Self-Organized Criticality

Table of Contents

1. #What Is Self-Organized Criticality?|What Is Self-Organized Criticality?
   • #Ordinary Criticality vs. SOC|Ordinary Criticality vs. SOC
   • #The Three Empirical Hallmarks|The Three Empirical Hallmarks
   • #Natural Examples and Connection to Scaling Laws|Natural Examples and Connection to Scaling Laws
2. #The BTW Sandpile Model|The BTW Sandpile Model
   • #2.1 Formal Definition|2.1 Formal Definition
   • #2.2 Time-Scale Separation|2.2 Time-Scale Separation
   • #2.3 Avalanche Observables|2.3 Avalanche Observables
   • #2.4 The 1/f Noise Connection|2.4 The 1/f Noise Connection
3. #Why Self-Organized Criticality?|Why Self-Organized Criticality?
   • #3.1 The Attractor Argument|3.1 The Attractor Argument
   • #3.2 Connection to Absorbing State Phase Transitions|3.2 Connection to Absorbing State Phase Transitions
4. #The Abelian Sandpile (Dhar 1990)|The Abelian Sandpile (Dhar 1990)
   • #4.1 Setup on a General Graph|4.1 Setup on a General Graph
   • #4.2 The Abelian Property|4.2 The Abelian Property
   • #4.3 The Toppling Lemma|4.3 The Toppling Lemma
   • #4.4 Recurrent Configurations and the Sandpile Group|4.4 Recurrent Configurations and the Sandpile Group
5. #Scaling Theory and Exponents|Scaling Theory and Exponents
   • #5.1 Finite-Size Scaling|5.1 Finite-Size Scaling
   • #5.2 Scaling Relations|5.2 Scaling Relations
   • #5.3 Known Values for 2D BTW|5.3 Known Values for 2D BTW
   • #5.4 Universality|5.4 Universality
6. #References|References

What Is Self-Organized Criticality?

Ordinary Criticality vs. SOC

In equilibrium statistical mechanics, a critical point — for instance the Ising model at temperature \(T = T_c\) — requires fine-tuning a control parameter to a special value. The free energy landscape has a non-analytic point precisely at the critical temperature, and generically, the system is off-critical. Away from \(T_c\), the correlation length \(\xi\) is finite:

\[\xi \sim |T - T_c|^{-\nu}\]

and fluctuations are characterized by an intrinsic length scale \(\xi\). The divergence of \(\xi\) at \(T_c\) is what produces scale-free behavior (power-law correlations, critical opalescence, etc.), but this behavior is a set of measure zero in parameter space.

Self-organized criticality (SOC), introduced by Bak, Tang, and Wiesenfeld (1987), asserts that a broad class of complex systems — those with many slowly-driven, weakly-dissipative components — can spontaneously evolve to a critical state exhibiting scale-free behavior without any external fine-tuning of parameters.

The “self-organized” aspect is the key conceptual departure: the dynamics itself tunes the system to the critical point. The critical state is an attractor of the dynamics, not a fine-tuned boundary condition imposed by an external agent. The control parameter (analogous to \(T\)) is not fixed externally but is instead an emergent property of the stationary state reached by the dynamics under slow drive and weak dissipation.

The Three Empirical Hallmarks

A system exhibiting SOC is characterized by three signatures, all of which reflect the absence of a characteristic scale:

1. Power-law event-size distributions: the probability of an event (avalanche, earthquake, neural cascade) of size \(s\) satisfies

   \[P(s) \sim s^{-\tau}, \quad s \to \infty\]

   with no exponential cutoff. Compare to an exponential distribution \(P(s) \sim e^{-s/s_0}\), which has a characteristic scale \(s_0\) and arises generically away from criticality. The power law implies that large events are not exponentially rare — they are merely power-law suppressed.

2. \(1/f\) noise: the power spectral density of the system’s output signal \(\phi(t)\) satisfies

   \[S(f) = |\hat{\phi}(f)|^2 \sim f^{-\beta}, \quad \beta \approx 1\]

   This is “pink noise,” intermediate between white noise (\(\beta = 0\), flat spectrum, no temporal correlations) and Brownian motion (\(\beta = 2\), corresponding to a random walk). Pink noise implies temporal correlations at all scales — no characteristic time.

