Interferometric all-optical switches for ultrafast signal processing

Transcription

1 Interferometric all-optical switches for ultrafast signal processing Naimish S. Patel, Katherine L. Hall, and Kristin A. Rauschenbach All-optical switches find applications in ultrahigh-speed network user interfaces and in specialized high-speed processors, such as data regenerators and encryptors and microwave signal generators. We describe a semiconductor optical-amplifier-based single-arm interferometric switch called the ultrafast nonlinear interferometer. We discuss how the gain and the refractive-index nonlinearities in semiconductor optical amplifiers have an impact on the all-optical switch design, and we review experimental results obtained with the ultrafast nonlinear interferometer Optical Society of America OCIS codes: , , Introduction All-optical switches and logic gates have been studied for many years, primarily because of their potential for high-speed operation. High-speed time-division multiaccess networks would clearly benefit from the availability of high-speed all-optical switches that could perform essential network functions such as clock recovery and buffering at the optical bus data rate. 1,2 In addition to network applications, specialized processors requiring a limited number of optical gates, such as data regenerators and encryptors and microwave signal generators, could be feasible in the near term. Of the optical switches that have been demonstrated, those that use optically induced nonlinearities in the transmission and the refractive index of semiconductor optical amplifiers SOA s are the most promising because they have the potential to be integrated and someday mass produced. However, the bit-by-bit switching speed of SOA-based devices has been limited to 40 Gbits s. 3,4 In this paper we discuss optically induced nonlinearities in SOA s and describe how these nonlinearities can be exploited in various optical switch designs. We concentrate our discussion on interferometric switches because they are the most versatile. In When this study was performed the authors were with the MIT Lincoln Laboratory, C-237, 244 Wood Street, Lexington, Massachusetts. N. S. Patel is currently with Sycamore Networks, 1000 Winter Street, Suite 4610, Waltham, Massachusetts Received 11 August 1997; revised manuscript received 25 September $ Optical Society of America particular, we describe a single-arm ultrafast nonlinear interferometer UNI that has been demonstrated at the record speed of 40 Gbits s, and we discuss ways of extending the operating speed to 100 Gbits s. As is the case for electronic logic gates, all-optical switches fundamentally rely on nonlinearities. Fortunately a multitude of nonlinear optical mechanisms are at one s disposal for use in switches. Most switching demonstrations to date have relied on nonlinearities that can be characterized by second- and third-order susceptibility tensors, 2 and 3, respectively. Parametric processes such as sum and difference frequency generation 5 and four-wave mixing 6 8 based primarily on inefficient but fast intraband dynamics have been used to obtain AND functionality since these mechanisms essentially perform temporal correlation. The focus of this paper is on taking advantage of cross-phase modulation XPM and crossgain saturation XGS in conjunction with interferometric structures to obtain a variety of logic functions. In terms of nonlinear materials that have been studied in the context of optical switches, silica fiber has been the most widely demonstrated and, as a result, is the most well understood. Many switching demonstrations have been performed with the nonlinear optical loop-mirror NOLM geometry 9 15 with a fiber nonlinearity. Fiber has the advantage in that its 3 nonlinearity is composed primarily of bound-electron processes that result in very short relaxation times of tens of femtoseconds. This fact leads to the potential of fiber s providing very high switching rates, possibly exceeding terabits per second. The disadvantages of pure silica fiber, however, are that its nonlinearity is weak and long interaction lengths are necessary to achieve reasonable switching energies. 10 May 1998 Vol. 37, No. 14 APPLIED OPTICS 2831

2 A by-product of the long interaction length is that group-velocity dispersion, compounded by the spectral broadening effects of self-phase modulation, can result in severe distortion of short pulses, thereby imposing practical limits on achievable data rates. Operating one or both input signals in the anomalous dispersion regime, taking advantage of soliton effects, may alleviate this problem to a small degree. Still, the long interaction length of fiber has two inevitable practical consequences: large physical size and microsecond propagation delays. Although a recently developed chalcogenide-doped fiber with large 3 nonlinearities has been used successfully in demultiplexing experiments, 16 the required interaction lengths are still of the order of meters, corresponding to nanosecond propagation delays. Ideally switch delays should be comparable with the bit period, e.g., picoseconds for bit rates in excess of 100 Gbits s. Ultimately optical switches will be feasible in larger systems only if they can be reliably and repeatably manufactured. This fact has motivated much research into using semiconductor nonlinearities for optical switching. Typically, relevant semiconductor nonlinearities are 4 orders of magnitude larger than those in fiber and the required interaction lengths are of the order of millimeters. This comparatively small interaction length in conjunction with monolithic integration techniques can realize potentially huge gains in manufacturability and reliability. Semiconductor materials are not without their drawbacks, however. In addition to subpicosecond components of the semiconductor nonlinearity resulting from two-photon absorption, carrier heating, and virtual electronic processes, there are customarily free-carrier components with relaxation times of tens of picoseconds to milliseconds. Comprehensive pump probe studies have characterized the various nonlinear mechanisms in InGaAsP waveguides. 