Wei-Ren Peng,Itsuro Morita,Hidenori Takahashi,and Takehiro Tsuritani

(KDDIR&D Laboratories Inc.,Fujimino-shi,Saitama 356-8502,Japan)

Abstract In this paper,we propose direct-detection optical orthogonal frequency division multiplexing superchannel(DDO-OFDM-S)and optical multiband receiving method(OMBR)to support a greater than 200 Gb/s data rate and longer distance for direct-detection systems.For the new OMBR,we discuss the optimum carrier-to-sideband power ratio(CSPR)in the cases of back-to-back and post transmission.We derive the analytical form for CSPRand theoretically verify it.A low overhead training method for estimating I/Q imbalance is also introduced in order to improve performance and maintain high system throughput.The experiment results show that these proposals enable an unprecedented data rate of 214 Gb/s(190 Gb/s without overhead)per wavelength over an unprecedented distance of 720 km SSMFin greater than 100 Gb/s DDO-OFDM systems.

Keyw ords orthogonal frequency division multiplexing(OFDM);direct detection;multiband transmission

1 Introduction

F or years,papers have been written on direct-detection transmission because it uses a very simple receiver that usually only requires photodiodes for detection[1],[2].Most research on direct-detection has been focused only on the dispersion-managed links,and not until recently has electronic equalization has been introduced into optical communications to compensate for chromatic dispersion(CD)[3].Electronic equalization has ushered in a new era of dispersion-unmanaged links that do not require dispersion map design and offer a lower-nonlinear transmission medium[4].Joint direct-detection receiving and dispersion-unmanaged links is an interesting topic related to implementation cost.

To incorporate direct-detection in a dispersion-unmanaged link,single-carrier transmission with pre-equalization[3]and multicarrier transmission with post-equalization[5]have been proposed.The pre-equalization in[3]could work very well to counter CD;however,when used in dynamic routing networks,it requires the link information(the exact information of the accumulated CD)beforehand via complex signaling methods[6],[7].The proposalin[5],commonly referred to as direct-detection opticalorthogonal frequency division multiplexing(DDO-OFDM),automatically estimates and compensates for CD at the receiver,without any link information.Therefore,DDO-OFDM could easily support both point-to-point and dynamic routing networks.DDO-OFDM should be repositioned because of its transparency to different kinds of network architectures.

Even with these advantages,conventional DDO-OFDM has limited transmission performance that might restrict its application only to short-reach networks[8],[9].Among the many reported DDO-OFDM systems,a self-coherent OFDM[10]system extends both the capacity and reach of conventional DDO-OFDM and significantly increases the receiver’s complexity,which diminishes the inherent benefits of using direct detection.Therefore,a laudable goal for a DDO-OFDM system is to improve transmission with simple solutions while keeping the cost of direct detection low.

In this paper,we propose DDO-OFDM superchannel(DDO-OFDM-S)with a simple opticalmultiband receiving(OMBR)[11],[12]method.DDO-OFDM-Shas dual carriers at both sides to reduce receiver bandwidth,and OMBR detects the signal band-by-band to relax the sampling rate of the receiver and to reduce the frequency gap between carriers and superchannelfor high spectralusage.In this paper,an optimum carrier-to-sideband power ratio(CSPR)is discussed for back-to-back and post transmission,and its analytical form is theoretically derived and verified.We also describe a low-overhead I/Q imbalance estimation approach that can be used to adaptively compensate for I/Q imbalances while maintaining the system’s high capacity.In greater than 100 Gb/s DDO-OFDM systems,the proposed DDO-OFDM-Senables an unprecedented 214 Gb/s data rate per wavelength over 720 km standard single-mode fiber(SSMF)with EDAF-only amplification.

In section 2,the working principles of DDO-OFDM-Sand OMBRare given,and a solution for generating DDO-OFDM-Swith only one wavelength is proposed.In section 3,a CSPRfor the new OMBRis suggested that is optimized for minimum required optical signal-to-noise ratio(OSNR)in back-to-back transmission(conventional definition)and better signalperformance after transmission.An analytical form of optimum CSPRin back-to-back transmission is also derived.In section 4,we describe the setup for our 16-QAM,214 Gb/s DDO-OFDM-S transmission experiment.In section 5,we propose a low-overhead I/Q imbalance training method and introduce a zero-overhead phase noise compensator.In section 6,numerical results for CSPRand experiment results for transmissions are presented.Section 7 concludes the paper.

