a

r
X
i
v
:
m
a
t
h
-
p
h
/
0
1
0
1
0
0
5
v
1


5

J
a
n

2
0
0
1
The Abelian Sandpile Model on an
Innite Tree
Christian Maes, K.U.Leuven
Frank Redig, T.U. Eindhoven
Ellen Saada, C.N.R.S., Rouen
February 7, 2008
Abstract: We consider the standard Abelian sandpile process on the Bethe
lattice. We show the existence of the thermodynamic limit for the nite volume
stationary measures and the existence of a unique innite volume Markov process
exhibiting features of self-organized criticality
1
.
1 Introduction
Global Markov processes for spatially extended systems have been around for
about 30 years now and interacting particle systems have become a branch of
probability theory with an increasing number of connections with the natural
and human sciences. While standard techniques and general results have been
collected in a number of books such as [Liggett (1985), Chen (1992), Toom (1990)]
and are capable to treat the innite volume construction for stochastic systems
with locally interacting components, some of the most elementary questions for
long range and nonlocal dynamics have remained wide open. We have in mind
the class of stochastic interacting systems that during the last decade have in-
vaded the soft condensed matter literature and are sometimes placed under the
common denominator of self-organizing systems.
Since the appearance of the paper [BTW (1988)], the concept of self-organized
criticality (SOC) has suscited much interest, and is applied in a great variety of
domains (see e.g. [Turcotte (1999)] for an overview). From the mathematical
point of view, the situation is however quite unsatisfactory. The models ex-
hibiting SOC are in general very boundary condition dependent (especially the
BTW model in dimension 2), which suggests that the denition of an innite
volume dynamics poses a serious problem. Even the existence of a (unique)
thermodynamic limit of the nite volume stationary measure is not clear. From
1
MSC 2000: Primary-82C22; secondary-60K35.
Key-words: Sandpile dynamics, Nonlocal interactions, Interacting particle systems, Ther-
modynamic limit.
1
Abelian Sandpile Model 2
the point of view of interacting particle systems no standard theorems are at
our disposal. The innite volume processes we are looking for will be non-Feller
and cannot be constructed by monotonicity arguments as in the case of the
one-dimensional BTW model (see [MRSV (2000)]) or the long-range exclusion
process (see [Liggett (1980)]). On the other hand in order to make mathemati-
cally exact statements about SOC, it is necessary to have some kind of innite
volume limit, both for statics and for dynamics.
In this paper we continue our study of the BTW-model for the case of the
Bethe lattice, this is the abelian sandpile model on an innite tree. For this
system, many exact results were obtained in [DM (1990b)]. In contrast to the
one-dimensional case this system has a non-trivial stationary measure. We show
here that the nite volume stationary measures converge to a unique measure
which is not Dirac and exhibits all the properties of a SOC-state. We then turn
to the construction of a stationary Markov process starting from this measure
. The main diculty to overcome is the strong non-locality: Adding a grain at
some lattice site x can inuence the conguration far from x. In fact the cluster
of sites inuenced by adding at some xed site has to be thought of as a critical
percolation cluster which is almost surely nite but not of integrable size. The
process we construct is intuitively described as follows: At each site x of the
Bethe lattice we have an exponential clock which rings at rate (x). At the
ringing of the clock we add a grain at x. Depending on the addition rate (x),
we show existence of a stationary Markov process which corresponds to this
description. We also extend this stationary dynamics to initial congurations
which are typical for a measure

that is stochastically below .

The paper is organized as follows. In section 2 we introduce standard results
on nite volume abelian sandpile models, and summarize some specic results
of [DM (1990b)] for the Bethe lattice which we need for the innite volume
construction. In section 3 we present the results on the thermodynamic limit
of the nite volume stationary measures and on the existence of innite vol-
ume Markovian dynamics. Section 4 is devoted to proofs and contains some
additional remarks.
2 Finite Volume Abelian Sandpiles
In this section we collect some results on abelian sandpiles on nite graphs which
we will need later on. Most of these results are contained in the review paper
[Dhar (1999)], or in [IP (1998)].
2.1 Toppling Matrix
Let V denote a nite set of sites and
V
= (
V
x,y
)
x,yV
a matrix indexed by
the elements of V satisfying the conditions :
1. For all x, y V , x ,= y,
V
x,y
=
V
y,x
0,
2. For all x V ,
V
x,x
1,
3. For all x V ,

yV

V
x,y
0,
4.

x,yV

V
x,y
> 0.
Abelian Sandpile Model 3
Such a matrix
V
is called a toppling matrix. The fourth condition ensures that
there are sites (so called dissipative sites) for which the inequality in the third
condition is strict. This is fundamental for having a well dened toppling rule
later on. In the rest of the paper we will choose
V
to be the lattice Laplacian
with open boundary conditions on a nite simply-connected set V S, where
S is a regular graph, like the d-dimensional lattice
d
, or the innite rootless
tree I T
d
of degree d + 1. More explicitly:

V
x,x
= 2d if V
d
,
= d + 1 if V I T
d
,

V
x,y
= 1 if x and y are nearest neighbors. (2.1)
The dissipative sites then correspond to the boundary sites of V .
2.2 Congurations
A height conguration is a mapping from V to IN = 1, 2, ... assigning to
each site a natural number (x) 1 (the number of sand grains at site x). A
conguration IN
V
is called stable if, for all x V , (x)
V
x,x
. Otherwise
is unstable. We denote by
V
the set of all stable height congurations. For
IN
V
and V

V , [
V
denotes the restriction of to V

.
2.3 Toppling Rule
The toppling rule corresponding to the toppling matrix
V
is the mapping
T

V : IN
V
V IN
V
dened by
T

V (, x)(y) = (y)
V
x,y
if (x) >
V
x,x
,
= (y) otherwise. (2.2)
In words, site x topples if and only if its height is strictly larger than
V
x,x
, by
transferring
V
x,y
grains to site y ,= x and losing itself
V
x,x
grains. Toppling
rules commute on unstable congurations. This means for x, z V and such
that (x) >
V
x,x
, (z) >
V
z,z
,
T

V (T

V (, x), z) = T

V (T

V (, z), x) (2.3)
We write [T

V (, z)T

V (, x)]() = [T

V (, x)T

V (, z)]().
Choose some enumeration x
1
, , x
n
of the set V . The toppling transfor-
mation is the mapping
T

