Experimental results are reported demonstrating non-line of sight short-range ultraviolet communication link losses, and performance of photon counting detectors, operating in the solar blind spectrum regime. We employ light emitting diodes with divergent beams, a solar blind filter, and a wide field-of-view detector. Signal and noise statistics are characterized, and receiver performance is demonstrated. The effects of transmitter and receiver elevation angles, separation distance, and path loss are included.
©2008 Optical Society of America
Ultraviolet (UV) communication has a rich history (e.g., see the recent survey in ). Relevant studies date back to the 1960s [2,3], mainly for outdoor communications. A pioneering experimental work focused on the characterization of a scattering-based non-line of sight (NLOS) UV link . Abundant atmospheric scattering of the UV radiation by molecules and aerosols provided a major mechanism for establishing a communications link by redirecting the transmitted signal towards a NLOS receiver. A high-power xenon UV flashtube was used in  as the light source, with the shortest wavelength at 280 nm. The receiver employed a photomultiplier tube (PMT) at a range of 26 km. In addition to this experimental work, an elegant analytical channel response model was developed to describe the temporal characteristics of scattered radiation in the middle UV (wavelength 200~300 nm) . This channel model was further extended to examine angular spectra and path losses [6,7]. Around the same time, a point-to-point NLOS UV system was also demonstrated based on an isotropic radiating mercury arc lamp at a modulation rate of 40 kHz . An improved UV local area network test-bed covering a 1 km range that used a collimated mercury-xenon lamp was constructed at an increased modulation rate (up to 400 kHz) and a wavelength of 265 nm . UV pulsed laser communication systems operating at 266 nm have also been reported [10,11], with rates of few hundred hertz.
All the above UV systems used flashtubes/lamps/lasers as light sources. These devices tend to be bulky, power hungry, or bandwidth limited. Semiconductor-based UV optical source technologies offer a potential for low cost, small size, low power, high reliability, and high bandwidth sources. State-of-the-art commercial (research-grade) deep UV light emitting diodes (LEDs) have recently become available . These include deep UV LEDs at peak wavelengths 247~365 nm and with a spectral width of less than 20 nm. A single LED typically consumes electrical power of 150 mW and radiates an average optical power of 1 mW. Although much less than commercial infrared LEDs that can produce up to tens of milli-watts for the same input power, improvements to the power output, efficiency and reliability can be expected [13,14].
To develop an effective UV communication transceiver that can operate under exposure to solar radiation, solar-blind UV detectors and filters with high sensitivity, gain, and out-of-band rejection are crucial. The requirements of high sensitivity and high gain make photomultiplier tubes (PMTs) and avalanche photodiodes (APDs) suitable candidates for such communication systems. Hamamatsu and PerkinElmer are major vendors for deep UV PMTs with off-the-shelf commercial products attaining high multiplication gains of 105~107, high responsivity of 62 A/W, large detection area of few cm2, reasonable quantum efficiency of η=15%, low dark count rate of few hertz, and low dark current of 0.1 nA/cm2. The response time is typically on the order of 20 ns. For solar blind applications, PMTs are typically combined with solar-blind filters resulting in an enhanced out-of-band rejection ratio of about 108. These features enable PMTs to detect very weak signals even in the presence of high-levels of background radiation, down to single photon counting resolution. Although PMTs are capable of providing the best performance of any commercial product, they tend to be fragile, somewhat bulky, costly, and sensitive to magnetic interference. They also require high voltage supplies, and need to be integrated with expensive external filters due to their limited out-of-band suppression capability. These PMT attributes inhibit low cost and compact designs.
Research in solid-state solar-blind deep UV APDs is also rapidly maturing. Emerging devices are potentially small, low cost, and most importantly, can be intrinsically made solar blind [15–20]. Thus far, the GaN-based work by Dupuis et al. has resulted in responsivity of 0.15A/W, gain of 104, and dark current of 100 nA/cm2, and the SiC-based work by Campbell et al. has yielded gain of 103, dark current of 64 nA/cm2, quantum efficiency of 45%, and single photon sensitivity. Results by many other researchers have also been reported, with excellent specifications in some categories [21–23]. For example, SiC APDs with a gain of 106 and outstanding visible rejection ratio of 108 have been experimentally demonstrated . DARPA’s recent Deep Ultraviolet Avalanche Photodetectors (DUVAP) program aims to demonstrate APD arrays operating in the UV band centered at 280 nm, with effective Geiger mode gain of 106, an effective aperture up to 1 cm2, field-of-view (FOV) up to 60 degrees, dark count rate below 10 kHz, and a solar rejection ratio exceeding 106. Except for the dark count rate, these goals are comparable to PMT performance.
