GEOPHYSICS. VOL. 44, NO. 2 (FEBRUARY 1979): P. 179-184.

A signal processing method for circular arrays

John D. Henstridge*

The theory of spectral representationsof stationary random processescan be a useful tool in signal pro-
cessing. Using this theory, it is possible to derive simple methods of estimating the phase velocity of surface
waves from observationsmade with a circular array. The methods are totally nondirectional, thus allowing the
use of microseismsfor exploration seismology. Furthermore, the methods can be extended to yield directional
information about both correlated and uncorrelated signals.

INTRODUCTION t is time and t is the location in the plane. These will

Measurement of the dispersion of surface waves is have a spectral representation (see Yaglom, 1961),
potentially a method of gaining information about
the structureof the earth immediately below the point
of measurement.
ToksGz (1964) investigated this by observing the
dispersion of microseisms. However, when using
where w is the frequency (in radian\ per unit time),
most arrays, this problem is made difficult by the
K the vector wavenumber (in radians per unit dis-
need to estimate simultaneouslythe direction of travel
tance), and 5’ is a random spectral process wsithun-
and the velocity of the microseismic waves, an expe-
correlated increments. While this representationwas
cially difficult task when waves are traveling in
derived by Yaglom as a purely mathematical result,
several directions. Aki (19.57, 1964) pointed out that
it can be given a physical interpretaticln.It statesthat
the use of a special design of array, consisting of
any stationaryprocess in time and space can be con-
seismometers equally spaced on a circle and one
sidered as a continuous sum of independent waves
seismometer at the center, could simplify processing
with different frequenciesw and wavenumbersK. The
to overcome these difficulties.
random spectral measure 5’ gives the amplitude and
In this paper, we first provide a mathematical back-
phase of each of these waves. While all stationary
ground for the problem by employing the theory of
processescan be represented in this way, we shall
stationary stochastic processes and their spectral
see below that it is not always the mo\t useful repre-
representations. With this background it is then
sentation.
shown that Aki’s method, which involved averaging
It will be convenient to consider thi\ representation
the observed correlations between the center seis-
in polar form. If 5 = r(cos 0, sin 0) and K = k(cos c$,
mometer and each seismometeron the circle, can be
sin r$), then
extended to the situation where waves arriving from
several directions are correlated. Furthermore, a m cc 7r
z(t,r,O)= exp{- iwt - irk.
method is given for testing the assumptionswhich JII 0
must be made in applying the technique. The final
. ,cS -;I] {(dw. dk, d4).
section gives some simple examples.
We can make an assumptionthat at each frequency
STATIONARY PROCESSES IN THE PLANE
CO,the energy is concentratedat a single wavenumber,
We begin by considering stationary random pro- that is, the velocity is a single valued function of
cesses z(t, [j defined in time and the plane, where frequency. This corresponds to the \pcctral process

Manuscriptreceivedby the EditorJune27, 1977;revisedmanuscriptreceivedJuly 24, 1978

*Dept. of Mathematicc,Univel-sity(of‘Hestem Austrdiia,?+zilanu’s‘&
, ‘. A. 6889, Ausrraiia.

179
180 Henstridge

5 being concentrated on a curve [w, k(w)], and extra seismometer at the center. At each frequency
we have of interest, the coherence between each of the seis-
X 7r mometers on the circle and the one in the middle was
z(t,r,0)= exp{-iwr - irk(o). averaged to obtain an estimate b(w. r) of the aver-
li aged complex coherence. Then an estimate l(o)
. Zs(B” qb)}<(dw, d$). can be formed for the wavenumber by letting