17 On time scales of hundreds of femtoseconds, carrier heating is dominant, whereas free-carrier recombination and carrier diffusion across the active region of the device are important on time scales of 10 to 1000 ps. In addition, thermal effects may have an impact on device performance on time scales as long as milliseconds, but these effects have yet to limit the performance of all-optical switching devices. The primary effect that limits the speed of SOAbased all-optical switches is the free-carrier recombination time. At high carrier densities the carrier lifetime may be reduced to a few tens of picoseconds by nonradiative recombination mechanisms, such as Auger recombination, 18 and by the presence of optical holding beams. 19 Still, switches based on these freecarrier effects will be limited to operating rates of approximately 100 Gbits s in the best case. This limitation can potentially be overcome by use of semiconductor materials with primarily ultrafast nonlinearities, such as passive semiconductor waveguides. Alternatively, it can be circumvented by use of innovative optical switch geometries, as are described in this paper. One of the first demonstrations of all-optical switching was conducted by Lattes et al., 20 who used an integrated Mach Zehnder interferometer. In this demonstration, a clock stream entered the device and was split into two components, which traveled along distinct paths and subsequently were interferometrically recombined at the output of the device. Two logical inputs were coupled into the two arms and were used to gate the clock stream. Varying functionality was obtained by appropriate biasing of each arm. As an application, a pseudorandom bit-stream generator based on a XOR gate with feedback was proposed. For all-optical switching Mach Zehnder and other interferometric switches require at least two optical input streams. Typically, one stream is a low-power clock stream that is split into the two arms at the input to the interferometer and is recombined at the output. This low-power stream is often referred to as the signal stream. The second stream is a high-power stream that is used to induce transmission and phase changes in the nonlinear material. This high-power stream is often referred to as the control stream. In some cases the signal and the control streams are the two logical inputs to the optical logic gate, while in other cases two control streams are the logical inputs and the signal is an optical clock stream. Recent switch demonstrations have used interferometric structures encompassing materials with refractive-index, gain, or both refractive-index and gain nonlinearities. These devices can exhibit nonlinear transmission characteristics with high contrast ratios that stand-alone nonlinear materials typically cannot achieve. Note, however, that gain differentials between distinct interferometric light paths lead to output contrast-ratio degradation because full destructive interference becomes impossible. The latter is simple to verify if we note that, for a Mach Zehnder interferometer with a splitter and combiner ratios with unity field gain in one arm and a multiplicative field gain of g in the second arm, the ratio of the output intensity to the input intensity is I out 1 I in 4 1 g2 2g cos, (1) where is the accrued phase shift between each arm. Equation 1 achieves a minimum of g 2 when. Clearly full destructive interference is possible only when g 1, namely, when no gain differential exists between the two arms. Thus, to extract the greatest value from an interferometric device geometry, one should maximize the relative magnitude of phase i.e., reactive nonlinearities with respect to gain nonlinearities. Unfortunately, this goal is typically difficult to attain in the case of active semiconductor nonlinearities, in which the gain and the refractive index are related by a linewidth enhancement factor Simulations An additional problem associated with using semiconductor nonlinearities in a two-arm interferometer optical switch geometry is related to components of 2832 APPLIED OPTICS Vol. 37, No May 1998

3 the nonlinear response that relax on time scales longer than the bit period, as alluded to in Section 1. More specifically, the control stream induces carrierdensity changes in the semiconductor with recombination times of up to a few nanoseconds. Thus, if the bit period of the control stream is less than the freecarrier recombination time, the phase and the gain modulations imposed on one of the signal components are pattern dependent, which leads ultimately to fluctuating output-signal-pulse intensities. Stated another way, the value of a particular output bit no longer depends exclusively on its associated inputsignal and control bits but rather on the exact pattern of control bits over a time approximately equal to the carrier recombination time, i.e., a form of nonlinear intersymbol interference. We have performed simulations of switching based on unbalanced geometries, in which unbalanced refers to any geometry susceptible, in the manner above, to long-lived nonlinearities. For the purposes of simulation we adopt a model of the semiconductor nonlinearities described in Ref. 17, in which the relevant nonlinear response is characterized by a weighted sum of exponential responses with various time constants. Mathematically the impulse response is represented by 17 h t a i exp t i. (2) i The total nonlinear response resulting from a particular control stream I c t is then obtained by convolution of h t with I c t. One may raise the issue that using linear superposition for an inherently nonlinear process is invalid. However, as long as the optical intensity of the control stream does not saturate the physical nonlinear mechanism, Eq. 2 is valid. Finally, the output-signal intensity I out t is obtained by multiplication of the input-signal