The availability of deep UV LEDs, solar blind PMTs, filters, and primitive APDs has inspired recent work on LED-based UV system performance for applications from short range communications [24–31], to sensing , and wireless imaging , in either line of sight (LOS) or NLOS channels. In this paper we present an experimental study of solar-blind UV communication link outdoor performance. In particular, we discuss channel losses, received signal and channel characteristics, and communication system bit-error-rate (BER) performance. We examine the effect of communication range, transmitter and receiver elevation (pointing, or apex) angles, and signal to noise ratio (SNR) utilizing an experimental test-bed that was developed at the University of California, Riverside (UCR), and described in detail in the next section.
2. Ultraviolet communication test-bed and experimental conditions
Our experimental results are based on the solar blind UV communication test-bed as depicted in Fig. 1. All measurements were taken in an open field close to the UCR campus, under clear skies on a sunny day in February of 2008. The high and low temperatures were 6°C and 20°C respectively, wind speed was 5 miles per hour, and humidity was below 40%. The molecular distribution was relatively stable regardless of the temperature or pressure variation. The aerosol distribution was unknown, but could be described by an appropriately fitted model, such as a continental type . Our experimental measurements and analysis have not yet included the effect of different atmospheric conditions on the communication link and channel performance, but future studies on this aspect will be conducted and reported.
The transmitter used a waveform generator feeding binary sequences to current driver circuitry that powered an array of 7 ball-lens UV LEDs. Each LED received a driving current of 30 mA, yielding an average radiated optical power of 0.3 mW. The beam angular distribution was found to follow a superposition of multiple Gaussian functions with a full divergence angle of 10°. The LEDs (UVTOP250 with nominal center wavelengths of 250 nm) were mounted on a calibrated plate. At the receiver, either a PMT or APD detector may be used. For this study, we employed a solar-blind filter combined with a PMT for photon detection. The solar blind filter was placed in front of the sensing window of a PerkinElmer PMT module MP1922 (head-on window). The filter has a full-width half-maximum (FWHM) bandwidth of 15 nm with peak transmission of 10.4% at 255 nm. The spectral mismatch between the LED and the filter was found to be less than 30%. The PMT has a circular sensing window with a diameter of 1.5 cm, resulting in an active detection area of 1.77 cm2, and it has an average of 10 dark counts per second (10 Hz). The composite in-band UV transmission of PMT plus filter was found to be 1%. The detector’s effective FOV was estimated to be about 30°. The PMT output current was directed to a low noise amplifier followed by a photon counter unit. Note that some spectral mismatch loss between the LEDs and the filter was unavoidable due to practical device constraints. A mechanical module at Tx/Rx used two perpendicular rotation stages to achieve high-resolution angular control in both azimuth and zenith directions.
3. Experimental results
In the following, we extend some preliminary test results [31,35,36], and first characterize solar background noise and signal count distributions. Then we further link key system parameters such as path loss and communication BER under different transmitter (Tx) and receiver (Rx) geometries.
3.1 Solar irradiance and signal count distributions
Solar noise and signal count distributions are helpful for power budget calculations and system design. The dark count rate of the PMT was negligible when compared with the count rates due to solar radiation in the deep UV band and the received signal. Measurements of the maximum solar radiation noise counts were recorded as the PMT was aimed directly towards the sun at noon. The time interval for measurements was set to be 200 µs (a rate of 5 kHz). This value was set to achieve reasonably high signal levels (i.e., the number of signal counts) per pulse for a variety of test geometries. Each observation window was segmented into several time intervals. The received solar noise counts within each time window were then recorded. Measurements from dozens of time windows were used to obtain the distribution of the random noise photon counts. In Fig. 2, experimental results are compared against a best-fit Poisson distribution and a best-fit truncated Gaussian distribution. The Gaussian distribution is found to have a mean of 2.9 counts and a standard deviation of 1.8 counts per interval. This yields an average noise count rate of 14.5 kHz with standard deviation equal to 9 kHz. The Poisson distribution has the same mean. It is found that both fitting errors are below 2%, but the Poisson fit is somewhat better. However, for simplicity of communication performance analysis, we adopt a Gaussian distribution as a reasonable approximation. It is worth mentioning that the measured solar background count represents the total contribution of solar radiation over a range of wavelengths below 320 nm, and is not necessarily only due to in-band noise, because the PMT and filter still have out-of-band leakage. Consequently, the system detected non-negligible solar radiation.
