in the passive network core
through the use of ultra-fast wavelength tunable lasers at the edge.
1. Introduction
End-to-end applications almost always require only a small fraction of the capacity of a wavelength.
Optical Circuit Switching (OCS) with wavelength granularity is therefore not economically viable unless
accompanied by optical-to-electronic conversion and traffic grooming. While electronic traffic grooming
can effectively increase network efficiency, per-unit cost of bandwidth cannot be kept low due to the
conversion and the electronic processing in the network.
Fractional wavelength switching performed entirely in an optical domain has gained interest due to its
potential in lowering the transmission cost. One prominent example is the Optical Burst Switching (OBS)
[1]. OBS improves bandwidth efficiency over OCS through statistical multiplexing in an optical domain.
However, OBS requires fast switching and header processing at each node, and may also require optical
buffering for better performance. Fast switching and optical buffering currently present significant
technological challenges.
Ultra-fast tunable lasers (UFTLs) that can switch wavelengths in nanoseconds, wavelength-selective cross-
connects (WSXCs) and burst-mode receivers are becoming commercially viable. We propose an optical
network architecture, Time-domain Wavelength Interleaved Networking (TWIN), that exploits UFTLs to
emulate fast switching (micro-second packet length) in the network core, WSXCs to perform self-routing
of optical signals (bursts) and burst-mode receivers to synchronize receiver’s clock to incoming signal in
tens of nanoseconds. This architecture enables cost-effective transfer of information across a fiber link,
typically containing tens to hundreds of wavelength each operating at the rate of 2.5 Gbps to 40 Gbps.
TWIN improves bandwidth efficiency over OCS while avoiding the technological barriers inherent in OBS.
Fast switching in the network core is avoided by rapid tuning of wavelengths at the sources. Tunable lasers
and burst-mode receivers at the network edge enable fractional services to applications without optical-to-
electronic conversion inside the network. Furthermore, WSXCs enable passive routing of optical signals to
intended destinations. Table 1 summarizes the main characteristics of the three network architectures
described above.
Table 1 Comparison among different optical network architectures.
Optical Bandwidth Core switching Optical Data Response to
network efficiency speed buffering in processing in dynamic
architecture requirement the core the core traffic
OCS Low Low No No Slow
OBS High High Possibly Yes Fast
TWIN High Low No No Fast
2. Network Architecture
Figure 1 shows the architecture of TWIN. Each source is equipped with a fast tunable laser and each
destination is assigned a unique (set of) wavelength(s). When a source has data to send to a destination, the
source tunes its laser to the wavelength assigned to that destination for the duration of the data
transmission. Each intermediate node performs self-routing of optical bursts without buffering to the
intended destination based solely on the wavelength of the burst. Self-routing is effected through use of
TuH4
WSXCs. No label/address lookup processing is needed in forwarding bursts from one node to another,
thereby making the network core transparent and simple. One can implement an intermediate node with a
passive combiner and a 1xK WSXC [2]. The intermediate nodes are pre-configured so that any incoming
optical signal of a given wavelength will be routed to the appropriate destination. One example is to pre-
configure the routes that form an optical multipoint-to-point tree for each destination, as shown in Figure 1.
Since there is no buffering in the network core, each source needs to perform media-access control for burst
transmission. TWIN uses a scheduling approach to arbitrate burst transmissions at tunable lasers and ensure
that conflicts (potential burst overlaps) do not occur in the network. The multipoint-to-point routing
considerably simplifies scheduling, as conflicts should be observed only at the transmitters and receivers,
but not in the network core. One important objective of the scheduler is to maximize the achievable
throughput of the network. In the context of a single node, switch scheduling in an input-queued crossbar
switch has been investigated extensively in the past (e.g., [3]). However, TWIN requires network
scheduling that deals with the entire network consisting of multiple nodes. More importantly, network
scheduling also has to explicitly take into account the propagation delays between various nodes.
Intermediate
Destination
Source
Figure 1 TWIN architecture consisting of destination-based optical trees.

3. Network Scheduling
Network scheduling can be centralized or distributed. A centralized scheduler knows the traffic demand
information for each node pair and uses this information to compute the schedules for all node pairs. A
centralized scheduler can be effective when the traffic demand information is quasi-static, but may not be
able to respond to fast changes in traffic demand, as it needs constant communications to all nodes. A
distributed scheduler resides at each destination and only computes the schedules for sources that have
burst transmissions to it. In general, a distributed scheduler can better respond to dynamic changes to traffic
demand than a centralized one. Centralized scheduling has been proposed in [4]. This paper presents a
distributed scheduling protocol. We assume that burst transmissions are contained in repetitive scheduling
cycles.
Figure 2 Blocking probability versus number of sources as the batch size (number of transmissions
per source) varies.
In its simplest form, distributed scheduling involves uncoordinated transmissions of bursts among sources.
Each source, however, ensures that the bursts it transmits are not scheduled for transmission at the same
time. The performance of this simple scheduler can be obtained by means of analysis or simulation (due to
space constraint, the analysis cannot be provided in this paper). Figure 2 shows the performance of this
scheduler in terms of burst blocking probability with a scheduling cycle of length B =150. Consider a
particular tagged destination. Each of the N sources is assumed to transmit d bursts in a cycle to the tagged
TuH4
destination. Some of these bursts will be blocked when they collide at the destination. The blocking
probability is defined as the ratio of the number of bursts that are blocked to the total number of bursts
transmitted by all sources (Nd). When Nd is fixed, we expect that the number of bursts that are blocked
increases as N increases, as shown in the figure. When Nd = 150 (or ρ =1.0), observe that the blocking
probability is about 35% as N becomes large. This also corresponds to the case when a source is completely
uncoordinated. Note that the blocking probability drops to about 5% when Nd = 16 (ρ = 0.11).
We now describe an improvement to the distributed scheduler given above when delays in the order of one
round-trip time can be tolerated. The protocol relies on request and grant message exchanges to
communicate schedules between a source and a destination, learns when conflicts occur, and reassign time
slots upon learning a conflict. The basic idea of the protocol is illustrated in Figure 3. In this example, the
source initially requests burst transmission at the rate of three bursts per cycle. Upon receiving the request,
the destination grants time slots 1, 6 and 9 for the source to use in the subsequent cycles. Upon receiving
the grant, the source checks whether it can transmit the bursts at the designated departure times. Here, it is
assumed that the source can only transmit bursts on time slots 1 and 9, but not on 6 because of a conflict
with another transmission (to a different destination). The destination learns about the conflict when an
allocated time slot (which is 6 in this case) does not contain a burst. Upon detection of a conflict, the
destination reassigns time slot 8 to the source. It is assumed that time slot 8 does not create another conflict,
and thus the source eventually transmits at its desired rate of three bursts per cycle.
Source Destination
Request 3
time slots
Conflict ts 1, 6, 9
on slot 6 Grant slo
Learn
Transmits conflict
on slots 1,
9 on slot 6
nge slot 6 to 8
Cha
Transmits
on slots 1,
9
Time
Transmits
on slots 1, 8, 9
Figure 3 Basic operation of the distributed Figure 4 Requests and grants as function of
protocol with learning. time.
It is of interest to evaluate the performance of the distributed protocol under dynamic traffic (bandwidth
requests change over time). Figure 4 shows the dynamics of the distributed protocol with learning. Notice
that the grants track the dynamics of the requests quite well, demonstrating the responsiveness of protocol
with dynamic traffic. Due to space limitations, the description of the distributed protocol is necessarily
brief. Other performance issues of interest will be presented at the conference.
4. Conclusion
We have presented an all-optical network architecture that provides fractional wavelength services by
exploiting joint time and wavelength interleaving. The architecture does not rely on fast optical packet
switching or buffering in the network core. These functions are emulated at the network edge through the
use of ultra-fast wavelength tunable lasers. We have also described a distributed scheduler that can arbitrate
bursts in the network and respond to dynamic traffic.
1. M. Yoo and C. Qiao, “A novel switching paradigm for buffer-less WDM networks,” OFC 1999, paper
ThM6, pp. 177-179, 1999.
2. D. M. Marom et al, "Wavelength-selective 1x4 switch for 128 WDM channels at 50 GHz spacing,"
OFC2002, postdeadline paper FB7, Los Angeles.
3. N. McKeown, “The iSLIP scheduling algorithm for input-queue switches,” IEEE Trans. on
Networking, vol. 7, pp. 188-201, Apr. 1999.
4. K. Ross, N. Bambos, K. Kumaran, I. Saniee and I. Widjaja, “Scheduling Bursts in Time-Domain
Wavelength Interleaved Networks,” to appear in IEEE J. on Selected Areas in Communications.
TuH5
Blocking Probability of a Centralized Switch with Shared

