368

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 20, NO. 3, MARCH 2002

A Comparison Between Different PMD Compensation Techniques

AbstractWe quantify the benefits of using different techniques for compensation of polarization mode dispersion (PMD) in fiberoptic communication systems by means of numerical simulations. This is done both with respect to PMD-induced pulse broadening and in terms of system outage probability for different data formats [nonreturn-to-zero (NRZ) and return-to-zero (RZ)]. Attention is focused on simple and relevant single- and double-stage posttransmission compensators with a few degrees of freedom (DOF). It is generally believed that a PMD compensator with a polarization controller and a variable delay line can only compensate the PMD to the first order. We show, from analytical results, the counterintuitive fact that this scheme can also partially compensate for higher order PMD. We also investigate the benefit of using a polarizer as compensation element where the optical average power can be used as a feedback signal. Index TermsCompensation, data formats, equalization, fiberoptic communication, mitigation, nonreturn-to-zero (NRZ), outage probability, polarization mode dispersion (PMD), pulse broadening, return-to-zero (RZ), transmission, ultrahigh bit rates.

I. INTRODUCTION OLARIZATION-MODE dispersion (PMD) is an obstacle in many fiber-optic transmission systems, and as the bit rates increase to 40 Gb/s and beyond, its impact will be even more great. Consequently, there is a large interest in techniques to compensate or mitigate the effects of PMD, and a number of methods have been proposed. The main problem is the temporal drift of the PMD characteristics, which forces any compensation technique to dynamically adapt to changes while the system is running. Launching the signal into a principal state of polarization (PSP) [1] is a straightforward method to optically reduce the effects of PMD. This is defined as a first-order compensation because it is able to compensate the PMD only to the first order. However, in practice, a feedback signal must be connected back to the transmitter, which makes the compensation inherently slow. Optical post-transmission compensation is currently an intense research field where both first-order and higher order compensators have been suggested [2][7]. It is practical to use a compensator with just a few control parameters because this
Manuscript received April 3, 2001; revised December 13, 2001. This work was supported by Chalmers Centre for High Speed Technology (CHACH). H. Sunnerud, C. Xie, M. Karlsson, and R. Samuelsson are with the Photonics Laboratory, Department of Microelectronics, Chalmers University of Technology, Gteborg SE-412 96, Sweden (e-mail: [email protected]). P. A. Andrekson is with the Photonics Laboratory, Department of Microelectronics, Chalmers University of Technology, Gteborg SE-412 96, Sweden, and also with CENiX, Inc., Allentown, PA 18106 USA. Publisher Item Identifier S 0733-8724(02)02193-X.

limits the complexity and speeds up the response. However, a properly designed compensator with a large number of control parameters [or degrees of freedom (DOF)] is more flexible and able to compensate also for higher orders of PMD [5]. There is no complete consensus about what first-order compensation really means. It is generally believed that optical postcompensators with a polarization controller (PC) and a variable time delay only can compensate the PMD to the first order. We like to think of first-order compensation, from a mathematical point of view, as either aligning the light along a PSP (the PSP method) or cancelling the PMD vector at the carrier frequency, as in [8], where it also is shown that these two methods have very similar performances. However, that is not to say that this is the optimum strategy of operation. In this paper, we quantify the benefit of using different PMD compensation techniques by means of numerical simulations. The investigated methods are all simple and relevant, and some may become successful commercial products. First, we study the average pulse-broadening reduction associated with the different techniques and compare that with the first-order compensation technique where the PMD vector is cancelled at the carrier frequency. This method is henceforth referred to as first order, which we use as the benchmark method. We also present analytical results which show that an optical postcompensator with a PC and a variable time delay (previously known as firstorder compensation) is also able to compensate for higher orders of PMD. Quantifying the compensators in terms of average pulse broadening entails a high grade of generality and simplicity, and allows for some analytical predictions. However, it does not reveal the whole truth of how efficient a compensator is. Therefore, we also take the whole statistical picture into account by investigating the corresponding system outage probability [9], in terms of bit error rate (BER) degradation, which is a very relevant measure of the system performance. This is evaluated for both NRZ and RZ data formats in a realistic transmission system. II. SIMULATION PRINCIPLE The investigated techniques are schematically shown in Fig. 1. We find it useful to classify the compensators in terms of their number of DOF because this is a measure of their complexity, rather than their alleged order of compensation (such as first or second). The PSP method shown in Fig. 1(a) is a pretransmission compensation technique in which a PC is used to align the state of polarization (SOP) with one of the input PSPs of the fiber link [1], [8]. This corresponds to only retarder followed by a two DOF because it suffices with a

0733-8724/02$17.00 2002 IEEE

SUNNERUD et al.: DIFFERENT PMD COMPENSATION TECHNIQUES

369

Fig. 1. Compensator schemes. (a) PSP method (two DOF). (b) PC fixed delay (two DOF). (c) PC variable delay (three DOF). (d) Two PCs and two fixed delays (four DOF). (e) Two PCs, one fixed and one variable delay (five DOF). (f) PC polarizer (two DOF).