The assumptionabout the velocity is a key one. Later

J,[rk^(o)] = Real(g(w, r)].
we give a method which tests whether it holds in
practice. In practice, we would want to be able to observe
A frequency-direction spectral density f(w. c$) as wide a frequency range as possible while keeping
can be defined by the radius of the array fixed. The function J,[rk(w)]
is one-to-one for t%(o) over the range 0 to ~3.8.
But in practice, not all of this range can be used since
which gives the average energy at frequency w arriv at each end any errors are greatly magnified because
ing from direction 6.
Variance[k(w)] = Variance[c(w, r)] .
Using this density, we may define a spatial covari-
. {rJ1 [rk(w)]}-*.
ante function
This usually means that rk(w) must lie in the range
T(w, r, 0) = * exp{irk(w)cos(O - 4)). .4 to 3.2, or alternatively the wavelengths must lie
I -77 between 2r and 15r.
Aki suggestedthat p(w, r) could be estimated by
considering the coherence between the center seis-
which measures the covariance at frequency o be-
mometer and the average of the outputs of the seis-
tween the signals observed at (say) the origin and
mometers on the circle. We now show that this is
(r. 0). It is a quantity easily measured in practice
not correct, but it does lead to a method of testing for
using either analog or digital correlation techniques.
nonpropagating noise or multiple velocities.
We shall be more interested in an averaged covari-
We shall define the center sigal by
ante function,
x,(t) = =(r, 0, O),
and the circle signal by

Jo(t) = (27r-’ li -_(t, r, O)d/3.

I -77
Expressed in terms of the spectral process 5, they are
r
x,(t) = exp{ -iwt} _: <(dw, d+)>
I I
and
where J, is the zero-order Bessel function of the first
kind and y”(t) = 1’ ST (2~)~’ IX exp{-iwr

f”(W) = I” f(w, $)d&. -yrkLi) cos(t3 --i)jdO [(do, d$)

pli
= x exp{-imt}J,[rk(w)] .
It can be seen that the averaged covariance function
J -cc
is not dependentupon the directional properties of the
process. It should make no difference whether there
-jr <(dwd+).
is a single signal or many. -77
The spectrum of the process at a single point in If we let
spaceis f,,(w) so that the averagedcomplex coherence
would be p(w, r) = J,[rk(w)]. Aki (1957) suggested n 5(w> d+) = d<,(w),
I -7r
using this fact by having an array of seismometers
equally spaced on a circle of radius r and having an then these simplify to
 Circular Arrays 181

x0(t) =
Jx
-CC
exp{-iwt}d<,(w),
,?I=--r
and and

).0(r) = x exp{-iwr}J,[rk(w)] d<,(w).

I --3c
This implies that the circle signal Y”(r) can be re- Furthermore, the original process :(t, r, 0) can be
garded as a linearly filtered form of the center signal expressed in terms of
X,(I), with gain function J,,[rk(w)] (see, for example, z r

Koopmans, 1974, p. 86). Except when J,[rk(o)]

0, the complex coherence between the two signals
= z(t, r, 0) =
J 2
--r ,)1=--r
exp{ -iwr} *

will be 1 or -1 depending on the sign of J,,[rk(w)].

Furthermore, the introduction of either noise or an- . (27r-’ 1: exp{-irk(w) .
other significant velocity will reduce this coherence
to a value less than one. . cos(0 - 6) + im&}d$d<,,,(w)
There are two practical outcomes of this. First, by x r

estimating the coherence between the center signal

and the circle signal (in practice obtained by averag-
=

I C
--r m=--r
exp{-iwr + im0).

ing the outputs of the seismometerson the circle), a

check can be made on the assumptionsnecessary to
Now we define what we shall call the rnth order
apply these techniques. Second, if the gain function
circle process:
is estimated [considering x,(t) as the input to the
Z
filter and y,(f) as the output], it could be used as an
y,(t) = (27T)-t exp{-im0}=(t, r, 19)dB
alternative estimate of J,[rk(o)], and hence k(w) I -?7
may be estimated. Any standard procedure for
estimating the gain may be used (see, for example, =
z exp{-iot}J,,,(rk(w))d5,,,(o).
Brillinger, 1975, p. 302). However, this method of I -x

