Coherent detection in optical fiber systems

: The drive for higher performance in optical fiber systems has renewed interest in coherent detection. We review detection methods, including noncoherent, differentially coherent, and coherent detection, as well as a hybrid method. We compare modulation methods encoding information in various degrees of freedom (DOF). Polarization-multiplexed quadrature-amplitude modulation maximizes spectral efficiency and power efficiency, by utilizing all four available DOF, the two field quadratures in the two polarizations. Dual-polarization homodyne or heterodyne downconversion are linear processes that can fully recover the received signal field in these four DOF. When downconverted signals are sampled at the Nyquist rate, compensation of transmission impairments can be performed using digital signal processing (DSP). Linear impairments, including chromatic dispersion and polarization-mode dispersion, can be compensated quasi-exactly using finite impulse response filters. Some nonlinear impairments, such as intra-channel four-wave mixing and nonlinear phase noise, can be compensated partially. Carrier phase recovery can be performed using feedforward methods, even when phase-locked loops may fail due to delay constraints. DSP-based compensation enables a receiver to adapt to time-varying impairments, and facilitates use of advanced forward-error-correction codes. We discuss both single- and multi-carrier system implementations. For a given modulation format, using coherent detection, they offer fundamentally the same spectral efficiency and power efficiency, but may differ in practice, because of different impairments and implementation details. With anticipated advances in analog-to-digital converters and integrated circuit technology, DSP-based coherent receivers at bit rates up to 100 Gbit/s should become practical within the next few years. Abstract: We correct a table showing the signal-to-noise ratio obtained in coherent detection. We also clarify a table showing the bandwidth requirements for homodyne and heterodyne downconverters. Otherwise, our previous results and conclusions are unchanged.


Introduction
An important goal of a long-haul optical fiber system is to transmit the highest data throughput over the longest distance without signal regeneration. Given constraints on the bandwidth imposed by optical amplifiers and ultimately by the fiber itself, it is important to maximize spectral efficiency, measured in bit/s/Hz. But given constraints on signal power imposed by fiber nonlinearity, it is also important to maximize power (or SNR) efficiency, i.e., to minimize the required average transmitted energy per bit (or the required signal-tonoise ratio per bit). Most current systems use binary modulation formats, such as on-off keying or differential phase-shift keying, which encode one bit per symbol. Given practical constraints on filters for dense wavelength-division multiplexing (DWDM), these are able to achieve spectral efficiencies of 0.8 bit/s/Hz per polarization. Spectral efficiency limits for various detection and modulation methods have been studied in the linear [1−3] and nonlinear regimes [4,5]. Noncoherent detection and differentially coherent detection offer good power efficiency only at low spectral efficiency, because they limit the degrees of freedom available for encoding of information [5].
The most promising detection technique for achieving high spectral efficiency while maximizing power (or SNR) efficiency, is coherent detection with polarization multiplexing, as symbol decisions are made using the in-phase (I) and quadrature (Q) signals in the two field polarizations, allowing information to be encoded in all the available degrees of freedom. When the outputs of an optoelectronic downconverter are sampled at Nyquist rate, the digitized waveform retains full information of the electric field, which enables compensation of transmission impairments by digital signal processing (DSP). A DSP-based receiver is highly advantageous because adaptive algorithms can be used to compensate time-varying transmission impairments. Advanced forward error-correction coding can also be implemented. Moreover, digitized signals can be delayed, split and amplified without degradation in signal quality. DSP-based receivers are ubiquitous in wireless and digital subscriber line (DSL) systems at lower data rates. In such systems, computationally intensive techniques have been demonstrated, such as orthogonal frequency-division multiplexing (OFDM) with multiple-input-multiple-output (MIMO) transmission in a real-time 1 Gbit/s wireless link [6]. Continued hardware improvements will enable deployment of DSP-based coherent optical systems in the next few years.
Experimental results in coherent optical communication have been promising. Kikuchi demonstrated polarization-multiplexed 4-ary quadrature-amplitude modulation (4-QAM) transmission at 40 Gbit/s with a channel bandwidth of 16 GHz (2.5 bit/s/Hz) [7]. This experiment used a high-speed sampling oscilloscope to record the output of a homodyne downconverter. DSP was performed offline because of the unavailability of sufficiently fast processing hardware. The first demonstration of real-time coherent detection occurred in 2006, when an 800 Mbit/s 4-QAM signal was coherently detected using a receiver with 5-bit analog-to-digital converters (ADC) followed by a field programmable gate array [8]. This paper provides an overview of detection and modulation methods, with emphasis on coherent detection and digital compensation of channel impairments. The paper is organized as follows: in Section 3, we review signal detection methods, including noncoherent, differentially coherent and coherent detection. We compare these techniques and the modulation formats they enable. In Section 4, we compare the bit-error ratio (BER) performance of different modulation formats in the presence of additive white Gaussian noise (AWGN). In Section 5, we review the major channel impairments in long-haul optical transmission. We show how these can be compensated digitally in single-carrier systems, and we compare digital compensation to traditional compensation methods. In Section 6, we review OFDM, which is a multi-carrier modulation format. Finally, in Section 7, we compare implementation complexities of single-and multi-carrier systems.

Notation
In this paper, we represent optical signal and noise electric fields as complex-valued, baseband quantities, which are denoted as ( ) t E subscript . In Section 3, photocurrents are represented as real passband signals as ( ) t I subscript . After we derive the canonical model for a coherent receiver, in all subsequent sections, we write all signals and noises as complexvalued baseband quantities, with the convention that the output of the optoelectronic downconverter is denoted by y, the transmitted symbol as x, the output of the linear equalizer as x , and the phase de-rotated output prior to symbol decisions as x . Signals that occupy one polarization only are represented as scalars and are written in italics as shown; while dually polarized signals are denoted as vectors and are written in bold face, such as k y . Unless stated otherwise, we employ a linear (x, y) basis for decomposing dually polarized signals. In Section 5 dealing with impairment compensation, matrix and vector quantities are denoted in bold face. In noncoherent detection, a receiver computes decision variables based on a measurement of signal energy. An example of noncoherent detection is direct detection of on-off-keying (OOK) using a simple photodiode ( Fig. 1(a)). To encode more than one bit per symbol, multilevel amplitude-shift keying (ASK) -also known as pulse-amplitude modulation -can be used. Another example of noncoherent detection is frequency-shift keying (FSK) with wide frequency separation between the carriers. Fig. 1(b) shows a noncoherent receiver for binary FSK.

