JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 102, NO.

B3, PAGES 5369-5384, MARCH 10, 1997

Empirical Green's function analysis: Taking the next

step
S. E Hough
U.S. GeologicalSurvey,Pasadena,California

Abstract. An extensionof the empirical Green'sfunction (EGF) method is

presentedthat involvesdetermination of sourceparameters using standard EGF
aleconvolution,followed by inversion for a common attenuation parameter for a set
of colocatedevents. Recordingsof three or more colocatedeventscan thus be used
to constraina singlepath attenuation estimate. I apply this method to recordings
from the 1995-1996 Ridgecrest, California, earthquake sequence;I analyze four
clustersconsistingof 13 total eventswith magnitudesbetween2.6 and 4.9. I first
obtain corner frequencies,which are used to infer Brune stressdrop estimates. I
obtainstressdropvaluesof 0.3-53MPa (with all but onebetween0.3 and 11 MPa),
with no resolvedincreaseof stressdrop with moment. With the corner frequencies
constrained,the inferred attenuation parametersare very consistent;they imply
an averageshearwavequality factor of approximately20-25 for alluvial sediments
within the Indian Wells Valley. Althoughthe resultantspectralfitting (usingcorner
frequencyand •) is good,the residualsare consistentamongthe clustersanalyzed.
Their spectral shape is similar to the the theoretical one-dimensionalresponseof
a layeredlow-velocitystructurein the valley (an absolutesite responsecannotbe
determined by this method, becauseof an ambiguity between absolute response
and sourcespectralamplitudes). I showthat even this subtle site responsecan
significantlybias estimatesof corner frequencyand •, if it is ignored in an inversion
for only source and path effects. The multiple-EGF method presented in this
paper is analogousto a joint inversionfor source,path, and site effects;the useof
colocatedsets of earthquakesappears to offer significantadvantagesin improving
resolutionof all three estimates,especiallyif data are from a singlesite or siteswith
similar site response.

1. Introduction Alternatively, several studies have presentedjoint si-

Independentresolutionof source,path, and site ef- multaneous inversionsfor source,path, and site terms
fects from waveform recordingsof small-to-moderate [Boatwrightand Fletcher, 1991;Humphreyand Ander-
eventshas proven to be an elusivegoal. While several son,1995].Thisapproachis attractivebecausedata are
early studies relied on subjective determination of cor- not limited to earthquakesfor which a suitable EGF
ner frequencies[e.gBakun et al., 1980;Archuletaet al., recordingis available.
1982], many of thesestudieswereshownto be flawed In spite of advancesin methodology,fundamental
becauseof an incompleteaccountingof the attenuation; questionsremain unanswered,including the scalingre-
in particular,near-surface
attenuation[Anderson,1986; lationship betweenstressdrop and seismicmoment. Be-
Houghet al., 1989]. Recentstudieshaveachievedmore causethis issue is critical to various aspectsof earth-
reliable independent resolutionof sourceand attenua- quake source mechanicsand seismichazard, it is im-
tion characteristicsby usingan empiricalGreen'sfunc- portant to extract as much information as possiblefrom
tion (EGF) method, in which attenuation effectsare small-to-moderate earthquake recordings.
deconvolvedfrom an earthquakerecordingby using a The scaling of stressdrop with seismicmoment has
been debated at some length in previousstudies. Be-
nearbysmallerevent[e.g.Mueller,1985;Frankelet al.,
causea wide range of moments is required to deter-
1986];however,resolutionof cornerfrequencycan still
mine the scaling,resolutionof this issuenecessarilyin-
be limited, especiallyfor smallerevents[Hough,1996].
volvesthe determination of stressdrop for small events.
The method also requiresa specializeddata set: nearly
colocated events with similar waveforms. Most studies have focusedon Brune stressdrop rrbde-
termined from the inferred corner frequency or pulse
width, assumingit to be representativeof static stress
This paper is not subject to U.S. copyright. Published in
1997 by the American GeophysicalUnion.
drop. As discussed by Boatwright[1984],absolutestatic
stressdrop rrs will depend on both the initial and final
Paper number 96JB03488. stressstate of the fault, which may be impossibleto
5369
5370 HOUGH: EMPIRICAL
determine. The Brune stress drop, which is fundamen-
tally an estimate of dynamic stressdrop for a given rup-
ture model, may not equal the static stressdrop if the
earthquake rupture deviates from the assumedmodel.
However,if earthquakerupture processesdo not depend
systematicallyon sourcemagnitude, the use of ab to in-
vestigate static stressdrop scaling would be justified.
That is, departures from an rupture model would con-
tribute uncertainty to the static stressdrop estimates
but no systematicbias. If, on the other hand, an analy-
sis of ab revealssomethingother than constantscaling,
this wouldsuggesta dependenceof rupture processes on
magnitude, which, in turn, would require a reexamina-
tion of the assumption that ab can be used to estimate
While some empirical Green's function studies have
suggestedan increaseof stressdrop with moment in
isolatedtectonicregimes[e.g. Guo et al., 1997],a com-
pilation of resultsfrom different studiessuggestsno cor-
relation[Abercrombie andLeafy,1993].Althoughthere
is compelling evidencethat no nonconstantscaling of
stress drop can be extrapolated to events with mag-
nitude greaterthan 6 [Abercrombie and Leafy, 1993;
Houghand Dreger,1995;Hough,1996]arguesthat res-
olution of sourcescalingfor M ( 4 eventsis limited
even with empirical Green's function analysis. The per-
vasivelack of high signal-to-noiseratio at frequencies
above 60 Hz precludesresolutionof high stressdrop
(greaterthan 10 MPa) smallevents;bandwidthlimita-
tions of recordingsystemsprovide a further limitation.
Although high signal-to-noiseratios can be recordedat
higher frequenciesfrom borehole seismometers,many
suchinstrumentshave a limited low-frequencyresponse.
Boreholerecordingsare also considerablylessabundant
than surfacerecordings.
In light of the limitations discussedabove, it is im-
portant to extract as much information as possiblefrom
the available data and to identify the resolution limits
of the results. In this paper, I present an extensionof
the empirical Green's function analysisin which source
parametersare first obtained using standard EGF de-
convolutionon sets of nearly colocatedevents. Once
the corner frequenciesare determined, sets of three or
four spectra from colocatedevents are inverted jointly
to determine attenuation. Residual spectra are then
determined.
The method shares the same fundamental limita-
tion of the standard EGF approachrequiring a spe-
cialized data set; however, within dense aftershock se-
quences,clustersof multiple colocated(or nearly colo-
cared) eventsare not uncommon.In this study,I use
broadband recordingsof M2.5-4.9 events within the
1995 Ridgecrest, California, sequencefor illustration,
as well as one cluster from the 1992 JoshuaTree, Cali-
fornia, aftershocksequencefor comparison.
GREEN'S FUNCTION ANALYSIS
2. Method
I first assumethat the far-field seismicdisplacement
sourcespectrumfl(f) can be modeledby a one-corner
model:
_
l +(f/fc)• (1)
where ao is the low-frequencyspectral level, fc 'mthe
cornerfrequency[e.g.Brune,1970],and c is the asymp-
totic high-frequencyspectral decay. c is bounded be-
tween 1.5 (requiredfor conservationof energy) and 3;
because
the seismic
momentscalesas 1If 3, the omega-
cube model predictsthat abovesomefrequency,large
events will radiate the same high-frequencyenergy as
small events,a prediction that is generally contradicted
by observation [Walter et al., 1988]. The spectralfalloff
has been the subject of considerabledebate, with dif-
ferent theoreticalmodelspredicting a range of values
between2 and 3. Haskell [1966]proposeda model
which yields c- 3, although this model was shownby
Madariaga[1978]to produceinfinite stressdrop, while
a circular crack model with a finite rupture velocity
yieldsc = 2 [e.g.Madariaga,1976]. Frankel[1991a]
showedthat an omega-squarespectrum will result from
a self-similar distribution of subevents and a constant
subeventstressdrop. Most observationalevidenceap-
pearsto favor the w-square(c--2) model [e.g.Hanks,
1979;Andrews,1986;Houghand Seebet,1991]. In this
study, both the w-squareand w-cube modelsare tested.
The presenceof apparentattenuationand/or siteam-
plificationswill alter equation(1) to give an observed
spectrumA(f, r) that canbe generallywritten:
A(f,r) - Ao(f)S(f)G(r)e
-'• f •-•-•" (2)
whereS(f) is the frequency-dependent siteterm, G(r)
is geometricalspreading,Ao(f) is the sourceaccelera-
tionspectrum(f2•(f)), f](r) is the shearwavevelocity
along the ray path, and Q is a quality factor that may
includeboth intrinsicand scatteringlosses.In general,
Q is alsoexpectedto varyalongthe raypath [e.g.Ander-
son and Hough,1984]. As discussed by Andersonand
Hough[1984],if observedaccelerationspectraare char-
acterizedby exponentialdecay (i.e., a linear decayon
log-linear
axes),
thenthepathintegral
of(Qf](r))-x can
be equated with the observedspectral decayparameter
•, definedto be 1/•r timesthe slopeof the acceleration
spectrum above the corner frequency when plotted on
a log-linear scale. In this case, the spectral shape will
be insensitiveto any contribution to attenuation from
a componentof Q that is strictly proportional to fre-
quency,Q•,. That is, if 1/Qt = 1/Qi + l/Q/, whereQt
is total Q and Qi is the frequency-independent
compo-
nent,then Qy will remainunresolved.As discussedby
Frankel[1991b],there tendsto be strongtrade-offbe-
 HOUGH: EMPIRICAL GREEN'S FUNCTION ANALYSIS 5371

