Spatial correlations in SIS processes on random regular graphs

This paper develops a hierarchical system of ordinary differential equations to model higher-order spatial correlations in SIS processes on random regular graphs, significantly improving the accuracy of infection density predictions compared to traditional mean-field and pairwise approximations.

The Gist

Imagine a game of "Hot Potato" played on a giant, invisible web of friends. In this game, the "potato" is a virus, and the friends are people. Some people have the potato (they are Infected), and some don't (they are Susceptible).

Every time a person with the potato shakes hands with a friend who doesn't have it, there's a chance the friend catches it. If someone has the potato long enough, they eventually get bored, throw it away, and become safe again (they Recover).

This is the SIS Model (Susceptible-Infected-Susceptible). It's how scientists try to predict how diseases like the common cold or the flu spread.

The Problem: The "Well-Mixed Soup" Mistake

For a long time, scientists used a simple way to predict the spread called the Mean-Field Approximation.

The Analogy: Imagine you drop a drop of red dye into a giant, perfectly stirred pot of soup. You assume the dye spreads instantly and evenly everywhere. In this "soup" model, everyone is equally likely to meet anyone else. It assumes the world is a giant, perfectly mixed cocktail party where everyone shakes hands with everyone else at the same rate.

Why it fails: In real life, we don't live in a soup. We live in neighborhoods. You only catch a cold from the people you actually hang out with (your neighbors, coworkers, family). If your best friend is sick, you are much more likely to get sick than a stranger on the other side of the country.

The "soup" model ignores this. It assumes that if your friend is sick, your other friends are just as likely to be sick as anyone else. But in reality, sickness tends to cluster. If your friend is sick, their other friends are probably sick too, because they all hang out together. This creates "hot spots" of infection.

The Old Fix: The "Neighbor Watch"

Scientists tried to fix this with a method called the Pairwise Model. This is like a "Neighbor Watch." Instead of just looking at the whole crowd, it looks at pairs of people: "Is Person A sick? Is their neighbor Person B sick?"

This is better than the soup model, but it still has a blind spot. It only looks at immediate neighbors. It doesn't realize that if Person A is sick, and Person B is sick, then Person C (who is friends with both A and B) is very likely to be sick too, even if C isn't directly connected to the original source in the model's simple view. It misses the "ripple effect" of infection spreading further out.

The New Solution: The "Multi-Layer Shell" Map

The authors of this paper (Leibenzon, Johnson, Baker, and Assaf) came up with a smarter way to look at the web. They call it the Multi-Shell Pairwise Model (MPM).

The Analogy: Imagine you are standing in the center of a forest.

• Shell 1: These are the trees (people) you can touch right now.
• Shell 2: These are the trees you can reach by touching a tree in Shell 1.
• Shell 3: The trees you can reach by going through Shell 2, and so on.

The old models only looked at Shell 1. They asked, "Is my immediate neighbor sick?"

The new model looks at all the shells. It asks:

• "Is my neighbor sick?"
• "Is my neighbor's neighbor sick?"
• "Is my neighbor's neighbor's neighbor sick?"

By tracking these "shells" of distance, the model understands how the infection clusters. It realizes that if the virus is spreading, it's not just jumping randomly; it's building a "fortress" of infection in specific areas.

Why Does This Matter?

The researchers tested this on Random Regular Graphs. Think of this as a social network where everyone has the exact same number of friends (say, 4 or 5), but who those friends are is random.

They found that:

1. The "Soup" model is too optimistic: It thinks the disease will spread everywhere, everywhere, everywhere.
2. The "Neighbor Watch" model is okay, but not great: It gets the basics right but misses the deeper connections.
3. The "Multi-Shell" model is the winner: It matches real computer simulations almost perfectly.

The Big Discovery:
The paper shows that in networks with fewer connections (like a small town where everyone knows a few people), the "clustering" of the virus is very strong. The virus gets stuck in small groups. This actually slows down the spread compared to the "soup" model's prediction.

However, if you add more random connections (making the network more like a chaotic city), the virus spreads faster and the "clustering" effect weakens.

The Takeaway

This paper is like upgrading from a blurry, black-and-white map to a high-definition, 3D GPS.

• Old Way: "The virus is everywhere, so everyone is at risk." (Too scary, too simple).
• New Way: "The virus is building up in specific clusters. If we understand how these clusters form and how they connect to the next layer of friends, we can predict exactly how big the outbreak will get and how long it will last."

By understanding these "spatial correlations" (how the virus groups itself), scientists can make much better predictions about infectious diseases, helping us prepare for outbreaks without panicking or being too complacent. It's about seeing the forest and the trees, and even the trees behind the trees.

Technical Summary

1. Problem Statement

The Susceptible-Infected-Susceptible (SIS) model is a fundamental framework for describing infectious diseases where individuals do not acquire lasting immunity. Traditional mean-field approximations assume a well-mixed population, ignoring spatial structure. While mathematically tractable, these models fail to capture spatial correlations inherent in network-based transmission, where infection spreads only through direct links.