retarder to transform any given input SOP to an arbitrary output SOP. However, in practice, rotating waveplates are slow, and fast electrooptical crystals (like LiNbO ) are preferable, but more than two segments must be used to achieve endless control in such crystals [10]. Various optical postcompensators have been studied. One is a PC and a fixed time delay (two DOF) [2], as shown in Fig. 1(b). Another is a PC and a variable delay (three DOF) [3], as shown in Fig. 1(c). In Fig. 1(d), a double-stage compensator with fixed delays (four DOF) [4] is demonstrated. Fig. 1(e) shows a double-stage compensator with one fixed and one variable delay (five DOF), slightly similar to [7]. Fig. 1(f) shows a PC and a polarizer in combination (two DOF) [11], where the average power through the polarizer could act as an error signal. The exact simulation principle for modeling the compensator schemes in a transmission system in terms of Jones matrices is given in Appendix I. In order to quantify the benefit of applying the different compensation techniques, we first investigated the reduced pulse broadening associated with the compensators. We launched single Gaussian pulses into a transmission system (i.e., a fiber link and a compensator) and calculated the output root mean square (rms) pulsewidth. The operation of the PSP method in Fig. 1(a) was given analytically (see Appendix I-B), but corresponds well to optimizing the input SOP with respect to the minimum output pulsewidth. The postcompensators using birefringent elements [Fig. 1(b)(e)] were numerically optimized with respect to the output pulsewidth (the simulation details are given in Appendix I-C). The polarizer in Fig. 1(f) was aligned so that maximum average power was transmitted which also was given analytically (see Appendix I-D); the fiber was modeled as a concatenation of randomly oriented birefrin-

gent segments (see Appendix I-A). From each fiber realization, one sample of the output pulsewidth was obtained. In order to establish a reliable average, 2000 randomly independent realizations were tested in a Monte Carlo simulation. This is described in detail in Section III. The normalized average pulse broadening is a convenient measure of the advantage of using different compensation schemes because it is generic, independent of the bit rate, and allows for some analytic predictions. However, it does not reveal the whole truth of how efficient a compensator is. It is very relevant to compare the PMD compensation techniques with respect to their performances in a transmission system with digitally encoded data relating the average differential group delay (DGD) in the system to the actual bit rate. It is also very useful to quantify the performance in terms of outage probability, i.e., the probability that the system stops functioning properly. Hence, we launched pseudorandom bit sequences (PRBS) of Gaussian RZ pulses or raised cosine NRZ pulses into the system (i.e., the fiber link with PMD in combination with a compensator) and detected the output sequences with an erbium-doped fiber amplifier (EDFA)-preamplified receiver. Subsequently, the corresponding BER was calculated . In where an outage was defined as whenever BER order to establish a reliable outage probability, a large number (10 000) of randomly independent realizations were tested. The results from this approach are presented in Section IV. III. PULSE BROADENING First, we investigated the improvement from using the different compensation techniques with respect to PMD-induced

370

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 20, NO. 3, MARCH 2002

Fig. 2. Average broadening factors for the uncompensated case, first-order compensation and the PSP method, a comparison between simulations (symbols) and analytical results (lines) from [8].

Fig. 3. A comparison between first-order compensation and a real compensator with a PC and a variable delay (three DOF), which are optimized with respect to the output pulsewidth (dots). The solid line represents the approximate broadening factor E fb g  . Ef  g =

pulse broadening. We launched single Gaussian pulses into the system and calculated the output rms pulsewidth , and compared it with the input pulsewidth . The pulse broadening was and the average pulse-broadening factor defined as , where represents the average of the statisas tical data. We used the second moment of the broadening factor for the purpose of comparing the simulations with the analytical results in [8] for the uncompensated case, the PSP-method, and first-order compensation (canceling the PMD vector at the carrier frequency), which we will use henceforth as the benchmark method. This comparison is shown in Fig. 2 as a function of the normalized average DGD, i.e., divided by the pulsewidth with excellent agreement. It was pointed out in [8] that the PSP method and the first-order compensation schemes , are identical in the low PMD region where whereas the PSP method is slightly better at high average DGD, . This can be understood when considering where that the part of the pulse spectrum that is uncompensated (i.e., located outside the operational bandwidth of the compensator, [12]) will have an extra contribution to the typically PMD from the compensating element and, hence, a higher average PMD. Note that it is not trivial to extract an error signal for the first-order operation. However, some progress in this direction has been made recently through polarization scrambling [13], [14]. An advantage of the first-order technique is that it only has to adapt to the PMD vector of the fiber, rather than the actual SOP, which varies approximately twice as fast [15]. Also, note that the first-order compensation technique actually is a three-DOF method, but the parameters are fixed relative to the PMD vector and are not optimized with respect to pulsewidth. Hence, one can expect that some improvement is achieved by optimizing the parameters properly, which will be discussed. A. Variable Delay (Three DOF) Cancellation of the PMD vector at the carrier frequency seems to represent a somewhat constructed case, and it is not obvious that this is the best strategy of operation. Usually,

