PHYSiCA

ELSEVIER Physica A 224 (1996) 48-67

The dynamics of sandpiles:

The competing roles of grains and clusters
Anita Mehta a'l, G.C. Barker b'2, J.M. Luck c'3, R.J. Needs d'4
a S.N. Bose National Centre for Basic Sciences, DB 17, Sector 1, Salt Lake City, Calcutta 700 064, India
b Institute of Food Research, Norwich Research Park, Colney, Norwich NR4 7UA, UK
c Service de Physique Thgorique, C.E. Saclay, 91191 Gif-sur-Yvette cedex, France
d TCM Group, Cavendish Laboratory, Madingley Road, Cambridge CB30HE, UK

Abstract

The competing roles of grains and clusters in the dynamics of a driven sandpile are investigated
via two models. First, we present a coupled-map lattice model of a slowly rotating sandpile:
we find that our inclusion of parameters to represent grain reorganisation and inertia leads to a
rich diversity of behaviour, which we interpret via a simple model of grains and voids. Next,
we present a set of coupled nonlinear stochastic equations which model the interactions between
immobile clusters and mobile grains; in particular, we are able to model the effect of rotation via
the inclusion of a 'tilt' term. We find that these nonfinearities are relevant and we present the new
and nontrivial exponents we obtain in a number of different cases.

1. I n t r o d u c t i o n

Sandpile models are now being widely studied in the context of granular media [ 1-
3]. While most theoretical approaches thus far [4] have focused on the response of
a driven pile via flowing grains down its surface, our approaches to this problem thus
far [5,6] have always emphasised the coupling between the moving grains and the
relatively immobile clusters and its impact on sandpile dynamics. To illustrate this, we
here present two models: a coupled-map lattice (CML) model of a sandpile subjected

t E-mail: anita@veccai.emet.inor anita@bose.emet.in

2 E-mail: barker@bbsrc.ac.uk
3 E-mail: luck@amoco.saclay.cea.fr
4 E-mail: rnl I @phy.cam.ac.uk

0378-4371/96/$15.00 (~) 1996 Elsevier Science B.V. All rights reserved

SSDI0378-4371 (95)00314-2
 A. Mehta et al./Physica A 224 (1996) 48-67 49

tO constant tilt in a rotating cylinder [7] and a set of coupled stochastic nonlinear
equations to model a driven sandpile [8].

2. The coupled-map lattice model

Sandpiles in rotating cylinders are a case of great experimental importance [3,9] ; we

have therefore focused on this system in our CML model, which illustrates the effects of
grain reorganisation and inertia on the avalanche spectra of driven sandpiles. We were
motivated by the following anomaly: while the earliest models of sandpiles [4] predicted
a power law for the avalanche size distribution function, thus indicating scale-invariant
dynamics, experiments [ 10] showed a deviation from this due to the presence of a
disproportionately large number of large avalanches. To explain this, we identified [ 1,5 ]
two distinct relaxational mechanisms in a reorganising sandpile: a faster dynamical
mechanism corresponding to the motion of grains moving independently of their clusters,
and a slower mechanism corresponding to the collective reorganisation of grains within
their clusters. In earlier work [ 11 ], we showed that the avalanche spectrum obtained in
a cellular-automaton model based upon these ideas agreed with experiment [ 10]; this
indicated that the observed breakdown of scale invariance in real systems was due to
the competition between the two mechanisms.
In the present paper, we use the idea of competing grain-cluster mechanisms and focus
on the situation where cluster reorganisation dominates grain flow in a sandpile subjected
to a very slow constant rotation; the grains form part of a continuum so that the column
heights, hi, are real variables, while column widths, i = 1 . . . . . L, are kept discrete as
usual. In addition to the normal surface flow, our granular driving forces include a
contribution proportional to the deviation of the column height from an 'ideal' height;
this ideal height is a simple representation of the limit of random close packing, so
that columns which are taller (shorter) than ideal would be relatively loosely (closely)
packed, and driven to consolidate (dilate) when the sandpile is perturbed externally.
Our motivation for this choice comes from previous work [ 12] where we have shown
that a shaken sandpile tends to consolidate or dilate, depending on its initial state; we
generalise this here to include the cases of rotation and deposition. Thus the force fi
acting on column i reads

fi=kl(hi-iaSo) +k2(hi-hi-1 -aSo) (i ~ 1), (1)

where hi are the column heights, kl and k2 are constants, a is the lattice spacing and
iaSo is the 'equilibrium' height of column i. We note that
• the first term, which depends on the absolute height of the sandpile, corresponds to a
force that drives column compression or expansion towards the reference height. Since
we normally deal with columns which are more dilated than their normal height, we
will henceforth talk principally about column compression;
• the second term is the usual term driving surface flow, which depends on slope, or
height differences; the offset of So is the reference slope from which differences are
50 A. Mehta et al./Physica A 224 (1996) 48-67

measured.
The definition of relative heights zi = hi - iaSo leads to the dimensionless represen-
tation

fi = ( k l / k 2 ) z i + zi -- Z i - I (i ~ 1). (2)

When column i is subject to a force f i greater than or equal to the threshold force
fth, the height changes are as follows:

Zi ~ Zi -- ~Z,

zi-I --~ zi-l + 8 z ' (i -~ 1). (3)

Thus, the height 8z removed from column i leads to:

• a flow of grains, with total height increment 8z ~, from column i onto column i - 1,
and
• a consolidation of the grains in column i which reduces the column height by 8z - 8z ~.
The coupling between the column heights expressed in Eq. (3) may lead to the
propagation of instabilities along the sandpile and hence to avalanches. Since avalanches
have also been discussed widely in the context of earthquakes, we draw some useful
analogies, choosing for this purpose a discrete model of earthquakes put forward by
Nakanishi [ 13]. In our model [7] as well as in his, the redistribution of force at a
relaxation event is governed by the parameter A = 2 ( 1 + k l / k2 ) / ( 1 + ( 1 + k l / k2 ) 2); the
undistributed force 1 - A is dissipated. Note, importantly, that in our sandpile model,
this dissipation is linked to non-conservation of the sandpile volume arising from the
compression of columns towards their ideal heights, in keeping with our modelling of
granular reorganisation.
Our force relaxation function has a minimum value when fi = fth, and increases
monotonically with increasing fi. This models the stick-slip friction associated with
sandpiles and earthquakes, whereby nothing happens below the threshold force fth, but
after this threshold is exceeded, the relaxation increases in proportion to the excess
force. Accordingly, the minimum value of this function is known as the minimum event
size and its initial rate of increase a is called the amplification [ 13]. In physical terms,
amplification refers to the phenomenon whereby grains collide with each other during
an avalanche, and their inertial motion contributes to the amplification of the avalanche;
thus a is a measure of granular inertia. Finally, we can rewrite the map in terms of the
forces as
f i -- fli = 2 t ~ Z / A , (4)
8z' = 8z/(1 + kl/ka), (5)
f ; - I - f , = f;+l - f,+, = - d ( f ~ - f,)/2 (i~ 1 or L). (6)
This prescription classifies our model as a local and unlimited sandpile [ 14]. It can be
shown [7] that for all A ~ 1, the largest part of the volume change occurs as a result
of consolidation; the quantity of interest is thus the difference between the old and new
configurations, rather than the mass exiting the sandpile [ 11 ]. A measure of this change
 A. Mehta et al./Physica A 224 (1996) 48-67 51

