Pattern Formation and Tropical Geometry

Sandpile models exhibit fascinating pattern structures: patches, characterized by quadratic functions, and line-shaped patterns (also called solitons, webs, or linear defects). It was predicted by Dhar and Sadhu that sandpile patterns with line-like features may be described in terms of tropical geometry. We explain the main ideas and technical tools—tropical geometry and discrete superharmonic functions—used to rigorously establish certain properties of these patterns. It seems that the aforementioned tools have great potential for generalization and application in a variety of situations.


PATTERN FORMATION AND CELLULAR AUTOMATA
Animals show beautiful skin and wing patterns. Explaining how these come about has been a longstanding puzzle. In line with the Darwinian paradigm, an evolutionary biologist may suggest that formations of patterns on the skin of animals are visual traces of certain biological mechanisms that help survival in terms of natural selection.
In his seminal book, however, Thomson [1] argues that the geometry of patterns may be mainly dictated by chemical forces, albeit it is known that patterns may benefit their owners in certain cases. In his famous paper on morphogenesis [2], Turing speculated on the mechanism behind pattern formation on the skin of animals and proposed the famous reaction diffusion system, which consists of an inhibitor and an activator with different diffusion rates. Historically, this was the first explicit model of pattern formation, though it is purely theoretical.
An important example of self-oscillating patterns in the real world was discovered soon after, see the Belousov-Zhabotinsky reaction [3,4]. Both the Turing model and the Belousov-Zhabotinsky reaction produce beautiful spatiotemporal patterns with quasi-ordered strips and spots, see a popular concise exposition of both topics in Ball [5]. A recent discussion about parallels in emerging complexities and patterns in biological systems and physical glass-like models can be found in Wolf et al. [6].
A possibility to obtain all sorts of patterns starting from local interactions suggests trying relatively simple models to explore patterns by pure or computer mathematics. Assuming that on a big scale all coarse grained functions are smooth and continuous, one may use differential equations in the study of patterns; see comprehensive reviews [7,8]. But, appealing to the discrete nature of pattern formation, we shift our attention to cellular automata.
In this article we focus on so-called sandpile models, and firstly, discuss in section 2 how the patterns in that model were obtained in experimental computer physics, and secondly, we survey the main ideas permitting to study these patterns with mathematical rigor: discrete harmonic analysis (section 3), tropical geometry (section 4), toppling function (section 5), and the most technical part of the proofs, the lower bound (section 6). We mention open problems and new research directions when appropriate.

Definitions
A sandpile model, which we consider here, consists of the standard integer lattice grid inside a compact convex domain ⊂ R 2 , i.e., the graph Ŵ = ∩ Z 2 .
A state of a sandpile model is a function ϕ(i, j), representing the number of grains of sand at the vertex (i, j) ∈ Ŵ. A state, thus, is an integer-valued function ϕ : Ŵ → Z ≥0 .
A vertex (i, j) is unstable if there are four or more grains of sand at (i, j), i.e., ϕ(i, j) ≥ 4. The evolution proceeds as follows: any unstable vertex (i, j) topples by sending one grain of sand to each of its four neighbors (i+1, j+1), (i+1, j−1), (i−1, j+1), (i− 1, j − 1). The sand, falling outside , disappears from the system. Vertices outside of (formally they do not even belong to Ŵ) and stable vertices are never toppled. Equivalently, one may think that all the lattice points outside of are sinks; sand, falling to sinks, disappears. Given an initial state ϕ, the state ϕ • denotes the stable state reached after all possible topplings have been performed. It is a classical fact that ϕ • does not depend on the order of topplings [19,20].

