JOURNAL OF GEOPHYSICAL RESEARCH, VOL. 90, NO.

B2, PAGES 1894-1900,FEBRUARY 10, 1985

A Renormalization Group Approach to the Stick-Slip Behavior of Faults

R. F. SMALLEY,JR., AND D. L. TURCOTTE

Department of Geolo•TicalSciences,Cornell University, Ithaca, New York

SARA A. SOI•I•A

Laboratory of Atomic and Solid State Physics

Cornell University,Ithaca, New York

We treat a fault as an array of asperitieswith a prescribedstatisticaldistribution of strengths.When an

asperity fails, the stresson the failed asperity is transferredto one or more adjacent asperities.For a
linear array the stressis transferredto a singleadjacentasperityand for a two-dimensionalarray to three
adjacentasperities.Using a renormalizationgroup (RG) method,we investigatethe propertiesof a scale
invariant hierarchicalmodel for the stochasticgrowth of fault breaks through induced failure by stress
transfer. An extrapolation to arbitrarily large scalesshowsthe existenceof a critical applied stressat
which the solutionsbifurcate.At stresseslessthan the critical stress,virtually no asperitiesfail on a large
scale,and the fault is locked.Above the critical stress,asperityfailure cascadesaway from the nucleusof
failure; we interpret this catastrophicfailure as an earthquake and it correspondsto the transition from
stick to slip behavior on the fault. Thus the stick-slipbehaviorof most faults can be attributed to the
distribution of asperitieson the fault. We proposeour stochasticmechanismas an alternativeto the
traditional hypothesesfor the stick-slipbehavior of faults. A major advantageof our approachis the
inclusionof scaleinvariance.Thus the observedfrequency-magnitude relation for seismicityis a natural
consequenceof our basichypothesis.

