IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, VOL. ASSP-34, NO.
3, JUNE 1986 393
Practical Supergain HENRY COX, FELLOW, IEEE, ROBERT M. ZESKIND, SENIOR MEMBER, IEEE, AND THEO KOOIJ, MEMBER, IEEE Abstract-The problem considered is that of designing endfire line array shadings which provide a useful amount of supergain without extreme sensitivity to random errors. Optimum shading weights are obtained subject to a constraint on the gain against uncorrelated white noise. The results of optimum array gain versus white noise gain con- straint aye presented parametrically for arrays of different interele- ment spacings, and different noise fields. Results are presented for spherically and cylindrically isotropic noise, and other wavenumber limited noise fields, used in modeling ocean ambient noise. It is found that nearly optimum performance can be obtained in a simple delay and sum beamformer by shading to reduce sidelobes and modest oversteering to reduce mainlobe width without too large a reduction in mainlobe sensitivity. I INTRODUCTION T has long been known that a significant increase in the endfire beam array gain of a line array over that of con- ventional beamforming can be achieved in isotropic noise fields when adjacent elements are separated by less than one-half wavelength 11-[4]. This phenomenon has been called supergain at endfire. For example, for N very closely spaced elements, endfire array gain approaching N 2 can be achieved in spherically isotropic noise [3] and a corresponding array gain o f 2 N 1 can be achieved for a cylindrically isotropic noise field 121. Early work involved steering Dolph-Chebyshev shaded arrays past endfire to narrow their mainlobes [6], [7]. For closely spaced elements, this can be done without a grat- ing lobe entering visible space. It soon became apparent [5] that the theoretical gains of optimum performance were not obtainable in practice due to the extreme sensitivity of these highly tuned systems to small random errors en- countered in real world applications. Important errors, such as amplitude and phase errors in sensor channels, are nearly uncorrelated from sensor to sensor and affect the beamformer in a similar manner to spatially white noise. Hence, the array gain against uncorrelated or spatially white noise is a good measure the array processor's robustness to errors. Gilbert and Morgan [8] first addressed the problem of maximizing array gain against isotropic noise with a con- straint on gain against uncorrelated noise and showed a monotonic dependence between the gain and the con- straint. A closely related problem was examined by Uz- soky and Solymar [lo]. We consider the generalized ver- Manuscript received April 22, 1985; revised October 14, 1985. H. and R. M. Zeskind are with the BBN Laboratories, Inc., Ar- T. Kooij is with the Office of Naval Technology, Arlington, VA 222 IEEE Log Number 8407205. lington, VA 22209. sion 1 13of that problem; that of finding the set of optimal complex weights which maximizes the array gain against correlated noise subject to an inequality constraint on the gain against spatially white noise. In this problem, the white noise gain is specified for a desired level of insen- sitivity to random errors. The solution of the constrained optimization problem yields endfire array gain which is less than the uncon- strained supergain value but greater than for conventional beamforming. Constrained optimal solutions represent a good tradeoff between increased array gain and sensitivity to errors.' The complex.weights may be implemented in a frequency domain beamformer. These weights generally do not have a linear phase relationship associated with the simple delay and beamformer. In this paper we first consider the sensitivity con- strained optimization problem and present results for sev- eral types of background noise. Then the problem of achieving near optimum performance with a delay and sum beamformer is examined and a design technique pre- sented. It is found that near optimum performance can be achieved using a simple delay and sum beamformer with proper delay selection and amplitude shading. SENSITIVITY CONSTRAINED OPTIMAL BEAMFORMING Consider a line array of N equally spaced sensors. The Hermitian cross-spectral density matrix of the sensor out- puts at a frequency f is given by where a: is the noise power spectral density averaged over the N sensors, and is the average signal power spectral density of the N sensors. The signal cross-spectral density matrix is P and that for the noise 1s Q. Both P and Q are normalized to have their traces equal to N. The quantity o:/aZ is the input signal-to-noise spectral ratio. The output spectral density of a beamformer z( f ) can be expressed as the Hermitian form z ( f ) w*( f ) R(f) w f f ) (2) where w( f ) is the vector of complex beamformer weights at frequencyfand the asterisk denotes complex conjugate transpose. The array gain G is defined as the output sig- nal-to-noise ratio divided by the input signal-to-noise ra- tio and is given by 0096-3518/86/0600-0393$01.00 O 1986 IEEE