intensity I s t by the standard interferometer expression including the phase, gain, or both phase and gain modulation induced by the control stream. It is implicitly assumed that I s t is sufficiently small that it does not induce significant refractive-index changes in the nonlinear material. Furthermore, we assume that the interferometer is initially biased OFF so that an output pulse should appear only in the presence of both a signal and a control pulse, i.e., the AND operation. Therefore the output stream is computed by means of I out t g2 t 2g t cos t, g t h g I c t I c h g t d, (3a) (3b) Fig Gbit s AND gate simulations for unbalanced and balanced switches. Signal and control are both 40-Gbit s pseudorandom bit streams generated with 12.5-ps pulses. The unbalanced output corresponds to a Mach Zehnder interferometer output, while the balanced output corresponds to the single-arm interferometer output. Note the contrast between the distorted unbalanced output stream and the uniform balanced output stream. t h I c t I c h t d, (3c) where h g t and h t are the one-sided causal impulse responses for the gain and the phase nonlinearities, respectively. These gain and phase responses are related through the Kramers Krönig relations. For the current purposes we assume a value of g t 1 and consider the effects of only long-lived refractive-index nonlinearities. We model the nonlinear phase response as a sum of three components. The first two are an instantaneous bound-electron component and a carrierheating component with a relaxation time of 600 fs. These are the nonlinear mechanisms we would like to exploit in switching. The third is a carrierrecombination component with a relaxation time of 100 ps. This long-lived response is the unavoidable effect of optically induced carrier-density changes. Last, we choose the peak intensity of an isolated control pulse such that it induces a phase shift of over the interaction length of the semiconductor. Figure 1 shows the simulation of the induced index change and the 40-Gbit s AND gate functions of a traditional interferometer and a balanced interferometer described below. The pulse width is 12.5 ps and assumed to be Gaussian. Portions of the resulting patterns are displayed. Note that long run lengths of 1 s in the control stream compound the effects of long-lived refractive-index nonlinearities, leading to peak phase shifts exceeding and output-pulse distortion. Considerably more detrimental to switch performance, however, is the fact, shown in the fourth trace, that output pulses appear in time slots where they should not because of nonzero phase shifts, even in the absence of a control pulse. This is the result of remnants of the long carrier-recombination time in the traditional, Mach Zehnder-type unbalanced interferometer. The practical consequence of this phenomenon is a reduction in the contrast ratio of the output bit stream and a commensurate increase in the biterror rate BER produced by the device. Note that 10 May 1998 Vol. 37, No. 14 APPLIED OPTICS 2833

4 Fig. 2. Block diagram of a SAI: PSD s, polarization-sensitive delays; NLM, nonlinear medium. A filter is required at the output of the device for extinguishing the control stream. one may attempt to reduce the probability that phase shifts exceed by prebiasing the interferometer by a phase shift equal to the average dc value of the phase modulation induced by a random control stream. Although this does, in fact, mitigate the problem on average, transient bit errors will still occur when the input control stream departs from average behavior. 3. Single-Arm Interferometer A device geometry that circumvents the effects of long-lived refractive-index nonlinearities is the single-arm interferometer SAI, originally developed in the context of femtosecond pump probe studies of various nonlinear processes in semiconductor waveguides. Figure 2 shows a block diagram of the SAI. An input-signal pulse enters the device and is split, by a polarization-sensitive delay element, into orthogonal polarizations separated by approximately a pulse width. These two signal components travel through a nonlinear medium in which a control pulse, coupled in by a splitter, is temporally coincident with the lagging signal pulse. The two signal components are retimed to overlap in a second polarization-sensitive delay and are subsequently interfered by use of a polarizer set at 45 with respect to the orthogonal signal polarizations. The control pulse is filtered out at the output of the device. One advantage of this geometry is interferometric stability, since all signals travel along the same path and thus are exposed to identical optical length variations, if any. In fact, throughout our experiments with the SAI we have found that active interferometer bias stabilization is not required. The more important characteristic of the SAI from the perspective of optical switching, however, is immunity to longlived refractive-index changes in the nonlinear medium. More specifically, since both signal components traverse the nonlinear medium, they each accrue phase changes that occur over time scales longer than twice the temporal pulse width, whereas only the lagging signal pulse experiences the ultrafast components of the refractive-index nonlinearity induced by the control pulse. Thus ultrafast differential phase modulation between orthogonal signal components can be achieved with the SAI. This is precisely the dynamic equalization of long-lived nonlinear responses referred to above. We have performed simulations, shown in Fig. 1 balanced case trace, of switching by using balanced devices like the SAI. For purposes of comparison we assumed the same nonlinear refractive-index impulse response that was used for the unbalanced devices. In addition, we explicitly accounted for the small differences in long-lived phase shifts experienced by each orthogonal signal component as a result of their 12.5-ps temporal offset. As can be seen from the wellbehaved output stream of the