A theoretical study of the use of arrays for the P(k,, k,, f) has been estimated from both long
analysis of seismic noise fields has been completed. and shortperiod noise recorded by the Large
The frequencywavenumber power spectral den Aperture Seismic Array in Montana. At fre
sity P(k,, k,, f) is defined and techniques for esti quencies higher than 0.3 Hz, a compressional
mating it are given. The estimates require that bodywave component which correlates with
the auto and crosspower spectral densities be atmospheric disturbances over distant oceans has
estimated for all elements in the array. Subject to been detected. In the frequency range of 0.2 and
certain asymptotic properties of these auto and 0.3 Hz both body waves and higher mode Ray
crosspower spectral density estimates, expressions leigh waves are observed. At frequencies below
for both the mean and variance of the estimates 0.15 Hz the organized vertical component of
of P(k,, ky,f) have been obtained. It has been de microseisms consists primarily of fundamental
monstrated that if P(k,,k,,f) is estimated by the mode Rayleigh waves. Appreciable amounts of
Frequency Domain Beamforming Method, then fundamental mode Love wave energy may also be
the estimate has the same stability as the esti present on horizontal instruments at these low
mates of auto and crosspower spectral density. frequencies.
i Manuscript received by the Editor June 6, 1968; revisions received October 12, 1968.
* Lincoln Laboratory, Massachusetts Institute of Technology, Lexington, Massachusetts.
** Department of Geology and Geophysics, Massachusetts Institute of Technology, Cambridge, Massachusetts.
21
22 Lacoss et al
sider P(k,, k,, f) to represent the noise power at putational aspects of estimates have been given
frequency f propagating as a plane wave with considerable attention. ‘4s a result, some com
phase velocity components ments and suggestions relating to this aspect of
estimates will be found in the remaining sections
I’, =  fiZ,/(k% + k;) of this paper.
It should be noted that both space and wave
and (3) number could be introduced as vector quantities
and vector notation used. Rather than do this
we have generally considered the twodimen
sional case and explicitly treated all coordinates.
The definition of P(k,, ky,f) involves a number
of assumptions concerning the time and space DEFINITION AND PROPERTIES OF ESTIMATORS
stationarity of the seismic noise field. It has been OF NOISE FIELD STRUCTURE
SEISMOMETERS
LOCATED AT (Xi, y,)
sl+=+ll

DELAY T*
BANDPASS
+ FILTER
CENTERED AT f,
DELAY r3
DELAY r4
s4
P (v,,vy I fo)
xi vx+ YiVy
r, =
(v; + $1
The weights can be chosen quite arbitrarily to P,,,,(f). Note that 8 is real since Pmn(f) = P:,(f)
achieve velocity sensitivity. For example, op where * denotes complex conjugate. Equation (7)
timum weights yielding maximum likelihood esti is the basis for the frequencydomainbeamforming
mates of the noise (Capon, Greenfield, and Kolker method (FDBM) Afor the study of seismic noise.
1967) might be used, or all the IV, may be Suppose that Pmn(f) is some reasonable esti
made equal, resulting in simple delay and sum mate of P,,,%(f) with the property that pmn(f) =
processing. The beam shown is bandpass filtered, P,*,,,(j). Then a reasonable estimateL@(vZ, zb,f), of
squared, and averaged to obtain an estimate of 6(v,, z+,,f) can be defined by using Pm&/) in place
the power at frequency10 which is propagating at of Pmn(f) in equation (7). Now assume that
velocity (tk, ~1~).The final frequency analysis step vZ,z’,, k,, k, andJ are related according to equation
could also be done using the indirect method of (3). We can then define
Blackman and Tukey (1958), using a smoothed
periodogram as described by Jones (1965), or
some other related method, such as the direct
Using equation (7) then gives
segment method suggested by Capon, Greenfield,
and Kolker (1967). In any case, we shall refer to
the above as a timedomain beamforming method
for the analysis of noise.
It is possible to study a noise field using a .exp I _2a[k,(rc,,  x,)
method which has the same physical and intui
+ &(Y?Z  Y?n)l). (9)
tive appeal of timedomain beamforming but
which avoids the formation and spectral analysis It is clear that, rather than doing timedomain
of a large number of beams. Such a method, re beamforming and spectral analysis of each beam,
ferred to as a frequencydomain beamforming one can estimate the P,,(f) and then study the
method (FDBM), is developed below. velocity structure of the noise using equation (9).
