0% found this document useful (0 votes)
73 views

Frequency-Wavenumber Array Processing

.

Uploaded by

Fakhrur Novianto
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
0% found this document useful (0 votes)
73 views

Frequency-Wavenumber Array Processing

.

Uploaded by

Fakhrur Novianto
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 6

Frequency-wavenumber array processing

Melvin J. Hinich')
FirginiaPolytechnic
Instituteand State Unioersity.
Blacksburg.Virginia24061
{Received21 August 1980;acceptedfor publication14 November 1980)

Mostarraysignal
processing
systems
usedelay-and-sum
beamforming
toestimate
source
bearings.
Thispaper
demonstratesthe closerelationshipbetweenbeamformingand frequency-wavenumber spectrumanalysis.
The latterapproachhascomputational advantages overbeamforming whenthenoiseis spatiallycorrelated.
The wavenumber approachis usedto'derivethe array response
of a generallinearor planararray to plane
wavesignals.Thestatistical
propertiesofthe maximum-likelihoodestimatorsof sourcebearingandamplitude
arepresentedfor an arraywith manyelements.Optimalarraydesignis alsodiscussed.

PACS numbers: 43.60.Ok, 43.30.Vh

INTRODUCTION K = (%/c) cos0 '


Supposethat an array of sensors is receiving coher- B(t,O)=Aexp(i•ot)
• exp[i(g
-
ent radiation from a distant source. Most array signal
processing systems use delay-and-sum beamforming to
estimate the sourcebearing.t For sonar arrays, accu- =• s(t,x,)
e•i•) . (3)
rate bearing estimation of single or multiple targets is
the main design goal; array design and signal process- • other words, bemform• is the sine as comput•
ing are subproblems. the spatMl Fo•ier transform• the M smuts from the
array. The spatM•fr•uency g=(%/c) cos8corre-
There is a close relationship between beamforming sponds• •e look angie •.
and frequency-wavenumber Fourier analysis, which
has computational advantages over beamforming when • actml practice, a beam is computedfrom a f•ite
the noise is spatially correlated. This relationship is record of the M e•els, the bern outp• is fDter•-•
described in the first part of this paper. Then I show a •rrowb•d a•ut %, andthe filtered sM•l is
how to use the wavenumber approach to easily calculate •mred andsmmed to givethe averse energy• the
the array response for any linear or planar array ge- b• got •e • set. • fr•uency-wavenmber •l-
ometry, and to mitigate the jamming effects of a co- ysis•eache•el is fiRer• •d thentilespatMl
herentinterferingsignal. The final secti6ndealswith Fo•ier tr•sform is compute. ff the receiv• sMnal
bearing estimation of a broadband wave when the noise is a wave, the •mre of the m•nit•e of this tr•s-
is spatially correlated. form•s a p• of heMht(M B I) wheng=go
= (dc) cosOo.
I. WAVENUMBER SPECTRUM AND BEAMFORMING
Now consider a p•r array of M sensors, and a
Let us begin with the simplest model: A linear array p•e wave
of M sensors and a single frequency plane wave signal
-- [. [. xneos0o+yn
m complexvariableform. Let Oodenotethe wave'sdi-
rection of arrival with respect to the array axis, let c
denote
thewave'svelocity,andA= {A]exp(iq•)
is its where • is wav&s directionwi• respectto •e x •is,
amplitude. The s•gnal at the kth sensor when there is ' andthe kth sensor is locatedat {x•, yJ. The s•l •
no noise is a be• po•ted •t a•le • is
s(•,x,)=Aexp
[iwo(t
- x,cos
Odc)
]. (1)
wherex• is the locationof the kth sensor(x•<xa<ßß'
<xu). The signalin a beampointedat angle0 (and-0)
is from (4) •d (5) tMt

B(t,01=• s(t+,,.x,). (21 B(t,O)=• s(t,x,,y) e•[i(g•x,+•0], (6)


wherethekthdelayis r,=x, cos0/c. Sincea linearar- a• lB(t,8){a•s a p• ofheMht (M{A[)awhen•
ray cannot identify between eoand -•o, let us arbitrar- =(o•c) coseo•d •={mo/C)s•O• Th• B(t,•) is •e
ily assume that 0o> 0. •o-d•ensio•l s•t•l •o•ier transform • the •.
Since the wavenumber component on the array axis is • terms of •e wavenmbercomponents,
•= t•'ag•/•.
%= (•Oo/C)
cdS0o,it followsfrom (1) and(2) that for II. ARRAY RESPONSE FOR RECEIVED WAVES

a)Thia work was supportedby the Office of Naval Research Let • •e •e wavenmber appr•ch to compute •e
(Statisticsand Probability Program) under contract. res•nse of a given array, s•rt• with a l•ear geom-