INTRODUCTION acceleration of gravity, and z the depth. A wide variety of

The worldwide distribution of seismicactivity is largely ex-
plained by the plate tectonic hypothesis.The relative motion
of two surfaceplates imposesa constant velocity displacement
condition on the boundary zone between two plates. This dis-
placementis accommodatedby displacementsin near-surface
fault zones.A fault zone generallyincludesa number of identi-
fiable faults. A large fraction of the active faults appear to
behave in a stick-slip rather than a stable sliding manner.
Stresson a fault builds up until a critical condition is reached
and a displacement occurs on the fault; this results in an
earthquake.
One approach to the understanding of earthquakes is to
considerthe fracture of a uniform elasticmedium subjectedto
an applied deviatoric stress.Since there is a high density of
faults and joints in the earth's crust, it is clearly not appropri-
ate to consider the fracture of pristine rock; however, it may
be argued that faults heal, so that a fracture model may be
applicable.
An alternative model for the behavior of faults is a planar
surface between two elastic half spaces,and the shear resist-
ance to displacementon the boundary is modeled in terms of
a coefficient of friction. Several authors have modeled
stick-slipbehavior of faults in terms of frictional effects.Weert-
mann [1979] modeled the instability in terms of a frictional
stresson a fault that decreaseswith increasingslip velocity.
Stuart and Mavko [1979] modeled the instability in terms of a
strain-softeningconstitutiverelation for the fault zone.
The direct application of the friction coefficientmodel to
faultspredictsthat the shearstressz is given by
ß =fpz
wheref is the friction coefficient,p the crustal density, g the
Copyright1985by the AmericanGeophysicalUnion.
Paper number 4B1324.
0148-0227/85/004B-1324505.00
experimentalstudieson friction predict that the coefficientof
friction on a fault should be near 0.8 [Byeflee, 1978]. Since
earthquakesextend to a depth of 15 km on the San Andreas
fault, a model for the stick-slip behavior of a fault must be
applicableto this depth. Taking f= 0.8, p = 2.8 x 103 kg
m-3, •/ = 10 m s-2, and z = 15 km, we find • = 340 MPa.
Any stress associatedwith fracture and rehealing must be
larger than the frictional stress.When applied to the San An-
dreas fault, the stress level given by the friction hypothesis
gives a heat flow anomaly that is much larger than the ob-
served values [Lachenbruchand Sass, 1980; Turcotte et al.,
1980]. One way to reduce the frictional stressis to require a
fluid pressureon the fault that is nearly equal to the lithostatic
pressure.However, this is a difficult condition to satisfyin a
porous fault zone. Another alternative is to attribute the loss
of heat to hydrothermal flows. This alternative has been con-
sideredby Lachenbruchand Sass [1980] who argue that it is
unlikely. Another difficulty of the friction hypothesisis the
predictionof a small fractional stressdrop during a slip event.
Stressdrops of about 10 MPa are required by seismicobser-
vations and studies of strain fields associated with observed
the surfaceoffsets on faults. Since the heat flow studiesprovide
evidence that the mean stresson the fault is less than 20 MPa,
a significantfractional stressdrop is required.The apparent
contradiction between the heat flow observations and the lab-
oratory friction measurementsis the basisfor this paper.
Why might the use of laboratory-derived friction coef-
ficients be an inadequate basis for fault models? We argue
that real faults are scale invariant, whereas laboratory friction
experimentsprescribea scale.The prescribedlaboratory scale
is the size of the asperitieson the surfacesin contact or the
sizeof the particulatematter placed betweenthe surfaces.An
actual fault has asperitiesand barriers on a wide range of
scales. Evidence for scale invariance comes from the wide
range of applicability of a power law relationshipbetween
earthquakefrequencyand earthquakemagnitude.
1894
 SMALLEY ET AL.: RENORMALIZATION APPROACH TO FAULT BEHAVIOR
We compare the behavior of faults to the behavior of the
atmosphereand oceans.The laboratory friction experiments
may be analogousto laminar flow studiesin the laboratory.
Laminar flow is not applicableto the atmosphereand oceans
becausethe flow is turbulent. We argue that the behavior of
real faults is analogousto the turbulent behavior of the atmo-
sphereand oceans.Over a wide range of scales,turbulence is
scaleinvariant. There are flow eddieswith a wide range of
sizesin turbulent flow just as there are asperitiesand barriers
with a wide range of sizeson real faults.
An expressionof scaleinvariance is the fractal dimension of
a process[Mandelbrot, 1982]. The simplestexample of scale
invarianceis the lengthof the coastline of a rocky island.The
length scale I is the length of the rod used to make the
measurement.The perimeterPt (lengthof the coastline) of the
islandobtainedusinga measuringrod of length l is related to
the length of the yardstickby
Pt • I 1- D
where D is the fractal dimension of the coast line. Note
the length of the coast line increaseswhen measured with a
smaller rod becausemany more small indentations are then
included. The fractal dimension of the coast line of Great
Britain over a wide range of scalesis D _• 1.25 [Mandelbrot,