Line-Shaped Patterns in the Literature
Line-shaped patterns [which can also go under different names: solitons, linear defects, and (p, q)-webs] can be found in the pictures in Liu et al. [21] and Ostojic [22], but the main subject of the latter article was quadratic patches, recently explained in Pegden and Smart [23,24] and Levine at al. [25,26] using Apollonian circle packings.
Let us put three grains in all the vertices of the intersection between the standard grid Z 2 and a planar domain . Let us choose several vertices and add one more grain to each of them. An example of the relaxation of such a state for being a square is shown in Figure 1. We sequentially drop grains to blue points, performing a relaxation after each dropping (thus we have one blue point on the first pictures and four blue points on the fourth picture, where blue points indicate the positions of additional grains). One may easily guess a graph with straight edges along non-white parts of the pictures. With each new blue point such a graph changes, but its edges always pass through the blue points. In Figure 2 we added grains to all blue points simultaneously and took snapshots of the subsequent relaxation (since the order of topplings does not influence the final picture, we might add grains sequentially).
Line-shaped patterns, clearly recognizable in Figures 1, 2 along straight edges of the imaginable graph, explicitly came in the sight of researchers in Dhar et al. [27] and Dhar and Sadhu [28] with an emphasis on a proportional growth phenomenon, and later, in Caracciolo et al. [29] (see also [30,31]). These papers performed the analysis from the point of view of theoretical physics and explained the pictures based on experimental evidence.
The use of tropical geometry was predicted in Sadhu and Dhar [32] and later implemented with rather involved mathematical proofs in a series of articles [33][34][35]. A certain limit of the sandpile model gives a continuous limiting piece-wise linear model that also exhibits power-law behavior [36]; the statistical properties of the latter model can be found in Kalinin and Prieto [37]. Line-shaped patterns may be viewed as spacetime diagrams of two incoming particles that join to form one particle. It turns out that we can associate the "energy of the particle" with each world line such that total energy is conserved in these collisions. As it was recently shown (experimentally), quadratic patches may be thought of as a limit of many line-shaped patterns coming together during a relaxation [38].

Our Main Problem: Small Perturbation of the Maximal Stable State
Let us formalize the observations in Figure 1 in the following way. Let be a non-degenerate compact domain with nonempty interior and P be a finite non-empty subset of . For each N ∈ N consider a set Ŵ N = ∩ 1 N Z 2 , i.e., the intersection of with the lattice of mesh 1 N . Define the state ϕ N = (3 + p∈P δ p ) • on Ŵ N , i.e., we put three grains everywhere on Ŵ N and dropped one additional grain to each of the points p ∈ P or to a near vertex in Ŵ N if p / ∈ Ŵ N , and then we performed a relaxation. Define the deviation set Experimental evidence suggests that when N grows, the sets C N ⊂ converge to a thin balanced graph (see Figure 1). Theorem 1. announced in [33] and proven in [33][34][35] The sequence of sets C N ⊂ converges (in the Hausdorff sense) to a setC ,P . The setC ,P is a planar graph passing through the points P. Each edge ofC ,P is a straight segment with a rational slope.
At least in this setting, we have proven that the limiting/asymptotic patterns exist, though there is no close description of the shape and exact amount of grains for the pattern in the direction (p, q) ∈ Z 2 .
More intricate instances of the line-shaped patterns, namely, in the identity element of the sandpile group of , await an explanation. Recall that the recurrent states of a sandpile model on a given graph form an Abelian group, the sandpile group of a graph [19,39]. In Figure 3, the identity of the sandpile group for a cylinder consists of linear-shaped patterns. In Figure 4, we see similar patterns (along with quadratic patches) on the identity of the sandpile group for a non-convex domain. It has not yet been mathematically proven that the sandpile identity of such graphs indeed contains these linear patterns. A remarkably simple pattern for the identity on a circular base was found by Melchionna [40] [see Figures 7,10 in [40]], the sandpile identity on ellipses of certain type consists of a unique pattern, up to "linear defects." In the further text we briefly explain the main ideas behind the proof of Theorem 2. All the details of the actual proofs, FIGURE 1 | is a square [0, 100] × [0, 100], we put 3 grains at every vertex of the lattice inside and drop 1 additional grain to each blue point (sequentially). The result of the relaxation after one additional grain is on the top-left picture. Then we added a grain of sand to the second blue point and performed a relaxation, the result of which is the top-right picture. The third picture is the bottom left one, and the last one is the bottom right one. White is three grains, green is two, yellow is one, and red is zero. Crosses mark the sinks in the model. exact notation, omitted conditions, etc., can be found in Kalinin and Shkolnikov [33][34][35]. We believe that our tools are worth generalizing for other situations and can be used to prove the aforementioned appearance of linear patterns in different setups.