(1 + 1

(144 ))

the PC is optimized with respect to a certain error signal, as in Fig. 1(c) (three DOF). Fig. 3 shows a comparison of the first-order compensation and the real compensator, which was operated to minimize the output rms pulsewidth. This strategy of operation is better than the first-order compensation, as can be expected. The solid line represents the approximate analytical broadening factor for this scheme. It should be noted that if the operation were optimized with respect to full-width at half-maximum (FWMH) pulse broadening, the relative improvement from the three-DOF method would be further enhanced [16]. It was pointed out in [17] that, for a compensator with fixed length, there exists one more useful point of operation where the total PMD vector (the PMD vector of the compensator in combination with that of the fiber) is aligned with the state of polarization, and an alternative PSP method is, in that way, obtained. To investigate how this alternative process behaves for the variable compensator, we calculated the angle between the and the SOP in Stokes space, . It should total PMD vector be noted that this angle is calculated at the fiber output, before the compensator, where all vectors must be transposed by the appropriate Mller matrices to the same position [18]. Here, the , is constant with respect to PMD vector of the compensator, , and the output PMD frequency while both the output SOP, , vary with frequency. To obtain convector of the fiber, stant average vectors, they are integrated over the normalized pulse spectrum, which is indicated by the bracket notation (see Appendix I-D). The angle is defined by the scalar product as

(1)

In Fig. 4, it is clearly shown that this angle has a large probability to be 0 or , which implies that the total PMD vector very often is located along the line of the signal SOP, rather than

SUNNERUD et al.: DIFFERENT PMD COMPENSATION TECHNIQUES

371

Fig. 4. Histogram of the angle  between the total PMD vector and the SOP at the optimum point for the three-DOF compensator E f  g= : .

= 0 67

being null, which was the definition of the first-order compensator. Note that when the total PMD vector is very small, it is difficult to define the angle . Therefore, samples with a total PMD vector shorter than the numerical resolution in our simulations have been filtered out. In order to analytically capture the performance of the three-DOF compensation method with a PC and a variable delay, we use a similar approach as in [8] (see Appendix II for the full derivation). The PMD-induced rms pulse broadcan be written in terms of the PMD vector as ening , where is the Stokes vector of the signal after transmission. By inserting the total PMD into this expression, we can minimize vector . In this minimization the pulse broadening with respect to process, we notice that there is only one minimum present, i.e., no local suboptima arise, a problem for the double-stage compensators with more DOFs, which will be discussed in Section III-C. We are then able to analytically express the , corresponding to the minimum compensation vector, broadening, together with the minimum as (2) (3) is a parameter that is nonzero where is a function only if higher order PMD is present, for which of frequency. In the expression for the minimum broadening, (3), we note that the first two terms are identical with the performance of the PSP method [8], but the last term arises from the higher order PMD because it is proportional to , and will give an extra reduction of the pulsewidth. Hence, the three-DOF method is capable of improvements over the first-order PMD compensators. In Fig. 5, we show the PMD vector addition diagram for the averaged vectors (integrated over the pulse specand , and possible compensator vectors . All trum), fully comcompensation vectors that lie on the line through , is the optimum pensate for first-order PMD. One case, that also will reduce some amount of higher order PMD. We should also mention that it is possible to go a step further and calculate the statistical average of this result, as was done in [8], but then only for the limiting case of small PMD relative . The result is to the pulsewidth (solid line in Fig. 3), which is to be that was found compared with for first-order PMD compensation and the PSP- method [8].

Fig. 5. PMD vector addition diagram for the fiber h i and the compensato, in relation to the Stokes vector of the signal hs i. The compensation vectors (dashed) that lie on the line through hsi fully compensate for first-order PMD. One optimum case (solid) will also reduce some amount of higher order PMD.

Fig. 6. The pulse-broadening factor for the fixed (optimum) delay compensator (two DOF) compared with the benchmark method. Inset shows the optimum delay.

This implies that the system can accept more PMD than the first-order techniques for the same average pulse broadening in this limit, using the three-DOF technique. B. Fixed Delay (Two DOF) If a fixed delay is applied, as in Fig. 1(b) (two DOF), the PMD vector of the compensator is not able to fully cancel that of the fiber. This scheme optimizes the PC like the three-DOF method, but without varying the delay, it can still decrease the total DGD and, in many cases, align the total PMD vector with the SOP to achieve the alternative PSP transmission operation. The efficiency of this process depends on the length of the fixed delay element. Fig. 6 shows the average pulse broadening for this scheme with optimum (fixed) length compared with the benchmark method. The performance is worse than first-order compensation at low PMD, but it is slightly better at larger PMD. The inset shows the optimum length of the fixed compensator as a function of the average DGD of the fiber . In the approaches 1.15, low-PMD limit, the ratio , this value is 0.75. whereas at high PMD At large PMD, where the operational bandwidth of the compensator is low, the improved performance from a long delay decreases as the added DGD in the compensator increases the average PMD. Note that this is the optimum ratio with respect

372

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 20, NO. 3, MARCH 2002

Fig. 7. Histogram of the angle  between the total PMD vector and the SOP at the optimum point for the two-DOF compensator. E f  g= : ,X  =E f  g : .

1 = 1 08

= 0 67

Fig. 9. The pulse broadening factor for the polarizer compensator (two DOF) aligned so that maximum power is transmitted.

Fig. 8. The pulse-broadening factor for the double-stage compensators: four DOF with fixed delays   E f  g ; five DOF with one fixed delay  E f  g ; and one variable delay.