estimating k(w) has one disadvantageover the method Hence, the circle signalsprovide a method of isolating
that Aki proposed-the seismometers would have the Fourier components of the spectral process.
to be precisely calibrated. If the seismometer gains Before examining this further, we shall introduce
vary, it is even possible to have an estimated gain a slight extension to cover some cases where energy
function which is greater than unity at some fre- arriving from different directions is correlated. Two
quencies. Obviously, no sense can be made of this. obvious cases are that of a signal and its echo and
However, as we shall see in the next section, a slight that of a scattering medium. In both these cases, it is
extension of the technique makes this disadvantage reasonableto assumethat the observed signals are the
worth overcoming. result of unobserveduncorrelated signals undergoing
some form of scattering which alters the directional
distribution of energy. The only model for which the
FOURIER-BESSEL ANALYSIS USING A
mathematics is workable is that where this scattering
CIRCULAR ARRAY
is linear and isotropic (that is, the same for signals
The termsfo(w) and &,(o) may be regarded as the arriving from all directions). With such a model, the
zero-order terms in the Fourier series of f(o, 4) spectral processof the observed processcan be repre-
and ((0, $), respectively. The higher order terms sented as a convolution of the original spectral pro-
will be defined by cess with some suitable function.
We define a scarrering jiincrion W(o,$,), and a
.fn?(w) = j--L exp{-im+If(w, +I@, new spectral process to, by

and CO(% 4) = J_;= W(w3 6 - $) 5(% Q).

i,(w) = 1-L exp{-im415(~, d4). This equation implies that d{“(w, 4), the observed
energy at frequency w from direction 4, is generated
Then the inverse relations are by summing the contributions of an original un-
182 Henstridge

scattered process 5. The sum is weighted by the

function W. Note that in general co does not have
Since we can assume that C,,(w) = I, the center
orthogonal or uncorrelated increments in I$. Hence,
signal will be defined as before. Then the covariance
our representation is no longer the standard one of
relations between the circle signals and the center
Yaglom.
signals will be
Usually W(o, I,!J), considered as a function of $,
would have a single main peak at zero, implying
that most of the energy is only slightly scattered. It
could be complex valued if the scattering involves
time delays. We shall assume that W(w, $j has a
Fourier expansion with respect to IJJ,that is

W(0, $) = (2~)~’ 2 exp(im$t) C,(w).

VI=-x
Since the unscattered process is not observed, we
may assume that C,(wj = 1. If the scattering is
significant and W(o, +) is thus a smooth function in
$, we would expect the higher order coefficients
C,,,(w) to be relatively small.
Since {“(o,$) is defined by a convolution, we E[x,(s)x,(s + t)] = m exp{iwt} f0 (w) dw.
I --oc
have

loco, d4) = (23--l 2 exp(im4) . As in the previous section, it seemsmost useful to

1)1=--a compare the center signal with each of the circle
. C,(w) 5,(w)4, signals. The bivariate signal [x,,(t), v,,(t)] has a
spectral density matrix G(w) given by

.fo(o)
G(w) = f

Note that in the case m = 0, G(o) is not affected

a z
z(t, r, 0) =
J x
--m m=-m
exp{--iwt + imf3) - by the presence of scattering [since C,(w) = I] and
hence the results of the previous section are indepen-
dent of this type of correlation in the observedsignals.
*J,[rk(w)l C,(w) G,(o),
and PRACTICAL CONSIDERATIONS
Consider a practical array with p seismometers
Y,(f) = m
I --m
exp{--iwt} . equally spaced on a circle. Obviously, the integral
definition of the circle signals cannot be used and
.J,[Mw)l C,(o)&,(w). instead a finite sum must be formed. If we let the
polar coordinatesof the seismometersbe (r, 27rup-‘),
Now we can proceed to examine the properties of
a = 1, 2, , p, then the rnth ol-dcr circle process
theae circle signals. First, we note that
will be
-__

jm(f) =p-l i z(t, r, 2nyt I) .

u= 1
. exp{-i8$ + im4’).
. exp{-i2nmu~~~‘}.
. E(Ww &J) 5kfw 4b’))
This need only be calculated for 111= -hp, , &p
E/T exp{-i(l-tn)}. since a form of directional aliasing will occur, and