L
is the quantity In M = In ~-i--I ( f i - f : ) , which is the analogue of the event magnitude in
earthquake models [ 13]. We will now discuss the variation of this quantity as a function
of the model parameters.
For a sandpile in a rotating cylinder, the tilting of the sandpile results in changes
of slope over the complete surface in a continuous manner, in contrast to the case of
random deposition where the slopes change locally and discontinuously at a deposition
event [ 11 ]. Our CML model is driven continuously; from a configuration in which all
forces f i are less than the threshold force, elements of height, z+, are added onto each
column with

z + = i ( f t h -- f j ) / ( 1 +jkl/k2) (i = 1 ..... L), (7)

where f j = m a x f i . This transformation describes the effect of rotating the base of the
sandpile with a constant angular speed until a threshold force arises at column j. We use
boundary conditions appropriate to a sandpile in a rotating cylinder open at L = 0, in
which the height of the zeroth column is pinned at zero and the height of the (L + 1 )-st
column is (L + 1)aS0. The response to the tilting is, as described above, a flow of
particles down the slope as well as an internal reorganisation of particles.
For all A < 1, the critical sandpile starts with a slope greater than So, subsequently
decreases and then becomes steady at So towards the top of the sandpile; this state
is an attractor [7] for the system, distinct from that generated by standard lattice
sandpiles [ 11 ], and is close to the S-shaped sandpile shape observed in rotating cylinder
experiments [ 10]. The steady-state response of the driven sandpile may be represented
as a sequence of events each of which corresponds to a set of column height changes.
Each avalanche is considered to be instantaneous; also, we set fth = So = a = 1. In
Fig. 1 we plot the distribution function per unit time and length R(lnM), against In M
for sandpiles with size L = 512 and parameter values 8 f = 0.01, a = 2, 3, 4 and
A = 0.65, 0.85, 0.95. The distribution functions in Fig. 1 indicate a scaling behaviour
in the region of small events and, for larger event sizes, frequencies that are larger than
would be expected from extension of the same power law. The phase diagram in the
d-a-plane indicates qualitatively distinct behaviour for:
• low inertia, strongly consolidating (low ot and A) systems where the magnitude
distribution function has a single peak, and
• high inertia, weakly consolidating (high a and A) systems for which the magnitude
distribution has a clearly distinct second peak.
These results are in accord with the Nakanishi earthquake automaton [ 13 ]. The effect
of a large avalanche on the configuration of the sandpile is illustrated in Fig. 2, where the
relative column heights zi are plotted against the distance of the column from the axis
of rotation. The solid line denotes the configuration before, and the dotted line that after,
the avalanche; we note that a section of the sandpile has 'slipped' quite considerably.
We now explain the results in Fig. 1. The rotation of the sandpile causes a uniform
increase of the local slopes and, in the upper region of the sandpile, a preferential
increase of absolute column heights. The sandpile is thus driven towards its critical
shape where local forces are driven above threshold, and relaxation events are triggered.
52 A. Mehta et al./Physica A 224 (1996) 48-67

2- a_~3..O a----4.0
6=0.65

~o-
-2-

-4 I I I I I I I I i

2- a=3.O Or=4.0

~o-
tr

-2-

.4
L
I I I i i

a=2.0 a=3.0 a--4.0

t A=0.95 I &=o.95 &=0.95

~o-
:E

,,'r"

-2-

-4 i
i i i i i i i
-4 -2 0 2 -2 0 2 -4 -2 0 2
Log(M) Log(M) Log(M)

Fig. I. A logarithmic plot of the distribution function of event sizes, R ( l n M ) , for 107 consecutive events in
a C M L model sandpile with L = 512 and parameter values • f = 0.01, a = 2, 3, 4, a n d A = 0.65, 0.85, 0.95.

These events will be localised ( 'small' ) or cooperative ('large') depending on a and A.

For strongly consolidating systems with small amplification a, a lot of excess volume is
lost via consolidation, and the effect of granular inertia is small; in these circumstances,
the progressive build-up of an instability is unlikely so that events are, in general,
localised, uncorrelated and hence 'small', leading to the appearance of the single peak
in the graph on the upper left of Fig. 1. Alternatively, for weakly consolidating systems
with large amplification, the inertial effect is large and excess volume in the shape of
voids is not lost via consolidation.
Fig. 3 shows the time series of avalanche locations that occur in our model sandpile
in the two-peak region. The large events are almost periodic and each is preceded by
many small precursor events; this is in accord with previous work [ 11 ] on sandpiles as
well as earthquakes [ 13]. In addition it is apparent that large avalanches tend to occur
repeatedly at or around the same regions of the sandpile, whose location changes only
very slowly compared to the interval between the large avalanches; these correlations in
both the positions and the times of large events are often referred to as 'memory' [ 1 ].
To conclude this section, our main result here is that for large te and A, we see
a preferred size for large avalanches, which is manifested as a second peak in the
 A. Mehta et al./Physica A 224 (1996) 48-67 53

.~.
._~
0.0-

-0.1
?
"l-
r-
E -0.2
0
0
(I)
.>
•~ -0.3-
n-

-0.4-

0
I
20
I
40 60
I
80
I I
100
I
120
I
Position
Fig. 2. A plot of column heights, relative to their critical heights, for a CML model sandpile with L = 128
and parameter values tSf = 0.01, a = 3 and zl = 0.85. The full (dotted) line shows the configuration before
(after) a large event.