Wavelength Conversion
Aradhana Narula-Tam, Mark H. Brady, Steven G. Finn
MIT Lincoln Laboratory, 244 Wood Street, Lexington, MA 02420
781-981-0210, 781-981-2874, arad@ll.mit.edu
Philip J. Lin
Draper Laboratory, 555 Technology Square, Cambridge, MA 02139
Abstract: This paper presents a new blocking probability model for a centralized switch with shared
wavelength conversion. The ingress and egress link wavelengths are selectable. A small number of con-
verters achieves the performance of full conversion.
c 2003 Optical Society of America
OCIS codes: (060.4250) Networks; (060.4510) Optical Communication
1 Introduction
Wavelength converters in high speed WDM networks can reduce or eliminate the wavelength continuity requirement.
We investigate the number of shared wavelength converters needed to achieve the blocking probability performance of
full wavelength conversion. We consider a centralized switch network topology where N access stations are connected
via optical fiber links to a hub optical wavelength crossconnect equipped with a set of c shared wavelength converters.
Each connection must traverse two fibers, denoted as the ingress fiber to and the egress fiber from the hub crossconnect.
Multiple connection requests are generated at each access node and they can be carried on any wavelength. If a
wavelength converter is employed, the wavelength used on the egress fiber may be different from that used on the
ingress fiber. A new analytical approximation for blocking probability under shared wavelength conversion is presented
and verified via simulations. It is shown that a limited number of wavelength converters, far fewer than the switch size,
are sufficient to make blocking probability due to wavelength mismatch negligible.
It has been shown that a relatively small amount of conversion is sufficient for some network architectures to
perform close to the capability of networks with full wavelength conversion[1],[2]. Previous studies assume large
networks, hence their techniques are not readily applicable to a single switch network. For example most studies on
sparse wavelength conversion assume some optical crossconnect switches are equipped with full wavelength conver-
sion while others have no wavelength conversion capabilities, e.g. [1]. We consider a shared bank of wavelength
converters on the single crossconnect switch. Mitra et al. [2] also consider a shared bank of wavelength converters on
each crossconnect switch, however, they assume that calls arrive on a random wavelength on the ingress link, whereas
in our model the wavelength on both the ingress and egress links is selectable.
Several approximations for blocking probability comparing full wavelength conversion to no wavelength conver-
sion have been formulated [3], [4]. In this work we generalize the analysis in [3] to include the benefits of shared
wavelength conversion. Our model assumes that both the link load and wavelength utilization statistics of each link
are mutually independent. Although the independence assumption is inaccurate for small switches, the error is propor-
tional to 1=∆ [4], where ∆ is the number of ports; hence it is appropriate for large switches as we consider here.
2 Network Model
In the centralized switch network topology, each connection requires one wavelength on each link. Each physical
link carries W wavelengths. Access nodes are equipped with W tunable transmitters and receivers and are capable
of generating and and receiving up to W signals. Connection requests arriving at each access node are modeled as a
Poisson process. Connections are blocked if they can not be established and do not reenter the system. The wavelength
assignment of each connection is fixed for its duration, which is exponentially distributed.
Shared wavelength conversion can be implemented by increasing the crossconnect switch size from NW NW to
(NW + c) (NW + c) optical switch and adding a bank of c wavelength converters. The shared wavelength converters
are connected in feedback fashion so that an arriving signal can be converted and then fed back to the input and
switched to the appropriate output. All NW connections share the bank of c converters.
Without wavelength conversion, each connection must utilize the same wavelength on both links between the
source and destination access nodes. If each source access node knows which wavelengths are available on each link
that the connection traverses, the connection can be established without a wavelength converter whenever a common
wavelength is free on both links of the path. However, disseminating up-to-date wavelength utilization information
requires overhead and complexity. If each access node is only aware of the wavelength utilization on the ingress
physical link to the switch, then a wavelength must be selected on the first link without knowing which wavelengths
are available on the second link. We investigate the benefits of shared wavelength conversion in both scenarios, i.e., (1)
Coordinated wavelength assignment: which assumes the source node is aware of network wavelength assignment on
TuH5
all links of the path and (2)Autonomous wavelength assignment: which assumes the source node only has information
about the wavelength assignment on the ingress fiber.
3 Blocking Probability Analysis and Results
Call requests arrive at each access node with rate λin and are destined for another (random) access node. For the cases of
limited or no wavelength conversion, the wavelength utilized to establish the connection is randomly selected from the
set of wavelengths available on both the ingress and egress fiber links. If there is no single wavelength that is available
on both links, but disjoint wavelengths are available on the ingress and egress links and a wavelength converter is
available, the wavelength on the ingress and egress links is randomly selected from the set of wavelengths available on
each respective link and a converter is used to establish the connection. We also simulated other wavelength assignment
strategies such as first fit and found the results were similar to the random wavelength assignment strategy.
We consider a crossconnect equipped with a bank of c shared wavelength converters. Thus up to c of the NW
connections between the N nodes can use different wavelengths on the ingress and egress fibers. A connection is
blocked for one of three reasons: (1) there are no wavelengths available on the ingress link (input blocking), (2) there
are no wavelengths available on the egress link (output blocking), or (3) the sets of wavelengths available on the ingress
and egress links are disjoint and no wavelength converter is available (limited converter availability).
Without loss of generality, let T = 1 be the normalized average duration of each connection. Define λ i as the arrival
rate of connections on the ith link of the path. The average load on the i th link of the path is thus Li = λi T = λi . Note
that the arrival rate on the ingress link is not equal to the arrival rate at each access node, i.e., λ 1 6= λin , since blocking
reduces the offered load on each link. We calculate the reduced load L i using the approximation algorithm developed
in [7]. We also make the common approximation that the statistics of the link loads on each link are independent.
3.1 Coordinated Wavelength Assignment
In coordinated wavelength assignment, the wavelength utilization of each link is known. The probability that a wave-
length converter is needed, Pwc is the probability that there is a wavelength available on both the ingress and egress
links, however, the same wavelength is not available on both. Let p busyi (k) be the probability that there are k busy
wavelengths on the ith link. The probability that a wavelength converter is needed is
k

W 1 W 1
Pwc = ∑ ∑ W l
W
pbusy1 (k) pbusy2 (l ) (1)
k=1 l =W k l
where pbusyi (k) = (Li )k =k! , for i = 1 2.

l =0 (Li ) =l!
∑W l
We approximate the bank of c shared wavelength converters as an M =M =c=c queue with offered load weighted
by Pwc . This allows us to use standard queuing results to approximate the probability that all converters are in use.
Specifically, the load entering the bank of converters is the crossconnect load L 1 N multiplied by the probability a
connection requires a wavelength converter, i.e., L wc = L1 NPwc . The probability that a request for a converter will find
all c converters busy is given by the Erlang B formula:
Pconv busy (c) =

(Lwc )c c!
=
(2)
∑cn=0 (Lwc )n n!
=
:
Note that when c = NW , the probability that a request finds all converters busy should be zero although Pconv busy (NW )
is non-zero. However, our results indicate that this approximation error is insignificant.
The overall network blocking probability is the sum of the blocking probability due to channel limitation and due
to lack of wavelength conversion,
PB (c) = 1 (1 pbusy1 (W ))(1 pbusy2 (W ))]+ PwcPconv busy (c): (3)
We compare our analytical approximations for blocking probability to simulations for a centralized switch with
N = 16 nodes and W = 16 wavelengths in Figure 1(a) below as a function of the input load, ρ = λ in T for several values
of c. Analytical curves for full wavelength conversion and no wavelength conversion are computed according to [3].
The analytical computations are very close to simulations, but become slightly less accurate at higher loads because
of the independence assumption. As expected, blocking probability for a network with shared wavelength conversion
increases with load and lies between the blocking probability for a network with full wavelength conversion and that
of a network with no wavelength conversion. Also, we see that a small number of wavelength converters can reduce
blocking probability to the level achieved with full wavelength conversion. For example, at a load of 0.5, only 8
wavelength converters (rather than 256) are sufficient.
TuH5
Coordinated Wavelength Assignment, N=16, W=16 Autonomous Wavelength Assignment, N=16, W=16
ï1
10
ï1
10
ï2
10
ï2
10
ï3
10 Anl. No WC
Anl. 16 WC
Blocking Probability
Blocking Probability
Anl. 32 WC
ï3 Anl. 48 WC
10 ï4 Anl. 64 WC
10
Anl. 80 WC
Anl. No WC Anl. 96 WC
Anl. 2 WC Anl. 112 WC
ï4 Anl. 4 WC Anl Full WC
10 ï5
10
Anl. 8 WC Sim. No WC
Anl. 16 WC Sim. 16 WC
Anl. Full WC Sim. 32 WC
Sim. No WC Sim. 48 WC
ï6
ï5 Sim. 2 WC 10 Sim. 64 WC
10
Sim. 4 WC Sim. 80 WC
Sim. Full WC ï7
10 Sim. Full WC
ï6
10
0.2 0.25 0.3 0.35 0.4 0.45 0.5 0.55 0.6 0.65 0.7 0.2 0.3 0.4 0.5 0.6 0.7
Input Load (ρ) Input Load (ρ)
(a) Coordinated wavelength assignment (b) Autonomous wavelength assignment
Fig. 1. Analytical and simulation results for the blocking probability of a centralized switch with various numbers of shared
wavelength converters. Shared conversion blocking probabilities are compared to the blocking probability of a switch with
no wavelength conversion and full wavelength conversion [3].
3.2 Autonomous Wavelength Assignment

With autonomous wavelength assignment, the wavelength utilization of the second link is unknown at the access node.
Thus the access node randomly selects a wavelength from those available on link 1. The probability that a wavelength
converter is needed, Pwc is the probability that there is at least one wavelength available on links 1 and 2, and the
wavelength selected on link 1 is not available on link 2. The probability that a wavelength converter is needed is
W 1W 1
l
Pwc = ∑ ∑ pbusy1 (k) pbusy2 (l ): (4)
k=0 l =1 W
Again, the overall network blocking probability is the sum of the blocking probability due to input and output blocking
and due to lack of a sufficient number of converters and can be computed by substituting (4) into (3).
The analytical approximations for blocking probability are compared to simulations for a centralized switch with
N = 16 nodes and W = 16 wavelengths in Figure 1(b) as a function of the input load, ρ = λ in T and the number of
converters available. Again the analytical results match the simulations. Note that the blocking probability with full
wavelength conversion is equal under coordinated and autonomous wavelength assignment. As expected, the blocking
probability without wavelength conversion is greater under autonomous wavelength assignment; hence we see a larger
benefit from both shared and full wavelength conversion. However, a larger number of converters are needed to attain
the performance of full wavelength conversion. At a load of 0.5, 96 wavelength converters (out of 256) are sufficient.
Interestingly, under both coordinated and autonomous wavelength assignment, the number of wavelength convert-
ers needed is roughly PB (0)NW . This can be explained by noting that for an M =M =c=c queuing system, the number
of servers needed is approximately equal to the load.
4 Conclusions
We have developed a model that accurately predicts the blocking probability of a centralized switch network. We have
shown that far fewer than NW wavelength converters are needed to attain the benefits of full wavelength conversion
both when full wavelength state information is available (coordinated wavelength assignment) and when wavelength
state is unavailable (autonomous wavelength assignment). In future work we expand our model to investigate shared
wavelength converter requirements in mesh networks.
References
1. S. Subramaniam, M. Azizoglu, and A. K. Somani, “All-optical networks with sparse wavelength conversion,” IEEE Trans. on Net., 1996.
2. D. Mitra, C. Nuzman, and I. Saniee, “Optical crossconnect with shared wavelength conversion under dynamic loading,” in OFC, 2002.
3. M. Kovacevic and A. Acampora, “Benefits of wavelength translation in all-optical clear-channel networks,” IEEE JSAC, 1996.
4. R. Barry and P. Humblet, “Models of blocking probability in all-optical networks w/ and w/o wavelength changers,” IEEE JSAC, 1996.
5. R. Ramaswami and K. N. Sivarajan, Optical Networks: A Practical Perspective, Morgan Kaufmann Publishers,Inc., San Francisco, 1998.
6. S. Rahman and Dan Blumenthal, “Scalable switching in optical networks,” Electronic Products, Feb. 2002.
7. F. Kelly, “Blocking probabilities in large circuit switched networks,” Advances in Applied Probability, vol. 18, pp. 473–505, 1986.
TuH6
Towards a meshed ultra high speed TDM optical network :