tweena frequency-dependent Q and geometricalspread- f• 1 0 ... 0

ing, and soneithertendsto be resolved.In this study,I ß

will henceforthassumea frequency-independent Q. Be-

f,• I 0 -.- 0
causeQ is then determinedonly from spectralshape,
f• 0 I ... 0
absoluteamplitudesare not considered.
Equation(2) assumes correctionfor an instrumentre- ß

sponseterm or that instrumentresponseis fiat within f,• 0 I ... 0

the frequencyband of interest. For a set of colocated ß

eventsand spectraat discretefrequencypoints,oneob-

tains f• 0 0 -.. 1
ß

- f,• 0 0 ... I

for the ith frequencyand the jth event;the distance,r, The inversionof equation(6) thus yieldsn sourceam-
site and path terms are constant. plitudes and one • value.
To determinecornerfrequency,spectrafrom pairs of As a final step, residuals between the observedraw
nearly colocatedeventsare deconvolved: spectra and the modelsare estimated. This procedure
resultsin one unavoidableambiguity: any commonsite
amplification(amongall events)cannotbe resolved.
(4) The method described above is henceforth referred
to as the multiple-empiricalGreen'sfunction(MEGF)
That is, the common path and site effectscancelto iso- method. As an alternative, I also consideran approach
late the relative sourceterms. If the corner frequencies that bypassesthe initial EGF deconvolutionand thus
are distinctfrom oneanother,estimatesof equation(4) avoids the numerical instabilities associated with de-
can be fit by a ratio of w-squareor w-cube sourcespec- convolution. In the absenceof significantsite effects,
tra to obtain best fitting cornerfrequencyestimatesfor the above procedure can be revised to consider a hy-
both events(or deconvolved in the time domainto ob- brid inverse(for brevity,henceforthreferredto as HI)
tain a relativesourcetime functionfor the largerevent). method in which a suite of spectra from a cluster are
Once a set of fc valuesis constrained,one can return fit jointly for corner frequenciesand •. This problem
to equation(3) and multiply the observedspectraby is nonlinearin fc/; howeverit is possibleto considera
the shape of the sourcespectrum to obtain a spectrum, HI/parameter searchapproach,in which the observed
Pj (f i) correctedfor sourcespectralshape: spectra are multiplied by a sourcespectral shapeswith
a rangeof cornerfrequencies(that is, a rangefor each
P1(fi) - Aj(fi,r)(l + (f /f•j)•) - Coj$(fi)e event) to obtain the suite of fcj valuesand single•
(5) value that is optimal. This involvesa proceduresimilar
to that outlined above, except that the observedspec-
wherefci isthecornerfrequency ofthejth event.C'olis tra are multiplied by a suite of sourcespectral shapes
now a constantamplitude term that will be affectedby rather than one that is previously constrained:
the sourceamplitude, geometricalspreading,and any
Q/that is proportionalto frequency.For the Ridgecrest P1(fi) - (1 + (fi/f•1,t)•)Aj(fi, r) (8)
data, it will be shown that spectra corrected in this
manner do not exhibit strong apparent site effects;I wherefcj,t isthe trial cornerfrequencyfor the jth event.
thereforeproceedinitially to neglectS(fi) and invert If weallowa trial of C cornerfrequencies for eachevent,
each clusterof eventsfor a common,unknown• value. then this inversionis performedC'ntimes. However,the
Takingthe logarithmof equation(5) for a suiteof events parametersearchcan be done first over a coarsegrid
leadsto a matrix equationthat can be invertedlinearly of cornerfrequenciesand then repeatedwith a more
for Ao• and •: narrow rangeof trial event-specific
cornerfrequencies.
- Ag The optimalset of cornerfrequencies (and•) is chosen
to be the onethat optimizesthe collectiveleastsquares
b is given by residual.

( og(Co) 3. Case Study: The Ridgecrest Sequence

where n denotes the number of spectra. If m is the The 1995 Ridgecrestsequencebegan with a M5.4
numberof frequencypoints, A is a matrix with (mn) eventonAugust17, 1995[Hauksson et al., 1995](Figure
rowsand (n+l) columns
ß la). This sequence,
whichalsoproduceda M5.8 event
5372 HOUGH: EMPIRICAL GREEN'S FUNCTION ANALYSIS

...............
I ß• ::::::::::::::::::::::::::::::::::::::::::::.:•:.
...............................
•'-'::•:!i
iii•::i'i•
ii•i•ii
•iii::
i•i•::
i::iiii
ilii::11
i?:i
i711
'!I'-.................. ::
..............
.._.....
.......:'.'
:•!!i!i!•!:.:i•
.:.::. ..:.::.:
.:......... ....................................
35' 40'N ............
.::...•:?:;:!•i•?:?:•;•!::•::::•ii•::!•i•!iiii!•!i•;ii•ii•;;•:•:!i•;•i!•;;•:•!ii•iiiii!•ii•i•iiiiiiiiiiiiii;!:•iii?;;:i•.•.•?
............................
...

