# The Abelian Sandpile Model on a Random Binary Tree

- 590 Downloads
- 1 Citations

## Abstract

We study the abelian sandpile model on a random binary tree. Using a transfer matrix approach introduced by Dhar and Majumdar, we prove exponential decay of correlations, and in a small supercritical region (i.e., where the branching process survives with positive probability) exponential decay of avalanche sizes. This shows a phase transition phenomenon between exponential decay and power law decay of avalanche sizes. Our main technical tools are: (1) A recursion for the ratio between the numbers of weakly and strongly allowed configurations which is proved to have a well-defined stochastic solution; (2) quenched and annealed estimates of the eigenvalues of a product of *n* random transfer matrices.

## Keywords

Sandpile models Random binary trees Phase transition## 1 Introduction

The abelian sandpile model (ASM) is a thoroughly studied model both in the physics and in the mathematics literature see e.g. [3, 4, 9, 12, 15] for recent review papers on the subject. In physics, it serves as a paradigmatic model of self-organized criticality (SOC). SOC is usually referred to as the phenomenon that the model exhibits power law decay of correlations or avalanche sizes, without fine-tuning any external parameters such as temperature or magnetic field. In mathematics, the ASM is connected to several combinatorial objects such as spanning trees, graph-orientations, dimers, and it has an interesting abelian group structure.

The ASM has been studied on the Bethe lattice (i.e., the rootless binary tree) in [5]. Via a recursive analysis, based on a transfer matrix method, the authors in [5] arrive at exact expressions of various quantities of interest, such as the single height distribution, correlation functions of height variables, and avalanche size distribution.

There are various motivations to consider the ASM on *random graphs*. As an example, we mention integrate-and-fire models in neuroscience (see e.g. [10] and related papers), where the connections between neurons are updated after a neuron has fired. The typical connection structure of a network of firing neurons is therefore generically not translation invariant, and time dependent. As a first approximation, one can quench the randomness of the connection graph and study the firing of neurons on the derived random graph. So far, the ASM model has been studied on small world graphs from the physicist’s perspective, using a renormalization group approach [8]. In the mathematics literature, there are recent studies on so-called “cactus” graphs [7, 13].

In this paper, we start this study of the ASM on random graphs with the ASM on a random tree, for the sake of simplicity chosen to be a realization of a binary branching process with branching probability *p*∈[0,1]. We use the transfer-matrix method of [5] to express relevant quantities such as the correlation of height variables and the avalanche size distribution in terms of the eigenvalues of an ad hoc product of *random* matrices. This is the fundamental difference between the Bethe lattice case and the random tree, namely the fact that the transfer matrices depend randomly on the vertices and instead of having to deal with the *n*-th power of a simple two by two matrix, one has to control the product of *n* random matrices.

The crucial quantity entering the transfer matrices is the so-called characteristic ratio, which is the ratio between the numbers of weakly and strongly allowed configurations. This ratio is equal to 1 in the infinite Bethe lattice for every vertex and it is close to 1 for vertices belonging to a finite subset of the Bethe lattice which are far away from the “boundary” (see later on for precise statements). In our case, we show that for an infinite random tree the characteristic ratio is a well-defined random variable, uniquely determined by a stochastic recursion. The transfer matrices will then contain elements with that distribution. We also consider deterministic trees that are strict subsets of the binary tree where the characteristic ratio can be computed explicitly. Next, we prove the exponential decay of correlation of height variables (as in the Bethe lattice case), and show that for a branching probability *p* sufficiently small, but still supercritical (*p*>1/2), i.e., the branching process survives with positive probability, avalanche sizes decay exponentially. This shows a transition between exponential decay of avalanche sizes, for *p* small, and power law decay for *p* close to (possibly only equal to) one.

Our paper is organized as follows. First we recall some basic material about the ASM on trees and the recursive technique developed in [5]. Second we study the recursion for the characteristic ratio and show it has a unique solution for the random binary tree. Finally we give quenched and annealed estimates of the eigenvalues of the product of *n* random matrices, which we apply in the study of correlation of height variables and avalanche sizes.

## 2 Abelian Sandpile Model on Subtrees of the Full Binary Tree

We summarize here the basic and standard objects of the abelian sandpile model on a (general) tree. More details can be found e.g. in [5, 11].