concept, OADM architecture and proof of principle
H. Rohde, G. Lehmann, and W. Schairer
Siemens AG, CT IC 2 Otto Hahn Ring 6, 81370 Munich, Germany, Fax: +49 89 636 5115
Harald.Rohde@siemens.com
Abstract: Meshed ultra-high-speed OTDM networks give rise to a couple of challenges concerning the architecture
of network elements. We present an architecture of an OTDM-ADM and its experimental proof of concept.
1. Introduction
Optical data communication networks of the future will operate with much higher transfer capacities per fiber than
current ones. It is still an open issue wether the increase of capacity will be accomplished by a higher number of
wavelengths per fiber or by higher bitrates per wavelength or - most probably - a combination of both.
This paper reviews aspects of ultra high speed OTDM networks as part of future networks. Evolution of todays
networks tends towards meshed networks. The use of meshed ultra-high-speed OTDM networks implies the need of
additional elements in the network nodes to compensate for differential time shifts between different paths through
the network. As these timing elements are not necessary in todays ordinary (D)WDM networks they add another
layer to the architecture of ADM network nodes in the case of additional use of OTDM techniques.
The current paper surveys different ADM techniques together with their qualification in real transmission systems
and discusses the necessary enhancements for the use in a meshed network.
2. Realisation of an OTDM-ADM
In the scope of this work, three realisations of ADMs have been experimentally revised:
x In a Semiconductor-Optical-Amplifier Mach-Zehnder-Interferometer (SOA-MZI) differential nonlinear phase
shifts on the data signal are induced by control laser pulses. Careful adaption of these phase shifts directs the
through channels to one port of the MZI and the dropped channel(s) to the other port of the MZI. A detailled
description can be found in [1].
While the drop functionality was satisfying in the revised setup, the add functionality was not given due to
improper depletion of the dropped channel in the through signal. Further investigations and optimization of the
design of the intergated MZI structure is necessary.
x A Four Wave Mixing based ADM uses the former effect to extract single bits out of the data stream with a control
pulse. With one nonlinear element for the through and the drop channel respectively, ADM functionality can be
shown. First results are very promising [2].
x The Gain-Transparent Ultrafast-Nonlinear-Interferometer (GT-UNI) also uses nonlinar phase shifts induced by
control laser pulses. These phase shifts are translated into polarization rotations. The resulting different polari-
zations are used to separate through and drop channels. For a detailed description of the principle please see [3,4].
First very auspicious results of the GT-UNI will be discussed in the next chapter.
3. Experimental realization of an OTDM-ADM in a network

75 km SSMF DCF 75 km SSMF DCF
160 Gbit/s Transmitter

GT-UNI
OSO
TX Mux OTDM-ADM
OTDM-OADM
10 Gbit/s 10G 160G
2x 2x
TX RX SOA RX
10 Gbit/s 10 Gbit/s 160 Gbit/s
Figure 1 Experimental setup: 160 Gbit/s transmitter, 2×75 km fiber links, OTDM-ADM and another 2×75 km fiber links.
TuH6
Figure 1 presents the experimental setup used for transmission of 160 Gbit/s over a fiber link, OTDM-add-drop
functionality and further transmission over another fiber link. A 10 Gbit/s data stream is multiplexed into a
160 Gbit/s signal and sent over two fiber spans consisting of 75 km SSMF and corresponding DCF. The dispersion at
the end of the two spans is compensated to 100%, the residual dispersion slope is less than ± 0.8 ps/nm2. After
transmission through this 150 km fiber link the signal is fed into a GT-UNI OTDM-ADM. Here one of the 10 Gbit/s
tributaries is coupled out and the empty bit slot is filled with another 10 Gbit/s signal. A SOA-Demultiplexer [2] is
used to demultiplex the signal after the GT-UNI to measure the quality of the single channels. After the GT-UNI the
signal is retransmitted over two further fiber spanss of about 75 km each, again at the end of the link with 100%
dispersion compensation and slope compensation to less than ± 0.8 ps/nm2.
As showed in Figure 2, error free transmission could be achieved for all dropped channels (direclty received with a
10 Gbit/s receiver) as well as for all through channels which are demultiplexed by an FWM based demultiplexer. The
retransmitted signal after the second link is recorded by an Optical Sampling Oscilloscope (OSO) and a clear open
eye pattern is observed.
-4 a
-5
log(Bit Error Rate)
B2B 10G Drop Through Add

-6
b
-7
-8
-9
-10 c
-11
-12
-40 -38 -36 -34 -32 -30 -28 -26 -24 -22 -20 -18 -16
Received optical power (dBm)
Figure 2 Bit error rates of the ADM: Back-to-Back 10Gbit/s reference, dropped 10Gbit/s channels, 10Gbit/s through channels and one 10 Gbit/s
add channel. The OSO pictures show: a) input signal into the ADM, B) dropped channel, c) through channels with added channel after
transmission over another 150 km; for identification purpose a clock signal has been used as add signal..
4. Meshed OTDM networks
As shown in the previous chapter a GT-UNI is well suited as an ADM node in an OTDM network. A series of
concatenated nodes can be linked into a ring structure or – more flexible – into a meshed network. The system under
the scope of this work is a single wavelength Optical Time Domain Multiplexed (OTDM) system with an aggregated
data rate of 160 Gbit/s and a tributary data rate of 10 Gbit/s. The smallest meshed network consists of three network
nodes, named A,B and C, interconnected as shown in Figure 1.
TX 1A OADMA A
ADM OADMB B
ADM
RX 1
A
Through Through
16x10 Gbit/s
Add Drop Add Drop
OADMC C
ADM
TX 2
B RX 2B
Through
Add Drop
Figure 3 Example of a meshed OTDM network with three ADM network nodes.
Transmitter TX 1 sends a 160 Gbit/s data stream over a fiber optical link to network node ADM A where one or
more of the 10 Gbit/s tributaries are dropped. The undropped channels are transmitted to ADM B while the dropped
channels are transmitted to ADM C. In ADM C these channels are added to another data stream collected in
transmitter TX 2. Part of the traffic passing ADM C is sent over another fiber link to ADM B.
The network element ADM B now has to combine the two data streams; one coming from ADM A and another
coming from ADM C. For OTDM network nodes some additional effects in comparison to standard WDM nodes
have to be considered: a typical fiber link may consist of 75 km SSMF. With a typical change of the optical length of
10-7/K a temperature change of 1 K induces a change of the link length of 6 bits at 160 Gbit/s. Two data streams
coming in from different network nodes through two different fiber links will experience different temperature
fluctuations on the links and therefore a substantial shift of the two data streams with respect to each other. Even in
TuH6
an air conditioned laboratory a shift of several bits per minute was observed at the end of a link of 150 km SSMF. In
(D)WDM systems time shifts of the data in one wavelength compared to the data in another wavelength do not effect
system performance as each wavelength is received seperately. This situation changes in a high-bitrate OTDM
system when the different channels are coming from different sources and are multiplexed into a single data stream.
Each channel has to fit very precisely with a tolerance of less than a picosecond into its timeslot. Therefore, a
mechanism to detect and to correct phase shifts due to fluctuating lengths of optical fiber links has to be applied.
Figure 2 schematically shows the resulting structure of an OTDM-ADM for meshed optical networks.
Back to Back OTDM -OADM
incoming data ADM A

OADM out-bound data
Data Timeslot
Through
Control
Monitor
Add Drop
Clock
Variable
delay line Control
W
incoming data
Figure 4 schematic structure of an OTDM-ADM for a meshed network
An OTDM-ADM, like the one described above, is supplemented by a variable delay line (which can be for example
a motorized free space delay line or a fiber wound onto a piezoelectric cristal) for one of the incoming data channels.
After the ADM a timeslot control electronics permanently monitors the correct timing of all bit slots. This
information if fed into a control circuit and used to set the delay line to its optimal position. With the help of that
mechanism phase drifts of the different data streams relative to each other are compensated.
The timeslot monitor can be realized by any phase detector which works at bitrates of 160 Gbit/s, i.e. cascaded
EAMs [5], a nonlinearly driven single EAM [6] or devices based on four wave mixing in SOAs [7] or other
nonlinear materials [8]. The phase detector scans the 160 Gbit/s data signal against the 10 GHz base data rate clock.
Depending on the resolution of the phase detector a signal with distinct peaks for each single channel is generated
which can be used to adjust the delay line of one of the incoming data channels for optimal ADM performance
This enhancement of the architecture of an OTDM-ADM node enables the use of these elements in meshed OTDM
ultra high speed networks.
5. Conclusion
The functionality of an OTDM-ADM network node has been demonstrated. All necessary elements to expand the
node for the use in a meshed OTDM network are avaiable and experimentally tested. Therefore, from the
technological point of view meshed OTDM networks of the future are ready to be realized.
Acknowledgement
The work was supported by European Commission under the IST-2000-28765 project FASHION (ultra Fast
Switching in High-speed OTDM Networks). We would like to thank E. Tangiongga, J. P. Turkiewicz, H. de Waardt
from TU Eindhoven for their contribution to the experiments and helpful discussions.
References
1 M. Heid et al. “160 Gbit/s demultiplexing based on a monolithically integrated Mach-Zehnder interferometer“, ECOC 2001 PD.B.1.8
2 H. Rohde et al. “All-Optical Add/Drop Multiplexer for High Speed Optical Networks based on Four-Wave Mixing in SOA”, ECOC 2003
WE4.P127
3 J.P.Turkiewicz et al., “Simultaneous high speed OTDM add-drop multiplexing using GT-UNI switch”, Electron. Lett., 39, pp. 795 (2003)
4 C. Schubert et al., “ Error-free all-optical add-drop multiplexing at 160 Gbit/s” , Electron. Lett., 39, pp. 1074. (2003)
5 D.T.K. Tong et al., “160Gbit/s clock recovery using electroabsorption modulator based phase locked loop“, Electron. Lett. 36, p.1951 (2000)
6 J.P. Turkiewicz et al., “160 Gbit/s clock recovery using a single unidirectional electroabsorption modulator and its applications in the
transmission experiments”, submitted to OFC 2004
7 Jansen et al, Electron. Lett., OFC 2003, ThO5
8 Y. Fukuchi et al., “All-optical time division demultiplexing of 160 Gbit/s signal using cascaded second order nonlinear effect quasi phase
matched LiNbO3 waveguide device”, Electron. Lett. 39, pp. 768, (2003)
TuH7
An Overview of the Network Global Expectation Model

Steven K. Korotky
Lucent Technologies, 101 Crawfords Corner Road, Holmdel, NJ 07733-3030, USA
Email:skk@lucent.com
(Invited Paper)
Abstract: I describe and illustrate the application of the Network Global Expectation Model, in
which expectation values evaluated over the network provide both exact and approximate analytic
descriptions of the required quantities and variances of network resources and commensurate
costs.

