Granular Subdomain Coverage in SBMs

• Granular subdomain coverage is the explicit capture and analysis of fine-grained structures in stochastic block models, focusing on small, dense clusters and degree-separated classes.
• It advances inference by delineating phase transitions in recovery thresholds and applying spectral, convex, and combinatorial methods for precise subdomain detection.
• Algorithmic strategies like the graph pencil method and bias correction ensure reliable granular inference in complex, high-dimensional network settings.

Granular subdomain coverage refers to the explicit capture, detection, and recovery of fine-grained structural, spectral, or labeling features within subpopulations or structural blocks of a stochastic block model (SBM), particularly those corresponding to small, dense communities or degree-separated classes in high-dimensional or dense random graphs. This concept is central in advancing inference guarantees, computational thresholds, and methodological strategies for heterogeneous, multi-scale, and degree-separated block models. Theoretical work in recent years has characterized the statistical and algorithmic regimes in which granular subdomain recovery is possible, impossible, or sharply transitionary.

1. Formal Setting and Motivations

Granular subdomain coverage arises mainly in the study of SBMs, which model networks with $n$ nodes divided into $K$ communities, each specified by size $n_k$ and within/between-community edge probabilities $p_k$, $q_{k\ell}$, or, more generally, a $K \times K$ symmetric matrix $P$. In dense regimes ($p_{ab} = \Theta(1)$ for all $a,b$), accurate detection and recovery of small or heterogeneously sized communities—which constitute granular subdomains—are essential for high-resolution graph partitioning and inference. The challenge is heightened in the heterogeneous SBM, where both the number of communities and their sizes may scale nontrivially with $n$, and the separation in connection probabilities between small clusters and the bulk is critical.

The coverage of such granular domains is not only of theoretical interest, e.g., in determining phase transitions and computability results, but also underpins algorithmic design: when can convex or spectral methods recover extremely small dense clusters? Are there settings where only information-theoretic (intractable) procedures can achieve this level of granularity?

2. Recovery Regimes and Thresholds for Dense Heterogeneous SBMs

Jalali, Han, Dumitriu, and Fazel provide a comprehensive analysis of exact recovery thresholds in heterogeneous and dense SBMs (Jalali et al., 2015). Two central parameters for granular subdomain recovery are the relative cluster density

$\rho_k = n_k (p_k - q)$

and the chi-square divergence

$(p_k, q) = \frac{(p_k - q)^2}{q(1 - q)},$

which together quantify the signal-to-noise ratio (SNR) for each block. Efficient and exact coverage is then governed by three classes of thresholds:

• Information-theoretic possibility: Recovery is impossible if the total signal $\sum_k n_k^2(p_k, q)$ falls below an entropy term $\sum_k n_k \log(n/n_k)$. For dense, heterogeneous models, this enforces $n_k \gtrsim \log n$ for possibility, regardless of the number of clusters.
• Convex/programmatic recovery: Semidefinite programming (SDP) relaxations succeed if $\rho_k^2 \gtrsim n_k \log n$ and only a polylogarithmic number of $n_k = \Theta(\sqrt{\log n})$ clusters are present. This is the granular subdomain regime where convex methods are powerful but depend critically on relative density.
• Simple-counting recovery: With very dense clusters (e.g., $p_k - q = \Theta(1)$), even basic thresholding on degrees and common neighbors can recover subdomains of size $n_k \gtrsim \sqrt{\log n}$. However, this is restricted to settings with a small number of such "granule" clusters.

The granularity of subdomain recovery here is seen directly in the capacity to recover blocks orders of magnitude smaller than $n$ (down to $\sqrt{\log n}$), provided inter- and intra-block contrasts are sufficiently high and the overall number of small clusters does not overwhelm noise accumulation (Jalali et al., 2015).

3. Degree-Separated Models and Graph Pencil Methods

Gunderson et al. introduce a constructive method for granular subdomain inference—mapping counts of small subgraph densities (stars and bistars) onto the full set of SBM model parameters in the "degree-separated" regime (Gunderson et al., 31 Jan 2024).

• Degree separation is defined as all normalized block degrees $d_k = \sum_{\ell=1}^K \pi_\ell p_{\ell k}$ being distinct, ensuring invertibility of the associated moment systems.
• The "graph pencil method" (editor’s term) proves that, under degree separation and knowledge of $K$, the $2K-1$ star densities (moments) and $K^2$ bistar densities uniquely determine the entire SBM parameters $(\boldsymbol{\pi}, P)$. This explicit mapping leverages generalized eigenvalue problems on Hankel matrices formed by moments, guaranteeing granular resolution at the community level, even for small, distinct-degree blocks.
• In dense SBMs, subgraph densities concentrate at a $O_p(n^{-1/2})$ rate, rendering empirical estimation of granular subdomain structure consistent, non-iterative, and computationally trivial once densities are available (Gunderson et al., 31 Jan 2024).

