GEOPHYSICAL

RESEARCH
LETTERS,
VOL.19,NO.4,PAGES
333-336,
FEBRUARY
21,1992

TttETIME-DOMAIN
BEHAVIOR
OFPOWER-LAW
NOISES
DuncanCarrAgnew

Institute
ofGeophysics
andPlanetary
Physics,
University
ofCalifornia,
San
Diego

Abstract.The power spectraof many geophysical phenomena models andthespectral form(1) aremathematicalidealizations,
arewellapproximated by a power-law dependence onfrequency andin differentsituations
either
one might be thebetter
descrip-
or wavenumber.I derivea simpleexpression for the root-mean- tionof actual
data.I havechosen to use(1) asa modelbecause
squarevariability
of a processwithsucha spectrum overaninter- my mainpurpose is theinterpretation
of spectra,
for which(1)
valof timeor space.The resulting expressionyieldsthepower- caneasilybe generalized (Section
4). It is worthnotingthat
lawtimedependence characteristic
of fractalprocesses,butcanbe recentstatistical
studies[Mohr, 1981;Graf, !983] haveshown
generalized
to givethetemporal variability
formoregeneral spec- thatthebestmethod fordeterminingH istocompute thepower
tralbehaviors.The methodis appliedto spectraof crustalstrain
spectorre,
fit a functionof the form (1) to it, andthenfind H
(toshowwhatsizeof straineventscanbe detected
overperiods fromthev sodetermined; in thatviewv could beregarded asthe
of months
to seconds)
andof sealevel(to showthedifficultyof morefundamental parmeter.
extracting
long-termratesfrom shortrecords). However, it should
benotedthat(1) is notin generala com-
1. Introduction
pletespecification
of a process;the powerspectrum is onlya
summary of second
moments (variance
versusfrequency).Sto-
Manytypesof geophysical
datacomefrom processes
so com- chasticprocesseswithidentical powerspectra canhavevery
plexthattheiroutcome
is besttakento be random,
evenif the different
appearances
in thetimedomain[Press,1978]. I address
thispointmorefully below.
underlying
physicsis not; the mostefficientcharacterizations
of
suchdata are likely to be a statisticalmodel. The aim of this
2. Time-Domain Variation
paperis to developa usefulrelationfor a particular(butcommon)
classof suchmodels,and showseveralapplications
if it.
The particulartype of model consideredmight be calledthe Thequestion tobe addressedhereis howto gofromthepower
power-law
process.This is a one-dimensionalstochastic spectrum
process to (1) to thevariationof the process
overtime T; thatis,
the statistics of
whosebehaviorin the time domain(or spacedomainif appropri-
ate)we denoteby x (t); this time-domainbehavioris suchthat its yr(t) = x(t+T) - x(t) (2)
powerspecmanhasthe form
This quantity
wasintroduced
by Kolmogorov
in studiesof the
Pz(f) = PoOe/fo)v (1) theoryof turbulence,
andits secondmoment,underthe nameof
wheref is spatialor temporalfrequency,P o andfo are normal-
structure
function
of x, hasseenwideusein meteorology
and
izingconstants, andv is the spectralindex. The reasonfor adopt-
elsewhere[LindsayandChic,1976].Theattraction of lookingat
thevariation
of x overa fixedtimeT is thatyr(t) is oftensta-
ingthisformis that it is observedto be a goodfit to the spectra
tionary,
andthuseasilycharacterized
statistically,
evenwhenx(t)
of a wide variety of geophysicalphenomena, often over many is nonstationary
(asmustbe truefor v < -1). Butthe statistics of