1967]. Fractal dimensionshave been found for a variety of
turbulentphenomena.One exampleis the fractal dimensionof
the surfaceof cloudsin the atmosphere,with D • 2.35 [Hents-
chel and Procaccia, 1984]. Turbulent atmospheric processes
apparently lead to scaleinvariance for clouds.
Using frequency-magnitudeand moment-magnitude re-
lationships.Aki [1981] has shown that the fractal dimension
of a fault is D = 3b/c,whereb is the slopeof the log frequency-
magnitude relation and c is the slope of the log moment-
magnituderelation. For c = 1.5 [Hanks and Kanarnori, 1979]
and b = 1 the fractal dimensionis 2, the sameas the topologi-
cal dimensionof a plane [Aki, 1981].
Recently,renormalizationgroup (RG) techniqueshave been
successfully applied to scaleinvariant natural systems[Wilson
and Kogut, 1974; Fisher, 1974]; an exampleis a systemunder-
going a continuousphase transition. A characteristicfeature
of a phasetransitionis a discontinuous(catastrophic)change
of macroscopicparameters of the systemunder a continuous
changein the systemsstatevariables.Clearly, an earthquakeis
a catastrophicchangein the macroscopicresponseof a fault
system.
RG techniqueshave been usedby Madden [1983] to relate
the macroscopicelectricalconductivity and fracture of rocks
to the microcrackpopulation and by Allegre et al. [1982] to
studythe coalescence of fractures.Newmanand Knopoff[1982,
1983] have also studied the coalescenceof fractures, and while
they use the term renormalization in their work, no rescaling
is done, so their approach is substantiallydifferent from the
usual RG methods.
FORMULATION OF THE PROBLEM
In this paper we model a fault as an array of asperitieswith
a statistical distribution of strengths.We will consider both
linear and two-dimensional arrays of asperities.In order to
illustrate the approach we will first considera linear array as
illustrated in Figure 1. This model should be appropriate for
large-scaleasperities(barriers) on a long fault. The fault is
broken into n elements of length •ix, and each element is as-
signedan asperityfailure strengthas; the asperitywill fail
when the stress on the asperity reaches this value. The as-
1895
15
O0.... ,•n.
r'--I ....
r--2 .....
I 2 3 4 5 6 7 8
I 2 3 4
I 2
I
Fig. 1. ]||ustration of the statistical distribution of asperity
strengths
as for a lineararray;a0 is a reference
asperitystrength.
An
asperityis assignedto each unit length ax. Also shown are the cell
sizes for orders r = 1 to 4.
peritieshave a distribution of strengthswhich will be specified
by a statistical distribution function. When a stressa is ap-
(2)
pliedto the fault,all asperities
with a failurestrengthas < a
that will fail.
A hierarchical model is now constructed by dividing the
linear array of n asperitiesinto n/2 cells,each containing two
asperitiesas illustrated in Figure 1. The statisticsof cell failure
follow from the statisticsof asperity failure, since a cell fails
only when both asperitiesin the cell fail. Two alternative pro-
cesses lead to cell failure. Direct failure results when both
asperitiesbreak under an applied stressa. When only one
asperity in the cell fails, cell failure may be induced by stress
transferto the unbrokenasperity.(The strengthas of the un-
broken asperity must be greater than 2a if it is to survive the
transfer of stressa from the broken asperity.)The mechanism
for induced failure simulatesthe transfer of stressto adjacent
unbroken regionsthat occurswhen a crack is introduced into
an elasticsolid. By restrictingstresstransfer to the interior of
a cell we introduce an approximation which rendersthe prob-
lem tractable.A crucial featureincludedin this approximation
is that the size of the unbroken region within which stressis
redistributedis of the order of the sizeof the failed region.
Such a simple mechanismfor stressredistribution leads to a
complex behavior if used as a basis for a hierarchical model.
Subsequentorders in the hierarchy are obtained by treating
cells of order r as asperities of order r + 1, which are then
combined into pairs to obtain cells of order r + 1 as illus-
trated in Figure 1. Confining the redistribution of stressto
within cells yields cells that are independent. The use of a
hierarchicalmodel introducesa couplingamong adjacentcells
when they are treated as asperitiesat the next level of iter-
ation. The hypothesisof scaleinvariance implies the existence
of a unique model descriptionof the systemwhich is indepen-
dent of length scale except for a prescribed change in the
pertinent parameters.Such invariance is the basis for the RG
approach. In this caseit is implementedthrough a statistical
distribution of asperitystrengthswhich is invariant in that it is
the same for asperitiesof all orders except for an overall re-
normalization of stress.
The processof stresstransfer and induced failure tends to
increasethe lengths of segmentsOf broken asperities.For a
given applied stressthe length of the maximum segment of
broken asperities is the correlation length. As the applied
stress is increased, a value of stress is reached at which the
correlation length becomesinfinite and failure of an infinite
length of asperitiesoccurs.This stressis the critical stressand
the value at which the behavior changescatastrophicallyfrom
stick to slip. The stressat which this change occurs is analo-
1896 SMALLEY ET AL.' RENORMALIZATION APPROACH TO FAULT BEHAVIOR