Let (X(5:, pi, r) be the autocorrelation function This is the essenceof the FDBM.
Of b(r:, i’y, t) and let Z&(T) be the crosscorrelation It is possible to generalize equation (9) by re
function between seismometers m and 1~. It placing the product W, W, with a single function
follows iroin equations (4) and (5) that of two indices W,,. Thus
Let cP(o=, ~5,j) and Pmn(f) be the onedimensional In order to be sure that this &k,, k,,j) is real, we
Fourier transforms of @(v~,v,,~) and Rmn(r), re assume that W,, = W:,. This general Pfunction
spectively. Thus U’(rr, v,, I) is the power spectral defined by equation (10) has lost any simple
density of h(v,, v,, t) and Pmn(f) is the crosspower relationship to timedomain beamforming. It
spectral density between seismometers m and n. nevertheless can constitute a good estimate of
Taking the onedimensional Fourier transform of P(k,, k,, f).
equation (6) then gives In the preceding discussion of the FDBM the
weights W, were all real. Thus W,,= W,W,
satisfied the requirement that W,,= Wz,. How
ever, if the IV, are allowed to be complex, and
there are situations where this is desirable, then
W,,,,= W,,,Wz should be used in equation (10)
in order to satisfy the requirement that P be real.
In subsequent discussions, FDBM will refer to
Thus the power spectral density of b(v,, v,, t) can this slightly generalized situation where the W,
be obtained as a complex weighted sum of the might be complex. The quality and properties of
24 Lacoss et al
In the case of ordinary onedimensional Fourier rently under study. In our work up to the present
transforms, it is well known that the transform of time we have used the FDBM with UT,,= 1 for all
a product of two functions is the convolution of N. In that case, the wavenumber window function
their transforms. The same property holds for is simply the square of the magnitude of the
higher dimensions.
_ Thus, as a result of equation straight sum beam pattern of the array. The win
(27), E P(k,, k,, j) is a convolution of the trans dow functions have been adjusted by choosing
forms (with respect to pz and py) of w(pz, py) and different sets of seismometers available in the
of V(pI, pu,f). The transform of w(pz,pJ is usually Large Aperture Seismic Array (LAS.\) in Mon
called the spatial window function and is given by tana. Figures 2 and 3 show two shortperiod
seismometer configurations and wavenumber win
W(kz,6,) dows which have been used.
ss
cc
The resolution of a wavenumber window func
=
4P=, PA tion, that is, the size of the region around k,=
*
k,=O over which W is large, is roughly determined
.esp [iWk,p, + kupu)]dpzdpu by the spatial extent of the array being used. A
general rule of thumb is that the size of that cen
= l5
K? 7?7.?L=l
W,, exp ( i2~[k,(x,  x,) tral lobe in cycles/km is the inverse of the array
diameter in kilometers. This gives an order of
+ k&n  rm>ll (28) magnitude for the array resolution for an\ reason
According to equation (24) the twodimensional able choice of W,,. The window functions shown
transform of V(pz, pu,/) is the onedimensional in Figures 2 and 3 give some indication of the
convolution of h(f) and P(k,, k,, j). Thus validity of the rule.
It is often possible to obtain a more complete
@(k,, k,, f) understanding of general expressions by the
sss
00 consideration of simple idealized special cases.
=
W&  kz,E,  k,)h(F,f) With this in mind the expected value of estimates
Do have been obtained for completely disorganized
and for completely organized noise. This is done
on the assumption that h, the frequency window
is the theoretical relationship between E j and P. function, is a good approximation to an impulse
This equation demonstrates that E P is related to function. That is, Pmn= Pm,.
P by a threedimensional window function which
0 SEISMOMETER CONFIGURATION I
is the product of the wavenumber window W and
l SEISMOMETER CONFIGURATION II
the frequency window h. .
If the frequency resolution of the estimates of .
. 0 .
bias in p%,,(f) can be neglected. .
0 0 0
Ideally, one would like W(k,, k,) to act as a two O . .
0
dimensional impulse function. If it did, then 0 0 O Oo
E ?(k,, k,,f) would equal P(k,, k,,f), on the as 0 00 O O
0 o. o” .
sumption that h(f) also acts as an impulse func 0 0
0 0 0
tion. This is not possible. In general, one must be 0 0
satisfied with a W(k,, k,) which is large near 0 0 xix
0 0
k,= k, =0 and small for all other k,, k, which
000
might be of interest. The choice of seismometer 0
locations and weights W,,,, to yield good window I;IG. 2. Two arm\s of shortperiod
functions is an important problem which is cur verticalSinstruments.