V : IN
V

V
dened by
T

V () = lim
N
_
n

i=1
T

V (, x
i
)
_
N
(). (2.4)
In [IP (1998)] it is recalled that
Abelian Sandpile Model 4
1. The limit in (2.4) exists, i.e. there are no innite cycles, due to the presence
of dissipative sites.
2. The stable conguration T

V () is independent of the chosen enumeration

of V . This is the abelian property and follows from (2.3).
2.4 Addition Operators
For IN
V
and x V , let
x
denote the conguration obtained from by
adding one grain to site x, i.e.
x
(y) = (y) +
x,y
. The addition operator
dened by
a
x,V
:
V

V
; a
x,V
= T

V (
x
) (2.5)
represents the eect of adding a grain to the stable conguration and letting
the system topple until a new stable conguration is obtained. By (2.3), the
composition of addition operators is commutative: For all
V
, x, y V ,
a
x,V
(a
y,V
) = a
y,V
(a
x,V
).
2.5 Finite Volume Dynamics
Let p denote a non-degenerate probability measure on V , i.e. numbers p
x
,
0 < p
x
< 1 with

xV
p
x
= 1. We dene a discrete time Markov chain

n
: n 0 on
V
by picking a point x V according to p at each discrete
time step and applying the addition operator a
x,V
to the conguration. This
Markov chain has the transition operator
P
V
f() =

xV
p
x
f(a
x,V
). (2.6)
We can equally dene a continuous time Markov process
t
: t 0 with
innitesimal generator
L

V
f() =

xV
(x)[f(a
x,V
) f()], (2.7)
generating a pure jump process on
V
, with addition rate (x) > 0 at site x.
2.6 Recurrent Congurations, Stationary Measure
We see here that the Markov chain
n
, n 0 has only one recurrent class and
its stationary measure is the uniform measure on that class.
Let us call
V
the set of recurrent congurations for
n
, n 0, i.e. those
for which P

(
n
= innitely often) = 1, where P

denotes the distribution of

n
, n 0 starting from
0
=
V
. In the following proposition we list
some properties of
V
. For the sake of completeness we include a proof which
we could not nd worked out completely in the literature.
Proposition 2.1
1.
V
contains only one recurrent class.
2. The composition of the addition operators a
x,V
restricted to
V
denes
an abelian group G.
Abelian Sandpile Model 5
3. [G[ = [
V
[.
4. For any x V , there exists n
x
such that for any
V
, a
nx
x,V
= .
5. [
V
[ = det
V
.
Proof: 1. We write if in the Markov chain can be reached from with
positive probability. Since sand is added with positive probability on all sites
(p
x
> 0), the maximal conguration
max
dened by

max
(x) =
V
x,x
can be reached from any other conguration. Hence, if
V
then
max
,
therefore
max

V
and
max
(see e.g. [Chung (1960)] p.19).
2. Fix
V
; then there exist n
y
1 such that

yV
a
ny
y,V
= ,
and
g
x
= a
nx1
x,V

yV,y=x
a
ny
y,V
satises (a
x,V
g
x
)() = (g
x
a
x,V
)() = . The set

x
=
V
: (a
x,V
g
x
)() =
is closed under the action of a
x,V
, contains , hence also
max
: it is a recurrent
class. By part 1,
x
=
V
, a
x,V
g
x
is the neutral element e, and g
x
= a
1
x,V
if
we restrict a
x,V
to
V
.
3. Fix
V
and put

: G
V
; g g(). As before

(G) is a
recurrent class, hence

(G) =
V
. If for g, h G,

(g) =

(h), then
gh
1
() = , and by commutativity gh
1
(g

) = g

for any g

G. Therefore
gh
1
() = for all
V
, thus g = h. This proves that

is a bijection from
G to
V
.
4. Since G is a nite group, for any x V there exists n
x
1 such that
a
nx
x,V
= e.
5. Adding
V
x,x
particles at a site x V makes the site topple, and
V
x,y
particles are transferred to y. This gives
a

V
x,x
x,V
=

y=x
a

V
x,y
y,V
.
On
V
the a
x,V
can be inverted and we obtain the closure relation

yV
a

V
x,y
y,V
= e,
which completely determines the one-dimensional representations of the group
of addition operators, and in particular the cardinality of the latter, as obtained
in [Dhar (1990a)].
Remark.
V
does not depend on the p
x
, and does not change by going from
discrete to continuous time, i.e. from (2.6) to (2.7).
The main consequence of the group property of G is the fact that the unique
stationary measure is uniform on
V
.
Abelian Sandpile Model 6
Proposition 2.2
1. The measure

V
=

RV
1
[
V
[

(2.8)
is invariant under the action of a
x,V
, x V (

is the Dirac measure on

conguration ).
2. On L
2
(
V
) the adjoint of a
x,V
is
a

x,V
= a
1
x,V
. (2.9)
Proof: Since a
x,V
:
V

V
can be inverted, we have

RV
f(a
x,V
)g() =

RV
f()g(a
1
x,V
),
hence (2.9). By choosing g 1, part 1 follows.
Remark. This shows that
V
is invariant under the Markov processes gener-
ated by (2.6) and (2.7).
2.7 Burning Algorithm
The burning algorithm determines whether a stable conguration
V
is
recurrent or not. It is described as follows: Pick
V
and erase all sites
x V satisfying the inequality
(x) >

yV,y=x
(
V
x,y
).
This means erase the set E
1
of all sites x V with a height strictly larger than
the number of neighbors of that site in V . Iterate this procedure for the new
volume V E
1
, and the new matrix
V \E1
dened by

V \E1
x,y
=
V
x,y
if x, y V E
1
= 0 otherwise,
and so on. If at the end some non-empty subset V
f
is left, satises, for all
x V
f
,
(x)

yV
f
,y=x
(
V
x,y
).
The restriction [
V
f
is called a forbidden subconguration. If V
f
is empty, the
conguration is called allowed, and the set /
V
of allowed congurations satises
Proposition 2.3
/
V
=
V
.
Abelian Sandpile Model 7
The main ingredient to prove this result (see [IP (1998)], [Speer (1993)]) is
the fact that toppling or adding cannot create a forbidden subconguration.
The set /
V
is thus closed under the dynamics and contains the maximal con-
guration
max
.
Remark that the burning algorithm implies that for V

