Sie sind auf Seite 1von 7

How narrow is narrowbland?

M.Zatman

Indexing term: Narrowband, Cramer-Ruo bound, Super-resolution algorithms, Array signal processing

Abstract: The 'narrowband' assumption is often made in the analysis of array signal processing algorithms. The author provides a definition for the notion of narrowband, and the derivation of an expression which is useful in determining if a particular scenario qualifies as narrowband. The expression derived correctly predicts where the narrowband assumption fails for some superresolution algorithms, the Cramer-Rao bound on angle estimation and the signal-to-interference plus noise performance of adaptive beamformers.

the inter-element phase, q,depends on both the AOA and frequency, for a linear array, a non-zero-bandwidth signal appears as an extended angular source, whereas a zero-bandwidth signal appears to come from a discrete angle of arrival. This effect is known as dispersion and is pictured qualitatively in Fig. 1.

frequency
A zero-bandwidth signal appears to come from a single angle of arrival. However, a non-zero-bandwidth signals appear to come from an angular extent commensurate with its bandwidth. This effect is known as dispersion

Fig. 1

Qualitative picture of dispersion

Introduction

The 'narrowband' assumption is often made when analysing the performance of an adaptive array signal processing or estimation scheme. A vague definition of narrowband given in the literature [l] is that 'the]-e is essentially no decorrelation between signals received on opposite ends of the array'. In this paper, a more precise analytical expression is derived for classifying when a signal environment may be described as narrowband. The expression derived, motivated by the signal model, provides insight into the performance of a number of super-resolution algorithms, the Cramer-Rao lower bound (CRB) and the signal-to-interference-plus-noise ratio (SINR) performance of adaptive beamformer:;.
2

For the zero-bandwidth case, the exact covariance 4 matrix of the data received by the array from A signals arriving from AOAs 01, ..., 13, and the (zero mean and spatially white) noise is typically modelled as

Data model

Consider a uniform linear array (ULA) of N antennas with an inter-element spacing d, (the results in this paper are easily extended to an arbitrary array geometry). The transfer function between the angle of arrival (AOA) from the array broadside 6' and the output of the array is represented by the steering vector where

(3) where A = [a(6',), ..., a(@,>], I is an identity matrix, o2 is the noise power and represents the Hermitian (complex) transpose operator. It is also assumed that o2= 1 and that the signals are temporally uncorrelated, thus S is an M x A4 matrix with each signal's SNR along its leading diagonal and zeros elsewhere. R may be simply decomposed into the sum of the covariance matrices for each signal RI, ..., R, plus the noise covariance matrix 1. The covariance matrix for the mth signal may be written as

R = ASAH + a21

"

R = s,a(L)a(Q,)H , (4) The kith element of R, corresponding to the correlation of the mth signal between the kth and ith elements of the array is
(5) where zk[ is the time delay between the kth and lth ele, ments of the array, and s is the SNR of the mth signal. From eqn. 2 it is easily shown that
r k l = (1 - k)dsin(Q)/c For the non-zero-bandwidth case, the covariance matrix may be obtained by integrating eqn. 4 over the desired range of instantaneous frequency, i.e.

Tkl

= S,eJ2.rrmfo

a(8) = [l,eJ+,. . . , eJ(N-l)+]T represents the transpose operator and


27rdf
$I=--

(1)

sin(8)

f i s the frequency of the signal incident on the array, and c the propagation velocity. For convenience, the array is assumed to have half-wavelength spacing at the chosen operating frequency, f o . In array processing, a fixed phase is typically used to approximate the time delay of a signal between elements of an array. Since
IEE Proceedings online no. 19981670

Rm =

f"+k
sm(f

fo-%

).(e, f)a(Q, n f f

(6)

Paper first received 25th March and in revised form 11th August 1997 The author is with MIT Lincoln Laboratory, 244 Wood St., Lexington, MA 02138, USA
IEE Proc.-Radar, Sonar Navin.. Vol. 145. No. 2, April 1998

s,cr> is the received signal power as a function of frequency. Eqn. 6 suggests that, apart from the case where 6' = 0, a non-zero-bandwidth signal always produces a full-

where 0 is the bandwidth of the signal and

85

rank covariance matrix, since it may be viewed as an infinite sum of zero-bandwidth signals. However, some of the eigenvalues of such a signal may be small compared to the noise level, and will have little effect on the adaptive beamformer performance. Thus we define an effective rank, which is the number of signal subspace eigenvalues of the signal-only covariance matrix greater than 1. In the signal-plus-noise covariance matrix of eqn. 3, the effective rank is the number of eigenvalues greater than 2 (3dB above the noise floor). The klth element of R, is the integral of eqn. 5 over the desired bandwidth,
rkl

so+$
s,(f)eJzTTkl

fo-5

cif

(7)