732 J. Acoust.Soc.Am. 69(3), March1981 0001-4966/81/030732-06500.80 ¸ 1981 AcousticalSocietvof America 732

Downloaded 16 Jul 2010 to 146.6.201.36. Redistribution subject to ASA license or copyright; see http://asadl.org/journals/doc/ASALIB-home/info/terms.jsp
etry. To simplifythe discreteFourier transforms,
selectthe origin of the array axis so that x• = 0. Let us
make the modest assumption that there exists a dis-
S(
m)=•. Aexp
(-i2• expi
3• ,
rance d suchthat x,= n,d, where nk is a positive integer =HA, ff m=/o,
for eachk= 2,...,M. This will hold ff the x, are ra-
tional numbers. =0, ffm•l o. (11)

The question of array ali•sing (grating lobes) now Normalizing by settingA = 1, it follows from (9), (10),
and (11) that the responseis
arises? Let D denotethe greatest commondivisor of
the integers{nz,..., n•t}. From HinichandWeber,3the
discrete Fourier transform DFT (3) is periodic with
period 2g/Dd. Its principle domain is either -•/Dd
B(0,)=•
•j)exp(i2•r
j(/•lø)) (12)
The peak response is
• g • 7ffDd or 0 -<• -.<2g/ Dd, dependingon the convention
used. This means that a real wave is not aliaseal if its
wavelength
ko=2•rc/O•o>•2Dd.
For example,ff xx= 0,
x2= 11d, x3= 19d, and x= 30d, the wave is not aliaseal ff
which is clearly •epe•en• of •e sp•c[• of •e M
Xo>2d. Assume that the n, are such that D= 1.
Lets(g)denotethefilteredplanewaveA exp[-i(2n/Xo)x
ff •ere does not exist sucha lo for • given N, then
x coseo]at a pointx on the axis. Definethe counting
•e pe• responseis smeared over several adjacent
sequence
waven•bers. This smear• can be elected or re-
•(j) = 1, if there is a sensorat duc• • e•e• the •) s•uence
= 0, otherwise. (?)
thus •creasi• N. ff N is s•ficienfly large, •ere
istslosucht•t [2glo/N-go[is eEherzeroor ve•
From (3), the filtered transform can be written smaa,

Let us now t•n to the response of a p•r array.


o): exp(ijJ), (8) A•l•ous • •e l•ear 'geomet•, ass•e
where N an•. This transform can be computedfor the = n• a• y,= m• where n• andm, are positive integers
for k= 2,...,M, a• x•=y•=0.
grid {•t= 2•l/Nd: l = 0,1,...,N - !} usingthe FFT
algorithm. Defining L to be the integer closest to Let s(x,y) denotethe filter• p•ne wave s•l at lo-
to, theassociatedlookanglegrid is {e•= cos'•(ko//Nd) cation(x, y). From (6), the filtered •o-d•ens•o•l
if O.<l•L, •mde•=cos-•[lo(/-N)/Nd]if N-L </<N-1}. transform b g•ven by
For exampleeo=•/2 ande•= cos-•(1)=0 if Nd is divi-
sible by Xo.
These arccosineeqnationsgive a mappingbetweeathe
• anda set of N lookanglesin the interval[0,n] that
canaccomodate anyXo>•2d. For exampleif Xo=2d,
thenL = N/2 (N even)or L = (N - 1)/2 (N odd).
where •a• • •nd
Since the right-hand side of (8) is

= 0• o•e•ise.
- m)s(m)
•,•=0,1, .... •-l}, it followsfrom(Z3)
where

(9a)
B(e,..)=• • n(1
-m,k-u)S(m,.), (14)
where

• [i2•rjm\
(9b) and

then

(10) As lo• as k•+ l•=L •, the look a•le is as follows:

The responseat lookangie 0• is definedto be B(O•)for =•n'•k/(1-N), • N-L <I•N-I.O•k•L.