(0)
gous to the temperature at which a phase transition occurs.
The statisticaldistribution of energiesin a solid, liquid, or gas 1.0

is analogousto the statisticaldistribution of asperity strengths

in our model. The utilization of the RG method allows us to 0.5
study the developmentof failed segmentsas the characteristic
lengthsof the failed segmentsincreasewith increasedapplied
stress. O0 0.5 1.0 1.5 '2.0 2.5
Our model for fault failure is similar to the one for failure of
a fibrous composite material (or a stranded cable). If one (b)
strand fails, the stresscarried by that strand is transferredto 1.0
adjacent unbroken strands. Extensive numerical simulations dP,
of the failure of composite materials have been carried out d(•oo)
[Harlow and Phoenix, 1982-].Two results are of interest. The
first is that a compositematerial will fail catastrophicallyafter
only a few strands have failed. Second, the load that can be -0 0.5 I0 1.5 2.0 25
carried is less than the load that can be carried if the strands
%
had been combinedinto a single,load carrying member.How-
Fig. 2. (a) Dependenceof the probability P1 that failure of an
ever, an important differencebetweena fault and a composite asperitywill have occurredon the normalizedstressa/a o. (b) Depen-
material is scaleinvariance. The scaleof a compositematerial dence of the probability that failure will occur at the normalized
is prescribedby the size of the fibers. Thus an RG approach stressa/a o. This is the changein the probabilityof failure gP1 when
would not be applicableto its failure. thereis a changein the normalizedstressrS(a/ao).
We next consider the distribution of asperity strengths.
Clearly, a wide variety of strengthson a wide variety of scales
must exist on any real fault. For example, fault bends and averagestrengthof the asperities.It is clear that stresstransfer
offsets correspond to strong asperities. However, data on to the adjacent unbroken sectionsof a fault will occur on a
actual distributionsof asperitystrengthsare not available.It is real fault.
possiblethat studiesof the type given in this paper may allow In order to quantify the failure of asperitiesdue to the
asperity distributions to be inferred from such seismicobser- transferof stresswe introducethe conditionalprobability
vations as the dependenceof earthquake frequency on mag- that failure will occur when a stress(a-b)a is transferredto an
nitude. In the absenceof applicable data we assumea qua- unbroken asperity supporting a stressba, so that the final
dratic Weibull distribution for the probability Pa that the fail- stresson the asperity is aa. This conditional probability is
urestrengthaœof an asperityis lessthanthestress
art relatedto the probabilitythat aœis both largerthan ba and
smallerthan aa by [Meyer, 1970]
Prob(aœ< aa)-- Pa= 1 - e-(•x)2 (3)
where Prob (ba < aœ< aa)
Pa,b= (5)
Prob(aœ> ha)
X -- G/GO
A simple geometrical interpretation of both probabilities
and a0 is a referenceasperitystrength.Wiebull distributions neededto obtainPa,bfollowsfromtheidentity
are often used to representa statisticaldistribution of failure
strengths[Harlow and Phoenix, 1982]. It should be empha-
sized that our approach can be applied to any continuous
Pa =
j0
•
dx
dx a,/,o
dP•(x)
distributionof asperitystrengths.The probability
whichidentifiesPa as the area underthe dP•/dx curvebetween
Pi= 1--e-X2 (4) x = 0 and x = aa/ao, shownin Figure 3a. Then
that aœ< a is shownin Figure2a as a functionof a/ao.Ten
percent of the asperitieshave failed when a/ao = 0.32, fifty Prob(ha < af < aa)= Ia./.o
d•/•o
dP•(x)
dx
dx = Pa- Po (6)
percent of the asperitieshave failed when a/ao = 0.83, and
ninety percentof the asperitieshave failed when a/ao = 1.52. is the area under the dP1/dx curve between x = ba/ao and
Since the relation betweenP• and a is invertible, P• can be x = aa/ao, shownin Figure 3b' and
usedas a measureof the applied stress.The probability den-
Prob(% > ba)= 1 - Prob(% < ba)= 1 - Po (7)
sity that failure will occurat the appliedstressa/ao is givenby
dP•/dx and is shown in Figure 2b. The probability density is the area under the dPi/dx curve betweenx = ba/ao and
thataœ= a is zeroat zerostress
andincreases
to a maximum infinity, shownin Figure 3c. Substituting(6) and (7) into (5)
at a-0.71 a0. The mean strength of an asperity is d = yields
(x/•/2)a0
= 0.8862
a0.
An essential feature of our model is the transfer
from a failed asperity to its nearest neighbors.Without this
of stress
P"'o
= 1--Pb (8)
transfer of stress the behavior of the system is simple and Note that for a probabilityfunctionof the form (3),
uninteresting. Sincestrongasperitieswill not break until large
P, = 1 -- (1 -- P•)": (9)
stressesare applied,they can block the propagationof broken
segments,and there is no changefrom stick to slip behavior In principle,this problemcouldbe solvedwithoutthe useof
on the fault. It is the transfer of stressfrom broken asperities the RG technique.However,the range of scalesthat could be
onto the remaining unbroken asperitiesthat leads to cata- studiedis quite limited evenwith the largestcomputersavail-
strophic behavior at an applied stressthat is less than the able.Althoughwe will utilize the RG methodin the standard
 SMALLEY ET AL.: RENORMALIZATIONAPPROACH TO FAULT BEHAVIOR 1897