For example, if sc) is rectangular between b/2 and -bl ,f 2, then the solution to the above integral eqn. 7 is given by r h l = sinc(brk1)smej21Tkl fo (8) (note that we defined , = bs,vo) = Js,v> dJ). Eqn. 7 is s also the inverse Fourier transform of the signals power spectrum, which is of course the autocorrelation function. This information is used by Compton to derive eqn. 8 [l]. By noting that eqn. 8 is simply eqn. 5 multiplied by a term representing signal decorrelation between the elements of the array, Compton describes a signal as narrowband when

of the signal subspace is the same as the number of signals present, i.e. there is a rank-one representation of each signal present. For a sufficiently wide bandwidth, the effective rank of the signal subspace is larger than the number of signals present, and the zero-bandwidth signal model no longer applies. Thus we will define the notion of a narrowband signal as follows: If the bandwidth of a signal is such that the second eigenvalue of the signals noise-free covariance matrix is larger than the noise level in the signal-plus-noise covariance matrix, then that signal may not be described as narrowband. Simulations may be used to show that as a signals bandwidth is increased, eigenvalues pop up from the noise floor one at a time. Hence, an expression for the 2 largest eigenvalues is sufficient for the purpose of this paper. Methods have been proposed for determining the eigenvalues of a covariance matrix [2, 31; the derivation used here is a special case of the general result of [2]. For the problem posed here, consider the covariance matrix of two uncorrelated zero bandwidth signals with powers ,ul and p2:

R = p1a1a? p . aza? u 2 (13) The eigen-problem is to find values of A and v such that

sinc(brlN) N 1 (9) , where z is the time delay between the 1st and Nth elements of the array, i.e. there is essentially no decorrelation between signals received at opposite ends of the array. The covariance matrix of a number of non-zerobandwidth signals is calculated by first computing all the elements of the covariance matrix for each signal according to eqn. 8, summing the individual matrices and then adding noise as in eqn. 3. Instead of referring to specific values off0 and b, we will also refer to the fractional bandwidth which we define as b bf = f0

Rv = AV (14) Each of the signal subspace eigenvectors is a linear combination of the signals present, so for the two signal case v may be expressed as
v = pal ya2 (15) Thus from eqns. 13 and 15 the product Rv can be written as

Rv = piPNai

+ piy$Nai + p2P$*Na2 + p2yNaz


(16)

The definition of narrowband

A covariance matrix R may be represented by its eigenvalues and eigenvectors such that
(11) where E is the matrix of eigenvectors, and A is the diagonal matrix of eigenvalues. It is well known that the eigenvectors may be partitioned into two subspaces, a signal subspace and a noise subspace, i.e. (12) refer to the signal and noise subspaces, respectively. In super-resolution, E, is a minimum-rank orthonormal basis for the signals whose parameters we wish to estimate. In adaptive beamforming, the E, is called the interference subspace, since it contains an orthonormal basis for the interference we wish to suppress [Note I]. For the zero-bandwidth signal model, it follows from eqn. 3 that the rank

where aHa = N ( N is the number of array elements, equal to the array gain assuming equal power sensors) and aIHa, = NI). The quantity I) is the cosine of the angle between the two vectors in N-dimensional space. From eqns. 14 and 15 we can rearrange eqn. 16 as the following pair of simultaneous equations in the form of a 2 x 2 eigen-problem XP = PlPN pLly?LN

A = P2P$*N y

+ PzyN

(17)

R = EhEH

Thus the size-N eigen-problem has been reduced to a size-two eigen-problem with roots at

R = E,A,E,H

+ E,AnE,H

Solving the resulting quadratic in h gives the following eigenvalues

where the subscripts s and

y1

(19) For equal power sources, 1 e. p 1 = p2, eqn. 19 simplifies to

Note 1: In this paper the radar scenario, where the interference-plus-noise covariance matrix is estimated from range gates which exclude the target signal, is assumed
86