394 lEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, VOL. ASSP-34, NO. 3, JUNE 1986 The maximum gain is achieved by the complex weight vector w which maximizes ( 3 ) . The problem of maximizing the array gain ( 3 ) subject to a constraint on the white noise gain, defined as is most easily formulated as that of finding the vector w which minimizes 1/G with a constraint on l/G,,,, That is, minimizing where E is a Lagrange multiplier. Finding the w which minimizes ( 5) is equivalent to finding the w which maxi- mizes the reciprocal of ( 3 , which in turn is an eigenvalue problem. The optimum w is chosen to be proportional to the eigenvector corresponding to the largest eigenvalue of [Q EZ] - ' P. When there is a single coherent wavefront from direction 0, P is a dyad or rank one matrix which may be written as P rn (f, 0) m* (f, 0). Then the sen- sitivity constrained optimum w is any nonzero'scalar rnul- tiple of [Q ~11-I 0). (6) The Lagrange multiplier provides a continuous param- eterization between the array gain and the white noise gain. The relationship between these gains is monotonic in that increasing the white noise gain from its uncon- strained optimum value at E 0 to its maximum value, 10 log N, E causes the constrained optimum array gain to decrease monotonically from its unconstrained maximum to the array gain of a uniformly shaded con- ventional beamformer. As an example, consider an eight-element uniformly spaced endfire line array [ 12] . Figs. 1-4 present curves of constrained optimal array gain versus the white noise gain for several values of the ratio of element spacing s to plane wave wavelength X. The Lagrange multiplier E is a parameter along the curves. Fig. 1 shows the results for a spherically isotropic noise field, and Fig. 2 shows re- sults for a cylindrically isotropic noise field. Figs. 3 and show similar results for other models of ocean ambient noise. For very large E , the sensitivity constrained weight vec- tor of (6) will tend to m, since the matrix I will dominate Q. But m is just conventional endfire bearnforming with uniform shading and plane wave phasing. Therefore, as the desired white noise gain constraint is increased, the weight vector of (6) and the array gain approach that of conventional beamforming. The upper limit on G,, for this example, is 10 log 8 9 dB, the conventional array gain against uncorrelated noise. As the minimum accept- able value of G, is decreased, the constrained optimal array gain increases approaching the asymptotic iimit for closely spaced elements of N 2 for spherically isotropic noise field and 2N for cylindrically isotropic G 1 -50 -30 -20 70 Fig. 1. Array gain versus G,, for an eight-element endfire line array in spherically isotropic noise. 1' 01 -50 -40 -20 -10 0 70 Fig. 2. Array gain versus G, for an eight-element endfire line array in cylindrically isotropic noise. 0 -50 -40 - 3 0 - 2 0 -70 0 10 Idis] Fig. 3 . Array gain versus G, for an eight-element endfire line array in surface noise. noise field. Notice that there is an initial very sharp rise in the curves that occurs as G, is allowed to decrease from 9 dB. We will use 3 dB as a design constraint for G , which permits a significant increased endfire array gain
COX et al.: PRACTICAL SUPERGAIN 395 G 3 -SO 7.3 6.2 t -40 -30 0 WHITE Fig. 4. Array gain versus G, for an eight-element endfire line array in distant shipping noise. 1 8.8 Fig. 5. Constrained optimum weights for cylindrically isotropic noise where SIX 0.3, G, 3 dB. (a) l W , 2 1 . (b) Phase perturbation. SIA Fig. 6 . Optimum beampattern (cylindrically isotropic noise) G 6.9 =9 Fig. 7. Conventional beampattern (cylindrically isotropic noise). against isotropic noise or ocean ambient noise without ex- cessive sensitivity to random errors. Fig. presents t hew which is the solution to (6) when G, 3 dB, 0.3, in a cylindrically isotropic noise field. The upper curve in Fig. is the magnitude of the weights, while the lower curve is the phase departure from plane wave phasing. The endfire beampattern resulting from these constrained optimal weights is shown in Fig. 6. Comparing Fig. 6 to