35' 20'N

118' O0'W 117' 40'W 117' 20'W

Figure la. Settingof Ridgecrestearthquakesequence, includingregionalfaultsand topography

(greyscale).The September 20, 1995,M5.8 mainshock isindicatedby a largestar;smalldiamond
indicatesstationRC1. Open squareis centeredat samelocationas map shownin Figure lb but
not drawn to scale. The Indian Wells Valley is flanked by the southerntip of the Sierra Nevada
to the westand the ArgusRangeto the east. Inset at lowerleft indicatespositionof Ridgecrest
(star) within California.

on September20, 1995, and a M5.2 event on January cludinga Mark-ProductsL-22 2-Hz weak-motionsensor
7, 1996, representsthe largest seismiceventsto be ob- and a Kinemetricsforce-balance accelerometer (FBA).
servedin the immediate region sincethe inceptionof The samplerate was 100 samplesper secondon all six
the SouthernCalifornia SeismicNetwork (SCSN) cat- channels. One of these sites was still in place at the
alog. The sequenceoccurred within the Indian Wells time of the M5.8 September20, 1995, event, although
Valley, south of the Owens Valley and flanked by the that event was not recordeddue to an unrelated previ-
Sierrarangeto the westand the Argusrangeto the east ous power failure at the site. In this study, I analyze
(Figure1). Boreholelogsreveala depth-to-basement of data from the one longer-term site, RC1, becauseof
approximately1.6 km in the epicentralregion,with the the significantlyhigher number of recordingsavailable
upper layerscomposedprimarily of alluvium with some from this station. Because both the M5.4 and the later
interbedded basalt. Events in the 1995-1996 sequence M5.8 events had extensiveaftershocksequences,RC1
arealignedwith a broadN-NW trendingsystemof small recordedan abundanceof eventsof M3 and greater;the
faults. Focal mechanismsreveal a mixture of mostly largestrecordedevent was a September24, 1995, M4.9
right-lateral/normal/oblique eventsconsistent with the aftershockof the September20, 1995 mainshock. The
expected regional right-lateral shear and extensionof recorded events are associated with events in the SCSN
the Owens/Indian Wells Valley region. Following catalog;the network locationsand magnitudeswill be
the August 17, 1995, event, three portabledigital seis- used in this study.
mometerswere deployedin the epicentral region. The In this study, I use recordingsof events with high
siteswere instrumentedwith United States Geological signal-to-noiseratios on the FBA channelsto avoid
Survey(USGS)developed GEOS recorders [Borcherdt uncertainties associated with the L-22 instrument re-
et al., 1985]and a dual gain sensorconfiguration
in- sponse at frequenciesnear and just below its natural
 HOUGH: EMPIRICAL GREEN'S FUNCTION ANALYSIS 5373

35 50'N-

MAGNITUDES

0.0+

1.0+
•r RCl• ¸
2.0+

3.0+

4.0+

20 KM
I I I I , , I
35 40'N-
I ' ' I ' ' ' ' I

117 40'W 35'

A !

I I I I I I I I I I I I I

o•+•-

-10 i i i i i i i i i I i i i

0 km 10

Figurelb. Portable station,

RC1,thatprovided datausedin thisstudy(smallstar).Locations
ofeventsanalyzed areshown byothersymbols, withsizescaled to magnitude
asshown at right.
Forclarity,depthswithineachclusterarefixedto a representativeeventdepthforeachcluster;
actualvaluesarelistedin Table1. Clusters1, 2, 3, and4 arerepresented
by crosses,
triangles,
diamonds,
andsquares,
respectively.
Oneeventisusedforbothclusters
2 and4; it isshownhere
as a triangle.

frequency.Becauseof the shortepicentraldistance, largestof the recordeddata set (M4.9), no early after-
events with magnitudesas small as M • 2.4 can be shocksof this eventwererecordedbecause the trigger
used.
itselffilled the availablespaceon the GEOS data tape
Figure 2 showsone componentof motion for four (in general,the very earliestaftershocks often provide
eventclustersthat werejudgedto providesuitableevent the bestempiricalGreen'sfunctionevents).
suitesfor analysis.Parametersfor the eventsare listed The foursuitesspanmagnituderangesof 2.6-3.9,2.6-
in Table 1. To referto a givenevent,I will henceforth 3.8, 2.4-4.9, and 2.5-3.8 (a fifth possiblecluster was
useJulianday, hour, and minute(e.g., 2680427=day not used becauseof conspicuousnear-field phasesin
268,0427GMT). Figurelb shows eventlocations, with the waveforms).To calculatespectra,I use 4-s to 5-
depthsfixedto berepresentative
oftheaverage
foreach s windows(consistent
overa givencluster)bracketing
cluster;Table 1 lists actual depthsfor eachevent. In the direct S arrivalsusinga cosinetaper over 10% of
mostcases,eventswithin a clusterare separatedby less the serieslength at both ends. A standardfast Fourier
than a few kilometersin hypocentraldistance.One ex- transform(FFT) is usedto calculatespectrausinga
ceptionis cluster2, whichincludes oneevent(2680427) boxcar smoothingwindow of 0.3 Hz. I calculateroot-
that is 4-5 km shallower than the other two events in mean-square averagesof the two horizontalcomponents
the cluster. In spiteof this depthdifference, Figure for all analysisand focuson a frequencyrangeof 1-20
2 revealsa goodlevelof waveformsimilaritywith the Hz.
other two events,and so it is included. I note that al- The first clusterof eventsconsistsof four events,
thoughthe largesteventin cluster2 (2680447)is the three with magnitudesof 2.6, 2.7, and 3.8 that occurred
5374 HOUGH: EMPIRICAL GREEN'S FUNCTION ANALYSIS

Cluster 1 Cluster 2

lO
15X

15X

2.5 •
o

i i i i i I ' ' i , i i i
5 lO 5 10

time (s) time (s)

Cluster 3 Cluster 4
..... I ......... I'''
10

5
10X

iiiii iiiiiiiiiIiii
5 I0
..... I .........
5
I,,,
10

time(s) time(s)

Figure 2. Scaled,offsetnorth-southcomponentsof seismograms for earthquakesin eachcluster

usedin this study. Acceleration(FBA) channelsare integratedto obtain velocityand high-pass
filteredabove0.3 Hz. Time seriesare scaledby peakgroundmotionby factorsshown(or unscaled
if no factor is given) and offsetto illustratewaveformsimilarity.