balanced device, this is indeed a very small effect. The SAI also possesses a beneficial feature concerning long-lived gain nonlinearities. As discussed above, any gain differentials between signal components in the Mach Zehnder interferometer brought about, for example, by XGS in one arm lead to a reduction in the achievable extinction ratio. This is particularly important in the case in which a good OFF signal is desired, as is the case for an inverter, because we saw that the ratio of the output to the input intensities is lower bounded by g 2, where g is the ratio of the gains in each arm. Thus, since long-lived gain modulation is associated with longlived phase modulation, the Mach Zehnder interferometer will suffer from an extinction ratio that depends on the bit pattern of the control stream. Potentially more serious is the fact that the dc value of the gain modulation results in a timeaveraged reduction in the extinction ratio. In the case of the SAI, on the other hand, each signal component experiences the same long-lived gain modulation, leading simply to a pattern-dependent reduction in the maximum output intensity but not a pattern-dependent output extinction ratio. Note, however, that the ultrafast components of the gain nonlinearity still do have an impact on the SAI since they impart differential gain changes between the two signal components. We have performed simulations of the effect of gain modulation on the SAI. As above, we generated 40-Gbit s pseudorandom signal and control streams by using 12.5-ps Gaussian pulses. We chose a gain nonlinearity with the same time constants as the phase nonlinearity used above. Moreover, the ratio of the phase to the gain weighting coefficients was chosen such that a phase shift of induced by an isolated control pulse corresponded to a reduction in gain of 3 db, which corresponded to a 2834 APPLIED OPTICS Vol. 37, No May 1998

5 Fig. 4. Block diagram of the UNI: PSD s, polarization-sensitive delays; NLM, nonlinear medium. Note the absence of a filter at the output. Fig Gbit s AND gate simulation of the balanced device with gain modulation. Signal and control are both 40-Gbit s pseudorandom bit streams generated with 12.5-ps pulses. Output is the balanced output stream of the simulated AND gate. Note that the gain nonlinearity imparts a pattern-dependent amplitude modulation on the output stream. linewidth enhancement factor of approximately eight. We compute the output intensity with I out t I s t g t g t T 2 t t T 2 cos I s t g t g t T 2 2, (4) where T 12.5 ps is the temporal separation between orthogonal signal polarization components. Note that the last term in Eq. 4 represents a transient interferometer bias offset. Figure 3 displays the simulated output of the SAI for a 40-Gbit s AND operation with 12.5-ps pulses when these effects are included. Note that, when the interferometer is biased OFF, the effect of gain modulation is simply to impart a pattern-dependent amplitude to the output stream, but the quality of the OFF signal is unaffected. In other words, despite the long-lived gain nonlinearity, a pulse will still not appear at the output of the SAI unless both a signal and a control pulse are present in the same time slot, obviously a key requirement for proper AND operation. 4. Balanced Interferometers It should be noted that the SAI is not the only balanced geometry. The Mach Zehnder geometry can be modified to include two judiciously placed identical nonlinear media, one in each arm, such that the longlived nonlinear refractive-index responses are canceled at the output. 28 The difficulty with such a device, however, is that its performance is critically dependent on not only the precise placement of the nonlinear medium but also the extent to which the two nonlinear media exhibit identical responses. As such, manufacturing this device in a repeatable fashion may present obstacles. Another balanced device is the NOLM with an asymmetrically placed nonlinear medium. 29,30 Demultiplexing of 160-Gbit s bit streams at a frame rate of 10 Gbits s has been performed with the semiconductor NOLM. 31 Note that, in this case, the nonlinearity is perturbed on time scales equal to the inverse frame rate 100 ps and not at the bit period of the aggregate stream. Certain functional requirements are fundamental from the perspective of building combinatorial logic circuits. One important such feature is cascadability, that is, an output of a switch should be able to feed directly the signal or the control port or both with the use of a splitter of a subsequent switch. More important is the fact that cascadability should be attainable without the need to make specific changes to individual switches in the circuit design process. The latter is of primary importance in easing the design of combinatorial optical logic circuits. This requirement has many implications for the architecture of optical switches. First, the switch performance should be highly insensitive to the wavelength of the signal or the control, at least over the entire range of anticipated input wavelengths. Moreover, switching with degenerate signal and control wavelengths must also be possible to allow fan-out to both signal and control ports of other switches. It is clear that optical switches that use filters of any kind will not meet the above requirements. Consequently, none of the switches described thus far are cascadable. 5. Ultrafast Nonlinear Interferometer Adopting a simple modification to the SAI allows cascadability to be attained. In particular, one may couple the control pulse into the nonlinear medium in a counterpropagating configuration, as shown in Fig. 4, thus obviating the use of a filter at the output. We call this device the UNI. It is interesting to note that with cascadability comes no performance degradation, as long as the signal- and the control-pulse widths are comparable with the transit time through the semiconductor nonlinearity. If the nonlinear medium is longer than the spatial widths of the pulses, then a single control pulse can potentially interact with multiple signal pulses. Alternatively, if the nonlinear medium is shorter than the spatial pulse widths, a loss in switching efficiency is incurred because the effective interaction length is limited by the physical interaction length. To date, most all-optical switching experiments have used semiconductor optical amplifiers with 10 May 1998 Vol. 37, No. 14 APPLIED OPTICS 2835