V , and
V
,
,
V
, then any
V
such that [
V
= satises ,
V
. Indeed,
the property of having a forbidden subconguration in V
f
only depends on
the heights at sites x V
f
. Therefore
V
implies [
V

V
. This
consistency property will enable us to dene allowed congurations on innite
sets.
2.8 Expected Toppling Numbers
For x, y V and
V
, let n
V
(x, y, ) denote the number of topplings at site
y V by adding a grain at x V , i.e. the number of times we have to apply
the operator T

V (, y) to relax
x
. Dene
G
V
(x, y) =
_

V
(d) n
V
(x, y, ). (2.10)
Writing down balance between inow and outow at site y, one obtains (cf.
[Dhar (1990a)])

zV

V
x,z
G
V
(z, y) =
x,y
,
which yields
G
V
(x, y) = (
V
)
1
x,y
.
In the limit V S (where S is
d
or the innite tree), G
V
converges to the
Greens function of the simple random walk on S.
2.9 Some specic results for the tree
When V
n
is a binary tree of n generations, many explicit results have been
obtained in [DM (1990b)]. We summarize here the results we need for the
construction in innite volume.
1. When adding a grain on a particular site 0 V
n
of height 3, the set of
toppled sites is the connected cluster C
3
(0, ) of sites including 0 having
height 3. This cluster is distributed as a random animal (i.e. its distribu-
tion only depends on its cardinality, not on its form). Moreover
lim
n

Vn
([C
3
(0, )[ = k) Ck
3/2
(2.11)
as k goes to innity. The notation means that if we multiply the left
hand side of (2.11) by k
3/2
, then the limit k is some strictly positive
constant C.
2. When adding a grain on site x, the expected number of topplings at site
y satises
lim
n
_

Vn
(d) n
Vn
(x, y, ) = G(x, y), (2.12)
Abelian Sandpile Model 8
where G(x, y) is the Greens function of the simple random walk on the
innite tree, i.e.
G(0, x) = C2
|x|
, (2.13)
and [x[ is the generation number of x in the tree.
3. The correlations in the nite volume measures
V
k
can be estimated in
terms of the eigenvalues of a product of transfer matrices. This formalism
is explained in detail in [DM (1990b)], section 5: Let f, g be two local func-
tions whose dependence sets (see below a precise denition) are separated
by n generations. To estimate the truncated correlation function

V
k
(f; g) =
_
fgd
V
k

_
fd
V
k
_
gd
V
k
, (2.14)
consider the product of matrices
M
k
n
=
n

i=1
_
1 +
k,n
i
1 +
k,n
i
1 2 +
k,n
i
_
, (2.15)
where
k,n
i
[0, 1]. Let
n,k
m
(resp.
n,k
M
) denote the smallest (resp.
largest) eigenvalue of M
k
n
. Then

V
k
(f; g) C(f, g)

n,k
m

n,k
M
. (2.16)
If f and g have dependence sets deep within V
k
, then
k,n
i
is very close
to one, and the correlations are governed by the maximal and minimal
eigenvalues of M
n
=
_
2 2
1 3
_
n
.
3 Main results
3.1 Notation, denitions
From now on, S denotes the innite rootless binary tree, V S a nite subset
of S;
V
is the set of stable congurations in V , i.e.
V
= : V 1, 2, 3,
and the set of all innite volume stable congurations is = 1, 2, 3
S
. The
set is endowed with the product topology, making it into a compact metric
space. For , [
V
is its restriction to V , and for , ,
V

V
c denotes the
conguration whose restriction to V (resp. V
c
) coincides with [
V
(resp. [
V
c ).
As in the previous section,
V

V
is the set of all allowed (or recurrent)
congurations in V , and we dene
= : V S nite, [
V

V
. (3.17)
A function f : IR is local if there is a nite V S such that [
V
= [
V
implies f() = f(). The minimal (in the sense of set ordering) such V is called
dependence set of f and is denoted by D
f
. A local function can be seen as a
function on
V
for all V D
f
and every function on
V
can be seen as a local
function on . The set L of all local functions is uniformly dense in the set (()
of all continuous functions on .
All along the paper, we use the following notion of limit by inclusion for a
function f on the nite subsets of the tree with values in a metric space (K, d):
Abelian Sandpile Model 9
Denition 3.1 Let o = V S, V nite, and f : o (K, d). Then
lim
V S
f(V ) =
if for all > 0, there exists V
0
S such that for all V V
0
, d(f(V ), ) < .
Denition 3.2 A collection of probability measures
V
on
V
is a Cauchy net
if for any local f and for any > 0 there exists V
0
D
f
such that for any
V, V

V
0
[
_
f()
V
(d)
_
f()
V
(d)[ .
A Cauchy net converges to a probability measure in the following sense:
The mapping
: L IR; f (f) = lim
V S
_
fd
V
denes a continuous linear functional on L (hence on (()) which satises
(f) 0 for f 0 and (1) = 1. Thus by Riesz representation theorem
there exists a unique probability measure on such that (f) =
_
fd. We
denote
V
, and call this the innite volume limit of
V
.
We will also often consider an enumeration of the tree S, x
0
, x
1
, . . . , x
n
, . . .,
and put
T
n
= x
0
, . . . , x
n
. (3.18)
3.2 Thermodynamic limit of stationary measures
Theorem 3.1 The set dened in (3.17) is an uncountable perfect set, i.e.
1. is compact,
2. The interior of is empty,
3. For all there exists a sequence
n
,= ,
n
, converging to .
For , we denote by C
3
(0, ) the nearest neighbor connected cluster of
sites containing the origin and having height 3.
Theorem 3.2 The nite volume stationary measures
V
dened in (2.8) form
a Cauchy net. Their innite volume limit satises
1. () = 1,
2. is translation invariant and exponentially mixing,
3. ( : [C
3
(0, )[ < ) = 1,
4.
_
[C
3
(0, )[(d) = .
Remark: Point 3. above remains true for the set C
1
(0, ), the nearest neighbor
connected cluster of sites containing the origin and having height 1, and probably
also for C
2
(0, ) but this we have not been able to prove.
Abelian Sandpile Model 10
3.3 Innite volume dynamics
The nite volume addition operators a
x,V
(cf. (2.5)) can be extended to via
a
x,V
: ; a
x,V
= (a
x,V
[
V
)
V