within a 7-min span on Julian day 242 and another The other three clusters are analyzed similarly us-
magnitude2.6 eventthat occurredon day 243. Figure ing standardEGF analysis(Figure3); in eachcase,no
3 showsdeconvolutions for sourcespectrausingeachof compellingresults are found to support the •o-cubed
the three smaller events as the EGF; in each case, in- model. I thus assumethe •o-squaremodel for these
versionresultsusingboth •o-squareand •o-cubemodels clusters.One pair of events(2711136and 2720015,clus-
are shown. The least squaresmisfit is approximately ter 4) yieldeda deconvolved sourcespectrumestimate
20- 70% lower using the •o-cubemodel. Figure 3 re- that was observed to be noisy, with prominent high-
veals that this model is better able to fit the relatively frequencypeaks resulting from holes in the empirical
sharp inferred corner frequencyof the larger event. In Green's function spectrum. Based on a similarity in
one case, the •o-squaremodel is unable to match the low frequency(f < 10 Hz) spectralshapeof this ratio
shape of the deconvolvedsourcespectrum. As can be to the ratio between events 2750010 and 2720015, I as-
seenin Figure 4, this suite of eventsis consistentwith signto event 2711136the samefc value, 3.5, obtained
an •o-cubemodel, becausethe high-frequencylevelsare for event 2750010.
nearly indistinguishablefor all events. Once the corner frequenciesare constrained, the un-
 HOUGH: EMPIRICAL GREEN'S FUNCTION ANALYSIS 5375

Table 1. Event Clusters

Cluster Day Time, GMT Latitude Longitude Depth, km log(Mo) fcs as •

I 242 1554 35.7957 -117.6400 3.72 15.95 2.2 3.7 0.075

242 1551 35.7867 -117.6387 3.36 14.00 9.6 3.4
242 1558 35.7853 -117.6330 3.62 14.15 5.8 1.1
243 0255 35.7990 -117.6358 4.83 14.00 14.0 10.6
2 268 0421 35.8080 -117.6185 8.75 15.34 5.0 53.3 0.057
268 0427 35.8115 -117.6413 4.72 13.70 16.0 10.6
268 0447 35.8087 -117.6162 9.40 17.45 1.7 7.9
3 264 2348 35.7613 -117.6425 5.44 16.11 2.0 4.0 0.052
264 2353 35.7631 -117.6365 5.26 14.15 3.9 0.3
265 0006 35.7653 -117.6263 5.44 16.11 2.0 4.0
4 271 1136 35.8098 -117.6402 4.40 15.34 3.5 3.7 0.062
272 0015 35.8023 -117.6413 3.80 14.00 8.0 2.0
275 0010 35.8095 -117.6455 4.72 15.80 3.5 10.5
268 0427 35.8115 -117.6413 4.72 13.70 19.0 13.3

Assignedclusternumber;Julianday(1995),andhourandminuteof event. Moments aredetermined

by scaling
relationship
proposed by [HanksandBoore(1984)]
usingSCSNmagnitudes;
valuelistedis1og(Mo),
withMo in Newton-
meters.fcoand•roarecorner
frequencies
andstressdrops(inmegapascals)
determined fromS waves
(parentheses
indicate
assigned
value);• is inferredspectraldecayparameterfor eachcluster

corrected velocity spectra are used in an inversion to of the three eventsanalyzed correspondto events22 and
find amplitude terms Aoj for eachspectrum and a com- 80 of HD95: a M2.3 eventon Julian day 118 (1992) and
mon decay parameter, •. This procedure yields the a M5.0 event on day 139. A third event analyzed oc-
spectral fits shown in Figure 4, which are observedto curred approximately 47 secondsafter the M5.0 event;
be (subjectively)quite good in all cases. Optimal • it is not listed in the SCSN catalog and was not ana-
values of 0.075, 0.057, 0.052, and 0.062 are found for lyzed by HD95. Its inclusion was initially based on an
clusters1, 2, 3, and 4, respectively;the implications of observationfrom the Ridgecrestsequence,that tempo-
thesevaluesfor Q structure are discussedbelow. rally clustered events seem to correspondto spatially
As a final step, residualsare calculated between all clusteredevents(i.e., goodEGF events). Basedon a
spectra and models shown in Figure 4. This is done comparisonof long-perioddisplacementspectrallevels,
for each individual spectrum and the results are then I estimate a M3.6 for this event.
averagedwithin each cluster. Figure 5 showsthe final Figure 7 presentsone componentof the waveforms
inferred residuals,which are observedto be very con- for this cluster of events. Figures 8 and 9 show the
sistent among the clusters. deconvolved sourcespectra,velocityspectra(with best
I now apply the HI procedure described in the sec- fitting models),and residuals,analogousto Figures4,
tion above. Figure 6 showsspectra and model fits for 5, and 6.
cluster3, indicatingthe resultingcornerfrequencyes- As a further illustration, ! analyze this cluster using
timates from both methods. It is observedthat using the second, HI, method. Corner frequency estimates
the HI method, corner frequencyestimatesare gener- from the two methods are shown in Figure 9. In this
ally higherfor the large eventsand lowerfor the smaller case,the biasesassociatedwith the HI method are even
ones. This effect is seenconsistentlyamongthe other morepronounced:while the MEGf method resolvesdis-
clusters;these results are not shown. tinct corner frequenciesof 2 Hz, 6.3 Hz, and 11 Hz cor-
respondingto the magnitude 5.0, 3.6, and 2.3 events,
4. Results From the Joshua Tree respectively,the HI method yields corner frequencies
near 6.5 Hz for all three events. This bias results from
Sequence
apparent resonancepeaks in the residual spectrum near
For the purpose of comparison,I repeat the above 3.5 and 10 Hz; ! will discussthese results in the follow-
analysis on one cluster of aftershocks from the 1992 ing section.
M6.1 JoshuaTree, California, earthquake[Hauksson
et al., 1993]. The full set of portable GEOS aftershock 5. Interpretation: Source, Path, and
data for this event have been analyzed previouslyus- Site Effects
ing a standard eGf approachto obtain corner frequen-
cies[HoughandDreger,1995,hereinafterreferredto as 5.1. Stress Drop
HD95]. HereI userecordings fromstationFVS, on the To interpret the inferred corner frequencyvaluesfor
northern edge of the CoachellaValley to the south of sourceparameters,I usethe standardformulation[Keilis-
the JoshuaTree epicenter(seeHD95 for details). Two Borok,1959]
5376 HOUGH:EMPIRICALGREEN'SFUNCTIONANALYSIS

Cluster 1 Cluster 2

1 ooo

1 oooo

lOO

1ooo
lO

1 oo
1

lO
o.1

I , . , , . , ,I
0.01
1 lO lO

frequency(Hz) frequency(Hz)

Cluster 3 Cluster 4
1000 10000 ........ I

1, 1000

1 oo

100

lO

10

i i i i , i .I
lO 1 10

frequency
(Hz) frequency
(Hz)
Figure 3. Deconvolvedrelative source(rms-averaged)spectrafrom clusters1-4 usingsmaller
eventsas empirical Green'sfunctions. For cluster 1, the light solidline showsthe best fitting ratio
of ca-squaresourcemodels;dark solid line showsbest fitting ratio of ca-cubemodels. The latter
is observedto provide a significantly better fit to the sharp cornersevident in the deconvolved
sourcespectral ratio. For clusters 2-4, only the best fitting ca-squareratio is shown. No fit is
shown for one of the events in cluster 4 becausehigh-frequencynoise levels preclude a stable
regression;however, the spectral shape below 10 Hz is observedto be extremely similar to the
other two deconvolvedsource spectra in this cluster.