Ym(l) = 2 Ym+ph4)
.f;o, $)d+ dw k
 Circular Arrays 183

The asymptotic values of the functions urn, 0,,

and S, are given by

*C m+Pk(W)4?n+pk(W). (T,(w) = 1,
0,(w) = -m&i,
However, J,(X) tends to zero very rapidly as n tends
to infinity so that this aliasing will not be significant and
for reasonablevalues of p except at those frequencies S,(w) = J,(rk(w))‘.
where J,[rk(w)] is close to zero. At those fre-
quencies, the higher order abasedterms will dominate In practice these functions can he calculated, and
and confuse the results. it would be easy to decide if they are of the above
We now consider the bivariate signal [x,(t), form. If so, the direction of the signal can be easily
v,(t)]. This may either be analyzed in real time using estimated from 0,(w). It is interesting to compare
narrow-band filters or alternatively be recorded and this method with virtually all other methodsof finding
processed using the fast Fourier transform. Either directions of signals. Most of them involve finding
way, an estimate of the spectral density matrix G(w) the maximum of some function (an estimated wave-
can be obtained (see, for example, Koopmans, 1974). number spectrum, for example) wjhile the above
Using this estimate of G(o) we may define the method only requires evaluation of an explicit
following functions: function.

the mth order circle coherence,

Example 2
cm(w) = lG12(0)/[G11(~)G22(~)1-~“, If two signals are present, the problem is much
more difficult. We would have
the mth order circle phase,

0,(o) = phase G12(co).

and the rnth order scatterfunction, fo(oJ) =f(w) + g(w),

S,(w) = Gz(~)/G,lb). and

Observe that asymptotically (as the spectral density

estimates tend to the true values) o,(w) approaches
If, (0) Ilfo(o)9 Brnapproaches phase[c,(o)1 + The functions u,, 0,,, and S, will now be hard to
phase [f-,,,(w)], and S,(w) approachesJ,[rk(u)]*
interpret. However, it is obvious that o,(w) < I,
[C,(O)/~. Unfortunately, it is not possible to esti-
M # 0 at most frequencies, and in practice this could
mate the phasesof f,(w) and C,(w) separately since
be used as an indication of the presenceof more than
the unscatteredprocessis not observed. The following
one signal.
examples will demonstrate the usefulness of these
functions.
Example 3
Example 1 If the scattering function is simply assumedto be
The first example worth considering is that of a real and symmetric about zero, then the situation is
single signal arriving from direction & observed only slightly modified. In this case the coefficients
without scattering. Then C, are real and C,(w) = C_,(w), and the functions
urn(o) and 0,(w) will be unchanged. The presence
f(w, 4) =f(w) 6(4 - #Jo),
of scattering will be demonstrated by the change in
and S,(o). If the function has a smooth broad peak at
zero, it will usually be found that IC,(w)( < 1 for
W(w, $) = S(lcI), m f 0, and consequently S,(o) wpill be small for
where 6 is the delta function. This gives m f 0. In general, small observed values of S,(w)
will imply a high degree of scattering.
fo(oJ) =f(w),

fm(o) =f(w) I exp{-im+] S(4 - &)dJI

=f(w) exp{-im&], CONCLUSIONS

and This paper developed first a physical interpretation

of the spectral representation of stationary process
C,(w) = 1. to provide a context in which earliet- results of Aki
184 Henstridge

could be analyzed. It was then possible to obtain an REFERENCES

alternative method of estimating wavenumber func- Aki, K.. 1957, Space and time spectra of \tationarq sto-
chastic waves, with special reference to microtremours:
tions and a test for the validity of the methods. Bull. Earthquake Res. Inst., v. 35, p. 415-457.
The second part extended the methods to obtain 1964, k note on the use of micrbsei\ms in determin-
ing the shallow structures of the earth’s crust: Geophysics,
directional information from the data. This was done
v. 30, p. 665-666.
by deriving explicitly defined functions which are Brillinger, D. R., 1975, time series: Data analysis and
often readily interpreted. This is in contrastto almost theory: New York, Holt Rinehart and Winston, Inc.
Koopmans, L. H., 1974, The spectral analysis of time
all other methods of array processing which require series: New York, Academic Press.
nonlinear maximization of functions and which do Toksiiz, M. N., 1964, Microseisms and an attempted apoli-
not identify correlations between waves from differ- cation to exploration: Geophysics, v. 29, p.‘154-‘17 ’ 7.
Y aglom, A. N., 1961, Second order homogeneous random
ent directions. The appropriate context for these fields: J. Neyman, editor, Proc. 4th Berkeley Sympos.
methods appearsto be in obtaining a quick analysis of Probability and Statistics, Berkeley, University of Cali-
fornia Press, p, 593-622.
data, and as a starting point for more refined methods.