Explicit characterization of the identity configuration in an Abelian Sandpile Model
aa r X i v : . [ c ond - m a t . s t a t - m ec h ] O c t Explicit characterization of the identity configurationin an Abelian Sandpile Model
Sergio Caracciolo, Guglielmo Paoletti, and Andrea Sportiello Dip. Fisica, Universit`a degli Studi di Milano and INFN, via G. Celoria 16, 20133 Milano, Italy Dip. Fisica, Universit`a degli Studi di Pisa and INFN, largo B. Pontecorvo 3, 56127 Pisa, Italy (Dated: June 18, 2018)Since the work of Creutz, identifying the group identities for the Abelian Sandpile Model (ASM) ona given lattice is a puzzling issue: on rectangular portions of Z complex quasi–self-similar structuresarise. We study the ASM on the square lattice, in different geometries, and a variant with directededges. Cylinders, through their extra symmetry, allow an easy determination of the identity, whichis a homogeneous function. The directed variant on square geometry shows a remarkable exactstructure, asymptotically self-similar. INTRODUCTION
Bak, Tang and Wiesenfeld (BTW) introduced theAbelian Sandpile Model (with a different name) in [1] asa simple model for self-organized criticality. The modelhas been soon generalized from the square lattice to ar-bitrary graphs [2]. It is a lattice automata, described interms of local height variables z i ∈ N , i.e. the number of“particles” at site i , so that a configuration is given byset C = { z i } . Particles fall into the system through arandom memoryless process, and, as soon as some height z i is higher than the local threshold value ¯ z i , it topples tothe neighbouring sites, i.e. the value z i decreases by the(integer) amount ∆ ii , while, for each site j = i , increasesby the (integer) amount − ∆ ij . This relaxation process(the avalanche ) is guaranteed to stop in a finite numberof steps if b − i := P j ∆ ij ≥ i , and a cer-tain technical condition is satisfied (the matrix ∆ mustbe a strictly dissipative diffusion kernel). All the heightsremain non-negative if this is the case in the startingconfiguration, and ¯ z i ≥ ∆ ii . For future convenience wealso define b + i := P j ∆ Tij , and assume that also thesequantities be non-negative.A graph, in general directed and with multiple edges,is thus defined through the non-diagonal part of − ∆,seen as an adjacency matrix, while the non-zero values b ± i are regarded as (in- or out-coming) “connections tothe border” (and the sites i with b + i or b − i non-zero aresaid to be “on the border”). If ∆ is symmetric, we havean undirected graph, and automatically b + i = b − i We justwrite b i for b + i and b − i when clear that they are equal.This is the case in the original BTW model, which isdefined on a portion of the Z lattice. We have in thiscase ¯ z i = ∆ ii = 4 on each site, and ∆ ij = − L x and L y , we have b i = 1 for vertices on thesides of the rectangle, and b i = 2 for the corners.In [1] it is described how, already for this case, theavalanches show an unpredictable dynamics, with power-law size distribution, which candidated the model as a toydescription of many interesting features, such as the 1 /f noise.Afterwords, Deepak Dhar and collaborators, in a seriesof relevant works [2–4], elucidated a deep structure of theMarkov chain related to the general ASM: the set of stable configurations S = N i { , , . . . , ¯ z i − } , the ones whereno toppling can occur, is divided into a set T of transient and a set R of recurrent ones, the latter being visitedan infinite number of times by the chain, and forming asingle ergodic basin. Given the natural partial ordering, C C ′ iff z i ≤ z ′ i for all i , then R is in a sense “higher”than T , more precisely ∄ ( C, C ′ ) ∈ T × R : C ≻ C ′ ; (1) ∀ C ∈ T ∃ C ′ ∈ R : C ≺ C ′ . (2)In particular, the maximally-filled configuration C max = { ¯ z i − } is in R , and higher than any other stable config-uration.The set R has an underlying abelian structure, forwhich a presentation is explicitly constructed in termsof the matrix ∆, through the (heavy) study of its Smithnormal form. Shortcuts of the construction and more ex-plicit analytical results are achieved in the special caseof a rectangular L x × L y portion of the square lattice,and still stronger results are obtained for the case of L x = L y [4].Call ˜ a i the operator which adds a particle at site i toa configuration C , and a i the formal operator which ap-plies ˜ a i , followed by a sequence of topplings which makesthe configuration stable. Remarkably, the final configu-ration a i C is independent from the sequence of topplings,and also, applying two operators, the two configurations a j a i C and a i a j C coincide, so that at a formal level a i and a j do commute.More precisely, if a i acting on C consists of the fall of aparticle in i , ˜ a i , and the sequence of topplings t i , . . . , t i k on sites i , . . . , i k , the univocal definition of a i and thecommutation of a i and a j follow from the two facts: z j ≥ ¯ z j : t j ˜ a i C = ˜ a i t j C ; (3) z i ≥ ¯ z i and z j ≥ ¯ z j : t i t j C = t j t i C . (4)Another consequence is that, instead of doing all the top-plings immediately, we can postpone some of them afterthe following ˜ a ’s, and still get the same result. Similarmanipulations show that the relation a ∆ ii i = Y j = i a − ∆ ij j (5)holds when applied to an arbitrary configuration.These facts lead to the definition of an