Q.

i i , .111ollll~l~,llllll8 l, I

I I I I I I
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4
Time

Fig. 3. A plot showing the locations of relaxation events (changes in column heights), that occur during an
interval of length 1.5 which begins at t = l04, for a CML model sandpile with L = 256 and parameter values
6 f = 0.01, ct = 3, and A = 0.85.
54 A. Mehta et al./Physica A 224 (1996) 48-67

Fig. 4. A schematic diagram illustrating the mechanism for large-avalancheformation. When A is large, there
is a great deal of undissipated volume in the cluster, resulting in the upper (shaded) grains being unstable
to small perturbations. When o~ is large, the black grain hitting the cluster has large inertia so that a large
avalanche results when it dislodges the shaded grains.

distribution of event sizes (Fig. 1 ). In terms of our simple picture above, this is because
of the following: for large t~ and d, grains have enough inertia to speed past available
voids, and there is consequently a large amount of unrelaxed excess volume on the
surface. This excess volume can be visualised for instance as a precariously balanced
cluster (Fig. 4); the oncoming (dark) grain will knock off the shaded grains when it
hits them, unleashing a large avalanche. For small ot and A, we see by contrast mainly
small events leading to a single peak in Fig. 1; we visualise this by imagining slowly
moving grains (low inertia) drifting down the surface, locking into available voids and
dissipating excess volume efficiently. This picture also indicates that initiated avalanches
will be terminated relatively rapidly, leading to many small events. The large avalanches
mentioned above are quasi-periodic and tend to occur repeatedly around the same regions
of the surface (Fig. 3), providing an important quantification of configurational memory.
In terms of the simple picture above, this is because of the following: regions of the
sandpile which look like Fig. 4 are 'wiped clean' by the effect of the large avalanche,
so that further driving has no effect for a while. However, the effect of large a and
zl mean that once again, unrelaxed volume will be created around the same region
after high-inertia grains flow down the surface, leaving in their wake space-wasting
configurations such as the one in Fig. 4; this will be the case after a number of
small events have occurred (Fig. 3). These spatial and temporal correlations result in
a quasi-periodic repetition of large avalanches around the same regions of the sandpile,
resulting in configurational memory [ 1 ]. Thus, the inertia of moving grains leads to the
amplification ot of avalanches; 1 - A measures the volume lost when grains are able to
occupy voids wholly or partially, so that A represents the effect of reorganisation.

3. Coupled local equations for a driven sandpile

We now change gear to present continuum stochastic equations designed to model

a driven sandpile [8]; these contain the same basic ingredients of surface and bulk
exchange which have been a feature of much of our previous work [ 1 ]. Here, as in
the C M L model presented above, the underlying picture is of currents of grains moving
down the slope, knocking out bumps and filling in holes in clusters on the landscape.
 A. Mehta et al./Physica A 224 (1996) 48-67 55

While this picture is embodied in the CML model via the parameters a and 4, its
description here is via nonlinear equations involving a coupling between explicit grain
and cluster coordinates.
There have been a number of recent approaches [ 15-17], which have attempted to
model the surfaces of driven sandpiles; they are, however, unified by the fact that they
involve only one variable, the local height of the surface h ( x , t), rather than any form
of grain-cluster coupling 5.
In our approach the effective coordinate representing clusters is the local height
h ( x , t), whose variations manifest the geometric fluctuations of clusters, while the
moving grains are represented by their local density p ( x , t). In a driven sandpile these
coordinates will clearly be coupled; we choose to represent this coupling via the coupled
nonlinear Langevin equations [ 1,8] whose analysis concerns us here.
Before presenting the details of our model, we review some general facts about
rough interfaces. Three critical exponents, a, fl, and z, characterise the spatial and
temporal behaviour of a rough interface. They are conveniently defined by considering
the two-point correlation function of the heights G( x - x ' , t - t t) = ( h ( x, t) h ( x ' , t' )) -
(h(x, t ) ) ( h ( x ' , t')). We have

a ( x , O ) ~ l x l 2~ (Ixl--,e~), G ( O , t ) ~ l t ] 2/3 (Itl~c~), (8)

and more generally

a ( x , t) ,.~ Ixl2~F(Itl/Ixl z) (9)

in the whole long-distance scaling regime (x and t large). The scaling function F is
universal; t~ and z = ot/fl are respectively referred to as the roughness exponent and the
dynamical exponent of the problem.
The stochastic dynamical equations [ 1,8] under investigation here concern a one-
dimensional sandpile, described by two coupled variables as follows: h ( x , t) represents
the profile of relatively immobile material, measured from a fixed negative critical slope
- p c . In other words the actual height of immobile particles above a horizontal reference
line reads

y ( x , t) = - p o x + h ( x , t), (10)

whereas p ( x , t) represents the density of the layer of moving particles.

The dynamical equations have the following general form:
Oh/Ot = Dh X72h - T + ~h(X, t), (lla)
Op/Ot = - V j + D p V 2 p + T + rip(X, t). (llb)

For the time being we set

5 A recent approach due to Bouchaud et al. [ 18], which appeared after the publication of our equations in
Ref. [ 1 ], also features a form of grain-cluster coupling. However the nonlinear couplings added there are
different from ours, and in particular are irrelevant in that no crossover to a different universality class is
observed in that model. Also, they do not include the effect of flit, which is, in our view, one of the most
crucial physical ingredients of our model.
56 A. Mehta et al./Physica A 224 (1996) 48-67

T = - x p V 2 h - hp(~Th)+ - tzp•h - u ( V h ) _ (12)

and

j = -yp(Vh)_. (13)

We have introduced the notations ~7 = 8/Ox, and

{z, for z > 0 , {0, for z > 0 ,

z+= 0, for z <_ 0, z-= z, -
for z_<0. (14)

The content of our dynamical equations is as follows.