DISCRETE HARMONIC ANALYSIS
The laplacian F of a function F : Z 2 → Z is defined as Recall the Liouville theorem: a non-negative harmonic function on Z 2 must be a constant (for several proofs see Theorem 9.24 in [41]).
Assume that Ŵ is an intersection of a big convex subset ⊂ R 2 with Z 2 . Fix an arbitrary linear function L : Z 2 → Z. The following lemmata are close in spirit to Buhovsky et al. [42] where an improvement of the Liouville theorem is presented.

Lemma 1.
A positive integer-valued harmonic function F which is less than L on a large enough subdomain Ŵ ′ of Ŵ, is linear itself   195]. Let sinks be the vertices with less than four neighbors. The picture presents the identity of the sandpile group of this graph. White is three, green is two, red is one, and blue is zero. If we would take a flat cylinder, i.e., we would identify (x, 0) with (x, 12), then, in the picture for the identity of the sandpile group, we would have 3 everywhere except several green vertical lines, i.e., columns with two grains. on a (smaller, but still large) subdomain Ŵ ′′ of Ŵ ′ , i.e., there exists Lemma 2. Fix a constant c > 0. Consider a positive integervalued superharmonic function F such that the sum of its laplacian at points in Ŵ ′ ⊂ Ŵ linearly depends on the diameter of Ŵ ′ with an a priori bound c, i.e., Then, if F is less than L and the domain is large enough then F is linear itself on a large subdomain Ŵ ′′ ⊂ Ŵ ′ .
In other words, given an upper estimate of a natural-valued function F by a linear function L, we may deduce that F is linear on a large subdomain provided F is harmonic or almost harmonic. Precise formulations can be found in Kalinin and Shkolnikov [34]. The two main ideas used in the proofs are as follows.
• The green function (harmonic at all points except one) on the plane grows as the logarithm. • For a positive discrete harmonic function F on a ball of radius R with the center O, the discrete derivative of F at O (i.e., is at most the maximum of F on the ball, divided by R. If F has only integer values, then We conjecture that the line-shaped patterns show up in the relaxation of a perturbation of the maximal stable state in the sandpile model on a certain graph, if there is a notion of a linear function on such a graph and both lemmata above hold. To perform a "scaling" one needs a graph that is self-similar on different scales, such as Z 2 . A natural candidate is a Cayley graph of a group. Recall that given a group G and a set S of its generators, one may construct the so-called Cayley graph of G, whose vertices are elements of G and two vertices u, v ∈ G are connected by an then the Cayley graph is the standard grid Z 2 with all vertices of valency four. If G = Z 2 , S = {(1, 0), (0, 1), (1, 1)} then the Cayley graph is Z 2 and each vertex (i, j) is connected by edges On a Cayley graph, the generators of the group play the role of coordinates [modulo relations, as (1, 0) + (0, 1) = (1, 1) in the example above] so the notion of a linear function can be easily extended. The discrete harmonic function theory is quite developed for several classes of groups [43][44][45][46].
Question. Do Lemmata 1,2 hold for the harmonic and superharmonic functions on Cayley graphs of amenable groups? If yes, then, under an appropriately chosen scaling procedure, one should be able to prove convergence of the deviation sets of the relaxations of a slightly perturbed maximal stable states (on a big bounded polygonal-shaped part of the Cayley graph) to the corner locus of a piecewise linear function on the scaling limit of these polygonal shaped parts of the Cayley graphs. The first thing to prove is that the toppling function has a piecewise linear estimate from above.
The Cayley graphs of abelian groups are composed of Z k and cylinders as in Figure 3. The simplest non-abelian group, which is not much different from Z 3 , is the Heisenberg group. Consider a maximal stable state (i.e., 5 grains at every vertex) and add one grain to several vertices. One expects that the relaxation of such a state on Ŵ should be not a very complicated "extension" of a relaxation of a perturbation of the maximal stable sandpile on domains in Z 2 .

