IEEE JOURNAL OF SELECTED TOPICS IN QUANTUM ELECTRONICS, VOL. 16, NO.

5, SEPTEMBER/OCTOBER 2010

1217

Mitigation of Fiber Nonlinearity Using a Digital Coherent Receiver

David S. Millar, Student Member, IEEE, Sergejs Makovejs, Student Member, IEEE, Carsten Behrens, Student Member, IEEE, Stephan Hellerbrand, Student Member, IEEE, Robert I. Killey, Member, IEEE, Polina Bayvel, Fellow, IEEE, and Seb J. Savory, Member, IEEE

AbstractCoherent detection with receiver-based DSP has recently enabled the mitigation of ber nonlinear effects. We investigate the performance benets available from the backpropagation algorithm for polarization division multiplexed quadrature amplitude phase-shift keying (PDM-QPSK) and 16-state quadrature amplitude modulation (PDM-QAM16). The performance of the receiver using a digital backpropagation algorithm with varying nonlinear step size is characterized to determine an upper bound on the suppression of intrachannel nonlinearities in a singlechannel system. The results show that for the system under investigation PDM-QPSK and PDM-QAM16 have maximum step sizes for optimal performance of 160 and 80 km, respectively. Whilst the optimal launch power is increased by 2 and 2.5 dB for PDM-QPSK and PDM-QAM16, respectively, the Q-factor is correspondingly increased by 1.6 and 1 dB, highlighting the importance of studying nonlinear compensation for higher level modulation formats. Index TermsCoherent detection, digital backpropagation, nonlinearity compensation, quadrature amplitude modulation (QAM).

I. INTRODUCTION

XPONENTIAL growth in capacity requirements in recent years has led to rapid improvements in the spectral efciency of optical communications systems [1]. While this growth was previously sustained by the introduction of wavelength division multiplexing (WDM), with on-off-keyed systems, this approach yields a theoretical maximum of 1 bits/s/Hz over the bandwidth of the optical channel. By utilizing coherent detection with phase and polarization diversity, it becomes possible to detect the full 4-D signal space of amplitude and phase on two orthogonal polarizations rather than the single dimension of total power used with direct detection (DD). As all four dimensions of the optical eld are detected by a phase and

Manuscript received December 21, 2009; revised February 9, 2010; accepted March 4, 2010. Date of publication May 20, 2010; date of current version October 6, 2010. This work was supported by the Engineering and Physical Sciences Research Council, Building the Future Optical Network in Europe, a Network of Excellence funded by the European Commission through the 7th Information and Communications Technology-Framework Programme, Yokogawa Electric Corporation, Oclaro, and The Royal Society. D. S. Millar, S. Makovejs, C. Behrens, R. I. Killey, P. Bayvel, and S. J. Savory are with the Department of Electronic and Electrical Engineering, Optical Networks Group, University College London, London, WC1E 7JE, U.K. (e-mail: d.millar@ee.ucl.ac.uk; s.makovey@ee.ucl.ac.uk; c.behrens@ee.ucl.ac.uk; r.killey@ee.ucl.ac.uk; p.bayvel@ee.ucl.ac.uk; s.savory@ee.ucl.ac.uk). S. Hellerbrand is with the Institute for Communications Engineering, Technische Universit t M nchen, M nchen, D-80290, Germany (e-mail: a u u stephan.hellerbrand@mytum.de). Digital Object Identier 10.1109/JSTQE.2010.2047247

polarization diverse coherent receiver, all four of these can be used for modulation and improving the achievable spectral efciency. The detection of all four dimensions of the optical eld also enables the equalization of previously limiting linear transmission impairments, such as group velocity dispersion (GVD) and polarization mode dispersion (PMD). With the elimination of linear transmission impairments, attention has turned to the mitigation of nonlinear impairments, which digital coherent receivers cannot completely compensate stimulating research into nonlinearity mitigating receiver subsystems. Recently, theoretical research has been undertaken to assess the ultimate nonlinear capacity of optical bers [2], where both high-level modulation formats and intrachannel nonlinearity compensating DSP has been assumed. High-level modulation formats have been a topic of much research, resulting in spectral efciency for PDM-QAM16 in excess of 7 bits/s/Hz [3]. As higher level modulation formats have relatively lower optical signal to noise ratio (OSNR) tolerance; therefore, higher launch powers are required, resulting in greater nonlinear penalties. Although research has focused on the study into both the techniques for high-order modulation and nonlinearity compensating DSP, little work has been done into the intersection of these two areas: the comparative benets achievable with nonlinearity compensation when the order of modulation is increased. While the mitigation of interchannel ber nonlinearities remains an active research topic [4], this paper describes the study of a digital coherent receiver and associated algorithms for the intrachannel nonlinearity compensation. To maximize the efcacy of the nonlinear DSP, a phase and polarization diverse digital coherent receiver is employed, allowing the full optical eld within the receiver bandwidth to be reconstructed in the digital domain. This allows us to exploit our knowledge of the physical nature of the optical channel, and design our DSP accordingly. This investigation is performed by a series of single-channel experiments and simulations to determine both the possible benets and the necessary spatial resolution when using the digital backpropagation algorithm to mitigate intrachannel ber nonlinearity. We compare the effects of nonlinearity compensation on two widely investigated high-level modulation formats: polarization division multiplexed quaternary phase-shift keying (PDM-QPSK), which yields 4 bits/symbol; and PDM 16state quadrature amplitude modulation (PDM-QAM16), which yields 8 bits/symbol. Both modulation formats are investigated at 10.7 GBd, such that the benets of nonlinearity compensation may be compared with a doubling of modulation density, while the mitigating effects of dispersion and the signal bandwidth remain the same.