V
c. (3.19)
Proposition 3.1
1. There exists a subset

of with (

) = 1 on which the limit

lim
V S
a
x,V
= a
x
(3.20)
exists, and a
x

.
2. The measure of theorem 3.2 is invariant under the action of a
x
.
3. For every

, a
x
(a
y
) = a
y
(a
x
), for all x, y S.
Part 2 implies that the innite volume addition operators a
x
(cf. (3.20))
dene norm 1 operators on L
p
(), for 1 p via
(a
x
f)() = f(a
x
).
We now construct a Markov process on -typical innite volume congura-
tions which can be described intuitively as follows. Let : S (0, ); this
function will be the addition rate function. To each site x S we associate
a Poisson proces N
t,x

(for dierent sites these Poisson processes are mutually

independent) with rate (x). At the event times of N
t,x

we add a grain at x,
i.e. we apply the addition operator a
x
to the conguration. Then L

V
introduced
in (2.7) generates a pure jump Markov process on . Indeed, this operator is
well-dened and bounded on any L
p
() space by Proposition 3.1, which implies
Proposition 3.2 L

V
is the L
p
() generator of the stationary Markov process
dened by
exp(tL

V
)f =
_
_

xV
a
N
t,x

x
f
_
dIP,
where IP denotes the joint distribution of the independent Poisson processes
N
t,x

, and f L
p
().
The following condition on the addition rate is crucial in our construction.
Remember [x[ is the generation number of x:
Summability Condition:

xS
(x)2
|x|
< (3.21)
This condition ensures that the number of topplings at any site x S
remains nite during the addition process.
Theorem 3.3 If satises condition (3.21), then we have
Abelian Sandpile Model 11
1. The semigroups S

V
(t) = exp(tL

V
) converge strongly in L
1
() to a semi-
group S

(t).
2. S

(t) is the L
1
() semigroup of a stationary Markov process
t
: t 0
on .
3. For any f L,
lim
t0
S

(t)f f
t
= L

f =

xS
(x)[a
x
f f],
where the limit is taken in L
1
().
Remarks.
1. In Proposition 4.1, we show that S

(t) is a strongly continuous function of

.
2. In Proposition 4.2, we show that condition (3.21) is in some sense optimal.
Theorem 3.4 The process
t
: t 0 of Theorem 3.3 admits a cadlag version
(right-continuous with left limits).
The intuitive description of the process
t
: t 0 is actually correct
under condition (3.21), i.e. the process has a representation in terms of Poisson
processes:
Theorem 3.5 If satises condition (3.21), for IP almost every (, ) the
limit
lim
V S

xV
a
N
t,x

()
x
=
t
exists. The process
t
: t 0 is a version of the process of Theorem 3.3, i.e.
its L
1
() semigroup coincides with S

(t).
Finally, we can slightly generalize Theorem 3.5 in order to dene the dy-
namics starting from a measure stochastically below . For , ,
if for all x S, (x) (x). A function f : IR is monotone if
implies f() f(). Two probability measures and satisfy if for all
monotone functions,
_
fd
_
fd.
Theorem 3.6 Let

. If satises condition (3.21), for

IP almost
every (, ) the limit
lim
V S

xV
a
N
t,x

()
x
=
t
exists. The process
t
: t 0 is Markovian with
0
distributed according to

.
Remark. The last Theorem implies that 1 can be taken as initial congu-
ration.
Abelian Sandpile Model 12
4 Proofs
This section is devoted to the proofs of the results described above. Some of
them will be put in a slightly more general framework so that they can be ap-
plied to other cases (where S is not a binary tree or where we have other addition
operators a
x
) as soon as the existence of a thermodynamic limit of the nite
volume stationary measures is guaranteed. The essential cause of diculty is
the non-locality of the addition operators. The essential simplication is the
abelian property which enables us to think of the a
x
as complex numbers of
modulus one.
4.1 Thermodynamic limit of stationary measures
Proof of Theorem 3.1:
1,2. If and , then (by the burning algorithm [
V
[
V
implies that [
V

V
). Since 2 is in (again by the burning algorithm), we
conclude that is uncountable. To see that has empty interior, notice that if
, there does not exist x, y S nearest neighbors such that (x) = (y) = 1
(that way, [
{x,y}
would be a forbidden subconguration). Finally is closed
as intersection of closed sets.
3. Let
max
be the maximal conguration,
max
(x) = 3 for all x S. If
[
V

V
, then
V
(
max
)
V
c . Therefore any containing an in-
nite number of sites x for which (x) ,= 3 has property 3 of Theorem 3.1. If
contains only a nite number of sites having height 1 or 2, then we
choose a sequence = x
n
: n IN x S : (x) = 3 and (y) =
3 for any neighbor of x such that two elements of are never nearest neigh-
bors, and [x
n
[ is strictly increasing in n. We then dene
n
(x) = (x) for
x S x
k
: 0 k n and
n
(x
k
) = 2 for x
k
, k n + 1. These
n
belong to by the burning algorithm, and
n
.
Proof of Theorem 3.2:
We use T
n
introduced in (3.18), but with the x
i
such that n m implies that
the generation numbers satisfy [x
n
[ [x
m
[. Then we have
[x
n
[ log
2
n. (4.22)
To prove that the probability measures
V
form a Cauchy net, it is sucient to
show that for any local function f : IR we have

n
[
_
fd
Tn
fd
Tn+1
[ < . (4.23)
We do it for f() = (x
0
) (a general local function can be treated in the same
way), by giving an upper bound of the dierence
_
fd
Tn