• The first term on the right hand side of Eq. (1 la) represents the rearrangement of
clusters in the presence of an applied noise; the associated coupling D h is a diffusivity.
This corresponds to collective relaxation, denotes strictly intracluster motion, and
exists in the quasistatic as well as the grain-inertial regimes [ 1]. The absence of this
term would lead to an unphysical situation in the limit T = 0, when there are no
flowing grains; the clusters would be configurationally frozen and a vibrated sandpile
(finite noise in h) would have no means of relaxation! This is clearly inappropriate
given that clusters have been observed to rearrange [ 1-3 ] under conditions of low-
intensity vibration even when no mobile grains flow down the pile; in fact this
process of cluster rearrangement at low vibrational intensities provides a mechanism
for collective relaxation and has been shown [12] to be crucial for the compaction
of a pile to high densities.
• The second block of terms, T, represents the transfer of clusters to flowing grains,
and vice-versa. This transfer process is crucial to sandpile dynamics, and includes
representations of evolving disorder and inertia. Included in T are the following
terms:
- The term Kp~72h represents intercluster motion initiated by the flowing grains. The
essential difference between this term and the diffusive term for h is that, whereas
this mechanism stops in the absence of avalanches, the diffusive process continues
even in this limit. We visualise this as being due to a current of grains moving
down the slope, knocking out bumps and filling in holes, so that it is a term that
can exist in the grain-inertial regimes but not in the quasistatic one. Note that
this is a representation of inertia since it is a mechanism for amplifying sandpile
avalanches independently of slope.
- The term tzp~Th also represents intercluster motion of grains, which is mediated by
the motion of grains moving independently of each other down the slope. Thus, a
current of grains moves down the slope, accumulating at points of low slope and
knocking out grains from regions of large slope. Also, whereas Kp~72h smooths
out bumps and dips on the pile irrespective of slope, the IzpVh term exists purely
to smooth out deviations from the critical slope. This slope dependence allows
us to view this term as a representation of evolving configurational disorder and
memory [ 1 ] ; thus, an overly bumpy section of this pile (large deviation from critical
slope) will lead to a large avalanche even if very few grains (small p) hit it.
 A. Mehta et al./Physica A 224 (1996) 48-67 57

- The term u (XTh)_ represents the spontaneous generation of flowing grains whenever
the local slope is larger than critical; this exists even in the absence of flowing grains
and is meant to represent the effect of tilting a stationary sandpile. Note that, because
we have chosen the critical slope to be negative by definition, a negative sign of
XTh implies that the overall slope is steeper than critical. This term is a simple
representation of a crucially important effect in sandpile dynamics, and, as we will
see later, is in large part responsible for the novelty of our model [8] from the
viewpoint of critical phenomena. However we emphasise here that in its absence a
noiseless (undeposited) pile subjected to slow tilt would stay frozen. Its physics is
therefore crucial in that it is able to reproduce the everyday phenomenon whereby
an originally static pile, when tilted, generates mobile (flowing) grains.
- The term ,~p(XTh)+ can be viewed as a crude representation of the effect of the
boundary layer (whose width represents the maximal range for cluster-grain ex-
change to occur) [ 1 ] since its action is to limit the release of flowing grains
generated by the effect of tilt; it thus acts as a regulator on the generation of mobile
grains. While there is in principle no restriction on the amount of p (up to the size
of the pile) that can be generated in a real sandpile subjected to tilting or applied
noise, we restrict ourselves to the situation where tilting is a moderate perturbation,
and only grains occupying the boundary layer can be liberated to flow down the pile
via the h - p conversion in the tilt term. After a short transient the system relaxes to
a stationary state where (Op/Ot) = 0. We thus have (p) ~ ~,/A finite at saturation.
This also makes quantities such as h and p, as well as their fluctuations, finite and
measurable, which is essential for the numerical simulations described hereafter.
Hence ,~ plays the role of an experimental (and numerical) cutoff: we replicate
existing experimental approaches [ 3] where vibration and/or tilt are perturbations
rather than catastrophes, and it is mainly the boundary layer that is affected by
these [2].
• The first term in Eq. ( 1 lb), -xYj, represents the variation in p due to the nonunifor-
mity of the current of flowing grains, in such a way that the total number of particles
is conserved. The current j ( x , t) is proportional to the number of mobile grains and
to their velocity. As shown by Eq. (13), the latter is assumed to vanish for slopes
greater than critical, and to be proportional to the driving field which is the magnitude
of the local slope.
• The second term in Eq. (1 lb), X72p, represents the relaxation of the flowing grains,
and is a crude way of representing intergraln collisions. The coupling Dp is again a
diffusivity.
• Finally, to a discussion of the source terms rib(x, t) and rio(x, t). They will depend on
the physical situation under consideration. We shall often take them as two indepen-
dent Gaussian white noises, characterised by their widths Zlh, Zlo, defined according
to

(rlh(X, t)rlh(X', t')) = d 2 8(X -- X ' ) 6 ( t -- t'),

(15)
(rip(X, t )rlp( X', t') ) = d2p t3( x -- x') 6( t -- t').
58 A. Mehta et al./Physica A 224 (1996) 48-67

Thus, pouring grains onto a sandpile should be represented by noise only in p;

alternatively one might imagine that the sandpile is being subjected to vibration at its
base, the chief effect of which is transmitted to the surface clusters via the bulk, and
in this case one should include only noise in h. For the moment, we consider both
noise terms to be present.
To close this presentation of our dynamical equations, we briefly present their explicit
solution in the absence of noise. Let the initial situation be that of a uniform slope
P = - P c + e of immobile grains: y ( x , O ) = px, i.e., h ( x , O ) = ex, together with a
constant density of mobile particles: p ( x , O ) = Po. We thus have a uniform XTh = e. In
the subsequent evolution this slope remains constant, and the amounts of mobile and
immobile particles are related by the conservation law h ( x , t ) = h ( x , O ) + Po - p ( t ) .
The evolution of p ( t ) crucially depends on the sign of e, namely
• Subcritical case ([Pl < Pc, i.e., e > 0): the density of mobile particles relaxes
exponentially to zero, according to

p( t) = poe -(a+jz)Et. (16a)

• Supercritical case (IPl > Pc, i.e., e < 0): the density of mobile particles diverges
exponentially, according to

p ( t ) = - P l + (Po + Pl )e ~L~lt, (16b)

with pl = v / ~ .
• Critical case (JpJ = pc, i.e., e = 0 ) : t h e system is entirely frozen, namely

p ( t ) =P0. (16c)

The dynamics in the critical case is thus driven by the fluctuations generated by the
noisy source terms.
It is worth noticing that the characteristic times associated with the relaxation law
(16a) and with the law of divergence (16b) both diverge as ~" ~ 1/lel as the critical
slope e = 0 is approached.
We now investigate the critical behaviour of our dynamical equations in the presence
of noise.
It turns out that the most important new ingredient in our dynamical equations is the
tilt term v ( V h ) _ describing the physics of tilting a pile so that, quite simply, clusters
of grains which appear frozen and stationary when the pile is horizontal release grains
which flow down the pile when tilted. However, this is not the only term that induces
new universality classes. For instance, as discussed later, the effect of symmetry between
x and - x , and different combinations of the noise terms, induce changes of universality
class. These phenomena will be discussed under the relevant headings of this section;
the corresponding estimated critical exponents are listed in Table 1.
More specifically, in order to get an idea of which terms of our equations are the most
relevant, in the sense of the renormalisation-group approach to critical phenomena, we
 A. Mehta et al./Physica A 224 (1996) 48-67 59
Table 1
Critical exponents a,/3, and z for both fields h(x, t) and p(x, t), measured from numerical simulations of
the coupled stochastic equations in the various cases described in text

