This invention relates generally to communication networks, and more particularly to equalizing and decoding coherently received optical signals.
A desire to increase the data rate and transmission distance of communication channels has driven engineers and designers to consider the use of coherent signal transmissions, e.g., in optical networks. Conventionally, optical communication networks have relied on the use of simple signalling methods to encode data bits onto an optical carrier.
The most common signalling method is intensity modulation, in which a laser is gated to allow high intensity light to enter a fiber optical cable when a ‘1’ bit is transmitted, and low intensity light when a ‘0’ bit is transmitted. This is called on-off keying. This signalling method has the advantage that it is easily demodulated by a simple detector including a photodetector (typically a photo-diode) and an appropitate threshold.
The main drawback of intensity signalling is that bandwidth efficiency is low, due to the fact that data are transmitted only in a single dimension, i.e., signal intensity. Coherent signalling methods allow for the transmission of multidimensional signals, by modulating both the intensity and the phase of the light emitted by the laser. This increases bandwidth efficiency.
Optical Communication Network
The transmitter 110 typically includes a forward error correction (FEC) encoder, and an FEC decoder 107 in the receiver 120, to ensure reliability in the presence of noise, because advanced modulation schemes reduce Euclidean distances between symbols. The coherent receiver 120 includes another laser light source 101, optical hybrid, demodulator and photo detectors (termed a “coherent detector”) 106.
Several impairments affect the performance of such coherent optical transmission systems. The fiber channel exhibits Chromatic Dispersion (CD), Polarization Mode Dispersion (PMD), non-linear distortion such as Self-Phase Modulation (SPM), and so on. Nonlinear impairments have become a major limiting factor for high-rate data transmissions in long-haul optical fiber channels.
In the prior-art, Digital Back-Propagation (DBP) inverts the channel linear and nonlinear effects using a technique similar to the conventional split-step Fourier method (SSFM) for optical fiber modeling. However, the DBP suffers from high complexity in implementations and has reduced effectiveness in the presence of Amplified Spontaneous Emission (ASE) noise in optical amplifiers. Parameters used in the DBP generally need to be manually adjusted to obtain the best performance. Other nonlinear compensation techniques include Regular Perturbation and Volterra series expansion. Nevertheless, performance and implementation complexity remain a challenge.
FEC coding can reduce the bit error rate (BER) in channels with impairments. Soft-input Low-Density Parity-Check (LDPC) codes have been used for high-rate optical communications. A 2-bit soft-input LDPC code achieves over 9 dB net coding gain with 20% overhead.
Turbo Equalization
Turbo Equalization (TEQ) was originally developed to deal with inter-symbol interference (ISI) in wireless channels, and is very effective, and can approach channel capacity with low-complexity implementations. A “turbo loop” is formed between a Maximum A posteriori (MAP) equalizer and a Soft-Input Soft-Output (SISO) decoder that exchange belief messages, termed extrinsic information. TEQ, for non-coherent fiber-optic nonlinear transmissions, uses a Bahl-Cocke-Jelinek-Raviv (BCJR) MAP equalizer with probability functions obtained from training sequences. Significant performance improvements have been obtained in simulations, but the complexity is too high to be realistically implemented in high-rate applications, such fiber-optic communications.
A reduced-complexity symbol detector uses a training sequence to generate mean levels at the receiver for each of the possible patterns of consecutive symbols. After training, each symbol is decoded by determining a minimum Euclidean distance of an L-symbol received sequence to each of the possible transmitted patterns. An increase in nonlinear tolerance of 2 dB can be obtained. Such system uses only the first-order statistics (mean values), and therefore offers limited performance improvement.
Embodiments of the invention provide a low-complexity receiver in a communications network. The receiver uses a “sliding window” equalizer. The sliding window equalizer estimates a likelihood of transmitted symbols based on a received symbol sequence using statistics of the optical channel. The equalizer can be symbol-spaced, or fractional-spaced. The statistics include but are not limited to the mean, the variance and covariance of the signals.
In one embodiment of the invention, the sliding window MAP equalizer is combined with a SISO FEC decoder to form a Turbo Equalization (TEQ) structure in the receiver. The MAP equalizer and the SISO decoder operate iteratively on the received symbol sequence for a number of iterations, or until a termination condition is met. The sliding window MAP equalizer enables much lower complexity implementations than the conventional BOR MAP equalizer.