3. Fractal spatial geometry: the spatial footprint of events has no characteristic length scale and is statistically self-similar across scales. Formally, if \(M(r)\) is the mass (number of active sites) within radius \(r\) of the event origin, then

   \[M(r) \sim r^{D_f}\]

   for some fractal dimension \(D_f < d\) (the embedding dimension). This is to be contrasted with compact events (\(D_f = d\)) or point-like events (\(D_f = 0\)).

All three hallmarks are consequences of the same underlying criticality: the system sits at a second-order phase transition point where the correlation length diverges, but the system arrives there dynamically rather than by external tuning.

Natural Examples and Connection to Scaling Laws

Earthquakes. The Gutenberg-Richter law states

\[\log_{10} N(M) = a - bM, \quad b \approx 1\]

where \(N(M)\) is the number of earthquakes with magnitude \(\geq M\) and \(M = \frac{2}{3}\log_{10} E + \text{const}\) is the moment magnitude. Substituting, this implies a power law for seismic energy release:

\[P(E) \sim E^{-\tau_E}, \quad \tau_E = 1 + \frac{2b}{3} \approx \frac{5}{3}\]

Forest fires. In certain ecosystems, the distribution of fire sizes is approximately power-law, consistent with the forest-fire model of Drossel and Schwabl (1992), itself an SOC model.

Biological evolution. Bak and Sneppen (1993) proposed SOC as a mechanistic explanation for punctuated equilibrium in the fossil record — long periods of stasis interrupted by rapid bursts of evolutionary change. The Bak-Sneppen model produces power-law distributions of extinction cascade sizes.

Neural avalanches. Beggs and Plenz (2003) recorded local field potentials in cortical slices and observed that spontaneous bursts of neural activity (neural avalanches) have power-law size distributions with exponent \(\tau \approx 3/2\), consistent with a mean-field branching process at criticality. This suggests that neural networks may operate near an SOC critical point, potentially to maximize dynamic range and information transmission.

Caveat. The causal attribution of empirical power laws to SOC is debated. Many mechanisms can produce heavy-tailed or power-law distributions without SOC: preferential attachment (Barabasi-Albert), multiplicative noise, mixtures of exponentials, and simple self-similar geometry. The presence of a power law is necessary but not sufficient for SOC.

Connection to neural scaling laws. Bahri et al. (2021) provided a statistical mechanics derivation of neural network scaling laws (of the form \(L(N) \sim N^{-\alpha}\) for loss \(L\) as a function of model size \(N\)), arguing that the scaling exponent \(\alpha\) is inversely proportional to the intrinsic dimension \(d\) of the data manifold:

\[\alpha \propto \frac{1}{d}\]

This mirrors how SOC exponents depend on the geometry of the underlying system — in both cases, a power law emerges from an underlying geometric or statistical structure, and the exponent encodes dimensionality information. This is a deep structural analogy: in SOC, the critical exponents are determined by the dimension of the lattice and the universality class of the phase transition; in neural scaling, the loss exponents are determined by the intrinsic dimension of the data distribution.

The BTW Sandpile Model

The Bak-Tang-Wiesenfeld (BTW) sandpile model, introduced in Bak et al. (1987), is the canonical model of SOC. We give a precise mathematical definition.

2.1 Formal Definition

Lattice. Let \(\Lambda = \{1, \ldots, L\}^2 \subset \mathbb{Z}^2\) be an \(L \times L\) square lattice with open (absorbing) boundaries. Denote the set of boundary sites as \(\partial\Lambda\) (those with fewer than 4 neighbors in \(\Lambda\)) and interior sites as \(\Lambda^\circ = \Lambda \setminus \partial\Lambda\).