Seismic Noise Structure from Arrays 27
ifm = II
(30)
otherwise.
7
In this case, e
2 0
f
E&K,, k,,f) = Q(f) 5 IT’,,,,,iK? (31)
ll=l
(43)
and
. {E[L(f)~:“Cf)l Pmn(fK(f)]
.exp ( i2r[k,(x,  xm  x, + x,) P,,,P,*, = 2nPW (44)
+ k,(m  ym  yu+ rJ1). (36) Thus
+ Pm,P,**  Pm,P,*u)
Note that if IV,, = W,W$, it follows from equa
+ i Im ( PLPvn  PLP,, tions (48) and (31) that (E >)z/a2=M for this
spatially disorganized noise. If seismic noise is
 PIT1,,Pm+ PE.PnJ. (41) completely organized in the sense previously dis
Observe that cussed, then
= Q"
M
2
K”
2
,,r,,,=l
w,, times its complex conjugate. But reference to
equation (22) now yields the result
seems appropriate to include one very special where the seismometer numbering is indicated in
example to indicate just how much the stability of Figure I. The window functions of these two
some good estimators can vary from M. In par estimators are the same and are proportional to
ticular, we will show estimators with the same the Fourier transform of a set of impulse functions
windo\\ functions which have significantly differ located on the lag pattern of the regular array
ent stability. and having magnitudes given by the times the
Consider the two linear arrays shown in Figure lag is repeated. The stability for the FDBM
4. Only .r, and k, need be considered in this linear applied to the regular array is, of course, M.
case. The lag patterns show a point for every Direct calculation of the stability of the other
possible .r,.s, and the number of times that the estimator, assuming noise which is spatially dis
lag is repeated is also indicated. The nonredun organized, gives M/3.96. Thus the stability is
dant configuration might be considered in practice worse by a factor of four. One can interpret the
m order to reduce the number of instruments loss of stability as due to the loss of redundancy in
required to complete a noise survey. Sow suppose the lag pattern. Of course, the stability can be in
that the FDBM with IV,,= 1 is used with the creased by using more time to estimate the P,,,,(f).
regular linear array and the general estimator is It can be shown for a broad class of estimates of
used with the nonredundant array with the threedimensional frequencywavenumber
spectrum that the stability depends upon one
[i/l 6 3 1 1
product of two factors, one factor determined b\
the redundancy in time (temporal smoothing) and
one determined by the redundancy of the array
pattern in lag space. Thus, in general, a tradeoff
is possible between these factors. If one believes
1234567
. . . . . . .
4Ck
1 2 3 4 5 6 7 6 5 4 3 2 l+ TIMES REPEATED
. . . . . . . . . . . . .
e
LAG ZERO t
1 2 3 4
. . . .
++3c+/ 2c +
LAG ZERO t
that the noise is spatially stationary, then there AMPLITUDE “I GROUND MOTION
is no special requirement for spatial redundancy ‘8  AMPLITUDE
and the array designer would deploy his sensors
so as to obtain maximum coverage in lag space
with minimum redundancy. Stability would be
achieved by adequate temporal averaging. Xr
rays such as ell, tee, cross, or circular arrays have
all been discussed (Haubrich, 1967) in terms of E, = RESPONSE AT lcps
their lagspace coverage, which offers good uni E, = 0.019 VOLTS/mg
formity and low redundancy. Arrays which are OR E, = 0.003 VOLTS/mp/sec
efficient in this sense usually have very poor beam
patterns when used as antennas with simple de 24 I I I \I
0.5 1 2 5 10
lay and sum processing. However, with the gen
FREQUENCY kpd
eralization represented by equation (lo), quite
satisfactory windows can be obtained. Analogous FIG. 5,. Instrument characteristicsshortperiod.
techniques have been used in radio astronomy
(Wild, 1961) with circular arrays and in optics
the window function can be controlled onl)
(apodization).
by the placement of seismometers. In addition to
FREQUENCYWAVENUMBER STRUCTURE OF choosing locations to yield satisfactory windolz
MICROSEISMS AT LASA
functions, a number of noise samples have been
By use of the estimation methods defined and analyzed using several different window functions.
analyzed in the preceding section, a moderately This has been one method of determining which
complete study of the frequencywavenumber of the peaks of j(k,, IQ,,f) corresponds to trul!
structure of seismic noise fields in the vicinity of organized noise.