manner, it should be recognizedthat the approach is semiem- 1,0

pirical and has been principallyjustified by its successin solv- dPi
ing a variety of fundamental unsolved problems in physics. dx 0.5
These problems fall in a broad class in which a continuous
systemon the microscopicscaleexhibitsdiscontinuousbehav- 0
ior on the macroscopicscale. We argue that the stick-slip 0 o' x
behavior of faults falls in this generalclassificationof physical
l,
problems.
We will first illustrate the application of the RG technique
to the failure of a linear array of asperities,using a basic cell dx 0,5 (b)
composedof two "asperities"and the probability distribution
given in (2). Note that the (r + 1)th-order "asperities"which 0
•' •' x
result from r iterations of the RG transformation
actual first-order asperities.The first three renormalizations
contain 2•
o ø
are illustrated in Figure 1. For a cell containing two asperities
which are either broken or unbroken,four statesare possible: dx O. dPi
I'01 (C)
(1) [bb], (2) [bu], (3) [ub], and (4) [uu], where b representsa
broken asperity and u representsan unbroken asperity.Note
that states2 and 3 are equivalent and can be combinedinto a 0 bo" x
singlestate with a multiplicity of 2. The probabilitiesfor each %
of these states,neglectingany interactions between asperities, Fig. 3. (a) The probabilityP. that the failure strengthas of a
are given by given asperityis lessthan aa correspondsto the area under the dPx/
dx curve between x = 0 and x = aa/a o. (b) The probability P,- Pb
[bb] [ub] [uu] thatthefailurestrengthas of a givenasperityliesbetweenba andaa
P! 2 2P•(1
• P•) (1• V•)2 (10) correspondsto the area under the dP•/dx curve between x- ba/ao
andx = aa/ao.(c) The probability1 - Pbthat the failurestrengthas
of a given asperity exceedsba correspondsto the area under the
wherethe probability of the failure of an individual asperityis dP•/dx curvebetweenx + ba/ao and infinity.
P1.
Next, it is necessaryto consider the influence of a broken
asperity on an adjacent unbroken asperity. We use the con- determine
P•(•+i) fromP•(•),wherer is theorderof thecell
ditional probabilityP2.x that an unbrokenasperityalready beingconsidered.The generalform of (15) is
supportinga stressa will fail when an additional stressa is
Pz(•+z)= 2Pl(')[1 - (1 - pz(•))4]_ (p•(r))2 (16)
transferredto it from an adjacent broken asperity.This mech-
anism for transfer of stress leads to induced cell failures. We Equation (16) implies that a relation of the form (9) between
must prescribe a condition for determining whether an rth P• and P2 is valid at all orders r, as follows from the assump-
order cell is broken or unbroken. We assume that an rth order tion of invariance of the statistical distribution of asperity
cell is broken only if both "asperities"in the cell are broken. strengths.The probabilityPx(•) for failure of an rth-levelas-
Probabilities for asperity and cell behavior with stressinterac- perity is given by a Weibull relation of the form (4) in terms of
tions are a renormalized
stressa(•)= x(•)aoif

[bb] [ub] [bb] [ub] [uu]

p12 2P(1 -- P)P2, 2P(1 -- P1)(1-- P2,) (1 -- p)2 (11)
[bb] + {[ub]-, [bb]} [ub] + [uu]
+ 2P(1 -- Vl)V2, (1 -- V)2 + 2P(1 -- P•X1- V2,1)

The conditionalprobabilityfrom (8) is givenby

(x(•+•))2= _ In {2 exp[ -- 4(x(•))
2]
P2- P•
- -- 2 exp[-5(x(•))2] + exp[-2(X(r))2]} (17)
P2,•
-- 1- P•
Thus the functional form of the probability distribution is
Substitutionof (12) into (11) givesthe probability that a cell
preservedexcept for an overall stressrenormalization, as re-
hasfailed [b2] or not failed [u2]
quired by the hypothesisof scaleinvariance.Our use of a scale
invariant model is motivated by evidence of scaling in the
[b2]
2p•p2 _ p12
[u2]
1 + p•2 _ 2p•p 2
(13) nucleationand growth of cracks[Allegre et al., 1982].
The dependence
of Px('+x)on Px(')is givenin Figure4. The
Under this conditionthe probability that a first-ordercell is points 0 and 1 are stable fixed points of the system. The
broken,P •(2),is givenby straightline corresponding
to Pl•'+ •= P•'• is alsoincluded
p•(2)= 2plp 2 _ p•2 (14) in Figure 4. The iterative relation crossesthis straight line at
Px•'•= P*= 0.2063. We now show that P* is an unstable
and substitutionof P2 from (9) gives fixedpoint that separatesthe regionof stick behaviorfrom the
region of slip behavior.The RG iteration can be performed
p•(2) = 2P•[1 -- (1 -- P•)4] -- p12 (15)
graphicallyusingFigure 4. For example,we take P• = 0.6 and
For higher order cells,(15) is usedas an iteration equation to from (15) find pl(2)= 0.8093.This cell behaviorat order 1
1898 SMALLE¾ ET AL.: RENORMALIZATION APPROACH TO FAULT BEHAVIOR

1,0 I0

0,8

p•(r)
0,6

•
(r+•)

0,4 0• O•
02.
øo I 2 3 4 5
0
Fig. 5. Dependence of the probabilityof failureP•(') on the order
0 0,•- 0,4 0,6 0,8 1,0
r for severalvaluesof P•. The bifurcationof the solutionat P• = P*
= 0.2063 is clearly illustrated.