Noncoherent detection
The limitations of noncoherent detection are: (a) detection based on energy measurement allows signals to encode only one degree of freedom (DOF) per polarization per carrier, reducing spectral efficiency and power efficiency, (b) the loss of phase information during detection is an irreversible transformation that prevents full equalization of linear channel impairments like CD and PMD by linear filters. Although maximum-likelihood sequence detection (MLSD) can be used to find the best estimate of the transmitted sequence given only a sequence of received intensities, the achievable performance is suboptimal compared with optical or electrical equalization making use of the full electric field [13]. In differentially coherent detection, a receiver computes decision variables based on a measurement of differential phase between the symbol of interest and one or more reference symbol(s). In differential phase-shift keying (DPSK), the phase reference is provided by the previous symbol; in polarization-shift keying (PolSK), the phase reference is provided by the signal in the adjacent polarization. A binary DPSK receiver is shown in Fig. 2(a). Its output photocurrent is:

Differentially coherent detection
is the received signal, R is the responsivity of each photodiode, and T s is the symbol period. This receiver can also be used to detect continuous-phase frequency-shift keying (CPFSK), as the delay interferometer is equivalent to a delay-and-multiply demodulator. A receiver for M-ary DPSK, M > 2, can similarly be constructed as shown in Fig. 2(b). Its output photocurrents are:

Re
, and

Im .
A key motivation for employing differentially coherent detection is that binary DPSK has 2.8 dB higher sensitivity than noncoherent OOK at a BER of 10 −9 [5]. However, the constraint that signal points can only differ in phase allows only one DOF per polarization per carrier, the same as noncoherent detection. As the photocurrents in Eq. (1) to (3) are not linear functions of the E-field, linear impairments, such as CD and PMD, also cannot be compensated fully in the electrical domain after photodetection.
A more advanced detector for M-ary DPSK is the multichip DPSK receiver, which has multiple DPSK receivers arranged in parallel, each with a different interferometer delay that is an integer multiple of T s [14,15]. Since a multichip receiver compares the phase of the current symbol to a multiplicity of previous symbols, the extra information available to the detector enables higher sensitivity. In the limit that the number of parallel DPSK receivers is infinite, the performance of multi-chip DPSK approaches coherent PSK [15]. In practice, the number of parallel DPSK receivers required for good performance needs to be equal to the impulse duration of the channel divided by T s . Although multi-chip DPSK does not require a local oscillator (LO) laser, carrier synchronization and polarization control, the hardware complexity can be a significant disadvantage.

Hybrid of noncoherent and differentially coherent detection
A hybrid of noncoherent and differentially coherent detection can be used to recover information from both amplitude and differential phase. One such format is polarization-shift keying (PolSK), which encodes information in the Stokes parameter. If we let be the E-fields in the two polarizations, the Stokes parameters are is free of phase noise. In practice, the phase noise immunity of PolSK is limited by the bandwidth of the photodetectors [16]. It has been shown that 8-PolSK can tolerate laser linewidths as large as 01 , which is about 100 times greater than the phase noise tolerance of coherent 8-QAM (Section 5.3.3). This was a significant advantage in the early 1990s, when symbol rates were only in the low GHz range. In modern systems, symbol rates of tens of GHz, in conjunction with tunable laser having linewidths <100 kHz, has diminished the advantages of PolSK. Recent results have shown that feedforward carrier synchronization enables coherent detection of 16-QAM at b T ν Δ ~10 −5 [18], which is within the limits of current technology. As systems are increasingly driven by the need for high spectral efficiency, polarizationmultiplexed QAM is likely to be more attractive because of its higher sensitivity (Section 4). (2)

Coherent detection
The most advanced detection method is coherent detection, where the receiver computes decision variables based on the recovery of the full electric field, which contains both amplitude and phase information. Coherent detection thus allows the greatest flexibility in modulation formats, as information can be encoded in amplitude and phase, or alternatively in both in-phase (I) and quadrature (Q) components of a carrier. Coherent detection requires the receiver to have knowledge of the carrier phase, as the received signal is demodulated by a LO that serves as an absolute phase reference. Traditionally, carrier synchronization has been performed by a phase-locked loop (PLL). Optical systems can use (i) an optical PLL (OPLL) that synchronizes the frequency and phase of the LO laser with the TX laser, or (ii) an electrical PLL where downconversion using a free-running LO laser is followed by a secondstage demodulation by an analog or digital electrical VCO whose frequency and phase are synchronized. Use of an electrical PLL can be advantageous in duplex systems, as the transceiver may use one laser as both TX and LO. PLLs are sensitive to propagation delay in the feedback path, and the delay requirement can be difficult to satisfy (Section 5.3.1). Feedforward (FF) carrier synchronization overcomes this problem. In addition, as a FF synchronizer uses both past and future symbols to estimate the carrier phase, it can achieve better performance than a PLL which, as a feedback system, can only employ past symbols. Recently, DSP has enabled polarization alignment and carrier synchronization to be performed in software.  Fig. 4. At the transmitter, Mach-Zehnder (MZ) modulators encode data symbols onto an optical carrier and perform pulse shaping. If polarization multiplexing is used, the TX laser output is split into Channel Receiver two orthogonal polarization components, which are modulated separately and combined in a polarization beam splitter (PBS). We can write the transmitted signal as: where s T is the symbol period, t P is the average transmitted power, ( ) t b is the pulse shape (e.g., non-return-to-zero (NRZ) or return-to-zero (RZ)) with the normalization are the frequency and phase noise of the TX laser, and is a 2×1 complex vector representing the k-th transmitted symbol. We assume that symbols have normalized energy: For single-polarization transmission, we can set the unused polarization component k x , 2 to zero. The channel consists of N A spans of fiber, with inline amplification and DCF after each span. In the absence of nonlinear effects, we can model the channel as a 2×2 matrix: Under the assumption of Fig. 4 where inline amplification completely compensates propagation loss, t r P P = is the average received power, is a normalized pulse shape, and is ASE noise in the l-th polarization. Assuming the N A fiber spans are identical and all inline amplifiers have gain G and spontaneous emission factor n sp , the two-sided power spectral density (psd) of The first stage of a coherent receiver is a dual-polarization optoelectronic downconverter that recovers the baseband modulated signal. In a digital implementation, the analog outputs are lowpass filtered and sampled at is a rational oversampling ratio. Channel impairments can then be compensated digitally before symbol detection. We first consider a single-polarization downconverter, where the LO laser is aligned in the lth polarization. Downconversion from optical passband to electrical baseband can be achieved in two ways: in a homodyne receiver, the frequency of the LO laser is tuned to that of the TX laser so the photoreceiver output is at baseband. In a heterodyne receiver, the LO and TX lasers differ by an intermediate frequency (IF), and an electrical LO is used to downconvert the IF signal to baseband. Both implementations are shown in Fig. 5. Although we show the optical hybrids as 3-dB fiber couplers, the same networks can be implemented in free-space optics using 50/50 beamsplitters; this was the approach taken by Tsukamoto [21]. Since a beamsplitter has the same transfer function as a fiber coupler, there is no difference in their performances.