The above measurement results indicate that, in order for our receiver to achieve an SNR of 10 dB or more, the received signal count rate should be greater than F=145 kHz, on average. This rate can be translated into an average received power for a given transmission wavelength. Each photon carries a total energy, E=hc/λ where h is the Planck constant, c is the speed of light, and λ is transmission wavelength. For example, a photon with λ=250 nm carries a total energy E=7.956*10-19 J. Hence, to achieve 10 dB SNR the average received power must be no less than P r=E*F=1.15*10-13 W=1.15*10-10 mW. For a single LED transmitting an average power of P t=0.3 mW, the total system loss is thus required to be within 2.6*109, or equivalently 94 dB. This includes propagation loss, filter loss, and PMT loss. If we consider the system to have a constant loss budget, however, increasing the number of LEDs helps to increase the received signal power in order to reach a desired level of photon count.
Under daytime operating conditions, the noise count rate varies from late morning to early afternoon by at most a factor of two, down to a photon count rate of 8 kHz. The rate, however, varies significantly with time over the course of one day. During several experiments it dropped to below 1 kHz in the early morning or late afternoon. Knowledge of noise count is needed to determine acceptable signal levels and maintain desired SNR at the receiver, especially when using photon counting receivers. If the data rate is increased, both the noise count and the signal count per pulse decrease. More transmit power is also needed if the pulse duration is made shorter to keep a constant number of signal photons per pulse. Note also that, at night, dark counts become the dominant noise source.
Signal counts were also recorded. The signal photon count measurements were obtained during the same times of day, with Tx/Rx elevation angles of 30°/30° and a separation distance of 70 m. Figure 3 shows both experimental data points and a Poisson fitting curve for the signal photon counts. Very good agreement is apparent.
These measurements, along with their approximate distributions, provide a guideline for design of advanced statistical signal detection schemes. Additional performance measures, including path loss and BER, are discussed next.
3.2 Path loss
Path loss measurements were obtained for different Tx/Rx geometries and separation distances. The path loss was calculated as the ratio between the transmitted photons radiated from the UV LEDs and the signal photons impinging upon the receiver. The former was calculated based on the measured source radiated power, and the latter was calculated from measured received photons divided by the total percentage loss from the filter and PMT. The receiving area was 1.77 cm2. If path loss per unit area is of interest, then results can be normalized by this area. Figure 4 presents the path loss at different distances on a logarithmic scale, for different Tx and Rx elevation angles. We observe that the path loss increases by about 18 dB for each order of magnitude increase in distance r, i.e., path loss is proportional to r 1.8 under this geometry (the path loss exponent is 1.8). For other geometries, the path loss exponent may change. For example, for a very short range up to 10 m and Tx/Rx angle of 90° , it was found to be close to 1. However, the effect of geometry on the path loss exponent is still under investigation. For a fixed Rx angle, the loss is not very sensitive to the change in the Tx angle at these moderate angle values. A total variation of only a few decibels is observed when the Tx angle is changed from 30° to 45°. If we fix the Tx angle, however, the loss is found to depend highly on the Rx angle, with a 10 dB difference between Rx angles 30° and 45°. In general, as expected, we observe that the loss increases as either the Tx or Rx angle increases. This is due to the longer propagation path as well as the inherent scattering loss. In our experiments, the beam divergence and receiver FOV were fixed. They might also contribute to path loss variations, although their effects can only be observed if additional optical modules to control those angles are designed and integrated with the LEDs and the filter.
It is worth mentioning that the separation distance in our measurements is relatively short (up to 100 m). Because the attenuation coefficient is typically in the range of 1~10 km-1 , losses due to atmospheric attenuation were insignificant and thus not reflected in our measurements. However, if the separation distance is increased to multiple kilometers, atmospheric attenuation may become dominant, following the typical exponential power decay law assumed in the literature . Such observation also suggests that attenuation effects can be neglected for short range communication systems (<1 km), and that scattering loss is dominant in this case.