2 Principle of Operation

The proposed DDO-OFDM-Scan be better understood from its spectrum(Fig.1a).The spectrum consists of one OFDM superchannel at the center and two optical carriers,inserted at both sides,with a frequency gap,Ng,from the superchannel.In this paper,the superchannel and carriers are assumed to be aligned in the same polarization,and only this polarization is used for transmission.The use of dual-polarization transmission,that is,polarization division multiplexing(PDM),would be a powerful means of doubling the channel capacity;however,in this paper,we conduct a preliminary investigation into single-polarization transmission.The superchannel itself comprises closely spaced OFDM bands[13],denoted 1,2,3,...,N in Fig.1(a).The surrounding two carriers,1 and 2 in Fig.1(b),demodulate the half that is lower-frequency bands and the other half that is upper-frequency bands.Carrier 1 is responsible for bands 1 to N/2,and carrier 2 is responsible for bands N/2+1 to N.When N is an odd number,that is,when odd-numbered OFDM bands are transmitted,the central band should have similar spectral distances from both carriers so that it can be demodulated by either of them.

To demodulate a DDO-OFDM-S,we use the proposed OMBRat the receiver to process the signal in a band-by-band manner,as shown in Fig.1(b).An optical coupler splits the received signalinto N parallel paths,each of which consists of one dual-passband filter(DPF)that has its two passbands targeting the desired OFDM band and one of the carriers.Therefore,in the m th path in Fig.1b(top),only the m th OFDM band,and carrier p(where p=1 if m≤N/2 or p=2 if m＞N/2)reach the photodiode for detection.After the photodiode,the converted electrical current in each path contains the desired passband signal and unwanted(signal-signal)beat interference near the direct current(DC)frequency.After down-conversion,the desired baseband signal in each path is lowpass filtered,digitally-sampled,and processed with a regular OFDM equalizer.In this way,all the bands in the superchannelcan be demodulated with a bank of cheaper low-bandwidth receivers.

▲Figure 1.(a)Direct-detection optical OFDMsuperchannel(DDO-OFDM-S),(b)optical multiband receiving(OMBR),and(c)a possible transmitter solution to implementthe proposed DD?O-OFDM-S with single laser source.

Compared with a conventional DDO-OFDM signal[14],our DDO-OFDM-Shas unique characteristics:It uses dual carriers,and the width of the frequency gap is smaller.Dual carriers are used to reduce the bandwidth of the receivers.If only carrier 1 were used with the superchannel,the carrier would have to cover all the transmitted bands from 1 to N at the receivers.This means the photodiode in N th branch would need to have an ultrahigh bandwidth to cover the frequency between carrier 1 and band N.Therefore,the dual-carrier arrangement,in which each carrier takes care of 50%of the bands in the superchannel,can relax the high-bandwidth requirement of the receiver by a factor of two.The smaller gap width is created by optical pre-filtering before detection.In conventional systems,the required gap width has to be at least equal to the sideband bandwidth,and this leads to poor spectral efficiency.With OMBR,the DPF only allows one desired band and the carrier to be detected,so the required gap width can be reduced to a size similar to that of a single band.The gap width does not need to be the same as that of the whole superchannel,and this improves spectral efficiency.

The proposed DDO-OFDM-Sand OMBRhas the following advantages:

·better spectral efficiency because of the reduced gap width

·relaxed bandwidth requirement for receivers because of the dual-carrier arrangement and band-by-band demodulation

·simple receiver architecture because only one filter is used for each band

However,an issue to be considered is how to generate the DDO-OFDM-S.There might be a number of options,including multiple lasers,to generate the DDO-OFDM-S.Fig.1(c)shows the transmitter architecture of a possible solution that uses only one laser.The output from the laser is first split into two tones by an intensity modulator.These tones are later coupled into the upper and lower branches using one wavelength interleaver.The two tones are individually modulated with electrical multiband OFDM signals[15]using two intensity modulators that are biased slightly away from the null.This processing results in a double-sideband OFDM signal with a carrier on each branch.An optical coupler combines the signals from the two branches(that form the DDO-OFDM-S)with two residual sidebands from the carriers(Fig.1c,insets).A following optical bandpass filter(OBPF)with appropriate bandwidth removes the residual sidebands.Finally,the output signalof the filter has a central superchannel surrounded by two optical carriers,which is the proposed DDO-OFDM-S.