the wave signal. The beam power pattern is the se-
quence
(
For example, supposethere is an integer lo suchthat =[an'•(k-N)/(l-N), EN-L I<N-1,N-L<k
eo=cos'•(Xolo/Nd)
or 0o=cos'•[ko(l
-N/Nd]. Then •<N-1. (16)

733 J.Acoust.
Soc.Am.,Vol.69,No.3, March
1981 Melvin
J. Hinich:Frequency-wavenumber
arrayprocessing 733

Downloaded 16 Jul 2010 to 146.6.201.36. Redistribution subject to ASA license or copyright; see http://asadl.org/journals/doc/ASALIB-home/info/terms.jsp
Theseequations
providea mappingbetWeenthe/{,tand look angles, then the geometry must be circular. The
%•anda setofN: lookanglesintheinterval(0,2•r) analogyto the tWosubarraydesignis a disk where the
that can accomodateany wave with Xo>•2d. sensorsare placedaroundits outer edge. In generaI,
we wantto select the r(j,j •) to maximize the concavity
For example,supposethat the signaliS exp[-i(2*r/Xo) of the main lobe. Using similar calculations as were
(x ½os8o+y sin8o)
] , where>,•zcos8o=lo/NdandX•xsin8o used to derive (19), we want to maximize.
=ko/Nd for some(lo,ko), i.e., lo=L ½os8oandko=L
sin8o. Then from (15b), S(m,n)=N 2 for m= lo,n=k o
and S(m,n)= O, otherwise. From (14), the responseto
Z Z
y_-o .4'--o
Z
$_-o y'--o

this wavefor lookangleSt.• is


(21)
B(8,.•)=a(l - lo,k - ko). (17)
where
III. ARRAY DESIGN N-1

If there is no a priori knowledge


of the bearingsof J=M't•'• Z (22a)
y=o
sources of interest, it will now be shown that the opti-
mal geometry for a fixed aperture linear array with a
fixed number of sensors is tWo subarrays, one on each (22b)
endof the axis? Each subarray has its sensorsequally
spaced. By optimal, I mean the designconformingto and

the constraintson aperture and sensorsthat gives the


maximumaccuracyfor the estimateof a sourcebearing •. jj'r( j, j'>; (22c)
when noise is present. As is shown in Hinich and $=0
Shaman,
s the accuracyin terms of meansquarederror In the next section, we deal with estimating source
depends
onthesharpness
(theconcavity)
of IR(K)[:at bearings when the received signal is a sum of coherent
K= 0, where waves plus noise.

R(g)
=• r(j) exp(igjd)
. IV. SIGNAL PLUS NOISE
Thus we want to select the r(j)'s to maximize The major advantage of the frequency-wavenumber
approach is the ease with which it handles spatlally
c(0)=(a'ya, I. (18) correlated noise. Spatlally correlated noise makes the
It is easy to show that signal-to-noise ratio (SNR) direction dependent. Com-
plicated calculations are needed to achieve optimal
tO) frO)- jr(j) , bearing estimation using time delay methods in this
t--o •=o casefi The calculationsare relatively easy for the
½o-gapproach.

N-I N-I Once again it is easier to explain the method for a


linear array and a single frequency wave. Suppose that
Simplif7ing roarers by making 3// even, it follows from the signal at x} is a plane wave plus stationary, zero-
(1•) that ½(0) is maximizedby settingr(y)= 1 for y= 0, meanGaussiannoisedenoted½(t,x•). Filtering in a
...,M/2-1,...,N-1 andr(y)=0, otherwise. This is narrowband about Wo, the signal is
the tWo subarray geometry mentioned above. From
(12), its beampatternfor a wavewith direct[on8o is s(x•)=A exp[-i(2•r/Xo)X
• cOSSol
+ ½(x}),
= A exp(-i•ox•) + ½(x•), (23)
where ((x,) is the filtered noise. If the noise is spa-
=4cosZ
[r(•_-•)(/''3sin2[•rM(l-to)/2N]
sin---Zo-WY ' ttally correlated, then the wavenumber spectrum of the
noise is not flat. Let us go into this in some depth.
(•.0)
If N >>M, the peakto sideloberatio at g -• % • 2•/Md In most applications the•signals are transient, and it
is approximately (2/•)%-3.9 riB. At g-• Ko• 6•/Md, is then easy to observe the noise alone at each sensor
the ratio is approximately (2/3•)•: -13.5 dB. prior to the onset of the signals. From (8), (9), and
(10), the spatial Fourier transform of the noise along is
, By appropriately
taperingthearray (weighting
the
channels), the peak to first sideloberatio can be re-
dueed, but at the expenseof the height and width of the
$(8,)--NqZ R(I-m)U(m),
rn=O