the conventional endfire beam- pattern in Fig. 7, we observe that the weights of (6) result in a much narrower mainlobe and reduced sidelobe levels. RELATIONSHIP TO OVERSTEERING The 3 dB beamwidth of a uniformly spaced line array steered to endfire is approximately 2J o886x I L. If the elements are spaced at one-half wavelength, a grating lobe appears at the opposite endfire. By decreasing the element spacing slightly, the grating lobe vanishes from visible space. There is little corresponding broadening of the mainlobe since, at endfire, the beamwidth depends on the square root of array length. The net result is a nearly 3 dB increase in endfire array gain which is sometimes called the endfire anomaly. It is possible to oversteer or steer past endfire by using time delays phases) matched to a propagation velocity which is less than the speed propagation c . The delay r, applied to the nth element at position is 7, (1 which is matched to a wave propagating at a velocity (1 where 6 is a small positive constant which deter- mines the amount of oversteering. Oversteering has the effects of reducing the mainlobe width in visible space, reducing mainlobe sensitivity, and bringing into visible space new lobes at the opposite endfire. The sidelobe sen- sitivities are not changed but their levels relative to the reduced mainlobe are increased. The closeness of the ele- ment spacing determines the amount of oversteering which can be introduced without bringing in a grating lobe at the
396 IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, VOL. ASSP-34, NO. 3, JUNE 1986 opposite endfire. Early work in superdirective arrays in- volved extreme shading for greatly reduced sidelobes, and extreme oversteering of arrays with very closely spaced elements resulting in a very narrow mainlobe of greatly reduced sensitivity. These designs were soon found to be extremely sensitive to various errors making them im- practical. In our parlance, the extreme oversteering re- sulted in an unsatisfactorily low gain against uncorrelated noise. The question naturally arises: Can one achieve near constrained optimum performance using mild oversteer- ing? An examination of Fig. 5 is encouraging since the phases of the optimum weights are nearly linear in ele- ment position. Fig. 6 shows that the optimum beam pat- tern of an eight-element array with s/ h 0.3 and having a 3 dB white noise gain constraint has a peak sidelobe level of about 13 dB. This 13 dB plus a 6 dB loss in mainlobe sensitivity which reduced the white noise gain from 9 dB (10 log 8) to 3 dB suggests that a design which has 19 dB sidelobe before oversteering might give similar performance. As a candidate design, we will place slightly more emphasis on sidelobe control. We proceed as fol- lows. 1) Shade the array for 25 dB sidelobes before over steering. This should result in 19 dB sidelobes after over- steering. For simplicity we use 25 dB Taylor [9] shading. 2) Oversteer the array a limited amount such that the resulting white noise gain is 3 dB. This approximately corresponds to oversteering by an amount such that the 6 dB down point on the mainlobe is at endfire. Since we oversteer only a fraction of the mainlobe, only a fraction of an additional lobe enters visible space at the opposite endfire. The appearance of a grating lobe will be avoided if N - 2 2 N so that is slightly less than X/2. The design procedure is illustrated in Figs. 7, 8, and 9. Fig. 7 shows the beampattern of an unshaded array steered to endfire. Fig. 8 shows the corresponding beampattern when 25 dB shading is used. The shading has cost about 0.4 dB loss in white noise gain. Fig. 9 shows the beam- pattern of the shaded array with oversteering such that G, is 3 dB. The gain of this array against two-dimensional isotropic noise is only 0.2 dB lower than the constrained optimum of Fig. 6 . Nearly uniform 19 dB down sidelobes have been achieved. Thus, the simple shaded delay and sum oversteered beamformer is nearly optimum for this example. The time delays of (7) were adjusted i n the previous example to meet a 3 dB white noise gain constraint for a cylindrically isotropic noise field. The question naturally arises as to the performance of this simple delay and sum beamformer when the noise field is not the one assumed in adjusting the time delay, i.e., the effect of model mis- match. For the uniformly spaced eight-element line array with G, 3 dB and s/X 0.3, the constrained optimal SIX ~ 0 . 