6 Fig. 5. Implementation of the UNI. The output PSI is used to interfere the orthogonal signal-pulse components. The 12 angle polish is required for preventing backreflection of the control stream into the output port. lengths of approximately 1 mm, commensurate with the record 40-Gbit s switching demonstrations. At higher rates the length of the nonlinear semiconductor element must be reduced. The same nonlinear phase shift may be induced in shorter devices by use of higher-power switching pulses and by an increase in the spatial overlap of the optical mode and the active region of the device. With these changes, alloptical switching at data rates in the range of hundreds of gigabits per second is feasible. Finally, a favorable by-product of a counterpropagating controlpulse geometry is that, as the ratio of the pulse width to the transit time through the nonlinear medium decreases, the switching window becomes more squarelike. In the copropagating geometry, on the other hand, the shape of the switching window resembles the control-pulse shape. A. Demultiplexing To demonstrate that the copropagating and counterpropagating versions of the UNI yield similar switching capabilities, we performed demultiplexing of 20- and 40-Gbit s data streams at a frame rate of 10 Gbits s. 32 A diagram of the UNI constructed for this experiment is shown in Fig. 5. The polarizationsensitive delays were implemented with 7.5 m of birefringent fiber BRF corresponding to a walk-off between orthogonal signal polarizations of 12.5 ps half the bit period at 40 Gbits s. The signal components were interfered at the output with a polarization-sensitive isolator PSI. An additional PSI was used at the input of the UNI to maintain a linear polarization for signal pulses entering the BRF. The control pulse was coupled into the nonlinear material in either a copropagating or a counterpropagating direction. Recall that, in the former case, an optical bandpass filter is required at the output of the UNI for eliminating the control pulse. Finally, a 1-mm-long commercially available antireflection-coated multi-quantum-well SOA with a gain peak of approximately 1500 nm was used as the nonlinear medium. The experimental setup is Fig. 6. Experimental setup for demultiplexing with the UNI: ML-ECL s, mode-locked external-cavity lasers; PRBS, pseudorandom bit stream; EDFA s, erbium-doped fiber amplifiers. The electro-optic modulator is a commercially available, differentially driven lithium niobate modulator. The UNI is biased OFF and configured as an AND gate. The output of the UNI is detected with a 45-GHz 3-dBbandwidth photodiode and displayed on a 50-GHz 3-dB-bandwidth digital sampling oscilloscope APPLIED OPTICS Vol. 37, No May 1998

7 Fig. 7. Oscilloscope trace of 40-Gbit s demultiplexing. The 40- Gbit s data and the 10-Gbit s clock streams enter the signal and the control ports of the UNI, respectively. depicted in Fig. 6. To generate the signal and the control streams, we used mode-locked external-cavity semiconductor lasers generating 8-ps pulses at a repetition rate of 10 GHz. The output of one laser was electro-optically modulated with a lithium niobate modulator driven by a 10-Gbit s pseudorandom bit stream with a word length of This modulated stream was then passively multiplexed, with an optical bit interleaver, to rates of 20 and 40 Gbits s before it entered the signal port of the UNI. The output of the other laser was amplified in an erbiumdoped preamplifier before it entered the control port of the UNI. The output of the UNI was detected with a 45-GHz, 3-dB-bandwidth photodiode and subsequently displayed on a 50-GHz, 3-dB-bandwidth digital sampling oscilloscope in addition to being presented to the BER tester. For the copropagating geometry, the signal and the control wavelengths were 1562 and 1565 nm, respectively. For the counterpropagating case, the signal and the control wavelengths were both chosen to be 1560 nm to demonstrate the important degenerate-wavelength case. The average control power entering the UNI was mw in both cases. Figure 7 shows an oscilloscope trace of 40-Gbit s demultiplexing with the counterpropagating geometry. As expected, the control stream samples every fourth pulse of the data stream. Also note the uniformity of the demultiplexed output stream. This implies that the carrier-recombination time is less than the 100-ps frame period. Figure 8 shows the BER curves for the copropagating and the counterpropagating cases. Note that, for 40-Gbit s demultiplexing, the BER curves for the copropagating and the counterpropagating cases are essentially equivalent, and an approximately 3-dB power penalty is incurred for both geometries. This indicates that no significant performance difference is observed. We believe that the power penalty is associated with amplified spontaneous emission in the SOA. We have found that the switching performance of the UNI is relatively insensitive to the polarization of the control pulses, allowing output-port to control- Fig. 8. BER measurements for demultiplexing operation: a Copropagating control-pulse geometry. b Counterpropagating control-pulse geometry. The baseline measurements are obtained from the 10-Gbit s data stream before being passively multiplexed to 20 and 40 Gbits s. The power penalty for demultiplexing a 10-Gbit s stream from a 40-Gbit s stream is approximately 3 db in both device geometries. port cascadability without the use of polarizationalignment devices between switches. The level of insensitivity to control polarization depends primarily on the degree of anisotropy of the steady-state semiconductor gain and on the anisotropy of the 3 nonlinearity responsible for XPM and cross-gain modulation. For example, in materials like optical fiber, in which the 3 nonlinearity is essentially instantaneous, there is a factor of 3 difference between the XPM induced by parallel-polarized pulses versus orthogonally polarized pulses. 33 However, if the 3 nonlinearity is due primarily to free-carrier effects, 10 May 1998 Vol. 37, No. 14 APPLIED OPTICS 2837