We next investigate the path loss as a function of different Tx/Rx geometry. Measured results are shown in both Figs. 5 and 6, where the separation distance is fixed to 25 m, and we obtain the path loss as we vary the Tx and Rx elevation angles, respectively. Several observations can be made from each figure with regard to the properties of NLOS scattering communication systems. A total path-loss difference of over 50 dB is observed as the angle is changed from 0o (LOS) to 90o. The rate of change of the path loss decreases when the angle is increased and saturates at approximately 103 dB when the Tx/Rx angle is increased over 75°. Thus, path loss is more sensitive to angle variation when angles are small. The path loss is also observed to increase semi-monotonically with the variation of the Tx/Rx angle (x-axis) for a fixed Rx/Tx angle, with the exception of an observed minimum at around 50° Tx angle. This may be attributed to angular dependent scattering in the common volume, as represented by the scattering phase function [5,30]. The phase function which describes the angular distribution of the scattered beam is typically calculated as a weighted sum of both Rayleigh (molecular) and Mie (aerosol) scattering phase functions . The weights can be chosen to be proportional to the corresponding scattering coefficients. Due to individual behaviors of Rayleigh and Mie scattering phase functions in the deep UV band, the averaged phase function shows a maximum at a certain scattering angle, indicating the strongest scattered UV radiation (or the smallest path loss) in that direction relative to the beam incident angle. This property suggests that the phase function may have a direct impact on the angular path loss behavior. While quantitative and qualitative analysis would be more convincing than the current reasoning, this approach is unavailable because it depends on a realistic phase function model and a path loss model. The path loss is also observed to be less sensitive to the dependent angle in the range of 30°~60° versus other values, partially explaining the angle insensitivity phenomenon in Fig. 4.
3.3 BER performance
Experiments were also conducted to measure the communication BER using on-off keying (OOK) modulation. The received signal model is described by y = x+n where x is signal and n is noise. Demodulation was performed off-line after the received counts were recorded. The threshold used to decide whether a pulse was received or not was optimized based on the background noise and signal photon counts. Note that the SNR of the received signal is affected by the different geometric parameters described earlier. Therefore, to present the BER for different SNRs, we chose to vary the Tx and Rx angles in order to vary the SNR as desired. Considering the randomness of received signal and noise photons, the BER and SNR presented below are measured averages. Figure 7 compares measured and predicted BER, where the prediction is based on the SNR and the Gaussian Q-function formula valid for Gaussian noise , which approximates measured noise count distribution. Predicted and measured results show good agreement. This figure also reveals how much received SNR is required to achieve a certain BER, or equivalently the average required received signal photon count for a given noise environment. For example, at SNR=10 dB, a BER of 10-2 is achievable; and as SNR increases to 15 dB, a BER below 10-4 is achievable. To see how Rx elevation angle explicitly impacts BER, we fixed the Tx elevation angle at 30°, 40°, 50°, and 60°, respectively, at a communication distance of 35 m. Corresponding BER results are plotted in Fig. 8. The figure illustrates that the BER with Tx angle fixed at 30° can drop from 10-1 to about 10-6 when the Rx angle decreases from 40° to 20°, with further reductions in BER when pointing approaches line-of-sight.
4. Conclusions and future work
This paper presented various experimental results for NLOS UV communications based on low power divergent LED source arrays. Solar background noise and signal count distributions were characterized. Path loss and BER of corresponding photon counting detectors were studied under different system geometries determined by Tx and Rx elevation angles and communication distances. These experimental results are valuable for the design of practical receivers for NLOS systems.
Additional studies have also been conducted  showing that path loss predictions based on the single scattering model  are only applicable to very limited geometries and significantly deviate from measurements under many other geometries. An appropriate multiple scattering model may prove to be more generally applicable. The path loss exponent may depend on geometry, beam profile, and Rx FOV. It may vary from a value close to 1 as reported in [28,30], to a value close to 2 reported in the current work. Our future studies will focus on developing scattering and phase function models based on measurements under different meteorological conditions, incorporating the effects of beam angle and FOV. We are also studying the channel impulse response and atmospheric attenuation effects using a high power UV source.