Model Species a 13 z

EW h 1/2 1/4 2
KPZ h 1/2 1/3 3/2
(l) asymmetric h 0.94 4- 0.07 0.43 4- 0.04 2.2 4- 0.4
noise m h p 0.22 4- 0.08 0.07 4- 0.02 *
(la) no-tilt h 0.36 4- 0.02 0.41 4- 0.08 0.9 4- 0.2
home m h and in p p 0.39 4- 0.05 0.27 4- 0.04 1.4 4- 0.4
(2) asymmetric h 0.80 4- 0.06 0.42 4- 0.07 1.9 4- 0.4
noise m p p 0.33 4- 0.10 0.19 4- 0.04 1.7 4- 0.9
(3) asymmetric h 0.97 4- 0.07 0.45 4- 0.03 2.2 4- 0.3
noise in h and in p p 0.12 4- 0.08 0.07 4- 0.03 *
(4) symmetric h 0.404-0.06 0.37 4-0.04 1.1 4- 0.3
norse m h p -- _ -

(5) symmetric h - - -

noise in p p 1/2 1/4 2

(6) symmetric h 0.37 4- 0.05 0.39 4- 0.05 1.0 4- 0.2
nome m h and in p p -- --

The exact EW and KPZ values are recalled for comparison. A hyphen means that the corresponding quantity
is in our opinion not critical. An asterisk means that the exponent z cannot be accurately evaluated from the
available data on a and ft.

can use the a r g u m e n t s o f d i m e n s i o n a l analysis referred to as p o w e r - c o u n t i n g [ 19]. For

the sake o f clarity, we first perform this perturbative analysis on the K P Z equation [20]

Oh/Ot = D • 2 h + g ( V h ) 2 + r / ( x , t). (17)

We rescale space according to x ~ b x , and time according to t ~ b z t, with z being the

u n k n o w n d y n a m i c a l exponent. We make the natural hypothesis that the noise r / ( x , t)
is d i m e n s i o n l e s s , in the sense that its width A, defined in analogy with Eq. ( 1 5 ) , is
i n d e p e n d e n t o f b. We can then determine iteratively the power o f the linear scaling
factor b which affects every quantity or parameter o f the model under rescaling, in any
d i m e n s i o n d. By definition the negative o f this power is called the (classical or naive)
d i m e n s i o n o f the quantity u n d e r consideration. Starting from the d i m e n s i o n s [ x ] = - I ,
It] = - z , [A] = 0, we o b t a i n

[hl=(d-z)/2, [D]=z-2, [g]=(3z-d-4)/2. (18)

The general ideas o f the r e n o r m a l i s a t i o n - g r o u p approach imply that an operator is

relevant (respectively, irrelevant) whenever the d i m e n s i o n of the associated c o u p l i n g
constant is positive (respectively, negative).
• C o n s i d e r first the linear ( E W ) theory [ 2 1 ] , obtained for g = 0. This linear theory
is scale invariant for [ D ] = 0, w h e n c e z = 2. We thus recover the k n o w n scaling
60 A. Mehta et al./Physica A 224 (1996) 48-67

properties of the EW theory, in particular [hi = ( d - 2 ) / 2 , whence the EW exponents

a = [h]/[x] = (2 - d ) / 2 , and fl = [h]/[t] = (2 - d)/4.
• Consider now the full KPZ theory. With z = 2 we have [g] = ( 2 - d ) / 2 , implying that
a weak nonlinearity is irrelevant for d > 2, relevant for d < 2, so that the perturbative
critical dimension reads dc = 2. This kind of power-counting analysis is, however,
unable to predict that the KPZ exponents are nontrivial in any finite dimension, due
to nonperturbative effects [ 22 ].
The power-counting analysis of our dynamical equations ( 11 ) goes as follows. Under
the assumption that the noise is again dimensionless, starting from Ix] = - 1 , [t] = - z ,
[ d h ] ---- [ A p ] -- 0, we obtain

[h] = [p] = (d - z ) / 2 , [Oh] = [Dp] = z - 2, [~'] = z - 1,

[ a ] = [/z] = (3z - d - 2)/2, [K] = [y] = (3z - d - 4 ) / 2 . (19)

• The pseudo-linear theory obtained for t< = a = / z = y = 0 is formally scale invariant if

we set z = 1. We have then a mean-field-like scaling with [v] = 0, where diffusion is
irrelevant, since [ O h ] = [ O p ] = - - 1 . The other mean-field exponents read a h = flh =
[h]/[x] = (1 - d ) / 2 , and similarly for p.
• The full theory then has critical dimension dc = 1, where we have [A] = [/.t] = 0
(marginal) and [K] = [y] = [Dh] = [Dp] = - 1 (irrelevant).
Perturbative power-counting is even more questionable in the present case than for the
KPZ problem, since our pseudo-linear dynamical equations are already fully nonlinear.
The mean-field-like scaling exponents mentioned above may well not be observable in
any dimension. It nevertheless provides a hint concerning which perturbations are likely
to be the most relevant ones.
We now give some general information on the numerical simulations which yield the
results discussed below. The simulations have been performed by discretising Eqs. ( 11 )
both in space and in time. Since we are mainly interested in critical behaviour, we have
set the step in the spatial direction, i.e., the lattice spacing, equal to unity (a = 1).
We have employed a finite system of size L lattice points, with periodic boundary
conditions. The discretisation in the temporal direction requires more care, because of
the strong instabilities which are intrinsically present in nonlinear growth equations in
discrete time. Just as previous authors [23], we have had to use currently small values
of the time step, of order 6t = 10 -3, in order to avoid instabilities, and to generate
physically acceptable, well-behaved and stable solutions to our equations.
For each of the cases detailed below, we have evaluated the exponents of both h and
p at criticality. The exponents a and fl are a priori different for both species, whereas
a common value of the dynamical exponent z is expected. The actual evaluation of
the exponents has been done by means of the structure factors Sh (q, to) and Sp (q, to),
which are defined as the double Fourier transforms of the correlation functions G h ( X --
x', t - t') = (h(x, t ) h ( x ' , t') I and Gp(x - x', t - t') = (p(x, t ) p ( x ' , t')). The scaling
laws (8) and (9), recalled above, can be respectively recast as

S(q, to) ~ q-'-2'~crP(to/qZ), (20)

 A. Mehta et al./Physica A 224 (1996) 48-67 61

for q and to small, and especially

S(q,O),"~q -1-2a (q--*O), S(O, to) ".' to - l - 2 f l (to-+O). (21)

