Walk-summability in Gaussian Graphical Models

• Walk-summability is a spectral condition on the partial-correlation matrix ensuring the convergence of the Neumann series in Gaussian graphical models.
• It guarantees exponential decay of long-range dependencies, enabling efficient loopy belief propagation and optimal structure recovery through thresholding methods.
• The property underpins rigorous sample complexity and sparsistency guarantees, making it instrumental for reliable high-dimensional model design.

Walk-summability is a central concept in the theory of high-dimensional Gaussian graphical models, characterizing a class of Gaussian Markov random fields for which both efficient inference algorithms (such as loopy belief propagation) and statistically consistent structure learning methods (notably, thresholding-based estimators) are guaranteed to perform optimally. The walk-summability property is defined via spectral conditions on the normalized partial-correlation matrix associated with the Gaussian graphical model and underpins the exponential decay of correlations with graph distance. It plays an instrumental role in rigorously establishing sample complexity bounds and sparsistency guarantees for graph recovery procedures such as conditional covariance thresholding (Anandkumar et al., 2011).

1. Formal Definition of Walk-Summability

Let $G = (V, E)$ be an undirected graph with $p$ nodes, and let $\mathbf{X} = (X_1, ..., X_p)^\top$ be a zero-mean Gaussian Markov random field on $G$, with information-form density

$$f_{\mathbf{X}}(x) \propto \exp\left\{-\frac{1}{2}x^\top \mathbf{J} x\right\}$$

where the precision (potential) matrix $\mathbf{J}$ is positive definite and matches the sparsity structure of $G$ (i.e., $J_{ij}=0$ iff $(i,j)\notin E$), and $\mathbf{\Sigma} = \mathbf{J}^{-1}$ is the covariance matrix.

Normalize the diagonal via $p$0 for all $p$1, and set $p$2. Here, $p$3 for $p$4, $p$5, and the matrix $p$6 collects the entrywise absolute values, $p$7.

Definition (α-walk-summability):

The Gaussian model is called α-walk-summable if there exists $p$8 such that

$p$9

where $\mathbf{X} = (X_1, ..., X_p)^\top$0 denotes the spectral norm (i.e., largest absolute eigenvalue).

Equivalently, $\mathbf{X} = (X_1, ..., X_p)^\top$1 suffices. This ensures the Neumann series expansion converges: $\mathbf{X} = (X_1, ..., X_p)^\top$2 Each entry $\mathbf{X} = (X_1, ..., X_p)^\top$3 expands as a sum over all walks from $\mathbf{X} = (X_1, ..., X_p)^\top$4 to $\mathbf{X} = (X_1, ..., X_p)^\top$5, each contributing a product of $\mathbf{X} = (X_1, ..., X_p)^\top$6's along the walk.

2. Spectral Characterization and Walk-Sum Interpretation

The spectral radius characterization for walk-summability is

$\mathbf{X} = (X_1, ..., X_p)^\top$7

where $\mathbf{X} = (X_1, ..., X_p)^\top$8 denotes the spectral radius. This guarantees positive-definiteness even if all entries are replaced by their absolute values, and yields convergence of $\mathbf{X} = (X_1, ..., X_p)^\top$9.

Walk-summability admits a graph-theoretic interpretation: each covariance $G$0 can be expressed as a sum over all walks between $G$1 and $G$2, with each walk of length $G$3 weighted by a multiplicative product of the partial correlations along the path. This interprets the covariance as being built up from local walk contributions, and reveals that under walk-summability, long walks are suppressed exponentially in the walk length (Anandkumar et al., 2011).

3. Walk-Summability and Structure Recovery via Thresholding

Walk-summability is foundational in ensuring the success of simple structure estimation algorithms, particularly conditional covariance thresholding. For any non-edge $G$4, walk-summability ensures the existence of a small separator $G$5 so that the conditional covariance satisfies

$G$6

where $G$7 is a path-length parameter and $G$8. For true edges, $G$9 for $$f_{\mathbf{X}}(x) \propto \exp\left\{-\frac{1}{2}x^\top \mathbf{J} x\right\}$$0 not containing $$f_{\mathbf{X}}(x) \propto \exp\left\{-\frac{1}{2}x^\top \mathbf{J} x\right\}$$1 or $$f_{\mathbf{X}}(x) \propto \exp\left\{-\frac{1}{2}x^\top \mathbf{J} x\right\}$$2. A separating threshold $$f_{\mathbf{X}}(x) \propto \exp\left\{-\frac{1}{2}x^\top \mathbf{J} x\right\}$$3 such that

$$f_{\mathbf{X}}(x) \propto \exp\left\{-\frac{1}{2}x^\top \mathbf{J} x\right\}$$4