The authors would like to thank Haipeng Ding and Qunfeng He for their invaluable help with experiments. This work was supported in part by the Army Research Office under Grants W911NF-06-1-0364 and W911NF-06-1-0173, and the Army Research Laboratory under the Collaborative Technology Alliance Program, Cooperative Agreement DAAD19-01-2-0011.
References and links
1. Z. Xu and B. M. Sadler, “Ultraviolet communications: potential and state-of-the-art,” IEEE Commun. Mag. 46, (2008).
2. G. L. Harvey, “A survey of ultraviolet communication systems,” Naval Research Laboratory Technical Report , Washington D.C., March 13, 1964.
4. D. E. Sunstein, “A scatter communications link at ultraviolet frequencies,” B.S. Thesis, MIT, Cambridge, MA , 1968.
5. D. M. Reilly, “Atmospheric optical communications in the middle ultraviolet,” M.S. Thesis, MIT, Cambridge, MA, 1976.
6. D. M. Reilly and C. Warde, “Temporal characteristics of single-scatter radiation,” J. Opt. Soc. Am. A 69, 464–470 (1979). [CrossRef]
7. M. R. Luettgen, J. H. Shapiro, and D. M. Reilly, “Non-line-of-sight single-scatter propagation model,” J. Opt. Soc. Am. A 8, 1964–1972 (1991). [CrossRef]
8. E. S. Fishburne, M. E. Neer, and G. Sandri, “Voice communication via scattered ultraviolet radiation,” final report of Aeronautical Research Associates of Princeton, Inc., NJ , February 1976.
9. J. J. Puschell and R. Bayse, “High data rate ultraviolet communication systems for the tactical battlefield,” in Proceedings of Tactical Communications Conf. (April 1990), pp. 253–267.
10. B. Charles, B. Hughes, A. Erickson, J. Wilkins, and E. Teppo, “An ultraviolet laser based communication system for short range tactical applications,” Proc. SPIE 2115, 79–86 (1994). [CrossRef]
11. R. D. Shute, “Electrodeless ultraviolet communications system,” IEEE Aerosp. Electron. Syst. Mag. 10(11), 2–7 (1995). [CrossRef]
12. M. Shatalov, J. Zhang, A. S. Chitnis, V. Adivarahan, J. Yang, G. Simin, and M. Asif. Khan, “Deep ultraviolet light-emitting diodes using quaternary AlInGaN multiple quantum wells,” IEEE J. Sel. Top. Quantum Electron. 8, 302–309 (2002). [CrossRef]
13. V. Adivarahan, Q. Fareed, S. Srivastava, T. Katona, M. Gaevski, and Asif Khan, “Robust 285 nm deep UV light emitting diodes over metal organic hydride vapor phase epitaxially grown AlN/sapphire templates,” Jpn. J. Appl. Phys. 46, 537–539 (2007). [CrossRef]
14. H. Hirayama, T. Yatabe, N. Noguchi, T. Ohashi, and N. Kamata, “231–261 nm AlGaN deep-ultraviolet light-emitting diodes fabricated on AlN multilayer buffers grown by ammonia pulse-flow method on sapphire,” Appl. Phys. Lett. 91, 071901 (2007). [CrossRef]
15. X. Bai, D. Mcintosh, H. Liu, and J. C. Campbell, “Ultraviolet single photon detection with Geiger-mode 4H-SiC avalanche photodiodes,” IEEE Photon. Technol. Lett. 19, 1822–1824 (2007). [CrossRef]
16. J. C. Campbell, H. D. Liu, D. McIntosh, and X. Bai, “SiC avalanche photodiodes,” presented at the Intl. Semiconductor Device Research Symposium, College Park, Maryland, December 12–14, 2007.