Single-polarization downconverter
In the heterodyne downconverter of Fig. 5(a), the optical frequency bands around both map to the same IF. In order to avoid DWDM crosstalk and to avoid excess ASE from the unwanted image band, optical filtering is required before the downconverter. The output current of the balanced photodetector in Fig. 5(a) is: is the E-field of the LO laser, and , , , is the carrier phase, and ( )  It can similarly be shown that the currents at the outputs of the balanced photodetectors in the homodyne downconverter ( Fig. 5(b)) are: , , 2 2 2 1 , , I   lq  sh  lq  sp  lq  r  l  LO  q  l  hom  ,  '  ,  ,   2  4   2  3 , ,  10) and (11). In long-haul systems, the psd of LO-spontaneous beat noise is typically much larger than that of LO shot noise; such systems are thus ASE-limited. Conversely, unamplified systems do not have ASE, and are therefore LO shot-noise-limited.
If one were to demodulate Eq. . Its output photocurrent is . Compared to Eq. (11), the signal term is doubled (four times the power), while the shot noise power is only increased by two, thus yielding a sensitivity improvement of 3 dB compared to heterodyne or two-quadrature homodyne downconversion. This case is not of practical interest in this paper, however, as long haul systems are ASE-limited, not LO shot-noise-limited. Also, for good spectral and power efficiencies, modulation formats that encode information in both I and Q are preferred. Hence, there is no performance difference between a homodyne and a heterodyne downconverter provided optical filtering is used to reject image-band ASE for the heterodyne downconverter.
Since the two downconverters in Fig. 5 ultimately recover the same baseband signals, we can combine Eqs. (10) and (11) and write a normalized, canonical equation for their outputs as:  Table 1. For the shotnoise limited regime using a heterodyne or two-quadrature homodyne downconverter, s γ is simply the number of detected photons per symbol. We note that Eq. (12) is complex-valued, and its real and imaginary parts are the two baseband signals recovered in Fig. 5. For the remainder of this paper, it is understood that all complex arithmetic operations are ultimately implemented using these real and imaginary signals. The advantages of heterodyne downconversion are that it uses only one balanced photodetector and has a simpler optical hybrid. However, the photocurrent in Eq. (8) has a bandwidth of , where BW is the signal bandwidth ( Fig. 6(a)). To avoid signal distortion caused by overlapping side lobes, ω IF needs to be sufficiently large. Typically, , thus a heterodyne downconverter requires a balanced photodetector with at least twice the bandwidth of a homodyne downconverter, whose output photocurrents in Eqs. (10) and (11) only have bandwidths of BW ( Fig. 6(b)). This extra bandwidth requirement is a major disadvantage. In addition, it is also difficult to obtain electrical mixers with baseband bandwidth as large as the IF. A summary of homodyne and heterodyne receivers is given in Table 2. A comparison of carrier synchronization options is given in Table 3. Table 1. SNR per symbol for various receiver configurations. For the ASE-limited cases, s n is the average number of photons received per symbol, N A is the number of fiber spans, and n sp is the spontaneous emission noise factor of the inline amplifiers. For the LO shot-noise-limited cases,

Dual-polarization downconverter
In the single-polarization downconverter, the LO needed to be in the same polarization as the received signal. One way to align the LO polarization with the received signal is with a polarization controller (PC). There are several drawbacks with this: first, the received polarization can be time-varying due to random birefringence, so polarization tracking is required. Secondly, PMD causes the received Stokes vector to be frequency-dependent. When BW ω a single-polarization receiver is used, frequency-selective fading occurs unless PMD is first compensated in the optical domain. Thirdly, a single-polarization receiver precludes the use of polarization multiplexing to double the spectral efficiency. A dual-polarization downconverter is shown inside the receiver of Fig. 4(a). The LO laser is polarized at 45° relative to the PBS, and the received signal is separately demodulated by each LO component using two single-polarization downconverters in parallel, each of which can be heterodyne or homodyne. The four outputs are the I and Q of the two polarizations, which has the full information of ( ) . CD and PMD are linear distortions that can be compensated quasi-exactly by a linear filter. The signals E x (t) and E y (t) are the complex-valued analog outputs described by Eq. (12) for each polarization. We note that in the case of the heterodyne downconverter, the nonlinear operations shown can be performed at the IF output(s) of the balanced photoreceiver.
The lossless transformation from the optical to the electrical domain also allows the receiver to emulate noncoherent and differentially coherent detection by nonlinear signal processing in the electrical domain (Fig. 7). In long-haul transmission where ASE is the dominant noise source, these receivers have the same performance as those in Fig. 1−3. A summary of the detection methods discussed in this section is shown in Table 4.