_
fd
Tn+1
by a
truncated correlation function (cf. (2.14)). Then we estimate the latter by the
transfer matrix method (cf. Section 2.9, part 3). We abbreviate in what follows

n
=
Tn
,
n
=
Tn
.
Abelian Sandpile Model 13
Lemma 4.1
[
n+1
[(x
0
)]
n
[(x
0
)][ C
n+1
[(x
0
); I((x
n+1
) = 3)]
Proof: By the burning algorithm, every
n
can be extended to an element
of
n+1
by putting (x
n+1
) = 3. Moreover
[
Tn
:
n+1
, (x
n+1
) = 3 =
n
,
thus

n+1
[(x
n+1
) = 3] =
[
n
[
[
n+1
[
, (4.24)
which yields

n
((x
0
)) =

Rn
1
[
n
[
(x
0
)
=

Rn+1
1
[
n+1
[
(x
0
)I((x
n+1
= 3)
[
n+1
[
[
n
[
=

Rn+1
1
[
n+1
[
(x
0
)I((x
n+1
) = 3)
1

n+1
[(x
n+1
) = 3]
.
Therefore
[
n+1
((x
0
))
n
((x
0
))[

n+1
[(x
0
); I((x
n+1
) = 3)]

n+1
[(x
n
+ 1) = 3]
The Lemma follows now from (4.24), and the fact that [
n
[ grows like e
cn
for
some c log 2.
Recalling Section 2.9, part 3, we have

n
[(x
0
); I((x
n
) = 3)] C

|xn|,|xn|
m

|xn|,|xn|
M
(4.25)
Lemma 4.2
+

n=1

|xn|,|xn|
m

|xn|,|xn|
M
< +
Proof: We abbreviate
(n)
m
=
|xn|,|xn|
m
,
(n)
M
=
|xn|,|xn|
M
, M(n) = M
n
|xn|
,
i
=

|xn|,|xn|
i
. Remember 0
|xn|,|xn|
i
1 is close to one for i n and n large. In
terms of the trace and the determinant of M(n) we have

(n)
M
=
1
2
_
Tr(M(n)) +
_
[Tr(M(n))]
2
4det(M(n))
_

(n)
m
=
1
2
_
Tr(M(n))
_
[Tr(M(n))]
2
4det(M(n))
_
.
Therefore,
lim
n
_

(n)
m

(n)
M
_
_
[Tr(M(n))]
2
det(M(n))
_
= 1.
Abelian Sandpile Model 14
To prove the Lemma we show that (cf. (4.22))
_
det(M(n))
[Tr(M(n))]
2
_
(
4
9
)
|xn|
. (4.26)
Use
det(M(n)) =
|xn|

i=1
(1 +
i
)
2
,
and
Tr(M(n)) Tr
_
_
|xn|

i=1
_
1 +
i
0
0 2 +
i
_
_
_
=
|xn|

i=1
(1 +
i
) +
|xn|

i=1
(2 +
i
),
to estimate (for 1 i [x
n
[, 2(2 +
i
) 3(1 +
i
))
_
det(M(n))
[Tr(M(n))]
2
_

_
_
1 + 2
|xn|

i=1
_
2 +
i
1 +
i
_
+
|xn|

i=1
_
2 +
i
1 +
i
_
2
_
_
1
(1 + 2.(3/2)
|xn|
+ (3/2)
2|xn|
)
1
(4/9)
|xn|
.
For a general local function f, we have to replace [x
n
[ by [x
n
[ N
0
, where
N
0
is the number of generations involved in the dependence set of f. Since f is
local, N
0
is nite, hence the convergence in (4.23) is unaected.
4.2 Innite volume toppling operators
Denition 4.1 Given the nite volume addition operators a
x,V
(dened in
(3.19)) acting on , we call a conguration normal if for every x S
there exists a minimal nite set V
x
() S such that for all V

V V
x
()
a
x,V
= a
x,V
.
In other words, for a normal , outside V
x
(), no sites are aected by adding
a grain at x. In our case, when a particle is added at some site x S, the
cluster of toppled sites coincides with the cluster C
3
(x, ) of sites having height
3 including x, thus
V
x
() = C
3
(x, )
e
C
3
(x, ), (4.27)
where
e
denotes the exterior boundary. Notice that for a normal conguration
, by denition,
a
x
() = lim
V S
a
x,V
() = a
x,Vx()
() (4.28)
Abelian Sandpile Model 15
Proof of Proposition 3.1:
1. We show that there is a full measure set

of normal congurations. From

(2.11) and Theorem 3.2,
_
I([C
3
(x, )[ = n)d Cn
3/2
.
Therefore concentrates on the set

of congurations for which all the clusters

C
3
(x, ) are nite, hence for which is normal. Moreover this set

is closed
under the action of the addition operators a
y
, since (cf. (4.27))
C
3
(x, a
y
) V
x
() V
y
(). (4.29)
2. Choose > 0, pick a local function f, x V
n
S and n
0
such that n n
0
implies
: V
x
() , V
n


4|f|

+ 1
. (4.30)
This n
0
exists since concentrates on normal congurations. We estimate

_
f(a
x
)d f()d

_
f(a
x,Vn
)d
_
f()d

+ 2|f|

: V
x
() , V
n

lim
m

_
f(a
x,Vn
)d
Vm

_
f()d
Vm

+

2

2
+ 2|f|

lim
m

Vm
(a
x,Vn
() ,= a
x,Vm
())
=

2
+ 2|f|

_
1 lim
m

Vm
(V
x
() V
n
)
_
=

2
+ 2|f|

(1 (V
x
() V
n
)) .
In the last step we used that the indicator I(V
x
() V
n
) is a local function.
3. Let

, x, y S be two dierent sites and V V

x
() V
x
(a
x
)
V
x
(a
y
). Since a
x,V
and a
y,V
commute, we have
a
x
(a
y
) = a
x
(a
y,V
) = a
x,V
(a
y,V
)
= a
y,V
(a
x,V
) = a
y,V
(a
x
) = a
y
(a
x
).
4.3 Innite volume semigroup
We now turn to the proofs of Theorems 3.3 and 3.4.
Denition 4.2 We dene the cluster of at x S as
((x, ) = y S : a
y
(x) ,= (x), (4.31)
and put
G

(x, y) =
_
I(y ((x, ))d(). (4.32)
Abelian Sandpile Model 16
Finally for : S [0, ), write
|f|

xS
(x)
_
(d)[f(a
x
) f()[,
B

= f : IR : f bounded, |f|

< .
Lemma 4.3 If

xS
(x)G

(y, x) < for all y S, (4.33)

then all local functions are in B

.
Proof: Let f be a local function with dependence set D
f
. Then f(a
x
) ,= f()
if for y D
f
, a
x
(y) ,= (y), i.e. x ((y, ):
|f|

xS
(x)
_
[a
x
f f[d
=
_

xyD
f
C(y,)
(x)[a
x
f f[d
2|f|

xS
(x)
_
I(x
yD
f
((y, ))d
2|f|

yD
f

xS
G

(y, x)(x) < .