decades of frequency.The indexv oftenfalls in the range-3 to Yr mayoftenbeasof muchinterest asthoseof x; in particxt!ar,if
-1, meaningthat the energyat low frequencies exceedsthat at we wantto decidewhether
somerecentfluctuation in x (t) is con-
highfrequencies (a "red" spectrum).Spectraof this formhave
beenfoundfor bathymetrywith v = -2.3 [Malinaverno,1989];
sistent
withitspastbehavior
it is tothedistribution
of Yr thatwe
faultandjoint geometry, with v = -2 [Poweret al., 1987];and
mustturn. Unfommately,
if we choose
to summarize
thispast
behavioras havinga spectrum
of the form (1), the fractional
crustaldeformation, with v = -2.7 [Wyattet al., 1988]. Brownianmotionsof Mandelbrotand Van Ness [1968] tam out
Despitethe ubiquityof stochastic processes with power-law
SlYecrm,theyhavereceivedrelativelylittle attentionin the statisti-
to beveryinconvenient,
sincetheexpression
forthepowerspec-
trum of suchprocesses
is extremelycomplicated
[Graf, 1983;
cal literature.The only exceptionhas been the casev =-2,
whichcorresponds to a randomwalk (Brownianmotion):thisis Geweke
andPorter-Hudake,
1983].Thereis thusnosimplerela-
tionshipbetweenthe spectrumof such processesand their
theintegral
of whitenoise(for whichv = 0), the spectral
index variation in the time domain.
beingshifted
to -2 by theoperation
of integration(andsquaring
to get power). Mandelbrot and Van Ness [1968] developed
Thereis howevera relatively
simplemethodwherebywe can
relatethespecUal
level(1) to thedistribution
of yr(t), provided
mathematical
formsfor processes
thathavepowerspectra
closeto thatwe onlyaim to findonly the secondmoment,or variance,of
(1),those
with-3 < v <-1 beingtermed"fractional
BrownJanYr (denoted
by <yr2>),
which,asnotedabove,
is thestructure
motions,"and those with-1 <v < 1, "fractionalGaussian
function.Thisrestrictionis of courseunavoidablegiventhatour
noises."
Mandelbrot[1983]provides a generaldiscussion
of basicdescription is the powerspecman.Specifyingonly the
the•,andFeder[1988]a readable
introduction. secondmoment is adequate to definethedistribution
of Yr if it is
Forclarity,
it shouldbe notedthatthemodels introduced
by Gaussian.This restriction
will not usuallybe seriouslyviolated
Mandelbrot
donothaveexact power-law spectra
[Graf1983],
and for real data,but shouldbe kept in mind beforeinferencesabout
indeedaremoreusually discussed
in termsof theirbehaviorthecomplete distribution
aremade fromthevalueof<yrZ>.
underchangesin timescale(assming themto representa time To determine <yr2> from thespecmun,weobserve thaty(t) is
series).
In thisview,theimportantpmameterbecomes a number derivedfromx (t) by convolution;
we canrewrite(2), in thenota-
H, called
byMandelbrot [1983]theHurstexponent;aswewill tion of Bracewell[1965], as
seebelow,H = -•/2(v + 1). In practice,both Mandelbrot's
y(t) = x(t), [•t-T) - •t)] (3)
Copyright
1992bytheAmerican
Geophysical
Union. Straightforward
application
of the resultsof Fouriertheorythen
showsthatthepowerspectrum
of Yr is givenby
Paper
number
91GL02832
0094-8534/92/9!GL.02832503.00 Py(f) ---IGr(f)l:Px(f)
333
 19448007, 1992, 4, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/91GL02832 by Inst of Microbial Technology, Wiley Online Library on [10/08/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
334 Agnew:The Time-Domain
Behaviorof Power-lawNoises