Modulation formats
In this section, we compare the BER performance of different modulation methods for singlecarrier transmission corrupted by AWGN. We assume that all channel impairments other than AWGN -including CD, PMD, laser phase noise and nonlinear phase noise -have been compensated using techniques discussed in Section 5. Since ASE and LO shot noise are Gaussian, the performance equations obtained are valid for both long haul and back-to-back systems, when the definition of SNR defined by Table 1 is used. Owing to fiber nonlinearity, it is desirable to use modulation formats that maximize power efficiency. Unless otherwise stated (PolSK being the only exception), the formulae provided assume transmission in one polarization, where noise in the unused polarization has been filtered. This condition is naturally satisfied when a homodyne or heterodyne downconverter is used. For noncoherent detection, differentially coherent detection and hybrid detection, the received optical signal needs to be passed through a polarization controller followed by a linear polarizer.
Since the two polarizations in fiber are orthogonal channels with statistically independent noises, there is no loss in performance by modulating and detecting them separately. The BER formulae provided are thus valid for polarization-multiplexed transmission provided there is no polarization crosstalk. Polarization multiplexing doubles the capacity while maintaining the same receiver sensitivity in SNR per bit. We write the received signal as: where k x is the transmitted symbol and k n is AWGN. For the remainder of this paper, our notation shall be as follows: M is the number of signal points in the constellation.
is the SNR per symbol in single-polarization transmission.
is the SNR per symbol in dual-polarization transmission (e.g. polarization-multiplexed or PolSK).
The maximum achievable spectral efficiency (bit/s/Hz) of a linear AWGN channel is governed by the Shannon capacity [1]: If N D identical channels are available for transmission, and we utilize them all by dividing the available power equally amongst the channels, the total capacity is , which is an increasing function of N D . Hence, the best transmission strategy is to use all the dimensions available. For example, suppose a target spectral efficiency of 4 bits per symbol is needed. Polarization-multiplexed 4-QAM, which uses the inphase and quadrature components of both polarizations, has better sensitivity than single-polarization 16-QAM.

ASK with noncoherent detection
Optical M-ary ASK with noncoherent detection has signal points evenly spaced in nonnegative amplitude [25]. The photocurrents for different signal levels thus form a (15)   (14) quadratic series. It can be shown that the optimal decision thresholds are approximately the geometric means of the intensities of neighboring symbols. Assuming the use of Gray coding, it can be shown that for large M and γ b , the BER is approximated by [5]:

DPSK with differentially coherent detection
Assuming the use of Gray coding, the BER for M-ary DPSK employing differentially coherent detection is [26]: For binary DPSK, the above formula is exact, and can be simplified to [27]: For quaternary DPSK, we have [27]: function and the modified Bessel function of the zeroth order, respectively.

Polarization-Shift Keying (PolSK)
PolSK is the special case in this section where the transmitted signal naturally occupies both polarizations. Thus, polarization multiplexing cannot be employed to double system capacity.
The BER for binary PolSK is [16]: For higher-order PolSK, the BER is well-approximated by [28], n , 0 θ and 1 θ are related to the number of nearest neighbors and the shape of the decision boundaries on the Poincaré sphere.

PSK with coherent detection
Assuming the use of Gray coding, the BER for M-ary PSK employing coherent detection is given approximately by [29]: For the special cases of BPSK and QPSK, we have the exact expressions:

QAM with coherent detection
Assuming the use of Gray coding, the BER for a square M-QAM constellation with coherent detection is approximated by [29]: The BER performance of 8-QAM with the signals points arranged as shown in Fig. 8   Using Eq. (16) to (27), we compute the SNR per bit required for each modulation format discussed to achieve a target BER of 10 −3 , which is a typical threshold for receivers employing forward error-correction coding (FEC). The results are shown in Table 6. In Fig. 9, we plot spectral efficiency vs SNR per bit with polarization multiplexing to obtain fair comparison with PolSK (we also show results for 12-PolSK and 20-PolSK Bit encoding uses all four available DOF for encoding information, it has better SNR efficiency than the other formats, and exhibits a steeper slope at high spectral efficiency.  9. Spectral efficiency vs. SNR per bit required for different modulation formats at a target BER of 10 −3 . We assume polarization multiplexing for all schemes except PolSK. Also shown is the Shannon limit (15), which corresponds to zero BER.

Channel impairments and compensation techniques in single-carrier systems
In this section, we review the major channel impairments in fiber-optic transmission. We present the traditional methods of combating these, and show how compensation can be done electronically with coherent detection in single-carrier systems. Impairment compensation in multi-carrier systems is discussed in Section 6. Gbit/s or higher [30]. In reconfigurable networks, data can be routed dynamically through different fibers, so the residual dispersion can be time-varying. This necessitates tunable dispersion compensators. Polarization-mode dispersion (PMD) is caused by random birefringence in the fiber. In firstorder PMD, a fiber possesses a "fast axis" along in polarization and a "slow axis" in the orthogonal polarization (Fig. 10). These states of polarization are known as the principal states of polarization (PSPs), and can be any vector in Stokes space in general. First-order PMD can be written as [31]:

Polarization-mode dispersion
is a diagonal matrix with DGD τ being the differential group delay between the two PSPs, and 1 R and 2 R are unitary matrices that rotate the reference polarizations to the fiber's PSPs, which are elliptical in general. When a signal is launched in any polarization state other than a PSP, the receiver will detect two pulses at each reference polarization. Ignoring CD and other effects, the impulse response measured by a polarization-insensitive direct-detection receiver is , where a 2 is the proportion of transmitted energy falling in the slow PSP. In this simple two-path model, we see that PMD can lead to frequency-selective fading [32]. In contrast to CD, which is relatively static, PMD (both the PSPs and the DGD) can fluctuate on time scales on the order of a millisecond [33]. Thus PMD compensators thus need to be rapidly adjustable. The statistical properties of PMD have been studied in [34−36], and it has been shown that DGD τ has a Maxwellian distribution, whose mean value DGD τ grows as the square-root of fiber length.
depends on the degree of tunability, and increasing the number degrees of freedom can require costly hardware. However, optical PMD compensators are transparent to the data rate and modulation format of the transmitted signal, and have been successfully employed for very high-data-rate systems, where digital compensation is currently impossible.
Electronic PMD equalization has gained considerable recent interest. Buchali and Bülow studied the use of a feedforward equalizer (FFE) with decision feedback equalizer (DFE) to combat PMD systems using direct detection of OOK [39]. As with electronic CD compensation in direct detection of OOK [43], the loss of phase during detection prevents quasi-exact compensation of PMD.