1077-260X/$26.00 2010 IEEE

1218

IEEE JOURNAL OF SELECTED TOPICS IN QUANTUM ELECTRONICS, VOL. 16, NO. 5, SEPTEMBER/OCTOBER 2010

Section II gives an introduction to digital backpropagation, the use of the Manakov equation in forming split-step solutions and forms algorithms for different step sizes and two possible nonlinear cascade models for backpropagation. A detailed description of other DSP algorithms used in the receiver is given in Section III. This is followed by a description of the experimental setup used in Section IV, and simulations performed in Section V, both of which we used to emulate nonlinear distortion occurring in transmission. Results are given in Section VI with accompanying discussion, and conclusions in Section VII. II. DIGITAL BACKPROPAGATION A. Nonlinear Channel Models Digital backpropagation is a method of nonlinearity compensation, which has generated much interest recently [5][8]. It exploits the knowledge of the physical behavior of the optical ber as a nonlinear channel, by approximating the inverse nonlinear channel, most commonly described by the nonlinear Schr dinger equation (NLSE). The solution of the NLSE o is approximated by the split-step Fourier method (SSFM), commonly used to simulate nonlinear transmission in optical bers. While this algorithm may be used to approximate an inverse channel, the channel is inherently limited by additive Gaussian white noise (AGWN), which stems from amplied spontaneous emission (ASE) due to inline optical amplication. This noise cannot be removed by any lter due to its random nature, and will additionally result in nonlinear phase noise (due to the GordonMollenauer effect). Receiver-based digital channel inversion may be performed with many channels within the receiver bandwidth [5], thus also mitigating interchannel nonlinear effects. It should be noted, however, that the required nonlinear step size in previous experimental work [8] was much smaller in this case and the discussion of large step sizes in this paper may not be applicable. Rather than the coupled polarization NLSE (1) [9], in this paper, we have used the Manakov equation (2) [10] for the inverse channel model. The Manakov equation is more applicable for long distance simulations, where total transmission distance is greater than 1000 km [11]. This model accounts for the residual birefringence of the ber, and the effect that this has on the state of polarization and nonlinearity within the ber. Since the residual birefringence scatters the signal state of polarization on a signicantly smaller scale than the ber nonlinear length, ber nonlinearity acts on both polarizations equally as described by the rightmost term in (2). Here, E = [EX , EY ] is the optical eld in delayed time, subscripts X and Y denote orthogonal linear polarization states, is the ber loss parameter, 2 is the GVD parameter, and is the nonlinearity parameter. Though the Manakov equation is widely used for simulation of bers with residual birefringence, application of this equation for digital backpropagation algorithms has only recently been made in [12] and formalized in [13] j2 2 E X = EX + EX z 2 2 t2 2 j |EX |2 + |EY |2 3

j2 2 E Y = EY + EY z 2 2 t2 2 j |EY |2 + |EX |2 3

EY

j 2 E E 3 Y X

(1)

j2 2 8 E X = EX + EX j (|EX |2 + |EY |2 )EX z 2 2 t2 9 j2 2 8 E Y = EY + EY j (|EY |2 + |EX |2 )EY . z 2 2 t2 9 (2) Equation (2) may be split into its constituent linear and nonlinear parts, resulting in E = (D + N )E, z where

j2 2 D= , 2 t2 8 N = j EH E 9 2 E = [ EX EY ] T (3)

From (3), we may derive an exact solution of the Manakov equation, which forms the basis of the split-step type solutions: E(z + h, T ) = exp(h(D + N ))E(z, T ). B. Split-Step Methods The following approximation is central to all SSFM numerical solutions to both the coupled polarization nonlinear NLSE (1) and the Manakov equation (2), which we will investigate here. We may say that for sufciently small step h exp(h(D + N ))E(z, T ) exp(hD) exp(hN )E(z, T ). (5) While the approximation in (5) is very basic, this approximation of the exact solution in (4) forms the basis of all split-step type solutions. A common renement to (5) is to evaluate the nonlinear part of the solution with a constant envelope prole and varying intensity. This modication allows larger step sizes to be used as the solution does not imply constant power throughout the step as (5) does. By assuming that the only change in the electric eld over the nonlinear step is loss (and consequently that dispersion may be considered to act separately), we may normalize the solution of the nonlinear part of the Manakov equation to the varying power prole within the step and remove the loss term N (E , z ) =
z z +h

(4)

N (E , z )dz

1 exp (h) N (E, z ) = LE N (E, z ), where (6)

E (z + h, T ) = exp(h/2)E(z, T ). EX j 2 E E 3 X Y

The approximation in (6) gives us a nonlinear step, which includes loss and the total nonlinear phase shift over the spatial

MILLAR et al.: MITIGATION OF FIBER NONLINEARITY USING A DIGITAL COHERENT RECEIVER

1219

the order of the blocks is reversed, and the bulk-step model, therefore, refers to a cascade of Wiener systems. D. Channel Models and Accuracy Considerations
Fig. 1. Wiener model.

Fig. 2.

Hammerstein model.

The improvement in accuracy that the symmetric split-step method offers over the bulk step method given in (5) may be quantied by examining the error term generated by the approximations in (5) and (7), by use of the BakerCampbell Hausdorff formula [15] for the commutability of operators (8). This formula gives us an analytical insight into both the relative accuracy of the symmetric-step and bulk-step variants of the nonlinear channel model, and the effects of step size and channel model on the accuracy of the digital backpropagation algorithm exp(hD) exp(hN )

Fig. 3.

WienerHammerstein model.