L.G.1 has been completed in the period range The stability factor r ior all results described
between 1 and 30 sec. The analysis was completed below was between 30 and 50. That is, between
using FDBM and the data from shortperiod ver 30 and 50 blocks of data were used to estimate
tical instruments and longperiod threecom power spectral matrices. The frequency resolution
ponent instruments in L.AS.4. Figure 5 sho\vs the
response functions of the instruments.
iooo
It was found that gains of different instru
ments varied and these variations correlated well
nith variations in measured noise power. To cor
rect for this the FDBM was slightly modified by
F
0‘
.c
E 1
1
t
making z 100
5
TV, = 1 /‘[P,,,(f)]““. (57)
6
1
This has no first order effect upon the preceding
.t;
E
\
analysis but should remove some of the effects of i0
E
seismometer gain variations. As a practical mat 0
.L
ter, the results of several runs with W,= 1.0 were z
compared with those given with W, modified and E

no significant differences were noted. It is worth
noting that with W, given by equation (57) the
E 4.0
largest possible value of&k,, k,J) is unity, and it
2
i
is achieved if the noise is a single plane wave.
j(k,, k,, /) can be interpreted as the percent of 3 1
seismic noise at frequency f which would be
passed by a beam steered to the velocity given by
FREQUENCY ( Hz)
equation (3).
If the W, are constrained by equation (57), PIG. 51,. Instrument characteristicslongperiod.
32 Lacoss et al
N N 3
3
I I I I
3 2 4 6 6
PERIOD (SW)
FIG. 7. Estimated wavenumberstructure of December
2, 1965shortperiodnoisesampleat 0.2 Hz and 0.3 Hz. FIG. 8. Theoretical dispersioncurvesfor LAS.4
Seismic Noise Structure from Arrays 33
Hz. The locus of points corresponding to a hori Table 1. LASA crustal model
zontal phase velocity of 13.5 km/set has been
Layer
drawn on each of the plots. The presence of noise thickness Pvelocity Svelocity Density
propagating at about that phase velocity from 01(km/set) (3 (km/set) p (gm/cm3)
D (km)
two different directions is clearly indicated by the
0.5 2.0 1.15 2.1
contours of the plots. This highvelocity noise 2.5 3.7 2.10
corresponds to teleseismic Pwave energy. From 19.0 6.1 3.55 ::;t
15.0 7.0 3.90 3.0
the phase velocities and directions given in Figure
13.0 7.6 4.35 3.3
9 it is possible to determine the source areas by 30.0 4.50 3.45
projecting the rays. This has been done for the 20.0 KL 4.54 3.47
60.0 s:os 4.45 3.40
above case and the source regions shown in ‘Ml.0 8.10 4.55 3.49
Figure 10. Both areas are in the oceans, and a 50.0 8.15 4.60 3.50
check of surface weather maps showed atmo 50.0 8.20 4.65 3.55
m 8.30 4.70 3.60
spheric low pressure systems located at these posi
tions. This indicates, that at least for this case, P
wave noise was generated by the coupling of
energy from the atmosphere to the ocean and corresponding to phase velocity of 8 km/set has
then to the crust. been drawn to indicate the clear velocity separa
Noise structure plots similar to those on Figures tion of the data.
7 and 9 have been obtained for several different In general, more low velocity noise seems to
days and for each tenth of a Hz up to 1 Hz. Figure come from the east than from the west. Kate in
11 shows one way of recording the occurrence of particular the large number of peaks recorded at
noise peaks corresponding to propagation from low velocities for 0.2, 0.3, and 0.4 Hz. Figure 12
the two eastern quadrants. The magnitudes of the shows a similar plot obtained for the western two
wavenumbers for any peaks corresponding to quadrants by using the same noise samples. Again
propagation from the east were computed and the separation into two velocity ranges is clear.
plotted versus frequency on the figure. A line Also, the plot indicates fewer occurrences of low
160
75 ~
D5
60
.
., 
45
307
..___
I5
FIG. 10. Projection of Pwave raypaths for December 2, 1965 shortperiod highvelocity noise.
velocity noise from the west than from the east. The directional preference of the low velocity
Xlthough the figure does not show it, it was also noise at relatively high frequencies is probably
true that peaks for noise from the east contained due to the crustal structure rather than to the
more power than those from the west. source of the noise. The crustal structure to the
b
IO .