main lobe. This means that a tapered array gives a where


less accurateestimateof 8owhennoiseis presentthan
doesthe tintaperedarray discussedabove. This array m.z /i2•rjm\
gives the best resolution and accuracy for weak sources U(m)=
Z ,<jd)exp[--•). (24>
when M and N are large.
WhenN is large, {U(0),..., U(N-1)} are (approximate-
These results are easily extendedto planar arrays. ly) uneorrelated complex Gaussian varigres with zero
ßIf we want the array's responseto be the same for all means.? Moreover,

734 J. Acoust.Soc. Am., Vol. 69, No. 3, March 1981 MelvinJ. Hinich: Frequency-wavenumber,array
processing 734

ownloaded 16 Jul 2010 to 146.6.201.36. Redistribution subject to ASA license or copyright; see http://asadl.org/journals/doc/ASALIB-home/info/terms.js
EN" lain)p-•s.(•.), uncorrelatedx•variates. Thecumulativedistribution
function(cdf) of this maximumis {F{X)]•,where
thewavenumber
spectrumof the noiseat %= 2•m/Nd.s F(x) is the cdf of a X• vat(ate. The cumulative
Thus
rejectedat the a level if X>xo, wherexosatisfies
E[B(9,)[=•-N
'tZ IR(I-m)]=S.(K=). (26) a= t - [•Xo)]". (3•)
Assume that the array has a sidelobe pattern similar
Notethatin (30), IB(9•)12is divided
bythenoise
wavenumberspectrum. This division is the spatial
to the one given by (20). SinceR(0)=M, it follows from
analogyof prewhitening.If the beamoutputsare to be
(26) that
½p-i )/2
visually inspectedfor the presenceof a wave, then
P(O•)=]B(O•)[•/S,(g•)
should
beplotted
for l=O,...,
N-1.

wherep=(N/M) is assumedto be oddfor simplicity. Suppose thatP(O•)hasa peakat 0•othatis statisti-


Assuming$,(•) is slowingvaryingin the 2gp/Nd•-29/ callysignificant.ThenO•ois thenaturalestimateof
Md hand about g•, 0o. If (2v/lo)cosO
ois notequalto a %0'butfallsbe-
tweea two grid points, there is a quantizationerror in
(28) d•ooforder1/Nd.
Using the complexGaussiannature of the U{m), it can
WhenN is large, the root mean-equateerror of •to
due to noise is approximated by
beshown
thatthedistribution
of 2]B(0•)]2/MS,(%)
is
approximately chi-squared with two degrees of freedom rmse0•o
-•Xo/2•(2•pM)
t/ •x•sin0o, (32)
(X} andthusthevariance
ofM 'x[B(d•)[2is approxi- where
matelyS•g•). Thismeansthatwe mustaveragethe
M 'x[B(O•)
[• in somewayto obtain
anaccurate
estimate
of s,(•,).
If the noise is truly stationary, an accurate estimate is a geometryfactor, x• is the array length, and p is
canbeobtained byaveragingM '• ]B(•) Izcomputed at the powerSNRin a narrowband
about•o.'ø Since
differenttimes. For example,if M'• IB(•) I• is com- limMt/•rmse0•o
is equalto theCramer-Rao
bound
for
putedfor each I from 100 nonoveriapp'_mg
records of the the asymptoticvarianceof a consistentestimatorbf
array output,thevarianceof theaverage
is S•)/100. Let usassumethatM,N, andthesamp- •o, 0•ois approximately
maximum
likelihood
forlarge
M and N. For a large aperture array with many sen-
ling time are sufficiently large so that we can treat
sors, this bearingerror component
is oftenless than
$•(•) as kuourn
for i=0,...,N-1. the quantization error.
Returningto the signal, it followsfrom (12)• (23)• Now consider the problem of estL-natingthe complex
and (28) that
ampUtude
A. If thenoiseis spaticilywncorrelated,
it is
E,r' IAI=M"I•(t -to)p+ s.(•,), (39) shownin the Appendixthat the maximum-likelibood es-
timator of A is
if there is an lo suchthat 0o=cos'X(xo/•/Nd).
Since
IAI'+ thenorma/izedbeam ,•=(1/M)B(O,o)
. 433)
pattern
hasa peak oforder•/IAI=against
a backgro•d Since
EB(0•o)•AR(0)=AM,
itfollows
that
,• isunbi-
oforder$,(•) Ior •helookangle0•o=do,provided
that ased,i_e.,EA=A. From(29)thevariance
of,• is
• IAp>>s.(•)forg ina band
about
It is often useful to have a test statistic to,determine EI• -A 12=
•/M, (34)
the statistical significance of the maximum energy peak where•=
in the beam output. Let the null hypothesis be A = 0,
i.e., the peak is due to noise alone._ Consider the test If $.(•)isnotflat,,• isstillunbiased.
t• Itsvariance
statistic is approximately given by