In another embodiment of the invention, the sliding window equalizer is used in a Maximum-Likelihood Sequence Estimator (MLSE) using Viterbi decoding, or other sequence estimation procedures, such as Fano sequential decoding.
The sliding window equalizer can be combined with other pre-, and post-equalization schemes, such as a channel shortening linear equalizer or a DBP to further enhance the performance of the receiver.
Embodiments of invention also provide a method for establishing high-order statistics of the optical channel, and updating the high-order channel statistics periodically or continuously over successive symbols or over iterations of turbo loops. In addition, the statistics include first-order, second and higher orders, e.g., mean, covariance, skewness and kurtosis.
Coherent Fiber-Optic Network
A transmitter 150 transmits symbols sk via a nonlinear fiber optical channel 151 to a coherent receiver 160. After digitizing to, e.g., two samples per symbol, residual dispersion is removed using linear frequency-domain equalizer (FDE) 170. The oversampling signal is fed into a shift register 180 to obtain subsequences, and then a statistical Maximum-Likelihood Sequence Estimator (MLSE) equalizer 190 to determine an estimate of the transmitted signal. The MLSE detector 190 uses channel statistics 176, learned by a training processor 175. The channel statistics include high-order statistics, such as mean, covariance, skewness and kurtosis.
The invention is based on the realization that nonlinear distortion highly depends on patterns in the transmitted signal. Therefore, the statistical sequence equalizer first acquires such pattern dependent distortion characteristics by averaging the received sequence with training data, or an on-line learning process. The trained mean signals are then used to decode by searching for a minimum Euclidean distance from the received sequence. We use second and higher-order statistics (covariance) in addition to the first-order statistics (mean) to reduce residual nonlinear noise. In addition, we use fractionally-spaced processing with expanded window size to improve the performance through the use of the correlation over adjacent received samples.
Furthermore, we provide fractionally-spaced processing with a larger window to improve the performance by using correlation over adjacent received samples.
Sliding Window Estimator (SWE)
A sequence of discrete symbols r(n) 201 is received at an output of an optical channel or a pre-processing unit. The symbols are either symbol-spaced samples, or fractionally-spaced samples. The symbols arc fed to an N-length shift register 202, where N is a size of a sliding window, in teens of symbols. The shift register produces subsequences that are overlapping and time shifted
The SWE generates likelihood information of a transmitted symbol s(m) based on the received symbol sequence. The symbol position m can be any arbitrary posit-ion within the sequence. The position m is typically selected to be the middle symbol of the sequence, i.e., m=n−N/2+1.
At a likelihood calculator 203, the SWE estimates a likelihood Pr(R|S=Pj) of transmitted symbols S=s(n)s(n−1) . . . , s(n−N−1), given the received symbol sequence R=r(n), . . . , r(n−N+1), for all possible N-symbol patterns Pj. Note that there are MN patterns, where M is the number of modulation constellations, e.g., for QPSK, M=4.
Although the example is described for a single polarization system, it is understood that the invention can be extended to a dual-polarization multiplexed system. The dual-polarization application can include combined and individual use of the SWE, where the number of patterns for the combined case is M(2N), and that of the individual case is 2MN. Another embodiment uses individual SWE for x/y-polarizations. Decisions are fed into the combined SWE as a successive polarization nonlinearity canceller.
The number of patterns can be reduced by a kernel filter and clustering for higher-level modulations. The window size of the transmitted symbol sequence can differ from that of the received symbol sequence, specifically, S=s(n)s(n−1), . . . , s(n−Ns−1) and R=r(n), . . . , r(n−Nr+1) for window sizes Ns and Nr. Typically, the window size of the transmitted symbol sequence is no longer than an over-sampling factor multiplied by the window size of the transmitted symbol sequence. The window size can be adaptively optimized by tracking an effective memory of the channel.
The SWE uses channel statistics of the channel to estimate the likelihood. The channel statistics, including the pattern-dependent covariance, are obtained by a channel statistics analyzer at the receiver, see
Given the channel statistics for the jth pattern such as the first order mean μj and the second order covariance Σj and its inverse Σ−1j, a likelihood Pr(R|S=Pj) 210 is estimated as
where Tis a transpose operator, and σj are covariances.