A, = NPl(1+ 14) (20) Of interest for the problem addressed here is the smaller eigenvalue (the - of the 2 in eqns. 19 and 20) Now that an expression for the second eigenvalue of the covariance matrix has been obtained, the wideband
IEE Proc -Radar Sonar Navig Vol 145 No 2 April 1996

problem needs to be formulated in such a way that eqn. 20 may be applied. By the definition earlier, a signal ceases to be narrowband when its effective rank increases from one to two. Therefore, the rank-two approximation of a nonzero bandwidth by two discrete uncorrelated equal power sources is an adequate model for the purposes of this paper. The discrete sources need to be arranged so that the mean and variance of their instantaneous frequency spectrum correspond to the mean and variance for the non-zero-bandwidth signal they are meant to model. This is easily accomplished for an arbitrary signal pass-band. For the example of a rectangular power spectrum with bandwidth b mentioned earlier, the variance of the spectrum is given by

noting that each of the two signals has a power of 0 . 5 ~ Owing to the formulation of eqn. 13 the value of ~. A has been calculated without the inclusion of noise. If 2 noise is included, then the value of A2 will, assuming the model of eqn. 3 , be the result of eqn. 28 plus 1. For a signal to be classified as narrowband, i.e. for a signals effective rank to be 1, A2 must be smaller than 1, (equivalent to A2 < 2 or < 3dB above the noise level if the noise is included), i.e.

A two frequency representation of the signal would consist of two delta functions spaced a frequency K either side of the centre frequency. The varianclz of . such a representation is K ~ For the rank-two representation of the finite-bandwidth signal, the variance of the model and reality should be the same, thus
I&

From inspection of eqns. 26-29 the value of A2 is a function of the fractional bandwidth, AOA, power (SNR), the array gain and the dimensions of the array. Comptons result is neither a function of power (SNR) nor the array gain and has no precise threshold. Thus we claim our result is a more precise definition of narrowband. A similar result to eqn. 28 for the case of a rotating array was recently derived by Hayward [4], utilising Lees method for determining the eigenvalues [ 3 ] . For the case of a linear array, the apparent spreading of a signals AOA due to bandwidth effects is analogous to spreading due to array rotation [5].
30 I

and the spacing of the two delta functions is given by b 2K = Given both the two signal model of a non-zero-bandwidth signal, and the expression for the eigenvalues of a two-signal covariance matrix, both parts need to be fused. For an arbitrary array, IqA may be computed from the inner product of the two steering vectors a(@ + K) and a(@, f - K), where J is the mean of the received signals power spectral density, i.e.

-1 0

I
0

I
0.5
1
1.5

2.5

bandwidth, MHz

(24) For several regular array configurations eqn. 24 has a precise analytical form. For the ULA example used in this paper, the dependence of the inter-element phase on both frequency and AOA means the model of eqn. 23 corresponds to having two signals arriving from bearings equivalent to

Predicted and simulated values ofthe second eigenvalue as a function of bandwidth for 20-element array with AOA of 30and SNR of SOdB at a centre frequency of SOOMHz _ _ predicted ~ ~ __ predicted + noise

Fig.2

simulation

20
10

.
/
/ / / /

ui o x x x x x a The two signals are spaced in sin(6) by sin(0) b f


E=----

2 -10 . a , 0) .a,
-20 .
-30

/
/

/ /

For the ULA example, the value of /q( eqn. 20 is a in Dirichlet function of E

./

For the purposes of this paper, the smaller of the two eigenvalues of eqn. 20 IS of interest. Given , (the SNR s of the signal) and from eqn. 27 (or eqn. 24 for an arbitrary array), substituting the two signal model into eqn. 20 results in
IEE Pvoc -Radar, Sonar Navig , Vol 145. N o 2, April 1998

In order to demonstrate the accuracy of eqn. 29 for the second eigenvalue, the results of some simulations are shown in Figs. 2 and 3. A 20-element half-wave87

length spaced ULA was simulated at a centre frequency of 500MHz with a signal present from 30". In Figs. 2 and 3, either the SNR or bandwidth is fixed while the other parameter is varied. The predicted and measured value of the second eigenvalue is plotted as a function of the changing parameter. It is clear from Figs. 2 and 3 that the prediction of the second eigenvalue is accurate. In the example scenario the narrowband criterion is met under the following two conditions: with an SNR of 50dB, for bandwidths less than 0.135MHz (0.027% fractional bandwidth), and with a bandwidth of lMHz (0.2% fractional bandwidth), for SNRs of less than 32dB. Tiny fractional bandwidths and low SNRs are a requirement for the narrowband criterion to be met.
The non-zero-bandwidth properties of ESPRIT 4 and root-MUSIC

priate. Predicted and simulation results are plotted in Fig. 5 for a single signal incident on a 10-element array from 30", for fractional bandwidths of 0 to 5% with an SNR of 30dB. Based on eqn. 29, the narrowband signal model becomes invalid at a fractional bandwidth of about 0.8%. For fractional bandwidths above 0.8%, the predicted AOAs based on eqn. 26 agree with the simulation results for both algorithms.
0.51