21B(0•)I • E1,•- A1'=[$.(%o)/M]. (35)


x=max '•S.(g•)' (30) Thus,• is a preciseestimator
of A whenM is large,
and is useful in removing the effect of a coherent
jamm/ng signal.
which is analogous to the statistic of the Fisher test
for the presenceof a sinusoddin additive noise.9 The The planar array processingis a simple straightfor-
distribution of X is needed to compute the threshold for ward extension of the linear case, using two indices
an a-level test of the null hypothesis. This distribution andtwo sums. For example,applying(14), (15), and
is hardto obtainsincethe beamoutputsB(O•)are cor- (17) to (33), the estimatorof A usinga planararray is
related. But if the sidelobes rapidly diminish as M in-
creases, as is the case for the •ptimal array or an • =(1/•)s(• ,o.•o), (•6)
equaUy
spaced
array, {B(O•):j= 1,.... M} are uncor- where(lo,ko)jointlymaximize{B(0•,•){2/S,(•=•,
•),
related for large M. Then the distribution of X is approxi- andS,(•=•,•=m)
is thenoise'swavenumberspectrum at
mately the same as the distribution of the maximum of M (•,o•)- Thevariance
of,• is approximately
735 J. Acoust.SOC.Am., Vol. 69, NO.3, March1981 MelvinJ. Hinich: Frequency-wavenumber
arrayprocessing 735

Downloaded 16 Jul 2010 to 146.6.201.36. Redistribution subject to ASA license or copyright; see http://asadl.org/journals/doc/ASALIB-home/info/terms.js
ß/"S.(K.,o.
%0)' endeffects
are negligible
if xu/c<<T. Thenfrom(40),
NT-!
Theestimatorof 0•(0%,%)
is givenbytheappropri-
ateequation
in (16)withl andk replaced
byloandko. %(xn)
=N•,
• E s(n/x
,xn)
exp(-iw•na)
n=O

When
N is large,x:0%.%
is approximately
maximum
likelihood and =A(%)exp(-i%x•)+%(x•), (42)
rmseOzo.%
• Xo/2•
(g•PM)
•/2x
u' (37) where%=(wt/c)cos0
oand½t(x•)
is a zero-mean
com-
piex
Gaussian
variate.
Itsvariance
forlarge
Nr is
V. BLOCKING A JAMMING SIGNAL
E I(fixa)I:---N;tS,(co•), (43)
Supposethe filtered signal for the kth sensor is
whereS,(w), the powerspectrumof the noise, is inde-
s(xa)=A exp[-i( 2n/Xo)X
acos0o] pendentof xa. Thus the DFT of the receivedSignals
yields H single frequency waves plus filtered'noise.
+Ar exp[-i(2;r/Xr)x
acosOr]+ •(xa), For each j= 1,...,H, compute
whereAt, Xr, and0r are the amplitude,wavelength,
and direction, respectively, of a wave that is interlet- B(O,,%)=•.%(xa)
expi•,xa
•=1
, (44)
hagwiththe waveof interest. If Xr•Xo, thenall or
most of the jamming energy will be filtered out. Thus for the •l grid discussedin See. 1I. Concentratingon
let Xr = Xo,andassumethatA• and0r are unknown. bearing estimationlet•$=8•o(W fl denote
thelook angle
ff IArI>>IAI, thenthebeampattern
will bedomi-
associatedwith the maximum ]B(•.,.,fi
}:/S•(•).For
large M and N the maximum-likelihood estimator of
natedbythejammer. If thisis thecase,then0r is. •o, denoted
•o, is approximated
by
estimated
bytheangle0,rsuch
thatIB(O,•)
I2/S•(•c•r)
ts
a maximumfor l = 0,..., N - 1. The amplitudeA• is
estimatedusing(33) with l• in placeof lo. /