(1

(1 = 1 )

=1

= 1 )

to average pulse broadening. However, with respect to outage probability, this ratio is likely to be higher because the compensator then must be able to compensate for the highest DGD values in the tail of the Maxwell distribution to avoid an outage, rather than minimizing the average pulse broadening. This will be discussed further in Section IV. We also calculated the angle between the total PMD vector and the SOP for this method. The histogram in Fig. 7 shows that this angle has another distribution compared to the variable compensator. The total PMD vector does not reach to the line of the SOP as often, but a large correlation remains. C. Double-Stage Compensators (Four and Five DOF) It is well known that double-stage compensators are able to compensate also for higher order PMD. However, a compensator with fixed delays is mainly applied to facilitate the variation of the effective delay because PCs often are faster and easier to implement than a variable delay line. The double-stage compensators were investigated and optimized with respect to the rms pulse broadening. The results are shown in Fig. 8 and compared with the analytical benchmark method. Because we use a trial-and-error algorithm to find the optimum parameters (which also can be used in practice), the simulations are complicated here by suboptima that arise for the double-stage compensators. We have also found that there is a small probability

(typically 10 10 ) that the compensator will be trapped in a suboptimum that slightly degrades the performance with respect to the uncompensated case. This seems to occur only at high-average PMD while the local DGD at the carrier frequency is very low, so that the uncompensated pulses are more or less undistorted. However, this degradation is not serious and will not make the pulses worse than they would be at high local DGD. Fig. 8 shows the pulse broadening for the double-stage compensators both operating at an arbitrary optimum (local or global) and at the best out of four independent optimizations, which we use as an approximation of the global optimum. At large PMD, the double-stage compensators perform better than first-order compensation, and, as expected, the five-DOF compensator is better than the four-DOF compensator. Both fixed delays of the four-DOF scheme and the single fixed delay of the five-DOF scheme were chosen to be equal to the average PMD of the fiber, which turned out to be near optimal in terms of pulse broadening. In practice, the operation could be trapped in a local optimum unless an efficient way to find the global optimum is proposed. Apparently, the performance is worse when operating at a suboptimum. It is also somewhat worse than first-order compensation at low PMD, which is barely seen in this scale. However, this will be discussed further and is more clearly shown in Section III-E. D. Polarizer (Two DOF) The polarizer can be applied as a PMD compensation element [11], where the average power can be used as a feedback signal. This strategy is inexpensive and, therefore, an attractive way of reducing the pulse broadening. The compensation algorithm is simple because only one optimum is present, and furthermore, this scheme is independent of the amount of DGD in the link, making it a generic implementation. In Fig. 9, the average pulse broadening is shown for the polarizer compensator. The performance is better than first-order compensation at large PMD because it does not add any DGD. This indicates that this technique may be used advantageously in RZ systems.

SUNNERUD et al.: DIFFERENT PMD COMPENSATION TECHNIQUES

373

TABLE I PMD-ENHANCEMENT FACTORS IN DECIBELS FOR NRZ AND RZ PULSES

Fig. 10. PMD-enhancement factors as a function of the average DGD for the uncompensated case for different compensators compared with first-order compensation (dashed line). (a) Single-stage compensator with variable delay (three DOF) and fixed delay (two DOF). (b) Double-stage compensators with one fixed and one variable delay (five DOF), fixed delays (four DOF), and polarizer aligned for maximum average power transmission (two DOF).

However, the signal must be restored before too much PMD-induced depolarization has occurred for the feedback circuit to find an optimum point. The maximum signal loss is 3 dB, in this case. The performance is worse than first-order compensation at low PMD, but an improvement is achieved, compared to the uncompensated case. E. Comparison Another way of quantifying the efficiency of the compensators is with respect to the PMD-enhancement factor [8]. This factor is defined for a given average pulse broadening as the number of times the average PMD is allowed to increase for the compensated systems to give the same broadening as the uncompensated case. In Fig. 10(a), the PMD-enhancement factors are shown for the single-stage compensators with fixed (two DOF) and variable delay (three DOF) designs. In Fig. 10(b), this is shown for the double-stage compensators (four and five

DOF), for operation at both an arbitrary (local or global) optimum, and at the best out of four optima, which we use as an approximation of the global optimum. The PMD-enhancement factor is also shown for the polarizer compensator. The performance of all compensators is compared with first-order compensation (dashed line). The PMD-enhancement factor also clearly shows the performance at low PMD, which was not clearly seen in the pulse-broadening plots. Obviously, this factor depends very much on the pulsewidth in relation to the average DGD, and the compensation of long NRZ pulses will be more efficient compared to shorter RZ pulses. One exception is the characteristic behavior of the polarizer aligned so that maximum power is transmitted, which is different compared to all other methods because the improvement is virtually independent of the pulsewidth. The PMD-induced system degradation was studied for NRZ and RZ data formats in [19], and similar results will be presented in Section IV . For a certain outage probability and power margin, the NRZ system can accept 17% average DGD per bit slot, whereas, for the same conditions, the RZ format (with 30% duty factor) can handle more (22%). If we insert those numbers in the PMD-enhancement model, we could predict how efficient the different compensators are in NRZ and RZ systems. For exps ps NRZ pulses are used ample, if in a 40-Gb/s system, and 17% (4.25 ps) average DGD per bit . Correspondingly, for RZ slot is acceptable, modulation, with shorter pulses, say 7.5 ps (30% duty factor), . In Table I, the PMD-enhancement factors, which predict how much more PMD can be accepted for the same performance (i.e., the same average pulse broadening), are shown for all compensators for both NRZ and RZ data formats. dB corresponds to four times more acceptFor example, able PMD, which represents 16 times longer transmission distance. The PMD-enhancement factors could, in principle, also be used to predict how much PMD can be accepted per bit slot in a system environment. However, this is evaluated properly in Section IV, where we study the benefit of using the different compensators in terms of system outage probability for both NRZ and RZ coding.