The SWE can be used as a standalone hard-output Maximum-Likelihood (ML) symbol detector. In such a case, the switch 205 is connected to the block 206, which searches for the most likely estimate ŝ(m) with the maximum likelihood value as
The SWE can also generate the soft-output likelihood of the symbol L(s(m)) when switching to block 207. This soft information can be used as the input to the following blocks that accepts soft input e.g., for SISO FEC decoder. Typically, the soft-output detector provides better performance than the hard-output detector.
Channel Statistics Analyzer
For a given pattern s, the mean is
The covariance matrix is
is
where is the number of received sequences corresponding to transmitted pattern s, with represent element-wise real and imaginary parts, respectively (·) and ℑ(·) represent element-wise real and imaginary parts, respectively. The mean and the covariance matrix, as well as its inverse version, can be updated 305 sequentially with low-complexity processing. In
In one embodiment, the channel statistics analyzer determines the statistics using a training sequence. If the channel is stationary during operation, then the statistics remain unchanged.
In another embodiment, a receiver periodically receives training sequences and subsequently activates the channel statistics analyzer to update 305 the channel statistics. Thus, the channel statistics are adjusted for time-variation of channel characteristics.
Turbo Equalization Receiver
As shown in
In the turbo equalization receiver, the SW MAP estimator 600 is connected to an SISO decoder. The SWE MAP estimator outputs the log-likelihood ratio of symbols L(ŝ(m)). A de-interleaver (Π′) 402 decorrelates the symbols in the sequence and produces likelihood L(Ĉ(m)) corresponding to the de-interleaved sequence Ĉ=Π−1(Ŝ). The SISO decoder decodes L(Ĉ(m)) and either outputs the hard-decoded symbol sequence {circumflex over (D)} 401, or soft-output data sequence likelihood L({circumflex over (D)}), often called extrinsic information. Then, L({circumflex over (D)}) is re-encoded 403 and re-interleaved (Π) 404 to generate the a priori probability of the transmitted sequence, denoted as L−1(Ŝ), which is fed back to the MAP estimator 600.
Iterative TEQ Receiver
As shown in
The output of the SWE MAP estimator is fed into the SISO decoder and the output of the SISO decoder is fed back into the SWE MAP estimator. This iterative process continues until the number of iteration K reaches a pre-defined threshold, or a termination condition is met. The iterative TEQ receiver can be implemented in a pipeline manner as shown in
In one embodiment, the TEQ receiver can include any pre-processing unit, such as prior-art DBP, or a frequency-domain chromatic dispersion equalizer. This can yield a substantial performance gain, while the window size can be relatively small for low-complexity implementation. Similarly, the transmitter can use any pre-compensation techniques including pre-distortion, pre-coding, pre-DBP, pre-equalizer for performance enhancement. In one embodiment, the TEQ receiver can be simplified to a MLSE equalizer for hard-decision decoder, instead of using the SISO decoder.
Sliding Window MAP Estimator
The sliding window MAP estimator 600 is shown in
The a priori probability of the sequence is determined based on the a priori likelihood of the transmitted sequence, which is re-encoded from the soft output 601 of the SISO decoder 407. The a priori probability of each encoded symbol. L−(ŝ) is determined 602 first. The a priori probability of the encoded symbol sequence L−(ŝn . . . ŝn−L+1) 610 is then computed 603, and the a priori probability of all possible transmitted sequence 610 is computed and stored in a table 204.
The a priori probability 204 is then combined with the probability of the received sequence 203 to produced a posteriori probability of the received sequence, which is then used to compute the a posteriori probability of the received symbol in the sequence L(ŝ) 605. The a posteriori probability of the bits is then determined 604.
In particular, for all possible sequences, the a priori likelihood 610
L
−1(s(n)s(n−1) . . . s(n−L+1)=i) (5)
is derived directly from the likelihood of the bit, or symbol sequence re-encoded from the data sequence likelihood L({circumflex over (D)}).
The a posteriori likelihood is therefore determined 604 as
L(s|{circumflex over (r)})=L(r|s)+L−(i)+c, (6)
where c is a constant, and does not need to be determined.
For the MAP symbol hard-decision detector, the estimated transmitted symbol is
Similarly, the soft likelihood L(S=i|R) of the symbol S(m), can be calculated based on the a posteriori probability of the sequence.