8 Fig. 9. Experimental setup for CSR experiment: S, signal port; O, output port; C, control port; EDFA, erbium-doped fiber amplifier. such as carrier heating, there may be no polarization dependence of the nonlinearity. 34 In these alloptical switching experiments, the optical pulses are typically 10 ps in duration and primarily induce carrier-heating effects. 17 Therefore we do not expect the induced nonlinear gain and phase changes to be polarization dependent. However, a slight dependence of the switching energy on the control-pulse polarization may be attributed to polarizationdependent coupling to the SOA or other components in the control arm of the UNI. B. Circulating Shift Register with an Inverter To demonstrate the cascadability of the UNI explicitly, we have performed a circulating shift register CSR experiment by using the UNI as a regenerative switch, 34 the experimental setup for which is shown in Fig. 9. The output and the control ports of the UNI are connected to form a feedback loop, while the signal port of the UNI is fed by a 40-GHz pulse train generated by a soliton compression source. 35 Furthermore, the UNI is biased ON and configured as an inverter. The feedback loop contains a variable delay stage to adjust the timing of output pulses arriving at the control port, an erbium-doped fiber amplifier EDFA, a filter to remove the amplified spontaneous emission introduced by the EDFA and the SOA within the UNI, and finally a coupler to observe the recirculating data. The control-pulse energy required for switching the state of the UNI in this experiment was approximately 100 fj at the input to the SOA. Initially there are no control pulses in the feedback loop, and the signal pulses are transmitted by the UNI. When the output pulses are fed back into the control port of the UNI, they impart phase and gain modulation to the 40-GHz signal stream pulses, thereby turning them OFF at the output of the UNI. Thus one would expect to see a repeating pattern consisting of a block of 1 s and a block of 0 s, each of a length equal to the feedback delay time of the CSR. Indeed, this is what is observed in Fig. 10, which shows the output of the CSR as displayed on a 1-GHzbandwidth analog oscilloscope. It is evident that the feedback time is approximately 142 ns, corresponding to the loop length of 28 m. The CSR was stable, storing the alternating pattern for hours at a time, corresponding to billions of circulations. Temperature, length, or both temperature and length stabilization was not required. C. AND and INVERT Operations A second functional requirement for optical switches is logical versatility: the ability of a single device to perform a suite of logic functions with virtually no change in its fundamental design. We have used the counterpropagating UNI to perform bitwise AND and INVERT operations at 40 Gbits s in addition to OR and NOR operations at 10 Gbits s. 3 The experimental setup for AND and INVERT is shown in Fig. 11. The signal port of the UNI was fed by a 40-Gbit s clock stream CLK of 4-ps pulses at 1545 nm generated by a soliton compression source. A 40-Gbit s pseudorandom stream A at 1560 nm entered the control port of the UNI. The required switching energy per bit of the control stream was less than 2 pj. Figure 12 shows an oscilloscope trace of the 40-Gbit s clock stream and results of A CLK and A. It is implicit that we additionally have performed noninverting and inverting wavelength conversions from 1560 to 1545 nm in this experiment. Note that the 45-GHz, 3-dB bandwidth of the photodiode is limiting the resolution of the bit streams, as is particularly evident in the top trace, which actually represents a stream of 4-ps pulses. Moreover, ringing in the detector is likely the cause of undulations on the bottom two traces. It is evident, however, that gain dynamics are affecting the unifor- Fig. 10. Analog oscilloscope trace of CSR output APPLIED OPTICS Vol. 37, No May 1998

9 Fig. 11. Experimental setup for 40-Gbit s AND and INVERT operations notation is the same as that of Fig. 6. The output of the UNI is detected in a 45-GHz-bandwidth photodiode and displayed on a 50-GHz-bandwidth digital sampling oscilloscope. mity of the output stream. As mentioned previously, the carrier-recombination time associated with the gain transient recovery may be reduced by use of an out-of-band holding beam to increase the rate of stimulated emission 19 or larger SOA bias currents to increase the Auger recombination rate. 18 To ascertain the effects on the output extinction ratio of transient gain differentials between signal components induced by ultrafast gain nonlinearities, we performed an inverting operation at 10 Gbits s, a rate that enabled us to segregate the issue of patterning effects brought about by long-lived gain nonlinearities from that of transient differential gain modulation. Figure 13 shows an oscilloscope trace of the inverter operation. It is clear that a good OFF signal can be obtained despite the transient differential gain modulation. This assertion is corroborated by the observed 2.9-dB power penalty for the inverting operation shown in Fig. 14. The latter would seem to indicate that the weighting coefficient for the ultrafast components of the gain nonlinearity is small compared with the free-carrier contributions since long-lived patterndependent gain modulation is apparent in both the 40-Gbit s and the 10-Gbit s experiments. A pump probe study of the relevant nonlinearities would be required for confirming this conjecture. D. OR and NOR Operations