3 Carrier to Sideband Power Ratio

In this section,we focus on CSPRfor DDO-OFDM-S.CSPRis the power of both carriers over the power of the superchannel,that is,CSPR=2Pc/(NPs),where Pcis the power of each carrier and Psis the power of each signal band.Typically,there is an optimum CSPR,denoted CSPRMRO.The CSPRMROdemands the minimum required OSNRat a target bit error rate(BER),usually BER=1e-3.This optimum CSPRhas been theoretically and experimentally proven to be approximately 0 d B for conventional DDO-OFDM systems[14],[16].However,CSPRMROitself is a function of the optical filters[17],[18],and thus its value is subject to change when a different receiving technique,such as the proposed OMBR,is used.In the following,we derive the analytical form of CSPRMROspecifically for our DDO-OFDM-Ssignal and OMBR.

A higher electrical signal-to-noise ratio(ESNR)usually leads to a lower BER;there is a one-to-one relationship between BERand ESNR.Therefore,the minimum required OSNRfor a target BERbecomes the minimum required OSNR for a target ESNR.Before CSPRMROcan be derived,the relationship between ESNRand OSNRhas to be determined.We assume the DDO-OFDM-Scontains an N-band superchannel at the center with two carriers at both sides and that both the passbands of the DPFhave a brick-wall shape and have the same bandwidths,B0.These bandwidths should be equal to or slightly greater than the bandwidth of one OFDM band,Bs.The received ESNRfor each desired band can be expressed as a function of OSNR:

where OSNR=(2Pc+NPs)/(NoBo),Nois the noise spectral density,and BNis the noise bandwidth.With Cauchy-Schwarz’s inequality,the upper-bound of the ENSR and the lower bound of the OSNRcan be obtained from(1):

Therefore,the minimum OSNRgiven an ESNR(or a target BER)can be achieved withwhich implies that CSPRMROis proportionalto the inverse of the square root of the band number.For instance,considering a nine-band DDO-OFDM-S,.In section 6,we verify this theory.

We have derived the CSPRoptimized according to minimum required OSNRin back-to-back transmission where only linear noise is considered.Because a lower required OSNRtypically offers a larger noise margin and longer transmission distance,CSPRMROis defined in terms of longer distance.However,for DDO-OFDM systems,we propose a new transmission strategy that uses a higher CSPR,other than CSPRMRO,to achieve a longer distance even though it may not meet the minimum required OSNRconstraint.This idea comes from a CO-OFDM system where the carrier is offered by the local oscillator(LO)that has a high carrier-to-noise ratio(CNR).This concept is shown in Fig.2 for two cases—lower CSPR(CSPRL)and even higher CSPR(CSPRH)—where sideband powers are assumed to be equal.If both signals are transmitted over the same link with the same amount of accumulated ASEnoise,the signal with CSPRHhas a higher CNRthan the signal with CSPRL(Fig.2).However,the benefit of high CNRcomes at the price of larger fiber nonlinearities caused by the use of higher carrier power(under the same sideband power constraint).Therefore,the optimum CSPRshould trade off the better CNRfor lower fiber nonlinearities,which may not be equal to CSPRMRO.In section 6,we show that the optimum CSPR,defined in terms of longer distance,is greater than CSPRMROwhen both the linear noise and fiber nonlinearities are taken into account.