OCIS codes: (060.0060, 060.4250) fiber optics and communications, networks
1. Introduction
Fundamental to the comparison and selection of network architectures and their technological implementations is
the total cost of ownership of the network, which includes the expenses for capital equipment, network operation,
and network management. While operational and management expenses represent the largest share of the total cost
of ownership, capital costs are a considerable and highly visible portion of the initial investment, and so are a very
important factor in the choice of architecture and technology. The Network Global Expectation Model provides a
versatile means to quickly gauge the network equipment needs and costs using only modest computational
resources, and therefore can be of value in achieving the desired performance at minimum capital expense. Here I
provide an overview of the concept, formalism, results, and application of this model, which has recently been
developed and detailed [1].
In the Network Global Expectation Model, expectation values of network variables and functions of these
variables are formally evaluated by averaging over the entire network to establish either exact or approximate
analytic relationships between dependent and independent variables. As such the model constitutes a multi-moment
description of the required quantities of key network and network element resources and commensurate network
costs. This approach naturally and accurately connects the global (network) and local (network element) views of
the communication system, and results for a very wide range of network sizes and large number of variations can be
computed very fast with useful accuracy. Consequently, the model can be used as a tool to gain insight, evaluate
preliminary network designs, establish element feature requirements, determine costs, provide sensitivity analyses,
assess scaling performance, make comparisons, define products, and identify application domains. The
uncomplicated and transparent accounting of network elements, systems, and costs inherent in the model can also
constitute a framework for roadmapping and the cooperative exchange of critical planning information on evolving
network needs across the many sectors of the communication business.
2. Costs and Expectation Values
As the cost of the network for a specified set of features is considered the metric for comparison of architectures and
technologies, the model is constructed from this perspective. The fundamental accounting of network costs may be
expressed as
C T ≡ ¦ ci , (1)
i
where CT is the total network cost and ci is the unit cost of the ith component. The costs contributing to this sum
may be arranged without approximation into subtotals to reflect categories of costs. As a case of interest we
consider link costs and node costs. The summation of costs may subsequently be expressed exactly as expectation
values of cost over the members of the sets of the categories, in this case the links and nodes.
CT = L ¢cl² + N ¢cn² . (2)

TuH7
The global expectation values of ¢cl ² and ¢cn ² are themselves sums of products of expectation values of the number
of network elements and subsystems required for the links and nodes and the costs of these element and subsystems.
cl = ¦ vi l ci (3a)
i
and
cn = ¦ v j n c j . (3b)
j
As the costs of the individual network elements and subsystems are considered known, the objective of the
model is to formulate accurate representations of the expectation values of the number of network elements and
subsystems. The advantage of this approach is that by giving up detailed knowledge of the placement of each
network element and subsystem within the network while retaining an accurate count, an enormous reduction in the
required computation resources and time is achieved by avoiding repeatedly solving the time-consuming routing and
capacity-deployment optimization problem.
3. Network Definition, Variables, and Relationships
To illustrate representative results of the model, following we summarize some of the key input and output network
variables and present without derivation formulae that relate the expectations values of these variables.
We define a communication network as the combination of a network graph, denoted G, consisting of a set of N
nodes {ni} and set of L connecting two-way links, or edges, {li}, and a network traffic. The network graph may be
represented by the symmetric matrix [g] with elements gij . The pair-wise, two-way communication traffic between
terminals located at different nodes may be represented by the symmetric demand matrix [d] with elements dij with
total number of two-way demands D and total ingress/egress traffic T. The primary model input variables are taken
to be G (N,L), D, and T together with the demand model. All other variables of interest may be determined from
these. For specificity, here we consider individual demands to be carried on an individual channel, eg. wavelength.
The traffic demand and channel bit-rate, τ, can be computed exactly as the ratio of the total ingress/egress traffic,
T, and total number of two-way network demands, D, terminating at all nodes. We state
τ ≡ T/D . (4)
The average degree of a node ¢δ² is calculated straightforwardly by summing the number of one-way (directed) links
and by dividing by the number of nodes yielding
¢δ² = 2L/N . (5)
The number of hops between a pair of terminals is defined as the minimum number of links traversed by a demand
between the terminating node pair. The expectation value of the number of hops is over the set of demands and for
minimum hop routing of uniform demand over two-dimensional networks we find that ¢h² is accurately
approximated by
¢h² = [(N - 2)/(¢δ² - 1)]1/2 . (6)
The mean number of channels, ¢Wo², appearing on a link of the network has been derived and is given by
¢Wo² = ¢d² ¢h² / ¢δ² , (7)
where ¢d² is the mean number of one-way demands terminating at a node. This important new result is exact and
valid independent of the demand model. Note, however, the value of ¢h² is implicitly dependent upon the demand
model, as specified above. For uniform unit demand ¢d² = N-1. For one estimate of the variance of Wo we calculate
σ2(Wo) ≤ ¢Wo² [1 - 1/¢h²] . (8)
Extra capacity for restoration is expressed as a fractional increase, ¢κ² , in the total deployed capacity to service
the traffic and provide survivability relative to the case without survivability and using minimum hop routing. For a
heuristic for shared mesh restoration that limited the additional number of hops of the restoration path relative to the
working path, we found that the fractional extra capacity can be estimated by
¢κ²= 2/¢δ² . (9)
TuH7
Recently we have established accurate analytic approximations for the fractional extra capacity for path-disjoint
restoration without routing restrictions for general, nominally planar, mesh networks [2].
The average number of channels on a link including extra capacity for restoration can be expressed as
¢Wκ² ≡ ¢Wo² (1 + ¢κ²) . (10)

The average traffic carried on a link, ¢β², is the product of the mean number of channels on a link ¢W² and the bit-
rate per demand,τ, i.e.
¢β² ≡ ¢W²τ . (11)
The mean number of ports, ¢Pκ², required on a cross-connect to service the working and line-side restoration
channels present at a node is
¢Pκ² = ¢d² + ¢Wo² (1 + ¢κ²) ¢δ² , (12)
which scales as ¢Pκ² ≈ N3/2 for large N. The average traffic handled by a cross-connect ¢χ², measured in bits/second
for example, is computed straightforwardly from the average number of ports ¢P² and the traffic demand bit-rate τ:
¢χ² ≡ ¢P² τ . (13)

Independent of the demand model, the ratio, ¢ρ'², of the number of terminated channels to total (termination + thru)
channels present at a node is given by
¢ρ'² = 2/[1 + ¢h²] . (14)
Considering Eq.'s 6 and 14 we observe that ¢ρ'² scales as 2/√N for large N. It is also worth noting that Eq.14 may be
inverted to express ¢h² as a function of ¢ρ'², viz.
¢h² = [2/¢ρ'² - 1] . (15)
The mean length of a link may be estimated given the geographic area, A, serviced by the network using the
approximation
s ≅ A /( N − 1) . (16)
4. Network Costs and Applications

Given a cost structure for the network elements, the above relationships can be used to compute costs for categories
of the network elements, as well as total network costs. For illustration, examples of rudimentary cost structures, γ,
for transmission line systems, electronic cross-connects and optical cross-connects are γB-s ≈ $30/Gbps/km, γep ≈
$1K/Gbps, and γop ≈ $2.5K/port, respectively. Sample network costs using these forms will be discussed.
5. Summary
Here I have presented an overview of the Network Global Expectation Model, which is a comprehensive and
structured statistical formalism for estimating the number of network elements, network element characteristics, and
costs of communication networks using analytic formulae. The model includes the calculation of both the mean
value and variance of all key network quantities and may be applied to a wide range of topologies, architectures, and
demand profiles. Currently the general approach has been established, applied to single-tier mesh networks and
location-independent demands, and many of the results have been shown to be valid and applicable independent of
the demand model. For uniform demands it as also been shown that the number of nodes, the degrees of the
network nodes, the total ingress/egress traffic, the geographic extent of the network, and the equipment cost
structures are sufficient to estimate the network variables and costs of interest. Either exact or approximate semi-
empirical functions and closed form expressions for the network variables, which are easily incorporated into
software spreadsheet calculators, have been formulated. This methodology and corresponding analytic tool can
quickly provide insight and approximate results for preliminary network evaluation, design, and optimization.
References
1. S.K. Korotky, "Network global expectation model: A statistical formalism for quickly quantifying network needs
and costs," submitted to J. Lightwave Technol., July, 2003.
2. ,M. Bhardwaj, L. McCaughan, S.K. Korotky, and I. Saniee, "Global expectation values of shared restoration
capacity for general mesh networks," submitted to this conference, September, 2003.
TuI2
Low-loss and flat/wide-passband CWDM demultiplexer

using silica-based AWG with multi-mode output waveguides
S. Kamei, Y. Doi, Y. Hida, Y. Inoue, and S. Suzuki
NTT Photonics Laboratories, NTT Corporation
3-1 Morinosato-Wakamiya, Atsugi, Kanagawa, 243-0198, Japan.
E-mail: kamei@aecl.ntt.co.jp
K. Okamoto
NTT R&D Fellow, NTT Corporation
3-1 Morinosato-Wakamiya, Atsugi, Kanagawa, 243-0198, Japan.
Abstract: We demonstrated compact and cost-effective 8-channel filters for unidirectional and
bidirectional CWDM systems with a low loss of 1.7 dB, a wide 1-dB bandwidth of 14 nm, and
low crosstalk by using an AWG with multi-mode waveguides.
©2003 Optical society of America
OCIS codes: (130.3120) Integrated optics devices, (230.7390) waveguides, planar
1. Introduction
The recent rapid diversification of services based on IP technologies will move wavelength division
multiplexing (WDM) systems beyond backbone and metropolitan area use into access networks. For access
networks, coarse WDM (CWDM) with a large channel spacing of 20 nm is now available with a view to realizing
cost-effective systems, where the wavelength tolerance of the laser transmitter is relaxed and no optical amplifier is
used [1]. This wide tolerance and amplifier-free configuration requires a multi/demultiplexing filter with both low
loss and a wide passband. Moreover, the filter must be inexpensive and compact, because the filter to total system
cost ratio is greater in CWDM systems that have a relatively small number of channels.
There are several types of CWDM filter. Thin film filters (TFFs) have a good spectral response with a low loss
and a rectangular shape. However, it is difficult to reduce the production cost, because TFFs require complicated
fiber handling. In contrast, planar waveguide filters including the lattice-form filter [2] and arrayed-waveguide
grating (AWG) [3,4] are superior as regards cost-effectiveness thanks to their mass-producibility and simplicity of
fiber assembly. In particular, the AWGs, which have already been installed in practical dense WDM systems, also
have advantages of compactness and the capacity for increasing the channel number. However, the optical
performance of conventional AWGs is somewhat inferior to that of the other filters because of the trade-off
relationship between their loss and bandwidth.
To overcome this disadvantage, an AWG with multi-mode waveguides (MM-AWG) has been proposed [5].
Although its function is premised on the use of a demultiplexer with receivers through multi-mode fibers (MMFs),
the MM-AWG can realize both low loss and a wide spectral response while maintaining the advantages of
conventional AWGs. Therefore this filter is one of the most attractive candidates for use with CWDM
demultiplexers. However, as yet only some basic studies have been undertaken on MM-AWGs tuned for CWDM.
Here, we describe two types of 20-nm-spacing 8-channel MM-AWG filter for CWDM that employ a silica-
based planar lightwave circuit (PLC). One is a simple demultiplexer for unidirectional transmission and the other is
a multi/demultiplexer for mono-fiber bidirectional transmission. We designed the AWGs to be compact by using
high index contrast waveguides, and carefully avoided any crosstalk degradation by optimizing the layout of the
AWG components. Consequently, we achieved both excellent optical performance and a compact size for these
AWGs, which are suitable characteristics for a practical CWDM filter.
2. Design and fabrication
The schematic configuration of the MM-AWG and an enlarged representation of the output side are shown in
Fig. 1(a) and (b), respectively. The MM-AWG has a configuration in which the single-mode output waveguides
found in a conventional AWG are replaced with relatively wide multi-mode waveguides. In the MM-AWG, the
focusing field on the imaging plane traverses the plane with a lateral displacement that depends on the input light
wavelength. Then the spectral response of the AWG is determined by the coupling between the field and the
eigenmodes of the output multi-mode waveguide. This coupling is lossless as long as the position of the focusing
field is within the range of the multi-mode waveguide width, although the power distribution to these modes varies
depending on the position. Therefore we obtain a rectangular spectral shape. Here, we optimized the width of the
multi-mode waveguides so that they would couple efficiently with 50-Pm-diameter graded-index (GI50) MMFs and
set the lateral displacement of the AWG so that the 1-dB bandwidths would be > 14 nm.
TuI2
(a) (b) Arrayed waveguides Slab waveguides