### 2.1 Rooted and Unrooted Random Trees

We denote by Open image in new window the rooted binary tree of *n* generations, and by Open image in new window the rooted infinite binary tree. For a more general tree Open image in new window we write Open image in new window if we want to indicate that the tree has root *i*. The rootless infinite binary tree or Bethe lattice is then obtained by joining two infinite rooted binary trees by a single edge connecting their roots.

A random binary tree of *N* generations with branching probability *p*∈[0,1] is a random subset Open image in new window of Open image in new window obtained as follows. Starting from the root, we add two new vertices, each connected with a single edge to the root (resp. no vertices), with probability *p* (resp. (1−*p*)), and we iterate this from every new vertex independently for *N* generations. By letting *N*→∞ we obtain the full binary branching process. Joining two independent infinite copies of this process by a single edge connecting their roots creates the rootless random binary tree. This last procedure is of course identical to create the non-random rootless binary tree from non-random rooted binary trees.

### 2.2 Height Configurations and Legal Topplings

For Open image in new window a finite subtree of the Bethe lattice, height configurations on Open image in new window are elements Open image in new window . For Open image in new window and Open image in new window , *η* _{ u } denotes the height at vertex *u*. A height configuration Open image in new window is *stable* if *η* _{ u }∈{1,2,3} for all Open image in new window . Stable configurations are collected in the set Open image in new window .

*toppling operator*

*T*

_{ u }via

*toppling matrix*, indexed by vertices Open image in new window and defined by (

*u*,

*v*neighbors in Open image in new window means that an edge of Open image in new window connects

*u*to

*v*). In words, in a toppling at

*u*, 3 grains are removed from

*u*, and every neighbor of

*u*receives one grain.

A toppling at Open image in new window is called *legal* if *η* _{ u }>3. A sequence of legal topplings is a composition \(T_{u_{n}}\circ\cdots \circ T_{u_{1}} (\eta)\) such that for all *k*=1,…,*n* the toppling at *u* _{ k } is legal in \(T_{u_{k-1}}\circ\cdots\circ T_{u_{1}} (\eta)\). The *stabilization* of a configuration Open image in new window is defined as the unique stable configuration Open image in new window that arises from *η* by a sequence of legal topplings.

### 2.3 Addition Operator and Markovian Dynamics

*addition operator*is the map Open image in new window defined via where Open image in new window is such that

*δ*

_{ u }(

*u*)=1 and

*δ*

_{ u }(

*z*)=0 for Open image in new window . In other words,

*a*

_{ u }

*η*is the effect of an addition of a single grain at

*u*in

*η*, followed by stabilization.

The addition operators commute, i.e., *a* _{ u } *a* _{ v }=*a* _{ v } *a* _{ u }. This is the well-known and crucial *abelian property* of the sandpile model.

*η*(

*n*),

*n*∈ℕ} on Open image in new window defined via

*X*

_{ i }are i.i.d. uniformly chosen vertices of Open image in new window .

Given a stable height configuration *η* and Open image in new window , we define the *avalanche* *Av*(*u*,*η*) induced by addition at *u* in *η* to be the set of vertices in Open image in new window that have to be toppled in the course of the stabilization of *η*+*δ* _{ u }.

### 2.4 Recurrent Configurations and Stationary Measure

The *recurrent* configurations of the sandpile model form a subset of the stable configurations defined as follows. A configuration Open image in new window contains a *forbidden subconfiguration* (FSC) if there exists a subset Open image in new window such that for all *u*∈*S*, the height in *u* is less than or equal to the number of neighbors of *u* in *S*. The restriction of *η* to *S* is then called a FSC. A configuration is *allowed* if and only if it does not contain a FSC. Recurrent configurations coincide with allowed ones, and are collected in the set Open image in new window .

*δ*

_{ η }is the point mass concentrated on the configuration

*η*.

### 2.5 Specific Properties of the Sandpile Model on a Tree

In this section, for the sake of self-containdness, we briefly summarize some basic facts from the paper [5] which we need later on. In a subset of the Bethe lattice, the *distance between two vertices* is defined as the length (i.e., the number of edges) of the shortest path joining them. A vertex is a *surface vertex* if it has a number of neighbors strictly less than 3.