374

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 20, NO. 3, MARCH 2002

IV. OUTAGE PROBABILITY The average pulse broadening and the corresponding PMDenhancement factor is a convenient measure of the advantage of using different compensation schemes. However, it does not reveal the whole truth of how efficient a compensator is. It is often at high DGD values in the tail of the Maxwell distribution that large broadening occurs. These severe pulse distortions give rise to outages in the transmission, i.e., the system no longer works properly. Thus, it is relevant to compare the PMD-compensation techniques with respect to their outage probabilities in a transmission system with digitally encoded data and relate the average DGD in the system to the actual bit rate. We find it very relevant to quantify the system performance in terms of BER, because this puts the situation in a somewhat more practical context. Our simulation conditions are almost identical to those presented in [19] and similar to those presented in the analytical theory in [9]. However, the simulations take higher orders of PMD into account, and we also launch PRBS sequences (word length ) with Gaussian RZ pulses or raised cosine NRZ pulses into the fiber distorted by PMD and detect the output sequences with an EDFA-preamplified (noise figure 3 dB and a 1.3-nm optical filter) receiver. The electrical signal is low-pass filtered with a , was opthird-order Butterworth filter where the bandwidth, for both timized for best sensitivity without PMD ( RZ and NRZ, where is the bit rate). The EDFA adds noise to the signal, which then is detected at the receiver. The standard deviation of the noise in the detected signal as a function of time is calculated from 10 runs of detection with different amplifier noise seeds. Subsequently, the total BER is calculated (using Gaussian statistics [20]) as a function of decision level and sampling time (which is an average of more than 31 bits, similar to [21]), and minimized to give the optimum BER for the sequence. Only PMD and signal-to-noise ratio (SNR) were taken into account so that chromatic dispersion and nonlinearities were neglected. To be able to compare the compensators operating at different data formats, i.e., NRZ or RZ, in a fair way, we operated at 2-dB power margin above the receiver sensitivity at BER for both formats, as in [19]. The compensators, which previously were optimized with respect to the output pulsewidth, were optimized here with respect to a feedback signal given in [3], which was extracted from the received radiofrequency (RF) spectrum. The corresponding method given in [5] gave similar results. The operation of both the PSP method and the polarizer methods were still given analytically, as in the pulse broadening and, chapter. An outage was defined as whenever BER subsequently, the outage probability was calculated from 10 000 random fiber realizations. We use 40-Gb/s transmission, but the results could be scaled to any bit rate. The outage probability for the single-stage compensators with variable delay (three DOF) and fixed delay (two DOF) is shown for NRZ data in Fig. 11(a) and are compared with the uncompensated case and first-order compensation. The latter two agree well with the results from [22]. The variable compensator shows the best performance, as expected, where the average DGD is allowed to increase from 17% average DGD per bit slot for the uncompensated case to 43% at an outage probability of 10 . For the two-DOF compensator, the performance depends on the

Fig. 11. Outage probability as a function of the average DGD per bit slot for the variable (three DOF) and fixed-delay (two DOF) compensators compared with the uncompensated case and first-order compensation (dashed lines). (a) NRZ. (b) RZ.

fixed length of the delay. Two examples are shown for the relaand . For a long tive length, delay, the performance is worse at high PMD. However, the curve falls off rapidly as the PMD decreases and the performance is better at low PMD. The opposite holds for the short delay. Hence, there is an optimum delay that depends on the average DGD and the acceptable outage probability. The results for the two-DOF scheme is also shown at its optimum (fixed) length. The relative length of this delay, , is shown in the inset. This factor increases at low average DGD, as expected from the results of the pulse broadening in Section III-B. It is also larger when considering outage probability instead of the average pulse broadening, as expected. It should also be noted that at a lower acceptable outage probability and, hence, a lower acceptable average DGD, this optimum length will be even longer in relation to the average DGD. When operating at the optimum length, its performance is similar to the first-order compensation scheme near the acceptable outage probability of 10 . Thus, the average DGD is allowed to increase from 17% for the uncompensated case to 36% for the same outage probability.

SUNNERUD et al.: DIFFERENT PMD COMPENSATION TECHNIQUES

375

Fig. 12. Outage probability as a function of the average DGD per bit slot for the four-DOF double-stage compensator   Ef  g compared with the uncompensated case and first-order compensation (dashed lines). (a) NRZ. (b) RZ.

(1

=1

1 )

Fig. 13. Outage probability as a function of the average DGD per bit slot for the polarizer compensator (aligned for maximum power) compared with the uncompensated case and first-order compensation (dashed lines). (a) NRZ. (b) RZ.