State space. A configuration is a function \(z : \Lambda \to \mathbb{Z}_{\geq 0}\), where \(z_i\) is the “height” (number of sand grains) at site \(i \in \Lambda\). The state space is \(\Omega = \mathbb{Z}_{\geq 0}^{|\Lambda|}\).

Threshold. The critical threshold is \(z_c = 4\) (the degree of an interior site on \(\mathbb{Z}^2\)).

Toppling rule. A site \(i\) is unstable if \(z_i \geq z_c = 4\). When an unstable site \(i\) topples:

\[z_i \to z_i - 4, \qquad z_j \to z_j + 1 \quad \forall j \in \mathcal{N}(i)\]

where \(\mathcal{N}(i)\) is the set of nearest neighbors of \(i\) in \(\Lambda\) (with \(|\mathcal{N}(i)| \leq 4\)). If \(i \in \partial\Lambda\), it has \(|\mathcal{N}(i)| < 4\) neighbors; the \(4 - |\mathcal{N}(i)|\) “missing” grains are lost to the boundary. This loss constitutes the dissipation mechanism — the only way grains leave the system.

Stable configurations. A configuration \(z\) is stable if \(z_i < 4\) for all \(i \in \Lambda\). The set of stable configurations is \(\mathcal{S} = \{0,1,2,3\}^{|\Lambda|}\).

Relaxation operator. Given any configuration \(z \in \Omega\), the relaxation \(\mathcal{R}(z)\) is the unique stable configuration reached by repeatedly toppling any unstable site. (Uniqueness is guaranteed by the Abelian property; see Section 4.2.)

2.2 Time-Scale Separation

The BTW dynamics consists of two processes operating on separated time scales:

Slow drive. At discrete time \(t = 1, 2, 3, \ldots\), add one grain to a uniformly random site \(i_t \sim \text{Uniform}(\Lambda)\):

\[z \to z + \mathbf{e}_{i_t}\]

where \(\mathbf{e}_i\) is the unit vector at site \(i\).

Fast relaxation. Immediately after each addition, apply the relaxation operator:

\[z(t) = \mathcal{R}(z(t-1) + \mathbf{e}_{i_t})\]

The relaxation (avalanche) runs to completion before the next grain is added.

Time-scale separation condition. The drive rate \(h\) (grains per unit time) is taken to zero, \(h \to 0\), relative to the relaxation rate. This ensures: 1. At most one avalanche is active at any instant. 2. The system is always in a stable configuration between additions. 3. The avalanche statistics are not contaminated by overlap between successive perturbations.

This separation is not merely a technical convenience — it is physically essential. If \(h\) is too large (the “fast-drive” limit), avalanches overlap and the critical behavior is destroyed.

Stationary distribution. Under this dynamics, the system reaches a unique stationary distribution \(\mu^*\) over stable configurations. The SOC phenomenology (power-law avalanche statistics) is a property of this stationary distribution. Computing \(\mu^*\) exactly is the content of Dhar’s abelian sandpile theory (Section 4).

2.3 Avalanche Observables

After adding a grain at time \(t\), the resulting avalanche is characterized by three observables:

Size \(s\): the total number of topplings summed over all sites and all time steps of the avalanche,

\[s = \sum_{i \in \Lambda} n_i\]

where \(n_i\) is the number of times site \(i\) topples during the avalanche.

Duration \(T\): the number of parallel update steps (synchronous updates) until stability. In each step, all currently unstable sites topple simultaneously.

Area \(a\): the number of distinct sites that topple at least once,

\[a = |\{i \in \Lambda : n_i \geq 1\}|\]

Note the inequalities \(a \leq s\) (since multiple topplings at a single site count once in \(a\) but multiple times in \(s\)) and \(T \leq s\) (since each step involves at least one toppling).