As mentioned in the introduction to Section 5, we have augmented the set of logic functions achievable with the UNI to include OR and NOR by making a minor modification to the switch geometry, as shown in Fig. 15. Instead of one control stream s entering the UNI, we couple two logical control streams, A and B, into the SOA by using an additional splitter. When a combination of XGS Fig. 12. Oscilloscope traces of 40-Gbit s AND and INVERT operations. The clock stream CLK and the data stream A enter the signal and the control ports of the UNI, respectively. In the case of the AND operation, the UNI is biased OFF. In the case of INVERT, the UNI is biased ON. Fig. 13. Oscilloscope traces of the 10-Gbit s INVERT operation. The clock stream and the data stream enter the signal and the control ports of the UNI, respectively. In the case of the AND operation, the UNI is biased OFF. In the case of INVERT, the UNI is biased ON. The good extinction ratio of the inverted stream suggests that transient interferometer bias offset caused by ultrafast gain nonlinearities is not a serious problem. 10 May 1998 Vol. 37, No. 14 APPLIED OPTICS 2839

10 Fig. 16. Oscilloscope traces of OR and NOR operations. The A and the B streams are coupled into the control port of the UNI while a clock stream enters the signal port. Note the relative uniformity of the output-pulse amplitudes. Fig. 14. BER measurement for 10-Gbit s demultiplexing. Approximately a 3-dB power penalty is incurred for the INVERT operation, corroborating the assertion that the degree of differential gain modulation between orthogonal signal pulse components is small. and XPM is used, the switch acts as a logical OR gate if the interferometer is biased OFF and as a logical NOR gate when the interferometer is biased ON. The logical A and B streams are obtained by the splitting of the 10-Gbit s pseudorandom stream into two bit-phase-synchronized delayed paths before being coupled into the SOA. The average powers of the A and B streams were 26 and 24 mw, respectively. The oscilloscope trace in Fig. 16 shows the switching results. The degree of pulseamplitude uniformity and a reasonably high extinction ratio are somewhat surprising, given the complexity of the underlying physical mechanisms. In particular, the switched-out pulses in the OR case have similar amplitudes, irrespective of whether one or two control pulses were present. The most probable explanation for this concerns gain saturation within the SOA, namely, the energy of a single control pulse approximately 5 pj in this case is sufficient to saturate the gain of the SOA. Thus the temporal coincidence of two control pulses corresponding to the logical streams A and B has a relatively small additional effect on the nonlinear response because the combined intensity of the two control pulses within the SOA is only slightly larger than the intensity of a single pulse within the SOA. Last, the relatively good OFF state observed in the NOR trace again suggests that the magnitude of the transient gain modulation is small compared with that of long-lived gain modulation. The exact contributions of XPM and XGS in both the OR and the NOR cases are currently under investigation. It is important to note that the pulse energy required for switching the state of any interferometer based on SOA nonlinearities depends on many factors. For example, the length of the SOA, the saturated and the unsaturated gains of the SOA, the SOA Fig. 15. Implementation of OR and NOR gates based on the UNI. The 12 angle polish is required for preventing backreflection of the control stream into the output port APPLIED OPTICS Vol. 37, No May 1998

11 peak gain wavelength relative to the switching-pulse wavelength, and the confinement factor all affect the pulse energy required for inducing a nonlinear phase shift of. Throughout this text we have quoted the switching energy as the control energy coupled into the SOA, assuming a 50% coupling efficiency to the active region of the device. Typical switching-pulse energies range from hundreds of femtojoules to a few picojoules. 6. Summary In summary, we have described a stable, cascadable, and functionally versatile all-optical switch: the ultrafast nonlinear interferometer UNI. The switching experiments performed to date have indicated that the UNI has the potential for use in a variety of high-performance network and signal-processing applications. Scaling the bit rate of the UNI to 100 Gbits s should be possible if the signal wavelength is operated closer to the band edge of the semiconductor, where the relative effects of gain nonlinearities are small compared with refractive-index nonlinearities. The use of higher bias currents, higher optical intensities, or both to reduce carrier-recombination times would likely also be required. Scaling of the bit rate beyond 100 Gbits s demands a more detailed understanding of the subpicosecond nonlinearities present in semiconductor waveguides and their associated effects on switch design. Additionally, if a counterpropagating geometry is desired to obtain cascadability at high bit rates, the length of the SOA must be reduced to be commensurate with the signaland the control-pulse widths. To compensate for the consequent reduction in the nonlinear strength of the optical interaction in the SOA, one would need to operate the SOA at higher bias currents, use semiconductor waveguides with larger overlap integrals, or both. Although the UNI described in this paper was implemented with a BRF for the polarizationsensitive delays, full device integration could be achieved with polarization-sensitive couplers and delay lines. The inherently linear and symmetric geometry of the UNI should facilitate device integration. The authors gratefully acknowledge stimulating conversations with Erich P. Ippen and experimental assistance from Claudia I. Fennelly. Much of the research that is described in this paper was supported by the U.S. Advanced Research Projects Agency. Opinions, interpretations, conclusions, and recommendations are those of the authors and are not necessarily endorsed by the U.S. Air Force. References 1. R. A. Barry, V. W. S. Chan, K. L. Hall, E. S. Kintzer, J. D. Moores, K. A. Rauschenbach, E. A. Swanson, L. E. Adams, C. R. Doerr, S. G. Finn, H. A. Haus, E. P. Ippen, and W. S. Wong, All-optical network consortium-ultrafast TDM networks, J. Lightwave Technol. 