Arrayed waveguides
Output multi-mode waveguides
Input single-mode waveguide Output multi-mode waveguides
Multi-mode fibers Multi-mode fibers
Single-mode fiber Str

ay
ligh
t Focusing field
Slab waveguides
Enlargement: Fig1(b) Imaging plane
Fig. 1. (a) Schematic configuration of MM-AWG and (b) enlarged representation of output side.
Although the MM-AWG provides the desired rectangular spectral shape, there is a problem associated with the
multi-mode waveguides, which is the crosstalk caused by the re-coupling with the output MMFs of faint stray lights
that are mainly generated as a result of the coupling loss between the arrayed waveguides and the second slab
waveguide. To avoid this crosstalk problem we carefully optimized the layout of the AWG and output multi-mode
waveguides. Since we found that the stray lights mainly propagate along the elongated part of the second slab as
also shown in Fig. 1(a), we purposely arranged the multi-mode waveguides to bend so that their connecting edges
with the MMFs were not positioned on that elongated part.
We designed 20-nm-spacing 8-channel MM-AWG filters for unidirectional and bidirectional CWDM systems.
Figure 2(a) and (b) show schematic views of the two types of 8-channel CWDM system. In the unidirectional
system, the MM-AWG is employed as a demultiplexer whose outputs are all multi-mode waveguides connected to
receivers through MMFs. The bidirectional system is expected to be more economical and more suitable for access
networks than the unidirectional system because the upstream and downstream channels are transmitted together
through one fiber. Here, the MM-AWG is applicable to each end of the system as a multi/demultiplexer. The AWG
has a 2 x 9-port configuration where the odd output ports are single-mode waveguides for multiplexing, and the even
output ports are multi-mode waveguides for demultiplexing. As shown in Fig. 2(b), we can use the same MM-AWG
for both ends of the system by choosing one of two input ports.
(a) Multiplexer
MM-AWG
(b) MM-AWG
O 1 Tx 1 O 1 Tx 1 1
Rx O 1
O 2 Tx 2 O 2 Rx 2 2 Rx O 1
Rx O 2
3 O 3 Tx 3 3 Tx O 2
O 3 Tx Rx O 3
O 4 Tx 4 Rx O 4 O 4 Rx 4 4 Rx O 3
1 1 1
O 5 Tx SMF for transmission 5 Rx O 5 O 5 Tx 5 5 Tx O 4
2 SMF for transmission 2
O 6 Tx 6 Rx O 6 O 6 Rx 6 6 Rx O 5
O 7 Tx 7 Rx O 7 O 7 Tx 7 7 Tx O 6
O 8 Tx 8 Rx O 8 O 8 Rx 8 8 Rx O 7
Single-mode waveguide 9 9 Tx O 8
SMF MMF SMF Multi-mode waveguides
Tx: Transmitter
Multi-mode waveguides MMF
Rx: Receiver Single-mode waveguides
Fig. 2. Schematic views of (a) unidirectional and (b) bidirectional CWDM systems
We fabricated the two types of AWG using conventional silica-based PLC technology [6]. For compactness we
employed waveguides with a high index difference ' of 1.5% that we recently developed [7]. The respective core
sizes and minimum-bending radii were ~ 4.5 x 4.5 Pm2 and 2 mm for the single-mode waveguide, and ~ 20.0 x 4.5
Pm2 and 5 mm for the multi-mode waveguide. The AWGs were fabricated with vertically tapered waveguides [8] to
reduce the coupling loss between the arrayed and slab waveguides. The chips were 8.0 x 40.0 mm2 for both MM-
AWGs and attached to single-mode optical fibers (SMFs) and GI50-MMFs.
3. Experimental results
We first describe the optical performance of our unidirectional MM-AWG. Figure 3 shows the 8-channel
transmission spectra. We successfully achieved a low loss, and a flat and wide passband for every channel. The
insertion losses were < 1.7 dB and the 1-dB bandwidths were > 14.0 nm. These excellent rectangular shapes confirm
the suitability of our spectral optimization. We also obtained steep suppression slopes at the channel boundaries and
a low background level, which was due to our layout optimization of the AWG and multi-mode output waveguides.
The crosstalk was suppressed to < -29 dB for adjacent channels and < -35 dB for the background level.
We also evaluated the power penalty associated with the fluctuation in the power distribution between the
eigenmodes of the multi-mode waveguides and MMFs caused by the wavelength drift of a transmitter. Figure 4
shows the experimental setup and the obtained bit error rate characteristics. We used a single transmitter with a 2.5
Gb/s modulation. The signal was demultiplexed in the MM-AWG to port no. 5 and guided to a receiver thorough a
3-m-long GI50-MMF. We operated the transmitter, assuming its wavelength drift, at wavelengths of 1550 and 1545
nm, which correspond to the respective wavelengths of the center and edge of the AWG passband at port no. 5. We
TuI2
obtained error free operation and no power penalty can be seen in the figure. Consequently, we confirm that the
application of the MM-AWG to a CWDM system poses no problems in relation to wavelength drift.
0
-10
Transmittance [dB]
-20
-30
-40
1450 1470 1490 1510 1530 1550 1570 1590 1610 1630
Wavelength [nm]
Fig. 3. 8-channel transmission spectra of fabricated unidirectional MM-AWG

MM-AWG
-3
10
Back to back
1 10
-4 1550 nm
1545 nm
2
2.5 Gb/s Power -5
3 10
PRBS 223-1 meter
Bit error rate

O=1550 nm 4 10
-6
䇭 1545 nm Tx Modulator
5 Rx -7
10
SMF 6
MMF
3m 10
-8
7
Error 10
-9
8
detector 10
-10
-11
10
-32 -28 -24 -20 -16
Back to back Optical received power [dBm]
Fig. 4. Experimental setup for testing fabricated unidirectional MM-AWG and obtained bit error rate characteristics
We next describe the results we obtained for the bidirectional MM-AWG. Figure 5 shows the 8-channel
transmission spectra at input port no. 1. We also obtained good optical performance for this AWG. The insertion
losses and 1-dB bandwidths were < 1.6 dB and > 13.5 nm for demultiplexing multi-mode ports, and < 4.3 dB and >
5 nm for multiplexing single-mode ports, respectively. The crosstalk between adjacent demultiplexing ports was < -
35 dB and that between adjacent multiplexing ports was < -40 dB.
0
-10
Transmittance [dB]
-20
-30
-40
1450 1470 1490 1510 1530 1550 1570 1590 1610 1630
Wavelength [nm]
Fig. 5. 8-channel transmission spectra of fabricated bidirectional MM-AWG

5. Conclusion
We used silica-based PLC fabrication technology to develop two types of 20-nm-spacing 8-channel MM-AWG
filter for unidirectional and bidirectional CWDM systems. We demonstrated their advantages of low loss, and a flat
and wide passband resulting from the response of the multi-mode waveguides, compactness achieved by using high
index contrast waveguides, and low crosstalk obtained by optimizing the circuit layout. Along with the cost-
effectiveness of the PLC filters, this performance confirms that these MM-AWGs are suitable and attractive filters
for general application to CWDM systems. We believe that these MM-AWG filters will be key devices in future
optical access networks.
References
[1] P. P. Iannoue, et al., “WDM access networks,” in Proc. of ECOC2002, paper 9.2.1 (2002).
[2] Y. Inoue, et al., “Low-crosstalk 4-channel coarse WDM filter using silica-based planar-lightwave circuit,” in Proc. of OFC2002, paper TuK6,
pp. 75-76 (2002).
[3] K. Okamoto, et al., “Recent progress of integrated optics planar lightwave circuits,” Opt. Quantum Electron., 31, 107-129 (1999).
[4] R. Adar, et al., “Broad-band array multiplexers made with silica waveguides on silicon,” IEEE J. Lightwave Technol., 11, 212-219 (1993).
[5] M. R. Amersfoort, et al., “Phased-array wavelength demultiplexer with flattened wavelength response,” Electron. Lett., 30, 300-302 (1994).
[6] M. Kawachi, et al., “Silica waveguides on silicon and their application to integrated-optic components,” Opt. Quantum Electron., 22, 391–416
(1990).
[7] Y. Hibino, “Recent advances in high-density and large-scale AWG multi/demultiplexers with higher index-contrast silica-based PLCs,” IEEE
J. Selected Topics Quantum Electron., 8, 1090-1101 (2002).
[8] A. Sugita, et al., “Very low insertion loss arrayed-waveguide grating with vertically tapered waveguides,” IEEE Photon. Technol. Lett., 12,
1180-1182, (2000).
TuI3
Spectral engineering of wavelength-division multiplexers

based on planar holographic Bragg reflectors
C. Greiner, D. Iazikov and T. W. Mossberg
LightSmyth Technologies, Inc., 860 W. Park St. Ste 250, Eugene, OR 97041
Phone: (541)-431-0026, FAX: (541)-284-5607, e-mail: cgreiner@lightsmyth.com
Abstract: We report wavelength-division-multiplexing based on lithographically-fabricated slab-waveguide-