Empirical distributions in the stationary regime. For large \(L\), the distributions of \(s\), \(T\), and \(a\) in the stationary state follow power laws with exponential cutoffs imposed by the finite system size:

\[P(s) \sim s^{-\tau_s} \, g_s(s / L^{D_s}), \qquad \tau_s \approx 1.20, \quad D_s \approx 2.75\]

\[P(T) \sim T^{-\tau_T} \, g_T(T / L^{z}), \qquad \tau_T \approx 1.37, \quad z \approx 1.57\]

\[P(a) \sim a^{-\tau_a} \, g_a(a / L^{2}), \qquad \tau_a \approx 1.14\]

where \(g_s, g_T, g_a\) are scaling functions that decay rapidly for argument \(\gg 1\) and approach constants for argument \(\ll 1\). The exponent \(D_s\) is the avalanche fractal dimension (relating avalanche mass to its spatial extent), and \(z\) is the dynamical exponent (relating duration to spatial extent via \(T \sim r^z\)).

Remark on exact values. The exponents \(\tau_s, \tau_T, D_s, z\) for the 2D BTW sandpile are not known exactly and are determined numerically. This is in contrast to the 1D sandpile (trivially solved) and the mean-field case (\(d \geq 4\), where exact exponents are known via branching process arguments: \(\tau_s = 3/2\), \(\tau_T = 2\)).

2.4 The 1/f Noise Connection

Consider the time series \(\{\phi(t)\}_{t \geq 1}\) where \(\phi(t)\) is the number of grains dissipated through the boundary during the avalanche triggered at time \(t\). Bak, Tang, and Wiesenfeld argued that this signal exhibits 1/f noise.

Heuristic argument. Model each avalanche as a rectangular pulse of height \(h_0\) and duration \(T\). The Fourier transform of a single pulse of duration \(T\) has power concentrated at frequencies \(f \lesssim 1/T\):

\[|\hat{\phi}_T(f)|^2 \approx h_0^2 T^2 \cdot \mathbf{1}[f \lesssim 1/T]\]

The aggregate power spectrum is the average over avalanche durations drawn from \(P(T) \sim T^{-\tau_T}\):

\[S(f) \sim \int_0^\infty P(T) \, |\hat{\phi}_T(f)|^2 \, dT \sim \int_{1/f}^\infty T^{-\tau_T} \cdot T^2 \, dT \sim f^{-(3 - \tau_T)}\]

where the lower limit of integration is \(1/f\) because pulses shorter than \(1/f\) contribute negligibly at frequency \(f\). Therefore:

\[\boxed{S(f) \sim f^{-(3-\tau_T)}}\]

For \(\tau_T \approx 1.37\): \(S(f) \sim f^{-1.63}\), which is close to but not exactly \(1/f\). The exact \(\beta = 1\) (pink noise) would require \(\tau_T = 2\), which is the mean-field value.

Caveat. The \(1/f\) label is used loosely. More precisely, BTW predicts \(S(f) \sim f^{-\beta}\) with \(\beta = 3 - \tau_T \in (1, 2)\) depending on the model. Furthermore, \(1/f\) noise is not unique to SOC: any superposition of relaxation processes (Lorentzians) with a power-law distribution of relaxation times \(P(\tau) \sim \tau^{-1}\) produces \(S(f) \sim f^{-1}\) without any SOC dynamics. The presence of \(1/f\) noise is consistent with SOC but does not establish it.

Why Self-Organized Criticality?

3.1 The Attractor Argument

Argue that the critical state is the unique stationary state. Consider the average height \(\langle z \rangle\):

• Subcritical regime (\(\langle z \rangle \ll z_c\)): most sites are far from threshold. Avalanches are small — a grain addition typically causes 0 or 1 topplings. The output flux (grains leaving at the boundary) is much less than the input flux (one grain per step). Therefore \(\langle z \rangle\) increases monotonically. The subcritical state is not stationary — it drifts upward.

• Supercritical regime (\(\langle z \rangle \gtrsim z_c\)): many sites are near threshold. Grain additions trigger large avalanches that reach the boundary and dissipate many grains. The output flux exceeds the input flux. Therefore \(\langle z \rangle\) decreases. The supercritical state is also not stationary.

• Critical state (\(\langle z \rangle = \langle z \rangle_c\)): the unique average height at which input flux equals output flux in the stationary sense. This is a balance point that the dynamics drives the system toward from both sides — a dynamical attractor. No external control is needed to reach or maintain it.