I .

i '8 0.50
1

LI .Lo

Having established a criterion for narrowband, we first investigate how it applies to some super-resolution algorithms. The results presented here are asymptotic, in that they are for the known (infinite sample) covariance case and are intended to give insight into the performance of the algorithms. It is assumed that the reader is already familiar with the essence of these algorithms, the details of which may be found in [6-91. Since we are assuming a known rather than estimated covariance matrix, the results for the various implementations of ESPRIT and MUSIC are the same. In the implementation of the algorithms we define the signal subspace using the effective rank criterion described earlier. Before showing simulation results, we will first attempt to predict the performance of these algorithms as the zero-bandwidth model begins to break down. Consider a single signal incident on the array from an angle other than array broadside. According to eqn. 29, as the bandwidth is increased, at some point the effective rank of the signal subspace will increase from one to two. Both the MUSIC and ESPRIT algorithms attempt to find poles on the unit circle which are related to the AOAs of any signals present. In the case of non-zero-bandwidth signals, these poles attempt to form a low-rank approximation of the true environment. For the case of a rank-two approximation the two poles will be separated in the AOA as described by eqns. 23-26. This is pictured qualitatively in Fig. 4.

0.49

I .

I
2 3
4

fractional bandwidth, %

Fi .5 Predicted and simulated AOAs for root-MUSIC and ESPRIT as aj8nction of bandwidth
After the transition to the rank-two signal model both algorithms gave biased AOA measurements as predicted _ _ _ _ uredicted transition predicted AOA 0 root-MUSIC x ESPRIT
~

The erroneous results for root-MUSIC and ESPRIT are caused by the mismatch between the rank of the signal subspace and the number of signals present. It is well known in the literature that both MUSIC, ESPRIT and many other super-resolution algorithms are asymptotically unbiased for the zero-bandwidth signal model. We have just seen that this is not the case for the non-zero-bandwidth signal model.
5 Adaptive beamformer performance and the Cramer-Rao bound

One common measure of adaptive beamformer performance is the signal-to-interference-plus-noise-ratio (SINR) loss. SINR loss is defined as the ratio of the SINR achieved by the adaptive beamformer to that which could be achieved by a matched filter in the absence of interference. If w is the adaptive weight vector and v the target steering vector, then SINR loss is described by the equation

Fig.4

Non-zero-bandwidth signal occupying a region on the unit circle

In this Section, we will analyse the performance of the optimal adaptive weight vector, w = R-'v, i.e. the weight vector which maximises the SINR performance of the adaptive beamformer. As in Section 4, the results presented here assume a known covariance
matrix.

Many super-resolution algorithms form a low-rank approximation to the true environment, and find the poles corresponding to the low-rank approximation (i) Low-rank approximation of a non-zero-bandwidth signal; (ii) section of circle covered by the signal's bandwidth; (iii) unit circle

Thus, when the zero-bandwidth model begins to fail, we expect root-MUSIC and ESPRIT to begin to wrongly estimate two AOAs for a single signal instead of one. The estimated AOAs will be given by eqn. 25 for the combinations of bandwidth, SNR and AOA where a rank-two approximation of the signal is appro88

Fig. 6 shows plots of SINR loss against azimuth for a single 'jammer' incident on a 20-element array from 30" with an SNR of 50dB (the same scenario as Fig. 2). As the bandwidth is increased, the angular region in which the jammer denies coverage grows. Here we will define a second performance metric, the useable beamspace fraction (UBSF). The UBSF is the proportion of the beamspace (evaluated either in sin(0) or wavenumber rather than ") where the SINR loss is less than 5dB. The 5-dB-loss figure is somewhat
IEE Proc -Radar Sonar Navig
,

Vol 145, No 2.April 1998

arbitrary, but is used because a 5dB loss limits a radar's performance to about 75% of its interferericefree detection range.
0
-5

m
U
0 ! I z -15
v)

-10

-20 -25 0 0.2 0.4 0.6 0.8 sin (azimuth)