3.1. The asymmetric situation (Cases 1 to 3)

This situation is the most commonly encountered one, of a (sloping) sandpile with a
preferred direction of flow. We refer to it as the asymmetric case because it describes the
physics of fluctuations with respect to a uniform slope -Pc, so that there is no x ~ - x
symmetry.
In numerical simulations we have set the irrelevant couplings y and K equal to zero,
and chosen D h = Dp = 1. The transfer term T deserves some more attention. The effect
of the term Ap(Vh)+, as mentioned above, is to limit the fluctuations of p around its
finite mean value of order (p) ,,~ ~,/A. Critical fluctuations can therefore only develop
for ~, >> ~. We choose to set A = 1,/z = 0 (again for simplicity), keeping u >> 1 as a
free parameter, besides Ah and zip. In practice values of ~, of order 10 to 50 turn out to
be large enough in order not to alter the critical fluctuations, for sizes (L < 1000) and
observation times (t _< 105) actually used in our numerical simulations.
The dynamical equations thus read

ah/at = V2h - T + rib(X, t),

Cases 1 to 3: ap/c~t = V2p + T + 71p(x, t), (22)
T= -p(Vh)+ - p(Vh)_.

We have considered the following three cases, according to the nature of the noise:
• Case 1: noise in h (A h > 0, Ap = 0);
• Case 2: noise in p (Ah = 0 , Zlp> 0);
• Case 3: noise in h and p (Ah > 0, zip > 0).
Non-trivial long-range spatial and temporal critical fluctuations are observed for both
species h and p in these three cases. Case 1 is of particular interest, as it is the most
directly comparable with EW and KPZ. We observe a rougher behaviour of the h-
profile than the two aforenamed. This effect is very pronounced in the spatial direction
(Ceh = 0.94 + 0.07 is to be compared with 1/2 in both cases), and still appreciable in
the temporal one (flh = 0.43 ± 0.04 is tO be compared with 1/3 and 1/4). Intuitively,
this appears to be due to the roughening effect of the tilt term, whose chief role is
to cause a generation of flowing grains at points of excessively high slope on the
surface. The effect is thus of an enhanced deposition, so that we expect the values of
the exponent flh to lie between 1/4 (deposition and relaxation) and 1/2 (deposition
only, no relaxation). This is indeed what we have measured. However, given that the
roughening effect appears to manifest itself more directly in the temporal exponent,
since a KPz = a Ew, but flKPZ > flEW, we will restrict ourselves to discussing it in this
context.
In order to check further the validity of the above intuitive idea, we compare flh in
this case with that of the model where the tilt term is replaced by a noise in p. In this
62 A. Mehta et aL/Physica A 224 (1996) 48-67

situation, referred to as Case la, the dynamical equations read

Oh/Ot = ~72h - T + rib(X, t),

Case la: Op/Ot = ~72p + T + 71p(X, t), (23)
T = -Izp•h.

We find that, indeed, flh is unchanged, within error bars, with respect to Case 1 (and, as
it turns out, Cases 2 and 3 as well) verifying that the role of the tilt term is analogous
to that of a local generator of flowing grains, i.e., a noise in p.
We mention that, for the purpose of numerical stability, we have actually replaced the
transfer term T of Eq. (23) by an odd nonlinear function F ( T ) of the same quantity,
linear at small T and saturating at large T; we have chosen

-1, for T_<_ - 1 ,

F(T) = T, for - 1 < T < 1, (24)
1, forT_> 1,

in order to achieve a fast numerical evaluation.

Case 2 describes the situation of a noise in p, in the presence of tilt. This corresponds
to the case of pouring grains onto a sandpile, whereas Case 1 corresponded to shaking
one. In order to initiate nontrivial dynamics, we have to subject the pile to a transient
noise in h for a period (0 < t < to). The exponents in this case show a small, although
perhaps significant difference from the previous case. The p-exponents are slightly larger
than previously, while the h-exponents are smaller than in Case 1. Intuitively, this should
be expected, since in this case the major perturbation corresponds to a deposition of
flowing grains. However, and somewhat surprisingly, we still see that the h-exponents
in this case are larger than the corresponding p-exponents. Our tentative explanation for
this is that, whatever the nature of the noise, the transfer term T predominates; thus,
while the difference in the noise terms accounts for the relative values of, say tip, in
Cases 1 and 2, the nonlinear couplings in T ensure that in any given case, flh > tip.
Another way of saying this is that our dynamical equations result in a sandpile surface
whose roughness is due more to embedded clusters than flowing grains, which is clearly
sensible; also, a comparison of Cases 1 and 2 shows that shaking the pile is a more
efficient way of generating roughness than pouring grains down it, which also seems
intuitively plausible.
These speculations are reinforced to some extent by Case 3, which corresponds to a
noise in h and in p. Within error bars, we observe no significant difference between
this case and Case l (noise in h alone), indicating that noise in p does not materially
affect the situation of a noise in h in the presence of the tilt term. In other words, once
we have generated the 'maximal' roughness of the pile by shaking its clusters, we will
not materially change the observed roughness of the surface by, in addition, depositing
grains on it.
We have so far expressed all the surface heights as deviations from one unique critical
slope - P c . However, we know [ 1-3] that there is in reality a range of angles of repose
for a sandpile and we should have included these to describe the different thresholds for
 A. Mehta et al./Physica A 224 (1996) 48-67 63

onset and continuation of avalanches. We propose to do this via the introduction of a

threshold C, both in the tilt and the current terms. The incorporation of two thresholds,
pc and C, in our equations models the situation in real sandpiles where the former
corresponds to the minimum angle of repose (onset of avalanches) and the latter to the
maximum angle of stability (threshold for continuous avalanching). We thus write

3 h / 3 t = D h V 2 h -- T + rib(X, t),
3p/Ot = - - V j + Dp~Y2p + T + rio(x, t),
T = - K p V 2 h - Ap(XYh + C)+ - tzpXTh - v(XYh) _, (25)
j = -yp(Vh + C)_.

The threshold C can be viewed as an extra coupling constant; its dimensional content
as dictated by power counting reads [C] = (d - z + 2 ) / 2 --- (d + 1)/2, with the
mean-field-like dynamical exponent z = 1. The introduction of a nonzero threshold thus
drives the system away from criticality, at least within the perturbative power-counting
approach. This prediction has been confirmed by numerical simulations, which show
a definite crossover to noncritical fluctuations when C is switched on in a progressive
way.
From a more physical point of view, this implies that the introduction of the second
threshold C takes our system away from the critical behaviour of a second-order phase
transition to the more realistic first-order behaviour characteristic of real sandpiles [ 1-
3]. Our choice of the 'simplified' critical equations in this paper was made for the
simple purpose of demonstrating the novel critical exponents we obtain and so showing
the novelty and validity of the physics we have included; this being done, we emphasise
that in any experimental application, the full Eqs. (25) with two thresholds should be
used.