TROPICAL CURVES
A tropical polynomial is a piecewise linear function f : R 2 → R of the form f (x, y) = min{ix + jy + a ij |(i, j) ∈ A}, Frontiers in Physics | www.frontiersin.org where A is a finite subset of Z 2 and a ij ∈ R are coefficients. Each term ix + jy + a ij is called a monomial and should be thought of log(t a ij x i y j ), f should be thought of the limit of a certain logarithmic rescaling of To each tropical polynomial f there is a corresponding tropical curve C(f ), which, by definition, is the corner locus of f , i.e., the set of points (x, y) where f is not smooth. An equivalent definition follows. More on algebra-geometric aspects of tropical curves can be found in Brugallé et al. [47], Itenberg and Mikhalkin [48], and Maclagan and Sturmfels [49] along with recent applications in symplectic topology [50][51][52][53]. In this set-up, tropical curves should be thought of Riemann surfaces, and each vertex A of a tropical curve corresponds to a small surface S A with the boundary, the valency of A is equal to the number of the boundary components of S A , and each edge AB of the tropical curve corresponds to a very long thin cylinder connecting small surfaces S A and S B . Unfortunately, we found no connection between tropical curves in sandpiles and tropical curves in algebraic or symplectic geometry.

Tropical Series
Pick a convex compact set ⊂ R 2 with non-empty interior. Let P be a finite subset of . [35] an -tropical series is a piecewise linear function in given by the following:

Definition 4. Kalinin and Shkolnikov
where the set A is not necessarily finite and F| ∂ = 0. See an example in Figure 5.
Consider the family F P of -tropical series that are not smooth at every point of P.
Note that all functions in F P are concave and thus superharmonic. Let F P be the pointwise minimum of functions in F P . In Kalinin and Shkolnikov [35] it is proven that this pointwise minimum exists (that is easy) and belongs to F P (a bit more involved, because it may be not continuous or not a tropical series).
For each F ∈ F P we may consider the set It is easy to see that C(F) is the corner locus of the function F, i.e., exactly those points where F is not linear but changes its slope. The set C(F) is called the -tropical curve defined by F, and C(F) is a planar graph with straight edges of rational directions, the sum of directions of outgoing edges is zero for every vertex, and this is called the balancing condition.

Theorem 2. (elaboration)
The sequence of sets C N ⊂ converges (in the Hausdorff sense) to the -tropical curve C(F P ).
Let be a disk {x 2 + y 2 ≤ 1}. An example of an -tropical series is min{ix + jy + a ij |(i, j) ∈ Z 2 } with |a ij | = i 2 + j 2 is presented on the left in Figure 5, and its corresponding -tropical curve, which is an infinitely branching tree, is presented on the right. See details in Kalinin and Shkolnikov [54]. Question. The sum of the values of the above -tropical series for a circle (or other plane curve) gives interesting formulae in number theory [54] which are related to Mordell-Tornheim and Witten zeta functions [55,56]. These formulae take as input the coefficients of the equations of the tangent lines to a given plane curve, and they are thus easy to compute and may provoke interesting questions in the experimental computer mathematics. However, no analogs of these formulae for three dimensional bodies are known.