The corresponding results for RZ are shown in Fig. 11(b). The uncompensated case is better for RZ than for NRZ, which was emphasized in [19]. The main reason for this is that the energy is more confined in the center of each bit slot in the RZ case; also, more DGD is required before the energy leaks out of the bit slot to result in intersymbol interference. However, the benefit from using first-order PMD compensation is smaller in RZ systems, which is due to the broader spectrum. This was also mentioned in [23]. Accordingly, the benefit is reduced for both the fixed (two DOF) and the variable (three DOF) compensator, as expected, also from the pulse-broadening simulations. The optimum fixed length for the fixed compensator (inset) is smaller compared to the NRZ format, which also is expected. The PMD can be increased from 22% to 33% for the fixed compensator and to 35% for the variable compensator. The outage probability can be further decreased by using a double-stage compensator, which is capable of reducing a more significant amount of second-order PMD than the three-DOF method. The results for the four-DOF compensator are shown in Fig. 12 for operation at an arbitrary local optimum, and for

the best out of four independent optimizations that we use as an approximation of the global optimum. The acceptable average PMD is further enhanced from 17% to 49% in the NRZ case, and from 22% to 43% for the RZ data, when operating at the global optimum. The more interesting case is, perhaps, when the compensator is trapped in a local optimum, which also can occur in practice. In this case, the benefit is somewhat decreased, but still the performance is better than first-order compensation and, in fact, is very similar to the three-DOF method. The outage probability is shown in Fig. 13 for the polarizer compensator (two DOF), which is aligned so that maximum power is transmitted. For NRZ coding, the outage probability decreases when some intersymbol interference is reduced. At the same time, half of the amplified spontaneous emission (ASE) noise from the EDFA is removed, which decreases the spontaneousspontaneous beat noise in the detector and further improves the performance. However, the polarizer compensator gives rise to a power loss that degrades the performance. This is more evident for RZ modulation because a shorter pulse gener-

376

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 20, NO. 3, MARCH 2002

TABLE II ACCEPTABLE AVERAGE DGD PER BIT SLOT  FOR NRZ AND RZ WHEN USING VARIOUS PMD COMPENSATORS AT AN OUTAGE PROBABILITY OF 10

worse than the first-order compensation at low PMD, but still better at high PMD. The whole set of compensators were then quantified in terms of outage probability in a system with digitally encoded NRZ and RZ data. The order of precedence for the compensators is similar to the pulse broadening quantification, but with some exceptions. Real numbers are put on the amount of PMD that can be handled in terms of average DGD per bit slot. The RZ format performs better in the uncompensated case, while all compensators are more efficient for NRZ modulation, which is due to the narrower spectrum. Hence, when operating near the PMD limit in the NRZ case, some improvement can be achieved by going to the RZ format, and may, in some cases, eliminate the need for active PMD compensation. However, if that is not sufficient, NRZ in combination with PMD compensation is a better choice. APPENDIX I SIMULATION DETAILS A. Fiber Modeling The optical input signal to the fiber link, which can be a single pulse or a sequence of pulses, is represented by an electrical . The fiber is modeled as field (Jones) vector randomly oriented birefringent a concatenation of segments, which results in a frequency dependent Jones matrix

ally is more depolarized and results in almost no improvement at all for RZ modulation. In Table II, the acceptable average DGD per bit slot is shown for the uncompensated case and for various PMD compensators for NRZ and RZ data formats. The acceptable outage probability is 10 , in this case. Although RZ performs better in the uncompensated case, all compensators are more efficient for NRZ modulation, which is due to the narrower spectrum. Hence, when operating near the PMD limit in the NRZ case, some improvement can be achieved by going to the RZ format, which in some cases may eliminate the need for active PMD compensation. However, if that is not sufficient, NRZ in combination with PMD compensation is a better choice. For example, by going from NRZ to RZ, the acceptable average DGD can be increased from 17% to 22%, which corresponds to 67% longer transmission distance, whereas by applying a four-DOF compensator, this can be increased to 49%, or 8.3 times longer transmission distance. V. CONCLUSION By means of numerical simulations, we have quantified the advantage of using some simple and relevant PMD compensation techniques, first regarding the reduction of the PMD-induced pulse broadening. The compensators were compared with the analytically calculated broadening factors of the uncompensated case, the first-order compensation, and the PSP method. We have shown that a single-stage compensator with a PC and a variable delay (a real compensator) performs better than the (analytically given) first-order compensators, which is due to a higher order compensation effect. A single-stage compensator with fixed delay is also beneficial, but performs slightly worse than first-order compensation at low PMD; however, its performance is better at large PMD. Two different double-stage compensators (with fixed delays and with one fixed and one variable) were investigated. They are better than first-order compensation when operating at the global optimum. However, in reality, a double-stage compensator could be trapped in a local optimum unless a good way to reach the global optimum was found. As a consequence, in terms of pulse broadening, both double-stage compensators (four and five DOF) will be slightly

(4) Each segment is represented by (5) the delay of each where is the relative rotation angle and segment, respectively, and where is the angular frequency. , The output field is then given by is the Fourier transform of the input field. The where spectrum is centered around the carrier frequency , which is set equal to zero in the simulations, for simplicity. The output , is then obtained by inverse transformafield in time, . tion; finally, the optical power is given by From each fiber realization, one sample of, for example, the output pulsewidth or the BER is obtained. In order to establish a reliable average or a reliable outage probability, many randomly independent realizations must be tested in a Monte Carlo simulation. B. PSP Transmission Method In the PSP method, a PC is introduced before transmission, which is represented by the Jones matrix , and results is a total Jones matrix (6) and . The angles and can where now be optimized with respect to a given error signal, for example, the degree of polarization (DOP) [17], an electrical

