Erdős–Rényi Subgraph Pair Model

• The Erdős–Rényi Subgraph Pair Model is a formal framework that couples random graphs via subgraph extraction and vertex correspondence for network matching.
• It establishes sharp information-theoretic phase transitions for both exact and partial recovery, guiding practical and adversarial network de-anonymization.
• The model underpins studies in graph alignment and community detection using methods like brute-force MAP estimators and tail degree signature algorithms.

The Erdős–Rényi Subgraph Pair Model is a formal framework for studying random graphs coupled through subgraph extraction, vertex correspondence, and, crucially, for quantifying the information-theoretic limits of subgraph alignment and network matching. It forms the mathematical substrate for a wide array of statistical and computational analyses on network alignment, planted subgraph recovery, and correlated random graph pairs. This framework now plays a central role in the rigorous treatment of exact and partial graph alignment, particularly for assessing the feasibility and optimality of recovery in both practical and adversarial regimes (Shiu et al., 8 Jan 2026, Du, 17 Feb 2025, Bozorg et al., 2019).

1. Formal Model Definitions

1.1 Subgraph-Pair Model (Alignment/Recovery Setting)

Let $n\in\mathbb{N}$, $m=m(n)<n$, and $p=p(n)\in[0,1]$. A base random graph $G\sim ER(n,p)$ is sampled on vertex set $[n]=\{1,\ldots,n\}$. An $m$-subset $S\subset [n]$ is chosen uniformly; $H=G[S]$ is the induced subgraph, which is then anonymized by a uniformly random permutation $\pi:S\to[m]$ to produce $H_\pi$. The observer sees $(G,H_\pi)$ but neither $S$ nor $\pi$, and aims to recover $S$ (set recovery) and/or $\pi$ (permutation recovery) (Shiu et al., 8 Jan 2026).

1.2 Correlated Erdős–Rényi Subgraph Pair (Graph Matching Setting)

A "parent" graph $G_0\sim \mathcal{G}(n,p)$ is generated. Two edge-subsampled graphs $G_1,G_2$ are created by independently including each parent edge with probability $s$. One of the graphs is vertex-permuted by an unknown bijection $\pi^*\in S_n$. The analyst receives $(G_1,\pi^*(G_2))$ and aims to recover $\pi^*$ (Du, 17 Feb 2025, Bozorg et al., 2019).

1.3 Agglomerated Subgraph-Pair (Super-vertex Construction)

Given a partition of $[n]$ into $N$ disjoint nonempty subsets ("super-vertices"), a subgraph–pair model is defined on the super-vertex set: two super-vertices are connected iff at least one edge exists between their constituent nodes in the original $G(n,p)$ graph. This construction creates an effective inhomogeneous random graph on the super-vertex level, with edge probabilities depending on the subset sizes (Kang et al., 2013).

2. Information-Theoretic Recovery Thresholds

Sharp information-theoretic phase transitions delimit when exact or partial recovery is possible:

2.1 Exact Subgraph Set and Permutation Recovery

• Set Recovery: Achievable iff $\frac{1}{2} m h(p) - \ln n \to +\infty$, impossible (converse) if $\frac{1}{2} m h(p) - \ln(n/m) \to -\infty$, where $h(p) = -p\ln p - (1-p)\ln(1-p)$. Under mild conditions, the sharp threshold is $\frac{1}{2}m h(p)\asymp \ln n$ (Shiu et al., 8 Jan 2026).
• Permutation Recovery: Requires, in addition, $mp - \ln m \to +\infty$ (unique labeling). Fails if either the set recovery converse applies or $mp-\ln m\to -\infty$.

2.2 Partial Recovery in Correlated Graphs

For correlated pairs with $p = n^{-\alpha+o(1)}$, $\alpha\in(0,1]$, and $nps^2=\lambda=O(1)$, one cannot recover all vertices, but the fraction of recoverable correspondences is bounded tightly in terms of a limiting "balanced-load" distribution $\mu_\lambda$:

• The maximal fraction of accurately aligned vertices approaches $F_\lambda(1/\alpha)$, with $F_\lambda(x) = \mu_\lambda((x,\infty))$ (Du, 17 Feb 2025).

These thresholds delineate computational and information-theoretic feasibility in subgraph alignment and network de-anonymization.

