Network Science Notes

Topic Map


1. Why Networks Instead of Flat Euclidean Features

Classical pipelines focus on vectors $x_i\in\mathbb{R}^d$ and similarity metrics (Euclidean distance, cosine, Pearson correlation). That is sufficient when interaction structure is weak.

In many systems, the object of interest is relational:

A graph $G=(V,E)$ with adjacency $A$ captures these relations. The dynamics then live on topology, for example:

\[\dot z_i = f_i(z_i) + \sum_{j=1}^n A_{ij} g_{ij}(z_i,z_j).\]

So structure and dynamics are coupled; this is the central reason network science is not just “another clustering method.”


2. Propagation and Epidemic Modeling on Networks

2.1 SIS dynamics and threshold condition

For SIS on graph $A$, let $p_i(t)$ be infection probability of node $i$. The NIMFA form is

\[\dot p_i = -\delta p_i + \beta (1-p_i)\sum_j A_{ij}p_j.\]

Linearizing near disease-free state $p=0$:

\[\dot p \approx (\beta A - \delta I)p.\]

Disease-free equilibrium is unstable when

\[\beta\lambda_1(A)-\delta>0 \quad\Longleftrightarrow\quad \tau=\frac{\beta}{\delta}>\frac{1}{\lambda_1(A)}.\]

This spectral criterion explains why hub-heavy networks are vulnerable: large $\lambda_1(A)$ reduces threshold.

SIS spectral threshold sketch Source: Author-generated figure.

2.2 Structure effects

Given fixed $\beta,\delta$:

A useful control surrogate is spectral radius reduction: design interventions so that $\lambda_1(A_{\text{effective}})$ decreases.

2.3 Resource-limited spreading

For information diffusion with bounded attention $b_i$, transmission can be modeled as constrained flow:

\[\sum_j u_{ij}(t) \le b_i,\]

where $u_{ij}$ is outgoing transmission effort. This turns diffusion into a coupled dynamics-allocation problem, not pure SI/SIS.


3. Temporal Networks

A temporal network is an event sequence

\[\mathcal{E}=\{(i,j,t_k)\}_{k=1}^M.\]

A temporal path $i_0\to i_1\to\cdots\to i_r$ is valid only if times are nondecreasing:

\[t_1 \le t_2 \le \cdots \le t_r.\]

Hence static shortest-path calculations can be wrong for causality.

Key metrics:

Temporal path and causality Source: Author-generated figure.

In robotic communication networks, this distinction is critical because links appear/disappear with motion and occlusion.


4. Node Distances and Measures

4.1 Distances

Common notions (not interchangeable):

4.2 Centralities

Measure selection should match task dynamics (routing, influence, epidemic suppression, etc.).


5. Modularity Clustering and Louvain

5.1 Modularity

For partition labels $c_i$,

\[Q=\frac{1}{2m}\sum_{ij}\left(A_{ij}-\frac{k_i k_j}{2m}\right)\mathbf{1}[c_i=c_j].\]

This compares observed within-community edge mass against configuration-model expectation.

5.2 Louvain heuristic

Louvain iterates:

  1. local node moves maximizing modularity gain $\Delta Q$,
  2. aggregation of communities into supernodes,
  3. repeat on coarsened graph.

One useful local gain expression (node $i$ moved into community $C$) is:

\[\Delta Q = \frac{k_{i,\text{in}}}{m} - \frac{k_i\,\Sigma_{tot,C}}{2m^2},\]

with standard notation for $i$’s links into $C$ and total degree mass in $C$.

Modularity partition and Louvain aggregation Source: Author-generated figure.

5.3 Resolution limit

Modularity has a known scale bias: small true communities can be merged if global gain is larger. So “high $Q$” is not equivalent to “ground-truth recovery.”


6. Random Graphs and Community Detectability

Community claims should be benchmarked against random graph models.

For symmetric SBM with $q$ groups and average within/between expected degrees $c_{in},c_{out}$, detectability has a Kesten-Stigum-type boundary (tree/large-sparse limit intuition):

\[\frac{(c_{in}-c_{out})^2}{q\left(c_{in}+(q-1)c_{out}\right)} > 1.\]

Below threshold, no algorithm can recover labels better than chance asymptotically.

SBM detectability transition sketch Source: Author-generated figure.

This is the practical meaning of “universal lower bound” for community structure recovery in sparse random graphs.


7. Bayesian Inference in Networks

7.1 Posterior and model evidence

For model $\mathcal{M}$ with parameters $\Theta$:

\[p(\Theta\mid G,\mathcal M) \propto p(G\mid\Theta,\mathcal M)p(\Theta\mid\mathcal M),\] \[p(G\mid\mathcal M)=\int p(G\mid\Theta,\mathcal M)p(\Theta\mid\mathcal M)d\Theta.\]

Model evidence supports Bayesian model selection, balancing fit and complexity.

7.2 Inferential community detection

Instead of maximizing $Q$, infer latent block assignments $z$ in SBM/DC-SBM:

\[p(z,\Theta\mid G) \propto p(G\mid z,\Theta)p(z)p(\Theta).\]

Advantages:

Missing link score is posterior predictive probability:

\[p(A_{ij}=1\mid G_{obs}) = \int p(A_{ij}=1\mid\Theta) p(\Theta\mid G_{obs}) d\Theta.\]

8. Roles, Positions, and Blockmodeling

Roles and positions are related but not identical.

Blockmodeling builds mesoscale interaction matrices.

8.1 Indirect blockmodeling

Fit a block pattern by optimizing agreement between observed ties and idealized block image matrix.

8.2 Direct blockmodeling

Optimize a criterion directly on observed adjacency under allowed block types (null, dense, regular, etc.).

8.3 Generalized blockmodeling

Unifies multiple equivalence notions and block constraints; useful when role definitions are domain-specific.


9. Information-Theoretic Community Detection

Flow-based methods (e.g., map equation perspective) define good communities as those giving short description length for trajectories.

Optimization target is code length $L(\mathcal P)$ for partition $\mathcal P$:

\[\mathcal P^* = \arg\min_{\mathcal P} L(\mathcal P).\]

This differs from modularity:

So they can produce different, complementary partitions.


10. Multi-Robot Systems Mapping

Network science tools map naturally to multi-robot systems.

10.1 Communication and resilience

Model communication graph $G_t$ as temporal network. Maintain algebraic connectivity $\lambda_2(L_t)$ above threshold for consensus and robustness.

10.2 Failure and misinformation propagation

Use SIS/SIR-type processes over robot interaction network to evaluate containment policies and critical nodes.

10.3 Role-aware coordination

Use blockmodeling/community structure to assign:

10.4 Partial observability

When each robot sees only local neighborhoods, infer global structure probabilistically and propagate uncertainty into planner/controller decisions.


11. Decision-Based Models

Spreading is not always passive. Agents may act strategically under cost/utility:

\[\max_{a_i\in\mathcal A_i} \; U_i(a_i,a_{-i},G) - \lambda_i\,\text{risk}_i(a_i,G).\]

Coupling strategic decisions with network dynamics leads to adaptive graphs and nonstationary diffusion.

This is a key bridge to control and RL in networked systems.


12. Quick Reference Equations

These notes are intended as working technical references, not final polished lecture notes.