Other linear impairments
In addition to CD and PMD, a fiber optic link can also have polarization-dependent loss (PDL) due to anisotropy of network components such as couplers, isolators, filters, multiplexers, and amplifiers [44]. In DWDM transmission, arrayed waveguide gratings (AWG), interleavers and reconfigurable add-drop (de)multiplexers (ROADMs) cause attenuation at the band edges of a channel. When a signal has to pass through cascaded elements, bandwidth narrowing can be problematic. This is a major challenge in 40 Gbit/s RZ-DPSK at 50 GHz channel spacing [45]. Bandwidth narrowing can be equalized by tunable optical equalizers, but such devices are costly, introduce loss, and are difficult to make adaptive.

Compensation of linear impairments and computational complexity
Since a dual-polarization downconverter linearly recovers the full electric field, CD and PMD can be compensated quasi-exactly in the electronic domain after photodetection. One approach is to use a tunable analog filter. However, as in the case of optical compensators, it is difficult to implement the desired transfer function exactly, and analog filters are also difficult to make adaptive. In addition, parasitic effects like signal reflections can lead to signal degradation.
With improvements in DSP technology, digital equalization of CD and PMD is becoming feasible. When the outputs of a dual-polarization downconverter are sampled above the Nyquist rate, the digitized signal contains a full characterization of the received E-field, allowing compensation of linear distortions by a linear filter. CD compensation using a digital infinite impulse response (IIR) filter was studied by [46]. Although an IIR filter allows fewer taps, it is more difficult to analyze, and may require greater receiver complexity because of the need to implement time-reversal filters. In this paper, we concentrate on CD and PMD compensation using a finite impulse response (FIR) filter. Linear equalization using an FIR filter for dually polarized coherent systems was studied in [47,48]. The canonical system model is shown in Fig. 11. The analog outputs of a dualpolarization downconverter are passed through anti-aliasing filters with impulse responses ( ) t p and then sampled synchronously at a rate of is a rational oversampling ratio. We assume that the sampling clock has been synchronized using wellknown techniques [49]. In theory, the use of a matched filter in conjunction with symbol rate Suppose a linear equalizer takes the 1 2 + = L N samples closest to symbol k to computes the minimum-mean-square error (MMSE) estimate of the k-th symbol k x . We have: , and the anti-aliasing filter does not introduce amplitude distortion, the power penalty of a compensated CD/PMD channel approaches zero with respect to a pure AWGN channel at the same SNR. In practice, since H can be time-varying, an adaptive equalizer shown in Fig.  12 is desired. Owing to the quadratic nature of Eq. (33), we can use well-known algorithms such as least mean square (LMS) or recursive least squares (RLS) [53]. For LMS, we have the following update equation for the filter coefficients: is the equalizer coefficients use to compute k x , and μ is the step size parameter that needs to satisfy . When a non-integer oversampling ratio is used, K filters are required for all possible values of K kM mod . The index m indicates the m-th use of that particular filter.  , the required value of N is insensitive to whether NRZ or RZ pulses are used, that system performance is insensitive to sampling time errors. Typically, the required value of N is dominated by CD considerations. In Table 7 . We compare the complexity of single-carrier versus multi-carrier transmission (using OFDM) in Section 7.

Fiber nonlinearity
The dominant nonlinear impairments in fiber arise from the Kerr nonlinearity, which causes a refractive index change proportional to signal intensity. Signal propagation in the presence of fiber attenuation, CD and Kerr nonlinearity is described by the nonlinear Schrödinger equation (NLSE) α is the attenuation coefficient, β 2 is the dispersion parameter, γ is the nonlinear coefficient, and z and t are the propagation direction and time, respectively.
Nonlinear effects include deterministic and statistical components. The nonlinearity experienced by a signal due to its own intensity called self-phase modulation (SPM). In WDM systems, a signal also suffers nonlinear effects due to the fields of neighboring channels. These are cross-phase modulation (XPM) and four-wave-mixing (FWM) [22], and their impact can be reduced by allowing non-zero local dispersion. In the absence of ASE, and given knowledge of the transmitted data, all these nonlinear effects are deterministic, and it is possible, in principle, to pre-compensate them at the transmitter. At the receiver, it would be also be possible to employ joint multi-channel detection techniques, though complexity precludes their implementation at this time. Here, we consider only receiver-based compensation of SPM-induced impairments in the presence of CD, which leads to intrachannel nonlinear effects.
In long-haul systems, interaction between ASE noise and signal through the Kerr nonlinearity leads to nonlinear phase noise (NLPN). When caused by the ASE and signal in the channel of interest, this is called SPM-induced NLPN. When caused by the ASE and signal in neighboring channels, it is called XPM-induced NLPN. Here, we consider only receiver-based compensation of SPM-induced NLPN.
In following two sections, we discuss (i) SPM in the presence of CD, and (ii) SPMinduced NLPN. We show how these can be compensated by exploiting their correlation properties.

