Nonlinear Inverse Synthesis and Eigenvalue Division Multiplexing in Optical Fiber Channels

Jaroslaw E. Prilepsky, Stanislav A. Derevyanko, Keith J. Blow, Ildar Gabitov, and Sergei K. Turitsyn Aston Institute of Photonic Technologies, Aston University, B4 7ET Birmingham, United Kingdom Department of Physics of Complex Systems, Weizmann Institute of Science, Rehovot 76100, Israel Department of Mathematics, University of Arizona, 617 North Santa Rita Avenue, P.O. Box 210089, Tucson, Arizona 85721, USA (Received 13 February 2014; published 1 July 2014)

Optical fiber systems form the backbone of the global telecommunication networks and currently carry the majority of the world's information traffic, with the "fifth generation" of optical transmission systems operating with advanced modulation formats, e.g., orthogonal frequency division multiplexing (OFDM), digital signal processing techniques, etc. [1].Rapidly increasing data rates in the core fiber communication systems are currently approaching the limits of transmission technologies, many of which were originally developed for linear (radio) communication channels [2].The paramount limit on channel spectral efficiency is imposed by fiber nonlinearity [2][3][4][5][6][7][8].Therefore, the applications of "linear techniques" in a realistic fiber channel are limited by the nonlinear properties of the fiber medium itself.The increase in the signal power to improve the signal-to-noise ratio (SNR) leads to power-dependent nonlinear transmission distortions.Thus, nonlinearity is an essential component in the design of advanced fiber communication systems, but it is often shunned by engineers in view of its complexity.There is an evident need for radically different approaches to coding, transmission, and processing of information in fiber communication channels that would take into account the nonlinear properties of the optical fiber.In this Letter, we present a nonlinear inverse synthesis (NIS) method-a highly promising alternative transmission technique based on the integrability of the nonlinear channel.
The evolution of light in an optical fiber is modeled by the nonlinear Schrödinger equation (NLSE) [2,4], which accounts for the interplay between dispersion and nonlinearity.Averaging over periodic gain and loss variation in practical systems leads to effectively conservative signal dynamics governed by the renormalized NLSE [9].Moreover, a recent experimental demonstration of a quasilossless fiber span [10] shows that gain (loss) variations can be compensated continuously along the fiber.Importantly, the NLSE (without perturbations) belongs to the class of the integrable nonlinear systems [9,[11][12][13][14].The integrability of the NLSE means that one can present nonlinear field evolution in a special basis, within which the dynamics of individual "nonlinear normal modes" is effectively linear.A powerful method-the inverse scattering transform (IST) [9,[12][13][14], that is a nonlinear analog of the Fourier transform [14]-can be applied to find the solution of an integrable equation.Within this method, the first step (decomposition of initial conditions into spectral data) consists of solving a linear spectral Zakharov-Shabat problem (ZSP) [12].Any field evolving according to the NLSE can be described by a set of scattering data evolving in a linearlike manner, this fact is the foundation of the nonlinear Fourier transform (NFT) method [9,[12][13][14].The NFT does to the NLSE what the FT does to the linear equations: Just as the linear FT changes dispersion to a phase rotation in frequency space so the NFT leads to a trivial phase rotation of the spectral data.Thus, the fiber nonlinear effects are effectively included in the NFT [9,[12][13][14].This property constitutes the general idea of the eigenvalue communication introduced first in [15], the essence of which is to use invariant ZSP eigenvalues to encode and transmit information [16].The use of NFT decomposition opens fundamentally new possibilities for advanced coding and modulation schemes, resistant to nonlinear fiber effects.
Note that, in Ref. [15], only the discrete part of the ZSP spectrum was considered.This discrete spectrum corresponds to the soliton part of the NLSE solution [9,[12][13][14].On the other hand, the vast amount of available degrees of freedom contained in the continuous part of the spectral data has been generally overlooked so far.In recent years, the concept of transmission built on pure fundamental NLSE solitons has lost much of its appeal in contemporary wavelength division multiplexing systems due to problems caused by soliton collisions and corresponding cross talk [9,17] (though a recent study [18] indicated mechanisms for the mitigation of intersoliton cross talk).Very recently there has arisen a considerable renewed interest in transmission based on the integrability and IST principles [19][20][21], where, not the solitons themselves, but the IST spectral data associated with each solitary degree of freedom were utilized.In this Letter, we, however, apply the idea of integrability in a completely novel context of nonsoliton coherent optical communications.We introduce a new approach to digital signal processing based on the encoding of information directly onto the continuous nonlinear signal spectrum that evolves linearly along the transmission line in a nonlinear integrable channel.We call this approach nonlinear eigenvalue division multiplexing.The main result reported here is that, by applying the IST technique, one can develop a straightforward signal processing routine for compensating nonlinear distortions of arbitrary magnitude.