The response of the jammer can be then subtracted where(r•is thevariance


of•.x• Using
thelargesample
fromB(O,),andtheadjusted'
beampatterndefined
by approximationgivenby(32)with
=TulA•}•/S,(%)andXoby
I la(0,)-rn(l- l,)I
canbe usedto estimatethe bearing•o.•s ß J:l

If JArI is oftheorderof ]AI, weneed


toknow
e, and its root-mean-square error is
(if 8r½8o)to blockthejammer. If so, compute
(38)
withlr astheinteger
thatmakes
a•rclosest
toar.
VI. BROADBAND
SIGNALPROCESSING The SNRp• is estimablesinceA(wfl is preciselyesti-
Until nowthe signalhas beenassumedto be a single matedbyM'zB(O•o,
wflwhenM is large.
frequencyplanewave. To exposfiethe processingOfa Sincea maximum-likelihood
estimatorhasminimum
broadbandwave, let the array be linear and let •ean-square error whenthe samplesize (M) is large,
s(t,x•)=s(t- c7•x•
cos0
o)+ •(t,x•), (39) 0o is optimal in a mean-squaresensefor nonsparsear-
rays. There is no need for ad-hoc bearing estimators
wheres(t) is a bandlimited
signalwhoseupperfrequen- for such arrays.
cy is wu. Once again the noise is assumed to be Gaus-
sian andto be stationaryin time and space. This paper has shown'the connection between beam-
forming and frequency-wayønumberspectral analysis
All signalsare transient. Selectthe time origin so using discrete time and space measurements. A de-
thats(t)=0 for t< 0 andt> T, whereT is the signal signer of a robust and effective array processingsys-
duratio.n. Let H be the largest integer less than or tem shouldhavea eomp'leteunderstandingof the re-
equalto Twa/2•. The signalhasthe simpleFourier lationshipsbetweenphysicalmodelsof propagating
representation waves, backgroundnoise processes, and the statistical
properties of øst(rectors of the key parameters in the
s(t)=Z A(wflexp(iw•t)
, (40) signal models.

wherew•= 2•j/T and


APPENDIX
A(wfl=• s(t)exp(
-i wit)dr. (41)
If % is known, the maximum-likelihood estimator of
AssumeA(0!= 0. A for the statistical model (23) is given by (33) when
Suppose thats(t,x•) is sampled
at timest•= nZx(n=
0, [•(x•)] are independent
Gaussian
N(O,•) var'mtes.
ß.., Nr - 1)where• = •r/w,andNr =(T/•).x• If cos8
o -Proof: Since the errors are Gaussian, the least
> 0, part of the leading edgeof the signal is lost for k squares estimator of A is maximum likelihood. x• The
•1. The trailing edgeis lostwhentosco<0. These least squares estimator is