Sliding Window MAP Estimator
For each subsequence 701, determine 710 the likelihood for each pattern based on the channel statistics. Then, determine 702 the log Likelihood ratios (LLR) of the bits based on the a priori likelihood L−711 (produced by the decoder) to produce L(s) 704 for the decoder.
Sliding Window ML Estimator
As shown in
Although the invention has been described for an example optical network with single polarization, the embodiments can also be used for optical networks with polarization multiplexed signal, of for other wired and wireless communication systems.
Embodiments of the invention provide a fractionally-spaced equalizer second and higher-order statistics obtained by training to deal with nonlinear impairment in coherent optical communications. The equalizer improves the Q-factor by more than 2 dB for long-haul transmissions of 5,230 km.
The statistical sequence equalizer maintains 2 dBQ improvement even at 10,460 km for low dispersion case, whereas the improvement is considerably reduced for high dispersion case. It indicates that an equalizer with a small number of taps can work with other channel shortening methods for long-haul transmissions.
Using the likelihood described above, the statistical equalizer uses the maximum-likelihood sequence estimation (MLSE) to detect the transmitted symbols with a Viterbi algorithm. Because the computational complexity of MLSE grows exponentially with the channel memory, more specifically O[N4M], we can use a channel shortening equalizer including frequency-domain chromatic dispersion compensation or reduced-complexity DBP. We obtain higher than 2 dBQ with a short memory MLSE using just M=3 taps, that can outperform the DBP.
Another embodiment uses a low-complexity turbo equalizer with a sliding window (SW) MAP estimator, and a low overhead, small block-size SISO LDPC decoder. The ML estimator alone provides a 2.5˜4 dB gain in Q-factor over existing sliding window detector, and the turbo equalizer provides an additional ˜1 dB improvement in a nonlinear fiber channel over 5,000 km.
The equalizer outperforms conventional Digital Back-Propagation (DBP), which uses hundreds of SSF.M iterations, in low dispersion channels. Even for high-dispersion channels, the fractionally-spaced 3-tap equalizer achieved comparable performance in peak Q factor to the DBP.
The SW-ML detector out-performs the conventional SW-Minimum Distance (MD) detector by as much as 5 dB for the low dispersion channel, and 2-3.5 dB for a high dispersion channel. For an equalized linear channel, where symbols are considered independent and have equal variance, the SW-ML and SW-MD detectors have identical performance. This confirms that using the 2nd order statistics provides performance gain in non-linear channels. A performance improvement of 4 dB or higher can be achieved in 40 G bps non-return-to-zero (NRZ) quadrature phase-shift keying (QPSK) transmissions.
The turbo equalizer structure uses the SW-MAP estimator and the LDPC decoder with a short block size. The SW-MAP estimator utilizes multi-symbol sequence and second-order statistics to produce reliable likelihood information for a following SISO LDPC decoder. The complexity of the turbo equalizer is sufficiently low, and can be implemented in hardware. There is a significant BER performance and Q-factor improvement over prior art techniques.
The sliding window equalizer and receivers based on the SWEQ are effective in mitigating non-linear effect of the fiber channel.
For SWE-based TEQ receiver, we analyze the QPSK performance with a window size of L=3 symbols in a low local dispersion channel (1551.32 nm wavelength) and a high local dispersion channel (1561.01 nm wavelength).
Over the entire range of launch power simulated, the SW-ML detector out-performs the conventional Minimum Distance detector by as much as 5 dB for the 1551 nm channel and 2 to 3.5dB for the 1561 nm channel.
For an equalized linear channel, where symbols are considered independent and have equal variance, the SW-ML and SW-MD detectors have identical performance. This further confirms our analysis that using the 2nd order statistics provides performance gain in non-linear channels. For example, in high dispersion channel with 3.25 dBm launch power, the Q at the SW-MAP estimator is 6.92 dB, the decoder output is 10.02 dB after the initial iteration and improves to 10.31 dB and 10.93 dB after the 1st and 2nd iteration. The overall gain is greater than 7 dB at 0.5 dBm launch power for the low dispersion channel and 6.5 dB at 3.25 dBm launch power for the high dispersion channel.
Although the invention has been described by way of examples of preferred embodiments, it is to be understood that various other adaptations and modifications can be made within the spirit and scope of the invention. Therefore, it is the object of the appended claims to cover all such variations and modifications as come within the true spirit and scope of the invention.