abelian semi-group operation between two configurations, as the sumof the local height variables z i , followed by a relaxationprocess [5]: C ⊕ C ′ = (cid:18) Y i a z i i (cid:19) C ′ = (cid:18) Y i a z ′ i i (cid:19) C . (6)For a configuration C , we define multiplication by a pos-itive integer: k C = C ⊕ · · · ⊕ C | {z } k . (7)The set R of recurrent configurations is special, as eachoperator a i has an inverse in this set, so the operationabove, restricted to R , is raised to a group operation. Ac-cording to the Fundamental Theorem of Finite AbelianGroups, any such group must be a “discrete thorus” Z d × Z d ×· · ·× Z d g , for some integers d ≥ d ≥ · · · ≥ d g ,and such that d α +1 divides d α for each α = 1 , . . . , g − d α , called elementary divisors of ∆, and aset of generators e α with the proper periodicities, can beconstructed through the normal form decomposition [4].The composition of whatever C = { z i } with the set R acts then as a translation on this thoroidal geometry. Afurther consequence is that, for any recurrent configura-tion C , the inverse configuration ( − C ) is defined, so that k C is defined for k ∈ Z .Consider the product over sites i of equations (5) Y i a ∆ ii i = Y i Y j = i a − ∆ ij j = Y i a − P j = i ∆ ji i (8)On the set R , the inverses of the formal operators a i aredefined, so that we can simplify common factors in (8),recognize the expression for b + , and get Y i a b + i i = I (9)so that Q i a b + i i C = C is a necessary condition for C tobe recurrent (but it is also sufficient, as no transient con-figuration is found twice in the same realization of theMarkov chain), and goes under the name of identity test .This condition turns into an equivalent and computa-tionally-cheaper procedure, called burning test , of whicha side product, in case of positive answer, is a spanning Figure 1: Left: a portion of the square lattice. Right: aportion of the directed square lattice we considered in thiswork, with in- and out-edges alternated cyclically, and whiteand gray faces with arrows oriented clockwise and counter-clockwise. arborescence rooted on the vertices of the border. So, theburning test provides us a bijection between the two en-sembles, of recurrent sandpile configurations and rootedarborescences with roots on the border. This is in agree-ment with Kirchhoff Matrix-Tree theorem, which statesthat the number of such arborescences is given by det ∆,while the number of recurrent configurations is known tobe det ∆ as the first step of the procedure which deter-mines the elementary divisors d α [3].If, for the graph identified by ∆, a planar embeddingexists, with all sites i such that b + i > T , and needsno more words. If the graph is directed, the arborescenceis directed in the natural way, while the dual tree is con-strained in some complicated way (some combinations ofedge-occupations are forced to fixed values). However,the graphical picture simplifies if, on the planar embed-ding, in-coming and out-coming edges are cyclically al-ternating (and all plaquettes have consistent clockwise orcounter-clockwise perimeters), and all sites with either b + or b − positive are on the most external face, with in- andout-going arrows cyclically alternating (cfr. for exampleFigure 1). We will call a directed graph of this kind a planar alternating directed graph .The connection with uniform spanning trees and theKirchhoff theorem explains a posteriori the arising of self-organized criticality, i.e. the appearence of long-range be-haviour with no need of tuning any parameter. Indeed,uniform spanning trees on regular 2-dimensional latticesare a c = − q → |
2) non-linear σ -model [7]. If instead one considers the larger ensembleof spanning forests, in a parameter t counting the compo-nents (or describing the curvature of the OSP(1 |
2) super-sphere), the theory in two dimensions is scale-invariantfor three values: at t = 0 (the spanning trees, or theendpoint of the ferromagnetic critical line of Potts), atthe infinite-temperature point t = ∞ , and at some non-universal negative t corresponding to the endpoint of theanti-ferromagnetic critical line of Potts, being t c = − / σ -model, one can deduce at a perturbativelevel the renormalization group flow, and, in particular,that the system is asymptotically free for t > t = 0 + is an ultraviolet fixed point.An interesting connection between the height variablesin the ASM and the fields of the underlying CFT is per-formed in [3, 9–11], and especially in [11]. IDENTITIES
Given the algebraic relation (9), and the semi-group op-eration (6), one could define the frame configuration Id f as the one with z i = b + i for all i , and realize that it actsas an identity on recurrent configurations, Id f ⊕ C = C if C ∈ R . Conversely, in general it does not leave un-changed a transient configuration, and in particular, as,for any relevant extensive graph, Id f is itself transient,we have that Id f ⊕ Id f , and Id f ⊕ Id f ⊕ Id f , and so on,are all different, up to some number of repetitions k atwhich the configuration is sufficiently filled up with par-ticles to be recurrent. We call Id r this configuration, and k (∆) the minimum number of repetitions of Id f requiredin the ASM identified by ∆ (we name it the filling num-ber of ∆). The configuration Id r is the identity in theabelian group Z d × · · · × Z d g described above, and to-gether with a set of