Fig. 4. Dependenceof the probability of failure for the r + 1 cell

P•('+•) on theprobability
offailureof ther cellP•(')for cellscontain- the RG transformation
(16).Given the dependence
of P•('+ •)
ing two asperitieswith a quadraticWeibull distributionof strengths.
The proceduredescribedin the text for determiningthe probability
on P•{'),the slopeof the curvein Figure4 at P• = P* is given
iterationsis illustratedfor P• = 0.6, 0.1. by
of cell failure for successive
The critical probability of failure P* givesthe bifurcationpoint for
catastrophicfailure of the system.If P1 < P*, the solutioniteratesto
P•= 0 and no failure occurs.If P• >P*, the solution iterates to
P• • = 1, and the systemhas failed.
A -- ....
dP•(r+
•)1
dP• 09 t'•=1',
(19)

As long as (P*-P•'•)<< 1, a linear approximationto (16) is

valid, and
now becomesthe asperity behavior at order 2. To do this p, _ p•(r+ •)
graphically,a horizontalline is extendedto the line P•{'+ •) -
P•{') to reflectthe total cell behaviorat order 1 into the as-
A= P*-P•(r) (20)
perity behavior at order 2. Thus the probability of cell failure It then follows that [Wilson and Ko•7ut,1974]
at order2 is p•{3)= 0.9615.Thisprocedure is repeatedto give
A = b•/v (21)
p•{4)= 0.9985,etc.,and the probabilityof failurerapidlyap-
proachesunity as the order is increased.On the other hand, if where b is the linear rescalingfactor. For the b = 2 RG trans-
we take P• = 0.1, we find p•(2)= 0.05878,p•(3)= 0.02184, formation that led to (16) we obtain A = 1.6189, so that
p•m= 0.00322,etc.,and the probabilityof failuredecreases v = 1.4388.
towards zero as the order is increased. If P• > P*, failure So far, we have consideredonly a linear array of asperities.
occursfor infinite length scales,and slip behavior results.If We will next consider a two-dimensionalarray of asperities
P• < P*, the behavior is stable,and failure occursonly on the distributeduniformly on a planar fault as illustratedin Figure
smallestscales.Bifurcation of the solution occurs at P• = 6. We will divide the two-dimensionalarray of n asperities
P*-0.2063, and the critical stress leading to failure is into n/4 cells each containing four asperities.The failure of
a*= 0.4807 ao from (3). The correspondingvalue of x*= individual asperitieswill be treated in the sameway as in the
a*/ao - 0.4807 is the fixed point solution to the iterative rela- linear case,and (3) is assumedto be applicable.When one or
tion (17) for the renormalization of stress.The dependenceof more asperitiesin a cell fail, we assume that the stress on
P•(') on r for severalvaluesof P1 is givenin Figure 5. The those asperitiesis transferredequally to the remaining as-
bifurcation of the solution at P• - P* = 0.2063 is clearly illus- peritiesin the cell. That is, if one asperityfails, the stresson
trated. Note that the value of the critical stressis considerably the three remaining asperitiesis 4a/3. We choosefour as-
less than the value of the mean strength of an asperity ff = peritiesin a cell so that the stressin the field regionis applied
0.8862 ao.
The stable behavior of the system at a < a* can be
characterized by a correlation length L which measuresthe
maximum length over which failure occursfor P• < P*. The
rapid increaseof L as the thresholdis approachedfrom below
is describedby a power law
L oc(P* - p•)-v (18a)
or equivalently,

L oc(a* -- a)- v, (18b)

where v is the correlation length exponent [Wilson and Koqut,
1974]. According to this result the magnitude of precursory
seismicitywould be expected to increase as the critical stress 1
on the fault is approached.The onset of catastrophicbehavior
at a = a* correspondsto the divergence of the correlation Fig. 6. Illustration of the two-dimensionalarray of asperitieswith
length L. four asperitiesper cell. Second(2), third (3), and fourth (4) order cells
The correlation length exponent v is easily obtained from are also shown.
 $MALLEY ET AL.' RENORMALIZATIONAPPROACHTO FAULT BEHAVIOR 1899

over a length which is of the order of the length of the failed TABLE 1. Renormalization Group Resultsfor the Critical
region.We again assumethat the cell fails when all asperities Probability P*, the Critical Stressa*/a o, and the Correlation
in the cell fail. Length Exponent v in One and Two Dimensions

A second-ordercell is composedof four first-order cells or d P* a*/a o v

second-order asperities and therefore sixteen primary as-
perities, as illustrated in Figure 6. The statistics of failure of
the second-orderasperitiesand cells is the same as that of the
first-order asperitiesand cells. Again, the processis repeated
by iteration to infinite order. The RG transformation thus
constructedcorrespondsto a linear rescalingfactor b = 2 on a
two-dimensionalarray. This case is considerably more com-
plex than the linear array example consideredabove. Follow-
ing the same procedureillustrated in (10), (11), and (13) and
usingthe definition of the conditional probability, we find that
the probability that a cell fails is given by
1 0.2063 0.4807 1.4388
2 0.1707 0.4327 0.8084
theseresultsand those for the linear array, as listed in Table 1,
illustrate the effect of the physical dimensionality d on the
critical behavior of the system.Note the decreasein both the
critical probability P* and the correlation length exponent v
with increasingd. Simplerpercolationmodelsexhibit the same
trend when d is increasedfrom two to three [Stauffer, 1979].