The next lemma provides a link between G

and the Greens function for

simple random walk on S, i.e., between conditions (4.33) and (3.21).
Lemma 4.4
G

(x, y)

zx
G(y, z) =
x,y
+ 3G(x, y),
where z x means that z and x are neighbors.
Proof: We have to estimate the probability that a
x
(y) ,= (y). If by adding
a grain at x we inuence y, this can only be achieved by the toppling of one of
the nearest neighbor sites of y. Since concentrates on normal congurations,
(a
x
(y) ,= (y)) = lim
V S
(a
x
(y) ,= (y), V
x
() V
y
() V )
= lim
V S
lim
WS

W
(a
x,V
(y) ,= (y), V
x
() V
y
() V )
= lim
V S
lim
WS

W
(a
x,W
(y) ,= (y), V
x
() V
y
() V )
lim
WS

W
(a
x,W
(y) ,= (y))
lim
WS

W
( z W, z y, n
W
(z, y, ) 1)
lim
WS

zy
_
d
W
()n
W
(z, y, )
=

zy
G(z, y), (4.34)
Abelian Sandpile Model 17
where we used (2.10),(2.12), (4.27) and (4.29).
The following Lemma nishes the proof of Theorem 3.3 and shows that B

is
a natural core for the domain of the generator of the innite volume semigroup.
Lemma 4.5
1. For f B

the net
S

V
(t)f = exp(tL

V
)f = exp
_
t

xV
(x)(a
x
I)
_
f (4.35)
converges in L
1
() (as V S) to a function S

(t)f L
1
(). f S

(t)f
denes a semigroup on B

which is a contraction in both L

1
() and B

norms.
2. Under condition (4.33), the semigroup S

(t) corresponds to a unique Markov

process on .
Proof: We denote by |f| the L
1
()-norm of f, and we abbreviate S
V
(t) =
S

V
(t), S(t) = S

(t), L
V
= L

V
.
1. First note that S
V
(t) is well-dened on L
1
() by Proposition 3.2. By the
abelian property (Proposition 3.1, part 3) we can write for V V

S:
|S
V
(t)f S
V
(t)f| = |(S
V

\V
(t) I)S
V
(t)f|
By Proposition 3.2, S
V
(t) is the semigroup of a stationary Markov process and
hence a contraction on L
1
(). Therefore
|S
V
(t)(S
V

\V
(t) I)f| |(S
V

\V
(t) I)f|
= |
_
t
0
L
V

\V
S
V

\V
(s)fds|

_
t
0
|L
V

\V
f|ds
t

xV

\V
(x)
_
[(a
x
I)f[d 0 as V, V

S, (4.36)
where the last step follows from f B

. Hence S
V
(t)f S(t)f in L
1
(). We
show that S(t)f B

xS
(x)
_
[S(t)f(a
x
) S(t)f()[(d)

xS
(x)
_
S(t)[a
x
f f[d
=
_

xS
(x)[a
x
f f[d = |f|

.
Thus S(t) is also a contraction for the | |

-norm. We nish with the semigroup

property:
S(t)S(s)f = lim
V S
S
V
(t)[S(s)f]
= lim
V S
lim
WS
S
V
(t)S
W
(t)f,
Abelian Sandpile Model 18
and
S(t + s)f = lim
V S
S
V
(t)S
V
(s)f.
Then, since S
V
(t) is a contraction in L
1
(),
|S
V
(t)S
W
(t)f S
V
(t)S
V
(s)f| |S
W
(s)f S
V
(s)f|, (4.37)
By (4.36), the right hand side of (4.37) goes to zero as V, W S.
2. If condition (4.33) is met, then B

contains all local functions by Lemma

4.3. Therefore, by contractivity the semigroup S(t) on B

uniquely extends
to a semigroup of contractions on L
1
(). Since by Proposition 3.2, S
V
(t) is a
Markov semigroup, so is S(t), i.e. S(t)1 = 1, S(t)f 0 if f 0. Hence, by
Kolmogorovs theorem there is a unique Markov process with semigroup S(t).
Remark. When 1, condition (4.33) is equivalent to

xS
_
(d)I(x ((y, )) =
_
[((y, )[(d) < +,
i.e., the clusters must be integrable under . For models which exhibit self-
organized criticality, ((y, ) is usually a nite but critical percolation cluster,
implying that
_
[((y, )[d = (cf. Theorem 3.2 part 4, because ((y, )

e
C
3
(y, )). Therefore this formalism breaks down for addition rate 1.
The following Lemma proves Theorem 3.4.
Lemma 4.6 Under condition (4.33), the process
t
: t 0 of Theorem 3.3
is almost surely right-continuous, i.e.
IP

_
lim
t0
d(
t
,
0
)
_
= 0, (4.38)
where IP

is its path-space measure, and the distance d is dened below (in

(4.41)).
Proof: Pick a function : S (0, 1) such that

xS
(x) = 1, (4.39)
and

x,yS
(x)G

(x, y)(y) < (4.40)

The distance
d(, ) =

xS
[(x) (x)[(x) (4.41)
generates the product topology. Denote by IE

the expectation w.r.t. IP

. For
f
y
() = (y),
f
y
(
t
) f
y
(
0
) =
_
t
0
L

f
y
(
s
)ds + M
y
t
,
where M
y
t
is a centered martingale with quadratic variation
IE

_
(M
y
t
)
2

= IE

__
t
0
(L

f
2
y
(
s
) 2f
y
(
s
)L

f
y
(
s
))ds
_
. (4.42)
Abelian Sandpile Model 19
Using stationarity of
s
and
_
d[L

g[ 2|g|

xS

yDg
(x)G

(y, x),
for a local bounded function on , we obtain from (4.42)
IE

_
(M
y
t
)
2

Ct

xS
(x)G

(y, x).
Now we can estimate
IP

_
_
s t :

yS
[
s
(y)
0
(y)[(y)
_
_
IP

_
_
_
t
0
ds

yS
[L

f
y
(
s
)[(y) /2
_
_
+ IP

_
_
sup
0st

yS
M
y
s
(y)

/2
_
_
(12t/)

x,yS
(x)G

(y, x)(y) + (2/)

2
IE

yS
M
y
t
(y)

2
(12t/)

x,yS
(x)G

(y, x)(y) + (2/)

