Networks and Complex Systems Course

Processes on Networks

From structure to dynamics — Networks and Complex Systems

Resource for the Networks and Complex Systems course focused on dynamical processes on graphs. It covers SIS and SIR models, epidemic thresholds, intervention strategies, and temporal networks, combining mathematical intuition with interactive simulations.

Resource created by Cristian Candia at CRiSS Lab.

SIS vs SIR: two different questions

Both models define local transition rules on a network. But they answer different questions and prioritize different metrics.

AspectSISSIR
SequenceS → I → SS → I → R
Central questionDoes the infection persist?How large is the outbreak?
Theoretical objectStationary prevalenceFinal size / percolation
Structural intuitionHubs as reservoirsGeometric reach of the outbreak
Classical toolHMF, QMF, spectral radiusTransmissibility, percolation
The visualizations in this tool use small networks (≤100 nodes), discrete time, and synchronous updating. The results illustrate qualitative trends; they do not reproduce asymptotic theoretical predictions.

Interactive exploration: dynamics on a small graph

Watch how contagion spreads step by step. Each node changes color according to its state.

Susceptible
Infected
Recovered
0
Step
S
I
R

Microscopic rule

At each discrete step, every infected node attempts to infect each susceptible neighbor with probability β, and then recovers with probability μ. In SIR it moves to R; in SIS it returns to S.

The ratio λ = β/μ acts as an effective control parameter. But the threshold also depends on network structure.

In SIR, Newman (2002) showed that the final outbreak size is exactly equivalent to a bond percolation problem on random static networks with transmissibility T on each edge.

Epidemic threshold: two approximations

Is there a critical transmissibility value such that, below it, the outbreak dies out quickly, and above it, it can be sustained? That value is the epidemic threshold, and it is a joint property of dynamics + structure.

HMF: the role of ⟨k²⟩

λcHMF = ⟨k⟩ / ⟨k²⟩

If the network has hubs (large ⟨k²⟩), the threshold decreases: heterogeneity facilitates persistence because hubs act as dynamic reservoirs.

This formula is rigorous within the HMF scheme for SIS on uncorrelated networks. It is not a universal identity. In real finite networks, it is more accurate to speak of a low effective threshold, not necessarily zero.

QMF / Spectral: the role of λ₁(A)

λcQMF ≈ 1 / λ₁(A)

The spectral radius captures the structural amplification capacity of the whole network, not only the degree distribution.

1/λ₁(A) is a classical and informative approximation for SIS, not “the exact threshold.” Later literature (Castellano & Pastor-Satorras, 2010) shows that the asymptotic behavior can be more subtle.

Interactive structural comparison

Generate three networks of the same size and compare their threshold proxies.

If two networks have very different proxies, it is reasonable to expect different diffusion regimes. It is not correct to promise the exact threshold value from a single proxy.

SIR simulation: trajectories and Monte Carlo

A single run can be misleading: the initial seed and randomness dominate. That is why we repeat and average.

The networks generated here are small (N=150). The results illustrate qualitative trends. They do not replace an analytical threshold analysis and should not be extrapolated to networks of other sizes without caution.

Empirical transition: λ sweep

Vary λ = β/μ and observe where the outbreak stops being small. This shows an empirical transition zone, not an exact analytical threshold.

Intervention strategies

If I remove or immunize certain nodes, how much does the outbreak shrink? Three classical strategies:

StrategyCriterionIntuition
RandomAt randomBaseline
DegreeHighest degree firstCuts many potential contacts
k-coreHighest coreness firstBreaks the core where diffusion is sustained
There is no universal winner for all processes and all metrics (Bramson et al., 2016). The ranking depends on the process (SIR vs SIS), the network, and the evaluation criterion.

Interactive comparison

The comparison is on a BA network with 150 nodes and discrete-time SIR. Removing nodes changes size and connectivity; we always compare against an explicit baseline.

Temporal network vs aggregated network

The aggregated network answers: who was connected to whom at any time?
The temporal network answers: did a chronologically valid sequence for transmission exist?

These are not the same question. The order of contacts is part of the mechanism, not a formatting detail.

Interactive example

A chain of contacts with timestamps. Node 0 starts infected. Compare which nodes are reached temporally versus in the aggregated network.

Temporal network (ordered contacts)

Aggregated network (all contacts)

The aggregated network overestimates accessibility when it ignores temporal constraints. This is a robust methodological result (Holme, 2015).

Core idea of the class

"The importance of a node depends on the dynamical process taking place on the network, the time horizon, and the performance criterion."

There is no universally correct centrality. The appropriate measure depends on the propagation mechanism and the intervention objective.