#### 2.5.1 Weakly and Strongly Allowed Subconfigurations

**weakly**and

**strongly**allowed ones. Let

*T*be a rooted finite tree with root

*u*and extend it with one vertex

*v*and one edge 〈

*u*,

*v*〉. Consider an allowed configuration

*ξ*on

*T*. We put height 1 at vertex

*v*and investigate the derived configuration

*ξ*′ on

*T*∪{

*v*} such that

*ξ*is the restriction of

*ξ*′ to

*T*, and \(\xi'_{v}=1\). If

*ξ*′ has no FSC, then we call

*ξ*

*strongly*allowed on

*T*, otherwise

*weakly*allowed. We give in Fig. 1 an example of weakly and strongly allowed configurations.On one hand, if

*η*

_{ u }=2, then there exists a forbidden subconfiguration on

*S*={

*v*,

*u*,

*i*

_{1}}. On the other hand, if

*η*

_{ u }=3, there are no forbidden subconfigurations.

#### 2.5.2 Characteristic Ratio and Recursion

Let Open image in new window be a finite tree, rooted or not. A key quantity in the analysis of [5] is *the characteristic ratio* Open image in new window of Open image in new window between weakly and strongly allowed configurations. For the empty tree Open image in new window we put *x*(∅)=0.

For an infinite tree Open image in new window we say that the characteristic ratio is well-defined if the limit Open image in new window exists, where Open image in new window is taken along the net of finite subtrees. That is, Open image in new window means that for every *ε*>0, there exists Open image in new window a finite subtree of Open image in new window such that, for all finite subtrees Open image in new window , we have Open image in new window .

*T*

^{1},

*T*

^{2}are two non-intersecting subtrees of the tree

*T*defined as follows. If

*T*=

*T*

^{ o }is a rooted tree with root

*o*, then

*T*

^{1},

*T*

^{2}are obtained by deleting the root

*o*and splitting the tree into two subtrees whose roots are the two descendants of the root

*o*, see Fig. 2.If

*T*is an unrooted tree, we pick one of its edges, 〈

*u*,

*v*〉, we delete this edge and split

*T*into two rooted subtrees

*T*

^{1}and

*T*

^{2}whose roots are

*u*and

*v*, see Fig. 3.This recursion (4) holds for finite trees, and by passing to the limit for infinite trees, provided the limits defining

*x*(

*T*),

*x*(

*T*

^{1}),

*x*(

*T*

^{2}) exist. Dhar and Majumdar obtain from (4) that when the distance from the root of

*T*to the nearest surface vertex tends to infinity (what they call “deep in the lattice”), the characteristic ratio

*x*(

*T*) tends to 1. The authors then derive several explicit quantities (such as height probabilities of a vertex

*u*deep in the lattice) by replacing the characteristic ratio by 1.

#### 2.5.3 Transfer Matrix Approach

*u*,

*v*be two vertices in the tree at mutual distance

*n*. They determine

*n*+3 subtrees

*T*

_{1},…,

*T*

_{ n+3}, see Fig. 4. The number of allowed configurations when fixing the heights at

*u*and

*v*can then be obtained via a product Open image in new window of

*n*+3 two by two matrices with elements determined by the characteristic ratios of the subtrees

*T*

_{1},…,

*T*

_{ n+3}(see Sect. 4 below for the precise form of these matrices).In particular for the Bethe lattice, all the matrices involved in this product are equal to Open image in new window because the characteristic ratios of all the involved subtrees are equal to one. Let Open image in new window be a finite or infinite subtree of the full binary tree. The two-point correlation function, i.e., the probability that two vertices

*u*,

*v*at mutual distance

*n*have height

*i*resp.

*j*, for

*i*,

*j*∈{1,2,3}, is equal to (see [5, Sect. 5, Eq. (5.11)]) where Open image in new window is the ratio of the smallest and largest eigenvalues of the matrix Open image in new window and

*a*

_{ i,j }are some numerical constants depending on

*i*,

*j*. For the binary tree, Open image in new window and Open image in new window .

*u*, the probability that the avalanche

*Av*(

*u*,

*η*) is a given connected subset Open image in new window of cardinality

*n*containing

*u*is equal to for some constant

*C*, independent of the shape of the subset Open image in new window . Since there are 4

^{ n }

*n*

^{−3/2}(1+

*o*(1)) connected subsets of the Bethe lattice of cardinality

*n*containing

*u*, one concludes that for large

*n*(see [5, Sect. 6, Eqs. (6.13), (6.14)]), i.e., the tail of the avalanche size distribution decays like

*n*

^{−3/2}.