whereGr(f) is the frequencyresponseof the convolutionfilter, Px(f) to be described

by (1) forf > fb, andto be (say)constant
usedto producey (t) from x (t) namely at Po(ft,/f0)v forf < ft,, x (t) will be a stationary
process,the
[Gt(f)l2 = 4 sin2
a:fr operations leading from.(4)to (5) will bevalid,and,theresultin
(5) will be essentiallyunchanged aslongas T ncf/q. While
Finally, sincethe varianceof a randomprocessis equalto the introducing a cutofffrequency is in one sensearbitrary,
it must
integralof its powerspectrumwe find existfor anyactualprocess (see,for example, Keshner[1982]for
sucha modelfor 1If noise). Sincethe finite spanof our obser-
vationswill alwaysrenderus incapableof observing sucha
<YT•
--I Py
(f)df--'I4Px
(f)sin2
aftdf (4) cutoff,thereseemsto be no reasonto avoidintroducing it to
Other methodsthan (2) of fonning auxiliary series exist, and avoidthe difficulties
intowhicha too-strictadherenceto anideal
some have long been used in (for example)studiesof oscillator mathematicalmodelwouldotherwise leadus [Slepian,1978].
stability. Rutman[1978] discusses
manyof theseandshowshow An importantgeneralizationis to note that (4) appliesfora
the transfer-function
approach just discussedcanbe usedto derive generalspectralshapePx(f), providedthat at low frequendes
their behaviorfor differentspectra;while he discussesthe struc- Px(f) increases
lessrapidly
thanf-a, andat highfrequendes
ture function,it is not treatedvery fully becauseof its limited decreases
morerapidlythanf-•. Of course,
we thenwillnot
usefulness in oscillator studies. usually
be ableto finda closed-form
expression
for <yr2>,
but
If we now specializePx(f) to the form (1) and make the mustcalculateit numerically.This allowsus to proceedevenin
the face of the departuresfrom power-law (or fractal)behavior
changeof variableu = •fT, we find
noted,for example,
by Gilbert[1989]. A simplespectral
shape
whichfitsmanyspectra
quitewell is a piecewise
power-law
form;
<yT2>
= 4Pø
f•
T-(V+O
•v+• iuVsin2u
du=Cv
P
f• o (5)
T-(V+•)for i=! ....