3 =6.7 Fig. 8. Beampattern with 25 dB shading (cylindrically isotropic noise) =9.3 Fig. 9. Shaded and oversteered beampattern (cylindrically isotropic noise). array gain for spherically isotropic noise is found from Fig. 1 to be 14.1 dB. Using the Taylor shadings and time delays of the previous example, it was found that the array gain achieved by the simple delay and sum beamformer was 13.7 dB, only 0.4 dB lower than the constrained op- timum. Thus, the shaded and oversteered design gives nearly optimum performance in both cylindrically and spheri- cally isotropic noise fields. The reason for this is fairly obvious. In both cases oversteering exploits the fact that the noise fields are limited in wavenumber to those wave- numbers corresponding to real angles. The oversteered beamformer exploits in a slightly different way the same properties of the band-limited noise as does the con- strained optimum. As a second example, consider a 32-element line array with one-quarter wavelength spacing and a white noise constraint of 8 dB in a spherically isotropic noise field. The interelement spacing is 0.25 X. Conventional endfire steering with uniform weights yields the beampattern of
COX er ai.: PRACTICAL SUPERGAIN 397 Fig. 10. Conventional beampattern (spherically isotropic noise). Fig. 1 1 . Optimum beampattern (spherically isotropic noise). Fig. 12. Beampattern with 25 dB shading (spherically isotropic noise). Fig. 10 with an array gain of 15 dB and coincidently G, 15 dB. The constrained optimum beampattern for G,,, 8 dB is shown in Fig. 11. It achieves an array gain of 19.8 dB, a nearly 5 dB increase in array gain over con- Fig. 13. Shaded and oversteered beampattern (spherically isotropic noise). ventional beamforming while retaining reasonable ro- bustness against random errors. Fig. 12 shows the effect of 25 dB Taylor shading on the array steered to endfire. The value of G, is reduced by about 0.4 dB. Finally, Fig. 13 shows the shaded array oversteered such that G,,, 8 dB. The array gain of 19.3 dB, within dB of optimum, is achieved with a simple delay and sum beamformer. CONCLUSION Using constrained optimum beamforming, it is possible to achieve a significant increase in endfire array gain for a variety of noise fields consisting of plane waves. An essential feature of these noise fields is that they are band limited in wavenumber corresponding to waves propagat- ing at a fixed phase velocity. Thus, they have no energy corresponding to slow waves. The constraint on the white noise gain limits the sensitivity of the array gain to small deviations from idealized assumptions. The use of real shadings plus mild oversteering in a simple delay and sum beamformer has been found to give nearly optimum per- formance. ACKNOWLEDGMENT The authors express their appreciation to M. M. Owen, BBN Laboratories, Incorporated, who developed optimi- zation and simulation software which has been central to our recent work in optimum and adaptive beamforming. REFERENCES [l] W. W. Hansen and J. R. Woodyard, A new principle in directional antenna design, Proc. IRE, vol. 26, Mar. 1938. [2] S . A. Schelkunoff, A mathematical theory of linear arrays, Bell Tech. J . , vol. 2, Jan. 1943. [3] A. I. Uzkov, An approach to the problem of optimum directive an- tenna design, Comptes Rendus (Doklady) de lrlcademic des ences del LURSS, vol. LIII, no. 1946. H. J. Riblet, Note on the maximum directivity of an antenna, IRE, vol. 36, May 1948. [5] T. T. Taylor, A discussion of the maximum directivity of an an- tenna, Proc. IRE, vol. 36, Sept. 1948. [6] R. H. Duhamel, Optimum patterns for endfire arrays, Proc. IRE, vol. 41, May 1953. [7] R. L. Pritchard, Optimum directivity for linear point arrays, Proc. J . Acoust. Amer., vol. 25, Sept. 1953.
398 IEEE TRANSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESSING, VOL. ASSP-34, NO. 3, JUNE 1986 [8] E. N. Gilbert and S . P. Morgan, Optimum design of antenna arrays subject to random variations, Bell Syst. Tech. J . , vol. 34, May 1955. [9] T. T. Taylor, Design of line-source antennas for narrow beamwidth and low side lobes, IRE Trans. Antennas Propag., vol. AP-3, Jan. 1955. [lo] M. Uzsoky and L. Solymar, Theory of super-directive linear ar- rays, Acta Phys. Acad. Sci. Hungary, vol. 6 , no. 2, 1956. [!I] H. Cox, Sensitivity considerations in adaptive beamforming, in Signal Processing, J. W. R. Griffiths, P. L. Stocklin, and C. Van Schooneveld, Eds. New York and London: Academic, 1973. [I21 T. Kooij, Adaptive array processors for sensitivity constrained op- timization, Ph.D. dissertation, Catholic Univ. America, Washing- ton, DC, June 1977. [I31 H. Cox, Spatial correlation in arbitrary noise fields with application to ambient sea noise, J. Acoust. SOC. Amer., vol. 54, Nov. 1973.