4 Experiment Setup

Fig.3 shows the experiment setup for the 214 Gb/s DDO-OFDM-Ssystem.A 100 kHz line width external cavity laser(ECL)operated at approximately 192.76 THz is the transmitter light source,and this is followed by a 1×2 optical splitter that equally couples the laser output into the upper and lower branches.In the upper branch,the light is first modulated along with the electrical OFDM signal via an optical in-phase/quadrature(I/Q)modulator.The OFDM waveform is generated offline with MATLAB and comprises continuous frames,each of which contains two training symbols and 150 data symbols.In each OFDM symbol,binary data is randomly generated,mapped to 16-QAM format,and modulated onto 152 data subcarriers,which are later zero-padded to a fast Fourier transform(FFT)size of 256.A pilot is not used in order to minimize overhead.After inverse FFT(IFFT),a length of ten-point cyclic prefix(CP)is attached to each OFDM symbol,creating a totalof 266 points per symbol.This OFDM waveform is then loaded into an arbitrary waveform generator(AWG)that has its real and imaginary outputs driving the optical I/Q modulator with a 10 GSa/s sampling rate.Hence,the raw data rate of the output signal is 23.75 Gb/s and occupies a bandwidth of approximately 6 GHz.The output of the I/Q modulator is sent to a nine-comb generator,which cascades two intensity modulators driven by 6.5 GHz and 19.5 GHz sine wave signals in order to emulate a nine-band superchannel[6].The output superchannel occupies a bandwidth of approximately 58 GHz(with band spacing of 6.5 GHz)and has an aggregate data rate of 214 Gb/s.After removing the training,CP,and 7%FEC overhead,the net data rate is 190 Gb/s.An optical coupler follows to combine this superchannel with the signal from the lower branch.In the lower branch,the light is first modulated with a 40 GHz electrical sine wave signal with one intensity modulator biased at the null.This results in two strong carriers spaced at 80 GHz as well as one residualcarrier at the laser frequency.A 50:100 GHz wavelength interleaver(IL)is used to suppress this residual carrier,leaving only the two 80 GHz-spaced carriers.The two 80 GHz-spaced carriers are then combined with the 58 GHz superchannel,forming the DDO-OFDM-Sspectrum,as shown in Fig.3 insets(a),(b)and(c).The frequency gap between the carrier and the superchannel is approximately 11 GHz at both sides.The spectral distance between carriers(80 GHz here)can be further reduced at the expense of sacrificing the edge bands because of beat interference.In the experiments,this 80 GHz spacing is appropriate for balancing spectrum usage and receiving performance.At the transmitter output,the DDO-OFDM-Sis sent to a re-circulating fiber loop that comprises three EDFAs and three spools of 80 km SSMF.After 720 km transmission,the signal is fed to the optical multiband receiver.At the receiver,the signal is pre-amplified with an EDFA and then passed through a DPFthat has two10 GHz passbands targeting one of the carriers and the desired signal band.Because of the power loss of DPF,an EDFA and an 80 GHz optical bandpass filter(OBPF)raise the signal power before the signal enters the photodiode.In this experiment,the first to fifth bands are demodulated with the left carrier(carrier 1),and the sixth to ninth bands are demodulated with the right carrier(carrier 2).After the photodiode,the desired band is down-converted to its baseband via an electrical I/Q demodulator.This I/Q demodulator comprises one splitter,one synthesizer,one power amplifier,two mixers,and two electrical low-pass filters(3 d B bandwidth=3.7 GHz).The I/Q output signals are recorded by a real-time scope operated at 20 GSa/s.Synchronization,cyclic prefix removal,channel estimation,and equalization are performed offline using MATLAB.The BERis evaluated with an error-counting method,and for each BERanalysis,2 million sampling points are considered.

Figure 3.?Experiment setup for DDO-OFDM-S.(a)-(c)are optical spectra at point(a),(b)and(c),respectively.

▲Figure 2.Spectra of low-CSPRand high-CSPRDDO-OFDM-S.

Fig.3(a)shows the optical spectra of the transmitter output(resolution=20 MHz);Fig.3(b)shows the DPFoutput(targeting the 5th band);and Fig.3(c)shows the digital spectrum of the 5th band after the real-time scope(where the band index is defined in Fig.3a).In Fig.3(c),the power ripples on top of the signal come from the power reflection,at the RFports,of the mixers in the I/Q demodulator.These ripples are a function of frequency and introduce some OSNR penalty to the system,that is,implementation penalty.

Here,we highlight several points about the experiment setup.First,at the transmitter,the length of the optical paths between the sideband and carrier branches should be controlled so that they are as similar as possible.A significant difference in length would lead to strong phase incoherency between the carrier and sideband,and this would cause dramatic phase noise after the photodiode.In this experiment,we not only equalized the optical lengths of the carrier and sideband paths,but we also used a zero-overhead phase noise compensator at the receiver(section 5).Second,6.5 GHz and 19.5 GHz frequencies for the 9-comb generator are phase-locked in order to maintain the orthogonality and reduce linear crosstalk between the adjacent bands[13].However,because of the sufficiently large band spacing,the performance is hardly degraded,even if we remove the phase locking between the synthesizers.Third,to obtain a broadband bandwidth,we assemble the I/Q demodulator with discrete components rather than use an integrated I/Q mixer,which would introduce I/Q imbalances into the signal.This would mean I/Q imbalance estimation and compensation would be critically necessary in our system.In section 5,we introduce a low-overhead I/Q imbalance estimation approach that uses only two training symbols,that is,the same symbols for channel estimation.