This immediatelyimplies that the standarddeviationof Yr, Px(f) = Pi(f/f i)vt f i-1< f < f i (6)
or = (<YT2>)
v•,isproportional
to T-(v+ly2,
orTH in Mandelbrot'swith f0 beingsetto zero. The integral(4) thenbecomesthesum
notation;
forv =-2, wefindOTproportionalto T•, thefamiliar of integralsover eachfrequencyinterval. Becauseof the oscilla-
resultfor Brownianmotion. The definiteintegralin (5), and thus tory natureof the integrandandthe wide rangein frequency,
the
the coefficientCv, can be found in closedform: numerical integration must be done with some care. For
-1 •rfT n: 1, we may approximatesin•rfT by xfT; for largevalues
Cv --' of thisquantity
weusesin2•rfT
= 1/2(1- cos2•rfT),
which
leads
2v+l•vr(-v)cos(v•2)
to two integrals,one of which can be doneanalytically,theoth•
which
forv =-2 gives
Cv= 2•. Figure
1 shows
C•aoverthe of which, though still requiringnumericalintegration,becomes
range-3 < v < -1, and illustrateshow the expression(4) goesto small for •rfT large.
infinityat both limits of this rangebecauseof the divergence in
the integralin (5). Thesedivergences occurat oppositelimits of 3. Comparison
with SpectralCrossover
Approach
the integral;put crudely, as v approaches-3, the low-frequency
fluctuationsin x(t) become so large that y(T) becomesnonsta- Agnew[1987] described
anotherproblemrelatingto power-law
tionary,while as v approaches-1,the high-frequency fluctuations processes:
how to comparea recordwhose errorsare of thisfore
in ;•(t) approachan ultravioletcatastrophe,with infinitevariance with measurements with independent erroro madeat regular
at high frequencies.This latter divergencewill not be a concern intervalsA. Cllfis characterizesthe problemof comparing crustal
in practice,and couldbe eliminatedin the theoryby replacingthe deformationmeasurements made usingstrainmetersand tiltme•rs
•-functionsin (3) with finite-widthsamplingfunctions. with thosemadeby geodeticmethods.)For a spectralindexless
A majorassumption hasbeenpassedoverin usingequation(1) than-1, therewill be somefrequencyat which the fluctuations 'm
to go •fromequation(4) to equation(5); namely,thatwhile (4) the power-law errors equal the error gotten by averaging the
presupposes x(t) to possess a powerspectrum,any processwith independent measurements. At a higherfrequencythe power-law
an apparentspectralindexless than-1 mustbe nonstationary, so errorswill be smallerandusingrecordswith sucherrorswillgive
that its spectrumdoes not exist: a contradictionmore apparent a betterresult;at lowerfrequency the independent
errors,suitably
than real. It is true that (1), with v <-1, cannotdescribethe averaged,will be superior.
spectnanof a stationaryprocess. However, if we supposethe This problemis easily solvedif east in spectralform (Figu•
2). The spectrum of the independent-error
measurements mustbe
constant,
withlevelP,,,,from0 totheNyquist
frequency,(2A)
-t.
Since
theintegral
of thespectrum is thevariance,
P,,,= 2o2'•
This equalsthe power-lawspecmnn(1) at a crossoverfrequency