= exp hD + hN +

h2 (DN N D) + . . . . (8) 2

step. This is effectively a multiplication by the effective nonlinear length LE , as dened in [9]. The accuracy of this solution to the Manakov equation can be improved by applying the dispersion operator in two equal parts, before and after the nonlinear operator. This leads to the symmetric SSFM [9] exp(h(D + N ))E(z, T ) exp h D exp(LE N ) exp 2 h D E(z, T ). 2 (7)

In the BakerCampbellHausdorff formula (8), for convenience, we have included the dominant (rst) term only of the error series. This is equivalent to the error in the bulk-step approximation given in (5). The error term resulting from the symmetric split-step method may be described as follows: exp where
2 h h N, D + . . . = hN + D + 2 2 2

h D exp(hN ) exp 2

h D 2

= exp

h D exp() 2

C. Two- and Three-Block Nonlinear Models The two variations of the split-step method may be described in terms of two- and three-block nonlinear models from nonlinear systems theory [14], for each short length of ber over which a split step is taken. Nonlinear models of particular interest are: the Wiener model, which consists of a linear block followed by a memoryless nonlinear block (see Fig. 1), the Hammerstein model, consisting of a memoryless nonlinear block followed by a linear block (see Fig. 2), and the WienerHammerstein model, which represents the concatenation of the Wiener and Hammerstein models, that is, a linear block followed by a memoryless nonlinear block, followed by a second linear block (see Fig. 3). It is immediately apparent that the WienerHammerstein model (see Fig. 3) is used to represent the behavior of a section of ber in the symmetric split-step method. These three block systems are then cascaded to form an approximation of the entire channel. Similarly, the bulk-step approach refers to a cascaded two-block (Hammerstein model) approximation of the channel. The difference between the two- and three-block nonlinear models stems largely from their relative accuracy, which is analyzed in the Section II-D. It should be noted that while the bulk-step approximation of the forward channel corresponds to a cascade of Hammerstein systems, when this model is used for compensation (that is, approximation of the inverse channel),

exp

h D exp() 2 + ... . (9)

3 h N + D , N, D = exp hD + hN + 6 2 2

We note from (8) that the dominant error term is proportional to h2 , while in (9), the dominant is proportional to h3 . This indicates (a fact well known by those familiar with optical ber simulations using the SSFM) that the symmetric split-step method will give greater accuracy with an identical spatial step size. An essential factor in the application of this solution, however, is that the solution of the Manakov equation is to be performed at the receiver with a noisy signal and DSP. It is therefore in our interest to perform not the most accurate reverse propagation, but the least complex with sufcient accuracy. Much recent research [12], [16], [17] has demonstrated that for receiverbased digital backpropagation, a single nonlinear step per span with the bulk-step method may be considered a sufciently accurate solution of the Manakov equation. In this paper, we will extend the ideas proposed in [5] and [6] by examining both two- and three-block nonlinear models for step sizes ranging from subspan to an entire link.

1220

IEEE JOURNAL OF SELECTED TOPICS IN QUANTUM ELECTRONICS, VOL. 16, NO. 5, SEPTEMBER/OCTOBER 2010

Fig. 5. Recirculating loop setup used for transmission experiments, with optical front end of the phase and polarization diverse digital coherent receiver.

Fig. 4. Digital back end of phase and polarization diverse digital coherent receiver.

E. Digital Backpropagation Algorithms The coherent receiver-based nonlinearity compensation algorithms used in this investigation may be categorized as two variations of digital backpropagation. These are the Wiener and WienerHammerstein reverse-channel approximations (corresponding to the bulk-step and symmetric-step methods), each of which may be applied to step sizes of less than one span, or one span or more. Dispersive steps were performed in the frequency domain, using the dispersion operator dened in (3). This method describes a circular convolution of the dispersion operator (3) and the time-domain signal, and is given in (10), where F represents the Fourier transform exp(hD)E = F 1 {exp(hF{D})F{E}}. (10)

Nonlinear operations were performed in the time domain with the nonlinear operator normalized to both nonlinear step-size and launch power. For nonlinear step sizes of a single span or greater, the nonlinear operator is dened as follows: N (t, zNL ) = jzNL (|EX (t)|2 + |EY (t)|2 )PL . (11)

Here, PL represents launch power in milliwatt, zNL is the nonlinear step size, t represents the retarded time frame, and is the nonlinear phase-shifting coefcient, which is a constant to be optimized. All other symbols retain their conventional meaning or those dened previously. In the case of more than one nonlinear step per span, the nonlinear operator is modied to account for the exponentially varying power prole within the span. This leads to a nonlinear operator, which is dened as follows: N (t, zNL ) = j10(sL /10n ) zNL (|EX (t)|2 + |EY (t)|2 )PL (12) where n is the number of steps per span, s is the index of the step within the span, and L is the ber loss per span in decibels. III. RECEIVER DSP STRUCTURE The structure of the digital back end of our coherent receiver is as described in Fig. 4. Received digitized data is resampled to a rate of two samples per symbol, and then, prepared for processing by deskewing

and normalizing the signals to have unit power per polarization. The signal is then ltered with a stationary approximation of the inverse of the optical channel. In most of the previous work, this inverse channel model has compensated for dispersion only, which may be modeled as a nite-impulse response (FIR) lter [18]. Here, we investigate the different inverse nonlinear channel models described in Section II for this block. Adaptive equalization is performed using the constant modulus algorithm (CMA) equalizer for blind deconvolution of the signal and channel, and separation of the two orthogonal polarization states for PDM-QPSK signals, and the radially directed CMA equalizer for PDM-QAM16 [19]. Intradyne frequency offset is removed using a fourth-power nonlinearity on the symbol sequence to remove the effects of phase modulation, and the resulting complex signal is then examined in the frequency domain to nd the frequency of the resulting intradyne frequency tone as in [20]. Phase recovery is then performed using the well-known Viterbi and Viterbi algorithm for PDM-QPSK [21] modied from the normal fourth-power nonlinearity by raising the phase of each symbol to the power of four and the magnitude to the power of 1.3. This modication reduces the weighting in the averaging lter given to the highest power samples, which have undergone the most nonlinear distortion. A decision-directed digital phase locked loop was used in the case of PDM-QAM16 modulated signals [19]. Symbol estimation is then performed, followed by determination of bit error rate (BER) for quantifying system performance. IV. EXPERIMENTAL TRANSMISSION SETUP To characterize the functionality of our digital coherent receiver with nonlinearity compensation, we performed a set of transmission experiments to examine the effects of linear and nonlinear impairments. A particular focus was to investigate the effectiveness of nonlinear compensation techniques for PDMQPSK and PDM-QAM16 modulation formats and the impact of varying DSP complexity on the transmission performance. The optical signals were transmitted multiple times through a single-span recirculating ber loop, with the following coherent detection, digitization, and ofine DSP, as shown in Fig. 5. The loop consisted of 80.2 km single mode ber (SMF) with an overall chromatic dispersion of 1347 ps/nm and loss of 15.4 dB. The experimental procedure was similar to that of [17] and [22]. The polarization-multiplexed QPSK signal was generated using an IQ modulator, which was driven over 2V with respect to the minimum bias point of its transfer function. For the data,