2
IE

_
_

yS
(M
y
t
)
2
(y)
_
_
tC

x,yS
(x)G

(y, x)(y).
Here we used Markovs and Doobs inequalities in the second step and the
Cauchy-Schwarz inequality combined with (4.39) in the third step. The result
(4.38) follows.
4.4 Poisson representation
In this section we prove Theorems 3.5 and 3.6. Intuitively it is clear from the
abelian property that the process of which we showed existence in the previ-
ous subsection can be represented as

xS
a
N
t,x

x
, where N
t,x

are independent
Poisson processes of intensity (x).
We take T
n
as in (3.18). We say that the product

xS
a
nx
x
exists if for
every y S there exists N
y
such that for all m, n N
y

_

xTn
a
nx
x

_
(y)
_

xTm
a
nx
x

_
(y)

= 0.
This is equivalent to the convergence of the sequence

xTn
a
nx
x
in the product
topology.
Lemma 4.7 Under condition (4.33), the product

xS
a
N
t,x

x
=
t
Abelian Sandpile Model 20
exists for -almost every realization of N
t,x

and almost every . The process

t
: t 0 is a version of the Markov process of Lemma 4.5.
Proof: Choose a realization of N
t,x

such that

xS
N
t,x

(x, y) < (4.43)

for every y. This happens with probability one by condition (4.33). Dene for

Tn
(t) =

xTn
a
N
t,x

x
. (4.44)
Under ,
Tn
(t) is stationary in n and t. We have

(
Tn
(t))(y) (
Tn+1
(t))(y)

_
a
N
t,x
n+1

xn+1

Tn
(t)
_
(y) (
Tn
(t))(y)

(d)
=
_

_
a
N
t,x
n+1

xn+1

_
(y) (y)

(d)

_ N
t,x
n+1

j=1

_
a
j
xn+1

_
(y)
_
a
j1
xn+1

_
(y)

(d)
6N
t,xn+1

(x
n+1
, y).
In the second and last steps we used the invariance of under a
x
. By the
Borel Cantelli Lemma, (4.43) implies that for almost every realization of N
t,x

_
n
0
: n n
0
(
Tn
(t))(y) = (
Tn
0
(t))(y)

= 1.
This proves -a.s. convergence of the product. To see that
t
is a version of the
Markov process with semigroup S

(t), combine Proposition 3.2 with Theorem

3.3, part 1 to get, for any local function f,
_
d

_
dIPf(
t
) S

(t)f

= 0.
In the preceding argument we used a particular enumeration of the countable
set S. But changing it gives again a process with semigroup S

(t). Therefore
the limiting process will not depend (up to sets of measure zero) on the chosen
enumeration of S.
Proof of Theorem 3.6:
For

and y S we have the relation (remember (4.44))

V
(t)(y) = (y) + I(y V )

xV
N
t,x

n
t
V
(y), (4.45)
where n
t
V
(x), an integer valued random variable, is the number of topplings at
site x in the time interval [0, t], when sand is added in V . For T
n
dened in (3.18)
we will rst prove that n
t
Tn
increases IP almost surely to an L
1
(IP) random
Abelian Sandpile Model 21
variable n
t
, interpreted as the number of topplings in [0, t] when we add grains
according to N
t,x

. By the abelian property the sequence n

t
T
k
(0) is increasing in
k. The following estimate is similar to (4.34)
( IP) ( [n
t
T
k
(0) n
t
T
k+1
(0)[
_
= ( IP)
_
n
t
x
k+1
(0)
_

_
n
t
x
k+1
(0)(d) IP(d)

lim
V S
_
n
t
x
k+1
(0)I(V
x
k+1
() V
0
() V )(d) IP(d)

lim
WS
_
n
t
W
(x
k+1
, 0, )
W
(d) IP(d)

t(x
k+1
)G(0, x
k+1
). (4.46)
In the fourth line, n
t
W
(x
k+1
, 0, ) denotes the number of topplings up to time t
at site 0 W by adding grains at site x
k+1
W. By the Borel Cantelli Lemma,
condition (3.21) implies the a.s. convergence of n
t
T
k
(0), and analogously of every
n
t
T
k
(x). Pick (, ) such that n
t
T
k
(, ) converges, i.e. such that sup
k
n
t
T
k
(, ) =
n
t
(, ) is nite (indeed, n
t
T
k
(, ) is an integer). If

, then n
t
T
k
(

, )
n
t
T
k
(, ) because we can obtain from

by adding sand at sites x S for which

(x) < (x) thereby increasing the number of topplings. We thus conclude that
the convergence of n
t
T
k
(, ) implies the convergence of n
t
T
k
(

, ) for all

.
Now let

in the FKG sense. There is a coupling IP

12
of

IP and
IP such that
IP
12
(((
1
,
1
), (
2
,
2
)) :
1
=
2
,
1

2
) = 1,
i.e. we use the same Poisson events and couple

and according to the optimal

coupling (see [Strassen (1965)]). Then
(

IP)
_
n
t
T
k
(, ) n
t
(, )
_
= IP
12
_
n
t
T
k
(
1
,
1
) n
t
(
1
,
1
)
_
= IP
12
_
n
t
T
k
(
1
,
1
) n
t
(
1
,
1
), n
t
T
k
(
2
,
2
) n
t
(
2
,
2
),
1
=
2
,
1

2
_
IP
12
_
n
t
T
k
(
2
,
2
) n
t
(
2
,
2
)
_
= ( IP)
_
n
t
T
k
(, ) n
t
(, )
_
= 1.
This shows the

IP-almost sure convergence of n

t
T
k
, hence by (4.45) the prod-
uct

xS
a
N
t,x

()
x

converges

IP almost surely.
As a further result we show that the semigroup S

(t) is continuous as a
function of the addition rate . We dene

1
= : S [0, ) : || =

xS
(x)G(0, x) < .
It is a complete metric space (as a closed subset of a Banach space) with the
property: If
n