p•(2)= p• {4)+ 4P• 3(1-- P•)P4,• + 6p•2(1 -- P•)2[P2,•2

CONCLUSIONS
+ 2P2,•(1- P2,•)P4,23
+ 4P•(1-- e•)3{e4/3,13
We have proposed a model for the behavior of faults that is
+ 3P4/3,•2(1-P4/3,•)P4,4/3
+ 3P4/3,•(1-P4/3,•)
2 basedon the hypothesisof scaleinvariance.We postulate that
fault behavior is controlled by a distribution of asperitiesand
[P2,•/32+ 2P2,•/3(1-P2,•/,)P4,2]} (22) barriers on all scales. In addition to the frictional behavior of
and introducing(8), we obtain fault gougeon the smallestscale,the barriers include bendsin
p•(2)= P• 4 + 4P• 3(e4 -- e•) + 6p•2(e2 _ p1)2 the fault and offsets of fault strands. We also lump con-
centrations of stressand variations of material properties into
+ 12p•2(P2- P•)(P4- P2) our statistical approach. We postulate that the asperitiesand
barriers have a statistical distribution of strengths.Scale in-
+ 4P•(P4/3 - p•)3+ 12P•(P4/3 - P•)2(P4- P4/3) variance requires that the same statistical distribution of
+ 12P•(P4/3 - P•)(P2- P4/3) 2 strengthsbe applicable at all scales.This is the basis for the
applicability of the RG approach.We introduce a mechanism
+ 24P•(P4/3 - P•)(P2- P4/3)(P4-P2) (23) for transfer of stresswhen an asperity fails. This transfer of
The dependence of P•{•+•) on P•{•)shownin Figure7 follows stressto adjacent asperitiesis an essentialfeature of our ap-
from introducing Pa from (8) and using (23) as an iterative proach. However, our resultsare not qualitatively sensitiveto
relation. the details of the transfer process,i.e., to how many asperities
The general behavior of this two-dimensional case is the the stress is transferred to.
same as that of the linear example consideredabove. Again, The stick-slip behavior of faults is a natural consequenceof
an S-shapedcurve is generated.The points 0 and 1 are stable the renormalization group approach to the statistical model
fixed points. The crossingat P* = 0.1707 separatesstick from presentedhere for asperityfailure. A critical stressis found at
slip behavior. From (3) the bifurcation of the solution occurs which the catastrophicfailure occurs on a macroscopicscale.
at a*= 0.4327 ao. This is just about one half the mean Below this critical stress,very few precursoryfault breaks have
strength of the asperitiesa = 0.8862 ao. We also find that occurred. This is in agreement with observations.There is
A = 2.357; the correlation length exponent v = 0.8084 follows virtually no seismicactivity on the locked southern section of
from (20) with b = 2. The quantitative differencesbetween the San Andreasfault. We predict that there will be no signifi-
cant precursory activity on the fault prior to the next great
IO earthquake.We also find that the value of the critical stressis
lessthan the mean strength of the asperities.We believe that
0.8 this is an explanationfor the low stresslevelsassociatedwith
displacementson the San Andreasfault.
0.6
Clearly, the model presented in this paper involves many
simplifyingassumptionsand certainlydoesnot predict all as-
pectsof real fault behavior.Someexamplesare the following:
pi
(r'l)o4 1. Our analysis predicts that the catastrophic failure ex-
tends to infinity. That is, a fault rupture does not terminate.
o•
Also, we do not predict the transition from stable sliding to
stick-slipbehavior. Preliminary studiesindicate that both of
o
0 02 04 06 08 I0
these difficultiescan be overcomeby introducing a more com-
(r) plex form for the asperitystrengthdistribution.For instance,a
mechanismfor blocking the growth of fault breaks can be
Fig. 7. Dependenceof the probability of failure for the r + 1 cell incorporated into the model by allowing for several peaks
p•e+•) on theprobabilityof failureof ther cellP•(r) for cellscontain- rather than a single peak in the probability of failure for a
ing four asperitieswith a quadratic Weibull distribution of strengths. singleasperityshownin Figure 2b.
The critical probability of failure P* gives the bifurcation point for
2. Our analysisdoes not predict the fractional stressdrop
the system.If P• < P*, the solution iterates to P• © - 0 and no failure
occurs.If P• > P*, the solutioniteratesto P• © = 1 and the systemhas in an earthquake. Obviously, our simple assumptionof com-