0 0 0
ot<.JQ
I I 004 0.06
WAVE
I
006
NUMBER
0 IIO
(cycles
I
012
I,,“,)
I
014 .
0 002 004 006
WAVE NUMBER
008 010 012
(cycles /km)
I
014
I
OS
_
FIG. 11. Cumulative frequencywavenumber plot FIG. 12. Cumulative frequencywavenumber plot
of microseismic power arriving from the east. of microseismic power arriving from the nest.
Seismic Noise Structure from Arrays 35
.
. .
. .
l .
LPZ SEISMOMETERS
a
24 NOV 66
NOISE SAMPLE
kp 200 km ~4
frequency (Hz)
FIG. 15. One longperiod vertical array used for
E‘IG. 1.3. Typical longperiodpower spectraldensities. noisestudies.
36 Lacoss et al
wavenumber analysis of horizontal components Blackman,R. B., and Tukey, J. W., 1958,The measure
ment of power spectra from the pomt of view of
has demonstrated that appreciable amounts of communicationsengineering:Bell System Technical
fundamental Love wave energy are also present Journal, v. 38.
at these low frequencies. These results constitute Burg, J. P., 1964, Threedimensionalfiltering with an
array of seismometers:Geophysics,v. 29, p. 693713.
very comprehensive data about the mode struc Capon, J., Greenfield,R. J., and Kolker, R. J., 1967,
ture and composition of the microseisms. They Multidimensional maximumlikelihoodnrocessineof
a Large Aperture SeismicArray: Proc.‘I.E.E.E, ;I
cover a very broad frequency range from 0.04 Hz 55. D. 192211.
to 0.6 Hz. Capon, J., Greenfield, R. J.,. and Lacoss,R. T., 1967,
In this paper we did not discuss the mechanism Longperiodsignalprocessingresultsfor Large Aper
ture Seismic Array: M. I. T., Lincoln Laboratory
of generation of microseisms. Our results indicate Technical Note 196750.
that the sources of body wave components of Cooley, J. W., and Tukey, J. W., 1965, An algorithm
microseismic noise originated at sea and are asso for the machinecalculationof complex1:ourierseries:
Math. of Comp., v. 19, p. 297301.
ciated with storms. We could find no evidence of Douze, E. J., 1967, Shortperiod seismic noise: Bull.
prominent noise peaks associated with tectonic S. S. A., v. 57, p. 5581.
Haubrich, R. A., 1966, Semiannual technical summary
regions. Thus the contribution of creep and other Study of earth notse on land and sea bottom: Uni
tectonic activities to the microseismic field on a versity of California, Institute of Geophysics and
continuous basis seems to be small, if any. The Planetary Physics, La Jolla, February 11.
 1967, Array design, semiannual technical sum
presence of Love wave energy in the longperiod marystudy of earth noise on land and sea bottom:
microseisms cannot be explained in terms of University of California, Institute of Geophysics and
conventional hypotheses based on pressure Planetary Physics, La Jolla, July 1.
Jones, R. H., 1965, A reappraisal of the periodogram
fluctuations at ocean bottom. If we accept the in spectral analysis: Technometrics, v. 7, p. 531542.
oceanic origin, the Love waves must be produced Pakiser, L. C.., and Steinhart, J. S., 1964, Explosion
seismology m the Western Hemisphere: Research in
by shear forces at the ocean bottom or by Ray
Geophysics, v. 2, Cambridge, M. I. T. Press, p. 123
leigh to Love wave conversion due to strong crus _.
16<
tal irregularities along the propagation path. Toksoz, M. N., and Lacoss, R. T., 1968, Microseism:
Mode structure and sources: Science, v. 159, p. 872
One point about the above results should be 873.
clarified. They apply to LASA and probably to Vinnik, L. P., 1965, The structure of 46 second micro
the interior regions of the continents. At coastal seisms: Dokladi Academia Nauk. v. 162.~. 10411044.
1967, Structure of microseismsf LPapers pre
regions and near great lakes, fundamental mode sented at the Ninth Assembly of the European Seis
Rayleigh waves may be present at higher fre mological Commission, August 1966, H. Jensen, Ed.,
Copenhagen.
quencies, provided they can be generated nearby.
Wild, J. P., 1961, Circular aerial arrays for radio astron
omy: Proc. Roy. Sot. A (GB), v. 262, June, p. 84.
ACKNOWLEDGEMENTS