5 Low-Overhead Signal Processing Methods

To compensate for I/Q imbalance and phase noise,we propose a low-overhead training method to estimate I/Q imbalance and introduce a zero-overhead decision-directed phase noise compensator(DD-PNC)[19].

5.1 I/Q Imbalance Training Method

For each frame,the proposed method uses only two consecutive training symbols(the same as those for channel estimation).The first symbol is randomly generated,and the second symbolsimply copies the first one and inverts the signs of the data symbols on negative subcarriers.The two consecutive training symbols are denoted[ak,b-k]and[ak,-b-k],where akand b-kare the data symbols on k th and-k th subcarriers,respectively,and k is a positive integer ranging from 1 to Nd/2,with Ndbeing the data subcarrier number.At the receiver,the received training symbols disrupted by I/Q imbalances are denoted[pk,q-k]and[rk,l-k]for the first and second training symbols,respectively.Then,the input and output symbols can be expressed in 2×2 mutually coupled matrixes[20]:

The four elements in the 2×2 channel matrix H(Hij,which should contain both the channel response and I/Qimbalances)can be easily derived with H11=(pk+rk)/(2ak),;andfor which the estimation accuracy can be further improved using the intrachannel frequency-domain average method[15].

After obtaining the four elements,the inverse matrix of H can be derived for subsequent equalization.Because this technique uses the same training symbols(two per frame)for both channel and I/Q imbalance estimation,the training overhead is relatively lower than that in[20].

This data-aided adaptive approach is simple,has low overhead,and is a better way to estimate and compensate for I/Q imbalance.In the offline I/Q estimation proposal in[21],the I/Q imbalance in each individual receiver needs to be customized;however,using this new approach,the I/Q imbalances of each individual receiver can be handled adaptively using only two training symbols(which can the same ones for channel estimation).In particular,the adaptive method is suitable for lab experiments where I/Q imbalances vary over time because of the replacement of any component or the slow bias drift of the optical modulators.

Here we briefly discuss the computational complexity of this I/Q imbalance estimation method.At the training stage,to obtain the channel matrix for all subcarriers,2Ndcomplex multiplications are needed for each OFDM symbol,and the matrix inversions for equalization need an additional 3Ndmultiplications.Therefore,at the training stage,5Ndmultiplications are needed,which is 4Ndmore multiplications than in regular channel estimation.At the equalization stage,two subcarriers are jointly equalized via a 2×2 matrix so that the number of multiplications for each OFD symbol is 2Nd,which is Ndmore multiplications than in regular channel equalization(using a one-tap equalizer).

▲Figure 4.Zero-overhead decision-directed phase noise compensator(DD-PNC).

5.2 Phase Noise Compensation Method

DD-PNC uses all tentative decisions of the OFDM symbol being processed to estimate its common phase error(CPE).The CPEis later used for de-rotating the phase of the OFDM symbol[19].If noise and I/Q imbalance are ignored,the received data symbolon the k th subcarrier can be simply written as Yk=UXk,where Xkand Ykare the transmitted and equalized symbols(with no channel effect)on k th subcarrier,respectively,and U is the CPE,which is assumed to be independent of the subcarrier index k.Assuming CPEis not significant and the tentative decisions are statistically reliable,U can be estimated with U≈(1/Nd)Im{Σ?kYk/xk},where xkis the tentative decision of the equalized symbol,Yk,Ndis the number of data subcarriers,and Im{x}takes the real part of x.The output for finaldecision will be Ykexp(-ju),and in this output,CPEshould have been greatly mitigated.Fig.4[19]shows the corresponding processing.A more detailed discussion on,for example,computational complexity,can be found in[19].

The signal processing sequence is:1)synchronization,2)CPremoval and FFT,3)joint channel and I/Q imbalance estimation using proposed training symbols,4)estimation enhancement using intra-symbol frequency domain averaging method[22],5)joint channel and I/Q imbalance equalization with the inverse channel matrix,and 6)CPE mitigation using DD-PNC.