17. X. Guo, L. B. Rowland, G. T. Dunne, J. A. Fronheiser, P. M. Sandvik, A. L. Beck, and J. C. Campbell, “Demonstration of ultraviolet separate absorption and multiplication 4H-SiC avalanche photodiodes,” IEEE Photon. Technol. Lett. 18, 136–138 (2006). [CrossRef]
18. H. Liu, X. Guo, D. McIntosh, and J. C. Campbell, “Demonstration of ultraviolet 6H-SiC PIN avalanche photodiodes,” IEEE Photon. Technol. Lett. 18, 2508–2510 (2006). [CrossRef]
19. J. B. Limb, D. Yoo, J. H. Ryou, W. Lee, S. C. Shen, and R. D. Dupuis, “GaN ultraviolet avalanche photodiodes with optical gain greater than 1000 grown on GaN substrates by meal-organic chemical vapor deposition,” Appl. Phys. Lett. 89, 011112 (2006). [CrossRef]
20. S. C. Shen, Y. Zhang, D. Yoo, J. B. Limb, J. H. Ryou, P. D. Yoder, and R. D. Dupuis, “Performance of deep ultraviolet GaN avalanche photodiodes grown by MOCVD,” IEEE Photon. Technol. Lett. 19, 1744–1746 (2007). [CrossRef]
21. M. Razeghi, “Deep ultraviolet light-emitting diodes and photodetectors for UV communications,” Proc. SPIE 5729, 30–40 (2005). [CrossRef]
22. T. Tut, M. Gokkavas, A. Inal, and E. Ozbay, “AlxGa1-xN-based avalanche photodiodes with high reproducible avalanche gain,” Appl. Phys. Lett. 90, 163506 (2007). [CrossRef]
23. F. Yan, X. Xin, S. Aslam, Y. Zhao, D. Franz, J. H. Zhao, and M. Weiner, “4H-SiC UV photo detectors with large area and very high specific detectivity,” IEEE J. Quantum Electron . 40, 1315–1320 (2004). [CrossRef]
24. G. A. Shaw, M. L. Nischan, M. A. Iyengar, S. Kaushik, and M. K. Griffin, “NLOS UV communication for distributed sensor systems,” Proc. SPIE 4126, 83–96 (2000). [CrossRef]
25. G. A. Shaw and M. L. Nischan, “Short-range NLOS ultraviolet communication testbed and measurements,” Proc. SPIE 4396, 31–40 (2001). [CrossRef]
26. G. A. Shaw, J. Fitzgerald, M. L. Nischan, and P. W. Boettcher, “Collaborative sensing test bed and experiments,” Proc. SPIE 5101, 27–38 (2003). [CrossRef]
27. G. A. Shaw, A. M. Siegel, and M. L. Nischan, “Demonstration system and applications for compact wireless ultraviolet communications,” Proc. SPIE 5071, 241–252 (2003). [CrossRef]
28. G. A. Shaw, A. M. Siegel, J. Model, and D. Greisokh, “Recent progress in short-range ultraviolet communication,” Proc. SPIE 5796, 214–225 (2005). [CrossRef]
29. G. A. Shaw, A. M. Siegel, and J. Model, “Extending the range and performance of non-line-of-sight ultraviolet communication links,” Proc. SPIE 62310C, 1–12 (2006).
30. Z. Xu, H. Ding, B. M. Sadler, and G. Chen, “Analytical performance study of solar blind non-line-of-sight ultraviolet short-range communication links,” Opt. Lett. 33, 1860–1862 (2008). [CrossRef] [PubMed]
31. Z. Xu, G. Chen, F. Abou-Galala, and M. Leonardi, “Experimental performance evaluation of non-line-of-sight ultraviolet communication systems,” Proc. SPIE 67090Y, 1–12 (2007).
32. D. M. Reilly, D. T. Moriarty, and J. A. Maynard, “Unique properties of solar blind ultraviolet communication systems for unattended ground sensor networks,” Proc. SPIE 5611, 244–254 (2004). [CrossRef]
33. M. Razeghi, “Deep ultraviolet light-emitting diodes and photodetectors for UV communications,” Proc. SPIE 5729, 30–40 (2005). [CrossRef]
34. E. J. McCartney, Optics of the Atmosphere: Scattering by Molecules and Particles (John Wiley & Sons, New York, 1976).
35. Z. Xu, H. Ding, and G. Chen, “Non-line of sight atmospheric channel modeling in the solar blind ultraviolet regime,” presented at the SPIE Defense and Security Symposium, Orlando, Florida, March 17–20, 2008.
36. Z. Xu, G. Chen, and F. Abou-Galala, “Performance of ultraviolet communications through the atmosphere,” presented at the SPIE Defense and Security Symposium, Orlando, Florida, March 17–20, 2008.
37. R. M. Gagliardi and S. Karp, Optical Communications, 2nd ed. (John Wiley & Sons, New York, 1995).