MILLAR et al.: MITIGATION OF FIBER NONLINEARITY USING A DIGITAL COHERENT RECEIVER

1221

TABLE I FIBER AND LINK PARAMETERS

Fig. 6. Transmitter structure for PDM-QPSK, with optional QAM16 stage highlighted. Inset: optical eye-diagrams at the output of the transmitter for (top) PDM-QPSK and (bottom) PDM-QAM16.

two decorrelated 212 pseudorandom binary sequences were used from the output of the pulse pattern generator (PPG), which were subsequently amplied to 7Vpp (2V ) to separately drive the I and Q arms of the modulator. The transmitter DFB laser linewidth, wavelength, and output optical power were 1 MHz, 1554 nm, and 8 dBm, respectively. To emulate polarization multiplexing, we used a passive delay-line ber interferometer, where two data were again decorrelated, time and amplitude aligned, and nally, recombined via a polarization beam splitter (PBS), as shown in Fig. 6. To synthesize a PDM-QAM16 signal, we employed a recently developed method based on the interferometric optical processing of a QPSK signal [23]. To aid carrier phase estimation, an external cavity laser (ECL) with a linewidth of 100 kHz was used in the QAM16 transmitter. The initial QPSK signal is launched into a phase-stabilized ber interferometer, where the two signals are decorrelated, time-aligned, and attenuated with respect to each other by 6 dB (highlighted, Fig. 6). The phase between two arms was set to 90 and maintained utilizing a feedback circuit. For the feedback circuit, we used a ditherless bias control circuit, alternatively, a circuit design described in [24] may be used. Even though this method cannot be used to independently modulate different streams of data, this can be used to investigate transmission performance of PDM-QAM16 signals. In addition, this generation method allows suppression of the transfer of noise between the electrical and optical domains in the transmitter, owing to the nonlinear transfer function of the modulator. Polarization multiplexing emulation was performed as in the PDM-QPSK case. V. SIMULATION OF NONLINEAR EQUALIZER PERFORMANCE To verify experimental results, a 215 long symbol sequence has been simulated for 10.7 GBd PDM-QPSK and PDMQAM16 transmission. Laser phase noise was modeled as a Wiener process, leading to a transmitter laser linewidth of 1 MHz for PDM-QPSK and 100 kHz for PDM-QAM16, while the inuence of relative intensity noise (RIN) was neglected throughout the simulations. An electrical fth-order Bessel lter with a 3 dB bandwidth of 26 GHz was used to emulate the limited transmitter bandwidth. To ensure accurate simulation of the experimental setup, the transmission link shown in Fig. 5 has been modeled in as much detail as possible, while making the following assumptions. Each acousto-optic modulator (AOM) is assumed to introduce a loss of 3 dB, while erbium-doped ber ampliers are operated in saturation to give a xed output power of 17 dBm, and

add noise power to the signal corresponding to a noise gure of 4.5 dB. Further attenuation is then applied to attain the desired launch power into the ber. We simulated a span length of 80.2 km, modeling propagation inside the transmission ber with a symmetrical SSFM covering linear effects, the Kerr effect, PMD, and nonlinear polarization scattering. The ber parameters, which were used are detailed in Table I. The optical loop lter was modeled as a second-order Gaussian lter with a 3 dB bandwidth of 100 GHz. After transmission, the signal was detected with a singleended coherent receiver assuming a local oscillator (LO) to signal ratio of 24 dB and an LO linewidth of 100 kHz. Limited receiver bandwidth was largely determined by the p-i-n photodiodes, which are modeled with fth-order Bessel lters using a 3 dB bandwidth of 7 GHz. Receiver-side A/D converters introduce additional quantization noise and are modeled as having an effective resolution of 4 bits. Subsequent DSP was performed as described in Sections II and III. The residual implementation penalty is assumed to stem mainly from electrical noise in transmitter and receiver, and is modeled by adding additional electrical noise at the receiver to give back-to-back performance similar to the measured receiver sensitivity. VI. PDM-QPSK AND PDM-QAM16 TRANSMISSION RESULTS To examine the variation of system performance with various implementations of digital backpropagation, we performed experiments with PDM-QPSK and PDM-QAM16 near to their maximum reach without nonlinearity compensating DSP. A symbol rate of 10.7 GBd was chosen to exploit the full receiver bandwidth when using T /2 sampling for processing. The digital backpropagation algorithm was then investigated, in terms of both the nonlinear step size and the use of both Wiener and WienerHammerstein models. The performance of the Wiener cascade model backpropagation was experimentally characterized for PDM-QPSK and PDM-QAM16, and the results shown in Figs. 7 and 8, respectively. In Figs. 7 and 8, Q-factor in decibel is plotted as a contour graph against nonlinear step size in kilometers on the horizontal axis and launch power in dBm on the vertical axis. Algorithm performance was examined over 97 spans (7780 km) for PDMQPSK and 20 spans (1600 km) for PDM-QAM16. As these distances are close to maximum reach for the modulation formats investigated, both nonlinear effects and the possible benets of nonlinearity compensation are more signicant than for shorter distances.