3.2. The symmetric situation (Cases 4 to 6)

We now turn to the analysis of a version of our dynamical equations which has the
x +-. - x symmetry. This can be visualised as the surface of a sandpile which is flat on
average, and is subject as before to deposition and/or shaking. Physically this represents
an important difference in that flow is allowed in both directions, unlike in the previous
case of a background slope, where gravity imposes a preferred direction for grain flow.
The dynamical equations read in full generality

Oh~at = Dh•2h - T + rih(X, t),

Cases 4 to 6: a p / a t = - x T j + D p V 2 p + T + rip(x, t), (26)
T = - K p V Z h - apIVhl + u ( I V h l - C ) + ,
j = y p s i g n ( V h ) (IVhl - C)+.

In order to investigate the critical regime, we choose to simplify the above equations
for y = C = 0 as
64 A. Mehta et aL/Physica A 224 (1996) 48-67

Oh/Ot = xTZh - T + rlh(X, t),

Op/Ot = V 2 p + T + rip(x, t), (27)
T = - p V ~ h + ( u - p)]Vh].

An inspection of Eq. (27) shows the effect of symmetrising the tilt term; a finite
value of local slope now indicates a deviation from the equilibrium flat surface, so that
it is the absolute magnitude of the local slope which results in immobile clusters being
converted to flowing grains. This symmetrisation applies equally to the other term in T,
which arranges for cluster-grain exchange depending on X72h rather than XTh. We again
consider three cases, according to the number of components of the noise:
• Case 4: noise in h (Ah > 0, dp = 0);
• Case 5: noise in p (dh = O, dp > 0);
• Case 6: noise in h and in p (dh > 0, dp > 0).
In Cases 4 and 6, we observe ah ~ flh ~ 0.40 and zh = 1 within error bars, just
as in Case la, whereas p ( x , t) does not exhibit divergent fluctuations: the structure
factors S(q, 0) and S(0, to) rather saturate to constant values for small enough values
of wavevector q or frequency to, implying the fast decay of p - p correlations at long
separations in space and in time.
These results show, for the first time, that the dynamics corresponding to anisotropic
and isotropic sandpiles are quite different. Our interpretation of these results is based on
the symmetry of the pile. While the tilt term still performs its earlier role with respect
to flh (which is identical within error bars to Cases 1 to 3), there is an important
difference between what happens to the flowing grains once they are released from their
erstwhile clusters. With no preferred direction of flow, we would expect 'backflow' of
grains on a regular basis, contrary to the anisotropic case; so that we should expect
that different clumps of p generated by possibly anticorrelated bursts of j should have
decaying correlations in space and time. A way of verifying this conjecture consists in
checking j - j correlations in this case and to compare with the asymmetric case; we
have done this and find, as expected, that j is well correlated in the asymmetric case
(Case 1 ) [corresponding to unidirectional flow], but is weakly correlated (Cases 4 and
6) in the symmetric case.
Another interesting situation is Case 5, where noise is present only in the equation for
p. Just as for Case 2, we have to put noise in h too in an initial period. We find that h
gets frozen, soon after the initial period has elapsed, into a nontrivial rough landscape,
entirely inherited from the transient period, and therefore characterised by a roughness
exponent ce ~ 0.40. The evolution of p which then takes place, with this frozen height
configuration as a background, is effectively linear, and the EW exponents are observed,
in spite of the background. This is actually to be expected by simple physical reasoning:
the effect of pouring 'flowing' grains on a flat surface (which has no preferred direction
of flow for the flowing grains p) will, after transients, have relatively little effect on
h. In the absence of this coupling, Vh rapidly approaches a frozen profile across the
surface so that the tilt term becomes inactive; the only remaining effect is that build-ups
of flowing grains are gradually diffused away across the frozen h-landscape, leading
 A. Mehta et al./Physica A 224 (1996) 48-67 65

tO the observed EW exponent of t p = 1/4. Note that things would have been quite
different if we had had coupling terms involving Vp.
To sum up this section, we have found new, nontrivial critical exponents in a number
of different cases so that our leading nonlinearities are relevant. It is possible that there
is, in first approximation, a single independent exponent a tit ~ 0.40, such that we have

Cases 1, 2, and 3: ah "~ 2atilt, t h ~ atilt, Z ~ 2;

Cases la, 4, and 6: ah ~ t h "~ atilt, Z ~ 1. (28)

We definitely cannot rule out the existence of a fine structure, i.e., of small differences
between the exponents of various cases in each of the two broad universality classes
mentioned above. The situation of the exponents related to p is less clear. Furthermore,
we are unable to compare the nontrivial values of exponents with predictions existing
in the literature, since, to our knowledge, no similar coupled nonlinear equations, which
are nontrivial in their critical behaviour and correspond directly to similar physical
situations, have so far been investigated.
In Case 1, which is most directly comparable with EW and KPZ, that of a sandpile on
a slope in the presence of vibration, we observe a roughening of the surface compared
to the two above. We speculate that this could be due to the roughening effect of the
tilt term which acts as a local generator of flowing grains in areas of excess slope; and
have crudely verified this hypothesis by comparing the temporal exponent t h with that
obtained in Case la (coupled equations in the absence of tilt, with two noises).
Let us examine the implications of Cases 2 and 3 next: everyday experience tells us
that if we want to create a pile with a rough surface, shaking would be a more efficient
way to do this compared to pouring. The reason is that shaking directly affects cluster
shapes on the bulk and surface, whereas pouring would in general tend to cause grains
to flow down the pile. In the parlance of previous work [ 1 ], a noise in h predominantly
affects the collective motion of clusters, whereas a noise in p predominantly affects
the motion of independent grains. Also since clusters are in general able to sustain
roughness more than moving grains, we would expect that the roughness exponents of
the former would exceed those of the latter species. We find indeed that, in all of the
cases of interest (1 to 3), t h > tip; and that in Case 2 (corresponding to a noise in
p alone) while the roughness of the h-landscape is suppressed, the roughness of the
p-landscape is not overwhelmingly enhanced. Also, Cases 1 and 3 are not materially
different showing that the process of deposition of flowing grains does not make much
difference to the roughness profile of a sandpile that has already been subjected to
shaking. It is remarkable that, although we have in no sense built in these features into
our equations, we appear to get physically desirable and meaningful behaviour from
them.
So far, the new universality classes we have obtained are for a sloping pile, corre-
sponding to different combinations of noise. We feel that, in analogy with other critical
phenomena, symmetry could also be an important factor in determining universality
classes. For this reason we have examined the effect of deposition and/or shaking on a
fiat surface, where the important physical difference is that bidirectional flow of mobile
66 A. Mehta et al./Phy$ica A 224 (1996) 48-67