contained planar holographic Bragg reflectors. Partial diffractive contour writing and contour displacement are
successfully demonstrated to enable precise spectral engineering of multiplexer transfer functions and make possible
hologram overlay.
OCIS codes: (090.2890, 050.1970, 060.2340, 230.7400, 230.1480)
Planar holographic Bragg reflectors (HBRs) [1-3] are two-dimensional lithographically-scribed volume holograms
contained within a planar slab waveguide. In the slab waveguide, optical signals are free to propagate without
constraints in two dimensions – a geometry that enables 2D Bragg structures to provide powerful spectral and spatial
holographic functions. A single HBR can simultaneously spatially image an input signal to an output port (or from
one point within an integrated photonic circuit to another) while at the same time providing spectral filtering of the
signal. Unlike fiber and channel-waveguide gratings, where separation of the counter-propagating input and output
signals typically requires additional elements, planar HBRs provide spatially distinct and thus easily accessed
outputs. HBRs constitute the building blocks of unique integrated photonic circuits that operate entirely without
wire-analog channel waveguides, being based on HBR-mediated signal transport where signals freely overlap as
they are imaged from active element to active element. The HBR approach marries the power of free-space optics
and volume holography with a fully integrated environment. The powerful volume-holographic function provided
by HBR structures provides, via computer-generated complex-shaped diffractive contours, fully-optimized spatial
mapping of an arbitrary complex input field distribution to an entirely different output field mode. This holographic
imaging function is generally more powerful than that provided by simple confocal elliptical DBRs [4], whose
focusing power degrades when input and output optics deviate from the point source limit.
Recently [3], we demonstrated that photolithographic fabrication of HBRs in silica-on-silicon provides for the
highly accurate placement of constituent diffractive contours thus enabling fully coherent centimeter-scale planar
holographic structures. Additionally, a robust and fabrication-friendly method to control the reflective amplitude of
diffractive element contours via partial writing of the latter was presented [3]. Together, the precise
photolithographic feature placement and the partial contour writing approach provide control over the phase and
amplitude of diffractive elements on an individual line basis. This precision control not only offers a path to
unprecedented flexibility in the design of HBR spectral transfer functions but also enables the overlay of several
(a) (b)
y
Diffractive Element
x z z
x
Upper Cladding, nclad y
d
Output w
Core, ncore /
Input
Lower
LowerCladding,
Cladding,nnclad
clad
OUT
IN
Figure 1. Planar holographic Bragg reflector schematic. 1a, cross-Sectional View; 1b, top View.
1
TuI3
(partially written) planar holograms on the same substrate – highly enabling to the designing of (de)multiplexing
devices for multi-wavelength signals. In the present paper, we demonstrate, for the first time, the application of these
methods to the spectral and spatial engineering of multiplexing devices based on holographic Bragg reflectors.
The operational principle of an individual planar hologram is illustrated in Figure 1. Figure 1a is a schematic
cross section of the device. It consists of a silica slab waveguide with a single-mode core and bilateral 15-Pm-thick
cladding layers. The core-cladding index differential is 0.8%. The cross section also depicts representative
lithographically-scribed diffractive elements located at the upper core-cladding interface. The diffractive elements,
with depth d § 400 nm, are filled with cladding material. The device operates in first order with a diffractive
element period, / § 500 nm, i.e. half the in-medium wavelength of resonant light. In the schematic cross-section of
Figure 1a light enters the device from the left side and is coherently backscattered to the left by the diffractive
elements. Figure 1b is a top view of the simple planar hologram. An optical signal is coupled into the device via an
input port, IN, typically comprised of the endpoint of a channel waveguide that is used to connect the planar
waveguide chip to access optical fibers. The input beam expands in the slab region and is subsequently spectrally
filtered and spatially directed to the output port, OUT, by the planar grating. The HBR consists of diffractive
contours, represented by the dashed lines in Figure 1b, that may be designed individually to match the back-dif-
fracted input field to the output port. The detailed spacing and relative amplitude of the diffractive elements as a
function of position along the input direction determine the spectral transfer function of the device. Fully optimized
holographic contours, not implemented here, will improve input-output coupling through optimization of spatial
wavefront transformation and are equally consistent with overlay and apodization effects. All devices discussed here
with HBR and access waveguides occupy die areas of about 1.7 cm2. Multiplexer fabrication employed standard
photolithography using a DUV optical stepper.
In Fig 2, we present results on a 4-channel 100-GHz channel-spacing HBR-based wavelength-division
multiplexer with flat-top channel passbands. The device (Fig. 2a) comprises apodized individual-channel HBRs that
are staggered along the device depth dimension but are heavily overlapping as well. Fig 2b, upper (lower) graph,
depicts the measured (simulated) spectral transfer functions of the various multiplexer channels for TE-polarized
input light. The passbands clearly show the designed flat passband and channel spacing. The long-wavelength
shoulder on the measured passbands is believed to arise from a second-order apodization effect unaccounted for in
the present design comprising an effective refractive index variation concomitant to amplitude apodization.
Incorporation of this effect in design algorithms is straightforward. The multiplexer’s adjacent channel isolation
exceeds -22 dB. This is excellent for a first iteration design. Overall, the results shown in Figure 2a clearly
(a) (b)
0
-10
O4
Relative Insertion Loss (dB)
O3 -20
O2 -30
O1
-40
1.527 1.528 1.529 1.530
0
O1 O2 O3 O4
-10
O4 O2 O1 O3
In
-20
-30
-40
1.532 1.533 1.534 1.535
Wavelength (um)
Figure 2. 4-channel HBR-based multiplexer. 2a, device schematic; 2b, upper (lower) graph, measured (simulated) multiplexer spectral
transfer function.
2
TuI3
demonstrate the feasibility of spectral passband engineering as described and the ability to overlay HBRs. The
absolute multiplexer insertion loss (IL) through coupled fibers was ~6 dB implying a ~5.0 dB device intrinsic IL.
As constructed, the HBRs operate at low reflectivity. Calculations indicate that achievable alterations in diffractive
structure geometry and refractive-index contrast will provide strong reflectivity and thus low IL loss up to channel
counts of 16-32.
In Figure 3, we demonstrate demultiplexing of 8 wavelength channels with 100-GHz channel-spacing using an
HBR multiplexer consisting of eight stacked, apodized planar holograms. Similar to the device of Figure 2, the
multiplexer comprises a common input port and eight wavelength-specific outputs. In Figure 3, upper (lower) graph,
we show the measured (simulated) multiplexer spectral transfer function for TE-polarized input light. Agreement
between measured and designed performance is good. The average insertion loss was measured to be ~4 dB or about
3.0 dB device intrinsic. The bandpass profiles are slightly broader than the weak-reflection-limit simulation,
consistent with the measured 50 percent reflectivity. Long-wavelength sidelobes again are traced to uncompensated
second-order apodization effects as mentioned above.
In summary, we have demonstrated the viability of planar holographic Bragg reflectors as powerful building
blocks for wavelength-division multiplexers. For the first time, spectral engineering of multiplexer bandpass
functions via diffractive contour design was demonstrated. More generally, the powerful spectral and spatial beam
control inherent to the planar volume-holographic approach offers the possibility of channel-waveguide-free
integrated photonic circuits wherein signal routing and processing occurs entirely through interaction with
distributed diffractive structures like the HBR. Furthermore, as planar surface-relief structures, HBRs promise
consistency with low-cost, mass-production, nanoreplication techniques such as hot embossing or nanoimprint
lithography. In embossed/stamped formats, HBR’s present an economic route to volume production of high
performance optical communications components for datacom and access networks.
References
[1] T. W. Mossberg, “Planar holographic optical processing devices”, Opt. Lett. 26, 414 (2001).
[2] T. W. Mossberg, “Lithographic Holography in Planar Waveguides”, SPIE Holography Newsletter, Nov. 2001; see:
http://spie.org/web/techgroups/holography/pdfs/Holography2-12.pdf.
[3] C. Greiner, D. Iazikov, and T. W. Mossberg, “Lithographically-fabricated planar holographic Bragg reflectors”, J. Lightwave Technol.,
accepted for publication.
[4] C. H. Henry, R. F. Kazarinov, Y. Shani, R. C. Kistler, V. Pol, and K. J. Orlowsky, “Four-channel wavelength division multiplexers and
bandpass filters based on elliptical Bragg reflectors”, J. Lightwave Technol. 8, 748 (1990).
-10
Relative Insertion Loss (dB)
-20
-30
1.526 1.528 1.530 1.532
-10
-20
-30
1.530 1.532 1.534 1.536
Wavelength (um)
Figure 3. 8-channel HBR-based multiplexer. Upper (lower) graph, measured (simulated) multiplexer spectral transfer function.
3
TuI4
Low-loss and compact TFF-embedded silica-waveguide

WDM filter for video distribution services in FTTH systems
M. Yanagisawa, Y. Inoue, M. Ishii, T. Oguchi, and Y. Hida
NTT Photonics Laboratories, 3-1 Morinosato Wakamiya, Atsugi, Kanagawa 243-0198, Japan
m_yanagi@aecl.ntt.co.jp
H. Izumita1), N. Araki1), and T. Sugie2)

1) NTT Access Network Service Systems Laboratories, 1-7-1 Hanabatake, Tsukuba, Ibaraki 305-0805, Japan
2) NTT Access Network Service Systems Laboratories, 1-6 Nakase, Mihama, Chiba, Chiba 261-0023, Japan
Abstract: Silica-waveguide 1.50/1.55 Pm WDM filters with a TFF-embedded configuration have

been developed for video distribution services in FTTH systems. These filters exhibit low
insertion losses of less than 1 dB and high reliability.
OCIS codes: (230.0230) Optical devices; (230.7390) Waveguides, planar
1. Introduction
A broadband passive optical network (B-PON) system is the most promising approach to establishing a
cost-effective optical access network [1]. Three major U.S. carriers have adopted a set of common technical
requirements based on this system. In this enhanced system, specified in ITU-T G.983.3, the wavelength bands for
video distribution services and ATM-PON transport services are allocated at 1.55-1.56 Pm and 1.26-1.50 Pm,
respectively. While bulk-type modules have generally been used for this type of filter, cost-effective and compact
1.26-1.50/1.55-1.56 Pm WDM filters are strongly required. This is because bulk-type filters pose certain problems
as regards cost reduction at the mass production stage and the realization of the compact array modules required for
central offices.
To meet the above demand, WDM filters using a thin film filter (TFF) embedded silica-based planar lightwave
circuit (PLC) [2,3] are very attractive thanks to their compactness, mass-producibility and high reliability. However,
it has been considered very difficult to obtain a low insertion loss and high isolation when the TFF-embedded
waveguide WDM filter is applied to a filter whose guard band is narrower than 0.1 Pm (i.e. 100 nm). This is
because the flat and sharp TFF spectrum deteriorates due to the diffraction of the incident light in the TFF. An
L/U-band WDM coupler with a guard band of 25 nm, and composed of a TFF and PLC waveguides, has a high
insertion loss of 3 dB. [4]
Here we demonstrate a new waveguide technology for the TFF-embedded WDM filter designed to overcome the
guard band limitation. We have successfully realized 1.26-1.50/1.55-1.56 Pm (50-nm guard band) WDM filters
with an extremely low loss and a high isolation for both 1-ch and 8-ch arrays. We have shown for the first time that
the optical performance of a PLC-type filter with a 50-nm guard band can be comparable to that of bulk-type filters.
2. WDM filter configuration and fabrication
Figure 1 shows the schematic configuration of our proposed WDM filter. A TFF composed of a dielectric multilayer
evaporated on a polyimide substrate [5] is inserted into a groove formed at a cross-waveguides intersection. The
TFF is designed to have a passband at 1.26-1.50 Pm and a stopband at 1.55-1.56 Pm, because we require higher
isolation of the 1.50 Pm signal from the 1.55 Pm signal than for the opposite wavelength allocation. A 1.55 Pm
Groove Thin-film filter (TFF)

