🌀 nonlinear sciences

Sandpile Models on complex networks

This paper introduces a dissipation-aware branching-process framework to analyze sandpile models on complex networks, revealing that while dissipation induces exponential cutoffs, network features like clustering and sparse tree-like structures significantly alter avalanche scaling and invalidate classical independent-branch approximations.

Original authors: Komlan Fiagbe, Jean-François de Kemmeter, Timoteo Carletti

Published 2026-07-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Komlan Fiagbe, Jean-François de Kemmeter, Timoteo Carletti

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a giant, complex web of friends, computers, or neurons. Now, imagine dropping a single grain of sand onto one of these points. If that point gets too crowded, it "topples," dumping its sand onto its neighbors. Those neighbors might get too crowded and topple too, creating a chain reaction called an avalanche.

This paper is about studying how big these avalanches get when they happen on different kinds of webs (networks), and specifically, what happens when some of the sand disappears (dissipates) as it moves from one point to another.

Here is the breakdown of their findings using simple analogies:

1. The "Leaky Bucket" Effect (Dissipation)

In the old way of thinking about these avalanches, scientists assumed the sand was perfectly conserved—nothing ever got lost. It was like a game of "telephone" where every message is passed perfectly.

However, in the real world, sand often falls off the table or gets lost in the wind. The authors introduced a "leak" into their model: every time a grain of sand is passed to a neighbor, there is a chance it vanishes.

The Result: When you add this leak, the avalanches stop growing forever. Instead of a few massive, endless storms, you get a mix of small ripples and medium waves. The math shows that the "size" of these avalanches changes its pattern. If the leak is small, the pattern looks like a classic "power law" (a few huge events, many small ones). If the leak is big, the pattern changes, and huge events become much rarer.

2. The "Hubs" vs. The "Average Joe" (Scale-Free Networks)

Some networks are like a city with a few massive super-highways (hubs) and many tiny side streets. This is called a scale-free network.

The Finding: In these networks, the "hubs" (the super-highways) are so big that they can keep an avalanche going even if sand is leaking out. The authors found that when sand leaks, the mathematical rule describing the size of avalanches actually changes its number. It's as if the "recipe" for the storm changes depending on how leaky the system is.

3. The "Shortcuts" Problem (Clustering and Cycles)

The old math assumed that once an avalanche starts, it spreads out like a tree, with branches that never touch each other again. This works well if the network looks like a tree.

The Reality: Many real networks have "shortcuts" or loops (like a triangle of friends where A knows B, B knows C, and C knows A).
The Finding: When the authors added these loops (clustering), the avalanche behavior changed dramatically. The loops allowed the avalanche to "bounce back" and hit the same spots again. This created a feedback loop that made huge avalanches much more likely.
The Analogy: Imagine a rumor spreading. If everyone only talks to new people (a tree), the rumor dies out. But if people talk to their old friends too (loops), the rumor circles back, gets louder, and spreads much further than expected. The authors proved that these loops break the "independent branch" assumption used in older theories.

4. The "Empty Tree" Surprise (Trees and Sparse Networks)

You might think that a perfect tree (a network with no loops at all) would be the easiest place to predict avalanches.

The Surprise: The authors found that on very sparse trees (where there are very few connections and many "dead ends" or leaves), the avalanches behave strangely. They don't follow the neat, predictable power-law patterns that the math predicted.
The Reason: Because the tree is so sparse, the sand often hits a dead end (a leaf) and stops immediately, or it gets stuck in a small loop of just two nodes. The "tree" isn't big or connected enough to let the avalanche grow properly. It's like trying to start a forest fire in a sparse grove of trees with lots of gaps; the fire just sputters out rather than spreading in a predictable pattern.

5. The "Grid" vs. The "Random Web"

They also compared a perfectly organized grid (like a chessboard) to a random web of connections.

The Grid: Because a grid is full of loops and shortcuts, the avalanches didn't follow the "power law" rules at all.
The Random Web: A random web, even if everyone has the same number of friends, tends to look like a tree locally. On these, the avalanches did follow the predictable power-law rules.

The Big Takeaway

The main message of the paper is that structure matters more than we thought.

Dissipation (losing sand) changes the rules of the game.
Loops (clustering) make big disasters more likely than simple math predicts.
Sparse trees (too few connections) can actually stop avalanches from behaving in the "textbook" way.

The authors built a new mathematical toolkit that accounts for sand leaking out and the messy reality of loops in networks. This toolkit shows that you can't just treat every network as a simple tree; you have to look at how the connections are actually arranged to understand how big a "storm" (avalanche) might get.

Technical Summary: Sandpile Models on Complex Networks

Problem Statement
The sandpile model, a paradigmatic framework for self-organized criticality (SOC), has been extensively studied on regular Euclidean lattices and extended to complex networks. While classical analytical approaches often rely on branching-process approximations assuming conservative transport and locally tree-like independence, these methods frequently neglect dissipation. In finite systems, dissipation is necessary to prevent indefinite grain accumulation, yet its explicit incorporation into branching-process frameworks on complex networks remains underexplored. Furthermore, the impact of network topology—specifically clustering, short cycles, and sparse connectivity—on avalanche statistics and the validity of independent-branch approximations requires further investigation.