fc 3 J. Mori (personalcommunication,
1996) (Table2) and
a- Mo(0.49/•s
) (9) assumea Vv/V8ratio of 1.7, givinga ft8 of 2.8 km/s.
where/78 velocityat the source.The Stressdrop valuesfor the nine eventsare listed in Ta-
is the shear-wave
seismicmoment Mo, is estimated using the moment- ble 1 and shown in Figure 10; they are found to be
magnitude relationshipdevelopedby HanksandBoore consistent,generallyof the order of 1-10 MPa.
[1984],log(Mo)= 1.5M+ 16.1,whereI assume that The one stressdrop value appreciably over 10 MPa
the localnetworkmagnitudeprovidesan estimateof corresponds to the largest event in the data set, mag-
themoment-magnitude. Althoughuncertainty in abso- nitude 4.9. It is possiblethat the low-frequencyband-
lute stressdroplevelis introduced by severalfactors, width cutoff, caused mainly by low signal/noiseratios
including the choiceof/•8 in equation(9), in general, of the smaller events,resultsin a bias for this event.
it isexpectedthat relativestress
dropestimates willbe Although an increase of stressdrop with momentis
better resolved. I use a P wave velocity model from suggested, Figure 11 alsoindicatesthe limits in reso-
 HOUGH: EMPIRICAL GREEN'S FUNCTION ANALYSIS 5377

Cluster 1 Cluster 2
' ' ' ' ' ' ' ' I
lO lOO

lO

o.i

0.o•
o.1

o.o01
0.01

1 lO 1 lO

frequency(Hz) frequency(Hz)

Cluster 3 Cluster 4
i i i i i i i I I i • I i i i i i I
1 oo lOO

lO lO

o.1
o.1

O.Ol
0.0!
i i i i i i i iI
lO 1 lO

frequency
(Hz) frequency(Hz)

Figure 4, Estimated(uncorrected)
relativevelocityspectrafor all clustersare shown(rms
averageof two horizontalcomponents); solidline givesbestfitting modelwith cornerfrequency
constrainedfrom empirical Green's function analysisand common• value fit by simultaneous
inversionof all spectra in each cluster.

lution associatedwith a frequencybandwidth of 1-20 propriatefor 9 km depth), all valuesare reducedby a

Hz. Theselimits may be slightlygenerous in that is as- factor of 1.5. The M4.9 event remains as the highest
sumesresolutionof comerfrequencies all the way up to stressdropevent,but with a valueof 36 MPa, some-
the bandwidthbounds.Similarly,Figure 10 revealsan what more in line with other events(the value for the
inability to resolvestressdrop valuesappreciablyabove M3.5 eventin this clusteris reducedto 7.1 MPa).
10 MPa for the smallest events in the data set. The corner frequenciesfor the JoshuaTree eventsim-
i alsonote that the clusterwith the highestoverall ply somewhat higher stress drop values in general: 2,
stressdropvalues(cluster2) isalsothe deepest.Assug- 30, and 124 MPa for the magnitude 5.0, 3.6, and 2.3
gestedin previousstudies[e.g.Jonesand Helmberger, events,respectively,using/•, of 2.8 km/s. It is likely
1996],this couldindicatea tendencyof stressdropto that this/•, value is inappropriately low for the Joshua
increasewith depth. However, it could also, at least Tree events;using the preferredvalue of 3.64 km/s
in part, reflectan inappropriatechoiceof/•8 in equa- from HD95 lowersthe stressdropestimatesto 0.9, 14,
tion (9). In general,a constantvalueusedfor a suiteof and 56 MPa, respectively.Althoughit is again possi-
events
mayresultina biasofstress
dropvalues
(towardsble that the low-frequencycutoff biasesthe corner fre-
highvalues)for deepevents.If stressdropestimatesfor quency/stressdrop estimatesfor the MS.0 event,these
cluster2 are recalculated
using/•8of 3.2 km/s (moreap- resultsare consistentwith somewhathigherstressdrop
5378 HOUGH: EMPIPdCAL GREEN'S FUNCTION ANALYSIS

Average to basement(A. Katzenstein,personalcommunication,

2
I [ I I III ! I I I I I I I I I I I •ll I I I' I I I I I I I I I I
1996). Assuminga standard,shallowbedrockQ value
for 1.6-5.0 km depth, the dominant contribution to
1.5
will comefrom the portion of the ray path through the
1
valley itself. Differentiatingbetweena basementQ,
0.5
5 10 15 20 and a Q within the 1.6-km-thick alluvial sediments,
this dominance can be illustrated with a consideration
Cluster 4 of reasonablevaluesfor Qb. For example, a Qb of 500
i i i i i i i i i i i i i i i i •i i i i i i i i i i i i i i iI i l yieldsa Qa of 18-20(for a • of 0.06). Becauseanyplau-
sibleQb value for the basementrock is muchhigherthan
these values,the Q• resultsare fairly insensitiveto the
choiceof Qb. For example, lowering Q• to 200 only
5 10 15 20 raises Q• values to 20-25. These results are consistent
with other field and in situ attenuation measurements
Cluster 3 of shallowalluvial sediments[e.g. Joynet et al., 1976].
A Q• value on the order of 20-25 is somewhatlow for
the entire 1.6-km-thick valley. However, it is likely that
the inferred • value representsa path averageof lower
.
Q• values in the very uppermost layers and somewhat
5 10 15 20 highervaluesat depth. Because1/Qa is the physically
meaningfulquantity, even relatively thin layersof very
Cluster 2 low Q material will provide dominant overall contribu-
1 I I I I I I I I I 'l III I III I I I I I I I I I I' I I I I I I I 1 tions.
The four • estimates are actually consistentwith the
small differencesin ray paths for each cluster causedby
an epicentral distance range of 4-8.5 km and a source
5 10 15 20
depth rangeof roughly4-9 km. As revealedin the inset
to Figure 11, • correlatesnot with the overall hypocen-
Cluster 1 tral distancebut with the length of the ray path through
the valley itself. If I considereach ray path and • value
1.5 separately,assuminga Qb of 500 and estimating the
1
ray path length through the valley, I obtain separate
0.5
Q• valuesthat vary between 18.5 and 20.3.
5 10 15 2O
5.3. Site Response
frequency(Hz) The spectral residualspresentedin Figure 5 are ex-
Figure 5a. Residualsbetween spectra and modelsfor pected to incorporate a number of unmodeledeffects,
each cluster are shown in the bottom four panels. For includingpossibledifferencesin sourcedepth and/or
eachcluster,residualsbetweenspectraand modelsare focal mechanismbetween an earthquake and its EGf
averagedover all events. The top panel showsfinal event. However, the character of the residualsshownin
residualaveragedover all clusters. Figure 5 are quite consistentwith the expectedshape
of one-dimensionalresonancemodes caused by near-
surface sedimentary layers. A P wave velocity model
resultsfor the JoshuaTree eventsthan for Ridgecrest for the Ridgecrestregion was developedindependently
events. (J. Mori, personalcommunication, 1996)for earthquake
location. As a starting model,I modify the depth of the
5.2. Attenuation
top layer to reflect the local depth to basementand use
The inferred • valuesfrom the Ridgecresteventscan a standard Poisson's ratio to obtain the S wave veloc-
be interpretedusingthe plane layer velocity model, as ity model shown in Table 2. This model incorporates
illustrated schematicallyin figure 11. Assumingcon- Q values that are deliberately chosento be too high,
stant Q and a representativehypocentraldistanceof 8 because the MEGf method resolves shallow attenuation
km, a • of 0.06 impliesa shearwaveQ valueof approx- as a path, rather than a site, effect.
imately 70. More plausibly,attenuation is expectedto I use the reflector matrix technique of Kennett and
with depth[e.g.HoughandAnder- Kerry[1979]to compute
increasesignificantly thetheoreticalone-dimensional
son,!988]. The earthquakes occurredwithinthe Indian responseof the velocity model to vertically incident S
Wells Valley, in a region where depth to basementis waves. The full extent of the Indian Wells Valley causes
estimatedat approximately1.6 km (a numberof bore- a fundamentalresonancepeak at a frequencylowerthan
holes, flanking the epicentral region, have penetrated that observed.To matchthe inferredresiduals,I add a
 HOUGH: EMPIRICAL GREEN'S FUNCTION ANALYSIS 5379