4. Spectral Methods and Granular Detection in Dense Regimes

Granular spectral coverage is facilitated by spectral algorithms exploiting structural eigenvalues of various matrix representations.

• In the dense SBM, the adjacency, Laplacian, or modularity matrices have exactly $K$ leading eigenvalues that are separated from the bulk, directly enabling granular cluster recovery via spectral partitioning followed by $k$-means (Lei et al., 2013, Bolla et al., 2023).
• Consistency of spectral clustering requires that the minimum gap $(p_{\text{in}} - p_{\text{out}})$ is bounded below, and cluster sizes satisfy $K^3 \ll n$ for vanishing misclassification. Critically, for fixed $K$, granular coverage of all blocks is guaranteed if $n_k \gtrsim \sqrt{n}$ (in balanced models), with minimal loss for unbalanced cases (Lei et al., 2013).
• The inflation-deflation procedure, particularly in the modularity-based spectral approach, "inflates" the space by removing the all-ones direction and "deflates" into $\mathbb{R}^k$ for clustering, which is effective even in the presence of pronounced degree heterogeneity, supporting granular label recovery for all degree-separated subdomains in the dense limit (Bolla et al., 2023).
• Recent spectral results further reveal detectability phase transitions based on eigenvalue scaling (non-backtracking matrix or modularity): only clusters corresponding to sufficiently large (structural) eigenvalues emerge as detectable, underpinning a spectral granularity threshold indexed by eigenvalue magnitude (Bolla et al., 2023).

5. Non-Reconstruction Regimes and Contiguity Barriers

Banerjee and prior works precisely characterize sharp granular limits on when no information-theoretic method can achieve even partial subdomain coverage.

• For the two-block dense SBM, if $(a_n - b_n)^2 < 2(1-p)(a_n + b_n)$, then the model is contiguous to an Erdős–Rényi graph of the same average degree. No error rate improves on random guessing for block assignment, and local indistinguishability holds for any small subset of nodes; coverage at any granular scale is impossible (non-reconstruction) (Banerjee, 2016).
• The critical contour $(a_n - b_n)^2 \sim 2(1-p)(a_n + b_n)$ marks the transition. Above it, granular coverage becomes possible (asymptotic singularity implies strong recoverability), while below, statistical indistinguishability prohibits even coarse-grained detection.

A plausible implication is that even advanced algorithmic or moment-matching methods cannot surpass these information-theoretic boundaries; below the threshold, no granular subdomain coverage is achievable.

6. Computational-Statistical Gaps in Multi-layer and High-complexity Settings

Granular subdomain coverage in multi-layer stochastic block models (MLSBMs) is subject to distinct computational and statistical phase boundaries.

• Statistically, granular detection requires $nL\rho \to \infty$ (with $L$ layers and edge-probability difference $\rho$), enabling oracle (MLE) recovery of communities. However, for polynomial-time methods, the necessary threshold is relaxed only to $n\sqrt{L}\rho \gtrsim \sqrt{\log n}$, implying that as the number of layers increases, algorithmic coverage of fine-grained substructure is fundamentally limited (Lei et al., 2023).
• Under the low-degree polynomial conjecture, this computational barrier is nearly tight: in high-complexity multi-layer SBMs, the statistical and algorithmic capabilities to resolve fine-scale (granular) subdomains can diverge sharply.

7. Practical Implications, Algorithmic Recommendations, and Bias Correction

Algorithmic strategies for granular subdomain coverage vary with model details:

• For ultra-dense and degree-separated blocks, the graph pencil method is preferable due to its minimal computational cost and parameter identifiability without iterative optimization (Gunderson et al., 31 Jan 2024).
• In regimes featuring observation biases (e.g., random-walk explored subgraphs), bias-debiasing based on the stationary law or algebraic corrections remains critical for accurate granular inference (Tran et al., 2020). In particular, de-biasing standard SBM estimators by inverting the stationary CDF substantially improves mean-square error in block (subdomain) proportions, even for very small sample sizes.
• For very small and dense blocks, simple combinatorial thresholding is sufficient for granular coverage, provided the total number of such subdomains does not scale unfavorably (Jalali et al., 2015).
• Spectral and convex optimization methods are effective for a broad spectrum of dense or balanced models, but require adaptation or augmentation (e.g., bias correction, whitening) in heterogeneous settings or when the number of subdomains grows with $n$ (Lei et al., 2013, Bolla et al., 2023).

---

Granular subdomain coverage thus encapsulates the interplay between model heterogeneity, algorithmic sophistication, and the fundamental phase diagram defined by information-theoretic and computational thresholds. Modern results establish explicit spectral, convex, and moment-based procedures that attain these boundaries, while also clarifying when such fine-grained structural inference becomes impossible due to statistical indistinguishability or computational hardness.