3. Structural and Statistical Properties

3.1 Degree and Clustering Structure

• For two independent $G(n,p_1),G(n,p_2)$ on a common vertex set, their union is $G(n,p_\text{union})$ with $p_\text{union}=1-(1-p_1)(1-p_2)$. Degree distributions are binomial, clustering coefficient is $C(G)=p_\text{union}$ (Wen et al., 2012).
• Agglomerated super-vertex models produce inhomogeneous graphs, where connection probability between super-vertices of sizes $i,j$ is $1-(1-p)^{ij}$, enabling explicit degree and connectivity computations at the super-vertex level (Kang et al., 2013).

3.2 Emergence of Community and Heavy-Tailed Structures

When community sizes are heavy-tailed (e.g., $p_i \sim C i^{-\alpha}$), the induced super-vertex network has a scale-free (power-law) degree distribution, depending on the partition (Kang et al., 2013).

4. Methodologies and Algorithms

4.1 Brute-force (MAP) Estimator

For subgraph alignment, the optimal MAP estimator tests all $m$-subsets $S\subset[n]$ and bijections $\sigma:S\to[m]$, returning those for which relabeling $G[S]$ by $\sigma$ reproduces $H_\pi$. This is computationally intractable but achieves the information-theoretic threshold (Shiu et al., 8 Jan 2026).

4.2 Tail Degree Signature (TDS)

TDS is a polynomial-time, seedless matching algorithm exploiting the robustness of tail-degree statistics in correlated ER graphs. Feature vectors consist of sorted extremes of neighbor degree distributions across multiple neighborhood shells. Theoretical analysis shows it achieves the information-theoretic threshold $ps^2=\Omega(\log n/n)$ in regime $p=\Theta(\log n/n)$ (Bozorg et al., 2019).

Complexity

Algorithm	Time Complexity	Achieves IT Threshold
Brute-force MAP	Exponential ($m!{n\choose m}$)	Yes (exact recovery), not practical for large $n$
TDS–h (Hungarian)	$O(n^3)$	Yes (matching threshold for $G(n,p)$, sparse regime)
TDS–g (Greedy)	$O(n^2)$	Yes, with high probability under threshold conditions

5. Phase Transitions and Limit Theorems

5.1 Phase Diagrams in Alignment

Define $\alpha = [\frac{1}{2}mh(p)]/\ln n$. Set recovery is feasible for $\alpha>1+o(1)$, infeasible for $\alpha<1-o(1)$, with a grey zone in between. Sharp phase transitions demarcate algorithmic possibility from impossibility (Shiu et al., 8 Jan 2026).

5.2 Community Graph Phase Transitions

For agglomerated super-vertex graphs, thresholds for connectivity and giant component emergence follow from inhomogeneous random graph (IRG) theory (Kang et al., 2013). The key parameter is $c\tilde s_2$, the average squared community size times edge probability:

• Largest component vanishes if $c\tilde s_2\leq1$, occupies $\Theta(N)$ super-vertices if $c\tilde s_2>1$.

6. Connections to Broader Random Graph Models

The ER subgraph-pair model is a special case of subgraph generated models (SUGMs), where the only generated subgraphs are links ($k=1$ type), with SUGM reducing exactly to ER($n,p$). More general SUGMs encode dependency on motifs such as triangles, stars, and cliques, bridging ER structure and higher-order motif-based randomness (Chandrasekhar et al., 2016).

By tuning the types and rates of subgraph "atoms," the model generalizes ER, permitting tractable closed-form expressions for expectations, variances, and parameter inference.

7. Applications and Implications

The ER subgraph-pair model underpins rigorous analysis of biological network alignment, privacy and de-anonymization of social networks, and statistical models of network community structure. Its phase diagrams and thresholds provide foundational guarantees for algorithmic graph matching and motif-based inference. Recent advances demonstrate that truly seedless and polynomial-time algorithms can saturate the fundamental information-theoretic limits via robust local statistics, revealing new pathways for tractable recovery in high-noise regimes (Shiu et al., 8 Jan 2026, Du, 17 Feb 2025, Bozorg et al., 2019).

---

References:

(Shiu et al., 8 Jan 2026) Information-Theoretic Limits on Exact Subgraph Alignment Problem (Du, 17 Feb 2025) Optimal recovery of correlated Erdős-Rényi graphs (Bozorg et al., 2019) Seedless Graph Matching via Tail of Degree Distribution for Correlated Erdos-Renyi Graphs (Wen et al., 2012) Edge Union of Networks on the Same Vertex Set (Chandrasekhar et al., 2016) A Network Formation Model Based on Subgraphs (Kang et al., 2013) Evolution of a modified binomial random graph by agglomeration