Self-phase modulation with chromatic dispersion: intra-channel nonlinearity
We consider a noiseless transmitted signal is the linear solution to the NLSE (obtained by setting the right hand side of (37) to zero) and ΔE is the perturbation due to Kerr nonlinearity. The NLSE can be re-written as: For typical terrestrial links, the accumulated dispersion is such that the only the terms on the right hand side of (38) that will induce a sizeable effect on the pulse at symbol k are those with indices satisfying p m l k − + = . Without loss of generality, we focus on symbol k = 0. The NLSE can then be simplified to: is a deterministic distortion of a pulse by itself, and is referred to as intrachannel self-phase modulation (ISPM). The terms where l m l ≠ = , 0 (and m l m ≠ = , 0 ) are distortions to a pulse by neighboring pulses, and are known as intra-channel cross-phase modulation (IXPM), as they are analogous to signal distortion by neighboring channels in XPM. Finally, the remaining terms 0 , ≠ m l are called intra-channel four-wave mixing (IFWM), because they are analogous to the interacting frequencies in FWM. We emphasize that the "intra-channel" effects all originate from SPM. It is well-known that in OOK with direct detection, IXPM causes timing jitter to the pulses, while IFWM causes amplitude jitter (or "ghost pulses") in the zero bits [56,57]. One can minimize these effects by careful dispersion map design [58,59] which represents signal propagation through one span of fiber. The IFWM-induced phase noise at symbol 0 is given by: is the deterministic perturbation to   Fig. 13 shows ( ) t k R , for single-polarization RZ-QPSK at 80 Gbit/s, where we used Gaussian pulses with 33% duty cycle. We compare Eq. (43) with Monte Carlo simulations at different sampling times t. The simulations used 5000-trial propagations through a typical terrestrial system with a random 32-pulse sequence. We observe good agreement between theory and simulation results. It is thus possible to reduce effects of IFWM by implementing a linear noise predictor in DSP using knowledge of ( ) k R . Assuming the received symbols are sampled at the optimal instants, we have . A 1.8-dB improvement in performance was obtained using a linear predictor for IFWM when IFWM is the dominant system impairment.

Nonlinear phase noise
SPM-induced NLPN is often called the Gordon-Mollenauer (G-M) effect [64]. Fig. 14(b) illustrates the G-M effect for QPSK. The received constellation is spiral-shaped, as signal points with larger amplitude experiences larger phase shifts. , where G is the amplifier gain. In the absence of CD and multi-channel nonlinear effects, the NLPN of a system with N uniformly spaced identical amplifiers is: It can be shown that the variance of Eq. (44) is:  [76]. Hence, ML detection can be implemented by rotating the received signal by ML θ and then using straight decision boundaries (Fig. 14(c)). The phase rotation techniques discussed can also be applied to 16-QAM where it has been shown that ML detection is well-approximated by straight-line decision boundaries when phase precompensation and/or post-compensation is implemented at the transmitter and/or receiver [76].
When CD is also considered, NLPN becomes more complicated. For signal propagation over a single span of fiber with perfect dispersion compensation and at high OSNR, the nonlinear phase noise is given by is the impulse response due to the CD of the fiber from z to L.

Laser phase noise
Laser phase noise is caused by spontaneous emission [87], and is modeled as a Wiener process [88]: is the instantaneous phase, is frequency noise with zero mean and autocorrelation . It can be shown that the laser output has a Lorentzian spectrum with a 3-dB linewidth Δν. Schawlow and Townes showed that laser linewidth is inversely proportional to output power [89], so it is desirable to operate the TX and LO lasers at maximum power, attenuating their outputs as required.
Phase noise is an important impairment in coherent systems as it impacts carrier synchronization. In noncoherent detection, the carrier phase is unimportant because the receiver only measures energy. In DPSK, information is encoded by phase changes, and Δν only needs to be small enough such that the phase fluctuation over a symbol period is small.
In Eq. (12), we showed that the baseband signal y(t) is modulated by In the absence of other impairments, this manifests as a rotation of the received constellation. Carrier synchronization is required to ensure ( ) t φ is small so the transmitted symbols can be detected with low power penalty. Since phase noise is a Wiener process with temporal correlation, it can be mitigated by signal processing. In the next two subsections, we consider carrier synchronization in a single polarization using a PLL and FF carrier synchronizer.

Phase-locked loop
The traditional method of carrier synchronization is to use a PLL. A system diagram of a PLL is shown in Fig. 15(a). The phase estimator removes the data modulation so that ( ) t φ can be measured. This can be done by a number of methods, including raising the signal to the M-th power in M-ary PSK [23]. The phase estimator output is an error signal that is then passed through a loop filter, producing a control signal for the LO frequency. In an OPLL, the control signal drives the LO laser, whereas in an electrical PLL, the control signal drives the electrical VCO ( Fig. 5(a)). Both types of PLL can be analyzed in the same manner, and the design of PLLs has been extensively studied in [29,90]. The performance of a PLL is usually analyzed with the linear model shown in Fig. 15(b). We assume the LO has no phase noise, while the TX laser has phase noise equal to the sum of the linewidths of the two lasers. The signal to be tracked is ( ) , which evolves as Eq. (48).
Owing to AWGN, the phase estimator measures ( ) ( ) ( ) Fig. 15(c)), and produces a voltage performance. An electrical PLL is usually superior to an OPLL since according to Fig. 5(a), the optical path has to pass through the optical hybrid and balanced photodetectors. However, an LO laser is likely to have a larger tuning range than an electrical LO. If the frequency drift of the lasers is significant, an OPLL may be preferable.
Ignoring delay, the denominator of Eq. (49) gives the loop order. For a first-order loop, . The design parameter is the loop bandwidth is typically chosen as a compromise between a rapid response and low steady-state variance. For both first-and second-order PLLs, there is an optimal n ω that minimizes the phase-error variance:  The first term in Eq. (50) arises from phase noise and is proportional to , which is the ratio between the laser linewidth and signal bandwidth. The second term arises from AWGN, and is inversely proportional to the received OSNR. Assuming the use of a decision- is a penalty factor associated with the transmitted constellation [20]. A larger n ω allows the LO to adapt more quickly to phase fluctuations in the TX laser, but the loop becomes more susceptible to noise. These two conflicting requirements give rise to an optimum n ω , which needs to evaluated numerically. A typical plot of φ σ versus n ω is shown in Fig. 16 for a second-order PLL. We have assumed singlepolarization 16-QAM at 100 Gbit/s. The laser beat linewidth and received OSNR are 100 kHz and 11.5 dB, respectively, which is 1 dB above sensitivity for BER = 10 −3 ( Table 6). We observe that delay increases φ σ . Above , there is no n ω that can give 1-dB power penalty due to phase error ( 7 [20]). At 100 Gbit/s, this maximum delay corresponds to 1.25 ns, which corresponds to a transmission line only ~ 25 cm long. Even with careful circuit design, this is probably not feasible. Hence for high-data-rate systems, FF carrier recovery techniques are likely to be required.