TOPPLING FUNCTION
To understand the appearance of tropical geometry in sandpiles, consider the toppling function H(v) defined for every v in Ŵ N as follows: given an initial state ϕ on Ŵ and its relaxation ϕ • , the value H(v) equals the number of times that the vertex v toppled in the process taking ϕ to ϕ • .
The toppling function is clearly non-negative on Ŵ and vanishes at the boundary of Ŵ. The Laplacian H of H completely determines the final state ϕ • by the formula [22]: It can be shown by induction that the toppling function H satisfies the Least Action Principle [57,58]: if ϕ(v) + F(v) ≤ 3 is stable, then F(v) ≥ H(v). Ostojic noticed that H(i, j) is a piecewise quadratic function if we drop a lot of sand in the origin of the otherwise empty plane [22].

Piecewise Linearity of the Toppling Function in Our Main Problem
Consider a state ϕ P , which consists of three grains of sand at every vertex, except at a finite family of points P = {p 1 , . . . , p r } where we have four grains of sand: The state ϕ • and the evolution of the relaxation can be described by means of tropical geometry. This was discovered in Caracciolo et al. [29]. The crucial (experimental) observation is that the toppling function H of the state ϕ is almost harmonic everywhere since ϕ • = ϕ almost everywhere (see Figure 1). Even better, in this case the toppling function H is piecewise linear on the most part of and the line-shaped patterns belong to a finite neighborhood of the corner locus of H (in the next section we give a more detailed statement). It is easy to observe but tricky to prove. We provide an upper bound H u and a lower bound H l for H, which are close to H. These tight bounds force the set to belong to a small neighborhood of the set The upper bound H u is a piece-wise linear function, ϕ is equal to 3 everywhere except a small set P of points, the laplacian of a function is zero on the domains of its linearity. The set H u (v) = 0 (the corner locus of a piece-wise linear function) is thus close to the set H = 0, the deviation locus of ϕ • . This concludes the proof of the theorem.

Upper Bound for the Toppling Function
Denote by H(ϕ N ) the toppling function of the state ϕ N = 3 + δ p i on Ŵ N . Abusing notation we will write H(x, y) = 1 N H(ϕ N )(x, y) : → R for the rescaled toppling function without specifying N. Consider the pointwise minimal function F P in F P . Then F P ≥ H(ϕ N ) by the Least Action Principle (since F P ≤ 0, F P (p i ) < 0 for each i). Thus, F P ≥ H. Corollary. The total defect v∈Ŵ N (3 − ϕ • (v)) grows linearly in N.
Indeed, the total defect is equal to the amount of the sand fallen outside of the system, which, in turn, is equal to the sum of H(ϕ N ) near the boundary, which can be estimated as N · ∂ 1/N NF P , i.e., N times the integral of F P over the 1/Nneighborhood of the boundary of .
In order to study the dependence of the deviation set {ϕ • = 3} on and P (positions of points where we added grains), one may study F P because it determines the tropical curve. The dependence of F P on P is in no sense continuous: when P passes through degenerate configurations (e.g., several points on a vertical line), F P and the corresponding tropical curve drastically change. Similar phenomenon appears when we keep P fixed and change : no meaningful results about stability of the resulting picture are known.

LOWER BOUND. WAVE OPERATORS
Let ϕ be a sandpile state on a graph Ŵ. Given a fixed vertex p ∈ Ŵ, we define the wave operator W p acting on a sandpile state ϕ as the following: where T p is the operator that topples once the state ϕ at p if it is possible ( [59][60][61]) (see Figure 6). In a computer simulation, the application of this operator looks like one wave of topplings spreading from p, while each vertex topples at most once.
The first important property of W p is that, for the initial state ϕ : = 3 + δ P , we can achieve the final state ϕ • by successive applications of the operator W p 1 • · · · • W p r a large but finite number of times (we write ∞ in spite of the notation): This is not a deep theorem but a rather useful description of a relaxation. We thus decompose the total relaxation ϕ → ϕ • into layers of controlled avalanching These layers, in turn, can be described by means of tropical geometry. We only need to prove that the linear-shaped patterns, visible in the pictures, move toward the point where we apply a wave operator.