fc=
whichthussetstheboundary
[2o'2Af•l
between
Po vv
oneor another
process
hav-
inga lowerlevel.Thecrossover
frequency
canbeequally
easily
obtained
graphically
fora spectrum
ofmoregeneral
shape,
though
a closed-fomiexpression
becomescumbersome.To give a con-
creteexample,
we maycompare
thestrainspecmnn
shown
Figure
3 With
repeated
distance
measurements
withao of10
-7.
If these
weremade
weekly,
theequivalent
spectral
levelwould
-2.8 -2.4 -2.0 -• .6 -1.2
1.2x 10-8 e2/Hz,or -79 riB, givinga crossover
frequency
corresponding
to a period
of 300days;
if theyweremade
daily,
thisperiodbecomes200 days.
Fig. 1. The coefficientC v for the time-domainwander of a Thetheorydeveloped
in Section
2 givesanotherwayoflook-
purepower-lawnoise(equation5), as a functionof the spectral ingatthisproblem.
At a period
T• = f•4, thermsfluctuation
in
index v. thepower-lawprocess
will be (from(5) and(7))
 19448007, 1992, 4, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/91GL02832 by Inst of Microbial Technology, Wiley Online Library on [10/08/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Agnew:
TheTime-Domain
Behavior
ofPower-law
Noises 335

NW Laser Strain: Power Spectrum

, ,

-80 "'.

•N ,, Geodetic
Survey
Spectrum -100

nUnuous Deformation

! Itlilll! i r fillill ! i iiIllrl ! i Illill! I till.. ! ! ilir

! 0-• 10-s 10-• 10-2 1•o

Frequency(Hz)

I
fc fN = (2A)-1 LogFreq NW Laser Strain: Wander
10 3 ...........

Fig.2. Cartoonof the relationshipbetweenthe powerspec-

trumof a sampledmeasuremere with independent
errors,anda ,-, lO 2
continuousmeasurement subjectto low-frequencydivergence.
See Section 3 for details. '•e 1 01

•e 1 0o
= 2(v+iy(=v)o(]+vyv
A(l+vy(2v)
,..' 10-1

Theerrorin the independent-measurement

series,suitablyaver- "'"' i min I hr Idly
aged,
will be N-•o whereN is thenumber
of measurements; 10 -2
10o 102 10•' 10s
obviously
for regularsampling
N = T,:IA. Againusing(7), we
Time betweenmeasurements(sac)
oN-« 1
Fig. 3. The upperpanelshowsthe powerspeclrumof earth
<y•>• (2C•)• strain at Pi•on Flat Observatoryin southernCalifornia, as
measuredwith the NW-SE laser strainmeter there [Wyatt,
whichis alwaysless than one. This is as it shouldbe, sinceat
1988].Thespectrum
from3.3x10
-5 to 0.5 Hz is based
on 1-
thecrossoverfrequency
the independentmeasurements
shouldbe seconddata collectedbetweenday 31:17 and 33:14 of 1990;
capable
of resolving
thefluctuations
of thepower-law
series. this has beenprocessed with two differentsectionlengthsto
improvethestabilityof the specmm•estimateat the higherfre-
4. Applications quencies.
Thespectrum
from2.6x10
"sto !.5x10
"4Hz is based
on hourly data collectedbetween 1985:001 and 1990:.003,with
Thetoppanelof Figure3 shows
thespectrum
of earthstrainat the larger gravity and thermaltides removed. The dotted•
Pillon
FlatObservatory.
Thepeakathighfrequencies
(0.1Hz) is showsthe spectralmodelusedto computethe wandershownin
musedby microseisms;
the narrowpeaksaroundmultiplesof ! the lower panel. In the lower panel, the solid -line is the
cycle/day(1.16x 10-5 Hz) arecaused
by eattitfidesandthermal wander computedfrom the full •trmn and the dotted line
effects.Exceptfor thesenarrowpeaksthe specmnn is very well that computedif the microseisinpeak is falteredout (as indi-
it by equation(6), the piecewise
power-lawmodel. The lower catedby the solidline in the top panel).
panel
ofFigure
3 shows
thevalueof <yT:2>
• computed
fromthis
model,
andalsofor onein whichthespecmnn
is assumed
to fall
offasf-2 above0.1 Hz, as wouldbe trueif the datawere sampleAllan variancedescribedby Rutman[1978]) is givenby
highpassed
to removemicroseismic
energy.At periods
from10 anintegralsimilar
toequation (4),butwithsinnxft
if 2 replacing
the wanderis lessthan0.1 n•, whichmay be sin
to 100 seconds, 2•cft. Thedashedlinein thelowerpanelof Figure 3 shows
taken
to be theresolution such this quantity,and indicatesthat the standarddeviationof (for
limit of thisdatafor rapidchanges
asthecoseismic offsetsdiscussed by Wyatt [1988]. At times example)annualmeansof sea level is 4 cm. In this casewhat
longer
thana fewdaysthefluctuations
increase
steadily. may be of more interestis the rms rate of changeover time T,
Figure4 shows similar resultsfor sea level. In this casethe gottenby dividingthe wanderby T. Neither versionof wander
high-frequency by oceanswell,andnarrowpeaks in Figure4 increases
peakis caused fasterthanT, so the rmsrateof changewill
arecaused
by the fides,but the specumnas a wholecanbe fit by diminishwith increasing
T. While this lendssomeconfidence
to
afewpower-laws. If wecompute thewanderfromthisspecum•, attemptsto use thesedatato determinetrue secularincreases'm
weobtaina nearlyconstant linein thelowerpanel): sealevel (e.g., Barnett[1983]), it is nevertheless
value(dotteA clear that short
thevariation
causedby swelldominates out recordsare basically uselessfor this purpose;any rates deter-
overallotherchanges
to periodsof years(sincewe haveremovedthe fides). ff we mined from them will be dominatedby randomchangesfrom
ftlter
outtheswell,we obtaina wander
of about1 cmforperiods varioussources[Sturges,1987].
fromminutes
to hours, increasing
atlonger times.However, sea The wanderin strainshownin Figure3 increases with largerT
level
isnotusually evaluated asa pointmeasurement;
rather,the more rapidlythanthe sealevel wanderdoes,but still at a rateless
usual is to takemeanvaluesoverlongertimes. We can thanT, andin factabout
approach asTw. If welooked atrates, wewould
easily
usethetechniques of Section
2 to findthewanderof two againseethatthe longerthe time intervaloverwhichwe camIrate
adjoining
means,averaged overthe samelengthof time as strains,the slowerthe apparentrate. It hasbeena frequentobser-
selmares
them: the varianceof this quantity(twice the two- vation of geologists
that many rates of deformation,
examined
 19448007, 1992, 4, Downloaded from https://agupubs.onlinelibrary.wiley.com/doi/10.1029/91GL02832 by Inst of Microbial Technology, Wiley Online Library on [10/08/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
336 Agnew:
TheTime-Domain
Behavior
of Power-law
Noises