## 3 Some Characteristic Ratios

On the Bethe lattice Open image in new window for the (infinite) rooted subtrees (*i*=1,2,3) attached to every vertex since every Open image in new window is an infinite rooted binary tree (see Sect. 2.5.2). This property considerably simplifies the analysis of [5] and is no longer valid in the inhomogeneous or random cases studied here.

### 3.1 The Characteristic Ratio of the Random Binary Tree

*n*generations starting from a single individual (the “zero-th” generation) at time

*n*=0 (cf. Sect. 2.1). Then for

*n*>0, Open image in new window satisfies the recursive identity [5, Sect. 3, Eq. (3.12)] where

*x*(

*o*)=0). Notice that if

*u*,

*v*∈[0,1], then

*f*(

*u*,

*v*)≤1. Therefore, we view

*f*as a function from [0,1]

^{2}onto [1/2,1].

### Lemma 3.1

*For every finite subtree*Open image in new window

*of the Bethe lattice*,

*the ratio*Open image in new window .

*Moreover*,

*on*[0,1]

^{2}

*the function*

*f*

*defined by*(9)

*is symmetric*,

*i*.

*e*.,

*f*(

*u*,

*v*)=

*f*(

*v*,

*u*),

*and increasing in*

*u*

*and*

*v*,

*i*.

*e*.,

*for*

*u*

_{1}≤

*u*

_{2},

*and analogously in the other argument*.

### Proof

The proof is straightforward and left to the reader. □

### Proposition 3.1

*There exists a random variable* *X* _{∞} *such that* Open image in new window *in distribution as* *n*→∞.

### Proof

Denote by Open image in new window the distribution of Open image in new window . The recursion (8) induces a corresponding recursion on the distributions Open image in new window . We show that Open image in new window is a contraction on the set Open image in new window of probability measures on [1/2,1] endowed with the Wasserstein distance. This implies that it has a unique fixed point *μ* ^{∗} and from every initial *μ* _{0}, *μ* _{ n }→*μ* ^{∗} in Wasserstein distance and thus weakly.

*g*be a Lipschitz function. Then we have where

*T*

^{0},

*T*

^{1},

*T*

^{2}are the three trees which cannot be split into two subtrees both non-empty, see Fig. 5.Every other tree

*T*(appearing in the last sum of (12)) can be split into two subtrees both not reduced to a single point. Then using (8), the expression in (12) becomes

### Lemma 3.2

*The function*Open image in new window

*on*Open image in new window

*defined by*

*is a contraction on*Open image in new window

*endowed with the Wasserstein distance*.

*The contraction factor is bounded from above by*8/9.

### Proof

*g*:[1/2,1]→ℝ with Lipschitz constant less than or equal to one, i.e., such that |

*g*(

*x*)−

*g*(

*y*)|≤|

*x*−

*y*| for all

*x*,

*y*. We use the following two formulas for the Wasserstein distance of two elements

*μ*,

*ν*of Open image in new window [6]: where in the last right hand site the infimum is over all couplings ℙ with first marginal ℙ

_{1}(resp. second marginal ℙ

_{2}) equal to

*μ*(resp.

*ν*).

*f*and the fact that

*x*,

*y*∈[1/2,1] to estimate

*g*and a coupling ℙ of

*μ*and

*ν*: Taking now the infimum over all couplings ℙ, using (16) to bound (18), and taking the supremum over Open image in new window yields To estimate the term

*B*(

*μ*,

*ν*,

*g*), use the elementary bound (since

*x*,

*x*′,

*y*,

*y*′∈[1/2,1]):

*g*, and a coupling ℙ′=ℙ′(

*dxdx*′;

*dydy*′) of

*μ*⊗

*μ*and

*ν*⊗

*ν*Taking the supremum over Open image in new window , and infimum over the couplings ℙ′, we find Combining (19), (20) with (17) we arrive at where in the final inequality we used the elementary bound

This proof of Lemma 3.2 completes the proof of Proposition 3.1. □

### Remark 3.1

The random binomial tree is such that every vertex has two children with probability *p* ^{2}, 1 child with probability 2*p*(1−*p*) and 0 children with probability (1−*p*)^{2}, *p*∈[0,1]. For this case the same idea as in Lemma 3.2 gives a recursion leading to a contraction in the Wasserstein distance and hence Open image in new window has also a unique limiting distribution.