1
,
n
(pointwise), and
1
then
n
in
1
.
Proposition 4.1 The semigroup S

(t) of Theorem 3.3 is a strongly continuous

function of , i.e. if
n
in
1
, then for any local function f, S
n
(t)f
S

(t)f.
Abelian Sandpile Model 22
Proof: Let n
t

= lim
k
n
t
T
k
be the number of topplings in [0, t] from sand
addition at rate . In the proof of Theorem 3.5 we have shown that this random
variable is IP almost surely well dened and, after taking limits in (4.45),
satises

t
=
0
+ N
t

n
t

, (4.47)
where N
t

= lim
V S

xV
N
t,x

. Note that if
1

2
, the Poisson processes
N
t
1
and N
t
2
can be coupled in such a way that for all x S, N
t,x
1
N
t,x
2
, and
hence, by the abelian property, n
t
1
(x) n
t
2
(x). Consider a coupling of the
four Poisson processes N
t

, N
t
n
, N
t
n
and N
t
n
under which the inequalities
X
1
(t) X
2
(t), X
2
(t) X
3
(t), X
3
(t) X
4
(t), are satised with probability one.
Let

IP denote the law of the marginal (X
1
, X
4
). We have, by a reasoning similar
to (4.46),
_
d
_

IE[n
t
n
(0) n
t

(0)[
_

_
d
_

IE
_
n
t
n
(0) n
t
n
(0)
_
+

IE
_
n
t
n
(0) n
t
n
(0)
_
+

IE
_
n
t
n
(0) n
t

(0)
_
_
t

xS
[
n
(x) (x)[G(0, x), (4.48)
which tends to zero for
n
in
1
. Take now a local function f, and denote

D
f
= D
f

e
D
f
,
[S
n
(t)(f) S

(t)(f)[

IP
_
n
t
n
(x) ,= n
t

(x) for some x

D
f
_

x

D
f

IE[n
t

(x) n
t
n
(x)[.
Combining this with (4.48) concludes the proof.
One might ask whether we can go beyond condition (3.21), which essentially
guarantees that the expected number of topplings stays nite in the addition
process. In the following proposition we show that it is impossible to keep
integrable toppling numbers and rate 1 addition. The relation (4.49) should
be regarded as the innitesimal version of (4.47), where (x) replaces the rate
(x). We then show that has to depend on x.
Proposition 4.2 Let : S 0, 1 be a stationary and ergodic random eld
distributed according to . Denote by
_
(0)(d) = its density. Suppose
there exists a measurable transformation T : 0, 1
S
which satises the
conditions
1. The measure of Theorem 3.2 is invariant under T(, ) for any .
2.
T(, )(x) = (x) + (x) n(, , x), (4.49)
with n(, , .) L
1
() for almost every .
Then, = 0.
Abelian Sandpile Model 23
Proof: Taking expectation over in (4.49) gives
(, x) = (x), (4.50)
where (, x) =
_
n(, , x)(d). By stationarity of and , (, x) is a
stationary random eld. Let (x
t
: t 0) denote continuous time simple random
walk on S, starting at 0. From (4.50),
IE(, x
t
) = (, 0) + IE
_
t
0
(x
s
)ds.
Divide this last line by t and let t +. As is ergodic (making the last term
equal to ) and as the process (, x
t
) is stationary, we conclude that = 0.
References
[BTW (1988)] Bak, P., Tang, K. and Wiesenfeld, K., Self-Organized Criticality,
Phys. Rev. A 38, 364374 (1988).
[Chung (1960)] Chung, K.L., Markov Chains with Stationary Transition Prob-
abilities, Springer-Verlag, Berlin-Gottingen, Heidelberg, 1960.
[DR (1989)] Dhar, D. and Ramaswamy, R., Exactly Solved Model of Self-
Organized Critical Phenomena, Phys. Rev. Lett. 63, 16591662 (1989).
[Dhar (1990a)] Dhar, D., Self Organised Critical State of Sandpile Automaton
Models, Phys. Rev. Lett. 64, No.14, 16131616 (1990).
[DM (1990b)] Dhar, D. and Majumdar, S.N., Abelian Sandpile Models on the
Bethe Lattice, J. Phys. A 23, 43334350 (1990).
[Dhar (1999)] Dhar, D., The Abelian Sandpiles and Related Models, Physica A
263, 425 (1999).
[IP (1998)] E.V. Ivashkevich, Priezzhev, V.B., Introduction to the sandpile
model, Physica A 254, 97116 (1998).
[Liggett (1980)] Liggett, T.M., The Long Range Exclusion Process, Ann. Prob.
8, 861889 (1980).
[Liggett (1985)] Liggett, T.M., Interacting Particle Systems, Springer, Berlin,
Heidelberg, New York, 1985.
[MRSV (2000)] Maes, C., Redig, F., Saada E. and Van Moaert, A., On the
thermodynamic limit for a one-dimensional sandpile process, Markov Proc.
Rel. Fields, 6, 122 (2000).
[Chen (1992)] Mu Fa Chen, From Markov Chains to Non-Equilibrium Particle
Systems, World Scientic, 1992.
[Priezzhev (1994)] Priezzhev, V.B., Structure of Two Dimensional Sandpile.
I.Height Probabilities, J. Stat. Phys. 74, 955979 (1994).
Abelian Sandpile Model 24
[Speer (1993)] Speer, E., Asymmetric Abelian Sandpile Models, J. Stat. Phys.
71, 6174 (1993).
[Strassen (1965)] Strassen, V., The existence of probability measures with given
marginals, Ann. Math. Statist. 36, 423439 (1965).
[Toom (1990)] Toom, A.L., Vasilyev, N.B., Stavskaya, O.N., Mityushin, L.G.,
Kurdyumov, G.L., Pirogov, S.A., Discrete local Markov Systems. In: Do-
brushin, R.L., Kryukov V.I., Toom, A.L. (eds.), Stochastic cellular systems:
ergodicity, memory, morphogenesis, Manchester University Press, 1182,
1990.
[Turcotte (1999)] Turcotte, D.L., Self-Organized Criticality, Rep. Prog. Phys.
62, 13771429 (1999).
Adresses:
C.M.: Instituut voor Theoretische Fysica, K.U.Leuven, Celestijnenlaan 200D,
B-3001 Leuven, Belgium - email: christian.maes@fys.kuleuven.ac.be
F.R.: On leave from Instituut voor Theoretische Fysica, K.U. Leuven, Celestij-
nenlaan 200D, B-3001 Leuven, Belgium - email: f.h.j.redig@tue.nl
E.S.: CNRS, UMR 6085, Universite de Rouen, 76821 Mont-Saint-Aignan cedex,
France. - email: Ellen.Saada@univ-rouen.fr