1.55 Pm 1.50 Pm
< Port E > < Port A >
1.31 Pm
< Port C >
1.50+1.55 Pm
Taper PLC (cross waveguides)
1.31 Pm
Fig. 1 Configuration of TFF-embedded WDM filter and photograph of 1-ch filter module
TuI4
wavelength light input into the common port C is therefore reflected by the TFF and then output from the reflection
port E. A lateral waveguide taper is adopted to expand the mode-field of the incident light to the TFF and thus
suppress the diffraction of the incident light in the TFF, preventing any increase in the insertion loss or deterioration
of the transmission spectrum. This waveguide taper is very effective for realizing a WDM filter with a narrow guard
band.
The silica-based cross-waveguides were fabricated on a Si substrate by a combination of flame hydrolysis
deposition and reactive ion etching. The refractive index difference ' was set at 0.3 %, which is also effective in
decreasing the diffraction. Eight sets of cross-waveguides were arranged in a single chip for the 8-ch array filter. A
groove was cut at the waveguide intersection with a dicing saw. A 30 Pm-thick dielectric multilayered TFF was
then inserted into the groove and fixed in place with adhesive. The WDM filters were connected with fibers and
then housed in a 35 mm-long and 5 mm-diameter package for the 1-ch module, and a 60 mm x 8 mm x 6 mm
package for the 8-ch array module.
3. Results
Figure 2 shows the wavelength response of the fabricated 1-ch module. Figure 2(a) shows the insertion loss and
PDL spectra. The insertion losses were less than 0.85 dB at 1.26-1.50 Pm in the transmission port (C-A) and less
than 0.55 dB at 1.55-1.56 Pm in the reflection port (C-E). The PDL was less than 0.1 dB for all signal wavelength
ranges. The isolation characteristics are also shown in Fig. 2(b) and an extremely high isolation of more than 60 dB
was obtained at 1.55-1.56 Pm in the C-A port.
0.0 1.2
Transmittance (dB) 0
Transmittance (dB)
-0.5 1.0 -10

-1.0 0.8 C-E
PDL (dB)
-20
-1.5 0.6
C-A C-E -30
-2.0 0.4
-2.5 0.2 -40 C-A
-3.0 0.0 -50
1.45 1.50 1.55 1.60 1.45 1.50 1.55 1.60
Wavelength (Pm) Wavelength (Pm)
(a) Insertion loss (solid line) and PDL (dashed line) (b) Isolation characteristics
Fig. 2 Wavelength responses of 1-ch WDM filter module
The loss spectra for the 8-ch array module are also shown in Fig. 3(a) and (b). We also obtained low insertion losses
of less than 0.82 dB in the 1.26-1.50 Pm range and of less than 0.66 dB in the 1.55-1.56 Pm range. The insertion
loss uniformities for both wavelength ranges were as low as less than 0.2 dB. A flat wavelength response was
realized in the 1.26-1.50 Pm wide wavelength range. A high isolation property of more than 42 dB at 1.55-1.56 Pm
in the C-A port was also achieved. These low and uniform loss characteristics confirm that our TFF-embedded
PLC-type filter is very promising for realizing the array modules.
0.0 0
Transmittance (dB)
Transmittance (dB)
-0.5 -10 C-A

-1.0 -20
-1.5 C-A C-E -30
-2.0 -40 C-E
-2.5 -50
-3.0 -60
1.25 1.35 1.45 1.55 1.65 1.45 1.50 1.55 1.60
Wavelength (Pm) Wavelength (Pm)
(a) Insertion loss (b) Isolation characteristics
Fig. 3 Wavelength responses of 8-ch array filter module
TuI4
The optical performances of both 1-ch and 8-ch array filters is summarized in Table 1. The target values are
estimated from the system performance described in ITU-T G.983.3. The isolation in the A-E port shows the
directivity, which is the crosstalk of the 1.31 Pm wavelength signal into the port E. Good performance with high
clearance to the target is achieved for both the 1-ch and 8-ch modules. These results show that a PLC-type WDM
filter can be fabricated with a high yield leading to cost reduction.
Table 1 Optical performance of 1-ch and 8-ch array WDM filter modules
Wavelength Target Measured (dB)
Item Port
(Pm) (dB) 1-ch 8-ch
A-C 1.48-1.50 0.85 0.68-0.79
Insertion loss C-A 1.31 < 1.5 0.75 0.64-0.82
C-E 1.55-1.56 0.55 0.51-0.66
A-C 1.48-1.50 0.05 0.06-0.11
PDL C-A 1.31 < 0.3 0.05 0.11-0.18
C-E 1.55-1.56 0.04 0.06-0.13
C-A 1.55-1.56 > 32 61.4 43.0-50.1
Isolation C-E 1.48-1.50 > 22 23.4 (18.7-25.0)
A-E 1.31 > 50 57.7 51.5-56.5
The reliability of the 1-ch modules has been also investigated based on Telcordia GR-1221. Figure 4 shows the loss
changes during damp heat storage at 75 degrees and 90 %RH for 2000 hours, which is one of the most severe tests
for PLC modules. Changes in the insertion losses for the C-A and C-E ports were both less than +/- 0.2 dB, and the
isolation loss at 1.55 Pm in the C-A port was kept sufficiently high enough at over 60 dB during the test. These
results indicate that the proposed WDM filter can be used for practical applications.
1 80
0.8
Loss change (dB)
0.6 C-A (@1.50 Pm)

0.4 60
Loss (dB)
0.2
0 40
-0.2
-0.4 C-E (@1.55 Pm)
-0.6 20 C-A (@1.55 Pm)
-0.8
-1 0
0 500 1000 1500 2000 0 500 1000 1500 2000
Elapsed time (hrs) Elapsed time (hrs)
(a) Insertion loss (b) Isolation characteristics
Fig. 4 Loss change in 1-ch modules during damp heat storage test
4. Conclusions
We have developed silica-waveguide TFF-embedded 1.26-1.50/1.55-1.56 Pm WDM filters that exhibit low
insertion losses of less than 1 dB, high isolation losses of more than 42 dB, and good reliability. We have shown for
the first time that the optical characteristics are comparable to those of a bulk-type filter. These compact and high
performance PLC filters will enable us to provide cost-effective video distribution services in FTTH systems.
References
[1] Y. Maeda, K. Okada, and D. Faulkner, “FSAN OAN-WG and future issues for broadband optical access networks,” IEEE Communication
Magazine, Dec, 126-132 (2001)
[2] Y. Inoue, T. Oguchi, Y. Hibino, S. Suzuki, M. Yanagisawa, K. Moriwaki, and Y. Yamada, “Filter-embedded wavelength-division multiplexer
for hybrid-integrated transceiver based on silica-based PLC,” Electron. Lett., 32, 847-848 (1996)
[3] Y. Hida, Y. Inoue, F. Hanawa, T. Fukumitu, Y. Enomoto, and N. Takato, “Silica-based 1 x 32 splitter integrated with 32 WDM couplers using
multilayer dielectric filters for fiber line testing at 1.65 Pm,” IEEE Photon. Technol. Lett., 1, 96-98 (1999).
[4] N. Araki, H. Izumita, N. Honda, and M. Nakamura, “Extended optical fiber line testing system using L/U-band crossed optical waveguide
coupler for L-band WDM transmission,” in Tech. Dig. OFC’02, 165-166 (2002)
[5] T. Oguchi, J. Noda, H. Hanafusa, and S. Nishi, “Dielectric multilayered interference filters deposited on polyimide films,” Electron. Lett., 25,
706-707 (1991)
TuI5
Flexible pulse waveform generation using a silica waveguide

based spectrum synthesis circuit
Koichi Takiguchi, Toshimi Kominato, Hiroshi Takahashi, and Tomohiro Shibata
NTT Photonics Laboratories, NTT Corporation, 3-1 Morinosato Wakamiya, Atsugi, Kanagawa 243-0198, Japan
E-mail: taki@aecl.ntt.co.jp
Katsunari Okamoto
NTT R&D Fellow, NTT Corporation, 3-1 Morinosato Wakamiya, Atsugi, Kanagawa 243-0198, Japan
Abstract: We demonstrate pulse waveform shaping with a circuit employing an arrayed

waveguide grating pair and amplitude-phase controllers. We flexibly synthesized the frequency
components of a mode-locked laser to produce 40 GHz square pulses with various widths.
OCIS codes: (130.3120) Integrated optics devices; (060.4510) Optical communications
1. Introduction
The bit rate of optical communication systems is being steadily increased to 40 Gbit/s and more. Moreover, their
functions are being improved to achieve photonic routing in addition to conventional point-to-point transmissions
[1]. We are developing optical functional devices based on a planar lightwave circuit (PLC) for these advanced
optical networks [2]. Flexible shaping of the temporal pulse profile is also an important function that must be
investigated. For example, the manipulation of signal chirping is indispensable in high-speed transmissions, and
particular waveforms including square pulses are required in optical signal processing. It was difficult, however, to
generate these signals directly with conventional semiconductor or fiber light sources. Spectral amplitude and
phase control of the pulses is attractive because this would allow us to produce various kinds of waveforms without
using high-speed modulators or optical nonlinear processes. This procedure corresponds to the manipulation of
each complex frequency component. A bulk grating pair and masks were adopted as a dispersive element and
frequency component controllers, respectively [3, 4]. The use of the grating pair made the apparatus large (~m2),
and it was difficult to generate a variety of pulses with mask patterns where the amplitudes and phases were fixed.
We have reported a spectrum synthesis circuit that we fabricated using PLC technology [5]. The circuit
consists of two arrayed-waveguide gratings (AWGs), variable optical attenuators (VOAs), and phase shifters. We
described the preliminary frequency responses to clarify its operation, but did not show its pulse responses and
variable functions. In this paper, we demonstrate, for the first time, temporal waveform shaping using this type of
PLC circuit. We were able to generate square (flat-top) pulses with various pulse widths, whose repetition rate was
40 GHz. The square pulses produce less nonlinearity in the transmission than conventional pulse waveforms, and
can be used for optical gating [6] or metrology [4].
2. Operating principle and experimental results