1222

IEEE JOURNAL OF SELECTED TOPICS IN QUANTUM ELECTRONICS, VOL. 16, NO. 5, SEPTEMBER/OCTOBER 2010

Fig. 7. Contour plot of experimentally determined Q-factor in decibels against launch power and nonlinear step size for Wiener cascade compensation of 97 spans transmission PDM-QPSK at 10.7 GBd. Nonlinear step size of a single span lies at 80 km.

Fig. 9. Q-factor in decibels for transmission of 10.7 GBd PDM-QPSK over 97 spans. Broken lines denote experimental data, while solid markers correspond to simulated results. Various nonlinear step sizes are demonstrated.

Fig. 8. Contour plot of experimentally determined Q-factor in decibels against launch power and nonlinear step size for Wiener cascade compensation of 20 spans transmission PDM-QAM16 at 10.7 GBd. Nonlinear step size of a single span lies at 80 km. Fig. 10. Q-factor in decibels for transmission of 10.7 GBd PDM-QAM16 over 20 spans. Broken lines denote experimental data, while points correspond to simulated results. Various nonlinear step sizes are demonstrated.

In Figs. 7 and 8, we observe that for lower powers, performance is limited by the accumulated optical noise. For launch powers of below 5 dBm, there is an insignicant improvement in performance for either modulation format with any step size. As launch power is increased, we note that an improvement in performance is available for reduced step sizes up to 160 km. In the case of PDM-QPSK, the optimum launch power is improved by some 2 dB: from approximately 4.5 to 2.5 dBm. In the case of PDM-QAM16, optimum launch power increases by 2.5 dB: from approximately 3.5 to 1 dBm. For PDMQPSK, the benets available with decreasing nonlinear step size become saturated at 160 km, while with PDM-QAM16, the improvement in performance becomes saturated at approximately 80 km. A common characteristic of these curves is that the improvement in maximum Q-factor may be modest (in the region of 1 dB in each case), but the increase in input dynamic range (that is, the range of launch powers for which the BER is less the forward error correction limit) may be dramatic: approximately 4 dB for PDM-QPSK and greater than 5 dB for PDM-QAM16.

Figs. 9 and 10 show the variation of the Q-factor against launch power with varying nonlinear step sizes for PDM-QPSK and PDM-QAM16, respectively. Experimental data is denoted by broken lines, while Monte Carlo simulations (as described in Section IV) are denoted by points in Figs. 9 and 10. In both cases, there is a very good agreement between the experimental and simulated results. Linear compensation denotes compensation for chromatic dispersion only, while 1 step per span and 16 steps per span correspond to nonlinear step sizes of 80 and 5 km, respectively. One step per link and four steps per link correspond to a rst-order and a fourth-order nonlinear cascade, respectively. For PDM-QPSK over 97 spans, this corresponds to nonlinear step sizes of 7780 and 1945 km, respectively. For PDM-QAM16 over 20 spans, this corresponds to nonlinear step sizes of 1600 and 400 km, respectively. The inuence of the distribution of dispersion between the two linear blocks in the WienerHammerstein cascade was

MILLAR et al.: MITIGATION OF FIBER NONLINEARITY USING A DIGITAL COHERENT RECEIVER

1223

Fig. 11. Variation of experimental Q-factor in decibels with distribution of dispersion in WienerHammerstein cascade compensation of 97 spans transmission PDM-QPSK at 10.7 GBd. For the purposes of illustration, 0.5 dBm launch power and an 80 km nonlinear step size are shown.

Fig. 12. Variation of Q-factor in decibels with launch power for 20 spans PDM-QAM16 for different nonlinear step sizes with Wiener model compensation. Experimental data shown as points and t shown as solid lines.

investigated. This corresponds to varying the position of the nonlinear block within the section of ber, which is approximated by each three block system. The optimum was found to be 85% of the dispersion in the rst block and 15% in the second block. This corresponds to applying the nonlinearity at half of the effective nonlinear length of the ber section. A graph illustrating this is shown in Fig. 11, which shows the variation of Q-factor with the length of the rst dispersive step. Fig. 11 shows an 80 km nonlinear step size for 0.5 dBm launch power. Other launch powers and nonlinear step sizes also exhibit a maximum when the split of dispersion is 85%15%, although the maximum improvement in Q-factor is smaller for lower launch powers and smaller nonlinear step sizes. This dispersive split was used for all subsequent uses of WienerHammerstein model backpropagation. To effectively represent and compare the results between the Wiener and WienerHammerstein models, we used tted curves to quantify improvements in launch power and Q-factor with reduction in step size. A least-squares t was used to produce the polynomial approximation of Q-factor as a function of launch power given in [24]. The quality of the polynomial t is illustrated for PDM-QAM16 with several orders of Wiener model nonlinearity compensation in Fig. 12. This tting process was carried out for both the Wiener and WienerHammerstein models of nonlinearity compensation, varying the nonlinear step size between 5 km and the entire link for each modulation format. Using this tting process, we were able to extrapolate the optimum launch power for different models and spatial resolutions of nonlinearity compensation. This inferred optimum launch power is useful as the experimental measurements have a granularity of 1 dB in launch power, and the change in optimum launch power may be less than this, i.e., between different nonlinear step sizes. From the polynomial approximation of variation in Q-factor with launch power, we are also able to infer Q-factor at the optimum launch power. To account for the difference in complexity between the two nonlinear models and allow a more direct comparison, performance is characterized for the mean dispersive block length,

Fig. 13. Plot of improvement in inferred maximum Q-factor against mean dispersive block length for PDM-QPSK at 10.7 GBd using Wiener and Wiener Hammerstein model nonlinearity compensation.

which we dene as follows:1 Mean dispersive block length = Total link length . Number of dispersive blocks

The characterization of PDM-QPSK with varying mean dispersive block length is given later in Figs. 13 and 14. Here, we plot improvement in Q-factor at the optimum launch power (see Figs. 13 and 14) using the polynomial tting process with the experimental measurements as previously described. It is noted that in both Figs. 13 and 14 both nonlinear compensation models offer similar potential benets for a given mean dispersive block length. Maximum Q-factor is improved by approximately 1.6 dB, and optimum launch power is increased
1 Where there are two adjacent dispersive blocks in the WienerHammerstein system, we assume that they may be incorporated into a single block. Therefore, for a very high-order cascade, the mean dispersive block length of the Wiener Hammerstein system will approach that of the Wiener system. Conversely, for a rst-order nonlinear model, the Wiener model will have a mean dispersive block length, which is twice that of the WienerHammerstein model.