generators e α , completely identifiesin a constructive way the group structure of the statis-tical ensemble. The relevance of this configuration hasbeen stressed first by M. Creutz [5], so that we shall callit Creutz identity . See fig. 2 for an example.Unfortunately, despite many efforts, it has not beenpossible to give a closed-formula recipe for this identitystate on given large lattices, not even in the case of a L × L square, and the direct numerical investigation ofthese configurations at various sizes has produced pecu-liar puzzling pictures [12].In a large-side limit, we have the formation of curvi-linear triangular regions of extensive size (of order L ),filled with regular patterns, and occasionally crossedby straigth “defect lines”, of widths of order 1, which,furthermore, occasionally meet at Y-shaped “scatteringpoints”, satisfying peculiar conservation laws [13].Similar features emerge also for the filling numbers,e.g. on the square lattice of size L , the index k L ≡ k (∆ ( L ) ) is not badly fitted, for even L , by a parabola k L ≃ L / o ( L ), but showing fluctuations due tounpredictable number-theoretical properties of L . Thechallenging sequence of these numbers, for L up to 64, isgiven in Table I. It should be noted that, conversely, oddsizes 2 L + 1 are related to 2 L through a property provenin [4, sec. 7]. Figure 2: The top-left corners of the identities Id ( L )r for L =50 and L = 100 in the BTW model (the other quadrants arerelated by symmetry). Heights from 0 to 3 correspond to graytones from dark to light. The smaller-size identity is partiallyreproduced at the corner of the larger one, in a fashion whichresembles the results of Theorem 1. L ⊲ k L ◮⊲
24 26 28 30 32 34 36 38 40 42 44 ⊲◮
103 119 138 156 180 198 226 248 276 305 334 ◮⊲
46 48 50 52 54 56 58 60 62 64 ◮
367 397 430 464 499 538 572 615 653 699Table I: Values of k L for the BTK Abelian Sandpile on squaregeometries of even size, for L = 2 , . . . , The determination of Id r for a given graph is a proce-dure polynomial in the size of the graph. E.g. one couldprove that k (∆) is sub-exponential, and that relaxing C ⊕ C ′ for C and C ′ both stable is polynomial, thenone can produce the powers 2 s Id f recursively in s , up toget twice the same configuration. Better algorithms existhowever, see for example [5].Still, one would like to have a better understanding ofthese identity configurations. This motivates us to thestudy of the ASM on different regular graphs, such thatthey resemble as much as possible the original model,but simplify the problem in some regards so that thefamily (over L ) of resulting identities can be understoodtheoretically.In this research, we keep in mind a few principles:(1) we wish to simplify the problem in some structuralway; (2) we want to preserve the property of the L × L lattice, discussed in [4, sec. 4], of reconstruction of theelementary divisors of ∆ from a suitable matrix n yy ′ ,outcome of a machinery for producing an “economic”presentation of the group, in terms of O ( L ) generators a i of the whole set of L ; (3) possibly, we want to pre-serve planarity, and the interpretation of dual spanningsubgraphs as spanning trees on the dual lattice, i.e. wewant to use either planar undirected graphs, or planaralternating directed graphs (according to our definitionabove). A PROLOG: CYLINDRIC GEOMETRIES
A first possible simplification comes from working on acylindrical geometry. We call respectively periodic , open and closed the three natural conditions on the boundariesof a L x × L y rectangle. For example, in the BTW Model,for site ( i, z i, → z i, − z i ± , → z i ± , + 1; z i, → z i, + 1; z i,L y → z i,L y + 1;open: z i, → z i, − z i ± , → z i ± , + 1; z i, → z i, + 1;closed: z i, → z i, − z i ± , → z i ± , + 1; z i, → z i, + 1;The external face, with b ± i = 0, is in correspondence ofopen boundaries, so, as we want a single external face,if we take periodic boundary conditions in one direction(say, along x ), the only possible choice in this frameworkis to take closed and open conditions on the two sides inthe y direction.Our notation is that ¯ z i takes the same value every-where: on open boundaries, b − i and b + i are determinedaccordingly, while on closed boundaries either we addsome extra “loop” edges, or we take ¯ z i − ∆ ii > L x × L y cylinder than on a L x × L y open-boundary rectangle), but , for what concernsthe determination of Id r through relaxation of k Id f , the system behaves as a quasi-unidimensional one, and Id r isin general trivially determined.Furthermore, in many cases Id r is just the maximally-filled configuration C max . This fact is easily proven, ei-ther by checking that Id r ⊕ Id f = Id r , which is easy inthe quasi-1-dimensional formulation (or in other words,by exploiting the translation symmetry in one direction),or with a simple burning-test argument, in the cases inwhich on each site there is a single incoming edge fromsites nearer to the border. Figure 3 shows a few exam-ples.Such a peculiar property has a desiderable consequenceon the issue of inversion of a recurrent configuration. In-deed, if C and C ′ are in R , we have that( − C ) ⊕ C ′ = Id r ⊕ ( − C ) ⊕ C ′ = ( Id r − C ) ⊕ C ′ (10)where, if C = { z i } , ( − C ) = { ˜ z i } is the inverse configu-ration we seek, but Id r − C is simply the configurationwith ˜ z i = ¯ z i − z i −
1, and has all non-negative heights.