The slow drive (\(h \to 0\)) and open boundaries (dissipation) are both essential: without slow drive, the system cannot evolve; without dissipation at the boundary, grains accumulate without limit and there is no attractor.

3.2 Connection to Absorbing State Phase Transitions

A more formal perspective (Dickman, Muñoz, Vespignani, Zapperi 2000): interpret the activity (number of active/unstable sites) as an order parameter. In the limit \(h \to 0\), \(\epsilon \to 0\) (where \(h\) is the drive rate and \(\epsilon\) is the per-toppling dissipation rate):

• For \(h > 0\) fixed and \(\epsilon > 0\): the system is always in the active phase — eventually all configurations are visited, including highly active ones.
• For \(h = 0\), \(\epsilon > 0\): the system falls into an absorbing state (no unstable sites, no driving). This is the absorbing phase.
• The boundary between these phases — the absorbing state phase transition — is where SOC lives.

Formally, SOC corresponds to the self-tuning of the system to this phase boundary via the feedback between drive and dissipation. The SOC critical point is in the universality class of the Manna model (for stochastic sandpiles) or a separate class for the deterministic BTW model. This connection to absorbing state criticality provides a field-theoretic framework for computing SOC exponents — though the calculations are technically difficult and exponents are generally known only numerically.

The Abelian Sandpile (Dhar 1990)

4.1 Setup on a General Graph

Let \(G = (V \cup \{s\}, E)\) be a finite connected undirected graph with vertex set \(V\) (non-sink vertices) and a distinguished sink vertex \(s\). The toppling matrix \(\Delta\) is the \(|V| \times |V|\) matrix: \[\Delta_{ij} = \begin{cases} \deg(i) & i = j \\ -1 & \{i,j\} \in E,\ i \neq j \\ 0 & \text{otherwise} \end{cases}\] where \(\deg(i)\) counts all edges from \(i\), including edges to \(s\). This is the graph Laplacian restricted to non-sink vertices — equivalently, the full Laplacian with the sink row and column deleted.

The height variable \(z_i \in \mathbb{Z}_{\geq 0}\) at each \(i \in V\). Site \(i\) is stable if \(z_i < \deg(i)\) and unstable if \(z_i \geq \deg(i)\). The toppling rule at site \(i\): \[z_i \to z_i - \deg(i), \qquad z_j \to z_j + 1 \quad \forall \{i,j\} \in E, j \neq s\] Grains sent to \(s\) are lost (dissipation). In matrix form: a toppling at \(i\) changes the height vector by \(-\Delta_{i\cdot}\) (subtract the \(i\)-th row of \(\Delta\)).

4.2 The Abelian Property

Theorem (Dhar 1990). Let \(\eta\) be a configuration (possibly unstable). If there exists a finite legal toppling sequence — a sequence of topplings of unstable sites — that stabilizes \(\eta\), then: 1. Every legal toppling sequence from \(\eta\) also stabilizes \(\eta\). 2. All stabilizing sequences produce the same final stable configuration \(\eta'\). 3. Each site \(i \in V\) topples the same number of times \(n_i\) in every stabilizing sequence.

Proof.

Let \(\mathbf{n}^\alpha \in \mathbb{Z}_{\geq 0}^{|V|}\) denote the toppling vector for sequence \(\alpha\) (number of times each site topples). After sequence \(\alpha\): \[\eta'_i = \eta_i - \sum_j \Delta_{ij} n_j^\alpha = \eta_i - (\Delta \mathbf{n}^\alpha)_i\] or in vector form \(\eta' = \eta - \Delta \mathbf{n}^\alpha\).

If \(\alpha\) and \(\beta\) both stabilize \(\eta\): \[\eta - \Delta \mathbf{n}^\alpha = \eta - \Delta \mathbf{n}^\beta \implies \Delta(\mathbf{n}^\alpha - \mathbf{n}^\beta) = \mathbf{0}\]

Key claim: \(\Delta\) is positive definite on \(\mathbb{R}^{|V|}\), hence \(\ker(\Delta) = \{\mathbf{0}\}\).