### 3.2 The Characteristic Ratio of Some Deterministic Trees

The recursion (8) allows also to compute the characteristic ratio for certain (deterministic) infinite subsets of the full binary tree in terms of iterations of sections of *f*.

#### 3.2.1 Infinite Branch

*n*, i.e., the tree Open image in new window consisting of a root and

*n*≥1 generations of two individuals each (see Fig. 6).Using (10) we obtain the recursion which by Lemma 3.1 gives that Open image in new window is monotonically increasing in

*n*with limit which is the unique positive solution of

### Remark 3.2

*T*is attached (it is a singleton in the backbone case). The characteristic ratio equals the positive solution of the fixed-point equation

*x*(

*T*) denotes the characteristic ratio of the tree

*T*, i.e.,

#### 3.2.2 Finite Perturbations of a Single Branch

*T*of the full binary tree Open image in new window at level

*n*in the infinite single branch Open image in new window leads to a tree Open image in new window with characteristic ratio where

*φ*(

*x*)=

*f*(1/2,

*x*) (cf. (23)) is applied

*n*times. This shows that inserting a finite tree

*T*at level

*n*has an effect on the characteristic ratio that vanishes in the limit

*n*→∞, exponentially fast in

*n*.

*x*(

*T*)≥1/2 and

*x*↦

*f*(

*x*,

*y*) is monotone for all

*y*(by Lemma 3.1), we have from (26) From (27) and (22) we conclude that for every

*infinite*subtree Open image in new window for which Open image in new window exists,

## 4 Transfer Matrix and Eigenvalues: Uniform Estimates

*x*

_{ i }’s are the characteristic ratios of some recursively defined subtrees.

*x*

_{ i }=1 for all

*i*, this exactly corresponds to the analysis in [5] (see the above Sect. 2.5.3). More precisely, in that case, one needs the minimal and maximal eigenvalues (denoted by

*λ*

_{−}and

*λ*

_{+}) of which are

*λ*

_{−}=1 and

*λ*

_{+}=4

^{ n }. This leads to a decay of the covariance proportional to 1/4

^{ n }=

*λ*

_{−}/

*λ*

_{+}and decay of avalanche size asymptotically proportional to

*n*containing the origin

*o*. In the general case the decay of covariance can be estimated by the ratio Open image in new window , and for the decay of the avalanche size distribution, one needs to estimate Open image in new window as well as the analogue of

*A*

_{ n }. In the case of a branching process, the

*x*

_{ i }’s appearing in the matrix Open image in new window are independent random variables with distribution

*μ*

^{∗}defined in the proof of Proposition 3.1.

In this section we therefore concentrate on the estimation of the eigenvalues of a matrix of the form Open image in new window for general *x* _{ i }’s.

### Lemma 4.1

- 1.
*For all**n**and all**x*_{ i }∈[1/2,1] (1≤*i*≤*n*),*the eigenvalues of*Open image in new window*are non*-*negative*. - 2.
*We have the inequality* - 3.
*We have*

*where*

*o*(1)

*tends to zero as*

*n*→∞,

*uniformly in the choice of the*

*x*

_{ i }

*’s*.

### Proof

*n*=1,

*a*

^{2}−4

*b*=5+4

*x*≥0. Hence

*n*≥2, we estimate, as in [11] and using we have

*b*≤

*a*

^{2}which implies that the eigenvalues are real and non-negative (

*a*≥0). Inequality (31) then follows from (36), (33).

*h*=

*b*/

*a*

^{2}tends to zero as

*n*→∞ at a speed at least

*C*(4/9)

^{ n }, we have which proves the third statement. □

Lemma 4.1 shows that the ratio Open image in new window behaves in leading order as Open image in new window . To estimate this ratio, we need to estimate Open image in new window from below. We start with a useful representation of Open image in new window .