1224

IEEE JOURNAL OF SELECTED TOPICS IN QUANTUM ELECTRONICS, VOL. 16, NO. 5, SEPTEMBER/OCTOBER 2010

Fig. 14. Plot of improvement in inferred optimum launch power against mean dispersive block length for PDM-QPSK at 10.7 GBd using Wiener and Wiener Hammerstein model nonlinearity compensation.

Fig. 15. Plot of improvement in inferred maximum Q-factor against mean dispersive block length for PDM-QAM16 at 10.7 GBd using Wiener and Wiener Hammerstein model nonlinearity compensation.

by approximately 1.9 dB for both compensation models. The improvement in both Q-factor and launch power saturates for a mean dispersive block length of approximately 160 km, or two spans. The similarity between the models when characterized in terms of mean dispersive block length may be considered with reference to the discussion of the accuracy of the models presented in Section II. When the step size is small, the accuracy of both models is good and there is agreement in accuracy between the two models as the error due to the commutability of the linear and nonlinear blocks is insignicant. As the nonlinear step size increases, the benet in accuracy of the Wiener Hammerstein model becomes more signicant. This difference is offset, however, due to the difference in mean dispersive block length between the two models. This becomes most noticeable for a nonlinear step size corresponding to the link length L: for this case, the mean dispersive block length is L for the Wiener cascade, but L/2 for the WienerHammerstein cascade. The performance of PDM-QAM16 is characterized at 20 spans transmission for the Wiener and WienerHammerstein models for nonlinearity compensation later in Figs. 15 and 16. Again, we plot improvement in Q-factor at the optimum launch power (see Figs. 15 and 16) using the polynomial tting process with the experimental measurements. We note that similarly to the PDM-QPSK case, the performance of the Wiener and WienerHammerstein models are remarkably similar. Benets in performance are saturated for a mean dispersive block length of less than 80 km, with a maximum improvement in Q-factor of approximately 1 dB, and an increase in optimum launch power of 2.5 dB. This reects the greater improvement in input dynamic range noted for PDMQAM16 in comparison to PDM-QPSK. The reduced number of required nonlinear blocks for significant improvement in performance for PDM-QAM16 in comparison with PDM-QPSK is shown in Fig. 17. For a 1.5 dB increase in optimum launch power, it is noted that PDM-QPSK over 97 spans requires approximately 25 nonlinear blocks over 7780 km, while PDM-QAM16 requires approximately 7 over 1600 km. This corresponds to a slightly smaller nonlinear step size for PDM-QAM16 (230 km compared to 310 km), as may

Fig. 16. Plot of improvement in inferred optimum launch power against mean dispersive block length for PDM-QAM16 at 10.7 GBd using Wiener and WienerHammerstein model nonlinearity compensation.

Fig. 17. Plot of improvement in inferred optimum launch power against number of nonlinear blocks for Wiener model nonlinearity compensation of PDMQPSK and PDM-QAM16.

MILLAR et al.: MITIGATION OF FIBER NONLINEARITY USING A DIGITAL COHERENT RECEIVER

1225

Fig. 18. Plot of improvement in inferred optimum launch power against nonlinear step size for Wiener model nonlinearity compensation of PDM-QPSK and PDM-QAM16.

be seen in Fig. 18, though this difference is dwarfed by the reduction of nonlinear blocks due to the reduced transmission distance. This is also in agreement with our earlier observation that PDM-QAM16 requires a smaller nonlinear step size than PDM-QPSK to gain the maximum available benets from backpropagation. Again, the data used in Figs. 17 and 18 is from polynomial tting of the experimental measurements as previously described. VII. CONCLUSION We have investigated the performance of a coherent receiver with nonlinearity compensating DSP and have shown that it can be successfully used to mitigate intrachannel nonlinearities in both PDM-QPSK and PDM-QAM16 over distances of 7780 and 1600 km, respectively. The impact of the key receiver DSP parameter, namely the nonlinear step size was investigated. It was shown that signicant improvements in performance may be achieved with resolution signicantly coarser than a single span. While performance in this long-step region may be improved with the use of a three block WienerHammerstein model rather than the more commonly used Wiener model, the increased computational effort this model requires offsets any benet when performance is examined in terms of the mean dispersive block length. For the examined receiver bandwidth and symbol rate, the benet of nonlinearity compensation saturated for a nonlinear step size of 160 km for PDM-QPSK and 80 km for PDM-QAM16. Additionally, nonlinear backpropagation appears to offer a greater benet for PDM-QAM16, which may be attributed to this formats higher susceptibility to ber nonlinear effects. This leads us to infer that nonlinearity compensation of this kind is considerably more attractive for modulation formats, which are highly spectrally efcient, and transmitted over short links, where the reduced memory due to dispersion and the increased benet available combine to produce greater benets from fewer blocks. REFERENCES
[1] A. Chraplyvy, The coming capacity crunch, in Proc. ECOC 2009, Paper Mo1.0.2.