PSEUDO-MANHATTAN LATTICE
From this section on, we will concentrate on square ge-ometries, on the square lattice with edges having a givenorientation, and all vertices having in- and out-degreeequal to 2.For this reason, we can work with ¯ z i = 2, so that S = { , } V instead of { , , , } V (a kind of simplification,as from “CMYK” colour printing to “black and white”).In this step we lose in general a bit of symmetry: e.g. ona square of size 2 L we have still the four rotations, butwe lose the reflections, which are arrow-reversing, whileon a square of size 2 L + 1 we lose rotations of an angle π/
2, and only have rotation of π .Square lattices with oriented edges have already beenconsidered in Statistical Mechanics, especially in thevariant called “Manhattan Lattice” (i.e. with horizontaledges oriented east- and west-bound alternately on con-secutive rows, and coherently within a row, and similarlyfor vertical edges), cfr. for example [14]. However thislattice in two dimensions is not a planar alternating di-rected graph, and the results for the related ASM modelwill be discussed only briefly in last section.We start by analysing a less common variant, whichis better behaving for what concerns the ASM model,and which we call pseudo-Manhattan lattice (it appears,for example, in the totally unrelated paper [15]). In thiscase, the horizontal edges are oriented east- and west-bound alternately in both directions (i.e. in a chequerdesign), and similarly for vertical edges. As a result,all plaquettes have cyclically oriented edges. A smallportion of this lattice is shown in Figure 1, where it isdepicted indeed as the prototype planar alternating di-rected graph. Conventionally, in all our examples (unless Figure 3: Reduction to a quasi-1-dimensional system for theASM on cylindric geometry, on a few examples all having Id r coinciding with C max . From top to bottom: a portion ofthe square lattice, in the two orientations; of the hexagonallattice, in the two orientations, of Manhattan and pseudo-Manhattan lattices. Plain edges correspond to ∆ ij = ∆ ji = −
1, while a directed edge (from i to j ) with k arrows corre-spond to ∆ ij = − k . In-(out-)coming bold arrows correspondto b − ( b + ) equal to 1, while the lozenge-shaped double-arrowscorrespond to b − = b + = 1. A loop with k arrows on i corre-sponds to ¯ z i − ∆ ii = k , otherwise ¯ z i = ∆ ii . otherwise specified) we fix the orientations at the top-right corner to be as in the top-right corner of Figure 1.Quite recently, in [19] both Manhattan and pseudo-Manhattan lattices have been considered as an interest-ing variant of the ASM model (the latter under the nameof F-lattice ), with motivations analogous to ours. Thiscorroborates our claim that these variants are naturalsimplifications of certain features in the original BTKmodel.The main feature of the Creutz identities, on square
Figure 4: The top-right corners of the identities Id ( L )r for L = 21 and L = 64 (remark: 64 = 3 ·
21 + 1). The smaller-size one is exactly reproduced at the corner of the larger one,while the rest of the latter has an evident regular structure,according to theorem 1. portions of the pseudo-Manhattan lattice with even side,is self-similarity for sides best approximating the ratio1 /
3, up to a trivial part, as illustrated in Figure 4.The precise statement is in the following Theorem 1.Call Id ( L )r the set of heights in the Creutz identity for thesquare of side 2 L , encoded as a L × L matrix for one ofthe four portions related by rotation symmetry. Say thatindex (1 ,
1) is at a corner of the lattice, and index (
L, L )is in the middle. We have
Theorem 1
Say L = 3 ℓ + s , with s = 1 , , . Then,Id ( L )r is determined from Id ( ℓ )r and a closed formula, andthus, recursively, by a deterministic telescopic procedurein at most ⌊ log L ⌋ steps.For s = 1 or , we have (cid:0) Id ( L )r (cid:1) ij = (cid:0) Id ( ℓ )r (cid:1) ij if i, j ≤ ℓ ,otherwise (cid:0) Id ( L )r (cid:1) ij = 0 iff, for i + j + s even, i < ℓ, | L − ℓ − j | − i >
0; (11a) j > ℓ, | ℓ + 1 − i | + j < ℓ + s ; (11b) i ≤ ℓ, j = L − ℓ ; (11c) and, for i + j + s odd, i > ℓ, | L − ℓ − i | − j >
0; (11d) i > ℓ − , | ℓ + s − j | + i < ℓ ; (11e) If s = 2 the same holds with i and j transposed in Id ( L )r and Id ( ℓ )r (but not in (11)). Furthermore, k L = L ( L +1)2 . The statement of equations (11) is graphically repre-sented in Figure 5.The understanding of the Creutz identity on squareportions of the pseudo-Manhattan lattice is completedby the following theorem, relating the identity at side2 L + 1 to the one at side 2 L . We encoded the identity ateven sides in a L × L matrix Id ( L )r , exploiting the rotationsymmetry, such that the extended 2 L × L matrix hasthe property ( Id ( L )r ) i,j = ( Id ( L )r ) L +1 − j,i . (12)We can similarly encode the identity at odd sides in astructure of almost 1 / L +1) × ( L + 1) matrix ˆ Id ( L )r , using the fact that( ˆ Id ( L )r ) i,j = ( ˆ Id ( L )r ) L +2 − i,j = ( ˆ Id ( L )r ) i, L +2 − j . (13) Theorem 2