### Lemma 4.2

*Define*

*Then we have*

*where*

*y*

_{ i }=(1+

*x*

_{ i })

*and where*

### Proof

### Lemma 4.3

- 1.
*For all**n*≥1,$$ E_1^n= \left ( \begin{array}{c@{\quad}c} 1&n\\ 0&1 \end{array} \right ),\qquad E_2^n =E_2, \qquad E_1^n E_2= \left ( \begin{array}{c@{\quad}c} n&n\\ 1&1 \end{array} \right ). $$(40) - 2.
*For all**r*≥1,*k*_{1},…,*k*_{ r },*k*_{ r+1}≥0,$$ \operatorname{Tr} \Biggl(\prod_{i=1}^r \bigl(E_1^{k_i}E_2\bigr) \Biggr)= \prod _{i=1}^r (1+k_i) $$(41)*and*$$ \operatorname{Tr} \Biggl( \Biggl(\prod _{i=1}^r \bigl(E_1^{k_i}E_2 \bigr) \Biggr) E_1^{k_{r+1}} \Biggr)= \Biggl(\prod _{i=2}^r (1+k_i) \Biggr) (1+k_1+k_{r+1}). $$(42)

### Proof

### Proposition 4.1

*For all*

*x*

_{1},…,

*x*

_{ n }∈[1/2,1]

*we have*

*As a consequence we have the following uniform upper bound*

*where*

### Proof

*α*=(

*α*

_{1},…,

*α*

_{ n })∈{0,1}

^{ n }. Using Lemma 4.3 we see that where Open image in new window —with the convention

*α*

_{ n+1}=1—is the number of intervals of successive 1’s in the configuration

*α*. Hence by( 38) we obtain the lower bound Next since \(\sum_{\alpha=(\alpha_{1},\ldots,\alpha_{n})\in\{0,1\} ^{n}} (\prod_{i=1}^{n}(2y_{i})^{\alpha_{i}} )=\prod_{i=1}^{n}(1+2y_{i})=:Z\), we rewrite

*y*

_{ i })

_{ i }dependent probability measure on the

*α*’s. Now apply Jensen’s inequality in (46) to obtain Finally, use

*y*

_{ i }=1+

*x*

_{ i }∈[3/2,2] to estimate

*y*

_{ i }=(1+

*x*

_{ i }) with

*x*

_{ i }∈[1/2,1] and the fact that

*x*↦(1+

*x*)(3+2

*x*)

^{−1}is increasing to estimate

## 5 Transfer Matrix: Annealed Estimates

In this section we look at the eigenvalues of Open image in new window where now the *x* _{ i }’s are i.i.d. with a law *μ* on [1/2,1]. We denote by ℙ the joint law of the *x* _{ i }’s and by \(\mathbb{E}\) the corresponding expectation.

We start with the following lemma.

### Lemma 5.1

*For all*

*γ*≥0

*the eigenvalues of*

*given by*

*are non*-

*negative*.

### Proof

Elementary computation. □

### Theorem 5.1

*Let*Open image in new window

*denote the largest*,

*resp*.

*smallest*,

*eigenvalues of*Open image in new window

*where the*

*x*

_{ i }

*’s are i*.

*i*.

*d*.

*with a law supported on*[1/2,1].

*Denote*

*Then we have*:

- 1.
*Concentration property*:*there exists**C*>0*such that for all**ε*>0$$ \mathbb{P} \bigl( \bigl(Y_n-\mathbb{E}(Y_n) \bigr)>\varepsilon \bigr)\leq e^{-C\varepsilon^2 n}. $$(49) - 2.
*The limits**exist and satisfy*\(L_{\pm}=\mathbb{E}(L_{\pm})\)*almost surely*;*moreover*$$ L_++L_-= 2\mathbb{E} \bigl(\log(1+x_1) \bigr). $$(51) - 3.
*Upper bound*$$ \lim_{n\to\infty} Y_n=L_+-L_- = \lim_{n\to\infty} \mathbb{E}(Y_n)\leq \log\frac{\varLambda_+ (\gamma)}{\varLambda_- (\gamma)} $$(52)*where**Λ*_{±}*are given by*(47)*and*\(\gamma= (\mathbb{E}((1+x_{1})^{-1}) )^{-1}\in[3/2,2]\).

### Proof

*y*

_{ i }=1+

*x*

_{ i },1≤

*i*≤

*n*, to obtain with \(\gamma^{-1}= \mathbb{E}(y_{1}^{-1})\), and

*Λ*

_{±}(

*γ*) given by (47). Because of (35), we finally derive (51) by the law of large numbers. □

## 6 Covariance and Avalanche Sizes

### 6.1 Quenched and Annealed Covariance

### Theorem 6.1

*Let*Open image in new window

*be a finite or infinite subtree of the full binary tree*Open image in new window .

*As before*,

*denote by*Open image in new window

*the uniform measure on the recurrent configurations*Open image in new window

*of the sandpile model on*Open image in new window

*and let*Open image in new window

*be at mutual distance*

*n*.