SUNNERUD et al.: DIFFERENT PMD COMPENSATION TECHNIQUES

377

signal extracted from the received RF spectrum [3], [5], or the output pulsewidth in the pulse broadening simulations. The can also be calculated output PSPs at the carrier frequency to the matrix product analytically as the eigenvectors . The input eigenvectors are then obtained by , from which the input polarization state should be chosen for PSP transmission. C. Postcompensation Using Birefringent Elements An optical postcompensator is generally composed by a number of stages which are represented by individual Jones matrices. The schemes we have investigated, a single-stage compensator [Fig. 1(b) and (c)] and a double-stage compenand sator [Fig. 1(d) and (e)], are represented by , respectively. Hence, this results in a total . The PC matrix was given in (6) Jones matrix and a delay element is represented by (7) is variable in the general case. Subwhere the time delay sequently, the variables can be optimized with respect to a certain error signal. PMD compensation with a polarization controller and a variable delay is achieved by canceling the PMD vector (in Stokes space) at the carrier frequency , as in [8]. for this operThe angles and of the PC and the delay ) are obtained in Jones space through ation (where , the matrix product . which should be null, or equivalently One could argue that for the double-stage compensators, where higher order PMD is present, one should use a more general polarization controller with three DOF. In that way, the rotation of the PMD vector of the compensator with frequency could be better suited to the rotation of the fiber PMD vector. This was tried, but showed negligible improvement. We believe that this was because of the tendency for these compensators to getting trapped in suboptima. D. Polarizer A polarizer has been proposed as a compensation element [11] where the average power from the polarizer could work as a feedback signal. Hence, the polarizer should be oriented, in Stokes space, so that the average power is maximized. This , where implies that the polarizer Stokes vector equals is the frequency-dependent Stokes vector of the signal after indicates integration transmission and the bracket notation over the normalized pulse spectrum, i.e. (8) , equals We could also note that the modulus of this vector, the DOP. In Jones space, a general elliptic polarizer, i.e., a polarization controller and a linear polarizer, can be described by a Jones matrix (9)

where is the eigenvector of the polarizer that corresponds to maximum and are calculated from transmission. The angles . The total Jones matrix . for the fiber and the polarizer is given by APPENDIX II THREE-DOF MINIMUM PULSE BROADENING

DERIVATION

OF THE

In order to analytically capture the performance of the three-DOF compensation method with a PC and a variable delay, we use a similar approach as in [8]. The PMD-induced rms pulse broadening can be written in terms of the PMD vector as (10) denote integration over the normalized where the brackets pulse spectrum (see Appendix I-D). It is also crucial that the Stokes vectors within the scalar product are evaluated at the same position in the fiber. We find it convenient to select the vectors after transmission but before PMD compensation. At this position, we define to be the Stokes vector of the signal, and is the total PMD vector. The total PMD vector will be , the vector sum of the output PMD vector from the fiber, . Thus, we and the input PMD vector of the compensator, . This expression is now put into (10). By have viewing the pulse broadening as a function of the compensator (which is constant and can be moved outside the invector tegrals) we can write the pulse broadening in the form (11) where is a constant (independent of the compensator vector), , and . We now seek to minimize the pulse broadening with respect to the compensator vector. By putting the gradient (with respect to ) of the pulse broadening to zero, we get the vector equation (12) This equation has the unique solution (13) It may also be shown that this solution is always the unique global minimum to (11), implying that the three-DOF compensator has no suboptima, when using the pulsewidth as error signal. After some algebra, the minimum compensator vector may be expressed as (14) . It is a bit counterintuitive where that the minimum broadening does not arise for the nulling out of the fiber PMD vector (which would correspond to ), but this arises from the fact that higher orders of PMD are present. The presence of the parameter , which is nonzero