[2] R. J. Essiambre, G. J. Foschini, G. Kramer, and P. J. Winzer, Capacity limits of information transport in ber-optic networks, Phys. Rev. Lett., vol. 101, p. 163901, Oct. 2008. [3] P. J. Winzer, A. H. Gnauck, C. R. Doerr, M. Magarini, and L. L. Buhl, Spectrally efcient long-haul optical networking using 112-Gb/s polarization-multiplexed 16-QAM, J. Lightw. Technol., vol. 28, no. 4, pp. 547556, Feb. 2010. [4] C. Xie, Suppression of inter-channel nonlinearities in WDM coherent PDM-QPSK systems using periodic-group-delay dispersion compensators, in Proc. ECOC 2009, pp. 12, Paper P4.08. [5] X. Li et al., Electronic post-compensation of WDM transmission impairments using coherent detection and digital signal processing, Opt. Exp., vol. 16, no. 2, pp. 880888, Jan. 2008. [6] E. Ip and J. Kahn, Compensation of dispersion and nonlinear effects using digital backpropagation, J. Lightw. Technol., vol. 26, no. 20, pp. 3416 3425, Oct. 2008. [7] G. Charlet et al., 72 100 Gb/s transmission over transoceanic distance, using large effective area ber, hybrid Raman-erbium amplication and coherent detection, in Proc. OFC/NFOEC 2009, Paper PDPB6. [8] G. Goldfarb, M. G. Taylor, and G. Li, Experimental demonstration of ber impairment compensation using the split-step nite-impulse-response ltering method, IEEE Photon. Tech. Lett., vol. 20, no. 22, pp. 18871889, Nov. 2008. [9] G. P. Agrawal, Nonlinear Fiber Optics, 3rd ed. New York: Academic, 2001, pp. 4551. [10] D. Marcuse, C. R. Menyuk, and P. K. A. Wai, Application of the ManakovPMD equation to studies of signal propagation in optical bers with randomly varying birefringence, J. Lightw. Technol., vol. 15, no. 9, pp. 17351746, Oct. 1997. [11] C. R. Menyuk, Application of multiple-length-scale methods to the study of optical ber transmission, J. Eng. Math., vol. 36, no. 12, pp. 113136, 1999. [12] S. Oda et al., 112Gbps DP-QPSK transmission using a novel nonlinear compensator in digital coherent receiver, presented at the OFC/NFOEC 2009, San Diego, CA, Mar. 2226, Paper OThR6. [13] F. Yaman and G. Li, Nonlinear impairment compensation for polarization multiplexed WDM transmission using digital backward propagation, IEEE Photon. J., vol. 1, no. 2, pp. 143152, Aug. 2009. [14] R. Haber and H. Unbehaben, Structure identication of nonlinear dynamic systems A survey on input/output approaches, Automatica, vol. 26, no. 4, pp. 651677, 1990. [15] G. H. Weiss and A. A. Maradudin, The Baker-Hausdorff formula and a problem in crystal physics, J. Math. Phys., vol. 3, no. 4, pp. 771777, Jul./Aug. 1962. [16] E. Yamazaki et al., Multi-staged nonlinear compensation in coherent receiver for 16340-km transmission of 111-Gb/s no-guard-interval CoOFDM, presented at the ECOC 2009, Vienna, Austria, Paper Th9.4.6. [17] D. S. Millar, S. Makovejs, V. Mikhailov, R. I. Killey, P. Bayvel, and S. J. Savory, Experimental comparison of nonlinear compensation in longhaul PDM-QPSK transmission at 42.7 and 85.4 Gb/s, in Proc. ECOC 2009, pp. 12, Paper Th9.4.4. [18] S. J. Savory, Digital lters for coherent optical receivers, Opt. Exp., vol. 16, pp. 804817, Jan. 2008. [19] I. Fatadin, D. Ives, and S. J. Savory, Blind equalization and carrier phase recovery in a 16-QAM optical coherent system, J. Lightw. Technol., vol. 27, no. 15, pp. 30423049, Aug. 2009. [20] D. C. Rife and R. C. Boorstyn, Single-tone parameter estimation from discrete-time observations, IEEE Trans. Inf. Theory, vol. 20, no. 5, pp. 591598, Jul. 1974. [21] A. Viterbi and A. Viterbi, Nonlinear estimation of PSK modulated carrier phase with application to burst digital transmission, IEEE Trans. Inf. Theory, vol. 29, no. 4, pp. 543544, Jul. 1982. [22] S. Makovejs et al., Experimental investigation of PDM-QAM16 transmission at 112 Gbit/s over 2400 km, in Proc. OFC/NFOEC 2010, Paper OMJ6. [23] S. Makovejs et al., Novel method of generating QAM-16 signals at 21.3 Gbaud and transmission over 480 km, IEEE Photon. Tech. Lett., vol. 22, no. 1, pp. 3638, Nov. 2009. [24] S. Makovejs, G. Gavioli, V. Mikhailov, R. I. Killey, and P. Bayvel, Experimental and numerical investigation of bit-wise phase-control OTDM transmission, Opt. Exp., vol. 16, no. 23, pp. 1872518730, Oct. 2008. [25] S. J. Savory, Optimum electronic dispersion compensation strategies for nonlinear transmission, Electron. Lett., vol. 42, no. 7, pp. 407408, 2006.

1226

IEEE JOURNAL OF SELECTED TOPICS IN QUANTUM ELECTRONICS, VOL. 16, NO. 5, SEPTEMBER/OCTOBER 2010

David S. Millar (S07) was born in Manchester, U.K., in 1982. He received the M.Eng. degree in electronic and communications engineering from the University of Nottingham, Nottingham, U.K., in 2007. He is currently working toward the Ph.D. degree in digital signal processing for coherent optical communication in the Optical Networks Group, University College London, London, U.K. His current research interests include signalprocessing techniques for high-level modulation formats, and nonlinearity compensation. Mr. Millar was a Reviewer for several IEEE publications, including IEEE PHOTONICS TECHNOLOGY LETTERS and the IEEE JOURNAL OF LIGHTWAVE TECHNOLOGY.