Feedforward carrier recovery
The PLL is a feedback system as the control phase at time t can only depend on past input phases up to d t τ − . However, the laser phase noise process described by (48) has a symmetric autocorrelation function , so its value at time t has the same correlation with phases before and after t. The PLL is suboptimal as it does not exploit  FF carrier synchronization for an intradyne receiver using analog electronics was described in [24], while digital FF carrier synchronization was studied in [21,92−94]. In this section, we focus on digital FF carrier synchronization. Consider the system model shown in Fig. 17(a), where we assume all other channel impairments have been compensated by the digital coherent receiver, whose outputs are x is the transmitted symbol, and k φ and k n are the carrier phase and AWGN, respectively ( Fig. 17(b)). Instead of using a PLL to ensure that k φ is small, a FF phase estimator directly estimates the carrier phase and then de-rotates the received signal by this estimate so symbol decisions can be made at low BER. The FF phase estimator has a soft phase estimator that first computes a symbol-by-symbol estimate k ψ of k φ , followed by a MMSE filter ( ) z W p . A number of algorithms exist for finding k ψ . In the case of M-ary PSK, raising k y to the M-th power which removes the data modulation. For more general signaling formats, decision-directed (DD) phase estimation can be used [18]. The symbol-by-symbol estimate is corrupted by AWGN (Fig. 17(b)) so that  As with the PLL, the variance has two terms that arise from phase noise and AWGN.

Power Penalty from Phase Error
Regardless of whether a PLL or a FF carrier synchronizer is used, the coherent receiver makes symbol decisions on has Gaussian distribution. In the limit of high SNR, the Tikhonov is well-approximated by the Gaussian distribution. The method for computing BER for a given phase-error distribution can be found in [95]. With the phase error variances for the PLL and the FF carrier synchronizer given by Eqs. (50) and (54), the power penalty can be determined. In Table 8, we compare the linewidth requirements for receivers that use a PLL and a FF carrier synchronizer, assuming a 1-dB power penalty at a target BER of 10 −3 [18]. We observe that FF carrier recovery can tolerate 50% to 100% wider laser linewidth than a PLL, and is also insensitive to propagation delay.

Carrier synchronization in polarization-multiplexed systems
When polarization multiplexing is employed, the receiver has two independent signals from which to estimate carrier phase. Consider the system models shown in Fig. 18. In the PLLbased receiver, the baseband signals for each polarization are passed through phase error estimators that compute independent estimates However, to obtain the same symbol-error rate for each polarization, the SNR per polarization must be preserved. This requires the transmit energy per symbol (now a 2×1 complex vector) be doubled. Hence, the dependence of phase error variance on SNR per symbol is preserved, and Eqs. (50) and (54) hold for polarization-multiplexed transmission.

Integration of receiver functions
A functional receiver needs to integrate carrier synchronization, linear equalization, nonlinear impairment compensation and symbol synchronization. If we consider NLPN to be the dominant nonlinear effect, and we assume IFWM is negligible, a hypothetical receiver structure is shown in Fig. 19. The sampled outputs of a dual-polarization downconverter are first passed through an NLPN compensator, which performs phase rotation proportional to the received amplitude. Linear equalization follows, yielding an output at one sample per symbol for which FF carrier synchronization is performed, followed by symbol decisions. The ordering of linear equalization before carrier synchronization is possible because from Table   8, we expect the laser linewidth ν Δ to satisfy ~4 10 − < Δ s T ν . In contrast, the memory length of the linear equalizer is only 13 symbols for the worst case considered in Table 7. Since the carrier phase does not change significantly over the equalizer's memory, performance should not be compromised. The filter Fig. 19 exploits temporal correlation of laser phase noise and IFWM to mitigate both impairments. To make the system adaptive, we employ decision feedback to update equalizer coefficients, and the error signal is also passed through a timing error detector (TED) and loop filter to yield a control signal for the sampling clock. The design of the TED and loop filter for symbol synchronization can be found in [49]. A receiver structure with NLPN compensation first, followed by linear equalization and carrier synchronization has been employed in [72] and [73], and has been shown to reduce nonlinear phase variance for QPSK under certain transmission conditions. A receiver structure that can mitigate nonlinearities (including IFWM) for more general system designs is an ongoing research topic. Fig. 19. Possible integration of nonlinear phase noise compensation, adaptive digital equalization and carrier synchronization.

DD DD
Avg. Avg. DD DD

Introduction
Orthogonal frequency-division multiplexing (OFDM) is a multi-carrier technique that has been deployed in DSL and wireless systems [96,97], and is receiving increasing interest in optical fiber communications [98−101]. A multi-carrier system sends information over N c frequency-division-multiplexed (FDM) channels, each modulated by a different carrier (Fig.  20(a)). For a given modulation format, with coherent detection, single-carrier and multicarrier systems offer fundamentally the same spectral efficiency and power efficiency, but these may differ in practice, because of different impairments and implementation details. In OFDM, the modulation is implemented digitally by the inverse Fast Fourier Transform (IFFT) (Fig. 20(b)). We can write a multi-carrier signal as: (55) represent a periodic extension of the IFFT (Fig. 21), and are known as the cyclic prefix. The integers c N and pre N are thus the number of sub-carriers and the prefix length, respectively. We can generate Eq. (55) using the structure shown in Fig. 20(b). The inputs to the transmitter are the symbols modulating each sub-carrier, applied at a rate of  is shown in Fig. 22, where we assumed equal energy for all the sub-channels. The modulated sub-carriers are overlapping in frequency. The receiver is able to separate them by the Fast Fourier Transform (FFT) because d f 1 is an integer multiple of the chip period T c , i.e., the sub-carriers are "orthogonal" over the duration of an OFDM symbol.
In OFDM, oversampling can be implemented by not modulating the edge sub-carriers in Fig. 20. Supposing only u N of the c N sub-carriers are used, the effective oversampling rate is . Compared to single-carrier transmission, oversampling can be easier to implement in OFDM, as it is possible to employ arbitrary ratios where a corresponding multirate solution for single-carrier modulation is prohibitively complex. Another advantage of OFDM is that the signal spectrum (Fig. 22) falls off more rapidly at the band edges than for a single-carrier signal. This is because the spectrum at each sub-carrier is

Compensation of linear impairments and computational complexity
The cyclic prefix serves a special purpose in OFDM.   The number of sub-carriers required is determined by the channel's impulse response in conjunction with the target excess bandwidth factor (κ 1 ) and oversampling ratio (κ 2 ). Suppose a bit rate of R b is required and all sub-carriers (with polarization multiplexing considered) encode b bits per symbol. , and is the same as single-carrier transmission. If adaptive equalization is required however, OFDM has a simpler update algorithm for the equalizer coefficients, as the filter taps and transmitted symbols are both in the frequency domain. In an FFT-based equalizer for single-carrier modulation, the filter taps are in the frequency domain, while the symbols are in the time domain. The update equation thus requires computing an FFT. In order to keep the overhead manageable, we expect the coefficient can only be updated infrequently for single-carrier.