378

JOURNAL OF LIGHTWAVE TECHNOLOGY, VOL. 20, NO. 3, MARCH 2002

only if higher order PMD is present, reveals that it is possible to compensate also for the higher orders, to some extent, with in the expression the three-DOF method. After inserting for the pulse broadening and performing some algebra, one may express the minimum pulse broadening as (15) Finally, one might like to take the statistical average of the minimum broadening in the case of a Gaussian pulse, similar to what was done in [8]. However, we only manage to do this analytically in the limit of small PMD, where the average pulse broad, ening becomes is the rms pulsewidth of the Gaussian pulse. This where value should be compared with the PSP transmission method which, on average, gets a pulse broadening of , as found in [8]. REFERENCES
[1] T. Ono, S. Yamazaki, H. Shimizu, and H. Emura, Polarization control method for suppressing polarization mode dispersion in optical transmission systems, J. Lightwave Technol., vol. 12, pp. 891898, May 1994. [2] T. Takahashi, T. Imai, and M. Aiki, Automatic compensation technique for timewise fluctuation polarization mode dispersion in in-line amplifier systems, Electron. Lett, vol. 30, pp. 348349, Feb. 1994. [3] F. Heismann, D. A. Fishman, and D. L. Wilson, Automatic compensation of first-order polarization mode dispersion in a 10 Gb/s transmission system, in Proc. ECOC98, vol. 1, Madrid, Spain, 1998, pp. 529530. [4] W. Shieh, H. Haunstein, B. Mckay, D. Fishman, A. Golubchik, J. Diubaldi, C. Martell, V. Arya, R. Lee, and H. Choudhury, Dynamic polarization-mode dispersion compensation in WDM systems, in Proc. ECOC2000, vol. 2, Munich, Germany, 2000, pp. 4143. [5] R. No, D. Sandel, M. Yoshida-Dierolf, S. Hinz, V. Mirvoda, A. Schpflin, C. Glingener, E. Gottwald, C. Scheerer, G. Fisher, T. Weyrauch, and W. Haase, Polarization mode dispersion compensation at 10, 20 and 40 Gbit/s with various optical equalizers, J. Lightwave Technol., vol. 17, pp. 16021616, Sept. 1999. [6] M. Shtaif, A. Mecozzi, M. Tur, and J. Nagel, A compensator for the effects of high-order polarization mode dispersion in optical fibers, IEEE Photon. Technol. Lett., vol. 12, pp. 434436, Apr. 2000. [7] Q. Yu, L. Yan, S. Lee, Y. Xie, M. Hauer, Z. Pan, and A. E. Willner, Enhanced higher-order PMD compensation using a variable time delay between polarizations, in Proc. ECOC2000, vol. 2, Munich, Germany, 2000, pp. 4748. [8] H. Sunnerud, M. Karlsson, and P. A. Andrekson, Analytical theory for PMD-compensation, IEEE Photon. Technol. Lett., vol. 12, pp. 5052, Jan. 2000. [9] H. Blow, System outage probability due to first- and second-order PMD, IEEE Photon. Technol. Lett., vol. 10, p. 696, May 1998. [10] F. Heismann, Integrated-optic polarization transformer for reset-free endless polarization control, IEEE J. Quantum Electron., vol. 25, pp. 18981906, Aug. 1989. [11] M. Karlsson, H. Sunnerud, and P. A. Andrekson, A comparison of different PMD-compensation techniques, in Proc. ECOC2000, vol. 2, Munich, Germany, 2000, pp. 3335.

[12] M. Karlsson and J. Brentel, Autocorrelation function of the polarization mode dispersion vector, Opt. Lett., vol. 24, pp. 939941, July 1999. [13] H. Y. Pua, K. Peddanarappagari, B. Zhu, C. Allen, K. Demarest, and R. Hui, An adaptive first-order polarization-mode dispersion compensation system aided by polarization scrambling: Theory and demonstration, J. Lightwave Technol., vol. 18, pp. 832840, June 2000. [14] H. Rosenfeldt, C. Knothe, R. Ulrich, E. Brinkmeyer, U. Fieste, C. Schubert, J. Berger, R. Ludwig, H. G. Weber, and A. Ehrhardt, Automatic PMD compensation at 40 Gbit/s and 80 Gbit/s using 3-dimensional DOP evaluation for feedback, in Proc. OFC2001, Anaheim, CA, 2001, PD27. [15] M. Karlsson, J. Brentel, and P. A. Andrekson, Long-term measurement of PMD and polarization drift in installed fibers, J. Lightwave Technol., vol. 18, pp. 941951, July 2000. [16] M. Karlsson, C. Xie, H. Sunnerud, and P. A. Andrekson, Higher order polarization mode dispersion compensator with three degrees of freedom, in Proc. OFC2001, Anaheim, CA, 2001, MO1. [17] C. Francia, F. Bruyre, J.-P. Thiry, and D. Penninckx, Simple dynamic polarization mode dispersion compensator, Electron. Lett., vol. 35, pp. 414415, Mar. 1999. [18] M. Karlsson, Polarization mode dispersion-induced pulse broadening in optical fibers, Opt. Lett., vol. 23, pp. 688690, May 1998. [19] H. Sunnerud, M. Karlsson, and P. A. Andrekson, A comparison between RZ and NRZ data formats with respect to PMD-induced system outage probability, IEEE Photon. Technol. Lett., vol. 13, pp. 448450, May 2001. [20] G. P. Agrawal, Fiber-Optic Communication Systems. New York: Wiley, 1992. [21] C. J. Anderson and J. A. Lyle, Technique for evaluating system performance using Q in numerical simulations exhibiting intersymbol interference, Electron. Lett., vol. 30, pp. 7172, Jan. 1994. [22] H. Blow, Limitation of optical first-order PMD compensation, in Proc. OFC99, San Diego, CA, 1999, WE1, pp. 7476. , PMD mitigation techniques and their effectiveness in installed [23] fiber, in Proc. OFC00, Baltimore, MD, 2000, ThH1, pp. 110112.

Henrik Sunnerud, photograph and biography not available at the time of publication.

Chongjin Xie (M02), photograph and biography not available at the time of publication.

Magnus Karlsson, photograph and biography not available at the time of publication.

Robert Samuelsson, photograph and biography not available at the time of publication.

Peter A. Andrekson (S84M88), photograph and biography not available at the time of publication.