We have ( ˆ Id ( L )r ) i,j = ( Id ( L )r ) i,j if i, j ≤ L , ( ˆ Id ( L )r ) L +1 ,j = 0 if L − j is odd and 1 if it is even, and,for i ≤ L , ( ˆ Id ( L )r ) i,L +1 = 0 if L − i is even and 1 if it isodd. The statement of this theorem is illustrated in Figure 6.The proof of Theorem 2, given Theorem 1, is easilyachieved through arguments completely analogous to theones of [4, sec. 7]. We postpone the (harder) proof of The-orem 1 to the discussion of the equivalence with Theorem4 below.There are some similarities and differences with theidentities in the BTW model. In that case, besides theevident height-2 square in the middle, there arise somecurvilinear triangolar structures, mostly homogeneous atheight 3, but some others “texturized”, so that thereexists a variety of patterns which appear extensively atlarge sizes (in the metaphor of CMYK offset printing, likethe way in which composite colours are produced!). Afirst attempt of classification of these structures appearsin ?? . In our case, we only have “black and white”, but,as four zeroes in a square are a forbidden configuration(as well as many others too dense with zeroes), we cannot have extesive regions of zeroes (black, in our draw-ings of fig. 4). A big square in the middle is still there,rotated of 45 degrees, while triangoloids are replaced byexact “45-45-90” right triangles in a “texturized gray”coming from a chequered pattern. Indeed, to our knowl-edge, such a regular structure as in theorem 1 was not deducible a priori (in particular, not before the publica-tion of [19]), and our initial motivation was to study theemergence of patterns in a 2-colour case.The statement of Theorem 1 would suggest to look forsimilar features also in the BTW model. It turns outthat, while in the directed case the ⌊ ( L − / ⌋ size is fully contained in the corner of the L size, in the BTWthe ⌊ L/ ⌋ size is partially contained in the corner of the L size, in an empirical way which strongly fluctuates with L , but is in most cases more than 50% (cfr. figure 2 foran example).The heuristics in the case of the Manhattan lattice,under various boundary prescriptions, are somewhat in-termediate: the continuum-limit configuration exists andcoincides with the one for the pseudo-Manhattan, but the ⌊ ( L − / ⌋ corners are only partially reproduced by thesize- ℓ identities.How could one prove Theorem 1? An elegant algebraicproperty of the identity is that it is the only recurrentconfiguration for which all the “charges” are zero (see[4], eqs. (3.3) and (3.4)). More precisely, and in a slightlydifferent notation, given a whatever ordering of the sites,and any site i , and calling A i,j the minor ( i, j ) of a matrix A , we have Q j ( C ) := X i z i ( − i + j det ∆ i,j (14)and Q j ( Id r ) = q j det ∆ with q j ∈ Z , with Id r being theonly stable recurrent configuration with this property.A possible proof direction (that we do not follow inthis paper) could have been as follows. An algebraic re-statement of the expression for the charges is achieved inGrassmann calculus, through the introduction of a pairof anticommuting variables ¯ ψ i , ψ i per site. Then, byGrassmann Gaussian integration, we have that( − i + j det ∆ i,j = Z D ( ψ, ¯ ψ ) ¯ ψ i ψ j e ¯ ψ ∆ ψ (15)where there is a contribution b − k ¯ ψ k ψ k in the exponen-tial for each site on the border, and a contribution( ¯ ψ h − ¯ ψ k ) ψ k if a particle falls into h after a topplingin k , i.e. ∆ hk = − ψ and ψ ).So the configurations in the same equivalence class ofthe identity are the only ones such that, for each j , the“expectation value” D(cid:16) X i z i ¯ ψ i (cid:17) ψ j E := R D ( ψ, ¯ ψ ) ( P i z i ¯ ψ i ) ψ j e ¯ ψ ∆ ψ R D ( ψ, ¯ ψ ) e ¯ ψ ∆ ψ (16)is integer-valued.Furthermore, expressions as in the right hand side of(15) are related, through Kirchhoff theorem, to the com-binatorics of a collection of directed spanning trees, all s = 1 : ℓ ℓ − ℓ − ℓℓℓ Id ( ℓ )r s = 2 : ℓ ℓ ℓ − ℓℓ + 1 ℓ Id ( ℓ )r T s = 3 : ℓ ℓ + 1 3 ℓ − ℓℓ + 2 ℓ Id ( ℓ )r Figure 5: The non-recursive part of the identities Id ( L )r (top-right corners) for L = 3 ℓ + s , and s ∈ { , , } , illustrating the setdescribed by (11). For s = 2, we show the transposed of Id r . Black and white stand respectively for z i = 0 and 1; gray regiornscorresponds to chequered parts, starting with white on the cells cutted by π/ rooted on the boundary, with the exception of a singletree which instead contains both i and j , and the pathon the tree from i to j is directed consistently.So, a possible approach by combinatorial bijectionscould be to prove that, for any j , there is a suitable cor-respondence among the forests as above, and a number q j of copies of the original ensemble of rooted forests.Such a task is easily performed, even for a generic (ori-ented) graph, for what concerns the frame identity Id f ,for which all charges q j are 1. Unfortunately, for whatconcerns Id r , and with an eye to the proof performedin the following section, it seems difficult to pursue thisproject at least in the case of the pseudo-Manhattan lat-tice on a square geometry. Indeed, if the relaxation of k L Id f to Id r requires t j topplings on site j , it is easy tosee that q j ( Id r ) = k L − t j , and from the explicit expres-sions for k L and the values of the topplings (the latterare in the following Theorem 4) we see that the values ofthe charges q j are integers of order L . PROOF OF THE THEOREM