Average

• 2

I I I I I I I I I I I ! I I I i i I i
0 10 20

frequency(Hz)
Figure 5b. The solidline indicatesthe theoreticalone-dimensional
responseof the S-waveve-
locitymodelshownin Table2 (J. Mori, personalcommunication,
1996),with depth-to-basement
estimatedbasedon boreholeinformation(A. Katzenstein,personalcommunication,1996)and a
shallowlower-velocitylayer addedto match observedresidualshape.

shallowlayer(170m) with a lowervelocity(1200m/s). the four clusters. A further sourceof uncertainty(in

These values are poorly constrainedbecausethe reso- usingobservedresidualsto infer velocitystructure)re-
nance modes are controlled only by the ratio of velocity sults from the neglect of angle of incidencein the theo-
to layer thickness. In Figure 5b, I comparethe final retical responsecalculation. However, as illustrated by
model to a final spectral residual that is averagedover Figure 11, while there is a considerablerange in inci-

Cluster 3
10ooo ! i
' ' ' ' I

1ooo
1 I

lOO

lO

I lO
frequency(Hz)
Figure 6. Velocity spectra (rms averageof horizontals)from cluster3 are shownalong with
cornerfrequencyestimates(light tick marks) resultingfrom simultaneousinversionfor source
and attenuation terms; dark arrows indicate preferred corner frequencyestimatesfrom EGF
analysis.
5380 HOUGH: EMPIRICAL GREEN'S FUNCTION ANALYSIS

Joshua Tree with subtle site responsecausedby sedimentswithin the

..... I ......... I ......... I ....... CoachellaValley. Presumably,this sedimentaryvalley
will produce substantial site effects; however, station
FVS was located at the northern valley edge,presum-
ably on fairly thin sedimentcover.

lO 6. Discussion and Conclusions

I havepresentedan extensionof the empiricalGreen's

function method in which the constraint of corner fre-
quenciesis used to provide high-resolutionestimatesof
attenuation for a suite of eventseven for a single site.
This method can be applied to clustersof any size.
Applying the method to a set of M2.4-4.9 eventsfrom
the 1995-1996 Ridgecrest sequence,I find the results
to be stable and consistent. Corner frequency results
imply Brune stressdrop valuesalmost exclusivelyin the
range 1-10 MPa, with no resolvedscalingof stressdrop
with moment. The stress drop estimate of the M4.9
eventis high, 53 MPa, but may reflectborderlinesignal-
to-noiseat the longerperiods(in particular,signal-to-
noiselevelsof the Green'sfunctionevents).In general,
this method, as presentedhere, is best suited to analysis
..... I ......... I ......... I ....... of small-to-moderate events, perhaps M2-4.5, because,
5 7.5 10 at larger magnitudes, the point-source assumption is
no longer appropriate. However, the method could be
time (s)
extended to allow for finite-fault inversions,along the
Figure 7. N-S component
of velocityfor threeevents linesof previously
EGf analysis
ofsmalleventsfromthe
comprising
a clusterfromthe 1992JoshuaTreeafter-
shocksequence.
Eventsare scaledby factorshownand
offset for clarity. Joshua Tree
' ' ' ' ' ' ' I

denceanglewithin the data set, Figure 5a revealsthe

individual cluster residuals to be consistent. The choice
to averageall resultsstemsfrom this observedconsis- lOO

tency. The comparisonbetweenabsoluteamplification

levelsis reasonablygood (Figure5b); the low inferred
amplitude of the fundamentalmode may reflect rela-
tively low signal-to-noise
levelsof the empiricalGreen's
functioneventsanalyzedat the longerperiods.. Given
that • is determined before site responsein this inver-
lO
sionprocedure,it is alsopossiblethat a strongfunda-
mental mode could systematicallybias all • estimates
(towardshighervalues),thus effectivelymappingsite
responseinto the path term. However,plottingthe ac-
celerationspectraobtainedhereon log-linearaxes[e.g.
Andersonand Hough,1984]revealedthat the spectra
are well-characterizedby an exponential decay shape,
with no conspicuous peaksnear 1-3 Hz that wouldbias
• estimates.
1 lO
The averagefinal residualfor the JoshuaTree cluster
(Figure9) alsosuggests a siteresponse resonance,with frequency(Hz)
somewhatbroad harmonicpeaks near 1.5, 3.5, and 10
Figure 8. Deconvolved relativesourcespectra(rms-
Hz. The inferred site responsein this caseis lesswell averageof two horizontalcomponents) for JoshuaTree
resolvedthan that inferred for the Indian Wells Valley events;the two spectraresultfrom useof the M3.6 and
becauseit results from analysis of only a single clus- M2.3 eventsas empiricalGreen'sfunction. The solid
ter. However, the residual is qualitatively consistent line indicatesthe best fitting ratio of w-squaremodels.
 HOUGH: EMPIl•CAL GREEN'SFUNCTION ANALYSIS 5381

Joshua Tree
, , , , i ' ' I

0.1 i i i i i ]

' i i i i i i i

I ' I

1oo

1o

o.1

I
I .... i i a,i &i i i[ 1o

frequency(Hz)

Figure 9. (bottom)Joshua Treevelocityspectra'

(rmsaverageof horizontalcomponents) for the
threeevents;solidlinesindicatebestfitting modelsincorporating(preferred)cornerfrequencies
from figure8 and optimal • value. Smalldark arrowsindicatecornerfrequencyestimatesfrom
EGF analysis;light tick marksindicatecornerfrequencyestimatesfrom simultaneous inversion.
(top) The residualspectrum,averagedoverthe threeevents.