Fig.6
~

SINR loss against target AOA for a jammer at 30"

Note how the width of the null widens with increasing bandwidth 0% ___ 1Y" 5%)
100
98

Now let us consider the Cramer-Rao bound (CRB) for non-zero-bandwidth signals. We shall use the adaptive beamforming scenario used above, and look at the CRB for the angle error of a OdB SNR target in the presence of the 50dB jammer at 30". The unit of choice will be the beam-split ratio (frequently used in the radar community), which is given by the ratio of the array's beamwidth to the RMS angle error computed using the CRB. In order to compute the CRB, let us take the vector U of the unknown parameters. In our scenario, the unknown parameters are the target's AOA and magnitude [Note 21. The CRB for the error covariance matrix of an unbiased estimator of the real parameters U = [ul, u2, ...I is given by the inverse of the Fisher Information Matrix (FIM) J. Assuming that the data obeys the zero mean multivariate Gaussian p.d.f., the FIM for the stochastic signal model is conveniently computed using Bang's formula [l 11 as

aP
U -

96
94 92 90

where R, the exact covariance of all the observed data, is a function of U, and Jij is the ijth element of J. The details of the computation of J and the resulting CRB on the variance of the angle errors for the non-zero bandwidth case are given in the Appendix.
12,

0.1

0.2

0.3

0.4

0.5

fractional bandwidth, %

Fig.7
of 50dB

UBSF and estimated rank against bandwidth for a jummer SNR

Eqn. 29 predicts that the narrowband assumption fails at a fractional bandwidth of 0.027% UBSF - rank
-1

-0.5

0
sin (azimuth)

0.5

8
U "
v)

m
3

g 5 I1
95

Fig.9

CRB against target signal AOA with a jammer at 30"

I'

I - - - - - -

The width of the region denied by the jammer increases with bandwidth. Also, note how at the wider bandwidth the target accuracy degrades off broadside owing to the effects of dispersion 0% _ _ 1%> 5%)
~

90 90

-----

85 I 0

10

20

40 60 JNR, dB

80

100

Fig.8 UBSF and estimated rank against jammer SNR for a fractional bandwidth of 0.2%
Eqn. 29 predicts that the narrowband assumption fails at a jammer SNR of 32 dB

For the scenario of Fig. 6, Figs. 7 and 8 show plots of UBSF as a function of the bandwidth (see Fig. 2) and jammer SNR (see Fig. 3). Additionally, plots are shown of the measured estimated rank of the interference subspace. Clearly, there is a close relationship between the interference rank and the UBSF. It is also obvious that eqn. 29 correctly predicts where the zerobandwidth ceases to closely predict performance.
IEE Proc.-Rudar, Sonar Nuvig., Vol. 145,No. 2, April 1998

Fig. 9 shows plots of the CRB as a function of the target AOA for the radar and interference scenario described above using a OdB SNR target. As the bandwidth increases, the region around the jammer where good angle accuracy is denied increases with bandwidth. Additionally, at the higher bandwidths the effect of dispersion on the target away from the array broadside reduces the achievable angle accuracy. In a similar fashion to the SINR-loss case earlier, we can define a CRB UBSF. The CRB UBSF is the proportion of the beamspace where a beam-split ratio better than 5 can be achieved. Fig. 10 shows a plot of both the CRB UBSF and interference rank against bandwidth. As in the adaptive beamforming scenario results reported above, there is a close relationship between the CRB UBSF and the interference rank, and eqn. 29 correctly
Note 2: For the radar scenario assumed in [Note I], it can be shown that the CRB for the target's parameters is unaffected by the estimation of the interference-plus-noisecovariance matrix, since the cross terms in the FIM are zero [lo]
89

predicts the region where the zero-bandwidth model fails to accurately predict performance.
100

9 Appendix: Non-zero bandwidth Cramer-Rao bound on angle estimation

98

1
0
0.1
0.2

1 3

In this Appendix, we detail the computation of the non-zero bandwidth CRB for angle estimation used in Section 5. First, we will describe the computation of the elements of the Fisher information matrix (FIM) denoted by the symbol J, and then we will convert the resulting error variances into our unit of choice-beamsplit ratio. For the scenario considered, the two unknown parameters are the target signals AOA, 8, and power (SNR) s. However, it will prove convenient to use the inter-element phase
w = nsin(8) (32) (for a half-wavelength space array) rather than 0 itself. In order to compute the FIM using eqn. 3 1, the partial derivatives to the data covariance matrix R with respect to (w.r.t.) w and s are required. The covariance matrix of M uncorrelated signals plus noise is R = RI Ra + . . . Rn/r I (33) assuming a noise variance of 1. From eqn. 33 it is clear that the partial derivatives of R w.r.t. the ith signal are independent of the other M 1 signals present and the noise. Thus the partial derivatives required here may be obtained from the non-zero-bandwidth covariance matrix of the target signal only. Rewriting eqn. 8, the klth element of the target signal covariance matrix is given by

