Decomposing Degree Assortativity in Sparse Spatial Networks

Technical Summary: Decomposing Degree Assortativity in Sparse Spatial Networks

Problem Statement
In spatial networks, positive degree assortativity (the tendency for well-connected nodes to link to other well-connected nodes) is often interpreted as evidence of "sorting"—a mechanism where popular nodes concentrate in specific locations and seek each other out. However, a rival geometric explanation exists: nodes that are spatially close share a pool of potential neighbors. Even in the absence of sorting, this "shared opportunity" can induce positive correlations between the degrees of connected endpoints.

The central problem addressed is the inability to distinguish between these two mechanisms (sorting vs. shared opportunity) using standard network data. Existing literature lacks a population-level definition of these contributions, a proof that observable graph summaries can identify them, and valid inference methods that remain robust at the boundary of independence (where no sorting exists). Ignoring the geometric channel leads to attributing geometric effects to preference, while ignoring sorting leads to underestimating preference.

Methodology
The authors propose a specific model class: finite-range random-connection graphs where a node's latent popularity ($W$) and spatial position ($X$) are coupled by a copula that preserves both marginals. The dependence strength is parameterized by $\psi$, the squared $L^2$ norm of the copula perturbation.

1. Two-Channel Decomposition:
   The paper derives the limiting assortativity coefficient ($r$) in the sparse regime. It proves that $r$ splits exactly into two additive channels:

   • Intensity-Covariance Channel ($r_{int}$): Arises from the covariance of endpoint connection intensities. This aggregates all sources of dependence, including inhomogeneity of positions, and can be non-zero even at independence ($\psi=0$).
   • Shared-Neighbor Channel ($r_{overlap}$): Arises from the overlap of influence regions (shared neighbors) and depends on the tangent displacement of the edge.
     Crucially, the authors define the sorting contribution ($\Delta_{sort}$) not as the raw intensity channel, but as a counterfactual increment: the change in assortativity when the dependence parameter $\psi$ is introduced, holding the mean degree and transitivity fixed. This definition ensures $\Delta_{sort} = 0$ at independence.

2. Global Identification:
   The paper addresses whether the parameters $(\lambda, \eta, \psi)$ (connection intensity, range, and dependence strength) can be uniquely recovered from the three observable graph summaries: mean degree ($\ell$), transitivity ($C$), and assortativity ($r$).

   • Using a computer-assisted proof with exact rational arithmetic and interval analysis, the authors prove that the map from parameters to observables is globally injective on a certified region $B$, including the boundary $\psi=0$.
   • This global injectivity is stronger than standard local rank conditions, ruling out spurious solution branches.

3. Verified Inference:
   The authors develop a Generalized Method of Moments (GMM) framework with verified regularity conditions.

   • They prove a Central Limit Theorem (CLT) for motif counts and establish the asymptotic normality of the estimator.
   • They derive the specific asymptotic distribution for the estimator at the independence boundary ($\psi=0$), showing it follows a mixture of chi-squared distributions ($\frac{1}{2}\chi^2_0 + \frac{1}{2}\chi^2_1$) and that the amplitude estimator converges at a rate of $n^{1/4}$.
   • All model-side conditions (bias bounds, covariance non-degeneracy, identification) are verified via the same exact-rational interval certificates used for identification.

4. Two-View Design:
   In applications where node attributes (like activity rank) and positions are observed, the authors employ a "two-view" protocol:

   • Direct View: Estimates dependence directly from the $(X, W)$ pairs.
   • Graph-Only View: Estimates dependence from graph moments alone.
     Agreement between these views serves as a specification test. If the model class fits, the estimates should align; if not, the discrepancy localizes the missing mechanisms.

Key Results

• Theoretical Decomposition: The paper provides the first exact decomposition of limiting assortativity into intensity-covariance and shared-neighbor channels, carried by the edge-Palm law. It establishes that sorting is a compensated increment, distinct from raw intensity effects.
• Identification Certificate: The map $(\lambda, \eta, \psi) \to (\ell, C, r)$ is proven to be a diffeomorphism on the certified region $B$. This allows for the unique recovery of the sorting contribution from a single graph.
• Application to Location-Based Networks:
  • Analyzing 11 metropolitan networks (Brightkite and Gowalla), the model class is rejected in every city. The observed assortativity is significantly lower than what the model predicts even with zero dependence ($\psi=0$).
  • The "sorting" estimates are driven to the boundary ($\psi=0$) not because independence is true, but because the model cannot fit the data.
  • Direct estimates from location-rank pairs confirm no significant dependence in 10 of 11 cities.
  • The observed patterns are consistent with degree heterogeneity (configuration nulls) and non-spatial triadic closure, mechanisms absent from the short-range geometric model.
• Application to Co-authorship Network:
  • In a national physics co-authorship network, the direct view detects a clear, positive dependence (productive authors concentrate in high-density research hubs).
  • However, the graph-only view is refused by the goodness-of-fit gate at every model level. The model fails to reproduce the joint structure of mean degree, clustering, and assortativity.
  • The residual points to a missing mechanism: the "team structure" of multi-author papers (cliques formed in single events), which generates triangles without the degree-mixing side effects of iterated closure.

Significance and Claims
The paper claims to provide a rigorous framework for attributing assortativity in spatial networks, moving beyond the assumption that positive assortativity implies sorting. Its significance lies in:

1. Defining Sorting Rigorously: By defining sorting as a compensated increment, it separates preference from geometric density effects.
2. Valid Inference at the Boundary: It provides the first verified inference theory for spatial network parameters that remains valid when the null hypothesis of independence is true, a common scenario in real data.
3. Diagnostic Power: The methodology acts as a "gate." It does not produce sorting estimates unless the model fits. In the location-based applications, this gate prevents the false attribution of geometric effects to sorting. In the co-authorship case, it correctly identifies that while dependence exists, the standard spatial model is insufficient to explain the network structure, pointing instead to team-based formation.
4. Computational Verification: The use of exact-rational interval arithmetic to certify global identification and inference conditions sets a new standard for rigor in network model estimation, ensuring that results are not artifacts of numerical instability or local optima.

The authors emphasize that their results are specific to the certified submodel (one-harmonic dependence, hard-range kernels) and that the rejection of the model in real data serves as a diagnostic for missing mechanisms (like community structure or team formation) rather than a definitive measurement of independence.