Construction of Solitons
For each direction (p, q) ∈ Z 2 , gcd(p, q) = 1 we construct a function F p,q whose laplacian coincides with the linear pattern in the direction (p, q). To do that, consider a functioñ F : Z 2 → Z,F(x, y) = min(0, qx − py).
Frontiers in Physics | www.frontiersin.org FIGURE 6 | For a given sandpile state we apply several times the wave operator at the blue point p. One can see that the line-shaped pattern around p point creeps toward p until it belongs to one on them. Recall that white cells contain 3 grains, so the deviation set is the green, yellow, and red cells, which belong to a small neighborhood of a certain tropical curve. Let us present this tropical curve as a corner locus of a pointwise minimum of several linear functions (cf. Figure 5), i.e., as an -tropical series F. The planar graph is then the projection of edges of a three-dimensional polytope (the graph of F). Then the action of the wave operator corresponds to shifting one of the faces of this polytope, i.e., increasing by one the constant coefficient of the linear function, defining this face. On the level of planar graphs, we take the linear function in F (see Equation 1), which is the minimal at p, and increase its constant coefficient until p belongs to the corner locus of the new piecewise linear function.
Note that the corner locus l (the set of points in R 2 where min(0, qx − py) is not smooth) ofF is a line of direction (p, q).
Next, consider all the integer-valued superharmonic functions on Z 2 , which coincide withF outside of a finite neighborhood of l.
A non-trivial fact is that there exists a pointwise minimum F p,q among this family of functions [34]. The idea of the proof is as follows: instead of taking the pointwise minimum at once, we first prove that we can achieve it by "smoothings, " namely, by a sequence of stepsF = F 0 → F 1 → F 2 → . . . ; in each step F k → F k+1 we subtract the characteristic function of a certain set in a finite neighborhood of l (thus, a kind of inverse operator to the wave operator) and 0 ≤ F k − F k+1 ≤ 1. Then, sinceF is periodic, we may factor the plane by the action of the vector (p, q) and reduce the problem to a cylinder.
Then we use lemmata about superharmonic functions: if it would be possible to perform smoothings an infinite number of times, then Lemma 1 would imply that F k is linear with integer slope in a compact neighborhood of l, hence there exists a linear function with integer slope which is less thanF only on a finite neighborhood of the corner locus. This function would be periodic with respect to the shift on (p, q) and would therefore be like k(qx − py) + c (since gcd(p, q) = 1), but any such function (with an integer k) is less thanF outside of a finite neighborhood of l, which is a contradiction. Then we cannot perform smoothings an infinite number of times, and thus there is a pointwise minimum in the aforementioned family.
Once proved that the pointwise minimum exists, we may define solitons.

Definition 4.
A soliton [a linear-shaped pattern in the direction (p, q)] is ϕ pq = 3 + F p,q .
Then, from the Least Action principle and the minimality of F p,q it easily follows that sending a wave from one side of the deviation set of ϕ pq translates it [i.e., for certain p ′ , q ′ we have W x ϕ pq (i, j) = ϕ pq (i + p ′ , j + q ′ ) for all (i, j)], otherwise not changing. This is why we call them solitons.
The same can be done for three solitons of direction (p 1 , q 1 ), (p 2 , q 2 ), (p 3 , q 3 ), meeting at a point, provided that p i = q i = 0 and the triangle (p 1 , q 1 ), (p 2 , q 2 ), (p 3 , q 3 ) does not contain lattice points except vertices. The ideas of the proof are the same, we use Lemma 2 and the final step is that if a linear function px + qy is less than min(p 1 y − q 1 x, p 2 y − q 2 x, p 3 y − q 3 x) only in a compact neighborhood of the apex of the latter function, then (p, q) ∈ Z 2 must belong to the triangle (p 1 , −q 1 ), (p 2 , −q 2 ), (p 3 , −q 3 ), which is a contradiction.
We summarize the results as follows: there are certain functions f p,q,... ("at infinity" being described by piecewise functions, i.e., tropical functions), pointwise minimal in special families of superharmonic functions, such that 3 + f p,q,... models solitons, and three or four solitons coming to a point.
The crucial property of the wave operator W p is that its action on a state ϕ = 3 + f p,q,... has an interpretation in terms of tropical geometry; see the next section.