Honolulu Seo Level: Power Spectrum References

n- 4o Agnew,
D.C., Onnoise
levels
in deformation
measuremenw.
.-,. comparison
inthefrequency
domain,U.S.Geol.
Surv.
Open.
File Report,87-591, 839-844, 1987.
-• 20 Barnett,
T.P.,Possible
changes
in global
sealevelandair
causes,ClimateChange5, 15-38. 1983.
o 0 Bracewell,
R.,TheFourier
Tranbform
anditsApplications,
381
pp., McGraw-Hill, New York, 1965.
Feder,
$.V.,Fractals,
(Plenum
Press,
NewYork),1988.
Geweke,
J.,andS.Porter-Hudake,
Theestimation
andapp!icati0n
oflongmemorytimeseries
222-238, 1983.
models,
J. Time
Ser.Analy.,
4,
-40 ß I I1,111

10 -9 10 -7 10 -5
Frequency(Hz)
10 -3 10 -1
Gilbert,
L.E.,Aretopographic
data
sets
fractal?,
PureAppl.
phys., 131, 241-254, 1989.
Gilluly,
J., Distribution
of mountain-building
in geologic
time,
Honolulu Seo Level- Wonder Geol.Soc.Amer.Bull.,60, 561-590,1949.
Graf,H.-P, Long-range
correlations
andestimation
of theself-
,..., 10-1 similarity
parameter
Spectra
offractals,
Ph.D.
thesis,
205pp.,
E Eidegenoissche
Technische I-Iochschule,
Zurich,1983.
Keshner,
M.S., 1/f noise,Proc.!EEE,70, 212-2!8,1982.
Lindsay,
W.,andC. Chic,Theory
of oscillator
instability
based
uponstructure
functions,
Proc.IEEE,64, 1652-1666,
1976.
• 10-2 Malinavemo,
A.,Testing
linear
models
of sea-floor
topography,
PureAppl.Geophys.,
131, 139-155,1989.
' I rain I hr I•ø•' I)'" Mande!brot,
B.,TheFractal
Geometry
ofNature,
466pp.,W.I.I.
10 -3 -- I I illJill I i IllJill i iiii1111 I I JillIll I i iii!111 I I iiiii11 i I Jillill I I IIII111 I !1111I
Freeman,San Francisco,1983.
0ø 102 104 106 108 Mande!brot,
B., and$. VanNess,Fractional
Brownian
motions,
Time between measurements(sec) fractional
noises,andapplications,
SIAMRev.,10, 422-439,
1968.
Fig.4. Theupperpanelshows
thepowerspectrum
of sealevel Mohr,D.L., Modelingdataasa fractional
Gaussian
noise,
Ph3.
at Honolulu. The spectrumfrom 0.001 to 0.2 Hz is basedon
thesis,150pp.,Princeton
University,
1981.
2-secondaveragesof sea-bottompressurerecords,collected
between1964:216and 1964:226;from 3.5x10-• to 0.0017Hz Munk,W., andD. Cartwright,
Tidalspectroscopy
andprediction,
Phil. Trans.Roy.Soc.,Ser.A, 259, 533-581,1966.
it is based
on a similarrecordwith5-minute sampling
collected Power,W.L., T.E. Tullis,S.R.Brown,G.W. Boitnott,
andC.It.
between1964:246 and 1965:29 [Snodgrass et al., 1966;
Shodgrass 1964]. The spectra at longerfrequencies
arebased Scholz,
Roughness
of naturalfaultsurfaces
Geophys..
Res.
Lett., 14, 29-32, 1987.
on hourly tidal heightsmeasuredwith a conventionalfloat
Press,W.H., Flickernoisesin astronomy
andelsewhere,
Com-
gaugewiththe largestfidesremoved
[MunkandCartwright ments.Astrophys.7, 103-119, 1978.
1966];
from106 to 3.5x10
-• Hz thedataspanwas1905:001
to 1947:158,
andfrom3x10-9 to 106 Hz thedataspan
was Rutman,I., Characterization
of phase andfrequencyinstabilities
in precision
1905:007to 1973:214.The dottedline showsthe spectral frequency
sources:
fifteen
yearsof progress,
Proc.
IEEE, 66, 1048-1075, 1978.
modelusedto compute thewandershownin thelowerpanel. S!epian,
D., Onbandwidth,
Proc.IEEE,64, 292-300,1976.
In thelowerpanel,thedottedlineis thewandercomputedfrom
Snodgrass,
F. E., Precision
digitaltidegauge,
Science,
146,198-
thefixllspectrum
andthesolidlinethatcomputed if thehigh- 208, 1964.
frequencypeak is filteredout; the dashedline is the wanderfor
Snodgrass,
F. E., G. W. Groves,K. F. Hasselmann,
G. R. Miller,
adjacent
averages
overan intervalequalto the timebetween
measurements. W. H. Munk,andW. H. Powers,Propagation
of oceansw•
acrossthe Pacific,Phil. Trans.Roy.Soc.,Ser.A, 259, 431497,
1966.
over long times, appearto be much less than thosedetermined Sturges,
W., Large-scale
coherence of sealevelat verylowfre-
overshortertimes(for example,geodetically).But suchbehavior quencies,
J. Phys.Oceanog.,17, 2084-2094,1987.
is exactlywhatwouldbeexpected if thedeformations
beingcon- Wyatt, F., Measurementsof coseismicdeformation in southern
sidereal,
like thoseof shorterperiodshownin Figure3, had a California:1972-1982,or.Geophys.
Res.,93, 7923-7942,1988.
power-lawspectrumwith index between-3 and -1. Sucha sto- Wyatt,F., S.-T. Morrissey,
andD.C. Agnew,Shallowborehole
chastic
modelallows usto rexxmcile
apparent
changes of ratewith tilt-A reprise,J. Geophys.
Res.,93, 9197-9201,1988.
uniformitarianism
of rates(the "null" hypothesis of Gilluly
[1949]):the presentwill alwaysappearto be the mostactive D.C. Agnew,IGPP-0225,
University
of California,
La Jolla,
period,wheneverit happens
to be. CA 92093-0225

Acknowledgements.
Thisworkhasbeensupported
by NASA (ReceivedOctober8, 199!;
grantNAG-5-905. acceptedNovember7,1991)