14, K. L. Hall, K. A. Rauschenbach, S. G. Finn, R. A. Barry, N. S. Patel, and J. D. Moores, 100 Gbit s optical network technology, in Ultrafast Electronics and Optoelectronics, M. Nuss and J. Bowers, eds., Vol. 13 of OSA Trends in Optics and Photonics Series Optical Society of America, Washington, D.C., 1997, pp N. S. Patel, K. L. Hall, and K. A. Rauschenbach, 40-Gbit s cascadable all-optical logic with an ultrafast nonlinear interferometer, Opt. Lett. 21, S. L. Danielson, C. Joergensen, M. Vaa, B. Mikkelsen, K. E. Stubkjaer, P. Doussiere, F. Pommerau, L. Goldstein, and M. Gois, Bit error rate assessment of a 40 Gb s all-optical polarization independent wavelength converter, Electron. Lett. 32, H. Takara, S. Kawanishi, T. Morioka, K. Mori, and M. Saruwatari, 100 Gbit s optical waveform measurement with 0.6 ps resolution optical sampling using subpicosecond supercontinuum pulses, Electron. Lett. 30, P. A. Andrekson, Picosecond optical sampling using four wave mixing in fibre, Electron. Lett. 27, T. Morioka, S. Kawanishi, H. Takara, and M. Saruwatari, Multiple-output, 100 Gbit s all-optical demultiplexer based on multichannel four-wave-mixing pumped by a linearlychirped square pulse, Electron. Lett. 30, S. Kawanishi and O. Kamatani, All-optical time division multiplexing using four-wave-mixing, Electron. Lett. 30, N. J. Doran and D. Wood, Nonlinear-optical loop mirror, Opt. Lett. 13, P. A. Andrekson, N. A. Olsson, J. R. Simpson, D. J. DiGiovanni, P. A. Morton, T. Tanbun-Ek, R. A. Logan, and K. W. Wecht, 64 Gb s all-optical demultiplexing with a nonlinear optical loop mirror, IEEE Photon. Technol. Lett. 4, M. Jinno and T. Matsumoto, Ultrafast all-optical logic operations in a nonlinear Sagnac interferometer with two control beams, Opt. Lett. 16, K. L. Hall, K. A. Rauschenbach, E. A. Swanson, S. R. Chinn, and G. Raybon, Picosecond-accuracy all-optical bit phase sensing using a nonlinear optical loop mirror, IEEE Photon. Technol. Lett. 7, K. A. Rauschenbach, K. L. Hall, J. C. Livas, and G. Raybon, All-optical pulse width and wavelength conversion at 10 Gb s using a nonlinear optical loop mirror, IEEE Photon. Technol. Lett. 6, K. L. Hall and K. A. Rauschenbach, All-optical bit pattern generation and matching, Electron. Lett. 32, K. Uchiyama, H. Takara, S. Kawanishi, T. Morioka, M. Saruwatari, and T. Kitoh, 100 Gbit s all-optical demultiplexing using nonlinear optical loop mirror with gating-width control, Electron. Lett. 29, K. Uchiyama, T. Morioka, M. Saruwatari, M. Asobe, and T. Ohara, Error free all-optical demultiplexing using a chalcogenide glass fibre based nonlinear optical loop mirror, Electron. Lett. 32, K. L. Hall, G. Lenz, A. M. Darwish, and E. P. Ippen, Subpicosecond gain and index nonlinearities in InGaAsP diode lasers, Opt. Commun. 111, J. M. Wiesenfeld, Gain dynamics and associated nonlinearities in semiconductor optical amplifiers, Int. J. High Speed Electron. 7, R. J. Manning, D. A. O. Davies, D. Cotter, and J. K. Lucek, Enhanced recovery rates in semiconductor laser amplifiers using optical pumping, Electron. Lett. 30, A. Lattes, H. A. Haus, F. J. Leonberger, and E. P. Ippen, An ultrafast all-optical gate, IEEE J. Quantum Electron. 19, C. H. Henry, Theory of the linewidth of semiconductor lasers, IEEE J. Quantum Electron. 18, M. Osinski and J. Buus, Linewidth broadening factor in semi- 10 May 1998 Vol. 37, No. 14 APPLIED OPTICS 2841

12 conductor lasers an overview, IEEE J. Quantum Electron. QE-23, C. T. Hultgren and E. P. Ippen, Ultrafast refractive index dynamics in AlGaAs diode laser amplifiers, Appl. Phys. Lett. 59, M. Shirasaki, H. A. Haus, and D. L. Wong, Nonlinear fiber interferometer and logic gate, in Conference on Lasers and Electro-Optics, Vol. 14 of 1987 OSA Technical Digest Series Optical Society of America, Washington, D.C., 1987, paper THO M. J. LaGasse, D. Liu-Wong, J. G. Fujimoto, and H. A. Haus, Ultrafast switching with a single-fiber interferometer, Opt. Lett. 14, K. K. Anderson, M. J. LaGasse, C. A. Wang, and J. G. Fujimoto, Femtosecond dynamics of the nonlinear index near the band edge in AlGaAs waveguides, Appl. Phys. Lett. 56, M. J. LaGasse, K. K. Anderson, C. A. Wang, H. A. Haus, and J. G. Fujimoto, Femtosecond measurements of the nonresonant nonlinear index in AlGaAs, Appl. Phys. Lett. 56, K. I. Kang, T. G. Chang, I. Glesk, P. R. Prucnal, and R. K. Boncek, Demonstration of ultrafast, all-optical, low control energy, single wavelength, polarization independent, cascadable, and integratable switch, Appl. Phys. Lett. 67, M. Eiselt, W. Pieper, and H. G. Weber, SLALOM: semiconductor laser amplifier in a loop mirror, Electron. Lett. 29, I. Glesk, J. P. Sokoloff, and P. R. Prucnal, Demonstration of all-optical demultiplexing of TDM data at 250 Gbit s, Electron. Lett. 30, K. Suzuki, K. Iwatsuki, S. Nishi, and M. Saruwatari, Errorfree demultiplexing of 160 Gbit s pulse signal using optical loop mirror including semiconductor laser amplifier, Electron. Lett. 30, N. S. Patel, K. A. Rauschenbach, and K. L. Hall, 40 Gb s demultiplexing using an ultrafast nonlinear interferometer UNI, IEEE Photon. Technol. Lett. 8, Z. Vardeny and J. Tauc, Picosecond coherence coupling in the pump and probe technique, Opt. Commun. 39, K. L. Hall, J. P. Donnelly, S. H. Groves, C. I. Fennelly, R. J. Bailey, and A. Napoleone, 40-Gbit s all-optical circulating shift register with an inverter, Opt. Lett. 22, E. A. Swanson, S. R. Chinn, K. L. Hall, K. A. Rauschenbach, R. S. Bondurant, and J. W. Miller, 100 GHz soliton pulse train generation using soliton compression of two phase side bands from a single DFB laser, IEEE Photon. Technol. Lett. 6, APPLIED OPTICS Vol. 37, No May 1998