—combined with an empirical covariance estimator and sufficiently large sample size $$f_{\mathbf{X}}(x) \propto \exp\left\{-\frac{1}{2}x^\top \mathbf{J} x\right\}$$5—enables the recovery of the true graph structure with high probability. Walk-summability is crucial in establishing the exponential decay $$f_{\mathbf{X}}(x) \propto \exp\left\{-\frac{1}{2}x^\top \mathbf{J} x\right\}$$6 in spurious conditional covariances, thus separating edges from non-edges (Anandkumar et al., 2011).

4. Local Separation Criterion and Interplay with Walk-Summability

Practical application of thresholding requires that the underlying graph exhibits a sparse local-separation property: for every non-edge $$f_{\mathbf{X}}(x) \propto \exp\left\{-\frac{1}{2}x^\top \mathbf{J} x\right\}$$7, there must exist a minimal separator $$f_{\mathbf{X}}(x) \propto \exp\left\{-\frac{1}{2}x^\top \mathbf{J} x\right\}$$8 of size at most $$f_{\mathbf{X}}(x) \propto \exp\left\{-\frac{1}{2}x^\top \mathbf{J} x\right\}$$9 blocking all paths of length up to $\mathbf{J}$0 between $\mathbf{J}$1 and $\mathbf{J}$2.

Graph families such as bounded-degree, large-girth, Erdős–Rényi $\mathbf{J}$3, power-law, and small-world models generally satisfy $\mathbf{J}$4 and $\mathbf{J}$5 with high probability. The presence of such local separators, combined with walk-summability, ensures that all short contributing walks in the Neumann expansion are suppressed, and allows computationally feasible optimization over conditioning sets ($\mathbf{J}$6 complexity). The conditional covariance decays exponentially fast with $\mathbf{J}$7 when conditioned on these separators (Anandkumar et al., 2011).

5. Structural Consistency and Sample Complexity Bounds

Under the following conditions:

• $\mathbf{J}$8-walk-summability ($\mathbf{J}$9)
• Sparse local separators of size $G$0
• Path-length $G$1 chosen so that $G$2
• Edge non-degeneracy and appropriate thresholding

the Conditional Covariance Test (CCT) estimator recovers the true graph $G$3 asymptotically with probability tending to one, provided that the number of samples

$G$4

This rate is optimal up to constants, as any algorithm requires at least $G$5 samples, with $G$6 depending on $G$7.

Key lemmata show that:

• Covariances can be approximated to error $G$8 by truncating walks longer than $G$9.
• Conditional covariances for non-neighbor pairs decay as $J_{ij}=0$0 when conditioning on a local separator, whereas for edge pairs, a constant lower bound applies.

Uniform concentration of empirical conditional covariances is achieved when $J_{ij}=0$1 (Anandkumar et al., 2011).

6. Practical Guidance, Graph Classes, and Model Design

Walk-summability is a natural and directly verifiable condition, implementable as a spectral-radius bound on the absolute partial-correlation matrix. It implies exponential decay of long-range dependencies, making both inference and consistent structure learning feasible. Most random graph models of interest, such as sparse Erdős–Rényi, power-law, and small-world networks, satisfy the required local-separation property with high probability.

In practice, maintaining walk-summability for denser graphs (large maximum degree $J_{ij}=0$2) enforces a constraint on the maximum edge strength, with $J_{ij}=0$3 required for $J_{ij}=0$4. Increasing density necessitates downscaling edge weights, balancing learnability with statistical efficiency (Anandkumar et al., 2011).

7. Summary and Theoretical Importance

Walk-summability constitutes the key analytic device at the intersection of tractable inference (convergence of belief propagation) and sparsistent structure learning (successful recovery of graph structures by local threshold rules). It yields transparent verifiable conditions and establishes near-optimal sample complexity guarantees for high-dimensional Gaussian graphical model selection via local sparsity and decay-of-dependency structure. Its role is especially prominent in rigorous guarantees for thresholding-based algorithms and forms the basis for modern approaches to graphical model selection in high dimensions (Anandkumar et al., 2011).