grains is allowed in the absence of a biasing slope. One would expect intuitively that,
in the absence of noise, this would lead to the p-correlations becoming much weaker,
if not disappearing altogether. Cases 4 and 6 show that this is indeed the case. In Case
5, the absence of a noise in h causes the h-profile to stop fluctuating after a time; the
tilt term becomes inactive, and the noise in p generates a diffusive response from the
flowing grains across the frozen (but still rough) h-landscape.
Before summing up, we return to the subject of tilt, which has formed a leitmotif for
our work. Although it might appear from our menagerie of new critical exponents and
their obvious connection to the tilt term that this was a sufficient raison d'etre for the
latter, we emphasise that a very important feature resulting from its inclusion is to do
with experimental work. The rotating cylinder apparatus has been used for a very long
time [9] to investigate sandpile dynamics, and its use remains popular to this day [3].
Clearly, the driving force in this setup is the constant tilt to which the sand in the
cylinder is subjected, so that a modeling of the effect of tilt is essential in any equations
which seek to interpret such experiments. For us, the successful inclusion of tilt in our
dynamical equations brings with it the enormous potential benefit of being of use in
interpreting such a massive body of experimental work. In fact, we have recently become
aware of previous experimental work [24] which supports our theoretical predictions.
The authors of that work find, for the case of a rotated and shaken sandpile, exponents
ah = 0.92 + 0.05 and flh = 0.48 + 0.16, in remarkably close agreement with Case 1 in
our table.

4. Discussion

In conclusion, we have presented two studies [7,8] of sandpile dynamics, both of

them based on the idea of grain-cluster exchange. The first, a coupled-map lattice model
of a sandpile, focuses on the situation where reorganisation dominates flow in a sandpile
subjected to rotation. We have presented the event size distribution corresponding to
different regions of parameter space and explained our results in terms of the inertia of
the flowing grains and the reorganisation of clusters. The second, a set of phenomeno-
logical equations, models the dynamics of sandpile surfaces: they include nonlinear
couplings to represent the constant transfer between relatively immobile clusters and
mobile grains, incorporate the presence of tilt and contain representations of inertia and
evolving configurational disorder. We have looked at the response of these equations to
different perturbations, and presented our novel critical exponents for these, both in the
presence and the absence of a biasing slope. It is our hope that the work presented in
this paper will lead to further detailed experimental and theoretical investigations in this
very exciting and topical field.

References

[ 1] A. Mehta, ed., Granular Matter: An InterdisciplinaryApproach (Springer, New York, 1993).

[2] A. Mehta and G.C. Barker, Rep. Prog. Phys. 57 (1994) 383.
 A. Mehta et al./Physica A 224 (1996) 48-67 67

[31 H.M. Jaeger and S.R. Nagel, Science 255 (1992) 1523.
[4] P. Bak, C. Tang and K. Wiesenfeld, Phys. Rev. Lett. 59 (1987) 381; Phys. Rev. A 38 (1988) 364.
[5] A. Mehta, The physics of powders, in: Correlations and Connectivity, H.E. Stanley and N. Ostrowsky,
eds. (Kluwer Academic, Dordrecht, 1990);
A. Mehta and G.C. Barker, Phys. Rev. Lett. 67 (1991) 394;
G.C. Barker and A. Mehta, Phys. Rev. A 45 (1992) 3435.
[6] A. Mehta, R.J. Needs and S. Dattagupta, J. Stat. Phys. 68 (1992) 1131.
[7] G.C. Barker and A. Mehta, submitted to Phys. Rev. E (1995).
[8] A. Mehta, J.M. Luck and R.J. Needs, submitted to Phys. Rev. E (1995).
[9] R.L. Brown and J.C. Richards, Principles of Powder Mechanics (Pergamon, Oxford, 1966).
[10] H.M. Jaeger, C. Liu and S.R. Nagel, Phys. Rev. Lett. 62 (1989) 40.
[ 11 ] A. Mehta and G.C. Barker, Europhys. Lett. 27 (1994) 501.
112] G.C. Barker and A. Mehta, Phys. Rev. E 47 (1993) 184.
[13] H. Nakanishi, Phys. Rev. A 43 (1991) 6613.
[14] L.P. Kadanoff, S.R. Nagel, L. Wu and S.M. Zhou, Phys. Rev. A 39 (1989) 6524.
115] T. Hwa and M. Kardar, Phys. Rev. Lett. 62 (1989) 1813.
[ 16] G. Grinstein and D.H. Lee, Phys. Rev. Lett. 66 (1991) 177.
117] J. Toner, Phys. Rev. Lett. 66 (1991) 679.
1181 J.P. Bouchaud, M.E. Cates, J. Ravi Prakash and S.F. Edwards, J. Phys. 1 (Paris) 4 (1994) 1383; Phys.
Rev. Lett. 74 (1995) 1982.
1191 S.K. Ma, Modern Theory of Critical Phenomena (Benjamin, New York, 1976);
D. Amit, Field Theory, the Renormalization Group and Critical Phenomena (McGraw-Hill, New York,
1978; World Scientific, Singapore, 1984);
J. Zinn-Justin, Quantum Field Theory and Critical Phenomena (Clarendon, Oxford, 1989).
[201 M. Kardar, G. Parisi and Y. Zhang, Phys. Rev. Left. 56 (1986) 889.
[21] S.E Edwards and D.R. Wilkinson, Proc. Roy. Soc. A 381 (1982) 17.
[22] T. Sun and M. Plischke, Phys. Rev. E 49 (1994) 5046;
E. Frey and U.C. T~iuber, Phys. Rev. E 50 (1994) 1024, and the references therein.
[23] K. Moser, J. Kert6sz and D.E. Wolf, Physica A 178 (1991) 215;
J. Amar and E Family, Phys. Rev. A 45 (1992) 5318, and the references therein.
[24] M.L. Kurnaz, K.V. McCloud and J.V. Maher, Fractals 1 (1993) 1008;
M.L. Kurnaz and J.V. Maher, preprint.