736 J.Acoust.
Soc.
Am.,
Vol.
69,No.
3,March
1981 Melvin
J.Hinich:
Frequency-wavenumber
array
processing
736

ownloaded 16 Jul 2010 to 146.6.201.36. Redistribution subject to ASA license or copyright; see http://asadl.org/journals/doc/ASALIB-home/info/terms.js
lin, andC. vanSchoonereld
(Academic,NewYork, 1973),'
pp. 577-590.
?D.Brillinger,TimeSeries,DataAnalysis
andTheory(Holt,
r (j)s(jd) exp(iKava) Rinehart and Winston, New York, 1975), Sec. 4.4.
E exp(-i•oX,,)
exp(i•oX•) 8Thisspectrum is theone-dimensionalspatialanalogto the
spectrumof a Stationary
discrete-timestochasticprocess.
See C. M. Bennett, "A Directional Analysis of Sea Waves
from Bottom Pressure Measurements," in Transactions:
Ocean Sciences and Engineering of the Atlantic Shelf (Marine
TechnologySociety, 1968), pp. 71-87.
sp. Bloomfield,
FourierAnalysisof TimeSeries:An Intro-
duction(WHey,NewYork, 1976), Chap.5.
•øLe•inheuristically
derivesthemaximum-likelihood
bearing
•B. D. Steinberg,Principles of Aperture andA•ray System estimator and its properties for a general three-dimensional
Design(WHey, New York, 1976), See. 5.4. His use of array, assuming Gaussian noise. M. J. Levin, "Least-
symbolsdiffers from mine. For example, he uses 00 to Squares Array Processing for Signals of Unknown Form,
denote the direction of arriwal from the array normal. Radio Electron. Eng. 29, 213--222 (1965). These results
See also C. S. Clay and H. Medwin, Acoustical Ocean- are rigorously derived for a uniformly spaced array by
ography(WHey, New York, 1977), Sec. 5.3.2. Hinich and Shaman, Ref. 5. The Cramer-Rao bound for
2Aliasingis definedin M. J. Hialch, "ProcessingSpaticily the asymptotic bearing variance for a linear array is given
AliasedArrays," J. Acoust. Soc. Am. 64, 792-794 (1978). by V. H. MacDonald and P. J. Schultheiss, "Optimum
(Also see Steinberg,Sec. 5.2, Ref. 1). Passive Bearing Estimation in a Spat(ally Incoherent Noise
3M. J. Hinich and W. E. Weber, "Determination of the Nyqutst Environment,,' J. Acoust. Soc. Am. 46, 37-43 (1969). These
Frequency for Unequally SpacedData," ON-RTech. Rep. 17 rigorous derivations match Levin's results.
(revised), Virginia Tech. (1980). lithe maximum-likelihoodestimator of A for correlated noise
4Cartershows
thatthisdesign
minimizes
theasymptotic
bearing is given by J. Capon, R. J. Greenfield, and R. J. Kolker,
variance for a linear array with a given minimum sensor "Multidimensional Maximum-likelihood Processing of a Large
spacing. See G. C. Carter, "Variance Boundsfor Passively Aperture SeismicArray,,, Proc. IEEE 55, 192-211 (1967).
Locating an Acoustic Source with a Symmetric Line Array," 12C.S. Clay,M. J. Rinich,andP. Shaman,
"ErrorAna/ysis
J. Acoust. Soc. Am. 62, 922-926 (1977). This design is also of Velocity and Direction Measurements of Plane Waves
optimal for range estimation. M. J. Hinich, "Passive Range Using Thick Large Aperture Arrays," J. Acoust. Soc. Am.
Estimation Using Subarray Parallax," J. Acoust. Soc. Am. 53, 1161-1166 (1973).
65, 1229-1230 (1979). 13This
is equivalent
to steering
a nullat 0j, andthenesti-
5M. J. Hinich and P. Shaman, "Parameter Estimation for An mating 00. See V. C. Anderson and P. Budnick, "Reflection
r-dimensional Plane Wave Observed with Additive Indepen- of a Coherent Arrival of an Array,,' J. Acoust. Soc. Am. 45,
dent Gaussian Errors," Ann. Math. Statist. 43, 153-169 406-410 {1969).
(1972). I4inthis paper, Ix] denotes
theintegerclosestto x.
•G. C. Carter and C. B. Knapp, "Time Delay Estimation," in 15ff• and•2 are maximum-likelihood
estimatorsof•0 com-
t•roceedings of the 1976 IEEE Conference on Acoustics, puted fromtwoindependent
samples, the(q]2+•2).t (q•2•t
Speech
andSignalProcessing(IEEE, NewYork, 1976), +•2•2)is maximum-likelihood
forthecombined
sample.
pp. 357-360. W. J. Bangs and P.M. Schultheiss, "Space- Its varianceis (/r•2+•2)-•.
Time Processing for Optimal Parameter Estimation," in IøC.B. Rao,LinearStatistical
Interference
andIts Applica-
Signal Processing, edited by J. W. R. Griffiths, P. L. Stock- tions (WHey, New York, 1965), Chap. 4.

737 J.Acoust.
Soc.Am.,Vol.69,No.3, March1981 Melvin
J. Hinich:Frequency-wavenumber
arrayprocessing 737

ownloaded 16 Jul 2010 to 146.6.201.36. Redistribution subject to ASA license or copyright; see http://asadl.org/journals/doc/ASALIB-home/info/terms.js

You might also like