Here we perform the direct proof of Theorem 1. It isfully constructive, a bit technical, and maybe not spe-cially illuminating for what concerns the algebraic as-pects of the problem, but still, it makes the job.Now, in order to better exploit the geometry of oursquare lattice, we label a site through a pair ij denot-ing its coordinates. We can introduce the matrix T ( L ) ij ,which tells how many topplings site ij performed in therelaxation of k L Id f into Id r . Exploiting the rotationalsymmetry, we take it simply L × L instead of 2 L × L ,with ( i, j ) = (1 ,
1) for the site at the corner, analogouslyto what we have done for ( Id ( L )r ) ij . Of course, althoughwe call T ( L ) and Id ( L )r “matrices”, they have a singlesite-index, and are indeed vectors, for example, under the action of ∆.Clearly, T is a restatement of Id r , as( Id ( L )r ) ij = k L b + ij − ∆ ij,i ′ j ′ T ( L ) i ′ j ′ (17)so that Id r is determined from T , but also vice-versa, as∆ is invertible. Actually, through the locality of ∆, onecan avoid matrix inversion if one has some “boundarycondition” information on T , and the exact expressionfor Id r , e.g. if one knows T on two consecutive rows andtwo consecutive columns (and we have a guess of thiskind, as we discuss in the following).The constraint that ( Id r ) ij ∈ { , } gives that T is locally a parabola with small curvature. More-over, in the regions corresponding to homogeneousportions of Id r , T must correspond to a discretizedparabola through easy formulas. The telescopic nature of( Id ( L )r , Id ( ⌊ L − ⌋ )r , . . . ) has its origin in an analogous state-ment for ( T ( L ) , T ( ⌊ L − ⌋ ) , . . . ), and on the fact that k L hasa simple formula. As we will see, these facts are easier toprove.We start by defining a variation of T which takes inaccount explicitly both the height-1 square in the middleof Id r , and the spurious effects on the border. Define˜ k ( L ) = ⌊ ( L − L + 2) / ⌋ (which is approximatively k L / T ( L ) ij := T ( L ) ij − ˜ k ( L ) b + ij − (cid:0) L − i +12 (cid:1) − (cid:0) L − j +12 (cid:1) . (18)A first theorem is that Theorem 3 ˆ T ( L ) iL = ˆ T ( L ) Lj = 0 ; (19a)ˆ T ( L ) ij = 0 for i + j > L + ℓ ; (19b)ˆ T ( L ) ij ≤ for all i , j . (19c) Figure 6: The recurrent identity on the pseudo-Manhattanlattice of side 50 (up) and 51 (down), an example of how theidentity on side 2 L + 1 is trivially deduced from the one onside 2 L . This implicitly restates the claim about the middlesquare of height 1 in Id r , and is in accord with the upperbound on the curvature of T given by ( Id r ) ij ≤
1. Withabuse of notations, we denote by ˆ T ( ℓ ) ij also the L × L matrix corresponding to ˆ T ( ℓ ) ij in the ℓ × ℓ corner with i, j ≤ ℓ , and zero elsewhere. Then the rephrasing of thefull Theorem 1 is Theorem 4 If M ( L ) ij = − ( ˆ T ( L ) L − i L − j − ˆ T ( ℓ ) L − i L − j ) for s =1 , and the transpose of the latter for s = 2 , defin-ing θ ( n ) = 1 for n > and 0 otherwise, and the“quadratic + parity” function q ( n ) on positive integers q ( n ) = θ ( n ) (cid:18) n + 2 n + 1 − ( − n (cid:19) = n ≤ n + 1) / n positive odd; n ( n + 2) / n positive even; (20) then M ( L ) is a deterministic function, piecewise “qua- dratic + parity” on a finite number of triagular patches M ( L ) ij = θ ( i − L + ℓ ) h − (cid:0) j (cid:1) − max(0 , j − ℓ − q ( j − i + L − ℓ ) − q ( j + i − L − ℓ − i + θ ( j − L + ℓ ) h − (cid:0) i (cid:1) − θ ( L − ℓ − i + 1)+ q ( i − j + L − ℓ −
1) + q ( i + j − L + ℓ − q ( i + j − L − ℓ − i . (21)Clearly Theorem 3 is contained in Theorem 4, just bydirect inspection of the summands in (21). The equiva-lence among Theorems 1 and 4 is achieved still by directinspection of (21), with the help of some simple lem-mas. First remark that the use of − ˆ T instead of T makes us work “in false colours”, i.e. effectively inter-changes z into 1 − z in Id r . Then, defining ∇ x f ( i, j ) := f ( i + 1 , j )+ f ( i − , j ) − f ( i, j ), and analogously ∇ y with ± j , we have that ∇ x,y q ( ± i ± j − a ) = ( ± i ± j − a ≥ ∇ x max(0 , j − a ) = 1 at j = a only, while ∇ y max(0 , j − a ) = 0 always (this reproduces the ℓ extrazeroes out of the triangles depicted in figure 5).So, the explicit checks above can lead to the conclusionthat equation (17) holds at every L for the expressionsfor T and Id r given, in a special form, i.e. subtractingthe equation for size L with the one for size ℓ , but on the L × L system. The latter is then easily related to the oneon the ℓ × ℓ system, at the light of equation (19a), whichallows to state that there are no different contributionsfrom having splitted the four quadrants of ˆ T ( ℓ ) . Thus,equation (19a), guessing the exact expression for T ij onthe edges of each quadrant by mean of a simple formula,is crucial to the possibility of having the telescopic recon-struction procedure.In other words, even without doing the tedious checks,we have seen how, by a series of manipulations corre-sponding to the subtraction of ℓ × ℓ corners to the L × L matrices, the proof is restricted to the analysis of the“deterministic” part, with i > ℓ or j > ℓ . As, in thiscase, all the involved functions depend on L through itscongruence modulo 2, or 3, or 4, it suffices to check nu-merically