Tropical Wave Operators and the Lower Bound
Whenever, "at infinity" f is a piecewise linear function with integral slopes that, in a neighborhood of p, is expressed as where W(f ), another piece-wise linear "at infinity" function, has the same coefficients a ij as f , except one, namely a ′ i 0 j 0 = a i 0 j 0 + 1. This is to emulate the fact that the support of the wave (the set of vertices that toppled during the wave) is exactly the part of the plane where a i 0 j 0 + i 0 x + j 0 y is the leading part of f .
Consider an -tropical series f . We will write G p : = W ∞ p to denote the operator that "applies W p to 3 + f until p lies in the corner locus of f "; i.e., G p increases the coefficient a ij , corresponding to a neighborhood of p, by lifting the plane lying above p in the graph of f by integral steps until p belongs to the corner locus of G p f . Thus, G p has the effect of pushing the tropical curve closer toward p until it contains p (see Figure 6). From the properties of the wave operators, it follows immediately that (recall that F P is the upper bound): where 0 is the function which is identically zero on . Now we are ready to provide a lower bound in the main theorem.
Note that the upper bound can be obtained by (possibly infinite) series of applying tropical wave operators (which are nothing else but repetitive increasing of coefficients on linear parts in a piecewise linear function). Then, by the properties of the solitons, this can be emulated in the sandpile model, where wave operators are performed on the sandpile level, and instead of piecewise linear functions we have pictures as in Figure 6.
In other words, we choose an approximation of F P by a finite composition G p 1 G p 2 . . . of tropical wave operators, and we then choose an N big enough before starting from a state on Ŵ N with a tropical series and a collection of solitons, representing the corresponding -tropical curve. Then we perform the sandpile wave operators W p 1 W p 2 . . . as prescribed by tropical wave operators (see Figure 6). Since N is big enough we have full control on the picture and know that the solitons move exactly as edges of the tropical curve on the tropical pictures. By the nature of the construction, this will give us a lower bound for the relaxation of ϕ (constructed by the wave decomposition), which is close to the upper bound (given by an -tropical series) with any prescribed accuracy. The deviation set of ϕ • N consequently converges to the -tropical curve defined by F P .

DISCUSSION
We surveyed several mathematical tools which have been used for a concrete problem of sandpiles on a part of Z 2 . These tools may be generalized in several directions. One may consider sandpiles on other graphs, e.g., parts of the Cayley graphs for amenable groups. Also, one can take a part of hyperbolic tessellation [62] or other tiling of the plane [63] and ask similar questions about patterns and rescaling procedures.
It seems that the only tool to describe explicitly the picture for the sandpile identity is to compute the toppling function with high precision and controlled error. Above, we explained that the "smoothing" procedure allows us to show that there exists a pointwise minimal function in certain classes of superharmonic functions, and certain localization techniques (tropical geometry) could then be applied based of the properties of harmonic or almost harmonic functions on big domains with an explicit linear upper bound.
Tropical series for planar domains are connected to certain zeta functions. It would be nice to (at least, experimentally) compute series, similar to Kalinin and Shkolnikov [54], for higher dimensions with a good precision and guess what kind of numbers (e.g., polynomials in π if we start with a round sphere) will be obtained.
It would be interesting to find another decomposition of a relaxation into waves of higher magnitude, i.e., such a decomposition will allow us to control the change of not only linear-shaped pattern but the quadratic patches too.
Similar to Sadhu and Dhar [32] it would be nice to run the same research, and in particular, to establish continuity properties for the toppling functions of sandpiles on Cayley graphs and see whether one can get a kind of balancing conditions out of that.

DATA AVAILABILITY STATEMENT
The original contributions presented in the study are included in the article, further inquiries can be directed to the corresponding author.

AUTHOR CONTRIBUTIONS
The author confirms being the sole contributor of this work and has approved it for publication.