failed. plete stresstransfer to an adjacent asperity neglectsany re-
1900 SMALLEY ET AL.: RENORMALIZATION APPROACH TO FAULT BEHAVIOR
sidual asperity strengthand is an oversimplificationwhich can
be modified.The failure of a small asperityis equivalentto the
introductionof a small dislocationpair on the fault. This will
result in a small displacementon the fault whosemagnitude is
dependent upon the "stiffness"of the adjacent media. When
adjacentasperitiesfail, the size of the dislocationpatch grows.
This, in turn, allows a larger displacementto occur as long as
the failedasperitiesdo not lock.
We would like to emphasizethat the approachpresentedin
this paper can be expanded and improved. We considerit to
be a preliminary effort to introduce a new approach to the
understandingof fault behavior. Our primary purpose is to
introduce the concept of scale invariance and some of its im-
plications. Some of the features found in laboratory friction
experiments are incorporated through our simple model for
asperity behavior, but it is the concept of scale invariance
introduced by the RG approach which leads to the results
discussedhere for the macroscopicbehavior of faults.
Acknowledgments.We would like to acknowledgevaluable dis-
cussions with A. Ruina and F. Horowitz. This research has been
supported in part by the National Aeronautics and Space Adminis-
tration under contract NAS 5-27340 and grant NAG 5-319. This is
contribution 766 of the Department of Geological Sciences,Cornell
University.
REFERENCES
Aki, K., A probabilisticsynthesisof precursoryphenomena,in Earth-
quake Prediction: An International Review, Maurice Ewing Ser., vol.
4, editedby D. W. Simpsonand P. G. Richards,pp. 566-574, AGU,
Washington, D.C., 1981.
Allegre, C. J., J. L. Le Mouel, and A. Provost, Scalingrules in rock
fracture and possible implications for earthquake prediction,
Nature, 297, 47-49, 1982.
Byerlee,J., Friction of rocks,Pure Appl. Geophys.,116, 615-626, 1978.
Fisher, M. E., The renormalization group in the theory of critical
behavior, Rev. Mod. Phys.,46, 597-616, 1974.
Hanks, T. C., and H. Kanamori, A moment magnitude scale,J. Geo-
phys.Res.,84, 2348-2350, 1979.
Harlow, D. G., and S. L. Phoenix, Probability distributions for the
strengthof fibrous materials under local load sharing, 1, Two-level
failure and edgeeffects,Adv. Appl. Probab.14, 68-94, 1982.
Hentschel, H. G. E., and I. Procaccia, Relative diffusion in tubulent
media: The fractal dimension of clouds, Phys. Rev. A, in press,
1985.
Lachenbruch,A. H., and J. H. Sass,Heat flow and energeticsof the
SanAndreasfault zone,J. Geophys.
Res.,85, 6185-6222,1980.
Madden, T. R., Microcrack connectivityin rocks: A renormalization
group approach to the critical phenomena of conduction and fail-
ure in crystallinerocks,J. Geophys.Res.,88, 585-592, 1983.
Mandelbrot, B., How long is the coast of Britain? Statistical self-
similarity and fractional dimension,Science,156, 636-638, 1967.
Mandelbrot, B., The Fractal Geometryof Nature, W. H. Freeman,
San Francisco, Calif., 1982.
Meyer, P. L., Introductory Probability and Statistical Applications,
Addison-Wesley,Reading,Mass., 1970.
Newman, W. I., and L. Knopoff, Crack fusion dynamics:A model for
large earthquakes,Geophys.Res. Lett., 9, 735-738, 1982.
Newman, W. I., and L. Knopoff, A model for repetitivecyclesof large
earthquakes,Geophys.Res.Let., 10, 305-308, 1983.
Stauffer, D., Scaling theory of percolation clusters,Phys. Rep. 54,
1-74, 1979.
Stuart, W. D., and G. M. Mavko, Earthquake instability on a strike-
slip fault, J. Geophys.Res.,84, 2153-2160, 1979.
Turcotte, D. L., P. H. Tag, and R. F. Cooper, A steadystatemodel for
the distribution of stressand temperatureon the San Andreasfault,
J. Geophys.Res.,85, 6224-6230, 1980.
Weertman, J., Inherent instability of quasi-static creep slippageon a
fault, J. Geophys.Res.,84, 2146-2152, 1979.
Wilson, K. G., and J. Kogut, The renormalization group and the s
expansion,Phys. Rev. C, 12, 75-200, 1974.
R. F. Smalley,Jr., and D. L. Turcotte, Department of Geological
Sciences,Cornell University,Ithaca, NY 14853.
S. A. Solla, Laboratory of Atomic and Solid State Physics,Cornell
University,Ithaca, NY 14853.
(Received October 31, 1983;
revised March 27, 1984;
acceptedOctober 26, 1984.)