Figure 1 shows the configuration and operating principle of our spectrum synthesis circuit. Two frequency
multi/demultiplexing AWGs were interleaved with an array of 32 Mach-Zehnder interferometer (MZI) type VOAs
and thermo-optic (TO) phase shifters. Each VOA was composed of two multi-mode interference (MMI) couplers
with a 3 dB coupling ratio, and a cross-port was used to obtain a better extinction ratio. The AWGs had a Gaussian
transmittance, and their channel spacing, channel number, and free spectral range were 40 GHz, 32, and 3,200 GHz,
respectively. We inserted an adjustment part to equalize the waveguide lengths between the AWGs, and
monolithically integrated all the components on a silica waveguide on a silicon wafer (66 x 92 mm2) with a relative
index difference ' of 0.75 %. With a view to generating desired waveforms, we designed the VOAs and phase
shifters to accurately provide arbitrary amplitude and phase control, respectively, of 32 frequency components, that
were demultiplexed by the first AWG. Figure 2 shows the circuit transmittance in the TE mode when all the VOAs
were either open or closed. This demonstration of TE mode operation is sufficient because the circuit was installed
just behind a laser. The total fiber-to-fiber loss was 13 dB. The extinction ratio, namely, the dynamic range was
about 30 dB, and we were able to set the amplitude with an accuracy of about 0.1 dB because the TO phase shift
resolution is better than 0.01 S. This VOA performance demonstrates a wide dynamic range and precise resolution
meeting our requirements. Another type of integrated-optic circuit with a transversal-filter configuration has been
reported for these purposes [7]. However, its parameters are difficult to adjust because they do not directly
correspond to each frequency component, and the intrinsic loss varies greatly depending on the filter settings.
Figure 3 shows the measurement setup. We used a mode-locked fiber ring laser to provide a seed pulse train
TuI5
Fig. 1. Configuration and operating principle of spectrum synthesis circuit.
Fig. 2. Circuit transmittance. Fig. 3. Experimental setup for measuring pulse spectrum and waveform.
whose repetition rate and waveform were 40 GHz and sech2, respectively. We measured the waveform by utilizing
an auto- or cross-correlator. The seed pulse chirping was compensated with a single-mode fiber, and the tapped
pulse was used as a reference for the cross-correlation. We generated square pulses with different widths from a
seed light source to clarify the circuit validity. Figure 4 shows the original and obtained spectra and auto- or
cross-correlated pulse waveforms. The seed pulse had a full width at half maximum (FWHM) width 't of 0.9 ps,
whose auto-correlation is shown in Fig. 4 (b-1). Figure 4 (a-2, a-3) and (b-2, b-3) show two examples of
synthesized spectra and corresponding cross-correlated pulse waveforms for designed pulse widths 't of 11.9 and
4.4 ps, respectively. The broken and solid curves show calculated and experimental values, respectively. In the
spectra, the value of 0 or S indicates the required phase at each frequency component. We were able to synthesize
the line intensity spectra with an average accuracy of 0.6 dB. The ripples in the flat-top pulse region were caused
by the finite available bandwidth, and were estimated to be 0 to 0.1 dB compared with the calculated values of 0.1 to
0.2 dB. The obtained flat-top regions were generally smoother than the calculated regions. The rise and fall time
(10 to 90 %) was 2.9 to 5.4 ps, while the target was 1.5 ps. The deterioration in the rise and fall time was mainly
brought about by phase setting errors that originated in the thermal crosstalk among the phase shifters used for TO
phase adjustment. We must adopt heat-insulating grooves [8] to reduce the crosstalk and thus obtain better
performance. The results in Fig. 4 (b-2, b-3) show the trade-off relation between the ripples and the rise and fall
time. Figure 5 is an example of measured pulse trains (cross-correlation), corresponding to Fig. 4 (b-2), to
demonstrate that we could obtain square waveforms with a 40 GHz repetition rate. Figure 6 shows the relation
between the obtained and designed pulse widths. The obtained results (closed ellipses) agreed well with the
designed values. To conclude, we were able to achieve our target of generating square pulses with different widths.
3. Summary
We have shown the first temporal pulse shaping achieved using a PLC based spectrum synthesis circuit. The
circuit was composed of two AWGs and an array of 32 controllers, and enabled us to synthesize frequency
components much more accurately and flexibly than with other apparatus or devices. We successfully confirmed
its validity by carrying out experiments to generate square pulses with different widths and a 40 GHz repetition rate.
TuI5
The circuit has other applications including chirp manipulation and the generation of various modulation formats on
high-speed signals, and pulse generation for metrology. These various functions are difficult to achieve with a
conventional optical device, and our circuit will play an important role in the coming advanced optical networks.
4. References
[1] P. Green, “Progress in optical networking,” IEEE Commun. Mag. 39, 54-61 (2001).
[2] K. Takiguchi, “Photonic functional devices based on a silica planar lightwave circuit,” in Proc. ECIO ’03 FrB2, 169-180 (2003).
[3] J. P. Heritage et al., “Picosecond pulse shaping by spectral phase and amplitude manipulation,” OL 10, 609-611 (1985).
[4] A. M. Weiner et al., “Synthesis of phase-coherent, picosecond optical square pulses,” OL 11, 153-155 (1986).
[5] K. Okamoto et al., “Fabrication of frequency spectrum synthesizer consisting of arrayed-waveguide grating pair and thermo-optic amplitude
and phase controllers,” EL 35, 733-734 (1999).
[6] T. Morioka et al., “Multiple-output, 100 Gbit/s all-optical demultiplexer based on multichannel four-wave mixing pumped by a
linearly-chirped square pulse,” EL 30, 1959-1960 (1994).
[7] K. Kitayama et al., “Optical pulse train synthesis of arbitrary waveform using weight/phase-programmable 32-tapped delay line waveguide
filter,” in Proc. OFC ’01 WY3 (2001).
[8] S. Sohma et al., “Low switching power silica-based super high delta thermo-optic switch with heat insulating grooves,” EL 38, 127-128
(2002).
(a-1) Original spectrum. (a-2) Synthesized spectrum ('t=11.9 ps). (a-3) Synthesized spectrum ('t=4.4 ps).
(b-1) Original waveform. (b-2) Generated waveform ('t=11.9 ps). (b-3) Generated waveform ('t=4.4 ps).
Fig. 4. Spectra (a) and corresponding waveforms (b) of pulses. Broken curve : calculated and solid curve: experimental.

Fig. 5. Measured pulse train. Fig. 6. Obtained versus designed pulse width.

6 Bit S HZ Orthogonally Polarized CSRZ D PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

6 Bit S HZ Orthogonally Polarized CSRZ D PDF

Uploaded by

Copyright:

Available Formats

TuF1

1.6 bit/s/Hz orthogonally polarized CSRZ - DQPSK

Abstract: We report experimental results of 1.6 bit/s/Hz, 8u40 Gbit/s OP-CSRZ-DQPSK

Study on Optimum Pre-Filtering Condition

2. Back-to-back performance of pre-filtered CS-RZ DPSK signals

(1) Filter Detuning ASE

3. Tolerance evaluation of pre-filtered CS-RZ DPSK signals

Dispersion Comensation Error [ps/nm] Fiber Input Power [dBm]

polarization multiplexed. We measured the 11

Nonlinear tolerance of differential phase shift keying

N. Hecker-Denschlag, S. Calabrò, M. Herz, C.-J. Weiske, E.-D. Schmidt

D. van den Borne, G.-D. Khoe, H. de Waardt

Experimental comparison of non linear threshold and

2003 Optical Society of America

P OSNR RZ-33% RZ-50%

3. Results and discussion

Electronic Dispersion Compensation for Extended Reach

Binary MZI, Chirp = 0

Electronic Signal Processing for Differential Phase

Joint Symbol MLSE

Experimental verification of combined adaptive PMD and

Folkert Horst, Bert Jan Offrein, Gian Luca Bona

2004 Optical Society of America

3. Results and Discussion

-8 DGD=25ps & DGD=25ps & GVD=100ps/nm

Electrical PMD equalizer ICs for a 40-Gbit/s transmission

M. Nakamura, H. Nosaka, M. Ida, K. Kurishima, and M. Tokumitsu

2. FFE and DFE Equalizers

Delay1 Delay Subtractor

Tap1 Tap2 Tap3 Tap

3. Experiment and Results

Fig. 2 Experimental setup for PMD equalization using an emulator.

100 mV/div 100 mV/div

Mitigation of Nonlinear Impairments from Semiconductor

Abstract: We demonstrate and study significant improvements in the measured bit-error

Tunable Optical Error

Receiver Sensitivity (dBm)

The impact of nonlinearity on electronic dispersion

2. Model of the optical channel

Fig. 3. Required OSNR vs Dispersion

Efficient Physical Topologies for Regular WDM Networks

2. Optimal Network Connectivity Under the Uniform Traffic

Table 1. Optimal node degree '* as functions of N and t.

Topologies/ Optimal node degree as the function of N and t

0.8 Multi-stage Moore Graphs.

Using tunable optical transceivers for reducing the number

3. Minimum tunable transceiver provisioning

Asymmetric Reconfigurable OADMs for next generation

Abstract: Reconfigurable OADMs allow responding to unpredictable traffic demand. We propose

Network Growth and ROADM Design

Transmit vs. Receive Tunability

R ADD + RDROP R1*ADD + R1*DROP

Flexibility vs. Remote Reconfigurability

Iraj Saniee and Indra Widjaja

Figure 1 TWIN architecture consisting of destination-based optical trees.

Blocking Probability of a Centralized Switch with Shared

where pbusyi (k) = (Li )k =k! , for i = 1 2.

Pconv busy (c) =

(a) Coordinated wavelength assignment (b) Autonomous wavelength assignment

3.2 Autonomous Wavelength Assignment

Towards a meshed ultra high speed TDM optical network :

3. Experimental realization of an OTDM-ADM in a network

160 Gbit/s Transmitter

R ADD + RDROP R1ADD + R1DROP