We consider, as the master model, the NLSE with constant anomalous dispersion, describing the propagation of a slowly varying optical field envelope qðz; tÞ inside a single-mode nonlinear fiber [2] where z stands for the propagation distance and t is the time in the frame comoving with the envelope velocity.In Eq. ( 1), time is normalized to the characteristic duration of our input signal T s , distance-by the dispersion length: 67 ps 2 =km is the dispersion coefficient; the power is measured in units of ðγZ s Þ −1 , with the nonlinearity coefficient γ ¼ 1.27 ðW kmÞ −1 .The random quantity ηðt; zÞ describes the noise generation due to amplifier spontaneous emission (ASE) [2]: It is a symmetric additive complex Gaussian white noise with zero average, fully characterized by its autocorrelation intensity: hηðt; zÞηðt 0 ; z 0 Þi ¼ 2Dδðt − t 0 Þδðz − z 0 Þ [the overbar means complex conjugation].This model describes the amplification scheme, in which the distributed Raman gain exactly compensates for the fiber loss: Refs.[2] provide a detailed account of how noise intensity D is related to the line parameters.
As the NLSE with zero right-hand side is integrable, the explicit procedures for the forward and backward NFT are known [9,[12][13][14].The decomposition of the signal into spectral data (forward NFT) is achieved by solving the ZSP equations for auxiliary functions where the input pulse shape qðtÞ acts as a potential.Here ζ is a (generally complex) eigenvalue, , where the index n runs over all discrete eigenvalues of ZSP; rðξÞ plays the role of the nonlinear spectral profile.The nonlinear spectral function (NS) defined as NðωÞ ¼ −rðξÞj ξ¼−ω=2 , serves as the direct nonlinear analogue of the Fourier spectrum, tending to the ordinary FT of qðtÞ in the linear limit [22].The evolution of rðξÞ is trivial: rðξ; LÞ ¼ rðξÞe −2iξ 2 L (with L being the propagation distance), and substituting this expression into the definition of NS, one finds that NðωÞ obeys the linear dispersion law of the NLSE.Therefore, the orthogonality of nonlinear normal modes is preserved during signal propagation within the NLSE.
Transmission using the NIS.-The preliminary "proof of concept" demonstration that the nonsolitonic part of NS can be used for mitigating nonlinear distortions was presented in Refs.[21,23].However, these methods have limited applicability being confined to low-power initial signals or to the normal dispersion case, where no solitons generally appear.The case studied in our work is rather more challenging and practically important: For the focusing NLSE, the eigenvalue spectrum of ZSP generally brings about complex discrete eigenvalues corresponding to solitons; when the input is randomly coded, the eigenvalue portrait of the ZSP can be extremely involved.
To get rid of this issue, we suggest synthesizing the profile in the time domain starting from given encoded shapes in the nonlinear spectral domain (see also discussion in the Supplemental Material [24]).This idea is similar to that widely used for the syntheses of corrugated waveguides [25] and Bragg gratings [26,27]: one creates the input profile having the desired properties, starting from the nonlinear spectral data, and then, employs the backward NFT synthesizing the profile in the time domain.During the evolution, the spectral data undergo just a phase rotation, and hence, after unrolling this "nonlinear dispersion" at the receiver, the initial information can be recovered without nonlinear signal degradation.The scheme of the NIS method vs the simple conventional transmission is illustrated in Fig. 1; we show the true profiles for the transmission of a truncated harmonic: qðtÞ ¼ e iΩt , t ∈ ½0; T, Ω ¼ 2π=T.The NIS method involves two stages, see Fig. 1(b): (i) The backward NFT at the transmitter, providing the profile q GLM ð0; tÞ in the time domain corresponding to a desired initial NS NðωÞ; (ii) Recovery of the NS at the receiver by the forward NFT, i.e., by solving Eq. ( 2), and consequent linear dispersion removal (the latter is for recovering the initial phase information).For the conventional transmission, Fig. 1(a), we also removed the linear dispersion at the receiver.One observes, that while the profile for the conventional transmission is heavily degraded by the nonlinearity, the recovery of the initial waveform after using the NIS method is perfect.We now discuss the numerical complexity of the NIS as compared to the popular digital back propagation (DBP) method for the removal of the nonlinear distortions [28].In DBP, one reads the waveform at the receiver, inserts it in the noiseless NLSE, and then solves the NLSE in the backward direction.The numerical method for the NLSE integration is usually the split-step Fourier method [4,28], which requires N z M t log M t operations; M t being the number of discretization points in time, and N z -the number of steps in z, that grows with the transmission length and pulse power.In realistic transmission problems, one typically has N z ≫ 1 .The NIS method involves just two nonlinear transforms, each of those requiring ∼M 2 t operations for well-developed methods [27,29].Even with such an estimate, the complexity of the NIS can already be comparable to that of the DBP [21].However, the recent advancements in numerical NFT indicate that the complexity of NIS can be reduced even further.Namely, one of the fastest methods for the GLME solution is based on a single inversion of a Toeplitz matrix [27].Using the newest algorithms [30], this inversion can be performed in just M t log M t operations.For the ZSP, a recent study [31] suggests, that the recovery of continuous NS can be made in only ∼M t log 2 M t operations.With these estimates in mind, we believe that NIS-based transmission methods can significantly outperform the DBP in terms of numerical complexity for the digital signal processing.