6 Results and Discussions

6.1 Numerical Results for Optimum CSPRTransmission

The detailed system parameters in the simulations are almost the same as those in the experiment setup in section 4,except for CSPR,which is herein treated as a variable in order to determine its effect on system performance.The transmission link comprises nine spans of 80 km SSMF(720 km in total).The fiber loss,dispersion,and dispersion slope are 0.2 d B/km,16 ps/(nm.km),and 0.02 ps/(nm2.km),respectively.The effective area is 80μm2,and the nonlinearity coefficient is 1.3(W-1km-1).The EDFA’s noise figure is set to 6 d B,and the EDFA’s gain is set to 16 d B,which fully compensates for the fiber loss of each span.Each passband profile of DPFhas a second-order Gaussian shape and a 3 d B bandwidth of 10 GHz.The signal quality is expressed as Q2,obtained from the BERin(6):

where erfcinv{x}gives the inverse complimentary error function value of x,and BERis derived with the direct-error counting method.OSNRis defined with 0.1 nm noise bandwidth.

Fig.5 shows the Q2factor versus CSPRin back-to-back transmission in which OSNR=25 d B.The optimum CSPRin back-to-back transmission is approximately-3 d B,which conforms to our theoretical prediction that≈-3 d Band verifies our analytical analysis in section 3.This result differs from the previously derived optimum CSPRof 0 d B[12],[14]because of the OMBRreceiving method used in our system.However,this new CSPRMROis obtained in back-to-back transmission without nonlinear distortions being taken into account.Therefore,it is necessary to study the CSPRMROagain after transmission.In Fig.6(a)and(b),we show Q2as a function of launch power with different CSPRs.In Fig.6(a),the data sets for each band are mutually independent,that is,uncorrelated at the transmitter;however,in Fig.6(b),the data sets for all the bands are the same,that is,strongly correlated at the transmitter.The correlated data sets might lead to strong nonlinear phase noise at the beginning of the link[13],which is exactly the case in our experiment.Therefore,the case with correlated data sets has to be compared with the case with uncorrelated data sets.Because the curves in Fig.6(a)and(b)are similar with respect to CSPRs,we simply focus on the results shown in Fig.6(a).With a lower CSPR,for example,-9 d B,the optimum launch power is similar to that of CSPRMRO=-3 d B,and the optimum Q2has approximately 2 d B degradation relative to that of CSPRMRO.With a higher CSPR,for example,3 d B,the optimum launch power increases by approximately 3 d B.The raised power comes mostly from the carriers.More importantly,there is an approximately 0.7 d B improvement over CSPRMROby using 3 d B CSPR.This implies that with 3 d B CSPR,the improved CNRcan bring more benefits than the negative four-wave mixing(FWM)terms caused by the high power of the carriers.A further increase in CSPRto 9 d Bwould result in a slight sacrifice of signalquality but drastically enhance optimum launch power.Fig.6(c)shows the optimum Q2factors andlaunch power as a function of CSPR(with 3 d Bresolution).These functions can be used to determine the optimum CSPR(after transmission).Again,in Fig.6(c),the curves for uncorrelated and correlated data sets are shown for comparison.From Fig.6(c),we can draw the following conclusions:

Figure 5.?Q2 factor vs.CSPR with OSNR=25 dB.Except for CSPR,the simulation parameters are all similar to thoseof the 16-QAM,214 Gb/s experiment in section 3.

·The optimum CPSRs after transmission are similar for both cases and are approximately 3 d B to 6 d B,which is higher than the CSPRMROof-3 d B in back-to-back transmission.This means that higher carrier power,that is,higher CSPR,is encouraged in order to overcome linear noise and nonlinear distortions.

·At the optimum CSPRs,the uncorrelated bands give approximately the uncorrelated bands outperform the correlated bands by approximately 1.7 d B,which suggests that the distance achieved in our experiment could be further extended if uncorrelated data bands were used.

·Even though higher CSPRmight lead to better performance,the corresponding optimum launch power is relatively high.In a WDM system,a suitable CSPRshould be determined,taking into consideration both the channel number and the maximum output power of fiber amplifiers.

▲Figure 7.Experimental demonstration for 16-QAM,214 Gb/s DDO-OFDM-S:(a)OSNRtolerance,with 0 km SSMF,(b)Q2 vs.launch power with 720 kmSSMF,and(c)Q2 vs.band index with 720 km SSMF.