the theorems for 12 consecutive sizes in orderto have that the theorem must hold for all sizes. We didthe explicit check for sizes up to L = 64.The only thing that we need in order to conclude thatthe conjectured expression Id r corresponds to the iden-tity is to prove that it is indeed a recurrent configura-tion. Again, we do that in two steps, in order to dividethe behaviour on the self-similar ℓ × ℓ corner and on thedeterministic part.First, remark that in Id ( L )r the internal border of thequadrant (the sites ij with i or j = L ) is burnt in theBurning Test even without exploiting the other mirrorimages. This is true because, at every L and on bothborders, we have a substructure of the formwhich is burnt through the following cascade (we put theburning times, and denote with arrows from x to y atoppling occurring in x which triggles the toppling in y )
13 13 13 1312 12 1210 10 10 1411 11 11 16 17 20 2115 18 19 229 9 9 138 8 126 67 75 542 31 ······
We choose to show just an example, instead of giving theexplicit formula, but it should be clear from the picturesabove that a general- L regular procedure exists.So, Id r satisfies the burning test if and only if the de-terministic part of Id r (the L × L square minus the ℓ × ℓ corner) satisfies the burning test with the border of thecorner having a b ij = 1 every two sites. This check mustbe performed only on the deterministic part, and again isdone in a straightforward way, or, with conceptual econ-omy, implied by the explicit numerical check on 6 con-secutive sizes. This completes the proof of all the threetheorems. (cid:3) MANHATTAN LATTICE
For the Manhattan lattice, we performed an investiga-tion similar to the one above for the pseudo-Manhattanone, although, as we already said, the motivations areless strong, as this lattice is not planar alternating. Asmall-size example of Creutz identity, compared to thepseudo-Manhattan one, is shown in Figure 7.The numerics gave positive and negative results. Thepositive results concern the filling number k L , that, ac-cording to extensive tests (up to L ≃ k L = ( L ( L + 2) L even ( L + 1) L odd (23)Furthermore, the whole quadrant except for the ℓ × ( ℓ − Figure 7: The recurrent identity on a portion of side 50 of thesquare lattice, with pseudo-Manhattan (up) and Manhattan(down) orientation. of rules analogous to equations (11) (again, ℓ = ⌊ ( L +1) / ⌋ ), and, as in (11), depending from the congruenceof L modulo 3, and a transposition involved if s = 2.The negative result is that the telescopic exact self-similarity between side L and side ℓ seems to be lost inthis case. As we said, this property relies crucially on thesimplicity of the toppling matrix on the boundary of thequadrants, which seems to be an accidental fact of thepseudo-Manhattan lattice, and has few chances to showany universality. For this reason we did not attempt tostate and prove any theorem in the fashion of Theorem 1in this case. CONCLUSIONS
We have studied, numerically and analytically, theshape of the Creutz identity sandpile configurations, forvariants of the ASM, with directed edges on a squarelattice (pseudo-Manhattan and Manhattan), and squaregeometry. An original motivation for this study was thefact that heights are valued in { , } (while the original0BTW sandpile has heights in { , , , } ), and we conjec-tured that this could led to simplifications. The resultshave been even simpler than expected, and qualitativelydifferent from the ones in the BTW model.In the BTW model, the exact configuration seems to beunpredictable: although some general “coarse-grained”triangoloid shapes seem to have a definite large-volumelimit, similar in the square geometry and in the one ro-tated by π/
4, here and there perturbations arise in theconfiguration, along lines and of a width of order 1 inlattice spacing. We discuss the role of these structures invarious aspects of the ASM in a forthcoming paper [13].The triangoloids have precise shapes depending fromtheir position in the geometry, and are smaller andsmaller towards the corners. Understanding analyti-cally at least the limit shape (i.e. neglecting all the sub-extensive perturbations of the regular-pattern regions) isa task, at our knowledge, still not completed, althoughsome first important results have been obtained in [18],and further achievements in this direction have come withthe work of Levine and Peres [16, 17], both in the similarcontext of understanding the relaxation of a large pile ina single site (see in particular the image at page 10 ofLevine thesis, and the one at http://math.berkeley.edu/ ˜ levine/gallery/invertedsandpile1m10x.png ,and, for the directed model, the one at page 20 of newton.kias.re.kr/ ˜ nspcs08/Presentation/Dhar.pdf ).In our Manhattan-like lattices on square geometry,however, we show how the situation is much simpler, anddrastically different. Triangoloids are replaced by exacttriangles, all of the same shape (namely, shaped as half-squares), and with straight sides. All the sides of thetriangles are a fraction − k of the side of the lattice (inthe limit), where the integer k is a “generation” index de-pending on how near to a corner we are, and indeed eachquadrant of the configuration is self-similar under scal-ing of a factor 1 /