Table 2. Model SanJacinto,Californiaregion[Frankelet al., 1986]or

, i i ,

M5-6.7regionalevents[Mori andHartzell,1990;Dreger,
Z, km a, km/s •, km/s p, kg/ms Q 1994;Houghand Dreger,1995].
The attenuationand site responseresultsfrom this
0.17 3.00 1.2 1800 20 study yield a relativelylow Q• value, 20-25, for the
1.40 3.55 2.0 1800 50 mostly alluvial sediments within the Indian Wells Val-
1.40 3.76 2.17 2400 200 ley. It is likelythat this valuerepresents
an average
9.00 4.80 2.80 2800 500
overthe 1.6-kmextentof the valley,with lowerQ• over
10.00 5.52 3.20 2800 500
theupperfewhundredmetersandsomewhat higherQ•
at depth;sucha gradationcannotbe resolvedwith sur-
Thickness of layer(Z), P and$ wavevelocities(a and
fi), density(p), andqualityfactor(•). The• valueforthe facerecordings of earthquakesthat all occurbelowthe
first layeris determined fromobserved spectralshapes; Q bazement. TheinferredQ structurecorresponds to a t*
valuesare assumed for the deeperlayers. TheseQ values value of 0.05 (the minimumestimatedvalueof • from
will trade off to someextent, as discussedin the text. this study}.
5382 HOUGH: EMPIRICAL GREEN'S FUNCTION ANALYSIS

Ridgecrest/Joshua
Tree

lOO

I
lO & &

- I

0.1 I I
14 16 18

log moment(N-m)

Figure 10. PreferredBrune stressdrop resultsare shownfor Ridgecrestevents(triangles)

and JoshuaTree events(squares).The dashedlinesindicatesthe limits in stressdrop resolution
imposedby a minimum resolvedcornerfrequencyof ! Hz and a maximum of 20 Hz; any moment-
stressdrop pair outsidethese lines would not be resolvablewith the available data.

Q=20

=500

.062

.052
K• '
ø'ø81
o06 '

o.o
,,
Ra (km)
-lO I I
-lO -5 0

source-receiverdistance(km)

Figure 11. Cartoon illustrating raypaths for four analyzed event clustersand interpretation
of inferred• values(s) in termsof layeredQa structure.The dashedline indicatesthe depth
to basementof the Indian Wells Valley in the vicinity of the epicenter(in reality, the basement
is not flat, but • valuesare expectedto reflecta local depth to basement,as illustrated). The
basementQb value is assumed;the Qa value is inferred from the • results. A good correlation is
noted between• and the length of the ray path, Ra, within the Indian Wells Valley sediments;
increasedray path within the higher-Q basementdoes not appear to contribute significantly to
• estimates.Small inset shows• valuesversusestimatedraypath length throughthe valley (in
kilometer);solidline indicatesa Q• of 20. The line falls slightlybelowall four pointsbecauseit
neglectsthe small contribution to • from propagationthrough the basement;however,the slope
is well matched and the scatter is small.
 HOUGH: EMPIRICAL GREEN'S FUNCTION ANALYSIS 5383

Anderson[1986]showsthe bias in cornerfrequency tips; Art Frankel, Ned Field, Brian Cohee, an anonymousre-
that will result from neglectinga range of • values. His viewer, and the Associate Editor for their helpful reviews;
results show that, for •=0.05, corner frequency will be I thank Gary Glassmoyerand Chris Dietel for their pa-
tience and help in overcomingdata-playbackhurdles. Figure
underestimated by factors of about 8, 2, and 1.2 for la was generatedusingGMT software[Wesseland $raith,
momentsof 10•s, 10•5, and 10•7 Nrn, respectively
(i.e., 1991]. Finally, I thank William Hin•.e for unfailingprofes-
Mw valuesof 1.9, 3.9, and 5.9). Thus, neglectingthe sionalism,and Mary Sue Hough for making it a pleasureto
near-surface attenuation at site RC1 would bias stress travel to Ridgecrest.
drop values by a factor of nearly 6 for Mw4 events and
by a factor of 1.6 for M•6 earthquakes. This reveals References
significantpotential biasesin sourceparameter estima-
Abercrombie,R. and P. C. Leary, Sourceparametersof small
tion resulting from near-surfaceattenuation, even for earthquakes recorded at 2.5 km depth, Cajon Pass, Cali-
moderate magnitudesrecordedat short epicentral dis- fornia: Implications for earthquake scaling, Geophys.Res.
tance. Left., 20, 1511-1514, 1993.
It is instructive to considerpotential biasesrevealed Anderson, J. G, Implication of attenuation for studies of
earthquake source, in Earthquake SourceMechanics, Geo-
in this study, in particular, the effect of even subtle
phys. Monogr. Set., vol. 37, edited by S. Das and C. H.
site responseresonancepeaks on corner frequency esti- Schol•.,pp 311-318, AGU, Washington, D.C., 1986.
mates. Although standard EGf deconvolutionsshould Anderson, J. G. and S. E. Hough, A model for the shape
be successfulin removingany linear site response,stan- of the Fourier amplitude spectrum of accelerationat high
dard joint inversionsfor source, path, and site effects frequencies, Bull. Seism. Soc. Am. 74, 1969-1994. Bull.
may not be successfulin distinguishingcornerfrequen- $eisraol. $oc. Am., 7•, 1969-1994, 1984.
Andrews, D. J., Objective determination of sourceparam-
cies from resonance peaks. Applying the restricted eters and similarity of earthquakes of different size, in
source/pathinversionto the entireRidgecrestdata set, Earthquake Source Mechanics, Geophys. Monogr. Set.,
I observeda tendency for the small-event corner fre- vol. 37, edited by S. Das and C. H. Schol•.,pp 259-268,
quenciesto cluster near the 5-Hz resonancepeak (as AGU, Washington, D.C., 1986.
Archuleta, R. J., E. Cranswick, C. Mueller, and P. Spudich,
can be seenin the examplepresentedin Figure 7), even 1982. Source parameters of the 1980 Mammoth Lakes,
though the spectra are not characterizedby the domi- California, earthquake sequence, J. Geophys.Res., 87,
nant site responseeffectsseenin other regions.A simi- 4595-4608, 1982.
lar, perhapsstronger,bias is revealedin analysisof the Bakun, W. H., C. G. Bufe, and R. M. Stewart, Do P- and S-
Joshua Tree cluster. wave corner frequenciesmeasure different source param-
eters?, U.S. Geol. $urv. Open File Rep., 80-607, 1-10,
Previousjoint inversions[Boatwrightand Fletcher, 1980.
1991;Humphreyand Anderson,1995]haveincludedan Boatwright, J., Seismicestimates of stressrelease, J. Geo-
unknown site response term that is not incorporated phys. Res., 89, 6961-6968, 1984.
in the HI method. Although it would be possibleto in- Boatwright, J. and J. B. Fletcher, A generalinversion
clude an additional site responseterm in the HI method schemefor source,site, and propagationcharacteristics
as presented here, it is clear that if the data set in- using multiply recordedsets of moderate events, Bull.
$eisraol.$oc. Am., 81, 1754-1782,1991.
cludesrecordingsfrom a limited number of sites (or Borcherdt,R. G., J. B. Fletcher,E.G. Jensen,L. Maxwell,
siteswith similar site response),one wouldhave very J. R. Van Shaak,R. E. Warrick, E. Cranswick,J. S. John-
limited resolution between subtle resonancepeaks and ston, and R. McClearn, A generalearthquake-observation
an fm,x cornerfrequencyeffect[Hanks,1979].Ideally,a system,Bull. $eisraol.$oc. Am., 75,1783-1825, 1985.
joint inversioncould be made with the added constraint Brune, J. N., Tectonic stress and the seismic shear waves
that the residual spectrum, assumedto be site response, fromearthquakes, J. Geophys. Res.,75, 4997-5009,1970.
Dreger,D. S., EmpiricalGreen'sfunctionstudyof the Jan-
be modeled as a set of harmonic peaks; however, this uary 17, 1994 Northridge, California earthquake, Geo-
constraint is difficult to implement. In contrast, the phys.Res. Left., 21, 2633-2636, 1994.
stability of the • and site responseestimates obtained Frankel,A., Highfrequency spectralfalloffof earthquakes,
using the MEGf method suggeststhat modest site re- fractal dimensionof complexrupture,b-value,and the
sponsepeaksdo not significantlybias estimatesof path scalingof stresson faults, J. Geophys.Res., 96, 6291-
6302, 1991.
effectsif corner frequenciesare previously constrained. Frankel, A., Mechanisms of seismic attenuation in the crust
Conceptually, this result is not surprising;spectral de- - scattering and anelasticity in New York State, South
cay estimates will be constrainedby the shape of the Africa,and SouthernCalifornia,J. Geophys. Res., 96,
spectrumover the full bandwidth. The MEGF method 6269-6289, 1991.
thus appearsto offer significantadvantagesin the inde- Frankel,A., J. Fletcher,F. Vernon,L. Haar, J. Berger,T.
Hanks, and J. Brune, Rupture characteristicsand tomo-
pendentdeterminationof source,path, and site effects, graphicsourceimagingof M• • 3 earthquakes near An•.a,
even if data are available from only a single site. southernCalifornia,J. Geophys.Res., 91, 12633-12650,
1986.