Proof of positive definiteness: \(\Delta\) is the graph Laplacian restricted to \(V\), with the sink providing a “ground.” For any \(\mathbf{x} \in \mathbb{R}^{|V|}\): \[\mathbf{x}^\top \Delta \mathbf{x} = \sum_{\{i,j\} \in E, i,j \in V} (x_i - x_j)^2 + \sum_{i \in V,\ \{i,s\} \in E} x_i^2 \geq 0\] Since \(G\) is connected and \(s\) is adjacent to at least one vertex in \(V\), the quadratic form is strictly positive for \(\mathbf{x} \neq \mathbf{0}\).

Therefore \(\mathbf{n}^\alpha = \mathbf{n}^\beta\), and hence \(\eta'^\alpha = \eta'^\beta\). \(\square\)

The abelian property is why the sandpile is analytically tractable: the stabilization map \(\eta \mapsto \eta'\) and the toppling numbers \(\mathbf{n}(\eta)\) are well-defined functions, independent of implementation order.

4.3 The Toppling Lemma

Lemma. Let \(\eta\) be a configuration that can be stabilized. If \(n_i(\eta) \geq 1\) (site \(i\) topples at least once in the stabilization of \(\eta\)), then \(n_i(\eta + \mathbf{e}_j) \geq n_i(\eta)\) for any grain addition \(\mathbf{e}_j\) (adding a grain at site \(j\)).

More simply: adding grains can only increase the number of topplings, never decrease it.

Proof sketch: Adding a grain at \(j\) can only make \(\eta + \mathbf{e}_j\) “more unstable” than \(\eta\). Any toppling sequence that stabilizes \(\eta\) is also legal for \(\eta + \mathbf{e}_j\) (the extra grain at \(j\) never makes a previously legal toppling illegal). By the abelian property, the toppling numbers \(\mathbf{n}(\eta + \mathbf{e}_j) \geq \mathbf{n}(\eta)\) componentwise. \(\square\)

This monotonicity is essential for proving properties of the stationary distribution.

4.4 Recurrent Configurations and the Sandpile Group

A stable configuration \(\eta\) is recurrent if it appears with positive probability in the unique stationary distribution of the Markov chain (add grain at random site, stabilize, repeat). A configuration is transient if it is visited at most finitely often (probability zero in stationarity).

Dhar’s burning algorithm (criterion for recurrence): \(\eta\) is recurrent if and only if the following process terminates with all vertices burned: 1. Initialize: mark the sink \(s\) as “burned.” 2. Iterate: if an unburned vertex \(i\) has \(z_i \geq\) (number of unburned neighbors of \(i\)), mark \(i\) as burned. 3. Repeat until no more vertices can be burned. 4. \(\eta\) is recurrent iff all vertices in \(V\) are eventually burned.

Intuitively: a recurrent configuration is one “dense enough” that every site could topple at least once if grains arrived from burned (stabilized) neighbors — it cannot be “blocked” by any cluster of sites.

The sandpile group. The set \(\mathcal{R}\) of recurrent configurations forms an abelian group under the operation: \[\eta_1 \oplus \eta_2 = \text{Stab}(\eta_1 + \eta_2)\] (add the configurations componentwise, then stabilize). This group is: - Abelian: by the abelian property, \(\eta_1 \oplus \eta_2 = \eta_2 \oplus \eta_1\) - Finite: \(|\mathcal{R}| = \det(\Delta)\) (a remarkable identity proved by Dhar) - Cyclic structure: for the \(L \times L\) grid, \(\det(\Delta)\) grows as \(e^{c L^2}\) where \(c = \frac{4G}{\pi}\) and \(G\) is Catalan’s constant

The sandpile group is one of the rare examples of a non-trivial algebraic structure arising from a dynamical system, and its study connects to algebraic combinatorics, chip-firing games, and tropical geometry.