6.1.1 Discussion

The CSPRthat demands the minimum required OSNR,which can be derived easily in back-to-back,has long been considered the optimum value for DDO-OFDM systems.However,we have shown that even greater CSPR,with higher optimum launch power,can further improve the signal quality after transmission.The optimum CSPRshould be the value that promises better signal quality after transmission rather than in back-to-back only.

6.2 Experiment Results for DDO-OFDM-S Transmission

The band index is defined in the inset of Fig.3(b).OSNRis measured with 1.6 nm(200 GHz)resolution that covers the whole signal’s bandwidth and is later scaled to the presented value with 0.1 nm resolution.Q2factors(in decibels)are derived from the BERin(6).

Fig.7 shows the experiment results for 16-QAM,214 Gb/s DDO-OFDM-S.To achieve longer distance,we use the new high-CSPRstrategy in section 3.In this experiment,we use approximately 10 d B CSPR(optimum after 720 km transmission)to achieve a higher CNRafter transmission.Fig.7(a)shows the OSNRtolerance in back-to-back.The dashed line is the theoretical limit and is shown for comparison.The required OSNRfor a BERof 1e-3 is approximately 33.8 d B due to the use of 10 d B CSPR.Compared with the theoretical limit,the implementation penalty is approximately 3.4 d B.This implementation penalty can be attributed to imperfections in the components,power ripples(see section 4),residual phase noise,and linear crosstalk between bands.With 10 d BCSPR,the required OSNRfor the superchannel alone could be approximately 23.4 d B.This OSNRis similar to that for a coherent system[23],where the required OSNRat BER=1e-3 is approximately 23.4 d B with 224 Gb/s and would become approximately 23.2 d B after scaled to 214 Gb/s.The signal quality in our DDO-OFDM-Ssystem is almost dominated by the sideband because of the high applied CSPR.

Fig.7(b)shows the measured Q2as a function of the fiber launch power after 720 km transmission.The presented curves correspond to the first,fifth,and ninth bands,respectively.The optimum power is approximately 8 d Bm,which provides the best performance under the limitations of ASEand nonlinear distortion.If we consider the high CSPRof approximately 10 d B,the launch power for the superchannel alone is only about-2.4 d Bm.

Fig.7(c)shows measured Q2as a function of the band index in back-to-back and after 720 km SSMF.The optimum launch power of 8 d Bm is applied.In back-to-back,the fifth band shows the worst performance(Q2≈12 d B)caused by insufficient bandwidth of the receiver’s components;the other bands have Q2factors greater than 12 d B.After transmission,the worst performance occurs on the first band,which still yields a Q2higher than the 7%FEC threshold of 8.53 d B[24],and the Q2factor variation between the different bands is reduced because it is the noise and nonlinear distortion,not the receiver’s bandwidth,that dominate performance.

6.2.1 Discussion

Using our proposals,in the 16-QAM experiment,we generated and demodulated the record capacity per wavelength(214 Gb/s)for DDO-OFDM systems.We also achieved a record distance of 720 km SSMFfor greater than 100 Gb/s DDO-OFDM systems.Because of the strong nonlinear phase noise at the beginning of the link,a longer distance might be possible by using uncorrelated data bands,for example,approximately 2 d B Q2difference,as shown in Fig.6(c).To reach 720 km transmission,we used a high CSPRof approximately 10 d B,that is,high carrier power,to provide better CNRafter transmission.Such a high carrier power would induce strong FWM terms at both sides out of the carriers,which in our experiment,could be removed through moderate optical filtering.In a WDM system,these FWM terms would possibly overlap with the adjacent channels and become in-band nonlinear noise that results in some penalty.Therefore,in WDM systems,CSPRshould be carefully controlled and optimized so that strong FWM terms are not involved.Another simple way to avoid FWM is to use polarization interleaving for WDM neighboring channels.However,this would eliminate the possibility of using PDM.

7 Conclusion

We have proposed DDO-OFDM-Sand OMBR,a high-capacity long-reach solution for DDO-OFDM systems.We have discussed optimum CSPRin back-to-back and after transmission and found that a higher carrier power is needed to improve performance after transmission.We have also derived the analytical form of back-to-back CSPRand verified its correctness using numerical simulations.To maintain the system’s high throughput,we have proposed a low-overhead I/Q imbalance estimation method.Our experiment shows that in greater than 1 Gb/s DDO-OFDM systems,the proposed DDO-OFDM-Senables unprecedented capacity of 214 Gb/s per wavelength over an unprecedented distance of 720 km SSMF.