Acknowledgments. The author is most thankful to Guo,H., A. Lerner-Lain,S. E. Hough,andW. Menke,Em-

Allan Kat•.enstein and Frank Monastero for assistance with piricalGreen'sfunctionstudy of LomaPrieta aftershocks
challengingfieldwork logistics;I thank Jim Mori for provid- U.S. Geol.$urv. Prof. Pap., in press,1997.
ing his velocity model and helpful discussions/information Hanks,T. C., fmax,Bull. $eisraol.$oc.Am., 72, 1867-1880,
throughthe courseof this investigation;Dave Wald for GMT 1979.
5384 HOUGH: EMPIRICAL GREEN'S FUNCTION ANALYSIS

Hanks,T. C. and D. Boore,Moment-magnitude relations of seismogramsfrom a downholearray in sedimentsnear

in theoryand practiceJ. Geophys.
Res.,89, 6229-6235, San FranciscoBay, Bull. Seismol.Soc. Am., 66, 937-958,
1984. 1976..

Haskell,N. A., Totalenergyandenergyspectraldensityof Keilis-Borok, V.I., 1959. On the estimation of the displace-
elasticwavepropagation frompropagating
faults.Part II, ment of an earthquake source and of source dimension,
Bull. Seismol. Soc. Am., 56, 125-140, 1966. Ann. Geofis., 12, 205-214, 1959.
Hauksson,E., L. M. Jones,K. Hutton, and D. Eberhart- Kennett, B. L. N., and N.J. Kerry, Seismicwavesin a strat-
Phillips, The 1992 Landers earthquake sequence- ified half space, Geophys.J. R. Astron. Soc., 57, 557-583,
Seismological
observations,
J. Geophys.
Res., 98, 19835- 1979.
19858, 1993. Madariaga, R., Dynamics of an expanding circular fault,
Hauksson,
E., K. Hutton, H. Kanamori,L. Jones,J. Mori, Bull. Seisrnol. Soc. Am., 66, 639-666, 1976.
S. Hough,and G. Roquemore, Preliminaryreporton the Madariaga, R., The dynamic field of Haskell'srectangular
1995Ridgecrestearthquake sequencein easternCalifor- dislocationfault model, Bull Seismol. Soc. Am., 68, 869-
nia, Seismol.Res. Left., 66, 54-60, 1995. 887, 1978.
Hough, S. E., Observational constraintson earthquake Mori, J. and S. Hartzell, Sourceinversionof the 1988 Upland
sourcescaling: understanding the limits in resolution, earthquake: Determination of a fault plane for a small
Tectonophysics,
261, 83-96, 1996. event, Bull. Seismol. Soc. Am., 80, 507-518, 1990.
Hough,S. E., and J. G. Anderson,High-frequency spectra Mueller, C.S., Source pulse enhancementby deconvolution
observedat Anza, California: Implicationsfor Q struc- of an empirical Green's hmction, Geophys.Res. Left., 12,
ture, Bull. Seisrnol.Soc.Am., 78, 692-707, 1988. 33-36, 1985.
of the Thatcher, W. and T. C. Hanks, Sourceparametersof south-
Hough,S. E., and D. S. Dreger,Sourceparameters
4/23/92 Joshua Tree,California,earthquake
andits af- ern California earthquakes, J. Geophys.Res., 78, 8547-
tershocks:Empirical Green'sfunctionanalysisof TER- 8574, 1973.
RAscopeand GEOS data, Bull. Seismol.Soc.Am., 85, Walter,W. R., J.N. Brune, and K. Priestley,Observationsof
1576-1590, 1995. high frequencyP-wave earthquakeand explosionspectra
Hough,S. E., and L. Seebet,Seismological
constraints
on comparedwith w-s, w-2, and Sharpesourcemodels,J.
sourcepropertiesof the rnb = 4.0 Ardsley,New York Geophys.Res., 93, 6318-6324, 1988.
earthquake:A characteristicrupture?J. Geophys.Res., Wessel,P., and W.H.F. Smith, Free softwarehelps map and
96, 18183-18195., 1991. displaydata, Eos •rans. AGU, 72, 441-446, 1991.
Hough,S. E., K. Jacob,and R. Busby,Groundmotions Wyss, M. and J. N. Brune, Regional variationsof source
from a M=3.5 earthquakenear Massena,New York: Ev- propertiesin Southern California estimatedfrom the ratio
of short- to long-period amplitudes, Bull. Seisrnol.Soc.
idencefor the poor resolutionof cornerfrequencyfrom
Am., 61, 1153-1167, 1971.
small events,Seismol.Res. Left., 60, 95-99, 1989.
Humphrey,J. R., andJ. G. Anderson,Seismic
sourceparam-
eters from the Guerrero subduction zone, Bull. Seismol. S. E. Hough, 525 South Wilson Avenue, Pasadena,CA
Soc. Am., 8•, 1754-1769, 1995.. 91106(e-mail:hough@gps.caltech.edu)
Jones,L. E. and D. V. Helmberger,Seismicityand stress
dropin the easternTransverse ranges,southernCalifor-
nia, Geophys.Res. Left., 23, 233-236, 1996. (Received
March 18, 1996; revisedNovember5, 1996;
Joynet,W. B., R. E. Warrick,andA. A. Oliver,Analysis acceptedNovember8, 1996.)