Sergejs Makovejs (S08) was born in Volgograd, Russia, in 1983. He received the B.Sc. and M.Sc. degrees in telecommunications from Riga Technical University, Riga, Latvia, in 2004 and 2006, respectively. Since 2007, he has been working toward the Ph.D. degree from the University College London, London, U.K., where his research has been focused on high-speed optical bre communication systems. In 2006, he was at Siemens, Munich, Germany, where he was involved in designing of rail automation and signalling systems. He has authored and coauthored ve peer-reviewed conference and journal papers. Mr. Makovejs was nominated for the Corning Outstanding Student Paper Competition at OFC/NFOEC 2010.

Robert I. Killey (M00) received the B.Eng. degree in electronic and communications engineering from the University of Bristol, Bristol, U.K., in 1992, the M.Sc. degree in microwaves and optoelectronics from University College London (UCL), London, U.K., in 1994, and the D.Phil. degree from the University of Oxford, Oxford, U.K., in 1998. His doctoral work was on InGaAsP FabryP rot optical modulators e and their applications in soliton communications, in collaboration with Alcatel Submarine Systems Ltd., Greenwich, U.K. He was a Research Fellow in the Department of Electronic and Electrical Engineering, Optical Networks Group, UCL, and became a Lecturer in 2000. He has authored or coauthored more than 100 journal and conference papers. His current research interests include nonlinear ber effects in wavelength division multiplexing transmission systems, wavelength-routed optical networks, and applications of electronic signal processing in optical communications. Dr. Killey is a member of the IEEE Photonics Society and the Institution of Engineering and Technology. He has been involved in the Technical Program Committees for European Conference on Optical communication, IEEE Photonics Society Annual Meeting, and Optoelectronics and Communications Conference. Polina Bayvel (S87M89SM00F09) received the B.Sc. (Eng.) and Ph.D. degrees in electronic and electrical engineering from University College London (UCL), London, U.K., in 1986 and 1990, respectively. Her Ph.D. research focused on nonlinear ber optics and their applications. In 1990, she was at the Fiber Optics Laboratory, General Physics Institute, Moscow (Russian Academy of Sciences) under the Royal Society Postdoctoral Exchange Fellowship. She was a Principal Systems Engineer at STC Submarine Systems, Ltd., London, U.K., and Nortel Networks (Harlow, U.K., and Ottawa, Canada), where she was involved in the design and planning of optical ber transmission networks. During 19942004, she held a Royal Society University Research Fellowship at UCL, and in 2002, she became a Chairperson in Optical Communications and Networks. She is currently the Head of the Optical Networks Group, UCL. She has authored/coauthored more than 260 refereed journal and conference papers. Her research interests include optical networks, high-speed optical transmission, and the study and mitigation of ber nonlinearities. Prof. Bayvel is a member of the Technical Program Committee (TPC) of a number of conferences, including European Conference on Optical Communication (ECOC) and Co-Chairperson of the TPC for ECOC 2005. She is the 2002 recipient of the Institute of Physics Paterson Prize and Medal for her contributions to research on the fundamental aspects of nonlinear optics and their applications in optical communications systems. In 2007, she was the recipient of the Royal Society Wolfson Research Merit Award. She is a Fellow of the Royal Academy of Engineering (FREng), Optical Society of America, the U.K. Institute of Physics, and Institute of Engineering and Technology. Seb J. Savory (M07) was born in Stirling, U.K., in 1973. He received the M.Eng., M.A., and Ph.D. degrees in engineering from the University of Cambridge, Cambridge, U.K., in 1996, 1999, and 2001, respectively, and the M.Sc. degree in mathematics from the Open University, Milton Keynes, U.K., in 2007. His interest in optical communications began in 1991, when he joined Standard Telecommunications Laboratories, Harlow, U.K., prior to being sponsored though his undergraduate and postgraduate studies by Nortel. On completion of his Ph.D. degree in 2000, he joined Nortels Harlow Laboratories as Senior Research Engineer, where he was engaged in research into digital signal processing and advanced optical transmission systems. In 2005, he joined the Optical Networks Group at University College London (UCL), London, U.K., where he held a Leverhulme Trust Early Career Fellowship from 20052007, being appointed as a University Lecturer in 2007. From June 2009June 2010, he was also a Visiting Professor at the Politecnico di Torino, Italy. His research interests include digital coherent transceivers, optical transmission systems and subsystems, digital signal processing and nonlinear systems. Dr. Savory is a Chartered Engineer and an Associate Editor for IEEE PHOTONICS TECHNOLOGY LETTERS. He also serves on the technical program committee for the Optical Fiber Communication Conference, the European Conference on Optical Communication and the IEEE Photonics Society Annual Meeting.

Carsten Behrens (S09) was born in Potsdam, Germany, in 1981. He received the Dipl.-Ing. degree in electronic and electrical engineering from the Technische Universit t Berlin, Berlin, Germany, in a 2007. He is currently working toward the Ph.D. degree in the Optical Networks Group, University College London, London, U.K. His current research interests include coherent detection of higher order modulation formats and nonlinear effects in optical bers. Mr. Behrens was the recipient of the ErwinStephan Prize for excellent results in 2008.

Stephan Hellerbrand (S03) was born in Munich, Germany, in 1980. He received the B.Sc. degree in electrical engineering in 2003 from Technische Universit t M nchen (TUM), Munich, Germany, a u with a thesis on joint source-channel coding. In the same year he spent two terms at the University of Melbourne, Melbourne, Australia. In 2005, he received the Dipl.-Ing. degree in electrical engineering from TUM with a thesis on data compression, which was the result of a research visit to Lund Institute of Technology, Lund, Sweden. Since October 2005, he has been working towards the Dr.-Ing. degree as a full-time member of the research and teaching staff at the Institute for Communications Engineering of TUM. His research focus is on system modeling and signal processing for the compensation of transmission impairments in beroptic communication links.