Existing improvements, such as the Standard Pairwise Model (SPM), account for nearest-neighbor correlations but rely on closure approximations that assume correlations decay rapidly with distance. Consequently, the SPM neglects multi-scale spatial correlations beyond immediate neighbors. This leads to inaccurate predictions, particularly on Random Regular Graphs (RRGs) with low average node degrees ($k_0$), where the SPM systematically overestimates infection density and underestimates spatial clustering. There is a lack of an analytical framework that systematically accounts for higher-order spatial correlations on generic networks, specifically RRGs.

2. Methodology

The authors develop a Multi-Shell Pairwise Model (MPM) to address these limitations. The methodology involves:

• Hierarchical Shell Definition: Instead of Euclidean distance (which is ill-defined on RRGs), the network is conceptualized in "shells" based on shortest-path distance ($\ell$) from a reference node. Shell $\ell=1$ contains direct neighbors, $\ell=2$ contains neighbors of neighbors, etc.
• Correlation Functions: The authors define spatial correlation functions $F^{(\ell)}_{II}(t)$ representing the ratio of the actual fraction of infected-infected pairs at distance $\ell$ to the mean-field expectation.
• Derivation of Coupled ODEs:
  • They derive a system of coupled Ordinary Differential Equations (ODEs) governing the time evolution of $F^{(\ell)}_{II}$ for all shells $\ell$.
  • The derivation utilizes the Kirkwood Superposition Approximation (KSA) to close the hierarchy of higher-order moments (triplet probabilities) into pairwise terms.
  • Crucially, unlike the SPM which sets correlations to unity for $\ell > 1$, the MPM retains the dynamics of correlations at all distances.
• Topological Probabilities: To apply this to RRGs, the authors derive explicit expressions for the conditional probabilities $p(\ell'|\ell)$ (the probability that a neighbor of a node in shell $\ell$ lies in shell $\ell'$). These are calculated using combinatorial stub-matching arguments and the Gompertz distribution for shell populations on RRGs.
• Steady-State Analysis: The authors perform an asymptotic analysis for large $N$ and varying $k_0$, deriving closed-form expansions for the steady-state infection density ($\bar{x}_I$) and the pairwise correlation function.

3. Key Contributions

• Generalized Framework: The paper extends mean-field corrections from regular lattices to Random Regular Graphs, creating a general framework for equivalent corrections on random topologies.
• Multi-Shell Resolution: The MPM resolves the full hierarchy of spatial correlations at every network distance, capturing how infection clustering decays with distance.
• Analytical Corrections: The authors derive closed-form asymptotic expansions for the endemic infection density and pairwise correlations, explicitly showing how finite connectivity suppresses infection prevalence compared to classical mean-field predictions.
• Critical Threshold Refinement: The work provides a refined expression for the critical spreading rate ($\Lambda_c$) on RRGs, correcting the standard mean-field threshold ($\Lambda=1$) and the SPM threshold by accounting for higher-order spatial correlations.

4. Key Results

• Superior Accuracy: Numerical simulations demonstrate that the MPM predictions for global infection density and correlation functions agree significantly better with discrete simulations than both the mean-field model and the SPM.
  • The SPM systematically overestimates infection density, especially at low degrees ($k_0$).
  • The MPM accurately captures the suppression of infection density caused by local clustering.
• Dependence on Topology: Spatial correlations are shown to be highly sensitive to network structure. Increasing topological randomness (rewiring) or increasing the infection rate suppresses spatial correlations and accelerates epidemic dynamics.
• Asymptotic Behavior:
  • The steady-state pairwise correlation function $\bar{F}^{(\ell)}_{II}$ decays exponentially with distance $\ell$.
  • The corrected steady-state infection density is given by:
    $$\bar{x}_I \approx \bar{x}^{MF}_I - \frac{1}{k_0 \Lambda^2} - \frac{\Lambda^2 - \Lambda + 1}{k_0^2 \Lambda^4} + \dots$$
    This confirms that finite-connectivity corrections reduce the infection density below the mean-field prediction.
• Critical Threshold: The derived critical spreading rate is $\Lambda_c = 1 + k_0^{-1} + 2k_0^{-2} + O(k_0^{-3})$, which differs from the SPM threshold at the order of $O(k_0^{-2})$.

5. Significance

This work provides a robust analytical tool for forecasting infectious disease dynamics on networks where spatial structure plays a critical role. By moving beyond nearest-neighbor approximations, the MPM offers a more accurate characterization of epidemic trajectories, particularly in regimes with low connectivity where standard models fail.

The findings highlight that structural randomness in networks fundamentally alters the spatial organization of infections. The developed framework bridges the gap between simple mean-field theories and computationally expensive simulations, offering a tractable method to understand how network topology influences disease persistence and the epidemic threshold. Future applications could extend this framework to heterogeneous topologies and more complex closure approximations (e.g., Fisher-Kopeliovich) to further improve accuracy near critical thresholds.