0.3

0.4

0.5

__ UBSF

Fig. 10

fractional bandwidth, % CRB UBSF us ufunction of bandwidth

Conclusions

In this paper, an expression was derived which accurately defines the notion of narrowband as used in adaptive array signal processing and estimation. The resulting expression, which is a function of fractional bandwidth, AOA, SNR array gain and aperture, provides more insight into the validity of the narrowband assumption than previous definitions. Simulations were used to verify that the expression correctly predicts where the narrowband assumption fails in adaptive beamformer and super-resolution algorithm performance, and in computing the Cramer-Rao bound for AOA estimation. Additionally, it was shown that the root-MUSIC and ESPRIT algorithms are biased estimators of the AOA in the presence of non-narrowband signals. A by-product of the derivation of the narrowband expression predicts the bias.
7
Acknowledgments

rk1

= ssinc n f w

(?

ejnw

sin ( n g w )
(.%U)

ejnw = s

(34) where n = k 1 (recall that bf is the fractional bandwidth). The partial differential of eqn. 34 w.r.t. s is simply
-

This work was sponsored by the United States Department of the Navy under Air Force contract F19628-95C-0002. Opinions, interpretations, conclusions and recomendations are those of the author, and are not necessarily endorsed by the United States Air Force.
8
References
COMPTON, R.T.: Adaptive antennas (Prentice Hall, 1988) HUDSON, J.: Adaptive array principles (Peter Peregrinus, London, 1981) LEE, H.B.: Eigenvalues and eigenvectors of covariance matrices for signals closely spaced in frequency, IEEE Trans. Signal Process., 1992, 40, pp. 2518-2535 HAYWARD, S.D.: Effects of motion on adaptive arrays, IEE Proc. Radav Sonar Navig.,1997, 144, pp. 15-20 ZATMAN, M.A.: Production of adaptive array troughs by dispersion synthesis, Electron. Lett., 1995, 31, pp. 2141-2142 PAULRAJ, A., ROY, R., and KAILATH, T.: A subspace rotation approach to signal parameter estimation, Proc. IEEE, 1986, pp. 10441045 ROY, R., and KAILATH, T.: ESPRIT - estimation of signal parameters via rotational invariance techniques, IEEE Trans. Acoust. Speech Signal Process., 1989, 37, pp, 984995 BARABELL, A.J.: Improving the resolution performance of eigenstructure-based direction finding algorithms. Proceedings of ICASSP-83, Boston, MA, May 1983, pp. 336-339 RAO. B.D. . and HARI. K.V.S.: Performance analvsis of rootMUSIC, IEEE Trans. Acoust. Speech Signal Proceis., 1989, 37, pp, 1939-1949 WARD, J.: Personal communication BANGS, W.J.: Array processing with generalized beamformers. PhD dissertation, Yale University, New Haven, CT, 1971

i.e. the klth element of the target signals covariance matrix assuming an SNR of 1. After working through the product-quotient form of eqn. 34, its derivative w.r.t. o is
ark1 __ dW

n+eJnw sin ( n 2 u ) b
--s
(&W)

(36)

8 9 10 11
90

From eqn. 35, the partial derivative of R w.r.t. the target signals power is simply the target signals covariance matrix with an SNR of 1. For the partial derivative of R w.r.t. w, eqn. 36 must be computed for each combination of k and 1. Utilising eqn. 31, the elements of the FIM are then given by

IEE ProcRadar, Sonar Navig.. Vol. 145, No. 2,April 1998

Let us define C = J-I. The variance of w is then given by the first element of C, denoted cll. The RMS angle error in sin(0) space is given by

An N-element array produces N beams between sin(0) = -1 and sin(@ = 1. Thus the beam-split ratio is given by
n

beam-split ratio = N e

(39)

IEE Pror -Radur, Sonar Nuvig

Vot 145, N o 2,April 1998

91

Das könnte Ihnen auch gefallen