Methodology
The authors develop a generalized branching-process framework that explicitly incorporates dissipation during avalanche propagation.

Theoretical Framework: The model considers a network where grains are added to nodes. When a node's height exceeds its degree, it topples, distributing grains to neighbors. A probability $f$ is introduced for each transferred grain to be dissipated (removed from the system). The authors derive generalized probability generating functions (PGFs) for the offspring distribution, $Q(w|f)$ , which accounts for grain loss. They utilize functional relations between the PGF of the offspring distribution and the avalanche size distribution, $P(y) = y Q(P(y)|f)$ .
Analytical Tools: The study employs asymptotic analysis, singularity analysis, and Lagrange inversion theorems (specifically Theorem 1 and Theorem 2 from the literature) to derive the asymptotic behavior of the avalanche size distribution $p(s)$ for large $s$ .
Numerical Simulations: Extensive simulations are performed on various network topologies, including scale-free networks (Barabási–Albert and Holme–Kim models), Erdős–Rényi (ER) graphs, random regular graphs, Cayley trees, and lattice structures. The simulations use a driving mechanism proportional to network size and employ statistical fitting tools (e.g., the Python powerlaw package) with goodness-of-fit diagnostics (KS, Kuiper, Anderson–Darling statistics) to validate scaling behaviors.

Key Contributions and Results

Generalized Branching Process with Dissipation:
The paper derives explicit analytical expressions for avalanche statistics in dissipative regimes. It demonstrates that dissipation drives the system from a critical regime ( $\mu=1$ ) to a subcritical regime ( $\mu = 1-f < 1$ ).
- Scale-Free Networks: In the absence of dissipation ( $f=0$ ), the avalanche size distribution follows a power law $p(s) \sim s^{-\tau}$ with $\tau = \gamma/(\gamma-1)$ for $2 < \gamma < 3$ . When dissipation is introduced ( $0 < f < 1$ ), the exponent shifts to $\tau = \gamma$ . The distribution retains a power-law form but with a modified exponent, rather than simply acquiring an exponential cutoff, provided the network is sufficiently large and heterogeneous.
- Erdős–Rényi and Random Regular Graphs: For these networks with finite variance in degree distribution, the presence of dissipation leads to a distribution of the form $p(s) \sim s^{-3/2} \rho^{-s}$ . Here, the mean-field exponent $\tau = 3/2$ is preserved, but an exponential cutoff $\rho^{-s}$ is introduced intrinsically by the dissipative dynamics.
Impact of Clustering and Short Cycles:
Using Holme–Kim networks (which allow tuning of the clustering coefficient while maintaining a scale-free degree distribution with $\gamma=3$ ), the authors show that increasing the density of triangles systematically lowers the avalanche exponent $\tau$ .
- As clustering increases, the distribution becomes heavier-tailed, increasing the probability of large cascades.
- This finding invalidates the classical assumption that avalanche branches evolve independently; short cycles introduce strong correlations that alter the universality class of the dynamics.
Deviations in Tree-Like and Sparse Structures:
Contrary to the expectation that tree-like structures (which lack loops) perfectly satisfy branching-process assumptions, the study reveals significant deviations:
- Barabási–Albert Trees ( $m=1$ ): These sparse trees do not exhibit clear power-law scaling.
- Cayley Graphs: Despite being trees, Cayley graphs fail to follow the predicted $s^{-3/2}$ scaling.
- Cause: The authors attribute these failures to low edge density and the abundance of leaves. In sparse trees, grains tend to return to previously visited nodes or terminate quickly at leaves, creating correlations between branches and violating the independence assumption required for standard branching-process approximations.
Comparison of Topologies:
The results highlight that "local tree-likeness" (as found in random regular graphs) is more conducive to mean-field branching behavior than the mere absence of loops in sparse, structured trees. Highly structured topologies with short cycles (lattices, rings) or sparse connectivity (trees with many leaves) deviate significantly from ideal power-law predictions.

Significance and Claims
The paper claims to establish quantitative limits for branching-process descriptions of avalanche dynamics on networks. Its primary significance lies in:

Explicit Dissipation: Providing a rigorous analytical framework that incorporates grain loss directly into the offspring distribution, moving beyond conservative approximations.
Topology Dependence: Demonstrating that network heterogeneity (specifically the degree exponent $\gamma$ ) and local structural features (clustering, cycles, and leaf density) fundamentally reshape the avalanche size distribution.
Validity of Approximations: Showing that while branching-process theories are accurate for random, locally tree-like networks, they fail in highly structured or sparse systems where correlations and boundary effects (leaves) dominate.
Universality Shifts: Proving that dissipation can shift the universality class of scale-free networks (changing the exponent from $\gamma/(\gamma-1)$ to $\gamma$ ) and that clustering can continuously tune the critical exponent, thereby altering the probability of extreme events.

The authors conclude that dissipation, clustering, cycles, and sparse connectivity are not merely perturbative factors but are fundamental determinants of the sandpile model's behavior on complex networks.