*Then we have the following estimate for the two point correlation function*

*In particular*, Open image in new window

*is absolutely summable*,

*uniformly in the choice of the subtree*Open image in new window :

### Proof

The first statement (56) follows from the expression of the covariance in Sect. 2.5.3, Eq. (5), together with Proposition 4.1. Since the number of vertices at distance *n* in Open image in new window from the origin is bounded from above by 2^{ n }, and *γ*<1/2, (57) follows from (56). □

### Theorem 6.2

*Let*Open image in new window

*be the stationary binary branching process with reproduction probability*

*p*,

*conditioned to have a path from its root*

*o*

*to a vertex at distance*

*n*.

*Let*Open image in new window

*denote the uniform measure on recurrent configurations on*Open image in new window .

*Let*Open image in new window

*denote the covariance of the height variables at*

*o*

*and at a vertex*

*v*(

*n*)

*at distance*

*n*

*from the root*

*o*.

*Then we have the following annealed lower bound on the covariance*.

*where*\(\gamma^{-1}=\mathbb{E}(y_{1}^{-1})\)

*and*

*y*

_{1}=1+

*x*

_{1},

*with*

*x*

_{1}

*distributed according to the measure of Proposition*3.1,

*and*

*Λ*

_{±}

*given in*(47).

### 6.2 Avalanche Sizes

We start by defining a matrix associated to an avalanche cluster. Roughly speaking, the probability that the avalanche coincides with the cluster is the inverse of the maximal eigenvalue of this matrix. Let Open image in new window be a finite or infinite subtree of the full rootless binary tree, containing the origin *o*. Let Open image in new window be a connected subset of Open image in new window containing the origin.

### Definition 6.1

*n*+2 subtrees

*T*

_{1},…,

*T*

_{ n+2}with corresponding characteristic ratios

*x*(

*T*

_{ i }). Then

### Lemma 6.1

*As before*,

*for a stable height configuration*

*η*,

*let*

*Av*(

*o*,

*η*)

*denote the avalanche caused by addition of a single grain at the origin in*

*η*,

*and*Open image in new window

*the uniform measure on recurrent configurations*Open image in new window

*on*Open image in new window .

*Then there exist constants*

*c*

_{1},

*c*

_{2}>0

*such that*

*where*Open image in new window

*is the matrix of*(59)

*associated to*Open image in new window

*of Definition*6.1.

### Proof

It follows from the expression for Open image in new window in Sect. 2.5.3, Eq. (6). □

### Definition 6.2

*o*and of cardinality

*n*.

- 1.The growth rate is defined as
- 2.We define the averaged growth rate as

*κ*=log4, since see also [5], Sect. 2.5.2. For a stationary binary branching process we have the upper bound For an exact expression Open image in new window from which inequality (63) follows immediately, we refer to Appendix, Proposition 7.2.

### Theorem 6.3

- 1.
*There exists**C*>0*such that for all**n*≥1*we have**In particular if the tree*Open image in new window*has growth rate*Open image in new window ,*then the avalanche size decays exponentially*. - 2.
*For the stationary binary branching tree with branching probability**p*,*we have the estimate**In particular*,*for*$$p< 0.54511\ldots $$*the averaged*(*over the realization of the tree*)*probability of an avalanche size**n**decays exponentially in n*. - 3.
*For the binomial branching tree we have*

*In particular*,

*for*

*the averaged*(

*over the realization of the tree*)

*probability of an avalanche size*

*n*

*decays exponentially in n*.

### Proof

*p*is monotone, which yields that for

*p*<0.54511 the expected number of avalanches of size

*n*decays exponentially.

The third statement follows from Lemma 6.1, (67), (63), (62) and Remark 7.1. □

### Remark 6.1

From Theorem 6.3 we conclude that avalanche sizes decay exponentially for *p* small enough. We know that on the full binary tree (corresponding to *p*=1) avalanche sizes have a power law decay. It is a natural question whether the transition between exponential and power law decay occurs at some unique non-trivial value *p* _{ c }∈(1/2,1). This question is related to the behavior of a random walk on the full binary tree, killed upon exiting a random subtree. Would this random walk have a survival time with a finite exponential moment, then we would have *p* _{ c }=1. Notice that for the corresponding problem on the lattice ℤ^{ d }, this random walk *does not have a finite exponential moment* of its survival time, because the tail of the survival time is stretched exponential (Donsker-Varadhan tail \(e^{-c t^{d/(d+2)}}\)). However, on the tree in the annealed case corresponding to Theorem 6.3, the decay of the tail of the survival time is not straightforward. Formally taking the limit *d*→∞ in the Donsker-Varadhan tail suggests that on the tree, the survival time of this random walk has a finite exponential moment, which points into the direction *p* _{ c }=1. See also [2] for trapped random walk on a tree.