Scaling Theory and Exponents

5.1 Finite-Size Scaling

For a system of linear size \(L\), the avalanche size distribution is not a pure power law — it is cut off at large \(s\) by the system size. The finite-size scaling ansatz is:

\[P(s, L) = s^{-\tau_s}\, g\!\left(\frac{s}{L^D}\right)\]

where: - \(\tau_s > 1\) is the size exponent - \(D > 0\) is the avalanche fractal dimension — the exponent relating the typical maximum avalanche size to system size: \(s_{\max} \sim L^D\) - \(g : (0,\infty) \to (0,\infty)\) is a universal scaling function satisfying \(g(u) \approx \text{const}\) for \(u \ll 1\) and \(g(u) \to 0\) rapidly (faster than any power) for \(u \gg 1\)

In the \(L \to \infty\) limit, the cutoff moves to infinity and \(P(s) \sim s^{-\tau_s}\) for all \(s\) — a pure power law. For finite \(L\), the distribution follows the power law for \(s \ll L^D\) and is exponentially suppressed for \(s \gg L^D\).

The same ansatz applies to the duration and area distributions: \[P(T, L) = T^{-\tau_T} \tilde{g}\!\left(\frac{T}{L^z}\right), \qquad P(a, L) = a^{-\tau_a} \hat{g}\!\left(\frac{a}{L^{d_f}}\right)\] where \(z\) is the dynamical exponent and \(d_f\) is the spatial fractal dimension of avalanche footprints.

5.2 Scaling Relations

The exponents are not all independent — they are related by scaling laws derived from the internal consistency of the finite-size scaling ansatz.

Relation 1 — from \(s \sim T^{\sigma}\) (size scales as duration to a power): \[\tau_T = 1 + \sigma(\tau_s - 1), \qquad \sigma = D/z\]

Relation 2 — from \(a \sim s^{d_f/D}\) (area scales as a power of size): \[\tau_a = 1 + \frac{D}{d_f}(\tau_s - 1)\]

Relation 3 — from conservation: in the BTW sandpile, each toppling conserves grains in the interior. The total number of grains dissipated equals the avalanche size reaching the boundary. For \(d\)-dimensional systems, conservation implies: \[D = d + 2 - \tau_s \cdot d \quad \text{(heuristic)}\] This is an approximate relation; the exact form depends on the universality class.

These relations mean that, in principle, only two independent exponents characterize an SOC universality class (analogous to two independent exponents in equilibrium critical phenomena).

5.3 Known Values for 2D BTW

Numerically established values for the \(2\)-dimensional BTW sandpile on \(\mathbb{Z}^2\) (exact analytic values remain open):

Important caveats: - The 2D BTW sandpile is believed to have logarithmic corrections to pure power-law scaling, arising from its exact abelian structure. These corrections make numerical estimation of exponents particularly difficult. - Some exact results are known on special graphs (complete graph, trees) where the abelian structure can be exploited fully, but the 2D square lattice remains analytically open. - There is ongoing debate in the literature about the exact values of exponents and whether the 2D BTW sandpile is truly critical in the thermodynamic limit or exhibits only quasi-critical behavior.

5.4 Universality

Different SOC models belong to different universality classes, characterized by different sets of exponents:

• BTW (deterministic): topplings are deterministic — a site always sends exactly one grain to each neighbor. This determinism may produce logarithmic corrections and place BTW in a distinct class.
• Manna model (stochastic): when site \(i\) topples, it sends 2 grains to randomly chosen neighbors. This stochasticity places the Manna model in the “Manna universality class” — believed to be the generic class for stochastic sandpiles, with \(\tau_s^{\text{Manna}} \approx 1.28\) in 2D.
• Conserved vs. non-conserved: the BTW sandpile is locally conserved (grains only lost at boundary). Non-conserved models (e.g., Olami-Feder-Christensen earthquake model) generally exhibit different scaling.

The universality class is determined by symmetries of the toppling rule (deterministic vs. stochastic, conserved vs. non-conserved, spatial dimension), not merely by the qualitative presence of power-law avalanche statistics.

References