To illustrate how NIS performs in a realistic digital transmission scheme, we consider the popular OFDM encoding [32] of the input waveform both in the linear and nonlinear spectral domain.The first question concerning the NIS implementation is what do the elementary OFDM base functions look like when its spectral shape is used in the NS domain?We present our results for two kinds of base functions in Fig. 2 base function is utilized in the Nyquist-format technique, which provides the highest spectral efficiency [33].As seen in Fig. 2(b), the corresponding profiles in the time domain, q GLM ðtÞ, are already not symmetric, and the asymmetry grows with the increase of c (see the Supplemental Material [24] for analytical considerations).In the OFDM scheme, the elementary base in the time domain is simply a rectangle (i.e., a sinc function in the ω domain) and, without loss of generality, in what follows, we will focus on this format [24].The inverse NFT of the single OFDM spectral tone, NðωÞ ¼ csincðωÞ is given in Fig. 2(c), (d) for different values of amplitude c.We see that for a sufficiently large c the waveform of q GLM ðtÞ is significantly different from a rectangular profile occurring the linear case: While losing its symmetry, the profile develops an oscillatory advancing tail.The general form of the spectrum of an arbitrary OFDMencoded data sequence is [32] QðωÞ where c jk is the jth informational symbol in the kth subcarrier, N sc is the total number of subcarriers, Ω k is the frequency of the kth subcarrier.The frequencies are spaced at multiples of the inverse of the symbol rate: Ω k ¼ 2πðk − 1Þ (normalized slot duration T s ¼ 1), which ensures the orthogonality of subcarriers.In Figs.2(e)-2(f) we present the spectrum of a single OFDM slot with 10 subcarriers with c k ¼ 1, and a corresponding inverse NFT, q GLM ðtÞ.In this case, the structure of the advancing tail is more involved and reflects the structure of the pulse itself.
For illustrative reasons, we now employ random quadrature phase shift keying (QPSK) encoding of the coefficients c jk [34]: The absolute value of c jk is the same for each coefficient, jc jk j ¼ c, and the phase of each c jk randomly takes one of four discrete values Argfc jk g ¼ ϕ jk ¼ 2πp=4, p ¼ 0; …; 3. Let us turn to the performance of the digital NIS transmission as opposed to the ordinary one, Fig. 1, using the inverse linear FT of Eq. ( 4), qðtÞ, as an input for the ordinary transmission, and its inverse NFT counterpart, q GLM ðtÞ, for the NIS transmission.We quantify the noise influence on the transmission by optical signal-tonoise ratio (OSNR), defined as OSNR ¼ P s =2DLB ref [2,32], where P s is the average power of the input signal and B ref is a reference bandwidth (it is the simulation bandwidth in our case).The lower the value of OSNR, the stronger is the signal degradation by the noise.We show the results of our simulations on the constellation diagram indicating positions of the normalized coefficients C jk ¼ c jk =c after the transmission.For ordinary transmission, we use the scheme given in Fig. 1(a), the coefficients are recovered by a linear FT in the time domain.For NIS transmission, the scheme in Fig. 1(b) is used: The QPSK-OFDM coefficients are recovered from the NS domain.Our results are given in Fig. 3 for the OSNR ¼ 7 dB (medium noise).We observe that the conventional transmission performs poorly leading to an almost complete smearing of the constellation picture.In contrast, the locus of coefficients C jk for the NIS-based transmission is still close to the initial values (blue rectangles), indicating that the NIS method significantly outperforms its linear counterpart.The plots for different OSNR levels and lower input power are given in the Supplemental Material [24], indicating that our results are general.In Figs.3(b)-3(c) we present the probability density functions (PDFs) for the angular and radial distributions of C jk : These distributions also demonstrate the superior performance of the NIS as compared to the ordinary transmission.
In this Letter, we have demonstrated that the nonlinear spectral data can be used for high-quality long-haul information transmission.The advantage of using the NS is that it plays the role of the linear spectrum for the nonlinear integrable problem, and the encoded spectral content undergoes just the trivial phase rotation obeying the linear dispersion law.We have introduced the nonlinear inverse synthesis method that has its roots in the theory of fiber Bragg grating synthesis and consists in the direct encoding of the nonlinear spectral domain followed by reconstruction of the field profile in the time domain by solving the GLME.We illustrated the robustness of the NSbased transmission not only against nonlinearity, but also against the ASE noise using a popular OFDM modulation format as an example.

PRL 113 ,
013901 (2014) P H Y S I C A L R E V I E W L E T T E FIG. 1 (color online).The flow chart depicting the operations for an example initial waveform qð0; tÞ ¼ e 2πit=T if t ∈ ½0; T, 0 otherwise.The upper part (a) shows the conventional transmission scheme, and the lower part (b) corresponds to the NIS-based transmission.(a),(b) display the true profiles for T ¼ 1 ns, transmission length L ¼ 2000 km (noiseless case).