3. The corresponding “infinite-volume”limit configuration is depicted in Figure 8. A restatementof the self-similarity structure, in a language resemblingthe z → /z conformal transformation in Ostojic [18]and Levine and Peres [16], is the fact that, under themap z → ln z , a quadrant of the identity (centered at thecorner) is mapped in a quasi–doubly-periodic structure.Also the filling numbers (i.e. the minimal number offrame identities relaxing to the recurrent identity) havesimple parabolic formulas, while in the original BTWmodel a parabola is not exact, but only a good fittingformula.These features reach the extreme consequences in thepseudo-Manhattan lattice, where the exact configura-tion at some size is deterministically obtained, througha ratio-1 / ln 3 π Figure 8: Up: the limit Creutz Identity configuration on ourManhattan and pseudo-Manhattan lattices. White regionscorrespond to have height 1 almost everywhere. Gray regionscorrespond to have height 0 and 1 in a chequered patternalmost everywhere. Down: the image of a quadrant underthe map z → ln z . natorial way. Acknowledgements
We are grateful to an anonymous referee for many use-ful comments, and in particular for signaling us a veryrecent paper by Dhar and collaborators [19] in which thesame directed variants of the ASM as in our paper areconsidered, for the related problem of understanding therelaxation of a large pile of sand in a single site. This isa further motivation for our work, and is quite likely tobe a precious reference for future research.1 [1] P. Bak, C. Tang and K. Wiesenfeld,
Self-organized criti-cality: An explanation of the /f noise, Phys. Rev. Lett. , 381-384 (1987).[2] D. Dhar, Self-organized critical state of sandpile automa-ton models,
Phys. Rev. Lett. , 1613-1616 (1990)., The abelian sandpile and related models,
PhysicaA arXiv:cond-mat/9808047 [3] S.N. Majumdar and D. Dhar,
Height correlations in theAbelian sandpile model,
J. Phys. A L357-L362 (1991),
Equivalence between the Abelian sandpile model andthe q → limit of the Potts model, Physica A
Algebraicaspects of Abelian sandpile models,
J. Phys. A , 805-831(1995), arXiv:cond-mat/9408022 [5] M. Creutz, Abelian Sandpiles,
Comp. in Physics Abelian Sandpiles,
Nucl. Phys. B (Proc. Suppl.) Fractals and self-organized criti-cality,
Fractals in science, A. Bunde and S. Havlin eds.,26-47 (Springer-Verlag, 1994).[6] A.D. Sokal,
The multivariate Tutte polynomial (aliasPotts model) for graphs and matroids, arXiv:math/0503607 [7] S. Caracciolo, J.L. Jacobsen, H. Saleur, A.D. Sokal andA. Sportiello,
Fermionic field theory for trees and forests,
Phys. Rev. Lett. arXiv:cond-mat/0403271 [8] S. Caracciolo, C. De Grandi, A. Sportiello, Renormal-ization flow for unrooted forests on a triangular lattice,
Nucl. Phys. B arXiv:0705.3891 [9] V.B. Priezzhev,
Exact height probabilities in the Abeliansandpile model,
Phys. Scr.
T49B , 663-666 (1993).,
Structure of two-dimensional sandpile. I. Heightprobabilities,
J. Stat. Phys. , 955-979 (1994).[10] S. Mahieu, P. Ruelle, Scaling fields in the two-dimensional abelian sandpile model,
Phys. Rev. E arXiv:hep-th/0107150 P. Ruelle, A c = − boundary changing operator for theAbelian sandpile, Phys. Lett. B arXiv:hep-th/0203105
M. Jeng,
Conformal field theory correlations in the Abe-lian sandpile model,
Phys. Rev. E arXiv:cond-mat/0407115 S. Moghimi-Araghi, M.A. Rajabpour, S. Rouhani,
Abe-lian Sandpile Model: a Conformal Field Theory point ofview,
Nucl. Phys. B arXiv:cond-mat/0410434
G. Piroux and P. Ruelle,
Pre-logarithmic and logarith-mic fields in a sandpile model,
J. Stat. Mech.
P005(2004), arXiv:hep-th/0407143 , Boundary height fields in the Abelian sandpilemodel,
J. Phys. A arXiv:hep-th/0409126 , Logarithmic scaling for height variables in theAbelian sandpile model,
Phys. Lett. B arXiv:cond-mat/0410253 [11] M. Jeng, G. Piroux and P. Ruelle,
Height variables in theAbelian sandpile model: scaling fields and correlations,
J.Stat. Mech.
P015 (2006), arXiv:cond-mat/0609284
V.S. Poghosyan, S.Y. Grigorev, V.B. Priezzhev, P. Ru-elle,
Pair correlations in sandpile model: a check of loga-rithmic conformal field theory,
Phys. Lett. B arXiv:0710.3051 [12] M. Creutz,
Xtoys: cellular automata on xwindows,
Nucl.Phys. B (Proc. Suppl.) , 846-849 (1996) arXiv:hep-lat/9508029 code: [13] S. Caracciolo, G. Paoletti, A. Sportiello, Pseudo-prop-agators and other geometric structures in the AbelianSandpile, in preparation.[14] S. Caracciolo, M.S. Causo, P. Grassberger and A. Pelis-setto,
Determination of the exponent γ for SAWs on thetwo-dimensional Manhattan lattice, Journ. of Phys. A:Math. and Gen. (16), 2931-2948 (1999), arXiv:cond-mat/9812267 [15] J.T. Chalker, P.D. Coddington, Percolation, quantumtunnelling and the integer Hall effect,
J. Phys. C: SolidState Phys. Scaling Limits for Internal Aggrega-tion Models with Multiple Sources, arXiv:0712.3378 [17] L. Levine, PhD dissertation thesis, Univ. California atBerkeley, fall 2007, arXiv:0712.4358http://math.berkeley.edu/ ˜ levine/levine-thesis.pdf [18] S. Ostojic, Patterns formed by addition of grains to onlyone site of an abelian sandpile,
Physica A
Pattern formationin growing sandpiles, arXiv: 0808.1732arXiv: 0808.1732