## Notes

### Acknowledgements

We would like to thank Remco van der Hofstad for useful discussions. F.R. and W.M.R. thank NWO for financial support in the project “sandpile models in neuroscience” within the STAR cluster. This work was supported by ANR-2010-BLAN-0108 and by NWO. We also thank laboratoire MAP5 at Université Paris Descartes, Universities of Leiden and Delft for financial support and hospitality.

### Open Access

This article is distributed under the terms of the Creative Commons Attribution License which permits any use, distribution, and reproduction in any medium, provided the original author(s) and the source are credited.

## References

- 1.Andrews, G.E., Askey, R., Ranjan, R.: Special Functions. Encyclopedia of Mathematics and Its Applications, vol. 71. Cambridge University Press, Cambridge (1999). ISBN 978-0-521-62321-6 zbMATHGoogle Scholar
- 2.Chen, M., Yan, S., Zhou, X.: The range of random walk on trees and related trapping problem. Acta Math. Appl. Sin.
**13**, 1–16 (1997) MathSciNetGoogle Scholar - 3.Dhar, D.: Theoretical studies of self-organized criticality. Physica, A
**369**, 29–70 (2006) MathSciNetADSCrossRefGoogle Scholar - 4.Dhar, D.: Self-organized critical state of sandpile automaton models. Phys. Rev. Lett.
**64**, 1613–1616 (1990) MathSciNetADSzbMATHCrossRefGoogle Scholar - 5.Dhar, D., Majumdar, S.N.: Abelian sandpile model on the Bethe lattice. J. Phys. A, Math. Gen.
**23**, 4333–4350 (1990) MathSciNetADSCrossRefGoogle Scholar - 6.Dudley, R.M.: Real Analysis and Probability. Cambridge University Press, Cambridge (2002) zbMATHCrossRefGoogle Scholar
- 7.Gauthier, G.: Avalanche dynamics of the Abelian sandpile model on the expanded cactus graph. Preprint (2011). Available at http://arxiv.org/abs/1110.6263
- 8.Goh, K.I., Lee, D.S., Kahng, B., Kim, D.: Sandpile on scale-free networks. Phys. Rev. Lett.
**91**, 148701 (2003) ADSCrossRefGoogle Scholar - 9.Járai, A.A.: Thermodynamic limit of the abelian sandpile model on ℤ
^{d}. Markov Process. Relat. Fields**11**, 313–336 (2005) zbMATHGoogle Scholar - 10.Levina, A., Herrmann, J.M., Geisel, T.: Dynamical synapses causing self-organized criticality in neural networks. Nat. Phys.
**3**, 857–860 (2007) CrossRefGoogle Scholar - 11.Maes, C., Redig, F., Saada, E.: The Abelian sandpile model on an infinite tree. Ann. Probab.
**30**, 2081–2107 (2002) MathSciNetzbMATHCrossRefGoogle Scholar - 12.Maes, C., Redig, F., Saada, E.: Abelian sandpile models in infinite volume. Sankhya,
**67**(4), 634–661 (2005) MathSciNetzbMATHGoogle Scholar - 13.Matter, M., Nagnibeda, T.: Abelian sandpile model on randomly rooted graphs and self-similar groups. Preprint (2010). Available at http://arxiv.org/abs/1105.4036
- 14.Osedelec, V.L.: A multiplicative ergodic theorem: Lyapunov characteristic number for dynamical systems. Trans. Mosc. Math. Soc.
**19**, 197–231 (1968) Google Scholar - 15.Redig, F.: Mathematical aspects of the abelian sandpile model. In: Mathematical Statistical Physics, Les Houches Summer School, pp. 657–729. Elsevier, Amsterdam (1989) Google Scholar
- 16.Temme, N.M.: Special Functions: An Introduction to the Classical Functions of Mathematical Physics. Wiley, New York (1996). ISBN 0-471-11313-1 zbMATHCrossRefGoogle Scholar