Dynamic power allocation
For a frequency-selective channel, the optimum signal spectrum is given by water filling [105]. OFDM allows greater control of the signal spectrum as the power of each sub-carrier can be varied. In the limit of large u N , the OFDM spectrum can approach continuous water filling. Sub-carriers transmitted at higher power can use denser constellations, such that similar BERs are obtained in all sub-carriers. OFDM with dynamic power allocation has been successfully employed in wireless systems to combat signal multi-path [106−108].
Frequency-selective channels are less fundamental to optical fiber than to wireless systems, because a fiber's main linear impairments, CD and PMD, are unitary transformations in a dual-polarization coherent receiver. In particular, PMD merely shifts energy from one polarization to another. Nevertheless, the ability to optimize power allocation may be useful when transmitting through cascaded optical elements, such as (de)interleavers, (de)multiplexers and ROADMs. In reconfigurable networks, OFDM would enable optimal power allocation across the sub-carriers based on instantaneous channel conditions.

Laser Phase Noise and Nonlinear Effects
Laser phase noise destroys the orthogonality of the sub-carriers and causes inter-channel interference (ICI), which has noise-like characteristics. Wu and Bar-Ness showed that the variance of ICI grows linearly with N c [109]. Thus, the parameter that governs performance is the linewidth-to-sub-carrier-spacing ratio ( ). Phase noise considerations favor a smaller number of sub-carriers. At low phase noise, Armada and Calvo proposed using unmodulated pilot carriers to estimate the common phase error (CPE) for all the sub-channels [110]. At higher phase noise, the carrier phase can no longer be assumed as constant over an OFDM symbol. Wu and Bar-Ness found that using an MMSE equalizer that took into account both ICI and AWGN improves system performance [111]. However, this technique does not estimate or compensate for the phase noise process, unlike FF carrier synchronization for single-carrier modulation.
Carrier synchronization in OFDM may require an iterative algorithm. Suppose at the beginning of an OFDM symbol, an initial estimate of phase k φ is known from the previous symbol. We can first de-rotate the entire OFDM symbol (N c chips) by ( ) k jφ exp − , and then perform the FFT, equalization and symbol detection. We can multiply the symbol decisions by the channel's frequency response and then take the IFFT to compute what the time samples should have been without phase noise. This allows the receiver to compute estimates ψ . We can de-rotate the OFDM symbol again, and a second iteration follows. This process can be repeated until convergence is achieved. The performance of such an algorithm has not been characterized.
OFDM is potentially susceptible to fiber nonlinearity due to its high peak-to-average power ratio (PAPR) [112]. Electrical pre-and post-compensation of nonlinearity for long-haul OFDM transmission has been studied in [113,114], while the effects of FWM have been investigated by [115].

Implementation complexity: Comparison of single-and multi-carrier approaches
Digital coherent detection is, at present, limited mainly by available hardware. Optical systems transmit at symbol rates of tens of GHz, which requires very fast ADCs and powerful DSPs that can achieve the implementation complexity required. Parallelization and pipelining can be aided by block-based DSP algorithms. FFT-based implementations of both the singlecarrier and OFDM systems would facilitate this. Block-based digital FF carrier recovery has been studied in [116].
An important design parameter is the number of bits required for the ADCs, and this is determined by the tolerable clipping probability and quantization noise variance [27]. Suppose we require quantization noise to be at most a fraction k 1 of the AWGN variance per quadrature  Table 10 shows the required ADC resolution for 4-, 8-and 16-QAM assuming the receiver operates at 2 dB above sensitivity limit for a target BER of 10 −3 . We set the clipping probability to be 10 −4 , and the quantization noise variance to be 0.1 times the AWGN variance ( 1 k = 0.1, 2 k = 3.89). Finally, we consider computational complexity. Using the results of Tables 7 and 9, we compute the number of operations required for linear equalization of in single-carrier transmission and in OFDM, assuming polarization-multiplexed 4-QAM (Table 11). For single-carrier transmission, we calculated the complexity of direct implementation and FFTbased implementation using an optimized block size B (subject to N+B−1 being an integer power of two). We observe negligible difference between single-carrier transmission and (61) OFDM. In particular, for distances greater than 1,000 km, single-carrier equalization is more efficiently implemented by the FFT.

Conclusions
We have reviewed the principles of coherent detection in optical communications, and described digital techniques for compensating channel impairments.
Two types of channel impairments were identified: linear and nonlinear. Linear channel impairments include CD, PMD and optical filtering by network components. When the outputs of a dual-polarization downconverter are sampled above the Nyquist rate, CD, PMD and sampling time errors are unitary transformations that can be compensated by linear equalizers with negligible power penalty.
Nonlinear impairment compensation relies on exploiting known correlation structures for nonlinear phase processes. To combat NLPN, the receiver exploits the correlation between received amplitude and nonlinear phase shift. And to combat IFWM, the receiver uses the correlation of IFWM phase shift over neighboring symbols. In similar fashion, to combat laser phase noise, a FF carrier synchronizer relies on the temporal correlation of a Wiener process. In all three cases, the receiver performs ML phase correction.
Digital signal processing enables multi-carrier transmission using OFDM. The ability to dynamically manipulate the signal spectrum, the ease of implementing oversampling and the ease of adaptive equalization are key advantages of OFDM.